xref: /openbmc/qemu/block/file-posix.c (revision e3a99063)
1 /*
2  * Block driver for RAW files (posix)
3  *
4  * Copyright (c) 2006 Fabrice Bellard
5  *
6  * Permission is hereby granted, free of charge, to any person obtaining a copy
7  * of this software and associated documentation files (the "Software"), to deal
8  * in the Software without restriction, including without limitation the rights
9  * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
10  * copies of the Software, and to permit persons to whom the Software is
11  * furnished to do so, subject to the following conditions:
12  *
13  * The above copyright notice and this permission notice shall be included in
14  * all copies or substantial portions of the Software.
15  *
16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
19  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
21  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
22  * THE SOFTWARE.
23  */
24 
25 #include "qemu/osdep.h"
26 #include "qemu-common.h"
27 #include "qapi/error.h"
28 #include "qemu/cutils.h"
29 #include "qemu/error-report.h"
30 #include "block/block_int.h"
31 #include "qemu/module.h"
32 #include "qemu/option.h"
33 #include "trace.h"
34 #include "block/thread-pool.h"
35 #include "qemu/iov.h"
36 #include "block/raw-aio.h"
37 #include "qapi/qmp/qdict.h"
38 #include "qapi/qmp/qstring.h"
39 
40 #include "scsi/pr-manager.h"
41 #include "scsi/constants.h"
42 
43 #if defined(__APPLE__) && (__MACH__)
44 #include <paths.h>
45 #include <sys/param.h>
46 #include <IOKit/IOKitLib.h>
47 #include <IOKit/IOBSD.h>
48 #include <IOKit/storage/IOMediaBSDClient.h>
49 #include <IOKit/storage/IOMedia.h>
50 #include <IOKit/storage/IOCDMedia.h>
51 //#include <IOKit/storage/IOCDTypes.h>
52 #include <IOKit/storage/IODVDMedia.h>
53 #include <CoreFoundation/CoreFoundation.h>
54 #endif
55 
56 #ifdef __sun__
57 #define _POSIX_PTHREAD_SEMANTICS 1
58 #include <sys/dkio.h>
59 #endif
60 #ifdef __linux__
61 #include <sys/ioctl.h>
62 #include <sys/param.h>
63 #include <sys/syscall.h>
64 #include <linux/cdrom.h>
65 #include <linux/fd.h>
66 #include <linux/fs.h>
67 #include <linux/hdreg.h>
68 #include <scsi/sg.h>
69 #ifdef __s390__
70 #include <asm/dasd.h>
71 #endif
72 #ifndef FS_NOCOW_FL
73 #define FS_NOCOW_FL                     0x00800000 /* Do not cow file */
74 #endif
75 #endif
76 #if defined(CONFIG_FALLOCATE_PUNCH_HOLE) || defined(CONFIG_FALLOCATE_ZERO_RANGE)
77 #include <linux/falloc.h>
78 #endif
79 #if defined (__FreeBSD__) || defined(__FreeBSD_kernel__)
80 #include <sys/disk.h>
81 #include <sys/cdio.h>
82 #endif
83 
84 #ifdef __OpenBSD__
85 #include <sys/ioctl.h>
86 #include <sys/disklabel.h>
87 #include <sys/dkio.h>
88 #endif
89 
90 #ifdef __NetBSD__
91 #include <sys/ioctl.h>
92 #include <sys/disklabel.h>
93 #include <sys/dkio.h>
94 #include <sys/disk.h>
95 #endif
96 
97 #ifdef __DragonFly__
98 #include <sys/ioctl.h>
99 #include <sys/diskslice.h>
100 #endif
101 
102 #ifdef CONFIG_XFS
103 #include <xfs/xfs.h>
104 #endif
105 
106 #include "trace.h"
107 
108 /* OS X does not have O_DSYNC */
109 #ifndef O_DSYNC
110 #ifdef O_SYNC
111 #define O_DSYNC O_SYNC
112 #elif defined(O_FSYNC)
113 #define O_DSYNC O_FSYNC
114 #endif
115 #endif
116 
117 /* Approximate O_DIRECT with O_DSYNC if O_DIRECT isn't available */
118 #ifndef O_DIRECT
119 #define O_DIRECT O_DSYNC
120 #endif
121 
122 #define FTYPE_FILE   0
123 #define FTYPE_CD     1
124 
125 #define MAX_BLOCKSIZE	4096
126 
127 /* Posix file locking bytes. Libvirt takes byte 0, we start from higher bytes,
128  * leaving a few more bytes for its future use. */
129 #define RAW_LOCK_PERM_BASE             100
130 #define RAW_LOCK_SHARED_BASE           200
131 
132 typedef struct BDRVRawState {
133     int fd;
134     bool use_lock;
135     int type;
136     int open_flags;
137     size_t buf_align;
138 
139     /* The current permissions. */
140     uint64_t perm;
141     uint64_t shared_perm;
142 
143     /* The perms bits whose corresponding bytes are already locked in
144      * s->fd. */
145     uint64_t locked_perm;
146     uint64_t locked_shared_perm;
147 
148     int perm_change_fd;
149     int perm_change_flags;
150     BDRVReopenState *reopen_state;
151 
152 #ifdef CONFIG_XFS
153     bool is_xfs:1;
154 #endif
155     bool has_discard:1;
156     bool has_write_zeroes:1;
157     bool discard_zeroes:1;
158     bool use_linux_aio:1;
159     bool use_linux_io_uring:1;
160     bool page_cache_inconsistent:1;
161     bool has_fallocate;
162     bool needs_alignment;
163     bool drop_cache;
164     bool check_cache_dropped;
165     struct {
166         uint64_t discard_nb_ok;
167         uint64_t discard_nb_failed;
168         uint64_t discard_bytes_ok;
169     } stats;
170 
171     PRManager *pr_mgr;
172 } BDRVRawState;
173 
174 typedef struct BDRVRawReopenState {
175     int fd;
176     int open_flags;
177     bool drop_cache;
178     bool check_cache_dropped;
179 } BDRVRawReopenState;
180 
181 static int fd_open(BlockDriverState *bs);
182 static int64_t raw_getlength(BlockDriverState *bs);
183 
184 typedef struct RawPosixAIOData {
185     BlockDriverState *bs;
186     int aio_type;
187     int aio_fildes;
188 
189     off_t aio_offset;
190     uint64_t aio_nbytes;
191 
192     union {
193         struct {
194             struct iovec *iov;
195             int niov;
196         } io;
197         struct {
198             uint64_t cmd;
199             void *buf;
200         } ioctl;
201         struct {
202             int aio_fd2;
203             off_t aio_offset2;
204         } copy_range;
205         struct {
206             PreallocMode prealloc;
207             Error **errp;
208         } truncate;
209     };
210 } RawPosixAIOData;
211 
212 #if defined(__FreeBSD__) || defined(__FreeBSD_kernel__)
213 static int cdrom_reopen(BlockDriverState *bs);
214 #endif
215 
216 #if defined(__NetBSD__)
217 static int raw_normalize_devicepath(const char **filename, Error **errp)
218 {
219     static char namebuf[PATH_MAX];
220     const char *dp, *fname;
221     struct stat sb;
222 
223     fname = *filename;
224     dp = strrchr(fname, '/');
225     if (lstat(fname, &sb) < 0) {
226         error_setg_file_open(errp, errno, fname);
227         return -errno;
228     }
229 
230     if (!S_ISBLK(sb.st_mode)) {
231         return 0;
232     }
233 
234     if (dp == NULL) {
235         snprintf(namebuf, PATH_MAX, "r%s", fname);
236     } else {
237         snprintf(namebuf, PATH_MAX, "%.*s/r%s",
238             (int)(dp - fname), fname, dp + 1);
239     }
240     *filename = namebuf;
241     warn_report("%s is a block device, using %s", fname, *filename);
242 
243     return 0;
244 }
245 #else
246 static int raw_normalize_devicepath(const char **filename, Error **errp)
247 {
248     return 0;
249 }
250 #endif
251 
252 /*
253  * Get logical block size via ioctl. On success store it in @sector_size_p.
254  */
255 static int probe_logical_blocksize(int fd, unsigned int *sector_size_p)
256 {
257     unsigned int sector_size;
258     bool success = false;
259     int i;
260 
261     errno = ENOTSUP;
262     static const unsigned long ioctl_list[] = {
263 #ifdef BLKSSZGET
264         BLKSSZGET,
265 #endif
266 #ifdef DKIOCGETBLOCKSIZE
267         DKIOCGETBLOCKSIZE,
268 #endif
269 #ifdef DIOCGSECTORSIZE
270         DIOCGSECTORSIZE,
271 #endif
272     };
273 
274     /* Try a few ioctls to get the right size */
275     for (i = 0; i < (int)ARRAY_SIZE(ioctl_list); i++) {
276         if (ioctl(fd, ioctl_list[i], &sector_size) >= 0) {
277             *sector_size_p = sector_size;
278             success = true;
279         }
280     }
281 
282     return success ? 0 : -errno;
283 }
284 
285 /**
286  * Get physical block size of @fd.
287  * On success, store it in @blk_size and return 0.
288  * On failure, return -errno.
289  */
290 static int probe_physical_blocksize(int fd, unsigned int *blk_size)
291 {
292 #ifdef BLKPBSZGET
293     if (ioctl(fd, BLKPBSZGET, blk_size) < 0) {
294         return -errno;
295     }
296     return 0;
297 #else
298     return -ENOTSUP;
299 #endif
300 }
301 
302 /* Check if read is allowed with given memory buffer and length.
303  *
304  * This function is used to check O_DIRECT memory buffer and request alignment.
305  */
306 static bool raw_is_io_aligned(int fd, void *buf, size_t len)
307 {
308     ssize_t ret = pread(fd, buf, len, 0);
309 
310     if (ret >= 0) {
311         return true;
312     }
313 
314 #ifdef __linux__
315     /* The Linux kernel returns EINVAL for misaligned O_DIRECT reads.  Ignore
316      * other errors (e.g. real I/O error), which could happen on a failed
317      * drive, since we only care about probing alignment.
318      */
319     if (errno != EINVAL) {
320         return true;
321     }
322 #endif
323 
324     return false;
325 }
326 
327 static void raw_probe_alignment(BlockDriverState *bs, int fd, Error **errp)
328 {
329     BDRVRawState *s = bs->opaque;
330     char *buf;
331     size_t max_align = MAX(MAX_BLOCKSIZE, qemu_real_host_page_size);
332     size_t alignments[] = {1, 512, 1024, 2048, 4096};
333 
334     /* For SCSI generic devices the alignment is not really used.
335        With buffered I/O, we don't have any restrictions. */
336     if (bdrv_is_sg(bs) || !s->needs_alignment) {
337         bs->bl.request_alignment = 1;
338         s->buf_align = 1;
339         return;
340     }
341 
342     bs->bl.request_alignment = 0;
343     s->buf_align = 0;
344     /* Let's try to use the logical blocksize for the alignment. */
345     if (probe_logical_blocksize(fd, &bs->bl.request_alignment) < 0) {
346         bs->bl.request_alignment = 0;
347     }
348 #ifdef CONFIG_XFS
349     if (s->is_xfs) {
350         struct dioattr da;
351         if (xfsctl(NULL, fd, XFS_IOC_DIOINFO, &da) >= 0) {
352             bs->bl.request_alignment = da.d_miniosz;
353             /* The kernel returns wrong information for d_mem */
354             /* s->buf_align = da.d_mem; */
355         }
356     }
357 #endif
358 
359     /*
360      * If we could not get the sizes so far, we can only guess them. First try
361      * to detect request alignment, since it is more likely to succeed. Then
362      * try to detect buf_align, which cannot be detected in some cases (e.g.
363      * Gluster). If buf_align cannot be detected, we fallback to the value of
364      * request_alignment.
365      */
366 
367     if (!bs->bl.request_alignment) {
368         int i;
369         size_t align;
370         buf = qemu_memalign(max_align, max_align);
371         for (i = 0; i < ARRAY_SIZE(alignments); i++) {
372             align = alignments[i];
373             if (raw_is_io_aligned(fd, buf, align)) {
374                 /* Fallback to safe value. */
375                 bs->bl.request_alignment = (align != 1) ? align : max_align;
376                 break;
377             }
378         }
379         qemu_vfree(buf);
380     }
381 
382     if (!s->buf_align) {
383         int i;
384         size_t align;
385         buf = qemu_memalign(max_align, 2 * max_align);
386         for (i = 0; i < ARRAY_SIZE(alignments); i++) {
387             align = alignments[i];
388             if (raw_is_io_aligned(fd, buf + align, max_align)) {
389                 /* Fallback to request_alignment. */
390                 s->buf_align = (align != 1) ? align : bs->bl.request_alignment;
391                 break;
392             }
393         }
394         qemu_vfree(buf);
395     }
396 
397     if (!s->buf_align || !bs->bl.request_alignment) {
398         error_setg(errp, "Could not find working O_DIRECT alignment");
399         error_append_hint(errp, "Try cache.direct=off\n");
400     }
401 }
402 
403 static void raw_parse_flags(int bdrv_flags, int *open_flags, bool has_writers)
404 {
405     bool read_write = false;
406     assert(open_flags != NULL);
407 
408     *open_flags |= O_BINARY;
409     *open_flags &= ~O_ACCMODE;
410 
411     if (bdrv_flags & BDRV_O_AUTO_RDONLY) {
412         read_write = has_writers;
413     } else if (bdrv_flags & BDRV_O_RDWR) {
414         read_write = true;
415     }
416 
417     if (read_write) {
418         *open_flags |= O_RDWR;
419     } else {
420         *open_flags |= O_RDONLY;
421     }
422 
423     /* Use O_DSYNC for write-through caching, no flags for write-back caching,
424      * and O_DIRECT for no caching. */
425     if ((bdrv_flags & BDRV_O_NOCACHE)) {
426         *open_flags |= O_DIRECT;
427     }
428 }
429 
430 static void raw_parse_filename(const char *filename, QDict *options,
431                                Error **errp)
432 {
433     bdrv_parse_filename_strip_prefix(filename, "file:", options);
434 }
435 
436 static QemuOptsList raw_runtime_opts = {
437     .name = "raw",
438     .head = QTAILQ_HEAD_INITIALIZER(raw_runtime_opts.head),
439     .desc = {
440         {
441             .name = "filename",
442             .type = QEMU_OPT_STRING,
443             .help = "File name of the image",
444         },
445         {
446             .name = "aio",
447             .type = QEMU_OPT_STRING,
448             .help = "host AIO implementation (threads, native, io_uring)",
449         },
450         {
451             .name = "locking",
452             .type = QEMU_OPT_STRING,
453             .help = "file locking mode (on/off/auto, default: auto)",
454         },
455         {
456             .name = "pr-manager",
457             .type = QEMU_OPT_STRING,
458             .help = "id of persistent reservation manager object (default: none)",
459         },
460 #if defined(__linux__)
461         {
462             .name = "drop-cache",
463             .type = QEMU_OPT_BOOL,
464             .help = "invalidate page cache during live migration (default: on)",
465         },
466 #endif
467         {
468             .name = "x-check-cache-dropped",
469             .type = QEMU_OPT_BOOL,
470             .help = "check that page cache was dropped on live migration (default: off)"
471         },
472         { /* end of list */ }
473     },
474 };
475 
476 static const char *const mutable_opts[] = { "x-check-cache-dropped", NULL };
477 
478 static int raw_open_common(BlockDriverState *bs, QDict *options,
479                            int bdrv_flags, int open_flags,
480                            bool device, Error **errp)
481 {
482     BDRVRawState *s = bs->opaque;
483     QemuOpts *opts;
484     Error *local_err = NULL;
485     const char *filename = NULL;
486     const char *str;
487     BlockdevAioOptions aio, aio_default;
488     int fd, ret;
489     struct stat st;
490     OnOffAuto locking;
491 
492     opts = qemu_opts_create(&raw_runtime_opts, NULL, 0, &error_abort);
493     qemu_opts_absorb_qdict(opts, options, &local_err);
494     if (local_err) {
495         error_propagate(errp, local_err);
496         ret = -EINVAL;
497         goto fail;
498     }
499 
500     filename = qemu_opt_get(opts, "filename");
501 
502     ret = raw_normalize_devicepath(&filename, errp);
503     if (ret != 0) {
504         goto fail;
505     }
506 
507     if (bdrv_flags & BDRV_O_NATIVE_AIO) {
508         aio_default = BLOCKDEV_AIO_OPTIONS_NATIVE;
509 #ifdef CONFIG_LINUX_IO_URING
510     } else if (bdrv_flags & BDRV_O_IO_URING) {
511         aio_default = BLOCKDEV_AIO_OPTIONS_IO_URING;
512 #endif
513     } else {
514         aio_default = BLOCKDEV_AIO_OPTIONS_THREADS;
515     }
516 
517     aio = qapi_enum_parse(&BlockdevAioOptions_lookup,
518                           qemu_opt_get(opts, "aio"),
519                           aio_default, &local_err);
520     if (local_err) {
521         error_propagate(errp, local_err);
522         ret = -EINVAL;
523         goto fail;
524     }
525 
526     s->use_linux_aio = (aio == BLOCKDEV_AIO_OPTIONS_NATIVE);
527 #ifdef CONFIG_LINUX_IO_URING
528     s->use_linux_io_uring = (aio == BLOCKDEV_AIO_OPTIONS_IO_URING);
529 #endif
530 
531     locking = qapi_enum_parse(&OnOffAuto_lookup,
532                               qemu_opt_get(opts, "locking"),
533                               ON_OFF_AUTO_AUTO, &local_err);
534     if (local_err) {
535         error_propagate(errp, local_err);
536         ret = -EINVAL;
537         goto fail;
538     }
539     switch (locking) {
540     case ON_OFF_AUTO_ON:
541         s->use_lock = true;
542         if (!qemu_has_ofd_lock()) {
543             warn_report("File lock requested but OFD locking syscall is "
544                         "unavailable, falling back to POSIX file locks");
545             error_printf("Due to the implementation, locks can be lost "
546                          "unexpectedly.\n");
547         }
548         break;
549     case ON_OFF_AUTO_OFF:
550         s->use_lock = false;
551         break;
552     case ON_OFF_AUTO_AUTO:
553         s->use_lock = qemu_has_ofd_lock();
554         break;
555     default:
556         abort();
557     }
558 
559     str = qemu_opt_get(opts, "pr-manager");
560     if (str) {
561         s->pr_mgr = pr_manager_lookup(str, &local_err);
562         if (local_err) {
563             error_propagate(errp, local_err);
564             ret = -EINVAL;
565             goto fail;
566         }
567     }
568 
569     s->drop_cache = qemu_opt_get_bool(opts, "drop-cache", true);
570     s->check_cache_dropped = qemu_opt_get_bool(opts, "x-check-cache-dropped",
571                                                false);
572 
573     s->open_flags = open_flags;
574     raw_parse_flags(bdrv_flags, &s->open_flags, false);
575 
576     s->fd = -1;
577     fd = qemu_open(filename, s->open_flags, 0644);
578     ret = fd < 0 ? -errno : 0;
579 
580     if (ret < 0) {
581         error_setg_file_open(errp, -ret, filename);
582         if (ret == -EROFS) {
583             ret = -EACCES;
584         }
585         goto fail;
586     }
587     s->fd = fd;
588 
589     s->perm = 0;
590     s->shared_perm = BLK_PERM_ALL;
591 
592 #ifdef CONFIG_LINUX_AIO
593      /* Currently Linux does AIO only for files opened with O_DIRECT */
594     if (s->use_linux_aio) {
595         if (!(s->open_flags & O_DIRECT)) {
596             error_setg(errp, "aio=native was specified, but it requires "
597                              "cache.direct=on, which was not specified.");
598             ret = -EINVAL;
599             goto fail;
600         }
601         if (!aio_setup_linux_aio(bdrv_get_aio_context(bs), errp)) {
602             error_prepend(errp, "Unable to use native AIO: ");
603             goto fail;
604         }
605     }
606 #else
607     if (s->use_linux_aio) {
608         error_setg(errp, "aio=native was specified, but is not supported "
609                          "in this build.");
610         ret = -EINVAL;
611         goto fail;
612     }
613 #endif /* !defined(CONFIG_LINUX_AIO) */
614 
615 #ifdef CONFIG_LINUX_IO_URING
616     if (s->use_linux_io_uring) {
617         if (!aio_setup_linux_io_uring(bdrv_get_aio_context(bs), errp)) {
618             error_prepend(errp, "Unable to use io_uring: ");
619             goto fail;
620         }
621     }
622 #else
623     if (s->use_linux_io_uring) {
624         error_setg(errp, "aio=io_uring was specified, but is not supported "
625                          "in this build.");
626         ret = -EINVAL;
627         goto fail;
628     }
629 #endif /* !defined(CONFIG_LINUX_IO_URING) */
630 
631     s->has_discard = true;
632     s->has_write_zeroes = true;
633     if ((bs->open_flags & BDRV_O_NOCACHE) != 0) {
634         s->needs_alignment = true;
635     }
636 
637     if (fstat(s->fd, &st) < 0) {
638         ret = -errno;
639         error_setg_errno(errp, errno, "Could not stat file");
640         goto fail;
641     }
642 
643     if (!device) {
644         if (S_ISBLK(st.st_mode)) {
645             warn_report("Opening a block device as a file using the '%s' "
646                         "driver is deprecated", bs->drv->format_name);
647         } else if (S_ISCHR(st.st_mode)) {
648             warn_report("Opening a character device as a file using the '%s' "
649                         "driver is deprecated", bs->drv->format_name);
650         } else if (!S_ISREG(st.st_mode)) {
651             error_setg(errp, "A regular file was expected by the '%s' driver, "
652                        "but something else was given", bs->drv->format_name);
653             ret = -EINVAL;
654             goto fail;
655         } else {
656             s->discard_zeroes = true;
657             s->has_fallocate = true;
658         }
659     } else {
660         if (!(S_ISCHR(st.st_mode) || S_ISBLK(st.st_mode))) {
661             error_setg(errp, "'%s' driver expects either "
662                        "a character or block device", bs->drv->format_name);
663             ret = -EINVAL;
664             goto fail;
665         }
666     }
667 
668     if (S_ISBLK(st.st_mode)) {
669 #ifdef BLKDISCARDZEROES
670         unsigned int arg;
671         if (ioctl(s->fd, BLKDISCARDZEROES, &arg) == 0 && arg) {
672             s->discard_zeroes = true;
673         }
674 #endif
675 #ifdef __linux__
676         /* On Linux 3.10, BLKDISCARD leaves stale data in the page cache.  Do
677          * not rely on the contents of discarded blocks unless using O_DIRECT.
678          * Same for BLKZEROOUT.
679          */
680         if (!(bs->open_flags & BDRV_O_NOCACHE)) {
681             s->discard_zeroes = false;
682             s->has_write_zeroes = false;
683         }
684 #endif
685     }
686 #ifdef __FreeBSD__
687     if (S_ISCHR(st.st_mode)) {
688         /*
689          * The file is a char device (disk), which on FreeBSD isn't behind
690          * a pager, so force all requests to be aligned. This is needed
691          * so QEMU makes sure all IO operations on the device are aligned
692          * to sector size, or else FreeBSD will reject them with EINVAL.
693          */
694         s->needs_alignment = true;
695     }
696 #endif
697 
698 #ifdef CONFIG_XFS
699     if (platform_test_xfs_fd(s->fd)) {
700         s->is_xfs = true;
701     }
702 #endif
703 
704     bs->supported_zero_flags = BDRV_REQ_MAY_UNMAP | BDRV_REQ_NO_FALLBACK;
705     if (S_ISREG(st.st_mode)) {
706         /* When extending regular files, we get zeros from the OS */
707         bs->supported_truncate_flags = BDRV_REQ_ZERO_WRITE;
708     }
709     ret = 0;
710 fail:
711     if (filename && (bdrv_flags & BDRV_O_TEMPORARY)) {
712         unlink(filename);
713     }
714     qemu_opts_del(opts);
715     return ret;
716 }
717 
718 static int raw_open(BlockDriverState *bs, QDict *options, int flags,
719                     Error **errp)
720 {
721     BDRVRawState *s = bs->opaque;
722 
723     s->type = FTYPE_FILE;
724     return raw_open_common(bs, options, flags, 0, false, errp);
725 }
726 
727 typedef enum {
728     RAW_PL_PREPARE,
729     RAW_PL_COMMIT,
730     RAW_PL_ABORT,
731 } RawPermLockOp;
732 
733 #define PERM_FOREACH(i) \
734     for ((i) = 0; (1ULL << (i)) <= BLK_PERM_ALL; i++)
735 
736 /* Lock bytes indicated by @perm_lock_bits and @shared_perm_lock_bits in the
737  * file; if @unlock == true, also unlock the unneeded bytes.
738  * @shared_perm_lock_bits is the mask of all permissions that are NOT shared.
739  */
740 static int raw_apply_lock_bytes(BDRVRawState *s, int fd,
741                                 uint64_t perm_lock_bits,
742                                 uint64_t shared_perm_lock_bits,
743                                 bool unlock, Error **errp)
744 {
745     int ret;
746     int i;
747     uint64_t locked_perm, locked_shared_perm;
748 
749     if (s) {
750         locked_perm = s->locked_perm;
751         locked_shared_perm = s->locked_shared_perm;
752     } else {
753         /*
754          * We don't have the previous bits, just lock/unlock for each of the
755          * requested bits.
756          */
757         if (unlock) {
758             locked_perm = BLK_PERM_ALL;
759             locked_shared_perm = BLK_PERM_ALL;
760         } else {
761             locked_perm = 0;
762             locked_shared_perm = 0;
763         }
764     }
765 
766     PERM_FOREACH(i) {
767         int off = RAW_LOCK_PERM_BASE + i;
768         uint64_t bit = (1ULL << i);
769         if ((perm_lock_bits & bit) && !(locked_perm & bit)) {
770             ret = qemu_lock_fd(fd, off, 1, false);
771             if (ret) {
772                 error_setg(errp, "Failed to lock byte %d", off);
773                 return ret;
774             } else if (s) {
775                 s->locked_perm |= bit;
776             }
777         } else if (unlock && (locked_perm & bit) && !(perm_lock_bits & bit)) {
778             ret = qemu_unlock_fd(fd, off, 1);
779             if (ret) {
780                 error_setg(errp, "Failed to unlock byte %d", off);
781                 return ret;
782             } else if (s) {
783                 s->locked_perm &= ~bit;
784             }
785         }
786     }
787     PERM_FOREACH(i) {
788         int off = RAW_LOCK_SHARED_BASE + i;
789         uint64_t bit = (1ULL << i);
790         if ((shared_perm_lock_bits & bit) && !(locked_shared_perm & bit)) {
791             ret = qemu_lock_fd(fd, off, 1, false);
792             if (ret) {
793                 error_setg(errp, "Failed to lock byte %d", off);
794                 return ret;
795             } else if (s) {
796                 s->locked_shared_perm |= bit;
797             }
798         } else if (unlock && (locked_shared_perm & bit) &&
799                    !(shared_perm_lock_bits & bit)) {
800             ret = qemu_unlock_fd(fd, off, 1);
801             if (ret) {
802                 error_setg(errp, "Failed to unlock byte %d", off);
803                 return ret;
804             } else if (s) {
805                 s->locked_shared_perm &= ~bit;
806             }
807         }
808     }
809     return 0;
810 }
811 
812 /* Check "unshared" bytes implied by @perm and ~@shared_perm in the file. */
813 static int raw_check_lock_bytes(int fd, uint64_t perm, uint64_t shared_perm,
814                                 Error **errp)
815 {
816     int ret;
817     int i;
818 
819     PERM_FOREACH(i) {
820         int off = RAW_LOCK_SHARED_BASE + i;
821         uint64_t p = 1ULL << i;
822         if (perm & p) {
823             ret = qemu_lock_fd_test(fd, off, 1, true);
824             if (ret) {
825                 char *perm_name = bdrv_perm_names(p);
826                 error_setg(errp,
827                            "Failed to get \"%s\" lock",
828                            perm_name);
829                 g_free(perm_name);
830                 return ret;
831             }
832         }
833     }
834     PERM_FOREACH(i) {
835         int off = RAW_LOCK_PERM_BASE + i;
836         uint64_t p = 1ULL << i;
837         if (!(shared_perm & p)) {
838             ret = qemu_lock_fd_test(fd, off, 1, true);
839             if (ret) {
840                 char *perm_name = bdrv_perm_names(p);
841                 error_setg(errp,
842                            "Failed to get shared \"%s\" lock",
843                            perm_name);
844                 g_free(perm_name);
845                 return ret;
846             }
847         }
848     }
849     return 0;
850 }
851 
852 static int raw_handle_perm_lock(BlockDriverState *bs,
853                                 RawPermLockOp op,
854                                 uint64_t new_perm, uint64_t new_shared,
855                                 Error **errp)
856 {
857     BDRVRawState *s = bs->opaque;
858     int ret = 0;
859     Error *local_err = NULL;
860 
861     if (!s->use_lock) {
862         return 0;
863     }
864 
865     if (bdrv_get_flags(bs) & BDRV_O_INACTIVE) {
866         return 0;
867     }
868 
869     switch (op) {
870     case RAW_PL_PREPARE:
871         if ((s->perm | new_perm) == s->perm &&
872             (s->shared_perm & new_shared) == s->shared_perm)
873         {
874             /*
875              * We are going to unlock bytes, it should not fail. If it fail due
876              * to some fs-dependent permission-unrelated reasons (which occurs
877              * sometimes on NFS and leads to abort in bdrv_replace_child) we
878              * can't prevent such errors by any check here. And we ignore them
879              * anyway in ABORT and COMMIT.
880              */
881             return 0;
882         }
883         ret = raw_apply_lock_bytes(s, s->fd, s->perm | new_perm,
884                                    ~s->shared_perm | ~new_shared,
885                                    false, errp);
886         if (!ret) {
887             ret = raw_check_lock_bytes(s->fd, new_perm, new_shared, errp);
888             if (!ret) {
889                 return 0;
890             }
891             error_append_hint(errp,
892                               "Is another process using the image [%s]?\n",
893                               bs->filename);
894         }
895         /* fall through to unlock bytes. */
896     case RAW_PL_ABORT:
897         raw_apply_lock_bytes(s, s->fd, s->perm, ~s->shared_perm,
898                              true, &local_err);
899         if (local_err) {
900             /* Theoretically the above call only unlocks bytes and it cannot
901              * fail. Something weird happened, report it.
902              */
903             warn_report_err(local_err);
904         }
905         break;
906     case RAW_PL_COMMIT:
907         raw_apply_lock_bytes(s, s->fd, new_perm, ~new_shared,
908                              true, &local_err);
909         if (local_err) {
910             /* Theoretically the above call only unlocks bytes and it cannot
911              * fail. Something weird happened, report it.
912              */
913             warn_report_err(local_err);
914         }
915         break;
916     }
917     return ret;
918 }
919 
920 static int raw_reconfigure_getfd(BlockDriverState *bs, int flags,
921                                  int *open_flags, uint64_t perm, bool force_dup,
922                                  Error **errp)
923 {
924     BDRVRawState *s = bs->opaque;
925     int fd = -1;
926     int ret;
927     bool has_writers = perm &
928         (BLK_PERM_WRITE | BLK_PERM_WRITE_UNCHANGED | BLK_PERM_RESIZE);
929     int fcntl_flags = O_APPEND | O_NONBLOCK;
930 #ifdef O_NOATIME
931     fcntl_flags |= O_NOATIME;
932 #endif
933 
934     *open_flags = 0;
935     if (s->type == FTYPE_CD) {
936         *open_flags |= O_NONBLOCK;
937     }
938 
939     raw_parse_flags(flags, open_flags, has_writers);
940 
941 #ifdef O_ASYNC
942     /* Not all operating systems have O_ASYNC, and those that don't
943      * will not let us track the state into rs->open_flags (typically
944      * you achieve the same effect with an ioctl, for example I_SETSIG
945      * on Solaris). But we do not use O_ASYNC, so that's fine.
946      */
947     assert((s->open_flags & O_ASYNC) == 0);
948 #endif
949 
950     if (!force_dup && *open_flags == s->open_flags) {
951         /* We're lucky, the existing fd is fine */
952         return s->fd;
953     }
954 
955     if ((*open_flags & ~fcntl_flags) == (s->open_flags & ~fcntl_flags)) {
956         /* dup the original fd */
957         fd = qemu_dup(s->fd);
958         if (fd >= 0) {
959             ret = fcntl_setfl(fd, *open_flags);
960             if (ret) {
961                 qemu_close(fd);
962                 fd = -1;
963             }
964         }
965     }
966 
967     /* If we cannot use fcntl, or fcntl failed, fall back to qemu_open() */
968     if (fd == -1) {
969         const char *normalized_filename = bs->filename;
970         ret = raw_normalize_devicepath(&normalized_filename, errp);
971         if (ret >= 0) {
972             assert(!(*open_flags & O_CREAT));
973             fd = qemu_open(normalized_filename, *open_flags);
974             if (fd == -1) {
975                 error_setg_errno(errp, errno, "Could not reopen file");
976                 return -1;
977             }
978         }
979     }
980 
981     return fd;
982 }
983 
984 static int raw_reopen_prepare(BDRVReopenState *state,
985                               BlockReopenQueue *queue, Error **errp)
986 {
987     BDRVRawState *s;
988     BDRVRawReopenState *rs;
989     QemuOpts *opts;
990     int ret;
991     Error *local_err = NULL;
992 
993     assert(state != NULL);
994     assert(state->bs != NULL);
995 
996     s = state->bs->opaque;
997 
998     state->opaque = g_new0(BDRVRawReopenState, 1);
999     rs = state->opaque;
1000 
1001     /* Handle options changes */
1002     opts = qemu_opts_create(&raw_runtime_opts, NULL, 0, &error_abort);
1003     qemu_opts_absorb_qdict(opts, state->options, &local_err);
1004     if (local_err) {
1005         error_propagate(errp, local_err);
1006         ret = -EINVAL;
1007         goto out;
1008     }
1009 
1010     rs->drop_cache = qemu_opt_get_bool_del(opts, "drop-cache", true);
1011     rs->check_cache_dropped =
1012         qemu_opt_get_bool_del(opts, "x-check-cache-dropped", false);
1013 
1014     /* This driver's reopen function doesn't currently allow changing
1015      * other options, so let's put them back in the original QDict and
1016      * bdrv_reopen_prepare() will detect changes and complain. */
1017     qemu_opts_to_qdict(opts, state->options);
1018 
1019     rs->fd = raw_reconfigure_getfd(state->bs, state->flags, &rs->open_flags,
1020                                    state->perm, true, &local_err);
1021     if (local_err) {
1022         error_propagate(errp, local_err);
1023         ret = -1;
1024         goto out;
1025     }
1026 
1027     /* Fail already reopen_prepare() if we can't get a working O_DIRECT
1028      * alignment with the new fd. */
1029     if (rs->fd != -1) {
1030         raw_probe_alignment(state->bs, rs->fd, &local_err);
1031         if (local_err) {
1032             error_propagate(errp, local_err);
1033             ret = -EINVAL;
1034             goto out_fd;
1035         }
1036     }
1037 
1038     s->reopen_state = state;
1039     ret = 0;
1040 out_fd:
1041     if (ret < 0) {
1042         qemu_close(rs->fd);
1043         rs->fd = -1;
1044     }
1045 out:
1046     qemu_opts_del(opts);
1047     return ret;
1048 }
1049 
1050 static void raw_reopen_commit(BDRVReopenState *state)
1051 {
1052     BDRVRawReopenState *rs = state->opaque;
1053     BDRVRawState *s = state->bs->opaque;
1054 
1055     s->drop_cache = rs->drop_cache;
1056     s->check_cache_dropped = rs->check_cache_dropped;
1057     s->open_flags = rs->open_flags;
1058 
1059     qemu_close(s->fd);
1060     s->fd = rs->fd;
1061 
1062     g_free(state->opaque);
1063     state->opaque = NULL;
1064 
1065     assert(s->reopen_state == state);
1066     s->reopen_state = NULL;
1067 }
1068 
1069 
1070 static void raw_reopen_abort(BDRVReopenState *state)
1071 {
1072     BDRVRawReopenState *rs = state->opaque;
1073     BDRVRawState *s = state->bs->opaque;
1074 
1075      /* nothing to do if NULL, we didn't get far enough */
1076     if (rs == NULL) {
1077         return;
1078     }
1079 
1080     if (rs->fd >= 0) {
1081         qemu_close(rs->fd);
1082         rs->fd = -1;
1083     }
1084     g_free(state->opaque);
1085     state->opaque = NULL;
1086 
1087     assert(s->reopen_state == state);
1088     s->reopen_state = NULL;
1089 }
1090 
1091 static int sg_get_max_transfer_length(int fd)
1092 {
1093 #ifdef BLKSECTGET
1094     int max_bytes = 0;
1095 
1096     if (ioctl(fd, BLKSECTGET, &max_bytes) == 0) {
1097         return max_bytes;
1098     } else {
1099         return -errno;
1100     }
1101 #else
1102     return -ENOSYS;
1103 #endif
1104 }
1105 
1106 static int sg_get_max_segments(int fd)
1107 {
1108 #ifdef CONFIG_LINUX
1109     char buf[32];
1110     const char *end;
1111     char *sysfspath = NULL;
1112     int ret;
1113     int sysfd = -1;
1114     long max_segments;
1115     struct stat st;
1116 
1117     if (fstat(fd, &st)) {
1118         ret = -errno;
1119         goto out;
1120     }
1121 
1122     sysfspath = g_strdup_printf("/sys/dev/block/%u:%u/queue/max_segments",
1123                                 major(st.st_rdev), minor(st.st_rdev));
1124     sysfd = open(sysfspath, O_RDONLY);
1125     if (sysfd == -1) {
1126         ret = -errno;
1127         goto out;
1128     }
1129     do {
1130         ret = read(sysfd, buf, sizeof(buf) - 1);
1131     } while (ret == -1 && errno == EINTR);
1132     if (ret < 0) {
1133         ret = -errno;
1134         goto out;
1135     } else if (ret == 0) {
1136         ret = -EIO;
1137         goto out;
1138     }
1139     buf[ret] = 0;
1140     /* The file is ended with '\n', pass 'end' to accept that. */
1141     ret = qemu_strtol(buf, &end, 10, &max_segments);
1142     if (ret == 0 && end && *end == '\n') {
1143         ret = max_segments;
1144     }
1145 
1146 out:
1147     if (sysfd != -1) {
1148         close(sysfd);
1149     }
1150     g_free(sysfspath);
1151     return ret;
1152 #else
1153     return -ENOTSUP;
1154 #endif
1155 }
1156 
1157 static void raw_refresh_limits(BlockDriverState *bs, Error **errp)
1158 {
1159     BDRVRawState *s = bs->opaque;
1160 
1161     if (bs->sg) {
1162         int ret = sg_get_max_transfer_length(s->fd);
1163 
1164         if (ret > 0 && ret <= BDRV_REQUEST_MAX_BYTES) {
1165             bs->bl.max_transfer = pow2floor(ret);
1166         }
1167 
1168         ret = sg_get_max_segments(s->fd);
1169         if (ret > 0) {
1170             bs->bl.max_transfer = MIN(bs->bl.max_transfer,
1171                                       ret * qemu_real_host_page_size);
1172         }
1173     }
1174 
1175     raw_probe_alignment(bs, s->fd, errp);
1176     bs->bl.min_mem_alignment = s->buf_align;
1177     bs->bl.opt_mem_alignment = MAX(s->buf_align, qemu_real_host_page_size);
1178 }
1179 
1180 static int check_for_dasd(int fd)
1181 {
1182 #ifdef BIODASDINFO2
1183     struct dasd_information2_t info = {0};
1184 
1185     return ioctl(fd, BIODASDINFO2, &info);
1186 #else
1187     return -1;
1188 #endif
1189 }
1190 
1191 /**
1192  * Try to get @bs's logical and physical block size.
1193  * On success, store them in @bsz and return zero.
1194  * On failure, return negative errno.
1195  */
1196 static int hdev_probe_blocksizes(BlockDriverState *bs, BlockSizes *bsz)
1197 {
1198     BDRVRawState *s = bs->opaque;
1199     int ret;
1200 
1201     /* If DASD, get blocksizes */
1202     if (check_for_dasd(s->fd) < 0) {
1203         return -ENOTSUP;
1204     }
1205     ret = probe_logical_blocksize(s->fd, &bsz->log);
1206     if (ret < 0) {
1207         return ret;
1208     }
1209     return probe_physical_blocksize(s->fd, &bsz->phys);
1210 }
1211 
1212 /**
1213  * Try to get @bs's geometry: cyls, heads, sectors.
1214  * On success, store them in @geo and return 0.
1215  * On failure return -errno.
1216  * (Allows block driver to assign default geometry values that guest sees)
1217  */
1218 #ifdef __linux__
1219 static int hdev_probe_geometry(BlockDriverState *bs, HDGeometry *geo)
1220 {
1221     BDRVRawState *s = bs->opaque;
1222     struct hd_geometry ioctl_geo = {0};
1223 
1224     /* If DASD, get its geometry */
1225     if (check_for_dasd(s->fd) < 0) {
1226         return -ENOTSUP;
1227     }
1228     if (ioctl(s->fd, HDIO_GETGEO, &ioctl_geo) < 0) {
1229         return -errno;
1230     }
1231     /* HDIO_GETGEO may return success even though geo contains zeros
1232        (e.g. certain multipath setups) */
1233     if (!ioctl_geo.heads || !ioctl_geo.sectors || !ioctl_geo.cylinders) {
1234         return -ENOTSUP;
1235     }
1236     /* Do not return a geometry for partition */
1237     if (ioctl_geo.start != 0) {
1238         return -ENOTSUP;
1239     }
1240     geo->heads = ioctl_geo.heads;
1241     geo->sectors = ioctl_geo.sectors;
1242     geo->cylinders = ioctl_geo.cylinders;
1243 
1244     return 0;
1245 }
1246 #else /* __linux__ */
1247 static int hdev_probe_geometry(BlockDriverState *bs, HDGeometry *geo)
1248 {
1249     return -ENOTSUP;
1250 }
1251 #endif
1252 
1253 #if defined(__linux__)
1254 static int handle_aiocb_ioctl(void *opaque)
1255 {
1256     RawPosixAIOData *aiocb = opaque;
1257     int ret;
1258 
1259     ret = ioctl(aiocb->aio_fildes, aiocb->ioctl.cmd, aiocb->ioctl.buf);
1260     if (ret == -1) {
1261         return -errno;
1262     }
1263 
1264     return 0;
1265 }
1266 #endif /* linux */
1267 
1268 static int handle_aiocb_flush(void *opaque)
1269 {
1270     RawPosixAIOData *aiocb = opaque;
1271     BDRVRawState *s = aiocb->bs->opaque;
1272     int ret;
1273 
1274     if (s->page_cache_inconsistent) {
1275         return -EIO;
1276     }
1277 
1278     ret = qemu_fdatasync(aiocb->aio_fildes);
1279     if (ret == -1) {
1280         /* There is no clear definition of the semantics of a failing fsync(),
1281          * so we may have to assume the worst. The sad truth is that this
1282          * assumption is correct for Linux. Some pages are now probably marked
1283          * clean in the page cache even though they are inconsistent with the
1284          * on-disk contents. The next fdatasync() call would succeed, but no
1285          * further writeback attempt will be made. We can't get back to a state
1286          * in which we know what is on disk (we would have to rewrite
1287          * everything that was touched since the last fdatasync() at least), so
1288          * make bdrv_flush() fail permanently. Given that the behaviour isn't
1289          * really defined, I have little hope that other OSes are doing better.
1290          *
1291          * Obviously, this doesn't affect O_DIRECT, which bypasses the page
1292          * cache. */
1293         if ((s->open_flags & O_DIRECT) == 0) {
1294             s->page_cache_inconsistent = true;
1295         }
1296         return -errno;
1297     }
1298     return 0;
1299 }
1300 
1301 #ifdef CONFIG_PREADV
1302 
1303 static bool preadv_present = true;
1304 
1305 static ssize_t
1306 qemu_preadv(int fd, const struct iovec *iov, int nr_iov, off_t offset)
1307 {
1308     return preadv(fd, iov, nr_iov, offset);
1309 }
1310 
1311 static ssize_t
1312 qemu_pwritev(int fd, const struct iovec *iov, int nr_iov, off_t offset)
1313 {
1314     return pwritev(fd, iov, nr_iov, offset);
1315 }
1316 
1317 #else
1318 
1319 static bool preadv_present = false;
1320 
1321 static ssize_t
1322 qemu_preadv(int fd, const struct iovec *iov, int nr_iov, off_t offset)
1323 {
1324     return -ENOSYS;
1325 }
1326 
1327 static ssize_t
1328 qemu_pwritev(int fd, const struct iovec *iov, int nr_iov, off_t offset)
1329 {
1330     return -ENOSYS;
1331 }
1332 
1333 #endif
1334 
1335 static ssize_t handle_aiocb_rw_vector(RawPosixAIOData *aiocb)
1336 {
1337     ssize_t len;
1338 
1339     do {
1340         if (aiocb->aio_type & QEMU_AIO_WRITE)
1341             len = qemu_pwritev(aiocb->aio_fildes,
1342                                aiocb->io.iov,
1343                                aiocb->io.niov,
1344                                aiocb->aio_offset);
1345          else
1346             len = qemu_preadv(aiocb->aio_fildes,
1347                               aiocb->io.iov,
1348                               aiocb->io.niov,
1349                               aiocb->aio_offset);
1350     } while (len == -1 && errno == EINTR);
1351 
1352     if (len == -1) {
1353         return -errno;
1354     }
1355     return len;
1356 }
1357 
1358 /*
1359  * Read/writes the data to/from a given linear buffer.
1360  *
1361  * Returns the number of bytes handles or -errno in case of an error. Short
1362  * reads are only returned if the end of the file is reached.
1363  */
1364 static ssize_t handle_aiocb_rw_linear(RawPosixAIOData *aiocb, char *buf)
1365 {
1366     ssize_t offset = 0;
1367     ssize_t len;
1368 
1369     while (offset < aiocb->aio_nbytes) {
1370         if (aiocb->aio_type & QEMU_AIO_WRITE) {
1371             len = pwrite(aiocb->aio_fildes,
1372                          (const char *)buf + offset,
1373                          aiocb->aio_nbytes - offset,
1374                          aiocb->aio_offset + offset);
1375         } else {
1376             len = pread(aiocb->aio_fildes,
1377                         buf + offset,
1378                         aiocb->aio_nbytes - offset,
1379                         aiocb->aio_offset + offset);
1380         }
1381         if (len == -1 && errno == EINTR) {
1382             continue;
1383         } else if (len == -1 && errno == EINVAL &&
1384                    (aiocb->bs->open_flags & BDRV_O_NOCACHE) &&
1385                    !(aiocb->aio_type & QEMU_AIO_WRITE) &&
1386                    offset > 0) {
1387             /* O_DIRECT pread() may fail with EINVAL when offset is unaligned
1388              * after a short read.  Assume that O_DIRECT short reads only occur
1389              * at EOF.  Therefore this is a short read, not an I/O error.
1390              */
1391             break;
1392         } else if (len == -1) {
1393             offset = -errno;
1394             break;
1395         } else if (len == 0) {
1396             break;
1397         }
1398         offset += len;
1399     }
1400 
1401     return offset;
1402 }
1403 
1404 static int handle_aiocb_rw(void *opaque)
1405 {
1406     RawPosixAIOData *aiocb = opaque;
1407     ssize_t nbytes;
1408     char *buf;
1409 
1410     if (!(aiocb->aio_type & QEMU_AIO_MISALIGNED)) {
1411         /*
1412          * If there is just a single buffer, and it is properly aligned
1413          * we can just use plain pread/pwrite without any problems.
1414          */
1415         if (aiocb->io.niov == 1) {
1416             nbytes = handle_aiocb_rw_linear(aiocb, aiocb->io.iov->iov_base);
1417             goto out;
1418         }
1419         /*
1420          * We have more than one iovec, and all are properly aligned.
1421          *
1422          * Try preadv/pwritev first and fall back to linearizing the
1423          * buffer if it's not supported.
1424          */
1425         if (preadv_present) {
1426             nbytes = handle_aiocb_rw_vector(aiocb);
1427             if (nbytes == aiocb->aio_nbytes ||
1428                 (nbytes < 0 && nbytes != -ENOSYS)) {
1429                 goto out;
1430             }
1431             preadv_present = false;
1432         }
1433 
1434         /*
1435          * XXX(hch): short read/write.  no easy way to handle the reminder
1436          * using these interfaces.  For now retry using plain
1437          * pread/pwrite?
1438          */
1439     }
1440 
1441     /*
1442      * Ok, we have to do it the hard way, copy all segments into
1443      * a single aligned buffer.
1444      */
1445     buf = qemu_try_blockalign(aiocb->bs, aiocb->aio_nbytes);
1446     if (buf == NULL) {
1447         nbytes = -ENOMEM;
1448         goto out;
1449     }
1450 
1451     if (aiocb->aio_type & QEMU_AIO_WRITE) {
1452         char *p = buf;
1453         int i;
1454 
1455         for (i = 0; i < aiocb->io.niov; ++i) {
1456             memcpy(p, aiocb->io.iov[i].iov_base, aiocb->io.iov[i].iov_len);
1457             p += aiocb->io.iov[i].iov_len;
1458         }
1459         assert(p - buf == aiocb->aio_nbytes);
1460     }
1461 
1462     nbytes = handle_aiocb_rw_linear(aiocb, buf);
1463     if (!(aiocb->aio_type & QEMU_AIO_WRITE)) {
1464         char *p = buf;
1465         size_t count = aiocb->aio_nbytes, copy;
1466         int i;
1467 
1468         for (i = 0; i < aiocb->io.niov && count; ++i) {
1469             copy = count;
1470             if (copy > aiocb->io.iov[i].iov_len) {
1471                 copy = aiocb->io.iov[i].iov_len;
1472             }
1473             memcpy(aiocb->io.iov[i].iov_base, p, copy);
1474             assert(count >= copy);
1475             p     += copy;
1476             count -= copy;
1477         }
1478         assert(count == 0);
1479     }
1480     qemu_vfree(buf);
1481 
1482 out:
1483     if (nbytes == aiocb->aio_nbytes) {
1484         return 0;
1485     } else if (nbytes >= 0 && nbytes < aiocb->aio_nbytes) {
1486         if (aiocb->aio_type & QEMU_AIO_WRITE) {
1487             return -EINVAL;
1488         } else {
1489             iov_memset(aiocb->io.iov, aiocb->io.niov, nbytes,
1490                       0, aiocb->aio_nbytes - nbytes);
1491             return 0;
1492         }
1493     } else {
1494         assert(nbytes < 0);
1495         return nbytes;
1496     }
1497 }
1498 
1499 static int translate_err(int err)
1500 {
1501     if (err == -ENODEV || err == -ENOSYS || err == -EOPNOTSUPP ||
1502         err == -ENOTTY) {
1503         err = -ENOTSUP;
1504     }
1505     return err;
1506 }
1507 
1508 #ifdef CONFIG_FALLOCATE
1509 static int do_fallocate(int fd, int mode, off_t offset, off_t len)
1510 {
1511     do {
1512         if (fallocate(fd, mode, offset, len) == 0) {
1513             return 0;
1514         }
1515     } while (errno == EINTR);
1516     return translate_err(-errno);
1517 }
1518 #endif
1519 
1520 static ssize_t handle_aiocb_write_zeroes_block(RawPosixAIOData *aiocb)
1521 {
1522     int ret = -ENOTSUP;
1523     BDRVRawState *s = aiocb->bs->opaque;
1524 
1525     if (!s->has_write_zeroes) {
1526         return -ENOTSUP;
1527     }
1528 
1529 #ifdef BLKZEROOUT
1530     /* The BLKZEROOUT implementation in the kernel doesn't set
1531      * BLKDEV_ZERO_NOFALLBACK, so we can't call this if we have to avoid slow
1532      * fallbacks. */
1533     if (!(aiocb->aio_type & QEMU_AIO_NO_FALLBACK)) {
1534         do {
1535             uint64_t range[2] = { aiocb->aio_offset, aiocb->aio_nbytes };
1536             if (ioctl(aiocb->aio_fildes, BLKZEROOUT, range) == 0) {
1537                 return 0;
1538             }
1539         } while (errno == EINTR);
1540 
1541         ret = translate_err(-errno);
1542         if (ret == -ENOTSUP) {
1543             s->has_write_zeroes = false;
1544         }
1545     }
1546 #endif
1547 
1548     return ret;
1549 }
1550 
1551 static int handle_aiocb_write_zeroes(void *opaque)
1552 {
1553     RawPosixAIOData *aiocb = opaque;
1554 #ifdef CONFIG_FALLOCATE
1555     BDRVRawState *s = aiocb->bs->opaque;
1556     int64_t len;
1557 #endif
1558 
1559     if (aiocb->aio_type & QEMU_AIO_BLKDEV) {
1560         return handle_aiocb_write_zeroes_block(aiocb);
1561     }
1562 
1563 #ifdef CONFIG_FALLOCATE_ZERO_RANGE
1564     if (s->has_write_zeroes) {
1565         int ret = do_fallocate(s->fd, FALLOC_FL_ZERO_RANGE,
1566                                aiocb->aio_offset, aiocb->aio_nbytes);
1567         if (ret == -EINVAL) {
1568             /*
1569              * Allow falling back to pwrite for file systems that
1570              * do not support fallocate() for an unaligned byte range.
1571              */
1572             return -ENOTSUP;
1573         }
1574         if (ret == 0 || ret != -ENOTSUP) {
1575             return ret;
1576         }
1577         s->has_write_zeroes = false;
1578     }
1579 #endif
1580 
1581 #ifdef CONFIG_FALLOCATE_PUNCH_HOLE
1582     if (s->has_discard && s->has_fallocate) {
1583         int ret = do_fallocate(s->fd,
1584                                FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE,
1585                                aiocb->aio_offset, aiocb->aio_nbytes);
1586         if (ret == 0) {
1587             ret = do_fallocate(s->fd, 0, aiocb->aio_offset, aiocb->aio_nbytes);
1588             if (ret == 0 || ret != -ENOTSUP) {
1589                 return ret;
1590             }
1591             s->has_fallocate = false;
1592         } else if (ret != -ENOTSUP) {
1593             return ret;
1594         } else {
1595             s->has_discard = false;
1596         }
1597     }
1598 #endif
1599 
1600 #ifdef CONFIG_FALLOCATE
1601     /* Last resort: we are trying to extend the file with zeroed data. This
1602      * can be done via fallocate(fd, 0) */
1603     len = bdrv_getlength(aiocb->bs);
1604     if (s->has_fallocate && len >= 0 && aiocb->aio_offset >= len) {
1605         int ret = do_fallocate(s->fd, 0, aiocb->aio_offset, aiocb->aio_nbytes);
1606         if (ret == 0 || ret != -ENOTSUP) {
1607             return ret;
1608         }
1609         s->has_fallocate = false;
1610     }
1611 #endif
1612 
1613     return -ENOTSUP;
1614 }
1615 
1616 static int handle_aiocb_write_zeroes_unmap(void *opaque)
1617 {
1618     RawPosixAIOData *aiocb = opaque;
1619     BDRVRawState *s G_GNUC_UNUSED = aiocb->bs->opaque;
1620     int ret;
1621 
1622     /* First try to write zeros and unmap at the same time */
1623 
1624 #ifdef CONFIG_FALLOCATE_PUNCH_HOLE
1625     ret = do_fallocate(s->fd, FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE,
1626                        aiocb->aio_offset, aiocb->aio_nbytes);
1627     if (ret != -ENOTSUP) {
1628         return ret;
1629     }
1630 #endif
1631 
1632     /* If we couldn't manage to unmap while guaranteed that the area reads as
1633      * all-zero afterwards, just write zeroes without unmapping */
1634     ret = handle_aiocb_write_zeroes(aiocb);
1635     return ret;
1636 }
1637 
1638 #ifndef HAVE_COPY_FILE_RANGE
1639 static off_t copy_file_range(int in_fd, off_t *in_off, int out_fd,
1640                              off_t *out_off, size_t len, unsigned int flags)
1641 {
1642 #ifdef __NR_copy_file_range
1643     return syscall(__NR_copy_file_range, in_fd, in_off, out_fd,
1644                    out_off, len, flags);
1645 #else
1646     errno = ENOSYS;
1647     return -1;
1648 #endif
1649 }
1650 #endif
1651 
1652 static int handle_aiocb_copy_range(void *opaque)
1653 {
1654     RawPosixAIOData *aiocb = opaque;
1655     uint64_t bytes = aiocb->aio_nbytes;
1656     off_t in_off = aiocb->aio_offset;
1657     off_t out_off = aiocb->copy_range.aio_offset2;
1658 
1659     while (bytes) {
1660         ssize_t ret = copy_file_range(aiocb->aio_fildes, &in_off,
1661                                       aiocb->copy_range.aio_fd2, &out_off,
1662                                       bytes, 0);
1663         trace_file_copy_file_range(aiocb->bs, aiocb->aio_fildes, in_off,
1664                                    aiocb->copy_range.aio_fd2, out_off, bytes,
1665                                    0, ret);
1666         if (ret == 0) {
1667             /* No progress (e.g. when beyond EOF), let the caller fall back to
1668              * buffer I/O. */
1669             return -ENOSPC;
1670         }
1671         if (ret < 0) {
1672             switch (errno) {
1673             case ENOSYS:
1674                 return -ENOTSUP;
1675             case EINTR:
1676                 continue;
1677             default:
1678                 return -errno;
1679             }
1680         }
1681         bytes -= ret;
1682     }
1683     return 0;
1684 }
1685 
1686 static int handle_aiocb_discard(void *opaque)
1687 {
1688     RawPosixAIOData *aiocb = opaque;
1689     int ret = -EOPNOTSUPP;
1690     BDRVRawState *s = aiocb->bs->opaque;
1691 
1692     if (!s->has_discard) {
1693         return -ENOTSUP;
1694     }
1695 
1696     if (aiocb->aio_type & QEMU_AIO_BLKDEV) {
1697 #ifdef BLKDISCARD
1698         do {
1699             uint64_t range[2] = { aiocb->aio_offset, aiocb->aio_nbytes };
1700             if (ioctl(aiocb->aio_fildes, BLKDISCARD, range) == 0) {
1701                 return 0;
1702             }
1703         } while (errno == EINTR);
1704 
1705         ret = -errno;
1706 #endif
1707     } else {
1708 #ifdef CONFIG_FALLOCATE_PUNCH_HOLE
1709         ret = do_fallocate(s->fd, FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE,
1710                            aiocb->aio_offset, aiocb->aio_nbytes);
1711 #endif
1712     }
1713 
1714     ret = translate_err(ret);
1715     if (ret == -ENOTSUP) {
1716         s->has_discard = false;
1717     }
1718     return ret;
1719 }
1720 
1721 /*
1722  * Help alignment probing by allocating the first block.
1723  *
1724  * When reading with direct I/O from unallocated area on Gluster backed by XFS,
1725  * reading succeeds regardless of request length. In this case we fallback to
1726  * safe alignment which is not optimal. Allocating the first block avoids this
1727  * fallback.
1728  *
1729  * fd may be opened with O_DIRECT, but we don't know the buffer alignment or
1730  * request alignment, so we use safe values.
1731  *
1732  * Returns: 0 on success, -errno on failure. Since this is an optimization,
1733  * caller may ignore failures.
1734  */
1735 static int allocate_first_block(int fd, size_t max_size)
1736 {
1737     size_t write_size = (max_size < MAX_BLOCKSIZE)
1738         ? BDRV_SECTOR_SIZE
1739         : MAX_BLOCKSIZE;
1740     size_t max_align = MAX(MAX_BLOCKSIZE, qemu_real_host_page_size);
1741     void *buf;
1742     ssize_t n;
1743     int ret;
1744 
1745     buf = qemu_memalign(max_align, write_size);
1746     memset(buf, 0, write_size);
1747 
1748     do {
1749         n = pwrite(fd, buf, write_size, 0);
1750     } while (n == -1 && errno == EINTR);
1751 
1752     ret = (n == -1) ? -errno : 0;
1753 
1754     qemu_vfree(buf);
1755     return ret;
1756 }
1757 
1758 static int handle_aiocb_truncate(void *opaque)
1759 {
1760     RawPosixAIOData *aiocb = opaque;
1761     int result = 0;
1762     int64_t current_length = 0;
1763     char *buf = NULL;
1764     struct stat st;
1765     int fd = aiocb->aio_fildes;
1766     int64_t offset = aiocb->aio_offset;
1767     PreallocMode prealloc = aiocb->truncate.prealloc;
1768     Error **errp = aiocb->truncate.errp;
1769 
1770     if (fstat(fd, &st) < 0) {
1771         result = -errno;
1772         error_setg_errno(errp, -result, "Could not stat file");
1773         return result;
1774     }
1775 
1776     current_length = st.st_size;
1777     if (current_length > offset && prealloc != PREALLOC_MODE_OFF) {
1778         error_setg(errp, "Cannot use preallocation for shrinking files");
1779         return -ENOTSUP;
1780     }
1781 
1782     switch (prealloc) {
1783 #ifdef CONFIG_POSIX_FALLOCATE
1784     case PREALLOC_MODE_FALLOC:
1785         /*
1786          * Truncating before posix_fallocate() makes it about twice slower on
1787          * file systems that do not support fallocate(), trying to check if a
1788          * block is allocated before allocating it, so don't do that here.
1789          */
1790         if (offset != current_length) {
1791             result = -posix_fallocate(fd, current_length,
1792                                       offset - current_length);
1793             if (result != 0) {
1794                 /* posix_fallocate() doesn't set errno. */
1795                 error_setg_errno(errp, -result,
1796                                  "Could not preallocate new data");
1797             } else if (current_length == 0) {
1798                 /*
1799                  * posix_fallocate() uses fallocate() if the filesystem
1800                  * supports it, or fallback to manually writing zeroes. If
1801                  * fallocate() was used, unaligned reads from the fallocated
1802                  * area in raw_probe_alignment() will succeed, hence we need to
1803                  * allocate the first block.
1804                  *
1805                  * Optimize future alignment probing; ignore failures.
1806                  */
1807                 allocate_first_block(fd, offset);
1808             }
1809         } else {
1810             result = 0;
1811         }
1812         goto out;
1813 #endif
1814     case PREALLOC_MODE_FULL:
1815     {
1816         int64_t num = 0, left = offset - current_length;
1817         off_t seek_result;
1818 
1819         /*
1820          * Knowing the final size from the beginning could allow the file
1821          * system driver to do less allocations and possibly avoid
1822          * fragmentation of the file.
1823          */
1824         if (ftruncate(fd, offset) != 0) {
1825             result = -errno;
1826             error_setg_errno(errp, -result, "Could not resize file");
1827             goto out;
1828         }
1829 
1830         buf = g_malloc0(65536);
1831 
1832         seek_result = lseek(fd, current_length, SEEK_SET);
1833         if (seek_result < 0) {
1834             result = -errno;
1835             error_setg_errno(errp, -result,
1836                              "Failed to seek to the old end of file");
1837             goto out;
1838         }
1839 
1840         while (left > 0) {
1841             num = MIN(left, 65536);
1842             result = write(fd, buf, num);
1843             if (result < 0) {
1844                 if (errno == EINTR) {
1845                     continue;
1846                 }
1847                 result = -errno;
1848                 error_setg_errno(errp, -result,
1849                                  "Could not write zeros for preallocation");
1850                 goto out;
1851             }
1852             left -= result;
1853         }
1854         if (result >= 0) {
1855             result = fsync(fd);
1856             if (result < 0) {
1857                 result = -errno;
1858                 error_setg_errno(errp, -result,
1859                                  "Could not flush file to disk");
1860                 goto out;
1861             }
1862         }
1863         goto out;
1864     }
1865     case PREALLOC_MODE_OFF:
1866         if (ftruncate(fd, offset) != 0) {
1867             result = -errno;
1868             error_setg_errno(errp, -result, "Could not resize file");
1869         } else if (current_length == 0 && offset > current_length) {
1870             /* Optimize future alignment probing; ignore failures. */
1871             allocate_first_block(fd, offset);
1872         }
1873         return result;
1874     default:
1875         result = -ENOTSUP;
1876         error_setg(errp, "Unsupported preallocation mode: %s",
1877                    PreallocMode_str(prealloc));
1878         return result;
1879     }
1880 
1881 out:
1882     if (result < 0) {
1883         if (ftruncate(fd, current_length) < 0) {
1884             error_report("Failed to restore old file length: %s",
1885                          strerror(errno));
1886         }
1887     }
1888 
1889     g_free(buf);
1890     return result;
1891 }
1892 
1893 static int coroutine_fn raw_thread_pool_submit(BlockDriverState *bs,
1894                                                ThreadPoolFunc func, void *arg)
1895 {
1896     /* @bs can be NULL, bdrv_get_aio_context() returns the main context then */
1897     ThreadPool *pool = aio_get_thread_pool(bdrv_get_aio_context(bs));
1898     return thread_pool_submit_co(pool, func, arg);
1899 }
1900 
1901 static int coroutine_fn raw_co_prw(BlockDriverState *bs, uint64_t offset,
1902                                    uint64_t bytes, QEMUIOVector *qiov, int type)
1903 {
1904     BDRVRawState *s = bs->opaque;
1905     RawPosixAIOData acb;
1906 
1907     if (fd_open(bs) < 0)
1908         return -EIO;
1909 
1910     /*
1911      * When using O_DIRECT, the request must be aligned to be able to use
1912      * either libaio or io_uring interface. If not fail back to regular thread
1913      * pool read/write code which emulates this for us if we
1914      * set QEMU_AIO_MISALIGNED.
1915      */
1916     if (s->needs_alignment && !bdrv_qiov_is_aligned(bs, qiov)) {
1917         type |= QEMU_AIO_MISALIGNED;
1918 #ifdef CONFIG_LINUX_IO_URING
1919     } else if (s->use_linux_io_uring) {
1920         LuringState *aio = aio_get_linux_io_uring(bdrv_get_aio_context(bs));
1921         assert(qiov->size == bytes);
1922         return luring_co_submit(bs, aio, s->fd, offset, qiov, type);
1923 #endif
1924 #ifdef CONFIG_LINUX_AIO
1925     } else if (s->use_linux_aio) {
1926         LinuxAioState *aio = aio_get_linux_aio(bdrv_get_aio_context(bs));
1927         assert(qiov->size == bytes);
1928         return laio_co_submit(bs, aio, s->fd, offset, qiov, type);
1929 #endif
1930     }
1931 
1932     acb = (RawPosixAIOData) {
1933         .bs             = bs,
1934         .aio_fildes     = s->fd,
1935         .aio_type       = type,
1936         .aio_offset     = offset,
1937         .aio_nbytes     = bytes,
1938         .io             = {
1939             .iov            = qiov->iov,
1940             .niov           = qiov->niov,
1941         },
1942     };
1943 
1944     assert(qiov->size == bytes);
1945     return raw_thread_pool_submit(bs, handle_aiocb_rw, &acb);
1946 }
1947 
1948 static int coroutine_fn raw_co_preadv(BlockDriverState *bs, uint64_t offset,
1949                                       uint64_t bytes, QEMUIOVector *qiov,
1950                                       int flags)
1951 {
1952     return raw_co_prw(bs, offset, bytes, qiov, QEMU_AIO_READ);
1953 }
1954 
1955 static int coroutine_fn raw_co_pwritev(BlockDriverState *bs, uint64_t offset,
1956                                        uint64_t bytes, QEMUIOVector *qiov,
1957                                        int flags)
1958 {
1959     assert(flags == 0);
1960     return raw_co_prw(bs, offset, bytes, qiov, QEMU_AIO_WRITE);
1961 }
1962 
1963 static void raw_aio_plug(BlockDriverState *bs)
1964 {
1965     BDRVRawState __attribute__((unused)) *s = bs->opaque;
1966 #ifdef CONFIG_LINUX_AIO
1967     if (s->use_linux_aio) {
1968         LinuxAioState *aio = aio_get_linux_aio(bdrv_get_aio_context(bs));
1969         laio_io_plug(bs, aio);
1970     }
1971 #endif
1972 #ifdef CONFIG_LINUX_IO_URING
1973     if (s->use_linux_io_uring) {
1974         LuringState *aio = aio_get_linux_io_uring(bdrv_get_aio_context(bs));
1975         luring_io_plug(bs, aio);
1976     }
1977 #endif
1978 }
1979 
1980 static void raw_aio_unplug(BlockDriverState *bs)
1981 {
1982     BDRVRawState __attribute__((unused)) *s = bs->opaque;
1983 #ifdef CONFIG_LINUX_AIO
1984     if (s->use_linux_aio) {
1985         LinuxAioState *aio = aio_get_linux_aio(bdrv_get_aio_context(bs));
1986         laio_io_unplug(bs, aio);
1987     }
1988 #endif
1989 #ifdef CONFIG_LINUX_IO_URING
1990     if (s->use_linux_io_uring) {
1991         LuringState *aio = aio_get_linux_io_uring(bdrv_get_aio_context(bs));
1992         luring_io_unplug(bs, aio);
1993     }
1994 #endif
1995 }
1996 
1997 static int raw_co_flush_to_disk(BlockDriverState *bs)
1998 {
1999     BDRVRawState *s = bs->opaque;
2000     RawPosixAIOData acb;
2001     int ret;
2002 
2003     ret = fd_open(bs);
2004     if (ret < 0) {
2005         return ret;
2006     }
2007 
2008     acb = (RawPosixAIOData) {
2009         .bs             = bs,
2010         .aio_fildes     = s->fd,
2011         .aio_type       = QEMU_AIO_FLUSH,
2012     };
2013 
2014 #ifdef CONFIG_LINUX_IO_URING
2015     if (s->use_linux_io_uring) {
2016         LuringState *aio = aio_get_linux_io_uring(bdrv_get_aio_context(bs));
2017         return luring_co_submit(bs, aio, s->fd, 0, NULL, QEMU_AIO_FLUSH);
2018     }
2019 #endif
2020     return raw_thread_pool_submit(bs, handle_aiocb_flush, &acb);
2021 }
2022 
2023 static void raw_aio_attach_aio_context(BlockDriverState *bs,
2024                                        AioContext *new_context)
2025 {
2026     BDRVRawState __attribute__((unused)) *s = bs->opaque;
2027 #ifdef CONFIG_LINUX_AIO
2028     if (s->use_linux_aio) {
2029         Error *local_err = NULL;
2030         if (!aio_setup_linux_aio(new_context, &local_err)) {
2031             error_reportf_err(local_err, "Unable to use native AIO, "
2032                                          "falling back to thread pool: ");
2033             s->use_linux_aio = false;
2034         }
2035     }
2036 #endif
2037 #ifdef CONFIG_LINUX_IO_URING
2038     if (s->use_linux_io_uring) {
2039         Error *local_err;
2040         if (!aio_setup_linux_io_uring(new_context, &local_err)) {
2041             error_reportf_err(local_err, "Unable to use linux io_uring, "
2042                                          "falling back to thread pool: ");
2043             s->use_linux_io_uring = false;
2044         }
2045     }
2046 #endif
2047 }
2048 
2049 static void raw_close(BlockDriverState *bs)
2050 {
2051     BDRVRawState *s = bs->opaque;
2052 
2053     if (s->fd >= 0) {
2054         qemu_close(s->fd);
2055         s->fd = -1;
2056     }
2057 }
2058 
2059 /**
2060  * Truncates the given regular file @fd to @offset and, when growing, fills the
2061  * new space according to @prealloc.
2062  *
2063  * Returns: 0 on success, -errno on failure.
2064  */
2065 static int coroutine_fn
2066 raw_regular_truncate(BlockDriverState *bs, int fd, int64_t offset,
2067                      PreallocMode prealloc, Error **errp)
2068 {
2069     RawPosixAIOData acb;
2070 
2071     acb = (RawPosixAIOData) {
2072         .bs             = bs,
2073         .aio_fildes     = fd,
2074         .aio_type       = QEMU_AIO_TRUNCATE,
2075         .aio_offset     = offset,
2076         .truncate       = {
2077             .prealloc       = prealloc,
2078             .errp           = errp,
2079         },
2080     };
2081 
2082     return raw_thread_pool_submit(bs, handle_aiocb_truncate, &acb);
2083 }
2084 
2085 static int coroutine_fn raw_co_truncate(BlockDriverState *bs, int64_t offset,
2086                                         bool exact, PreallocMode prealloc,
2087                                         BdrvRequestFlags flags, Error **errp)
2088 {
2089     BDRVRawState *s = bs->opaque;
2090     struct stat st;
2091     int ret;
2092 
2093     if (fstat(s->fd, &st)) {
2094         ret = -errno;
2095         error_setg_errno(errp, -ret, "Failed to fstat() the file");
2096         return ret;
2097     }
2098 
2099     if (S_ISREG(st.st_mode)) {
2100         /* Always resizes to the exact @offset */
2101         return raw_regular_truncate(bs, s->fd, offset, prealloc, errp);
2102     }
2103 
2104     if (prealloc != PREALLOC_MODE_OFF) {
2105         error_setg(errp, "Preallocation mode '%s' unsupported for this "
2106                    "non-regular file", PreallocMode_str(prealloc));
2107         return -ENOTSUP;
2108     }
2109 
2110     if (S_ISCHR(st.st_mode) || S_ISBLK(st.st_mode)) {
2111         int64_t cur_length = raw_getlength(bs);
2112 
2113         if (offset != cur_length && exact) {
2114             error_setg(errp, "Cannot resize device files");
2115             return -ENOTSUP;
2116         } else if (offset > cur_length) {
2117             error_setg(errp, "Cannot grow device files");
2118             return -EINVAL;
2119         }
2120     } else {
2121         error_setg(errp, "Resizing this file is not supported");
2122         return -ENOTSUP;
2123     }
2124 
2125     return 0;
2126 }
2127 
2128 #ifdef __OpenBSD__
2129 static int64_t raw_getlength(BlockDriverState *bs)
2130 {
2131     BDRVRawState *s = bs->opaque;
2132     int fd = s->fd;
2133     struct stat st;
2134 
2135     if (fstat(fd, &st))
2136         return -errno;
2137     if (S_ISCHR(st.st_mode) || S_ISBLK(st.st_mode)) {
2138         struct disklabel dl;
2139 
2140         if (ioctl(fd, DIOCGDINFO, &dl))
2141             return -errno;
2142         return (uint64_t)dl.d_secsize *
2143             dl.d_partitions[DISKPART(st.st_rdev)].p_size;
2144     } else
2145         return st.st_size;
2146 }
2147 #elif defined(__NetBSD__)
2148 static int64_t raw_getlength(BlockDriverState *bs)
2149 {
2150     BDRVRawState *s = bs->opaque;
2151     int fd = s->fd;
2152     struct stat st;
2153 
2154     if (fstat(fd, &st))
2155         return -errno;
2156     if (S_ISCHR(st.st_mode) || S_ISBLK(st.st_mode)) {
2157         struct dkwedge_info dkw;
2158 
2159         if (ioctl(fd, DIOCGWEDGEINFO, &dkw) != -1) {
2160             return dkw.dkw_size * 512;
2161         } else {
2162             struct disklabel dl;
2163 
2164             if (ioctl(fd, DIOCGDINFO, &dl))
2165                 return -errno;
2166             return (uint64_t)dl.d_secsize *
2167                 dl.d_partitions[DISKPART(st.st_rdev)].p_size;
2168         }
2169     } else
2170         return st.st_size;
2171 }
2172 #elif defined(__sun__)
2173 static int64_t raw_getlength(BlockDriverState *bs)
2174 {
2175     BDRVRawState *s = bs->opaque;
2176     struct dk_minfo minfo;
2177     int ret;
2178     int64_t size;
2179 
2180     ret = fd_open(bs);
2181     if (ret < 0) {
2182         return ret;
2183     }
2184 
2185     /*
2186      * Use the DKIOCGMEDIAINFO ioctl to read the size.
2187      */
2188     ret = ioctl(s->fd, DKIOCGMEDIAINFO, &minfo);
2189     if (ret != -1) {
2190         return minfo.dki_lbsize * minfo.dki_capacity;
2191     }
2192 
2193     /*
2194      * There are reports that lseek on some devices fails, but
2195      * irc discussion said that contingency on contingency was overkill.
2196      */
2197     size = lseek(s->fd, 0, SEEK_END);
2198     if (size < 0) {
2199         return -errno;
2200     }
2201     return size;
2202 }
2203 #elif defined(CONFIG_BSD)
2204 static int64_t raw_getlength(BlockDriverState *bs)
2205 {
2206     BDRVRawState *s = bs->opaque;
2207     int fd = s->fd;
2208     int64_t size;
2209     struct stat sb;
2210 #if defined (__FreeBSD__) || defined(__FreeBSD_kernel__)
2211     int reopened = 0;
2212 #endif
2213     int ret;
2214 
2215     ret = fd_open(bs);
2216     if (ret < 0)
2217         return ret;
2218 
2219 #if defined (__FreeBSD__) || defined(__FreeBSD_kernel__)
2220 again:
2221 #endif
2222     if (!fstat(fd, &sb) && (S_IFCHR & sb.st_mode)) {
2223 #ifdef DIOCGMEDIASIZE
2224         if (ioctl(fd, DIOCGMEDIASIZE, (off_t *)&size))
2225 #elif defined(DIOCGPART)
2226         {
2227                 struct partinfo pi;
2228                 if (ioctl(fd, DIOCGPART, &pi) == 0)
2229                         size = pi.media_size;
2230                 else
2231                         size = 0;
2232         }
2233         if (size == 0)
2234 #endif
2235 #if defined(__APPLE__) && defined(__MACH__)
2236         {
2237             uint64_t sectors = 0;
2238             uint32_t sector_size = 0;
2239 
2240             if (ioctl(fd, DKIOCGETBLOCKCOUNT, &sectors) == 0
2241                && ioctl(fd, DKIOCGETBLOCKSIZE, &sector_size) == 0) {
2242                 size = sectors * sector_size;
2243             } else {
2244                 size = lseek(fd, 0LL, SEEK_END);
2245                 if (size < 0) {
2246                     return -errno;
2247                 }
2248             }
2249         }
2250 #else
2251         size = lseek(fd, 0LL, SEEK_END);
2252         if (size < 0) {
2253             return -errno;
2254         }
2255 #endif
2256 #if defined(__FreeBSD__) || defined(__FreeBSD_kernel__)
2257         switch(s->type) {
2258         case FTYPE_CD:
2259             /* XXX FreeBSD acd returns UINT_MAX sectors for an empty drive */
2260             if (size == 2048LL * (unsigned)-1)
2261                 size = 0;
2262             /* XXX no disc?  maybe we need to reopen... */
2263             if (size <= 0 && !reopened && cdrom_reopen(bs) >= 0) {
2264                 reopened = 1;
2265                 goto again;
2266             }
2267         }
2268 #endif
2269     } else {
2270         size = lseek(fd, 0, SEEK_END);
2271         if (size < 0) {
2272             return -errno;
2273         }
2274     }
2275     return size;
2276 }
2277 #else
2278 static int64_t raw_getlength(BlockDriverState *bs)
2279 {
2280     BDRVRawState *s = bs->opaque;
2281     int ret;
2282     int64_t size;
2283 
2284     ret = fd_open(bs);
2285     if (ret < 0) {
2286         return ret;
2287     }
2288 
2289     size = lseek(s->fd, 0, SEEK_END);
2290     if (size < 0) {
2291         return -errno;
2292     }
2293     return size;
2294 }
2295 #endif
2296 
2297 static int64_t raw_get_allocated_file_size(BlockDriverState *bs)
2298 {
2299     struct stat st;
2300     BDRVRawState *s = bs->opaque;
2301 
2302     if (fstat(s->fd, &st) < 0) {
2303         return -errno;
2304     }
2305     return (int64_t)st.st_blocks * 512;
2306 }
2307 
2308 static int coroutine_fn
2309 raw_co_create(BlockdevCreateOptions *options, Error **errp)
2310 {
2311     BlockdevCreateOptionsFile *file_opts;
2312     Error *local_err = NULL;
2313     int fd;
2314     uint64_t perm, shared;
2315     int result = 0;
2316 
2317     /* Validate options and set default values */
2318     assert(options->driver == BLOCKDEV_DRIVER_FILE);
2319     file_opts = &options->u.file;
2320 
2321     if (!file_opts->has_nocow) {
2322         file_opts->nocow = false;
2323     }
2324     if (!file_opts->has_preallocation) {
2325         file_opts->preallocation = PREALLOC_MODE_OFF;
2326     }
2327 
2328     /* Create file */
2329     fd = qemu_open(file_opts->filename, O_RDWR | O_CREAT | O_BINARY, 0644);
2330     if (fd < 0) {
2331         result = -errno;
2332         error_setg_errno(errp, -result, "Could not create file");
2333         goto out;
2334     }
2335 
2336     /* Take permissions: We want to discard everything, so we need
2337      * BLK_PERM_WRITE; and truncation to the desired size requires
2338      * BLK_PERM_RESIZE.
2339      * On the other hand, we cannot share the RESIZE permission
2340      * because we promise that after this function, the file has the
2341      * size given in the options.  If someone else were to resize it
2342      * concurrently, we could not guarantee that.
2343      * Note that after this function, we can no longer guarantee that
2344      * the file is not touched by a third party, so it may be resized
2345      * then. */
2346     perm = BLK_PERM_WRITE | BLK_PERM_RESIZE;
2347     shared = BLK_PERM_ALL & ~BLK_PERM_RESIZE;
2348 
2349     /* Step one: Take locks */
2350     result = raw_apply_lock_bytes(NULL, fd, perm, ~shared, false, errp);
2351     if (result < 0) {
2352         goto out_close;
2353     }
2354 
2355     /* Step two: Check that nobody else has taken conflicting locks */
2356     result = raw_check_lock_bytes(fd, perm, shared, errp);
2357     if (result < 0) {
2358         error_append_hint(errp,
2359                           "Is another process using the image [%s]?\n",
2360                           file_opts->filename);
2361         goto out_unlock;
2362     }
2363 
2364     /* Clear the file by truncating it to 0 */
2365     result = raw_regular_truncate(NULL, fd, 0, PREALLOC_MODE_OFF, errp);
2366     if (result < 0) {
2367         goto out_unlock;
2368     }
2369 
2370     if (file_opts->nocow) {
2371 #ifdef __linux__
2372         /* Set NOCOW flag to solve performance issue on fs like btrfs.
2373          * This is an optimisation. The FS_IOC_SETFLAGS ioctl return value
2374          * will be ignored since any failure of this operation should not
2375          * block the left work.
2376          */
2377         int attr;
2378         if (ioctl(fd, FS_IOC_GETFLAGS, &attr) == 0) {
2379             attr |= FS_NOCOW_FL;
2380             ioctl(fd, FS_IOC_SETFLAGS, &attr);
2381         }
2382 #endif
2383     }
2384 
2385     /* Resize and potentially preallocate the file to the desired
2386      * final size */
2387     result = raw_regular_truncate(NULL, fd, file_opts->size,
2388                                   file_opts->preallocation, errp);
2389     if (result < 0) {
2390         goto out_unlock;
2391     }
2392 
2393 out_unlock:
2394     raw_apply_lock_bytes(NULL, fd, 0, 0, true, &local_err);
2395     if (local_err) {
2396         /* The above call should not fail, and if it does, that does
2397          * not mean the whole creation operation has failed.  So
2398          * report it the user for their convenience, but do not report
2399          * it to the caller. */
2400         warn_report_err(local_err);
2401     }
2402 
2403 out_close:
2404     if (qemu_close(fd) != 0 && result == 0) {
2405         result = -errno;
2406         error_setg_errno(errp, -result, "Could not close the new file");
2407     }
2408 out:
2409     return result;
2410 }
2411 
2412 static int coroutine_fn raw_co_create_opts(BlockDriver *drv,
2413                                            const char *filename,
2414                                            QemuOpts *opts,
2415                                            Error **errp)
2416 {
2417     BlockdevCreateOptions options;
2418     int64_t total_size = 0;
2419     bool nocow = false;
2420     PreallocMode prealloc;
2421     char *buf = NULL;
2422     Error *local_err = NULL;
2423 
2424     /* Skip file: protocol prefix */
2425     strstart(filename, "file:", &filename);
2426 
2427     /* Read out options */
2428     total_size = ROUND_UP(qemu_opt_get_size_del(opts, BLOCK_OPT_SIZE, 0),
2429                           BDRV_SECTOR_SIZE);
2430     nocow = qemu_opt_get_bool(opts, BLOCK_OPT_NOCOW, false);
2431     buf = qemu_opt_get_del(opts, BLOCK_OPT_PREALLOC);
2432     prealloc = qapi_enum_parse(&PreallocMode_lookup, buf,
2433                                PREALLOC_MODE_OFF, &local_err);
2434     g_free(buf);
2435     if (local_err) {
2436         error_propagate(errp, local_err);
2437         return -EINVAL;
2438     }
2439 
2440     options = (BlockdevCreateOptions) {
2441         .driver     = BLOCKDEV_DRIVER_FILE,
2442         .u.file     = {
2443             .filename           = (char *) filename,
2444             .size               = total_size,
2445             .has_preallocation  = true,
2446             .preallocation      = prealloc,
2447             .has_nocow          = true,
2448             .nocow              = nocow,
2449         },
2450     };
2451     return raw_co_create(&options, errp);
2452 }
2453 
2454 static int coroutine_fn raw_co_delete_file(BlockDriverState *bs,
2455                                            Error **errp)
2456 {
2457     struct stat st;
2458     int ret;
2459 
2460     if (!(stat(bs->filename, &st) == 0) || !S_ISREG(st.st_mode)) {
2461         error_setg_errno(errp, ENOENT, "%s is not a regular file",
2462                          bs->filename);
2463         return -ENOENT;
2464     }
2465 
2466     ret = unlink(bs->filename);
2467     if (ret < 0) {
2468         ret = -errno;
2469         error_setg_errno(errp, -ret, "Error when deleting file %s",
2470                          bs->filename);
2471     }
2472 
2473     return ret;
2474 }
2475 
2476 /*
2477  * Find allocation range in @bs around offset @start.
2478  * May change underlying file descriptor's file offset.
2479  * If @start is not in a hole, store @start in @data, and the
2480  * beginning of the next hole in @hole, and return 0.
2481  * If @start is in a non-trailing hole, store @start in @hole and the
2482  * beginning of the next non-hole in @data, and return 0.
2483  * If @start is in a trailing hole or beyond EOF, return -ENXIO.
2484  * If we can't find out, return a negative errno other than -ENXIO.
2485  */
2486 static int find_allocation(BlockDriverState *bs, off_t start,
2487                            off_t *data, off_t *hole)
2488 {
2489 #if defined SEEK_HOLE && defined SEEK_DATA
2490     BDRVRawState *s = bs->opaque;
2491     off_t offs;
2492 
2493     /*
2494      * SEEK_DATA cases:
2495      * D1. offs == start: start is in data
2496      * D2. offs > start: start is in a hole, next data at offs
2497      * D3. offs < 0, errno = ENXIO: either start is in a trailing hole
2498      *                              or start is beyond EOF
2499      *     If the latter happens, the file has been truncated behind
2500      *     our back since we opened it.  All bets are off then.
2501      *     Treating like a trailing hole is simplest.
2502      * D4. offs < 0, errno != ENXIO: we learned nothing
2503      */
2504     offs = lseek(s->fd, start, SEEK_DATA);
2505     if (offs < 0) {
2506         return -errno;          /* D3 or D4 */
2507     }
2508 
2509     if (offs < start) {
2510         /* This is not a valid return by lseek().  We are safe to just return
2511          * -EIO in this case, and we'll treat it like D4. */
2512         return -EIO;
2513     }
2514 
2515     if (offs > start) {
2516         /* D2: in hole, next data at offs */
2517         *hole = start;
2518         *data = offs;
2519         return 0;
2520     }
2521 
2522     /* D1: in data, end not yet known */
2523 
2524     /*
2525      * SEEK_HOLE cases:
2526      * H1. offs == start: start is in a hole
2527      *     If this happens here, a hole has been dug behind our back
2528      *     since the previous lseek().
2529      * H2. offs > start: either start is in data, next hole at offs,
2530      *                   or start is in trailing hole, EOF at offs
2531      *     Linux treats trailing holes like any other hole: offs ==
2532      *     start.  Solaris seeks to EOF instead: offs > start (blech).
2533      *     If that happens here, a hole has been dug behind our back
2534      *     since the previous lseek().
2535      * H3. offs < 0, errno = ENXIO: start is beyond EOF
2536      *     If this happens, the file has been truncated behind our
2537      *     back since we opened it.  Treat it like a trailing hole.
2538      * H4. offs < 0, errno != ENXIO: we learned nothing
2539      *     Pretend we know nothing at all, i.e. "forget" about D1.
2540      */
2541     offs = lseek(s->fd, start, SEEK_HOLE);
2542     if (offs < 0) {
2543         return -errno;          /* D1 and (H3 or H4) */
2544     }
2545 
2546     if (offs < start) {
2547         /* This is not a valid return by lseek().  We are safe to just return
2548          * -EIO in this case, and we'll treat it like H4. */
2549         return -EIO;
2550     }
2551 
2552     if (offs > start) {
2553         /*
2554          * D1 and H2: either in data, next hole at offs, or it was in
2555          * data but is now in a trailing hole.  In the latter case,
2556          * all bets are off.  Treating it as if it there was data all
2557          * the way to EOF is safe, so simply do that.
2558          */
2559         *data = start;
2560         *hole = offs;
2561         return 0;
2562     }
2563 
2564     /* D1 and H1 */
2565     return -EBUSY;
2566 #else
2567     return -ENOTSUP;
2568 #endif
2569 }
2570 
2571 /*
2572  * Returns the allocation status of the specified offset.
2573  *
2574  * The block layer guarantees 'offset' and 'bytes' are within bounds.
2575  *
2576  * 'pnum' is set to the number of bytes (including and immediately following
2577  * the specified offset) that are known to be in the same
2578  * allocated/unallocated state.
2579  *
2580  * 'bytes' is the max value 'pnum' should be set to.
2581  */
2582 static int coroutine_fn raw_co_block_status(BlockDriverState *bs,
2583                                             bool want_zero,
2584                                             int64_t offset,
2585                                             int64_t bytes, int64_t *pnum,
2586                                             int64_t *map,
2587                                             BlockDriverState **file)
2588 {
2589     off_t data = 0, hole = 0;
2590     int ret;
2591 
2592     assert(QEMU_IS_ALIGNED(offset | bytes, bs->bl.request_alignment));
2593 
2594     ret = fd_open(bs);
2595     if (ret < 0) {
2596         return ret;
2597     }
2598 
2599     if (!want_zero) {
2600         *pnum = bytes;
2601         *map = offset;
2602         *file = bs;
2603         return BDRV_BLOCK_DATA | BDRV_BLOCK_OFFSET_VALID;
2604     }
2605 
2606     ret = find_allocation(bs, offset, &data, &hole);
2607     if (ret == -ENXIO) {
2608         /* Trailing hole */
2609         *pnum = bytes;
2610         ret = BDRV_BLOCK_ZERO;
2611     } else if (ret < 0) {
2612         /* No info available, so pretend there are no holes */
2613         *pnum = bytes;
2614         ret = BDRV_BLOCK_DATA;
2615     } else if (data == offset) {
2616         /* On a data extent, compute bytes to the end of the extent,
2617          * possibly including a partial sector at EOF. */
2618         *pnum = MIN(bytes, hole - offset);
2619 
2620         /*
2621          * We are not allowed to return partial sectors, though, so
2622          * round up if necessary.
2623          */
2624         if (!QEMU_IS_ALIGNED(*pnum, bs->bl.request_alignment)) {
2625             int64_t file_length = raw_getlength(bs);
2626             if (file_length > 0) {
2627                 /* Ignore errors, this is just a safeguard */
2628                 assert(hole == file_length);
2629             }
2630             *pnum = ROUND_UP(*pnum, bs->bl.request_alignment);
2631         }
2632 
2633         ret = BDRV_BLOCK_DATA;
2634     } else {
2635         /* On a hole, compute bytes to the beginning of the next extent.  */
2636         assert(hole == offset);
2637         *pnum = MIN(bytes, data - offset);
2638         ret = BDRV_BLOCK_ZERO;
2639     }
2640     *map = offset;
2641     *file = bs;
2642     return ret | BDRV_BLOCK_OFFSET_VALID;
2643 }
2644 
2645 #if defined(__linux__)
2646 /* Verify that the file is not in the page cache */
2647 static void check_cache_dropped(BlockDriverState *bs, Error **errp)
2648 {
2649     const size_t window_size = 128 * 1024 * 1024;
2650     BDRVRawState *s = bs->opaque;
2651     void *window = NULL;
2652     size_t length = 0;
2653     unsigned char *vec;
2654     size_t page_size;
2655     off_t offset;
2656     off_t end;
2657 
2658     /* mincore(2) page status information requires 1 byte per page */
2659     page_size = sysconf(_SC_PAGESIZE);
2660     vec = g_malloc(DIV_ROUND_UP(window_size, page_size));
2661 
2662     end = raw_getlength(bs);
2663 
2664     for (offset = 0; offset < end; offset += window_size) {
2665         void *new_window;
2666         size_t new_length;
2667         size_t vec_end;
2668         size_t i;
2669         int ret;
2670 
2671         /* Unmap previous window if size has changed */
2672         new_length = MIN(end - offset, window_size);
2673         if (new_length != length) {
2674             munmap(window, length);
2675             window = NULL;
2676             length = 0;
2677         }
2678 
2679         new_window = mmap(window, new_length, PROT_NONE, MAP_PRIVATE,
2680                           s->fd, offset);
2681         if (new_window == MAP_FAILED) {
2682             error_setg_errno(errp, errno, "mmap failed");
2683             break;
2684         }
2685 
2686         window = new_window;
2687         length = new_length;
2688 
2689         ret = mincore(window, length, vec);
2690         if (ret < 0) {
2691             error_setg_errno(errp, errno, "mincore failed");
2692             break;
2693         }
2694 
2695         vec_end = DIV_ROUND_UP(length, page_size);
2696         for (i = 0; i < vec_end; i++) {
2697             if (vec[i] & 0x1) {
2698                 break;
2699             }
2700         }
2701         if (i < vec_end) {
2702             error_setg(errp, "page cache still in use!");
2703             break;
2704         }
2705     }
2706 
2707     if (window) {
2708         munmap(window, length);
2709     }
2710 
2711     g_free(vec);
2712 }
2713 #endif /* __linux__ */
2714 
2715 static void coroutine_fn raw_co_invalidate_cache(BlockDriverState *bs,
2716                                                  Error **errp)
2717 {
2718     BDRVRawState *s = bs->opaque;
2719     int ret;
2720 
2721     ret = fd_open(bs);
2722     if (ret < 0) {
2723         error_setg_errno(errp, -ret, "The file descriptor is not open");
2724         return;
2725     }
2726 
2727     if (!s->drop_cache) {
2728         return;
2729     }
2730 
2731     if (s->open_flags & O_DIRECT) {
2732         return; /* No host kernel page cache */
2733     }
2734 
2735 #if defined(__linux__)
2736     /* This sets the scene for the next syscall... */
2737     ret = bdrv_co_flush(bs);
2738     if (ret < 0) {
2739         error_setg_errno(errp, -ret, "flush failed");
2740         return;
2741     }
2742 
2743     /* Linux does not invalidate pages that are dirty, locked, or mmapped by a
2744      * process.  These limitations are okay because we just fsynced the file,
2745      * we don't use mmap, and the file should not be in use by other processes.
2746      */
2747     ret = posix_fadvise(s->fd, 0, 0, POSIX_FADV_DONTNEED);
2748     if (ret != 0) { /* the return value is a positive errno */
2749         error_setg_errno(errp, ret, "fadvise failed");
2750         return;
2751     }
2752 
2753     if (s->check_cache_dropped) {
2754         check_cache_dropped(bs, errp);
2755     }
2756 #else /* __linux__ */
2757     /* Do nothing.  Live migration to a remote host with cache.direct=off is
2758      * unsupported on other host operating systems.  Cache consistency issues
2759      * may occur but no error is reported here, partly because that's the
2760      * historical behavior and partly because it's hard to differentiate valid
2761      * configurations that should not cause errors.
2762      */
2763 #endif /* !__linux__ */
2764 }
2765 
2766 static void raw_account_discard(BDRVRawState *s, uint64_t nbytes, int ret)
2767 {
2768     if (ret) {
2769         s->stats.discard_nb_failed++;
2770     } else {
2771         s->stats.discard_nb_ok++;
2772         s->stats.discard_bytes_ok += nbytes;
2773     }
2774 }
2775 
2776 static coroutine_fn int
2777 raw_do_pdiscard(BlockDriverState *bs, int64_t offset, int bytes, bool blkdev)
2778 {
2779     BDRVRawState *s = bs->opaque;
2780     RawPosixAIOData acb;
2781     int ret;
2782 
2783     acb = (RawPosixAIOData) {
2784         .bs             = bs,
2785         .aio_fildes     = s->fd,
2786         .aio_type       = QEMU_AIO_DISCARD,
2787         .aio_offset     = offset,
2788         .aio_nbytes     = bytes,
2789     };
2790 
2791     if (blkdev) {
2792         acb.aio_type |= QEMU_AIO_BLKDEV;
2793     }
2794 
2795     ret = raw_thread_pool_submit(bs, handle_aiocb_discard, &acb);
2796     raw_account_discard(s, bytes, ret);
2797     return ret;
2798 }
2799 
2800 static coroutine_fn int
2801 raw_co_pdiscard(BlockDriverState *bs, int64_t offset, int bytes)
2802 {
2803     return raw_do_pdiscard(bs, offset, bytes, false);
2804 }
2805 
2806 static int coroutine_fn
2807 raw_do_pwrite_zeroes(BlockDriverState *bs, int64_t offset, int bytes,
2808                      BdrvRequestFlags flags, bool blkdev)
2809 {
2810     BDRVRawState *s = bs->opaque;
2811     RawPosixAIOData acb;
2812     ThreadPoolFunc *handler;
2813 
2814 #ifdef CONFIG_FALLOCATE
2815     if (offset + bytes > bs->total_sectors * BDRV_SECTOR_SIZE) {
2816         BdrvTrackedRequest *req;
2817         uint64_t end;
2818 
2819         /*
2820          * This is a workaround for a bug in the Linux XFS driver,
2821          * where writes submitted through the AIO interface will be
2822          * discarded if they happen beyond a concurrently running
2823          * fallocate() that increases the file length (i.e., both the
2824          * write and the fallocate() happen beyond the EOF).
2825          *
2826          * To work around it, we extend the tracked request for this
2827          * zero write until INT64_MAX (effectively infinity), and mark
2828          * it as serializing.
2829          *
2830          * We have to enable this workaround for all filesystems and
2831          * AIO modes (not just XFS with aio=native), because for
2832          * remote filesystems we do not know the host configuration.
2833          */
2834 
2835         req = bdrv_co_get_self_request(bs);
2836         assert(req);
2837         assert(req->type == BDRV_TRACKED_WRITE);
2838         assert(req->offset <= offset);
2839         assert(req->offset + req->bytes >= offset + bytes);
2840 
2841         end = INT64_MAX & -(uint64_t)bs->bl.request_alignment;
2842         req->bytes = end - req->offset;
2843         req->overlap_bytes = req->bytes;
2844 
2845         bdrv_mark_request_serialising(req, bs->bl.request_alignment);
2846     }
2847 #endif
2848 
2849     acb = (RawPosixAIOData) {
2850         .bs             = bs,
2851         .aio_fildes     = s->fd,
2852         .aio_type       = QEMU_AIO_WRITE_ZEROES,
2853         .aio_offset     = offset,
2854         .aio_nbytes     = bytes,
2855     };
2856 
2857     if (blkdev) {
2858         acb.aio_type |= QEMU_AIO_BLKDEV;
2859     }
2860     if (flags & BDRV_REQ_NO_FALLBACK) {
2861         acb.aio_type |= QEMU_AIO_NO_FALLBACK;
2862     }
2863 
2864     if (flags & BDRV_REQ_MAY_UNMAP) {
2865         acb.aio_type |= QEMU_AIO_DISCARD;
2866         handler = handle_aiocb_write_zeroes_unmap;
2867     } else {
2868         handler = handle_aiocb_write_zeroes;
2869     }
2870 
2871     return raw_thread_pool_submit(bs, handler, &acb);
2872 }
2873 
2874 static int coroutine_fn raw_co_pwrite_zeroes(
2875     BlockDriverState *bs, int64_t offset,
2876     int bytes, BdrvRequestFlags flags)
2877 {
2878     return raw_do_pwrite_zeroes(bs, offset, bytes, flags, false);
2879 }
2880 
2881 static int raw_get_info(BlockDriverState *bs, BlockDriverInfo *bdi)
2882 {
2883     BDRVRawState *s = bs->opaque;
2884 
2885     bdi->unallocated_blocks_are_zero = s->discard_zeroes;
2886     return 0;
2887 }
2888 
2889 static BlockStatsSpecificFile get_blockstats_specific_file(BlockDriverState *bs)
2890 {
2891     BDRVRawState *s = bs->opaque;
2892     return (BlockStatsSpecificFile) {
2893         .discard_nb_ok = s->stats.discard_nb_ok,
2894         .discard_nb_failed = s->stats.discard_nb_failed,
2895         .discard_bytes_ok = s->stats.discard_bytes_ok,
2896     };
2897 }
2898 
2899 static BlockStatsSpecific *raw_get_specific_stats(BlockDriverState *bs)
2900 {
2901     BlockStatsSpecific *stats = g_new(BlockStatsSpecific, 1);
2902 
2903     stats->driver = BLOCKDEV_DRIVER_FILE;
2904     stats->u.file = get_blockstats_specific_file(bs);
2905 
2906     return stats;
2907 }
2908 
2909 static BlockStatsSpecific *hdev_get_specific_stats(BlockDriverState *bs)
2910 {
2911     BlockStatsSpecific *stats = g_new(BlockStatsSpecific, 1);
2912 
2913     stats->driver = BLOCKDEV_DRIVER_HOST_DEVICE;
2914     stats->u.host_device = get_blockstats_specific_file(bs);
2915 
2916     return stats;
2917 }
2918 
2919 static QemuOptsList raw_create_opts = {
2920     .name = "raw-create-opts",
2921     .head = QTAILQ_HEAD_INITIALIZER(raw_create_opts.head),
2922     .desc = {
2923         {
2924             .name = BLOCK_OPT_SIZE,
2925             .type = QEMU_OPT_SIZE,
2926             .help = "Virtual disk size"
2927         },
2928         {
2929             .name = BLOCK_OPT_NOCOW,
2930             .type = QEMU_OPT_BOOL,
2931             .help = "Turn off copy-on-write (valid only on btrfs)"
2932         },
2933         {
2934             .name = BLOCK_OPT_PREALLOC,
2935             .type = QEMU_OPT_STRING,
2936             .help = "Preallocation mode (allowed values: off"
2937 #ifdef CONFIG_POSIX_FALLOCATE
2938                     ", falloc"
2939 #endif
2940                     ", full)"
2941         },
2942         { /* end of list */ }
2943     }
2944 };
2945 
2946 static int raw_check_perm(BlockDriverState *bs, uint64_t perm, uint64_t shared,
2947                           Error **errp)
2948 {
2949     BDRVRawState *s = bs->opaque;
2950     BDRVRawReopenState *rs = NULL;
2951     int open_flags;
2952     int ret;
2953 
2954     if (s->perm_change_fd) {
2955         /*
2956          * In the context of reopen, this function may be called several times
2957          * (directly and recursively while change permissions of the parent).
2958          * This is even true for children that don't inherit from the original
2959          * reopen node, so s->reopen_state is not set.
2960          *
2961          * Ignore all but the first call.
2962          */
2963         return 0;
2964     }
2965 
2966     if (s->reopen_state) {
2967         /* We already have a new file descriptor to set permissions for */
2968         assert(s->reopen_state->perm == perm);
2969         assert(s->reopen_state->shared_perm == shared);
2970         rs = s->reopen_state->opaque;
2971         s->perm_change_fd = rs->fd;
2972         s->perm_change_flags = rs->open_flags;
2973     } else {
2974         /* We may need a new fd if auto-read-only switches the mode */
2975         ret = raw_reconfigure_getfd(bs, bs->open_flags, &open_flags, perm,
2976                                     false, errp);
2977         if (ret < 0) {
2978             return ret;
2979         } else if (ret != s->fd) {
2980             s->perm_change_fd = ret;
2981             s->perm_change_flags = open_flags;
2982         }
2983     }
2984 
2985     /* Prepare permissions on old fd to avoid conflicts between old and new,
2986      * but keep everything locked that new will need. */
2987     ret = raw_handle_perm_lock(bs, RAW_PL_PREPARE, perm, shared, errp);
2988     if (ret < 0) {
2989         goto fail;
2990     }
2991 
2992     /* Copy locks to the new fd */
2993     if (s->perm_change_fd) {
2994         ret = raw_apply_lock_bytes(NULL, s->perm_change_fd, perm, ~shared,
2995                                    false, errp);
2996         if (ret < 0) {
2997             raw_handle_perm_lock(bs, RAW_PL_ABORT, 0, 0, NULL);
2998             goto fail;
2999         }
3000     }
3001     return 0;
3002 
3003 fail:
3004     if (s->perm_change_fd && !s->reopen_state) {
3005         qemu_close(s->perm_change_fd);
3006     }
3007     s->perm_change_fd = 0;
3008     return ret;
3009 }
3010 
3011 static void raw_set_perm(BlockDriverState *bs, uint64_t perm, uint64_t shared)
3012 {
3013     BDRVRawState *s = bs->opaque;
3014 
3015     /* For reopen, we have already switched to the new fd (.bdrv_set_perm is
3016      * called after .bdrv_reopen_commit) */
3017     if (s->perm_change_fd && s->fd != s->perm_change_fd) {
3018         qemu_close(s->fd);
3019         s->fd = s->perm_change_fd;
3020         s->open_flags = s->perm_change_flags;
3021     }
3022     s->perm_change_fd = 0;
3023 
3024     raw_handle_perm_lock(bs, RAW_PL_COMMIT, perm, shared, NULL);
3025     s->perm = perm;
3026     s->shared_perm = shared;
3027 }
3028 
3029 static void raw_abort_perm_update(BlockDriverState *bs)
3030 {
3031     BDRVRawState *s = bs->opaque;
3032 
3033     /* For reopen, .bdrv_reopen_abort is called afterwards and will close
3034      * the file descriptor. */
3035     if (s->perm_change_fd && !s->reopen_state) {
3036         qemu_close(s->perm_change_fd);
3037     }
3038     s->perm_change_fd = 0;
3039 
3040     raw_handle_perm_lock(bs, RAW_PL_ABORT, 0, 0, NULL);
3041 }
3042 
3043 static int coroutine_fn raw_co_copy_range_from(
3044         BlockDriverState *bs, BdrvChild *src, uint64_t src_offset,
3045         BdrvChild *dst, uint64_t dst_offset, uint64_t bytes,
3046         BdrvRequestFlags read_flags, BdrvRequestFlags write_flags)
3047 {
3048     return bdrv_co_copy_range_to(src, src_offset, dst, dst_offset, bytes,
3049                                  read_flags, write_flags);
3050 }
3051 
3052 static int coroutine_fn raw_co_copy_range_to(BlockDriverState *bs,
3053                                              BdrvChild *src,
3054                                              uint64_t src_offset,
3055                                              BdrvChild *dst,
3056                                              uint64_t dst_offset,
3057                                              uint64_t bytes,
3058                                              BdrvRequestFlags read_flags,
3059                                              BdrvRequestFlags write_flags)
3060 {
3061     RawPosixAIOData acb;
3062     BDRVRawState *s = bs->opaque;
3063     BDRVRawState *src_s;
3064 
3065     assert(dst->bs == bs);
3066     if (src->bs->drv->bdrv_co_copy_range_to != raw_co_copy_range_to) {
3067         return -ENOTSUP;
3068     }
3069 
3070     src_s = src->bs->opaque;
3071     if (fd_open(src->bs) < 0 || fd_open(dst->bs) < 0) {
3072         return -EIO;
3073     }
3074 
3075     acb = (RawPosixAIOData) {
3076         .bs             = bs,
3077         .aio_type       = QEMU_AIO_COPY_RANGE,
3078         .aio_fildes     = src_s->fd,
3079         .aio_offset     = src_offset,
3080         .aio_nbytes     = bytes,
3081         .copy_range     = {
3082             .aio_fd2        = s->fd,
3083             .aio_offset2    = dst_offset,
3084         },
3085     };
3086 
3087     return raw_thread_pool_submit(bs, handle_aiocb_copy_range, &acb);
3088 }
3089 
3090 BlockDriver bdrv_file = {
3091     .format_name = "file",
3092     .protocol_name = "file",
3093     .instance_size = sizeof(BDRVRawState),
3094     .bdrv_needs_filename = true,
3095     .bdrv_probe = NULL, /* no probe for protocols */
3096     .bdrv_parse_filename = raw_parse_filename,
3097     .bdrv_file_open = raw_open,
3098     .bdrv_reopen_prepare = raw_reopen_prepare,
3099     .bdrv_reopen_commit = raw_reopen_commit,
3100     .bdrv_reopen_abort = raw_reopen_abort,
3101     .bdrv_close = raw_close,
3102     .bdrv_co_create = raw_co_create,
3103     .bdrv_co_create_opts = raw_co_create_opts,
3104     .bdrv_has_zero_init = bdrv_has_zero_init_1,
3105     .bdrv_has_zero_init_truncate = bdrv_has_zero_init_1,
3106     .bdrv_co_block_status = raw_co_block_status,
3107     .bdrv_co_invalidate_cache = raw_co_invalidate_cache,
3108     .bdrv_co_pwrite_zeroes = raw_co_pwrite_zeroes,
3109     .bdrv_co_delete_file = raw_co_delete_file,
3110 
3111     .bdrv_co_preadv         = raw_co_preadv,
3112     .bdrv_co_pwritev        = raw_co_pwritev,
3113     .bdrv_co_flush_to_disk  = raw_co_flush_to_disk,
3114     .bdrv_co_pdiscard       = raw_co_pdiscard,
3115     .bdrv_co_copy_range_from = raw_co_copy_range_from,
3116     .bdrv_co_copy_range_to  = raw_co_copy_range_to,
3117     .bdrv_refresh_limits = raw_refresh_limits,
3118     .bdrv_io_plug = raw_aio_plug,
3119     .bdrv_io_unplug = raw_aio_unplug,
3120     .bdrv_attach_aio_context = raw_aio_attach_aio_context,
3121 
3122     .bdrv_co_truncate = raw_co_truncate,
3123     .bdrv_getlength = raw_getlength,
3124     .bdrv_get_info = raw_get_info,
3125     .bdrv_get_allocated_file_size
3126                         = raw_get_allocated_file_size,
3127     .bdrv_get_specific_stats = raw_get_specific_stats,
3128     .bdrv_check_perm = raw_check_perm,
3129     .bdrv_set_perm   = raw_set_perm,
3130     .bdrv_abort_perm_update = raw_abort_perm_update,
3131     .create_opts = &raw_create_opts,
3132     .mutable_opts = mutable_opts,
3133 };
3134 
3135 /***********************************************/
3136 /* host device */
3137 
3138 #if defined(__APPLE__) && defined(__MACH__)
3139 static kern_return_t GetBSDPath(io_iterator_t mediaIterator, char *bsdPath,
3140                                 CFIndex maxPathSize, int flags);
3141 static char *FindEjectableOpticalMedia(io_iterator_t *mediaIterator)
3142 {
3143     kern_return_t kernResult = KERN_FAILURE;
3144     mach_port_t     masterPort;
3145     CFMutableDictionaryRef  classesToMatch;
3146     const char *matching_array[] = {kIODVDMediaClass, kIOCDMediaClass};
3147     char *mediaType = NULL;
3148 
3149     kernResult = IOMasterPort( MACH_PORT_NULL, &masterPort );
3150     if ( KERN_SUCCESS != kernResult ) {
3151         printf( "IOMasterPort returned %d\n", kernResult );
3152     }
3153 
3154     int index;
3155     for (index = 0; index < ARRAY_SIZE(matching_array); index++) {
3156         classesToMatch = IOServiceMatching(matching_array[index]);
3157         if (classesToMatch == NULL) {
3158             error_report("IOServiceMatching returned NULL for %s",
3159                          matching_array[index]);
3160             continue;
3161         }
3162         CFDictionarySetValue(classesToMatch, CFSTR(kIOMediaEjectableKey),
3163                              kCFBooleanTrue);
3164         kernResult = IOServiceGetMatchingServices(masterPort, classesToMatch,
3165                                                   mediaIterator);
3166         if (kernResult != KERN_SUCCESS) {
3167             error_report("Note: IOServiceGetMatchingServices returned %d",
3168                          kernResult);
3169             continue;
3170         }
3171 
3172         /* If a match was found, leave the loop */
3173         if (*mediaIterator != 0) {
3174             trace_file_FindEjectableOpticalMedia(matching_array[index]);
3175             mediaType = g_strdup(matching_array[index]);
3176             break;
3177         }
3178     }
3179     return mediaType;
3180 }
3181 
3182 kern_return_t GetBSDPath(io_iterator_t mediaIterator, char *bsdPath,
3183                          CFIndex maxPathSize, int flags)
3184 {
3185     io_object_t     nextMedia;
3186     kern_return_t   kernResult = KERN_FAILURE;
3187     *bsdPath = '\0';
3188     nextMedia = IOIteratorNext( mediaIterator );
3189     if ( nextMedia )
3190     {
3191         CFTypeRef   bsdPathAsCFString;
3192     bsdPathAsCFString = IORegistryEntryCreateCFProperty( nextMedia, CFSTR( kIOBSDNameKey ), kCFAllocatorDefault, 0 );
3193         if ( bsdPathAsCFString ) {
3194             size_t devPathLength;
3195             strcpy( bsdPath, _PATH_DEV );
3196             if (flags & BDRV_O_NOCACHE) {
3197                 strcat(bsdPath, "r");
3198             }
3199             devPathLength = strlen( bsdPath );
3200             if ( CFStringGetCString( bsdPathAsCFString, bsdPath + devPathLength, maxPathSize - devPathLength, kCFStringEncodingASCII ) ) {
3201                 kernResult = KERN_SUCCESS;
3202             }
3203             CFRelease( bsdPathAsCFString );
3204         }
3205         IOObjectRelease( nextMedia );
3206     }
3207 
3208     return kernResult;
3209 }
3210 
3211 /* Sets up a real cdrom for use in QEMU */
3212 static bool setup_cdrom(char *bsd_path, Error **errp)
3213 {
3214     int index, num_of_test_partitions = 2, fd;
3215     char test_partition[MAXPATHLEN];
3216     bool partition_found = false;
3217 
3218     /* look for a working partition */
3219     for (index = 0; index < num_of_test_partitions; index++) {
3220         snprintf(test_partition, sizeof(test_partition), "%ss%d", bsd_path,
3221                  index);
3222         fd = qemu_open(test_partition, O_RDONLY | O_BINARY | O_LARGEFILE);
3223         if (fd >= 0) {
3224             partition_found = true;
3225             qemu_close(fd);
3226             break;
3227         }
3228     }
3229 
3230     /* if a working partition on the device was not found */
3231     if (partition_found == false) {
3232         error_setg(errp, "Failed to find a working partition on disc");
3233     } else {
3234         trace_file_setup_cdrom(test_partition);
3235         pstrcpy(bsd_path, MAXPATHLEN, test_partition);
3236     }
3237     return partition_found;
3238 }
3239 
3240 /* Prints directions on mounting and unmounting a device */
3241 static void print_unmounting_directions(const char *file_name)
3242 {
3243     error_report("If device %s is mounted on the desktop, unmount"
3244                  " it first before using it in QEMU", file_name);
3245     error_report("Command to unmount device: diskutil unmountDisk %s",
3246                  file_name);
3247     error_report("Command to mount device: diskutil mountDisk %s", file_name);
3248 }
3249 
3250 #endif /* defined(__APPLE__) && defined(__MACH__) */
3251 
3252 static int hdev_probe_device(const char *filename)
3253 {
3254     struct stat st;
3255 
3256     /* allow a dedicated CD-ROM driver to match with a higher priority */
3257     if (strstart(filename, "/dev/cdrom", NULL))
3258         return 50;
3259 
3260     if (stat(filename, &st) >= 0 &&
3261             (S_ISCHR(st.st_mode) || S_ISBLK(st.st_mode))) {
3262         return 100;
3263     }
3264 
3265     return 0;
3266 }
3267 
3268 static int check_hdev_writable(BDRVRawState *s)
3269 {
3270 #if defined(BLKROGET)
3271     /* Linux block devices can be configured "read-only" using blockdev(8).
3272      * This is independent of device node permissions and therefore open(2)
3273      * with O_RDWR succeeds.  Actual writes fail with EPERM.
3274      *
3275      * bdrv_open() is supposed to fail if the disk is read-only.  Explicitly
3276      * check for read-only block devices so that Linux block devices behave
3277      * properly.
3278      */
3279     struct stat st;
3280     int readonly = 0;
3281 
3282     if (fstat(s->fd, &st)) {
3283         return -errno;
3284     }
3285 
3286     if (!S_ISBLK(st.st_mode)) {
3287         return 0;
3288     }
3289 
3290     if (ioctl(s->fd, BLKROGET, &readonly) < 0) {
3291         return -errno;
3292     }
3293 
3294     if (readonly) {
3295         return -EACCES;
3296     }
3297 #endif /* defined(BLKROGET) */
3298     return 0;
3299 }
3300 
3301 static void hdev_parse_filename(const char *filename, QDict *options,
3302                                 Error **errp)
3303 {
3304     bdrv_parse_filename_strip_prefix(filename, "host_device:", options);
3305 }
3306 
3307 static bool hdev_is_sg(BlockDriverState *bs)
3308 {
3309 
3310 #if defined(__linux__)
3311 
3312     BDRVRawState *s = bs->opaque;
3313     struct stat st;
3314     struct sg_scsi_id scsiid;
3315     int sg_version;
3316     int ret;
3317 
3318     if (stat(bs->filename, &st) < 0 || !S_ISCHR(st.st_mode)) {
3319         return false;
3320     }
3321 
3322     ret = ioctl(s->fd, SG_GET_VERSION_NUM, &sg_version);
3323     if (ret < 0) {
3324         return false;
3325     }
3326 
3327     ret = ioctl(s->fd, SG_GET_SCSI_ID, &scsiid);
3328     if (ret >= 0) {
3329         trace_file_hdev_is_sg(scsiid.scsi_type, sg_version);
3330         return true;
3331     }
3332 
3333 #endif
3334 
3335     return false;
3336 }
3337 
3338 static int hdev_open(BlockDriverState *bs, QDict *options, int flags,
3339                      Error **errp)
3340 {
3341     BDRVRawState *s = bs->opaque;
3342     Error *local_err = NULL;
3343     int ret;
3344 
3345 #if defined(__APPLE__) && defined(__MACH__)
3346     /*
3347      * Caution: while qdict_get_str() is fine, getting non-string types
3348      * would require more care.  When @options come from -blockdev or
3349      * blockdev_add, its members are typed according to the QAPI
3350      * schema, but when they come from -drive, they're all QString.
3351      */
3352     const char *filename = qdict_get_str(options, "filename");
3353     char bsd_path[MAXPATHLEN] = "";
3354     bool error_occurred = false;
3355 
3356     /* If using a real cdrom */
3357     if (strcmp(filename, "/dev/cdrom") == 0) {
3358         char *mediaType = NULL;
3359         kern_return_t ret_val;
3360         io_iterator_t mediaIterator = 0;
3361 
3362         mediaType = FindEjectableOpticalMedia(&mediaIterator);
3363         if (mediaType == NULL) {
3364             error_setg(errp, "Please make sure your CD/DVD is in the optical"
3365                        " drive");
3366             error_occurred = true;
3367             goto hdev_open_Mac_error;
3368         }
3369 
3370         ret_val = GetBSDPath(mediaIterator, bsd_path, sizeof(bsd_path), flags);
3371         if (ret_val != KERN_SUCCESS) {
3372             error_setg(errp, "Could not get BSD path for optical drive");
3373             error_occurred = true;
3374             goto hdev_open_Mac_error;
3375         }
3376 
3377         /* If a real optical drive was not found */
3378         if (bsd_path[0] == '\0') {
3379             error_setg(errp, "Failed to obtain bsd path for optical drive");
3380             error_occurred = true;
3381             goto hdev_open_Mac_error;
3382         }
3383 
3384         /* If using a cdrom disc and finding a partition on the disc failed */
3385         if (strncmp(mediaType, kIOCDMediaClass, 9) == 0 &&
3386             setup_cdrom(bsd_path, errp) == false) {
3387             print_unmounting_directions(bsd_path);
3388             error_occurred = true;
3389             goto hdev_open_Mac_error;
3390         }
3391 
3392         qdict_put_str(options, "filename", bsd_path);
3393 
3394 hdev_open_Mac_error:
3395         g_free(mediaType);
3396         if (mediaIterator) {
3397             IOObjectRelease(mediaIterator);
3398         }
3399         if (error_occurred) {
3400             return -ENOENT;
3401         }
3402     }
3403 #endif /* defined(__APPLE__) && defined(__MACH__) */
3404 
3405     s->type = FTYPE_FILE;
3406 
3407     ret = raw_open_common(bs, options, flags, 0, true, &local_err);
3408     if (ret < 0) {
3409         error_propagate(errp, local_err);
3410 #if defined(__APPLE__) && defined(__MACH__)
3411         if (*bsd_path) {
3412             filename = bsd_path;
3413         }
3414         /* if a physical device experienced an error while being opened */
3415         if (strncmp(filename, "/dev/", 5) == 0) {
3416             print_unmounting_directions(filename);
3417         }
3418 #endif /* defined(__APPLE__) && defined(__MACH__) */
3419         return ret;
3420     }
3421 
3422     /* Since this does ioctl the device must be already opened */
3423     bs->sg = hdev_is_sg(bs);
3424 
3425     if (flags & BDRV_O_RDWR) {
3426         ret = check_hdev_writable(s);
3427         if (ret < 0) {
3428             raw_close(bs);
3429             error_setg_errno(errp, -ret, "The device is not writable");
3430             return ret;
3431         }
3432     }
3433 
3434     return ret;
3435 }
3436 
3437 #if defined(__linux__)
3438 static int coroutine_fn
3439 hdev_co_ioctl(BlockDriverState *bs, unsigned long int req, void *buf)
3440 {
3441     BDRVRawState *s = bs->opaque;
3442     RawPosixAIOData acb;
3443     int ret;
3444 
3445     ret = fd_open(bs);
3446     if (ret < 0) {
3447         return ret;
3448     }
3449 
3450     if (req == SG_IO && s->pr_mgr) {
3451         struct sg_io_hdr *io_hdr = buf;
3452         if (io_hdr->cmdp[0] == PERSISTENT_RESERVE_OUT ||
3453             io_hdr->cmdp[0] == PERSISTENT_RESERVE_IN) {
3454             return pr_manager_execute(s->pr_mgr, bdrv_get_aio_context(bs),
3455                                       s->fd, io_hdr);
3456         }
3457     }
3458 
3459     acb = (RawPosixAIOData) {
3460         .bs         = bs,
3461         .aio_type   = QEMU_AIO_IOCTL,
3462         .aio_fildes = s->fd,
3463         .aio_offset = 0,
3464         .ioctl      = {
3465             .buf        = buf,
3466             .cmd        = req,
3467         },
3468     };
3469 
3470     return raw_thread_pool_submit(bs, handle_aiocb_ioctl, &acb);
3471 }
3472 #endif /* linux */
3473 
3474 static int fd_open(BlockDriverState *bs)
3475 {
3476     BDRVRawState *s = bs->opaque;
3477 
3478     /* this is just to ensure s->fd is sane (its called by io ops) */
3479     if (s->fd >= 0)
3480         return 0;
3481     return -EIO;
3482 }
3483 
3484 static coroutine_fn int
3485 hdev_co_pdiscard(BlockDriverState *bs, int64_t offset, int bytes)
3486 {
3487     BDRVRawState *s = bs->opaque;
3488     int ret;
3489 
3490     ret = fd_open(bs);
3491     if (ret < 0) {
3492         raw_account_discard(s, bytes, ret);
3493         return ret;
3494     }
3495     return raw_do_pdiscard(bs, offset, bytes, true);
3496 }
3497 
3498 static coroutine_fn int hdev_co_pwrite_zeroes(BlockDriverState *bs,
3499     int64_t offset, int bytes, BdrvRequestFlags flags)
3500 {
3501     int rc;
3502 
3503     rc = fd_open(bs);
3504     if (rc < 0) {
3505         return rc;
3506     }
3507 
3508     return raw_do_pwrite_zeroes(bs, offset, bytes, flags, true);
3509 }
3510 
3511 static BlockDriver bdrv_host_device = {
3512     .format_name        = "host_device",
3513     .protocol_name        = "host_device",
3514     .instance_size      = sizeof(BDRVRawState),
3515     .bdrv_needs_filename = true,
3516     .bdrv_probe_device  = hdev_probe_device,
3517     .bdrv_parse_filename = hdev_parse_filename,
3518     .bdrv_file_open     = hdev_open,
3519     .bdrv_close         = raw_close,
3520     .bdrv_reopen_prepare = raw_reopen_prepare,
3521     .bdrv_reopen_commit  = raw_reopen_commit,
3522     .bdrv_reopen_abort   = raw_reopen_abort,
3523     .bdrv_co_create_opts = bdrv_co_create_opts_simple,
3524     .create_opts         = &bdrv_create_opts_simple,
3525     .mutable_opts        = mutable_opts,
3526     .bdrv_co_invalidate_cache = raw_co_invalidate_cache,
3527     .bdrv_co_pwrite_zeroes = hdev_co_pwrite_zeroes,
3528 
3529     .bdrv_co_preadv         = raw_co_preadv,
3530     .bdrv_co_pwritev        = raw_co_pwritev,
3531     .bdrv_co_flush_to_disk  = raw_co_flush_to_disk,
3532     .bdrv_co_pdiscard       = hdev_co_pdiscard,
3533     .bdrv_co_copy_range_from = raw_co_copy_range_from,
3534     .bdrv_co_copy_range_to  = raw_co_copy_range_to,
3535     .bdrv_refresh_limits = raw_refresh_limits,
3536     .bdrv_io_plug = raw_aio_plug,
3537     .bdrv_io_unplug = raw_aio_unplug,
3538     .bdrv_attach_aio_context = raw_aio_attach_aio_context,
3539 
3540     .bdrv_co_truncate       = raw_co_truncate,
3541     .bdrv_getlength	= raw_getlength,
3542     .bdrv_get_info = raw_get_info,
3543     .bdrv_get_allocated_file_size
3544                         = raw_get_allocated_file_size,
3545     .bdrv_get_specific_stats = hdev_get_specific_stats,
3546     .bdrv_check_perm = raw_check_perm,
3547     .bdrv_set_perm   = raw_set_perm,
3548     .bdrv_abort_perm_update = raw_abort_perm_update,
3549     .bdrv_probe_blocksizes = hdev_probe_blocksizes,
3550     .bdrv_probe_geometry = hdev_probe_geometry,
3551 
3552     /* generic scsi device */
3553 #ifdef __linux__
3554     .bdrv_co_ioctl          = hdev_co_ioctl,
3555 #endif
3556 };
3557 
3558 #if defined(__linux__) || defined(__FreeBSD__) || defined(__FreeBSD_kernel__)
3559 static void cdrom_parse_filename(const char *filename, QDict *options,
3560                                  Error **errp)
3561 {
3562     bdrv_parse_filename_strip_prefix(filename, "host_cdrom:", options);
3563 }
3564 #endif
3565 
3566 #ifdef __linux__
3567 static int cdrom_open(BlockDriverState *bs, QDict *options, int flags,
3568                       Error **errp)
3569 {
3570     BDRVRawState *s = bs->opaque;
3571 
3572     s->type = FTYPE_CD;
3573 
3574     /* open will not fail even if no CD is inserted, so add O_NONBLOCK */
3575     return raw_open_common(bs, options, flags, O_NONBLOCK, true, errp);
3576 }
3577 
3578 static int cdrom_probe_device(const char *filename)
3579 {
3580     int fd, ret;
3581     int prio = 0;
3582     struct stat st;
3583 
3584     fd = qemu_open(filename, O_RDONLY | O_NONBLOCK);
3585     if (fd < 0) {
3586         goto out;
3587     }
3588     ret = fstat(fd, &st);
3589     if (ret == -1 || !S_ISBLK(st.st_mode)) {
3590         goto outc;
3591     }
3592 
3593     /* Attempt to detect via a CDROM specific ioctl */
3594     ret = ioctl(fd, CDROM_DRIVE_STATUS, CDSL_CURRENT);
3595     if (ret >= 0)
3596         prio = 100;
3597 
3598 outc:
3599     qemu_close(fd);
3600 out:
3601     return prio;
3602 }
3603 
3604 static bool cdrom_is_inserted(BlockDriverState *bs)
3605 {
3606     BDRVRawState *s = bs->opaque;
3607     int ret;
3608 
3609     ret = ioctl(s->fd, CDROM_DRIVE_STATUS, CDSL_CURRENT);
3610     return ret == CDS_DISC_OK;
3611 }
3612 
3613 static void cdrom_eject(BlockDriverState *bs, bool eject_flag)
3614 {
3615     BDRVRawState *s = bs->opaque;
3616 
3617     if (eject_flag) {
3618         if (ioctl(s->fd, CDROMEJECT, NULL) < 0)
3619             perror("CDROMEJECT");
3620     } else {
3621         if (ioctl(s->fd, CDROMCLOSETRAY, NULL) < 0)
3622             perror("CDROMEJECT");
3623     }
3624 }
3625 
3626 static void cdrom_lock_medium(BlockDriverState *bs, bool locked)
3627 {
3628     BDRVRawState *s = bs->opaque;
3629 
3630     if (ioctl(s->fd, CDROM_LOCKDOOR, locked) < 0) {
3631         /*
3632          * Note: an error can happen if the distribution automatically
3633          * mounts the CD-ROM
3634          */
3635         /* perror("CDROM_LOCKDOOR"); */
3636     }
3637 }
3638 
3639 static BlockDriver bdrv_host_cdrom = {
3640     .format_name        = "host_cdrom",
3641     .protocol_name      = "host_cdrom",
3642     .instance_size      = sizeof(BDRVRawState),
3643     .bdrv_needs_filename = true,
3644     .bdrv_probe_device	= cdrom_probe_device,
3645     .bdrv_parse_filename = cdrom_parse_filename,
3646     .bdrv_file_open     = cdrom_open,
3647     .bdrv_close         = raw_close,
3648     .bdrv_reopen_prepare = raw_reopen_prepare,
3649     .bdrv_reopen_commit  = raw_reopen_commit,
3650     .bdrv_reopen_abort   = raw_reopen_abort,
3651     .bdrv_co_create_opts = bdrv_co_create_opts_simple,
3652     .create_opts         = &bdrv_create_opts_simple,
3653     .mutable_opts        = mutable_opts,
3654     .bdrv_co_invalidate_cache = raw_co_invalidate_cache,
3655 
3656     .bdrv_co_preadv         = raw_co_preadv,
3657     .bdrv_co_pwritev        = raw_co_pwritev,
3658     .bdrv_co_flush_to_disk  = raw_co_flush_to_disk,
3659     .bdrv_refresh_limits = raw_refresh_limits,
3660     .bdrv_io_plug = raw_aio_plug,
3661     .bdrv_io_unplug = raw_aio_unplug,
3662     .bdrv_attach_aio_context = raw_aio_attach_aio_context,
3663 
3664     .bdrv_co_truncate    = raw_co_truncate,
3665     .bdrv_getlength      = raw_getlength,
3666     .has_variable_length = true,
3667     .bdrv_get_allocated_file_size
3668                         = raw_get_allocated_file_size,
3669 
3670     /* removable device support */
3671     .bdrv_is_inserted   = cdrom_is_inserted,
3672     .bdrv_eject         = cdrom_eject,
3673     .bdrv_lock_medium   = cdrom_lock_medium,
3674 
3675     /* generic scsi device */
3676     .bdrv_co_ioctl      = hdev_co_ioctl,
3677 };
3678 #endif /* __linux__ */
3679 
3680 #if defined (__FreeBSD__) || defined(__FreeBSD_kernel__)
3681 static int cdrom_open(BlockDriverState *bs, QDict *options, int flags,
3682                       Error **errp)
3683 {
3684     BDRVRawState *s = bs->opaque;
3685     Error *local_err = NULL;
3686     int ret;
3687 
3688     s->type = FTYPE_CD;
3689 
3690     ret = raw_open_common(bs, options, flags, 0, true, &local_err);
3691     if (ret) {
3692         error_propagate(errp, local_err);
3693         return ret;
3694     }
3695 
3696     /* make sure the door isn't locked at this time */
3697     ioctl(s->fd, CDIOCALLOW);
3698     return 0;
3699 }
3700 
3701 static int cdrom_probe_device(const char *filename)
3702 {
3703     if (strstart(filename, "/dev/cd", NULL) ||
3704             strstart(filename, "/dev/acd", NULL))
3705         return 100;
3706     return 0;
3707 }
3708 
3709 static int cdrom_reopen(BlockDriverState *bs)
3710 {
3711     BDRVRawState *s = bs->opaque;
3712     int fd;
3713 
3714     /*
3715      * Force reread of possibly changed/newly loaded disc,
3716      * FreeBSD seems to not notice sometimes...
3717      */
3718     if (s->fd >= 0)
3719         qemu_close(s->fd);
3720     fd = qemu_open(bs->filename, s->open_flags, 0644);
3721     if (fd < 0) {
3722         s->fd = -1;
3723         return -EIO;
3724     }
3725     s->fd = fd;
3726 
3727     /* make sure the door isn't locked at this time */
3728     ioctl(s->fd, CDIOCALLOW);
3729     return 0;
3730 }
3731 
3732 static bool cdrom_is_inserted(BlockDriverState *bs)
3733 {
3734     return raw_getlength(bs) > 0;
3735 }
3736 
3737 static void cdrom_eject(BlockDriverState *bs, bool eject_flag)
3738 {
3739     BDRVRawState *s = bs->opaque;
3740 
3741     if (s->fd < 0)
3742         return;
3743 
3744     (void) ioctl(s->fd, CDIOCALLOW);
3745 
3746     if (eject_flag) {
3747         if (ioctl(s->fd, CDIOCEJECT) < 0)
3748             perror("CDIOCEJECT");
3749     } else {
3750         if (ioctl(s->fd, CDIOCCLOSE) < 0)
3751             perror("CDIOCCLOSE");
3752     }
3753 
3754     cdrom_reopen(bs);
3755 }
3756 
3757 static void cdrom_lock_medium(BlockDriverState *bs, bool locked)
3758 {
3759     BDRVRawState *s = bs->opaque;
3760 
3761     if (s->fd < 0)
3762         return;
3763     if (ioctl(s->fd, (locked ? CDIOCPREVENT : CDIOCALLOW)) < 0) {
3764         /*
3765          * Note: an error can happen if the distribution automatically
3766          * mounts the CD-ROM
3767          */
3768         /* perror("CDROM_LOCKDOOR"); */
3769     }
3770 }
3771 
3772 static BlockDriver bdrv_host_cdrom = {
3773     .format_name        = "host_cdrom",
3774     .protocol_name      = "host_cdrom",
3775     .instance_size      = sizeof(BDRVRawState),
3776     .bdrv_needs_filename = true,
3777     .bdrv_probe_device	= cdrom_probe_device,
3778     .bdrv_parse_filename = cdrom_parse_filename,
3779     .bdrv_file_open     = cdrom_open,
3780     .bdrv_close         = raw_close,
3781     .bdrv_reopen_prepare = raw_reopen_prepare,
3782     .bdrv_reopen_commit  = raw_reopen_commit,
3783     .bdrv_reopen_abort   = raw_reopen_abort,
3784     .bdrv_co_create_opts = bdrv_co_create_opts_simple,
3785     .create_opts         = &bdrv_create_opts_simple,
3786     .mutable_opts       = mutable_opts,
3787 
3788     .bdrv_co_preadv         = raw_co_preadv,
3789     .bdrv_co_pwritev        = raw_co_pwritev,
3790     .bdrv_co_flush_to_disk  = raw_co_flush_to_disk,
3791     .bdrv_refresh_limits = raw_refresh_limits,
3792     .bdrv_io_plug = raw_aio_plug,
3793     .bdrv_io_unplug = raw_aio_unplug,
3794     .bdrv_attach_aio_context = raw_aio_attach_aio_context,
3795 
3796     .bdrv_co_truncate    = raw_co_truncate,
3797     .bdrv_getlength      = raw_getlength,
3798     .has_variable_length = true,
3799     .bdrv_get_allocated_file_size
3800                         = raw_get_allocated_file_size,
3801 
3802     /* removable device support */
3803     .bdrv_is_inserted   = cdrom_is_inserted,
3804     .bdrv_eject         = cdrom_eject,
3805     .bdrv_lock_medium   = cdrom_lock_medium,
3806 };
3807 #endif /* __FreeBSD__ */
3808 
3809 static void bdrv_file_init(void)
3810 {
3811     /*
3812      * Register all the drivers.  Note that order is important, the driver
3813      * registered last will get probed first.
3814      */
3815     bdrv_register(&bdrv_file);
3816     bdrv_register(&bdrv_host_device);
3817 #ifdef __linux__
3818     bdrv_register(&bdrv_host_cdrom);
3819 #endif
3820 #if defined(__FreeBSD__) || defined(__FreeBSD_kernel__)
3821     bdrv_register(&bdrv_host_cdrom);
3822 #endif
3823 }
3824 
3825 block_init(bdrv_file_init);
3826