xref: /openbmc/qemu/block/file-posix.c (revision 7c08eefc)
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 has_laio_fdsync:1;
163     bool use_linux_io_uring:1;
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 && !(s->open_flags & O_DIRECT)) {
717         error_setg(errp, "aio=native was specified, but it requires "
718                          "cache.direct=on, which was not specified.");
719         ret = -EINVAL;
720         goto fail;
721     }
722     if (s->use_linux_aio) {
723         s->has_laio_fdsync = laio_has_fdsync(s->fd);
724     }
725 #else
726     if (s->use_linux_aio) {
727         error_setg(errp, "aio=native was specified, but is not supported "
728                          "in this build.");
729         ret = -EINVAL;
730         goto fail;
731     }
732 #endif /* !defined(CONFIG_LINUX_AIO) */
733 
734 #ifndef CONFIG_LINUX_IO_URING
735     if (s->use_linux_io_uring) {
736         error_setg(errp, "aio=io_uring was specified, but is not supported "
737                          "in this build.");
738         ret = -EINVAL;
739         goto fail;
740     }
741 #endif /* !defined(CONFIG_LINUX_IO_URING) */
742 
743     s->has_discard = true;
744     s->has_write_zeroes = true;
745 
746     if (fstat(s->fd, &st) < 0) {
747         ret = -errno;
748         error_setg_errno(errp, errno, "Could not stat file");
749         goto fail;
750     }
751 
752     if (!device) {
753         if (!S_ISREG(st.st_mode)) {
754             error_setg(errp, "'%s' driver requires '%s' to be a regular file",
755                        bs->drv->format_name, bs->filename);
756             ret = -EINVAL;
757             goto fail;
758         } else {
759             s->has_fallocate = true;
760         }
761     } else {
762         if (!(S_ISCHR(st.st_mode) || S_ISBLK(st.st_mode))) {
763             error_setg(errp, "'%s' driver requires '%s' to be either "
764                        "a character or block device",
765                        bs->drv->format_name, bs->filename);
766             ret = -EINVAL;
767             goto fail;
768         }
769     }
770 #ifdef CONFIG_BLKZONED
771     /*
772      * The kernel page cache does not reliably work for writes to SWR zones
773      * of zoned block device because it can not guarantee the order of writes.
774      */
775     if ((bs->bl.zoned != BLK_Z_NONE) &&
776         (!(s->open_flags & O_DIRECT))) {
777         error_setg(errp, "The driver supports zoned devices, and it requires "
778                          "cache.direct=on, which was not specified.");
779         return -EINVAL; /* No host kernel page cache */
780     }
781 #endif
782 
783     if (S_ISBLK(st.st_mode)) {
784 #ifdef __linux__
785         /* On Linux 3.10, BLKDISCARD leaves stale data in the page cache.  Do
786          * not rely on the contents of discarded blocks unless using O_DIRECT.
787          * Same for BLKZEROOUT.
788          */
789         if (!(bs->open_flags & BDRV_O_NOCACHE)) {
790             s->has_write_zeroes = false;
791         }
792 #endif
793     }
794 #ifdef __FreeBSD__
795     if (S_ISCHR(st.st_mode)) {
796         /*
797          * The file is a char device (disk), which on FreeBSD isn't behind
798          * a pager, so force all requests to be aligned. This is needed
799          * so QEMU makes sure all IO operations on the device are aligned
800          * to sector size, or else FreeBSD will reject them with EINVAL.
801          */
802         s->force_alignment = true;
803     }
804 #endif
805     s->needs_alignment = raw_needs_alignment(bs);
806 
807     bs->supported_zero_flags = BDRV_REQ_MAY_UNMAP | BDRV_REQ_NO_FALLBACK;
808     if (S_ISREG(st.st_mode)) {
809         /* When extending regular files, we get zeros from the OS */
810         bs->supported_truncate_flags = BDRV_REQ_ZERO_WRITE;
811     }
812     ret = 0;
813 fail:
814     if (ret < 0 && s->fd != -1) {
815         qemu_close(s->fd);
816     }
817     if (filename && (bdrv_flags & BDRV_O_TEMPORARY)) {
818         unlink(filename);
819     }
820     qemu_opts_del(opts);
821     return ret;
822 }
823 
824 static int raw_open(BlockDriverState *bs, QDict *options, int flags,
825                     Error **errp)
826 {
827     BDRVRawState *s = bs->opaque;
828 
829     s->type = FTYPE_FILE;
830     return raw_open_common(bs, options, flags, 0, false, errp);
831 }
832 
833 typedef enum {
834     RAW_PL_PREPARE,
835     RAW_PL_COMMIT,
836     RAW_PL_ABORT,
837 } RawPermLockOp;
838 
839 #define PERM_FOREACH(i) \
840     for ((i) = 0; (1ULL << (i)) <= BLK_PERM_ALL; i++)
841 
842 /* Lock bytes indicated by @perm_lock_bits and @shared_perm_lock_bits in the
843  * file; if @unlock == true, also unlock the unneeded bytes.
844  * @shared_perm_lock_bits is the mask of all permissions that are NOT shared.
845  */
846 static int raw_apply_lock_bytes(BDRVRawState *s, int fd,
847                                 uint64_t perm_lock_bits,
848                                 uint64_t shared_perm_lock_bits,
849                                 bool unlock, Error **errp)
850 {
851     int ret;
852     int i;
853     uint64_t locked_perm, locked_shared_perm;
854 
855     if (s) {
856         locked_perm = s->locked_perm;
857         locked_shared_perm = s->locked_shared_perm;
858     } else {
859         /*
860          * We don't have the previous bits, just lock/unlock for each of the
861          * requested bits.
862          */
863         if (unlock) {
864             locked_perm = BLK_PERM_ALL;
865             locked_shared_perm = BLK_PERM_ALL;
866         } else {
867             locked_perm = 0;
868             locked_shared_perm = 0;
869         }
870     }
871 
872     PERM_FOREACH(i) {
873         int off = RAW_LOCK_PERM_BASE + i;
874         uint64_t bit = (1ULL << i);
875         if ((perm_lock_bits & bit) && !(locked_perm & bit)) {
876             ret = qemu_lock_fd(fd, off, 1, false);
877             if (ret) {
878                 raw_lock_error_setg_errno(errp, -ret, "Failed to lock byte %d",
879                                           off);
880                 return ret;
881             } else if (s) {
882                 s->locked_perm |= bit;
883             }
884         } else if (unlock && (locked_perm & bit) && !(perm_lock_bits & bit)) {
885             ret = qemu_unlock_fd(fd, off, 1);
886             if (ret) {
887                 error_setg_errno(errp, -ret, "Failed to unlock byte %d", off);
888                 return ret;
889             } else if (s) {
890                 s->locked_perm &= ~bit;
891             }
892         }
893     }
894     PERM_FOREACH(i) {
895         int off = RAW_LOCK_SHARED_BASE + i;
896         uint64_t bit = (1ULL << i);
897         if ((shared_perm_lock_bits & bit) && !(locked_shared_perm & bit)) {
898             ret = qemu_lock_fd(fd, off, 1, false);
899             if (ret) {
900                 raw_lock_error_setg_errno(errp, -ret, "Failed to lock byte %d",
901                                           off);
902                 return ret;
903             } else if (s) {
904                 s->locked_shared_perm |= bit;
905             }
906         } else if (unlock && (locked_shared_perm & bit) &&
907                    !(shared_perm_lock_bits & bit)) {
908             ret = qemu_unlock_fd(fd, off, 1);
909             if (ret) {
910                 error_setg_errno(errp, -ret, "Failed to unlock byte %d", off);
911                 return ret;
912             } else if (s) {
913                 s->locked_shared_perm &= ~bit;
914             }
915         }
916     }
917     return 0;
918 }
919 
920 /* Check "unshared" bytes implied by @perm and ~@shared_perm in the file. */
921 static int raw_check_lock_bytes(int fd, uint64_t perm, uint64_t shared_perm,
922                                 Error **errp)
923 {
924     int ret;
925     int i;
926 
927     PERM_FOREACH(i) {
928         int off = RAW_LOCK_SHARED_BASE + i;
929         uint64_t p = 1ULL << i;
930         if (perm & p) {
931             ret = qemu_lock_fd_test(fd, off, 1, true);
932             if (ret) {
933                 char *perm_name = bdrv_perm_names(p);
934 
935                 raw_lock_error_setg_errno(errp, -ret,
936                                           "Failed to get \"%s\" lock",
937                                           perm_name);
938                 g_free(perm_name);
939                 return ret;
940             }
941         }
942     }
943     PERM_FOREACH(i) {
944         int off = RAW_LOCK_PERM_BASE + i;
945         uint64_t p = 1ULL << i;
946         if (!(shared_perm & p)) {
947             ret = qemu_lock_fd_test(fd, off, 1, true);
948             if (ret) {
949                 char *perm_name = bdrv_perm_names(p);
950 
951                 raw_lock_error_setg_errno(errp, -ret,
952                                           "Failed to get shared \"%s\" lock",
953                                           perm_name);
954                 g_free(perm_name);
955                 return ret;
956             }
957         }
958     }
959     return 0;
960 }
961 
962 static int raw_handle_perm_lock(BlockDriverState *bs,
963                                 RawPermLockOp op,
964                                 uint64_t new_perm, uint64_t new_shared,
965                                 Error **errp)
966 {
967     BDRVRawState *s = bs->opaque;
968     int ret = 0;
969     Error *local_err = NULL;
970 
971     if (!s->use_lock) {
972         return 0;
973     }
974 
975     if (bdrv_get_flags(bs) & BDRV_O_INACTIVE) {
976         return 0;
977     }
978 
979     switch (op) {
980     case RAW_PL_PREPARE:
981         if ((s->perm | new_perm) == s->perm &&
982             (s->shared_perm & new_shared) == s->shared_perm)
983         {
984             /*
985              * We are going to unlock bytes, it should not fail. If it fail due
986              * to some fs-dependent permission-unrelated reasons (which occurs
987              * sometimes on NFS and leads to abort in bdrv_replace_child) we
988              * can't prevent such errors by any check here. And we ignore them
989              * anyway in ABORT and COMMIT.
990              */
991             return 0;
992         }
993         ret = raw_apply_lock_bytes(s, s->fd, s->perm | new_perm,
994                                    ~s->shared_perm | ~new_shared,
995                                    false, errp);
996         if (!ret) {
997             ret = raw_check_lock_bytes(s->fd, new_perm, new_shared, errp);
998             if (!ret) {
999                 return 0;
1000             }
1001             error_append_hint(errp,
1002                               "Is another process using the image [%s]?\n",
1003                               bs->filename);
1004         }
1005         /* fall through to unlock bytes. */
1006     case RAW_PL_ABORT:
1007         raw_apply_lock_bytes(s, s->fd, s->perm, ~s->shared_perm,
1008                              true, &local_err);
1009         if (local_err) {
1010             /* Theoretically the above call only unlocks bytes and it cannot
1011              * fail. Something weird happened, report it.
1012              */
1013             warn_report_err(local_err);
1014         }
1015         break;
1016     case RAW_PL_COMMIT:
1017         raw_apply_lock_bytes(s, s->fd, new_perm, ~new_shared,
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     }
1027     return ret;
1028 }
1029 
1030 /* Sets a specific flag */
1031 static int fcntl_setfl(int fd, int flag)
1032 {
1033     int flags;
1034 
1035     flags = fcntl(fd, F_GETFL);
1036     if (flags == -1) {
1037         return -errno;
1038     }
1039     if (fcntl(fd, F_SETFL, flags | flag) == -1) {
1040         return -errno;
1041     }
1042     return 0;
1043 }
1044 
1045 static int raw_reconfigure_getfd(BlockDriverState *bs, int flags,
1046                                  int *open_flags, uint64_t perm, Error **errp)
1047 {
1048     BDRVRawState *s = bs->opaque;
1049     int fd = -1;
1050     int ret;
1051     bool has_writers = perm &
1052         (BLK_PERM_WRITE | BLK_PERM_WRITE_UNCHANGED | BLK_PERM_RESIZE);
1053     int fcntl_flags = O_APPEND | O_NONBLOCK;
1054 #ifdef O_NOATIME
1055     fcntl_flags |= O_NOATIME;
1056 #endif
1057 
1058     *open_flags = 0;
1059     if (s->type == FTYPE_CD) {
1060         *open_flags |= O_NONBLOCK;
1061     }
1062 
1063     raw_parse_flags(flags, open_flags, has_writers);
1064 
1065 #ifdef O_ASYNC
1066     /* Not all operating systems have O_ASYNC, and those that don't
1067      * will not let us track the state into rs->open_flags (typically
1068      * you achieve the same effect with an ioctl, for example I_SETSIG
1069      * on Solaris). But we do not use O_ASYNC, so that's fine.
1070      */
1071     assert((s->open_flags & O_ASYNC) == 0);
1072 #endif
1073 
1074     if (*open_flags == s->open_flags) {
1075         /* We're lucky, the existing fd is fine */
1076         return s->fd;
1077     }
1078 
1079     if ((*open_flags & ~fcntl_flags) == (s->open_flags & ~fcntl_flags)) {
1080         /* dup the original fd */
1081         fd = qemu_dup(s->fd);
1082         if (fd >= 0) {
1083             ret = fcntl_setfl(fd, *open_flags);
1084             if (ret) {
1085                 qemu_close(fd);
1086                 fd = -1;
1087             }
1088         }
1089     }
1090 
1091     /* If we cannot use fcntl, or fcntl failed, fall back to qemu_open() */
1092     if (fd == -1) {
1093         const char *normalized_filename = bs->filename;
1094         ret = raw_normalize_devicepath(&normalized_filename, errp);
1095         if (ret >= 0) {
1096             fd = qemu_open(normalized_filename, *open_flags, errp);
1097             if (fd == -1) {
1098                 return -1;
1099             }
1100         }
1101     }
1102 
1103     if (fd != -1 && (*open_flags & O_RDWR)) {
1104         ret = check_hdev_writable(fd);
1105         if (ret < 0) {
1106             qemu_close(fd);
1107             error_setg_errno(errp, -ret, "The device is not writable");
1108             return -1;
1109         }
1110     }
1111 
1112     return fd;
1113 }
1114 
1115 static int raw_reopen_prepare(BDRVReopenState *state,
1116                               BlockReopenQueue *queue, Error **errp)
1117 {
1118     BDRVRawState *s;
1119     BDRVRawReopenState *rs;
1120     QemuOpts *opts;
1121     int ret;
1122 
1123     assert(state != NULL);
1124     assert(state->bs != NULL);
1125 
1126     s = state->bs->opaque;
1127 
1128     state->opaque = g_new0(BDRVRawReopenState, 1);
1129     rs = state->opaque;
1130 
1131     /* Handle options changes */
1132     opts = qemu_opts_create(&raw_runtime_opts, NULL, 0, &error_abort);
1133     if (!qemu_opts_absorb_qdict(opts, state->options, errp)) {
1134         ret = -EINVAL;
1135         goto out;
1136     }
1137 
1138     rs->drop_cache = qemu_opt_get_bool_del(opts, "drop-cache", true);
1139     rs->check_cache_dropped =
1140         qemu_opt_get_bool_del(opts, "x-check-cache-dropped", false);
1141 
1142     /* This driver's reopen function doesn't currently allow changing
1143      * other options, so let's put them back in the original QDict and
1144      * bdrv_reopen_prepare() will detect changes and complain. */
1145     qemu_opts_to_qdict(opts, state->options);
1146 
1147     /*
1148      * As part of reopen prepare we also want to create new fd by
1149      * raw_reconfigure_getfd(). But it wants updated "perm", when in
1150      * bdrv_reopen_multiple() .bdrv_reopen_prepare() callback called prior to
1151      * permission update. Happily, permission update is always a part
1152      * (a separate stage) of bdrv_reopen_multiple() so we can rely on this
1153      * fact and reconfigure fd in raw_check_perm().
1154      */
1155 
1156     s->reopen_state = state;
1157     ret = 0;
1158 
1159 out:
1160     qemu_opts_del(opts);
1161     return ret;
1162 }
1163 
1164 static void raw_reopen_commit(BDRVReopenState *state)
1165 {
1166     BDRVRawReopenState *rs = state->opaque;
1167     BDRVRawState *s = state->bs->opaque;
1168 
1169     s->drop_cache = rs->drop_cache;
1170     s->check_cache_dropped = rs->check_cache_dropped;
1171     s->open_flags = rs->open_flags;
1172     g_free(state->opaque);
1173     state->opaque = NULL;
1174 
1175     assert(s->reopen_state == state);
1176     s->reopen_state = NULL;
1177 }
1178 
1179 
1180 static void raw_reopen_abort(BDRVReopenState *state)
1181 {
1182     BDRVRawReopenState *rs = state->opaque;
1183     BDRVRawState *s = state->bs->opaque;
1184 
1185      /* nothing to do if NULL, we didn't get far enough */
1186     if (rs == NULL) {
1187         return;
1188     }
1189 
1190     g_free(state->opaque);
1191     state->opaque = NULL;
1192 
1193     assert(s->reopen_state == state);
1194     s->reopen_state = NULL;
1195 }
1196 
1197 static int hdev_get_max_hw_transfer(int fd, struct stat *st)
1198 {
1199 #ifdef BLKSECTGET
1200     if (S_ISBLK(st->st_mode)) {
1201         unsigned short max_sectors = 0;
1202         if (ioctl(fd, BLKSECTGET, &max_sectors) == 0) {
1203             return max_sectors * 512;
1204         }
1205     } else {
1206         int max_bytes = 0;
1207         if (ioctl(fd, BLKSECTGET, &max_bytes) == 0) {
1208             return max_bytes;
1209         }
1210     }
1211     return -errno;
1212 #else
1213     return -ENOSYS;
1214 #endif
1215 }
1216 
1217 /*
1218  * Get a sysfs attribute value as character string.
1219  */
1220 #ifdef CONFIG_LINUX
1221 static int get_sysfs_str_val(struct stat *st, const char *attribute,
1222                              char **val) {
1223     g_autofree char *sysfspath = NULL;
1224     size_t len;
1225 
1226     if (!S_ISBLK(st->st_mode)) {
1227         return -ENOTSUP;
1228     }
1229 
1230     sysfspath = g_strdup_printf("/sys/dev/block/%u:%u/queue/%s",
1231                                 major(st->st_rdev), minor(st->st_rdev),
1232                                 attribute);
1233     if (!g_file_get_contents(sysfspath, val, &len, NULL)) {
1234         return -ENOENT;
1235     }
1236 
1237     /* The file is ended with '\n' */
1238     char *p;
1239     p = *val;
1240     if (*(p + len - 1) == '\n') {
1241         *(p + len - 1) = '\0';
1242     }
1243     return 0;
1244 }
1245 #endif
1246 
1247 #if defined(CONFIG_BLKZONED)
1248 static int get_sysfs_zoned_model(struct stat *st, BlockZoneModel *zoned)
1249 {
1250     g_autofree char *val = NULL;
1251     int ret;
1252 
1253     ret = get_sysfs_str_val(st, "zoned", &val);
1254     if (ret < 0) {
1255         return ret;
1256     }
1257 
1258     if (strcmp(val, "host-managed") == 0) {
1259         *zoned = BLK_Z_HM;
1260     } else if (strcmp(val, "host-aware") == 0) {
1261         *zoned = BLK_Z_HA;
1262     } else if (strcmp(val, "none") == 0) {
1263         *zoned = BLK_Z_NONE;
1264     } else {
1265         return -ENOTSUP;
1266     }
1267     return 0;
1268 }
1269 #endif /* defined(CONFIG_BLKZONED) */
1270 
1271 /*
1272  * Get a sysfs attribute value as a long integer.
1273  */
1274 #ifdef CONFIG_LINUX
1275 static long get_sysfs_long_val(struct stat *st, const char *attribute)
1276 {
1277     g_autofree char *str = NULL;
1278     const char *end;
1279     long val;
1280     int ret;
1281 
1282     ret = get_sysfs_str_val(st, attribute, &str);
1283     if (ret < 0) {
1284         return ret;
1285     }
1286 
1287     /* The file is ended with '\n', pass 'end' to accept that. */
1288     ret = qemu_strtol(str, &end, 10, &val);
1289     if (ret == 0 && end && *end == '\0') {
1290         ret = val;
1291     }
1292     return ret;
1293 }
1294 #endif
1295 
1296 static int hdev_get_max_segments(int fd, struct stat *st)
1297 {
1298 #ifdef CONFIG_LINUX
1299     int ret;
1300 
1301     if (S_ISCHR(st->st_mode)) {
1302         if (ioctl(fd, SG_GET_SG_TABLESIZE, &ret) == 0) {
1303             return ret;
1304         }
1305         return -ENOTSUP;
1306     }
1307     return get_sysfs_long_val(st, "max_segments");
1308 #else
1309     return -ENOTSUP;
1310 #endif
1311 }
1312 
1313 #if defined(CONFIG_BLKZONED)
1314 /*
1315  * If the reset_all flag is true, then the wps of zone whose state is
1316  * not readonly or offline should be all reset to the start sector.
1317  * Else, take the real wp of the device.
1318  */
1319 static int get_zones_wp(BlockDriverState *bs, int fd, int64_t offset,
1320                         unsigned int nrz, bool reset_all)
1321 {
1322     struct blk_zone *blkz;
1323     size_t rep_size;
1324     uint64_t sector = offset >> BDRV_SECTOR_BITS;
1325     BlockZoneWps *wps = bs->wps;
1326     unsigned int j = offset / bs->bl.zone_size;
1327     unsigned int n = 0, i = 0;
1328     int ret;
1329     rep_size = sizeof(struct blk_zone_report) + nrz * sizeof(struct blk_zone);
1330     g_autofree struct blk_zone_report *rep = NULL;
1331 
1332     rep = g_malloc(rep_size);
1333     blkz = (struct blk_zone *)(rep + 1);
1334     while (n < nrz) {
1335         memset(rep, 0, rep_size);
1336         rep->sector = sector;
1337         rep->nr_zones = nrz - n;
1338 
1339         do {
1340             ret = ioctl(fd, BLKREPORTZONE, rep);
1341         } while (ret != 0 && errno == EINTR);
1342         if (ret != 0) {
1343             error_report("%d: ioctl BLKREPORTZONE at %" PRId64 " failed %d",
1344                     fd, offset, errno);
1345             return -errno;
1346         }
1347 
1348         if (!rep->nr_zones) {
1349             break;
1350         }
1351 
1352         for (i = 0; i < rep->nr_zones; ++i, ++n, ++j) {
1353             /*
1354              * The wp tracking cares only about sequential writes required and
1355              * sequential write preferred zones so that the wp can advance to
1356              * the right location.
1357              * Use the most significant bit of the wp location to indicate the
1358              * zone type: 0 for SWR/SWP zones and 1 for conventional zones.
1359              */
1360             if (blkz[i].type == BLK_ZONE_TYPE_CONVENTIONAL) {
1361                 wps->wp[j] |= 1ULL << 63;
1362             } else {
1363                 switch(blkz[i].cond) {
1364                 case BLK_ZONE_COND_FULL:
1365                 case BLK_ZONE_COND_READONLY:
1366                     /* Zone not writable */
1367                     wps->wp[j] = (blkz[i].start + blkz[i].len) << BDRV_SECTOR_BITS;
1368                     break;
1369                 case BLK_ZONE_COND_OFFLINE:
1370                     /* Zone not writable nor readable */
1371                     wps->wp[j] = (blkz[i].start) << BDRV_SECTOR_BITS;
1372                     break;
1373                 default:
1374                     if (reset_all) {
1375                         wps->wp[j] = blkz[i].start << BDRV_SECTOR_BITS;
1376                     } else {
1377                         wps->wp[j] = blkz[i].wp << BDRV_SECTOR_BITS;
1378                     }
1379                     break;
1380                 }
1381             }
1382         }
1383         sector = blkz[i - 1].start + blkz[i - 1].len;
1384     }
1385 
1386     return 0;
1387 }
1388 
1389 static void update_zones_wp(BlockDriverState *bs, int fd, int64_t offset,
1390                             unsigned int nrz)
1391 {
1392     if (get_zones_wp(bs, fd, offset, nrz, 0) < 0) {
1393         error_report("update zone wp failed");
1394     }
1395 }
1396 
1397 static void raw_refresh_zoned_limits(BlockDriverState *bs, struct stat *st,
1398                                      Error **errp)
1399 {
1400     BDRVRawState *s = bs->opaque;
1401     BlockZoneModel zoned;
1402     int ret;
1403 
1404     ret = get_sysfs_zoned_model(st, &zoned);
1405     if (ret < 0 || zoned == BLK_Z_NONE) {
1406         goto no_zoned;
1407     }
1408     bs->bl.zoned = zoned;
1409 
1410     ret = get_sysfs_long_val(st, "max_open_zones");
1411     if (ret >= 0) {
1412         bs->bl.max_open_zones = ret;
1413     }
1414 
1415     ret = get_sysfs_long_val(st, "max_active_zones");
1416     if (ret >= 0) {
1417         bs->bl.max_active_zones = ret;
1418     }
1419 
1420     /*
1421      * The zoned device must at least have zone size and nr_zones fields.
1422      */
1423     ret = get_sysfs_long_val(st, "chunk_sectors");
1424     if (ret < 0) {
1425         error_setg_errno(errp, -ret, "Unable to read chunk_sectors "
1426                                      "sysfs attribute");
1427         goto no_zoned;
1428     } else if (!ret) {
1429         error_setg(errp, "Read 0 from chunk_sectors sysfs attribute");
1430         goto no_zoned;
1431     }
1432     bs->bl.zone_size = ret << BDRV_SECTOR_BITS;
1433 
1434     ret = get_sysfs_long_val(st, "nr_zones");
1435     if (ret < 0) {
1436         error_setg_errno(errp, -ret, "Unable to read nr_zones "
1437                                      "sysfs attribute");
1438         goto no_zoned;
1439     } else if (!ret) {
1440         error_setg(errp, "Read 0 from nr_zones sysfs attribute");
1441         goto no_zoned;
1442     }
1443     bs->bl.nr_zones = ret;
1444 
1445     ret = get_sysfs_long_val(st, "zone_append_max_bytes");
1446     if (ret > 0) {
1447         bs->bl.max_append_sectors = ret >> BDRV_SECTOR_BITS;
1448     }
1449 
1450     ret = get_sysfs_long_val(st, "physical_block_size");
1451     if (ret >= 0) {
1452         bs->bl.write_granularity = ret;
1453     }
1454 
1455     /* The refresh_limits() function can be called multiple times. */
1456     g_free(bs->wps);
1457     bs->wps = g_malloc(sizeof(BlockZoneWps) +
1458             sizeof(int64_t) * bs->bl.nr_zones);
1459     ret = get_zones_wp(bs, s->fd, 0, bs->bl.nr_zones, 0);
1460     if (ret < 0) {
1461         error_setg_errno(errp, -ret, "report wps failed");
1462         goto no_zoned;
1463     }
1464     qemu_co_mutex_init(&bs->wps->colock);
1465     return;
1466 
1467 no_zoned:
1468     bs->bl.zoned = BLK_Z_NONE;
1469     g_free(bs->wps);
1470     bs->wps = NULL;
1471 }
1472 #else /* !defined(CONFIG_BLKZONED) */
1473 static void raw_refresh_zoned_limits(BlockDriverState *bs, struct stat *st,
1474                                      Error **errp)
1475 {
1476     bs->bl.zoned = BLK_Z_NONE;
1477 }
1478 #endif /* !defined(CONFIG_BLKZONED) */
1479 
1480 static void raw_refresh_limits(BlockDriverState *bs, Error **errp)
1481 {
1482     BDRVRawState *s = bs->opaque;
1483     struct stat st;
1484 
1485     s->needs_alignment = raw_needs_alignment(bs);
1486     raw_probe_alignment(bs, s->fd, errp);
1487 
1488     bs->bl.min_mem_alignment = s->buf_align;
1489     bs->bl.opt_mem_alignment = MAX(s->buf_align, qemu_real_host_page_size());
1490 
1491     /*
1492      * Maximum transfers are best effort, so it is okay to ignore any
1493      * errors.  That said, based on the man page errors in fstat would be
1494      * very much unexpected; the only possible case seems to be ENOMEM.
1495      */
1496     if (fstat(s->fd, &st)) {
1497         return;
1498     }
1499 
1500 #if defined(__APPLE__) && (__MACH__)
1501     struct statfs buf;
1502 
1503     if (!fstatfs(s->fd, &buf)) {
1504         bs->bl.opt_transfer = buf.f_iosize;
1505         bs->bl.pdiscard_alignment = buf.f_bsize;
1506     }
1507 #endif
1508 
1509     if (bdrv_is_sg(bs) || S_ISBLK(st.st_mode)) {
1510         int ret = hdev_get_max_hw_transfer(s->fd, &st);
1511 
1512         if (ret > 0 && ret <= BDRV_REQUEST_MAX_BYTES) {
1513             bs->bl.max_hw_transfer = ret;
1514         }
1515 
1516         ret = hdev_get_max_segments(s->fd, &st);
1517         if (ret > 0) {
1518             bs->bl.max_hw_iov = ret;
1519         }
1520     }
1521 
1522     raw_refresh_zoned_limits(bs, &st, errp);
1523 }
1524 
1525 static int check_for_dasd(int fd)
1526 {
1527 #ifdef BIODASDINFO2
1528     struct dasd_information2_t info = {0};
1529 
1530     return ioctl(fd, BIODASDINFO2, &info);
1531 #else
1532     return -1;
1533 #endif
1534 }
1535 
1536 /**
1537  * Try to get @bs's logical and physical block size.
1538  * On success, store them in @bsz and return zero.
1539  * On failure, return negative errno.
1540  */
1541 static int hdev_probe_blocksizes(BlockDriverState *bs, BlockSizes *bsz)
1542 {
1543     BDRVRawState *s = bs->opaque;
1544     int ret;
1545 
1546     /* If DASD or zoned devices, get blocksizes */
1547     if (check_for_dasd(s->fd) < 0) {
1548         /* zoned devices are not DASD */
1549         if (bs->bl.zoned == BLK_Z_NONE) {
1550             return -ENOTSUP;
1551         }
1552     }
1553     ret = probe_logical_blocksize(s->fd, &bsz->log);
1554     if (ret < 0) {
1555         return ret;
1556     }
1557     return probe_physical_blocksize(s->fd, &bsz->phys);
1558 }
1559 
1560 /**
1561  * Try to get @bs's geometry: cyls, heads, sectors.
1562  * On success, store them in @geo and return 0.
1563  * On failure return -errno.
1564  * (Allows block driver to assign default geometry values that guest sees)
1565  */
1566 #ifdef __linux__
1567 static int hdev_probe_geometry(BlockDriverState *bs, HDGeometry *geo)
1568 {
1569     BDRVRawState *s = bs->opaque;
1570     struct hd_geometry ioctl_geo = {0};
1571 
1572     /* If DASD, get its geometry */
1573     if (check_for_dasd(s->fd) < 0) {
1574         return -ENOTSUP;
1575     }
1576     if (ioctl(s->fd, HDIO_GETGEO, &ioctl_geo) < 0) {
1577         return -errno;
1578     }
1579     /* HDIO_GETGEO may return success even though geo contains zeros
1580        (e.g. certain multipath setups) */
1581     if (!ioctl_geo.heads || !ioctl_geo.sectors || !ioctl_geo.cylinders) {
1582         return -ENOTSUP;
1583     }
1584     /* Do not return a geometry for partition */
1585     if (ioctl_geo.start != 0) {
1586         return -ENOTSUP;
1587     }
1588     geo->heads = ioctl_geo.heads;
1589     geo->sectors = ioctl_geo.sectors;
1590     geo->cylinders = ioctl_geo.cylinders;
1591 
1592     return 0;
1593 }
1594 #else /* __linux__ */
1595 static int hdev_probe_geometry(BlockDriverState *bs, HDGeometry *geo)
1596 {
1597     return -ENOTSUP;
1598 }
1599 #endif
1600 
1601 #if defined(__linux__)
1602 static int handle_aiocb_ioctl(void *opaque)
1603 {
1604     RawPosixAIOData *aiocb = opaque;
1605     int ret;
1606 
1607     ret = RETRY_ON_EINTR(
1608         ioctl(aiocb->aio_fildes, aiocb->ioctl.cmd, aiocb->ioctl.buf)
1609     );
1610     if (ret == -1) {
1611         return -errno;
1612     }
1613 
1614     return 0;
1615 }
1616 #endif /* linux */
1617 
1618 static int handle_aiocb_flush(void *opaque)
1619 {
1620     RawPosixAIOData *aiocb = opaque;
1621     BDRVRawState *s = aiocb->bs->opaque;
1622     int ret;
1623 
1624     if (s->page_cache_inconsistent) {
1625         return -s->page_cache_inconsistent;
1626     }
1627 
1628     ret = qemu_fdatasync(aiocb->aio_fildes);
1629     if (ret == -1) {
1630         trace_file_flush_fdatasync_failed(errno);
1631 
1632         /* There is no clear definition of the semantics of a failing fsync(),
1633          * so we may have to assume the worst. The sad truth is that this
1634          * assumption is correct for Linux. Some pages are now probably marked
1635          * clean in the page cache even though they are inconsistent with the
1636          * on-disk contents. The next fdatasync() call would succeed, but no
1637          * further writeback attempt will be made. We can't get back to a state
1638          * in which we know what is on disk (we would have to rewrite
1639          * everything that was touched since the last fdatasync() at least), so
1640          * make bdrv_flush() fail permanently. Given that the behaviour isn't
1641          * really defined, I have little hope that other OSes are doing better.
1642          *
1643          * Obviously, this doesn't affect O_DIRECT, which bypasses the page
1644          * cache. */
1645         if ((s->open_flags & O_DIRECT) == 0) {
1646             s->page_cache_inconsistent = errno;
1647         }
1648         return -errno;
1649     }
1650     return 0;
1651 }
1652 
1653 #ifdef CONFIG_PREADV
1654 
1655 static bool preadv_present = true;
1656 
1657 static ssize_t
1658 qemu_preadv(int fd, const struct iovec *iov, int nr_iov, off_t offset)
1659 {
1660     return preadv(fd, iov, nr_iov, offset);
1661 }
1662 
1663 static ssize_t
1664 qemu_pwritev(int fd, const struct iovec *iov, int nr_iov, off_t offset)
1665 {
1666     return pwritev(fd, iov, nr_iov, offset);
1667 }
1668 
1669 #else
1670 
1671 static bool preadv_present = false;
1672 
1673 static ssize_t
1674 qemu_preadv(int fd, const struct iovec *iov, int nr_iov, off_t offset)
1675 {
1676     return -ENOSYS;
1677 }
1678 
1679 static ssize_t
1680 qemu_pwritev(int fd, const struct iovec *iov, int nr_iov, off_t offset)
1681 {
1682     return -ENOSYS;
1683 }
1684 
1685 #endif
1686 
1687 static ssize_t handle_aiocb_rw_vector(RawPosixAIOData *aiocb)
1688 {
1689     ssize_t len;
1690 
1691     len = RETRY_ON_EINTR(
1692         (aiocb->aio_type & (QEMU_AIO_WRITE | QEMU_AIO_ZONE_APPEND)) ?
1693             qemu_pwritev(aiocb->aio_fildes,
1694                            aiocb->io.iov,
1695                            aiocb->io.niov,
1696                            aiocb->aio_offset) :
1697             qemu_preadv(aiocb->aio_fildes,
1698                           aiocb->io.iov,
1699                           aiocb->io.niov,
1700                           aiocb->aio_offset)
1701     );
1702 
1703     if (len == -1) {
1704         return -errno;
1705     }
1706     return len;
1707 }
1708 
1709 /*
1710  * Read/writes the data to/from a given linear buffer.
1711  *
1712  * Returns the number of bytes handles or -errno in case of an error. Short
1713  * reads are only returned if the end of the file is reached.
1714  */
1715 static ssize_t handle_aiocb_rw_linear(RawPosixAIOData *aiocb, char *buf)
1716 {
1717     ssize_t offset = 0;
1718     ssize_t len;
1719 
1720     while (offset < aiocb->aio_nbytes) {
1721         if (aiocb->aio_type & (QEMU_AIO_WRITE | QEMU_AIO_ZONE_APPEND)) {
1722             len = pwrite(aiocb->aio_fildes,
1723                          (const char *)buf + offset,
1724                          aiocb->aio_nbytes - offset,
1725                          aiocb->aio_offset + offset);
1726         } else {
1727             len = pread(aiocb->aio_fildes,
1728                         buf + offset,
1729                         aiocb->aio_nbytes - offset,
1730                         aiocb->aio_offset + offset);
1731         }
1732         if (len == -1 && errno == EINTR) {
1733             continue;
1734         } else if (len == -1 && errno == EINVAL &&
1735                    (aiocb->bs->open_flags & BDRV_O_NOCACHE) &&
1736                    !(aiocb->aio_type & QEMU_AIO_WRITE) &&
1737                    offset > 0) {
1738             /* O_DIRECT pread() may fail with EINVAL when offset is unaligned
1739              * after a short read.  Assume that O_DIRECT short reads only occur
1740              * at EOF.  Therefore this is a short read, not an I/O error.
1741              */
1742             break;
1743         } else if (len == -1) {
1744             offset = -errno;
1745             break;
1746         } else if (len == 0) {
1747             break;
1748         }
1749         offset += len;
1750     }
1751 
1752     return offset;
1753 }
1754 
1755 static int handle_aiocb_rw(void *opaque)
1756 {
1757     RawPosixAIOData *aiocb = opaque;
1758     ssize_t nbytes;
1759     char *buf;
1760 
1761     if (!(aiocb->aio_type & QEMU_AIO_MISALIGNED)) {
1762         /*
1763          * If there is just a single buffer, and it is properly aligned
1764          * we can just use plain pread/pwrite without any problems.
1765          */
1766         if (aiocb->io.niov == 1) {
1767             nbytes = handle_aiocb_rw_linear(aiocb, aiocb->io.iov->iov_base);
1768             goto out;
1769         }
1770         /*
1771          * We have more than one iovec, and all are properly aligned.
1772          *
1773          * Try preadv/pwritev first and fall back to linearizing the
1774          * buffer if it's not supported.
1775          */
1776         if (preadv_present) {
1777             nbytes = handle_aiocb_rw_vector(aiocb);
1778             if (nbytes == aiocb->aio_nbytes ||
1779                 (nbytes < 0 && nbytes != -ENOSYS)) {
1780                 goto out;
1781             }
1782             preadv_present = false;
1783         }
1784 
1785         /*
1786          * XXX(hch): short read/write.  no easy way to handle the reminder
1787          * using these interfaces.  For now retry using plain
1788          * pread/pwrite?
1789          */
1790     }
1791 
1792     /*
1793      * Ok, we have to do it the hard way, copy all segments into
1794      * a single aligned buffer.
1795      */
1796     buf = qemu_try_blockalign(aiocb->bs, aiocb->aio_nbytes);
1797     if (buf == NULL) {
1798         nbytes = -ENOMEM;
1799         goto out;
1800     }
1801 
1802     if (aiocb->aio_type & QEMU_AIO_WRITE) {
1803         char *p = buf;
1804         int i;
1805 
1806         for (i = 0; i < aiocb->io.niov; ++i) {
1807             memcpy(p, aiocb->io.iov[i].iov_base, aiocb->io.iov[i].iov_len);
1808             p += aiocb->io.iov[i].iov_len;
1809         }
1810         assert(p - buf == aiocb->aio_nbytes);
1811     }
1812 
1813     nbytes = handle_aiocb_rw_linear(aiocb, buf);
1814     if (!(aiocb->aio_type & (QEMU_AIO_WRITE | QEMU_AIO_ZONE_APPEND))) {
1815         char *p = buf;
1816         size_t count = aiocb->aio_nbytes, copy;
1817         int i;
1818 
1819         for (i = 0; i < aiocb->io.niov && count; ++i) {
1820             copy = count;
1821             if (copy > aiocb->io.iov[i].iov_len) {
1822                 copy = aiocb->io.iov[i].iov_len;
1823             }
1824             memcpy(aiocb->io.iov[i].iov_base, p, copy);
1825             assert(count >= copy);
1826             p     += copy;
1827             count -= copy;
1828         }
1829         assert(count == 0);
1830     }
1831     qemu_vfree(buf);
1832 
1833 out:
1834     if (nbytes == aiocb->aio_nbytes) {
1835         return 0;
1836     } else if (nbytes >= 0 && nbytes < aiocb->aio_nbytes) {
1837         if (aiocb->aio_type & QEMU_AIO_WRITE) {
1838             return -EINVAL;
1839         } else {
1840             iov_memset(aiocb->io.iov, aiocb->io.niov, nbytes,
1841                       0, aiocb->aio_nbytes - nbytes);
1842             return 0;
1843         }
1844     } else {
1845         assert(nbytes < 0);
1846         return nbytes;
1847     }
1848 }
1849 
1850 #if defined(CONFIG_FALLOCATE) || defined(BLKZEROOUT) || defined(BLKDISCARD)
1851 static int translate_err(int err)
1852 {
1853     if (err == -ENODEV || err == -ENOSYS || err == -EOPNOTSUPP ||
1854         err == -ENOTTY) {
1855         err = -ENOTSUP;
1856     }
1857     return err;
1858 }
1859 #endif
1860 
1861 #ifdef CONFIG_FALLOCATE
1862 static int do_fallocate(int fd, int mode, off_t offset, off_t len)
1863 {
1864     do {
1865         if (fallocate(fd, mode, offset, len) == 0) {
1866             return 0;
1867         }
1868     } while (errno == EINTR);
1869     return translate_err(-errno);
1870 }
1871 #endif
1872 
1873 static ssize_t handle_aiocb_write_zeroes_block(RawPosixAIOData *aiocb)
1874 {
1875     int ret = -ENOTSUP;
1876     BDRVRawState *s = aiocb->bs->opaque;
1877 
1878     if (!s->has_write_zeroes) {
1879         return -ENOTSUP;
1880     }
1881 
1882 #ifdef BLKZEROOUT
1883     /* The BLKZEROOUT implementation in the kernel doesn't set
1884      * BLKDEV_ZERO_NOFALLBACK, so we can't call this if we have to avoid slow
1885      * fallbacks. */
1886     if (!(aiocb->aio_type & QEMU_AIO_NO_FALLBACK)) {
1887         do {
1888             uint64_t range[2] = { aiocb->aio_offset, aiocb->aio_nbytes };
1889             if (ioctl(aiocb->aio_fildes, BLKZEROOUT, range) == 0) {
1890                 return 0;
1891             }
1892         } while (errno == EINTR);
1893 
1894         ret = translate_err(-errno);
1895         if (ret == -ENOTSUP) {
1896             s->has_write_zeroes = false;
1897         }
1898     }
1899 #endif
1900 
1901     return ret;
1902 }
1903 
1904 static int handle_aiocb_write_zeroes(void *opaque)
1905 {
1906     RawPosixAIOData *aiocb = opaque;
1907 #ifdef CONFIG_FALLOCATE
1908     BDRVRawState *s = aiocb->bs->opaque;
1909     int64_t len;
1910 #endif
1911 
1912     if (aiocb->aio_type & QEMU_AIO_BLKDEV) {
1913         return handle_aiocb_write_zeroes_block(aiocb);
1914     }
1915 
1916 #ifdef CONFIG_FALLOCATE_ZERO_RANGE
1917     if (s->has_write_zeroes) {
1918         int ret = do_fallocate(s->fd, FALLOC_FL_ZERO_RANGE,
1919                                aiocb->aio_offset, aiocb->aio_nbytes);
1920         if (ret == -ENOTSUP) {
1921             s->has_write_zeroes = false;
1922         } else if (ret == 0 || ret != -EINVAL) {
1923             return ret;
1924         }
1925         /*
1926          * Note: Some file systems do not like unaligned byte ranges, and
1927          * return EINVAL in such a case, though they should not do it according
1928          * to the man-page of fallocate(). Thus we simply ignore this return
1929          * value and try the other fallbacks instead.
1930          */
1931     }
1932 #endif
1933 
1934 #ifdef CONFIG_FALLOCATE_PUNCH_HOLE
1935     if (s->has_discard && s->has_fallocate) {
1936         int ret = do_fallocate(s->fd,
1937                                FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE,
1938                                aiocb->aio_offset, aiocb->aio_nbytes);
1939         if (ret == 0) {
1940             ret = do_fallocate(s->fd, 0, aiocb->aio_offset, aiocb->aio_nbytes);
1941             if (ret == 0 || ret != -ENOTSUP) {
1942                 return ret;
1943             }
1944             s->has_fallocate = false;
1945         } else if (ret == -EINVAL) {
1946             /*
1947              * Some file systems like older versions of GPFS do not like un-
1948              * aligned byte ranges, and return EINVAL in such a case, though
1949              * they should not do it according to the man-page of fallocate().
1950              * Warn about the bad filesystem and try the final fallback instead.
1951              */
1952             warn_report_once("Your file system is misbehaving: "
1953                              "fallocate(FALLOC_FL_PUNCH_HOLE) returned EINVAL. "
1954                              "Please report this bug to your file system "
1955                              "vendor.");
1956         } else if (ret != -ENOTSUP) {
1957             return ret;
1958         } else {
1959             s->has_discard = false;
1960         }
1961     }
1962 #endif
1963 
1964 #ifdef CONFIG_FALLOCATE
1965     /* Last resort: we are trying to extend the file with zeroed data. This
1966      * can be done via fallocate(fd, 0) */
1967     len = raw_getlength(aiocb->bs);
1968     if (s->has_fallocate && len >= 0 && aiocb->aio_offset >= len) {
1969         int ret = do_fallocate(s->fd, 0, aiocb->aio_offset, aiocb->aio_nbytes);
1970         if (ret == 0 || ret != -ENOTSUP) {
1971             return ret;
1972         }
1973         s->has_fallocate = false;
1974     }
1975 #endif
1976 
1977     return -ENOTSUP;
1978 }
1979 
1980 static int handle_aiocb_write_zeroes_unmap(void *opaque)
1981 {
1982     RawPosixAIOData *aiocb = opaque;
1983     BDRVRawState *s G_GNUC_UNUSED = aiocb->bs->opaque;
1984 
1985     /* First try to write zeros and unmap at the same time */
1986 
1987 #ifdef CONFIG_FALLOCATE_PUNCH_HOLE
1988     int ret = do_fallocate(s->fd, FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE,
1989                            aiocb->aio_offset, aiocb->aio_nbytes);
1990     switch (ret) {
1991     case -ENOTSUP:
1992     case -EINVAL:
1993     case -EBUSY:
1994         break;
1995     default:
1996         return ret;
1997     }
1998 #endif
1999 
2000     /* If we couldn't manage to unmap while guaranteed that the area reads as
2001      * all-zero afterwards, just write zeroes without unmapping */
2002     return handle_aiocb_write_zeroes(aiocb);
2003 }
2004 
2005 #ifndef HAVE_COPY_FILE_RANGE
2006 static off_t copy_file_range(int in_fd, off_t *in_off, int out_fd,
2007                              off_t *out_off, size_t len, unsigned int flags)
2008 {
2009 #ifdef __NR_copy_file_range
2010     return syscall(__NR_copy_file_range, in_fd, in_off, out_fd,
2011                    out_off, len, flags);
2012 #else
2013     errno = ENOSYS;
2014     return -1;
2015 #endif
2016 }
2017 #endif
2018 
2019 /*
2020  * parse_zone - Fill a zone descriptor
2021  */
2022 #if defined(CONFIG_BLKZONED)
2023 static inline int parse_zone(struct BlockZoneDescriptor *zone,
2024                               const struct blk_zone *blkz) {
2025     zone->start = blkz->start << BDRV_SECTOR_BITS;
2026     zone->length = blkz->len << BDRV_SECTOR_BITS;
2027     zone->wp = blkz->wp << BDRV_SECTOR_BITS;
2028 
2029 #ifdef HAVE_BLK_ZONE_REP_CAPACITY
2030     zone->cap = blkz->capacity << BDRV_SECTOR_BITS;
2031 #else
2032     zone->cap = blkz->len << BDRV_SECTOR_BITS;
2033 #endif
2034 
2035     switch (blkz->type) {
2036     case BLK_ZONE_TYPE_SEQWRITE_REQ:
2037         zone->type = BLK_ZT_SWR;
2038         break;
2039     case BLK_ZONE_TYPE_SEQWRITE_PREF:
2040         zone->type = BLK_ZT_SWP;
2041         break;
2042     case BLK_ZONE_TYPE_CONVENTIONAL:
2043         zone->type = BLK_ZT_CONV;
2044         break;
2045     default:
2046         error_report("Unsupported zone type: 0x%x", blkz->type);
2047         return -ENOTSUP;
2048     }
2049 
2050     switch (blkz->cond) {
2051     case BLK_ZONE_COND_NOT_WP:
2052         zone->state = BLK_ZS_NOT_WP;
2053         break;
2054     case BLK_ZONE_COND_EMPTY:
2055         zone->state = BLK_ZS_EMPTY;
2056         break;
2057     case BLK_ZONE_COND_IMP_OPEN:
2058         zone->state = BLK_ZS_IOPEN;
2059         break;
2060     case BLK_ZONE_COND_EXP_OPEN:
2061         zone->state = BLK_ZS_EOPEN;
2062         break;
2063     case BLK_ZONE_COND_CLOSED:
2064         zone->state = BLK_ZS_CLOSED;
2065         break;
2066     case BLK_ZONE_COND_READONLY:
2067         zone->state = BLK_ZS_RDONLY;
2068         break;
2069     case BLK_ZONE_COND_FULL:
2070         zone->state = BLK_ZS_FULL;
2071         break;
2072     case BLK_ZONE_COND_OFFLINE:
2073         zone->state = BLK_ZS_OFFLINE;
2074         break;
2075     default:
2076         error_report("Unsupported zone state: 0x%x", blkz->cond);
2077         return -ENOTSUP;
2078     }
2079     return 0;
2080 }
2081 #endif
2082 
2083 #if defined(CONFIG_BLKZONED)
2084 static int handle_aiocb_zone_report(void *opaque)
2085 {
2086     RawPosixAIOData *aiocb = opaque;
2087     int fd = aiocb->aio_fildes;
2088     unsigned int *nr_zones = aiocb->zone_report.nr_zones;
2089     BlockZoneDescriptor *zones = aiocb->zone_report.zones;
2090     /* zoned block devices use 512-byte sectors */
2091     uint64_t sector = aiocb->aio_offset / 512;
2092 
2093     struct blk_zone *blkz;
2094     size_t rep_size;
2095     unsigned int nrz;
2096     int ret;
2097     unsigned int n = 0, i = 0;
2098 
2099     nrz = *nr_zones;
2100     rep_size = sizeof(struct blk_zone_report) + nrz * sizeof(struct blk_zone);
2101     g_autofree struct blk_zone_report *rep = NULL;
2102     rep = g_malloc(rep_size);
2103 
2104     blkz = (struct blk_zone *)(rep + 1);
2105     while (n < nrz) {
2106         memset(rep, 0, rep_size);
2107         rep->sector = sector;
2108         rep->nr_zones = nrz - n;
2109 
2110         do {
2111             ret = ioctl(fd, BLKREPORTZONE, rep);
2112         } while (ret != 0 && errno == EINTR);
2113         if (ret != 0) {
2114             error_report("%d: ioctl BLKREPORTZONE at %" PRId64 " failed %d",
2115                          fd, sector, errno);
2116             return -errno;
2117         }
2118 
2119         if (!rep->nr_zones) {
2120             break;
2121         }
2122 
2123         for (i = 0; i < rep->nr_zones; i++, n++) {
2124             ret = parse_zone(&zones[n], &blkz[i]);
2125             if (ret != 0) {
2126                 return ret;
2127             }
2128 
2129             /* The next report should start after the last zone reported */
2130             sector = blkz[i].start + blkz[i].len;
2131         }
2132     }
2133 
2134     *nr_zones = n;
2135     return 0;
2136 }
2137 #endif
2138 
2139 #if defined(CONFIG_BLKZONED)
2140 static int handle_aiocb_zone_mgmt(void *opaque)
2141 {
2142     RawPosixAIOData *aiocb = opaque;
2143     int fd = aiocb->aio_fildes;
2144     uint64_t sector = aiocb->aio_offset / 512;
2145     int64_t nr_sectors = aiocb->aio_nbytes / 512;
2146     struct blk_zone_range range;
2147     int ret;
2148 
2149     /* Execute the operation */
2150     range.sector = sector;
2151     range.nr_sectors = nr_sectors;
2152     do {
2153         ret = ioctl(fd, aiocb->zone_mgmt.op, &range);
2154     } while (ret != 0 && errno == EINTR);
2155 
2156     return ret < 0 ? -errno : ret;
2157 }
2158 #endif
2159 
2160 static int handle_aiocb_copy_range(void *opaque)
2161 {
2162     RawPosixAIOData *aiocb = opaque;
2163     uint64_t bytes = aiocb->aio_nbytes;
2164     off_t in_off = aiocb->aio_offset;
2165     off_t out_off = aiocb->copy_range.aio_offset2;
2166 
2167     while (bytes) {
2168         ssize_t ret = copy_file_range(aiocb->aio_fildes, &in_off,
2169                                       aiocb->copy_range.aio_fd2, &out_off,
2170                                       bytes, 0);
2171         trace_file_copy_file_range(aiocb->bs, aiocb->aio_fildes, in_off,
2172                                    aiocb->copy_range.aio_fd2, out_off, bytes,
2173                                    0, ret);
2174         if (ret == 0) {
2175             /* No progress (e.g. when beyond EOF), let the caller fall back to
2176              * buffer I/O. */
2177             return -ENOSPC;
2178         }
2179         if (ret < 0) {
2180             switch (errno) {
2181             case ENOSYS:
2182                 return -ENOTSUP;
2183             case EINTR:
2184                 continue;
2185             default:
2186                 return -errno;
2187             }
2188         }
2189         bytes -= ret;
2190     }
2191     return 0;
2192 }
2193 
2194 static int handle_aiocb_discard(void *opaque)
2195 {
2196     RawPosixAIOData *aiocb = opaque;
2197     int ret = -ENOTSUP;
2198     BDRVRawState *s = aiocb->bs->opaque;
2199 
2200     if (!s->has_discard) {
2201         return -ENOTSUP;
2202     }
2203 
2204     if (aiocb->aio_type & QEMU_AIO_BLKDEV) {
2205 #ifdef BLKDISCARD
2206         do {
2207             uint64_t range[2] = { aiocb->aio_offset, aiocb->aio_nbytes };
2208             if (ioctl(aiocb->aio_fildes, BLKDISCARD, range) == 0) {
2209                 return 0;
2210             }
2211         } while (errno == EINTR);
2212 
2213         ret = translate_err(-errno);
2214 #endif
2215     } else {
2216 #ifdef CONFIG_FALLOCATE_PUNCH_HOLE
2217         ret = do_fallocate(s->fd, FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE,
2218                            aiocb->aio_offset, aiocb->aio_nbytes);
2219         ret = translate_err(ret);
2220 #elif defined(__APPLE__) && (__MACH__)
2221         fpunchhole_t fpunchhole;
2222         fpunchhole.fp_flags = 0;
2223         fpunchhole.reserved = 0;
2224         fpunchhole.fp_offset = aiocb->aio_offset;
2225         fpunchhole.fp_length = aiocb->aio_nbytes;
2226         if (fcntl(s->fd, F_PUNCHHOLE, &fpunchhole) == -1) {
2227             ret = errno == ENODEV ? -ENOTSUP : -errno;
2228         } else {
2229             ret = 0;
2230         }
2231 #endif
2232     }
2233 
2234     if (ret == -ENOTSUP) {
2235         s->has_discard = false;
2236     }
2237     return ret;
2238 }
2239 
2240 /*
2241  * Help alignment probing by allocating the first block.
2242  *
2243  * When reading with direct I/O from unallocated area on Gluster backed by XFS,
2244  * reading succeeds regardless of request length. In this case we fallback to
2245  * safe alignment which is not optimal. Allocating the first block avoids this
2246  * fallback.
2247  *
2248  * fd may be opened with O_DIRECT, but we don't know the buffer alignment or
2249  * request alignment, so we use safe values.
2250  *
2251  * Returns: 0 on success, -errno on failure. Since this is an optimization,
2252  * caller may ignore failures.
2253  */
2254 static int allocate_first_block(int fd, size_t max_size)
2255 {
2256     size_t write_size = (max_size < MAX_BLOCKSIZE)
2257         ? BDRV_SECTOR_SIZE
2258         : MAX_BLOCKSIZE;
2259     size_t max_align = MAX(MAX_BLOCKSIZE, qemu_real_host_page_size());
2260     void *buf;
2261     ssize_t n;
2262     int ret;
2263 
2264     buf = qemu_memalign(max_align, write_size);
2265     memset(buf, 0, write_size);
2266 
2267     n = RETRY_ON_EINTR(pwrite(fd, buf, write_size, 0));
2268 
2269     ret = (n == -1) ? -errno : 0;
2270 
2271     qemu_vfree(buf);
2272     return ret;
2273 }
2274 
2275 static int handle_aiocb_truncate(void *opaque)
2276 {
2277     RawPosixAIOData *aiocb = opaque;
2278     int result = 0;
2279     int64_t current_length = 0;
2280     char *buf = NULL;
2281     struct stat st;
2282     int fd = aiocb->aio_fildes;
2283     int64_t offset = aiocb->aio_offset;
2284     PreallocMode prealloc = aiocb->truncate.prealloc;
2285     Error **errp = aiocb->truncate.errp;
2286 
2287     if (fstat(fd, &st) < 0) {
2288         result = -errno;
2289         error_setg_errno(errp, -result, "Could not stat file");
2290         return result;
2291     }
2292 
2293     current_length = st.st_size;
2294     if (current_length > offset && prealloc != PREALLOC_MODE_OFF) {
2295         error_setg(errp, "Cannot use preallocation for shrinking files");
2296         return -ENOTSUP;
2297     }
2298 
2299     switch (prealloc) {
2300 #ifdef CONFIG_POSIX_FALLOCATE
2301     case PREALLOC_MODE_FALLOC:
2302         /*
2303          * Truncating before posix_fallocate() makes it about twice slower on
2304          * file systems that do not support fallocate(), trying to check if a
2305          * block is allocated before allocating it, so don't do that here.
2306          */
2307         if (offset != current_length) {
2308             result = -posix_fallocate(fd, current_length,
2309                                       offset - current_length);
2310             if (result != 0) {
2311                 /* posix_fallocate() doesn't set errno. */
2312                 error_setg_errno(errp, -result,
2313                                  "Could not preallocate new data");
2314             } else if (current_length == 0) {
2315                 /*
2316                  * posix_fallocate() uses fallocate() if the filesystem
2317                  * supports it, or fallback to manually writing zeroes. If
2318                  * fallocate() was used, unaligned reads from the fallocated
2319                  * area in raw_probe_alignment() will succeed, hence we need to
2320                  * allocate the first block.
2321                  *
2322                  * Optimize future alignment probing; ignore failures.
2323                  */
2324                 allocate_first_block(fd, offset);
2325             }
2326         } else {
2327             result = 0;
2328         }
2329         goto out;
2330 #endif
2331     case PREALLOC_MODE_FULL:
2332     {
2333         int64_t num = 0, left = offset - current_length;
2334         off_t seek_result;
2335 
2336         /*
2337          * Knowing the final size from the beginning could allow the file
2338          * system driver to do less allocations and possibly avoid
2339          * fragmentation of the file.
2340          */
2341         if (ftruncate(fd, offset) != 0) {
2342             result = -errno;
2343             error_setg_errno(errp, -result, "Could not resize file");
2344             goto out;
2345         }
2346 
2347         buf = g_malloc0(65536);
2348 
2349         seek_result = lseek(fd, current_length, SEEK_SET);
2350         if (seek_result < 0) {
2351             result = -errno;
2352             error_setg_errno(errp, -result,
2353                              "Failed to seek to the old end of file");
2354             goto out;
2355         }
2356 
2357         while (left > 0) {
2358             num = MIN(left, 65536);
2359             result = write(fd, buf, num);
2360             if (result < 0) {
2361                 if (errno == EINTR) {
2362                     continue;
2363                 }
2364                 result = -errno;
2365                 error_setg_errno(errp, -result,
2366                                  "Could not write zeros for preallocation");
2367                 goto out;
2368             }
2369             left -= result;
2370         }
2371         if (result >= 0) {
2372             result = fsync(fd);
2373             if (result < 0) {
2374                 result = -errno;
2375                 error_setg_errno(errp, -result,
2376                                  "Could not flush file to disk");
2377                 goto out;
2378             }
2379         }
2380         goto out;
2381     }
2382     case PREALLOC_MODE_OFF:
2383         if (ftruncate(fd, offset) != 0) {
2384             result = -errno;
2385             error_setg_errno(errp, -result, "Could not resize file");
2386         } else if (current_length == 0 && offset > current_length) {
2387             /* Optimize future alignment probing; ignore failures. */
2388             allocate_first_block(fd, offset);
2389         }
2390         return result;
2391     default:
2392         result = -ENOTSUP;
2393         error_setg(errp, "Unsupported preallocation mode: %s",
2394                    PreallocMode_str(prealloc));
2395         return result;
2396     }
2397 
2398 out:
2399     if (result < 0) {
2400         if (ftruncate(fd, current_length) < 0) {
2401             error_report("Failed to restore old file length: %s",
2402                          strerror(errno));
2403         }
2404     }
2405 
2406     g_free(buf);
2407     return result;
2408 }
2409 
2410 static int coroutine_fn raw_thread_pool_submit(ThreadPoolFunc func, void *arg)
2411 {
2412     return thread_pool_submit_co(func, arg);
2413 }
2414 
2415 /*
2416  * Check if all memory in this vector is sector aligned.
2417  */
2418 static bool bdrv_qiov_is_aligned(BlockDriverState *bs, QEMUIOVector *qiov)
2419 {
2420     int i;
2421     size_t alignment = bdrv_min_mem_align(bs);
2422     size_t len = bs->bl.request_alignment;
2423     IO_CODE();
2424 
2425     for (i = 0; i < qiov->niov; i++) {
2426         if ((uintptr_t) qiov->iov[i].iov_base % alignment) {
2427             return false;
2428         }
2429         if (qiov->iov[i].iov_len % len) {
2430             return false;
2431         }
2432     }
2433 
2434     return true;
2435 }
2436 
2437 #ifdef CONFIG_LINUX_IO_URING
2438 static inline bool raw_check_linux_io_uring(BDRVRawState *s)
2439 {
2440     Error *local_err = NULL;
2441     AioContext *ctx;
2442 
2443     if (!s->use_linux_io_uring) {
2444         return false;
2445     }
2446 
2447     ctx = qemu_get_current_aio_context();
2448     if (unlikely(!aio_setup_linux_io_uring(ctx, &local_err))) {
2449         error_reportf_err(local_err, "Unable to use linux io_uring, "
2450                                      "falling back to thread pool: ");
2451         s->use_linux_io_uring = false;
2452         return false;
2453     }
2454     return true;
2455 }
2456 #endif
2457 
2458 #ifdef CONFIG_LINUX_AIO
2459 static inline bool raw_check_linux_aio(BDRVRawState *s)
2460 {
2461     Error *local_err = NULL;
2462     AioContext *ctx;
2463 
2464     if (!s->use_linux_aio) {
2465         return false;
2466     }
2467 
2468     ctx = qemu_get_current_aio_context();
2469     if (unlikely(!aio_setup_linux_aio(ctx, &local_err))) {
2470         error_reportf_err(local_err, "Unable to use Linux AIO, "
2471                                      "falling back to thread pool: ");
2472         s->use_linux_aio = false;
2473         return false;
2474     }
2475     return true;
2476 }
2477 #endif
2478 
2479 static int coroutine_fn raw_co_prw(BlockDriverState *bs, int64_t *offset_ptr,
2480                                    uint64_t bytes, QEMUIOVector *qiov, int type)
2481 {
2482     BDRVRawState *s = bs->opaque;
2483     RawPosixAIOData acb;
2484     int ret;
2485     uint64_t offset = *offset_ptr;
2486 
2487     if (fd_open(bs) < 0)
2488         return -EIO;
2489 #if defined(CONFIG_BLKZONED)
2490     if ((type & (QEMU_AIO_WRITE | QEMU_AIO_ZONE_APPEND)) &&
2491         bs->bl.zoned != BLK_Z_NONE) {
2492         qemu_co_mutex_lock(&bs->wps->colock);
2493         if (type & QEMU_AIO_ZONE_APPEND) {
2494             int index = offset / bs->bl.zone_size;
2495             offset = bs->wps->wp[index];
2496         }
2497     }
2498 #endif
2499 
2500     /*
2501      * When using O_DIRECT, the request must be aligned to be able to use
2502      * either libaio or io_uring interface. If not fail back to regular thread
2503      * pool read/write code which emulates this for us if we
2504      * set QEMU_AIO_MISALIGNED.
2505      */
2506     if (s->needs_alignment && !bdrv_qiov_is_aligned(bs, qiov)) {
2507         type |= QEMU_AIO_MISALIGNED;
2508 #ifdef CONFIG_LINUX_IO_URING
2509     } else if (raw_check_linux_io_uring(s)) {
2510         assert(qiov->size == bytes);
2511         ret = luring_co_submit(bs, s->fd, offset, qiov, type);
2512         goto out;
2513 #endif
2514 #ifdef CONFIG_LINUX_AIO
2515     } else if (raw_check_linux_aio(s)) {
2516         assert(qiov->size == bytes);
2517         ret = laio_co_submit(s->fd, offset, qiov, type,
2518                               s->aio_max_batch);
2519         goto out;
2520 #endif
2521     }
2522 
2523     acb = (RawPosixAIOData) {
2524         .bs             = bs,
2525         .aio_fildes     = s->fd,
2526         .aio_type       = type,
2527         .aio_offset     = offset,
2528         .aio_nbytes     = bytes,
2529         .io             = {
2530             .iov            = qiov->iov,
2531             .niov           = qiov->niov,
2532         },
2533     };
2534 
2535     assert(qiov->size == bytes);
2536     ret = raw_thread_pool_submit(handle_aiocb_rw, &acb);
2537     goto out; /* Avoid the compiler err of unused label */
2538 
2539 out:
2540 #if defined(CONFIG_BLKZONED)
2541     if ((type & (QEMU_AIO_WRITE | QEMU_AIO_ZONE_APPEND)) &&
2542         bs->bl.zoned != BLK_Z_NONE) {
2543         BlockZoneWps *wps = bs->wps;
2544         if (ret == 0) {
2545             uint64_t *wp = &wps->wp[offset / bs->bl.zone_size];
2546             if (!BDRV_ZT_IS_CONV(*wp)) {
2547                 if (type & QEMU_AIO_ZONE_APPEND) {
2548                     *offset_ptr = *wp;
2549                     trace_zbd_zone_append_complete(bs, *offset_ptr
2550                         >> BDRV_SECTOR_BITS);
2551                 }
2552                 /* Advance the wp if needed */
2553                 if (offset + bytes > *wp) {
2554                     *wp = offset + bytes;
2555                 }
2556             }
2557         } else {
2558             /*
2559              * write and append write are not allowed to cross zone boundaries
2560              */
2561             update_zones_wp(bs, s->fd, offset, 1);
2562         }
2563 
2564         qemu_co_mutex_unlock(&wps->colock);
2565     }
2566 #endif
2567     return ret;
2568 }
2569 
2570 static int coroutine_fn raw_co_preadv(BlockDriverState *bs, int64_t offset,
2571                                       int64_t bytes, QEMUIOVector *qiov,
2572                                       BdrvRequestFlags flags)
2573 {
2574     return raw_co_prw(bs, &offset, bytes, qiov, QEMU_AIO_READ);
2575 }
2576 
2577 static int coroutine_fn raw_co_pwritev(BlockDriverState *bs, int64_t offset,
2578                                        int64_t bytes, QEMUIOVector *qiov,
2579                                        BdrvRequestFlags flags)
2580 {
2581     return raw_co_prw(bs, &offset, bytes, qiov, QEMU_AIO_WRITE);
2582 }
2583 
2584 static int coroutine_fn raw_co_flush_to_disk(BlockDriverState *bs)
2585 {
2586     BDRVRawState *s = bs->opaque;
2587     RawPosixAIOData acb;
2588     int ret;
2589 
2590     ret = fd_open(bs);
2591     if (ret < 0) {
2592         return ret;
2593     }
2594 
2595     acb = (RawPosixAIOData) {
2596         .bs             = bs,
2597         .aio_fildes     = s->fd,
2598         .aio_type       = QEMU_AIO_FLUSH,
2599     };
2600 
2601 #ifdef CONFIG_LINUX_IO_URING
2602     if (raw_check_linux_io_uring(s)) {
2603         return luring_co_submit(bs, s->fd, 0, NULL, QEMU_AIO_FLUSH);
2604     }
2605 #endif
2606 #ifdef CONFIG_LINUX_AIO
2607     if (s->has_laio_fdsync && raw_check_linux_aio(s)) {
2608         return laio_co_submit(s->fd, 0, NULL, QEMU_AIO_FLUSH, 0);
2609     }
2610 #endif
2611     return raw_thread_pool_submit(handle_aiocb_flush, &acb);
2612 }
2613 
2614 static void raw_close(BlockDriverState *bs)
2615 {
2616     BDRVRawState *s = bs->opaque;
2617 
2618     if (s->fd >= 0) {
2619 #if defined(CONFIG_BLKZONED)
2620         g_free(bs->wps);
2621 #endif
2622         qemu_close(s->fd);
2623         s->fd = -1;
2624     }
2625 }
2626 
2627 /**
2628  * Truncates the given regular file @fd to @offset and, when growing, fills the
2629  * new space according to @prealloc.
2630  *
2631  * Returns: 0 on success, -errno on failure.
2632  */
2633 static int coroutine_fn
2634 raw_regular_truncate(BlockDriverState *bs, int fd, int64_t offset,
2635                      PreallocMode prealloc, Error **errp)
2636 {
2637     RawPosixAIOData acb;
2638 
2639     acb = (RawPosixAIOData) {
2640         .bs             = bs,
2641         .aio_fildes     = fd,
2642         .aio_type       = QEMU_AIO_TRUNCATE,
2643         .aio_offset     = offset,
2644         .truncate       = {
2645             .prealloc       = prealloc,
2646             .errp           = errp,
2647         },
2648     };
2649 
2650     return raw_thread_pool_submit(handle_aiocb_truncate, &acb);
2651 }
2652 
2653 static int coroutine_fn raw_co_truncate(BlockDriverState *bs, int64_t offset,
2654                                         bool exact, PreallocMode prealloc,
2655                                         BdrvRequestFlags flags, Error **errp)
2656 {
2657     BDRVRawState *s = bs->opaque;
2658     struct stat st;
2659     int ret;
2660 
2661     if (fstat(s->fd, &st)) {
2662         ret = -errno;
2663         error_setg_errno(errp, -ret, "Failed to fstat() the file");
2664         return ret;
2665     }
2666 
2667     if (S_ISREG(st.st_mode)) {
2668         /* Always resizes to the exact @offset */
2669         return raw_regular_truncate(bs, s->fd, offset, prealloc, errp);
2670     }
2671 
2672     if (prealloc != PREALLOC_MODE_OFF) {
2673         error_setg(errp, "Preallocation mode '%s' unsupported for this "
2674                    "non-regular file", PreallocMode_str(prealloc));
2675         return -ENOTSUP;
2676     }
2677 
2678     if (S_ISCHR(st.st_mode) || S_ISBLK(st.st_mode)) {
2679         int64_t cur_length = raw_getlength(bs);
2680 
2681         if (offset != cur_length && exact) {
2682             error_setg(errp, "Cannot resize device files");
2683             return -ENOTSUP;
2684         } else if (offset > cur_length) {
2685             error_setg(errp, "Cannot grow device files");
2686             return -EINVAL;
2687         }
2688     } else {
2689         error_setg(errp, "Resizing this file is not supported");
2690         return -ENOTSUP;
2691     }
2692 
2693     return 0;
2694 }
2695 
2696 #ifdef __OpenBSD__
2697 static int64_t raw_getlength(BlockDriverState *bs)
2698 {
2699     BDRVRawState *s = bs->opaque;
2700     int fd = s->fd;
2701     struct stat st;
2702 
2703     if (fstat(fd, &st))
2704         return -errno;
2705     if (S_ISCHR(st.st_mode) || S_ISBLK(st.st_mode)) {
2706         struct disklabel dl;
2707 
2708         if (ioctl(fd, DIOCGDINFO, &dl))
2709             return -errno;
2710         return (uint64_t)dl.d_secsize *
2711             dl.d_partitions[DISKPART(st.st_rdev)].p_size;
2712     } else
2713         return st.st_size;
2714 }
2715 #elif defined(__NetBSD__)
2716 static int64_t raw_getlength(BlockDriverState *bs)
2717 {
2718     BDRVRawState *s = bs->opaque;
2719     int fd = s->fd;
2720     struct stat st;
2721 
2722     if (fstat(fd, &st))
2723         return -errno;
2724     if (S_ISCHR(st.st_mode) || S_ISBLK(st.st_mode)) {
2725         struct dkwedge_info dkw;
2726 
2727         if (ioctl(fd, DIOCGWEDGEINFO, &dkw) != -1) {
2728             return dkw.dkw_size * 512;
2729         } else {
2730             struct disklabel dl;
2731 
2732             if (ioctl(fd, DIOCGDINFO, &dl))
2733                 return -errno;
2734             return (uint64_t)dl.d_secsize *
2735                 dl.d_partitions[DISKPART(st.st_rdev)].p_size;
2736         }
2737     } else
2738         return st.st_size;
2739 }
2740 #elif defined(__sun__)
2741 static int64_t raw_getlength(BlockDriverState *bs)
2742 {
2743     BDRVRawState *s = bs->opaque;
2744     struct dk_minfo minfo;
2745     int ret;
2746     int64_t size;
2747 
2748     ret = fd_open(bs);
2749     if (ret < 0) {
2750         return ret;
2751     }
2752 
2753     /*
2754      * Use the DKIOCGMEDIAINFO ioctl to read the size.
2755      */
2756     ret = ioctl(s->fd, DKIOCGMEDIAINFO, &minfo);
2757     if (ret != -1) {
2758         return minfo.dki_lbsize * minfo.dki_capacity;
2759     }
2760 
2761     /*
2762      * There are reports that lseek on some devices fails, but
2763      * irc discussion said that contingency on contingency was overkill.
2764      */
2765     size = lseek(s->fd, 0, SEEK_END);
2766     if (size < 0) {
2767         return -errno;
2768     }
2769     return size;
2770 }
2771 #elif defined(CONFIG_BSD)
2772 static int64_t raw_getlength(BlockDriverState *bs)
2773 {
2774     BDRVRawState *s = bs->opaque;
2775     int fd = s->fd;
2776     int64_t size;
2777     struct stat sb;
2778 #if defined (__FreeBSD__) || defined(__FreeBSD_kernel__)
2779     int reopened = 0;
2780 #endif
2781     int ret;
2782 
2783     ret = fd_open(bs);
2784     if (ret < 0)
2785         return ret;
2786 
2787 #if defined (__FreeBSD__) || defined(__FreeBSD_kernel__)
2788 again:
2789 #endif
2790     if (!fstat(fd, &sb) && (S_IFCHR & sb.st_mode)) {
2791         size = 0;
2792 #ifdef DIOCGMEDIASIZE
2793         if (ioctl(fd, DIOCGMEDIASIZE, (off_t *)&size)) {
2794             size = 0;
2795         }
2796 #endif
2797 #ifdef DIOCGPART
2798         if (size == 0) {
2799             struct partinfo pi;
2800             if (ioctl(fd, DIOCGPART, &pi) == 0) {
2801                 size = pi.media_size;
2802             }
2803         }
2804 #endif
2805 #if defined(DKIOCGETBLOCKCOUNT) && defined(DKIOCGETBLOCKSIZE)
2806         if (size == 0) {
2807             uint64_t sectors = 0;
2808             uint32_t sector_size = 0;
2809 
2810             if (ioctl(fd, DKIOCGETBLOCKCOUNT, &sectors) == 0
2811                && ioctl(fd, DKIOCGETBLOCKSIZE, &sector_size) == 0) {
2812                 size = sectors * sector_size;
2813             }
2814         }
2815 #endif
2816         if (size == 0) {
2817             size = lseek(fd, 0LL, SEEK_END);
2818         }
2819         if (size < 0) {
2820             return -errno;
2821         }
2822 #if defined(__FreeBSD__) || defined(__FreeBSD_kernel__)
2823         switch(s->type) {
2824         case FTYPE_CD:
2825             /* XXX FreeBSD acd returns UINT_MAX sectors for an empty drive */
2826             if (size == 2048LL * (unsigned)-1)
2827                 size = 0;
2828             /* XXX no disc?  maybe we need to reopen... */
2829             if (size <= 0 && !reopened && cdrom_reopen(bs) >= 0) {
2830                 reopened = 1;
2831                 goto again;
2832             }
2833         }
2834 #endif
2835     } else {
2836         size = lseek(fd, 0, SEEK_END);
2837         if (size < 0) {
2838             return -errno;
2839         }
2840     }
2841     return size;
2842 }
2843 #else
2844 static int64_t raw_getlength(BlockDriverState *bs)
2845 {
2846     BDRVRawState *s = bs->opaque;
2847     int ret;
2848     int64_t size;
2849 
2850     ret = fd_open(bs);
2851     if (ret < 0) {
2852         return ret;
2853     }
2854 
2855     size = lseek(s->fd, 0, SEEK_END);
2856     if (size < 0) {
2857         return -errno;
2858     }
2859     return size;
2860 }
2861 #endif
2862 
2863 static int64_t coroutine_fn raw_co_getlength(BlockDriverState *bs)
2864 {
2865     return raw_getlength(bs);
2866 }
2867 
2868 static int64_t coroutine_fn raw_co_get_allocated_file_size(BlockDriverState *bs)
2869 {
2870     struct stat st;
2871     BDRVRawState *s = bs->opaque;
2872 
2873     if (fstat(s->fd, &st) < 0) {
2874         return -errno;
2875     }
2876     return (int64_t)st.st_blocks * 512;
2877 }
2878 
2879 static int coroutine_fn
2880 raw_co_create(BlockdevCreateOptions *options, Error **errp)
2881 {
2882     BlockdevCreateOptionsFile *file_opts;
2883     Error *local_err = NULL;
2884     int fd;
2885     uint64_t perm, shared;
2886     int result = 0;
2887 
2888     /* Validate options and set default values */
2889     assert(options->driver == BLOCKDEV_DRIVER_FILE);
2890     file_opts = &options->u.file;
2891 
2892     if (!file_opts->has_nocow) {
2893         file_opts->nocow = false;
2894     }
2895     if (!file_opts->has_preallocation) {
2896         file_opts->preallocation = PREALLOC_MODE_OFF;
2897     }
2898     if (!file_opts->has_extent_size_hint) {
2899         file_opts->extent_size_hint = 1 * MiB;
2900     }
2901     if (file_opts->extent_size_hint > UINT32_MAX) {
2902         result = -EINVAL;
2903         error_setg(errp, "Extent size hint is too large");
2904         goto out;
2905     }
2906 
2907     /* Create file */
2908     fd = qemu_create(file_opts->filename, O_RDWR | O_BINARY, 0644, errp);
2909     if (fd < 0) {
2910         result = -errno;
2911         goto out;
2912     }
2913 
2914     /* Take permissions: We want to discard everything, so we need
2915      * BLK_PERM_WRITE; and truncation to the desired size requires
2916      * BLK_PERM_RESIZE.
2917      * On the other hand, we cannot share the RESIZE permission
2918      * because we promise that after this function, the file has the
2919      * size given in the options.  If someone else were to resize it
2920      * concurrently, we could not guarantee that.
2921      * Note that after this function, we can no longer guarantee that
2922      * the file is not touched by a third party, so it may be resized
2923      * then. */
2924     perm = BLK_PERM_WRITE | BLK_PERM_RESIZE;
2925     shared = BLK_PERM_ALL & ~BLK_PERM_RESIZE;
2926 
2927     /* Step one: Take locks */
2928     result = raw_apply_lock_bytes(NULL, fd, perm, ~shared, false, errp);
2929     if (result < 0) {
2930         goto out_close;
2931     }
2932 
2933     /* Step two: Check that nobody else has taken conflicting locks */
2934     result = raw_check_lock_bytes(fd, perm, shared, errp);
2935     if (result < 0) {
2936         error_append_hint(errp,
2937                           "Is another process using the image [%s]?\n",
2938                           file_opts->filename);
2939         goto out_unlock;
2940     }
2941 
2942     /* Clear the file by truncating it to 0 */
2943     result = raw_regular_truncate(NULL, fd, 0, PREALLOC_MODE_OFF, errp);
2944     if (result < 0) {
2945         goto out_unlock;
2946     }
2947 
2948     if (file_opts->nocow) {
2949 #ifdef __linux__
2950         /* Set NOCOW flag to solve performance issue on fs like btrfs.
2951          * This is an optimisation. The FS_IOC_SETFLAGS ioctl return value
2952          * will be ignored since any failure of this operation should not
2953          * block the left work.
2954          */
2955         int attr;
2956         if (ioctl(fd, FS_IOC_GETFLAGS, &attr) == 0) {
2957             attr |= FS_NOCOW_FL;
2958             ioctl(fd, FS_IOC_SETFLAGS, &attr);
2959         }
2960 #endif
2961     }
2962 #ifdef FS_IOC_FSSETXATTR
2963     /*
2964      * Try to set the extent size hint. Failure is not fatal, and a warning is
2965      * only printed if the option was explicitly specified.
2966      */
2967     {
2968         struct fsxattr attr;
2969         result = ioctl(fd, FS_IOC_FSGETXATTR, &attr);
2970         if (result == 0) {
2971             attr.fsx_xflags |= FS_XFLAG_EXTSIZE;
2972             attr.fsx_extsize = file_opts->extent_size_hint;
2973             result = ioctl(fd, FS_IOC_FSSETXATTR, &attr);
2974         }
2975         if (result < 0 && file_opts->has_extent_size_hint &&
2976             file_opts->extent_size_hint)
2977         {
2978             warn_report("Failed to set extent size hint: %s",
2979                         strerror(errno));
2980         }
2981     }
2982 #endif
2983 
2984     /* Resize and potentially preallocate the file to the desired
2985      * final size */
2986     result = raw_regular_truncate(NULL, fd, file_opts->size,
2987                                   file_opts->preallocation, errp);
2988     if (result < 0) {
2989         goto out_unlock;
2990     }
2991 
2992 out_unlock:
2993     raw_apply_lock_bytes(NULL, fd, 0, 0, true, &local_err);
2994     if (local_err) {
2995         /* The above call should not fail, and if it does, that does
2996          * not mean the whole creation operation has failed.  So
2997          * report it the user for their convenience, but do not report
2998          * it to the caller. */
2999         warn_report_err(local_err);
3000     }
3001 
3002 out_close:
3003     if (qemu_close(fd) != 0 && result == 0) {
3004         result = -errno;
3005         error_setg_errno(errp, -result, "Could not close the new file");
3006     }
3007 out:
3008     return result;
3009 }
3010 
3011 static int coroutine_fn GRAPH_RDLOCK
3012 raw_co_create_opts(BlockDriver *drv, const char *filename,
3013                    QemuOpts *opts, Error **errp)
3014 {
3015     BlockdevCreateOptions options;
3016     int64_t total_size = 0;
3017     int64_t extent_size_hint = 0;
3018     bool has_extent_size_hint = false;
3019     bool nocow = false;
3020     PreallocMode prealloc;
3021     char *buf = NULL;
3022     Error *local_err = NULL;
3023 
3024     /* Skip file: protocol prefix */
3025     strstart(filename, "file:", &filename);
3026 
3027     /* Read out options */
3028     total_size = ROUND_UP(qemu_opt_get_size_del(opts, BLOCK_OPT_SIZE, 0),
3029                           BDRV_SECTOR_SIZE);
3030     if (qemu_opt_get(opts, BLOCK_OPT_EXTENT_SIZE_HINT)) {
3031         has_extent_size_hint = true;
3032         extent_size_hint =
3033             qemu_opt_get_size_del(opts, BLOCK_OPT_EXTENT_SIZE_HINT, -1);
3034     }
3035     nocow = qemu_opt_get_bool(opts, BLOCK_OPT_NOCOW, false);
3036     buf = qemu_opt_get_del(opts, BLOCK_OPT_PREALLOC);
3037     prealloc = qapi_enum_parse(&PreallocMode_lookup, buf,
3038                                PREALLOC_MODE_OFF, &local_err);
3039     g_free(buf);
3040     if (local_err) {
3041         error_propagate(errp, local_err);
3042         return -EINVAL;
3043     }
3044 
3045     options = (BlockdevCreateOptions) {
3046         .driver     = BLOCKDEV_DRIVER_FILE,
3047         .u.file     = {
3048             .filename           = (char *) filename,
3049             .size               = total_size,
3050             .has_preallocation  = true,
3051             .preallocation      = prealloc,
3052             .has_nocow          = true,
3053             .nocow              = nocow,
3054             .has_extent_size_hint = has_extent_size_hint,
3055             .extent_size_hint   = extent_size_hint,
3056         },
3057     };
3058     return raw_co_create(&options, errp);
3059 }
3060 
3061 static int coroutine_fn raw_co_delete_file(BlockDriverState *bs,
3062                                            Error **errp)
3063 {
3064     struct stat st;
3065     int ret;
3066 
3067     if (!(stat(bs->filename, &st) == 0) || !S_ISREG(st.st_mode)) {
3068         error_setg_errno(errp, ENOENT, "%s is not a regular file",
3069                          bs->filename);
3070         return -ENOENT;
3071     }
3072 
3073     ret = unlink(bs->filename);
3074     if (ret < 0) {
3075         ret = -errno;
3076         error_setg_errno(errp, -ret, "Error when deleting file %s",
3077                          bs->filename);
3078     }
3079 
3080     return ret;
3081 }
3082 
3083 /*
3084  * Find allocation range in @bs around offset @start.
3085  * May change underlying file descriptor's file offset.
3086  * If @start is not in a hole, store @start in @data, and the
3087  * beginning of the next hole in @hole, and return 0.
3088  * If @start is in a non-trailing hole, store @start in @hole and the
3089  * beginning of the next non-hole in @data, and return 0.
3090  * If @start is in a trailing hole or beyond EOF, return -ENXIO.
3091  * If we can't find out, return a negative errno other than -ENXIO.
3092  */
3093 static int find_allocation(BlockDriverState *bs, off_t start,
3094                            off_t *data, off_t *hole)
3095 {
3096 #if defined SEEK_HOLE && defined SEEK_DATA
3097     BDRVRawState *s = bs->opaque;
3098     off_t offs;
3099 
3100     /*
3101      * SEEK_DATA cases:
3102      * D1. offs == start: start is in data
3103      * D2. offs > start: start is in a hole, next data at offs
3104      * D3. offs < 0, errno = ENXIO: either start is in a trailing hole
3105      *                              or start is beyond EOF
3106      *     If the latter happens, the file has been truncated behind
3107      *     our back since we opened it.  All bets are off then.
3108      *     Treating like a trailing hole is simplest.
3109      * D4. offs < 0, errno != ENXIO: we learned nothing
3110      */
3111     offs = lseek(s->fd, start, SEEK_DATA);
3112     if (offs < 0) {
3113         return -errno;          /* D3 or D4 */
3114     }
3115 
3116     if (offs < start) {
3117         /* This is not a valid return by lseek().  We are safe to just return
3118          * -EIO in this case, and we'll treat it like D4. */
3119         return -EIO;
3120     }
3121 
3122     if (offs > start) {
3123         /* D2: in hole, next data at offs */
3124         *hole = start;
3125         *data = offs;
3126         return 0;
3127     }
3128 
3129     /* D1: in data, end not yet known */
3130 
3131     /*
3132      * SEEK_HOLE cases:
3133      * H1. offs == start: start is in a hole
3134      *     If this happens here, a hole has been dug behind our back
3135      *     since the previous lseek().
3136      * H2. offs > start: either start is in data, next hole at offs,
3137      *                   or start is in trailing hole, EOF at offs
3138      *     Linux treats trailing holes like any other hole: offs ==
3139      *     start.  Solaris seeks to EOF instead: offs > start (blech).
3140      *     If that happens here, a hole has been dug behind our back
3141      *     since the previous lseek().
3142      * H3. offs < 0, errno = ENXIO: start is beyond EOF
3143      *     If this happens, the file has been truncated behind our
3144      *     back since we opened it.  Treat it like a trailing hole.
3145      * H4. offs < 0, errno != ENXIO: we learned nothing
3146      *     Pretend we know nothing at all, i.e. "forget" about D1.
3147      */
3148     offs = lseek(s->fd, start, SEEK_HOLE);
3149     if (offs < 0) {
3150         return -errno;          /* D1 and (H3 or H4) */
3151     }
3152 
3153     if (offs < start) {
3154         /* This is not a valid return by lseek().  We are safe to just return
3155          * -EIO in this case, and we'll treat it like H4. */
3156         return -EIO;
3157     }
3158 
3159     if (offs > start) {
3160         /*
3161          * D1 and H2: either in data, next hole at offs, or it was in
3162          * data but is now in a trailing hole.  In the latter case,
3163          * all bets are off.  Treating it as if it there was data all
3164          * the way to EOF is safe, so simply do that.
3165          */
3166         *data = start;
3167         *hole = offs;
3168         return 0;
3169     }
3170 
3171     /* D1 and H1 */
3172     return -EBUSY;
3173 #else
3174     return -ENOTSUP;
3175 #endif
3176 }
3177 
3178 /*
3179  * Returns the allocation status of the specified offset.
3180  *
3181  * The block layer guarantees 'offset' and 'bytes' are within bounds.
3182  *
3183  * 'pnum' is set to the number of bytes (including and immediately following
3184  * the specified offset) that are known to be in the same
3185  * allocated/unallocated state.
3186  *
3187  * 'bytes' is a soft cap for 'pnum'.  If the information is free, 'pnum' may
3188  * well exceed it.
3189  */
3190 static int coroutine_fn raw_co_block_status(BlockDriverState *bs,
3191                                             bool want_zero,
3192                                             int64_t offset,
3193                                             int64_t bytes, int64_t *pnum,
3194                                             int64_t *map,
3195                                             BlockDriverState **file)
3196 {
3197     off_t data = 0, hole = 0;
3198     int ret;
3199 
3200     assert(QEMU_IS_ALIGNED(offset | bytes, bs->bl.request_alignment));
3201 
3202     ret = fd_open(bs);
3203     if (ret < 0) {
3204         return ret;
3205     }
3206 
3207     if (!want_zero) {
3208         *pnum = bytes;
3209         *map = offset;
3210         *file = bs;
3211         return BDRV_BLOCK_DATA | BDRV_BLOCK_OFFSET_VALID;
3212     }
3213 
3214     ret = find_allocation(bs, offset, &data, &hole);
3215     if (ret == -ENXIO) {
3216         /* Trailing hole */
3217         *pnum = bytes;
3218         ret = BDRV_BLOCK_ZERO;
3219     } else if (ret < 0) {
3220         /* No info available, so pretend there are no holes */
3221         *pnum = bytes;
3222         ret = BDRV_BLOCK_DATA;
3223     } else if (data == offset) {
3224         /* On a data extent, compute bytes to the end of the extent,
3225          * possibly including a partial sector at EOF. */
3226         *pnum = hole - offset;
3227 
3228         /*
3229          * We are not allowed to return partial sectors, though, so
3230          * round up if necessary.
3231          */
3232         if (!QEMU_IS_ALIGNED(*pnum, bs->bl.request_alignment)) {
3233             int64_t file_length = raw_getlength(bs);
3234             if (file_length > 0) {
3235                 /* Ignore errors, this is just a safeguard */
3236                 assert(hole == file_length);
3237             }
3238             *pnum = ROUND_UP(*pnum, bs->bl.request_alignment);
3239         }
3240 
3241         ret = BDRV_BLOCK_DATA;
3242     } else {
3243         /* On a hole, compute bytes to the beginning of the next extent.  */
3244         assert(hole == offset);
3245         *pnum = data - offset;
3246         ret = BDRV_BLOCK_ZERO;
3247     }
3248     *map = offset;
3249     *file = bs;
3250     return ret | BDRV_BLOCK_OFFSET_VALID;
3251 }
3252 
3253 #if defined(__linux__)
3254 /* Verify that the file is not in the page cache */
3255 static void check_cache_dropped(BlockDriverState *bs, Error **errp)
3256 {
3257     const size_t window_size = 128 * 1024 * 1024;
3258     BDRVRawState *s = bs->opaque;
3259     void *window = NULL;
3260     size_t length = 0;
3261     unsigned char *vec;
3262     size_t page_size;
3263     off_t offset;
3264     off_t end;
3265 
3266     /* mincore(2) page status information requires 1 byte per page */
3267     page_size = sysconf(_SC_PAGESIZE);
3268     vec = g_malloc(DIV_ROUND_UP(window_size, page_size));
3269 
3270     end = raw_getlength(bs);
3271 
3272     for (offset = 0; offset < end; offset += window_size) {
3273         void *new_window;
3274         size_t new_length;
3275         size_t vec_end;
3276         size_t i;
3277         int ret;
3278 
3279         /* Unmap previous window if size has changed */
3280         new_length = MIN(end - offset, window_size);
3281         if (new_length != length) {
3282             munmap(window, length);
3283             window = NULL;
3284             length = 0;
3285         }
3286 
3287         new_window = mmap(window, new_length, PROT_NONE, MAP_PRIVATE,
3288                           s->fd, offset);
3289         if (new_window == MAP_FAILED) {
3290             error_setg_errno(errp, errno, "mmap failed");
3291             break;
3292         }
3293 
3294         window = new_window;
3295         length = new_length;
3296 
3297         ret = mincore(window, length, vec);
3298         if (ret < 0) {
3299             error_setg_errno(errp, errno, "mincore failed");
3300             break;
3301         }
3302 
3303         vec_end = DIV_ROUND_UP(length, page_size);
3304         for (i = 0; i < vec_end; i++) {
3305             if (vec[i] & 0x1) {
3306                 break;
3307             }
3308         }
3309         if (i < vec_end) {
3310             error_setg(errp, "page cache still in use!");
3311             break;
3312         }
3313     }
3314 
3315     if (window) {
3316         munmap(window, length);
3317     }
3318 
3319     g_free(vec);
3320 }
3321 #endif /* __linux__ */
3322 
3323 static void coroutine_fn GRAPH_RDLOCK
3324 raw_co_invalidate_cache(BlockDriverState *bs, Error **errp)
3325 {
3326     BDRVRawState *s = bs->opaque;
3327     int ret;
3328 
3329     ret = fd_open(bs);
3330     if (ret < 0) {
3331         error_setg_errno(errp, -ret, "The file descriptor is not open");
3332         return;
3333     }
3334 
3335     if (!s->drop_cache) {
3336         return;
3337     }
3338 
3339     if (s->open_flags & O_DIRECT) {
3340         return; /* No host kernel page cache */
3341     }
3342 
3343 #if defined(__linux__)
3344     /* This sets the scene for the next syscall... */
3345     ret = bdrv_co_flush(bs);
3346     if (ret < 0) {
3347         error_setg_errno(errp, -ret, "flush failed");
3348         return;
3349     }
3350 
3351     /* Linux does not invalidate pages that are dirty, locked, or mmapped by a
3352      * process.  These limitations are okay because we just fsynced the file,
3353      * we don't use mmap, and the file should not be in use by other processes.
3354      */
3355     ret = posix_fadvise(s->fd, 0, 0, POSIX_FADV_DONTNEED);
3356     if (ret != 0) { /* the return value is a positive errno */
3357         error_setg_errno(errp, ret, "fadvise failed");
3358         return;
3359     }
3360 
3361     if (s->check_cache_dropped) {
3362         check_cache_dropped(bs, errp);
3363     }
3364 #else /* __linux__ */
3365     /* Do nothing.  Live migration to a remote host with cache.direct=off is
3366      * unsupported on other host operating systems.  Cache consistency issues
3367      * may occur but no error is reported here, partly because that's the
3368      * historical behavior and partly because it's hard to differentiate valid
3369      * configurations that should not cause errors.
3370      */
3371 #endif /* !__linux__ */
3372 }
3373 
3374 static void raw_account_discard(BDRVRawState *s, uint64_t nbytes, int ret)
3375 {
3376     if (ret) {
3377         s->stats.discard_nb_failed++;
3378     } else {
3379         s->stats.discard_nb_ok++;
3380         s->stats.discard_bytes_ok += nbytes;
3381     }
3382 }
3383 
3384 /*
3385  * zone report - Get a zone block device's information in the form
3386  * of an array of zone descriptors.
3387  * zones is an array of zone descriptors to hold zone information on reply;
3388  * offset can be any byte within the entire size of the device;
3389  * nr_zones is the maximum number of sectors the command should operate on.
3390  */
3391 #if defined(CONFIG_BLKZONED)
3392 static int coroutine_fn raw_co_zone_report(BlockDriverState *bs, int64_t offset,
3393                                            unsigned int *nr_zones,
3394                                            BlockZoneDescriptor *zones) {
3395     BDRVRawState *s = bs->opaque;
3396     RawPosixAIOData acb = (RawPosixAIOData) {
3397         .bs         = bs,
3398         .aio_fildes = s->fd,
3399         .aio_type   = QEMU_AIO_ZONE_REPORT,
3400         .aio_offset = offset,
3401         .zone_report    = {
3402             .nr_zones       = nr_zones,
3403             .zones          = zones,
3404         },
3405     };
3406 
3407     trace_zbd_zone_report(bs, *nr_zones, offset >> BDRV_SECTOR_BITS);
3408     return raw_thread_pool_submit(handle_aiocb_zone_report, &acb);
3409 }
3410 #endif
3411 
3412 /*
3413  * zone management operations - Execute an operation on a zone
3414  */
3415 #if defined(CONFIG_BLKZONED)
3416 static int coroutine_fn raw_co_zone_mgmt(BlockDriverState *bs, BlockZoneOp op,
3417         int64_t offset, int64_t len) {
3418     BDRVRawState *s = bs->opaque;
3419     RawPosixAIOData acb;
3420     int64_t zone_size, zone_size_mask;
3421     const char *op_name;
3422     unsigned long zo;
3423     int ret;
3424     BlockZoneWps *wps = bs->wps;
3425     int64_t capacity = bs->total_sectors << BDRV_SECTOR_BITS;
3426 
3427     zone_size = bs->bl.zone_size;
3428     zone_size_mask = zone_size - 1;
3429     if (offset & zone_size_mask) {
3430         error_report("sector offset %" PRId64 " is not aligned to zone size "
3431                      "%" PRId64 "", offset / 512, zone_size / 512);
3432         return -EINVAL;
3433     }
3434 
3435     if (((offset + len) < capacity && len & zone_size_mask) ||
3436         offset + len > capacity) {
3437         error_report("number of sectors %" PRId64 " is not aligned to zone size"
3438                       " %" PRId64 "", len / 512, zone_size / 512);
3439         return -EINVAL;
3440     }
3441 
3442     uint32_t i = offset / bs->bl.zone_size;
3443     uint32_t nrz = len / bs->bl.zone_size;
3444     uint64_t *wp = &wps->wp[i];
3445     if (BDRV_ZT_IS_CONV(*wp) && len != capacity) {
3446         error_report("zone mgmt operations are not allowed for conventional zones");
3447         return -EIO;
3448     }
3449 
3450     switch (op) {
3451     case BLK_ZO_OPEN:
3452         op_name = "BLKOPENZONE";
3453         zo = BLKOPENZONE;
3454         break;
3455     case BLK_ZO_CLOSE:
3456         op_name = "BLKCLOSEZONE";
3457         zo = BLKCLOSEZONE;
3458         break;
3459     case BLK_ZO_FINISH:
3460         op_name = "BLKFINISHZONE";
3461         zo = BLKFINISHZONE;
3462         break;
3463     case BLK_ZO_RESET:
3464         op_name = "BLKRESETZONE";
3465         zo = BLKRESETZONE;
3466         break;
3467     default:
3468         error_report("Unsupported zone op: 0x%x", op);
3469         return -ENOTSUP;
3470     }
3471 
3472     acb = (RawPosixAIOData) {
3473         .bs             = bs,
3474         .aio_fildes     = s->fd,
3475         .aio_type       = QEMU_AIO_ZONE_MGMT,
3476         .aio_offset     = offset,
3477         .aio_nbytes     = len,
3478         .zone_mgmt  = {
3479             .op = zo,
3480         },
3481     };
3482 
3483     trace_zbd_zone_mgmt(bs, op_name, offset >> BDRV_SECTOR_BITS,
3484                         len >> BDRV_SECTOR_BITS);
3485     ret = raw_thread_pool_submit(handle_aiocb_zone_mgmt, &acb);
3486     if (ret != 0) {
3487         update_zones_wp(bs, s->fd, offset, nrz);
3488         error_report("ioctl %s failed %d", op_name, ret);
3489         return ret;
3490     }
3491 
3492     if (zo == BLKRESETZONE && len == capacity) {
3493         ret = get_zones_wp(bs, s->fd, 0, bs->bl.nr_zones, 1);
3494         if (ret < 0) {
3495             error_report("reporting single wp failed");
3496             return ret;
3497         }
3498     } else if (zo == BLKRESETZONE) {
3499         for (unsigned int j = 0; j < nrz; ++j) {
3500             wp[j] = offset + j * zone_size;
3501         }
3502     } else if (zo == BLKFINISHZONE) {
3503         for (unsigned int j = 0; j < nrz; ++j) {
3504             /* The zoned device allows the last zone smaller that the
3505              * zone size. */
3506             wp[j] = MIN(offset + (j + 1) * zone_size, offset + len);
3507         }
3508     }
3509 
3510     return ret;
3511 }
3512 #endif
3513 
3514 #if defined(CONFIG_BLKZONED)
3515 static int coroutine_fn raw_co_zone_append(BlockDriverState *bs,
3516                                            int64_t *offset,
3517                                            QEMUIOVector *qiov,
3518                                            BdrvRequestFlags flags) {
3519     assert(flags == 0);
3520     int64_t zone_size_mask = bs->bl.zone_size - 1;
3521     int64_t iov_len = 0;
3522     int64_t len = 0;
3523 
3524     if (*offset & zone_size_mask) {
3525         error_report("sector offset %" PRId64 " is not aligned to zone size "
3526                      "%" PRId32 "", *offset / 512, bs->bl.zone_size / 512);
3527         return -EINVAL;
3528     }
3529 
3530     int64_t wg = bs->bl.write_granularity;
3531     int64_t wg_mask = wg - 1;
3532     for (int i = 0; i < qiov->niov; i++) {
3533         iov_len = qiov->iov[i].iov_len;
3534         if (iov_len & wg_mask) {
3535             error_report("len of IOVector[%d] %" PRId64 " is not aligned to "
3536                          "block size %" PRId64 "", i, iov_len, wg);
3537             return -EINVAL;
3538         }
3539         len += iov_len;
3540     }
3541 
3542     trace_zbd_zone_append(bs, *offset >> BDRV_SECTOR_BITS);
3543     return raw_co_prw(bs, offset, len, qiov, QEMU_AIO_ZONE_APPEND);
3544 }
3545 #endif
3546 
3547 static coroutine_fn int
3548 raw_do_pdiscard(BlockDriverState *bs, int64_t offset, int64_t bytes,
3549                 bool blkdev)
3550 {
3551     BDRVRawState *s = bs->opaque;
3552     RawPosixAIOData acb;
3553     int ret;
3554 
3555     acb = (RawPosixAIOData) {
3556         .bs             = bs,
3557         .aio_fildes     = s->fd,
3558         .aio_type       = QEMU_AIO_DISCARD,
3559         .aio_offset     = offset,
3560         .aio_nbytes     = bytes,
3561     };
3562 
3563     if (blkdev) {
3564         acb.aio_type |= QEMU_AIO_BLKDEV;
3565     }
3566 
3567     ret = raw_thread_pool_submit(handle_aiocb_discard, &acb);
3568     raw_account_discard(s, bytes, ret);
3569     return ret;
3570 }
3571 
3572 static coroutine_fn int
3573 raw_co_pdiscard(BlockDriverState *bs, int64_t offset, int64_t bytes)
3574 {
3575     return raw_do_pdiscard(bs, offset, bytes, false);
3576 }
3577 
3578 static int coroutine_fn
3579 raw_do_pwrite_zeroes(BlockDriverState *bs, int64_t offset, int64_t bytes,
3580                      BdrvRequestFlags flags, bool blkdev)
3581 {
3582     BDRVRawState *s = bs->opaque;
3583     RawPosixAIOData acb;
3584     ThreadPoolFunc *handler;
3585 
3586 #ifdef CONFIG_FALLOCATE
3587     if (offset + bytes > bs->total_sectors * BDRV_SECTOR_SIZE) {
3588         BdrvTrackedRequest *req;
3589 
3590         /*
3591          * This is a workaround for a bug in the Linux XFS driver,
3592          * where writes submitted through the AIO interface will be
3593          * discarded if they happen beyond a concurrently running
3594          * fallocate() that increases the file length (i.e., both the
3595          * write and the fallocate() happen beyond the EOF).
3596          *
3597          * To work around it, we extend the tracked request for this
3598          * zero write until INT64_MAX (effectively infinity), and mark
3599          * it as serializing.
3600          *
3601          * We have to enable this workaround for all filesystems and
3602          * AIO modes (not just XFS with aio=native), because for
3603          * remote filesystems we do not know the host configuration.
3604          */
3605 
3606         req = bdrv_co_get_self_request(bs);
3607         assert(req);
3608         assert(req->type == BDRV_TRACKED_WRITE);
3609         assert(req->offset <= offset);
3610         assert(req->offset + req->bytes >= offset + bytes);
3611 
3612         req->bytes = BDRV_MAX_LENGTH - req->offset;
3613 
3614         bdrv_check_request(req->offset, req->bytes, &error_abort);
3615 
3616         bdrv_make_request_serialising(req, bs->bl.request_alignment);
3617     }
3618 #endif
3619 
3620     acb = (RawPosixAIOData) {
3621         .bs             = bs,
3622         .aio_fildes     = s->fd,
3623         .aio_type       = QEMU_AIO_WRITE_ZEROES,
3624         .aio_offset     = offset,
3625         .aio_nbytes     = bytes,
3626     };
3627 
3628     if (blkdev) {
3629         acb.aio_type |= QEMU_AIO_BLKDEV;
3630     }
3631     if (flags & BDRV_REQ_NO_FALLBACK) {
3632         acb.aio_type |= QEMU_AIO_NO_FALLBACK;
3633     }
3634 
3635     if (flags & BDRV_REQ_MAY_UNMAP) {
3636         acb.aio_type |= QEMU_AIO_DISCARD;
3637         handler = handle_aiocb_write_zeroes_unmap;
3638     } else {
3639         handler = handle_aiocb_write_zeroes;
3640     }
3641 
3642     return raw_thread_pool_submit(handler, &acb);
3643 }
3644 
3645 static int coroutine_fn raw_co_pwrite_zeroes(
3646     BlockDriverState *bs, int64_t offset,
3647     int64_t bytes, BdrvRequestFlags flags)
3648 {
3649     return raw_do_pwrite_zeroes(bs, offset, bytes, flags, false);
3650 }
3651 
3652 static int coroutine_fn
3653 raw_co_get_info(BlockDriverState *bs, BlockDriverInfo *bdi)
3654 {
3655     return 0;
3656 }
3657 
3658 static ImageInfoSpecific *raw_get_specific_info(BlockDriverState *bs,
3659                                                 Error **errp)
3660 {
3661     ImageInfoSpecificFile *file_info = g_new0(ImageInfoSpecificFile, 1);
3662     ImageInfoSpecific *spec_info = g_new(ImageInfoSpecific, 1);
3663 
3664     *spec_info = (ImageInfoSpecific){
3665         .type = IMAGE_INFO_SPECIFIC_KIND_FILE,
3666         .u.file.data = file_info,
3667     };
3668 
3669 #ifdef FS_IOC_FSGETXATTR
3670     {
3671         BDRVRawState *s = bs->opaque;
3672         struct fsxattr attr;
3673         int ret;
3674 
3675         ret = ioctl(s->fd, FS_IOC_FSGETXATTR, &attr);
3676         if (!ret && attr.fsx_extsize != 0) {
3677             file_info->has_extent_size_hint = true;
3678             file_info->extent_size_hint = attr.fsx_extsize;
3679         }
3680     }
3681 #endif
3682 
3683     return spec_info;
3684 }
3685 
3686 static BlockStatsSpecificFile get_blockstats_specific_file(BlockDriverState *bs)
3687 {
3688     BDRVRawState *s = bs->opaque;
3689     return (BlockStatsSpecificFile) {
3690         .discard_nb_ok = s->stats.discard_nb_ok,
3691         .discard_nb_failed = s->stats.discard_nb_failed,
3692         .discard_bytes_ok = s->stats.discard_bytes_ok,
3693     };
3694 }
3695 
3696 static BlockStatsSpecific *raw_get_specific_stats(BlockDriverState *bs)
3697 {
3698     BlockStatsSpecific *stats = g_new(BlockStatsSpecific, 1);
3699 
3700     stats->driver = BLOCKDEV_DRIVER_FILE;
3701     stats->u.file = get_blockstats_specific_file(bs);
3702 
3703     return stats;
3704 }
3705 
3706 #if defined(HAVE_HOST_BLOCK_DEVICE)
3707 static BlockStatsSpecific *hdev_get_specific_stats(BlockDriverState *bs)
3708 {
3709     BlockStatsSpecific *stats = g_new(BlockStatsSpecific, 1);
3710 
3711     stats->driver = BLOCKDEV_DRIVER_HOST_DEVICE;
3712     stats->u.host_device = get_blockstats_specific_file(bs);
3713 
3714     return stats;
3715 }
3716 #endif /* HAVE_HOST_BLOCK_DEVICE */
3717 
3718 static QemuOptsList raw_create_opts = {
3719     .name = "raw-create-opts",
3720     .head = QTAILQ_HEAD_INITIALIZER(raw_create_opts.head),
3721     .desc = {
3722         {
3723             .name = BLOCK_OPT_SIZE,
3724             .type = QEMU_OPT_SIZE,
3725             .help = "Virtual disk size"
3726         },
3727         {
3728             .name = BLOCK_OPT_NOCOW,
3729             .type = QEMU_OPT_BOOL,
3730             .help = "Turn off copy-on-write (valid only on btrfs)"
3731         },
3732         {
3733             .name = BLOCK_OPT_PREALLOC,
3734             .type = QEMU_OPT_STRING,
3735             .help = "Preallocation mode (allowed values: off"
3736 #ifdef CONFIG_POSIX_FALLOCATE
3737                     ", falloc"
3738 #endif
3739                     ", full)"
3740         },
3741         {
3742             .name = BLOCK_OPT_EXTENT_SIZE_HINT,
3743             .type = QEMU_OPT_SIZE,
3744             .help = "Extent size hint for the image file, 0 to disable"
3745         },
3746         { /* end of list */ }
3747     }
3748 };
3749 
3750 static int raw_check_perm(BlockDriverState *bs, uint64_t perm, uint64_t shared,
3751                           Error **errp)
3752 {
3753     BDRVRawState *s = bs->opaque;
3754     int input_flags = s->reopen_state ? s->reopen_state->flags : bs->open_flags;
3755     int open_flags;
3756     int ret;
3757 
3758     /* We may need a new fd if auto-read-only switches the mode */
3759     ret = raw_reconfigure_getfd(bs, input_flags, &open_flags, perm, errp);
3760     if (ret < 0) {
3761         return ret;
3762     } else if (ret != s->fd) {
3763         Error *local_err = NULL;
3764 
3765         /*
3766          * Fail already check_perm() if we can't get a working O_DIRECT
3767          * alignment with the new fd.
3768          */
3769         raw_probe_alignment(bs, ret, &local_err);
3770         if (local_err) {
3771             error_propagate(errp, local_err);
3772             return -EINVAL;
3773         }
3774 
3775         s->perm_change_fd = ret;
3776         s->perm_change_flags = open_flags;
3777     }
3778 
3779     /* Prepare permissions on old fd to avoid conflicts between old and new,
3780      * but keep everything locked that new will need. */
3781     ret = raw_handle_perm_lock(bs, RAW_PL_PREPARE, perm, shared, errp);
3782     if (ret < 0) {
3783         goto fail;
3784     }
3785 
3786     /* Copy locks to the new fd */
3787     if (s->perm_change_fd && s->use_lock) {
3788         ret = raw_apply_lock_bytes(NULL, s->perm_change_fd, perm, ~shared,
3789                                    false, errp);
3790         if (ret < 0) {
3791             raw_handle_perm_lock(bs, RAW_PL_ABORT, 0, 0, NULL);
3792             goto fail;
3793         }
3794     }
3795     return 0;
3796 
3797 fail:
3798     if (s->perm_change_fd) {
3799         qemu_close(s->perm_change_fd);
3800     }
3801     s->perm_change_fd = 0;
3802     return ret;
3803 }
3804 
3805 static void raw_set_perm(BlockDriverState *bs, uint64_t perm, uint64_t shared)
3806 {
3807     BDRVRawState *s = bs->opaque;
3808 
3809     /* For reopen, we have already switched to the new fd (.bdrv_set_perm is
3810      * called after .bdrv_reopen_commit) */
3811     if (s->perm_change_fd && s->fd != s->perm_change_fd) {
3812         qemu_close(s->fd);
3813         s->fd = s->perm_change_fd;
3814         s->open_flags = s->perm_change_flags;
3815     }
3816     s->perm_change_fd = 0;
3817 
3818     raw_handle_perm_lock(bs, RAW_PL_COMMIT, perm, shared, NULL);
3819     s->perm = perm;
3820     s->shared_perm = shared;
3821 }
3822 
3823 static void raw_abort_perm_update(BlockDriverState *bs)
3824 {
3825     BDRVRawState *s = bs->opaque;
3826 
3827     /* For reopen, .bdrv_reopen_abort is called afterwards and will close
3828      * the file descriptor. */
3829     if (s->perm_change_fd) {
3830         qemu_close(s->perm_change_fd);
3831     }
3832     s->perm_change_fd = 0;
3833 
3834     raw_handle_perm_lock(bs, RAW_PL_ABORT, 0, 0, NULL);
3835 }
3836 
3837 static int coroutine_fn GRAPH_RDLOCK raw_co_copy_range_from(
3838         BlockDriverState *bs, BdrvChild *src, int64_t src_offset,
3839         BdrvChild *dst, int64_t dst_offset, int64_t bytes,
3840         BdrvRequestFlags read_flags, BdrvRequestFlags write_flags)
3841 {
3842     return bdrv_co_copy_range_to(src, src_offset, dst, dst_offset, bytes,
3843                                  read_flags, write_flags);
3844 }
3845 
3846 static int coroutine_fn GRAPH_RDLOCK
3847 raw_co_copy_range_to(BlockDriverState *bs,
3848                      BdrvChild *src, int64_t src_offset,
3849                      BdrvChild *dst, int64_t dst_offset,
3850                      int64_t bytes, BdrvRequestFlags read_flags,
3851                      BdrvRequestFlags write_flags)
3852 {
3853     RawPosixAIOData acb;
3854     BDRVRawState *s = bs->opaque;
3855     BDRVRawState *src_s;
3856 
3857     assert(dst->bs == bs);
3858     if (src->bs->drv->bdrv_co_copy_range_to != raw_co_copy_range_to) {
3859         return -ENOTSUP;
3860     }
3861 
3862     src_s = src->bs->opaque;
3863     if (fd_open(src->bs) < 0 || fd_open(dst->bs) < 0) {
3864         return -EIO;
3865     }
3866 
3867     acb = (RawPosixAIOData) {
3868         .bs             = bs,
3869         .aio_type       = QEMU_AIO_COPY_RANGE,
3870         .aio_fildes     = src_s->fd,
3871         .aio_offset     = src_offset,
3872         .aio_nbytes     = bytes,
3873         .copy_range     = {
3874             .aio_fd2        = s->fd,
3875             .aio_offset2    = dst_offset,
3876         },
3877     };
3878 
3879     return raw_thread_pool_submit(handle_aiocb_copy_range, &acb);
3880 }
3881 
3882 BlockDriver bdrv_file = {
3883     .format_name = "file",
3884     .protocol_name = "file",
3885     .instance_size = sizeof(BDRVRawState),
3886     .bdrv_needs_filename = true,
3887     .bdrv_probe = NULL, /* no probe for protocols */
3888     .bdrv_parse_filename = raw_parse_filename,
3889     .bdrv_open      = raw_open,
3890     .bdrv_reopen_prepare = raw_reopen_prepare,
3891     .bdrv_reopen_commit = raw_reopen_commit,
3892     .bdrv_reopen_abort = raw_reopen_abort,
3893     .bdrv_close = raw_close,
3894     .bdrv_co_create = raw_co_create,
3895     .bdrv_co_create_opts = raw_co_create_opts,
3896     .bdrv_has_zero_init = bdrv_has_zero_init_1,
3897     .bdrv_co_block_status = raw_co_block_status,
3898     .bdrv_co_invalidate_cache = raw_co_invalidate_cache,
3899     .bdrv_co_pwrite_zeroes = raw_co_pwrite_zeroes,
3900     .bdrv_co_delete_file = raw_co_delete_file,
3901 
3902     .bdrv_co_preadv         = raw_co_preadv,
3903     .bdrv_co_pwritev        = raw_co_pwritev,
3904     .bdrv_co_flush_to_disk  = raw_co_flush_to_disk,
3905     .bdrv_co_pdiscard       = raw_co_pdiscard,
3906     .bdrv_co_copy_range_from = raw_co_copy_range_from,
3907     .bdrv_co_copy_range_to  = raw_co_copy_range_to,
3908     .bdrv_refresh_limits = raw_refresh_limits,
3909 
3910     .bdrv_co_truncate                   = raw_co_truncate,
3911     .bdrv_co_getlength                  = raw_co_getlength,
3912     .bdrv_co_get_info                   = raw_co_get_info,
3913     .bdrv_get_specific_info             = raw_get_specific_info,
3914     .bdrv_co_get_allocated_file_size    = raw_co_get_allocated_file_size,
3915     .bdrv_get_specific_stats = raw_get_specific_stats,
3916     .bdrv_check_perm = raw_check_perm,
3917     .bdrv_set_perm   = raw_set_perm,
3918     .bdrv_abort_perm_update = raw_abort_perm_update,
3919     .create_opts = &raw_create_opts,
3920     .mutable_opts = mutable_opts,
3921 };
3922 
3923 /***********************************************/
3924 /* host device */
3925 
3926 #if defined(HAVE_HOST_BLOCK_DEVICE)
3927 
3928 #if defined(__APPLE__) && defined(__MACH__)
3929 static kern_return_t GetBSDPath(io_iterator_t mediaIterator, char *bsdPath,
3930                                 CFIndex maxPathSize, int flags);
3931 
3932 #if !defined(MAC_OS_VERSION_12_0) \
3933     || (MAC_OS_X_VERSION_MIN_REQUIRED < MAC_OS_VERSION_12_0)
3934 #define IOMainPort IOMasterPort
3935 #endif
3936 
3937 static char *FindEjectableOpticalMedia(io_iterator_t *mediaIterator)
3938 {
3939     kern_return_t kernResult = KERN_FAILURE;
3940     mach_port_t mainPort;
3941     CFMutableDictionaryRef  classesToMatch;
3942     const char *matching_array[] = {kIODVDMediaClass, kIOCDMediaClass};
3943     char *mediaType = NULL;
3944 
3945     kernResult = IOMainPort(MACH_PORT_NULL, &mainPort);
3946     if ( KERN_SUCCESS != kernResult ) {
3947         printf("IOMainPort returned %d\n", kernResult);
3948     }
3949 
3950     int index;
3951     for (index = 0; index < ARRAY_SIZE(matching_array); index++) {
3952         classesToMatch = IOServiceMatching(matching_array[index]);
3953         if (classesToMatch == NULL) {
3954             error_report("IOServiceMatching returned NULL for %s",
3955                          matching_array[index]);
3956             continue;
3957         }
3958         CFDictionarySetValue(classesToMatch, CFSTR(kIOMediaEjectableKey),
3959                              kCFBooleanTrue);
3960         kernResult = IOServiceGetMatchingServices(mainPort, classesToMatch,
3961                                                   mediaIterator);
3962         if (kernResult != KERN_SUCCESS) {
3963             error_report("Note: IOServiceGetMatchingServices returned %d",
3964                          kernResult);
3965             continue;
3966         }
3967 
3968         /* If a match was found, leave the loop */
3969         if (*mediaIterator != 0) {
3970             trace_file_FindEjectableOpticalMedia(matching_array[index]);
3971             mediaType = g_strdup(matching_array[index]);
3972             break;
3973         }
3974     }
3975     return mediaType;
3976 }
3977 
3978 kern_return_t GetBSDPath(io_iterator_t mediaIterator, char *bsdPath,
3979                          CFIndex maxPathSize, int flags)
3980 {
3981     io_object_t     nextMedia;
3982     kern_return_t   kernResult = KERN_FAILURE;
3983     *bsdPath = '\0';
3984     nextMedia = IOIteratorNext( mediaIterator );
3985     if ( nextMedia )
3986     {
3987         CFTypeRef   bsdPathAsCFString;
3988     bsdPathAsCFString = IORegistryEntryCreateCFProperty( nextMedia, CFSTR( kIOBSDNameKey ), kCFAllocatorDefault, 0 );
3989         if ( bsdPathAsCFString ) {
3990             size_t devPathLength;
3991             strcpy( bsdPath, _PATH_DEV );
3992             if (flags & BDRV_O_NOCACHE) {
3993                 strcat(bsdPath, "r");
3994             }
3995             devPathLength = strlen( bsdPath );
3996             if ( CFStringGetCString( bsdPathAsCFString, bsdPath + devPathLength, maxPathSize - devPathLength, kCFStringEncodingASCII ) ) {
3997                 kernResult = KERN_SUCCESS;
3998             }
3999             CFRelease( bsdPathAsCFString );
4000         }
4001         IOObjectRelease( nextMedia );
4002     }
4003 
4004     return kernResult;
4005 }
4006 
4007 /* Sets up a real cdrom for use in QEMU */
4008 static bool setup_cdrom(char *bsd_path, Error **errp)
4009 {
4010     int index, num_of_test_partitions = 2, fd;
4011     char test_partition[MAXPATHLEN];
4012     bool partition_found = false;
4013 
4014     /* look for a working partition */
4015     for (index = 0; index < num_of_test_partitions; index++) {
4016         snprintf(test_partition, sizeof(test_partition), "%ss%d", bsd_path,
4017                  index);
4018         fd = qemu_open(test_partition, O_RDONLY | O_BINARY | O_LARGEFILE, NULL);
4019         if (fd >= 0) {
4020             partition_found = true;
4021             qemu_close(fd);
4022             break;
4023         }
4024     }
4025 
4026     /* if a working partition on the device was not found */
4027     if (partition_found == false) {
4028         error_setg(errp, "Failed to find a working partition on disc");
4029     } else {
4030         trace_file_setup_cdrom(test_partition);
4031         pstrcpy(bsd_path, MAXPATHLEN, test_partition);
4032     }
4033     return partition_found;
4034 }
4035 
4036 /* Prints directions on mounting and unmounting a device */
4037 static void print_unmounting_directions(const char *file_name)
4038 {
4039     error_report("If device %s is mounted on the desktop, unmount"
4040                  " it first before using it in QEMU", file_name);
4041     error_report("Command to unmount device: diskutil unmountDisk %s",
4042                  file_name);
4043     error_report("Command to mount device: diskutil mountDisk %s", file_name);
4044 }
4045 
4046 #endif /* defined(__APPLE__) && defined(__MACH__) */
4047 
4048 static int hdev_probe_device(const char *filename)
4049 {
4050     struct stat st;
4051 
4052     /* allow a dedicated CD-ROM driver to match with a higher priority */
4053     if (strstart(filename, "/dev/cdrom", NULL))
4054         return 50;
4055 
4056     if (stat(filename, &st) >= 0 &&
4057             (S_ISCHR(st.st_mode) || S_ISBLK(st.st_mode))) {
4058         return 100;
4059     }
4060 
4061     return 0;
4062 }
4063 
4064 static void hdev_parse_filename(const char *filename, QDict *options,
4065                                 Error **errp)
4066 {
4067     bdrv_parse_filename_strip_prefix(filename, "host_device:", options);
4068 }
4069 
4070 static bool hdev_is_sg(BlockDriverState *bs)
4071 {
4072 
4073 #if defined(__linux__)
4074 
4075     BDRVRawState *s = bs->opaque;
4076     struct stat st;
4077     struct sg_scsi_id scsiid;
4078     int sg_version;
4079     int ret;
4080 
4081     if (stat(bs->filename, &st) < 0 || !S_ISCHR(st.st_mode)) {
4082         return false;
4083     }
4084 
4085     ret = ioctl(s->fd, SG_GET_VERSION_NUM, &sg_version);
4086     if (ret < 0) {
4087         return false;
4088     }
4089 
4090     ret = ioctl(s->fd, SG_GET_SCSI_ID, &scsiid);
4091     if (ret >= 0) {
4092         trace_file_hdev_is_sg(scsiid.scsi_type, sg_version);
4093         return true;
4094     }
4095 
4096 #endif
4097 
4098     return false;
4099 }
4100 
4101 static int hdev_open(BlockDriverState *bs, QDict *options, int flags,
4102                      Error **errp)
4103 {
4104     BDRVRawState *s = bs->opaque;
4105     int ret;
4106 
4107 #if defined(__APPLE__) && defined(__MACH__)
4108     /*
4109      * Caution: while qdict_get_str() is fine, getting non-string types
4110      * would require more care.  When @options come from -blockdev or
4111      * blockdev_add, its members are typed according to the QAPI
4112      * schema, but when they come from -drive, they're all QString.
4113      */
4114     const char *filename = qdict_get_str(options, "filename");
4115     char bsd_path[MAXPATHLEN] = "";
4116     bool error_occurred = false;
4117 
4118     /* If using a real cdrom */
4119     if (strcmp(filename, "/dev/cdrom") == 0) {
4120         char *mediaType = NULL;
4121         kern_return_t ret_val;
4122         io_iterator_t mediaIterator = 0;
4123 
4124         mediaType = FindEjectableOpticalMedia(&mediaIterator);
4125         if (mediaType == NULL) {
4126             error_setg(errp, "Please make sure your CD/DVD is in the optical"
4127                        " drive");
4128             error_occurred = true;
4129             goto hdev_open_Mac_error;
4130         }
4131 
4132         ret_val = GetBSDPath(mediaIterator, bsd_path, sizeof(bsd_path), flags);
4133         if (ret_val != KERN_SUCCESS) {
4134             error_setg(errp, "Could not get BSD path for optical drive");
4135             error_occurred = true;
4136             goto hdev_open_Mac_error;
4137         }
4138 
4139         /* If a real optical drive was not found */
4140         if (bsd_path[0] == '\0') {
4141             error_setg(errp, "Failed to obtain bsd path for optical drive");
4142             error_occurred = true;
4143             goto hdev_open_Mac_error;
4144         }
4145 
4146         /* If using a cdrom disc and finding a partition on the disc failed */
4147         if (strncmp(mediaType, kIOCDMediaClass, 9) == 0 &&
4148             setup_cdrom(bsd_path, errp) == false) {
4149             print_unmounting_directions(bsd_path);
4150             error_occurred = true;
4151             goto hdev_open_Mac_error;
4152         }
4153 
4154         qdict_put_str(options, "filename", bsd_path);
4155 
4156 hdev_open_Mac_error:
4157         g_free(mediaType);
4158         if (mediaIterator) {
4159             IOObjectRelease(mediaIterator);
4160         }
4161         if (error_occurred) {
4162             return -ENOENT;
4163         }
4164     }
4165 #endif /* defined(__APPLE__) && defined(__MACH__) */
4166 
4167     s->type = FTYPE_FILE;
4168 
4169     ret = raw_open_common(bs, options, flags, 0, true, errp);
4170     if (ret < 0) {
4171 #if defined(__APPLE__) && defined(__MACH__)
4172         if (*bsd_path) {
4173             filename = bsd_path;
4174         }
4175         /* if a physical device experienced an error while being opened */
4176         if (strncmp(filename, "/dev/", 5) == 0) {
4177             print_unmounting_directions(filename);
4178         }
4179 #endif /* defined(__APPLE__) && defined(__MACH__) */
4180         return ret;
4181     }
4182 
4183     /* Since this does ioctl the device must be already opened */
4184     bs->sg = hdev_is_sg(bs);
4185 
4186     return ret;
4187 }
4188 
4189 #if defined(__linux__)
4190 static int coroutine_fn
4191 hdev_co_ioctl(BlockDriverState *bs, unsigned long int req, void *buf)
4192 {
4193     BDRVRawState *s = bs->opaque;
4194     RawPosixAIOData acb;
4195     int ret;
4196 
4197     ret = fd_open(bs);
4198     if (ret < 0) {
4199         return ret;
4200     }
4201 
4202     if (req == SG_IO && s->pr_mgr) {
4203         struct sg_io_hdr *io_hdr = buf;
4204         if (io_hdr->cmdp[0] == PERSISTENT_RESERVE_OUT ||
4205             io_hdr->cmdp[0] == PERSISTENT_RESERVE_IN) {
4206             return pr_manager_execute(s->pr_mgr, qemu_get_current_aio_context(),
4207                                       s->fd, io_hdr);
4208         }
4209     }
4210 
4211     acb = (RawPosixAIOData) {
4212         .bs         = bs,
4213         .aio_type   = QEMU_AIO_IOCTL,
4214         .aio_fildes = s->fd,
4215         .aio_offset = 0,
4216         .ioctl      = {
4217             .buf        = buf,
4218             .cmd        = req,
4219         },
4220     };
4221 
4222     return raw_thread_pool_submit(handle_aiocb_ioctl, &acb);
4223 }
4224 #endif /* linux */
4225 
4226 static coroutine_fn int
4227 hdev_co_pdiscard(BlockDriverState *bs, int64_t offset, int64_t bytes)
4228 {
4229     BDRVRawState *s = bs->opaque;
4230     int ret;
4231 
4232     ret = fd_open(bs);
4233     if (ret < 0) {
4234         raw_account_discard(s, bytes, ret);
4235         return ret;
4236     }
4237     return raw_do_pdiscard(bs, offset, bytes, true);
4238 }
4239 
4240 static coroutine_fn int hdev_co_pwrite_zeroes(BlockDriverState *bs,
4241     int64_t offset, int64_t bytes, BdrvRequestFlags flags)
4242 {
4243     int rc;
4244 
4245     rc = fd_open(bs);
4246     if (rc < 0) {
4247         return rc;
4248     }
4249 
4250     return raw_do_pwrite_zeroes(bs, offset, bytes, flags, true);
4251 }
4252 
4253 static BlockDriver bdrv_host_device = {
4254     .format_name        = "host_device",
4255     .protocol_name        = "host_device",
4256     .instance_size      = sizeof(BDRVRawState),
4257     .bdrv_needs_filename = true,
4258     .bdrv_probe_device  = hdev_probe_device,
4259     .bdrv_parse_filename = hdev_parse_filename,
4260     .bdrv_open          = hdev_open,
4261     .bdrv_close         = raw_close,
4262     .bdrv_reopen_prepare = raw_reopen_prepare,
4263     .bdrv_reopen_commit  = raw_reopen_commit,
4264     .bdrv_reopen_abort   = raw_reopen_abort,
4265     .bdrv_co_create_opts = bdrv_co_create_opts_simple,
4266     .create_opts         = &bdrv_create_opts_simple,
4267     .mutable_opts        = mutable_opts,
4268     .bdrv_co_invalidate_cache = raw_co_invalidate_cache,
4269     .bdrv_co_pwrite_zeroes = hdev_co_pwrite_zeroes,
4270 
4271     .bdrv_co_preadv         = raw_co_preadv,
4272     .bdrv_co_pwritev        = raw_co_pwritev,
4273     .bdrv_co_flush_to_disk  = raw_co_flush_to_disk,
4274     .bdrv_co_pdiscard       = hdev_co_pdiscard,
4275     .bdrv_co_copy_range_from = raw_co_copy_range_from,
4276     .bdrv_co_copy_range_to  = raw_co_copy_range_to,
4277     .bdrv_refresh_limits = raw_refresh_limits,
4278 
4279     .bdrv_co_truncate                   = raw_co_truncate,
4280     .bdrv_co_getlength                  = raw_co_getlength,
4281     .bdrv_co_get_info                   = raw_co_get_info,
4282     .bdrv_get_specific_info             = raw_get_specific_info,
4283     .bdrv_co_get_allocated_file_size    = raw_co_get_allocated_file_size,
4284     .bdrv_get_specific_stats = hdev_get_specific_stats,
4285     .bdrv_check_perm = raw_check_perm,
4286     .bdrv_set_perm   = raw_set_perm,
4287     .bdrv_abort_perm_update = raw_abort_perm_update,
4288     .bdrv_probe_blocksizes = hdev_probe_blocksizes,
4289     .bdrv_probe_geometry = hdev_probe_geometry,
4290 
4291     /* generic scsi device */
4292 #ifdef __linux__
4293     .bdrv_co_ioctl          = hdev_co_ioctl,
4294 #endif
4295 
4296     /* zoned device */
4297 #if defined(CONFIG_BLKZONED)
4298     /* zone management operations */
4299     .bdrv_co_zone_report = raw_co_zone_report,
4300     .bdrv_co_zone_mgmt = raw_co_zone_mgmt,
4301     .bdrv_co_zone_append = raw_co_zone_append,
4302 #endif
4303 };
4304 
4305 #if defined(__linux__) || defined(__FreeBSD__) || defined(__FreeBSD_kernel__)
4306 static void cdrom_parse_filename(const char *filename, QDict *options,
4307                                  Error **errp)
4308 {
4309     bdrv_parse_filename_strip_prefix(filename, "host_cdrom:", options);
4310 }
4311 
4312 static void cdrom_refresh_limits(BlockDriverState *bs, Error **errp)
4313 {
4314     bs->bl.has_variable_length = true;
4315     raw_refresh_limits(bs, errp);
4316 }
4317 #endif
4318 
4319 #ifdef __linux__
4320 static int cdrom_open(BlockDriverState *bs, QDict *options, int flags,
4321                       Error **errp)
4322 {
4323     BDRVRawState *s = bs->opaque;
4324 
4325     s->type = FTYPE_CD;
4326 
4327     /* open will not fail even if no CD is inserted, so add O_NONBLOCK */
4328     return raw_open_common(bs, options, flags, O_NONBLOCK, true, errp);
4329 }
4330 
4331 static int cdrom_probe_device(const char *filename)
4332 {
4333     int fd, ret;
4334     int prio = 0;
4335     struct stat st;
4336 
4337     fd = qemu_open(filename, O_RDONLY | O_NONBLOCK, NULL);
4338     if (fd < 0) {
4339         goto out;
4340     }
4341     ret = fstat(fd, &st);
4342     if (ret == -1 || !S_ISBLK(st.st_mode)) {
4343         goto outc;
4344     }
4345 
4346     /* Attempt to detect via a CDROM specific ioctl */
4347     ret = ioctl(fd, CDROM_DRIVE_STATUS, CDSL_CURRENT);
4348     if (ret >= 0)
4349         prio = 100;
4350 
4351 outc:
4352     qemu_close(fd);
4353 out:
4354     return prio;
4355 }
4356 
4357 static bool coroutine_fn cdrom_co_is_inserted(BlockDriverState *bs)
4358 {
4359     BDRVRawState *s = bs->opaque;
4360     int ret;
4361 
4362     ret = ioctl(s->fd, CDROM_DRIVE_STATUS, CDSL_CURRENT);
4363     return ret == CDS_DISC_OK;
4364 }
4365 
4366 static void coroutine_fn cdrom_co_eject(BlockDriverState *bs, bool eject_flag)
4367 {
4368     BDRVRawState *s = bs->opaque;
4369 
4370     if (eject_flag) {
4371         if (ioctl(s->fd, CDROMEJECT, NULL) < 0)
4372             perror("CDROMEJECT");
4373     } else {
4374         if (ioctl(s->fd, CDROMCLOSETRAY, NULL) < 0)
4375             perror("CDROMEJECT");
4376     }
4377 }
4378 
4379 static void coroutine_fn cdrom_co_lock_medium(BlockDriverState *bs, bool locked)
4380 {
4381     BDRVRawState *s = bs->opaque;
4382 
4383     if (ioctl(s->fd, CDROM_LOCKDOOR, locked) < 0) {
4384         /*
4385          * Note: an error can happen if the distribution automatically
4386          * mounts the CD-ROM
4387          */
4388         /* perror("CDROM_LOCKDOOR"); */
4389     }
4390 }
4391 
4392 static BlockDriver bdrv_host_cdrom = {
4393     .format_name        = "host_cdrom",
4394     .protocol_name      = "host_cdrom",
4395     .instance_size      = sizeof(BDRVRawState),
4396     .bdrv_needs_filename = true,
4397     .bdrv_probe_device	= cdrom_probe_device,
4398     .bdrv_parse_filename = cdrom_parse_filename,
4399     .bdrv_open          = cdrom_open,
4400     .bdrv_close         = raw_close,
4401     .bdrv_reopen_prepare = raw_reopen_prepare,
4402     .bdrv_reopen_commit  = raw_reopen_commit,
4403     .bdrv_reopen_abort   = raw_reopen_abort,
4404     .bdrv_co_create_opts = bdrv_co_create_opts_simple,
4405     .create_opts         = &bdrv_create_opts_simple,
4406     .mutable_opts        = mutable_opts,
4407     .bdrv_co_invalidate_cache = raw_co_invalidate_cache,
4408 
4409     .bdrv_co_preadv         = raw_co_preadv,
4410     .bdrv_co_pwritev        = raw_co_pwritev,
4411     .bdrv_co_flush_to_disk  = raw_co_flush_to_disk,
4412     .bdrv_refresh_limits    = cdrom_refresh_limits,
4413 
4414     .bdrv_co_truncate                   = raw_co_truncate,
4415     .bdrv_co_getlength                  = raw_co_getlength,
4416     .bdrv_co_get_allocated_file_size    = raw_co_get_allocated_file_size,
4417 
4418     /* removable device support */
4419     .bdrv_co_is_inserted    = cdrom_co_is_inserted,
4420     .bdrv_co_eject          = cdrom_co_eject,
4421     .bdrv_co_lock_medium    = cdrom_co_lock_medium,
4422 
4423     /* generic scsi device */
4424     .bdrv_co_ioctl      = hdev_co_ioctl,
4425 };
4426 #endif /* __linux__ */
4427 
4428 #if defined (__FreeBSD__) || defined(__FreeBSD_kernel__)
4429 static int cdrom_open(BlockDriverState *bs, QDict *options, int flags,
4430                       Error **errp)
4431 {
4432     BDRVRawState *s = bs->opaque;
4433     int ret;
4434 
4435     s->type = FTYPE_CD;
4436 
4437     ret = raw_open_common(bs, options, flags, 0, true, errp);
4438     if (ret) {
4439         return ret;
4440     }
4441 
4442     /* make sure the door isn't locked at this time */
4443     ioctl(s->fd, CDIOCALLOW);
4444     return 0;
4445 }
4446 
4447 static int cdrom_probe_device(const char *filename)
4448 {
4449     if (strstart(filename, "/dev/cd", NULL) ||
4450             strstart(filename, "/dev/acd", NULL))
4451         return 100;
4452     return 0;
4453 }
4454 
4455 static int cdrom_reopen(BlockDriverState *bs)
4456 {
4457     BDRVRawState *s = bs->opaque;
4458     int fd;
4459 
4460     /*
4461      * Force reread of possibly changed/newly loaded disc,
4462      * FreeBSD seems to not notice sometimes...
4463      */
4464     if (s->fd >= 0)
4465         qemu_close(s->fd);
4466     fd = qemu_open(bs->filename, s->open_flags, NULL);
4467     if (fd < 0) {
4468         s->fd = -1;
4469         return -EIO;
4470     }
4471     s->fd = fd;
4472 
4473     /* make sure the door isn't locked at this time */
4474     ioctl(s->fd, CDIOCALLOW);
4475     return 0;
4476 }
4477 
4478 static bool coroutine_fn cdrom_co_is_inserted(BlockDriverState *bs)
4479 {
4480     return raw_getlength(bs) > 0;
4481 }
4482 
4483 static void coroutine_fn cdrom_co_eject(BlockDriverState *bs, bool eject_flag)
4484 {
4485     BDRVRawState *s = bs->opaque;
4486 
4487     if (s->fd < 0)
4488         return;
4489 
4490     (void) ioctl(s->fd, CDIOCALLOW);
4491 
4492     if (eject_flag) {
4493         if (ioctl(s->fd, CDIOCEJECT) < 0)
4494             perror("CDIOCEJECT");
4495     } else {
4496         if (ioctl(s->fd, CDIOCCLOSE) < 0)
4497             perror("CDIOCCLOSE");
4498     }
4499 
4500     cdrom_reopen(bs);
4501 }
4502 
4503 static void coroutine_fn cdrom_co_lock_medium(BlockDriverState *bs, bool locked)
4504 {
4505     BDRVRawState *s = bs->opaque;
4506 
4507     if (s->fd < 0)
4508         return;
4509     if (ioctl(s->fd, (locked ? CDIOCPREVENT : CDIOCALLOW)) < 0) {
4510         /*
4511          * Note: an error can happen if the distribution automatically
4512          * mounts the CD-ROM
4513          */
4514         /* perror("CDROM_LOCKDOOR"); */
4515     }
4516 }
4517 
4518 static BlockDriver bdrv_host_cdrom = {
4519     .format_name        = "host_cdrom",
4520     .protocol_name      = "host_cdrom",
4521     .instance_size      = sizeof(BDRVRawState),
4522     .bdrv_needs_filename = true,
4523     .bdrv_probe_device	= cdrom_probe_device,
4524     .bdrv_parse_filename = cdrom_parse_filename,
4525     .bdrv_open          = cdrom_open,
4526     .bdrv_close         = raw_close,
4527     .bdrv_reopen_prepare = raw_reopen_prepare,
4528     .bdrv_reopen_commit  = raw_reopen_commit,
4529     .bdrv_reopen_abort   = raw_reopen_abort,
4530     .bdrv_co_create_opts = bdrv_co_create_opts_simple,
4531     .create_opts         = &bdrv_create_opts_simple,
4532     .mutable_opts       = mutable_opts,
4533 
4534     .bdrv_co_preadv         = raw_co_preadv,
4535     .bdrv_co_pwritev        = raw_co_pwritev,
4536     .bdrv_co_flush_to_disk  = raw_co_flush_to_disk,
4537     .bdrv_refresh_limits    = cdrom_refresh_limits,
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