src/journal/journald-kmsg.c

   1 /* SPDX-License-Identifier: LGPL-2.1+ */
   2 /***
   3   This file is part of systemd.
   4
   5   Copyright 2011 Lennart Poettering
   6 ***/
   7
   8 #include <fcntl.h>
   9 #include <sys/epoll.h>
  10 #include <sys/mman.h>
  11 #include <sys/socket.h>
  12 #include <unistd.h>
  13
  14 #include "libudev.h"
  15 #include "sd-messages.h"
  16
  17 #include "alloc-util.h"
  18 #include "escape.h"
  19 #include "fd-util.h"
  20 #include "format-util.h"
  21 #include "io-util.h"
  22 #include "journald-kmsg.h"
  23 #include "journald-server.h"
  24 #include "journald-syslog.h"
  25 #include "parse-util.h"
  26 #include "process-util.h"
  27 #include "stdio-util.h"
  28 #include "string-util.h"
  29
  30 void server_forward_kmsg(
  31         Server *s,
  32         int priority,
  33         const char *identifier,
  34         const char *message,
  35         const struct ucred *ucred) {
  36
  37         _cleanup_free_ char *ident_buf = NULL;
  38         struct iovec iovec[5];
  39         char header_priority[DECIMAL_STR_MAX(priority) + 3],
  40              header_pid[STRLEN("[]: ") + DECIMAL_STR_MAX(pid_t) + 1];
  41         int n = 0;
  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 */
  59         xsprintf(header_priority, "<%i>", priority);
  60         iovec[n++] = IOVEC_MAKE_STRING(header_priority);
  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
  69                 xsprintf(header_pid, "["PID_FMT"]: ", ucred->pid);
  70
  71                 if (identifier)
  72                         iovec[n++] = IOVEC_MAKE_STRING(identifier);
  73
  74                 iovec[n++] = IOVEC_MAKE_STRING(header_pid);
  75         } else if (identifier) {
  76                 iovec[n++] = IOVEC_MAKE_STRING(identifier);
  77                 iovec[n++] = IOVEC_MAKE_STRING(": ");
  78         }
  79
  80         /* Fourth: message */
  81         iovec[n++] = IOVEC_MAKE_STRING(message);
  82         iovec[n++] = IOVEC_MAKE_STRING("\n");
  83
  84         if (writev(s->dev_kmsg_fd, iovec, n) < 0)
  85                 log_debug_errno(errno, "Failed to write to /dev/kmsg for logging: %m");
  86 }
  87
  88 static bool is_us(const char *identifier, const char *pid) {
  89         pid_t pid_num;
  90
  91         if (!identifier || !pid)
  92                 return false;
  93
  94         if (parse_pid(pid, &pid_num) < 0)
  95                 return false;
  96
  97         return pid_num == getpid_cached() &&
  98                streq(identifier, program_invocation_short_name);
  99 }
 100
 101 static void dev_kmsg_record(Server *s, const char *p, size_t l) {
 102
 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;
 104         struct iovec iovec[N_IOVEC_META_FIELDS + 7 + N_IOVEC_KERNEL_FIELDS + 2 + N_IOVEC_UDEV_FIELDS];
 105         char *kernel_device = NULL;
 106         unsigned long long usec;
 107         size_t n = 0, z = 0, j;
 108         int priority, r;
 109         char *e, *f, *k;
 110         uint64_t serial;
 111         size_t pl;
 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
 128         if (s->forward_to_kmsg && LOG_FAC(priority) != LOG_KERN)
 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)
 149                         server_driver_message(s, 0,
 150                                               "MESSAGE_ID=" SD_MESSAGE_JOURNAL_MISSED_STR,
 151                                               LOG_MESSAGE("Missed %"PRIu64" kernel messages",
 152                                                           serial - *s->kernel_seqnum),
 153                                               NULL);
 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
 173         r = safe_atollu(p, &usec);
 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;
 190                 /* Metadata fields attached */
 191
 192                 if (*k != ' ')
 193                         break;
 194
 195                 k++, l--;
 196
 197                 e = memchr(k, '\n', l);
 198                 if (!e)
 199                         return;
 200
 201                 *e = 0;
 202
 203                 if (cunescape_length_with_prefix(k, e - k, "_KERNEL_", UNESCAPE_RELAX, &m) < 0)
 204                         break;
 205
 206                 if (startswith(m, "_KERNEL_DEVICE="))
 207                         kernel_device = m + 15;
 208
 209                 iovec[n++] = IOVEC_MAKE_STRING(m);
 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) {
 229                                         iovec[n++] = IOVEC_MAKE_STRING(b);
 230                                         z++;
 231                                 }
 232                         }
 233
 234                         g = udev_device_get_sysname(ud);
 235                         if (g) {
 236                                 b = strappend("_UDEV_SYSNAME=", g);
 237                                 if (b) {
 238                                         iovec[n++] = IOVEC_MAKE_STRING(b);
 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);
 251                                 if (g) {
 252                                         b = strappend("_UDEV_DEVLINK=", g);
 253                                         if (b) {
 254                                                 iovec[n++] = IOVEC_MAKE_STRING(b);
 255                                                 z++;
 256                                         }
 257                                 }
 258
 259                                 j++;
 260                         }
 261
 262                         udev_device_unref(ud);
 263                 }
 264         }
 265
 266         if (asprintf(&source_time, "_SOURCE_MONOTONIC_TIMESTAMP=%llu", usec) >= 0)
 267                 iovec[n++] = IOVEC_MAKE_STRING(source_time);
 268
 269         iovec[n++] = IOVEC_MAKE_STRING("_TRANSPORT=kernel");
 270
 271         if (asprintf(&syslog_priority, "PRIORITY=%i", priority & LOG_PRIMASK) >= 0)
 272                 iovec[n++] = IOVEC_MAKE_STRING(syslog_priority);
 273
 274         if (asprintf(&syslog_facility, "SYSLOG_FACILITY=%i", LOG_FAC(priority)) >= 0)
 275                 iovec[n++] = IOVEC_MAKE_STRING(syslog_facility);
 276
 277         if (LOG_FAC(priority) == LOG_KERN)
 278                 iovec[n++] = IOVEC_MAKE_STRING("SYSLOG_IDENTIFIER=kernel");
 279         else {
 280                 pl -= syslog_parse_identifier((const char**) &p, &identifier, &pid);
 281
 282                 /* Avoid any messages we generated ourselves via
 283                  * log_info() and friends. */
 284                 if (is_us(identifier, pid))
 285                         goto finish;
 286
 287                 if (identifier) {
 288                         syslog_identifier = strappend("SYSLOG_IDENTIFIER=", identifier);
 289                         if (syslog_identifier)
 290                                 iovec[n++] = IOVEC_MAKE_STRING(syslog_identifier);
 291                 }
 292
 293                 if (pid) {
 294                         syslog_pid = strappend("SYSLOG_PID=", pid);
 295                         if (syslog_pid)
 296                                 iovec[n++] = IOVEC_MAKE_STRING(syslog_pid);
 297                 }
 298         }
 299
 300         if (cunescape_length_with_prefix(p, pl, "MESSAGE=", UNESCAPE_RELAX, &message) >= 0)
 301                 iovec[n++] = IOVEC_MAKE_STRING(message);
 302
 303         server_dispatch_message(s, iovec, n, ELEMENTSOF(iovec), NULL, NULL, priority, 0);
 304
 305 finish:
 306         for (j = 0; j < z; j++)
 307                 free(iovec[j].iov_base);
 308 }
 309
 310 static int server_read_dev_kmsg(Server *s) {
 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) {
 325                         s->dev_kmsg_event_source = sd_event_source_unref(s->dev_kmsg_event_source);
 326                         return 0;
 327                 }
 328
 329                 if (IN_SET(errno, EAGAIN, EINTR, EPIPE))
 330                         return 0;
 331
 332                 return log_error_errno(errno, "Failed to read from kernel: %m");
 333         }
 334
 335         dev_kmsg_record(s, buffer, l);
 336         return 1;
 337 }
 338
 339 int 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
 350         log_debug("Flushing /dev/kmsg...");
 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
 364 static 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
 380 int server_open_dev_kmsg(Server *s) {
 381         mode_t mode;
 382         int r;
 383
 384         assert(s);
 385
 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);
 392         if (s->dev_kmsg_fd < 0) {
 393                 log_full(errno == ENOENT ? LOG_DEBUG : LOG_WARNING,
 394                          "Failed to open /dev/kmsg, ignoring: %m");
 395                 return 0;
 396         }
 397
 398         if (!s->read_kmsg)
 399                 return 0;
 400
 401         r = sd_event_add_io(s->event, &s->dev_kmsg_event_source, s->dev_kmsg_fd, EPOLLIN, dispatch_dev_kmsg, s);
 402         if (r < 0) {
 403
 404                 /* This will fail with EPERM on older kernels where
 405                  * /dev/kmsg is not readable. */
 406                 if (r == -EPERM) {
 407                         r = 0;
 408                         goto fail;
 409                 }
 410
 411                 log_error_errno(r, "Failed to add /dev/kmsg fd to event loop: %m");
 412                 goto fail;
 413         }
 414
 415         r = sd_event_source_set_priority(s->dev_kmsg_event_source, SD_EVENT_PRIORITY_IMPORTANT+10);
 416         if (r < 0) {
 417                 log_error_errno(r, "Failed to adjust priority of kmsg event source: %m");
 418                 goto fail;
 419         }
 420
 421         s->dev_kmsg_readable = true;
 422
 423         return 0;
 424
 425 fail:
 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);
 428
 429         return r;
 430 }
 431
 432 int server_open_kernel_seqnum(Server *s) {
 433         _cleanup_close_ int fd;
 434         uint64_t *p;
 435         int r;
 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
 441          * persistent and automatically flushed at reboot. */
 442
 443         fd = open("/run/systemd/journal/kernel-seqnum", O_RDWR|O_CREAT|O_CLOEXEC|O_NOCTTY|O_NOFOLLOW, 0644);
 444         if (fd < 0) {
 445                 log_error_errno(errno, "Failed to open /run/systemd/journal/kernel-seqnum, ignoring: %m");
 446                 return 0;
 447         }
 448
 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");
 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) {
 457                 log_error_errno(errno, "Failed to map sequential number file, ignoring: %m");
 458                 return 0;
 459         }
 460
 461         s->kernel_seqnum = p;
 462
 463         return 0;
 464 }