xref: /openbmc/qemu/block/file-posix.c (revision 77d361b1)
1 /*
2  * Block driver for RAW files (posix)
3  *
4  * Copyright (c) 2006 Fabrice Bellard
5  *
6  * Permission is hereby granted, free of charge, to any person obtaining a copy
7  * of this software and associated documentation files (the "Software"), to deal
8  * in the Software without restriction, including without limitation the rights
9  * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
10  * copies of the Software, and to permit persons to whom the Software is
11  * furnished to do so, subject to the following conditions:
12  *
13  * The above copyright notice and this permission notice shall be included in
14  * all copies or substantial portions of the Software.
15  *
16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
19  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
21  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
22  * THE SOFTWARE.
23  */
24 
25 #include "qemu/osdep.h"
26 #include "qapi/error.h"
27 #include "qemu/cutils.h"
28 #include "qemu/error-report.h"
29 #include "block/block_int.h"
30 #include "qemu/module.h"
31 #include "qemu/option.h"
32 #include "trace.h"
33 #include "block/thread-pool.h"
34 #include "qemu/iov.h"
35 #include "block/raw-aio.h"
36 #include "qapi/qmp/qdict.h"
37 #include "qapi/qmp/qstring.h"
38 
39 #include "scsi/pr-manager.h"
40 #include "scsi/constants.h"
41 
42 #if defined(__APPLE__) && (__MACH__)
43 #include <paths.h>
44 #include <sys/param.h>
45 #include <IOKit/IOKitLib.h>
46 #include <IOKit/IOBSD.h>
47 #include <IOKit/storage/IOMediaBSDClient.h>
48 #include <IOKit/storage/IOMedia.h>
49 #include <IOKit/storage/IOCDMedia.h>
50 //#include <IOKit/storage/IOCDTypes.h>
51 #include <IOKit/storage/IODVDMedia.h>
52 #include <CoreFoundation/CoreFoundation.h>
53 #endif
54 
55 #ifdef __sun__
56 #define _POSIX_PTHREAD_SEMANTICS 1
57 #include <sys/dkio.h>
58 #endif
59 #ifdef __linux__
60 #include <sys/ioctl.h>
61 #include <sys/param.h>
62 #include <sys/syscall.h>
63 #include <linux/cdrom.h>
64 #include <linux/fd.h>
65 #include <linux/fs.h>
66 #include <linux/hdreg.h>
67 #include <scsi/sg.h>
68 #ifdef __s390__
69 #include <asm/dasd.h>
70 #endif
71 #ifndef FS_NOCOW_FL
72 #define FS_NOCOW_FL                     0x00800000 /* Do not cow file */
73 #endif
74 #endif
75 #if defined(CONFIG_FALLOCATE_PUNCH_HOLE) || defined(CONFIG_FALLOCATE_ZERO_RANGE)
76 #include <linux/falloc.h>
77 #endif
78 #if defined (__FreeBSD__) || defined(__FreeBSD_kernel__)
79 #include <sys/disk.h>
80 #include <sys/cdio.h>
81 #endif
82 
83 #ifdef __OpenBSD__
84 #include <sys/ioctl.h>
85 #include <sys/disklabel.h>
86 #include <sys/dkio.h>
87 #endif
88 
89 #ifdef __NetBSD__
90 #include <sys/ioctl.h>
91 #include <sys/disklabel.h>
92 #include <sys/dkio.h>
93 #include <sys/disk.h>
94 #endif
95 
96 #ifdef __DragonFly__
97 #include <sys/ioctl.h>
98 #include <sys/diskslice.h>
99 #endif
100 
101 #ifdef CONFIG_XFS
102 #include <xfs/xfs.h>
103 #endif
104 
105 //#define DEBUG_BLOCK
106 
107 #ifdef DEBUG_BLOCK
108 # define DEBUG_BLOCK_PRINT 1
109 #else
110 # define DEBUG_BLOCK_PRINT 0
111 #endif
112 #define DPRINTF(fmt, ...) \
113 do { \
114     if (DEBUG_BLOCK_PRINT) { \
115         printf(fmt, ## __VA_ARGS__); \
116     } \
117 } while (0)
118 
119 /* OS X does not have O_DSYNC */
120 #ifndef O_DSYNC
121 #ifdef O_SYNC
122 #define O_DSYNC O_SYNC
123 #elif defined(O_FSYNC)
124 #define O_DSYNC O_FSYNC
125 #endif
126 #endif
127 
128 /* Approximate O_DIRECT with O_DSYNC if O_DIRECT isn't available */
129 #ifndef O_DIRECT
130 #define O_DIRECT O_DSYNC
131 #endif
132 
133 #define FTYPE_FILE   0
134 #define FTYPE_CD     1
135 
136 #define MAX_BLOCKSIZE	4096
137 
138 /* Posix file locking bytes. Libvirt takes byte 0, we start from higher bytes,
139  * leaving a few more bytes for its future use. */
140 #define RAW_LOCK_PERM_BASE             100
141 #define RAW_LOCK_SHARED_BASE           200
142 
143 typedef struct BDRVRawState {
144     int fd;
145     int lock_fd;
146     bool use_lock;
147     int type;
148     int open_flags;
149     size_t buf_align;
150 
151     /* The current permissions. */
152     uint64_t perm;
153     uint64_t shared_perm;
154 
155 #ifdef CONFIG_XFS
156     bool is_xfs:1;
157 #endif
158     bool has_discard:1;
159     bool has_write_zeroes:1;
160     bool discard_zeroes:1;
161     bool use_linux_aio:1;
162     bool page_cache_inconsistent:1;
163     bool has_fallocate;
164     bool needs_alignment;
165     bool check_cache_dropped;
166 
167     PRManager *pr_mgr;
168 } BDRVRawState;
169 
170 typedef struct BDRVRawReopenState {
171     int fd;
172     int open_flags;
173     bool check_cache_dropped;
174 } BDRVRawReopenState;
175 
176 static int fd_open(BlockDriverState *bs);
177 static int64_t raw_getlength(BlockDriverState *bs);
178 
179 typedef struct RawPosixAIOData {
180     BlockDriverState *bs;
181     int aio_fildes;
182     union {
183         struct iovec *aio_iov;
184         void *aio_ioctl_buf;
185     };
186     int aio_niov;
187     uint64_t aio_nbytes;
188 #define aio_ioctl_cmd   aio_nbytes /* for QEMU_AIO_IOCTL */
189     off_t aio_offset;
190     int aio_type;
191     int aio_fd2;
192     off_t aio_offset2;
193 } RawPosixAIOData;
194 
195 #if defined(__FreeBSD__) || defined(__FreeBSD_kernel__)
196 static int cdrom_reopen(BlockDriverState *bs);
197 #endif
198 
199 #if defined(__NetBSD__)
200 static int raw_normalize_devicepath(const char **filename)
201 {
202     static char namebuf[PATH_MAX];
203     const char *dp, *fname;
204     struct stat sb;
205 
206     fname = *filename;
207     dp = strrchr(fname, '/');
208     if (lstat(fname, &sb) < 0) {
209         fprintf(stderr, "%s: stat failed: %s\n",
210             fname, strerror(errno));
211         return -errno;
212     }
213 
214     if (!S_ISBLK(sb.st_mode)) {
215         return 0;
216     }
217 
218     if (dp == NULL) {
219         snprintf(namebuf, PATH_MAX, "r%s", fname);
220     } else {
221         snprintf(namebuf, PATH_MAX, "%.*s/r%s",
222             (int)(dp - fname), fname, dp + 1);
223     }
224     fprintf(stderr, "%s is a block device", fname);
225     *filename = namebuf;
226     fprintf(stderr, ", using %s\n", *filename);
227 
228     return 0;
229 }
230 #else
231 static int raw_normalize_devicepath(const char **filename)
232 {
233     return 0;
234 }
235 #endif
236 
237 /*
238  * Get logical block size via ioctl. On success store it in @sector_size_p.
239  */
240 static int probe_logical_blocksize(int fd, unsigned int *sector_size_p)
241 {
242     unsigned int sector_size;
243     bool success = false;
244     int i;
245 
246     errno = ENOTSUP;
247     static const unsigned long ioctl_list[] = {
248 #ifdef BLKSSZGET
249         BLKSSZGET,
250 #endif
251 #ifdef DKIOCGETBLOCKSIZE
252         DKIOCGETBLOCKSIZE,
253 #endif
254 #ifdef DIOCGSECTORSIZE
255         DIOCGSECTORSIZE,
256 #endif
257     };
258 
259     /* Try a few ioctls to get the right size */
260     for (i = 0; i < (int)ARRAY_SIZE(ioctl_list); i++) {
261         if (ioctl(fd, ioctl_list[i], &sector_size) >= 0) {
262             *sector_size_p = sector_size;
263             success = true;
264         }
265     }
266 
267     return success ? 0 : -errno;
268 }
269 
270 /**
271  * Get physical block size of @fd.
272  * On success, store it in @blk_size and return 0.
273  * On failure, return -errno.
274  */
275 static int probe_physical_blocksize(int fd, unsigned int *blk_size)
276 {
277 #ifdef BLKPBSZGET
278     if (ioctl(fd, BLKPBSZGET, blk_size) < 0) {
279         return -errno;
280     }
281     return 0;
282 #else
283     return -ENOTSUP;
284 #endif
285 }
286 
287 /* Check if read is allowed with given memory buffer and length.
288  *
289  * This function is used to check O_DIRECT memory buffer and request alignment.
290  */
291 static bool raw_is_io_aligned(int fd, void *buf, size_t len)
292 {
293     ssize_t ret = pread(fd, buf, len, 0);
294 
295     if (ret >= 0) {
296         return true;
297     }
298 
299 #ifdef __linux__
300     /* The Linux kernel returns EINVAL for misaligned O_DIRECT reads.  Ignore
301      * other errors (e.g. real I/O error), which could happen on a failed
302      * drive, since we only care about probing alignment.
303      */
304     if (errno != EINVAL) {
305         return true;
306     }
307 #endif
308 
309     return false;
310 }
311 
312 static void raw_probe_alignment(BlockDriverState *bs, int fd, Error **errp)
313 {
314     BDRVRawState *s = bs->opaque;
315     char *buf;
316     size_t max_align = MAX(MAX_BLOCKSIZE, getpagesize());
317 
318     /* For SCSI generic devices the alignment is not really used.
319        With buffered I/O, we don't have any restrictions. */
320     if (bdrv_is_sg(bs) || !s->needs_alignment) {
321         bs->bl.request_alignment = 1;
322         s->buf_align = 1;
323         return;
324     }
325 
326     bs->bl.request_alignment = 0;
327     s->buf_align = 0;
328     /* Let's try to use the logical blocksize for the alignment. */
329     if (probe_logical_blocksize(fd, &bs->bl.request_alignment) < 0) {
330         bs->bl.request_alignment = 0;
331     }
332 #ifdef CONFIG_XFS
333     if (s->is_xfs) {
334         struct dioattr da;
335         if (xfsctl(NULL, fd, XFS_IOC_DIOINFO, &da) >= 0) {
336             bs->bl.request_alignment = da.d_miniosz;
337             /* The kernel returns wrong information for d_mem */
338             /* s->buf_align = da.d_mem; */
339         }
340     }
341 #endif
342 
343     /* If we could not get the sizes so far, we can only guess them */
344     if (!s->buf_align) {
345         size_t align;
346         buf = qemu_memalign(max_align, 2 * max_align);
347         for (align = 512; align <= max_align; align <<= 1) {
348             if (raw_is_io_aligned(fd, buf + align, max_align)) {
349                 s->buf_align = align;
350                 break;
351             }
352         }
353         qemu_vfree(buf);
354     }
355 
356     if (!bs->bl.request_alignment) {
357         size_t align;
358         buf = qemu_memalign(s->buf_align, max_align);
359         for (align = 512; align <= max_align; align <<= 1) {
360             if (raw_is_io_aligned(fd, buf, align)) {
361                 bs->bl.request_alignment = align;
362                 break;
363             }
364         }
365         qemu_vfree(buf);
366     }
367 
368     if (!s->buf_align || !bs->bl.request_alignment) {
369         error_setg(errp, "Could not find working O_DIRECT alignment");
370         error_append_hint(errp, "Try cache.direct=off\n");
371     }
372 }
373 
374 static void raw_parse_flags(int bdrv_flags, int *open_flags)
375 {
376     assert(open_flags != NULL);
377 
378     *open_flags |= O_BINARY;
379     *open_flags &= ~O_ACCMODE;
380     if (bdrv_flags & BDRV_O_RDWR) {
381         *open_flags |= O_RDWR;
382     } else {
383         *open_flags |= O_RDONLY;
384     }
385 
386     /* Use O_DSYNC for write-through caching, no flags for write-back caching,
387      * and O_DIRECT for no caching. */
388     if ((bdrv_flags & BDRV_O_NOCACHE)) {
389         *open_flags |= O_DIRECT;
390     }
391 }
392 
393 static void raw_parse_filename(const char *filename, QDict *options,
394                                Error **errp)
395 {
396     bdrv_parse_filename_strip_prefix(filename, "file:", options);
397 }
398 
399 static QemuOptsList raw_runtime_opts = {
400     .name = "raw",
401     .head = QTAILQ_HEAD_INITIALIZER(raw_runtime_opts.head),
402     .desc = {
403         {
404             .name = "filename",
405             .type = QEMU_OPT_STRING,
406             .help = "File name of the image",
407         },
408         {
409             .name = "aio",
410             .type = QEMU_OPT_STRING,
411             .help = "host AIO implementation (threads, native)",
412         },
413         {
414             .name = "locking",
415             .type = QEMU_OPT_STRING,
416             .help = "file locking mode (on/off/auto, default: auto)",
417         },
418         {
419             .name = "pr-manager",
420             .type = QEMU_OPT_STRING,
421             .help = "id of persistent reservation manager object (default: none)",
422         },
423         {
424             .name = "x-check-cache-dropped",
425             .type = QEMU_OPT_BOOL,
426             .help = "check that page cache was dropped on live migration (default: off)"
427         },
428         { /* end of list */ }
429     },
430 };
431 
432 static int raw_open_common(BlockDriverState *bs, QDict *options,
433                            int bdrv_flags, int open_flags, Error **errp)
434 {
435     BDRVRawState *s = bs->opaque;
436     QemuOpts *opts;
437     Error *local_err = NULL;
438     const char *filename = NULL;
439     const char *str;
440     BlockdevAioOptions aio, aio_default;
441     int fd, ret;
442     struct stat st;
443     OnOffAuto locking;
444 
445     opts = qemu_opts_create(&raw_runtime_opts, NULL, 0, &error_abort);
446     qemu_opts_absorb_qdict(opts, options, &local_err);
447     if (local_err) {
448         error_propagate(errp, local_err);
449         ret = -EINVAL;
450         goto fail;
451     }
452 
453     filename = qemu_opt_get(opts, "filename");
454 
455     ret = raw_normalize_devicepath(&filename);
456     if (ret != 0) {
457         error_setg_errno(errp, -ret, "Could not normalize device path");
458         goto fail;
459     }
460 
461     aio_default = (bdrv_flags & BDRV_O_NATIVE_AIO)
462                   ? BLOCKDEV_AIO_OPTIONS_NATIVE
463                   : BLOCKDEV_AIO_OPTIONS_THREADS;
464     aio = qapi_enum_parse(&BlockdevAioOptions_lookup,
465                           qemu_opt_get(opts, "aio"),
466                           aio_default, &local_err);
467     if (local_err) {
468         error_propagate(errp, local_err);
469         ret = -EINVAL;
470         goto fail;
471     }
472     s->use_linux_aio = (aio == BLOCKDEV_AIO_OPTIONS_NATIVE);
473 
474     locking = qapi_enum_parse(&OnOffAuto_lookup,
475                               qemu_opt_get(opts, "locking"),
476                               ON_OFF_AUTO_AUTO, &local_err);
477     if (local_err) {
478         error_propagate(errp, local_err);
479         ret = -EINVAL;
480         goto fail;
481     }
482     switch (locking) {
483     case ON_OFF_AUTO_ON:
484         s->use_lock = true;
485         if (!qemu_has_ofd_lock()) {
486             fprintf(stderr,
487                     "File lock requested but OFD locking syscall is "
488                     "unavailable, falling back to POSIX file locks.\n"
489                     "Due to the implementation, locks can be lost "
490                     "unexpectedly.\n");
491         }
492         break;
493     case ON_OFF_AUTO_OFF:
494         s->use_lock = false;
495         break;
496     case ON_OFF_AUTO_AUTO:
497         s->use_lock = qemu_has_ofd_lock();
498         break;
499     default:
500         abort();
501     }
502 
503     str = qemu_opt_get(opts, "pr-manager");
504     if (str) {
505         s->pr_mgr = pr_manager_lookup(str, &local_err);
506         if (local_err) {
507             error_propagate(errp, local_err);
508             ret = -EINVAL;
509             goto fail;
510         }
511     }
512 
513     s->check_cache_dropped = qemu_opt_get_bool(opts, "x-check-cache-dropped",
514                                                false);
515 
516     s->open_flags = open_flags;
517     raw_parse_flags(bdrv_flags, &s->open_flags);
518 
519     s->fd = -1;
520     fd = qemu_open(filename, s->open_flags, 0644);
521     if (fd < 0) {
522         ret = -errno;
523         error_setg_errno(errp, errno, "Could not open '%s'", filename);
524         if (ret == -EROFS) {
525             ret = -EACCES;
526         }
527         goto fail;
528     }
529     s->fd = fd;
530 
531     s->lock_fd = -1;
532     if (s->use_lock) {
533         fd = qemu_open(filename, s->open_flags);
534         if (fd < 0) {
535             ret = -errno;
536             error_setg_errno(errp, errno, "Could not open '%s' for locking",
537                              filename);
538             qemu_close(s->fd);
539             goto fail;
540         }
541         s->lock_fd = fd;
542     }
543     s->perm = 0;
544     s->shared_perm = BLK_PERM_ALL;
545 
546 #ifdef CONFIG_LINUX_AIO
547      /* Currently Linux does AIO only for files opened with O_DIRECT */
548     if (s->use_linux_aio && !(s->open_flags & O_DIRECT)) {
549         error_setg(errp, "aio=native was specified, but it requires "
550                          "cache.direct=on, which was not specified.");
551         ret = -EINVAL;
552         goto fail;
553     }
554 #else
555     if (s->use_linux_aio) {
556         error_setg(errp, "aio=native was specified, but is not supported "
557                          "in this build.");
558         ret = -EINVAL;
559         goto fail;
560     }
561 #endif /* !defined(CONFIG_LINUX_AIO) */
562 
563     s->has_discard = true;
564     s->has_write_zeroes = true;
565     if ((bs->open_flags & BDRV_O_NOCACHE) != 0) {
566         s->needs_alignment = true;
567     }
568 
569     if (fstat(s->fd, &st) < 0) {
570         ret = -errno;
571         error_setg_errno(errp, errno, "Could not stat file");
572         goto fail;
573     }
574     if (S_ISREG(st.st_mode)) {
575         s->discard_zeroes = true;
576         s->has_fallocate = true;
577     }
578     if (S_ISBLK(st.st_mode)) {
579 #ifdef BLKDISCARDZEROES
580         unsigned int arg;
581         if (ioctl(s->fd, BLKDISCARDZEROES, &arg) == 0 && arg) {
582             s->discard_zeroes = true;
583         }
584 #endif
585 #ifdef __linux__
586         /* On Linux 3.10, BLKDISCARD leaves stale data in the page cache.  Do
587          * not rely on the contents of discarded blocks unless using O_DIRECT.
588          * Same for BLKZEROOUT.
589          */
590         if (!(bs->open_flags & BDRV_O_NOCACHE)) {
591             s->discard_zeroes = false;
592             s->has_write_zeroes = false;
593         }
594 #endif
595     }
596 #ifdef __FreeBSD__
597     if (S_ISCHR(st.st_mode)) {
598         /*
599          * The file is a char device (disk), which on FreeBSD isn't behind
600          * a pager, so force all requests to be aligned. This is needed
601          * so QEMU makes sure all IO operations on the device are aligned
602          * to sector size, or else FreeBSD will reject them with EINVAL.
603          */
604         s->needs_alignment = true;
605     }
606 #endif
607 
608 #ifdef CONFIG_XFS
609     if (platform_test_xfs_fd(s->fd)) {
610         s->is_xfs = true;
611     }
612 #endif
613 
614     bs->supported_zero_flags = s->discard_zeroes ? BDRV_REQ_MAY_UNMAP : 0;
615     ret = 0;
616 fail:
617     if (filename && (bdrv_flags & BDRV_O_TEMPORARY)) {
618         unlink(filename);
619     }
620     qemu_opts_del(opts);
621     return ret;
622 }
623 
624 static int raw_open(BlockDriverState *bs, QDict *options, int flags,
625                     Error **errp)
626 {
627     BDRVRawState *s = bs->opaque;
628 
629     s->type = FTYPE_FILE;
630     return raw_open_common(bs, options, flags, 0, errp);
631 }
632 
633 typedef enum {
634     RAW_PL_PREPARE,
635     RAW_PL_COMMIT,
636     RAW_PL_ABORT,
637 } RawPermLockOp;
638 
639 #define PERM_FOREACH(i) \
640     for ((i) = 0; (1ULL << (i)) <= BLK_PERM_ALL; i++)
641 
642 /* Lock bytes indicated by @perm_lock_bits and @shared_perm_lock_bits in the
643  * file; if @unlock == true, also unlock the unneeded bytes.
644  * @shared_perm_lock_bits is the mask of all permissions that are NOT shared.
645  */
646 static int raw_apply_lock_bytes(int fd,
647                                 uint64_t perm_lock_bits,
648                                 uint64_t shared_perm_lock_bits,
649                                 bool unlock, Error **errp)
650 {
651     int ret;
652     int i;
653 
654     PERM_FOREACH(i) {
655         int off = RAW_LOCK_PERM_BASE + i;
656         if (perm_lock_bits & (1ULL << i)) {
657             ret = qemu_lock_fd(fd, off, 1, false);
658             if (ret) {
659                 error_setg(errp, "Failed to lock byte %d", off);
660                 return ret;
661             }
662         } else if (unlock) {
663             ret = qemu_unlock_fd(fd, off, 1);
664             if (ret) {
665                 error_setg(errp, "Failed to unlock byte %d", off);
666                 return ret;
667             }
668         }
669     }
670     PERM_FOREACH(i) {
671         int off = RAW_LOCK_SHARED_BASE + i;
672         if (shared_perm_lock_bits & (1ULL << i)) {
673             ret = qemu_lock_fd(fd, off, 1, false);
674             if (ret) {
675                 error_setg(errp, "Failed to lock byte %d", off);
676                 return ret;
677             }
678         } else if (unlock) {
679             ret = qemu_unlock_fd(fd, off, 1);
680             if (ret) {
681                 error_setg(errp, "Failed to unlock byte %d", off);
682                 return ret;
683             }
684         }
685     }
686     return 0;
687 }
688 
689 /* Check "unshared" bytes implied by @perm and ~@shared_perm in the file. */
690 static int raw_check_lock_bytes(int fd, uint64_t perm, uint64_t shared_perm,
691                                 Error **errp)
692 {
693     int ret;
694     int i;
695 
696     PERM_FOREACH(i) {
697         int off = RAW_LOCK_SHARED_BASE + i;
698         uint64_t p = 1ULL << i;
699         if (perm & p) {
700             ret = qemu_lock_fd_test(fd, off, 1, true);
701             if (ret) {
702                 char *perm_name = bdrv_perm_names(p);
703                 error_setg(errp,
704                            "Failed to get \"%s\" lock",
705                            perm_name);
706                 g_free(perm_name);
707                 error_append_hint(errp,
708                                   "Is another process using the image?\n");
709                 return ret;
710             }
711         }
712     }
713     PERM_FOREACH(i) {
714         int off = RAW_LOCK_PERM_BASE + i;
715         uint64_t p = 1ULL << i;
716         if (!(shared_perm & p)) {
717             ret = qemu_lock_fd_test(fd, off, 1, true);
718             if (ret) {
719                 char *perm_name = bdrv_perm_names(p);
720                 error_setg(errp,
721                            "Failed to get shared \"%s\" lock",
722                            perm_name);
723                 g_free(perm_name);
724                 error_append_hint(errp,
725                                   "Is another process using the image?\n");
726                 return ret;
727             }
728         }
729     }
730     return 0;
731 }
732 
733 static int raw_handle_perm_lock(BlockDriverState *bs,
734                                 RawPermLockOp op,
735                                 uint64_t new_perm, uint64_t new_shared,
736                                 Error **errp)
737 {
738     BDRVRawState *s = bs->opaque;
739     int ret = 0;
740     Error *local_err = NULL;
741 
742     if (!s->use_lock) {
743         return 0;
744     }
745 
746     if (bdrv_get_flags(bs) & BDRV_O_INACTIVE) {
747         return 0;
748     }
749 
750     assert(s->lock_fd > 0);
751 
752     switch (op) {
753     case RAW_PL_PREPARE:
754         ret = raw_apply_lock_bytes(s->lock_fd, s->perm | new_perm,
755                                    ~s->shared_perm | ~new_shared,
756                                    false, errp);
757         if (!ret) {
758             ret = raw_check_lock_bytes(s->lock_fd, new_perm, new_shared, errp);
759             if (!ret) {
760                 return 0;
761             }
762         }
763         op = RAW_PL_ABORT;
764         /* fall through to unlock bytes. */
765     case RAW_PL_ABORT:
766         raw_apply_lock_bytes(s->lock_fd, s->perm, ~s->shared_perm,
767                              true, &local_err);
768         if (local_err) {
769             /* Theoretically the above call only unlocks bytes and it cannot
770              * fail. Something weird happened, report it.
771              */
772             error_report_err(local_err);
773         }
774         break;
775     case RAW_PL_COMMIT:
776         raw_apply_lock_bytes(s->lock_fd, new_perm, ~new_shared,
777                              true, &local_err);
778         if (local_err) {
779             /* Theoretically the above call only unlocks bytes and it cannot
780              * fail. Something weird happened, report it.
781              */
782             error_report_err(local_err);
783         }
784         break;
785     }
786     return ret;
787 }
788 
789 static int raw_reopen_prepare(BDRVReopenState *state,
790                               BlockReopenQueue *queue, Error **errp)
791 {
792     BDRVRawState *s;
793     BDRVRawReopenState *rs;
794     QemuOpts *opts;
795     int ret = 0;
796     Error *local_err = NULL;
797 
798     assert(state != NULL);
799     assert(state->bs != NULL);
800 
801     s = state->bs->opaque;
802 
803     state->opaque = g_new0(BDRVRawReopenState, 1);
804     rs = state->opaque;
805     rs->fd = -1;
806 
807     /* Handle options changes */
808     opts = qemu_opts_create(&raw_runtime_opts, NULL, 0, &error_abort);
809     qemu_opts_absorb_qdict(opts, state->options, &local_err);
810     if (local_err) {
811         error_propagate(errp, local_err);
812         ret = -EINVAL;
813         goto out;
814     }
815 
816     rs->check_cache_dropped = qemu_opt_get_bool(opts, "x-check-cache-dropped",
817                                                 s->check_cache_dropped);
818 
819     if (s->type == FTYPE_CD) {
820         rs->open_flags |= O_NONBLOCK;
821     }
822 
823     raw_parse_flags(state->flags, &rs->open_flags);
824 
825     int fcntl_flags = O_APPEND | O_NONBLOCK;
826 #ifdef O_NOATIME
827     fcntl_flags |= O_NOATIME;
828 #endif
829 
830 #ifdef O_ASYNC
831     /* Not all operating systems have O_ASYNC, and those that don't
832      * will not let us track the state into rs->open_flags (typically
833      * you achieve the same effect with an ioctl, for example I_SETSIG
834      * on Solaris). But we do not use O_ASYNC, so that's fine.
835      */
836     assert((s->open_flags & O_ASYNC) == 0);
837 #endif
838 
839     if ((rs->open_flags & ~fcntl_flags) == (s->open_flags & ~fcntl_flags)) {
840         /* dup the original fd */
841         rs->fd = qemu_dup(s->fd);
842         if (rs->fd >= 0) {
843             ret = fcntl_setfl(rs->fd, rs->open_flags);
844             if (ret) {
845                 qemu_close(rs->fd);
846                 rs->fd = -1;
847             }
848         }
849     }
850 
851     /* If we cannot use fcntl, or fcntl failed, fall back to qemu_open() */
852     if (rs->fd == -1) {
853         const char *normalized_filename = state->bs->filename;
854         ret = raw_normalize_devicepath(&normalized_filename);
855         if (ret < 0) {
856             error_setg_errno(errp, -ret, "Could not normalize device path");
857         } else {
858             assert(!(rs->open_flags & O_CREAT));
859             rs->fd = qemu_open(normalized_filename, rs->open_flags);
860             if (rs->fd == -1) {
861                 error_setg_errno(errp, errno, "Could not reopen file");
862                 ret = -1;
863             }
864         }
865     }
866 
867     /* Fail already reopen_prepare() if we can't get a working O_DIRECT
868      * alignment with the new fd. */
869     if (rs->fd != -1) {
870         raw_probe_alignment(state->bs, rs->fd, &local_err);
871         if (local_err) {
872             qemu_close(rs->fd);
873             rs->fd = -1;
874             error_propagate(errp, local_err);
875             ret = -EINVAL;
876         }
877     }
878 
879 out:
880     qemu_opts_del(opts);
881     return ret;
882 }
883 
884 static void raw_reopen_commit(BDRVReopenState *state)
885 {
886     BDRVRawReopenState *rs = state->opaque;
887     BDRVRawState *s = state->bs->opaque;
888 
889     s->check_cache_dropped = rs->check_cache_dropped;
890     s->open_flags = rs->open_flags;
891 
892     qemu_close(s->fd);
893     s->fd = rs->fd;
894 
895     g_free(state->opaque);
896     state->opaque = NULL;
897 }
898 
899 
900 static void raw_reopen_abort(BDRVReopenState *state)
901 {
902     BDRVRawReopenState *rs = state->opaque;
903 
904      /* nothing to do if NULL, we didn't get far enough */
905     if (rs == NULL) {
906         return;
907     }
908 
909     if (rs->fd >= 0) {
910         qemu_close(rs->fd);
911         rs->fd = -1;
912     }
913     g_free(state->opaque);
914     state->opaque = NULL;
915 }
916 
917 static int hdev_get_max_transfer_length(BlockDriverState *bs, int fd)
918 {
919 #ifdef BLKSECTGET
920     int max_bytes = 0;
921     short max_sectors = 0;
922     if (bs->sg && ioctl(fd, BLKSECTGET, &max_bytes) == 0) {
923         return max_bytes;
924     } else if (!bs->sg && ioctl(fd, BLKSECTGET, &max_sectors) == 0) {
925         return max_sectors << BDRV_SECTOR_BITS;
926     } else {
927         return -errno;
928     }
929 #else
930     return -ENOSYS;
931 #endif
932 }
933 
934 static int hdev_get_max_segments(const struct stat *st)
935 {
936 #ifdef CONFIG_LINUX
937     char buf[32];
938     const char *end;
939     char *sysfspath;
940     int ret;
941     int fd = -1;
942     long max_segments;
943 
944     sysfspath = g_strdup_printf("/sys/dev/block/%u:%u/queue/max_segments",
945                                 major(st->st_rdev), minor(st->st_rdev));
946     fd = open(sysfspath, O_RDONLY);
947     if (fd == -1) {
948         ret = -errno;
949         goto out;
950     }
951     do {
952         ret = read(fd, buf, sizeof(buf) - 1);
953     } while (ret == -1 && errno == EINTR);
954     if (ret < 0) {
955         ret = -errno;
956         goto out;
957     } else if (ret == 0) {
958         ret = -EIO;
959         goto out;
960     }
961     buf[ret] = 0;
962     /* The file is ended with '\n', pass 'end' to accept that. */
963     ret = qemu_strtol(buf, &end, 10, &max_segments);
964     if (ret == 0 && end && *end == '\n') {
965         ret = max_segments;
966     }
967 
968 out:
969     if (fd != -1) {
970         close(fd);
971     }
972     g_free(sysfspath);
973     return ret;
974 #else
975     return -ENOTSUP;
976 #endif
977 }
978 
979 static void raw_refresh_limits(BlockDriverState *bs, Error **errp)
980 {
981     BDRVRawState *s = bs->opaque;
982     struct stat st;
983 
984     if (!fstat(s->fd, &st)) {
985         if (S_ISBLK(st.st_mode) || S_ISCHR(st.st_mode)) {
986             int ret = hdev_get_max_transfer_length(bs, s->fd);
987             if (ret > 0 && ret <= BDRV_REQUEST_MAX_BYTES) {
988                 bs->bl.max_transfer = pow2floor(ret);
989             }
990             ret = hdev_get_max_segments(&st);
991             if (ret > 0) {
992                 bs->bl.max_transfer = MIN(bs->bl.max_transfer,
993                                           ret * getpagesize());
994             }
995         }
996     }
997 
998     raw_probe_alignment(bs, s->fd, errp);
999     bs->bl.min_mem_alignment = s->buf_align;
1000     bs->bl.opt_mem_alignment = MAX(s->buf_align, getpagesize());
1001 }
1002 
1003 static int check_for_dasd(int fd)
1004 {
1005 #ifdef BIODASDINFO2
1006     struct dasd_information2_t info = {0};
1007 
1008     return ioctl(fd, BIODASDINFO2, &info);
1009 #else
1010     return -1;
1011 #endif
1012 }
1013 
1014 /**
1015  * Try to get @bs's logical and physical block size.
1016  * On success, store them in @bsz and return zero.
1017  * On failure, return negative errno.
1018  */
1019 static int hdev_probe_blocksizes(BlockDriverState *bs, BlockSizes *bsz)
1020 {
1021     BDRVRawState *s = bs->opaque;
1022     int ret;
1023 
1024     /* If DASD, get blocksizes */
1025     if (check_for_dasd(s->fd) < 0) {
1026         return -ENOTSUP;
1027     }
1028     ret = probe_logical_blocksize(s->fd, &bsz->log);
1029     if (ret < 0) {
1030         return ret;
1031     }
1032     return probe_physical_blocksize(s->fd, &bsz->phys);
1033 }
1034 
1035 /**
1036  * Try to get @bs's geometry: cyls, heads, sectors.
1037  * On success, store them in @geo and return 0.
1038  * On failure return -errno.
1039  * (Allows block driver to assign default geometry values that guest sees)
1040  */
1041 #ifdef __linux__
1042 static int hdev_probe_geometry(BlockDriverState *bs, HDGeometry *geo)
1043 {
1044     BDRVRawState *s = bs->opaque;
1045     struct hd_geometry ioctl_geo = {0};
1046 
1047     /* If DASD, get its geometry */
1048     if (check_for_dasd(s->fd) < 0) {
1049         return -ENOTSUP;
1050     }
1051     if (ioctl(s->fd, HDIO_GETGEO, &ioctl_geo) < 0) {
1052         return -errno;
1053     }
1054     /* HDIO_GETGEO may return success even though geo contains zeros
1055        (e.g. certain multipath setups) */
1056     if (!ioctl_geo.heads || !ioctl_geo.sectors || !ioctl_geo.cylinders) {
1057         return -ENOTSUP;
1058     }
1059     /* Do not return a geometry for partition */
1060     if (ioctl_geo.start != 0) {
1061         return -ENOTSUP;
1062     }
1063     geo->heads = ioctl_geo.heads;
1064     geo->sectors = ioctl_geo.sectors;
1065     geo->cylinders = ioctl_geo.cylinders;
1066 
1067     return 0;
1068 }
1069 #else /* __linux__ */
1070 static int hdev_probe_geometry(BlockDriverState *bs, HDGeometry *geo)
1071 {
1072     return -ENOTSUP;
1073 }
1074 #endif
1075 
1076 static ssize_t handle_aiocb_ioctl(RawPosixAIOData *aiocb)
1077 {
1078     int ret;
1079 
1080     ret = ioctl(aiocb->aio_fildes, aiocb->aio_ioctl_cmd, aiocb->aio_ioctl_buf);
1081     if (ret == -1) {
1082         return -errno;
1083     }
1084 
1085     return 0;
1086 }
1087 
1088 static ssize_t handle_aiocb_flush(RawPosixAIOData *aiocb)
1089 {
1090     BDRVRawState *s = aiocb->bs->opaque;
1091     int ret;
1092 
1093     if (s->page_cache_inconsistent) {
1094         return -EIO;
1095     }
1096 
1097     ret = qemu_fdatasync(aiocb->aio_fildes);
1098     if (ret == -1) {
1099         /* There is no clear definition of the semantics of a failing fsync(),
1100          * so we may have to assume the worst. The sad truth is that this
1101          * assumption is correct for Linux. Some pages are now probably marked
1102          * clean in the page cache even though they are inconsistent with the
1103          * on-disk contents. The next fdatasync() call would succeed, but no
1104          * further writeback attempt will be made. We can't get back to a state
1105          * in which we know what is on disk (we would have to rewrite
1106          * everything that was touched since the last fdatasync() at least), so
1107          * make bdrv_flush() fail permanently. Given that the behaviour isn't
1108          * really defined, I have little hope that other OSes are doing better.
1109          *
1110          * Obviously, this doesn't affect O_DIRECT, which bypasses the page
1111          * cache. */
1112         if ((s->open_flags & O_DIRECT) == 0) {
1113             s->page_cache_inconsistent = true;
1114         }
1115         return -errno;
1116     }
1117     return 0;
1118 }
1119 
1120 #ifdef CONFIG_PREADV
1121 
1122 static bool preadv_present = true;
1123 
1124 static ssize_t
1125 qemu_preadv(int fd, const struct iovec *iov, int nr_iov, off_t offset)
1126 {
1127     return preadv(fd, iov, nr_iov, offset);
1128 }
1129 
1130 static ssize_t
1131 qemu_pwritev(int fd, const struct iovec *iov, int nr_iov, off_t offset)
1132 {
1133     return pwritev(fd, iov, nr_iov, offset);
1134 }
1135 
1136 #else
1137 
1138 static bool preadv_present = false;
1139 
1140 static ssize_t
1141 qemu_preadv(int fd, const struct iovec *iov, int nr_iov, off_t offset)
1142 {
1143     return -ENOSYS;
1144 }
1145 
1146 static ssize_t
1147 qemu_pwritev(int fd, const struct iovec *iov, int nr_iov, off_t offset)
1148 {
1149     return -ENOSYS;
1150 }
1151 
1152 #endif
1153 
1154 static ssize_t handle_aiocb_rw_vector(RawPosixAIOData *aiocb)
1155 {
1156     ssize_t len;
1157 
1158     do {
1159         if (aiocb->aio_type & QEMU_AIO_WRITE)
1160             len = qemu_pwritev(aiocb->aio_fildes,
1161                                aiocb->aio_iov,
1162                                aiocb->aio_niov,
1163                                aiocb->aio_offset);
1164          else
1165             len = qemu_preadv(aiocb->aio_fildes,
1166                               aiocb->aio_iov,
1167                               aiocb->aio_niov,
1168                               aiocb->aio_offset);
1169     } while (len == -1 && errno == EINTR);
1170 
1171     if (len == -1) {
1172         return -errno;
1173     }
1174     return len;
1175 }
1176 
1177 /*
1178  * Read/writes the data to/from a given linear buffer.
1179  *
1180  * Returns the number of bytes handles or -errno in case of an error. Short
1181  * reads are only returned if the end of the file is reached.
1182  */
1183 static ssize_t handle_aiocb_rw_linear(RawPosixAIOData *aiocb, char *buf)
1184 {
1185     ssize_t offset = 0;
1186     ssize_t len;
1187 
1188     while (offset < aiocb->aio_nbytes) {
1189         if (aiocb->aio_type & QEMU_AIO_WRITE) {
1190             len = pwrite(aiocb->aio_fildes,
1191                          (const char *)buf + offset,
1192                          aiocb->aio_nbytes - offset,
1193                          aiocb->aio_offset + offset);
1194         } else {
1195             len = pread(aiocb->aio_fildes,
1196                         buf + offset,
1197                         aiocb->aio_nbytes - offset,
1198                         aiocb->aio_offset + offset);
1199         }
1200         if (len == -1 && errno == EINTR) {
1201             continue;
1202         } else if (len == -1 && errno == EINVAL &&
1203                    (aiocb->bs->open_flags & BDRV_O_NOCACHE) &&
1204                    !(aiocb->aio_type & QEMU_AIO_WRITE) &&
1205                    offset > 0) {
1206             /* O_DIRECT pread() may fail with EINVAL when offset is unaligned
1207              * after a short read.  Assume that O_DIRECT short reads only occur
1208              * at EOF.  Therefore this is a short read, not an I/O error.
1209              */
1210             break;
1211         } else if (len == -1) {
1212             offset = -errno;
1213             break;
1214         } else if (len == 0) {
1215             break;
1216         }
1217         offset += len;
1218     }
1219 
1220     return offset;
1221 }
1222 
1223 static ssize_t handle_aiocb_rw(RawPosixAIOData *aiocb)
1224 {
1225     ssize_t nbytes;
1226     char *buf;
1227 
1228     if (!(aiocb->aio_type & QEMU_AIO_MISALIGNED)) {
1229         /*
1230          * If there is just a single buffer, and it is properly aligned
1231          * we can just use plain pread/pwrite without any problems.
1232          */
1233         if (aiocb->aio_niov == 1) {
1234              return handle_aiocb_rw_linear(aiocb, aiocb->aio_iov->iov_base);
1235         }
1236         /*
1237          * We have more than one iovec, and all are properly aligned.
1238          *
1239          * Try preadv/pwritev first and fall back to linearizing the
1240          * buffer if it's not supported.
1241          */
1242         if (preadv_present) {
1243             nbytes = handle_aiocb_rw_vector(aiocb);
1244             if (nbytes == aiocb->aio_nbytes ||
1245                 (nbytes < 0 && nbytes != -ENOSYS)) {
1246                 return nbytes;
1247             }
1248             preadv_present = false;
1249         }
1250 
1251         /*
1252          * XXX(hch): short read/write.  no easy way to handle the reminder
1253          * using these interfaces.  For now retry using plain
1254          * pread/pwrite?
1255          */
1256     }
1257 
1258     /*
1259      * Ok, we have to do it the hard way, copy all segments into
1260      * a single aligned buffer.
1261      */
1262     buf = qemu_try_blockalign(aiocb->bs, aiocb->aio_nbytes);
1263     if (buf == NULL) {
1264         return -ENOMEM;
1265     }
1266 
1267     if (aiocb->aio_type & QEMU_AIO_WRITE) {
1268         char *p = buf;
1269         int i;
1270 
1271         for (i = 0; i < aiocb->aio_niov; ++i) {
1272             memcpy(p, aiocb->aio_iov[i].iov_base, aiocb->aio_iov[i].iov_len);
1273             p += aiocb->aio_iov[i].iov_len;
1274         }
1275         assert(p - buf == aiocb->aio_nbytes);
1276     }
1277 
1278     nbytes = handle_aiocb_rw_linear(aiocb, buf);
1279     if (!(aiocb->aio_type & QEMU_AIO_WRITE)) {
1280         char *p = buf;
1281         size_t count = aiocb->aio_nbytes, copy;
1282         int i;
1283 
1284         for (i = 0; i < aiocb->aio_niov && count; ++i) {
1285             copy = count;
1286             if (copy > aiocb->aio_iov[i].iov_len) {
1287                 copy = aiocb->aio_iov[i].iov_len;
1288             }
1289             memcpy(aiocb->aio_iov[i].iov_base, p, copy);
1290             assert(count >= copy);
1291             p     += copy;
1292             count -= copy;
1293         }
1294         assert(count == 0);
1295     }
1296     qemu_vfree(buf);
1297 
1298     return nbytes;
1299 }
1300 
1301 #ifdef CONFIG_XFS
1302 static int xfs_write_zeroes(BDRVRawState *s, int64_t offset, uint64_t bytes)
1303 {
1304     struct xfs_flock64 fl;
1305     int err;
1306 
1307     memset(&fl, 0, sizeof(fl));
1308     fl.l_whence = SEEK_SET;
1309     fl.l_start = offset;
1310     fl.l_len = bytes;
1311 
1312     if (xfsctl(NULL, s->fd, XFS_IOC_ZERO_RANGE, &fl) < 0) {
1313         err = errno;
1314         DPRINTF("cannot write zero range (%s)\n", strerror(errno));
1315         return -err;
1316     }
1317 
1318     return 0;
1319 }
1320 
1321 static int xfs_discard(BDRVRawState *s, int64_t offset, uint64_t bytes)
1322 {
1323     struct xfs_flock64 fl;
1324     int err;
1325 
1326     memset(&fl, 0, sizeof(fl));
1327     fl.l_whence = SEEK_SET;
1328     fl.l_start = offset;
1329     fl.l_len = bytes;
1330 
1331     if (xfsctl(NULL, s->fd, XFS_IOC_UNRESVSP64, &fl) < 0) {
1332         err = errno;
1333         DPRINTF("cannot punch hole (%s)\n", strerror(errno));
1334         return -err;
1335     }
1336 
1337     return 0;
1338 }
1339 #endif
1340 
1341 static int translate_err(int err)
1342 {
1343     if (err == -ENODEV || err == -ENOSYS || err == -EOPNOTSUPP ||
1344         err == -ENOTTY) {
1345         err = -ENOTSUP;
1346     }
1347     return err;
1348 }
1349 
1350 #ifdef CONFIG_FALLOCATE
1351 static int do_fallocate(int fd, int mode, off_t offset, off_t len)
1352 {
1353     do {
1354         if (fallocate(fd, mode, offset, len) == 0) {
1355             return 0;
1356         }
1357     } while (errno == EINTR);
1358     return translate_err(-errno);
1359 }
1360 #endif
1361 
1362 static ssize_t handle_aiocb_write_zeroes_block(RawPosixAIOData *aiocb)
1363 {
1364     int ret = -ENOTSUP;
1365     BDRVRawState *s = aiocb->bs->opaque;
1366 
1367     if (!s->has_write_zeroes) {
1368         return -ENOTSUP;
1369     }
1370 
1371 #ifdef BLKZEROOUT
1372     do {
1373         uint64_t range[2] = { aiocb->aio_offset, aiocb->aio_nbytes };
1374         if (ioctl(aiocb->aio_fildes, BLKZEROOUT, range) == 0) {
1375             return 0;
1376         }
1377     } while (errno == EINTR);
1378 
1379     ret = translate_err(-errno);
1380 #endif
1381 
1382     if (ret == -ENOTSUP) {
1383         s->has_write_zeroes = false;
1384     }
1385     return ret;
1386 }
1387 
1388 static ssize_t handle_aiocb_write_zeroes(RawPosixAIOData *aiocb)
1389 {
1390 #if defined(CONFIG_FALLOCATE) || defined(CONFIG_XFS)
1391     BDRVRawState *s = aiocb->bs->opaque;
1392 #endif
1393 #ifdef CONFIG_FALLOCATE
1394     int64_t len;
1395 #endif
1396 
1397     if (aiocb->aio_type & QEMU_AIO_BLKDEV) {
1398         return handle_aiocb_write_zeroes_block(aiocb);
1399     }
1400 
1401 #ifdef CONFIG_XFS
1402     if (s->is_xfs) {
1403         return xfs_write_zeroes(s, aiocb->aio_offset, aiocb->aio_nbytes);
1404     }
1405 #endif
1406 
1407 #ifdef CONFIG_FALLOCATE_ZERO_RANGE
1408     if (s->has_write_zeroes) {
1409         int ret = do_fallocate(s->fd, FALLOC_FL_ZERO_RANGE,
1410                                aiocb->aio_offset, aiocb->aio_nbytes);
1411         if (ret == 0 || ret != -ENOTSUP) {
1412             return ret;
1413         }
1414         s->has_write_zeroes = false;
1415     }
1416 #endif
1417 
1418 #ifdef CONFIG_FALLOCATE_PUNCH_HOLE
1419     if (s->has_discard && s->has_fallocate) {
1420         int ret = do_fallocate(s->fd,
1421                                FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE,
1422                                aiocb->aio_offset, aiocb->aio_nbytes);
1423         if (ret == 0) {
1424             ret = do_fallocate(s->fd, 0, aiocb->aio_offset, aiocb->aio_nbytes);
1425             if (ret == 0 || ret != -ENOTSUP) {
1426                 return ret;
1427             }
1428             s->has_fallocate = false;
1429         } else if (ret != -ENOTSUP) {
1430             return ret;
1431         } else {
1432             s->has_discard = false;
1433         }
1434     }
1435 #endif
1436 
1437 #ifdef CONFIG_FALLOCATE
1438     /* Last resort: we are trying to extend the file with zeroed data. This
1439      * can be done via fallocate(fd, 0) */
1440     len = bdrv_getlength(aiocb->bs);
1441     if (s->has_fallocate && len >= 0 && aiocb->aio_offset >= len) {
1442         int ret = do_fallocate(s->fd, 0, aiocb->aio_offset, aiocb->aio_nbytes);
1443         if (ret == 0 || ret != -ENOTSUP) {
1444             return ret;
1445         }
1446         s->has_fallocate = false;
1447     }
1448 #endif
1449 
1450     return -ENOTSUP;
1451 }
1452 
1453 #ifndef HAVE_COPY_FILE_RANGE
1454 static off_t copy_file_range(int in_fd, off_t *in_off, int out_fd,
1455                              off_t *out_off, size_t len, unsigned int flags)
1456 {
1457 #ifdef __NR_copy_file_range
1458     return syscall(__NR_copy_file_range, in_fd, in_off, out_fd,
1459                    out_off, len, flags);
1460 #else
1461     errno = ENOSYS;
1462     return -1;
1463 #endif
1464 }
1465 #endif
1466 
1467 static ssize_t handle_aiocb_copy_range(RawPosixAIOData *aiocb)
1468 {
1469     uint64_t bytes = aiocb->aio_nbytes;
1470     off_t in_off = aiocb->aio_offset;
1471     off_t out_off = aiocb->aio_offset2;
1472 
1473     while (bytes) {
1474         ssize_t ret = copy_file_range(aiocb->aio_fildes, &in_off,
1475                                       aiocb->aio_fd2, &out_off,
1476                                       bytes, 0);
1477         if (ret == -EINTR) {
1478             continue;
1479         }
1480         if (ret < 0) {
1481             if (errno == ENOSYS) {
1482                 return -ENOTSUP;
1483             } else {
1484                 return -errno;
1485             }
1486         }
1487         if (!ret) {
1488             /* No progress (e.g. when beyond EOF), fall back to buffer I/O. */
1489             return -ENOTSUP;
1490         }
1491         bytes -= ret;
1492     }
1493     return 0;
1494 }
1495 
1496 static ssize_t handle_aiocb_discard(RawPosixAIOData *aiocb)
1497 {
1498     int ret = -EOPNOTSUPP;
1499     BDRVRawState *s = aiocb->bs->opaque;
1500 
1501     if (!s->has_discard) {
1502         return -ENOTSUP;
1503     }
1504 
1505     if (aiocb->aio_type & QEMU_AIO_BLKDEV) {
1506 #ifdef BLKDISCARD
1507         do {
1508             uint64_t range[2] = { aiocb->aio_offset, aiocb->aio_nbytes };
1509             if (ioctl(aiocb->aio_fildes, BLKDISCARD, range) == 0) {
1510                 return 0;
1511             }
1512         } while (errno == EINTR);
1513 
1514         ret = -errno;
1515 #endif
1516     } else {
1517 #ifdef CONFIG_XFS
1518         if (s->is_xfs) {
1519             return xfs_discard(s, aiocb->aio_offset, aiocb->aio_nbytes);
1520         }
1521 #endif
1522 
1523 #ifdef CONFIG_FALLOCATE_PUNCH_HOLE
1524         ret = do_fallocate(s->fd, FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE,
1525                            aiocb->aio_offset, aiocb->aio_nbytes);
1526 #endif
1527     }
1528 
1529     ret = translate_err(ret);
1530     if (ret == -ENOTSUP) {
1531         s->has_discard = false;
1532     }
1533     return ret;
1534 }
1535 
1536 static int aio_worker(void *arg)
1537 {
1538     RawPosixAIOData *aiocb = arg;
1539     ssize_t ret = 0;
1540 
1541     switch (aiocb->aio_type & QEMU_AIO_TYPE_MASK) {
1542     case QEMU_AIO_READ:
1543         ret = handle_aiocb_rw(aiocb);
1544         if (ret >= 0 && ret < aiocb->aio_nbytes) {
1545             iov_memset(aiocb->aio_iov, aiocb->aio_niov, ret,
1546                       0, aiocb->aio_nbytes - ret);
1547 
1548             ret = aiocb->aio_nbytes;
1549         }
1550         if (ret == aiocb->aio_nbytes) {
1551             ret = 0;
1552         } else if (ret >= 0 && ret < aiocb->aio_nbytes) {
1553             ret = -EINVAL;
1554         }
1555         break;
1556     case QEMU_AIO_WRITE:
1557         ret = handle_aiocb_rw(aiocb);
1558         if (ret == aiocb->aio_nbytes) {
1559             ret = 0;
1560         } else if (ret >= 0 && ret < aiocb->aio_nbytes) {
1561             ret = -EINVAL;
1562         }
1563         break;
1564     case QEMU_AIO_FLUSH:
1565         ret = handle_aiocb_flush(aiocb);
1566         break;
1567     case QEMU_AIO_IOCTL:
1568         ret = handle_aiocb_ioctl(aiocb);
1569         break;
1570     case QEMU_AIO_DISCARD:
1571         ret = handle_aiocb_discard(aiocb);
1572         break;
1573     case QEMU_AIO_WRITE_ZEROES:
1574         ret = handle_aiocb_write_zeroes(aiocb);
1575         break;
1576     case QEMU_AIO_COPY_RANGE:
1577         ret = handle_aiocb_copy_range(aiocb);
1578         break;
1579     default:
1580         fprintf(stderr, "invalid aio request (0x%x)\n", aiocb->aio_type);
1581         ret = -EINVAL;
1582         break;
1583     }
1584 
1585     g_free(aiocb);
1586     return ret;
1587 }
1588 
1589 static int paio_submit_co_full(BlockDriverState *bs, int fd,
1590                                int64_t offset, int fd2, int64_t offset2,
1591                                QEMUIOVector *qiov,
1592                                int bytes, int type)
1593 {
1594     RawPosixAIOData *acb = g_new(RawPosixAIOData, 1);
1595     ThreadPool *pool;
1596 
1597     acb->bs = bs;
1598     acb->aio_type = type;
1599     acb->aio_fildes = fd;
1600     acb->aio_fd2 = fd2;
1601     acb->aio_offset2 = offset2;
1602 
1603     acb->aio_nbytes = bytes;
1604     acb->aio_offset = offset;
1605 
1606     if (qiov) {
1607         acb->aio_iov = qiov->iov;
1608         acb->aio_niov = qiov->niov;
1609         assert(qiov->size == bytes);
1610     }
1611 
1612     trace_paio_submit_co(offset, bytes, type);
1613     pool = aio_get_thread_pool(bdrv_get_aio_context(bs));
1614     return thread_pool_submit_co(pool, aio_worker, acb);
1615 }
1616 
1617 static inline int paio_submit_co(BlockDriverState *bs, int fd,
1618                                  int64_t offset, QEMUIOVector *qiov,
1619                                  int bytes, int type)
1620 {
1621     return paio_submit_co_full(bs, fd, offset, -1, 0, qiov, bytes, type);
1622 }
1623 
1624 static BlockAIOCB *paio_submit(BlockDriverState *bs, int fd,
1625         int64_t offset, QEMUIOVector *qiov, int bytes,
1626         BlockCompletionFunc *cb, void *opaque, int type)
1627 {
1628     RawPosixAIOData *acb = g_new(RawPosixAIOData, 1);
1629     ThreadPool *pool;
1630 
1631     acb->bs = bs;
1632     acb->aio_type = type;
1633     acb->aio_fildes = fd;
1634 
1635     acb->aio_nbytes = bytes;
1636     acb->aio_offset = offset;
1637 
1638     if (qiov) {
1639         acb->aio_iov = qiov->iov;
1640         acb->aio_niov = qiov->niov;
1641         assert(qiov->size == acb->aio_nbytes);
1642     }
1643 
1644     trace_paio_submit(acb, opaque, offset, bytes, type);
1645     pool = aio_get_thread_pool(bdrv_get_aio_context(bs));
1646     return thread_pool_submit_aio(pool, aio_worker, acb, cb, opaque);
1647 }
1648 
1649 static int coroutine_fn raw_co_prw(BlockDriverState *bs, uint64_t offset,
1650                                    uint64_t bytes, QEMUIOVector *qiov, int type)
1651 {
1652     BDRVRawState *s = bs->opaque;
1653 
1654     if (fd_open(bs) < 0)
1655         return -EIO;
1656 
1657     /*
1658      * Check if the underlying device requires requests to be aligned,
1659      * and if the request we are trying to submit is aligned or not.
1660      * If this is the case tell the low-level driver that it needs
1661      * to copy the buffer.
1662      */
1663     if (s->needs_alignment) {
1664         if (!bdrv_qiov_is_aligned(bs, qiov)) {
1665             type |= QEMU_AIO_MISALIGNED;
1666 #ifdef CONFIG_LINUX_AIO
1667         } else if (s->use_linux_aio) {
1668             LinuxAioState *aio = aio_get_linux_aio(bdrv_get_aio_context(bs));
1669             assert(qiov->size == bytes);
1670             return laio_co_submit(bs, aio, s->fd, offset, qiov, type);
1671 #endif
1672         }
1673     }
1674 
1675     return paio_submit_co(bs, s->fd, offset, qiov, bytes, type);
1676 }
1677 
1678 static int coroutine_fn raw_co_preadv(BlockDriverState *bs, uint64_t offset,
1679                                       uint64_t bytes, QEMUIOVector *qiov,
1680                                       int flags)
1681 {
1682     return raw_co_prw(bs, offset, bytes, qiov, QEMU_AIO_READ);
1683 }
1684 
1685 static int coroutine_fn raw_co_pwritev(BlockDriverState *bs, uint64_t offset,
1686                                        uint64_t bytes, QEMUIOVector *qiov,
1687                                        int flags)
1688 {
1689     assert(flags == 0);
1690     return raw_co_prw(bs, offset, bytes, qiov, QEMU_AIO_WRITE);
1691 }
1692 
1693 static void raw_aio_plug(BlockDriverState *bs)
1694 {
1695 #ifdef CONFIG_LINUX_AIO
1696     BDRVRawState *s = bs->opaque;
1697     if (s->use_linux_aio) {
1698         LinuxAioState *aio = aio_get_linux_aio(bdrv_get_aio_context(bs));
1699         laio_io_plug(bs, aio);
1700     }
1701 #endif
1702 }
1703 
1704 static void raw_aio_unplug(BlockDriverState *bs)
1705 {
1706 #ifdef CONFIG_LINUX_AIO
1707     BDRVRawState *s = bs->opaque;
1708     if (s->use_linux_aio) {
1709         LinuxAioState *aio = aio_get_linux_aio(bdrv_get_aio_context(bs));
1710         laio_io_unplug(bs, aio);
1711     }
1712 #endif
1713 }
1714 
1715 static BlockAIOCB *raw_aio_flush(BlockDriverState *bs,
1716         BlockCompletionFunc *cb, void *opaque)
1717 {
1718     BDRVRawState *s = bs->opaque;
1719 
1720     if (fd_open(bs) < 0)
1721         return NULL;
1722 
1723     return paio_submit(bs, s->fd, 0, NULL, 0, cb, opaque, QEMU_AIO_FLUSH);
1724 }
1725 
1726 static void raw_close(BlockDriverState *bs)
1727 {
1728     BDRVRawState *s = bs->opaque;
1729 
1730     if (s->fd >= 0) {
1731         qemu_close(s->fd);
1732         s->fd = -1;
1733     }
1734     if (s->lock_fd >= 0) {
1735         qemu_close(s->lock_fd);
1736         s->lock_fd = -1;
1737     }
1738 }
1739 
1740 /**
1741  * Truncates the given regular file @fd to @offset and, when growing, fills the
1742  * new space according to @prealloc.
1743  *
1744  * Returns: 0 on success, -errno on failure.
1745  */
1746 static int raw_regular_truncate(int fd, int64_t offset, PreallocMode prealloc,
1747                                 Error **errp)
1748 {
1749     int result = 0;
1750     int64_t current_length = 0;
1751     char *buf = NULL;
1752     struct stat st;
1753 
1754     if (fstat(fd, &st) < 0) {
1755         result = -errno;
1756         error_setg_errno(errp, -result, "Could not stat file");
1757         return result;
1758     }
1759 
1760     current_length = st.st_size;
1761     if (current_length > offset && prealloc != PREALLOC_MODE_OFF) {
1762         error_setg(errp, "Cannot use preallocation for shrinking files");
1763         return -ENOTSUP;
1764     }
1765 
1766     switch (prealloc) {
1767 #ifdef CONFIG_POSIX_FALLOCATE
1768     case PREALLOC_MODE_FALLOC:
1769         /*
1770          * Truncating before posix_fallocate() makes it about twice slower on
1771          * file systems that do not support fallocate(), trying to check if a
1772          * block is allocated before allocating it, so don't do that here.
1773          */
1774         if (offset != current_length) {
1775             result = -posix_fallocate(fd, current_length, offset - current_length);
1776             if (result != 0) {
1777                 /* posix_fallocate() doesn't set errno. */
1778                 error_setg_errno(errp, -result,
1779                                  "Could not preallocate new data");
1780             }
1781         } else {
1782             result = 0;
1783         }
1784         goto out;
1785 #endif
1786     case PREALLOC_MODE_FULL:
1787     {
1788         int64_t num = 0, left = offset - current_length;
1789         off_t seek_result;
1790 
1791         /*
1792          * Knowing the final size from the beginning could allow the file
1793          * system driver to do less allocations and possibly avoid
1794          * fragmentation of the file.
1795          */
1796         if (ftruncate(fd, offset) != 0) {
1797             result = -errno;
1798             error_setg_errno(errp, -result, "Could not resize file");
1799             goto out;
1800         }
1801 
1802         buf = g_malloc0(65536);
1803 
1804         seek_result = lseek(fd, current_length, SEEK_SET);
1805         if (seek_result < 0) {
1806             result = -errno;
1807             error_setg_errno(errp, -result,
1808                              "Failed to seek to the old end of file");
1809             goto out;
1810         }
1811 
1812         while (left > 0) {
1813             num = MIN(left, 65536);
1814             result = write(fd, buf, num);
1815             if (result < 0) {
1816                 result = -errno;
1817                 error_setg_errno(errp, -result,
1818                                  "Could not write zeros for preallocation");
1819                 goto out;
1820             }
1821             left -= result;
1822         }
1823         if (result >= 0) {
1824             result = fsync(fd);
1825             if (result < 0) {
1826                 result = -errno;
1827                 error_setg_errno(errp, -result,
1828                                  "Could not flush file to disk");
1829                 goto out;
1830             }
1831         }
1832         goto out;
1833     }
1834     case PREALLOC_MODE_OFF:
1835         if (ftruncate(fd, offset) != 0) {
1836             result = -errno;
1837             error_setg_errno(errp, -result, "Could not resize file");
1838         }
1839         return result;
1840     default:
1841         result = -ENOTSUP;
1842         error_setg(errp, "Unsupported preallocation mode: %s",
1843                    PreallocMode_str(prealloc));
1844         return result;
1845     }
1846 
1847 out:
1848     if (result < 0) {
1849         if (ftruncate(fd, current_length) < 0) {
1850             error_report("Failed to restore old file length: %s",
1851                          strerror(errno));
1852         }
1853     }
1854 
1855     g_free(buf);
1856     return result;
1857 }
1858 
1859 static int raw_truncate(BlockDriverState *bs, int64_t offset,
1860                         PreallocMode prealloc, Error **errp)
1861 {
1862     BDRVRawState *s = bs->opaque;
1863     struct stat st;
1864     int ret;
1865 
1866     if (fstat(s->fd, &st)) {
1867         ret = -errno;
1868         error_setg_errno(errp, -ret, "Failed to fstat() the file");
1869         return ret;
1870     }
1871 
1872     if (S_ISREG(st.st_mode)) {
1873         return raw_regular_truncate(s->fd, offset, prealloc, errp);
1874     }
1875 
1876     if (prealloc != PREALLOC_MODE_OFF) {
1877         error_setg(errp, "Preallocation mode '%s' unsupported for this "
1878                    "non-regular file", PreallocMode_str(prealloc));
1879         return -ENOTSUP;
1880     }
1881 
1882     if (S_ISCHR(st.st_mode) || S_ISBLK(st.st_mode)) {
1883         if (offset > raw_getlength(bs)) {
1884             error_setg(errp, "Cannot grow device files");
1885             return -EINVAL;
1886         }
1887     } else {
1888         error_setg(errp, "Resizing this file is not supported");
1889         return -ENOTSUP;
1890     }
1891 
1892     return 0;
1893 }
1894 
1895 #ifdef __OpenBSD__
1896 static int64_t raw_getlength(BlockDriverState *bs)
1897 {
1898     BDRVRawState *s = bs->opaque;
1899     int fd = s->fd;
1900     struct stat st;
1901 
1902     if (fstat(fd, &st))
1903         return -errno;
1904     if (S_ISCHR(st.st_mode) || S_ISBLK(st.st_mode)) {
1905         struct disklabel dl;
1906 
1907         if (ioctl(fd, DIOCGDINFO, &dl))
1908             return -errno;
1909         return (uint64_t)dl.d_secsize *
1910             dl.d_partitions[DISKPART(st.st_rdev)].p_size;
1911     } else
1912         return st.st_size;
1913 }
1914 #elif defined(__NetBSD__)
1915 static int64_t raw_getlength(BlockDriverState *bs)
1916 {
1917     BDRVRawState *s = bs->opaque;
1918     int fd = s->fd;
1919     struct stat st;
1920 
1921     if (fstat(fd, &st))
1922         return -errno;
1923     if (S_ISCHR(st.st_mode) || S_ISBLK(st.st_mode)) {
1924         struct dkwedge_info dkw;
1925 
1926         if (ioctl(fd, DIOCGWEDGEINFO, &dkw) != -1) {
1927             return dkw.dkw_size * 512;
1928         } else {
1929             struct disklabel dl;
1930 
1931             if (ioctl(fd, DIOCGDINFO, &dl))
1932                 return -errno;
1933             return (uint64_t)dl.d_secsize *
1934                 dl.d_partitions[DISKPART(st.st_rdev)].p_size;
1935         }
1936     } else
1937         return st.st_size;
1938 }
1939 #elif defined(__sun__)
1940 static int64_t raw_getlength(BlockDriverState *bs)
1941 {
1942     BDRVRawState *s = bs->opaque;
1943     struct dk_minfo minfo;
1944     int ret;
1945     int64_t size;
1946 
1947     ret = fd_open(bs);
1948     if (ret < 0) {
1949         return ret;
1950     }
1951 
1952     /*
1953      * Use the DKIOCGMEDIAINFO ioctl to read the size.
1954      */
1955     ret = ioctl(s->fd, DKIOCGMEDIAINFO, &minfo);
1956     if (ret != -1) {
1957         return minfo.dki_lbsize * minfo.dki_capacity;
1958     }
1959 
1960     /*
1961      * There are reports that lseek on some devices fails, but
1962      * irc discussion said that contingency on contingency was overkill.
1963      */
1964     size = lseek(s->fd, 0, SEEK_END);
1965     if (size < 0) {
1966         return -errno;
1967     }
1968     return size;
1969 }
1970 #elif defined(CONFIG_BSD)
1971 static int64_t raw_getlength(BlockDriverState *bs)
1972 {
1973     BDRVRawState *s = bs->opaque;
1974     int fd = s->fd;
1975     int64_t size;
1976     struct stat sb;
1977 #if defined (__FreeBSD__) || defined(__FreeBSD_kernel__)
1978     int reopened = 0;
1979 #endif
1980     int ret;
1981 
1982     ret = fd_open(bs);
1983     if (ret < 0)
1984         return ret;
1985 
1986 #if defined (__FreeBSD__) || defined(__FreeBSD_kernel__)
1987 again:
1988 #endif
1989     if (!fstat(fd, &sb) && (S_IFCHR & sb.st_mode)) {
1990 #ifdef DIOCGMEDIASIZE
1991 	if (ioctl(fd, DIOCGMEDIASIZE, (off_t *)&size))
1992 #elif defined(DIOCGPART)
1993         {
1994                 struct partinfo pi;
1995                 if (ioctl(fd, DIOCGPART, &pi) == 0)
1996                         size = pi.media_size;
1997                 else
1998                         size = 0;
1999         }
2000         if (size == 0)
2001 #endif
2002 #if defined(__APPLE__) && defined(__MACH__)
2003         {
2004             uint64_t sectors = 0;
2005             uint32_t sector_size = 0;
2006 
2007             if (ioctl(fd, DKIOCGETBLOCKCOUNT, &sectors) == 0
2008                && ioctl(fd, DKIOCGETBLOCKSIZE, &sector_size) == 0) {
2009                 size = sectors * sector_size;
2010             } else {
2011                 size = lseek(fd, 0LL, SEEK_END);
2012                 if (size < 0) {
2013                     return -errno;
2014                 }
2015             }
2016         }
2017 #else
2018         size = lseek(fd, 0LL, SEEK_END);
2019         if (size < 0) {
2020             return -errno;
2021         }
2022 #endif
2023 #if defined(__FreeBSD__) || defined(__FreeBSD_kernel__)
2024         switch(s->type) {
2025         case FTYPE_CD:
2026             /* XXX FreeBSD acd returns UINT_MAX sectors for an empty drive */
2027             if (size == 2048LL * (unsigned)-1)
2028                 size = 0;
2029             /* XXX no disc?  maybe we need to reopen... */
2030             if (size <= 0 && !reopened && cdrom_reopen(bs) >= 0) {
2031                 reopened = 1;
2032                 goto again;
2033             }
2034         }
2035 #endif
2036     } else {
2037         size = lseek(fd, 0, SEEK_END);
2038         if (size < 0) {
2039             return -errno;
2040         }
2041     }
2042     return size;
2043 }
2044 #else
2045 static int64_t raw_getlength(BlockDriverState *bs)
2046 {
2047     BDRVRawState *s = bs->opaque;
2048     int ret;
2049     int64_t size;
2050 
2051     ret = fd_open(bs);
2052     if (ret < 0) {
2053         return ret;
2054     }
2055 
2056     size = lseek(s->fd, 0, SEEK_END);
2057     if (size < 0) {
2058         return -errno;
2059     }
2060     return size;
2061 }
2062 #endif
2063 
2064 static int64_t raw_get_allocated_file_size(BlockDriverState *bs)
2065 {
2066     struct stat st;
2067     BDRVRawState *s = bs->opaque;
2068 
2069     if (fstat(s->fd, &st) < 0) {
2070         return -errno;
2071     }
2072     return (int64_t)st.st_blocks * 512;
2073 }
2074 
2075 static int raw_co_create(BlockdevCreateOptions *options, Error **errp)
2076 {
2077     BlockdevCreateOptionsFile *file_opts;
2078     int fd;
2079     int perm, shared;
2080     int result = 0;
2081 
2082     /* Validate options and set default values */
2083     assert(options->driver == BLOCKDEV_DRIVER_FILE);
2084     file_opts = &options->u.file;
2085 
2086     if (!file_opts->has_nocow) {
2087         file_opts->nocow = false;
2088     }
2089     if (!file_opts->has_preallocation) {
2090         file_opts->preallocation = PREALLOC_MODE_OFF;
2091     }
2092 
2093     /* Create file */
2094     fd = qemu_open(file_opts->filename, O_RDWR | O_CREAT | O_BINARY, 0644);
2095     if (fd < 0) {
2096         result = -errno;
2097         error_setg_errno(errp, -result, "Could not create file");
2098         goto out;
2099     }
2100 
2101     /* Take permissions: We want to discard everything, so we need
2102      * BLK_PERM_WRITE; and truncation to the desired size requires
2103      * BLK_PERM_RESIZE.
2104      * On the other hand, we cannot share the RESIZE permission
2105      * because we promise that after this function, the file has the
2106      * size given in the options.  If someone else were to resize it
2107      * concurrently, we could not guarantee that.
2108      * Note that after this function, we can no longer guarantee that
2109      * the file is not touched by a third party, so it may be resized
2110      * then. */
2111     perm = BLK_PERM_WRITE | BLK_PERM_RESIZE;
2112     shared = BLK_PERM_ALL & ~BLK_PERM_RESIZE;
2113 
2114     /* Step one: Take locks */
2115     result = raw_apply_lock_bytes(fd, perm, shared, false, errp);
2116     if (result < 0) {
2117         goto out_close;
2118     }
2119 
2120     /* Step two: Check that nobody else has taken conflicting locks */
2121     result = raw_check_lock_bytes(fd, perm, shared, errp);
2122     if (result < 0) {
2123         goto out_close;
2124     }
2125 
2126     /* Clear the file by truncating it to 0 */
2127     result = raw_regular_truncate(fd, 0, PREALLOC_MODE_OFF, errp);
2128     if (result < 0) {
2129         goto out_close;
2130     }
2131 
2132     if (file_opts->nocow) {
2133 #ifdef __linux__
2134         /* Set NOCOW flag to solve performance issue on fs like btrfs.
2135          * This is an optimisation. The FS_IOC_SETFLAGS ioctl return value
2136          * will be ignored since any failure of this operation should not
2137          * block the left work.
2138          */
2139         int attr;
2140         if (ioctl(fd, FS_IOC_GETFLAGS, &attr) == 0) {
2141             attr |= FS_NOCOW_FL;
2142             ioctl(fd, FS_IOC_SETFLAGS, &attr);
2143         }
2144 #endif
2145     }
2146 
2147     /* Resize and potentially preallocate the file to the desired
2148      * final size */
2149     result = raw_regular_truncate(fd, file_opts->size, file_opts->preallocation,
2150                                   errp);
2151     if (result < 0) {
2152         goto out_close;
2153     }
2154 
2155 out_close:
2156     if (qemu_close(fd) != 0 && result == 0) {
2157         result = -errno;
2158         error_setg_errno(errp, -result, "Could not close the new file");
2159     }
2160 out:
2161     return result;
2162 }
2163 
2164 static int coroutine_fn raw_co_create_opts(const char *filename, QemuOpts *opts,
2165                                            Error **errp)
2166 {
2167     BlockdevCreateOptions options;
2168     int64_t total_size = 0;
2169     bool nocow = false;
2170     PreallocMode prealloc;
2171     char *buf = NULL;
2172     Error *local_err = NULL;
2173 
2174     /* Skip file: protocol prefix */
2175     strstart(filename, "file:", &filename);
2176 
2177     /* Read out options */
2178     total_size = ROUND_UP(qemu_opt_get_size_del(opts, BLOCK_OPT_SIZE, 0),
2179                           BDRV_SECTOR_SIZE);
2180     nocow = qemu_opt_get_bool(opts, BLOCK_OPT_NOCOW, false);
2181     buf = qemu_opt_get_del(opts, BLOCK_OPT_PREALLOC);
2182     prealloc = qapi_enum_parse(&PreallocMode_lookup, buf,
2183                                PREALLOC_MODE_OFF, &local_err);
2184     g_free(buf);
2185     if (local_err) {
2186         error_propagate(errp, local_err);
2187         return -EINVAL;
2188     }
2189 
2190     options = (BlockdevCreateOptions) {
2191         .driver     = BLOCKDEV_DRIVER_FILE,
2192         .u.file     = {
2193             .filename           = (char *) filename,
2194             .size               = total_size,
2195             .has_preallocation  = true,
2196             .preallocation      = prealloc,
2197             .has_nocow          = true,
2198             .nocow              = nocow,
2199         },
2200     };
2201     return raw_co_create(&options, errp);
2202 }
2203 
2204 /*
2205  * Find allocation range in @bs around offset @start.
2206  * May change underlying file descriptor's file offset.
2207  * If @start is not in a hole, store @start in @data, and the
2208  * beginning of the next hole in @hole, and return 0.
2209  * If @start is in a non-trailing hole, store @start in @hole and the
2210  * beginning of the next non-hole in @data, and return 0.
2211  * If @start is in a trailing hole or beyond EOF, return -ENXIO.
2212  * If we can't find out, return a negative errno other than -ENXIO.
2213  */
2214 static int find_allocation(BlockDriverState *bs, off_t start,
2215                            off_t *data, off_t *hole)
2216 {
2217 #if defined SEEK_HOLE && defined SEEK_DATA
2218     BDRVRawState *s = bs->opaque;
2219     off_t offs;
2220 
2221     /*
2222      * SEEK_DATA cases:
2223      * D1. offs == start: start is in data
2224      * D2. offs > start: start is in a hole, next data at offs
2225      * D3. offs < 0, errno = ENXIO: either start is in a trailing hole
2226      *                              or start is beyond EOF
2227      *     If the latter happens, the file has been truncated behind
2228      *     our back since we opened it.  All bets are off then.
2229      *     Treating like a trailing hole is simplest.
2230      * D4. offs < 0, errno != ENXIO: we learned nothing
2231      */
2232     offs = lseek(s->fd, start, SEEK_DATA);
2233     if (offs < 0) {
2234         return -errno;          /* D3 or D4 */
2235     }
2236 
2237     if (offs < start) {
2238         /* This is not a valid return by lseek().  We are safe to just return
2239          * -EIO in this case, and we'll treat it like D4. */
2240         return -EIO;
2241     }
2242 
2243     if (offs > start) {
2244         /* D2: in hole, next data at offs */
2245         *hole = start;
2246         *data = offs;
2247         return 0;
2248     }
2249 
2250     /* D1: in data, end not yet known */
2251 
2252     /*
2253      * SEEK_HOLE cases:
2254      * H1. offs == start: start is in a hole
2255      *     If this happens here, a hole has been dug behind our back
2256      *     since the previous lseek().
2257      * H2. offs > start: either start is in data, next hole at offs,
2258      *                   or start is in trailing hole, EOF at offs
2259      *     Linux treats trailing holes like any other hole: offs ==
2260      *     start.  Solaris seeks to EOF instead: offs > start (blech).
2261      *     If that happens here, a hole has been dug behind our back
2262      *     since the previous lseek().
2263      * H3. offs < 0, errno = ENXIO: start is beyond EOF
2264      *     If this happens, the file has been truncated behind our
2265      *     back since we opened it.  Treat it like a trailing hole.
2266      * H4. offs < 0, errno != ENXIO: we learned nothing
2267      *     Pretend we know nothing at all, i.e. "forget" about D1.
2268      */
2269     offs = lseek(s->fd, start, SEEK_HOLE);
2270     if (offs < 0) {
2271         return -errno;          /* D1 and (H3 or H4) */
2272     }
2273 
2274     if (offs < start) {
2275         /* This is not a valid return by lseek().  We are safe to just return
2276          * -EIO in this case, and we'll treat it like H4. */
2277         return -EIO;
2278     }
2279 
2280     if (offs > start) {
2281         /*
2282          * D1 and H2: either in data, next hole at offs, or it was in
2283          * data but is now in a trailing hole.  In the latter case,
2284          * all bets are off.  Treating it as if it there was data all
2285          * the way to EOF is safe, so simply do that.
2286          */
2287         *data = start;
2288         *hole = offs;
2289         return 0;
2290     }
2291 
2292     /* D1 and H1 */
2293     return -EBUSY;
2294 #else
2295     return -ENOTSUP;
2296 #endif
2297 }
2298 
2299 /*
2300  * Returns the allocation status of the specified offset.
2301  *
2302  * The block layer guarantees 'offset' and 'bytes' are within bounds.
2303  *
2304  * 'pnum' is set to the number of bytes (including and immediately following
2305  * the specified offset) that are known to be in the same
2306  * allocated/unallocated state.
2307  *
2308  * 'bytes' is the max value 'pnum' should be set to.
2309  */
2310 static int coroutine_fn raw_co_block_status(BlockDriverState *bs,
2311                                             bool want_zero,
2312                                             int64_t offset,
2313                                             int64_t bytes, int64_t *pnum,
2314                                             int64_t *map,
2315                                             BlockDriverState **file)
2316 {
2317     off_t data = 0, hole = 0;
2318     int ret;
2319 
2320     ret = fd_open(bs);
2321     if (ret < 0) {
2322         return ret;
2323     }
2324 
2325     if (!want_zero) {
2326         *pnum = bytes;
2327         *map = offset;
2328         *file = bs;
2329         return BDRV_BLOCK_DATA | BDRV_BLOCK_OFFSET_VALID;
2330     }
2331 
2332     ret = find_allocation(bs, offset, &data, &hole);
2333     if (ret == -ENXIO) {
2334         /* Trailing hole */
2335         *pnum = bytes;
2336         ret = BDRV_BLOCK_ZERO;
2337     } else if (ret < 0) {
2338         /* No info available, so pretend there are no holes */
2339         *pnum = bytes;
2340         ret = BDRV_BLOCK_DATA;
2341     } else if (data == offset) {
2342         /* On a data extent, compute bytes to the end of the extent,
2343          * possibly including a partial sector at EOF. */
2344         *pnum = MIN(bytes, hole - offset);
2345         ret = BDRV_BLOCK_DATA;
2346     } else {
2347         /* On a hole, compute bytes to the beginning of the next extent.  */
2348         assert(hole == offset);
2349         *pnum = MIN(bytes, data - offset);
2350         ret = BDRV_BLOCK_ZERO;
2351     }
2352     *map = offset;
2353     *file = bs;
2354     return ret | BDRV_BLOCK_OFFSET_VALID;
2355 }
2356 
2357 #if defined(__linux__)
2358 /* Verify that the file is not in the page cache */
2359 static void check_cache_dropped(BlockDriverState *bs, Error **errp)
2360 {
2361     const size_t window_size = 128 * 1024 * 1024;
2362     BDRVRawState *s = bs->opaque;
2363     void *window = NULL;
2364     size_t length = 0;
2365     unsigned char *vec;
2366     size_t page_size;
2367     off_t offset;
2368     off_t end;
2369 
2370     /* mincore(2) page status information requires 1 byte per page */
2371     page_size = sysconf(_SC_PAGESIZE);
2372     vec = g_malloc(DIV_ROUND_UP(window_size, page_size));
2373 
2374     end = raw_getlength(bs);
2375 
2376     for (offset = 0; offset < end; offset += window_size) {
2377         void *new_window;
2378         size_t new_length;
2379         size_t vec_end;
2380         size_t i;
2381         int ret;
2382 
2383         /* Unmap previous window if size has changed */
2384         new_length = MIN(end - offset, window_size);
2385         if (new_length != length) {
2386             munmap(window, length);
2387             window = NULL;
2388             length = 0;
2389         }
2390 
2391         new_window = mmap(window, new_length, PROT_NONE, MAP_PRIVATE,
2392                           s->fd, offset);
2393         if (new_window == MAP_FAILED) {
2394             error_setg_errno(errp, errno, "mmap failed");
2395             break;
2396         }
2397 
2398         window = new_window;
2399         length = new_length;
2400 
2401         ret = mincore(window, length, vec);
2402         if (ret < 0) {
2403             error_setg_errno(errp, errno, "mincore failed");
2404             break;
2405         }
2406 
2407         vec_end = DIV_ROUND_UP(length, page_size);
2408         for (i = 0; i < vec_end; i++) {
2409             if (vec[i] & 0x1) {
2410                 error_setg(errp, "page cache still in use!");
2411                 break;
2412             }
2413         }
2414     }
2415 
2416     if (window) {
2417         munmap(window, length);
2418     }
2419 
2420     g_free(vec);
2421 }
2422 #endif /* __linux__ */
2423 
2424 static void coroutine_fn raw_co_invalidate_cache(BlockDriverState *bs,
2425                                                  Error **errp)
2426 {
2427     BDRVRawState *s = bs->opaque;
2428     int ret;
2429 
2430     ret = fd_open(bs);
2431     if (ret < 0) {
2432         error_setg_errno(errp, -ret, "The file descriptor is not open");
2433         return;
2434     }
2435 
2436     if (s->open_flags & O_DIRECT) {
2437         return; /* No host kernel page cache */
2438     }
2439 
2440 #if defined(__linux__)
2441     /* This sets the scene for the next syscall... */
2442     ret = bdrv_co_flush(bs);
2443     if (ret < 0) {
2444         error_setg_errno(errp, -ret, "flush failed");
2445         return;
2446     }
2447 
2448     /* Linux does not invalidate pages that are dirty, locked, or mmapped by a
2449      * process.  These limitations are okay because we just fsynced the file,
2450      * we don't use mmap, and the file should not be in use by other processes.
2451      */
2452     ret = posix_fadvise(s->fd, 0, 0, POSIX_FADV_DONTNEED);
2453     if (ret != 0) { /* the return value is a positive errno */
2454         error_setg_errno(errp, ret, "fadvise failed");
2455         return;
2456     }
2457 
2458     if (s->check_cache_dropped) {
2459         check_cache_dropped(bs, errp);
2460     }
2461 #else /* __linux__ */
2462     /* Do nothing.  Live migration to a remote host with cache.direct=off is
2463      * unsupported on other host operating systems.  Cache consistency issues
2464      * may occur but no error is reported here, partly because that's the
2465      * historical behavior and partly because it's hard to differentiate valid
2466      * configurations that should not cause errors.
2467      */
2468 #endif /* !__linux__ */
2469 }
2470 
2471 static coroutine_fn BlockAIOCB *raw_aio_pdiscard(BlockDriverState *bs,
2472     int64_t offset, int bytes,
2473     BlockCompletionFunc *cb, void *opaque)
2474 {
2475     BDRVRawState *s = bs->opaque;
2476 
2477     return paio_submit(bs, s->fd, offset, NULL, bytes,
2478                        cb, opaque, QEMU_AIO_DISCARD);
2479 }
2480 
2481 static int coroutine_fn raw_co_pwrite_zeroes(
2482     BlockDriverState *bs, int64_t offset,
2483     int bytes, BdrvRequestFlags flags)
2484 {
2485     BDRVRawState *s = bs->opaque;
2486 
2487     if (!(flags & BDRV_REQ_MAY_UNMAP)) {
2488         return paio_submit_co(bs, s->fd, offset, NULL, bytes,
2489                               QEMU_AIO_WRITE_ZEROES);
2490     } else if (s->discard_zeroes) {
2491         return paio_submit_co(bs, s->fd, offset, NULL, bytes,
2492                               QEMU_AIO_DISCARD);
2493     }
2494     return -ENOTSUP;
2495 }
2496 
2497 static int raw_get_info(BlockDriverState *bs, BlockDriverInfo *bdi)
2498 {
2499     BDRVRawState *s = bs->opaque;
2500 
2501     bdi->unallocated_blocks_are_zero = s->discard_zeroes;
2502     return 0;
2503 }
2504 
2505 static QemuOptsList raw_create_opts = {
2506     .name = "raw-create-opts",
2507     .head = QTAILQ_HEAD_INITIALIZER(raw_create_opts.head),
2508     .desc = {
2509         {
2510             .name = BLOCK_OPT_SIZE,
2511             .type = QEMU_OPT_SIZE,
2512             .help = "Virtual disk size"
2513         },
2514         {
2515             .name = BLOCK_OPT_NOCOW,
2516             .type = QEMU_OPT_BOOL,
2517             .help = "Turn off copy-on-write (valid only on btrfs)"
2518         },
2519         {
2520             .name = BLOCK_OPT_PREALLOC,
2521             .type = QEMU_OPT_STRING,
2522             .help = "Preallocation mode (allowed values: off, falloc, full)"
2523         },
2524         { /* end of list */ }
2525     }
2526 };
2527 
2528 static int raw_check_perm(BlockDriverState *bs, uint64_t perm, uint64_t shared,
2529                           Error **errp)
2530 {
2531     return raw_handle_perm_lock(bs, RAW_PL_PREPARE, perm, shared, errp);
2532 }
2533 
2534 static void raw_set_perm(BlockDriverState *bs, uint64_t perm, uint64_t shared)
2535 {
2536     BDRVRawState *s = bs->opaque;
2537     raw_handle_perm_lock(bs, RAW_PL_COMMIT, perm, shared, NULL);
2538     s->perm = perm;
2539     s->shared_perm = shared;
2540 }
2541 
2542 static void raw_abort_perm_update(BlockDriverState *bs)
2543 {
2544     raw_handle_perm_lock(bs, RAW_PL_ABORT, 0, 0, NULL);
2545 }
2546 
2547 static int coroutine_fn raw_co_copy_range_from(BlockDriverState *bs,
2548                                                BdrvChild *src, uint64_t src_offset,
2549                                                BdrvChild *dst, uint64_t dst_offset,
2550                                                uint64_t bytes, BdrvRequestFlags flags)
2551 {
2552     return bdrv_co_copy_range_to(src, src_offset, dst, dst_offset, bytes, flags);
2553 }
2554 
2555 static int coroutine_fn raw_co_copy_range_to(BlockDriverState *bs,
2556                                              BdrvChild *src, uint64_t src_offset,
2557                                              BdrvChild *dst, uint64_t dst_offset,
2558                                              uint64_t bytes, BdrvRequestFlags flags)
2559 {
2560     BDRVRawState *s = bs->opaque;
2561     BDRVRawState *src_s;
2562 
2563     assert(dst->bs == bs);
2564     if (src->bs->drv->bdrv_co_copy_range_to != raw_co_copy_range_to) {
2565         return -ENOTSUP;
2566     }
2567 
2568     src_s = src->bs->opaque;
2569     if (fd_open(bs) < 0 || fd_open(bs) < 0) {
2570         return -EIO;
2571     }
2572     return paio_submit_co_full(bs, src_s->fd, src_offset, s->fd, dst_offset,
2573                                NULL, bytes, QEMU_AIO_COPY_RANGE);
2574 }
2575 
2576 BlockDriver bdrv_file = {
2577     .format_name = "file",
2578     .protocol_name = "file",
2579     .instance_size = sizeof(BDRVRawState),
2580     .bdrv_needs_filename = true,
2581     .bdrv_probe = NULL, /* no probe for protocols */
2582     .bdrv_parse_filename = raw_parse_filename,
2583     .bdrv_file_open = raw_open,
2584     .bdrv_reopen_prepare = raw_reopen_prepare,
2585     .bdrv_reopen_commit = raw_reopen_commit,
2586     .bdrv_reopen_abort = raw_reopen_abort,
2587     .bdrv_close = raw_close,
2588     .bdrv_co_create = raw_co_create,
2589     .bdrv_co_create_opts = raw_co_create_opts,
2590     .bdrv_has_zero_init = bdrv_has_zero_init_1,
2591     .bdrv_co_block_status = raw_co_block_status,
2592     .bdrv_co_invalidate_cache = raw_co_invalidate_cache,
2593     .bdrv_co_pwrite_zeroes = raw_co_pwrite_zeroes,
2594 
2595     .bdrv_co_preadv         = raw_co_preadv,
2596     .bdrv_co_pwritev        = raw_co_pwritev,
2597     .bdrv_aio_flush = raw_aio_flush,
2598     .bdrv_aio_pdiscard = raw_aio_pdiscard,
2599     .bdrv_co_copy_range_from = raw_co_copy_range_from,
2600     .bdrv_co_copy_range_to  = raw_co_copy_range_to,
2601     .bdrv_refresh_limits = raw_refresh_limits,
2602     .bdrv_io_plug = raw_aio_plug,
2603     .bdrv_io_unplug = raw_aio_unplug,
2604 
2605     .bdrv_truncate = raw_truncate,
2606     .bdrv_getlength = raw_getlength,
2607     .bdrv_get_info = raw_get_info,
2608     .bdrv_get_allocated_file_size
2609                         = raw_get_allocated_file_size,
2610     .bdrv_check_perm = raw_check_perm,
2611     .bdrv_set_perm   = raw_set_perm,
2612     .bdrv_abort_perm_update = raw_abort_perm_update,
2613     .create_opts = &raw_create_opts,
2614 };
2615 
2616 /***********************************************/
2617 /* host device */
2618 
2619 #if defined(__APPLE__) && defined(__MACH__)
2620 static kern_return_t GetBSDPath(io_iterator_t mediaIterator, char *bsdPath,
2621                                 CFIndex maxPathSize, int flags);
2622 static char *FindEjectableOpticalMedia(io_iterator_t *mediaIterator)
2623 {
2624     kern_return_t kernResult = KERN_FAILURE;
2625     mach_port_t     masterPort;
2626     CFMutableDictionaryRef  classesToMatch;
2627     const char *matching_array[] = {kIODVDMediaClass, kIOCDMediaClass};
2628     char *mediaType = NULL;
2629 
2630     kernResult = IOMasterPort( MACH_PORT_NULL, &masterPort );
2631     if ( KERN_SUCCESS != kernResult ) {
2632         printf( "IOMasterPort returned %d\n", kernResult );
2633     }
2634 
2635     int index;
2636     for (index = 0; index < ARRAY_SIZE(matching_array); index++) {
2637         classesToMatch = IOServiceMatching(matching_array[index]);
2638         if (classesToMatch == NULL) {
2639             error_report("IOServiceMatching returned NULL for %s",
2640                          matching_array[index]);
2641             continue;
2642         }
2643         CFDictionarySetValue(classesToMatch, CFSTR(kIOMediaEjectableKey),
2644                              kCFBooleanTrue);
2645         kernResult = IOServiceGetMatchingServices(masterPort, classesToMatch,
2646                                                   mediaIterator);
2647         if (kernResult != KERN_SUCCESS) {
2648             error_report("Note: IOServiceGetMatchingServices returned %d",
2649                          kernResult);
2650             continue;
2651         }
2652 
2653         /* If a match was found, leave the loop */
2654         if (*mediaIterator != 0) {
2655             DPRINTF("Matching using %s\n", matching_array[index]);
2656             mediaType = g_strdup(matching_array[index]);
2657             break;
2658         }
2659     }
2660     return mediaType;
2661 }
2662 
2663 kern_return_t GetBSDPath(io_iterator_t mediaIterator, char *bsdPath,
2664                          CFIndex maxPathSize, int flags)
2665 {
2666     io_object_t     nextMedia;
2667     kern_return_t   kernResult = KERN_FAILURE;
2668     *bsdPath = '\0';
2669     nextMedia = IOIteratorNext( mediaIterator );
2670     if ( nextMedia )
2671     {
2672         CFTypeRef   bsdPathAsCFString;
2673     bsdPathAsCFString = IORegistryEntryCreateCFProperty( nextMedia, CFSTR( kIOBSDNameKey ), kCFAllocatorDefault, 0 );
2674         if ( bsdPathAsCFString ) {
2675             size_t devPathLength;
2676             strcpy( bsdPath, _PATH_DEV );
2677             if (flags & BDRV_O_NOCACHE) {
2678                 strcat(bsdPath, "r");
2679             }
2680             devPathLength = strlen( bsdPath );
2681             if ( CFStringGetCString( bsdPathAsCFString, bsdPath + devPathLength, maxPathSize - devPathLength, kCFStringEncodingASCII ) ) {
2682                 kernResult = KERN_SUCCESS;
2683             }
2684             CFRelease( bsdPathAsCFString );
2685         }
2686         IOObjectRelease( nextMedia );
2687     }
2688 
2689     return kernResult;
2690 }
2691 
2692 /* Sets up a real cdrom for use in QEMU */
2693 static bool setup_cdrom(char *bsd_path, Error **errp)
2694 {
2695     int index, num_of_test_partitions = 2, fd;
2696     char test_partition[MAXPATHLEN];
2697     bool partition_found = false;
2698 
2699     /* look for a working partition */
2700     for (index = 0; index < num_of_test_partitions; index++) {
2701         snprintf(test_partition, sizeof(test_partition), "%ss%d", bsd_path,
2702                  index);
2703         fd = qemu_open(test_partition, O_RDONLY | O_BINARY | O_LARGEFILE);
2704         if (fd >= 0) {
2705             partition_found = true;
2706             qemu_close(fd);
2707             break;
2708         }
2709     }
2710 
2711     /* if a working partition on the device was not found */
2712     if (partition_found == false) {
2713         error_setg(errp, "Failed to find a working partition on disc");
2714     } else {
2715         DPRINTF("Using %s as optical disc\n", test_partition);
2716         pstrcpy(bsd_path, MAXPATHLEN, test_partition);
2717     }
2718     return partition_found;
2719 }
2720 
2721 /* Prints directions on mounting and unmounting a device */
2722 static void print_unmounting_directions(const char *file_name)
2723 {
2724     error_report("If device %s is mounted on the desktop, unmount"
2725                  " it first before using it in QEMU", file_name);
2726     error_report("Command to unmount device: diskutil unmountDisk %s",
2727                  file_name);
2728     error_report("Command to mount device: diskutil mountDisk %s", file_name);
2729 }
2730 
2731 #endif /* defined(__APPLE__) && defined(__MACH__) */
2732 
2733 static int hdev_probe_device(const char *filename)
2734 {
2735     struct stat st;
2736 
2737     /* allow a dedicated CD-ROM driver to match with a higher priority */
2738     if (strstart(filename, "/dev/cdrom", NULL))
2739         return 50;
2740 
2741     if (stat(filename, &st) >= 0 &&
2742             (S_ISCHR(st.st_mode) || S_ISBLK(st.st_mode))) {
2743         return 100;
2744     }
2745 
2746     return 0;
2747 }
2748 
2749 static int check_hdev_writable(BDRVRawState *s)
2750 {
2751 #if defined(BLKROGET)
2752     /* Linux block devices can be configured "read-only" using blockdev(8).
2753      * This is independent of device node permissions and therefore open(2)
2754      * with O_RDWR succeeds.  Actual writes fail with EPERM.
2755      *
2756      * bdrv_open() is supposed to fail if the disk is read-only.  Explicitly
2757      * check for read-only block devices so that Linux block devices behave
2758      * properly.
2759      */
2760     struct stat st;
2761     int readonly = 0;
2762 
2763     if (fstat(s->fd, &st)) {
2764         return -errno;
2765     }
2766 
2767     if (!S_ISBLK(st.st_mode)) {
2768         return 0;
2769     }
2770 
2771     if (ioctl(s->fd, BLKROGET, &readonly) < 0) {
2772         return -errno;
2773     }
2774 
2775     if (readonly) {
2776         return -EACCES;
2777     }
2778 #endif /* defined(BLKROGET) */
2779     return 0;
2780 }
2781 
2782 static void hdev_parse_filename(const char *filename, QDict *options,
2783                                 Error **errp)
2784 {
2785     bdrv_parse_filename_strip_prefix(filename, "host_device:", options);
2786 }
2787 
2788 static bool hdev_is_sg(BlockDriverState *bs)
2789 {
2790 
2791 #if defined(__linux__)
2792 
2793     BDRVRawState *s = bs->opaque;
2794     struct stat st;
2795     struct sg_scsi_id scsiid;
2796     int sg_version;
2797     int ret;
2798 
2799     if (stat(bs->filename, &st) < 0 || !S_ISCHR(st.st_mode)) {
2800         return false;
2801     }
2802 
2803     ret = ioctl(s->fd, SG_GET_VERSION_NUM, &sg_version);
2804     if (ret < 0) {
2805         return false;
2806     }
2807 
2808     ret = ioctl(s->fd, SG_GET_SCSI_ID, &scsiid);
2809     if (ret >= 0) {
2810         DPRINTF("SG device found: type=%d, version=%d\n",
2811             scsiid.scsi_type, sg_version);
2812         return true;
2813     }
2814 
2815 #endif
2816 
2817     return false;
2818 }
2819 
2820 static int hdev_open(BlockDriverState *bs, QDict *options, int flags,
2821                      Error **errp)
2822 {
2823     BDRVRawState *s = bs->opaque;
2824     Error *local_err = NULL;
2825     int ret;
2826 
2827 #if defined(__APPLE__) && defined(__MACH__)
2828     /*
2829      * Caution: while qdict_get_str() is fine, getting non-string types
2830      * would require more care.  When @options come from -blockdev or
2831      * blockdev_add, its members are typed according to the QAPI
2832      * schema, but when they come from -drive, they're all QString.
2833      */
2834     const char *filename = qdict_get_str(options, "filename");
2835     char bsd_path[MAXPATHLEN] = "";
2836     bool error_occurred = false;
2837 
2838     /* If using a real cdrom */
2839     if (strcmp(filename, "/dev/cdrom") == 0) {
2840         char *mediaType = NULL;
2841         kern_return_t ret_val;
2842         io_iterator_t mediaIterator = 0;
2843 
2844         mediaType = FindEjectableOpticalMedia(&mediaIterator);
2845         if (mediaType == NULL) {
2846             error_setg(errp, "Please make sure your CD/DVD is in the optical"
2847                        " drive");
2848             error_occurred = true;
2849             goto hdev_open_Mac_error;
2850         }
2851 
2852         ret_val = GetBSDPath(mediaIterator, bsd_path, sizeof(bsd_path), flags);
2853         if (ret_val != KERN_SUCCESS) {
2854             error_setg(errp, "Could not get BSD path for optical drive");
2855             error_occurred = true;
2856             goto hdev_open_Mac_error;
2857         }
2858 
2859         /* If a real optical drive was not found */
2860         if (bsd_path[0] == '\0') {
2861             error_setg(errp, "Failed to obtain bsd path for optical drive");
2862             error_occurred = true;
2863             goto hdev_open_Mac_error;
2864         }
2865 
2866         /* If using a cdrom disc and finding a partition on the disc failed */
2867         if (strncmp(mediaType, kIOCDMediaClass, 9) == 0 &&
2868             setup_cdrom(bsd_path, errp) == false) {
2869             print_unmounting_directions(bsd_path);
2870             error_occurred = true;
2871             goto hdev_open_Mac_error;
2872         }
2873 
2874         qdict_put_str(options, "filename", bsd_path);
2875 
2876 hdev_open_Mac_error:
2877         g_free(mediaType);
2878         if (mediaIterator) {
2879             IOObjectRelease(mediaIterator);
2880         }
2881         if (error_occurred) {
2882             return -ENOENT;
2883         }
2884     }
2885 #endif /* defined(__APPLE__) && defined(__MACH__) */
2886 
2887     s->type = FTYPE_FILE;
2888 
2889     ret = raw_open_common(bs, options, flags, 0, &local_err);
2890     if (ret < 0) {
2891         error_propagate(errp, local_err);
2892 #if defined(__APPLE__) && defined(__MACH__)
2893         if (*bsd_path) {
2894             filename = bsd_path;
2895         }
2896         /* if a physical device experienced an error while being opened */
2897         if (strncmp(filename, "/dev/", 5) == 0) {
2898             print_unmounting_directions(filename);
2899         }
2900 #endif /* defined(__APPLE__) && defined(__MACH__) */
2901         return ret;
2902     }
2903 
2904     /* Since this does ioctl the device must be already opened */
2905     bs->sg = hdev_is_sg(bs);
2906 
2907     if (flags & BDRV_O_RDWR) {
2908         ret = check_hdev_writable(s);
2909         if (ret < 0) {
2910             raw_close(bs);
2911             error_setg_errno(errp, -ret, "The device is not writable");
2912             return ret;
2913         }
2914     }
2915 
2916     return ret;
2917 }
2918 
2919 #if defined(__linux__)
2920 
2921 static BlockAIOCB *hdev_aio_ioctl(BlockDriverState *bs,
2922         unsigned long int req, void *buf,
2923         BlockCompletionFunc *cb, void *opaque)
2924 {
2925     BDRVRawState *s = bs->opaque;
2926     RawPosixAIOData *acb;
2927     ThreadPool *pool;
2928 
2929     if (fd_open(bs) < 0)
2930         return NULL;
2931 
2932     if (req == SG_IO && s->pr_mgr) {
2933         struct sg_io_hdr *io_hdr = buf;
2934         if (io_hdr->cmdp[0] == PERSISTENT_RESERVE_OUT ||
2935             io_hdr->cmdp[0] == PERSISTENT_RESERVE_IN) {
2936             return pr_manager_execute(s->pr_mgr, bdrv_get_aio_context(bs),
2937                                       s->fd, io_hdr, cb, opaque);
2938         }
2939     }
2940 
2941     acb = g_new(RawPosixAIOData, 1);
2942     acb->bs = bs;
2943     acb->aio_type = QEMU_AIO_IOCTL;
2944     acb->aio_fildes = s->fd;
2945     acb->aio_offset = 0;
2946     acb->aio_ioctl_buf = buf;
2947     acb->aio_ioctl_cmd = req;
2948     pool = aio_get_thread_pool(bdrv_get_aio_context(bs));
2949     return thread_pool_submit_aio(pool, aio_worker, acb, cb, opaque);
2950 }
2951 #endif /* linux */
2952 
2953 static int fd_open(BlockDriverState *bs)
2954 {
2955     BDRVRawState *s = bs->opaque;
2956 
2957     /* this is just to ensure s->fd is sane (its called by io ops) */
2958     if (s->fd >= 0)
2959         return 0;
2960     return -EIO;
2961 }
2962 
2963 static coroutine_fn BlockAIOCB *hdev_aio_pdiscard(BlockDriverState *bs,
2964     int64_t offset, int bytes,
2965     BlockCompletionFunc *cb, void *opaque)
2966 {
2967     BDRVRawState *s = bs->opaque;
2968 
2969     if (fd_open(bs) < 0) {
2970         return NULL;
2971     }
2972     return paio_submit(bs, s->fd, offset, NULL, bytes,
2973                        cb, opaque, QEMU_AIO_DISCARD|QEMU_AIO_BLKDEV);
2974 }
2975 
2976 static coroutine_fn int hdev_co_pwrite_zeroes(BlockDriverState *bs,
2977     int64_t offset, int bytes, BdrvRequestFlags flags)
2978 {
2979     BDRVRawState *s = bs->opaque;
2980     int rc;
2981 
2982     rc = fd_open(bs);
2983     if (rc < 0) {
2984         return rc;
2985     }
2986     if (!(flags & BDRV_REQ_MAY_UNMAP)) {
2987         return paio_submit_co(bs, s->fd, offset, NULL, bytes,
2988                               QEMU_AIO_WRITE_ZEROES|QEMU_AIO_BLKDEV);
2989     } else if (s->discard_zeroes) {
2990         return paio_submit_co(bs, s->fd, offset, NULL, bytes,
2991                               QEMU_AIO_DISCARD|QEMU_AIO_BLKDEV);
2992     }
2993     return -ENOTSUP;
2994 }
2995 
2996 static int coroutine_fn hdev_co_create_opts(const char *filename, QemuOpts *opts,
2997                                             Error **errp)
2998 {
2999     int fd;
3000     int ret = 0;
3001     struct stat stat_buf;
3002     int64_t total_size = 0;
3003     bool has_prefix;
3004 
3005     /* This function is used by both protocol block drivers and therefore either
3006      * of these prefixes may be given.
3007      * The return value has to be stored somewhere, otherwise this is an error
3008      * due to -Werror=unused-value. */
3009     has_prefix =
3010         strstart(filename, "host_device:", &filename) ||
3011         strstart(filename, "host_cdrom:" , &filename);
3012 
3013     (void)has_prefix;
3014 
3015     ret = raw_normalize_devicepath(&filename);
3016     if (ret < 0) {
3017         error_setg_errno(errp, -ret, "Could not normalize device path");
3018         return ret;
3019     }
3020 
3021     /* Read out options */
3022     total_size = ROUND_UP(qemu_opt_get_size_del(opts, BLOCK_OPT_SIZE, 0),
3023                           BDRV_SECTOR_SIZE);
3024 
3025     fd = qemu_open(filename, O_WRONLY | O_BINARY);
3026     if (fd < 0) {
3027         ret = -errno;
3028         error_setg_errno(errp, -ret, "Could not open device");
3029         return ret;
3030     }
3031 
3032     if (fstat(fd, &stat_buf) < 0) {
3033         ret = -errno;
3034         error_setg_errno(errp, -ret, "Could not stat device");
3035     } else if (!S_ISBLK(stat_buf.st_mode) && !S_ISCHR(stat_buf.st_mode)) {
3036         error_setg(errp,
3037                    "The given file is neither a block nor a character device");
3038         ret = -ENODEV;
3039     } else if (lseek(fd, 0, SEEK_END) < total_size) {
3040         error_setg(errp, "Device is too small");
3041         ret = -ENOSPC;
3042     }
3043 
3044     if (!ret && total_size) {
3045         uint8_t buf[BDRV_SECTOR_SIZE] = { 0 };
3046         int64_t zero_size = MIN(BDRV_SECTOR_SIZE, total_size);
3047         if (lseek(fd, 0, SEEK_SET) == -1) {
3048             ret = -errno;
3049         } else {
3050             ret = qemu_write_full(fd, buf, zero_size);
3051             ret = ret == zero_size ? 0 : -errno;
3052         }
3053     }
3054     qemu_close(fd);
3055     return ret;
3056 }
3057 
3058 static BlockDriver bdrv_host_device = {
3059     .format_name        = "host_device",
3060     .protocol_name        = "host_device",
3061     .instance_size      = sizeof(BDRVRawState),
3062     .bdrv_needs_filename = true,
3063     .bdrv_probe_device  = hdev_probe_device,
3064     .bdrv_parse_filename = hdev_parse_filename,
3065     .bdrv_file_open     = hdev_open,
3066     .bdrv_close         = raw_close,
3067     .bdrv_reopen_prepare = raw_reopen_prepare,
3068     .bdrv_reopen_commit  = raw_reopen_commit,
3069     .bdrv_reopen_abort   = raw_reopen_abort,
3070     .bdrv_co_create_opts = hdev_co_create_opts,
3071     .create_opts         = &raw_create_opts,
3072     .bdrv_co_invalidate_cache = raw_co_invalidate_cache,
3073     .bdrv_co_pwrite_zeroes = hdev_co_pwrite_zeroes,
3074 
3075     .bdrv_co_preadv         = raw_co_preadv,
3076     .bdrv_co_pwritev        = raw_co_pwritev,
3077     .bdrv_aio_flush	= raw_aio_flush,
3078     .bdrv_aio_pdiscard   = hdev_aio_pdiscard,
3079     .bdrv_co_copy_range_from = raw_co_copy_range_from,
3080     .bdrv_co_copy_range_to  = raw_co_copy_range_to,
3081     .bdrv_refresh_limits = raw_refresh_limits,
3082     .bdrv_io_plug = raw_aio_plug,
3083     .bdrv_io_unplug = raw_aio_unplug,
3084 
3085     .bdrv_truncate      = raw_truncate,
3086     .bdrv_getlength	= raw_getlength,
3087     .bdrv_get_info = raw_get_info,
3088     .bdrv_get_allocated_file_size
3089                         = raw_get_allocated_file_size,
3090     .bdrv_check_perm = raw_check_perm,
3091     .bdrv_set_perm   = raw_set_perm,
3092     .bdrv_abort_perm_update = raw_abort_perm_update,
3093     .bdrv_probe_blocksizes = hdev_probe_blocksizes,
3094     .bdrv_probe_geometry = hdev_probe_geometry,
3095 
3096     /* generic scsi device */
3097 #ifdef __linux__
3098     .bdrv_aio_ioctl     = hdev_aio_ioctl,
3099 #endif
3100 };
3101 
3102 #if defined(__linux__) || defined(__FreeBSD__) || defined(__FreeBSD_kernel__)
3103 static void cdrom_parse_filename(const char *filename, QDict *options,
3104                                  Error **errp)
3105 {
3106     bdrv_parse_filename_strip_prefix(filename, "host_cdrom:", options);
3107 }
3108 #endif
3109 
3110 #ifdef __linux__
3111 static int cdrom_open(BlockDriverState *bs, QDict *options, int flags,
3112                       Error **errp)
3113 {
3114     BDRVRawState *s = bs->opaque;
3115 
3116     s->type = FTYPE_CD;
3117 
3118     /* open will not fail even if no CD is inserted, so add O_NONBLOCK */
3119     return raw_open_common(bs, options, flags, O_NONBLOCK, errp);
3120 }
3121 
3122 static int cdrom_probe_device(const char *filename)
3123 {
3124     int fd, ret;
3125     int prio = 0;
3126     struct stat st;
3127 
3128     fd = qemu_open(filename, O_RDONLY | O_NONBLOCK);
3129     if (fd < 0) {
3130         goto out;
3131     }
3132     ret = fstat(fd, &st);
3133     if (ret == -1 || !S_ISBLK(st.st_mode)) {
3134         goto outc;
3135     }
3136 
3137     /* Attempt to detect via a CDROM specific ioctl */
3138     ret = ioctl(fd, CDROM_DRIVE_STATUS, CDSL_CURRENT);
3139     if (ret >= 0)
3140         prio = 100;
3141 
3142 outc:
3143     qemu_close(fd);
3144 out:
3145     return prio;
3146 }
3147 
3148 static bool cdrom_is_inserted(BlockDriverState *bs)
3149 {
3150     BDRVRawState *s = bs->opaque;
3151     int ret;
3152 
3153     ret = ioctl(s->fd, CDROM_DRIVE_STATUS, CDSL_CURRENT);
3154     return ret == CDS_DISC_OK;
3155 }
3156 
3157 static void cdrom_eject(BlockDriverState *bs, bool eject_flag)
3158 {
3159     BDRVRawState *s = bs->opaque;
3160 
3161     if (eject_flag) {
3162         if (ioctl(s->fd, CDROMEJECT, NULL) < 0)
3163             perror("CDROMEJECT");
3164     } else {
3165         if (ioctl(s->fd, CDROMCLOSETRAY, NULL) < 0)
3166             perror("CDROMEJECT");
3167     }
3168 }
3169 
3170 static void cdrom_lock_medium(BlockDriverState *bs, bool locked)
3171 {
3172     BDRVRawState *s = bs->opaque;
3173 
3174     if (ioctl(s->fd, CDROM_LOCKDOOR, locked) < 0) {
3175         /*
3176          * Note: an error can happen if the distribution automatically
3177          * mounts the CD-ROM
3178          */
3179         /* perror("CDROM_LOCKDOOR"); */
3180     }
3181 }
3182 
3183 static BlockDriver bdrv_host_cdrom = {
3184     .format_name        = "host_cdrom",
3185     .protocol_name      = "host_cdrom",
3186     .instance_size      = sizeof(BDRVRawState),
3187     .bdrv_needs_filename = true,
3188     .bdrv_probe_device	= cdrom_probe_device,
3189     .bdrv_parse_filename = cdrom_parse_filename,
3190     .bdrv_file_open     = cdrom_open,
3191     .bdrv_close         = raw_close,
3192     .bdrv_reopen_prepare = raw_reopen_prepare,
3193     .bdrv_reopen_commit  = raw_reopen_commit,
3194     .bdrv_reopen_abort   = raw_reopen_abort,
3195     .bdrv_co_create_opts = hdev_co_create_opts,
3196     .create_opts         = &raw_create_opts,
3197     .bdrv_co_invalidate_cache = raw_co_invalidate_cache,
3198 
3199 
3200     .bdrv_co_preadv         = raw_co_preadv,
3201     .bdrv_co_pwritev        = raw_co_pwritev,
3202     .bdrv_aio_flush	= raw_aio_flush,
3203     .bdrv_refresh_limits = raw_refresh_limits,
3204     .bdrv_io_plug = raw_aio_plug,
3205     .bdrv_io_unplug = raw_aio_unplug,
3206 
3207     .bdrv_truncate      = raw_truncate,
3208     .bdrv_getlength      = raw_getlength,
3209     .has_variable_length = true,
3210     .bdrv_get_allocated_file_size
3211                         = raw_get_allocated_file_size,
3212 
3213     /* removable device support */
3214     .bdrv_is_inserted   = cdrom_is_inserted,
3215     .bdrv_eject         = cdrom_eject,
3216     .bdrv_lock_medium   = cdrom_lock_medium,
3217 
3218     /* generic scsi device */
3219     .bdrv_aio_ioctl     = hdev_aio_ioctl,
3220 };
3221 #endif /* __linux__ */
3222 
3223 #if defined (__FreeBSD__) || defined(__FreeBSD_kernel__)
3224 static int cdrom_open(BlockDriverState *bs, QDict *options, int flags,
3225                       Error **errp)
3226 {
3227     BDRVRawState *s = bs->opaque;
3228     Error *local_err = NULL;
3229     int ret;
3230 
3231     s->type = FTYPE_CD;
3232 
3233     ret = raw_open_common(bs, options, flags, 0, &local_err);
3234     if (ret) {
3235         error_propagate(errp, local_err);
3236         return ret;
3237     }
3238 
3239     /* make sure the door isn't locked at this time */
3240     ioctl(s->fd, CDIOCALLOW);
3241     return 0;
3242 }
3243 
3244 static int cdrom_probe_device(const char *filename)
3245 {
3246     if (strstart(filename, "/dev/cd", NULL) ||
3247             strstart(filename, "/dev/acd", NULL))
3248         return 100;
3249     return 0;
3250 }
3251 
3252 static int cdrom_reopen(BlockDriverState *bs)
3253 {
3254     BDRVRawState *s = bs->opaque;
3255     int fd;
3256 
3257     /*
3258      * Force reread of possibly changed/newly loaded disc,
3259      * FreeBSD seems to not notice sometimes...
3260      */
3261     if (s->fd >= 0)
3262         qemu_close(s->fd);
3263     fd = qemu_open(bs->filename, s->open_flags, 0644);
3264     if (fd < 0) {
3265         s->fd = -1;
3266         return -EIO;
3267     }
3268     s->fd = fd;
3269 
3270     /* make sure the door isn't locked at this time */
3271     ioctl(s->fd, CDIOCALLOW);
3272     return 0;
3273 }
3274 
3275 static bool cdrom_is_inserted(BlockDriverState *bs)
3276 {
3277     return raw_getlength(bs) > 0;
3278 }
3279 
3280 static void cdrom_eject(BlockDriverState *bs, bool eject_flag)
3281 {
3282     BDRVRawState *s = bs->opaque;
3283 
3284     if (s->fd < 0)
3285         return;
3286 
3287     (void) ioctl(s->fd, CDIOCALLOW);
3288 
3289     if (eject_flag) {
3290         if (ioctl(s->fd, CDIOCEJECT) < 0)
3291             perror("CDIOCEJECT");
3292     } else {
3293         if (ioctl(s->fd, CDIOCCLOSE) < 0)
3294             perror("CDIOCCLOSE");
3295     }
3296 
3297     cdrom_reopen(bs);
3298 }
3299 
3300 static void cdrom_lock_medium(BlockDriverState *bs, bool locked)
3301 {
3302     BDRVRawState *s = bs->opaque;
3303 
3304     if (s->fd < 0)
3305         return;
3306     if (ioctl(s->fd, (locked ? CDIOCPREVENT : CDIOCALLOW)) < 0) {
3307         /*
3308          * Note: an error can happen if the distribution automatically
3309          * mounts the CD-ROM
3310          */
3311         /* perror("CDROM_LOCKDOOR"); */
3312     }
3313 }
3314 
3315 static BlockDriver bdrv_host_cdrom = {
3316     .format_name        = "host_cdrom",
3317     .protocol_name      = "host_cdrom",
3318     .instance_size      = sizeof(BDRVRawState),
3319     .bdrv_needs_filename = true,
3320     .bdrv_probe_device	= cdrom_probe_device,
3321     .bdrv_parse_filename = cdrom_parse_filename,
3322     .bdrv_file_open     = cdrom_open,
3323     .bdrv_close         = raw_close,
3324     .bdrv_reopen_prepare = raw_reopen_prepare,
3325     .bdrv_reopen_commit  = raw_reopen_commit,
3326     .bdrv_reopen_abort   = raw_reopen_abort,
3327     .bdrv_co_create_opts = hdev_co_create_opts,
3328     .create_opts        = &raw_create_opts,
3329 
3330     .bdrv_co_preadv         = raw_co_preadv,
3331     .bdrv_co_pwritev        = raw_co_pwritev,
3332     .bdrv_aio_flush	= raw_aio_flush,
3333     .bdrv_refresh_limits = raw_refresh_limits,
3334     .bdrv_io_plug = raw_aio_plug,
3335     .bdrv_io_unplug = raw_aio_unplug,
3336 
3337     .bdrv_truncate      = raw_truncate,
3338     .bdrv_getlength      = raw_getlength,
3339     .has_variable_length = true,
3340     .bdrv_get_allocated_file_size
3341                         = raw_get_allocated_file_size,
3342 
3343     /* removable device support */
3344     .bdrv_is_inserted   = cdrom_is_inserted,
3345     .bdrv_eject         = cdrom_eject,
3346     .bdrv_lock_medium   = cdrom_lock_medium,
3347 };
3348 #endif /* __FreeBSD__ */
3349 
3350 static void bdrv_file_init(void)
3351 {
3352     /*
3353      * Register all the drivers.  Note that order is important, the driver
3354      * registered last will get probed first.
3355      */
3356     bdrv_register(&bdrv_file);
3357     bdrv_register(&bdrv_host_device);
3358 #ifdef __linux__
3359     bdrv_register(&bdrv_host_cdrom);
3360 #endif
3361 #if defined(__FreeBSD__) || defined(__FreeBSD_kernel__)
3362     bdrv_register(&bdrv_host_cdrom);
3363 #endif
3364 }
3365 
3366 block_init(bdrv_file_init);
3367