src/journal/journald-kmsg.c

   1 /* SPDX-License-Identifier: LGPL-2.1-or-later */
   2
   3 #include <fcntl.h>
   4 #include <sys/epoll.h>
   5 #include <sys/mman.h>
   6 #include <sys/socket.h>
   7 #include <unistd.h>
   8
   9 #include "sd-device.h"
  10 #include "sd-messages.h"
  11
  12 #include "alloc-util.h"
  13 #include "device-util.h"
  14 #include "escape.h"
  15 #include "fd-util.h"
  16 #include "format-util.h"
  17 #include "fs-util.h"
  18 #include "io-util.h"
  19 #include "journal-internal.h"
  20 #include "journald-kmsg.h"
  21 #include "journald-server.h"
  22 #include "journald-syslog.h"
  23 #include "log.h"
  24 #include "parse-util.h"
  25 #include "process-util.h"
  26 #include "stdio-util.h"
  27 #include "string-util.h"
  28
  29 void server_forward_kmsg(
  30                 Server *s,
  31                 int priority,
  32                 const char *identifier,
  33                 const char *message,
  34                 const struct ucred *ucred) {
  35
  36         _cleanup_free_ char *ident_buf = NULL;
  37         struct iovec iovec[5];
  38         char header_priority[DECIMAL_STR_MAX(priority) + 3],
  39              header_pid[STRLEN("[]: ") + DECIMAL_STR_MAX(pid_t) + 1];
  40         size_t n = 0;
  41
  42         assert(s);
  43         assert(priority >= 0);
  44         assert(priority <= 999);
  45         assert(message);
  46
  47         if (_unlikely_(LOG_PRI(priority) > s->max_level_kmsg))
  48                 return;
  49
  50         if (_unlikely_(s->dev_kmsg_fd < 0))
  51                 return;
  52
  53         /* Never allow messages with kernel facility to be written to
  54          * kmsg, regardless where the data comes from. */
  55         priority = syslog_fixup_facility(priority);
  56
  57         /* First: priority field */
  58         xsprintf(header_priority, "<%i>", priority);
  59         iovec[n++] = IOVEC_MAKE_STRING(header_priority);
  60
  61         /* Second: identifier and PID */
  62         if (ucred) {
  63                 if (!identifier) {
  64                         (void) get_process_comm(ucred->pid, &ident_buf);
  65                         identifier = ident_buf;
  66                 }
  67
  68                 if (identifier)
  69                         iovec[n++] = IOVEC_MAKE_STRING(identifier);
  70
  71                 xsprintf(header_pid, "["PID_FMT"]: ", ucred->pid);
  72                 iovec[n++] = IOVEC_MAKE_STRING(header_pid);
  73         } else if (identifier) {
  74                 iovec[n++] = IOVEC_MAKE_STRING(identifier);
  75                 iovec[n++] = IOVEC_MAKE_STRING(": ");
  76         }
  77
  78         /* Fourth: message */
  79         iovec[n++] = IOVEC_MAKE_STRING(message);
  80         iovec[n++] = IOVEC_MAKE_STRING("\n");
  81
  82         if (writev(s->dev_kmsg_fd, iovec, n) < 0)
  83                 log_debug_errno(errno, "Failed to write to /dev/kmsg for logging, ignoring: %m");
  84 }
  85
  86 static bool is_us(const char *identifier, const char *pid) {
  87         pid_t pid_num;
  88
  89         if (!identifier || !pid)
  90                 return false;
  91
  92         if (parse_pid(pid, &pid_num) < 0)
  93                 return false;
  94
  95         return pid_num == getpid_cached() &&
  96                streq(identifier, program_invocation_short_name);
  97 }
  98
  99 void dev_kmsg_record(Server *s, char *p, size_t l) {
 100
 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;
 102         struct iovec iovec[N_IOVEC_META_FIELDS + 7 + N_IOVEC_KERNEL_FIELDS + 2 + N_IOVEC_UDEV_FIELDS];
 103         char *kernel_device = NULL;
 104         unsigned long long usec;
 105         size_t n = 0, z = 0, j;
 106         int priority, r;
 107         char *e, *f, *k;
 108         uint64_t serial;
 109         size_t pl;
 110         int saved_log_max_level = INT_MAX;
 111         ClientContext *c = NULL;
 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                         goto finish;
 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                 _cleanup_(sd_device_unrefp) sd_device *d = NULL;
 218
 219                 if (sd_device_new_from_device_id(&d, kernel_device) >= 0) {
 220                         const char *g;
 221                         char *b;
 222
 223                         if (sd_device_get_devname(d, &g) >= 0) {
 224                                 b = strjoin("_UDEV_DEVNODE=", g);
 225                                 if (b) {
 226                                         iovec[n++] = IOVEC_MAKE_STRING(b);
 227                                         z++;
 228                                 }
 229                         }
 230
 231                         if (sd_device_get_sysname(d, &g) >= 0) {
 232                                 b = strjoin("_UDEV_SYSNAME=", g);
 233                                 if (b) {
 234                                         iovec[n++] = IOVEC_MAKE_STRING(b);
 235                                         z++;
 236                                 }
 237                         }
 238
 239                         j = 0;
 240                         FOREACH_DEVICE_DEVLINK(d, link) {
 241
 242                                 if (j >= N_IOVEC_UDEV_FIELDS)
 243                                         break;
 244
 245                                 b = strjoin("_UDEV_DEVLINK=", link);
 246                                 if (b) {
 247                                         iovec[n++] = IOVEC_MAKE_STRING(b);
 248                                         z++;
 249                                 }
 250
 251                                 j++;
 252                         }
 253                 }
 254         }
 255
 256         if (asprintf(&source_time, "_SOURCE_MONOTONIC_TIMESTAMP=%llu", usec) >= 0)
 257                 iovec[n++] = IOVEC_MAKE_STRING(source_time);
 258
 259         iovec[n++] = IOVEC_MAKE_STRING("_TRANSPORT=kernel");
 260
 261         if (asprintf(&syslog_priority, "PRIORITY=%i", priority & LOG_PRIMASK) >= 0)
 262                 iovec[n++] = IOVEC_MAKE_STRING(syslog_priority);
 263
 264         if (asprintf(&syslog_facility, "SYSLOG_FACILITY=%i", LOG_FAC(priority)) >= 0)
 265                 iovec[n++] = IOVEC_MAKE_STRING(syslog_facility);
 266
 267         if (LOG_FAC(priority) == LOG_KERN)
 268                 iovec[n++] = IOVEC_MAKE_STRING("SYSLOG_IDENTIFIER=kernel");
 269         else {
 270                 pl -= syslog_parse_identifier((const char**) &p, &identifier, &pid);
 271
 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                 }
 282
 283                 if (identifier) {
 284                         syslog_identifier = strjoin("SYSLOG_IDENTIFIER=", identifier);
 285                         if (syslog_identifier)
 286                                 iovec[n++] = IOVEC_MAKE_STRING(syslog_identifier);
 287                 }
 288
 289                 if (pid) {
 290                         syslog_pid = strjoin("SYSLOG_PID=", pid);
 291                         if (syslog_pid)
 292                                 iovec[n++] = IOVEC_MAKE_STRING(syslog_pid);
 293                 }
 294         }
 295
 296         if (cunescape_length_with_prefix(p, pl, "MESSAGE=", UNESCAPE_RELAX, &message) >= 0)
 297                 iovec[n++] = IOVEC_MAKE_STRING(message);
 298
 299         server_dispatch_message(s, iovec, n, ELEMENTSOF(iovec), c, NULL, priority, 0);
 300
 301         if (saved_log_max_level != INT_MAX)
 302                 log_set_max_level(saved_log_max_level);
 303
 304 finish:
 305         for (j = 0; j < z; j++)
 306                 free(iovec[j].iov_base);
 307 }
 308
 309 static int server_read_dev_kmsg(Server *s) {
 310         char buffer[8192+1]; /* the kernel-side limit per record is 8K currently */
 311         ssize_t l;
 312
 313         assert(s);
 314         assert(s->dev_kmsg_fd >= 0);
 315
 316         l = read(s->dev_kmsg_fd, buffer, sizeof(buffer) - 1);
 317         if (l == 0)
 318                 return 0;
 319         if (l < 0) {
 320                 /* Old kernels which don't allow reading from /dev/kmsg return EINVAL when we try. So handle
 321                  * this cleanly, but don't try to ever read from it again. */
 322                 if (errno == EINVAL) {
 323                         s->dev_kmsg_event_source = sd_event_source_unref(s->dev_kmsg_event_source);
 324                         s->dev_kmsg_readable = false;
 325                         return 0;
 326                 }
 327
 328                 if (ERRNO_IS_TRANSIENT(errno) || errno == EPIPE)
 329                         return 0;
 330
 331                 return log_ratelimit_error_errno(errno, JOURNAL_LOG_RATELIMIT, "Failed to read from /dev/kmsg: %m");
 332         }
 333
 334         dev_kmsg_record(s, buffer, l);
 335         return 1;
 336 }
 337
 338 int server_flush_dev_kmsg(Server *s) {
 339         int r;
 340
 341         assert(s);
 342
 343         if (s->dev_kmsg_fd < 0)
 344                 return 0;
 345
 346         if (!s->dev_kmsg_readable)
 347                 return 0;
 348
 349         log_debug("Flushing /dev/kmsg...");
 350
 351         for (;;) {
 352                 r = server_read_dev_kmsg(s);
 353                 if (r < 0)
 354                         return r;
 355
 356                 if (r == 0)
 357                         break;
 358         }
 359
 360         return 0;
 361 }
 362
 363 static int dispatch_dev_kmsg(sd_event_source *es, int fd, uint32_t revents, void *userdata) {
 364         Server *s = ASSERT_PTR(userdata);
 365
 366         assert(es);
 367         assert(fd == s->dev_kmsg_fd);
 368
 369         if (revents & EPOLLERR)
 370                 log_ratelimit_warning(JOURNAL_LOG_RATELIMIT,
 371                                       "/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
 379 int server_open_dev_kmsg(Server *s) {
 380         mode_t mode;
 381         int r;
 382
 383         assert(s);
 384
 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);
 391         if (s->dev_kmsg_fd < 0) {
 392                 log_full_errno(errno == ENOENT ? LOG_DEBUG : LOG_WARNING,
 393                                errno, "Failed to open /dev/kmsg, ignoring: %m");
 394                 return 0;
 395         }
 396
 397         if (!s->read_kmsg)
 398                 return 0;
 399
 400         r = sd_event_add_io(s->event, &s->dev_kmsg_event_source, s->dev_kmsg_fd, EPOLLIN, dispatch_dev_kmsg, s);
 401         if (r == -EPERM) { /* This will fail with EPERM on older kernels where /dev/kmsg is not readable. */
 402                 r = 0;
 403                 goto finish;
 404         }
 405         if (r < 0) {
 406                 log_error_errno(r, "Failed to add /dev/kmsg fd to event loop: %m");
 407                 goto finish;
 408         }
 409
 410         r = sd_event_source_set_priority(s->dev_kmsg_event_source, SD_EVENT_PRIORITY_IMPORTANT+10);
 411         if (r < 0) {
 412                 log_error_errno(r, "Failed to adjust priority of kmsg event source: %m");
 413                 goto finish;
 414         }
 415
 416         s->dev_kmsg_readable = true;
 417         return 0;
 418
 419 finish:
 420         s->dev_kmsg_event_source = sd_event_source_unref(s->dev_kmsg_event_source);
 421         s->dev_kmsg_fd = safe_close(s->dev_kmsg_fd);
 422         return r;
 423 }
 424
 425 int server_open_kernel_seqnum(Server *s) {
 426         int r;
 427
 428         assert(s);
 429
 430         /* We store the seqnum we last read in an mmapped file. That way we can just use it like a variable,
 431          * but it is persistent and automatically flushed at reboot. */
 432
 433         if (!s->dev_kmsg_readable)
 434                 return 0;
 435
 436         r = server_map_seqnum_file(s, "kernel-seqnum", sizeof(uint64_t), (void**) &s->kernel_seqnum);
 437         if (r < 0)
 438                 return log_error_errno(r, "Failed to map kernel seqnum file: %m");
 439
 440         return 0;
 441 }