| 1 |
/* SPDX-License-Identifier: LGPL-2.1-or-later */ |
| 2 |
|
| 3 |
#include <fcntl.h> |
| 4 |
#include <stdio.h> |
| 5 |
#include <stdlib.h> |
| 6 |
#include <string.h> |
| 7 |
#include <sys/file.h> |
| 8 |
#include <unistd.h> |
| 9 |
|
| 10 |
#include "alloc-util.h" |
| 11 |
#include "errno-util.h" |
| 12 |
#include "fd-util.h" |
| 13 |
#include "fs-util.h" |
| 14 |
#include "lock-util.h" |
| 15 |
#include "log.h" |
| 16 |
#include "path-util.h" |
| 17 |
#include "pidref.h" |
| 18 |
#include "process-util.h" |
| 19 |
#include "string-util.h" |
| 20 |
#include "time-util.h" |
| 21 |
|
| 22 |
int make_lock_file_at(int dir_fd, const char *p, int operation, LockFile *ret) { |
| 23 |
_cleanup_close_ int fd = -EBADF, dfd = -EBADF; |
| 24 |
_cleanup_free_ char *t = NULL; |
| 25 |
|
| 26 |
assert(dir_fd >= 0 || dir_fd == AT_FDCWD); |
| 27 |
assert(p); |
| 28 |
assert(IN_SET(operation & ~LOCK_NB, LOCK_EX, LOCK_SH)); |
| 29 |
assert(ret); |
| 30 |
|
| 31 |
if (isempty(p)) |
| 32 |
return -EINVAL; |
| 33 |
|
| 34 |
/* We use UNPOSIX locks as they have nice semantics, and are mostly compatible with NFS. */ |
| 35 |
|
| 36 |
dfd = fd_reopen(dir_fd, O_CLOEXEC|O_PATH|O_DIRECTORY); |
| 37 |
if (dfd < 0) |
| 38 |
return dfd; |
| 39 |
|
| 40 |
t = strdup(p); |
| 41 |
if (!t) |
| 42 |
return -ENOMEM; |
| 43 |
|
| 44 |
fd = xopenat_lock_full(dfd, |
| 45 |
p, |
| 46 |
O_CREAT|O_RDWR|O_NOFOLLOW|O_CLOEXEC|O_NOCTTY, |
| 47 |
/* xopen_flags= */ 0, |
| 48 |
0600, |
| 49 |
LOCK_UNPOSIX, |
| 50 |
operation); |
| 51 |
if (fd < 0) |
| 52 |
return fd == -EAGAIN ? -EBUSY : fd; |
| 53 |
|
| 54 |
*ret = (LockFile) { |
| 55 |
.dir_fd = TAKE_FD(dfd), |
| 56 |
.path = TAKE_PTR(t), |
| 57 |
.fd = TAKE_FD(fd), |
| 58 |
.operation = operation, |
| 59 |
}; |
| 60 |
|
| 61 |
return 0; |
| 62 |
} |
| 63 |
|
| 64 |
int make_lock_file_for(const char *p, int operation, LockFile *ret) { |
| 65 |
_cleanup_free_ char *fn = NULL, *dn = NULL, *t = NULL; |
| 66 |
int r; |
| 67 |
|
| 68 |
assert(p); |
| 69 |
assert(ret); |
| 70 |
|
| 71 |
r = path_split_prefix_filename(p, &dn, &fn); |
| 72 |
if (r < 0) |
| 73 |
return r; |
| 74 |
|
| 75 |
if (dn) |
| 76 |
t = strjoin(dn, "/.#", fn, ".lck"); |
| 77 |
else |
| 78 |
t = strjoin(".#", fn, ".lck"); |
| 79 |
if (!t) |
| 80 |
return -ENOMEM; |
| 81 |
|
| 82 |
return make_lock_file(t, operation, ret); |
| 83 |
} |
| 84 |
|
| 85 |
void release_lock_file(LockFile *f) { |
| 86 |
if (!f) |
| 87 |
return; |
| 88 |
|
| 89 |
if (f->path) { |
| 90 |
|
| 91 |
/* If we are the exclusive owner we can safely delete |
| 92 |
* the lock file itself. If we are not the exclusive |
| 93 |
* owner, we can try becoming it. */ |
| 94 |
|
| 95 |
if (f->fd >= 0 && |
| 96 |
(f->operation & ~LOCK_NB) == LOCK_SH && |
| 97 |
unposix_lock(f->fd, LOCK_EX|LOCK_NB) >= 0) |
| 98 |
f->operation = LOCK_EX|LOCK_NB; |
| 99 |
|
| 100 |
if ((f->operation & ~LOCK_NB) == LOCK_EX) |
| 101 |
(void) unlinkat(f->dir_fd, f->path, 0); |
| 102 |
|
| 103 |
f->path = mfree(f->path); |
| 104 |
} |
| 105 |
|
| 106 |
f->dir_fd = safe_close(f->dir_fd); |
| 107 |
f->fd = safe_close(f->fd); |
| 108 |
f->operation = 0; |
| 109 |
} |
| 110 |
|
| 111 |
static int fcntl_lock(int fd, int operation, bool ofd) { |
| 112 |
int cmd, type, r; |
| 113 |
|
| 114 |
assert(fd >= 0); |
| 115 |
|
| 116 |
if (ofd) |
| 117 |
cmd = (operation & LOCK_NB) ? F_OFD_SETLK : F_OFD_SETLKW; |
| 118 |
else |
| 119 |
cmd = (operation & LOCK_NB) ? F_SETLK : F_SETLKW; |
| 120 |
|
| 121 |
switch (operation & ~LOCK_NB) { |
| 122 |
case LOCK_EX: |
| 123 |
type = F_WRLCK; |
| 124 |
break; |
| 125 |
case LOCK_SH: |
| 126 |
type = F_RDLCK; |
| 127 |
break; |
| 128 |
case LOCK_UN: |
| 129 |
type = F_UNLCK; |
| 130 |
break; |
| 131 |
default: |
| 132 |
assert_not_reached(); |
| 133 |
} |
| 134 |
|
| 135 |
r = RET_NERRNO(fcntl(fd, cmd, &(struct flock) { |
| 136 |
.l_type = type, |
| 137 |
.l_whence = SEEK_SET, |
| 138 |
.l_start = 0, |
| 139 |
.l_len = 0, |
| 140 |
})); |
| 141 |
|
| 142 |
/* If we are doing non-blocking operations, treat EACCES/EAGAIN the same as per man page. But if |
| 143 |
* not, propagate EACCES back, as it will likely be due to an LSM denying the operation (for example |
| 144 |
* LXC with AppArmor when running on kernel < 6.2), and in some cases we want to gracefully |
| 145 |
* fallback (e.g.: PrivateNetwork=yes). As per documentation, it's only the non-blocking operation |
| 146 |
* F_SETLK that might return EACCES on some platforms (although the Linux implementation doesn't |
| 147 |
* seem to), as F_SETLKW and F_OFD_SETLKW block so this is not an issue, and F_OFD_SETLK is documented |
| 148 |
* to only return EAGAIN if the lock is already held. */ |
| 149 |
if ((operation & LOCK_NB) && r == -EACCES) |
| 150 |
r = -EAGAIN; |
| 151 |
|
| 152 |
return r; |
| 153 |
} |
| 154 |
|
| 155 |
int posix_lock(int fd, int operation) { |
| 156 |
return fcntl_lock(fd, operation, /* ofd= */ false); |
| 157 |
} |
| 158 |
|
| 159 |
int unposix_lock(int fd, int operation) { |
| 160 |
return fcntl_lock(fd, operation, /* ofd= */ true); |
| 161 |
} |
| 162 |
|
| 163 |
void posix_unlockpp(int **fd) { |
| 164 |
assert(fd); |
| 165 |
|
| 166 |
if (!*fd || **fd < 0) |
| 167 |
return; |
| 168 |
|
| 169 |
(void) fcntl_lock(**fd, LOCK_UN, /* ofd= */ false); |
| 170 |
*fd = NULL; |
| 171 |
} |
| 172 |
|
| 173 |
void unposix_unlockpp(int **fd) { |
| 174 |
assert(fd); |
| 175 |
|
| 176 |
if (!*fd || **fd < 0) |
| 177 |
return; |
| 178 |
|
| 179 |
(void) fcntl_lock(**fd, LOCK_UN, /* ofd= */ true); |
| 180 |
*fd = NULL; |
| 181 |
} |
| 182 |
|
| 183 |
int lock_generic(int fd, LockType type, int operation) { |
| 184 |
assert(fd >= 0); |
| 185 |
|
| 186 |
switch (type) { |
| 187 |
case LOCK_NONE: |
| 188 |
return 0; |
| 189 |
case LOCK_BSD: |
| 190 |
return RET_NERRNO(flock(fd, operation)); |
| 191 |
case LOCK_POSIX: |
| 192 |
return posix_lock(fd, operation); |
| 193 |
case LOCK_UNPOSIX: |
| 194 |
return unposix_lock(fd, operation); |
| 195 |
default: |
| 196 |
assert_not_reached(); |
| 197 |
} |
| 198 |
} |
| 199 |
|
| 200 |
int lock_generic_with_timeout(int fd, LockType type, int operation, usec_t timeout) { |
| 201 |
int r; |
| 202 |
|
| 203 |
assert(fd >= 0); |
| 204 |
|
| 205 |
/* A version of lock_generic(), but with a timeout. We do this in a child process, since the kernel |
| 206 |
* APIs natively don't support a timeout. We set a SIGALRM timer that will kill the child after the |
| 207 |
* timeout is hit. Returns -ETIMEDOUT if the timeout is hit, and 0 on success. |
| 208 |
* |
| 209 |
* This only works for BSD and UNPOSIX locks, as only those are fd-bound, and hence can be acquired |
| 210 |
* from any process that has access to the fd. POSIX locks OTOH are process-bound, and hence if we'd |
| 211 |
* acquire them in a child process they'd remain unlocked in the parent. */ |
| 212 |
|
| 213 |
if (type == LOCK_NONE) |
| 214 |
return 0; |
| 215 |
if (!IN_SET(type, LOCK_BSD, LOCK_UNPOSIX)) /* Not for POSIX locks, see above. */ |
| 216 |
return -EOPNOTSUPP; |
| 217 |
|
| 218 |
/* First, try without forking anything off */ |
| 219 |
r = lock_generic(fd, type, operation | (timeout == USEC_INFINITY ? 0 : LOCK_NB)); |
| 220 |
if (r != -EAGAIN || timeout == 0 || FLAGS_SET(operation, LOCK_NB)) |
| 221 |
return r; |
| 222 |
|
| 223 |
/* If that didn't work, try with a child */ |
| 224 |
_cleanup_(pidref_done_sigkill_wait) PidRef pidref = PIDREF_NULL; |
| 225 |
r = pidref_safe_fork("(sd-flock)", FORK_RESET_SIGNALS|FORK_DEATHSIG_SIGKILL, &pidref); |
| 226 |
if (r < 0) |
| 227 |
return log_error_errno(r, "Failed to flock block device in child process: %m"); |
| 228 |
if (r == 0) { |
| 229 |
if (lock_generic(fd, type, operation) < 0) { |
| 230 |
log_error_errno(errno, "Unable to get an exclusive lock on the device: %m"); |
| 231 |
_exit(EXIT_FAILURE); |
| 232 |
} |
| 233 |
|
| 234 |
_exit(EXIT_SUCCESS); |
| 235 |
} |
| 236 |
|
| 237 |
siginfo_t status; |
| 238 |
r = pidref_wait_for_terminate_full(&pidref, timeout, &status); |
| 239 |
if (r < 0) |
| 240 |
return r; |
| 241 |
|
| 242 |
pidref_done(&pidref); |
| 243 |
|
| 244 |
switch (status.si_code) { |
| 245 |
|
| 246 |
case CLD_EXITED: |
| 247 |
if (status.si_status != EXIT_SUCCESS) |
| 248 |
return -EPROTO; |
| 249 |
|
| 250 |
return 0; |
| 251 |
|
| 252 |
case CLD_KILLED: |
| 253 |
case CLD_DUMPED: |
| 254 |
return -EPROTO; |
| 255 |
|
| 256 |
default: |
| 257 |
assert_not_reached(); |
| 258 |
} |
| 259 |
} |