xref: /openbmc/qemu/block/file-posix.c (revision 4b5d80f3)
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     int64_t *offset; /* offset of zone append operation */
164     int page_cache_inconsistent; /* errno from fdatasync failure */
165     bool has_fallocate;
166     bool needs_alignment;
167     bool force_alignment;
168     bool drop_cache;
169     bool check_cache_dropped;
170     struct {
171         uint64_t discard_nb_ok;
172         uint64_t discard_nb_failed;
173         uint64_t discard_bytes_ok;
174     } stats;
175 
176     PRManager *pr_mgr;
177 } BDRVRawState;
178 
179 typedef struct BDRVRawReopenState {
180     int open_flags;
181     bool drop_cache;
182     bool check_cache_dropped;
183 } BDRVRawReopenState;
184 
185 static int fd_open(BlockDriverState *bs)
186 {
187     BDRVRawState *s = bs->opaque;
188 
189     /* this is just to ensure s->fd is sane (its called by io ops) */
190     if (s->fd >= 0) {
191         return 0;
192     }
193     return -EIO;
194 }
195 
196 static int64_t raw_getlength(BlockDriverState *bs);
197 
198 typedef struct RawPosixAIOData {
199     BlockDriverState *bs;
200     int aio_type;
201     int aio_fildes;
202 
203     off_t aio_offset;
204     uint64_t aio_nbytes;
205 
206     union {
207         struct {
208             struct iovec *iov;
209             int niov;
210         } io;
211         struct {
212             uint64_t cmd;
213             void *buf;
214         } ioctl;
215         struct {
216             int aio_fd2;
217             off_t aio_offset2;
218         } copy_range;
219         struct {
220             PreallocMode prealloc;
221             Error **errp;
222         } truncate;
223         struct {
224             unsigned int *nr_zones;
225             BlockZoneDescriptor *zones;
226         } zone_report;
227         struct {
228             unsigned long op;
229         } zone_mgmt;
230     };
231 } RawPosixAIOData;
232 
233 #if defined(__FreeBSD__) || defined(__FreeBSD_kernel__)
234 static int cdrom_reopen(BlockDriverState *bs);
235 #endif
236 
237 /*
238  * Elide EAGAIN and EACCES details when failing to lock, as this
239  * indicates that the specified file region is already locked by
240  * another process, which is considered a common scenario.
241  */
242 #define raw_lock_error_setg_errno(errp, err, fmt, ...)                  \
243     do {                                                                \
244         if ((err) == EAGAIN || (err) == EACCES) {                       \
245             error_setg((errp), (fmt), ## __VA_ARGS__);                  \
246         } else {                                                        \
247             error_setg_errno((errp), (err), (fmt), ## __VA_ARGS__);     \
248         }                                                               \
249     } while (0)
250 
251 #if defined(__NetBSD__)
252 static int raw_normalize_devicepath(const char **filename, Error **errp)
253 {
254     static char namebuf[PATH_MAX];
255     const char *dp, *fname;
256     struct stat sb;
257 
258     fname = *filename;
259     dp = strrchr(fname, '/');
260     if (lstat(fname, &sb) < 0) {
261         error_setg_file_open(errp, errno, fname);
262         return -errno;
263     }
264 
265     if (!S_ISBLK(sb.st_mode)) {
266         return 0;
267     }
268 
269     if (dp == NULL) {
270         snprintf(namebuf, PATH_MAX, "r%s", fname);
271     } else {
272         snprintf(namebuf, PATH_MAX, "%.*s/r%s",
273             (int)(dp - fname), fname, dp + 1);
274     }
275     *filename = namebuf;
276     warn_report("%s is a block device, using %s", fname, *filename);
277 
278     return 0;
279 }
280 #else
281 static int raw_normalize_devicepath(const char **filename, Error **errp)
282 {
283     return 0;
284 }
285 #endif
286 
287 /*
288  * Get logical block size via ioctl. On success store it in @sector_size_p.
289  */
290 static int probe_logical_blocksize(int fd, unsigned int *sector_size_p)
291 {
292     unsigned int sector_size;
293     bool success = false;
294     int i;
295 
296     errno = ENOTSUP;
297     static const unsigned long ioctl_list[] = {
298 #ifdef BLKSSZGET
299         BLKSSZGET,
300 #endif
301 #ifdef DKIOCGETBLOCKSIZE
302         DKIOCGETBLOCKSIZE,
303 #endif
304 #ifdef DIOCGSECTORSIZE
305         DIOCGSECTORSIZE,
306 #endif
307     };
308 
309     /* Try a few ioctls to get the right size */
310     for (i = 0; i < (int)ARRAY_SIZE(ioctl_list); i++) {
311         if (ioctl(fd, ioctl_list[i], &sector_size) >= 0) {
312             *sector_size_p = sector_size;
313             success = true;
314         }
315     }
316 
317     return success ? 0 : -errno;
318 }
319 
320 /**
321  * Get physical block size of @fd.
322  * On success, store it in @blk_size and return 0.
323  * On failure, return -errno.
324  */
325 static int probe_physical_blocksize(int fd, unsigned int *blk_size)
326 {
327 #ifdef BLKPBSZGET
328     if (ioctl(fd, BLKPBSZGET, blk_size) < 0) {
329         return -errno;
330     }
331     return 0;
332 #else
333     return -ENOTSUP;
334 #endif
335 }
336 
337 /*
338  * Returns true if no alignment restrictions are necessary even for files
339  * opened with O_DIRECT.
340  *
341  * raw_probe_alignment() probes the required alignment and assume that 1 means
342  * the probing failed, so it falls back to a safe default of 4k. This can be
343  * avoided if we know that byte alignment is okay for the file.
344  */
345 static bool dio_byte_aligned(int fd)
346 {
347 #ifdef __linux__
348     struct statfs buf;
349     int ret;
350 
351     ret = fstatfs(fd, &buf);
352     if (ret == 0 && buf.f_type == NFS_SUPER_MAGIC) {
353         return true;
354     }
355 #endif
356     return false;
357 }
358 
359 static bool raw_needs_alignment(BlockDriverState *bs)
360 {
361     BDRVRawState *s = bs->opaque;
362 
363     if ((bs->open_flags & BDRV_O_NOCACHE) != 0 && !dio_byte_aligned(s->fd)) {
364         return true;
365     }
366 
367     return s->force_alignment;
368 }
369 
370 /* Check if read is allowed with given memory buffer and length.
371  *
372  * This function is used to check O_DIRECT memory buffer and request alignment.
373  */
374 static bool raw_is_io_aligned(int fd, void *buf, size_t len)
375 {
376     ssize_t ret = pread(fd, buf, len, 0);
377 
378     if (ret >= 0) {
379         return true;
380     }
381 
382 #ifdef __linux__
383     /* The Linux kernel returns EINVAL for misaligned O_DIRECT reads.  Ignore
384      * other errors (e.g. real I/O error), which could happen on a failed
385      * drive, since we only care about probing alignment.
386      */
387     if (errno != EINVAL) {
388         return true;
389     }
390 #endif
391 
392     return false;
393 }
394 
395 static void raw_probe_alignment(BlockDriverState *bs, int fd, Error **errp)
396 {
397     BDRVRawState *s = bs->opaque;
398     char *buf;
399     size_t max_align = MAX(MAX_BLOCKSIZE, qemu_real_host_page_size());
400     size_t alignments[] = {1, 512, 1024, 2048, 4096};
401 
402     /* For SCSI generic devices the alignment is not really used.
403        With buffered I/O, we don't have any restrictions. */
404     if (bdrv_is_sg(bs) || !s->needs_alignment) {
405         bs->bl.request_alignment = 1;
406         s->buf_align = 1;
407         return;
408     }
409 
410     bs->bl.request_alignment = 0;
411     s->buf_align = 0;
412     /* Let's try to use the logical blocksize for the alignment. */
413     if (probe_logical_blocksize(fd, &bs->bl.request_alignment) < 0) {
414         bs->bl.request_alignment = 0;
415     }
416 
417 #ifdef __linux__
418     /*
419      * The XFS ioctl definitions are shipped in extra packages that might
420      * not always be available. Since we just need the XFS_IOC_DIOINFO ioctl
421      * here, we simply use our own definition instead:
422      */
423     struct xfs_dioattr {
424         uint32_t d_mem;
425         uint32_t d_miniosz;
426         uint32_t d_maxiosz;
427     } da;
428     if (ioctl(fd, _IOR('X', 30, struct xfs_dioattr), &da) >= 0) {
429         bs->bl.request_alignment = da.d_miniosz;
430         /* The kernel returns wrong information for d_mem */
431         /* s->buf_align = da.d_mem; */
432     }
433 #endif
434 
435     /*
436      * If we could not get the sizes so far, we can only guess them. First try
437      * to detect request alignment, since it is more likely to succeed. Then
438      * try to detect buf_align, which cannot be detected in some cases (e.g.
439      * Gluster). If buf_align cannot be detected, we fallback to the value of
440      * request_alignment.
441      */
442 
443     if (!bs->bl.request_alignment) {
444         int i;
445         size_t align;
446         buf = qemu_memalign(max_align, max_align);
447         for (i = 0; i < ARRAY_SIZE(alignments); i++) {
448             align = alignments[i];
449             if (raw_is_io_aligned(fd, buf, align)) {
450                 /* Fallback to safe value. */
451                 bs->bl.request_alignment = (align != 1) ? align : max_align;
452                 break;
453             }
454         }
455         qemu_vfree(buf);
456     }
457 
458     if (!s->buf_align) {
459         int i;
460         size_t align;
461         buf = qemu_memalign(max_align, 2 * max_align);
462         for (i = 0; i < ARRAY_SIZE(alignments); i++) {
463             align = alignments[i];
464             if (raw_is_io_aligned(fd, buf + align, max_align)) {
465                 /* Fallback to request_alignment. */
466                 s->buf_align = (align != 1) ? align : bs->bl.request_alignment;
467                 break;
468             }
469         }
470         qemu_vfree(buf);
471     }
472 
473     if (!s->buf_align || !bs->bl.request_alignment) {
474         error_setg(errp, "Could not find working O_DIRECT alignment");
475         error_append_hint(errp, "Try cache.direct=off\n");
476     }
477 }
478 
479 static int check_hdev_writable(int fd)
480 {
481 #if defined(BLKROGET)
482     /* Linux block devices can be configured "read-only" using blockdev(8).
483      * This is independent of device node permissions and therefore open(2)
484      * with O_RDWR succeeds.  Actual writes fail with EPERM.
485      *
486      * bdrv_open() is supposed to fail if the disk is read-only.  Explicitly
487      * check for read-only block devices so that Linux block devices behave
488      * properly.
489      */
490     struct stat st;
491     int readonly = 0;
492 
493     if (fstat(fd, &st)) {
494         return -errno;
495     }
496 
497     if (!S_ISBLK(st.st_mode)) {
498         return 0;
499     }
500 
501     if (ioctl(fd, BLKROGET, &readonly) < 0) {
502         return -errno;
503     }
504 
505     if (readonly) {
506         return -EACCES;
507     }
508 #endif /* defined(BLKROGET) */
509     return 0;
510 }
511 
512 static void raw_parse_flags(int bdrv_flags, int *open_flags, bool has_writers)
513 {
514     bool read_write = false;
515     assert(open_flags != NULL);
516 
517     *open_flags |= O_BINARY;
518     *open_flags &= ~O_ACCMODE;
519 
520     if (bdrv_flags & BDRV_O_AUTO_RDONLY) {
521         read_write = has_writers;
522     } else if (bdrv_flags & BDRV_O_RDWR) {
523         read_write = true;
524     }
525 
526     if (read_write) {
527         *open_flags |= O_RDWR;
528     } else {
529         *open_flags |= O_RDONLY;
530     }
531 
532     /* Use O_DSYNC for write-through caching, no flags for write-back caching,
533      * and O_DIRECT for no caching. */
534     if ((bdrv_flags & BDRV_O_NOCACHE)) {
535         *open_flags |= O_DIRECT;
536     }
537 }
538 
539 static void raw_parse_filename(const char *filename, QDict *options,
540                                Error **errp)
541 {
542     bdrv_parse_filename_strip_prefix(filename, "file:", options);
543 }
544 
545 static QemuOptsList raw_runtime_opts = {
546     .name = "raw",
547     .head = QTAILQ_HEAD_INITIALIZER(raw_runtime_opts.head),
548     .desc = {
549         {
550             .name = "filename",
551             .type = QEMU_OPT_STRING,
552             .help = "File name of the image",
553         },
554         {
555             .name = "aio",
556             .type = QEMU_OPT_STRING,
557             .help = "host AIO implementation (threads, native, io_uring)",
558         },
559         {
560             .name = "aio-max-batch",
561             .type = QEMU_OPT_NUMBER,
562             .help = "AIO max batch size (0 = auto handled by AIO backend, default: 0)",
563         },
564         {
565             .name = "locking",
566             .type = QEMU_OPT_STRING,
567             .help = "file locking mode (on/off/auto, default: auto)",
568         },
569         {
570             .name = "pr-manager",
571             .type = QEMU_OPT_STRING,
572             .help = "id of persistent reservation manager object (default: none)",
573         },
574 #if defined(__linux__)
575         {
576             .name = "drop-cache",
577             .type = QEMU_OPT_BOOL,
578             .help = "invalidate page cache during live migration (default: on)",
579         },
580 #endif
581         {
582             .name = "x-check-cache-dropped",
583             .type = QEMU_OPT_BOOL,
584             .help = "check that page cache was dropped on live migration (default: off)"
585         },
586         { /* end of list */ }
587     },
588 };
589 
590 static const char *const mutable_opts[] = { "x-check-cache-dropped", NULL };
591 
592 static int raw_open_common(BlockDriverState *bs, QDict *options,
593                            int bdrv_flags, int open_flags,
594                            bool device, Error **errp)
595 {
596     BDRVRawState *s = bs->opaque;
597     QemuOpts *opts;
598     Error *local_err = NULL;
599     const char *filename = NULL;
600     const char *str;
601     BlockdevAioOptions aio, aio_default;
602     int fd, ret;
603     struct stat st;
604     OnOffAuto locking;
605 
606     opts = qemu_opts_create(&raw_runtime_opts, NULL, 0, &error_abort);
607     if (!qemu_opts_absorb_qdict(opts, options, errp)) {
608         ret = -EINVAL;
609         goto fail;
610     }
611 
612     filename = qemu_opt_get(opts, "filename");
613 
614     ret = raw_normalize_devicepath(&filename, errp);
615     if (ret != 0) {
616         goto fail;
617     }
618 
619     if (bdrv_flags & BDRV_O_NATIVE_AIO) {
620         aio_default = BLOCKDEV_AIO_OPTIONS_NATIVE;
621 #ifdef CONFIG_LINUX_IO_URING
622     } else if (bdrv_flags & BDRV_O_IO_URING) {
623         aio_default = BLOCKDEV_AIO_OPTIONS_IO_URING;
624 #endif
625     } else {
626         aio_default = BLOCKDEV_AIO_OPTIONS_THREADS;
627     }
628 
629     aio = qapi_enum_parse(&BlockdevAioOptions_lookup,
630                           qemu_opt_get(opts, "aio"),
631                           aio_default, &local_err);
632     if (local_err) {
633         error_propagate(errp, local_err);
634         ret = -EINVAL;
635         goto fail;
636     }
637 
638     s->use_linux_aio = (aio == BLOCKDEV_AIO_OPTIONS_NATIVE);
639 #ifdef CONFIG_LINUX_IO_URING
640     s->use_linux_io_uring = (aio == BLOCKDEV_AIO_OPTIONS_IO_URING);
641 #endif
642 
643     s->aio_max_batch = qemu_opt_get_number(opts, "aio-max-batch", 0);
644 
645     locking = qapi_enum_parse(&OnOffAuto_lookup,
646                               qemu_opt_get(opts, "locking"),
647                               ON_OFF_AUTO_AUTO, &local_err);
648     if (local_err) {
649         error_propagate(errp, local_err);
650         ret = -EINVAL;
651         goto fail;
652     }
653     switch (locking) {
654     case ON_OFF_AUTO_ON:
655         s->use_lock = true;
656         if (!qemu_has_ofd_lock()) {
657             warn_report("File lock requested but OFD locking syscall is "
658                         "unavailable, falling back to POSIX file locks");
659             error_printf("Due to the implementation, locks can be lost "
660                          "unexpectedly.\n");
661         }
662         break;
663     case ON_OFF_AUTO_OFF:
664         s->use_lock = false;
665         break;
666     case ON_OFF_AUTO_AUTO:
667         s->use_lock = qemu_has_ofd_lock();
668         break;
669     default:
670         abort();
671     }
672 
673     str = qemu_opt_get(opts, "pr-manager");
674     if (str) {
675         s->pr_mgr = pr_manager_lookup(str, &local_err);
676         if (local_err) {
677             error_propagate(errp, local_err);
678             ret = -EINVAL;
679             goto fail;
680         }
681     }
682 
683     s->drop_cache = qemu_opt_get_bool(opts, "drop-cache", true);
684     s->check_cache_dropped = qemu_opt_get_bool(opts, "x-check-cache-dropped",
685                                                false);
686 
687     s->open_flags = open_flags;
688     raw_parse_flags(bdrv_flags, &s->open_flags, false);
689 
690     s->fd = -1;
691     fd = qemu_open(filename, s->open_flags, errp);
692     ret = fd < 0 ? -errno : 0;
693 
694     if (ret < 0) {
695         if (ret == -EROFS) {
696             ret = -EACCES;
697         }
698         goto fail;
699     }
700     s->fd = fd;
701 
702     /* Check s->open_flags rather than bdrv_flags due to auto-read-only */
703     if (s->open_flags & O_RDWR) {
704         ret = check_hdev_writable(s->fd);
705         if (ret < 0) {
706             error_setg_errno(errp, -ret, "The device is not writable");
707             goto fail;
708         }
709     }
710 
711     s->perm = 0;
712     s->shared_perm = BLK_PERM_ALL;
713 
714 #ifdef CONFIG_LINUX_AIO
715      /* Currently Linux does AIO only for files opened with O_DIRECT */
716     if (s->use_linux_aio) {
717         if (!(s->open_flags & O_DIRECT)) {
718             error_setg(errp, "aio=native was specified, but it requires "
719                              "cache.direct=on, which was not specified.");
720             ret = -EINVAL;
721             goto fail;
722         }
723         if (!aio_setup_linux_aio(bdrv_get_aio_context(bs), errp)) {
724             error_prepend(errp, "Unable to use native AIO: ");
725             goto fail;
726         }
727     }
728 #else
729     if (s->use_linux_aio) {
730         error_setg(errp, "aio=native was specified, but is not supported "
731                          "in this build.");
732         ret = -EINVAL;
733         goto fail;
734     }
735 #endif /* !defined(CONFIG_LINUX_AIO) */
736 
737 #ifdef CONFIG_LINUX_IO_URING
738     if (s->use_linux_io_uring) {
739         if (!aio_setup_linux_io_uring(bdrv_get_aio_context(bs), errp)) {
740             error_prepend(errp, "Unable to use io_uring: ");
741             goto fail;
742         }
743     }
744 #else
745     if (s->use_linux_io_uring) {
746         error_setg(errp, "aio=io_uring was specified, but is not supported "
747                          "in this build.");
748         ret = -EINVAL;
749         goto fail;
750     }
751 #endif /* !defined(CONFIG_LINUX_IO_URING) */
752 
753     s->has_discard = true;
754     s->has_write_zeroes = true;
755 
756     if (fstat(s->fd, &st) < 0) {
757         ret = -errno;
758         error_setg_errno(errp, errno, "Could not stat file");
759         goto fail;
760     }
761 
762     if (!device) {
763         if (!S_ISREG(st.st_mode)) {
764             error_setg(errp, "'%s' driver requires '%s' to be a regular file",
765                        bs->drv->format_name, bs->filename);
766             ret = -EINVAL;
767             goto fail;
768         } else {
769             s->has_fallocate = true;
770         }
771     } else {
772         if (!(S_ISCHR(st.st_mode) || S_ISBLK(st.st_mode))) {
773             error_setg(errp, "'%s' driver requires '%s' to be either "
774                        "a character or block device",
775                        bs->drv->format_name, bs->filename);
776             ret = -EINVAL;
777             goto fail;
778         }
779     }
780 #ifdef CONFIG_BLKZONED
781     /*
782      * The kernel page cache does not reliably work for writes to SWR zones
783      * of zoned block device because it can not guarantee the order of writes.
784      */
785     if ((bs->bl.zoned != BLK_Z_NONE) &&
786         (!(s->open_flags & O_DIRECT))) {
787         error_setg(errp, "The driver supports zoned devices, and it requires "
788                          "cache.direct=on, which was not specified.");
789         return -EINVAL; /* No host kernel page cache */
790     }
791 #endif
792 
793     if (S_ISBLK(st.st_mode)) {
794 #ifdef __linux__
795         /* On Linux 3.10, BLKDISCARD leaves stale data in the page cache.  Do
796          * not rely on the contents of discarded blocks unless using O_DIRECT.
797          * Same for BLKZEROOUT.
798          */
799         if (!(bs->open_flags & BDRV_O_NOCACHE)) {
800             s->has_write_zeroes = false;
801         }
802 #endif
803     }
804 #ifdef __FreeBSD__
805     if (S_ISCHR(st.st_mode)) {
806         /*
807          * The file is a char device (disk), which on FreeBSD isn't behind
808          * a pager, so force all requests to be aligned. This is needed
809          * so QEMU makes sure all IO operations on the device are aligned
810          * to sector size, or else FreeBSD will reject them with EINVAL.
811          */
812         s->force_alignment = true;
813     }
814 #endif
815     s->needs_alignment = raw_needs_alignment(bs);
816 
817     bs->supported_zero_flags = BDRV_REQ_MAY_UNMAP | BDRV_REQ_NO_FALLBACK;
818     if (S_ISREG(st.st_mode)) {
819         /* When extending regular files, we get zeros from the OS */
820         bs->supported_truncate_flags = BDRV_REQ_ZERO_WRITE;
821     }
822     ret = 0;
823 fail:
824     if (ret < 0 && s->fd != -1) {
825         qemu_close(s->fd);
826     }
827     if (filename && (bdrv_flags & BDRV_O_TEMPORARY)) {
828         unlink(filename);
829     }
830     qemu_opts_del(opts);
831     return ret;
832 }
833 
834 static int raw_open(BlockDriverState *bs, QDict *options, int flags,
835                     Error **errp)
836 {
837     BDRVRawState *s = bs->opaque;
838 
839     s->type = FTYPE_FILE;
840     return raw_open_common(bs, options, flags, 0, false, errp);
841 }
842 
843 typedef enum {
844     RAW_PL_PREPARE,
845     RAW_PL_COMMIT,
846     RAW_PL_ABORT,
847 } RawPermLockOp;
848 
849 #define PERM_FOREACH(i) \
850     for ((i) = 0; (1ULL << (i)) <= BLK_PERM_ALL; i++)
851 
852 /* Lock bytes indicated by @perm_lock_bits and @shared_perm_lock_bits in the
853  * file; if @unlock == true, also unlock the unneeded bytes.
854  * @shared_perm_lock_bits is the mask of all permissions that are NOT shared.
855  */
856 static int raw_apply_lock_bytes(BDRVRawState *s, int fd,
857                                 uint64_t perm_lock_bits,
858                                 uint64_t shared_perm_lock_bits,
859                                 bool unlock, Error **errp)
860 {
861     int ret;
862     int i;
863     uint64_t locked_perm, locked_shared_perm;
864 
865     if (s) {
866         locked_perm = s->locked_perm;
867         locked_shared_perm = s->locked_shared_perm;
868     } else {
869         /*
870          * We don't have the previous bits, just lock/unlock for each of the
871          * requested bits.
872          */
873         if (unlock) {
874             locked_perm = BLK_PERM_ALL;
875             locked_shared_perm = BLK_PERM_ALL;
876         } else {
877             locked_perm = 0;
878             locked_shared_perm = 0;
879         }
880     }
881 
882     PERM_FOREACH(i) {
883         int off = RAW_LOCK_PERM_BASE + i;
884         uint64_t bit = (1ULL << i);
885         if ((perm_lock_bits & bit) && !(locked_perm & bit)) {
886             ret = qemu_lock_fd(fd, off, 1, false);
887             if (ret) {
888                 raw_lock_error_setg_errno(errp, -ret, "Failed to lock byte %d",
889                                           off);
890                 return ret;
891             } else if (s) {
892                 s->locked_perm |= bit;
893             }
894         } else if (unlock && (locked_perm & bit) && !(perm_lock_bits & bit)) {
895             ret = qemu_unlock_fd(fd, off, 1);
896             if (ret) {
897                 error_setg_errno(errp, -ret, "Failed to unlock byte %d", off);
898                 return ret;
899             } else if (s) {
900                 s->locked_perm &= ~bit;
901             }
902         }
903     }
904     PERM_FOREACH(i) {
905         int off = RAW_LOCK_SHARED_BASE + i;
906         uint64_t bit = (1ULL << i);
907         if ((shared_perm_lock_bits & bit) && !(locked_shared_perm & bit)) {
908             ret = qemu_lock_fd(fd, off, 1, false);
909             if (ret) {
910                 raw_lock_error_setg_errno(errp, -ret, "Failed to lock byte %d",
911                                           off);
912                 return ret;
913             } else if (s) {
914                 s->locked_shared_perm |= bit;
915             }
916         } else if (unlock && (locked_shared_perm & bit) &&
917                    !(shared_perm_lock_bits & bit)) {
918             ret = qemu_unlock_fd(fd, off, 1);
919             if (ret) {
920                 error_setg_errno(errp, -ret, "Failed to unlock byte %d", off);
921                 return ret;
922             } else if (s) {
923                 s->locked_shared_perm &= ~bit;
924             }
925         }
926     }
927     return 0;
928 }
929 
930 /* Check "unshared" bytes implied by @perm and ~@shared_perm in the file. */
931 static int raw_check_lock_bytes(int fd, uint64_t perm, uint64_t shared_perm,
932                                 Error **errp)
933 {
934     int ret;
935     int i;
936 
937     PERM_FOREACH(i) {
938         int off = RAW_LOCK_SHARED_BASE + i;
939         uint64_t p = 1ULL << i;
940         if (perm & p) {
941             ret = qemu_lock_fd_test(fd, off, 1, true);
942             if (ret) {
943                 char *perm_name = bdrv_perm_names(p);
944 
945                 raw_lock_error_setg_errno(errp, -ret,
946                                           "Failed to get \"%s\" lock",
947                                           perm_name);
948                 g_free(perm_name);
949                 return ret;
950             }
951         }
952     }
953     PERM_FOREACH(i) {
954         int off = RAW_LOCK_PERM_BASE + i;
955         uint64_t p = 1ULL << i;
956         if (!(shared_perm & p)) {
957             ret = qemu_lock_fd_test(fd, off, 1, true);
958             if (ret) {
959                 char *perm_name = bdrv_perm_names(p);
960 
961                 raw_lock_error_setg_errno(errp, -ret,
962                                           "Failed to get shared \"%s\" lock",
963                                           perm_name);
964                 g_free(perm_name);
965                 return ret;
966             }
967         }
968     }
969     return 0;
970 }
971 
972 static int raw_handle_perm_lock(BlockDriverState *bs,
973                                 RawPermLockOp op,
974                                 uint64_t new_perm, uint64_t new_shared,
975                                 Error **errp)
976 {
977     BDRVRawState *s = bs->opaque;
978     int ret = 0;
979     Error *local_err = NULL;
980 
981     if (!s->use_lock) {
982         return 0;
983     }
984 
985     if (bdrv_get_flags(bs) & BDRV_O_INACTIVE) {
986         return 0;
987     }
988 
989     switch (op) {
990     case RAW_PL_PREPARE:
991         if ((s->perm | new_perm) == s->perm &&
992             (s->shared_perm & new_shared) == s->shared_perm)
993         {
994             /*
995              * We are going to unlock bytes, it should not fail. If it fail due
996              * to some fs-dependent permission-unrelated reasons (which occurs
997              * sometimes on NFS and leads to abort in bdrv_replace_child) we
998              * can't prevent such errors by any check here. And we ignore them
999              * anyway in ABORT and COMMIT.
1000              */
1001             return 0;
1002         }
1003         ret = raw_apply_lock_bytes(s, s->fd, s->perm | new_perm,
1004                                    ~s->shared_perm | ~new_shared,
1005                                    false, errp);
1006         if (!ret) {
1007             ret = raw_check_lock_bytes(s->fd, new_perm, new_shared, errp);
1008             if (!ret) {
1009                 return 0;
1010             }
1011             error_append_hint(errp,
1012                               "Is another process using the image [%s]?\n",
1013                               bs->filename);
1014         }
1015         /* fall through to unlock bytes. */
1016     case RAW_PL_ABORT:
1017         raw_apply_lock_bytes(s, s->fd, s->perm, ~s->shared_perm,
1018                              true, &local_err);
1019         if (local_err) {
1020             /* Theoretically the above call only unlocks bytes and it cannot
1021              * fail. Something weird happened, report it.
1022              */
1023             warn_report_err(local_err);
1024         }
1025         break;
1026     case RAW_PL_COMMIT:
1027         raw_apply_lock_bytes(s, s->fd, new_perm, ~new_shared,
1028                              true, &local_err);
1029         if (local_err) {
1030             /* Theoretically the above call only unlocks bytes and it cannot
1031              * fail. Something weird happened, report it.
1032              */
1033             warn_report_err(local_err);
1034         }
1035         break;
1036     }
1037     return ret;
1038 }
1039 
1040 /* Sets a specific flag */
1041 static int fcntl_setfl(int fd, int flag)
1042 {
1043     int flags;
1044 
1045     flags = fcntl(fd, F_GETFL);
1046     if (flags == -1) {
1047         return -errno;
1048     }
1049     if (fcntl(fd, F_SETFL, flags | flag) == -1) {
1050         return -errno;
1051     }
1052     return 0;
1053 }
1054 
1055 static int raw_reconfigure_getfd(BlockDriverState *bs, int flags,
1056                                  int *open_flags, uint64_t perm, bool force_dup,
1057                                  Error **errp)
1058 {
1059     BDRVRawState *s = bs->opaque;
1060     int fd = -1;
1061     int ret;
1062     bool has_writers = perm &
1063         (BLK_PERM_WRITE | BLK_PERM_WRITE_UNCHANGED | BLK_PERM_RESIZE);
1064     int fcntl_flags = O_APPEND | O_NONBLOCK;
1065 #ifdef O_NOATIME
1066     fcntl_flags |= O_NOATIME;
1067 #endif
1068 
1069     *open_flags = 0;
1070     if (s->type == FTYPE_CD) {
1071         *open_flags |= O_NONBLOCK;
1072     }
1073 
1074     raw_parse_flags(flags, open_flags, has_writers);
1075 
1076 #ifdef O_ASYNC
1077     /* Not all operating systems have O_ASYNC, and those that don't
1078      * will not let us track the state into rs->open_flags (typically
1079      * you achieve the same effect with an ioctl, for example I_SETSIG
1080      * on Solaris). But we do not use O_ASYNC, so that's fine.
1081      */
1082     assert((s->open_flags & O_ASYNC) == 0);
1083 #endif
1084 
1085     if (!force_dup && *open_flags == s->open_flags) {
1086         /* We're lucky, the existing fd is fine */
1087         return s->fd;
1088     }
1089 
1090     if ((*open_flags & ~fcntl_flags) == (s->open_flags & ~fcntl_flags)) {
1091         /* dup the original fd */
1092         fd = qemu_dup(s->fd);
1093         if (fd >= 0) {
1094             ret = fcntl_setfl(fd, *open_flags);
1095             if (ret) {
1096                 qemu_close(fd);
1097                 fd = -1;
1098             }
1099         }
1100     }
1101 
1102     /* If we cannot use fcntl, or fcntl failed, fall back to qemu_open() */
1103     if (fd == -1) {
1104         const char *normalized_filename = bs->filename;
1105         ret = raw_normalize_devicepath(&normalized_filename, errp);
1106         if (ret >= 0) {
1107             fd = qemu_open(normalized_filename, *open_flags, errp);
1108             if (fd == -1) {
1109                 return -1;
1110             }
1111         }
1112     }
1113 
1114     if (fd != -1 && (*open_flags & O_RDWR)) {
1115         ret = check_hdev_writable(fd);
1116         if (ret < 0) {
1117             qemu_close(fd);
1118             error_setg_errno(errp, -ret, "The device is not writable");
1119             return -1;
1120         }
1121     }
1122 
1123     return fd;
1124 }
1125 
1126 static int raw_reopen_prepare(BDRVReopenState *state,
1127                               BlockReopenQueue *queue, Error **errp)
1128 {
1129     BDRVRawState *s;
1130     BDRVRawReopenState *rs;
1131     QemuOpts *opts;
1132     int ret;
1133 
1134     assert(state != NULL);
1135     assert(state->bs != NULL);
1136 
1137     s = state->bs->opaque;
1138 
1139     state->opaque = g_new0(BDRVRawReopenState, 1);
1140     rs = state->opaque;
1141 
1142     /* Handle options changes */
1143     opts = qemu_opts_create(&raw_runtime_opts, NULL, 0, &error_abort);
1144     if (!qemu_opts_absorb_qdict(opts, state->options, errp)) {
1145         ret = -EINVAL;
1146         goto out;
1147     }
1148 
1149     rs->drop_cache = qemu_opt_get_bool_del(opts, "drop-cache", true);
1150     rs->check_cache_dropped =
1151         qemu_opt_get_bool_del(opts, "x-check-cache-dropped", false);
1152 
1153     /* This driver's reopen function doesn't currently allow changing
1154      * other options, so let's put them back in the original QDict and
1155      * bdrv_reopen_prepare() will detect changes and complain. */
1156     qemu_opts_to_qdict(opts, state->options);
1157 
1158     /*
1159      * As part of reopen prepare we also want to create new fd by
1160      * raw_reconfigure_getfd(). But it wants updated "perm", when in
1161      * bdrv_reopen_multiple() .bdrv_reopen_prepare() callback called prior to
1162      * permission update. Happily, permission update is always a part (a seprate
1163      * stage) of bdrv_reopen_multiple() so we can rely on this fact and
1164      * reconfigure fd in raw_check_perm().
1165      */
1166 
1167     s->reopen_state = state;
1168     ret = 0;
1169 
1170 out:
1171     qemu_opts_del(opts);
1172     return ret;
1173 }
1174 
1175 static void raw_reopen_commit(BDRVReopenState *state)
1176 {
1177     BDRVRawReopenState *rs = state->opaque;
1178     BDRVRawState *s = state->bs->opaque;
1179 
1180     s->drop_cache = rs->drop_cache;
1181     s->check_cache_dropped = rs->check_cache_dropped;
1182     s->open_flags = rs->open_flags;
1183     g_free(state->opaque);
1184     state->opaque = NULL;
1185 
1186     assert(s->reopen_state == state);
1187     s->reopen_state = NULL;
1188 }
1189 
1190 
1191 static void raw_reopen_abort(BDRVReopenState *state)
1192 {
1193     BDRVRawReopenState *rs = state->opaque;
1194     BDRVRawState *s = state->bs->opaque;
1195 
1196      /* nothing to do if NULL, we didn't get far enough */
1197     if (rs == NULL) {
1198         return;
1199     }
1200 
1201     g_free(state->opaque);
1202     state->opaque = NULL;
1203 
1204     assert(s->reopen_state == state);
1205     s->reopen_state = NULL;
1206 }
1207 
1208 static int hdev_get_max_hw_transfer(int fd, struct stat *st)
1209 {
1210 #ifdef BLKSECTGET
1211     if (S_ISBLK(st->st_mode)) {
1212         unsigned short max_sectors = 0;
1213         if (ioctl(fd, BLKSECTGET, &max_sectors) == 0) {
1214             return max_sectors * 512;
1215         }
1216     } else {
1217         int max_bytes = 0;
1218         if (ioctl(fd, BLKSECTGET, &max_bytes) == 0) {
1219             return max_bytes;
1220         }
1221     }
1222     return -errno;
1223 #else
1224     return -ENOSYS;
1225 #endif
1226 }
1227 
1228 /*
1229  * Get a sysfs attribute value as character string.
1230  */
1231 #ifdef CONFIG_LINUX
1232 static int get_sysfs_str_val(struct stat *st, const char *attribute,
1233                              char **val) {
1234     g_autofree char *sysfspath = NULL;
1235     size_t len;
1236 
1237     if (!S_ISBLK(st->st_mode)) {
1238         return -ENOTSUP;
1239     }
1240 
1241     sysfspath = g_strdup_printf("/sys/dev/block/%u:%u/queue/%s",
1242                                 major(st->st_rdev), minor(st->st_rdev),
1243                                 attribute);
1244     if (!g_file_get_contents(sysfspath, val, &len, NULL)) {
1245         return -ENOENT;
1246     }
1247 
1248     /* The file is ended with '\n' */
1249     char *p;
1250     p = *val;
1251     if (*(p + len - 1) == '\n') {
1252         *(p + len - 1) = '\0';
1253     }
1254     return 0;
1255 }
1256 #endif
1257 
1258 #if defined(CONFIG_BLKZONED)
1259 static int get_sysfs_zoned_model(struct stat *st, BlockZoneModel *zoned)
1260 {
1261     g_autofree char *val = NULL;
1262     int ret;
1263 
1264     ret = get_sysfs_str_val(st, "zoned", &val);
1265     if (ret < 0) {
1266         return ret;
1267     }
1268 
1269     if (strcmp(val, "host-managed") == 0) {
1270         *zoned = BLK_Z_HM;
1271     } else if (strcmp(val, "host-aware") == 0) {
1272         *zoned = BLK_Z_HA;
1273     } else if (strcmp(val, "none") == 0) {
1274         *zoned = BLK_Z_NONE;
1275     } else {
1276         return -ENOTSUP;
1277     }
1278     return 0;
1279 }
1280 #endif /* defined(CONFIG_BLKZONED) */
1281 
1282 /*
1283  * Get a sysfs attribute value as a long integer.
1284  */
1285 #ifdef CONFIG_LINUX
1286 static long get_sysfs_long_val(struct stat *st, const char *attribute)
1287 {
1288     g_autofree char *str = NULL;
1289     const char *end;
1290     long val;
1291     int ret;
1292 
1293     ret = get_sysfs_str_val(st, attribute, &str);
1294     if (ret < 0) {
1295         return ret;
1296     }
1297 
1298     /* The file is ended with '\n', pass 'end' to accept that. */
1299     ret = qemu_strtol(str, &end, 10, &val);
1300     if (ret == 0 && end && *end == '\0') {
1301         ret = val;
1302     }
1303     return ret;
1304 }
1305 #endif
1306 
1307 static int hdev_get_max_segments(int fd, struct stat *st)
1308 {
1309 #ifdef CONFIG_LINUX
1310     int ret;
1311 
1312     if (S_ISCHR(st->st_mode)) {
1313         if (ioctl(fd, SG_GET_SG_TABLESIZE, &ret) == 0) {
1314             return ret;
1315         }
1316         return -ENOTSUP;
1317     }
1318     return get_sysfs_long_val(st, "max_segments");
1319 #else
1320     return -ENOTSUP;
1321 #endif
1322 }
1323 
1324 #if defined(CONFIG_BLKZONED)
1325 /*
1326  * If the reset_all flag is true, then the wps of zone whose state is
1327  * not readonly or offline should be all reset to the start sector.
1328  * Else, take the real wp of the device.
1329  */
1330 static int get_zones_wp(BlockDriverState *bs, int fd, int64_t offset,
1331                         unsigned int nrz, bool reset_all)
1332 {
1333     struct blk_zone *blkz;
1334     size_t rep_size;
1335     uint64_t sector = offset >> BDRV_SECTOR_BITS;
1336     BlockZoneWps *wps = bs->wps;
1337     unsigned int j = offset / bs->bl.zone_size;
1338     unsigned int n = 0, i = 0;
1339     int ret;
1340     rep_size = sizeof(struct blk_zone_report) + nrz * sizeof(struct blk_zone);
1341     g_autofree struct blk_zone_report *rep = NULL;
1342 
1343     rep = g_malloc(rep_size);
1344     blkz = (struct blk_zone *)(rep + 1);
1345     while (n < nrz) {
1346         memset(rep, 0, rep_size);
1347         rep->sector = sector;
1348         rep->nr_zones = nrz - n;
1349 
1350         do {
1351             ret = ioctl(fd, BLKREPORTZONE, rep);
1352         } while (ret != 0 && errno == EINTR);
1353         if (ret != 0) {
1354             error_report("%d: ioctl BLKREPORTZONE at %" PRId64 " failed %d",
1355                     fd, offset, errno);
1356             return -errno;
1357         }
1358 
1359         if (!rep->nr_zones) {
1360             break;
1361         }
1362 
1363         for (i = 0; i < rep->nr_zones; ++i, ++n, ++j) {
1364             /*
1365              * The wp tracking cares only about sequential writes required and
1366              * sequential write preferred zones so that the wp can advance to
1367              * the right location.
1368              * Use the most significant bit of the wp location to indicate the
1369              * zone type: 0 for SWR/SWP zones and 1 for conventional zones.
1370              */
1371             if (blkz[i].type == BLK_ZONE_TYPE_CONVENTIONAL) {
1372                 wps->wp[j] |= 1ULL << 63;
1373             } else {
1374                 switch(blkz[i].cond) {
1375                 case BLK_ZONE_COND_FULL:
1376                 case BLK_ZONE_COND_READONLY:
1377                     /* Zone not writable */
1378                     wps->wp[j] = (blkz[i].start + blkz[i].len) << BDRV_SECTOR_BITS;
1379                     break;
1380                 case BLK_ZONE_COND_OFFLINE:
1381                     /* Zone not writable nor readable */
1382                     wps->wp[j] = (blkz[i].start) << BDRV_SECTOR_BITS;
1383                     break;
1384                 default:
1385                     if (reset_all) {
1386                         wps->wp[j] = blkz[i].start << BDRV_SECTOR_BITS;
1387                     } else {
1388                         wps->wp[j] = blkz[i].wp << BDRV_SECTOR_BITS;
1389                     }
1390                     break;
1391                 }
1392             }
1393         }
1394         sector = blkz[i - 1].start + blkz[i - 1].len;
1395     }
1396 
1397     return 0;
1398 }
1399 
1400 static void update_zones_wp(BlockDriverState *bs, int fd, int64_t offset,
1401                             unsigned int nrz)
1402 {
1403     if (get_zones_wp(bs, fd, offset, nrz, 0) < 0) {
1404         error_report("update zone wp failed");
1405     }
1406 }
1407 
1408 static void raw_refresh_zoned_limits(BlockDriverState *bs, struct stat *st,
1409                                      Error **errp)
1410 {
1411     BDRVRawState *s = bs->opaque;
1412     BlockZoneModel zoned;
1413     int ret;
1414 
1415     ret = get_sysfs_zoned_model(st, &zoned);
1416     if (ret < 0 || zoned == BLK_Z_NONE) {
1417         goto no_zoned;
1418     }
1419     bs->bl.zoned = zoned;
1420 
1421     ret = get_sysfs_long_val(st, "max_open_zones");
1422     if (ret >= 0) {
1423         bs->bl.max_open_zones = ret;
1424     }
1425 
1426     ret = get_sysfs_long_val(st, "max_active_zones");
1427     if (ret >= 0) {
1428         bs->bl.max_active_zones = ret;
1429     }
1430 
1431     /*
1432      * The zoned device must at least have zone size and nr_zones fields.
1433      */
1434     ret = get_sysfs_long_val(st, "chunk_sectors");
1435     if (ret < 0) {
1436         error_setg_errno(errp, -ret, "Unable to read chunk_sectors "
1437                                      "sysfs attribute");
1438         goto no_zoned;
1439     } else if (!ret) {
1440         error_setg(errp, "Read 0 from chunk_sectors sysfs attribute");
1441         goto no_zoned;
1442     }
1443     bs->bl.zone_size = ret << BDRV_SECTOR_BITS;
1444 
1445     ret = get_sysfs_long_val(st, "nr_zones");
1446     if (ret < 0) {
1447         error_setg_errno(errp, -ret, "Unable to read nr_zones "
1448                                      "sysfs attribute");
1449         goto no_zoned;
1450     } else if (!ret) {
1451         error_setg(errp, "Read 0 from nr_zones sysfs attribute");
1452         goto no_zoned;
1453     }
1454     bs->bl.nr_zones = ret;
1455 
1456     ret = get_sysfs_long_val(st, "zone_append_max_bytes");
1457     if (ret > 0) {
1458         bs->bl.max_append_sectors = ret >> BDRV_SECTOR_BITS;
1459     }
1460 
1461     ret = get_sysfs_long_val(st, "physical_block_size");
1462     if (ret >= 0) {
1463         bs->bl.write_granularity = ret;
1464     }
1465 
1466     /* The refresh_limits() function can be called multiple times. */
1467     g_free(bs->wps);
1468     bs->wps = g_malloc(sizeof(BlockZoneWps) +
1469             sizeof(int64_t) * bs->bl.nr_zones);
1470     ret = get_zones_wp(bs, s->fd, 0, bs->bl.nr_zones, 0);
1471     if (ret < 0) {
1472         error_setg_errno(errp, -ret, "report wps failed");
1473         goto no_zoned;
1474     }
1475     qemu_co_mutex_init(&bs->wps->colock);
1476     return;
1477 
1478 no_zoned:
1479     bs->bl.zoned = BLK_Z_NONE;
1480     g_free(bs->wps);
1481     bs->wps = NULL;
1482 }
1483 #else /* !defined(CONFIG_BLKZONED) */
1484 static void raw_refresh_zoned_limits(BlockDriverState *bs, struct stat *st,
1485                                      Error **errp)
1486 {
1487     bs->bl.zoned = BLK_Z_NONE;
1488 }
1489 #endif /* !defined(CONFIG_BLKZONED) */
1490 
1491 static void raw_refresh_limits(BlockDriverState *bs, Error **errp)
1492 {
1493     BDRVRawState *s = bs->opaque;
1494     struct stat st;
1495 
1496     s->needs_alignment = raw_needs_alignment(bs);
1497     raw_probe_alignment(bs, s->fd, errp);
1498 
1499     bs->bl.min_mem_alignment = s->buf_align;
1500     bs->bl.opt_mem_alignment = MAX(s->buf_align, qemu_real_host_page_size());
1501 
1502     /*
1503      * Maximum transfers are best effort, so it is okay to ignore any
1504      * errors.  That said, based on the man page errors in fstat would be
1505      * very much unexpected; the only possible case seems to be ENOMEM.
1506      */
1507     if (fstat(s->fd, &st)) {
1508         return;
1509     }
1510 
1511 #if defined(__APPLE__) && (__MACH__)
1512     struct statfs buf;
1513 
1514     if (!fstatfs(s->fd, &buf)) {
1515         bs->bl.opt_transfer = buf.f_iosize;
1516         bs->bl.pdiscard_alignment = buf.f_bsize;
1517     }
1518 #endif
1519 
1520     if (bdrv_is_sg(bs) || S_ISBLK(st.st_mode)) {
1521         int ret = hdev_get_max_hw_transfer(s->fd, &st);
1522 
1523         if (ret > 0 && ret <= BDRV_REQUEST_MAX_BYTES) {
1524             bs->bl.max_hw_transfer = ret;
1525         }
1526 
1527         ret = hdev_get_max_segments(s->fd, &st);
1528         if (ret > 0) {
1529             bs->bl.max_hw_iov = ret;
1530         }
1531     }
1532 
1533     raw_refresh_zoned_limits(bs, &st, errp);
1534 }
1535 
1536 static int check_for_dasd(int fd)
1537 {
1538 #ifdef BIODASDINFO2
1539     struct dasd_information2_t info = {0};
1540 
1541     return ioctl(fd, BIODASDINFO2, &info);
1542 #else
1543     return -1;
1544 #endif
1545 }
1546 
1547 /**
1548  * Try to get @bs's logical and physical block size.
1549  * On success, store them in @bsz and return zero.
1550  * On failure, return negative errno.
1551  */
1552 static int hdev_probe_blocksizes(BlockDriverState *bs, BlockSizes *bsz)
1553 {
1554     BDRVRawState *s = bs->opaque;
1555     int ret;
1556 
1557     /* If DASD or zoned devices, get blocksizes */
1558     if (check_for_dasd(s->fd) < 0) {
1559         /* zoned devices are not DASD */
1560         if (bs->bl.zoned == BLK_Z_NONE) {
1561             return -ENOTSUP;
1562         }
1563     }
1564     ret = probe_logical_blocksize(s->fd, &bsz->log);
1565     if (ret < 0) {
1566         return ret;
1567     }
1568     return probe_physical_blocksize(s->fd, &bsz->phys);
1569 }
1570 
1571 /**
1572  * Try to get @bs's geometry: cyls, heads, sectors.
1573  * On success, store them in @geo and return 0.
1574  * On failure return -errno.
1575  * (Allows block driver to assign default geometry values that guest sees)
1576  */
1577 #ifdef __linux__
1578 static int hdev_probe_geometry(BlockDriverState *bs, HDGeometry *geo)
1579 {
1580     BDRVRawState *s = bs->opaque;
1581     struct hd_geometry ioctl_geo = {0};
1582 
1583     /* If DASD, get its geometry */
1584     if (check_for_dasd(s->fd) < 0) {
1585         return -ENOTSUP;
1586     }
1587     if (ioctl(s->fd, HDIO_GETGEO, &ioctl_geo) < 0) {
1588         return -errno;
1589     }
1590     /* HDIO_GETGEO may return success even though geo contains zeros
1591        (e.g. certain multipath setups) */
1592     if (!ioctl_geo.heads || !ioctl_geo.sectors || !ioctl_geo.cylinders) {
1593         return -ENOTSUP;
1594     }
1595     /* Do not return a geometry for partition */
1596     if (ioctl_geo.start != 0) {
1597         return -ENOTSUP;
1598     }
1599     geo->heads = ioctl_geo.heads;
1600     geo->sectors = ioctl_geo.sectors;
1601     geo->cylinders = ioctl_geo.cylinders;
1602 
1603     return 0;
1604 }
1605 #else /* __linux__ */
1606 static int hdev_probe_geometry(BlockDriverState *bs, HDGeometry *geo)
1607 {
1608     return -ENOTSUP;
1609 }
1610 #endif
1611 
1612 #if defined(__linux__)
1613 static int handle_aiocb_ioctl(void *opaque)
1614 {
1615     RawPosixAIOData *aiocb = opaque;
1616     int ret;
1617 
1618     ret = RETRY_ON_EINTR(
1619         ioctl(aiocb->aio_fildes, aiocb->ioctl.cmd, aiocb->ioctl.buf)
1620     );
1621     if (ret == -1) {
1622         return -errno;
1623     }
1624 
1625     return 0;
1626 }
1627 #endif /* linux */
1628 
1629 static int handle_aiocb_flush(void *opaque)
1630 {
1631     RawPosixAIOData *aiocb = opaque;
1632     BDRVRawState *s = aiocb->bs->opaque;
1633     int ret;
1634 
1635     if (s->page_cache_inconsistent) {
1636         return -s->page_cache_inconsistent;
1637     }
1638 
1639     ret = qemu_fdatasync(aiocb->aio_fildes);
1640     if (ret == -1) {
1641         trace_file_flush_fdatasync_failed(errno);
1642 
1643         /* There is no clear definition of the semantics of a failing fsync(),
1644          * so we may have to assume the worst. The sad truth is that this
1645          * assumption is correct for Linux. Some pages are now probably marked
1646          * clean in the page cache even though they are inconsistent with the
1647          * on-disk contents. The next fdatasync() call would succeed, but no
1648          * further writeback attempt will be made. We can't get back to a state
1649          * in which we know what is on disk (we would have to rewrite
1650          * everything that was touched since the last fdatasync() at least), so
1651          * make bdrv_flush() fail permanently. Given that the behaviour isn't
1652          * really defined, I have little hope that other OSes are doing better.
1653          *
1654          * Obviously, this doesn't affect O_DIRECT, which bypasses the page
1655          * cache. */
1656         if ((s->open_flags & O_DIRECT) == 0) {
1657             s->page_cache_inconsistent = errno;
1658         }
1659         return -errno;
1660     }
1661     return 0;
1662 }
1663 
1664 #ifdef CONFIG_PREADV
1665 
1666 static bool preadv_present = true;
1667 
1668 static ssize_t
1669 qemu_preadv(int fd, const struct iovec *iov, int nr_iov, off_t offset)
1670 {
1671     return preadv(fd, iov, nr_iov, offset);
1672 }
1673 
1674 static ssize_t
1675 qemu_pwritev(int fd, const struct iovec *iov, int nr_iov, off_t offset)
1676 {
1677     return pwritev(fd, iov, nr_iov, offset);
1678 }
1679 
1680 #else
1681 
1682 static bool preadv_present = false;
1683 
1684 static ssize_t
1685 qemu_preadv(int fd, const struct iovec *iov, int nr_iov, off_t offset)
1686 {
1687     return -ENOSYS;
1688 }
1689 
1690 static ssize_t
1691 qemu_pwritev(int fd, const struct iovec *iov, int nr_iov, off_t offset)
1692 {
1693     return -ENOSYS;
1694 }
1695 
1696 #endif
1697 
1698 static ssize_t handle_aiocb_rw_vector(RawPosixAIOData *aiocb)
1699 {
1700     ssize_t len;
1701 
1702     len = RETRY_ON_EINTR(
1703         (aiocb->aio_type & (QEMU_AIO_WRITE | QEMU_AIO_ZONE_APPEND)) ?
1704             qemu_pwritev(aiocb->aio_fildes,
1705                            aiocb->io.iov,
1706                            aiocb->io.niov,
1707                            aiocb->aio_offset) :
1708             qemu_preadv(aiocb->aio_fildes,
1709                           aiocb->io.iov,
1710                           aiocb->io.niov,
1711                           aiocb->aio_offset)
1712     );
1713 
1714     if (len == -1) {
1715         return -errno;
1716     }
1717     return len;
1718 }
1719 
1720 /*
1721  * Read/writes the data to/from a given linear buffer.
1722  *
1723  * Returns the number of bytes handles or -errno in case of an error. Short
1724  * reads are only returned if the end of the file is reached.
1725  */
1726 static ssize_t handle_aiocb_rw_linear(RawPosixAIOData *aiocb, char *buf)
1727 {
1728     ssize_t offset = 0;
1729     ssize_t len;
1730 
1731     while (offset < aiocb->aio_nbytes) {
1732         if (aiocb->aio_type & (QEMU_AIO_WRITE | QEMU_AIO_ZONE_APPEND)) {
1733             len = pwrite(aiocb->aio_fildes,
1734                          (const char *)buf + offset,
1735                          aiocb->aio_nbytes - offset,
1736                          aiocb->aio_offset + offset);
1737         } else {
1738             len = pread(aiocb->aio_fildes,
1739                         buf + offset,
1740                         aiocb->aio_nbytes - offset,
1741                         aiocb->aio_offset + offset);
1742         }
1743         if (len == -1 && errno == EINTR) {
1744             continue;
1745         } else if (len == -1 && errno == EINVAL &&
1746                    (aiocb->bs->open_flags & BDRV_O_NOCACHE) &&
1747                    !(aiocb->aio_type & QEMU_AIO_WRITE) &&
1748                    offset > 0) {
1749             /* O_DIRECT pread() may fail with EINVAL when offset is unaligned
1750              * after a short read.  Assume that O_DIRECT short reads only occur
1751              * at EOF.  Therefore this is a short read, not an I/O error.
1752              */
1753             break;
1754         } else if (len == -1) {
1755             offset = -errno;
1756             break;
1757         } else if (len == 0) {
1758             break;
1759         }
1760         offset += len;
1761     }
1762 
1763     return offset;
1764 }
1765 
1766 static int handle_aiocb_rw(void *opaque)
1767 {
1768     RawPosixAIOData *aiocb = opaque;
1769     ssize_t nbytes;
1770     char *buf;
1771 
1772     if (!(aiocb->aio_type & QEMU_AIO_MISALIGNED)) {
1773         /*
1774          * If there is just a single buffer, and it is properly aligned
1775          * we can just use plain pread/pwrite without any problems.
1776          */
1777         if (aiocb->io.niov == 1) {
1778             nbytes = handle_aiocb_rw_linear(aiocb, aiocb->io.iov->iov_base);
1779             goto out;
1780         }
1781         /*
1782          * We have more than one iovec, and all are properly aligned.
1783          *
1784          * Try preadv/pwritev first and fall back to linearizing the
1785          * buffer if it's not supported.
1786          */
1787         if (preadv_present) {
1788             nbytes = handle_aiocb_rw_vector(aiocb);
1789             if (nbytes == aiocb->aio_nbytes ||
1790                 (nbytes < 0 && nbytes != -ENOSYS)) {
1791                 goto out;
1792             }
1793             preadv_present = false;
1794         }
1795 
1796         /*
1797          * XXX(hch): short read/write.  no easy way to handle the reminder
1798          * using these interfaces.  For now retry using plain
1799          * pread/pwrite?
1800          */
1801     }
1802 
1803     /*
1804      * Ok, we have to do it the hard way, copy all segments into
1805      * a single aligned buffer.
1806      */
1807     buf = qemu_try_blockalign(aiocb->bs, aiocb->aio_nbytes);
1808     if (buf == NULL) {
1809         nbytes = -ENOMEM;
1810         goto out;
1811     }
1812 
1813     if (aiocb->aio_type & QEMU_AIO_WRITE) {
1814         char *p = buf;
1815         int i;
1816 
1817         for (i = 0; i < aiocb->io.niov; ++i) {
1818             memcpy(p, aiocb->io.iov[i].iov_base, aiocb->io.iov[i].iov_len);
1819             p += aiocb->io.iov[i].iov_len;
1820         }
1821         assert(p - buf == aiocb->aio_nbytes);
1822     }
1823 
1824     nbytes = handle_aiocb_rw_linear(aiocb, buf);
1825     if (!(aiocb->aio_type & (QEMU_AIO_WRITE | QEMU_AIO_ZONE_APPEND))) {
1826         char *p = buf;
1827         size_t count = aiocb->aio_nbytes, copy;
1828         int i;
1829 
1830         for (i = 0; i < aiocb->io.niov && count; ++i) {
1831             copy = count;
1832             if (copy > aiocb->io.iov[i].iov_len) {
1833                 copy = aiocb->io.iov[i].iov_len;
1834             }
1835             memcpy(aiocb->io.iov[i].iov_base, p, copy);
1836             assert(count >= copy);
1837             p     += copy;
1838             count -= copy;
1839         }
1840         assert(count == 0);
1841     }
1842     qemu_vfree(buf);
1843 
1844 out:
1845     if (nbytes == aiocb->aio_nbytes) {
1846         return 0;
1847     } else if (nbytes >= 0 && nbytes < aiocb->aio_nbytes) {
1848         if (aiocb->aio_type & QEMU_AIO_WRITE) {
1849             return -EINVAL;
1850         } else {
1851             iov_memset(aiocb->io.iov, aiocb->io.niov, nbytes,
1852                       0, aiocb->aio_nbytes - nbytes);
1853             return 0;
1854         }
1855     } else {
1856         assert(nbytes < 0);
1857         return nbytes;
1858     }
1859 }
1860 
1861 #if defined(CONFIG_FALLOCATE) || defined(BLKZEROOUT) || defined(BLKDISCARD)
1862 static int translate_err(int err)
1863 {
1864     if (err == -ENODEV || err == -ENOSYS || err == -EOPNOTSUPP ||
1865         err == -ENOTTY) {
1866         err = -ENOTSUP;
1867     }
1868     return err;
1869 }
1870 #endif
1871 
1872 #ifdef CONFIG_FALLOCATE
1873 static int do_fallocate(int fd, int mode, off_t offset, off_t len)
1874 {
1875     do {
1876         if (fallocate(fd, mode, offset, len) == 0) {
1877             return 0;
1878         }
1879     } while (errno == EINTR);
1880     return translate_err(-errno);
1881 }
1882 #endif
1883 
1884 static ssize_t handle_aiocb_write_zeroes_block(RawPosixAIOData *aiocb)
1885 {
1886     int ret = -ENOTSUP;
1887     BDRVRawState *s = aiocb->bs->opaque;
1888 
1889     if (!s->has_write_zeroes) {
1890         return -ENOTSUP;
1891     }
1892 
1893 #ifdef BLKZEROOUT
1894     /* The BLKZEROOUT implementation in the kernel doesn't set
1895      * BLKDEV_ZERO_NOFALLBACK, so we can't call this if we have to avoid slow
1896      * fallbacks. */
1897     if (!(aiocb->aio_type & QEMU_AIO_NO_FALLBACK)) {
1898         do {
1899             uint64_t range[2] = { aiocb->aio_offset, aiocb->aio_nbytes };
1900             if (ioctl(aiocb->aio_fildes, BLKZEROOUT, range) == 0) {
1901                 return 0;
1902             }
1903         } while (errno == EINTR);
1904 
1905         ret = translate_err(-errno);
1906         if (ret == -ENOTSUP) {
1907             s->has_write_zeroes = false;
1908         }
1909     }
1910 #endif
1911 
1912     return ret;
1913 }
1914 
1915 static int handle_aiocb_write_zeroes(void *opaque)
1916 {
1917     RawPosixAIOData *aiocb = opaque;
1918 #ifdef CONFIG_FALLOCATE
1919     BDRVRawState *s = aiocb->bs->opaque;
1920     int64_t len;
1921 #endif
1922 
1923     if (aiocb->aio_type & QEMU_AIO_BLKDEV) {
1924         return handle_aiocb_write_zeroes_block(aiocb);
1925     }
1926 
1927 #ifdef CONFIG_FALLOCATE_ZERO_RANGE
1928     if (s->has_write_zeroes) {
1929         int ret = do_fallocate(s->fd, FALLOC_FL_ZERO_RANGE,
1930                                aiocb->aio_offset, aiocb->aio_nbytes);
1931         if (ret == -ENOTSUP) {
1932             s->has_write_zeroes = false;
1933         } else if (ret == 0 || ret != -EINVAL) {
1934             return ret;
1935         }
1936         /*
1937          * Note: Some file systems do not like unaligned byte ranges, and
1938          * return EINVAL in such a case, though they should not do it according
1939          * to the man-page of fallocate(). Thus we simply ignore this return
1940          * value and try the other fallbacks instead.
1941          */
1942     }
1943 #endif
1944 
1945 #ifdef CONFIG_FALLOCATE_PUNCH_HOLE
1946     if (s->has_discard && s->has_fallocate) {
1947         int ret = do_fallocate(s->fd,
1948                                FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE,
1949                                aiocb->aio_offset, aiocb->aio_nbytes);
1950         if (ret == 0) {
1951             ret = do_fallocate(s->fd, 0, aiocb->aio_offset, aiocb->aio_nbytes);
1952             if (ret == 0 || ret != -ENOTSUP) {
1953                 return ret;
1954             }
1955             s->has_fallocate = false;
1956         } else if (ret == -EINVAL) {
1957             /*
1958              * Some file systems like older versions of GPFS do not like un-
1959              * aligned byte ranges, and return EINVAL in such a case, though
1960              * they should not do it according to the man-page of fallocate().
1961              * Warn about the bad filesystem and try the final fallback instead.
1962              */
1963             warn_report_once("Your file system is misbehaving: "
1964                              "fallocate(FALLOC_FL_PUNCH_HOLE) returned EINVAL. "
1965                              "Please report this bug to your file system "
1966                              "vendor.");
1967         } else if (ret != -ENOTSUP) {
1968             return ret;
1969         } else {
1970             s->has_discard = false;
1971         }
1972     }
1973 #endif
1974 
1975 #ifdef CONFIG_FALLOCATE
1976     /* Last resort: we are trying to extend the file with zeroed data. This
1977      * can be done via fallocate(fd, 0) */
1978     len = raw_getlength(aiocb->bs);
1979     if (s->has_fallocate && len >= 0 && aiocb->aio_offset >= len) {
1980         int ret = do_fallocate(s->fd, 0, aiocb->aio_offset, aiocb->aio_nbytes);
1981         if (ret == 0 || ret != -ENOTSUP) {
1982             return ret;
1983         }
1984         s->has_fallocate = false;
1985     }
1986 #endif
1987 
1988     return -ENOTSUP;
1989 }
1990 
1991 static int handle_aiocb_write_zeroes_unmap(void *opaque)
1992 {
1993     RawPosixAIOData *aiocb = opaque;
1994     BDRVRawState *s G_GNUC_UNUSED = aiocb->bs->opaque;
1995 
1996     /* First try to write zeros and unmap at the same time */
1997 
1998 #ifdef CONFIG_FALLOCATE_PUNCH_HOLE
1999     int ret = do_fallocate(s->fd, FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE,
2000                            aiocb->aio_offset, aiocb->aio_nbytes);
2001     switch (ret) {
2002     case -ENOTSUP:
2003     case -EINVAL:
2004     case -EBUSY:
2005         break;
2006     default:
2007         return ret;
2008     }
2009 #endif
2010 
2011     /* If we couldn't manage to unmap while guaranteed that the area reads as
2012      * all-zero afterwards, just write zeroes without unmapping */
2013     return handle_aiocb_write_zeroes(aiocb);
2014 }
2015 
2016 #ifndef HAVE_COPY_FILE_RANGE
2017 static off_t copy_file_range(int in_fd, off_t *in_off, int out_fd,
2018                              off_t *out_off, size_t len, unsigned int flags)
2019 {
2020 #ifdef __NR_copy_file_range
2021     return syscall(__NR_copy_file_range, in_fd, in_off, out_fd,
2022                    out_off, len, flags);
2023 #else
2024     errno = ENOSYS;
2025     return -1;
2026 #endif
2027 }
2028 #endif
2029 
2030 /*
2031  * parse_zone - Fill a zone descriptor
2032  */
2033 #if defined(CONFIG_BLKZONED)
2034 static inline int parse_zone(struct BlockZoneDescriptor *zone,
2035                               const struct blk_zone *blkz) {
2036     zone->start = blkz->start << BDRV_SECTOR_BITS;
2037     zone->length = blkz->len << BDRV_SECTOR_BITS;
2038     zone->wp = blkz->wp << BDRV_SECTOR_BITS;
2039 
2040 #ifdef HAVE_BLK_ZONE_REP_CAPACITY
2041     zone->cap = blkz->capacity << BDRV_SECTOR_BITS;
2042 #else
2043     zone->cap = blkz->len << BDRV_SECTOR_BITS;
2044 #endif
2045 
2046     switch (blkz->type) {
2047     case BLK_ZONE_TYPE_SEQWRITE_REQ:
2048         zone->type = BLK_ZT_SWR;
2049         break;
2050     case BLK_ZONE_TYPE_SEQWRITE_PREF:
2051         zone->type = BLK_ZT_SWP;
2052         break;
2053     case BLK_ZONE_TYPE_CONVENTIONAL:
2054         zone->type = BLK_ZT_CONV;
2055         break;
2056     default:
2057         error_report("Unsupported zone type: 0x%x", blkz->type);
2058         return -ENOTSUP;
2059     }
2060 
2061     switch (blkz->cond) {
2062     case BLK_ZONE_COND_NOT_WP:
2063         zone->state = BLK_ZS_NOT_WP;
2064         break;
2065     case BLK_ZONE_COND_EMPTY:
2066         zone->state = BLK_ZS_EMPTY;
2067         break;
2068     case BLK_ZONE_COND_IMP_OPEN:
2069         zone->state = BLK_ZS_IOPEN;
2070         break;
2071     case BLK_ZONE_COND_EXP_OPEN:
2072         zone->state = BLK_ZS_EOPEN;
2073         break;
2074     case BLK_ZONE_COND_CLOSED:
2075         zone->state = BLK_ZS_CLOSED;
2076         break;
2077     case BLK_ZONE_COND_READONLY:
2078         zone->state = BLK_ZS_RDONLY;
2079         break;
2080     case BLK_ZONE_COND_FULL:
2081         zone->state = BLK_ZS_FULL;
2082         break;
2083     case BLK_ZONE_COND_OFFLINE:
2084         zone->state = BLK_ZS_OFFLINE;
2085         break;
2086     default:
2087         error_report("Unsupported zone state: 0x%x", blkz->cond);
2088         return -ENOTSUP;
2089     }
2090     return 0;
2091 }
2092 #endif
2093 
2094 #if defined(CONFIG_BLKZONED)
2095 static int handle_aiocb_zone_report(void *opaque)
2096 {
2097     RawPosixAIOData *aiocb = opaque;
2098     int fd = aiocb->aio_fildes;
2099     unsigned int *nr_zones = aiocb->zone_report.nr_zones;
2100     BlockZoneDescriptor *zones = aiocb->zone_report.zones;
2101     /* zoned block devices use 512-byte sectors */
2102     uint64_t sector = aiocb->aio_offset / 512;
2103 
2104     struct blk_zone *blkz;
2105     size_t rep_size;
2106     unsigned int nrz;
2107     int ret;
2108     unsigned int n = 0, i = 0;
2109 
2110     nrz = *nr_zones;
2111     rep_size = sizeof(struct blk_zone_report) + nrz * sizeof(struct blk_zone);
2112     g_autofree struct blk_zone_report *rep = NULL;
2113     rep = g_malloc(rep_size);
2114 
2115     blkz = (struct blk_zone *)(rep + 1);
2116     while (n < nrz) {
2117         memset(rep, 0, rep_size);
2118         rep->sector = sector;
2119         rep->nr_zones = nrz - n;
2120 
2121         do {
2122             ret = ioctl(fd, BLKREPORTZONE, rep);
2123         } while (ret != 0 && errno == EINTR);
2124         if (ret != 0) {
2125             error_report("%d: ioctl BLKREPORTZONE at %" PRId64 " failed %d",
2126                          fd, sector, errno);
2127             return -errno;
2128         }
2129 
2130         if (!rep->nr_zones) {
2131             break;
2132         }
2133 
2134         for (i = 0; i < rep->nr_zones; i++, n++) {
2135             ret = parse_zone(&zones[n], &blkz[i]);
2136             if (ret != 0) {
2137                 return ret;
2138             }
2139 
2140             /* The next report should start after the last zone reported */
2141             sector = blkz[i].start + blkz[i].len;
2142         }
2143     }
2144 
2145     *nr_zones = n;
2146     return 0;
2147 }
2148 #endif
2149 
2150 #if defined(CONFIG_BLKZONED)
2151 static int handle_aiocb_zone_mgmt(void *opaque)
2152 {
2153     RawPosixAIOData *aiocb = opaque;
2154     int fd = aiocb->aio_fildes;
2155     uint64_t sector = aiocb->aio_offset / 512;
2156     int64_t nr_sectors = aiocb->aio_nbytes / 512;
2157     struct blk_zone_range range;
2158     int ret;
2159 
2160     /* Execute the operation */
2161     range.sector = sector;
2162     range.nr_sectors = nr_sectors;
2163     do {
2164         ret = ioctl(fd, aiocb->zone_mgmt.op, &range);
2165     } while (ret != 0 && errno == EINTR);
2166 
2167     return ret < 0 ? -errno : ret;
2168 }
2169 #endif
2170 
2171 static int handle_aiocb_copy_range(void *opaque)
2172 {
2173     RawPosixAIOData *aiocb = opaque;
2174     uint64_t bytes = aiocb->aio_nbytes;
2175     off_t in_off = aiocb->aio_offset;
2176     off_t out_off = aiocb->copy_range.aio_offset2;
2177 
2178     while (bytes) {
2179         ssize_t ret = copy_file_range(aiocb->aio_fildes, &in_off,
2180                                       aiocb->copy_range.aio_fd2, &out_off,
2181                                       bytes, 0);
2182         trace_file_copy_file_range(aiocb->bs, aiocb->aio_fildes, in_off,
2183                                    aiocb->copy_range.aio_fd2, out_off, bytes,
2184                                    0, ret);
2185         if (ret == 0) {
2186             /* No progress (e.g. when beyond EOF), let the caller fall back to
2187              * buffer I/O. */
2188             return -ENOSPC;
2189         }
2190         if (ret < 0) {
2191             switch (errno) {
2192             case ENOSYS:
2193                 return -ENOTSUP;
2194             case EINTR:
2195                 continue;
2196             default:
2197                 return -errno;
2198             }
2199         }
2200         bytes -= ret;
2201     }
2202     return 0;
2203 }
2204 
2205 static int handle_aiocb_discard(void *opaque)
2206 {
2207     RawPosixAIOData *aiocb = opaque;
2208     int ret = -ENOTSUP;
2209     BDRVRawState *s = aiocb->bs->opaque;
2210 
2211     if (!s->has_discard) {
2212         return -ENOTSUP;
2213     }
2214 
2215     if (aiocb->aio_type & QEMU_AIO_BLKDEV) {
2216 #ifdef BLKDISCARD
2217         do {
2218             uint64_t range[2] = { aiocb->aio_offset, aiocb->aio_nbytes };
2219             if (ioctl(aiocb->aio_fildes, BLKDISCARD, range) == 0) {
2220                 return 0;
2221             }
2222         } while (errno == EINTR);
2223 
2224         ret = translate_err(-errno);
2225 #endif
2226     } else {
2227 #ifdef CONFIG_FALLOCATE_PUNCH_HOLE
2228         ret = do_fallocate(s->fd, FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE,
2229                            aiocb->aio_offset, aiocb->aio_nbytes);
2230         ret = translate_err(ret);
2231 #elif defined(__APPLE__) && (__MACH__)
2232         fpunchhole_t fpunchhole;
2233         fpunchhole.fp_flags = 0;
2234         fpunchhole.reserved = 0;
2235         fpunchhole.fp_offset = aiocb->aio_offset;
2236         fpunchhole.fp_length = aiocb->aio_nbytes;
2237         if (fcntl(s->fd, F_PUNCHHOLE, &fpunchhole) == -1) {
2238             ret = errno == ENODEV ? -ENOTSUP : -errno;
2239         } else {
2240             ret = 0;
2241         }
2242 #endif
2243     }
2244 
2245     if (ret == -ENOTSUP) {
2246         s->has_discard = false;
2247     }
2248     return ret;
2249 }
2250 
2251 /*
2252  * Help alignment probing by allocating the first block.
2253  *
2254  * When reading with direct I/O from unallocated area on Gluster backed by XFS,
2255  * reading succeeds regardless of request length. In this case we fallback to
2256  * safe alignment which is not optimal. Allocating the first block avoids this
2257  * fallback.
2258  *
2259  * fd may be opened with O_DIRECT, but we don't know the buffer alignment or
2260  * request alignment, so we use safe values.
2261  *
2262  * Returns: 0 on success, -errno on failure. Since this is an optimization,
2263  * caller may ignore failures.
2264  */
2265 static int allocate_first_block(int fd, size_t max_size)
2266 {
2267     size_t write_size = (max_size < MAX_BLOCKSIZE)
2268         ? BDRV_SECTOR_SIZE
2269         : MAX_BLOCKSIZE;
2270     size_t max_align = MAX(MAX_BLOCKSIZE, qemu_real_host_page_size());
2271     void *buf;
2272     ssize_t n;
2273     int ret;
2274 
2275     buf = qemu_memalign(max_align, write_size);
2276     memset(buf, 0, write_size);
2277 
2278     n = RETRY_ON_EINTR(pwrite(fd, buf, write_size, 0));
2279 
2280     ret = (n == -1) ? -errno : 0;
2281 
2282     qemu_vfree(buf);
2283     return ret;
2284 }
2285 
2286 static int handle_aiocb_truncate(void *opaque)
2287 {
2288     RawPosixAIOData *aiocb = opaque;
2289     int result = 0;
2290     int64_t current_length = 0;
2291     char *buf = NULL;
2292     struct stat st;
2293     int fd = aiocb->aio_fildes;
2294     int64_t offset = aiocb->aio_offset;
2295     PreallocMode prealloc = aiocb->truncate.prealloc;
2296     Error **errp = aiocb->truncate.errp;
2297 
2298     if (fstat(fd, &st) < 0) {
2299         result = -errno;
2300         error_setg_errno(errp, -result, "Could not stat file");
2301         return result;
2302     }
2303 
2304     current_length = st.st_size;
2305     if (current_length > offset && prealloc != PREALLOC_MODE_OFF) {
2306         error_setg(errp, "Cannot use preallocation for shrinking files");
2307         return -ENOTSUP;
2308     }
2309 
2310     switch (prealloc) {
2311 #ifdef CONFIG_POSIX_FALLOCATE
2312     case PREALLOC_MODE_FALLOC:
2313         /*
2314          * Truncating before posix_fallocate() makes it about twice slower on
2315          * file systems that do not support fallocate(), trying to check if a
2316          * block is allocated before allocating it, so don't do that here.
2317          */
2318         if (offset != current_length) {
2319             result = -posix_fallocate(fd, current_length,
2320                                       offset - current_length);
2321             if (result != 0) {
2322                 /* posix_fallocate() doesn't set errno. */
2323                 error_setg_errno(errp, -result,
2324                                  "Could not preallocate new data");
2325             } else if (current_length == 0) {
2326                 /*
2327                  * posix_fallocate() uses fallocate() if the filesystem
2328                  * supports it, or fallback to manually writing zeroes. If
2329                  * fallocate() was used, unaligned reads from the fallocated
2330                  * area in raw_probe_alignment() will succeed, hence we need to
2331                  * allocate the first block.
2332                  *
2333                  * Optimize future alignment probing; ignore failures.
2334                  */
2335                 allocate_first_block(fd, offset);
2336             }
2337         } else {
2338             result = 0;
2339         }
2340         goto out;
2341 #endif
2342     case PREALLOC_MODE_FULL:
2343     {
2344         int64_t num = 0, left = offset - current_length;
2345         off_t seek_result;
2346 
2347         /*
2348          * Knowing the final size from the beginning could allow the file
2349          * system driver to do less allocations and possibly avoid
2350          * fragmentation of the file.
2351          */
2352         if (ftruncate(fd, offset) != 0) {
2353             result = -errno;
2354             error_setg_errno(errp, -result, "Could not resize file");
2355             goto out;
2356         }
2357 
2358         buf = g_malloc0(65536);
2359 
2360         seek_result = lseek(fd, current_length, SEEK_SET);
2361         if (seek_result < 0) {
2362             result = -errno;
2363             error_setg_errno(errp, -result,
2364                              "Failed to seek to the old end of file");
2365             goto out;
2366         }
2367 
2368         while (left > 0) {
2369             num = MIN(left, 65536);
2370             result = write(fd, buf, num);
2371             if (result < 0) {
2372                 if (errno == EINTR) {
2373                     continue;
2374                 }
2375                 result = -errno;
2376                 error_setg_errno(errp, -result,
2377                                  "Could not write zeros for preallocation");
2378                 goto out;
2379             }
2380             left -= result;
2381         }
2382         if (result >= 0) {
2383             result = fsync(fd);
2384             if (result < 0) {
2385                 result = -errno;
2386                 error_setg_errno(errp, -result,
2387                                  "Could not flush file to disk");
2388                 goto out;
2389             }
2390         }
2391         goto out;
2392     }
2393     case PREALLOC_MODE_OFF:
2394         if (ftruncate(fd, offset) != 0) {
2395             result = -errno;
2396             error_setg_errno(errp, -result, "Could not resize file");
2397         } else if (current_length == 0 && offset > current_length) {
2398             /* Optimize future alignment probing; ignore failures. */
2399             allocate_first_block(fd, offset);
2400         }
2401         return result;
2402     default:
2403         result = -ENOTSUP;
2404         error_setg(errp, "Unsupported preallocation mode: %s",
2405                    PreallocMode_str(prealloc));
2406         return result;
2407     }
2408 
2409 out:
2410     if (result < 0) {
2411         if (ftruncate(fd, current_length) < 0) {
2412             error_report("Failed to restore old file length: %s",
2413                          strerror(errno));
2414         }
2415     }
2416 
2417     g_free(buf);
2418     return result;
2419 }
2420 
2421 static int coroutine_fn raw_thread_pool_submit(ThreadPoolFunc func, void *arg)
2422 {
2423     return thread_pool_submit_co(func, arg);
2424 }
2425 
2426 /*
2427  * Check if all memory in this vector is sector aligned.
2428  */
2429 static bool bdrv_qiov_is_aligned(BlockDriverState *bs, QEMUIOVector *qiov)
2430 {
2431     int i;
2432     size_t alignment = bdrv_min_mem_align(bs);
2433     size_t len = bs->bl.request_alignment;
2434     IO_CODE();
2435 
2436     for (i = 0; i < qiov->niov; i++) {
2437         if ((uintptr_t) qiov->iov[i].iov_base % alignment) {
2438             return false;
2439         }
2440         if (qiov->iov[i].iov_len % len) {
2441             return false;
2442         }
2443     }
2444 
2445     return true;
2446 }
2447 
2448 static int coroutine_fn raw_co_prw(BlockDriverState *bs, uint64_t offset,
2449                                    uint64_t bytes, QEMUIOVector *qiov, int type)
2450 {
2451     BDRVRawState *s = bs->opaque;
2452     RawPosixAIOData acb;
2453     int ret;
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 {
2510     BlockZoneWps *wps = bs->wps;
2511     if (ret == 0) {
2512         if ((type & (QEMU_AIO_WRITE | QEMU_AIO_ZONE_APPEND)) &&
2513             bs->bl.zoned != BLK_Z_NONE) {
2514             uint64_t *wp = &wps->wp[offset / bs->bl.zone_size];
2515             if (!BDRV_ZT_IS_CONV(*wp)) {
2516                 if (type & QEMU_AIO_ZONE_APPEND) {
2517                     *s->offset = *wp;
2518                     trace_zbd_zone_append_complete(bs, *s->offset
2519                         >> BDRV_SECTOR_BITS);
2520                 }
2521                 /* Advance the wp if needed */
2522                 if (offset + bytes > *wp) {
2523                     *wp = offset + bytes;
2524                 }
2525             }
2526         }
2527     } else {
2528         if (type & (QEMU_AIO_WRITE | QEMU_AIO_ZONE_APPEND)) {
2529             update_zones_wp(bs, s->fd, 0, 1);
2530         }
2531     }
2532 
2533     if ((type & (QEMU_AIO_WRITE | QEMU_AIO_ZONE_APPEND)) &&
2534         bs->blk.zoned != BLK_Z_NONE) {
2535         qemu_co_mutex_unlock(&wps->colock);
2536     }
2537 }
2538 #endif
2539     return ret;
2540 }
2541 
2542 static int coroutine_fn raw_co_preadv(BlockDriverState *bs, int64_t offset,
2543                                       int64_t bytes, QEMUIOVector *qiov,
2544                                       BdrvRequestFlags flags)
2545 {
2546     return raw_co_prw(bs, offset, bytes, qiov, QEMU_AIO_READ);
2547 }
2548 
2549 static int coroutine_fn raw_co_pwritev(BlockDriverState *bs, int64_t offset,
2550                                        int64_t bytes, QEMUIOVector *qiov,
2551                                        BdrvRequestFlags flags)
2552 {
2553     return raw_co_prw(bs, offset, bytes, qiov, QEMU_AIO_WRITE);
2554 }
2555 
2556 static int coroutine_fn raw_co_flush_to_disk(BlockDriverState *bs)
2557 {
2558     BDRVRawState *s = bs->opaque;
2559     RawPosixAIOData acb;
2560     int ret;
2561 
2562     ret = fd_open(bs);
2563     if (ret < 0) {
2564         return ret;
2565     }
2566 
2567     acb = (RawPosixAIOData) {
2568         .bs             = bs,
2569         .aio_fildes     = s->fd,
2570         .aio_type       = QEMU_AIO_FLUSH,
2571     };
2572 
2573 #ifdef CONFIG_LINUX_IO_URING
2574     if (s->use_linux_io_uring) {
2575         return luring_co_submit(bs, s->fd, 0, NULL, QEMU_AIO_FLUSH);
2576     }
2577 #endif
2578     return raw_thread_pool_submit(handle_aiocb_flush, &acb);
2579 }
2580 
2581 static void raw_aio_attach_aio_context(BlockDriverState *bs,
2582                                        AioContext *new_context)
2583 {
2584     BDRVRawState __attribute__((unused)) *s = bs->opaque;
2585 #ifdef CONFIG_LINUX_AIO
2586     if (s->use_linux_aio) {
2587         Error *local_err = NULL;
2588         if (!aio_setup_linux_aio(new_context, &local_err)) {
2589             error_reportf_err(local_err, "Unable to use native AIO, "
2590                                          "falling back to thread pool: ");
2591             s->use_linux_aio = false;
2592         }
2593     }
2594 #endif
2595 #ifdef CONFIG_LINUX_IO_URING
2596     if (s->use_linux_io_uring) {
2597         Error *local_err = NULL;
2598         if (!aio_setup_linux_io_uring(new_context, &local_err)) {
2599             error_reportf_err(local_err, "Unable to use linux io_uring, "
2600                                          "falling back to thread pool: ");
2601             s->use_linux_io_uring = false;
2602         }
2603     }
2604 #endif
2605 }
2606 
2607 static void raw_close(BlockDriverState *bs)
2608 {
2609     BDRVRawState *s = bs->opaque;
2610 
2611     if (s->fd >= 0) {
2612 #if defined(CONFIG_BLKZONED)
2613         g_free(bs->wps);
2614 #endif
2615         qemu_close(s->fd);
2616         s->fd = -1;
2617     }
2618 }
2619 
2620 /**
2621  * Truncates the given regular file @fd to @offset and, when growing, fills the
2622  * new space according to @prealloc.
2623  *
2624  * Returns: 0 on success, -errno on failure.
2625  */
2626 static int coroutine_fn
2627 raw_regular_truncate(BlockDriverState *bs, int fd, int64_t offset,
2628                      PreallocMode prealloc, Error **errp)
2629 {
2630     RawPosixAIOData acb;
2631 
2632     acb = (RawPosixAIOData) {
2633         .bs             = bs,
2634         .aio_fildes     = fd,
2635         .aio_type       = QEMU_AIO_TRUNCATE,
2636         .aio_offset     = offset,
2637         .truncate       = {
2638             .prealloc       = prealloc,
2639             .errp           = errp,
2640         },
2641     };
2642 
2643     return raw_thread_pool_submit(handle_aiocb_truncate, &acb);
2644 }
2645 
2646 static int coroutine_fn raw_co_truncate(BlockDriverState *bs, int64_t offset,
2647                                         bool exact, PreallocMode prealloc,
2648                                         BdrvRequestFlags flags, Error **errp)
2649 {
2650     BDRVRawState *s = bs->opaque;
2651     struct stat st;
2652     int ret;
2653 
2654     if (fstat(s->fd, &st)) {
2655         ret = -errno;
2656         error_setg_errno(errp, -ret, "Failed to fstat() the file");
2657         return ret;
2658     }
2659 
2660     if (S_ISREG(st.st_mode)) {
2661         /* Always resizes to the exact @offset */
2662         return raw_regular_truncate(bs, s->fd, offset, prealloc, errp);
2663     }
2664 
2665     if (prealloc != PREALLOC_MODE_OFF) {
2666         error_setg(errp, "Preallocation mode '%s' unsupported for this "
2667                    "non-regular file", PreallocMode_str(prealloc));
2668         return -ENOTSUP;
2669     }
2670 
2671     if (S_ISCHR(st.st_mode) || S_ISBLK(st.st_mode)) {
2672         int64_t cur_length = raw_getlength(bs);
2673 
2674         if (offset != cur_length && exact) {
2675             error_setg(errp, "Cannot resize device files");
2676             return -ENOTSUP;
2677         } else if (offset > cur_length) {
2678             error_setg(errp, "Cannot grow device files");
2679             return -EINVAL;
2680         }
2681     } else {
2682         error_setg(errp, "Resizing this file is not supported");
2683         return -ENOTSUP;
2684     }
2685 
2686     return 0;
2687 }
2688 
2689 #ifdef __OpenBSD__
2690 static int64_t raw_getlength(BlockDriverState *bs)
2691 {
2692     BDRVRawState *s = bs->opaque;
2693     int fd = s->fd;
2694     struct stat st;
2695 
2696     if (fstat(fd, &st))
2697         return -errno;
2698     if (S_ISCHR(st.st_mode) || S_ISBLK(st.st_mode)) {
2699         struct disklabel dl;
2700 
2701         if (ioctl(fd, DIOCGDINFO, &dl))
2702             return -errno;
2703         return (uint64_t)dl.d_secsize *
2704             dl.d_partitions[DISKPART(st.st_rdev)].p_size;
2705     } else
2706         return st.st_size;
2707 }
2708 #elif defined(__NetBSD__)
2709 static int64_t raw_getlength(BlockDriverState *bs)
2710 {
2711     BDRVRawState *s = bs->opaque;
2712     int fd = s->fd;
2713     struct stat st;
2714 
2715     if (fstat(fd, &st))
2716         return -errno;
2717     if (S_ISCHR(st.st_mode) || S_ISBLK(st.st_mode)) {
2718         struct dkwedge_info dkw;
2719 
2720         if (ioctl(fd, DIOCGWEDGEINFO, &dkw) != -1) {
2721             return dkw.dkw_size * 512;
2722         } else {
2723             struct disklabel dl;
2724 
2725             if (ioctl(fd, DIOCGDINFO, &dl))
2726                 return -errno;
2727             return (uint64_t)dl.d_secsize *
2728                 dl.d_partitions[DISKPART(st.st_rdev)].p_size;
2729         }
2730     } else
2731         return st.st_size;
2732 }
2733 #elif defined(__sun__)
2734 static int64_t raw_getlength(BlockDriverState *bs)
2735 {
2736     BDRVRawState *s = bs->opaque;
2737     struct dk_minfo minfo;
2738     int ret;
2739     int64_t size;
2740 
2741     ret = fd_open(bs);
2742     if (ret < 0) {
2743         return ret;
2744     }
2745 
2746     /*
2747      * Use the DKIOCGMEDIAINFO ioctl to read the size.
2748      */
2749     ret = ioctl(s->fd, DKIOCGMEDIAINFO, &minfo);
2750     if (ret != -1) {
2751         return minfo.dki_lbsize * minfo.dki_capacity;
2752     }
2753 
2754     /*
2755      * There are reports that lseek on some devices fails, but
2756      * irc discussion said that contingency on contingency was overkill.
2757      */
2758     size = lseek(s->fd, 0, SEEK_END);
2759     if (size < 0) {
2760         return -errno;
2761     }
2762     return size;
2763 }
2764 #elif defined(CONFIG_BSD)
2765 static int64_t raw_getlength(BlockDriverState *bs)
2766 {
2767     BDRVRawState *s = bs->opaque;
2768     int fd = s->fd;
2769     int64_t size;
2770     struct stat sb;
2771 #if defined (__FreeBSD__) || defined(__FreeBSD_kernel__)
2772     int reopened = 0;
2773 #endif
2774     int ret;
2775 
2776     ret = fd_open(bs);
2777     if (ret < 0)
2778         return ret;
2779 
2780 #if defined (__FreeBSD__) || defined(__FreeBSD_kernel__)
2781 again:
2782 #endif
2783     if (!fstat(fd, &sb) && (S_IFCHR & sb.st_mode)) {
2784         size = 0;
2785 #ifdef DIOCGMEDIASIZE
2786         if (ioctl(fd, DIOCGMEDIASIZE, (off_t *)&size)) {
2787             size = 0;
2788         }
2789 #endif
2790 #ifdef DIOCGPART
2791         if (size == 0) {
2792             struct partinfo pi;
2793             if (ioctl(fd, DIOCGPART, &pi) == 0) {
2794                 size = pi.media_size;
2795             }
2796         }
2797 #endif
2798 #if defined(DKIOCGETBLOCKCOUNT) && defined(DKIOCGETBLOCKSIZE)
2799         if (size == 0) {
2800             uint64_t sectors = 0;
2801             uint32_t sector_size = 0;
2802 
2803             if (ioctl(fd, DKIOCGETBLOCKCOUNT, &sectors) == 0
2804                && ioctl(fd, DKIOCGETBLOCKSIZE, &sector_size) == 0) {
2805                 size = sectors * sector_size;
2806             }
2807         }
2808 #endif
2809         if (size == 0) {
2810             size = lseek(fd, 0LL, SEEK_END);
2811         }
2812         if (size < 0) {
2813             return -errno;
2814         }
2815 #if defined(__FreeBSD__) || defined(__FreeBSD_kernel__)
2816         switch(s->type) {
2817         case FTYPE_CD:
2818             /* XXX FreeBSD acd returns UINT_MAX sectors for an empty drive */
2819             if (size == 2048LL * (unsigned)-1)
2820                 size = 0;
2821             /* XXX no disc?  maybe we need to reopen... */
2822             if (size <= 0 && !reopened && cdrom_reopen(bs) >= 0) {
2823                 reopened = 1;
2824                 goto again;
2825             }
2826         }
2827 #endif
2828     } else {
2829         size = lseek(fd, 0, SEEK_END);
2830         if (size < 0) {
2831             return -errno;
2832         }
2833     }
2834     return size;
2835 }
2836 #else
2837 static int64_t raw_getlength(BlockDriverState *bs)
2838 {
2839     BDRVRawState *s = bs->opaque;
2840     int ret;
2841     int64_t size;
2842 
2843     ret = fd_open(bs);
2844     if (ret < 0) {
2845         return ret;
2846     }
2847 
2848     size = lseek(s->fd, 0, SEEK_END);
2849     if (size < 0) {
2850         return -errno;
2851     }
2852     return size;
2853 }
2854 #endif
2855 
2856 static int64_t coroutine_fn raw_co_getlength(BlockDriverState *bs)
2857 {
2858     return raw_getlength(bs);
2859 }
2860 
2861 static int64_t coroutine_fn raw_co_get_allocated_file_size(BlockDriverState *bs)
2862 {
2863     struct stat st;
2864     BDRVRawState *s = bs->opaque;
2865 
2866     if (fstat(s->fd, &st) < 0) {
2867         return -errno;
2868     }
2869     return (int64_t)st.st_blocks * 512;
2870 }
2871 
2872 static int coroutine_fn
2873 raw_co_create(BlockdevCreateOptions *options, Error **errp)
2874 {
2875     BlockdevCreateOptionsFile *file_opts;
2876     Error *local_err = NULL;
2877     int fd;
2878     uint64_t perm, shared;
2879     int result = 0;
2880 
2881     /* Validate options and set default values */
2882     assert(options->driver == BLOCKDEV_DRIVER_FILE);
2883     file_opts = &options->u.file;
2884 
2885     if (!file_opts->has_nocow) {
2886         file_opts->nocow = false;
2887     }
2888     if (!file_opts->has_preallocation) {
2889         file_opts->preallocation = PREALLOC_MODE_OFF;
2890     }
2891     if (!file_opts->has_extent_size_hint) {
2892         file_opts->extent_size_hint = 1 * MiB;
2893     }
2894     if (file_opts->extent_size_hint > UINT32_MAX) {
2895         result = -EINVAL;
2896         error_setg(errp, "Extent size hint is too large");
2897         goto out;
2898     }
2899 
2900     /* Create file */
2901     fd = qemu_create(file_opts->filename, O_RDWR | O_BINARY, 0644, errp);
2902     if (fd < 0) {
2903         result = -errno;
2904         goto out;
2905     }
2906 
2907     /* Take permissions: We want to discard everything, so we need
2908      * BLK_PERM_WRITE; and truncation to the desired size requires
2909      * BLK_PERM_RESIZE.
2910      * On the other hand, we cannot share the RESIZE permission
2911      * because we promise that after this function, the file has the
2912      * size given in the options.  If someone else were to resize it
2913      * concurrently, we could not guarantee that.
2914      * Note that after this function, we can no longer guarantee that
2915      * the file is not touched by a third party, so it may be resized
2916      * then. */
2917     perm = BLK_PERM_WRITE | BLK_PERM_RESIZE;
2918     shared = BLK_PERM_ALL & ~BLK_PERM_RESIZE;
2919 
2920     /* Step one: Take locks */
2921     result = raw_apply_lock_bytes(NULL, fd, perm, ~shared, false, errp);
2922     if (result < 0) {
2923         goto out_close;
2924     }
2925 
2926     /* Step two: Check that nobody else has taken conflicting locks */
2927     result = raw_check_lock_bytes(fd, perm, shared, errp);
2928     if (result < 0) {
2929         error_append_hint(errp,
2930                           "Is another process using the image [%s]?\n",
2931                           file_opts->filename);
2932         goto out_unlock;
2933     }
2934 
2935     /* Clear the file by truncating it to 0 */
2936     result = raw_regular_truncate(NULL, fd, 0, PREALLOC_MODE_OFF, errp);
2937     if (result < 0) {
2938         goto out_unlock;
2939     }
2940 
2941     if (file_opts->nocow) {
2942 #ifdef __linux__
2943         /* Set NOCOW flag to solve performance issue on fs like btrfs.
2944          * This is an optimisation. The FS_IOC_SETFLAGS ioctl return value
2945          * will be ignored since any failure of this operation should not
2946          * block the left work.
2947          */
2948         int attr;
2949         if (ioctl(fd, FS_IOC_GETFLAGS, &attr) == 0) {
2950             attr |= FS_NOCOW_FL;
2951             ioctl(fd, FS_IOC_SETFLAGS, &attr);
2952         }
2953 #endif
2954     }
2955 #ifdef FS_IOC_FSSETXATTR
2956     /*
2957      * Try to set the extent size hint. Failure is not fatal, and a warning is
2958      * only printed if the option was explicitly specified.
2959      */
2960     {
2961         struct fsxattr attr;
2962         result = ioctl(fd, FS_IOC_FSGETXATTR, &attr);
2963         if (result == 0) {
2964             attr.fsx_xflags |= FS_XFLAG_EXTSIZE;
2965             attr.fsx_extsize = file_opts->extent_size_hint;
2966             result = ioctl(fd, FS_IOC_FSSETXATTR, &attr);
2967         }
2968         if (result < 0 && file_opts->has_extent_size_hint &&
2969             file_opts->extent_size_hint)
2970         {
2971             warn_report("Failed to set extent size hint: %s",
2972                         strerror(errno));
2973         }
2974     }
2975 #endif
2976 
2977     /* Resize and potentially preallocate the file to the desired
2978      * final size */
2979     result = raw_regular_truncate(NULL, fd, file_opts->size,
2980                                   file_opts->preallocation, errp);
2981     if (result < 0) {
2982         goto out_unlock;
2983     }
2984 
2985 out_unlock:
2986     raw_apply_lock_bytes(NULL, fd, 0, 0, true, &local_err);
2987     if (local_err) {
2988         /* The above call should not fail, and if it does, that does
2989          * not mean the whole creation operation has failed.  So
2990          * report it the user for their convenience, but do not report
2991          * it to the caller. */
2992         warn_report_err(local_err);
2993     }
2994 
2995 out_close:
2996     if (qemu_close(fd) != 0 && result == 0) {
2997         result = -errno;
2998         error_setg_errno(errp, -result, "Could not close the new file");
2999     }
3000 out:
3001     return result;
3002 }
3003 
3004 static int coroutine_fn GRAPH_RDLOCK
3005 raw_co_create_opts(BlockDriver *drv, const char *filename,
3006                    QemuOpts *opts, Error **errp)
3007 {
3008     BlockdevCreateOptions options;
3009     int64_t total_size = 0;
3010     int64_t extent_size_hint = 0;
3011     bool has_extent_size_hint = false;
3012     bool nocow = false;
3013     PreallocMode prealloc;
3014     char *buf = NULL;
3015     Error *local_err = NULL;
3016 
3017     /* Skip file: protocol prefix */
3018     strstart(filename, "file:", &filename);
3019 
3020     /* Read out options */
3021     total_size = ROUND_UP(qemu_opt_get_size_del(opts, BLOCK_OPT_SIZE, 0),
3022                           BDRV_SECTOR_SIZE);
3023     if (qemu_opt_get(opts, BLOCK_OPT_EXTENT_SIZE_HINT)) {
3024         has_extent_size_hint = true;
3025         extent_size_hint =
3026             qemu_opt_get_size_del(opts, BLOCK_OPT_EXTENT_SIZE_HINT, -1);
3027     }
3028     nocow = qemu_opt_get_bool(opts, BLOCK_OPT_NOCOW, false);
3029     buf = qemu_opt_get_del(opts, BLOCK_OPT_PREALLOC);
3030     prealloc = qapi_enum_parse(&PreallocMode_lookup, buf,
3031                                PREALLOC_MODE_OFF, &local_err);
3032     g_free(buf);
3033     if (local_err) {
3034         error_propagate(errp, local_err);
3035         return -EINVAL;
3036     }
3037 
3038     options = (BlockdevCreateOptions) {
3039         .driver     = BLOCKDEV_DRIVER_FILE,
3040         .u.file     = {
3041             .filename           = (char *) filename,
3042             .size               = total_size,
3043             .has_preallocation  = true,
3044             .preallocation      = prealloc,
3045             .has_nocow          = true,
3046             .nocow              = nocow,
3047             .has_extent_size_hint = has_extent_size_hint,
3048             .extent_size_hint   = extent_size_hint,
3049         },
3050     };
3051     return raw_co_create(&options, errp);
3052 }
3053 
3054 static int coroutine_fn raw_co_delete_file(BlockDriverState *bs,
3055                                            Error **errp)
3056 {
3057     struct stat st;
3058     int ret;
3059 
3060     if (!(stat(bs->filename, &st) == 0) || !S_ISREG(st.st_mode)) {
3061         error_setg_errno(errp, ENOENT, "%s is not a regular file",
3062                          bs->filename);
3063         return -ENOENT;
3064     }
3065 
3066     ret = unlink(bs->filename);
3067     if (ret < 0) {
3068         ret = -errno;
3069         error_setg_errno(errp, -ret, "Error when deleting file %s",
3070                          bs->filename);
3071     }
3072 
3073     return ret;
3074 }
3075 
3076 /*
3077  * Find allocation range in @bs around offset @start.
3078  * May change underlying file descriptor's file offset.
3079  * If @start is not in a hole, store @start in @data, and the
3080  * beginning of the next hole in @hole, and return 0.
3081  * If @start is in a non-trailing hole, store @start in @hole and the
3082  * beginning of the next non-hole in @data, and return 0.
3083  * If @start is in a trailing hole or beyond EOF, return -ENXIO.
3084  * If we can't find out, return a negative errno other than -ENXIO.
3085  */
3086 static int find_allocation(BlockDriverState *bs, off_t start,
3087                            off_t *data, off_t *hole)
3088 {
3089 #if defined SEEK_HOLE && defined SEEK_DATA
3090     BDRVRawState *s = bs->opaque;
3091     off_t offs;
3092 
3093     /*
3094      * SEEK_DATA cases:
3095      * D1. offs == start: start is in data
3096      * D2. offs > start: start is in a hole, next data at offs
3097      * D3. offs < 0, errno = ENXIO: either start is in a trailing hole
3098      *                              or start is beyond EOF
3099      *     If the latter happens, the file has been truncated behind
3100      *     our back since we opened it.  All bets are off then.
3101      *     Treating like a trailing hole is simplest.
3102      * D4. offs < 0, errno != ENXIO: we learned nothing
3103      */
3104     offs = lseek(s->fd, start, SEEK_DATA);
3105     if (offs < 0) {
3106         return -errno;          /* D3 or D4 */
3107     }
3108 
3109     if (offs < start) {
3110         /* This is not a valid return by lseek().  We are safe to just return
3111          * -EIO in this case, and we'll treat it like D4. */
3112         return -EIO;
3113     }
3114 
3115     if (offs > start) {
3116         /* D2: in hole, next data at offs */
3117         *hole = start;
3118         *data = offs;
3119         return 0;
3120     }
3121 
3122     /* D1: in data, end not yet known */
3123 
3124     /*
3125      * SEEK_HOLE cases:
3126      * H1. offs == start: start is in a hole
3127      *     If this happens here, a hole has been dug behind our back
3128      *     since the previous lseek().
3129      * H2. offs > start: either start is in data, next hole at offs,
3130      *                   or start is in trailing hole, EOF at offs
3131      *     Linux treats trailing holes like any other hole: offs ==
3132      *     start.  Solaris seeks to EOF instead: offs > start (blech).
3133      *     If that happens here, a hole has been dug behind our back
3134      *     since the previous lseek().
3135      * H3. offs < 0, errno = ENXIO: start is beyond EOF
3136      *     If this happens, the file has been truncated behind our
3137      *     back since we opened it.  Treat it like a trailing hole.
3138      * H4. offs < 0, errno != ENXIO: we learned nothing
3139      *     Pretend we know nothing at all, i.e. "forget" about D1.
3140      */
3141     offs = lseek(s->fd, start, SEEK_HOLE);
3142     if (offs < 0) {
3143         return -errno;          /* D1 and (H3 or H4) */
3144     }
3145 
3146     if (offs < start) {
3147         /* This is not a valid return by lseek().  We are safe to just return
3148          * -EIO in this case, and we'll treat it like H4. */
3149         return -EIO;
3150     }
3151 
3152     if (offs > start) {
3153         /*
3154          * D1 and H2: either in data, next hole at offs, or it was in
3155          * data but is now in a trailing hole.  In the latter case,
3156          * all bets are off.  Treating it as if it there was data all
3157          * the way to EOF is safe, so simply do that.
3158          */
3159         *data = start;
3160         *hole = offs;
3161         return 0;
3162     }
3163 
3164     /* D1 and H1 */
3165     return -EBUSY;
3166 #else
3167     return -ENOTSUP;
3168 #endif
3169 }
3170 
3171 /*
3172  * Returns the allocation status of the specified offset.
3173  *
3174  * The block layer guarantees 'offset' and 'bytes' are within bounds.
3175  *
3176  * 'pnum' is set to the number of bytes (including and immediately following
3177  * the specified offset) that are known to be in the same
3178  * allocated/unallocated state.
3179  *
3180  * 'bytes' is a soft cap for 'pnum'.  If the information is free, 'pnum' may
3181  * well exceed it.
3182  */
3183 static int coroutine_fn raw_co_block_status(BlockDriverState *bs,
3184                                             bool want_zero,
3185                                             int64_t offset,
3186                                             int64_t bytes, int64_t *pnum,
3187                                             int64_t *map,
3188                                             BlockDriverState **file)
3189 {
3190     off_t data = 0, hole = 0;
3191     int ret;
3192 
3193     assert(QEMU_IS_ALIGNED(offset | bytes, bs->bl.request_alignment));
3194 
3195     ret = fd_open(bs);
3196     if (ret < 0) {
3197         return ret;
3198     }
3199 
3200     if (!want_zero) {
3201         *pnum = bytes;
3202         *map = offset;
3203         *file = bs;
3204         return BDRV_BLOCK_DATA | BDRV_BLOCK_OFFSET_VALID;
3205     }
3206 
3207     ret = find_allocation(bs, offset, &data, &hole);
3208     if (ret == -ENXIO) {
3209         /* Trailing hole */
3210         *pnum = bytes;
3211         ret = BDRV_BLOCK_ZERO;
3212     } else if (ret < 0) {
3213         /* No info available, so pretend there are no holes */
3214         *pnum = bytes;
3215         ret = BDRV_BLOCK_DATA;
3216     } else if (data == offset) {
3217         /* On a data extent, compute bytes to the end of the extent,
3218          * possibly including a partial sector at EOF. */
3219         *pnum = hole - offset;
3220 
3221         /*
3222          * We are not allowed to return partial sectors, though, so
3223          * round up if necessary.
3224          */
3225         if (!QEMU_IS_ALIGNED(*pnum, bs->bl.request_alignment)) {
3226             int64_t file_length = raw_getlength(bs);
3227             if (file_length > 0) {
3228                 /* Ignore errors, this is just a safeguard */
3229                 assert(hole == file_length);
3230             }
3231             *pnum = ROUND_UP(*pnum, bs->bl.request_alignment);
3232         }
3233 
3234         ret = BDRV_BLOCK_DATA;
3235     } else {
3236         /* On a hole, compute bytes to the beginning of the next extent.  */
3237         assert(hole == offset);
3238         *pnum = data - offset;
3239         ret = BDRV_BLOCK_ZERO;
3240     }
3241     *map = offset;
3242     *file = bs;
3243     return ret | BDRV_BLOCK_OFFSET_VALID;
3244 }
3245 
3246 #if defined(__linux__)
3247 /* Verify that the file is not in the page cache */
3248 static void check_cache_dropped(BlockDriverState *bs, Error **errp)
3249 {
3250     const size_t window_size = 128 * 1024 * 1024;
3251     BDRVRawState *s = bs->opaque;
3252     void *window = NULL;
3253     size_t length = 0;
3254     unsigned char *vec;
3255     size_t page_size;
3256     off_t offset;
3257     off_t end;
3258 
3259     /* mincore(2) page status information requires 1 byte per page */
3260     page_size = sysconf(_SC_PAGESIZE);
3261     vec = g_malloc(DIV_ROUND_UP(window_size, page_size));
3262 
3263     end = raw_getlength(bs);
3264 
3265     for (offset = 0; offset < end; offset += window_size) {
3266         void *new_window;
3267         size_t new_length;
3268         size_t vec_end;
3269         size_t i;
3270         int ret;
3271 
3272         /* Unmap previous window if size has changed */
3273         new_length = MIN(end - offset, window_size);
3274         if (new_length != length) {
3275             munmap(window, length);
3276             window = NULL;
3277             length = 0;
3278         }
3279 
3280         new_window = mmap(window, new_length, PROT_NONE, MAP_PRIVATE,
3281                           s->fd, offset);
3282         if (new_window == MAP_FAILED) {
3283             error_setg_errno(errp, errno, "mmap failed");
3284             break;
3285         }
3286 
3287         window = new_window;
3288         length = new_length;
3289 
3290         ret = mincore(window, length, vec);
3291         if (ret < 0) {
3292             error_setg_errno(errp, errno, "mincore failed");
3293             break;
3294         }
3295 
3296         vec_end = DIV_ROUND_UP(length, page_size);
3297         for (i = 0; i < vec_end; i++) {
3298             if (vec[i] & 0x1) {
3299                 break;
3300             }
3301         }
3302         if (i < vec_end) {
3303             error_setg(errp, "page cache still in use!");
3304             break;
3305         }
3306     }
3307 
3308     if (window) {
3309         munmap(window, length);
3310     }
3311 
3312     g_free(vec);
3313 }
3314 #endif /* __linux__ */
3315 
3316 static void coroutine_fn GRAPH_RDLOCK
3317 raw_co_invalidate_cache(BlockDriverState *bs, Error **errp)
3318 {
3319     BDRVRawState *s = bs->opaque;
3320     int ret;
3321 
3322     ret = fd_open(bs);
3323     if (ret < 0) {
3324         error_setg_errno(errp, -ret, "The file descriptor is not open");
3325         return;
3326     }
3327 
3328     if (!s->drop_cache) {
3329         return;
3330     }
3331 
3332     if (s->open_flags & O_DIRECT) {
3333         return; /* No host kernel page cache */
3334     }
3335 
3336 #if defined(__linux__)
3337     /* This sets the scene for the next syscall... */
3338     ret = bdrv_co_flush(bs);
3339     if (ret < 0) {
3340         error_setg_errno(errp, -ret, "flush failed");
3341         return;
3342     }
3343 
3344     /* Linux does not invalidate pages that are dirty, locked, or mmapped by a
3345      * process.  These limitations are okay because we just fsynced the file,
3346      * we don't use mmap, and the file should not be in use by other processes.
3347      */
3348     ret = posix_fadvise(s->fd, 0, 0, POSIX_FADV_DONTNEED);
3349     if (ret != 0) { /* the return value is a positive errno */
3350         error_setg_errno(errp, ret, "fadvise failed");
3351         return;
3352     }
3353 
3354     if (s->check_cache_dropped) {
3355         check_cache_dropped(bs, errp);
3356     }
3357 #else /* __linux__ */
3358     /* Do nothing.  Live migration to a remote host with cache.direct=off is
3359      * unsupported on other host operating systems.  Cache consistency issues
3360      * may occur but no error is reported here, partly because that's the
3361      * historical behavior and partly because it's hard to differentiate valid
3362      * configurations that should not cause errors.
3363      */
3364 #endif /* !__linux__ */
3365 }
3366 
3367 static void raw_account_discard(BDRVRawState *s, uint64_t nbytes, int ret)
3368 {
3369     if (ret) {
3370         s->stats.discard_nb_failed++;
3371     } else {
3372         s->stats.discard_nb_ok++;
3373         s->stats.discard_bytes_ok += nbytes;
3374     }
3375 }
3376 
3377 /*
3378  * zone report - Get a zone block device's information in the form
3379  * of an array of zone descriptors.
3380  * zones is an array of zone descriptors to hold zone information on reply;
3381  * offset can be any byte within the entire size of the device;
3382  * nr_zones is the maxium number of sectors the command should operate on.
3383  */
3384 #if defined(CONFIG_BLKZONED)
3385 static int coroutine_fn raw_co_zone_report(BlockDriverState *bs, int64_t offset,
3386                                            unsigned int *nr_zones,
3387                                            BlockZoneDescriptor *zones) {
3388     BDRVRawState *s = bs->opaque;
3389     RawPosixAIOData acb = (RawPosixAIOData) {
3390         .bs         = bs,
3391         .aio_fildes = s->fd,
3392         .aio_type   = QEMU_AIO_ZONE_REPORT,
3393         .aio_offset = offset,
3394         .zone_report    = {
3395             .nr_zones       = nr_zones,
3396             .zones          = zones,
3397         },
3398     };
3399 
3400     trace_zbd_zone_report(bs, *nr_zones, offset >> BDRV_SECTOR_BITS);
3401     return raw_thread_pool_submit(handle_aiocb_zone_report, &acb);
3402 }
3403 #endif
3404 
3405 /*
3406  * zone management operations - Execute an operation on a zone
3407  */
3408 #if defined(CONFIG_BLKZONED)
3409 static int coroutine_fn raw_co_zone_mgmt(BlockDriverState *bs, BlockZoneOp op,
3410         int64_t offset, int64_t len) {
3411     BDRVRawState *s = bs->opaque;
3412     RawPosixAIOData acb;
3413     int64_t zone_size, zone_size_mask;
3414     const char *op_name;
3415     unsigned long zo;
3416     int ret;
3417     BlockZoneWps *wps = bs->wps;
3418     int64_t capacity = bs->total_sectors << BDRV_SECTOR_BITS;
3419 
3420     zone_size = bs->bl.zone_size;
3421     zone_size_mask = zone_size - 1;
3422     if (offset & zone_size_mask) {
3423         error_report("sector offset %" PRId64 " is not aligned to zone size "
3424                      "%" PRId64 "", offset / 512, zone_size / 512);
3425         return -EINVAL;
3426     }
3427 
3428     if (((offset + len) < capacity && len & zone_size_mask) ||
3429         offset + len > capacity) {
3430         error_report("number of sectors %" PRId64 " is not aligned to zone size"
3431                       " %" PRId64 "", len / 512, zone_size / 512);
3432         return -EINVAL;
3433     }
3434 
3435     uint32_t i = offset / bs->bl.zone_size;
3436     uint32_t nrz = len / bs->bl.zone_size;
3437     uint64_t *wp = &wps->wp[i];
3438     if (BDRV_ZT_IS_CONV(*wp) && len != capacity) {
3439         error_report("zone mgmt operations are not allowed for conventional zones");
3440         return -EIO;
3441     }
3442 
3443     switch (op) {
3444     case BLK_ZO_OPEN:
3445         op_name = "BLKOPENZONE";
3446         zo = BLKOPENZONE;
3447         break;
3448     case BLK_ZO_CLOSE:
3449         op_name = "BLKCLOSEZONE";
3450         zo = BLKCLOSEZONE;
3451         break;
3452     case BLK_ZO_FINISH:
3453         op_name = "BLKFINISHZONE";
3454         zo = BLKFINISHZONE;
3455         break;
3456     case BLK_ZO_RESET:
3457         op_name = "BLKRESETZONE";
3458         zo = BLKRESETZONE;
3459         break;
3460     default:
3461         error_report("Unsupported zone op: 0x%x", op);
3462         return -ENOTSUP;
3463     }
3464 
3465     acb = (RawPosixAIOData) {
3466         .bs             = bs,
3467         .aio_fildes     = s->fd,
3468         .aio_type       = QEMU_AIO_ZONE_MGMT,
3469         .aio_offset     = offset,
3470         .aio_nbytes     = len,
3471         .zone_mgmt  = {
3472             .op = zo,
3473         },
3474     };
3475 
3476     trace_zbd_zone_mgmt(bs, op_name, offset >> BDRV_SECTOR_BITS,
3477                         len >> BDRV_SECTOR_BITS);
3478     ret = raw_thread_pool_submit(handle_aiocb_zone_mgmt, &acb);
3479     if (ret != 0) {
3480         update_zones_wp(bs, s->fd, offset, i);
3481         error_report("ioctl %s failed %d", op_name, ret);
3482         return ret;
3483     }
3484 
3485     if (zo == BLKRESETZONE && len == capacity) {
3486         ret = get_zones_wp(bs, s->fd, 0, bs->bl.nr_zones, 1);
3487         if (ret < 0) {
3488             error_report("reporting single wp failed");
3489             return ret;
3490         }
3491     } else if (zo == BLKRESETZONE) {
3492         for (unsigned int j = 0; j < nrz; ++j) {
3493             wp[j] = offset + j * zone_size;
3494         }
3495     } else if (zo == BLKFINISHZONE) {
3496         for (unsigned int j = 0; j < nrz; ++j) {
3497             /* The zoned device allows the last zone smaller that the
3498              * zone size. */
3499             wp[j] = MIN(offset + (j + 1) * zone_size, offset + len);
3500         }
3501     }
3502 
3503     return ret;
3504 }
3505 #endif
3506 
3507 #if defined(CONFIG_BLKZONED)
3508 static int coroutine_fn raw_co_zone_append(BlockDriverState *bs,
3509                                            int64_t *offset,
3510                                            QEMUIOVector *qiov,
3511                                            BdrvRequestFlags flags) {
3512     assert(flags == 0);
3513     int64_t zone_size_mask = bs->bl.zone_size - 1;
3514     int64_t iov_len = 0;
3515     int64_t len = 0;
3516     BDRVRawState *s = bs->opaque;
3517     s->offset = offset;
3518 
3519     if (*offset & zone_size_mask) {
3520         error_report("sector offset %" PRId64 " is not aligned to zone size "
3521                      "%" PRId32 "", *offset / 512, bs->bl.zone_size / 512);
3522         return -EINVAL;
3523     }
3524 
3525     int64_t wg = bs->bl.write_granularity;
3526     int64_t wg_mask = wg - 1;
3527     for (int i = 0; i < qiov->niov; i++) {
3528         iov_len = qiov->iov[i].iov_len;
3529         if (iov_len & wg_mask) {
3530             error_report("len of IOVector[%d] %" PRId64 " is not aligned to "
3531                          "block size %" PRId64 "", i, iov_len, wg);
3532             return -EINVAL;
3533         }
3534         len += iov_len;
3535     }
3536 
3537     trace_zbd_zone_append(bs, *offset >> BDRV_SECTOR_BITS);
3538     return raw_co_prw(bs, *offset, len, qiov, QEMU_AIO_ZONE_APPEND);
3539 }
3540 #endif
3541 
3542 static coroutine_fn int
3543 raw_do_pdiscard(BlockDriverState *bs, int64_t offset, int64_t bytes,
3544                 bool blkdev)
3545 {
3546     BDRVRawState *s = bs->opaque;
3547     RawPosixAIOData acb;
3548     int ret;
3549 
3550     acb = (RawPosixAIOData) {
3551         .bs             = bs,
3552         .aio_fildes     = s->fd,
3553         .aio_type       = QEMU_AIO_DISCARD,
3554         .aio_offset     = offset,
3555         .aio_nbytes     = bytes,
3556     };
3557 
3558     if (blkdev) {
3559         acb.aio_type |= QEMU_AIO_BLKDEV;
3560     }
3561 
3562     ret = raw_thread_pool_submit(handle_aiocb_discard, &acb);
3563     raw_account_discard(s, bytes, ret);
3564     return ret;
3565 }
3566 
3567 static coroutine_fn int
3568 raw_co_pdiscard(BlockDriverState *bs, int64_t offset, int64_t bytes)
3569 {
3570     return raw_do_pdiscard(bs, offset, bytes, false);
3571 }
3572 
3573 static int coroutine_fn
3574 raw_do_pwrite_zeroes(BlockDriverState *bs, int64_t offset, int64_t bytes,
3575                      BdrvRequestFlags flags, bool blkdev)
3576 {
3577     BDRVRawState *s = bs->opaque;
3578     RawPosixAIOData acb;
3579     ThreadPoolFunc *handler;
3580 
3581 #ifdef CONFIG_FALLOCATE
3582     if (offset + bytes > bs->total_sectors * BDRV_SECTOR_SIZE) {
3583         BdrvTrackedRequest *req;
3584 
3585         /*
3586          * This is a workaround for a bug in the Linux XFS driver,
3587          * where writes submitted through the AIO interface will be
3588          * discarded if they happen beyond a concurrently running
3589          * fallocate() that increases the file length (i.e., both the
3590          * write and the fallocate() happen beyond the EOF).
3591          *
3592          * To work around it, we extend the tracked request for this
3593          * zero write until INT64_MAX (effectively infinity), and mark
3594          * it as serializing.
3595          *
3596          * We have to enable this workaround for all filesystems and
3597          * AIO modes (not just XFS with aio=native), because for
3598          * remote filesystems we do not know the host configuration.
3599          */
3600 
3601         req = bdrv_co_get_self_request(bs);
3602         assert(req);
3603         assert(req->type == BDRV_TRACKED_WRITE);
3604         assert(req->offset <= offset);
3605         assert(req->offset + req->bytes >= offset + bytes);
3606 
3607         req->bytes = BDRV_MAX_LENGTH - req->offset;
3608 
3609         bdrv_check_request(req->offset, req->bytes, &error_abort);
3610 
3611         bdrv_make_request_serialising(req, bs->bl.request_alignment);
3612     }
3613 #endif
3614 
3615     acb = (RawPosixAIOData) {
3616         .bs             = bs,
3617         .aio_fildes     = s->fd,
3618         .aio_type       = QEMU_AIO_WRITE_ZEROES,
3619         .aio_offset     = offset,
3620         .aio_nbytes     = bytes,
3621     };
3622 
3623     if (blkdev) {
3624         acb.aio_type |= QEMU_AIO_BLKDEV;
3625     }
3626     if (flags & BDRV_REQ_NO_FALLBACK) {
3627         acb.aio_type |= QEMU_AIO_NO_FALLBACK;
3628     }
3629 
3630     if (flags & BDRV_REQ_MAY_UNMAP) {
3631         acb.aio_type |= QEMU_AIO_DISCARD;
3632         handler = handle_aiocb_write_zeroes_unmap;
3633     } else {
3634         handler = handle_aiocb_write_zeroes;
3635     }
3636 
3637     return raw_thread_pool_submit(handler, &acb);
3638 }
3639 
3640 static int coroutine_fn raw_co_pwrite_zeroes(
3641     BlockDriverState *bs, int64_t offset,
3642     int64_t bytes, BdrvRequestFlags flags)
3643 {
3644     return raw_do_pwrite_zeroes(bs, offset, bytes, flags, false);
3645 }
3646 
3647 static int coroutine_fn
3648 raw_co_get_info(BlockDriverState *bs, BlockDriverInfo *bdi)
3649 {
3650     return 0;
3651 }
3652 
3653 static ImageInfoSpecific *raw_get_specific_info(BlockDriverState *bs,
3654                                                 Error **errp)
3655 {
3656     ImageInfoSpecificFile *file_info = g_new0(ImageInfoSpecificFile, 1);
3657     ImageInfoSpecific *spec_info = g_new(ImageInfoSpecific, 1);
3658 
3659     *spec_info = (ImageInfoSpecific){
3660         .type = IMAGE_INFO_SPECIFIC_KIND_FILE,
3661         .u.file.data = file_info,
3662     };
3663 
3664 #ifdef FS_IOC_FSGETXATTR
3665     {
3666         BDRVRawState *s = bs->opaque;
3667         struct fsxattr attr;
3668         int ret;
3669 
3670         ret = ioctl(s->fd, FS_IOC_FSGETXATTR, &attr);
3671         if (!ret && attr.fsx_extsize != 0) {
3672             file_info->has_extent_size_hint = true;
3673             file_info->extent_size_hint = attr.fsx_extsize;
3674         }
3675     }
3676 #endif
3677 
3678     return spec_info;
3679 }
3680 
3681 static BlockStatsSpecificFile get_blockstats_specific_file(BlockDriverState *bs)
3682 {
3683     BDRVRawState *s = bs->opaque;
3684     return (BlockStatsSpecificFile) {
3685         .discard_nb_ok = s->stats.discard_nb_ok,
3686         .discard_nb_failed = s->stats.discard_nb_failed,
3687         .discard_bytes_ok = s->stats.discard_bytes_ok,
3688     };
3689 }
3690 
3691 static BlockStatsSpecific *raw_get_specific_stats(BlockDriverState *bs)
3692 {
3693     BlockStatsSpecific *stats = g_new(BlockStatsSpecific, 1);
3694 
3695     stats->driver = BLOCKDEV_DRIVER_FILE;
3696     stats->u.file = get_blockstats_specific_file(bs);
3697 
3698     return stats;
3699 }
3700 
3701 #if defined(HAVE_HOST_BLOCK_DEVICE)
3702 static BlockStatsSpecific *hdev_get_specific_stats(BlockDriverState *bs)
3703 {
3704     BlockStatsSpecific *stats = g_new(BlockStatsSpecific, 1);
3705 
3706     stats->driver = BLOCKDEV_DRIVER_HOST_DEVICE;
3707     stats->u.host_device = get_blockstats_specific_file(bs);
3708 
3709     return stats;
3710 }
3711 #endif /* HAVE_HOST_BLOCK_DEVICE */
3712 
3713 static QemuOptsList raw_create_opts = {
3714     .name = "raw-create-opts",
3715     .head = QTAILQ_HEAD_INITIALIZER(raw_create_opts.head),
3716     .desc = {
3717         {
3718             .name = BLOCK_OPT_SIZE,
3719             .type = QEMU_OPT_SIZE,
3720             .help = "Virtual disk size"
3721         },
3722         {
3723             .name = BLOCK_OPT_NOCOW,
3724             .type = QEMU_OPT_BOOL,
3725             .help = "Turn off copy-on-write (valid only on btrfs)"
3726         },
3727         {
3728             .name = BLOCK_OPT_PREALLOC,
3729             .type = QEMU_OPT_STRING,
3730             .help = "Preallocation mode (allowed values: off"
3731 #ifdef CONFIG_POSIX_FALLOCATE
3732                     ", falloc"
3733 #endif
3734                     ", full)"
3735         },
3736         {
3737             .name = BLOCK_OPT_EXTENT_SIZE_HINT,
3738             .type = QEMU_OPT_SIZE,
3739             .help = "Extent size hint for the image file, 0 to disable"
3740         },
3741         { /* end of list */ }
3742     }
3743 };
3744 
3745 static int raw_check_perm(BlockDriverState *bs, uint64_t perm, uint64_t shared,
3746                           Error **errp)
3747 {
3748     BDRVRawState *s = bs->opaque;
3749     int input_flags = s->reopen_state ? s->reopen_state->flags : bs->open_flags;
3750     int open_flags;
3751     int ret;
3752 
3753     /* We may need a new fd if auto-read-only switches the mode */
3754     ret = raw_reconfigure_getfd(bs, input_flags, &open_flags, perm,
3755                                 false, errp);
3756     if (ret < 0) {
3757         return ret;
3758     } else if (ret != s->fd) {
3759         Error *local_err = NULL;
3760 
3761         /*
3762          * Fail already check_perm() if we can't get a working O_DIRECT
3763          * alignment with the new fd.
3764          */
3765         raw_probe_alignment(bs, ret, &local_err);
3766         if (local_err) {
3767             error_propagate(errp, local_err);
3768             return -EINVAL;
3769         }
3770 
3771         s->perm_change_fd = ret;
3772         s->perm_change_flags = open_flags;
3773     }
3774 
3775     /* Prepare permissions on old fd to avoid conflicts between old and new,
3776      * but keep everything locked that new will need. */
3777     ret = raw_handle_perm_lock(bs, RAW_PL_PREPARE, perm, shared, errp);
3778     if (ret < 0) {
3779         goto fail;
3780     }
3781 
3782     /* Copy locks to the new fd */
3783     if (s->perm_change_fd && s->use_lock) {
3784         ret = raw_apply_lock_bytes(NULL, s->perm_change_fd, perm, ~shared,
3785                                    false, errp);
3786         if (ret < 0) {
3787             raw_handle_perm_lock(bs, RAW_PL_ABORT, 0, 0, NULL);
3788             goto fail;
3789         }
3790     }
3791     return 0;
3792 
3793 fail:
3794     if (s->perm_change_fd) {
3795         qemu_close(s->perm_change_fd);
3796     }
3797     s->perm_change_fd = 0;
3798     return ret;
3799 }
3800 
3801 static void raw_set_perm(BlockDriverState *bs, uint64_t perm, uint64_t shared)
3802 {
3803     BDRVRawState *s = bs->opaque;
3804 
3805     /* For reopen, we have already switched to the new fd (.bdrv_set_perm is
3806      * called after .bdrv_reopen_commit) */
3807     if (s->perm_change_fd && s->fd != s->perm_change_fd) {
3808         qemu_close(s->fd);
3809         s->fd = s->perm_change_fd;
3810         s->open_flags = s->perm_change_flags;
3811     }
3812     s->perm_change_fd = 0;
3813 
3814     raw_handle_perm_lock(bs, RAW_PL_COMMIT, perm, shared, NULL);
3815     s->perm = perm;
3816     s->shared_perm = shared;
3817 }
3818 
3819 static void raw_abort_perm_update(BlockDriverState *bs)
3820 {
3821     BDRVRawState *s = bs->opaque;
3822 
3823     /* For reopen, .bdrv_reopen_abort is called afterwards and will close
3824      * the file descriptor. */
3825     if (s->perm_change_fd) {
3826         qemu_close(s->perm_change_fd);
3827     }
3828     s->perm_change_fd = 0;
3829 
3830     raw_handle_perm_lock(bs, RAW_PL_ABORT, 0, 0, NULL);
3831 }
3832 
3833 static int coroutine_fn GRAPH_RDLOCK raw_co_copy_range_from(
3834         BlockDriverState *bs, BdrvChild *src, int64_t src_offset,
3835         BdrvChild *dst, int64_t dst_offset, int64_t bytes,
3836         BdrvRequestFlags read_flags, BdrvRequestFlags write_flags)
3837 {
3838     return bdrv_co_copy_range_to(src, src_offset, dst, dst_offset, bytes,
3839                                  read_flags, write_flags);
3840 }
3841 
3842 static int coroutine_fn GRAPH_RDLOCK
3843 raw_co_copy_range_to(BlockDriverState *bs,
3844                      BdrvChild *src, int64_t src_offset,
3845                      BdrvChild *dst, int64_t dst_offset,
3846                      int64_t bytes, BdrvRequestFlags read_flags,
3847                      BdrvRequestFlags write_flags)
3848 {
3849     RawPosixAIOData acb;
3850     BDRVRawState *s = bs->opaque;
3851     BDRVRawState *src_s;
3852 
3853     assert(dst->bs == bs);
3854     if (src->bs->drv->bdrv_co_copy_range_to != raw_co_copy_range_to) {
3855         return -ENOTSUP;
3856     }
3857 
3858     src_s = src->bs->opaque;
3859     if (fd_open(src->bs) < 0 || fd_open(dst->bs) < 0) {
3860         return -EIO;
3861     }
3862 
3863     acb = (RawPosixAIOData) {
3864         .bs             = bs,
3865         .aio_type       = QEMU_AIO_COPY_RANGE,
3866         .aio_fildes     = src_s->fd,
3867         .aio_offset     = src_offset,
3868         .aio_nbytes     = bytes,
3869         .copy_range     = {
3870             .aio_fd2        = s->fd,
3871             .aio_offset2    = dst_offset,
3872         },
3873     };
3874 
3875     return raw_thread_pool_submit(handle_aiocb_copy_range, &acb);
3876 }
3877 
3878 BlockDriver bdrv_file = {
3879     .format_name = "file",
3880     .protocol_name = "file",
3881     .instance_size = sizeof(BDRVRawState),
3882     .bdrv_needs_filename = true,
3883     .bdrv_probe = NULL, /* no probe for protocols */
3884     .bdrv_parse_filename = raw_parse_filename,
3885     .bdrv_file_open = raw_open,
3886     .bdrv_reopen_prepare = raw_reopen_prepare,
3887     .bdrv_reopen_commit = raw_reopen_commit,
3888     .bdrv_reopen_abort = raw_reopen_abort,
3889     .bdrv_close = raw_close,
3890     .bdrv_co_create = raw_co_create,
3891     .bdrv_co_create_opts = raw_co_create_opts,
3892     .bdrv_has_zero_init = bdrv_has_zero_init_1,
3893     .bdrv_co_block_status = raw_co_block_status,
3894     .bdrv_co_invalidate_cache = raw_co_invalidate_cache,
3895     .bdrv_co_pwrite_zeroes = raw_co_pwrite_zeroes,
3896     .bdrv_co_delete_file = raw_co_delete_file,
3897 
3898     .bdrv_co_preadv         = raw_co_preadv,
3899     .bdrv_co_pwritev        = raw_co_pwritev,
3900     .bdrv_co_flush_to_disk  = raw_co_flush_to_disk,
3901     .bdrv_co_pdiscard       = raw_co_pdiscard,
3902     .bdrv_co_copy_range_from = raw_co_copy_range_from,
3903     .bdrv_co_copy_range_to  = raw_co_copy_range_to,
3904     .bdrv_refresh_limits = raw_refresh_limits,
3905     .bdrv_attach_aio_context = raw_aio_attach_aio_context,
3906 
3907     .bdrv_co_truncate                   = raw_co_truncate,
3908     .bdrv_co_getlength                  = raw_co_getlength,
3909     .bdrv_co_get_info                   = raw_co_get_info,
3910     .bdrv_get_specific_info             = raw_get_specific_info,
3911     .bdrv_co_get_allocated_file_size    = raw_co_get_allocated_file_size,
3912     .bdrv_get_specific_stats = raw_get_specific_stats,
3913     .bdrv_check_perm = raw_check_perm,
3914     .bdrv_set_perm   = raw_set_perm,
3915     .bdrv_abort_perm_update = raw_abort_perm_update,
3916     .create_opts = &raw_create_opts,
3917     .mutable_opts = mutable_opts,
3918 };
3919 
3920 /***********************************************/
3921 /* host device */
3922 
3923 #if defined(HAVE_HOST_BLOCK_DEVICE)
3924 
3925 #if defined(__APPLE__) && defined(__MACH__)
3926 static kern_return_t GetBSDPath(io_iterator_t mediaIterator, char *bsdPath,
3927                                 CFIndex maxPathSize, int flags);
3928 
3929 #if !defined(MAC_OS_VERSION_12_0) \
3930     || (MAC_OS_X_VERSION_MIN_REQUIRED < MAC_OS_VERSION_12_0)
3931 #define IOMainPort IOMasterPort
3932 #endif
3933 
3934 static char *FindEjectableOpticalMedia(io_iterator_t *mediaIterator)
3935 {
3936     kern_return_t kernResult = KERN_FAILURE;
3937     mach_port_t mainPort;
3938     CFMutableDictionaryRef  classesToMatch;
3939     const char *matching_array[] = {kIODVDMediaClass, kIOCDMediaClass};
3940     char *mediaType = NULL;
3941 
3942     kernResult = IOMainPort(MACH_PORT_NULL, &mainPort);
3943     if ( KERN_SUCCESS != kernResult ) {
3944         printf("IOMainPort returned %d\n", kernResult);
3945     }
3946 
3947     int index;
3948     for (index = 0; index < ARRAY_SIZE(matching_array); index++) {
3949         classesToMatch = IOServiceMatching(matching_array[index]);
3950         if (classesToMatch == NULL) {
3951             error_report("IOServiceMatching returned NULL for %s",
3952                          matching_array[index]);
3953             continue;
3954         }
3955         CFDictionarySetValue(classesToMatch, CFSTR(kIOMediaEjectableKey),
3956                              kCFBooleanTrue);
3957         kernResult = IOServiceGetMatchingServices(mainPort, classesToMatch,
3958                                                   mediaIterator);
3959         if (kernResult != KERN_SUCCESS) {
3960             error_report("Note: IOServiceGetMatchingServices returned %d",
3961                          kernResult);
3962             continue;
3963         }
3964 
3965         /* If a match was found, leave the loop */
3966         if (*mediaIterator != 0) {
3967             trace_file_FindEjectableOpticalMedia(matching_array[index]);
3968             mediaType = g_strdup(matching_array[index]);
3969             break;
3970         }
3971     }
3972     return mediaType;
3973 }
3974 
3975 kern_return_t GetBSDPath(io_iterator_t mediaIterator, char *bsdPath,
3976                          CFIndex maxPathSize, int flags)
3977 {
3978     io_object_t     nextMedia;
3979     kern_return_t   kernResult = KERN_FAILURE;
3980     *bsdPath = '\0';
3981     nextMedia = IOIteratorNext( mediaIterator );
3982     if ( nextMedia )
3983     {
3984         CFTypeRef   bsdPathAsCFString;
3985     bsdPathAsCFString = IORegistryEntryCreateCFProperty( nextMedia, CFSTR( kIOBSDNameKey ), kCFAllocatorDefault, 0 );
3986         if ( bsdPathAsCFString ) {
3987             size_t devPathLength;
3988             strcpy( bsdPath, _PATH_DEV );
3989             if (flags & BDRV_O_NOCACHE) {
3990                 strcat(bsdPath, "r");
3991             }
3992             devPathLength = strlen( bsdPath );
3993             if ( CFStringGetCString( bsdPathAsCFString, bsdPath + devPathLength, maxPathSize - devPathLength, kCFStringEncodingASCII ) ) {
3994                 kernResult = KERN_SUCCESS;
3995             }
3996             CFRelease( bsdPathAsCFString );
3997         }
3998         IOObjectRelease( nextMedia );
3999     }
4000 
4001     return kernResult;
4002 }
4003 
4004 /* Sets up a real cdrom for use in QEMU */
4005 static bool setup_cdrom(char *bsd_path, Error **errp)
4006 {
4007     int index, num_of_test_partitions = 2, fd;
4008     char test_partition[MAXPATHLEN];
4009     bool partition_found = false;
4010 
4011     /* look for a working partition */
4012     for (index = 0; index < num_of_test_partitions; index++) {
4013         snprintf(test_partition, sizeof(test_partition), "%ss%d", bsd_path,
4014                  index);
4015         fd = qemu_open(test_partition, O_RDONLY | O_BINARY | O_LARGEFILE, NULL);
4016         if (fd >= 0) {
4017             partition_found = true;
4018             qemu_close(fd);
4019             break;
4020         }
4021     }
4022 
4023     /* if a working partition on the device was not found */
4024     if (partition_found == false) {
4025         error_setg(errp, "Failed to find a working partition on disc");
4026     } else {
4027         trace_file_setup_cdrom(test_partition);
4028         pstrcpy(bsd_path, MAXPATHLEN, test_partition);
4029     }
4030     return partition_found;
4031 }
4032 
4033 /* Prints directions on mounting and unmounting a device */
4034 static void print_unmounting_directions(const char *file_name)
4035 {
4036     error_report("If device %s is mounted on the desktop, unmount"
4037                  " it first before using it in QEMU", file_name);
4038     error_report("Command to unmount device: diskutil unmountDisk %s",
4039                  file_name);
4040     error_report("Command to mount device: diskutil mountDisk %s", file_name);
4041 }
4042 
4043 #endif /* defined(__APPLE__) && defined(__MACH__) */
4044 
4045 static int hdev_probe_device(const char *filename)
4046 {
4047     struct stat st;
4048 
4049     /* allow a dedicated CD-ROM driver to match with a higher priority */
4050     if (strstart(filename, "/dev/cdrom", NULL))
4051         return 50;
4052 
4053     if (stat(filename, &st) >= 0 &&
4054             (S_ISCHR(st.st_mode) || S_ISBLK(st.st_mode))) {
4055         return 100;
4056     }
4057 
4058     return 0;
4059 }
4060 
4061 static void hdev_parse_filename(const char *filename, QDict *options,
4062                                 Error **errp)
4063 {
4064     bdrv_parse_filename_strip_prefix(filename, "host_device:", options);
4065 }
4066 
4067 static bool hdev_is_sg(BlockDriverState *bs)
4068 {
4069 
4070 #if defined(__linux__)
4071 
4072     BDRVRawState *s = bs->opaque;
4073     struct stat st;
4074     struct sg_scsi_id scsiid;
4075     int sg_version;
4076     int ret;
4077 
4078     if (stat(bs->filename, &st) < 0 || !S_ISCHR(st.st_mode)) {
4079         return false;
4080     }
4081 
4082     ret = ioctl(s->fd, SG_GET_VERSION_NUM, &sg_version);
4083     if (ret < 0) {
4084         return false;
4085     }
4086 
4087     ret = ioctl(s->fd, SG_GET_SCSI_ID, &scsiid);
4088     if (ret >= 0) {
4089         trace_file_hdev_is_sg(scsiid.scsi_type, sg_version);
4090         return true;
4091     }
4092 
4093 #endif
4094 
4095     return false;
4096 }
4097 
4098 static int hdev_open(BlockDriverState *bs, QDict *options, int flags,
4099                      Error **errp)
4100 {
4101     BDRVRawState *s = bs->opaque;
4102     int ret;
4103 
4104 #if defined(__APPLE__) && defined(__MACH__)
4105     /*
4106      * Caution: while qdict_get_str() is fine, getting non-string types
4107      * would require more care.  When @options come from -blockdev or
4108      * blockdev_add, its members are typed according to the QAPI
4109      * schema, but when they come from -drive, they're all QString.
4110      */
4111     const char *filename = qdict_get_str(options, "filename");
4112     char bsd_path[MAXPATHLEN] = "";
4113     bool error_occurred = false;
4114 
4115     /* If using a real cdrom */
4116     if (strcmp(filename, "/dev/cdrom") == 0) {
4117         char *mediaType = NULL;
4118         kern_return_t ret_val;
4119         io_iterator_t mediaIterator = 0;
4120 
4121         mediaType = FindEjectableOpticalMedia(&mediaIterator);
4122         if (mediaType == NULL) {
4123             error_setg(errp, "Please make sure your CD/DVD is in the optical"
4124                        " drive");
4125             error_occurred = true;
4126             goto hdev_open_Mac_error;
4127         }
4128 
4129         ret_val = GetBSDPath(mediaIterator, bsd_path, sizeof(bsd_path), flags);
4130         if (ret_val != KERN_SUCCESS) {
4131             error_setg(errp, "Could not get BSD path for optical drive");
4132             error_occurred = true;
4133             goto hdev_open_Mac_error;
4134         }
4135 
4136         /* If a real optical drive was not found */
4137         if (bsd_path[0] == '\0') {
4138             error_setg(errp, "Failed to obtain bsd path for optical drive");
4139             error_occurred = true;
4140             goto hdev_open_Mac_error;
4141         }
4142 
4143         /* If using a cdrom disc and finding a partition on the disc failed */
4144         if (strncmp(mediaType, kIOCDMediaClass, 9) == 0 &&
4145             setup_cdrom(bsd_path, errp) == false) {
4146             print_unmounting_directions(bsd_path);
4147             error_occurred = true;
4148             goto hdev_open_Mac_error;
4149         }
4150 
4151         qdict_put_str(options, "filename", bsd_path);
4152 
4153 hdev_open_Mac_error:
4154         g_free(mediaType);
4155         if (mediaIterator) {
4156             IOObjectRelease(mediaIterator);
4157         }
4158         if (error_occurred) {
4159             return -ENOENT;
4160         }
4161     }
4162 #endif /* defined(__APPLE__) && defined(__MACH__) */
4163 
4164     s->type = FTYPE_FILE;
4165 
4166     ret = raw_open_common(bs, options, flags, 0, true, errp);
4167     if (ret < 0) {
4168 #if defined(__APPLE__) && defined(__MACH__)
4169         if (*bsd_path) {
4170             filename = bsd_path;
4171         }
4172         /* if a physical device experienced an error while being opened */
4173         if (strncmp(filename, "/dev/", 5) == 0) {
4174             print_unmounting_directions(filename);
4175         }
4176 #endif /* defined(__APPLE__) && defined(__MACH__) */
4177         return ret;
4178     }
4179 
4180     /* Since this does ioctl the device must be already opened */
4181     bs->sg = hdev_is_sg(bs);
4182 
4183     return ret;
4184 }
4185 
4186 #if defined(__linux__)
4187 static int coroutine_fn
4188 hdev_co_ioctl(BlockDriverState *bs, unsigned long int req, void *buf)
4189 {
4190     BDRVRawState *s = bs->opaque;
4191     RawPosixAIOData acb;
4192     int ret;
4193 
4194     ret = fd_open(bs);
4195     if (ret < 0) {
4196         return ret;
4197     }
4198 
4199     if (req == SG_IO && s->pr_mgr) {
4200         struct sg_io_hdr *io_hdr = buf;
4201         if (io_hdr->cmdp[0] == PERSISTENT_RESERVE_OUT ||
4202             io_hdr->cmdp[0] == PERSISTENT_RESERVE_IN) {
4203             return pr_manager_execute(s->pr_mgr, qemu_get_current_aio_context(),
4204                                       s->fd, io_hdr);
4205         }
4206     }
4207 
4208     acb = (RawPosixAIOData) {
4209         .bs         = bs,
4210         .aio_type   = QEMU_AIO_IOCTL,
4211         .aio_fildes = s->fd,
4212         .aio_offset = 0,
4213         .ioctl      = {
4214             .buf        = buf,
4215             .cmd        = req,
4216         },
4217     };
4218 
4219     return raw_thread_pool_submit(handle_aiocb_ioctl, &acb);
4220 }
4221 #endif /* linux */
4222 
4223 static coroutine_fn int
4224 hdev_co_pdiscard(BlockDriverState *bs, int64_t offset, int64_t bytes)
4225 {
4226     BDRVRawState *s = bs->opaque;
4227     int ret;
4228 
4229     ret = fd_open(bs);
4230     if (ret < 0) {
4231         raw_account_discard(s, bytes, ret);
4232         return ret;
4233     }
4234     return raw_do_pdiscard(bs, offset, bytes, true);
4235 }
4236 
4237 static coroutine_fn int hdev_co_pwrite_zeroes(BlockDriverState *bs,
4238     int64_t offset, int64_t bytes, BdrvRequestFlags flags)
4239 {
4240     int rc;
4241 
4242     rc = fd_open(bs);
4243     if (rc < 0) {
4244         return rc;
4245     }
4246 
4247     return raw_do_pwrite_zeroes(bs, offset, bytes, flags, true);
4248 }
4249 
4250 static BlockDriver bdrv_host_device = {
4251     .format_name        = "host_device",
4252     .protocol_name        = "host_device",
4253     .instance_size      = sizeof(BDRVRawState),
4254     .bdrv_needs_filename = true,
4255     .bdrv_probe_device  = hdev_probe_device,
4256     .bdrv_parse_filename = hdev_parse_filename,
4257     .bdrv_file_open     = hdev_open,
4258     .bdrv_close         = raw_close,
4259     .bdrv_reopen_prepare = raw_reopen_prepare,
4260     .bdrv_reopen_commit  = raw_reopen_commit,
4261     .bdrv_reopen_abort   = raw_reopen_abort,
4262     .bdrv_co_create_opts = bdrv_co_create_opts_simple,
4263     .create_opts         = &bdrv_create_opts_simple,
4264     .mutable_opts        = mutable_opts,
4265     .bdrv_co_invalidate_cache = raw_co_invalidate_cache,
4266     .bdrv_co_pwrite_zeroes = hdev_co_pwrite_zeroes,
4267 
4268     .bdrv_co_preadv         = raw_co_preadv,
4269     .bdrv_co_pwritev        = raw_co_pwritev,
4270     .bdrv_co_flush_to_disk  = raw_co_flush_to_disk,
4271     .bdrv_co_pdiscard       = hdev_co_pdiscard,
4272     .bdrv_co_copy_range_from = raw_co_copy_range_from,
4273     .bdrv_co_copy_range_to  = raw_co_copy_range_to,
4274     .bdrv_refresh_limits = raw_refresh_limits,
4275     .bdrv_attach_aio_context = raw_aio_attach_aio_context,
4276 
4277     .bdrv_co_truncate                   = raw_co_truncate,
4278     .bdrv_co_getlength                  = raw_co_getlength,
4279     .bdrv_co_get_info                   = raw_co_get_info,
4280     .bdrv_get_specific_info             = raw_get_specific_info,
4281     .bdrv_co_get_allocated_file_size    = raw_co_get_allocated_file_size,
4282     .bdrv_get_specific_stats = hdev_get_specific_stats,
4283     .bdrv_check_perm = raw_check_perm,
4284     .bdrv_set_perm   = raw_set_perm,
4285     .bdrv_abort_perm_update = raw_abort_perm_update,
4286     .bdrv_probe_blocksizes = hdev_probe_blocksizes,
4287     .bdrv_probe_geometry = hdev_probe_geometry,
4288 
4289     /* generic scsi device */
4290 #ifdef __linux__
4291     .bdrv_co_ioctl          = hdev_co_ioctl,
4292 #endif
4293 
4294     /* zoned device */
4295 #if defined(CONFIG_BLKZONED)
4296     /* zone management operations */
4297     .bdrv_co_zone_report = raw_co_zone_report,
4298     .bdrv_co_zone_mgmt = raw_co_zone_mgmt,
4299     .bdrv_co_zone_append = raw_co_zone_append,
4300 #endif
4301 };
4302 
4303 #if defined(__linux__) || defined(__FreeBSD__) || defined(__FreeBSD_kernel__)
4304 static void cdrom_parse_filename(const char *filename, QDict *options,
4305                                  Error **errp)
4306 {
4307     bdrv_parse_filename_strip_prefix(filename, "host_cdrom:", options);
4308 }
4309 
4310 static void cdrom_refresh_limits(BlockDriverState *bs, Error **errp)
4311 {
4312     bs->bl.has_variable_length = true;
4313     raw_refresh_limits(bs, errp);
4314 }
4315 #endif
4316 
4317 #ifdef __linux__
4318 static int cdrom_open(BlockDriverState *bs, QDict *options, int flags,
4319                       Error **errp)
4320 {
4321     BDRVRawState *s = bs->opaque;
4322 
4323     s->type = FTYPE_CD;
4324 
4325     /* open will not fail even if no CD is inserted, so add O_NONBLOCK */
4326     return raw_open_common(bs, options, flags, O_NONBLOCK, true, errp);
4327 }
4328 
4329 static int cdrom_probe_device(const char *filename)
4330 {
4331     int fd, ret;
4332     int prio = 0;
4333     struct stat st;
4334 
4335     fd = qemu_open(filename, O_RDONLY | O_NONBLOCK, NULL);
4336     if (fd < 0) {
4337         goto out;
4338     }
4339     ret = fstat(fd, &st);
4340     if (ret == -1 || !S_ISBLK(st.st_mode)) {
4341         goto outc;
4342     }
4343 
4344     /* Attempt to detect via a CDROM specific ioctl */
4345     ret = ioctl(fd, CDROM_DRIVE_STATUS, CDSL_CURRENT);
4346     if (ret >= 0)
4347         prio = 100;
4348 
4349 outc:
4350     qemu_close(fd);
4351 out:
4352     return prio;
4353 }
4354 
4355 static bool coroutine_fn cdrom_co_is_inserted(BlockDriverState *bs)
4356 {
4357     BDRVRawState *s = bs->opaque;
4358     int ret;
4359 
4360     ret = ioctl(s->fd, CDROM_DRIVE_STATUS, CDSL_CURRENT);
4361     return ret == CDS_DISC_OK;
4362 }
4363 
4364 static void coroutine_fn cdrom_co_eject(BlockDriverState *bs, bool eject_flag)
4365 {
4366     BDRVRawState *s = bs->opaque;
4367 
4368     if (eject_flag) {
4369         if (ioctl(s->fd, CDROMEJECT, NULL) < 0)
4370             perror("CDROMEJECT");
4371     } else {
4372         if (ioctl(s->fd, CDROMCLOSETRAY, NULL) < 0)
4373             perror("CDROMEJECT");
4374     }
4375 }
4376 
4377 static void coroutine_fn cdrom_co_lock_medium(BlockDriverState *bs, bool locked)
4378 {
4379     BDRVRawState *s = bs->opaque;
4380 
4381     if (ioctl(s->fd, CDROM_LOCKDOOR, locked) < 0) {
4382         /*
4383          * Note: an error can happen if the distribution automatically
4384          * mounts the CD-ROM
4385          */
4386         /* perror("CDROM_LOCKDOOR"); */
4387     }
4388 }
4389 
4390 static BlockDriver bdrv_host_cdrom = {
4391     .format_name        = "host_cdrom",
4392     .protocol_name      = "host_cdrom",
4393     .instance_size      = sizeof(BDRVRawState),
4394     .bdrv_needs_filename = true,
4395     .bdrv_probe_device	= cdrom_probe_device,
4396     .bdrv_parse_filename = cdrom_parse_filename,
4397     .bdrv_file_open     = cdrom_open,
4398     .bdrv_close         = raw_close,
4399     .bdrv_reopen_prepare = raw_reopen_prepare,
4400     .bdrv_reopen_commit  = raw_reopen_commit,
4401     .bdrv_reopen_abort   = raw_reopen_abort,
4402     .bdrv_co_create_opts = bdrv_co_create_opts_simple,
4403     .create_opts         = &bdrv_create_opts_simple,
4404     .mutable_opts        = mutable_opts,
4405     .bdrv_co_invalidate_cache = raw_co_invalidate_cache,
4406 
4407     .bdrv_co_preadv         = raw_co_preadv,
4408     .bdrv_co_pwritev        = raw_co_pwritev,
4409     .bdrv_co_flush_to_disk  = raw_co_flush_to_disk,
4410     .bdrv_refresh_limits    = cdrom_refresh_limits,
4411     .bdrv_attach_aio_context = raw_aio_attach_aio_context,
4412 
4413     .bdrv_co_truncate                   = raw_co_truncate,
4414     .bdrv_co_getlength                  = raw_co_getlength,
4415     .bdrv_co_get_allocated_file_size    = raw_co_get_allocated_file_size,
4416 
4417     /* removable device support */
4418     .bdrv_co_is_inserted    = cdrom_co_is_inserted,
4419     .bdrv_co_eject          = cdrom_co_eject,
4420     .bdrv_co_lock_medium    = cdrom_co_lock_medium,
4421 
4422     /* generic scsi device */
4423     .bdrv_co_ioctl      = hdev_co_ioctl,
4424 };
4425 #endif /* __linux__ */
4426 
4427 #if defined (__FreeBSD__) || defined(__FreeBSD_kernel__)
4428 static int cdrom_open(BlockDriverState *bs, QDict *options, int flags,
4429                       Error **errp)
4430 {
4431     BDRVRawState *s = bs->opaque;
4432     int ret;
4433 
4434     s->type = FTYPE_CD;
4435 
4436     ret = raw_open_common(bs, options, flags, 0, true, errp);
4437     if (ret) {
4438         return ret;
4439     }
4440 
4441     /* make sure the door isn't locked at this time */
4442     ioctl(s->fd, CDIOCALLOW);
4443     return 0;
4444 }
4445 
4446 static int cdrom_probe_device(const char *filename)
4447 {
4448     if (strstart(filename, "/dev/cd", NULL) ||
4449             strstart(filename, "/dev/acd", NULL))
4450         return 100;
4451     return 0;
4452 }
4453 
4454 static int cdrom_reopen(BlockDriverState *bs)
4455 {
4456     BDRVRawState *s = bs->opaque;
4457     int fd;
4458 
4459     /*
4460      * Force reread of possibly changed/newly loaded disc,
4461      * FreeBSD seems to not notice sometimes...
4462      */
4463     if (s->fd >= 0)
4464         qemu_close(s->fd);
4465     fd = qemu_open(bs->filename, s->open_flags, NULL);
4466     if (fd < 0) {
4467         s->fd = -1;
4468         return -EIO;
4469     }
4470     s->fd = fd;
4471 
4472     /* make sure the door isn't locked at this time */
4473     ioctl(s->fd, CDIOCALLOW);
4474     return 0;
4475 }
4476 
4477 static bool coroutine_fn cdrom_co_is_inserted(BlockDriverState *bs)
4478 {
4479     return raw_getlength(bs) > 0;
4480 }
4481 
4482 static void coroutine_fn cdrom_co_eject(BlockDriverState *bs, bool eject_flag)
4483 {
4484     BDRVRawState *s = bs->opaque;
4485 
4486     if (s->fd < 0)
4487         return;
4488 
4489     (void) ioctl(s->fd, CDIOCALLOW);
4490 
4491     if (eject_flag) {
4492         if (ioctl(s->fd, CDIOCEJECT) < 0)
4493             perror("CDIOCEJECT");
4494     } else {
4495         if (ioctl(s->fd, CDIOCCLOSE) < 0)
4496             perror("CDIOCCLOSE");
4497     }
4498 
4499     cdrom_reopen(bs);
4500 }
4501 
4502 static void coroutine_fn cdrom_co_lock_medium(BlockDriverState *bs, bool locked)
4503 {
4504     BDRVRawState *s = bs->opaque;
4505 
4506     if (s->fd < 0)
4507         return;
4508     if (ioctl(s->fd, (locked ? CDIOCPREVENT : CDIOCALLOW)) < 0) {
4509         /*
4510          * Note: an error can happen if the distribution automatically
4511          * mounts the CD-ROM
4512          */
4513         /* perror("CDROM_LOCKDOOR"); */
4514     }
4515 }
4516 
4517 static BlockDriver bdrv_host_cdrom = {
4518     .format_name        = "host_cdrom",
4519     .protocol_name      = "host_cdrom",
4520     .instance_size      = sizeof(BDRVRawState),
4521     .bdrv_needs_filename = true,
4522     .bdrv_probe_device	= cdrom_probe_device,
4523     .bdrv_parse_filename = cdrom_parse_filename,
4524     .bdrv_file_open     = cdrom_open,
4525     .bdrv_close         = raw_close,
4526     .bdrv_reopen_prepare = raw_reopen_prepare,
4527     .bdrv_reopen_commit  = raw_reopen_commit,
4528     .bdrv_reopen_abort   = raw_reopen_abort,
4529     .bdrv_co_create_opts = bdrv_co_create_opts_simple,
4530     .create_opts         = &bdrv_create_opts_simple,
4531     .mutable_opts       = mutable_opts,
4532 
4533     .bdrv_co_preadv         = raw_co_preadv,
4534     .bdrv_co_pwritev        = raw_co_pwritev,
4535     .bdrv_co_flush_to_disk  = raw_co_flush_to_disk,
4536     .bdrv_refresh_limits    = cdrom_refresh_limits,
4537     .bdrv_attach_aio_context = raw_aio_attach_aio_context,
4538 
4539     .bdrv_co_truncate                   = raw_co_truncate,
4540     .bdrv_co_getlength                  = raw_co_getlength,
4541     .bdrv_co_get_allocated_file_size    = raw_co_get_allocated_file_size,
4542 
4543     /* removable device support */
4544     .bdrv_co_is_inserted     = cdrom_co_is_inserted,
4545     .bdrv_co_eject           = cdrom_co_eject,
4546     .bdrv_co_lock_medium     = cdrom_co_lock_medium,
4547 };
4548 #endif /* __FreeBSD__ */
4549 
4550 #endif /* HAVE_HOST_BLOCK_DEVICE */
4551 
4552 static void bdrv_file_init(void)
4553 {
4554     /*
4555      * Register all the drivers.  Note that order is important, the driver
4556      * registered last will get probed first.
4557      */
4558     bdrv_register(&bdrv_file);
4559 #if defined(HAVE_HOST_BLOCK_DEVICE)
4560     bdrv_register(&bdrv_host_device);
4561 #ifdef __linux__
4562     bdrv_register(&bdrv_host_cdrom);
4563 #endif
4564 #if defined(__FreeBSD__) || defined(__FreeBSD_kernel__)
4565     bdrv_register(&bdrv_host_cdrom);
4566 #endif
4567 #endif /* HAVE_HOST_BLOCK_DEVICE */
4568 }
4569 
4570 block_init(bdrv_file_init);
4571