• Home
  • Features
  • Pricing
  • Docs
  • Announcements
  • Sign In

systemd / systemd / 27517269012

14 Jun 2026 11:20PM UTC coverage: 72.921% (-0.09%) from 73.015%
27517269012

push

github

yuwata
hwdb: Make Samsung download mode work out-of-the-box

https://github.com/Benjamin-Dobell/Heimdall/blob/3997d5cc6/heimdall/source/BridgeManager.h#L69-L79

337290 of 462539 relevant lines covered (72.92%)

1281143.83 hits per line

Source File
Press 'n' to go to next uncovered line, 'b' for previous

65.22
/src/shared/loop-util.c
1
/* SPDX-License-Identifier: LGPL-2.1-or-later */
2

3
#if HAVE_VALGRIND_MEMCHECK_H
4
#include <valgrind/memcheck.h>
5
#endif
6

7
#include <fcntl.h>
8
#include <linux/loop.h>
9
#include <sys/file.h>
10
#include <sys/ioctl.h>
11
#include <unistd.h>
12

13
#include "sd-device.h"
14

15
#include "alloc-util.h"
16
#include "blockdev-util.h"
17
#include "data-fd-util.h"
18
#include "device-private.h"
19
#include "device-util.h"
20
#include "devnum-util.h"
21
#include "dissect-image.h"
22
#include "env-util.h"
23
#include "errno-util.h"
24
#include "fd-util.h"
25
#include "fileio.h"
26
#include "fs-util.h"
27
#include "loop-util.h"
28
#include "parse-util.h"
29
#include "path-util.h"
30
#include "random-util.h"
31
#include "stat-util.h"
32
#include "stdio-util.h"
33
#include "string-util.h"
34
#include "time-util.h"
35

36
static void cleanup_clear_loop_close(int *fd) {
2,545✔
37
        assert(fd);
2,545✔
38

39
        if (*fd < 0)
2,545✔
40
                return;
41

42
        (void) ioctl(*fd, LOOP_CLR_FD);
×
43
        (void) safe_close(*fd);
×
44
}
45

46
static int loop_is_bound(int fd) {
2,545✔
47
        struct loop_info64 info;
2,545✔
48

49
        if (ioctl(ASSERT_FD(fd), LOOP_GET_STATUS64, &info) < 0) {
2,545✔
50
                if (errno == ENXIO)
2,545✔
51
                        return false; /* not bound! */
2,545✔
52

53
                return -errno;
×
54
        }
55

56
        return true; /* bound! */
57
}
58

59
static int open_lock_fd(int primary_fd, int operation) {
2,680✔
60
        _cleanup_close_ int lock_fd = -EBADF;
2,680✔
61

62
        assert(IN_SET(operation & ~LOCK_NB, LOCK_SH, LOCK_EX));
2,680✔
63

64
        lock_fd = fd_reopen(ASSERT_FD(primary_fd), O_RDONLY|O_CLOEXEC|O_NONBLOCK|O_NOCTTY);
2,680✔
65
        if (lock_fd < 0)
2,680✔
66
                return lock_fd;
67

68
        if (flock(lock_fd, operation) < 0)
2,680✔
69
                return -errno;
×
70

71
        return TAKE_FD(lock_fd);
72
}
73

74
static int loop_configure_verify_direct_io(int fd, const struct loop_config *c) {
2,545✔
75
        assert(fd >= 0);
2,545✔
76
        assert(c);
2,545✔
77

78
        if (FLAGS_SET(c->info.lo_flags, LO_FLAGS_DIRECT_IO)) {
2,545✔
79
                struct loop_info64 info;
2,544✔
80

81
                if (ioctl(fd, LOOP_GET_STATUS64, &info) < 0)
2,544✔
82
                        return log_debug_errno(errno, "Failed to issue LOOP_GET_STATUS64: %m");
×
83

84
#if HAVE_VALGRIND_MEMCHECK_H
85
                VALGRIND_MAKE_MEM_DEFINED(&info, sizeof(info));
86
#endif
87

88
                /* On older kernels (<= 5.3) it was necessary to set the block size of the loopback block
89
                 * device to the logical block size of the underlying file system. Since there was no nice
90
                 * way to query the value, we are not bothering to do this however. On newer kernels the
91
                 * block size is propagated automatically and does not require intervention from us. We'll
92
                 * check here if enabling direct IO worked, to make this easily debuggable however.
93
                 *
94
                 * (Should anyone really care and actually wants direct IO on old kernels: it might be worth
95
                 * enabling direct IO with iteratively larger block sizes until it eventually works.)
96
                 *
97
                 * On older kernels (e.g.: 5.10) when this is attempted on a file stored on a dm-crypt
98
                 * backed partition the kernel will start returning I/O errors when accessing the mounted
99
                 * loop device, so return a recognizable error that causes the operation to be started
100
                 * from scratch without the LO_FLAGS_DIRECT_IO flag. */
101
                if (!FLAGS_SET(info.lo_flags, LO_FLAGS_DIRECT_IO))
2,544✔
102
                        return log_debug_errno(
×
103
                                        SYNTHETIC_ERRNO(ENOANO),
104
                                        "Could not enable direct IO mode, retrying in buffered IO mode.");
105
        }
106

107
        return 0;
108
}
109

110
static int loop_configure_verify(int fd, const struct loop_config *c) {
2,545✔
111
        bool broken = false;
2,545✔
112
        int r;
2,545✔
113

114
        assert(fd >= 0);
2,545✔
115
        assert(c);
2,545✔
116

117
        if (c->block_size != 0) {
2,545✔
118
                uint32_t ssz;
2,545✔
119

120
                r = blockdev_get_sector_size(fd, &ssz);
2,545✔
121
                if (r < 0)
2,545✔
122
                        return r;
×
123

124
                if (ssz != c->block_size) {
2,545✔
125
                        log_debug("LOOP_CONFIGURE didn't honour requested block size %" PRIu32 ", got %" PRIu32 " instead. Ignoring.", c->block_size, ssz);
×
126
                        broken = true;
127
                }
128
        }
129

130
        if (c->info.lo_sizelimit != 0) {
2,545✔
131
                /* Kernel 5.8 vanilla doesn't properly propagate the size limit into the
132
                 * block device. If it's used, let's immediately check if it had the desired
133
                 * effect hence. And if not use classic LOOP_SET_STATUS64. */
134
                uint64_t z;
101✔
135

136
                r = blockdev_get_device_size(fd, &z);
101✔
137
                if (r < 0)
101✔
138
                        return r;
×
139

140
                if (z != c->info.lo_sizelimit) {
101✔
141
                        log_debug("LOOP_CONFIGURE is broken, doesn't honour .info.lo_sizelimit. Falling back to LOOP_SET_STATUS64.");
×
142
                        broken = true;
143
                }
144
        }
145

146
        if (FLAGS_SET(c->info.lo_flags, LO_FLAGS_PARTSCAN)) {
2,545✔
147
                /* Kernel 5.8 vanilla doesn't properly propagate the partition scanning flag
148
                 * into the block device. Let's hence verify if things work correctly here
149
                 * before returning. */
150

151
                r = blockdev_partscan_enabled_fd(fd);
×
152
                if (r < 0)
×
153
                        return r;
154
                if (r == 0) {
×
155
                        log_debug("LOOP_CONFIGURE is broken, doesn't honour LO_FLAGS_PARTSCAN. Falling back to LOOP_SET_STATUS64.");
×
156
                        broken = true;
157
                }
158
        }
159

160
        r = loop_configure_verify_direct_io(fd, c);
2,545✔
161
        if (r < 0)
2,545✔
162
                return r;
163

164
        return !broken;
2,545✔
165
}
166

167
static int loop_configure_fallback(int fd, const struct loop_config *c) {
×
168
        struct loop_info64 info_copy;
×
169
        int r;
×
170

171
        assert(fd >= 0);
×
172
        assert(c);
×
173

174
        /* Only some of the flags LOOP_CONFIGURE can set are also settable via LOOP_SET_STATUS64, hence mask
175
         * them out. */
176
        info_copy = c->info;
×
177
        info_copy.lo_flags &= LOOP_SET_STATUS_SETTABLE_FLAGS;
×
178

179
        /* Since kernel commit 5db470e229e22b7eda6e23b5566e532c96fb5bc3 (kernel v5.0) the LOOP_SET_STATUS64
180
         * ioctl can return EAGAIN in case we change the info.lo_offset field, if someone else is accessing the
181
         * block device while we try to reconfigure it. This is a pretty common case, since udev might
182
         * instantly start probing the device as soon as we attach an fd to it. Hence handle it in two ways:
183
         * first, let's take the BSD lock to ensure that udev will not step in between the point in
184
         * time where we attach the fd and where we reconfigure the device. Secondly, let's wait 50ms on
185
         * EAGAIN and retry. The former should be an efficient mechanism to avoid we have to wait 50ms
186
         * needlessly if we are just racing against udev. The latter is protection against all other cases,
187
         * i.e. peers that do not take the BSD lock. */
188

189
        for (unsigned n_attempts = 0;;) {
×
190
                if (ioctl(fd, LOOP_SET_STATUS64, &info_copy) >= 0)
×
191
                        break;
192

193
                if (errno != EAGAIN || ++n_attempts >= 64)
×
194
                        return log_debug_errno(errno, "Failed to configure loopback block device: %m");
×
195

196
                /* Sleep some random time, but at least 10ms, at most 250ms. Increase the delay the more
197
                 * failed attempts we see */
198
                (void) usleep_safe(UINT64_C(10) * USEC_PER_MSEC +
×
199
                              random_u64_range(UINT64_C(240) * USEC_PER_MSEC * n_attempts/64));
×
200
        }
201

202
        /* If a block size is requested then try to configure it. If that doesn't work, ignore errors, but
203
         * afterwards, let's validate what is in effect, and if it doesn't match what we want, fail */
204
        if (c->block_size != 0) {
×
205
                uint32_t ssz;
×
206

207
                if (ioctl(fd, LOOP_SET_BLOCK_SIZE, (unsigned long) c->block_size) < 0)
×
208
                        log_debug_errno(errno, "Failed to set sector size, ignoring: %m");
×
209

210
                r = blockdev_get_sector_size(fd, &ssz);
×
211
                if (r < 0)
×
212
                        return log_debug_errno(r, "Failed to read sector size: %m");
×
213
                if (ssz != c->block_size)
×
214
                        return log_debug_errno(SYNTHETIC_ERRNO(EIO), "Sector size of loopback device doesn't match what we requested, refusing.");
×
215
        }
216

217
        /* LO_FLAGS_DIRECT_IO is a flags we need to configure via explicit ioctls. */
218
        if (FLAGS_SET(c->info.lo_flags, LO_FLAGS_DIRECT_IO))
×
219
                if (ioctl(fd, LOOP_SET_DIRECT_IO, 1UL) < 0)
×
220
                        log_debug_errno(errno, "Failed to enable direct IO mode, ignoring: %m");
×
221

222
        return loop_configure_verify_direct_io(fd, c);
×
223
}
224

225
static int loop_configure(
2,545✔
226
                int nr,
227
                int open_flags,
228
                int lock_op,
229
                const struct loop_config *c,
230
                LoopDevice **ret) {
231

232
        static bool loop_configure_broken = false;
2,545✔
233

234
        _cleanup_(sd_device_unrefp) sd_device *dev = NULL;
2,545✔
235
        _cleanup_(cleanup_clear_loop_close) int loop_with_fd = -EBADF; /* This must be declared before lock_fd. */
×
236
        _cleanup_close_ int fd = -EBADF, lock_fd = -EBADF;
5,090✔
237
        _cleanup_free_ char *node = NULL;
2,545✔
238
        uint64_t diskseq = 0;
2,545✔
239
        dev_t devno;
2,545✔
240
        int r;
2,545✔
241

242
        assert(nr >= 0);
2,545✔
243
        assert(c);
2,545✔
244
        assert(ret);
2,545✔
245

246
        if (asprintf(&node, "/dev/loop%i", nr) < 0)
2,545✔
247
                return log_oom_debug();
×
248

249
        r = sd_device_new_from_devname(&dev, node);
2,545✔
250
        if (r < 0)
2,545✔
251
                return log_debug_errno(r, "Failed to create sd_device object for \"%s\": %m", node);
×
252

253
        r = sd_device_get_devnum(dev, &devno);
2,545✔
254
        if (r < 0)
2,545✔
255
                return log_device_debug_errno(dev, r, "Failed to get devnum: %m");
×
256

257
        fd = sd_device_open(dev, O_CLOEXEC|O_NONBLOCK|O_NOCTTY|open_flags);
2,545✔
258
        if (fd < 0)
2,545✔
259
                return log_device_debug_errno(dev, fd, "Failed to open device: %m");
×
260

261
        /* Let's lock the device before we do anything. We take the BSD lock on a second, separately opened
262
         * fd for the device. udev after all watches for close() events (specifically IN_CLOSE_WRITE) on
263
         * block devices to reprobe them, hence by having a separate fd we will later close() we can ensure
264
         * we trigger udev after everything is done. If we'd lock our own fd instead and keep it open for a
265
         * long time udev would possibly never run on it again, even though the fd is unlocked, simply
266
         * because we never close() it. It also has the nice benefit we can use the _cleanup_close_ logic to
267
         * automatically release the lock, after we are done. */
268
        lock_fd = open_lock_fd(fd, LOCK_EX);
2,545✔
269
        if (lock_fd < 0)
2,545✔
270
                return log_device_debug_errno(dev, lock_fd, "Failed to acquire lock: %m");
×
271

272
        log_device_debug(dev, "Acquired exclusive lock.");
2,577✔
273

274
        /* Let's see if backing file is really unattached. Someone may already attach a backing file without
275
         * taking BSD lock. */
276
        r = loop_is_bound(fd);
2,545✔
277
        if (r < 0)
2,545✔
278
                return log_device_debug_errno(dev, r, "Failed to check if the loopback block device is bound: %m");
×
279
        if (r > 0)
2,545✔
280
                return log_device_debug_errno(dev, SYNTHETIC_ERRNO(EBUSY),
×
281
                                              "The loopback block device is already bound, ignoring.");
282

283
        /* Let's see if the device is really detached, i.e. currently has no associated partition block
284
         * devices. On various kernels (such as 5.8) it is possible to have a loopback block device that
285
         * superficially is detached but still has partition block devices associated for it. Let's then
286
         * manually remove the partitions via BLKPG, and tell the caller we did that via EUCLEAN, so they try
287
         * again. */
288
        r = block_device_remove_all_partitions(dev, fd);
2,545✔
289
        if (r < 0)
2,545✔
290
                return log_device_debug_errno(dev, r, "Failed to remove partitions on the loopback block device: %m");
×
291
        if (r > 0)
2,545✔
292
                /* Removed all partitions. Let's report this to the caller, to try again, and count this as
293
                 * an attempt. */
294
                return log_device_debug_errno(dev, SYNTHETIC_ERRNO(EUCLEAN),
×
295
                                              "Removed partitions on the loopback block device.");
296

297
        if (!loop_configure_broken) {
2,545✔
298
                if (ioctl(fd, LOOP_CONFIGURE, c) < 0) {
2,545✔
299
                        /* Do fallback only if LOOP_CONFIGURE is not supported, propagate all other errors. */
300
                        if (!ERRNO_IS_IOCTL_NOT_SUPPORTED(errno))
×
301
                                return log_device_debug_errno(dev, errno, "ioctl(LOOP_CONFIGURE) failed: %m");
×
302

303
                        loop_configure_broken = true;
×
304
                } else {
305
                        loop_with_fd = TAKE_FD(fd);
2,545✔
306

307
                        r = loop_configure_verify(loop_with_fd, c);
2,545✔
308
                        if (r < 0)
2,545✔
309
                                return log_device_debug_errno(dev, r, "Failed to verify if loopback block device is correctly configured: %m");
×
310
                        if (r == 0) {
2,545✔
311
                                /* LOOP_CONFIGURE doesn't work. Remember that. */
312
                                loop_configure_broken = true;
×
313

314
                                /* We return EBUSY here instead of retrying immediately with LOOP_SET_FD,
315
                                 * because LOOP_CLR_FD is async: if the operation cannot be executed right
316
                                 * away it just sets the autoclear flag on the device. This means there's a
317
                                 * good chance we cannot actually reuse the loopback device right-away. Hence
318
                                 * let's assume it's busy, avoid the trouble and let the calling loop call us
319
                                 * again with a new, likely unused device. */
320
                                return -EBUSY;
×
321
                        }
322
                }
323
        }
324

325
        if (loop_configure_broken) {
2,545✔
326
                if (ioctl(fd, LOOP_SET_FD, c->fd) < 0)
×
327
                        return log_device_debug_errno(dev, errno, "ioctl(LOOP_SET_FD) failed: %m");
×
328

329
                loop_with_fd = TAKE_FD(fd);
×
330

331
                r = loop_configure_fallback(loop_with_fd, c);
×
332
                if (r < 0)
×
333
                        return r;
334
        }
335

336
        r = fd_get_diskseq(loop_with_fd, &diskseq);
2,545✔
337
        if (r < 0 && r != -EOPNOTSUPP)
2,545✔
338
                return log_device_debug_errno(dev, r, "Failed to get diskseq: %m");
×
339

340
        switch (lock_op & ~LOCK_NB) {
2,545✔
341
        case LOCK_EX: /* Already in effect */
342
                break;
343
        case LOCK_SH: /* Downgrade */
2,410✔
344
                if (flock(lock_fd, lock_op) < 0)
2,410✔
345
                        return log_device_debug_errno(dev, errno, "Failed to downgrade lock level: %m");
×
346
                break;
347
        case LOCK_UN: /* Release */
×
348
                lock_fd = safe_close(lock_fd);
×
349
                break;
350
        default:
×
351
                assert_not_reached();
×
352
        }
353

354
        uint64_t device_size;
2,545✔
355
        r = blockdev_get_device_size(loop_with_fd, &device_size);
2,545✔
356
        if (r < 0)
2,545✔
357
                return log_device_debug_errno(dev, r, "Failed to get loopback device size: %m");
×
358

359
        LoopDevice *d = new(LoopDevice, 1);
2,545✔
360
        if (!d)
2,545✔
361
                return log_oom_debug();
×
362

363
        *d = (LoopDevice) {
2,545✔
364
                .n_ref = 1,
365
                .fd = TAKE_FD(loop_with_fd),
2,545✔
366
                .lock_fd = TAKE_FD(lock_fd),
2,545✔
367
                .node = TAKE_PTR(node),
2,545✔
368
                .nr = nr,
369
                .devno = devno,
370
                .dev = TAKE_PTR(dev),
2,545✔
371
                .diskseq = diskseq,
372
                .sector_size = c->block_size,
2,545✔
373
                .device_size = device_size,
374
                .created = true,
375
        };
376

377
        *ret = TAKE_PTR(d);
2,545✔
378
        return 0;
2,545✔
379
}
380

381
static int fd_get_max_discard(int fd, uint64_t *ret) {
×
382
        char sysfs_path[STRLEN("/sys/dev/block/" ":" "/queue/discard_max_bytes") + DECIMAL_STR_MAX(dev_t) * 2 + 1];
×
383
        _cleanup_free_ char *buffer = NULL;
×
384
        struct stat st;
×
385
        int r;
×
386

387
        assert(ret);
×
388

389
        if (fstat(ASSERT_FD(fd), &st) < 0)
×
390
                return -errno;
×
391

392
        r = stat_verify_block(&st);
×
393
        if (r < 0)
×
394
                return r;
395

396
        xsprintf(sysfs_path, "/sys/dev/block/" DEVNUM_FORMAT_STR "/queue/discard_max_bytes", DEVNUM_FORMAT_VAL(st.st_rdev));
×
397

398
        r = read_one_line_file(sysfs_path, &buffer);
×
399
        if (r < 0)
×
400
                return r;
401

402
        return safe_atou64(buffer, ret);
×
403
}
404

405
static int fd_set_max_discard(int fd, uint64_t max_discard) {
×
406
        char sysfs_path[STRLEN("/sys/dev/block/" ":" "/queue/discard_max_bytes") + DECIMAL_STR_MAX(dev_t) * 2 + 1];
×
407
        struct stat st;
×
408
        int r;
×
409

410
        if (fstat(ASSERT_FD(fd), &st) < 0)
×
411
                return -errno;
×
412

413
        r = stat_verify_block(&st);
×
414
        if (r < 0)
×
415
                return r;
416

417
        xsprintf(sysfs_path, "/sys/dev/block/" DEVNUM_FORMAT_STR "/queue/discard_max_bytes", DEVNUM_FORMAT_VAL(st.st_rdev));
×
418

419
        return write_string_filef(sysfs_path, WRITE_STRING_FILE_DISABLE_BUFFER, "%" PRIu64, max_discard);
×
420
}
421

422
static int probe_sector_size_harder(int fd, uint32_t *ret) {
2,431✔
423
        _cleanup_close_ int non_direct_io_fd = -EBADF;
2,431✔
424
        int probe_fd, f_flags;
2,431✔
425

426
        assert(fd >= 0);
2,431✔
427
        assert(ret);
2,431✔
428

429
        /* Wraps probe_sector_size() but handles O_DIRECT: if the fd is opened with O_DIRECT there are
430
         * strict alignment requirements for reads, so we temporarily reopen it without O_DIRECT for the
431
         * probing logic. */
432

433
        f_flags = fcntl(fd, F_GETFL);
2,431✔
434
        if (f_flags < 0)
2,431✔
435
                return -errno;
×
436

437
        if (FLAGS_SET(f_flags, O_DIRECT)) {
2,431✔
438
                non_direct_io_fd = fd_reopen(fd, O_RDONLY|O_CLOEXEC|O_NONBLOCK);
2,424✔
439
                if (non_direct_io_fd < 0)
2,424✔
440
                        return non_direct_io_fd;
441

442
                probe_fd = non_direct_io_fd;
443
        } else
444
                probe_fd = fd;
445

446
        return probe_sector_size(probe_fd, ret);
2,431✔
447
}
448

449
static int loop_device_can_shortcut(
1✔
450
                int fd,
451
                uint64_t offset,
452
                uint64_t size,
453
                uint32_t sector_size,
454
                uint32_t device_ssz,
455
                uint32_t loop_flags) {
456

457
        int r;
1✔
458

459
        /* Returns whether we can hand back the original block device fd instead of allocating a real
460
         * loopback device for it: it must cover the whole device, the requested sector size must match the
461
         * device's sector size, and if partscan was requested it must already be enabled on the device
462
         * (otherwise e.g. partition block devices or loop devices created without LO_FLAGS_PARTSCAN would
463
         * be reused even though they cannot expose nested partitions). */
464

465
        assert(fd >= 0);
1✔
466

467
        if (offset != 0)
1✔
468
                return false;
469
        if (!IN_SET(size, 0, UINT64_MAX))
1✔
470
                return false;
471
        if (sector_size != device_ssz)
1✔
472
                return false;
473

474
        if (FLAGS_SET(loop_flags, LO_FLAGS_PARTSCAN)) {
1✔
475
                r = blockdev_partscan_enabled_fd(fd);
1✔
476
                if (r < 0)
1✔
477
                        return r;
478
                if (r == 0)
1✔
479
                        return false;
×
480
        }
481

482
        return true;
483
}
484

485
static int loop_device_make_internal(
2,566✔
486
                const char *path,
487
                int fd,
488
                int open_flags,
489
                uint64_t offset,
490
                uint64_t size,
491
                uint32_t sector_size,
492
                uint32_t loop_flags,
493
                int lock_op,
494
                LoopDevice **ret) {
495

496
        _cleanup_(loop_device_unrefp) LoopDevice *d = NULL;
×
497
        _cleanup_close_ int reopened_fd = -EBADF, control = -EBADF;
5,132✔
498
        _cleanup_free_ char *backing_file = NULL;
2,566✔
499
        struct loop_config config;
2,566✔
500
        int r, f_flags;
2,566✔
501
        struct stat st;
2,566✔
502

503
        assert(fd >= 0);
2,566✔
504
        assert(open_flags < 0 || IN_SET(open_flags, O_RDWR, O_RDONLY));
2,566✔
505
        assert(ret);
2,566✔
506

507
        /* sector_size interpretation:
508
         *   0          → use device sector size for block devices, 512 for regular files
509
         *   UINT32_MAX → probe GPT header to find the right sector size, fall back to 0 behavior
510
         *   other      → use the specified sector size explicitly */
511

512
        f_flags = fcntl(fd, F_GETFL);
2,566✔
513
        if (f_flags < 0)
2,566✔
514
                return -errno;
×
515

516
        if (open_flags < 0) {
2,566✔
517
                /* If open_flags is unset, initialize it from the open fd */
518
                if (FLAGS_SET(f_flags, O_PATH))
6✔
519
                        return log_debug_errno(SYNTHETIC_ERRNO(EBADFD), "Access mode of image file indicates O_PATH, cannot determine read/write flags.");
×
520

521
                open_flags = f_flags & O_ACCMODE_STRICT;
6✔
522
                if (!IN_SET(open_flags, O_RDWR, O_RDONLY))
6✔
523
                        return log_debug_errno(SYNTHETIC_ERRNO(EBADFD), "Access mode of image file is write only (?)");
×
524
        }
525

526
        if (sector_size == UINT32_MAX) {
2,566✔
527
                /* If sector size is specified as UINT32_MAX, we'll try to probe the right sector size
528
                 * by looking for the GPT partition header at various offsets. This of course only works
529
                 * if the image already has a disk label. */
530

531
                r = probe_sector_size_harder(fd, &sector_size);
2,431✔
532
                if (r < 0)
2,431✔
533
                        return r;
534
                if (r == 0)
2,431✔
535
                        sector_size = 0; /* If we can't probe anything, use default sector size. */
2,311✔
536
        }
537

538
        if (fstat(fd, &st) < 0)
2,566✔
539
                return -errno;
×
540

541
        if (S_ISBLK(st.st_mode)) {
2,566✔
542
                uint32_t device_ssz;
1✔
543
                r = blockdev_get_sector_size(fd, &device_ssz);
1✔
544
                if (r < 0)
1✔
545
                        return r;
1✔
546

547
                if (sector_size == 0)
1✔
548
                        sector_size = device_ssz;
×
549

550
                r = loop_device_can_shortcut(fd, offset, size, sector_size, device_ssz, loop_flags);
1✔
551
                if (r < 0)
1✔
552
                        return r;
553
                if (r > 0)
1✔
554
                        return loop_device_open_from_fd(fd, open_flags, lock_op, ret);
1✔
555
        } else {
556
                r = stat_verify_regular(&st);
2,565✔
557
                if (r < 0)
2,565✔
558
                        return r;
559

560
                if (sector_size == 0)
2,565✔
561
                        sector_size = 512;
2,311✔
562
        }
563

564
        if (path) {
2,565✔
565
                r = path_make_absolute_cwd(path, &backing_file);
344✔
566
                if (r < 0)
344✔
567
                        return r;
568

569
                path_simplify(backing_file);
344✔
570
        } else {
571
                r = fd_get_path(fd, &backing_file);
2,221✔
572
                if (r < 0)
2,221✔
573
                        return r;
574
        }
575

576
        if (FLAGS_SET(loop_flags, LO_FLAGS_DIRECT_IO) != FLAGS_SET(f_flags, O_DIRECT)) {
2,565✔
577
                /* If LO_FLAGS_DIRECT_IO is requested, then make sure we have the fd open with O_DIRECT, as
578
                 * that's required. Conversely, if it's off require that O_DIRECT is off too (that's because
579
                 * new kernels will implicitly enable LO_FLAGS_DIRECT_IO if O_DIRECT is set).
580
                 *
581
                 * Our intention here is that LO_FLAGS_DIRECT_IO is the primary knob, and O_DIRECT derived
582
                 * from that automatically. */
583

584
                reopened_fd = fd_reopen(fd, (FLAGS_SET(loop_flags, LO_FLAGS_DIRECT_IO) ? O_DIRECT : 0)|O_CLOEXEC|O_NONBLOCK|open_flags);
141✔
585
                if (reopened_fd < 0) {
141✔
586
                        if (!FLAGS_SET(loop_flags, LO_FLAGS_DIRECT_IO))
×
587
                                return log_debug_errno(reopened_fd, "Failed to reopen file descriptor without O_DIRECT: %m");
×
588

589
                        /* Some file systems might not support O_DIRECT, let's gracefully continue without it then. */
590
                        log_debug_errno(reopened_fd, "Failed to enable O_DIRECT for backing file descriptor for loopback device. Continuing without.");
×
591
                        loop_flags &= ~LO_FLAGS_DIRECT_IO;
×
592
                } else
593
                        fd = reopened_fd; /* From now on, operate on our new O_DIRECT fd */
594
        }
595

596
        control = open("/dev/loop-control", O_RDWR|O_CLOEXEC|O_NOCTTY|O_NONBLOCK);
2,565✔
597
        if (control < 0)
2,565✔
598
                return -errno;
20✔
599

600
        /* Strip LO_FLAGS_PARTSCAN from LOOP_CONFIGURE and enable it afterwards via
601
         * LOOP_SET_STATUS64 to work around a kernel race: LOOP_CONFIGURE sends a uevent with
602
         * GD_NEED_PART_SCAN set before calling loop_reread_partitions(). If udev opens the device in
603
         * response, blkdev_get_whole() triggers a first scan, then loop_reread_partitions() does a
604
         * second scan that briefly drops all partitions. By configuring without partscan,
605
         * GD_SUPPRESS_PART_SCAN stays set, making any concurrent open harmless. LOOP_SET_STATUS64
606
         * doesn't call disk_force_media_change() so it doesn't set GD_NEED_PART_SCAN.
607
         *
608
         * See: https://lore.kernel.org/linux-block/20260330081819.652890-1-daan@amutable.com/T/#u
609
         * Drop this workaround once the kernel fix is widely available. */
610
        bool deferred_partscan = FLAGS_SET(loop_flags, LO_FLAGS_PARTSCAN);
2,545✔
611

612
        config = (struct loop_config) {
5,090✔
613
                .fd = fd,
614
                .block_size = sector_size,
615
                .info = {
616
                        /* Use the specified flags, but configure the read-only flag from the open flags, and force autoclear */
617
                        .lo_flags = ((loop_flags & ~(LO_FLAGS_READ_ONLY|LO_FLAGS_PARTSCAN)) |
5,090✔
618
                                     ((open_flags & O_ACCMODE_STRICT) == O_RDONLY ? LO_FLAGS_READ_ONLY : 0) |
2,545✔
619
                                     LO_FLAGS_AUTOCLEAR),
620
                        .lo_offset = offset,
621
                        .lo_sizelimit = size == UINT64_MAX ? 0 : size,
2,545✔
622
                },
623
        };
624

625
        /* Loop around LOOP_CTL_GET_FREE, since at the moment we attempt to open the returned device it might
626
         * be gone already, taken by somebody else racing against us. */
627
        for (unsigned n_attempts = 0;;) {
2,545✔
628
                usec_t usec;
2,545✔
629
                int nr;
2,545✔
630

631
                /* Let's take a lock on the control device first. On a busy system, where many programs
632
                 * attempt to allocate a loopback device at the same time, we might otherwise keep looping
633
                 * around relatively heavy operations: asking for a free loopback device, then opening it,
634
                 * validating it, attaching something to it. Let's serialize this whole operation, to make
635
                 * unnecessary busywork less likely. Note that this is just something we do to optimize our
636
                 * own code (and whoever else decides to use LOCK_EX locks for this), taking this lock is not
637
                 * necessary, it just means it's less likely we have to iterate through this loop again and
638
                 * again if our own code races against our own code.
639
                 *
640
                 * Note: our lock protocol is to take the /dev/loop-control lock first, and the block device
641
                 * lock second, if both are taken, and always in this order, to avoid ABBA locking issues. */
642
                if (flock(control, LOCK_EX) < 0)
2,545✔
643
                        return -errno;
×
644

645
                nr = ioctl(control, LOOP_CTL_GET_FREE);
2,545✔
646
                if (nr < 0)
2,545✔
647
                        return -errno;
×
648

649
                r = loop_configure(nr, open_flags, lock_op, &config, &d);
2,545✔
650
                if (r >= 0)
2,545✔
651
                        break;
652

653
                /* -ENODEV or friends: Somebody might've gotten the same number from the kernel, used the
654
                 * device, and called LOOP_CTL_REMOVE on it. Let's retry with a new number.
655
                 * -EBUSY: a file descriptor is already bound to the loopback block device.
656
                 * -EUCLEAN: some left-over partition devices that were cleaned up.
657
                 * -ENOANO: we tried to use LO_FLAGS_DIRECT_IO but the kernel rejected it. */
658
                if (!ERRNO_IS_DEVICE_ABSENT(r) && !IN_SET(r, -EBUSY, -EUCLEAN, -ENOANO))
×
659
                        return r;
660

661
                /* OK, this didn't work, let's try again a bit later, but first release the lock on the
662
                 * control device */
663
                if (flock(control, LOCK_UN) < 0)
×
664
                        return -errno;
×
665

666
                if (++n_attempts >= 64) /* Give up eventually */
×
667
                        return -EBUSY;
668

669
                /* If we failed to enable direct IO mode, let's retry without it. We restart the process as
670
                 * on some combination of kernel version and storage filesystem, the kernel is very unhappy
671
                 * about a failed DIRECT_IO enablement and throws I/O errors. */
672
                if (r == -ENOANO && FLAGS_SET(config.info.lo_flags, LO_FLAGS_DIRECT_IO)) {
×
673
                        config.info.lo_flags &= ~LO_FLAGS_DIRECT_IO;
×
674
                        open_flags &= ~O_DIRECT;
×
675

676
                        int non_direct_io_fd = fd_reopen(config.fd, O_CLOEXEC|O_NONBLOCK|open_flags);
×
677
                        if (non_direct_io_fd < 0)
×
678
                                return log_debug_errno(
×
679
                                                non_direct_io_fd,
680
                                                "Failed to reopen file descriptor without O_DIRECT: %m");
681

682
                        safe_close(reopened_fd);
×
683
                        fd = config.fd = /* For cleanups */ reopened_fd = non_direct_io_fd;
×
684
                }
685

686
                /* Wait some random time, to make collision less likely. Let's pick a random time in the
687
                 * range 0ms…250ms, linearly scaled by the number of failed attempts. */
688
                usec = random_u64_range(UINT64_C(10) * USEC_PER_MSEC +
×
689
                                        UINT64_C(240) * USEC_PER_MSEC * n_attempts/64);
×
690
                log_debug("Trying again after %s.", FORMAT_TIMESPAN(usec, USEC_PER_MSEC));
×
691
                (void) usleep_safe(usec);
×
692
        }
693

694
        if (S_ISBLK(st.st_mode)) {
2,545✔
695
                /* Propagate backing device's discard byte limit to our loopback block device. We do this in
696
                 * order to avoid that (supposedly quick) discard requests on the loopback device get turned
697
                 * into (likely slow) zero-out requests on backing devices that do not support discarding
698
                 * natively, but do support zero-out. */
699
                uint64_t discard_max_bytes;
×
700

701
                r = fd_get_max_discard(fd, &discard_max_bytes);
×
702
                if (r < 0)
×
703
                        log_debug_errno(r, "Failed to read 'discard_max_bytes' of backing device, ignoring: %m");
×
704
                else {
705
                        r = fd_set_max_discard(d->fd, discard_max_bytes);
×
706
                        if (r < 0)
×
707
                                log_debug_errno(r, "Failed to write 'discard_max_bytes' of loop device, ignoring: %m");
×
708
                }
709
        }
710

711
        if (deferred_partscan) {
2,545✔
712
                /* Open+close to drain GD_NEED_PART_SCAN harmlessly (GD_SUPPRESS_PART_SCAN is still
713
                 * set so no partitions appear). Then enable partscan via LOOP_SET_STATUS64. */
714
                int tmp_fd = fd_reopen(d->fd, O_RDONLY|O_CLOEXEC|O_NONBLOCK);
2,366✔
715
                if (tmp_fd < 0)
2,366✔
716
                        return log_debug_errno(tmp_fd, "Failed to reopen loop device to drain partscan flag: %m");
×
717
                safe_close(tmp_fd);
2,366✔
718

719
                struct loop_info64 info;
2,366✔
720
                if (ioctl(d->fd, LOOP_GET_STATUS64, &info) < 0)
2,366✔
721
                        return log_debug_errno(errno, "Failed to get loop device status: %m");
×
722

723
                info.lo_flags |= LO_FLAGS_PARTSCAN;
2,366✔
724

725
                if (ioctl(d->fd, LOOP_SET_STATUS64, &info) < 0)
2,366✔
726
                        return log_debug_errno(errno, "Failed to enable partscan on loop device: %m");
×
727
        }
728

729
        d->backing_file = TAKE_PTR(backing_file);
2,545✔
730
        d->backing_inode = st.st_ino;
2,545✔
731
        d->backing_devno = st.st_dev;
2,545✔
732

733
        log_debug("Successfully acquired %s, devno=%u:%u, nr=%i, diskseq=%" PRIu64,
2,545✔
734
                  d->node,
735
                  major(d->devno), minor(d->devno),
736
                  d->nr,
737
                  d->diskseq);
738

739
        *ret = TAKE_PTR(d);
2,545✔
740
        return 0;
2,545✔
741
}
742

743
static uint32_t loop_flags_mangle(uint32_t loop_flags) {
2,566✔
744
        int r;
2,566✔
745

746
        r = getenv_bool("SYSTEMD_LOOP_DIRECT_IO");
2,566✔
747
        if (r < 0 && r != -ENXIO)
2,566✔
748
                log_debug_errno(r, "Failed to parse $SYSTEMD_LOOP_DIRECT_IO, ignoring: %m");
×
749

750
        return UPDATE_FLAG(loop_flags, LO_FLAGS_DIRECT_IO, r != 0); /* Turn on LO_FLAGS_DIRECT_IO by default, unless explicitly configured to off. */
2,566✔
751
}
752

753
int loop_device_make(
141✔
754
                int fd,
755
                int open_flags,
756
                uint64_t offset,
757
                uint64_t size,
758
                uint32_t sector_size,
759
                uint32_t loop_flags,
760
                int lock_op,
761
                LoopDevice **ret) {
762

763
        assert(fd >= 0);
141✔
764
        assert(ret);
141✔
765

766
        return loop_device_make_internal(
141✔
767
                        NULL,
768
                        fd,
769
                        open_flags,
770
                        offset,
771
                        size,
772
                        sector_size,
773
                        loop_flags_mangle(loop_flags),
774
                        lock_op,
775
                        ret);
776
}
777

778
int loop_device_make_by_path_at(
2,425✔
779
                int dir_fd,
780
                const char *path,
781
                int open_flags,
782
                uint32_t sector_size,
783
                uint32_t loop_flags,
784
                int lock_op,
785
                LoopDevice **ret) {
786

787
        int r, basic_flags, direct_flags, rdwr_flags;
2,425✔
788
        _cleanup_close_ int fd = -EBADF;
2,425✔
789
        bool direct = false;
2,425✔
790

791
        assert(dir_fd >= 0 || dir_fd == AT_FDCWD);
2,425✔
792
        assert(ret);
2,425✔
793
        assert(open_flags < 0 || IN_SET(open_flags, O_RDWR, O_RDONLY));
2,425✔
794

795
        /* Passing < 0 as open_flags here means we'll try to open the device writable if we can, retrying
796
         * read-only if we cannot. */
797

798
        loop_flags = loop_flags_mangle(loop_flags);
2,425✔
799

800
        /* Let's open with O_DIRECT if we can. But not all file systems support that, hence fall back to
801
         * non-O_DIRECT mode automatically, if it fails. */
802

803
        basic_flags = O_CLOEXEC|O_NONBLOCK|O_NOCTTY;
2,425✔
804
        direct_flags = FLAGS_SET(loop_flags, LO_FLAGS_DIRECT_IO) ? O_DIRECT : 0;
2,425✔
805
        rdwr_flags = open_flags >= 0 ? open_flags : O_RDWR;
2,425✔
806

807
        fd = xopenat(dir_fd, path, basic_flags|direct_flags|rdwr_flags);
2,425✔
808
        if (fd < 0 && direct_flags != 0) /* If we had O_DIRECT on, and things failed with that, let's immediately try again without */
2,425✔
809
                fd = xopenat(dir_fd, path, basic_flags|rdwr_flags);
1✔
810
        else
811
                direct = direct_flags != 0;
×
812
        if (fd < 0) {
2,425✔
813
                r = fd;
1✔
814

815
                /* Retry read-only? */
816
                if (open_flags >= 0 || !ERRNO_IS_NEG_FS_WRITE_REFUSED(r))
2,426✔
817
                        return r;
818

819
                fd = xopenat(dir_fd, path, basic_flags|direct_flags|O_RDONLY);
×
820
                if (fd < 0 && direct_flags != 0) /* as above */
×
821
                        fd = xopenat(dir_fd, path, basic_flags|O_RDONLY);
×
822
                else
823
                        direct = direct_flags != 0;
×
824
                if (fd < 0)
×
825
                        return r; /* Propagate original error */
826

827
                open_flags = O_RDONLY;
828
        } else if (open_flags < 0)
2,424✔
829
                open_flags = O_RDWR;
148✔
830

831
        log_debug("Opened %s in %s access mode%s, with O_DIRECT %s%s.",
11,564✔
832
                  path ?: "loop device",
833
                  open_flags == O_RDWR ? "O_RDWR" : "O_RDONLY",
834
                  open_flags != rdwr_flags ? " (O_RDWR was requested but not allowed)" : "",
835
                  direct ? "enabled" : "disabled",
836
                  direct != (direct_flags != 0) ? " (O_DIRECT was requested but not supported)" : "");
837

838
        return loop_device_make_internal(
4,503✔
839
                        dir_fd == AT_FDCWD ? path : NULL,
840
                        fd,
841
                        open_flags,
842
                        /* offset= */ 0,
843
                        /* size= */ 0,
844
                        sector_size,
845
                        loop_flags,
846
                        lock_op,
847
                        ret);
848
}
849

850
int loop_device_make_by_path_memory(
1✔
851
                const char *path,
852
                int open_flags,
853
                uint32_t sector_size,
854
                uint32_t loop_flags,
855
                int lock_op,
856
                LoopDevice **ret) {
857

858
        _cleanup_close_ int fd = -EBADF, mfd = -EBADF;
1✔
859
        _cleanup_free_ char *fn = NULL;
1✔
860
        int r;
1✔
861

862
        assert(path);
1✔
863
        assert(open_flags < 0 || IN_SET(open_flags, O_RDWR, O_RDONLY));
1✔
864
        assert(ret);
1✔
865

866
        /* memfds are always writable, so default to O_RDWR when auto-detecting. */
867
        if (open_flags < 0)
1✔
868
                open_flags = O_RDWR;
1✔
869

870
        loop_flags &= ~LO_FLAGS_DIRECT_IO; /* memfds don't support O_DIRECT, hence LO_FLAGS_DIRECT_IO can't be used either */
1✔
871

872
        fd = open(path, O_CLOEXEC|O_NONBLOCK|O_NOCTTY|O_RDONLY);
1✔
873
        if (fd < 0)
1✔
874
                return -errno;
×
875

876
        r = fd_verify_regular_or_block(fd);
1✔
877
        if (r < 0)
1✔
878
                return r;
879

880
        r = path_extract_filename(path, &fn);
1✔
881
        if (r < 0)
1✔
882
                return r;
883

884
        mfd = memfd_clone_fd(fd, fn, open_flags|O_CLOEXEC);
1✔
885
        if (mfd < 0)
1✔
886
                return mfd;
887

888
        fd = safe_close(fd); /* Let's close the original early */
1✔
889

890
        return loop_device_make_internal(NULL, mfd, open_flags, 0, 0, sector_size, loop_flags, lock_op, ret);
1✔
891
}
892

893
static LoopDevice* loop_device_free(LoopDevice *d) {
2,580✔
894
        _cleanup_close_ int control = -EBADF;
2,580✔
895
        int r;
2,580✔
896

897
        if (!d)
2,580✔
898
                return NULL;
899

900
        /* Release any lock we might have on the device first. We want to open+lock the /dev/loop-control
901
         * device below, but our lock protocol says that if both control and block device locks are taken,
902
         * the control lock needs to be taken first, the block device lock second — in order to avoid ABBA
903
         * locking issues. Moreover, we want to issue LOOP_CLR_FD on the block device further down, and that
904
         * would fail if we had another fd open to the device. */
905
        d->lock_fd = safe_close(d->lock_fd);
2,580✔
906

907
        /* Let's open the control device early, and lock it, so that we can release our block device and
908
         * delete it in a synchronized fashion, and allocators won't needlessly see the block device as free
909
         * while we are about to delete it. */
910
        if (!LOOP_DEVICE_IS_FOREIGN(d) && !d->relinquished) {
2,580✔
911
                control = open("/dev/loop-control", O_RDWR|O_CLOEXEC|O_NOCTTY|O_NONBLOCK);
2,269✔
912
                if (control < 0)
2,269✔
913
                        log_debug_errno(errno, "Failed to open loop control device, cannot remove loop device '%s', ignoring: %m", strna(d->node));
×
914
                else if (flock(control, LOCK_EX) < 0)
2,269✔
915
                        log_debug_errno(errno, "Failed to lock loop control device, ignoring: %m");
×
916
        }
917

918
        /* Then let's release the loopback block device */
919
        if (d->fd >= 0) {
2,580✔
920
                /* Implicitly sync the device, since otherwise in-flight blocks might not get written */
921
                if (fsync(d->fd) < 0)
2,580✔
922
                        log_debug_errno(errno, "Failed to sync loop block device, ignoring: %m");
×
923

924
                if (!LOOP_DEVICE_IS_FOREIGN(d) && !d->relinquished) {
2,580✔
925
                        /* We are supposed to clear the loopback device. Let's do this synchronously: lock
926
                         * the device, manually remove all partitions and then clear it. This should ensure
927
                         * udev doesn't concurrently access the devices, and we can be reasonably sure that
928
                         * once we are done here the device is cleared and all its partition children
929
                         * removed. Note that we lock our primary device fd here (and not a separate locking
930
                         * fd, as we do during allocation, since we want to keep the lock all the way through
931
                         * the LOOP_CLR_FD, but that call would fail if we had more than one fd open.) */
932

933
                        if (flock(d->fd, LOCK_EX) < 0)
2,269✔
934
                                log_debug_errno(errno, "Failed to lock loop block device, ignoring: %m");
×
935

936
                        r = block_device_remove_all_partitions(d->dev, d->fd);
2,269✔
937
                        if (r < 0)
2,269✔
938
                                log_debug_errno(r, "Failed to remove partitions of loopback block device, ignoring: %m");
×
939

940
                        if (ioctl(d->fd, LOOP_CLR_FD) < 0)
2,269✔
941
                                log_debug_errno(errno, "Failed to clear loop device, ignoring: %m");
×
942
                }
943

944
                safe_close(d->fd);
2,580✔
945
        }
946

947
        /* Now that the block device is released, let's also try to remove it */
948
        if (control >= 0) {
2,580✔
949
                useconds_t delay = 5 * USEC_PER_MSEC;  /* A total delay of 5090 ms between 39 attempts,
950
                                                        * (4*5 + 5*10 + 5*20 + … + 3*640) = 5090. */
951

952
                for (unsigned attempt = 1;; attempt++) {
421✔
953
                        if (ioctl(control, LOOP_CTL_REMOVE, d->nr) >= 0)
2,690✔
954
                                break;
955
                        if (errno != EBUSY || attempt > 38) {
422✔
956
                                log_debug_errno(errno, "Failed to remove device %s: %m", strna(d->node));
1✔
957
                                break;
958
                        }
959
                        if (attempt % 5 == 0) {
421✔
960
                                log_debug("Device is still busy after %u attempts…", attempt);
58✔
961
                                delay *= 2;
58✔
962
                        }
963

964
                        (void) usleep_safe(delay);
421✔
965
                }
966
        }
967

968
        free(d->node);
2,580✔
969
        sd_device_unref(d->dev);
2,580✔
970
        free(d->backing_file);
2,580✔
971
        return mfree(d);
2,580✔
972
}
973

974
DEFINE_TRIVIAL_REF_UNREF_FUNC(LoopDevice, loop_device, loop_device_free);
8,061✔
975

976
void loop_device_relinquish(LoopDevice *d) {
202✔
977
        assert(d);
202✔
978

979
        /* Don't attempt to clean up the loop device anymore from this point on. Leave the clean-ing up to the kernel
980
         * itself, using the loop device "auto-clear" logic we already turned on when creating the device. */
981

982
        d->relinquished = true;
202✔
983
}
202✔
984

985
void loop_device_unrelinquish(LoopDevice *d) {
25✔
986
        assert(d);
25✔
987
        d->relinquished = false;
25✔
988
}
25✔
989

990
int loop_device_open(
134✔
991
                sd_device *dev,
992
                int open_flags,
993
                int lock_op,
994
                LoopDevice **ret) {
995

996
        _cleanup_close_ int fd = -EBADF, lock_fd = -EBADF;
134✔
997
        _cleanup_free_ char *node = NULL, *backing_file = NULL;
134✔
998
        dev_t devnum, backing_devno = 0;
134✔
999
        struct loop_info64 info;
134✔
1000
        ino_t backing_inode = 0;
134✔
1001
        uint64_t diskseq = 0;
134✔
1002
        LoopDevice *d;
134✔
1003
        const char *s;
134✔
1004
        int r, nr = -1;
134✔
1005

1006
        assert(dev);
134✔
1007
        assert(IN_SET(open_flags, O_RDWR, O_RDONLY));
134✔
1008
        assert(ret);
134✔
1009

1010
        /* Even if fd is provided through the argument in loop_device_open_from_fd(), we reopen the inode
1011
         * here, instead of keeping just a dup() clone of it around, since we want to ensure that the
1012
         * O_DIRECT flag of the handle we keep is off, we have our own file index, and have the right
1013
         * read/write mode in effect. */
1014
        fd = sd_device_open(dev, O_CLOEXEC|O_NONBLOCK|O_NOCTTY|open_flags);
134✔
1015
        if (fd < 0)
134✔
1016
                return fd;
1017

1018
        if ((lock_op & ~LOCK_NB) != LOCK_UN) {
134✔
1019
                lock_fd = open_lock_fd(fd, lock_op);
134✔
1020
                if (lock_fd < 0)
134✔
1021
                        return lock_fd;
1022
        }
1023

1024
        if (ioctl(fd, LOOP_GET_STATUS64, &info) >= 0) {
134✔
1025
#if HAVE_VALGRIND_MEMCHECK_H
1026
                /* Valgrind currently doesn't know LOOP_GET_STATUS64. Remove this once it does */
1027
                VALGRIND_MAKE_MEM_DEFINED(&info, sizeof(info));
1028
#endif
1029
                nr = info.lo_number;
30✔
1030

1031
                if (device_get_sysattr_safe_string(dev, "loop/backing_file", &s) >= 0) {
30✔
1032
                        backing_file = strdup(s);
21✔
1033
                        if (!backing_file)
21✔
1034
                                return -ENOMEM;
1035
                }
1036

1037
                backing_devno = info.lo_device;
30✔
1038
                backing_inode = info.lo_inode;
30✔
1039
        }
1040

1041
        r = fd_get_diskseq(fd, &diskseq);
134✔
1042
        if (r < 0 && r != -EOPNOTSUPP)
134✔
1043
                return r;
1044

1045
        uint32_t sector_size;
134✔
1046
        r = blockdev_get_sector_size(fd, &sector_size);
134✔
1047
        if (r < 0)
134✔
1048
                return r;
1049

1050
        uint64_t device_size;
134✔
1051
        r = blockdev_get_device_size(fd, &device_size);
134✔
1052
        if (r < 0)
134✔
1053
                return r;
1054

1055
        r = sd_device_get_devnum(dev, &devnum);
134✔
1056
        if (r < 0)
134✔
1057
                return r;
1058

1059
        r = sd_device_get_devname(dev, &s);
134✔
1060
        if (r < 0)
134✔
1061
                return r;
1062

1063
        node = strdup(s);
134✔
1064
        if (!node)
134✔
1065
                return -ENOMEM;
1066

1067
        d = new(LoopDevice, 1);
134✔
1068
        if (!d)
134✔
1069
                return -ENOMEM;
1070

1071
        *d = (LoopDevice) {
268✔
1072
                .n_ref = 1,
1073
                .fd = TAKE_FD(fd),
134✔
1074
                .lock_fd = TAKE_FD(lock_fd),
134✔
1075
                .nr = nr,
1076
                .node = TAKE_PTR(node),
134✔
1077
                .dev = sd_device_ref(dev),
134✔
1078
                .backing_file = TAKE_PTR(backing_file),
134✔
1079
                .backing_inode = backing_inode,
1080
                .backing_devno = backing_devno,
1081
                .relinquished = true, /* It's not ours, don't try to destroy it when this object is freed */
1082
                .devno = devnum,
1083
                .diskseq = diskseq,
1084
                .sector_size = sector_size,
1085
                .device_size = device_size,
1086
                .created = false,
1087
        };
1088

1089
        *ret = d;
134✔
1090
        return 0;
134✔
1091
}
1092

1093
int loop_device_open_from_fd(
6✔
1094
                int fd,
1095
                int open_flags,
1096
                int lock_op,
1097
                LoopDevice **ret) {
1098

1099
        _cleanup_(sd_device_unrefp) sd_device *dev = NULL;
6✔
1100
        int r;
6✔
1101

1102
        r = block_device_new_from_fd(ASSERT_FD(fd), 0, &dev);
6✔
1103
        if (r < 0)
6✔
1104
                return r;
1105

1106
        return loop_device_open(dev, open_flags, lock_op, ret);
6✔
1107
}
1108

1109
int loop_device_open_from_path(
×
1110
                const char *path,
1111
                int open_flags,
1112
                int lock_op,
1113
                LoopDevice **ret) {
1114

1115
        _cleanup_(sd_device_unrefp) sd_device *dev = NULL;
×
1116
        int r;
×
1117

1118
        assert(path);
×
1119

1120
        r = block_device_new_from_path(path, 0, &dev);
×
1121
        if (r < 0)
×
1122
                return r;
1123

1124
        return loop_device_open(dev, open_flags, lock_op, ret);
×
1125
}
1126

1127
static int resize_partition(int partition_fd, uint64_t offset, uint64_t size) {
×
1128
        char sysfs[STRLEN("/sys/dev/block/:/partition") + 2*DECIMAL_STR_MAX(dev_t) + 1];
×
1129
        _cleanup_free_ char *buffer = NULL;
×
1130
        uint64_t current_offset, current_size, partno;
×
1131
        _cleanup_close_ int whole_fd = -EBADF;
×
1132
        struct stat st;
×
1133
        dev_t devno;
×
1134
        int r;
×
1135

1136
        /* Resizes the partition the loopback device refer to (assuming it refers to one instead of an actual
1137
         * loopback device), and changes the offset, if needed. This is a fancy wrapper around
1138
         * BLKPG_RESIZE_PARTITION. */
1139

1140
        if (fstat(ASSERT_FD(partition_fd), &st) < 0)
×
1141
                return -errno;
×
1142

1143
        assert(S_ISBLK(st.st_mode));
×
1144

1145
        xsprintf(sysfs, "/sys/dev/block/" DEVNUM_FORMAT_STR "/partition", DEVNUM_FORMAT_VAL(st.st_rdev));
×
1146
        r = read_one_line_file(sysfs, &buffer);
×
1147
        if (r == -ENOENT) /* not a partition, cannot resize */
×
1148
                return -ENOTTY;
1149
        if (r < 0)
×
1150
                return r;
1151
        r = safe_atou64(buffer, &partno);
×
1152
        if (r < 0)
×
1153
                return r;
1154

1155
        xsprintf(sysfs, "/sys/dev/block/" DEVNUM_FORMAT_STR "/start", DEVNUM_FORMAT_VAL(st.st_rdev));
×
1156

1157
        buffer = mfree(buffer);
×
1158
        r = read_one_line_file(sysfs, &buffer);
×
1159
        if (r < 0)
×
1160
                return r;
1161
        r = safe_atou64(buffer, &current_offset);
×
1162
        if (r < 0)
×
1163
                return r;
1164
        if (current_offset > UINT64_MAX/512U)
×
1165
                return -EINVAL;
1166
        current_offset *= 512U;
×
1167

1168
        r = blockdev_get_device_size(partition_fd, &current_size);
×
1169
        if (r < 0)
×
1170
                return r;
1171

1172
        if (size == UINT64_MAX && offset == UINT64_MAX)
×
1173
                return 0;
1174
        if (current_size == size && current_offset == offset)
×
1175
                return 0;
1176

1177
        xsprintf(sysfs, "/sys/dev/block/" DEVNUM_FORMAT_STR "/../dev", DEVNUM_FORMAT_VAL(st.st_rdev));
×
1178

1179
        buffer = mfree(buffer);
×
1180
        r = read_one_line_file(sysfs, &buffer);
×
1181
        if (r < 0)
×
1182
                return r;
1183
        r = parse_devnum(buffer, &devno);
×
1184
        if (r < 0)
×
1185
                return r;
1186

1187
        whole_fd = r = device_open_from_devnum(S_IFBLK, devno, O_RDWR|O_CLOEXEC|O_NONBLOCK|O_NOCTTY, NULL);
×
1188
        if (r < 0)
×
1189
                return r;
1190

1191
        return block_device_resize_partition(
×
1192
                        whole_fd,
1193
                        partno,
1194
                        offset == UINT64_MAX ? current_offset : offset,
1195
                        size == UINT64_MAX ? current_size : size);
1196
}
1197

1198
int loop_device_refresh_size(LoopDevice *d, uint64_t offset, uint64_t size) {
22✔
1199
        struct loop_info64 info;
22✔
1200

1201
        assert(d);
22✔
1202
        assert(d->fd >= 0);
22✔
1203

1204
        /* Changes the offset/start of the loop device relative to the beginning of the underlying file or
1205
         * block device. If this loop device actually refers to a partition and not a loopback device, we'll
1206
         * try to adjust the partition offsets instead.
1207
         *
1208
         * If either offset or size is UINT64_MAX we won't change that parameter. */
1209

1210
        if (d->nr < 0) /* not a loopback device */
22✔
1211
                return resize_partition(d->fd, offset, size);
×
1212

1213
        if (ioctl(d->fd, LOOP_GET_STATUS64, &info) < 0)
22✔
1214
                return -errno;
×
1215

1216
#if HAVE_VALGRIND_MEMCHECK_H
1217
        /* Valgrind currently doesn't know LOOP_GET_STATUS64. Remove this once it does */
1218
        VALGRIND_MAKE_MEM_DEFINED(&info, sizeof(info));
1219
#endif
1220

1221
        if ((size == UINT64_MAX || info.lo_sizelimit == size) &&
22✔
1222
            (offset == UINT64_MAX || info.lo_offset == offset))
×
1223
                return 0;
1224

1225
        if (size != UINT64_MAX)
22✔
1226
                info.lo_sizelimit = size;
22✔
1227
        if (offset != UINT64_MAX)
22✔
1228
                info.lo_offset = offset;
×
1229

1230
        return RET_NERRNO(ioctl(d->fd, LOOP_SET_STATUS64, &info));
22✔
1231
}
1232

1233
int loop_device_flock(LoopDevice *d, int operation) {
162✔
1234
        assert(IN_SET(operation & ~LOCK_NB, LOCK_UN, LOCK_SH, LOCK_EX));
162✔
1235
        assert(d);
162✔
1236

1237
        /* When unlocking just close the lock fd */
1238
        if ((operation & ~LOCK_NB) == LOCK_UN) {
162✔
1239
                d->lock_fd = safe_close(d->lock_fd);
160✔
1240
                return 0;
160✔
1241
        }
1242

1243
        /* If we had no lock fd so far, create one and lock it right-away */
1244
        if (d->lock_fd < 0) {
2✔
1245
                d->lock_fd = open_lock_fd(ASSERT_FD(d->fd), operation);
1✔
1246
                if (d->lock_fd < 0)
1✔
1247
                        return d->lock_fd;
1248

1249
                return 0;
1✔
1250
        }
1251

1252
        /* Otherwise change the current lock mode on the existing fd */
1253
        return RET_NERRNO(flock(d->lock_fd, operation));
1✔
1254
}
1255

1256
int loop_device_sync(LoopDevice *d) {
79✔
1257
        assert(d);
79✔
1258

1259
        /* We also do this implicitly in loop_device_unref(). Doing this explicitly here has the benefit that
1260
         * we can check the return value though. */
1261

1262
        return RET_NERRNO(fsync(ASSERT_FD(d->fd)));
79✔
1263
}
1264

1265
int loop_device_set_autoclear(LoopDevice *d, bool autoclear) {
13✔
1266
        struct loop_info64 info;
13✔
1267

1268
        assert(d);
13✔
1269

1270
        if (LOOP_DEVICE_IS_FOREIGN(d))
13✔
1271
                return 0;
13✔
1272

1273
        if (ioctl(ASSERT_FD(d->fd), LOOP_GET_STATUS64, &info) < 0)
13✔
1274
                return -errno;
×
1275

1276
        if (autoclear == FLAGS_SET(info.lo_flags, LO_FLAGS_AUTOCLEAR))
13✔
1277
                return 0;
1278

1279
        SET_FLAG(info.lo_flags, LO_FLAGS_AUTOCLEAR, autoclear);
13✔
1280

1281
        if (ioctl(d->fd, LOOP_SET_STATUS64, &info) < 0)
13✔
1282
                return -errno;
×
1283

1284
        return 1;
1285
}
1286

1287
int loop_device_set_filename(LoopDevice *d, const char *name) {
5✔
1288
        struct loop_info64 info;
5✔
1289

1290
        assert(d);
5✔
1291

1292
        /* Sets the .lo_file_name of the loopback device. This is supposed to contain the path to the file
1293
         * backing the block device, but is actually just a free-form string you can pass to the kernel. Most
1294
         * tools that actually care for the backing file path use the sysfs attribute file loop/backing_file
1295
         * which is a kernel generated string, subject to file system namespaces and such.
1296
         *
1297
         * .lo_file_name is useful since userspace can select it freely when creating a loopback block
1298
         * device, and we can use it for /dev/disk/by-loop-ref/ symlinks, and similar, so that apps can
1299
         * recognize their own loopback files. */
1300

1301
        if (name && strlen(name) >= sizeof(info.lo_file_name))
5✔
1302
                return -ENOBUFS;
5✔
1303

1304
        if (ioctl(ASSERT_FD(d->fd), LOOP_GET_STATUS64, &info) < 0)
5✔
1305
                return -errno;
×
1306

1307
        if (strneq((char*) info.lo_file_name, strempty(name), sizeof(info.lo_file_name)))
5✔
1308
                return 0;
1309

1310
        if (name) {
5✔
1311
                strncpy((char*) info.lo_file_name, name, sizeof(info.lo_file_name)-1);
5✔
1312
                info.lo_file_name[sizeof(info.lo_file_name)-1] = 0;
5✔
1313
        } else
1314
                memzero(info.lo_file_name, sizeof(info.lo_file_name));
×
1315

1316
        if (ioctl(d->fd, LOOP_SET_STATUS64, &info) < 0)
5✔
1317
                return -errno;
×
1318

1319
        return 1;
1320
}
STATUS · Troubleshooting · Open an Issue · Sales · Support · CAREERS · ENTERPRISE · START FREE · SCHEDULE DEMO
ANNOUNCEMENTS · TWITTER · TOS & SLA · Supported CI Services · What's a CI service? · Automated Testing

© 2026 Coveralls, Inc