]> git.ipfire.org Git - thirdparty/systemd.git/blame - src/journal/journald-kmsg.c
tree-wide: use ASSERT_PTR more
[thirdparty/systemd.git] / src / journal / journald-kmsg.c
CommitLineData
db9ecf05 1/* SPDX-License-Identifier: LGPL-2.1-or-later */
ef63833d 2
ef63833d 3#include <fcntl.h>
4f5dd394 4#include <sys/epoll.h>
ef63833d 5#include <sys/mman.h>
4871690d 6#include <sys/socket.h>
4f5dd394 7#include <unistd.h>
ef63833d 8
8bebb09c 9#include "sd-device.h"
4f5dd394 10#include "sd-messages.h"
ef63833d 11
e6a7ec4b 12#include "alloc-util.h"
8bebb09c 13#include "device-util.h"
4f5dd394 14#include "escape.h"
3ffd4af2 15#include "fd-util.h"
f97b34a6 16#include "format-util.h"
4c54768c 17#include "fs-util.h"
afc5dbf3 18#include "io-util.h"
3ffd4af2 19#include "journald-kmsg.h"
d025f1e4 20#include "journald-server.h"
35e2e347 21#include "journald-syslog.h"
9c416180 22#include "log.h"
6bedfcbb 23#include "parse-util.h"
0b452006 24#include "process-util.h"
15a5e950 25#include "stdio-util.h"
07630cea 26#include "string-util.h"
ef63833d
LP
27
28void server_forward_kmsg(
dbac2625
LP
29 Server *s,
30 int priority,
31 const char *identifier,
32 const char *message,
33 const struct ucred *ucred) {
ef63833d 34
e6a7ec4b 35 _cleanup_free_ char *ident_buf = NULL;
ef63833d 36 struct iovec iovec[5];
3b97fcbd 37 char header_priority[DECIMAL_STR_MAX(priority) + 3],
fbd0b64f 38 header_pid[STRLEN("[]: ") + DECIMAL_STR_MAX(pid_t) + 1];
ef63833d 39 int n = 0;
ef63833d
LP
40
41 assert(s);
42 assert(priority >= 0);
43 assert(priority <= 999);
44 assert(message);
45
46 if (_unlikely_(LOG_PRI(priority) > s->max_level_kmsg))
47 return;
48
49 if (_unlikely_(s->dev_kmsg_fd < 0))
50 return;
51
52 /* Never allow messages with kernel facility to be written to
53 * kmsg, regardless where the data comes from. */
54 priority = syslog_fixup_facility(priority);
55
56 /* First: priority field */
5ffa8c81 57 xsprintf(header_priority, "<%i>", priority);
e6a7ec4b 58 iovec[n++] = IOVEC_MAKE_STRING(header_priority);
ef63833d
LP
59
60 /* Second: identifier and PID */
61 if (ucred) {
62 if (!identifier) {
74d6421d 63 (void) get_process_comm(ucred->pid, &ident_buf);
ef63833d
LP
64 identifier = ident_buf;
65 }
66
5ffa8c81 67 xsprintf(header_pid, "["PID_FMT"]: ", ucred->pid);
ef63833d
LP
68
69 if (identifier)
e6a7ec4b 70 iovec[n++] = IOVEC_MAKE_STRING(identifier);
ef63833d 71
e6a7ec4b 72 iovec[n++] = IOVEC_MAKE_STRING(header_pid);
ef63833d 73 } else if (identifier) {
e6a7ec4b
LP
74 iovec[n++] = IOVEC_MAKE_STRING(identifier);
75 iovec[n++] = IOVEC_MAKE_STRING(": ");
ef63833d
LP
76 }
77
78 /* Fourth: message */
e6a7ec4b
LP
79 iovec[n++] = IOVEC_MAKE_STRING(message);
80 iovec[n++] = IOVEC_MAKE_STRING("\n");
ef63833d
LP
81
82 if (writev(s->dev_kmsg_fd, iovec, n) < 0)
56f64d95 83 log_debug_errno(errno, "Failed to write to /dev/kmsg for logging: %m");
ef63833d
LP
84}
85
fe167298
AJ
86static bool is_us(const char *identifier, const char *pid) {
87 pid_t pid_num;
ef63833d 88
fe167298
AJ
89 if (!identifier || !pid)
90 return false;
ef63833d 91
fe167298 92 if (parse_pid(pid, &pid_num) < 0)
ef63833d
LP
93 return false;
94
fe167298
AJ
95 return pid_num == getpid_cached() &&
96 streq(identifier, program_invocation_short_name);
ef63833d
LP
97}
98
8857fb9b 99void dev_kmsg_record(Server *s, char *p, size_t l) {
d3070fbd 100
e6a7ec4b 101 _cleanup_free_ char *message = NULL, *syslog_priority = NULL, *syslog_pid = NULL, *syslog_facility = NULL, *syslog_identifier = NULL, *source_time = NULL, *identifier = NULL, *pid = NULL;
d3070fbd
LP
102 struct iovec iovec[N_IOVEC_META_FIELDS + 7 + N_IOVEC_KERNEL_FIELDS + 2 + N_IOVEC_UDEV_FIELDS];
103 char *kernel_device = NULL;
e9f600f2 104 unsigned long long usec;
d3070fbd
LP
105 size_t n = 0, z = 0, j;
106 int priority, r;
e6a7ec4b 107 char *e, *f, *k;
ef63833d
LP
108 uint64_t serial;
109 size_t pl;
9c416180
DDM
110 int saved_log_max_level = INT_MAX;
111 ClientContext *c = NULL;
ef63833d
LP
112
113 assert(s);
114 assert(p);
115
116 if (l <= 0)
117 return;
118
119 e = memchr(p, ',', l);
120 if (!e)
121 return;
122 *e = 0;
123
124 r = safe_atoi(p, &priority);
125 if (r < 0 || priority < 0 || priority > 999)
126 return;
127
b3600346 128 if (s->forward_to_kmsg && LOG_FAC(priority) != LOG_KERN)
ef63833d
LP
129 return;
130
131 l -= (e - p) + 1;
132 p = e + 1;
133 e = memchr(p, ',', l);
134 if (!e)
135 return;
136 *e = 0;
137
138 r = safe_atou64(p, &serial);
139 if (r < 0)
140 return;
141
142 if (s->kernel_seqnum) {
143 /* We already read this one? */
144 if (serial < *s->kernel_seqnum)
145 return;
146
147 /* Did we lose any? */
148 if (serial > *s->kernel_seqnum)
13181942 149 server_driver_message(s, 0,
2b044526 150 "MESSAGE_ID=" SD_MESSAGE_JOURNAL_MISSED_STR,
8a03c9ef
ZJS
151 LOG_MESSAGE("Missed %"PRIu64" kernel messages",
152 serial - *s->kernel_seqnum),
153 NULL);
ef63833d
LP
154
155 /* Make sure we never read this one again. Note that
156 * we always store the next message serial we expect
157 * here, simply because this makes handling the first
158 * message with serial 0 easy. */
159 *s->kernel_seqnum = serial + 1;
160 }
161
162 l -= (e - p) + 1;
163 p = e + 1;
164 f = memchr(p, ';', l);
165 if (!f)
166 return;
167 /* Kernel 3.6 has the flags field, kernel 3.5 lacks that */
168 e = memchr(p, ',', l);
169 if (!e || f < e)
170 e = f;
171 *e = 0;
172
e9f600f2 173 r = safe_atollu(p, &usec);
ef63833d
LP
174 if (r < 0)
175 return;
176
177 l -= (f - p) + 1;
178 p = f + 1;
179 e = memchr(p, '\n', l);
180 if (!e)
181 return;
182 *e = 0;
183
184 pl = e - p;
185 l -= (e - p) + 1;
186 k = e + 1;
187
188 for (j = 0; l > 0 && j < N_IOVEC_KERNEL_FIELDS; j++) {
189 char *m;
dc61b7e4 190 /* Metadata fields attached */
ef63833d
LP
191
192 if (*k != ' ')
193 break;
194
313cefa1 195 k++, l--;
ef63833d
LP
196
197 e = memchr(k, '\n', l);
198 if (!e)
30eddcd5 199 goto finish;
ef63833d
LP
200
201 *e = 0;
202
527b7a42 203 if (cunescape_length_with_prefix(k, e - k, "_KERNEL_", UNESCAPE_RELAX, &m) < 0)
ef63833d
LP
204 break;
205
206 if (startswith(m, "_KERNEL_DEVICE="))
207 kernel_device = m + 15;
208
e6a7ec4b 209 iovec[n++] = IOVEC_MAKE_STRING(m);
ef63833d
LP
210 z++;
211
212 l -= (e - k) + 1;
213 k = e + 1;
214 }
215
216 if (kernel_device) {
8bebb09c 217 _cleanup_(sd_device_unrefp) sd_device *d = NULL;
ef63833d 218
8bebb09c 219 if (sd_device_new_from_device_id(&d, kernel_device) >= 0) {
ef63833d 220 const char *g;
ef63833d
LP
221 char *b;
222
8bebb09c 223 if (sd_device_get_devname(d, &g) >= 0) {
b910cc72 224 b = strjoin("_UDEV_DEVNODE=", g);
ef63833d 225 if (b) {
e6a7ec4b 226 iovec[n++] = IOVEC_MAKE_STRING(b);
ef63833d
LP
227 z++;
228 }
229 }
230
8bebb09c 231 if (sd_device_get_sysname(d, &g) >= 0) {
b910cc72 232 b = strjoin("_UDEV_SYSNAME=", g);
ef63833d 233 if (b) {
e6a7ec4b 234 iovec[n++] = IOVEC_MAKE_STRING(b);
ef63833d
LP
235 z++;
236 }
237 }
238
239 j = 0;
8bebb09c 240 FOREACH_DEVICE_DEVLINK(d, g) {
ef63833d 241
080d112c 242 if (j >= N_IOVEC_UDEV_FIELDS)
ef63833d
LP
243 break;
244
b910cc72 245 b = strjoin("_UDEV_DEVLINK=", g);
8bebb09c
YW
246 if (b) {
247 iovec[n++] = IOVEC_MAKE_STRING(b);
248 z++;
ef63833d
LP
249 }
250
251 j++;
252 }
ef63833d
LP
253 }
254 }
255
e9f600f2 256 if (asprintf(&source_time, "_SOURCE_MONOTONIC_TIMESTAMP=%llu", usec) >= 0)
e6a7ec4b 257 iovec[n++] = IOVEC_MAKE_STRING(source_time);
ef63833d 258
e6a7ec4b 259 iovec[n++] = IOVEC_MAKE_STRING("_TRANSPORT=kernel");
ef63833d
LP
260
261 if (asprintf(&syslog_priority, "PRIORITY=%i", priority & LOG_PRIMASK) >= 0)
e6a7ec4b 262 iovec[n++] = IOVEC_MAKE_STRING(syslog_priority);
ef63833d 263
36dd072c 264 if (asprintf(&syslog_facility, "SYSLOG_FACILITY=%i", LOG_FAC(priority)) >= 0)
e6a7ec4b 265 iovec[n++] = IOVEC_MAKE_STRING(syslog_facility);
36dd072c 266
b3600346 267 if (LOG_FAC(priority) == LOG_KERN)
e6a7ec4b 268 iovec[n++] = IOVEC_MAKE_STRING("SYSLOG_IDENTIFIER=kernel");
ef63833d 269 else {
e88baee8 270 pl -= syslog_parse_identifier((const char**) &p, &identifier, &pid);
ef63833d 271
9c416180
DDM
272 /* Avoid logging any new messages when we're processing messages generated by ourselves via
273 * log_info() and friends to avoid infinite loops. */
274 if (is_us(identifier, pid)) {
275 if (!ratelimit_below(&s->kmsg_own_ratelimit))
276 return;
277
278 saved_log_max_level = log_get_max_level();
279 c = s->my_context;
280 log_set_max_level(LOG_NULL);
281 }
ef63833d
LP
282
283 if (identifier) {
b910cc72 284 syslog_identifier = strjoin("SYSLOG_IDENTIFIER=", identifier);
ef63833d 285 if (syslog_identifier)
e6a7ec4b 286 iovec[n++] = IOVEC_MAKE_STRING(syslog_identifier);
ef63833d
LP
287 }
288
289 if (pid) {
b910cc72 290 syslog_pid = strjoin("SYSLOG_PID=", pid);
ef63833d 291 if (syslog_pid)
e6a7ec4b 292 iovec[n++] = IOVEC_MAKE_STRING(syslog_pid);
ef63833d 293 }
ef63833d
LP
294 }
295
527b7a42 296 if (cunescape_length_with_prefix(p, pl, "MESSAGE=", UNESCAPE_RELAX, &message) >= 0)
e6a7ec4b 297 iovec[n++] = IOVEC_MAKE_STRING(message);
ef63833d 298
9c416180
DDM
299
300 server_dispatch_message(s, iovec, n, ELEMENTSOF(iovec), c, NULL, priority, 0);
301
302 if (saved_log_max_level != INT_MAX)
303 log_set_max_level(saved_log_max_level);
ef63833d
LP
304
305finish:
306 for (j = 0; j < z; j++)
307 free(iovec[j].iov_base);
ef63833d
LP
308}
309
f9a810be 310static int server_read_dev_kmsg(Server *s) {
ef63833d
LP
311 char buffer[8192+1]; /* the kernel-side limit per record is 8K currently */
312 ssize_t l;
313
314 assert(s);
315 assert(s->dev_kmsg_fd >= 0);
316
317 l = read(s->dev_kmsg_fd, buffer, sizeof(buffer) - 1);
318 if (l == 0)
319 return 0;
320 if (l < 0) {
321 /* Old kernels who don't allow reading from /dev/kmsg
322 * return EINVAL when we try. So handle this cleanly,
323 * but don' try to ever read from it again. */
324 if (errno == EINVAL) {
f9a810be 325 s->dev_kmsg_event_source = sd_event_source_unref(s->dev_kmsg_event_source);
ef63833d
LP
326 return 0;
327 }
328
8add30a0 329 if (ERRNO_IS_TRANSIENT(errno) || errno == EPIPE)
ef63833d
LP
330 return 0;
331
a53f024d 332 return log_error_errno(errno, "Failed to read from /dev/kmsg: %m");
ef63833d
LP
333 }
334
335 dev_kmsg_record(s, buffer, l);
336 return 1;
337}
338
339int server_flush_dev_kmsg(Server *s) {
340 int r;
341
342 assert(s);
343
344 if (s->dev_kmsg_fd < 0)
345 return 0;
346
347 if (!s->dev_kmsg_readable)
348 return 0;
349
2b43f939 350 log_debug("Flushing /dev/kmsg...");
ef63833d
LP
351
352 for (;;) {
353 r = server_read_dev_kmsg(s);
354 if (r < 0)
355 return r;
356
357 if (r == 0)
358 break;
359 }
360
361 return 0;
362}
363
f9a810be 364static int dispatch_dev_kmsg(sd_event_source *es, int fd, uint32_t revents, void *userdata) {
99534007 365 Server *s = ASSERT_PTR(userdata);
f9a810be
LP
366
367 assert(es);
368 assert(fd == s->dev_kmsg_fd);
f9a810be
LP
369
370 if (revents & EPOLLERR)
371 log_warning("/dev/kmsg buffer overrun, some messages lost.");
372
373 if (!(revents & EPOLLIN))
374 log_error("Got invalid event from epoll for /dev/kmsg: %"PRIx32, revents);
375
376 return server_read_dev_kmsg(s);
377}
378
ef63833d 379int server_open_dev_kmsg(Server *s) {
b2392ff3 380 mode_t mode;
f9a810be 381 int r;
ef63833d
LP
382
383 assert(s);
384
b2392ff3
SS
385 if (s->read_kmsg)
386 mode = O_RDWR|O_CLOEXEC|O_NONBLOCK|O_NOCTTY;
387 else
388 mode = O_WRONLY|O_CLOEXEC|O_NONBLOCK|O_NOCTTY;
389
390 s->dev_kmsg_fd = open("/dev/kmsg", mode);
ef63833d 391 if (s->dev_kmsg_fd < 0) {
cf5a2ee8
ZJS
392 log_full_errno(errno == ENOENT ? LOG_DEBUG : LOG_WARNING,
393 errno, "Failed to open /dev/kmsg, ignoring: %m");
ef63833d
LP
394 return 0;
395 }
396
b2392ff3
SS
397 if (!s->read_kmsg)
398 return 0;
399
151b9b96 400 r = sd_event_add_io(s->event, &s->dev_kmsg_event_source, s->dev_kmsg_fd, EPOLLIN, dispatch_dev_kmsg, s);
f9a810be 401 if (r < 0) {
ef63833d
LP
402
403 /* This will fail with EPERM on older kernels where
404 * /dev/kmsg is not readable. */
c0f71f46
LP
405 if (r == -EPERM) {
406 r = 0;
407 goto fail;
408 }
ef63833d 409
da927ba9 410 log_error_errno(r, "Failed to add /dev/kmsg fd to event loop: %m");
c0f71f46 411 goto fail;
f9a810be
LP
412 }
413
414 r = sd_event_source_set_priority(s->dev_kmsg_event_source, SD_EVENT_PRIORITY_IMPORTANT+10);
415 if (r < 0) {
da927ba9 416 log_error_errno(r, "Failed to adjust priority of kmsg event source: %m");
c0f71f46 417 goto fail;
ef63833d
LP
418 }
419
420 s->dev_kmsg_readable = true;
421
422 return 0;
c0f71f46
LP
423
424fail:
03e334a1
LP
425 s->dev_kmsg_event_source = sd_event_source_unref(s->dev_kmsg_event_source);
426 s->dev_kmsg_fd = safe_close(s->dev_kmsg_fd);
c0f71f46
LP
427
428 return r;
ef63833d
LP
429}
430
431int server_open_kernel_seqnum(Server *s) {
a2735a45 432 _cleanup_close_ int fd = -1;
b1852c48 433 const char *fn;
ef63833d 434 uint64_t *p;
7bb87460 435 int r;
ef63833d
LP
436
437 assert(s);
438
162392b7 439 /* We store the seqnum we last read in an mmapped file. That way we can just use it like a variable,
a2735a45
LP
440 * but it is persistent and automatically flushed at reboot. */
441
442 if (!s->read_kmsg)
443 return 0;
ef63833d 444
b1852c48
LP
445 fn = strjoina(s->runtime_directory, "/kernel-seqnum");
446 fd = open(fn, O_RDWR|O_CREAT|O_CLOEXEC|O_NOCTTY|O_NOFOLLOW, 0644);
ef63833d 447 if (fd < 0) {
b1852c48 448 log_error_errno(errno, "Failed to open %s, ignoring: %m", fn);
ef63833d
LP
449 return 0;
450 }
451
4c54768c
IZ
452 r = posix_fallocate_loop(fd, 0, sizeof(uint64_t));
453 if (r < 0) {
7bb87460 454 log_error_errno(r, "Failed to allocate sequential number file, ignoring: %m");
ef63833d
LP
455 return 0;
456 }
457
458 p = mmap(NULL, sizeof(uint64_t), PROT_READ|PROT_WRITE, MAP_SHARED, fd, 0);
459 if (p == MAP_FAILED) {
56f64d95 460 log_error_errno(errno, "Failed to map sequential number file, ignoring: %m");
ef63833d
LP
461 return 0;
462 }
463
ef63833d
LP
464 s->kernel_seqnum = p;
465
466 return 0;
467}