]> git.ipfire.org Git - thirdparty/systemd.git/blame - src/journal/journald-kmsg.c
tree-wide: drop license boilerplate
[thirdparty/systemd.git] / src / journal / journald-kmsg.c
CommitLineData
53e1b683 1/* SPDX-License-Identifier: LGPL-2.1+ */
ef63833d
LP
2/***
3 This file is part of systemd.
4
5 Copyright 2011 Lennart Poettering
ef63833d
LP
6***/
7
ef63833d 8#include <fcntl.h>
4f5dd394 9#include <sys/epoll.h>
ef63833d 10#include <sys/mman.h>
4871690d 11#include <sys/socket.h>
4f5dd394 12#include <unistd.h>
ef63833d 13
b4bbcaa9 14#include "libudev.h"
4f5dd394 15#include "sd-messages.h"
ef63833d 16
e6a7ec4b 17#include "alloc-util.h"
4f5dd394 18#include "escape.h"
3ffd4af2 19#include "fd-util.h"
f97b34a6 20#include "format-util.h"
afc5dbf3 21#include "io-util.h"
3ffd4af2 22#include "journald-kmsg.h"
d025f1e4 23#include "journald-server.h"
35e2e347 24#include "journald-syslog.h"
6bedfcbb 25#include "parse-util.h"
0b452006 26#include "process-util.h"
15a5e950 27#include "stdio-util.h"
07630cea 28#include "string-util.h"
ef63833d
LP
29
30void server_forward_kmsg(
31 Server *s,
32 int priority,
33 const char *identifier,
34 const char *message,
3b3154df 35 const struct ucred *ucred) {
ef63833d 36
e6a7ec4b 37 _cleanup_free_ char *ident_buf = NULL;
ef63833d 38 struct iovec iovec[5];
3b97fcbd 39 char header_priority[DECIMAL_STR_MAX(priority) + 3],
fbd0b64f 40 header_pid[STRLEN("[]: ") + DECIMAL_STR_MAX(pid_t) + 1];
ef63833d 41 int n = 0;
ef63833d
LP
42
43 assert(s);
44 assert(priority >= 0);
45 assert(priority <= 999);
46 assert(message);
47
48 if (_unlikely_(LOG_PRI(priority) > s->max_level_kmsg))
49 return;
50
51 if (_unlikely_(s->dev_kmsg_fd < 0))
52 return;
53
54 /* Never allow messages with kernel facility to be written to
55 * kmsg, regardless where the data comes from. */
56 priority = syslog_fixup_facility(priority);
57
58 /* First: priority field */
5ffa8c81 59 xsprintf(header_priority, "<%i>", priority);
e6a7ec4b 60 iovec[n++] = IOVEC_MAKE_STRING(header_priority);
ef63833d
LP
61
62 /* Second: identifier and PID */
63 if (ucred) {
64 if (!identifier) {
65 get_process_comm(ucred->pid, &ident_buf);
66 identifier = ident_buf;
67 }
68
5ffa8c81 69 xsprintf(header_pid, "["PID_FMT"]: ", ucred->pid);
ef63833d
LP
70
71 if (identifier)
e6a7ec4b 72 iovec[n++] = IOVEC_MAKE_STRING(identifier);
ef63833d 73
e6a7ec4b 74 iovec[n++] = IOVEC_MAKE_STRING(header_pid);
ef63833d 75 } else if (identifier) {
e6a7ec4b
LP
76 iovec[n++] = IOVEC_MAKE_STRING(identifier);
77 iovec[n++] = IOVEC_MAKE_STRING(": ");
ef63833d
LP
78 }
79
80 /* Fourth: message */
e6a7ec4b
LP
81 iovec[n++] = IOVEC_MAKE_STRING(message);
82 iovec[n++] = IOVEC_MAKE_STRING("\n");
ef63833d
LP
83
84 if (writev(s->dev_kmsg_fd, iovec, n) < 0)
56f64d95 85 log_debug_errno(errno, "Failed to write to /dev/kmsg for logging: %m");
ef63833d
LP
86}
87
fe167298
AJ
88static bool is_us(const char *identifier, const char *pid) {
89 pid_t pid_num;
ef63833d 90
fe167298
AJ
91 if (!identifier || !pid)
92 return false;
ef63833d 93
fe167298 94 if (parse_pid(pid, &pid_num) < 0)
ef63833d
LP
95 return false;
96
fe167298
AJ
97 return pid_num == getpid_cached() &&
98 streq(identifier, program_invocation_short_name);
ef63833d
LP
99}
100
3b3154df 101static void dev_kmsg_record(Server *s, const char *p, size_t l) {
d3070fbd 102
e6a7ec4b 103 _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
104 struct iovec iovec[N_IOVEC_META_FIELDS + 7 + N_IOVEC_KERNEL_FIELDS + 2 + N_IOVEC_UDEV_FIELDS];
105 char *kernel_device = NULL;
e9f600f2 106 unsigned long long usec;
d3070fbd
LP
107 size_t n = 0, z = 0, j;
108 int priority, r;
e6a7ec4b 109 char *e, *f, *k;
ef63833d
LP
110 uint64_t serial;
111 size_t pl;
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)
199 return;
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) {
217 struct udev_device *ud;
218
219 ud = udev_device_new_from_device_id(s->udev, kernel_device);
220 if (ud) {
221 const char *g;
222 struct udev_list_entry *ll;
223 char *b;
224
225 g = udev_device_get_devnode(ud);
226 if (g) {
227 b = strappend("_UDEV_DEVNODE=", g);
228 if (b) {
e6a7ec4b 229 iovec[n++] = IOVEC_MAKE_STRING(b);
ef63833d
LP
230 z++;
231 }
232 }
233
234 g = udev_device_get_sysname(ud);
235 if (g) {
236 b = strappend("_UDEV_SYSNAME=", g);
237 if (b) {
e6a7ec4b 238 iovec[n++] = IOVEC_MAKE_STRING(b);
ef63833d
LP
239 z++;
240 }
241 }
242
243 j = 0;
244 ll = udev_device_get_devlinks_list_entry(ud);
245 udev_list_entry_foreach(ll, ll) {
246
247 if (j > N_IOVEC_UDEV_FIELDS)
248 break;
249
250 g = udev_list_entry_get_name(ll);
ef63833d 251 if (g) {
4b94f3b8
ZJS
252 b = strappend("_UDEV_DEVLINK=", g);
253 if (b) {
e6a7ec4b 254 iovec[n++] = IOVEC_MAKE_STRING(b);
4b94f3b8
ZJS
255 z++;
256 }
ef63833d
LP
257 }
258
259 j++;
260 }
261
262 udev_device_unref(ud);
263 }
264 }
265
e9f600f2 266 if (asprintf(&source_time, "_SOURCE_MONOTONIC_TIMESTAMP=%llu", usec) >= 0)
e6a7ec4b 267 iovec[n++] = IOVEC_MAKE_STRING(source_time);
ef63833d 268
e6a7ec4b 269 iovec[n++] = IOVEC_MAKE_STRING("_TRANSPORT=kernel");
ef63833d
LP
270
271 if (asprintf(&syslog_priority, "PRIORITY=%i", priority & LOG_PRIMASK) >= 0)
e6a7ec4b 272 iovec[n++] = IOVEC_MAKE_STRING(syslog_priority);
ef63833d 273
36dd072c 274 if (asprintf(&syslog_facility, "SYSLOG_FACILITY=%i", LOG_FAC(priority)) >= 0)
e6a7ec4b 275 iovec[n++] = IOVEC_MAKE_STRING(syslog_facility);
36dd072c 276
b3600346 277 if (LOG_FAC(priority) == LOG_KERN)
e6a7ec4b 278 iovec[n++] = IOVEC_MAKE_STRING("SYSLOG_IDENTIFIER=kernel");
ef63833d 279 else {
e88baee8 280 pl -= syslog_parse_identifier((const char**) &p, &identifier, &pid);
ef63833d
LP
281
282 /* Avoid any messages we generated ourselves via
283 * log_info() and friends. */
fe167298 284 if (is_us(identifier, pid))
ef63833d
LP
285 goto finish;
286
287 if (identifier) {
288 syslog_identifier = strappend("SYSLOG_IDENTIFIER=", identifier);
289 if (syslog_identifier)
e6a7ec4b 290 iovec[n++] = IOVEC_MAKE_STRING(syslog_identifier);
ef63833d
LP
291 }
292
293 if (pid) {
294 syslog_pid = strappend("SYSLOG_PID=", pid);
295 if (syslog_pid)
e6a7ec4b 296 iovec[n++] = IOVEC_MAKE_STRING(syslog_pid);
ef63833d 297 }
ef63833d
LP
298 }
299
527b7a42 300 if (cunescape_length_with_prefix(p, pl, "MESSAGE=", UNESCAPE_RELAX, &message) >= 0)
e6a7ec4b 301 iovec[n++] = IOVEC_MAKE_STRING(message);
ef63833d 302
22e3a02b 303 server_dispatch_message(s, iovec, n, ELEMENTSOF(iovec), NULL, NULL, priority, 0);
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
3742095b 329 if (IN_SET(errno, EAGAIN, EINTR, EPIPE))
ef63833d
LP
330 return 0;
331
e1427b13 332 return log_error_errno(errno, "Failed to read from kernel: %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
LP
364static int dispatch_dev_kmsg(sd_event_source *es, int fd, uint32_t revents, void *userdata) {
365 Server *s = userdata;
366
367 assert(es);
368 assert(fd == s->dev_kmsg_fd);
369 assert(s);
370
371 if (revents & EPOLLERR)
372 log_warning("/dev/kmsg buffer overrun, some messages lost.");
373
374 if (!(revents & EPOLLIN))
375 log_error("Got invalid event from epoll for /dev/kmsg: %"PRIx32, revents);
376
377 return server_read_dev_kmsg(s);
378}
379
ef63833d 380int server_open_dev_kmsg(Server *s) {
b2392ff3 381 mode_t mode;
f9a810be 382 int r;
ef63833d
LP
383
384 assert(s);
385
b2392ff3
SS
386 if (s->read_kmsg)
387 mode = O_RDWR|O_CLOEXEC|O_NONBLOCK|O_NOCTTY;
388 else
389 mode = O_WRONLY|O_CLOEXEC|O_NONBLOCK|O_NOCTTY;
390
391 s->dev_kmsg_fd = open("/dev/kmsg", mode);
ef63833d 392 if (s->dev_kmsg_fd < 0) {
445ea9be
LP
393 log_full(errno == ENOENT ? LOG_DEBUG : LOG_WARNING,
394 "Failed to open /dev/kmsg, ignoring: %m");
ef63833d
LP
395 return 0;
396 }
397
b2392ff3
SS
398 if (!s->read_kmsg)
399 return 0;
400
151b9b96 401 r = sd_event_add_io(s->event, &s->dev_kmsg_event_source, s->dev_kmsg_fd, EPOLLIN, dispatch_dev_kmsg, s);
f9a810be 402 if (r < 0) {
ef63833d
LP
403
404 /* This will fail with EPERM on older kernels where
405 * /dev/kmsg is not readable. */
c0f71f46
LP
406 if (r == -EPERM) {
407 r = 0;
408 goto fail;
409 }
ef63833d 410
da927ba9 411 log_error_errno(r, "Failed to add /dev/kmsg fd to event loop: %m");
c0f71f46 412 goto fail;
f9a810be
LP
413 }
414
415 r = sd_event_source_set_priority(s->dev_kmsg_event_source, SD_EVENT_PRIORITY_IMPORTANT+10);
416 if (r < 0) {
da927ba9 417 log_error_errno(r, "Failed to adjust priority of kmsg event source: %m");
c0f71f46 418 goto fail;
ef63833d
LP
419 }
420
421 s->dev_kmsg_readable = true;
422
423 return 0;
c0f71f46
LP
424
425fail:
03e334a1
LP
426 s->dev_kmsg_event_source = sd_event_source_unref(s->dev_kmsg_event_source);
427 s->dev_kmsg_fd = safe_close(s->dev_kmsg_fd);
c0f71f46
LP
428
429 return r;
ef63833d
LP
430}
431
432int server_open_kernel_seqnum(Server *s) {
03e334a1 433 _cleanup_close_ int fd;
ef63833d 434 uint64_t *p;
7bb87460 435 int r;
ef63833d
LP
436
437 assert(s);
438
439 /* We store the seqnum we last read in an mmaped file. That
440 * way we can just use it like a variable, but it is
b2e6df73 441 * persistent and automatically flushed at reboot. */
ef63833d
LP
442
443 fd = open("/run/systemd/journal/kernel-seqnum", O_RDWR|O_CREAT|O_CLOEXEC|O_NOCTTY|O_NOFOLLOW, 0644);
444 if (fd < 0) {
56f64d95 445 log_error_errno(errno, "Failed to open /run/systemd/journal/kernel-seqnum, ignoring: %m");
ef63833d
LP
446 return 0;
447 }
448
7bb87460
MS
449 r = posix_fallocate(fd, 0, sizeof(uint64_t));
450 if (r != 0) {
451 log_error_errno(r, "Failed to allocate sequential number file, ignoring: %m");
ef63833d
LP
452 return 0;
453 }
454
455 p = mmap(NULL, sizeof(uint64_t), PROT_READ|PROT_WRITE, MAP_SHARED, fd, 0);
456 if (p == MAP_FAILED) {
56f64d95 457 log_error_errno(errno, "Failed to map sequential number file, ignoring: %m");
ef63833d
LP
458 return 0;
459 }
460
ef63833d
LP
461 s->kernel_seqnum = p;
462
463 return 0;
464}