xref: /openbmc/qemu/block/file-posix.c (revision f0e48cbd)
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 "qemu-common.h"
27 #include "qapi/error.h"
28 #include "qemu/cutils.h"
29 #include "qemu/error-report.h"
30 #include "block/block_int.h"
31 #include "qemu/module.h"
32 #include "qemu/option.h"
33 #include "qemu/units.h"
34 #include "trace.h"
35 #include "block/thread-pool.h"
36 #include "qemu/iov.h"
37 #include "block/raw-aio.h"
38 #include "qapi/qmp/qdict.h"
39 #include "qapi/qmp/qstring.h"
40 
41 #include "scsi/pr-manager.h"
42 #include "scsi/constants.h"
43 
44 #if defined(__APPLE__) && (__MACH__)
45 #include <sys/ioctl.h>
46 #if defined(HAVE_HOST_BLOCK_DEVICE)
47 #include <paths.h>
48 #include <sys/param.h>
49 #include <sys/mount.h>
50 #include <IOKit/IOKitLib.h>
51 #include <IOKit/IOBSD.h>
52 #include <IOKit/storage/IOMediaBSDClient.h>
53 #include <IOKit/storage/IOMedia.h>
54 #include <IOKit/storage/IOCDMedia.h>
55 //#include <IOKit/storage/IOCDTypes.h>
56 #include <IOKit/storage/IODVDMedia.h>
57 #include <CoreFoundation/CoreFoundation.h>
58 #endif /* defined(HAVE_HOST_BLOCK_DEVICE) */
59 #endif
60 
61 #ifdef __sun__
62 #define _POSIX_PTHREAD_SEMANTICS 1
63 #include <sys/dkio.h>
64 #endif
65 #ifdef __linux__
66 #include <sys/ioctl.h>
67 #include <sys/param.h>
68 #include <sys/syscall.h>
69 #include <sys/vfs.h>
70 #include <linux/cdrom.h>
71 #include <linux/fd.h>
72 #include <linux/fs.h>
73 #include <linux/hdreg.h>
74 #include <linux/magic.h>
75 #include <scsi/sg.h>
76 #ifdef __s390__
77 #include <asm/dasd.h>
78 #endif
79 #ifndef FS_NOCOW_FL
80 #define FS_NOCOW_FL                     0x00800000 /* Do not cow file */
81 #endif
82 #endif
83 #if defined(CONFIG_FALLOCATE_PUNCH_HOLE) || defined(CONFIG_FALLOCATE_ZERO_RANGE)
84 #include <linux/falloc.h>
85 #endif
86 #if defined (__FreeBSD__) || defined(__FreeBSD_kernel__)
87 #include <sys/disk.h>
88 #include <sys/cdio.h>
89 #endif
90 
91 #ifdef __OpenBSD__
92 #include <sys/ioctl.h>
93 #include <sys/disklabel.h>
94 #include <sys/dkio.h>
95 #endif
96 
97 #ifdef __NetBSD__
98 #include <sys/ioctl.h>
99 #include <sys/disklabel.h>
100 #include <sys/dkio.h>
101 #include <sys/disk.h>
102 #endif
103 
104 #ifdef __DragonFly__
105 #include <sys/ioctl.h>
106 #include <sys/diskslice.h>
107 #endif
108 
109 #ifdef CONFIG_XFS
110 #include <xfs/xfs.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     int perm_change_fd;
154     int perm_change_flags;
155     BDRVReopenState *reopen_state;
156 
157 #ifdef CONFIG_XFS
158     bool is_xfs:1;
159 #endif
160     bool has_discard:1;
161     bool has_write_zeroes:1;
162     bool discard_zeroes:1;
163     bool use_linux_aio:1;
164     bool use_linux_io_uring:1;
165     int page_cache_inconsistent; /* errno from fdatasync failure */
166     bool has_fallocate;
167     bool needs_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     };
224 } RawPosixAIOData;
225 
226 #if defined(__FreeBSD__) || defined(__FreeBSD_kernel__)
227 static int cdrom_reopen(BlockDriverState *bs);
228 #endif
229 
230 /*
231  * Elide EAGAIN and EACCES details when failing to lock, as this
232  * indicates that the specified file region is already locked by
233  * another process, which is considered a common scenario.
234  */
235 #define raw_lock_error_setg_errno(errp, err, fmt, ...)                  \
236     do {                                                                \
237         if ((err) == EAGAIN || (err) == EACCES) {                       \
238             error_setg((errp), (fmt), ## __VA_ARGS__);                  \
239         } else {                                                        \
240             error_setg_errno((errp), (err), (fmt), ## __VA_ARGS__);     \
241         }                                                               \
242     } while (0)
243 
244 #if defined(__NetBSD__)
245 static int raw_normalize_devicepath(const char **filename, Error **errp)
246 {
247     static char namebuf[PATH_MAX];
248     const char *dp, *fname;
249     struct stat sb;
250 
251     fname = *filename;
252     dp = strrchr(fname, '/');
253     if (lstat(fname, &sb) < 0) {
254         error_setg_file_open(errp, errno, fname);
255         return -errno;
256     }
257 
258     if (!S_ISBLK(sb.st_mode)) {
259         return 0;
260     }
261 
262     if (dp == NULL) {
263         snprintf(namebuf, PATH_MAX, "r%s", fname);
264     } else {
265         snprintf(namebuf, PATH_MAX, "%.*s/r%s",
266             (int)(dp - fname), fname, dp + 1);
267     }
268     *filename = namebuf;
269     warn_report("%s is a block device, using %s", fname, *filename);
270 
271     return 0;
272 }
273 #else
274 static int raw_normalize_devicepath(const char **filename, Error **errp)
275 {
276     return 0;
277 }
278 #endif
279 
280 /*
281  * Get logical block size via ioctl. On success store it in @sector_size_p.
282  */
283 static int probe_logical_blocksize(int fd, unsigned int *sector_size_p)
284 {
285     unsigned int sector_size;
286     bool success = false;
287     int i;
288 
289     errno = ENOTSUP;
290     static const unsigned long ioctl_list[] = {
291 #ifdef BLKSSZGET
292         BLKSSZGET,
293 #endif
294 #ifdef DKIOCGETBLOCKSIZE
295         DKIOCGETBLOCKSIZE,
296 #endif
297 #ifdef DIOCGSECTORSIZE
298         DIOCGSECTORSIZE,
299 #endif
300     };
301 
302     /* Try a few ioctls to get the right size */
303     for (i = 0; i < (int)ARRAY_SIZE(ioctl_list); i++) {
304         if (ioctl(fd, ioctl_list[i], &sector_size) >= 0) {
305             *sector_size_p = sector_size;
306             success = true;
307         }
308     }
309 
310     return success ? 0 : -errno;
311 }
312 
313 /**
314  * Get physical block size of @fd.
315  * On success, store it in @blk_size and return 0.
316  * On failure, return -errno.
317  */
318 static int probe_physical_blocksize(int fd, unsigned int *blk_size)
319 {
320 #ifdef BLKPBSZGET
321     if (ioctl(fd, BLKPBSZGET, blk_size) < 0) {
322         return -errno;
323     }
324     return 0;
325 #else
326     return -ENOTSUP;
327 #endif
328 }
329 
330 /*
331  * Returns true if no alignment restrictions are necessary even for files
332  * opened with O_DIRECT.
333  *
334  * raw_probe_alignment() probes the required alignment and assume that 1 means
335  * the probing failed, so it falls back to a safe default of 4k. This can be
336  * avoided if we know that byte alignment is okay for the file.
337  */
338 static bool dio_byte_aligned(int fd)
339 {
340 #ifdef __linux__
341     struct statfs buf;
342     int ret;
343 
344     ret = fstatfs(fd, &buf);
345     if (ret == 0 && buf.f_type == NFS_SUPER_MAGIC) {
346         return true;
347     }
348 #endif
349     return false;
350 }
351 
352 /* Check if read is allowed with given memory buffer and length.
353  *
354  * This function is used to check O_DIRECT memory buffer and request alignment.
355  */
356 static bool raw_is_io_aligned(int fd, void *buf, size_t len)
357 {
358     ssize_t ret = pread(fd, buf, len, 0);
359 
360     if (ret >= 0) {
361         return true;
362     }
363 
364 #ifdef __linux__
365     /* The Linux kernel returns EINVAL for misaligned O_DIRECT reads.  Ignore
366      * other errors (e.g. real I/O error), which could happen on a failed
367      * drive, since we only care about probing alignment.
368      */
369     if (errno != EINVAL) {
370         return true;
371     }
372 #endif
373 
374     return false;
375 }
376 
377 static void raw_probe_alignment(BlockDriverState *bs, int fd, Error **errp)
378 {
379     BDRVRawState *s = bs->opaque;
380     char *buf;
381     size_t max_align = MAX(MAX_BLOCKSIZE, qemu_real_host_page_size);
382     size_t alignments[] = {1, 512, 1024, 2048, 4096};
383 
384     /* For SCSI generic devices the alignment is not really used.
385        With buffered I/O, we don't have any restrictions. */
386     if (bdrv_is_sg(bs) || !s->needs_alignment) {
387         bs->bl.request_alignment = 1;
388         s->buf_align = 1;
389         return;
390     }
391 
392     bs->bl.request_alignment = 0;
393     s->buf_align = 0;
394     /* Let's try to use the logical blocksize for the alignment. */
395     if (probe_logical_blocksize(fd, &bs->bl.request_alignment) < 0) {
396         bs->bl.request_alignment = 0;
397     }
398 #ifdef CONFIG_XFS
399     if (s->is_xfs) {
400         struct dioattr da;
401         if (xfsctl(NULL, fd, XFS_IOC_DIOINFO, &da) >= 0) {
402             bs->bl.request_alignment = da.d_miniosz;
403             /* The kernel returns wrong information for d_mem */
404             /* s->buf_align = da.d_mem; */
405         }
406     }
407 #endif
408 
409     /*
410      * If we could not get the sizes so far, we can only guess them. First try
411      * to detect request alignment, since it is more likely to succeed. Then
412      * try to detect buf_align, which cannot be detected in some cases (e.g.
413      * Gluster). If buf_align cannot be detected, we fallback to the value of
414      * request_alignment.
415      */
416 
417     if (!bs->bl.request_alignment) {
418         int i;
419         size_t align;
420         buf = qemu_memalign(max_align, max_align);
421         for (i = 0; i < ARRAY_SIZE(alignments); i++) {
422             align = alignments[i];
423             if (raw_is_io_aligned(fd, buf, align)) {
424                 /* Fallback to safe value. */
425                 bs->bl.request_alignment = (align != 1) ? align : max_align;
426                 break;
427             }
428         }
429         qemu_vfree(buf);
430     }
431 
432     if (!s->buf_align) {
433         int i;
434         size_t align;
435         buf = qemu_memalign(max_align, 2 * max_align);
436         for (i = 0; i < ARRAY_SIZE(alignments); i++) {
437             align = alignments[i];
438             if (raw_is_io_aligned(fd, buf + align, max_align)) {
439                 /* Fallback to request_alignment. */
440                 s->buf_align = (align != 1) ? align : bs->bl.request_alignment;
441                 break;
442             }
443         }
444         qemu_vfree(buf);
445     }
446 
447     if (!s->buf_align || !bs->bl.request_alignment) {
448         error_setg(errp, "Could not find working O_DIRECT alignment");
449         error_append_hint(errp, "Try cache.direct=off\n");
450     }
451 }
452 
453 static int check_hdev_writable(int fd)
454 {
455 #if defined(BLKROGET)
456     /* Linux block devices can be configured "read-only" using blockdev(8).
457      * This is independent of device node permissions and therefore open(2)
458      * with O_RDWR succeeds.  Actual writes fail with EPERM.
459      *
460      * bdrv_open() is supposed to fail if the disk is read-only.  Explicitly
461      * check for read-only block devices so that Linux block devices behave
462      * properly.
463      */
464     struct stat st;
465     int readonly = 0;
466 
467     if (fstat(fd, &st)) {
468         return -errno;
469     }
470 
471     if (!S_ISBLK(st.st_mode)) {
472         return 0;
473     }
474 
475     if (ioctl(fd, BLKROGET, &readonly) < 0) {
476         return -errno;
477     }
478 
479     if (readonly) {
480         return -EACCES;
481     }
482 #endif /* defined(BLKROGET) */
483     return 0;
484 }
485 
486 static void raw_parse_flags(int bdrv_flags, int *open_flags, bool has_writers)
487 {
488     bool read_write = false;
489     assert(open_flags != NULL);
490 
491     *open_flags |= O_BINARY;
492     *open_flags &= ~O_ACCMODE;
493 
494     if (bdrv_flags & BDRV_O_AUTO_RDONLY) {
495         read_write = has_writers;
496     } else if (bdrv_flags & BDRV_O_RDWR) {
497         read_write = true;
498     }
499 
500     if (read_write) {
501         *open_flags |= O_RDWR;
502     } else {
503         *open_flags |= O_RDONLY;
504     }
505 
506     /* Use O_DSYNC for write-through caching, no flags for write-back caching,
507      * and O_DIRECT for no caching. */
508     if ((bdrv_flags & BDRV_O_NOCACHE)) {
509         *open_flags |= O_DIRECT;
510     }
511 }
512 
513 static void raw_parse_filename(const char *filename, QDict *options,
514                                Error **errp)
515 {
516     bdrv_parse_filename_strip_prefix(filename, "file:", options);
517 }
518 
519 static QemuOptsList raw_runtime_opts = {
520     .name = "raw",
521     .head = QTAILQ_HEAD_INITIALIZER(raw_runtime_opts.head),
522     .desc = {
523         {
524             .name = "filename",
525             .type = QEMU_OPT_STRING,
526             .help = "File name of the image",
527         },
528         {
529             .name = "aio",
530             .type = QEMU_OPT_STRING,
531             .help = "host AIO implementation (threads, native, io_uring)",
532         },
533         {
534             .name = "locking",
535             .type = QEMU_OPT_STRING,
536             .help = "file locking mode (on/off/auto, default: auto)",
537         },
538         {
539             .name = "pr-manager",
540             .type = QEMU_OPT_STRING,
541             .help = "id of persistent reservation manager object (default: none)",
542         },
543 #if defined(__linux__)
544         {
545             .name = "drop-cache",
546             .type = QEMU_OPT_BOOL,
547             .help = "invalidate page cache during live migration (default: on)",
548         },
549 #endif
550         {
551             .name = "x-check-cache-dropped",
552             .type = QEMU_OPT_BOOL,
553             .help = "check that page cache was dropped on live migration (default: off)"
554         },
555         { /* end of list */ }
556     },
557 };
558 
559 static const char *const mutable_opts[] = { "x-check-cache-dropped", NULL };
560 
561 static int raw_open_common(BlockDriverState *bs, QDict *options,
562                            int bdrv_flags, int open_flags,
563                            bool device, Error **errp)
564 {
565     BDRVRawState *s = bs->opaque;
566     QemuOpts *opts;
567     Error *local_err = NULL;
568     const char *filename = NULL;
569     const char *str;
570     BlockdevAioOptions aio, aio_default;
571     int fd, ret;
572     struct stat st;
573     OnOffAuto locking;
574 
575     opts = qemu_opts_create(&raw_runtime_opts, NULL, 0, &error_abort);
576     if (!qemu_opts_absorb_qdict(opts, options, errp)) {
577         ret = -EINVAL;
578         goto fail;
579     }
580 
581     filename = qemu_opt_get(opts, "filename");
582 
583     ret = raw_normalize_devicepath(&filename, errp);
584     if (ret != 0) {
585         goto fail;
586     }
587 
588     if (bdrv_flags & BDRV_O_NATIVE_AIO) {
589         aio_default = BLOCKDEV_AIO_OPTIONS_NATIVE;
590 #ifdef CONFIG_LINUX_IO_URING
591     } else if (bdrv_flags & BDRV_O_IO_URING) {
592         aio_default = BLOCKDEV_AIO_OPTIONS_IO_URING;
593 #endif
594     } else {
595         aio_default = BLOCKDEV_AIO_OPTIONS_THREADS;
596     }
597 
598     aio = qapi_enum_parse(&BlockdevAioOptions_lookup,
599                           qemu_opt_get(opts, "aio"),
600                           aio_default, &local_err);
601     if (local_err) {
602         error_propagate(errp, local_err);
603         ret = -EINVAL;
604         goto fail;
605     }
606 
607     s->use_linux_aio = (aio == BLOCKDEV_AIO_OPTIONS_NATIVE);
608 #ifdef CONFIG_LINUX_IO_URING
609     s->use_linux_io_uring = (aio == BLOCKDEV_AIO_OPTIONS_IO_URING);
610 #endif
611 
612     locking = qapi_enum_parse(&OnOffAuto_lookup,
613                               qemu_opt_get(opts, "locking"),
614                               ON_OFF_AUTO_AUTO, &local_err);
615     if (local_err) {
616         error_propagate(errp, local_err);
617         ret = -EINVAL;
618         goto fail;
619     }
620     switch (locking) {
621     case ON_OFF_AUTO_ON:
622         s->use_lock = true;
623         if (!qemu_has_ofd_lock()) {
624             warn_report("File lock requested but OFD locking syscall is "
625                         "unavailable, falling back to POSIX file locks");
626             error_printf("Due to the implementation, locks can be lost "
627                          "unexpectedly.\n");
628         }
629         break;
630     case ON_OFF_AUTO_OFF:
631         s->use_lock = false;
632         break;
633     case ON_OFF_AUTO_AUTO:
634         s->use_lock = qemu_has_ofd_lock();
635         break;
636     default:
637         abort();
638     }
639 
640     str = qemu_opt_get(opts, "pr-manager");
641     if (str) {
642         s->pr_mgr = pr_manager_lookup(str, &local_err);
643         if (local_err) {
644             error_propagate(errp, local_err);
645             ret = -EINVAL;
646             goto fail;
647         }
648     }
649 
650     s->drop_cache = qemu_opt_get_bool(opts, "drop-cache", true);
651     s->check_cache_dropped = qemu_opt_get_bool(opts, "x-check-cache-dropped",
652                                                false);
653 
654     s->open_flags = open_flags;
655     raw_parse_flags(bdrv_flags, &s->open_flags, false);
656 
657     s->fd = -1;
658     fd = qemu_open(filename, s->open_flags, errp);
659     ret = fd < 0 ? -errno : 0;
660 
661     if (ret < 0) {
662         if (ret == -EROFS) {
663             ret = -EACCES;
664         }
665         goto fail;
666     }
667     s->fd = fd;
668 
669     /* Check s->open_flags rather than bdrv_flags due to auto-read-only */
670     if (s->open_flags & O_RDWR) {
671         ret = check_hdev_writable(s->fd);
672         if (ret < 0) {
673             error_setg_errno(errp, -ret, "The device is not writable");
674             goto fail;
675         }
676     }
677 
678     s->perm = 0;
679     s->shared_perm = BLK_PERM_ALL;
680 
681 #ifdef CONFIG_LINUX_AIO
682      /* Currently Linux does AIO only for files opened with O_DIRECT */
683     if (s->use_linux_aio) {
684         if (!(s->open_flags & O_DIRECT)) {
685             error_setg(errp, "aio=native was specified, but it requires "
686                              "cache.direct=on, which was not specified.");
687             ret = -EINVAL;
688             goto fail;
689         }
690         if (!aio_setup_linux_aio(bdrv_get_aio_context(bs), errp)) {
691             error_prepend(errp, "Unable to use native AIO: ");
692             goto fail;
693         }
694     }
695 #else
696     if (s->use_linux_aio) {
697         error_setg(errp, "aio=native was specified, but is not supported "
698                          "in this build.");
699         ret = -EINVAL;
700         goto fail;
701     }
702 #endif /* !defined(CONFIG_LINUX_AIO) */
703 
704 #ifdef CONFIG_LINUX_IO_URING
705     if (s->use_linux_io_uring) {
706         if (!aio_setup_linux_io_uring(bdrv_get_aio_context(bs), errp)) {
707             error_prepend(errp, "Unable to use io_uring: ");
708             goto fail;
709         }
710     }
711 #else
712     if (s->use_linux_io_uring) {
713         error_setg(errp, "aio=io_uring was specified, but is not supported "
714                          "in this build.");
715         ret = -EINVAL;
716         goto fail;
717     }
718 #endif /* !defined(CONFIG_LINUX_IO_URING) */
719 
720     s->has_discard = true;
721     s->has_write_zeroes = true;
722     if ((bs->open_flags & BDRV_O_NOCACHE) != 0 && !dio_byte_aligned(s->fd)) {
723         s->needs_alignment = true;
724     }
725 
726     if (fstat(s->fd, &st) < 0) {
727         ret = -errno;
728         error_setg_errno(errp, errno, "Could not stat file");
729         goto fail;
730     }
731 
732     if (!device) {
733         if (!S_ISREG(st.st_mode)) {
734             error_setg(errp, "'%s' driver requires '%s' to be a regular file",
735                        bs->drv->format_name, bs->filename);
736             ret = -EINVAL;
737             goto fail;
738         } else {
739             s->discard_zeroes = true;
740             s->has_fallocate = true;
741         }
742     } else {
743         if (!(S_ISCHR(st.st_mode) || S_ISBLK(st.st_mode))) {
744             error_setg(errp, "'%s' driver requires '%s' to be either "
745                        "a character or block device",
746                        bs->drv->format_name, bs->filename);
747             ret = -EINVAL;
748             goto fail;
749         }
750     }
751 
752     if (S_ISBLK(st.st_mode)) {
753 #ifdef BLKDISCARDZEROES
754         unsigned int arg;
755         if (ioctl(s->fd, BLKDISCARDZEROES, &arg) == 0 && arg) {
756             s->discard_zeroes = true;
757         }
758 #endif
759 #ifdef __linux__
760         /* On Linux 3.10, BLKDISCARD leaves stale data in the page cache.  Do
761          * not rely on the contents of discarded blocks unless using O_DIRECT.
762          * Same for BLKZEROOUT.
763          */
764         if (!(bs->open_flags & BDRV_O_NOCACHE)) {
765             s->discard_zeroes = false;
766             s->has_write_zeroes = false;
767         }
768 #endif
769     }
770 #ifdef __FreeBSD__
771     if (S_ISCHR(st.st_mode)) {
772         /*
773          * The file is a char device (disk), which on FreeBSD isn't behind
774          * a pager, so force all requests to be aligned. This is needed
775          * so QEMU makes sure all IO operations on the device are aligned
776          * to sector size, or else FreeBSD will reject them with EINVAL.
777          */
778         s->needs_alignment = true;
779     }
780 #endif
781 
782 #ifdef CONFIG_XFS
783     if (platform_test_xfs_fd(s->fd)) {
784         s->is_xfs = true;
785     }
786 #endif
787 
788     bs->supported_zero_flags = BDRV_REQ_MAY_UNMAP | BDRV_REQ_NO_FALLBACK;
789     if (S_ISREG(st.st_mode)) {
790         /* When extending regular files, we get zeros from the OS */
791         bs->supported_truncate_flags = BDRV_REQ_ZERO_WRITE;
792     }
793     ret = 0;
794 fail:
795     if (ret < 0 && s->fd != -1) {
796         qemu_close(s->fd);
797     }
798     if (filename && (bdrv_flags & BDRV_O_TEMPORARY)) {
799         unlink(filename);
800     }
801     qemu_opts_del(opts);
802     return ret;
803 }
804 
805 static int raw_open(BlockDriverState *bs, QDict *options, int flags,
806                     Error **errp)
807 {
808     BDRVRawState *s = bs->opaque;
809 
810     s->type = FTYPE_FILE;
811     return raw_open_common(bs, options, flags, 0, false, errp);
812 }
813 
814 typedef enum {
815     RAW_PL_PREPARE,
816     RAW_PL_COMMIT,
817     RAW_PL_ABORT,
818 } RawPermLockOp;
819 
820 #define PERM_FOREACH(i) \
821     for ((i) = 0; (1ULL << (i)) <= BLK_PERM_ALL; i++)
822 
823 /* Lock bytes indicated by @perm_lock_bits and @shared_perm_lock_bits in the
824  * file; if @unlock == true, also unlock the unneeded bytes.
825  * @shared_perm_lock_bits is the mask of all permissions that are NOT shared.
826  */
827 static int raw_apply_lock_bytes(BDRVRawState *s, int fd,
828                                 uint64_t perm_lock_bits,
829                                 uint64_t shared_perm_lock_bits,
830                                 bool unlock, Error **errp)
831 {
832     int ret;
833     int i;
834     uint64_t locked_perm, locked_shared_perm;
835 
836     if (s) {
837         locked_perm = s->locked_perm;
838         locked_shared_perm = s->locked_shared_perm;
839     } else {
840         /*
841          * We don't have the previous bits, just lock/unlock for each of the
842          * requested bits.
843          */
844         if (unlock) {
845             locked_perm = BLK_PERM_ALL;
846             locked_shared_perm = BLK_PERM_ALL;
847         } else {
848             locked_perm = 0;
849             locked_shared_perm = 0;
850         }
851     }
852 
853     PERM_FOREACH(i) {
854         int off = RAW_LOCK_PERM_BASE + i;
855         uint64_t bit = (1ULL << i);
856         if ((perm_lock_bits & bit) && !(locked_perm & bit)) {
857             ret = qemu_lock_fd(fd, off, 1, false);
858             if (ret) {
859                 raw_lock_error_setg_errno(errp, -ret, "Failed to lock byte %d",
860                                           off);
861                 return ret;
862             } else if (s) {
863                 s->locked_perm |= bit;
864             }
865         } else if (unlock && (locked_perm & bit) && !(perm_lock_bits & bit)) {
866             ret = qemu_unlock_fd(fd, off, 1);
867             if (ret) {
868                 error_setg_errno(errp, -ret, "Failed to unlock byte %d", off);
869                 return ret;
870             } else if (s) {
871                 s->locked_perm &= ~bit;
872             }
873         }
874     }
875     PERM_FOREACH(i) {
876         int off = RAW_LOCK_SHARED_BASE + i;
877         uint64_t bit = (1ULL << i);
878         if ((shared_perm_lock_bits & bit) && !(locked_shared_perm & bit)) {
879             ret = qemu_lock_fd(fd, off, 1, false);
880             if (ret) {
881                 raw_lock_error_setg_errno(errp, -ret, "Failed to lock byte %d",
882                                           off);
883                 return ret;
884             } else if (s) {
885                 s->locked_shared_perm |= bit;
886             }
887         } else if (unlock && (locked_shared_perm & bit) &&
888                    !(shared_perm_lock_bits & bit)) {
889             ret = qemu_unlock_fd(fd, off, 1);
890             if (ret) {
891                 error_setg_errno(errp, -ret, "Failed to unlock byte %d", off);
892                 return ret;
893             } else if (s) {
894                 s->locked_shared_perm &= ~bit;
895             }
896         }
897     }
898     return 0;
899 }
900 
901 /* Check "unshared" bytes implied by @perm and ~@shared_perm in the file. */
902 static int raw_check_lock_bytes(int fd, uint64_t perm, uint64_t shared_perm,
903                                 Error **errp)
904 {
905     int ret;
906     int i;
907 
908     PERM_FOREACH(i) {
909         int off = RAW_LOCK_SHARED_BASE + i;
910         uint64_t p = 1ULL << i;
911         if (perm & p) {
912             ret = qemu_lock_fd_test(fd, off, 1, true);
913             if (ret) {
914                 char *perm_name = bdrv_perm_names(p);
915 
916                 raw_lock_error_setg_errno(errp, -ret,
917                                           "Failed to get \"%s\" lock",
918                                           perm_name);
919                 g_free(perm_name);
920                 return ret;
921             }
922         }
923     }
924     PERM_FOREACH(i) {
925         int off = RAW_LOCK_PERM_BASE + i;
926         uint64_t p = 1ULL << i;
927         if (!(shared_perm & p)) {
928             ret = qemu_lock_fd_test(fd, off, 1, true);
929             if (ret) {
930                 char *perm_name = bdrv_perm_names(p);
931 
932                 raw_lock_error_setg_errno(errp, -ret,
933                                           "Failed to get shared \"%s\" lock",
934                                           perm_name);
935                 g_free(perm_name);
936                 return ret;
937             }
938         }
939     }
940     return 0;
941 }
942 
943 static int raw_handle_perm_lock(BlockDriverState *bs,
944                                 RawPermLockOp op,
945                                 uint64_t new_perm, uint64_t new_shared,
946                                 Error **errp)
947 {
948     BDRVRawState *s = bs->opaque;
949     int ret = 0;
950     Error *local_err = NULL;
951 
952     if (!s->use_lock) {
953         return 0;
954     }
955 
956     if (bdrv_get_flags(bs) & BDRV_O_INACTIVE) {
957         return 0;
958     }
959 
960     switch (op) {
961     case RAW_PL_PREPARE:
962         if ((s->perm | new_perm) == s->perm &&
963             (s->shared_perm & new_shared) == s->shared_perm)
964         {
965             /*
966              * We are going to unlock bytes, it should not fail. If it fail due
967              * to some fs-dependent permission-unrelated reasons (which occurs
968              * sometimes on NFS and leads to abort in bdrv_replace_child) we
969              * can't prevent such errors by any check here. And we ignore them
970              * anyway in ABORT and COMMIT.
971              */
972             return 0;
973         }
974         ret = raw_apply_lock_bytes(s, s->fd, s->perm | new_perm,
975                                    ~s->shared_perm | ~new_shared,
976                                    false, errp);
977         if (!ret) {
978             ret = raw_check_lock_bytes(s->fd, new_perm, new_shared, errp);
979             if (!ret) {
980                 return 0;
981             }
982             error_append_hint(errp,
983                               "Is another process using the image [%s]?\n",
984                               bs->filename);
985         }
986         /* fall through to unlock bytes. */
987     case RAW_PL_ABORT:
988         raw_apply_lock_bytes(s, s->fd, s->perm, ~s->shared_perm,
989                              true, &local_err);
990         if (local_err) {
991             /* Theoretically the above call only unlocks bytes and it cannot
992              * fail. Something weird happened, report it.
993              */
994             warn_report_err(local_err);
995         }
996         break;
997     case RAW_PL_COMMIT:
998         raw_apply_lock_bytes(s, s->fd, new_perm, ~new_shared,
999                              true, &local_err);
1000         if (local_err) {
1001             /* Theoretically the above call only unlocks bytes and it cannot
1002              * fail. Something weird happened, report it.
1003              */
1004             warn_report_err(local_err);
1005         }
1006         break;
1007     }
1008     return ret;
1009 }
1010 
1011 static int raw_reconfigure_getfd(BlockDriverState *bs, int flags,
1012                                  int *open_flags, uint64_t perm, bool force_dup,
1013                                  Error **errp)
1014 {
1015     BDRVRawState *s = bs->opaque;
1016     int fd = -1;
1017     int ret;
1018     bool has_writers = perm &
1019         (BLK_PERM_WRITE | BLK_PERM_WRITE_UNCHANGED | BLK_PERM_RESIZE);
1020     int fcntl_flags = O_APPEND | O_NONBLOCK;
1021 #ifdef O_NOATIME
1022     fcntl_flags |= O_NOATIME;
1023 #endif
1024 
1025     *open_flags = 0;
1026     if (s->type == FTYPE_CD) {
1027         *open_flags |= O_NONBLOCK;
1028     }
1029 
1030     raw_parse_flags(flags, open_flags, has_writers);
1031 
1032 #ifdef O_ASYNC
1033     /* Not all operating systems have O_ASYNC, and those that don't
1034      * will not let us track the state into rs->open_flags (typically
1035      * you achieve the same effect with an ioctl, for example I_SETSIG
1036      * on Solaris). But we do not use O_ASYNC, so that's fine.
1037      */
1038     assert((s->open_flags & O_ASYNC) == 0);
1039 #endif
1040 
1041     if (!force_dup && *open_flags == s->open_flags) {
1042         /* We're lucky, the existing fd is fine */
1043         return s->fd;
1044     }
1045 
1046     if ((*open_flags & ~fcntl_flags) == (s->open_flags & ~fcntl_flags)) {
1047         /* dup the original fd */
1048         fd = qemu_dup(s->fd);
1049         if (fd >= 0) {
1050             ret = fcntl_setfl(fd, *open_flags);
1051             if (ret) {
1052                 qemu_close(fd);
1053                 fd = -1;
1054             }
1055         }
1056     }
1057 
1058     /* If we cannot use fcntl, or fcntl failed, fall back to qemu_open() */
1059     if (fd == -1) {
1060         const char *normalized_filename = bs->filename;
1061         ret = raw_normalize_devicepath(&normalized_filename, errp);
1062         if (ret >= 0) {
1063             fd = qemu_open(normalized_filename, *open_flags, errp);
1064             if (fd == -1) {
1065                 return -1;
1066             }
1067         }
1068     }
1069 
1070     if (fd != -1 && (*open_flags & O_RDWR)) {
1071         ret = check_hdev_writable(fd);
1072         if (ret < 0) {
1073             qemu_close(fd);
1074             error_setg_errno(errp, -ret, "The device is not writable");
1075             return -1;
1076         }
1077     }
1078 
1079     return fd;
1080 }
1081 
1082 static int raw_reopen_prepare(BDRVReopenState *state,
1083                               BlockReopenQueue *queue, Error **errp)
1084 {
1085     BDRVRawState *s;
1086     BDRVRawReopenState *rs;
1087     QemuOpts *opts;
1088     int ret;
1089 
1090     assert(state != NULL);
1091     assert(state->bs != NULL);
1092 
1093     s = state->bs->opaque;
1094 
1095     state->opaque = g_new0(BDRVRawReopenState, 1);
1096     rs = state->opaque;
1097 
1098     /* Handle options changes */
1099     opts = qemu_opts_create(&raw_runtime_opts, NULL, 0, &error_abort);
1100     if (!qemu_opts_absorb_qdict(opts, state->options, errp)) {
1101         ret = -EINVAL;
1102         goto out;
1103     }
1104 
1105     rs->drop_cache = qemu_opt_get_bool_del(opts, "drop-cache", true);
1106     rs->check_cache_dropped =
1107         qemu_opt_get_bool_del(opts, "x-check-cache-dropped", false);
1108 
1109     /* This driver's reopen function doesn't currently allow changing
1110      * other options, so let's put them back in the original QDict and
1111      * bdrv_reopen_prepare() will detect changes and complain. */
1112     qemu_opts_to_qdict(opts, state->options);
1113 
1114     /*
1115      * As part of reopen prepare we also want to create new fd by
1116      * raw_reconfigure_getfd(). But it wants updated "perm", when in
1117      * bdrv_reopen_multiple() .bdrv_reopen_prepare() callback called prior to
1118      * permission update. Happily, permission update is always a part (a seprate
1119      * stage) of bdrv_reopen_multiple() so we can rely on this fact and
1120      * reconfigure fd in raw_check_perm().
1121      */
1122 
1123     s->reopen_state = state;
1124     ret = 0;
1125 
1126 out:
1127     qemu_opts_del(opts);
1128     return ret;
1129 }
1130 
1131 static void raw_reopen_commit(BDRVReopenState *state)
1132 {
1133     BDRVRawReopenState *rs = state->opaque;
1134     BDRVRawState *s = state->bs->opaque;
1135 
1136     s->drop_cache = rs->drop_cache;
1137     s->check_cache_dropped = rs->check_cache_dropped;
1138     s->open_flags = rs->open_flags;
1139     g_free(state->opaque);
1140     state->opaque = NULL;
1141 
1142     assert(s->reopen_state == state);
1143     s->reopen_state = NULL;
1144 }
1145 
1146 
1147 static void raw_reopen_abort(BDRVReopenState *state)
1148 {
1149     BDRVRawReopenState *rs = state->opaque;
1150     BDRVRawState *s = state->bs->opaque;
1151 
1152      /* nothing to do if NULL, we didn't get far enough */
1153     if (rs == NULL) {
1154         return;
1155     }
1156 
1157     g_free(state->opaque);
1158     state->opaque = NULL;
1159 
1160     assert(s->reopen_state == state);
1161     s->reopen_state = NULL;
1162 }
1163 
1164 static int hdev_get_max_hw_transfer(int fd, struct stat *st)
1165 {
1166 #ifdef BLKSECTGET
1167     if (S_ISBLK(st->st_mode)) {
1168         unsigned short max_sectors = 0;
1169         if (ioctl(fd, BLKSECTGET, &max_sectors) == 0) {
1170             return max_sectors * 512;
1171         }
1172     } else {
1173         int max_bytes = 0;
1174         if (ioctl(fd, BLKSECTGET, &max_bytes) == 0) {
1175             return max_bytes;
1176         }
1177     }
1178     return -errno;
1179 #else
1180     return -ENOSYS;
1181 #endif
1182 }
1183 
1184 static int hdev_get_max_segments(int fd, struct stat *st)
1185 {
1186 #ifdef CONFIG_LINUX
1187     char buf[32];
1188     const char *end;
1189     char *sysfspath = NULL;
1190     int ret;
1191     int sysfd = -1;
1192     long max_segments;
1193 
1194     if (S_ISCHR(st->st_mode)) {
1195         if (ioctl(fd, SG_GET_SG_TABLESIZE, &ret) == 0) {
1196             return ret;
1197         }
1198         return -ENOTSUP;
1199     }
1200 
1201     if (!S_ISBLK(st->st_mode)) {
1202         return -ENOTSUP;
1203     }
1204 
1205     sysfspath = g_strdup_printf("/sys/dev/block/%u:%u/queue/max_segments",
1206                                 major(st->st_rdev), minor(st->st_rdev));
1207     sysfd = open(sysfspath, O_RDONLY);
1208     if (sysfd == -1) {
1209         ret = -errno;
1210         goto out;
1211     }
1212     do {
1213         ret = read(sysfd, buf, sizeof(buf) - 1);
1214     } while (ret == -1 && errno == EINTR);
1215     if (ret < 0) {
1216         ret = -errno;
1217         goto out;
1218     } else if (ret == 0) {
1219         ret = -EIO;
1220         goto out;
1221     }
1222     buf[ret] = 0;
1223     /* The file is ended with '\n', pass 'end' to accept that. */
1224     ret = qemu_strtol(buf, &end, 10, &max_segments);
1225     if (ret == 0 && end && *end == '\n') {
1226         ret = max_segments;
1227     }
1228 
1229 out:
1230     if (sysfd != -1) {
1231         close(sysfd);
1232     }
1233     g_free(sysfspath);
1234     return ret;
1235 #else
1236     return -ENOTSUP;
1237 #endif
1238 }
1239 
1240 static void raw_refresh_limits(BlockDriverState *bs, Error **errp)
1241 {
1242     BDRVRawState *s = bs->opaque;
1243     struct stat st;
1244 
1245     raw_probe_alignment(bs, s->fd, errp);
1246     bs->bl.min_mem_alignment = s->buf_align;
1247     bs->bl.opt_mem_alignment = MAX(s->buf_align, qemu_real_host_page_size);
1248 
1249     /*
1250      * Maximum transfers are best effort, so it is okay to ignore any
1251      * errors.  That said, based on the man page errors in fstat would be
1252      * very much unexpected; the only possible case seems to be ENOMEM.
1253      */
1254     if (fstat(s->fd, &st)) {
1255         return;
1256     }
1257 
1258 #if defined(__APPLE__) && (__MACH__)
1259     struct statfs buf;
1260 
1261     if (!fstatfs(s->fd, &buf)) {
1262         bs->bl.opt_transfer = buf.f_iosize;
1263         bs->bl.pdiscard_alignment = buf.f_bsize;
1264     }
1265 #endif
1266 
1267     if (bs->sg || S_ISBLK(st.st_mode)) {
1268         int ret = hdev_get_max_hw_transfer(s->fd, &st);
1269 
1270         if (ret > 0 && ret <= BDRV_REQUEST_MAX_BYTES) {
1271             bs->bl.max_hw_transfer = ret;
1272         }
1273 
1274         ret = hdev_get_max_segments(s->fd, &st);
1275         if (ret > 0) {
1276             bs->bl.max_iov = ret;
1277         }
1278     }
1279 }
1280 
1281 static int check_for_dasd(int fd)
1282 {
1283 #ifdef BIODASDINFO2
1284     struct dasd_information2_t info = {0};
1285 
1286     return ioctl(fd, BIODASDINFO2, &info);
1287 #else
1288     return -1;
1289 #endif
1290 }
1291 
1292 /**
1293  * Try to get @bs's logical and physical block size.
1294  * On success, store them in @bsz and return zero.
1295  * On failure, return negative errno.
1296  */
1297 static int hdev_probe_blocksizes(BlockDriverState *bs, BlockSizes *bsz)
1298 {
1299     BDRVRawState *s = bs->opaque;
1300     int ret;
1301 
1302     /* If DASD, get blocksizes */
1303     if (check_for_dasd(s->fd) < 0) {
1304         return -ENOTSUP;
1305     }
1306     ret = probe_logical_blocksize(s->fd, &bsz->log);
1307     if (ret < 0) {
1308         return ret;
1309     }
1310     return probe_physical_blocksize(s->fd, &bsz->phys);
1311 }
1312 
1313 /**
1314  * Try to get @bs's geometry: cyls, heads, sectors.
1315  * On success, store them in @geo and return 0.
1316  * On failure return -errno.
1317  * (Allows block driver to assign default geometry values that guest sees)
1318  */
1319 #ifdef __linux__
1320 static int hdev_probe_geometry(BlockDriverState *bs, HDGeometry *geo)
1321 {
1322     BDRVRawState *s = bs->opaque;
1323     struct hd_geometry ioctl_geo = {0};
1324 
1325     /* If DASD, get its geometry */
1326     if (check_for_dasd(s->fd) < 0) {
1327         return -ENOTSUP;
1328     }
1329     if (ioctl(s->fd, HDIO_GETGEO, &ioctl_geo) < 0) {
1330         return -errno;
1331     }
1332     /* HDIO_GETGEO may return success even though geo contains zeros
1333        (e.g. certain multipath setups) */
1334     if (!ioctl_geo.heads || !ioctl_geo.sectors || !ioctl_geo.cylinders) {
1335         return -ENOTSUP;
1336     }
1337     /* Do not return a geometry for partition */
1338     if (ioctl_geo.start != 0) {
1339         return -ENOTSUP;
1340     }
1341     geo->heads = ioctl_geo.heads;
1342     geo->sectors = ioctl_geo.sectors;
1343     geo->cylinders = ioctl_geo.cylinders;
1344 
1345     return 0;
1346 }
1347 #else /* __linux__ */
1348 static int hdev_probe_geometry(BlockDriverState *bs, HDGeometry *geo)
1349 {
1350     return -ENOTSUP;
1351 }
1352 #endif
1353 
1354 #if defined(__linux__)
1355 static int handle_aiocb_ioctl(void *opaque)
1356 {
1357     RawPosixAIOData *aiocb = opaque;
1358     int ret;
1359 
1360     do {
1361         ret = ioctl(aiocb->aio_fildes, aiocb->ioctl.cmd, aiocb->ioctl.buf);
1362     } while (ret == -1 && errno == EINTR);
1363     if (ret == -1) {
1364         return -errno;
1365     }
1366 
1367     return 0;
1368 }
1369 #endif /* linux */
1370 
1371 static int handle_aiocb_flush(void *opaque)
1372 {
1373     RawPosixAIOData *aiocb = opaque;
1374     BDRVRawState *s = aiocb->bs->opaque;
1375     int ret;
1376 
1377     if (s->page_cache_inconsistent) {
1378         return -s->page_cache_inconsistent;
1379     }
1380 
1381     ret = qemu_fdatasync(aiocb->aio_fildes);
1382     if (ret == -1) {
1383         trace_file_flush_fdatasync_failed(errno);
1384 
1385         /* There is no clear definition of the semantics of a failing fsync(),
1386          * so we may have to assume the worst. The sad truth is that this
1387          * assumption is correct for Linux. Some pages are now probably marked
1388          * clean in the page cache even though they are inconsistent with the
1389          * on-disk contents. The next fdatasync() call would succeed, but no
1390          * further writeback attempt will be made. We can't get back to a state
1391          * in which we know what is on disk (we would have to rewrite
1392          * everything that was touched since the last fdatasync() at least), so
1393          * make bdrv_flush() fail permanently. Given that the behaviour isn't
1394          * really defined, I have little hope that other OSes are doing better.
1395          *
1396          * Obviously, this doesn't affect O_DIRECT, which bypasses the page
1397          * cache. */
1398         if ((s->open_flags & O_DIRECT) == 0) {
1399             s->page_cache_inconsistent = errno;
1400         }
1401         return -errno;
1402     }
1403     return 0;
1404 }
1405 
1406 #ifdef CONFIG_PREADV
1407 
1408 static bool preadv_present = true;
1409 
1410 static ssize_t
1411 qemu_preadv(int fd, const struct iovec *iov, int nr_iov, off_t offset)
1412 {
1413     return preadv(fd, iov, nr_iov, offset);
1414 }
1415 
1416 static ssize_t
1417 qemu_pwritev(int fd, const struct iovec *iov, int nr_iov, off_t offset)
1418 {
1419     return pwritev(fd, iov, nr_iov, offset);
1420 }
1421 
1422 #else
1423 
1424 static bool preadv_present = false;
1425 
1426 static ssize_t
1427 qemu_preadv(int fd, const struct iovec *iov, int nr_iov, off_t offset)
1428 {
1429     return -ENOSYS;
1430 }
1431 
1432 static ssize_t
1433 qemu_pwritev(int fd, const struct iovec *iov, int nr_iov, off_t offset)
1434 {
1435     return -ENOSYS;
1436 }
1437 
1438 #endif
1439 
1440 static ssize_t handle_aiocb_rw_vector(RawPosixAIOData *aiocb)
1441 {
1442     ssize_t len;
1443 
1444     do {
1445         if (aiocb->aio_type & QEMU_AIO_WRITE)
1446             len = qemu_pwritev(aiocb->aio_fildes,
1447                                aiocb->io.iov,
1448                                aiocb->io.niov,
1449                                aiocb->aio_offset);
1450          else
1451             len = qemu_preadv(aiocb->aio_fildes,
1452                               aiocb->io.iov,
1453                               aiocb->io.niov,
1454                               aiocb->aio_offset);
1455     } while (len == -1 && errno == EINTR);
1456 
1457     if (len == -1) {
1458         return -errno;
1459     }
1460     return len;
1461 }
1462 
1463 /*
1464  * Read/writes the data to/from a given linear buffer.
1465  *
1466  * Returns the number of bytes handles or -errno in case of an error. Short
1467  * reads are only returned if the end of the file is reached.
1468  */
1469 static ssize_t handle_aiocb_rw_linear(RawPosixAIOData *aiocb, char *buf)
1470 {
1471     ssize_t offset = 0;
1472     ssize_t len;
1473 
1474     while (offset < aiocb->aio_nbytes) {
1475         if (aiocb->aio_type & QEMU_AIO_WRITE) {
1476             len = pwrite(aiocb->aio_fildes,
1477                          (const char *)buf + offset,
1478                          aiocb->aio_nbytes - offset,
1479                          aiocb->aio_offset + offset);
1480         } else {
1481             len = pread(aiocb->aio_fildes,
1482                         buf + offset,
1483                         aiocb->aio_nbytes - offset,
1484                         aiocb->aio_offset + offset);
1485         }
1486         if (len == -1 && errno == EINTR) {
1487             continue;
1488         } else if (len == -1 && errno == EINVAL &&
1489                    (aiocb->bs->open_flags & BDRV_O_NOCACHE) &&
1490                    !(aiocb->aio_type & QEMU_AIO_WRITE) &&
1491                    offset > 0) {
1492             /* O_DIRECT pread() may fail with EINVAL when offset is unaligned
1493              * after a short read.  Assume that O_DIRECT short reads only occur
1494              * at EOF.  Therefore this is a short read, not an I/O error.
1495              */
1496             break;
1497         } else if (len == -1) {
1498             offset = -errno;
1499             break;
1500         } else if (len == 0) {
1501             break;
1502         }
1503         offset += len;
1504     }
1505 
1506     return offset;
1507 }
1508 
1509 static int handle_aiocb_rw(void *opaque)
1510 {
1511     RawPosixAIOData *aiocb = opaque;
1512     ssize_t nbytes;
1513     char *buf;
1514 
1515     if (!(aiocb->aio_type & QEMU_AIO_MISALIGNED)) {
1516         /*
1517          * If there is just a single buffer, and it is properly aligned
1518          * we can just use plain pread/pwrite without any problems.
1519          */
1520         if (aiocb->io.niov == 1) {
1521             nbytes = handle_aiocb_rw_linear(aiocb, aiocb->io.iov->iov_base);
1522             goto out;
1523         }
1524         /*
1525          * We have more than one iovec, and all are properly aligned.
1526          *
1527          * Try preadv/pwritev first and fall back to linearizing the
1528          * buffer if it's not supported.
1529          */
1530         if (preadv_present) {
1531             nbytes = handle_aiocb_rw_vector(aiocb);
1532             if (nbytes == aiocb->aio_nbytes ||
1533                 (nbytes < 0 && nbytes != -ENOSYS)) {
1534                 goto out;
1535             }
1536             preadv_present = false;
1537         }
1538 
1539         /*
1540          * XXX(hch): short read/write.  no easy way to handle the reminder
1541          * using these interfaces.  For now retry using plain
1542          * pread/pwrite?
1543          */
1544     }
1545 
1546     /*
1547      * Ok, we have to do it the hard way, copy all segments into
1548      * a single aligned buffer.
1549      */
1550     buf = qemu_try_blockalign(aiocb->bs, aiocb->aio_nbytes);
1551     if (buf == NULL) {
1552         nbytes = -ENOMEM;
1553         goto out;
1554     }
1555 
1556     if (aiocb->aio_type & QEMU_AIO_WRITE) {
1557         char *p = buf;
1558         int i;
1559 
1560         for (i = 0; i < aiocb->io.niov; ++i) {
1561             memcpy(p, aiocb->io.iov[i].iov_base, aiocb->io.iov[i].iov_len);
1562             p += aiocb->io.iov[i].iov_len;
1563         }
1564         assert(p - buf == aiocb->aio_nbytes);
1565     }
1566 
1567     nbytes = handle_aiocb_rw_linear(aiocb, buf);
1568     if (!(aiocb->aio_type & QEMU_AIO_WRITE)) {
1569         char *p = buf;
1570         size_t count = aiocb->aio_nbytes, copy;
1571         int i;
1572 
1573         for (i = 0; i < aiocb->io.niov && count; ++i) {
1574             copy = count;
1575             if (copy > aiocb->io.iov[i].iov_len) {
1576                 copy = aiocb->io.iov[i].iov_len;
1577             }
1578             memcpy(aiocb->io.iov[i].iov_base, p, copy);
1579             assert(count >= copy);
1580             p     += copy;
1581             count -= copy;
1582         }
1583         assert(count == 0);
1584     }
1585     qemu_vfree(buf);
1586 
1587 out:
1588     if (nbytes == aiocb->aio_nbytes) {
1589         return 0;
1590     } else if (nbytes >= 0 && nbytes < aiocb->aio_nbytes) {
1591         if (aiocb->aio_type & QEMU_AIO_WRITE) {
1592             return -EINVAL;
1593         } else {
1594             iov_memset(aiocb->io.iov, aiocb->io.niov, nbytes,
1595                       0, aiocb->aio_nbytes - nbytes);
1596             return 0;
1597         }
1598     } else {
1599         assert(nbytes < 0);
1600         return nbytes;
1601     }
1602 }
1603 
1604 #if defined(CONFIG_FALLOCATE) || defined(BLKZEROOUT) || defined(BLKDISCARD)
1605 static int translate_err(int err)
1606 {
1607     if (err == -ENODEV || err == -ENOSYS || err == -EOPNOTSUPP ||
1608         err == -ENOTTY) {
1609         err = -ENOTSUP;
1610     }
1611     return err;
1612 }
1613 #endif
1614 
1615 #ifdef CONFIG_FALLOCATE
1616 static int do_fallocate(int fd, int mode, off_t offset, off_t len)
1617 {
1618     do {
1619         if (fallocate(fd, mode, offset, len) == 0) {
1620             return 0;
1621         }
1622     } while (errno == EINTR);
1623     return translate_err(-errno);
1624 }
1625 #endif
1626 
1627 static ssize_t handle_aiocb_write_zeroes_block(RawPosixAIOData *aiocb)
1628 {
1629     int ret = -ENOTSUP;
1630     BDRVRawState *s = aiocb->bs->opaque;
1631 
1632     if (!s->has_write_zeroes) {
1633         return -ENOTSUP;
1634     }
1635 
1636 #ifdef BLKZEROOUT
1637     /* The BLKZEROOUT implementation in the kernel doesn't set
1638      * BLKDEV_ZERO_NOFALLBACK, so we can't call this if we have to avoid slow
1639      * fallbacks. */
1640     if (!(aiocb->aio_type & QEMU_AIO_NO_FALLBACK)) {
1641         do {
1642             uint64_t range[2] = { aiocb->aio_offset, aiocb->aio_nbytes };
1643             if (ioctl(aiocb->aio_fildes, BLKZEROOUT, range) == 0) {
1644                 return 0;
1645             }
1646         } while (errno == EINTR);
1647 
1648         ret = translate_err(-errno);
1649         if (ret == -ENOTSUP) {
1650             s->has_write_zeroes = false;
1651         }
1652     }
1653 #endif
1654 
1655     return ret;
1656 }
1657 
1658 static int handle_aiocb_write_zeroes(void *opaque)
1659 {
1660     RawPosixAIOData *aiocb = opaque;
1661 #ifdef CONFIG_FALLOCATE
1662     BDRVRawState *s = aiocb->bs->opaque;
1663     int64_t len;
1664 #endif
1665 
1666     if (aiocb->aio_type & QEMU_AIO_BLKDEV) {
1667         return handle_aiocb_write_zeroes_block(aiocb);
1668     }
1669 
1670 #ifdef CONFIG_FALLOCATE_ZERO_RANGE
1671     if (s->has_write_zeroes) {
1672         int ret = do_fallocate(s->fd, FALLOC_FL_ZERO_RANGE,
1673                                aiocb->aio_offset, aiocb->aio_nbytes);
1674         if (ret == -ENOTSUP) {
1675             s->has_write_zeroes = false;
1676         } else if (ret == 0 || ret != -EINVAL) {
1677             return ret;
1678         }
1679         /*
1680          * Note: Some file systems do not like unaligned byte ranges, and
1681          * return EINVAL in such a case, though they should not do it according
1682          * to the man-page of fallocate(). Thus we simply ignore this return
1683          * value and try the other fallbacks instead.
1684          */
1685     }
1686 #endif
1687 
1688 #ifdef CONFIG_FALLOCATE_PUNCH_HOLE
1689     if (s->has_discard && s->has_fallocate) {
1690         int ret = do_fallocate(s->fd,
1691                                FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE,
1692                                aiocb->aio_offset, aiocb->aio_nbytes);
1693         if (ret == 0) {
1694             ret = do_fallocate(s->fd, 0, aiocb->aio_offset, aiocb->aio_nbytes);
1695             if (ret == 0 || ret != -ENOTSUP) {
1696                 return ret;
1697             }
1698             s->has_fallocate = false;
1699         } else if (ret == -EINVAL) {
1700             /*
1701              * Some file systems like older versions of GPFS do not like un-
1702              * aligned byte ranges, and return EINVAL in such a case, though
1703              * they should not do it according to the man-page of fallocate().
1704              * Warn about the bad filesystem and try the final fallback instead.
1705              */
1706             warn_report_once("Your file system is misbehaving: "
1707                              "fallocate(FALLOC_FL_PUNCH_HOLE) returned EINVAL. "
1708                              "Please report this bug to your file sytem "
1709                              "vendor.");
1710         } else if (ret != -ENOTSUP) {
1711             return ret;
1712         } else {
1713             s->has_discard = false;
1714         }
1715     }
1716 #endif
1717 
1718 #ifdef CONFIG_FALLOCATE
1719     /* Last resort: we are trying to extend the file with zeroed data. This
1720      * can be done via fallocate(fd, 0) */
1721     len = bdrv_getlength(aiocb->bs);
1722     if (s->has_fallocate && len >= 0 && aiocb->aio_offset >= len) {
1723         int ret = do_fallocate(s->fd, 0, aiocb->aio_offset, aiocb->aio_nbytes);
1724         if (ret == 0 || ret != -ENOTSUP) {
1725             return ret;
1726         }
1727         s->has_fallocate = false;
1728     }
1729 #endif
1730 
1731     return -ENOTSUP;
1732 }
1733 
1734 static int handle_aiocb_write_zeroes_unmap(void *opaque)
1735 {
1736     RawPosixAIOData *aiocb = opaque;
1737     BDRVRawState *s G_GNUC_UNUSED = aiocb->bs->opaque;
1738 
1739     /* First try to write zeros and unmap at the same time */
1740 
1741 #ifdef CONFIG_FALLOCATE_PUNCH_HOLE
1742     int ret = do_fallocate(s->fd, FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE,
1743                            aiocb->aio_offset, aiocb->aio_nbytes);
1744     switch (ret) {
1745     case -ENOTSUP:
1746     case -EINVAL:
1747     case -EBUSY:
1748         break;
1749     default:
1750         return ret;
1751     }
1752 #endif
1753 
1754     /* If we couldn't manage to unmap while guaranteed that the area reads as
1755      * all-zero afterwards, just write zeroes without unmapping */
1756     return handle_aiocb_write_zeroes(aiocb);
1757 }
1758 
1759 #ifndef HAVE_COPY_FILE_RANGE
1760 static off_t copy_file_range(int in_fd, off_t *in_off, int out_fd,
1761                              off_t *out_off, size_t len, unsigned int flags)
1762 {
1763 #ifdef __NR_copy_file_range
1764     return syscall(__NR_copy_file_range, in_fd, in_off, out_fd,
1765                    out_off, len, flags);
1766 #else
1767     errno = ENOSYS;
1768     return -1;
1769 #endif
1770 }
1771 #endif
1772 
1773 static int handle_aiocb_copy_range(void *opaque)
1774 {
1775     RawPosixAIOData *aiocb = opaque;
1776     uint64_t bytes = aiocb->aio_nbytes;
1777     off_t in_off = aiocb->aio_offset;
1778     off_t out_off = aiocb->copy_range.aio_offset2;
1779 
1780     while (bytes) {
1781         ssize_t ret = copy_file_range(aiocb->aio_fildes, &in_off,
1782                                       aiocb->copy_range.aio_fd2, &out_off,
1783                                       bytes, 0);
1784         trace_file_copy_file_range(aiocb->bs, aiocb->aio_fildes, in_off,
1785                                    aiocb->copy_range.aio_fd2, out_off, bytes,
1786                                    0, ret);
1787         if (ret == 0) {
1788             /* No progress (e.g. when beyond EOF), let the caller fall back to
1789              * buffer I/O. */
1790             return -ENOSPC;
1791         }
1792         if (ret < 0) {
1793             switch (errno) {
1794             case ENOSYS:
1795                 return -ENOTSUP;
1796             case EINTR:
1797                 continue;
1798             default:
1799                 return -errno;
1800             }
1801         }
1802         bytes -= ret;
1803     }
1804     return 0;
1805 }
1806 
1807 static int handle_aiocb_discard(void *opaque)
1808 {
1809     RawPosixAIOData *aiocb = opaque;
1810     int ret = -EOPNOTSUPP;
1811     BDRVRawState *s = aiocb->bs->opaque;
1812 
1813     if (!s->has_discard) {
1814         return -ENOTSUP;
1815     }
1816 
1817     if (aiocb->aio_type & QEMU_AIO_BLKDEV) {
1818 #ifdef BLKDISCARD
1819         do {
1820             uint64_t range[2] = { aiocb->aio_offset, aiocb->aio_nbytes };
1821             if (ioctl(aiocb->aio_fildes, BLKDISCARD, range) == 0) {
1822                 return 0;
1823             }
1824         } while (errno == EINTR);
1825 
1826         ret = translate_err(-errno);
1827 #endif
1828     } else {
1829 #ifdef CONFIG_FALLOCATE_PUNCH_HOLE
1830         ret = do_fallocate(s->fd, FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE,
1831                            aiocb->aio_offset, aiocb->aio_nbytes);
1832         ret = translate_err(-errno);
1833 #elif defined(__APPLE__) && (__MACH__)
1834         fpunchhole_t fpunchhole;
1835         fpunchhole.fp_flags = 0;
1836         fpunchhole.reserved = 0;
1837         fpunchhole.fp_offset = aiocb->aio_offset;
1838         fpunchhole.fp_length = aiocb->aio_nbytes;
1839         if (fcntl(s->fd, F_PUNCHHOLE, &fpunchhole) == -1) {
1840             ret = errno == ENODEV ? -ENOTSUP : -errno;
1841         } else {
1842             ret = 0;
1843         }
1844 #endif
1845     }
1846 
1847     if (ret == -ENOTSUP) {
1848         s->has_discard = false;
1849     }
1850     return ret;
1851 }
1852 
1853 /*
1854  * Help alignment probing by allocating the first block.
1855  *
1856  * When reading with direct I/O from unallocated area on Gluster backed by XFS,
1857  * reading succeeds regardless of request length. In this case we fallback to
1858  * safe alignment which is not optimal. Allocating the first block avoids this
1859  * fallback.
1860  *
1861  * fd may be opened with O_DIRECT, but we don't know the buffer alignment or
1862  * request alignment, so we use safe values.
1863  *
1864  * Returns: 0 on success, -errno on failure. Since this is an optimization,
1865  * caller may ignore failures.
1866  */
1867 static int allocate_first_block(int fd, size_t max_size)
1868 {
1869     size_t write_size = (max_size < MAX_BLOCKSIZE)
1870         ? BDRV_SECTOR_SIZE
1871         : MAX_BLOCKSIZE;
1872     size_t max_align = MAX(MAX_BLOCKSIZE, qemu_real_host_page_size);
1873     void *buf;
1874     ssize_t n;
1875     int ret;
1876 
1877     buf = qemu_memalign(max_align, write_size);
1878     memset(buf, 0, write_size);
1879 
1880     do {
1881         n = pwrite(fd, buf, write_size, 0);
1882     } while (n == -1 && errno == EINTR);
1883 
1884     ret = (n == -1) ? -errno : 0;
1885 
1886     qemu_vfree(buf);
1887     return ret;
1888 }
1889 
1890 static int handle_aiocb_truncate(void *opaque)
1891 {
1892     RawPosixAIOData *aiocb = opaque;
1893     int result = 0;
1894     int64_t current_length = 0;
1895     char *buf = NULL;
1896     struct stat st;
1897     int fd = aiocb->aio_fildes;
1898     int64_t offset = aiocb->aio_offset;
1899     PreallocMode prealloc = aiocb->truncate.prealloc;
1900     Error **errp = aiocb->truncate.errp;
1901 
1902     if (fstat(fd, &st) < 0) {
1903         result = -errno;
1904         error_setg_errno(errp, -result, "Could not stat file");
1905         return result;
1906     }
1907 
1908     current_length = st.st_size;
1909     if (current_length > offset && prealloc != PREALLOC_MODE_OFF) {
1910         error_setg(errp, "Cannot use preallocation for shrinking files");
1911         return -ENOTSUP;
1912     }
1913 
1914     switch (prealloc) {
1915 #ifdef CONFIG_POSIX_FALLOCATE
1916     case PREALLOC_MODE_FALLOC:
1917         /*
1918          * Truncating before posix_fallocate() makes it about twice slower on
1919          * file systems that do not support fallocate(), trying to check if a
1920          * block is allocated before allocating it, so don't do that here.
1921          */
1922         if (offset != current_length) {
1923             result = -posix_fallocate(fd, current_length,
1924                                       offset - current_length);
1925             if (result != 0) {
1926                 /* posix_fallocate() doesn't set errno. */
1927                 error_setg_errno(errp, -result,
1928                                  "Could not preallocate new data");
1929             } else if (current_length == 0) {
1930                 /*
1931                  * posix_fallocate() uses fallocate() if the filesystem
1932                  * supports it, or fallback to manually writing zeroes. If
1933                  * fallocate() was used, unaligned reads from the fallocated
1934                  * area in raw_probe_alignment() will succeed, hence we need to
1935                  * allocate the first block.
1936                  *
1937                  * Optimize future alignment probing; ignore failures.
1938                  */
1939                 allocate_first_block(fd, offset);
1940             }
1941         } else {
1942             result = 0;
1943         }
1944         goto out;
1945 #endif
1946     case PREALLOC_MODE_FULL:
1947     {
1948         int64_t num = 0, left = offset - current_length;
1949         off_t seek_result;
1950 
1951         /*
1952          * Knowing the final size from the beginning could allow the file
1953          * system driver to do less allocations and possibly avoid
1954          * fragmentation of the file.
1955          */
1956         if (ftruncate(fd, offset) != 0) {
1957             result = -errno;
1958             error_setg_errno(errp, -result, "Could not resize file");
1959             goto out;
1960         }
1961 
1962         buf = g_malloc0(65536);
1963 
1964         seek_result = lseek(fd, current_length, SEEK_SET);
1965         if (seek_result < 0) {
1966             result = -errno;
1967             error_setg_errno(errp, -result,
1968                              "Failed to seek to the old end of file");
1969             goto out;
1970         }
1971 
1972         while (left > 0) {
1973             num = MIN(left, 65536);
1974             result = write(fd, buf, num);
1975             if (result < 0) {
1976                 if (errno == EINTR) {
1977                     continue;
1978                 }
1979                 result = -errno;
1980                 error_setg_errno(errp, -result,
1981                                  "Could not write zeros for preallocation");
1982                 goto out;
1983             }
1984             left -= result;
1985         }
1986         if (result >= 0) {
1987             result = fsync(fd);
1988             if (result < 0) {
1989                 result = -errno;
1990                 error_setg_errno(errp, -result,
1991                                  "Could not flush file to disk");
1992                 goto out;
1993             }
1994         }
1995         goto out;
1996     }
1997     case PREALLOC_MODE_OFF:
1998         if (ftruncate(fd, offset) != 0) {
1999             result = -errno;
2000             error_setg_errno(errp, -result, "Could not resize file");
2001         } else if (current_length == 0 && offset > current_length) {
2002             /* Optimize future alignment probing; ignore failures. */
2003             allocate_first_block(fd, offset);
2004         }
2005         return result;
2006     default:
2007         result = -ENOTSUP;
2008         error_setg(errp, "Unsupported preallocation mode: %s",
2009                    PreallocMode_str(prealloc));
2010         return result;
2011     }
2012 
2013 out:
2014     if (result < 0) {
2015         if (ftruncate(fd, current_length) < 0) {
2016             error_report("Failed to restore old file length: %s",
2017                          strerror(errno));
2018         }
2019     }
2020 
2021     g_free(buf);
2022     return result;
2023 }
2024 
2025 static int coroutine_fn raw_thread_pool_submit(BlockDriverState *bs,
2026                                                ThreadPoolFunc func, void *arg)
2027 {
2028     /* @bs can be NULL, bdrv_get_aio_context() returns the main context then */
2029     ThreadPool *pool = aio_get_thread_pool(bdrv_get_aio_context(bs));
2030     return thread_pool_submit_co(pool, func, arg);
2031 }
2032 
2033 static int coroutine_fn raw_co_prw(BlockDriverState *bs, uint64_t offset,
2034                                    uint64_t bytes, QEMUIOVector *qiov, int type)
2035 {
2036     BDRVRawState *s = bs->opaque;
2037     RawPosixAIOData acb;
2038 
2039     if (fd_open(bs) < 0)
2040         return -EIO;
2041 
2042     /*
2043      * When using O_DIRECT, the request must be aligned to be able to use
2044      * either libaio or io_uring interface. If not fail back to regular thread
2045      * pool read/write code which emulates this for us if we
2046      * set QEMU_AIO_MISALIGNED.
2047      */
2048     if (s->needs_alignment && !bdrv_qiov_is_aligned(bs, qiov)) {
2049         type |= QEMU_AIO_MISALIGNED;
2050 #ifdef CONFIG_LINUX_IO_URING
2051     } else if (s->use_linux_io_uring) {
2052         LuringState *aio = aio_get_linux_io_uring(bdrv_get_aio_context(bs));
2053         assert(qiov->size == bytes);
2054         return luring_co_submit(bs, aio, s->fd, offset, qiov, type);
2055 #endif
2056 #ifdef CONFIG_LINUX_AIO
2057     } else if (s->use_linux_aio) {
2058         LinuxAioState *aio = aio_get_linux_aio(bdrv_get_aio_context(bs));
2059         assert(qiov->size == bytes);
2060         return laio_co_submit(bs, aio, s->fd, offset, qiov, type);
2061 #endif
2062     }
2063 
2064     acb = (RawPosixAIOData) {
2065         .bs             = bs,
2066         .aio_fildes     = s->fd,
2067         .aio_type       = type,
2068         .aio_offset     = offset,
2069         .aio_nbytes     = bytes,
2070         .io             = {
2071             .iov            = qiov->iov,
2072             .niov           = qiov->niov,
2073         },
2074     };
2075 
2076     assert(qiov->size == bytes);
2077     return raw_thread_pool_submit(bs, handle_aiocb_rw, &acb);
2078 }
2079 
2080 static int coroutine_fn raw_co_preadv(BlockDriverState *bs, uint64_t offset,
2081                                       uint64_t bytes, QEMUIOVector *qiov,
2082                                       int flags)
2083 {
2084     return raw_co_prw(bs, offset, bytes, qiov, QEMU_AIO_READ);
2085 }
2086 
2087 static int coroutine_fn raw_co_pwritev(BlockDriverState *bs, uint64_t offset,
2088                                        uint64_t bytes, QEMUIOVector *qiov,
2089                                        int flags)
2090 {
2091     assert(flags == 0);
2092     return raw_co_prw(bs, offset, bytes, qiov, QEMU_AIO_WRITE);
2093 }
2094 
2095 static void raw_aio_plug(BlockDriverState *bs)
2096 {
2097     BDRVRawState __attribute__((unused)) *s = bs->opaque;
2098 #ifdef CONFIG_LINUX_AIO
2099     if (s->use_linux_aio) {
2100         LinuxAioState *aio = aio_get_linux_aio(bdrv_get_aio_context(bs));
2101         laio_io_plug(bs, aio);
2102     }
2103 #endif
2104 #ifdef CONFIG_LINUX_IO_URING
2105     if (s->use_linux_io_uring) {
2106         LuringState *aio = aio_get_linux_io_uring(bdrv_get_aio_context(bs));
2107         luring_io_plug(bs, aio);
2108     }
2109 #endif
2110 }
2111 
2112 static void raw_aio_unplug(BlockDriverState *bs)
2113 {
2114     BDRVRawState __attribute__((unused)) *s = bs->opaque;
2115 #ifdef CONFIG_LINUX_AIO
2116     if (s->use_linux_aio) {
2117         LinuxAioState *aio = aio_get_linux_aio(bdrv_get_aio_context(bs));
2118         laio_io_unplug(bs, aio);
2119     }
2120 #endif
2121 #ifdef CONFIG_LINUX_IO_URING
2122     if (s->use_linux_io_uring) {
2123         LuringState *aio = aio_get_linux_io_uring(bdrv_get_aio_context(bs));
2124         luring_io_unplug(bs, aio);
2125     }
2126 #endif
2127 }
2128 
2129 static int raw_co_flush_to_disk(BlockDriverState *bs)
2130 {
2131     BDRVRawState *s = bs->opaque;
2132     RawPosixAIOData acb;
2133     int ret;
2134 
2135     ret = fd_open(bs);
2136     if (ret < 0) {
2137         return ret;
2138     }
2139 
2140     acb = (RawPosixAIOData) {
2141         .bs             = bs,
2142         .aio_fildes     = s->fd,
2143         .aio_type       = QEMU_AIO_FLUSH,
2144     };
2145 
2146 #ifdef CONFIG_LINUX_IO_URING
2147     if (s->use_linux_io_uring) {
2148         LuringState *aio = aio_get_linux_io_uring(bdrv_get_aio_context(bs));
2149         return luring_co_submit(bs, aio, s->fd, 0, NULL, QEMU_AIO_FLUSH);
2150     }
2151 #endif
2152     return raw_thread_pool_submit(bs, handle_aiocb_flush, &acb);
2153 }
2154 
2155 static void raw_aio_attach_aio_context(BlockDriverState *bs,
2156                                        AioContext *new_context)
2157 {
2158     BDRVRawState __attribute__((unused)) *s = bs->opaque;
2159 #ifdef CONFIG_LINUX_AIO
2160     if (s->use_linux_aio) {
2161         Error *local_err = NULL;
2162         if (!aio_setup_linux_aio(new_context, &local_err)) {
2163             error_reportf_err(local_err, "Unable to use native AIO, "
2164                                          "falling back to thread pool: ");
2165             s->use_linux_aio = false;
2166         }
2167     }
2168 #endif
2169 #ifdef CONFIG_LINUX_IO_URING
2170     if (s->use_linux_io_uring) {
2171         Error *local_err = NULL;
2172         if (!aio_setup_linux_io_uring(new_context, &local_err)) {
2173             error_reportf_err(local_err, "Unable to use linux io_uring, "
2174                                          "falling back to thread pool: ");
2175             s->use_linux_io_uring = false;
2176         }
2177     }
2178 #endif
2179 }
2180 
2181 static void raw_close(BlockDriverState *bs)
2182 {
2183     BDRVRawState *s = bs->opaque;
2184 
2185     if (s->fd >= 0) {
2186         qemu_close(s->fd);
2187         s->fd = -1;
2188     }
2189 }
2190 
2191 /**
2192  * Truncates the given regular file @fd to @offset and, when growing, fills the
2193  * new space according to @prealloc.
2194  *
2195  * Returns: 0 on success, -errno on failure.
2196  */
2197 static int coroutine_fn
2198 raw_regular_truncate(BlockDriverState *bs, int fd, int64_t offset,
2199                      PreallocMode prealloc, Error **errp)
2200 {
2201     RawPosixAIOData acb;
2202 
2203     acb = (RawPosixAIOData) {
2204         .bs             = bs,
2205         .aio_fildes     = fd,
2206         .aio_type       = QEMU_AIO_TRUNCATE,
2207         .aio_offset     = offset,
2208         .truncate       = {
2209             .prealloc       = prealloc,
2210             .errp           = errp,
2211         },
2212     };
2213 
2214     return raw_thread_pool_submit(bs, handle_aiocb_truncate, &acb);
2215 }
2216 
2217 static int coroutine_fn raw_co_truncate(BlockDriverState *bs, int64_t offset,
2218                                         bool exact, PreallocMode prealloc,
2219                                         BdrvRequestFlags flags, Error **errp)
2220 {
2221     BDRVRawState *s = bs->opaque;
2222     struct stat st;
2223     int ret;
2224 
2225     if (fstat(s->fd, &st)) {
2226         ret = -errno;
2227         error_setg_errno(errp, -ret, "Failed to fstat() the file");
2228         return ret;
2229     }
2230 
2231     if (S_ISREG(st.st_mode)) {
2232         /* Always resizes to the exact @offset */
2233         return raw_regular_truncate(bs, s->fd, offset, prealloc, errp);
2234     }
2235 
2236     if (prealloc != PREALLOC_MODE_OFF) {
2237         error_setg(errp, "Preallocation mode '%s' unsupported for this "
2238                    "non-regular file", PreallocMode_str(prealloc));
2239         return -ENOTSUP;
2240     }
2241 
2242     if (S_ISCHR(st.st_mode) || S_ISBLK(st.st_mode)) {
2243         int64_t cur_length = raw_getlength(bs);
2244 
2245         if (offset != cur_length && exact) {
2246             error_setg(errp, "Cannot resize device files");
2247             return -ENOTSUP;
2248         } else if (offset > cur_length) {
2249             error_setg(errp, "Cannot grow device files");
2250             return -EINVAL;
2251         }
2252     } else {
2253         error_setg(errp, "Resizing this file is not supported");
2254         return -ENOTSUP;
2255     }
2256 
2257     return 0;
2258 }
2259 
2260 #ifdef __OpenBSD__
2261 static int64_t raw_getlength(BlockDriverState *bs)
2262 {
2263     BDRVRawState *s = bs->opaque;
2264     int fd = s->fd;
2265     struct stat st;
2266 
2267     if (fstat(fd, &st))
2268         return -errno;
2269     if (S_ISCHR(st.st_mode) || S_ISBLK(st.st_mode)) {
2270         struct disklabel dl;
2271 
2272         if (ioctl(fd, DIOCGDINFO, &dl))
2273             return -errno;
2274         return (uint64_t)dl.d_secsize *
2275             dl.d_partitions[DISKPART(st.st_rdev)].p_size;
2276     } else
2277         return st.st_size;
2278 }
2279 #elif defined(__NetBSD__)
2280 static int64_t raw_getlength(BlockDriverState *bs)
2281 {
2282     BDRVRawState *s = bs->opaque;
2283     int fd = s->fd;
2284     struct stat st;
2285 
2286     if (fstat(fd, &st))
2287         return -errno;
2288     if (S_ISCHR(st.st_mode) || S_ISBLK(st.st_mode)) {
2289         struct dkwedge_info dkw;
2290 
2291         if (ioctl(fd, DIOCGWEDGEINFO, &dkw) != -1) {
2292             return dkw.dkw_size * 512;
2293         } else {
2294             struct disklabel dl;
2295 
2296             if (ioctl(fd, DIOCGDINFO, &dl))
2297                 return -errno;
2298             return (uint64_t)dl.d_secsize *
2299                 dl.d_partitions[DISKPART(st.st_rdev)].p_size;
2300         }
2301     } else
2302         return st.st_size;
2303 }
2304 #elif defined(__sun__)
2305 static int64_t raw_getlength(BlockDriverState *bs)
2306 {
2307     BDRVRawState *s = bs->opaque;
2308     struct dk_minfo minfo;
2309     int ret;
2310     int64_t size;
2311 
2312     ret = fd_open(bs);
2313     if (ret < 0) {
2314         return ret;
2315     }
2316 
2317     /*
2318      * Use the DKIOCGMEDIAINFO ioctl to read the size.
2319      */
2320     ret = ioctl(s->fd, DKIOCGMEDIAINFO, &minfo);
2321     if (ret != -1) {
2322         return minfo.dki_lbsize * minfo.dki_capacity;
2323     }
2324 
2325     /*
2326      * There are reports that lseek on some devices fails, but
2327      * irc discussion said that contingency on contingency was overkill.
2328      */
2329     size = lseek(s->fd, 0, SEEK_END);
2330     if (size < 0) {
2331         return -errno;
2332     }
2333     return size;
2334 }
2335 #elif defined(CONFIG_BSD)
2336 static int64_t raw_getlength(BlockDriverState *bs)
2337 {
2338     BDRVRawState *s = bs->opaque;
2339     int fd = s->fd;
2340     int64_t size;
2341     struct stat sb;
2342 #if defined (__FreeBSD__) || defined(__FreeBSD_kernel__)
2343     int reopened = 0;
2344 #endif
2345     int ret;
2346 
2347     ret = fd_open(bs);
2348     if (ret < 0)
2349         return ret;
2350 
2351 #if defined (__FreeBSD__) || defined(__FreeBSD_kernel__)
2352 again:
2353 #endif
2354     if (!fstat(fd, &sb) && (S_IFCHR & sb.st_mode)) {
2355         size = 0;
2356 #ifdef DIOCGMEDIASIZE
2357         if (ioctl(fd, DIOCGMEDIASIZE, (off_t *)&size)) {
2358             size = 0;
2359         }
2360 #endif
2361 #ifdef DIOCGPART
2362         if (size == 0) {
2363             struct partinfo pi;
2364             if (ioctl(fd, DIOCGPART, &pi) == 0) {
2365                 size = pi.media_size;
2366             }
2367         }
2368 #endif
2369 #if defined(DKIOCGETBLOCKCOUNT) && defined(DKIOCGETBLOCKSIZE)
2370         if (size == 0) {
2371             uint64_t sectors = 0;
2372             uint32_t sector_size = 0;
2373 
2374             if (ioctl(fd, DKIOCGETBLOCKCOUNT, &sectors) == 0
2375                && ioctl(fd, DKIOCGETBLOCKSIZE, &sector_size) == 0) {
2376                 size = sectors * sector_size;
2377             }
2378         }
2379 #endif
2380         if (size == 0) {
2381             size = lseek(fd, 0LL, SEEK_END);
2382         }
2383         if (size < 0) {
2384             return -errno;
2385         }
2386 #if defined(__FreeBSD__) || defined(__FreeBSD_kernel__)
2387         switch(s->type) {
2388         case FTYPE_CD:
2389             /* XXX FreeBSD acd returns UINT_MAX sectors for an empty drive */
2390             if (size == 2048LL * (unsigned)-1)
2391                 size = 0;
2392             /* XXX no disc?  maybe we need to reopen... */
2393             if (size <= 0 && !reopened && cdrom_reopen(bs) >= 0) {
2394                 reopened = 1;
2395                 goto again;
2396             }
2397         }
2398 #endif
2399     } else {
2400         size = lseek(fd, 0, SEEK_END);
2401         if (size < 0) {
2402             return -errno;
2403         }
2404     }
2405     return size;
2406 }
2407 #else
2408 static int64_t raw_getlength(BlockDriverState *bs)
2409 {
2410     BDRVRawState *s = bs->opaque;
2411     int ret;
2412     int64_t size;
2413 
2414     ret = fd_open(bs);
2415     if (ret < 0) {
2416         return ret;
2417     }
2418 
2419     size = lseek(s->fd, 0, SEEK_END);
2420     if (size < 0) {
2421         return -errno;
2422     }
2423     return size;
2424 }
2425 #endif
2426 
2427 static int64_t raw_get_allocated_file_size(BlockDriverState *bs)
2428 {
2429     struct stat st;
2430     BDRVRawState *s = bs->opaque;
2431 
2432     if (fstat(s->fd, &st) < 0) {
2433         return -errno;
2434     }
2435     return (int64_t)st.st_blocks * 512;
2436 }
2437 
2438 static int coroutine_fn
2439 raw_co_create(BlockdevCreateOptions *options, Error **errp)
2440 {
2441     BlockdevCreateOptionsFile *file_opts;
2442     Error *local_err = NULL;
2443     int fd;
2444     uint64_t perm, shared;
2445     int result = 0;
2446 
2447     /* Validate options and set default values */
2448     assert(options->driver == BLOCKDEV_DRIVER_FILE);
2449     file_opts = &options->u.file;
2450 
2451     if (!file_opts->has_nocow) {
2452         file_opts->nocow = false;
2453     }
2454     if (!file_opts->has_preallocation) {
2455         file_opts->preallocation = PREALLOC_MODE_OFF;
2456     }
2457     if (!file_opts->has_extent_size_hint) {
2458         file_opts->extent_size_hint = 1 * MiB;
2459     }
2460     if (file_opts->extent_size_hint > UINT32_MAX) {
2461         result = -EINVAL;
2462         error_setg(errp, "Extent size hint is too large");
2463         goto out;
2464     }
2465 
2466     /* Create file */
2467     fd = qemu_create(file_opts->filename, O_RDWR | O_BINARY, 0644, errp);
2468     if (fd < 0) {
2469         result = -errno;
2470         goto out;
2471     }
2472 
2473     /* Take permissions: We want to discard everything, so we need
2474      * BLK_PERM_WRITE; and truncation to the desired size requires
2475      * BLK_PERM_RESIZE.
2476      * On the other hand, we cannot share the RESIZE permission
2477      * because we promise that after this function, the file has the
2478      * size given in the options.  If someone else were to resize it
2479      * concurrently, we could not guarantee that.
2480      * Note that after this function, we can no longer guarantee that
2481      * the file is not touched by a third party, so it may be resized
2482      * then. */
2483     perm = BLK_PERM_WRITE | BLK_PERM_RESIZE;
2484     shared = BLK_PERM_ALL & ~BLK_PERM_RESIZE;
2485 
2486     /* Step one: Take locks */
2487     result = raw_apply_lock_bytes(NULL, fd, perm, ~shared, false, errp);
2488     if (result < 0) {
2489         goto out_close;
2490     }
2491 
2492     /* Step two: Check that nobody else has taken conflicting locks */
2493     result = raw_check_lock_bytes(fd, perm, shared, errp);
2494     if (result < 0) {
2495         error_append_hint(errp,
2496                           "Is another process using the image [%s]?\n",
2497                           file_opts->filename);
2498         goto out_unlock;
2499     }
2500 
2501     /* Clear the file by truncating it to 0 */
2502     result = raw_regular_truncate(NULL, fd, 0, PREALLOC_MODE_OFF, errp);
2503     if (result < 0) {
2504         goto out_unlock;
2505     }
2506 
2507     if (file_opts->nocow) {
2508 #ifdef __linux__
2509         /* Set NOCOW flag to solve performance issue on fs like btrfs.
2510          * This is an optimisation. The FS_IOC_SETFLAGS ioctl return value
2511          * will be ignored since any failure of this operation should not
2512          * block the left work.
2513          */
2514         int attr;
2515         if (ioctl(fd, FS_IOC_GETFLAGS, &attr) == 0) {
2516             attr |= FS_NOCOW_FL;
2517             ioctl(fd, FS_IOC_SETFLAGS, &attr);
2518         }
2519 #endif
2520     }
2521 #ifdef FS_IOC_FSSETXATTR
2522     /*
2523      * Try to set the extent size hint. Failure is not fatal, and a warning is
2524      * only printed if the option was explicitly specified.
2525      */
2526     {
2527         struct fsxattr attr;
2528         result = ioctl(fd, FS_IOC_FSGETXATTR, &attr);
2529         if (result == 0) {
2530             attr.fsx_xflags |= FS_XFLAG_EXTSIZE;
2531             attr.fsx_extsize = file_opts->extent_size_hint;
2532             result = ioctl(fd, FS_IOC_FSSETXATTR, &attr);
2533         }
2534         if (result < 0 && file_opts->has_extent_size_hint &&
2535             file_opts->extent_size_hint)
2536         {
2537             warn_report("Failed to set extent size hint: %s",
2538                         strerror(errno));
2539         }
2540     }
2541 #endif
2542 
2543     /* Resize and potentially preallocate the file to the desired
2544      * final size */
2545     result = raw_regular_truncate(NULL, fd, file_opts->size,
2546                                   file_opts->preallocation, errp);
2547     if (result < 0) {
2548         goto out_unlock;
2549     }
2550 
2551 out_unlock:
2552     raw_apply_lock_bytes(NULL, fd, 0, 0, true, &local_err);
2553     if (local_err) {
2554         /* The above call should not fail, and if it does, that does
2555          * not mean the whole creation operation has failed.  So
2556          * report it the user for their convenience, but do not report
2557          * it to the caller. */
2558         warn_report_err(local_err);
2559     }
2560 
2561 out_close:
2562     if (qemu_close(fd) != 0 && result == 0) {
2563         result = -errno;
2564         error_setg_errno(errp, -result, "Could not close the new file");
2565     }
2566 out:
2567     return result;
2568 }
2569 
2570 static int coroutine_fn raw_co_create_opts(BlockDriver *drv,
2571                                            const char *filename,
2572                                            QemuOpts *opts,
2573                                            Error **errp)
2574 {
2575     BlockdevCreateOptions options;
2576     int64_t total_size = 0;
2577     int64_t extent_size_hint = 0;
2578     bool has_extent_size_hint = false;
2579     bool nocow = false;
2580     PreallocMode prealloc;
2581     char *buf = NULL;
2582     Error *local_err = NULL;
2583 
2584     /* Skip file: protocol prefix */
2585     strstart(filename, "file:", &filename);
2586 
2587     /* Read out options */
2588     total_size = ROUND_UP(qemu_opt_get_size_del(opts, BLOCK_OPT_SIZE, 0),
2589                           BDRV_SECTOR_SIZE);
2590     if (qemu_opt_get(opts, BLOCK_OPT_EXTENT_SIZE_HINT)) {
2591         has_extent_size_hint = true;
2592         extent_size_hint =
2593             qemu_opt_get_size_del(opts, BLOCK_OPT_EXTENT_SIZE_HINT, -1);
2594     }
2595     nocow = qemu_opt_get_bool(opts, BLOCK_OPT_NOCOW, false);
2596     buf = qemu_opt_get_del(opts, BLOCK_OPT_PREALLOC);
2597     prealloc = qapi_enum_parse(&PreallocMode_lookup, buf,
2598                                PREALLOC_MODE_OFF, &local_err);
2599     g_free(buf);
2600     if (local_err) {
2601         error_propagate(errp, local_err);
2602         return -EINVAL;
2603     }
2604 
2605     options = (BlockdevCreateOptions) {
2606         .driver     = BLOCKDEV_DRIVER_FILE,
2607         .u.file     = {
2608             .filename           = (char *) filename,
2609             .size               = total_size,
2610             .has_preallocation  = true,
2611             .preallocation      = prealloc,
2612             .has_nocow          = true,
2613             .nocow              = nocow,
2614             .has_extent_size_hint = has_extent_size_hint,
2615             .extent_size_hint   = extent_size_hint,
2616         },
2617     };
2618     return raw_co_create(&options, errp);
2619 }
2620 
2621 static int coroutine_fn raw_co_delete_file(BlockDriverState *bs,
2622                                            Error **errp)
2623 {
2624     struct stat st;
2625     int ret;
2626 
2627     if (!(stat(bs->filename, &st) == 0) || !S_ISREG(st.st_mode)) {
2628         error_setg_errno(errp, ENOENT, "%s is not a regular file",
2629                          bs->filename);
2630         return -ENOENT;
2631     }
2632 
2633     ret = unlink(bs->filename);
2634     if (ret < 0) {
2635         ret = -errno;
2636         error_setg_errno(errp, -ret, "Error when deleting file %s",
2637                          bs->filename);
2638     }
2639 
2640     return ret;
2641 }
2642 
2643 /*
2644  * Find allocation range in @bs around offset @start.
2645  * May change underlying file descriptor's file offset.
2646  * If @start is not in a hole, store @start in @data, and the
2647  * beginning of the next hole in @hole, and return 0.
2648  * If @start is in a non-trailing hole, store @start in @hole and the
2649  * beginning of the next non-hole in @data, and return 0.
2650  * If @start is in a trailing hole or beyond EOF, return -ENXIO.
2651  * If we can't find out, return a negative errno other than -ENXIO.
2652  */
2653 static int find_allocation(BlockDriverState *bs, off_t start,
2654                            off_t *data, off_t *hole)
2655 {
2656 #if defined SEEK_HOLE && defined SEEK_DATA
2657     BDRVRawState *s = bs->opaque;
2658     off_t offs;
2659 
2660     /*
2661      * SEEK_DATA cases:
2662      * D1. offs == start: start is in data
2663      * D2. offs > start: start is in a hole, next data at offs
2664      * D3. offs < 0, errno = ENXIO: either start is in a trailing hole
2665      *                              or start is beyond EOF
2666      *     If the latter happens, the file has been truncated behind
2667      *     our back since we opened it.  All bets are off then.
2668      *     Treating like a trailing hole is simplest.
2669      * D4. offs < 0, errno != ENXIO: we learned nothing
2670      */
2671     offs = lseek(s->fd, start, SEEK_DATA);
2672     if (offs < 0) {
2673         return -errno;          /* D3 or D4 */
2674     }
2675 
2676     if (offs < start) {
2677         /* This is not a valid return by lseek().  We are safe to just return
2678          * -EIO in this case, and we'll treat it like D4. */
2679         return -EIO;
2680     }
2681 
2682     if (offs > start) {
2683         /* D2: in hole, next data at offs */
2684         *hole = start;
2685         *data = offs;
2686         return 0;
2687     }
2688 
2689     /* D1: in data, end not yet known */
2690 
2691     /*
2692      * SEEK_HOLE cases:
2693      * H1. offs == start: start is in a hole
2694      *     If this happens here, a hole has been dug behind our back
2695      *     since the previous lseek().
2696      * H2. offs > start: either start is in data, next hole at offs,
2697      *                   or start is in trailing hole, EOF at offs
2698      *     Linux treats trailing holes like any other hole: offs ==
2699      *     start.  Solaris seeks to EOF instead: offs > start (blech).
2700      *     If that happens here, a hole has been dug behind our back
2701      *     since the previous lseek().
2702      * H3. offs < 0, errno = ENXIO: start is beyond EOF
2703      *     If this happens, the file has been truncated behind our
2704      *     back since we opened it.  Treat it like a trailing hole.
2705      * H4. offs < 0, errno != ENXIO: we learned nothing
2706      *     Pretend we know nothing at all, i.e. "forget" about D1.
2707      */
2708     offs = lseek(s->fd, start, SEEK_HOLE);
2709     if (offs < 0) {
2710         return -errno;          /* D1 and (H3 or H4) */
2711     }
2712 
2713     if (offs < start) {
2714         /* This is not a valid return by lseek().  We are safe to just return
2715          * -EIO in this case, and we'll treat it like H4. */
2716         return -EIO;
2717     }
2718 
2719     if (offs > start) {
2720         /*
2721          * D1 and H2: either in data, next hole at offs, or it was in
2722          * data but is now in a trailing hole.  In the latter case,
2723          * all bets are off.  Treating it as if it there was data all
2724          * the way to EOF is safe, so simply do that.
2725          */
2726         *data = start;
2727         *hole = offs;
2728         return 0;
2729     }
2730 
2731     /* D1 and H1 */
2732     return -EBUSY;
2733 #else
2734     return -ENOTSUP;
2735 #endif
2736 }
2737 
2738 /*
2739  * Returns the allocation status of the specified offset.
2740  *
2741  * The block layer guarantees 'offset' and 'bytes' are within bounds.
2742  *
2743  * 'pnum' is set to the number of bytes (including and immediately following
2744  * the specified offset) that are known to be in the same
2745  * allocated/unallocated state.
2746  *
2747  * 'bytes' is the max value 'pnum' should be set to.
2748  */
2749 static int coroutine_fn raw_co_block_status(BlockDriverState *bs,
2750                                             bool want_zero,
2751                                             int64_t offset,
2752                                             int64_t bytes, int64_t *pnum,
2753                                             int64_t *map,
2754                                             BlockDriverState **file)
2755 {
2756     off_t data = 0, hole = 0;
2757     int ret;
2758 
2759     assert(QEMU_IS_ALIGNED(offset | bytes, bs->bl.request_alignment));
2760 
2761     ret = fd_open(bs);
2762     if (ret < 0) {
2763         return ret;
2764     }
2765 
2766     if (!want_zero) {
2767         *pnum = bytes;
2768         *map = offset;
2769         *file = bs;
2770         return BDRV_BLOCK_DATA | BDRV_BLOCK_OFFSET_VALID;
2771     }
2772 
2773     ret = find_allocation(bs, offset, &data, &hole);
2774     if (ret == -ENXIO) {
2775         /* Trailing hole */
2776         *pnum = bytes;
2777         ret = BDRV_BLOCK_ZERO;
2778     } else if (ret < 0) {
2779         /* No info available, so pretend there are no holes */
2780         *pnum = bytes;
2781         ret = BDRV_BLOCK_DATA;
2782     } else if (data == offset) {
2783         /* On a data extent, compute bytes to the end of the extent,
2784          * possibly including a partial sector at EOF. */
2785         *pnum = MIN(bytes, hole - offset);
2786 
2787         /*
2788          * We are not allowed to return partial sectors, though, so
2789          * round up if necessary.
2790          */
2791         if (!QEMU_IS_ALIGNED(*pnum, bs->bl.request_alignment)) {
2792             int64_t file_length = raw_getlength(bs);
2793             if (file_length > 0) {
2794                 /* Ignore errors, this is just a safeguard */
2795                 assert(hole == file_length);
2796             }
2797             *pnum = ROUND_UP(*pnum, bs->bl.request_alignment);
2798         }
2799 
2800         ret = BDRV_BLOCK_DATA;
2801     } else {
2802         /* On a hole, compute bytes to the beginning of the next extent.  */
2803         assert(hole == offset);
2804         *pnum = MIN(bytes, data - offset);
2805         ret = BDRV_BLOCK_ZERO;
2806     }
2807     *map = offset;
2808     *file = bs;
2809     return ret | BDRV_BLOCK_OFFSET_VALID;
2810 }
2811 
2812 #if defined(__linux__)
2813 /* Verify that the file is not in the page cache */
2814 static void check_cache_dropped(BlockDriverState *bs, Error **errp)
2815 {
2816     const size_t window_size = 128 * 1024 * 1024;
2817     BDRVRawState *s = bs->opaque;
2818     void *window = NULL;
2819     size_t length = 0;
2820     unsigned char *vec;
2821     size_t page_size;
2822     off_t offset;
2823     off_t end;
2824 
2825     /* mincore(2) page status information requires 1 byte per page */
2826     page_size = sysconf(_SC_PAGESIZE);
2827     vec = g_malloc(DIV_ROUND_UP(window_size, page_size));
2828 
2829     end = raw_getlength(bs);
2830 
2831     for (offset = 0; offset < end; offset += window_size) {
2832         void *new_window;
2833         size_t new_length;
2834         size_t vec_end;
2835         size_t i;
2836         int ret;
2837 
2838         /* Unmap previous window if size has changed */
2839         new_length = MIN(end - offset, window_size);
2840         if (new_length != length) {
2841             munmap(window, length);
2842             window = NULL;
2843             length = 0;
2844         }
2845 
2846         new_window = mmap(window, new_length, PROT_NONE, MAP_PRIVATE,
2847                           s->fd, offset);
2848         if (new_window == MAP_FAILED) {
2849             error_setg_errno(errp, errno, "mmap failed");
2850             break;
2851         }
2852 
2853         window = new_window;
2854         length = new_length;
2855 
2856         ret = mincore(window, length, vec);
2857         if (ret < 0) {
2858             error_setg_errno(errp, errno, "mincore failed");
2859             break;
2860         }
2861 
2862         vec_end = DIV_ROUND_UP(length, page_size);
2863         for (i = 0; i < vec_end; i++) {
2864             if (vec[i] & 0x1) {
2865                 break;
2866             }
2867         }
2868         if (i < vec_end) {
2869             error_setg(errp, "page cache still in use!");
2870             break;
2871         }
2872     }
2873 
2874     if (window) {
2875         munmap(window, length);
2876     }
2877 
2878     g_free(vec);
2879 }
2880 #endif /* __linux__ */
2881 
2882 static void coroutine_fn raw_co_invalidate_cache(BlockDriverState *bs,
2883                                                  Error **errp)
2884 {
2885     BDRVRawState *s = bs->opaque;
2886     int ret;
2887 
2888     ret = fd_open(bs);
2889     if (ret < 0) {
2890         error_setg_errno(errp, -ret, "The file descriptor is not open");
2891         return;
2892     }
2893 
2894     if (!s->drop_cache) {
2895         return;
2896     }
2897 
2898     if (s->open_flags & O_DIRECT) {
2899         return; /* No host kernel page cache */
2900     }
2901 
2902 #if defined(__linux__)
2903     /* This sets the scene for the next syscall... */
2904     ret = bdrv_co_flush(bs);
2905     if (ret < 0) {
2906         error_setg_errno(errp, -ret, "flush failed");
2907         return;
2908     }
2909 
2910     /* Linux does not invalidate pages that are dirty, locked, or mmapped by a
2911      * process.  These limitations are okay because we just fsynced the file,
2912      * we don't use mmap, and the file should not be in use by other processes.
2913      */
2914     ret = posix_fadvise(s->fd, 0, 0, POSIX_FADV_DONTNEED);
2915     if (ret != 0) { /* the return value is a positive errno */
2916         error_setg_errno(errp, ret, "fadvise failed");
2917         return;
2918     }
2919 
2920     if (s->check_cache_dropped) {
2921         check_cache_dropped(bs, errp);
2922     }
2923 #else /* __linux__ */
2924     /* Do nothing.  Live migration to a remote host with cache.direct=off is
2925      * unsupported on other host operating systems.  Cache consistency issues
2926      * may occur but no error is reported here, partly because that's the
2927      * historical behavior and partly because it's hard to differentiate valid
2928      * configurations that should not cause errors.
2929      */
2930 #endif /* !__linux__ */
2931 }
2932 
2933 static void raw_account_discard(BDRVRawState *s, uint64_t nbytes, int ret)
2934 {
2935     if (ret) {
2936         s->stats.discard_nb_failed++;
2937     } else {
2938         s->stats.discard_nb_ok++;
2939         s->stats.discard_bytes_ok += nbytes;
2940     }
2941 }
2942 
2943 static coroutine_fn int
2944 raw_do_pdiscard(BlockDriverState *bs, int64_t offset, int bytes, bool blkdev)
2945 {
2946     BDRVRawState *s = bs->opaque;
2947     RawPosixAIOData acb;
2948     int ret;
2949 
2950     acb = (RawPosixAIOData) {
2951         .bs             = bs,
2952         .aio_fildes     = s->fd,
2953         .aio_type       = QEMU_AIO_DISCARD,
2954         .aio_offset     = offset,
2955         .aio_nbytes     = bytes,
2956     };
2957 
2958     if (blkdev) {
2959         acb.aio_type |= QEMU_AIO_BLKDEV;
2960     }
2961 
2962     ret = raw_thread_pool_submit(bs, handle_aiocb_discard, &acb);
2963     raw_account_discard(s, bytes, ret);
2964     return ret;
2965 }
2966 
2967 static coroutine_fn int
2968 raw_co_pdiscard(BlockDriverState *bs, int64_t offset, int bytes)
2969 {
2970     return raw_do_pdiscard(bs, offset, bytes, false);
2971 }
2972 
2973 static int coroutine_fn
2974 raw_do_pwrite_zeroes(BlockDriverState *bs, int64_t offset, int bytes,
2975                      BdrvRequestFlags flags, bool blkdev)
2976 {
2977     BDRVRawState *s = bs->opaque;
2978     RawPosixAIOData acb;
2979     ThreadPoolFunc *handler;
2980 
2981 #ifdef CONFIG_FALLOCATE
2982     if (offset + bytes > bs->total_sectors * BDRV_SECTOR_SIZE) {
2983         BdrvTrackedRequest *req;
2984 
2985         /*
2986          * This is a workaround for a bug in the Linux XFS driver,
2987          * where writes submitted through the AIO interface will be
2988          * discarded if they happen beyond a concurrently running
2989          * fallocate() that increases the file length (i.e., both the
2990          * write and the fallocate() happen beyond the EOF).
2991          *
2992          * To work around it, we extend the tracked request for this
2993          * zero write until INT64_MAX (effectively infinity), and mark
2994          * it as serializing.
2995          *
2996          * We have to enable this workaround for all filesystems and
2997          * AIO modes (not just XFS with aio=native), because for
2998          * remote filesystems we do not know the host configuration.
2999          */
3000 
3001         req = bdrv_co_get_self_request(bs);
3002         assert(req);
3003         assert(req->type == BDRV_TRACKED_WRITE);
3004         assert(req->offset <= offset);
3005         assert(req->offset + req->bytes >= offset + bytes);
3006 
3007         req->bytes = BDRV_MAX_LENGTH - req->offset;
3008 
3009         bdrv_check_request(req->offset, req->bytes, &error_abort);
3010 
3011         bdrv_make_request_serialising(req, bs->bl.request_alignment);
3012     }
3013 #endif
3014 
3015     acb = (RawPosixAIOData) {
3016         .bs             = bs,
3017         .aio_fildes     = s->fd,
3018         .aio_type       = QEMU_AIO_WRITE_ZEROES,
3019         .aio_offset     = offset,
3020         .aio_nbytes     = bytes,
3021     };
3022 
3023     if (blkdev) {
3024         acb.aio_type |= QEMU_AIO_BLKDEV;
3025     }
3026     if (flags & BDRV_REQ_NO_FALLBACK) {
3027         acb.aio_type |= QEMU_AIO_NO_FALLBACK;
3028     }
3029 
3030     if (flags & BDRV_REQ_MAY_UNMAP) {
3031         acb.aio_type |= QEMU_AIO_DISCARD;
3032         handler = handle_aiocb_write_zeroes_unmap;
3033     } else {
3034         handler = handle_aiocb_write_zeroes;
3035     }
3036 
3037     return raw_thread_pool_submit(bs, handler, &acb);
3038 }
3039 
3040 static int coroutine_fn raw_co_pwrite_zeroes(
3041     BlockDriverState *bs, int64_t offset,
3042     int bytes, BdrvRequestFlags flags)
3043 {
3044     return raw_do_pwrite_zeroes(bs, offset, bytes, flags, false);
3045 }
3046 
3047 static int raw_get_info(BlockDriverState *bs, BlockDriverInfo *bdi)
3048 {
3049     return 0;
3050 }
3051 
3052 static BlockStatsSpecificFile get_blockstats_specific_file(BlockDriverState *bs)
3053 {
3054     BDRVRawState *s = bs->opaque;
3055     return (BlockStatsSpecificFile) {
3056         .discard_nb_ok = s->stats.discard_nb_ok,
3057         .discard_nb_failed = s->stats.discard_nb_failed,
3058         .discard_bytes_ok = s->stats.discard_bytes_ok,
3059     };
3060 }
3061 
3062 static BlockStatsSpecific *raw_get_specific_stats(BlockDriverState *bs)
3063 {
3064     BlockStatsSpecific *stats = g_new(BlockStatsSpecific, 1);
3065 
3066     stats->driver = BLOCKDEV_DRIVER_FILE;
3067     stats->u.file = get_blockstats_specific_file(bs);
3068 
3069     return stats;
3070 }
3071 
3072 #if defined(HAVE_HOST_BLOCK_DEVICE)
3073 static BlockStatsSpecific *hdev_get_specific_stats(BlockDriverState *bs)
3074 {
3075     BlockStatsSpecific *stats = g_new(BlockStatsSpecific, 1);
3076 
3077     stats->driver = BLOCKDEV_DRIVER_HOST_DEVICE;
3078     stats->u.host_device = get_blockstats_specific_file(bs);
3079 
3080     return stats;
3081 }
3082 #endif /* HAVE_HOST_BLOCK_DEVICE */
3083 
3084 static QemuOptsList raw_create_opts = {
3085     .name = "raw-create-opts",
3086     .head = QTAILQ_HEAD_INITIALIZER(raw_create_opts.head),
3087     .desc = {
3088         {
3089             .name = BLOCK_OPT_SIZE,
3090             .type = QEMU_OPT_SIZE,
3091             .help = "Virtual disk size"
3092         },
3093         {
3094             .name = BLOCK_OPT_NOCOW,
3095             .type = QEMU_OPT_BOOL,
3096             .help = "Turn off copy-on-write (valid only on btrfs)"
3097         },
3098         {
3099             .name = BLOCK_OPT_PREALLOC,
3100             .type = QEMU_OPT_STRING,
3101             .help = "Preallocation mode (allowed values: off"
3102 #ifdef CONFIG_POSIX_FALLOCATE
3103                     ", falloc"
3104 #endif
3105                     ", full)"
3106         },
3107         {
3108             .name = BLOCK_OPT_EXTENT_SIZE_HINT,
3109             .type = QEMU_OPT_SIZE,
3110             .help = "Extent size hint for the image file, 0 to disable"
3111         },
3112         { /* end of list */ }
3113     }
3114 };
3115 
3116 static int raw_check_perm(BlockDriverState *bs, uint64_t perm, uint64_t shared,
3117                           Error **errp)
3118 {
3119     BDRVRawState *s = bs->opaque;
3120     int input_flags = s->reopen_state ? s->reopen_state->flags : bs->open_flags;
3121     int open_flags;
3122     int ret;
3123 
3124     /* We may need a new fd if auto-read-only switches the mode */
3125     ret = raw_reconfigure_getfd(bs, input_flags, &open_flags, perm,
3126                                 false, errp);
3127     if (ret < 0) {
3128         return ret;
3129     } else if (ret != s->fd) {
3130         Error *local_err = NULL;
3131 
3132         /*
3133          * Fail already check_perm() if we can't get a working O_DIRECT
3134          * alignment with the new fd.
3135          */
3136         raw_probe_alignment(bs, ret, &local_err);
3137         if (local_err) {
3138             error_propagate(errp, local_err);
3139             return -EINVAL;
3140         }
3141 
3142         s->perm_change_fd = ret;
3143         s->perm_change_flags = open_flags;
3144     }
3145 
3146     /* Prepare permissions on old fd to avoid conflicts between old and new,
3147      * but keep everything locked that new will need. */
3148     ret = raw_handle_perm_lock(bs, RAW_PL_PREPARE, perm, shared, errp);
3149     if (ret < 0) {
3150         goto fail;
3151     }
3152 
3153     /* Copy locks to the new fd */
3154     if (s->perm_change_fd && s->use_lock) {
3155         ret = raw_apply_lock_bytes(NULL, s->perm_change_fd, perm, ~shared,
3156                                    false, errp);
3157         if (ret < 0) {
3158             raw_handle_perm_lock(bs, RAW_PL_ABORT, 0, 0, NULL);
3159             goto fail;
3160         }
3161     }
3162     return 0;
3163 
3164 fail:
3165     if (s->perm_change_fd) {
3166         qemu_close(s->perm_change_fd);
3167     }
3168     s->perm_change_fd = 0;
3169     return ret;
3170 }
3171 
3172 static void raw_set_perm(BlockDriverState *bs, uint64_t perm, uint64_t shared)
3173 {
3174     BDRVRawState *s = bs->opaque;
3175 
3176     /* For reopen, we have already switched to the new fd (.bdrv_set_perm is
3177      * called after .bdrv_reopen_commit) */
3178     if (s->perm_change_fd && s->fd != s->perm_change_fd) {
3179         qemu_close(s->fd);
3180         s->fd = s->perm_change_fd;
3181         s->open_flags = s->perm_change_flags;
3182     }
3183     s->perm_change_fd = 0;
3184 
3185     raw_handle_perm_lock(bs, RAW_PL_COMMIT, perm, shared, NULL);
3186     s->perm = perm;
3187     s->shared_perm = shared;
3188 }
3189 
3190 static void raw_abort_perm_update(BlockDriverState *bs)
3191 {
3192     BDRVRawState *s = bs->opaque;
3193 
3194     /* For reopen, .bdrv_reopen_abort is called afterwards and will close
3195      * the file descriptor. */
3196     if (s->perm_change_fd) {
3197         qemu_close(s->perm_change_fd);
3198     }
3199     s->perm_change_fd = 0;
3200 
3201     raw_handle_perm_lock(bs, RAW_PL_ABORT, 0, 0, NULL);
3202 }
3203 
3204 static int coroutine_fn raw_co_copy_range_from(
3205         BlockDriverState *bs, BdrvChild *src, uint64_t src_offset,
3206         BdrvChild *dst, uint64_t dst_offset, uint64_t bytes,
3207         BdrvRequestFlags read_flags, BdrvRequestFlags write_flags)
3208 {
3209     return bdrv_co_copy_range_to(src, src_offset, dst, dst_offset, bytes,
3210                                  read_flags, write_flags);
3211 }
3212 
3213 static int coroutine_fn raw_co_copy_range_to(BlockDriverState *bs,
3214                                              BdrvChild *src,
3215                                              uint64_t src_offset,
3216                                              BdrvChild *dst,
3217                                              uint64_t dst_offset,
3218                                              uint64_t bytes,
3219                                              BdrvRequestFlags read_flags,
3220                                              BdrvRequestFlags write_flags)
3221 {
3222     RawPosixAIOData acb;
3223     BDRVRawState *s = bs->opaque;
3224     BDRVRawState *src_s;
3225 
3226     assert(dst->bs == bs);
3227     if (src->bs->drv->bdrv_co_copy_range_to != raw_co_copy_range_to) {
3228         return -ENOTSUP;
3229     }
3230 
3231     src_s = src->bs->opaque;
3232     if (fd_open(src->bs) < 0 || fd_open(dst->bs) < 0) {
3233         return -EIO;
3234     }
3235 
3236     acb = (RawPosixAIOData) {
3237         .bs             = bs,
3238         .aio_type       = QEMU_AIO_COPY_RANGE,
3239         .aio_fildes     = src_s->fd,
3240         .aio_offset     = src_offset,
3241         .aio_nbytes     = bytes,
3242         .copy_range     = {
3243             .aio_fd2        = s->fd,
3244             .aio_offset2    = dst_offset,
3245         },
3246     };
3247 
3248     return raw_thread_pool_submit(bs, handle_aiocb_copy_range, &acb);
3249 }
3250 
3251 BlockDriver bdrv_file = {
3252     .format_name = "file",
3253     .protocol_name = "file",
3254     .instance_size = sizeof(BDRVRawState),
3255     .bdrv_needs_filename = true,
3256     .bdrv_probe = NULL, /* no probe for protocols */
3257     .bdrv_parse_filename = raw_parse_filename,
3258     .bdrv_file_open = raw_open,
3259     .bdrv_reopen_prepare = raw_reopen_prepare,
3260     .bdrv_reopen_commit = raw_reopen_commit,
3261     .bdrv_reopen_abort = raw_reopen_abort,
3262     .bdrv_close = raw_close,
3263     .bdrv_co_create = raw_co_create,
3264     .bdrv_co_create_opts = raw_co_create_opts,
3265     .bdrv_has_zero_init = bdrv_has_zero_init_1,
3266     .bdrv_co_block_status = raw_co_block_status,
3267     .bdrv_co_invalidate_cache = raw_co_invalidate_cache,
3268     .bdrv_co_pwrite_zeroes = raw_co_pwrite_zeroes,
3269     .bdrv_co_delete_file = raw_co_delete_file,
3270 
3271     .bdrv_co_preadv         = raw_co_preadv,
3272     .bdrv_co_pwritev        = raw_co_pwritev,
3273     .bdrv_co_flush_to_disk  = raw_co_flush_to_disk,
3274     .bdrv_co_pdiscard       = raw_co_pdiscard,
3275     .bdrv_co_copy_range_from = raw_co_copy_range_from,
3276     .bdrv_co_copy_range_to  = raw_co_copy_range_to,
3277     .bdrv_refresh_limits = raw_refresh_limits,
3278     .bdrv_io_plug = raw_aio_plug,
3279     .bdrv_io_unplug = raw_aio_unplug,
3280     .bdrv_attach_aio_context = raw_aio_attach_aio_context,
3281 
3282     .bdrv_co_truncate = raw_co_truncate,
3283     .bdrv_getlength = raw_getlength,
3284     .bdrv_get_info = raw_get_info,
3285     .bdrv_get_allocated_file_size
3286                         = raw_get_allocated_file_size,
3287     .bdrv_get_specific_stats = raw_get_specific_stats,
3288     .bdrv_check_perm = raw_check_perm,
3289     .bdrv_set_perm   = raw_set_perm,
3290     .bdrv_abort_perm_update = raw_abort_perm_update,
3291     .create_opts = &raw_create_opts,
3292     .mutable_opts = mutable_opts,
3293 };
3294 
3295 /***********************************************/
3296 /* host device */
3297 
3298 #if defined(HAVE_HOST_BLOCK_DEVICE)
3299 
3300 #if defined(__APPLE__) && defined(__MACH__)
3301 static kern_return_t GetBSDPath(io_iterator_t mediaIterator, char *bsdPath,
3302                                 CFIndex maxPathSize, int flags);
3303 static char *FindEjectableOpticalMedia(io_iterator_t *mediaIterator)
3304 {
3305     kern_return_t kernResult = KERN_FAILURE;
3306     mach_port_t     masterPort;
3307     CFMutableDictionaryRef  classesToMatch;
3308     const char *matching_array[] = {kIODVDMediaClass, kIOCDMediaClass};
3309     char *mediaType = NULL;
3310 
3311     kernResult = IOMasterPort( MACH_PORT_NULL, &masterPort );
3312     if ( KERN_SUCCESS != kernResult ) {
3313         printf( "IOMasterPort returned %d\n", kernResult );
3314     }
3315 
3316     int index;
3317     for (index = 0; index < ARRAY_SIZE(matching_array); index++) {
3318         classesToMatch = IOServiceMatching(matching_array[index]);
3319         if (classesToMatch == NULL) {
3320             error_report("IOServiceMatching returned NULL for %s",
3321                          matching_array[index]);
3322             continue;
3323         }
3324         CFDictionarySetValue(classesToMatch, CFSTR(kIOMediaEjectableKey),
3325                              kCFBooleanTrue);
3326         kernResult = IOServiceGetMatchingServices(masterPort, classesToMatch,
3327                                                   mediaIterator);
3328         if (kernResult != KERN_SUCCESS) {
3329             error_report("Note: IOServiceGetMatchingServices returned %d",
3330                          kernResult);
3331             continue;
3332         }
3333 
3334         /* If a match was found, leave the loop */
3335         if (*mediaIterator != 0) {
3336             trace_file_FindEjectableOpticalMedia(matching_array[index]);
3337             mediaType = g_strdup(matching_array[index]);
3338             break;
3339         }
3340     }
3341     return mediaType;
3342 }
3343 
3344 kern_return_t GetBSDPath(io_iterator_t mediaIterator, char *bsdPath,
3345                          CFIndex maxPathSize, int flags)
3346 {
3347     io_object_t     nextMedia;
3348     kern_return_t   kernResult = KERN_FAILURE;
3349     *bsdPath = '\0';
3350     nextMedia = IOIteratorNext( mediaIterator );
3351     if ( nextMedia )
3352     {
3353         CFTypeRef   bsdPathAsCFString;
3354     bsdPathAsCFString = IORegistryEntryCreateCFProperty( nextMedia, CFSTR( kIOBSDNameKey ), kCFAllocatorDefault, 0 );
3355         if ( bsdPathAsCFString ) {
3356             size_t devPathLength;
3357             strcpy( bsdPath, _PATH_DEV );
3358             if (flags & BDRV_O_NOCACHE) {
3359                 strcat(bsdPath, "r");
3360             }
3361             devPathLength = strlen( bsdPath );
3362             if ( CFStringGetCString( bsdPathAsCFString, bsdPath + devPathLength, maxPathSize - devPathLength, kCFStringEncodingASCII ) ) {
3363                 kernResult = KERN_SUCCESS;
3364             }
3365             CFRelease( bsdPathAsCFString );
3366         }
3367         IOObjectRelease( nextMedia );
3368     }
3369 
3370     return kernResult;
3371 }
3372 
3373 /* Sets up a real cdrom for use in QEMU */
3374 static bool setup_cdrom(char *bsd_path, Error **errp)
3375 {
3376     int index, num_of_test_partitions = 2, fd;
3377     char test_partition[MAXPATHLEN];
3378     bool partition_found = false;
3379 
3380     /* look for a working partition */
3381     for (index = 0; index < num_of_test_partitions; index++) {
3382         snprintf(test_partition, sizeof(test_partition), "%ss%d", bsd_path,
3383                  index);
3384         fd = qemu_open(test_partition, O_RDONLY | O_BINARY | O_LARGEFILE, NULL);
3385         if (fd >= 0) {
3386             partition_found = true;
3387             qemu_close(fd);
3388             break;
3389         }
3390     }
3391 
3392     /* if a working partition on the device was not found */
3393     if (partition_found == false) {
3394         error_setg(errp, "Failed to find a working partition on disc");
3395     } else {
3396         trace_file_setup_cdrom(test_partition);
3397         pstrcpy(bsd_path, MAXPATHLEN, test_partition);
3398     }
3399     return partition_found;
3400 }
3401 
3402 /* Prints directions on mounting and unmounting a device */
3403 static void print_unmounting_directions(const char *file_name)
3404 {
3405     error_report("If device %s is mounted on the desktop, unmount"
3406                  " it first before using it in QEMU", file_name);
3407     error_report("Command to unmount device: diskutil unmountDisk %s",
3408                  file_name);
3409     error_report("Command to mount device: diskutil mountDisk %s", file_name);
3410 }
3411 
3412 #endif /* defined(__APPLE__) && defined(__MACH__) */
3413 
3414 static int hdev_probe_device(const char *filename)
3415 {
3416     struct stat st;
3417 
3418     /* allow a dedicated CD-ROM driver to match with a higher priority */
3419     if (strstart(filename, "/dev/cdrom", NULL))
3420         return 50;
3421 
3422     if (stat(filename, &st) >= 0 &&
3423             (S_ISCHR(st.st_mode) || S_ISBLK(st.st_mode))) {
3424         return 100;
3425     }
3426 
3427     return 0;
3428 }
3429 
3430 static void hdev_parse_filename(const char *filename, QDict *options,
3431                                 Error **errp)
3432 {
3433     bdrv_parse_filename_strip_prefix(filename, "host_device:", options);
3434 }
3435 
3436 static bool hdev_is_sg(BlockDriverState *bs)
3437 {
3438 
3439 #if defined(__linux__)
3440 
3441     BDRVRawState *s = bs->opaque;
3442     struct stat st;
3443     struct sg_scsi_id scsiid;
3444     int sg_version;
3445     int ret;
3446 
3447     if (stat(bs->filename, &st) < 0 || !S_ISCHR(st.st_mode)) {
3448         return false;
3449     }
3450 
3451     ret = ioctl(s->fd, SG_GET_VERSION_NUM, &sg_version);
3452     if (ret < 0) {
3453         return false;
3454     }
3455 
3456     ret = ioctl(s->fd, SG_GET_SCSI_ID, &scsiid);
3457     if (ret >= 0) {
3458         trace_file_hdev_is_sg(scsiid.scsi_type, sg_version);
3459         return true;
3460     }
3461 
3462 #endif
3463 
3464     return false;
3465 }
3466 
3467 static int hdev_open(BlockDriverState *bs, QDict *options, int flags,
3468                      Error **errp)
3469 {
3470     BDRVRawState *s = bs->opaque;
3471     int ret;
3472 
3473 #if defined(__APPLE__) && defined(__MACH__)
3474     /*
3475      * Caution: while qdict_get_str() is fine, getting non-string types
3476      * would require more care.  When @options come from -blockdev or
3477      * blockdev_add, its members are typed according to the QAPI
3478      * schema, but when they come from -drive, they're all QString.
3479      */
3480     const char *filename = qdict_get_str(options, "filename");
3481     char bsd_path[MAXPATHLEN] = "";
3482     bool error_occurred = false;
3483 
3484     /* If using a real cdrom */
3485     if (strcmp(filename, "/dev/cdrom") == 0) {
3486         char *mediaType = NULL;
3487         kern_return_t ret_val;
3488         io_iterator_t mediaIterator = 0;
3489 
3490         mediaType = FindEjectableOpticalMedia(&mediaIterator);
3491         if (mediaType == NULL) {
3492             error_setg(errp, "Please make sure your CD/DVD is in the optical"
3493                        " drive");
3494             error_occurred = true;
3495             goto hdev_open_Mac_error;
3496         }
3497 
3498         ret_val = GetBSDPath(mediaIterator, bsd_path, sizeof(bsd_path), flags);
3499         if (ret_val != KERN_SUCCESS) {
3500             error_setg(errp, "Could not get BSD path for optical drive");
3501             error_occurred = true;
3502             goto hdev_open_Mac_error;
3503         }
3504 
3505         /* If a real optical drive was not found */
3506         if (bsd_path[0] == '\0') {
3507             error_setg(errp, "Failed to obtain bsd path for optical drive");
3508             error_occurred = true;
3509             goto hdev_open_Mac_error;
3510         }
3511 
3512         /* If using a cdrom disc and finding a partition on the disc failed */
3513         if (strncmp(mediaType, kIOCDMediaClass, 9) == 0 &&
3514             setup_cdrom(bsd_path, errp) == false) {
3515             print_unmounting_directions(bsd_path);
3516             error_occurred = true;
3517             goto hdev_open_Mac_error;
3518         }
3519 
3520         qdict_put_str(options, "filename", bsd_path);
3521 
3522 hdev_open_Mac_error:
3523         g_free(mediaType);
3524         if (mediaIterator) {
3525             IOObjectRelease(mediaIterator);
3526         }
3527         if (error_occurred) {
3528             return -ENOENT;
3529         }
3530     }
3531 #endif /* defined(__APPLE__) && defined(__MACH__) */
3532 
3533     s->type = FTYPE_FILE;
3534 
3535     ret = raw_open_common(bs, options, flags, 0, true, errp);
3536     if (ret < 0) {
3537 #if defined(__APPLE__) && defined(__MACH__)
3538         if (*bsd_path) {
3539             filename = bsd_path;
3540         }
3541         /* if a physical device experienced an error while being opened */
3542         if (strncmp(filename, "/dev/", 5) == 0) {
3543             print_unmounting_directions(filename);
3544         }
3545 #endif /* defined(__APPLE__) && defined(__MACH__) */
3546         return ret;
3547     }
3548 
3549     /* Since this does ioctl the device must be already opened */
3550     bs->sg = hdev_is_sg(bs);
3551 
3552     return ret;
3553 }
3554 
3555 #if defined(__linux__)
3556 static int coroutine_fn
3557 hdev_co_ioctl(BlockDriverState *bs, unsigned long int req, void *buf)
3558 {
3559     BDRVRawState *s = bs->opaque;
3560     RawPosixAIOData acb;
3561     int ret;
3562 
3563     ret = fd_open(bs);
3564     if (ret < 0) {
3565         return ret;
3566     }
3567 
3568     if (req == SG_IO && s->pr_mgr) {
3569         struct sg_io_hdr *io_hdr = buf;
3570         if (io_hdr->cmdp[0] == PERSISTENT_RESERVE_OUT ||
3571             io_hdr->cmdp[0] == PERSISTENT_RESERVE_IN) {
3572             return pr_manager_execute(s->pr_mgr, bdrv_get_aio_context(bs),
3573                                       s->fd, io_hdr);
3574         }
3575     }
3576 
3577     acb = (RawPosixAIOData) {
3578         .bs         = bs,
3579         .aio_type   = QEMU_AIO_IOCTL,
3580         .aio_fildes = s->fd,
3581         .aio_offset = 0,
3582         .ioctl      = {
3583             .buf        = buf,
3584             .cmd        = req,
3585         },
3586     };
3587 
3588     return raw_thread_pool_submit(bs, handle_aiocb_ioctl, &acb);
3589 }
3590 #endif /* linux */
3591 
3592 static coroutine_fn int
3593 hdev_co_pdiscard(BlockDriverState *bs, int64_t offset, int bytes)
3594 {
3595     BDRVRawState *s = bs->opaque;
3596     int ret;
3597 
3598     ret = fd_open(bs);
3599     if (ret < 0) {
3600         raw_account_discard(s, bytes, ret);
3601         return ret;
3602     }
3603     return raw_do_pdiscard(bs, offset, bytes, true);
3604 }
3605 
3606 static coroutine_fn int hdev_co_pwrite_zeroes(BlockDriverState *bs,
3607     int64_t offset, int bytes, BdrvRequestFlags flags)
3608 {
3609     int rc;
3610 
3611     rc = fd_open(bs);
3612     if (rc < 0) {
3613         return rc;
3614     }
3615 
3616     return raw_do_pwrite_zeroes(bs, offset, bytes, flags, true);
3617 }
3618 
3619 static BlockDriver bdrv_host_device = {
3620     .format_name        = "host_device",
3621     .protocol_name        = "host_device",
3622     .instance_size      = sizeof(BDRVRawState),
3623     .bdrv_needs_filename = true,
3624     .bdrv_probe_device  = hdev_probe_device,
3625     .bdrv_parse_filename = hdev_parse_filename,
3626     .bdrv_file_open     = hdev_open,
3627     .bdrv_close         = raw_close,
3628     .bdrv_reopen_prepare = raw_reopen_prepare,
3629     .bdrv_reopen_commit  = raw_reopen_commit,
3630     .bdrv_reopen_abort   = raw_reopen_abort,
3631     .bdrv_co_create_opts = bdrv_co_create_opts_simple,
3632     .create_opts         = &bdrv_create_opts_simple,
3633     .mutable_opts        = mutable_opts,
3634     .bdrv_co_invalidate_cache = raw_co_invalidate_cache,
3635     .bdrv_co_pwrite_zeroes = hdev_co_pwrite_zeroes,
3636 
3637     .bdrv_co_preadv         = raw_co_preadv,
3638     .bdrv_co_pwritev        = raw_co_pwritev,
3639     .bdrv_co_flush_to_disk  = raw_co_flush_to_disk,
3640     .bdrv_co_pdiscard       = hdev_co_pdiscard,
3641     .bdrv_co_copy_range_from = raw_co_copy_range_from,
3642     .bdrv_co_copy_range_to  = raw_co_copy_range_to,
3643     .bdrv_refresh_limits = raw_refresh_limits,
3644     .bdrv_io_plug = raw_aio_plug,
3645     .bdrv_io_unplug = raw_aio_unplug,
3646     .bdrv_attach_aio_context = raw_aio_attach_aio_context,
3647 
3648     .bdrv_co_truncate       = raw_co_truncate,
3649     .bdrv_getlength	= raw_getlength,
3650     .bdrv_get_info = raw_get_info,
3651     .bdrv_get_allocated_file_size
3652                         = raw_get_allocated_file_size,
3653     .bdrv_get_specific_stats = hdev_get_specific_stats,
3654     .bdrv_check_perm = raw_check_perm,
3655     .bdrv_set_perm   = raw_set_perm,
3656     .bdrv_abort_perm_update = raw_abort_perm_update,
3657     .bdrv_probe_blocksizes = hdev_probe_blocksizes,
3658     .bdrv_probe_geometry = hdev_probe_geometry,
3659 
3660     /* generic scsi device */
3661 #ifdef __linux__
3662     .bdrv_co_ioctl          = hdev_co_ioctl,
3663 #endif
3664 };
3665 
3666 #if defined(__linux__) || defined(__FreeBSD__) || defined(__FreeBSD_kernel__)
3667 static void cdrom_parse_filename(const char *filename, QDict *options,
3668                                  Error **errp)
3669 {
3670     bdrv_parse_filename_strip_prefix(filename, "host_cdrom:", options);
3671 }
3672 #endif
3673 
3674 #ifdef __linux__
3675 static int cdrom_open(BlockDriverState *bs, QDict *options, int flags,
3676                       Error **errp)
3677 {
3678     BDRVRawState *s = bs->opaque;
3679 
3680     s->type = FTYPE_CD;
3681 
3682     /* open will not fail even if no CD is inserted, so add O_NONBLOCK */
3683     return raw_open_common(bs, options, flags, O_NONBLOCK, true, errp);
3684 }
3685 
3686 static int cdrom_probe_device(const char *filename)
3687 {
3688     int fd, ret;
3689     int prio = 0;
3690     struct stat st;
3691 
3692     fd = qemu_open(filename, O_RDONLY | O_NONBLOCK, NULL);
3693     if (fd < 0) {
3694         goto out;
3695     }
3696     ret = fstat(fd, &st);
3697     if (ret == -1 || !S_ISBLK(st.st_mode)) {
3698         goto outc;
3699     }
3700 
3701     /* Attempt to detect via a CDROM specific ioctl */
3702     ret = ioctl(fd, CDROM_DRIVE_STATUS, CDSL_CURRENT);
3703     if (ret >= 0)
3704         prio = 100;
3705 
3706 outc:
3707     qemu_close(fd);
3708 out:
3709     return prio;
3710 }
3711 
3712 static bool cdrom_is_inserted(BlockDriverState *bs)
3713 {
3714     BDRVRawState *s = bs->opaque;
3715     int ret;
3716 
3717     ret = ioctl(s->fd, CDROM_DRIVE_STATUS, CDSL_CURRENT);
3718     return ret == CDS_DISC_OK;
3719 }
3720 
3721 static void cdrom_eject(BlockDriverState *bs, bool eject_flag)
3722 {
3723     BDRVRawState *s = bs->opaque;
3724 
3725     if (eject_flag) {
3726         if (ioctl(s->fd, CDROMEJECT, NULL) < 0)
3727             perror("CDROMEJECT");
3728     } else {
3729         if (ioctl(s->fd, CDROMCLOSETRAY, NULL) < 0)
3730             perror("CDROMEJECT");
3731     }
3732 }
3733 
3734 static void cdrom_lock_medium(BlockDriverState *bs, bool locked)
3735 {
3736     BDRVRawState *s = bs->opaque;
3737 
3738     if (ioctl(s->fd, CDROM_LOCKDOOR, locked) < 0) {
3739         /*
3740          * Note: an error can happen if the distribution automatically
3741          * mounts the CD-ROM
3742          */
3743         /* perror("CDROM_LOCKDOOR"); */
3744     }
3745 }
3746 
3747 static BlockDriver bdrv_host_cdrom = {
3748     .format_name        = "host_cdrom",
3749     .protocol_name      = "host_cdrom",
3750     .instance_size      = sizeof(BDRVRawState),
3751     .bdrv_needs_filename = true,
3752     .bdrv_probe_device	= cdrom_probe_device,
3753     .bdrv_parse_filename = cdrom_parse_filename,
3754     .bdrv_file_open     = cdrom_open,
3755     .bdrv_close         = raw_close,
3756     .bdrv_reopen_prepare = raw_reopen_prepare,
3757     .bdrv_reopen_commit  = raw_reopen_commit,
3758     .bdrv_reopen_abort   = raw_reopen_abort,
3759     .bdrv_co_create_opts = bdrv_co_create_opts_simple,
3760     .create_opts         = &bdrv_create_opts_simple,
3761     .mutable_opts        = mutable_opts,
3762     .bdrv_co_invalidate_cache = raw_co_invalidate_cache,
3763 
3764     .bdrv_co_preadv         = raw_co_preadv,
3765     .bdrv_co_pwritev        = raw_co_pwritev,
3766     .bdrv_co_flush_to_disk  = raw_co_flush_to_disk,
3767     .bdrv_refresh_limits = raw_refresh_limits,
3768     .bdrv_io_plug = raw_aio_plug,
3769     .bdrv_io_unplug = raw_aio_unplug,
3770     .bdrv_attach_aio_context = raw_aio_attach_aio_context,
3771 
3772     .bdrv_co_truncate    = raw_co_truncate,
3773     .bdrv_getlength      = raw_getlength,
3774     .has_variable_length = true,
3775     .bdrv_get_allocated_file_size
3776                         = raw_get_allocated_file_size,
3777 
3778     /* removable device support */
3779     .bdrv_is_inserted   = cdrom_is_inserted,
3780     .bdrv_eject         = cdrom_eject,
3781     .bdrv_lock_medium   = cdrom_lock_medium,
3782 
3783     /* generic scsi device */
3784     .bdrv_co_ioctl      = hdev_co_ioctl,
3785 };
3786 #endif /* __linux__ */
3787 
3788 #if defined (__FreeBSD__) || defined(__FreeBSD_kernel__)
3789 static int cdrom_open(BlockDriverState *bs, QDict *options, int flags,
3790                       Error **errp)
3791 {
3792     BDRVRawState *s = bs->opaque;
3793     int ret;
3794 
3795     s->type = FTYPE_CD;
3796 
3797     ret = raw_open_common(bs, options, flags, 0, true, errp);
3798     if (ret) {
3799         return ret;
3800     }
3801 
3802     /* make sure the door isn't locked at this time */
3803     ioctl(s->fd, CDIOCALLOW);
3804     return 0;
3805 }
3806 
3807 static int cdrom_probe_device(const char *filename)
3808 {
3809     if (strstart(filename, "/dev/cd", NULL) ||
3810             strstart(filename, "/dev/acd", NULL))
3811         return 100;
3812     return 0;
3813 }
3814 
3815 static int cdrom_reopen(BlockDriverState *bs)
3816 {
3817     BDRVRawState *s = bs->opaque;
3818     int fd;
3819 
3820     /*
3821      * Force reread of possibly changed/newly loaded disc,
3822      * FreeBSD seems to not notice sometimes...
3823      */
3824     if (s->fd >= 0)
3825         qemu_close(s->fd);
3826     fd = qemu_open(bs->filename, s->open_flags, NULL);
3827     if (fd < 0) {
3828         s->fd = -1;
3829         return -EIO;
3830     }
3831     s->fd = fd;
3832 
3833     /* make sure the door isn't locked at this time */
3834     ioctl(s->fd, CDIOCALLOW);
3835     return 0;
3836 }
3837 
3838 static bool cdrom_is_inserted(BlockDriverState *bs)
3839 {
3840     return raw_getlength(bs) > 0;
3841 }
3842 
3843 static void cdrom_eject(BlockDriverState *bs, bool eject_flag)
3844 {
3845     BDRVRawState *s = bs->opaque;
3846 
3847     if (s->fd < 0)
3848         return;
3849 
3850     (void) ioctl(s->fd, CDIOCALLOW);
3851 
3852     if (eject_flag) {
3853         if (ioctl(s->fd, CDIOCEJECT) < 0)
3854             perror("CDIOCEJECT");
3855     } else {
3856         if (ioctl(s->fd, CDIOCCLOSE) < 0)
3857             perror("CDIOCCLOSE");
3858     }
3859 
3860     cdrom_reopen(bs);
3861 }
3862 
3863 static void cdrom_lock_medium(BlockDriverState *bs, bool locked)
3864 {
3865     BDRVRawState *s = bs->opaque;
3866 
3867     if (s->fd < 0)
3868         return;
3869     if (ioctl(s->fd, (locked ? CDIOCPREVENT : CDIOCALLOW)) < 0) {
3870         /*
3871          * Note: an error can happen if the distribution automatically
3872          * mounts the CD-ROM
3873          */
3874         /* perror("CDROM_LOCKDOOR"); */
3875     }
3876 }
3877 
3878 static BlockDriver bdrv_host_cdrom = {
3879     .format_name        = "host_cdrom",
3880     .protocol_name      = "host_cdrom",
3881     .instance_size      = sizeof(BDRVRawState),
3882     .bdrv_needs_filename = true,
3883     .bdrv_probe_device	= cdrom_probe_device,
3884     .bdrv_parse_filename = cdrom_parse_filename,
3885     .bdrv_file_open     = cdrom_open,
3886     .bdrv_close         = raw_close,
3887     .bdrv_reopen_prepare = raw_reopen_prepare,
3888     .bdrv_reopen_commit  = raw_reopen_commit,
3889     .bdrv_reopen_abort   = raw_reopen_abort,
3890     .bdrv_co_create_opts = bdrv_co_create_opts_simple,
3891     .create_opts         = &bdrv_create_opts_simple,
3892     .mutable_opts       = mutable_opts,
3893 
3894     .bdrv_co_preadv         = raw_co_preadv,
3895     .bdrv_co_pwritev        = raw_co_pwritev,
3896     .bdrv_co_flush_to_disk  = raw_co_flush_to_disk,
3897     .bdrv_refresh_limits = raw_refresh_limits,
3898     .bdrv_io_plug = raw_aio_plug,
3899     .bdrv_io_unplug = raw_aio_unplug,
3900     .bdrv_attach_aio_context = raw_aio_attach_aio_context,
3901 
3902     .bdrv_co_truncate    = raw_co_truncate,
3903     .bdrv_getlength      = raw_getlength,
3904     .has_variable_length = true,
3905     .bdrv_get_allocated_file_size
3906                         = raw_get_allocated_file_size,
3907 
3908     /* removable device support */
3909     .bdrv_is_inserted   = cdrom_is_inserted,
3910     .bdrv_eject         = cdrom_eject,
3911     .bdrv_lock_medium   = cdrom_lock_medium,
3912 };
3913 #endif /* __FreeBSD__ */
3914 
3915 #endif /* HAVE_HOST_BLOCK_DEVICE */
3916 
3917 static void bdrv_file_init(void)
3918 {
3919     /*
3920      * Register all the drivers.  Note that order is important, the driver
3921      * registered last will get probed first.
3922      */
3923     bdrv_register(&bdrv_file);
3924 #if defined(HAVE_HOST_BLOCK_DEVICE)
3925     bdrv_register(&bdrv_host_device);
3926 #ifdef __linux__
3927     bdrv_register(&bdrv_host_cdrom);
3928 #endif
3929 #if defined(__FreeBSD__) || defined(__FreeBSD_kernel__)
3930     bdrv_register(&bdrv_host_cdrom);
3931 #endif
3932 #endif /* HAVE_HOST_BLOCK_DEVICE */
3933 }
3934 
3935 block_init(bdrv_file_init);
3936