src/journal/journald-kmsg.c

   1 /***
   2   This file is part of systemd.
   3
   4   Copyright 2011 Lennart Poettering
   5
   6   systemd is free software; you can redistribute it and/or modify it
   7   under the terms of the GNU Lesser General Public License as published by
   8   the Free Software Foundation; either version 2.1 of the License, or
   9   (at your option) any later version.
  10
  11   systemd is distributed in the hope that it will be useful, but
  12   WITHOUT ANY WARRANTY; without even the implied warranty of
  13   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
  14   Lesser General Public License for more details.
  15
  16   You should have received a copy of the GNU Lesser General Public License
  17   along with systemd; If not, see <http://www.gnu.org/licenses/>.
  18 ***/
  19
  20 #include <fcntl.h>
  21 #include <sys/epoll.h>
  22 #include <sys/mman.h>
  23 #include <sys/socket.h>
  24 #include <unistd.h>
  25
  26 #include "libudev.h"
  27 #include "sd-messages.h"
  28
  29 #include "alloc-util.h"
  30 #include "escape.h"
  31 #include "fd-util.h"
  32 #include "format-util.h"
  33 #include "io-util.h"
  34 #include "journald-kmsg.h"
  35 #include "journald-server.h"
  36 #include "journald-syslog.h"
  37 #include "parse-util.h"
  38 #include "process-util.h"
  39 #include "stdio-util.h"
  40 #include "string-util.h"
  41
  42 void server_forward_kmsg(
  43         Server *s,
  44         int priority,
  45         const char *identifier,
  46         const char *message,
  47         const struct ucred *ucred) {
  48
  49         _cleanup_free_ char *ident_buf = NULL;
  50         struct iovec iovec[5];
  51         char header_priority[DECIMAL_STR_MAX(priority) + 3],
  52              header_pid[sizeof("[]: ")-1 + DECIMAL_STR_MAX(pid_t) + 1];
  53         int n = 0;
  54
  55         assert(s);
  56         assert(priority >= 0);
  57         assert(priority <= 999);
  58         assert(message);
  59
  60         if (_unlikely_(LOG_PRI(priority) > s->max_level_kmsg))
  61                 return;
  62
  63         if (_unlikely_(s->dev_kmsg_fd < 0))
  64                 return;
  65
  66         /* Never allow messages with kernel facility to be written to
  67          * kmsg, regardless where the data comes from. */
  68         priority = syslog_fixup_facility(priority);
  69
  70         /* First: priority field */
  71         xsprintf(header_priority, "<%i>", priority);
  72         iovec[n++] = IOVEC_MAKE_STRING(header_priority);
  73
  74         /* Second: identifier and PID */
  75         if (ucred) {
  76                 if (!identifier) {
  77                         get_process_comm(ucred->pid, &ident_buf);
  78                         identifier = ident_buf;
  79                 }
  80
  81                 xsprintf(header_pid, "["PID_FMT"]: ", ucred->pid);
  82
  83                 if (identifier)
  84                         iovec[n++] = IOVEC_MAKE_STRING(identifier);
  85
  86                 iovec[n++] = IOVEC_MAKE_STRING(header_pid);
  87         } else if (identifier) {
  88                 iovec[n++] = IOVEC_MAKE_STRING(identifier);
  89                 iovec[n++] = IOVEC_MAKE_STRING(": ");
  90         }
  91
  92         /* Fourth: message */
  93         iovec[n++] = IOVEC_MAKE_STRING(message);
  94         iovec[n++] = IOVEC_MAKE_STRING("\n");
  95
  96         if (writev(s->dev_kmsg_fd, iovec, n) < 0)
  97                 log_debug_errno(errno, "Failed to write to /dev/kmsg for logging: %m");
  98 }
  99
 100 static bool is_us(const char *pid) {
 101         pid_t t;
 102
 103         assert(pid);
 104
 105         if (parse_pid(pid, &t) < 0)
 106                 return false;
 107
 108         return t == getpid_cached();
 109 }
 110
 111 static void dev_kmsg_record(Server *s, const char *p, size_t l) {
 112         struct iovec iovec[N_IOVEC_META_FIELDS + 7 + N_IOVEC_KERNEL_FIELDS + 2 + N_IOVEC_UDEV_FIELDS];
 113         _cleanup_free_ char *message = NULL, *syslog_priority = NULL, *syslog_pid = NULL, *syslog_facility = NULL, *syslog_identifier = NULL, *source_time = NULL, *identifier = NULL, *pid = NULL;
 114         int priority, r;
 115         unsigned n = 0, z = 0, j;
 116         unsigned long long usec;
 117         char *e, *f, *k;
 118         uint64_t serial;
 119         size_t pl;
 120         char *kernel_device = NULL;
 121
 122         assert(s);
 123         assert(p);
 124
 125         if (l <= 0)
 126                 return;
 127
 128         e = memchr(p, ',', l);
 129         if (!e)
 130                 return;
 131         *e = 0;
 132
 133         r = safe_atoi(p, &priority);
 134         if (r < 0 || priority < 0 || priority > 999)
 135                 return;
 136
 137         if (s->forward_to_kmsg && (priority & LOG_FACMASK) != LOG_KERN)
 138                 return;
 139
 140         l -= (e - p) + 1;
 141         p = e + 1;
 142         e = memchr(p, ',', l);
 143         if (!e)
 144                 return;
 145         *e = 0;
 146
 147         r = safe_atou64(p, &serial);
 148         if (r < 0)
 149                 return;
 150
 151         if (s->kernel_seqnum) {
 152                 /* We already read this one? */
 153                 if (serial < *s->kernel_seqnum)
 154                         return;
 155
 156                 /* Did we lose any? */
 157                 if (serial > *s->kernel_seqnum)
 158                         server_driver_message(s,
 159                                               "MESSAGE_ID=" SD_MESSAGE_JOURNAL_MISSED_STR,
 160                                               LOG_MESSAGE("Missed %"PRIu64" kernel messages",
 161                                                           serial - *s->kernel_seqnum),
 162                                               NULL);
 163
 164                 /* Make sure we never read this one again. Note that
 165                  * we always store the next message serial we expect
 166                  * here, simply because this makes handling the first
 167                  * message with serial 0 easy. */
 168                 *s->kernel_seqnum = serial + 1;
 169         }
 170
 171         l -= (e - p) + 1;
 172         p = e + 1;
 173         f = memchr(p, ';', l);
 174         if (!f)
 175                 return;
 176         /* Kernel 3.6 has the flags field, kernel 3.5 lacks that */
 177         e = memchr(p, ',', l);
 178         if (!e || f < e)
 179                 e = f;
 180         *e = 0;
 181
 182         r = safe_atollu(p, &usec);
 183         if (r < 0)
 184                 return;
 185
 186         l -= (f - p) + 1;
 187         p = f + 1;
 188         e = memchr(p, '\n', l);
 189         if (!e)
 190                 return;
 191         *e = 0;
 192
 193         pl = e - p;
 194         l -= (e - p) + 1;
 195         k = e + 1;
 196
 197         for (j = 0; l > 0 && j < N_IOVEC_KERNEL_FIELDS; j++) {
 198                 char *m;
 199                 /* Metadata fields attached */
 200
 201                 if (*k != ' ')
 202                         break;
 203
 204                 k++, l--;
 205
 206                 e = memchr(k, '\n', l);
 207                 if (!e)
 208                         return;
 209
 210                 *e = 0;
 211
 212                 if (cunescape_length_with_prefix(k, e - k, "_KERNEL_", UNESCAPE_RELAX, &m) < 0)
 213                         break;
 214
 215                 if (startswith(m, "_KERNEL_DEVICE="))
 216                         kernel_device = m + 15;
 217
 218                 iovec[n++] = IOVEC_MAKE_STRING(m);
 219                 z++;
 220
 221                 l -= (e - k) + 1;
 222                 k = e + 1;
 223         }
 224
 225         if (kernel_device) {
 226                 struct udev_device *ud;
 227
 228                 ud = udev_device_new_from_device_id(s->udev, kernel_device);
 229                 if (ud) {
 230                         const char *g;
 231                         struct udev_list_entry *ll;
 232                         char *b;
 233
 234                         g = udev_device_get_devnode(ud);
 235                         if (g) {
 236                                 b = strappend("_UDEV_DEVNODE=", g);
 237                                 if (b) {
 238                                         iovec[n++] = IOVEC_MAKE_STRING(b);
 239                                         z++;
 240                                 }
 241                         }
 242
 243                         g = udev_device_get_sysname(ud);
 244                         if (g) {
 245                                 b = strappend("_UDEV_SYSNAME=", g);
 246                                 if (b) {
 247                                         iovec[n++] = IOVEC_MAKE_STRING(b);
 248                                         z++;
 249                                 }
 250                         }
 251
 252                         j = 0;
 253                         ll = udev_device_get_devlinks_list_entry(ud);
 254                         udev_list_entry_foreach(ll, ll) {
 255
 256                                 if (j > N_IOVEC_UDEV_FIELDS)
 257                                         break;
 258
 259                                 g = udev_list_entry_get_name(ll);
 260                                 if (g) {
 261                                         b = strappend("_UDEV_DEVLINK=", g);
 262                                         if (b) {
 263                                                 iovec[n++] = IOVEC_MAKE_STRING(b);
 264                                                 z++;
 265                                         }
 266                                 }
 267
 268                                 j++;
 269                         }
 270
 271                         udev_device_unref(ud);
 272                 }
 273         }
 274
 275         if (asprintf(&source_time, "_SOURCE_MONOTONIC_TIMESTAMP=%llu", usec) >= 0)
 276                 iovec[n++] = IOVEC_MAKE_STRING(source_time);
 277
 278         iovec[n++] = IOVEC_MAKE_STRING("_TRANSPORT=kernel");
 279
 280         if (asprintf(&syslog_priority, "PRIORITY=%i", priority & LOG_PRIMASK) >= 0)
 281                 iovec[n++] = IOVEC_MAKE_STRING(syslog_priority);
 282
 283         if (asprintf(&syslog_facility, "SYSLOG_FACILITY=%i", LOG_FAC(priority)) >= 0)
 284                 iovec[n++] = IOVEC_MAKE_STRING(syslog_facility);
 285
 286         if ((priority & LOG_FACMASK) == LOG_KERN)
 287                 iovec[n++] = IOVEC_MAKE_STRING("SYSLOG_IDENTIFIER=kernel");
 288         else {
 289                 pl -= syslog_parse_identifier((const char**) &p, &identifier, &pid);
 290
 291                 /* Avoid any messages we generated ourselves via
 292                  * log_info() and friends. */
 293                 if (pid && is_us(pid))
 294                         goto finish;
 295
 296                 if (identifier) {
 297                         syslog_identifier = strappend("SYSLOG_IDENTIFIER=", identifier);
 298                         if (syslog_identifier)
 299                                 iovec[n++] = IOVEC_MAKE_STRING(syslog_identifier);
 300                 }
 301
 302                 if (pid) {
 303                         syslog_pid = strappend("SYSLOG_PID=", pid);
 304                         if (syslog_pid)
 305                                 iovec[n++] = IOVEC_MAKE_STRING(syslog_pid);
 306                 }
 307         }
 308
 309         if (cunescape_length_with_prefix(p, pl, "MESSAGE=", UNESCAPE_RELAX, &message) >= 0)
 310                 iovec[n++] = IOVEC_MAKE_STRING(message);
 311
 312         server_dispatch_message(s, iovec, n, ELEMENTSOF(iovec), NULL, NULL, priority, 0);
 313
 314 finish:
 315         for (j = 0; j < z; j++)
 316                 free(iovec[j].iov_base);
 317 }
 318
 319 static int server_read_dev_kmsg(Server *s) {
 320         char buffer[8192+1]; /* the kernel-side limit per record is 8K currently */
 321         ssize_t l;
 322
 323         assert(s);
 324         assert(s->dev_kmsg_fd >= 0);
 325
 326         l = read(s->dev_kmsg_fd, buffer, sizeof(buffer) - 1);
 327         if (l == 0)
 328                 return 0;
 329         if (l < 0) {
 330                 /* Old kernels who don't allow reading from /dev/kmsg
 331                  * return EINVAL when we try. So handle this cleanly,
 332                  * but don' try to ever read from it again. */
 333                 if (errno == EINVAL) {
 334                         s->dev_kmsg_event_source = sd_event_source_unref(s->dev_kmsg_event_source);
 335                         return 0;
 336                 }
 337
 338                 if (IN_SET(errno, EAGAIN, EINTR, EPIPE))
 339                         return 0;
 340
 341                 return log_error_errno(errno, "Failed to read from kernel: %m");
 342         }
 343
 344         dev_kmsg_record(s, buffer, l);
 345         return 1;
 346 }
 347
 348 int server_flush_dev_kmsg(Server *s) {
 349         int r;
 350
 351         assert(s);
 352
 353         if (s->dev_kmsg_fd < 0)
 354                 return 0;
 355
 356         if (!s->dev_kmsg_readable)
 357                 return 0;
 358
 359         log_debug("Flushing /dev/kmsg...");
 360
 361         for (;;) {
 362                 r = server_read_dev_kmsg(s);
 363                 if (r < 0)
 364                         return r;
 365
 366                 if (r == 0)
 367                         break;
 368         }
 369
 370         return 0;
 371 }
 372
 373 static int dispatch_dev_kmsg(sd_event_source *es, int fd, uint32_t revents, void *userdata) {
 374         Server *s = userdata;
 375
 376         assert(es);
 377         assert(fd == s->dev_kmsg_fd);
 378         assert(s);
 379
 380         if (revents & EPOLLERR)
 381                 log_warning("/dev/kmsg buffer overrun, some messages lost.");
 382
 383         if (!(revents & EPOLLIN))
 384                 log_error("Got invalid event from epoll for /dev/kmsg: %"PRIx32, revents);
 385
 386         return server_read_dev_kmsg(s);
 387 }
 388
 389 int server_open_dev_kmsg(Server *s) {
 390         mode_t mode;
 391         int r;
 392
 393         assert(s);
 394
 395         if (s->read_kmsg)
 396                 mode = O_RDWR|O_CLOEXEC|O_NONBLOCK|O_NOCTTY;
 397         else
 398                 mode = O_WRONLY|O_CLOEXEC|O_NONBLOCK|O_NOCTTY;
 399
 400         s->dev_kmsg_fd = open("/dev/kmsg", mode);
 401         if (s->dev_kmsg_fd < 0) {
 402                 log_full(errno == ENOENT ? LOG_DEBUG : LOG_WARNING,
 403                          "Failed to open /dev/kmsg, ignoring: %m");
 404                 return 0;
 405         }
 406
 407         if (!s->read_kmsg)
 408                 return 0;
 409
 410         r = sd_event_add_io(s->event, &s->dev_kmsg_event_source, s->dev_kmsg_fd, EPOLLIN, dispatch_dev_kmsg, s);
 411         if (r < 0) {
 412
 413                 /* This will fail with EPERM on older kernels where
 414                  * /dev/kmsg is not readable. */
 415                 if (r == -EPERM) {
 416                         r = 0;
 417                         goto fail;
 418                 }
 419
 420                 log_error_errno(r, "Failed to add /dev/kmsg fd to event loop: %m");
 421                 goto fail;
 422         }
 423
 424         r = sd_event_source_set_priority(s->dev_kmsg_event_source, SD_EVENT_PRIORITY_IMPORTANT+10);
 425         if (r < 0) {
 426                 log_error_errno(r, "Failed to adjust priority of kmsg event source: %m");
 427                 goto fail;
 428         }
 429
 430         s->dev_kmsg_readable = true;
 431
 432         return 0;
 433
 434 fail:
 435         s->dev_kmsg_event_source = sd_event_source_unref(s->dev_kmsg_event_source);
 436         s->dev_kmsg_fd = safe_close(s->dev_kmsg_fd);
 437
 438         return r;
 439 }
 440
 441 int server_open_kernel_seqnum(Server *s) {
 442         _cleanup_close_ int fd;
 443         uint64_t *p;
 444         int r;
 445
 446         assert(s);
 447
 448         /* We store the seqnum we last read in an mmaped file. That
 449          * way we can just use it like a variable, but it is
 450          * persistent and automatically flushed at reboot. */
 451
 452         fd = open("/run/systemd/journal/kernel-seqnum", O_RDWR|O_CREAT|O_CLOEXEC|O_NOCTTY|O_NOFOLLOW, 0644);
 453         if (fd < 0) {
 454                 log_error_errno(errno, "Failed to open /run/systemd/journal/kernel-seqnum, ignoring: %m");
 455                 return 0;
 456         }
 457
 458         r = posix_fallocate(fd, 0, sizeof(uint64_t));
 459         if (r != 0) {
 460                 log_error_errno(r, "Failed to allocate sequential number file, ignoring: %m");
 461                 return 0;
 462         }
 463
 464         p = mmap(NULL, sizeof(uint64_t), PROT_READ|PROT_WRITE, MAP_SHARED, fd, 0);
 465         if (p == MAP_FAILED) {
 466                 log_error_errno(errno, "Failed to map sequential number file, ignoring: %m");
 467                 return 0;
 468         }
 469
 470         s->kernel_seqnum = p;
 471
 472         return 0;
 473 }