src/journal/journald-kmsg.c

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