1 /* 2 * CDDL HEADER START 3 * 4 * The contents of this file are subject to the terms of the 5 * Common Development and Distribution License (the "License"). 6 * You may not use this file except in compliance with the License. 7 * 8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE 9 * or http://www.opensolaris.org/os/licensing. 10 * See the License for the specific language governing permissions 11 * and limitations under the License. 12 * 13 * When distributing Covered Code, include this CDDL HEADER in each 14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE. 15 * If applicable, add the following below this CDDL HEADER, with the 16 * fields enclosed by brackets "[]" replaced with your own identifying 17 * information: Portions Copyright [yyyy] [name of copyright owner] 18 * 19 * CDDL HEADER END 20 */ 21 22 /* 23 * Copyright 2009 Sun Microsystems, Inc. All rights reserved. 24 * Use is subject to license terms. 25 */ 26 27 #include <sys/sysevent/eventdefs.h> 28 #include <sys/sysevent.h> 29 #include <sys/sysevent_impl.h> 30 #include <sys/fm/protocol.h> 31 #include <sys/sysmacros.h> 32 #include <sys/dumphdr.h> 33 #include <sys/dumpadm.h> 34 #include <sys/fm/util.h> 35 36 #include <libsysevent.h> 37 #include <libnvpair.h> 38 #include <alloca.h> 39 #include <limits.h> 40 #include <strings.h> 41 #include <unistd.h> 42 #include <fcntl.h> 43 #include <errno.h> 44 45 #undef MUTEX_HELD 46 #undef RW_READ_HELD 47 #undef RW_WRITE_HELD 48 49 #include <fmd_api.h> 50 #include <fmd_log.h> 51 #include <fmd_subr.h> 52 #include <fmd_dispq.h> 53 #include <fmd_dr.h> 54 #include <fmd_module.h> 55 #include <fmd_protocol.h> 56 #include <fmd_scheme.h> 57 #include <fmd_error.h> 58 59 #include <fmd.h> 60 61 static char *sysev_channel; /* event channel to which we are subscribed */ 62 static char *sysev_class; /* event class to which we are subscribed */ 63 static char *sysev_device; /* device path to use for replaying events */ 64 static char *sysev_sid; /* event channel subscriber identifier */ 65 static void *sysev_evc; /* event channel cookie from evc_bind */ 66 67 static fmd_xprt_t *sysev_xprt; 68 static int sysev_xprt_refcnt; 69 static fmd_hdl_t *sysev_hdl; 70 71 static struct sysev_stats { 72 fmd_stat_t dump_replay; 73 fmd_stat_t dump_lost; 74 fmd_stat_t bad_class; 75 fmd_stat_t bad_attr; 76 fmd_stat_t eagain; 77 } sysev_stats = { 78 { "dump_replay", FMD_TYPE_UINT64, "events replayed from dump device" }, 79 { "dump_lost", FMD_TYPE_UINT64, "events lost from dump device" }, 80 { "bad_class", FMD_TYPE_UINT64, "events dropped due to invalid class" }, 81 { "bad_attr", FMD_TYPE_UINT64, "events dropped due to invalid nvlist" }, 82 { "eagain", FMD_TYPE_UINT64, "events retried due to low memory" }, 83 }; 84 85 static pthread_cond_t sysev_cv = PTHREAD_COND_INITIALIZER; 86 static pthread_mutex_t sysev_mutex = PTHREAD_MUTEX_INITIALIZER; 87 static int sysev_replay_wait = 1; 88 static int sysev_exiting; 89 90 /* 91 * Entry point for legacy sysevents. This function is responsible for two 92 * things: passing off interesting events to the DR handler, and converting 93 * sysevents into resource events that modules can then subscribe to. 94 */ 95 static void 96 sysev_legacy(sysevent_t *sep) 97 { 98 const char *class = sysevent_get_class_name(sep); 99 const char *subclass = sysevent_get_subclass_name(sep); 100 char *fullclass; 101 size_t len; 102 nvlist_t *attr, *nvl; 103 fmd_event_t *e; 104 hrtime_t hrt; 105 106 /* notify the DR subsystem of the event */ 107 fmd_dr_event(sep); 108 109 /* get the matching sysevent name */ 110 len = snprintf(NULL, 0, "%s%s.%s", SYSEVENT_RSRC_CLASS, 111 class, subclass); 112 fullclass = alloca(len + 1); 113 (void) snprintf(fullclass, len + 1, "%s%s.%s", 114 SYSEVENT_RSRC_CLASS, class, subclass); 115 116 /* construct the event payload */ 117 (void) nvlist_xalloc(&nvl, NV_UNIQUE_NAME, &fmd.d_nva); 118 (void) nvlist_add_string(nvl, FM_CLASS, fullclass); 119 (void) nvlist_add_uint8(nvl, FM_VERSION, FM_RSRC_VERSION); 120 if (sysevent_get_attr_list(sep, &attr) == 0) { 121 (void) nvlist_merge(nvl, attr, 0); 122 nvlist_free(attr); 123 } 124 125 /* 126 * Dispatch the event. Ideally, we'd like to use the same transport 127 * interface as sysev_recv(), but because the legacy sysevent mechanism 128 * puts in a thread outside fmd's control, using the module APIs is 129 * impossible. 130 */ 131 sysevent_get_time(sep, &hrt); 132 (void) nvlist_lookup_string(nvl, FM_CLASS, &fullclass); 133 e = fmd_event_create(FMD_EVT_PROTOCOL, hrt, nvl, fullclass); 134 fmd_dispq_dispatch(fmd.d_disp, e, fullclass); 135 } 136 137 /* 138 * Receive an event from the SysEvent channel and post it to our transport. 139 * Under extreme low-memory situations where we cannot event unpack the event, 140 * we can request that SysEvent redeliver the event later by returning EAGAIN. 141 * If we do this too many times, the kernel will drop the event. Rather than 142 * keeping state per-event, we simply attempt a garbage-collect, hoping that 143 * enough free memory will be available by the time the event is redelivered. 144 */ 145 static int 146 sysev_recv(sysevent_t *sep, void *arg) 147 { 148 uint64_t seq = sysevent_get_seq(sep); 149 fmd_xprt_t *xp = arg; 150 nvlist_t *nvl; 151 hrtime_t hrt; 152 int rc = 0; 153 154 (void) pthread_mutex_lock(&sysev_mutex); 155 if (sysev_exiting == 1) { 156 while (sysev_xprt_refcnt > 0) 157 (void) pthread_cond_wait(&sysev_cv, &sysev_mutex); 158 (void) pthread_mutex_unlock(&sysev_mutex); 159 return (EAGAIN); 160 } 161 sysev_xprt_refcnt++; 162 while (sysev_replay_wait) 163 (void) pthread_cond_wait(&sysev_cv, &sysev_mutex); 164 (void) pthread_mutex_unlock(&sysev_mutex); 165 166 if (strcmp(sysevent_get_class_name(sep), EC_FM) != 0) { 167 fmd_hdl_error(sysev_hdl, "discarding event 0x%llx: unexpected" 168 " transport class %s\n", seq, sysevent_get_class_name(sep)); 169 sysev_stats.bad_class.fmds_value.ui64++; 170 } else if (sysevent_get_attr_list(sep, &nvl) != 0) { 171 if (errno == EAGAIN || errno == ENOMEM) { 172 fmd_modhash_tryapply(fmd.d_mod_hash, fmd_module_trygc); 173 fmd_scheme_hash_trygc(fmd.d_schemes); 174 sysev_stats.eagain.fmds_value.ui64++; 175 rc = EAGAIN; 176 } else { 177 fmd_hdl_error(sysev_hdl, "discarding event 0x%llx: " 178 "missing or invalid payload", seq); 179 sysev_stats.bad_attr.fmds_value.ui64++; 180 } 181 } else { 182 sysevent_get_time(sep, &hrt); 183 fmd_xprt_post(sysev_hdl, xp, nvl, hrt); 184 } 185 186 (void) pthread_mutex_lock(&sysev_mutex); 187 if (--sysev_xprt_refcnt == 0 && sysev_exiting == 1) 188 (void) pthread_cond_broadcast(&sysev_cv); 189 (void) pthread_mutex_unlock(&sysev_mutex); 190 191 return (rc); 192 } 193 194 /* 195 * Checksum algorithm used by the dump transport for verifying the content of 196 * error reports saved on the dump device (copy of the kernel's checksum32()). 197 */ 198 static uint32_t 199 sysev_checksum(void *cp_arg, size_t length) 200 { 201 uchar_t *cp, *ep; 202 uint32_t sum = 0; 203 204 for (cp = cp_arg, ep = cp + length; cp < ep; cp++) 205 sum = ((sum >> 1) | (sum << 31)) + *cp; 206 207 return (sum); 208 } 209 210 /* 211 * Replay saved events from the dump transport. This function is installed as 212 * the timer callback and is called only once during the module's lifetime. 213 */ 214 /*ARGSUSED*/ 215 static void 216 sysev_replay(fmd_hdl_t *hdl, id_t id, void *arg) 217 { 218 char *dumpdev; 219 off64_t off, off0; 220 int fd, err; 221 222 /* 223 * Determine the appropriate dump device to use for replaying pending 224 * error reports. If the device property is NULL (default), we 225 * open and query /dev/dump to determine the current dump device. 226 */ 227 if ((dumpdev = sysev_device) == NULL) { 228 if ((fd = open("/dev/dump", O_RDONLY)) == -1) { 229 fmd_hdl_error(hdl, "failed to open /dev/dump " 230 "to locate dump device for event replay"); 231 goto done; 232 } 233 234 dumpdev = alloca(PATH_MAX); 235 err = ioctl(fd, DIOCGETDEV, dumpdev); 236 (void) close(fd); 237 238 if (err == -1) { 239 if (errno != ENODEV) { 240 fmd_hdl_error(hdl, "failed to obtain " 241 "path to dump device for event replay"); 242 } 243 goto done; 244 } 245 } 246 247 if (strcmp(dumpdev, "/dev/null") == 0) 248 goto done; /* return silently and skip replay for /dev/null */ 249 250 /* 251 * Open the appropriate device and then determine the offset of the 252 * start of the ereport dump region located at the end of the device. 253 */ 254 if ((fd = open64(dumpdev, O_RDWR | O_DSYNC)) == -1) { 255 fmd_hdl_error(hdl, "failed to open dump transport %s " 256 "(pending events will not be replayed)", dumpdev); 257 goto done; 258 } 259 260 off = DUMP_OFFSET + DUMP_LOGSIZE + DUMP_ERPTSIZE; 261 off = off0 = lseek64(fd, -off, SEEK_END) & -DUMP_OFFSET; 262 263 if (off == (off64_t)-1LL) { 264 fmd_hdl_error(hdl, "failed to seek dump transport %s " 265 "(pending events will not be replayed)", dumpdev); 266 (void) close(fd); 267 goto done; 268 } 269 270 /* 271 * The ereport dump region is a sequence of erpt_dump_t headers each of 272 * which is followed by packed nvlist data. We iterate over them in 273 * order, unpacking and dispatching each one to our dispatch queue. 274 */ 275 for (;;) { 276 char nvbuf[ERPT_DATA_SZ]; 277 uint32_t chksum; 278 erpt_dump_t ed; 279 nvlist_t *nvl; 280 281 fmd_timeval_t ftv, tod; 282 hrtime_t hrt; 283 uint64_t ena; 284 285 if (pread64(fd, &ed, sizeof (ed), off) != sizeof (ed)) { 286 fmd_hdl_error(hdl, "failed to read from dump " 287 "transport %s (pending events lost)", dumpdev); 288 break; 289 } 290 291 if (ed.ed_magic == 0 && ed.ed_size == 0) 292 break; /* end of list: all zero */ 293 294 if (ed.ed_magic == 0) { 295 off += sizeof (ed) + ed.ed_size; 296 continue; /* continue searching */ 297 } 298 299 if (ed.ed_magic != ERPT_MAGIC) { 300 /* 301 * Stop reading silently if the first record has the 302 * wrong magic number; this likely indicates that we 303 * rebooted from non-FMA bits or paged over the dump. 304 */ 305 if (off == off0) 306 break; 307 308 fmd_hdl_error(hdl, "invalid dump transport " 309 "record at %llx (magic number %x, expected %x)\n", 310 (u_longlong_t)off, ed.ed_magic, ERPT_MAGIC); 311 break; 312 } 313 314 if (ed.ed_size > ERPT_DATA_SZ) { 315 fmd_hdl_error(hdl, "invalid dump transport " 316 "record at %llx size (%u exceeds limit)\n", 317 (u_longlong_t)off, ed.ed_size); 318 break; 319 } 320 321 if (pread64(fd, nvbuf, ed.ed_size, 322 off + sizeof (ed)) != ed.ed_size) { 323 fmd_hdl_error(hdl, "failed to read dump " 324 "transport event (offset %llx)", (u_longlong_t)off); 325 326 sysev_stats.dump_lost.fmds_value.ui64++; 327 goto next; 328 } 329 330 if ((chksum = sysev_checksum(nvbuf, 331 ed.ed_size)) != ed.ed_chksum) { 332 fmd_hdl_error(hdl, "dump transport event at " 333 "offset %llx is corrupt (checksum %x != %x)\n", 334 (u_longlong_t)off, chksum, ed.ed_chksum); 335 336 sysev_stats.dump_lost.fmds_value.ui64++; 337 goto next; 338 } 339 340 if ((err = nvlist_xunpack(nvbuf, 341 ed.ed_size, &nvl, &fmd.d_nva)) != 0) { 342 fmd_hdl_error(hdl, "failed to unpack dump " 343 "transport event at offset %llx: %s\n", 344 (u_longlong_t)off, fmd_strerror(err)); 345 346 sysev_stats.dump_lost.fmds_value.ui64++; 347 goto next; 348 } 349 350 /* 351 * If ed_hrt_nsec is set it contains the gethrtime() value from 352 * when the event was originally enqueued for the transport. 353 * If it is zero, we use the weaker bound ed_hrt_base instead. 354 */ 355 if (ed.ed_hrt_nsec != 0) 356 hrt = ed.ed_hrt_nsec; 357 else 358 hrt = ed.ed_hrt_base; 359 360 /* 361 * If this is an FMA protocol event of class "ereport.*" that 362 * contains valid ENA, we can improve the precision of 'hrt'. 363 */ 364 if (nvlist_lookup_uint64(nvl, FM_EREPORT_ENA, &ena) == 0) 365 hrt = fmd_time_ena2hrt(hrt, ena); 366 367 /* 368 * Now convert 'hrt' to an adjustable TOD based on the values 369 * in ed_tod_base which correspond to one another and are 370 * sampled before reboot using the old gethrtime() clock. 371 * fmd_event_recreate() will use this TOD value to re-assign 372 * the event an updated gethrtime() value based on the current 373 * value of the non-adjustable gethrtime() clock. Phew. 374 */ 375 tod.ftv_sec = ed.ed_tod_base.sec; 376 tod.ftv_nsec = ed.ed_tod_base.nsec; 377 fmd_time_hrt2tod(ed.ed_hrt_base, &tod, hrt, &ftv); 378 379 (void) nvlist_remove_all(nvl, FMD_EVN_TOD); 380 (void) nvlist_add_uint64_array(nvl, 381 FMD_EVN_TOD, (uint64_t *)&ftv, 2); 382 383 fmd_xprt_post(hdl, sysev_xprt, nvl, 0); 384 sysev_stats.dump_replay.fmds_value.ui64++; 385 386 next: 387 /* 388 * Reset the magic number for the event record to zero so that 389 * we do not replay the same event multiple times. 390 */ 391 ed.ed_magic = 0; 392 393 if (pwrite64(fd, &ed, sizeof (ed), off) != sizeof (ed)) { 394 fmd_hdl_error(hdl, "failed to mark dump " 395 "transport event (offset %llx)", (u_longlong_t)off); 396 } 397 398 off += sizeof (ed) + ed.ed_size; 399 } 400 401 (void) close(fd); 402 done: 403 (void) pthread_mutex_lock(&sysev_mutex); 404 sysev_replay_wait = 0; 405 (void) pthread_cond_broadcast(&sysev_cv); 406 (void) pthread_mutex_unlock(&sysev_mutex); 407 } 408 409 static const fmd_prop_t sysev_props[] = { 410 { "class", FMD_TYPE_STRING, EC_ALL }, /* event class */ 411 { "device", FMD_TYPE_STRING, NULL }, /* replay device */ 412 { "channel", FMD_TYPE_STRING, FM_ERROR_CHAN }, /* channel name */ 413 { "sid", FMD_TYPE_STRING, "fmd" }, /* subscriber id */ 414 { NULL, 0, NULL } 415 }; 416 417 static const fmd_hdl_ops_t sysev_ops = { 418 NULL, /* fmdo_recv */ 419 sysev_replay, /* fmdo_timeout */ 420 NULL, /* fmdo_close */ 421 NULL, /* fmdo_stats */ 422 NULL, /* fmdo_gc */ 423 NULL, /* fmdo_send */ 424 }; 425 426 static const fmd_hdl_info_t sysev_info = { 427 "SysEvent Transport Agent", "1.0", &sysev_ops, sysev_props 428 }; 429 430 /* 431 * Bind to the sysevent channel we use for listening for error events and then 432 * subscribe to appropriate events received over this channel. Setup the 433 * legacy sysevent handler for creating sysevent resources and forwarding DR 434 * events. 435 */ 436 void 437 sysev_init(fmd_hdl_t *hdl) 438 { 439 uint_t flags; 440 const char *subclasses[] = { EC_SUB_ALL }; 441 442 if (fmd_hdl_register(hdl, FMD_API_VERSION, &sysev_info) != 0) 443 return; /* invalid property settings */ 444 445 (void) fmd_stat_create(hdl, FMD_STAT_NOALLOC, sizeof (sysev_stats) / 446 sizeof (fmd_stat_t), (fmd_stat_t *)&sysev_stats); 447 448 sysev_channel = fmd_prop_get_string(hdl, "channel"); 449 sysev_class = fmd_prop_get_string(hdl, "class"); 450 sysev_device = fmd_prop_get_string(hdl, "device"); 451 sysev_sid = fmd_prop_get_string(hdl, "sid"); 452 453 if (sysev_channel == NULL) 454 fmd_hdl_abort(hdl, "channel property must be defined\n"); 455 456 if (sysev_sid == NULL) 457 fmd_hdl_abort(hdl, "sid property must be defined\n"); 458 459 if ((errno = sysevent_evc_bind(sysev_channel, &sysev_evc, 460 EVCH_CREAT | EVCH_HOLD_PEND)) != 0) { 461 fmd_hdl_abort(hdl, "failed to bind to event transport " 462 "channel %s", sysev_channel); 463 } 464 465 sysev_xprt = fmd_xprt_open(hdl, FMD_XPRT_RDONLY | 466 FMD_XPRT_CACHE_AS_LOCAL, NULL, NULL); 467 sysev_hdl = hdl; 468 469 /* 470 * If we're subscribing to the default channel, keep our subscription 471 * active even if we die unexpectedly so we continue queuing events. 472 * If we're not (e.g. running under fmsim), do not specify SUB_KEEP so 473 * that our event channel will be destroyed if we die unpleasantly. 474 */ 475 if (strcmp(sysev_channel, FM_ERROR_CHAN) == 0) 476 flags = EVCH_SUB_KEEP | EVCH_SUB_DUMP; 477 else 478 flags = EVCH_SUB_DUMP; 479 480 errno = sysevent_evc_subscribe(sysev_evc, 481 sysev_sid, sysev_class, sysev_recv, sysev_xprt, flags); 482 483 if (errno != 0) { 484 if (errno == EEXIST) { 485 fmd_hdl_abort(hdl, "another fault management daemon is " 486 "active on transport channel %s\n", sysev_channel); 487 } else { 488 fmd_hdl_abort(hdl, "failed to subscribe to %s on " 489 "transport channel %s", sysev_class, sysev_channel); 490 } 491 } 492 493 /* 494 * Once the transport is open, install a single timer to fire at once 495 * in the context of the module's thread to run sysev_replay(). This 496 * thread will block in its first fmd_xprt_post() until fmd is ready. 497 */ 498 fmd_hdl_debug(hdl, "transport '%s' open\n", sysev_channel); 499 (void) fmd_timer_install(hdl, NULL, NULL, 0); 500 501 /* 502 * Open the legacy sysevent handle and subscribe to all events. These 503 * are automatically converted to "resource.sysevent.*" events so that 504 * modules can manage these events without additional infrastructure. 505 */ 506 if (geteuid() != 0) 507 return; 508 509 if ((fmd.d_sysev_hdl = 510 sysevent_bind_handle(sysev_legacy)) == NULL) 511 fmd_hdl_abort(hdl, "failed to bind to legacy sysevent channel"); 512 513 if (sysevent_subscribe_event(fmd.d_sysev_hdl, EC_ALL, 514 subclasses, 1) != 0) 515 fmd_hdl_abort(hdl, "failed to subscribe to legacy sysevents"); 516 } 517 518 /* 519 * Close the channel by unsubscribing and unbinding. We only do this when a 520 * a non-default channel has been selected. If we're using FM_ERROR_CHAN, 521 * the system default, we do *not* want to unsubscribe because the kernel will 522 * remove the subscriber queue and any events published in our absence will 523 * therefore be lost. This scenario may occur when, for example, fmd is sent 524 * a SIGTERM by init(1M) during reboot but an error is detected and makes it 525 * into the sysevent channel queue before init(1M) manages to call uadmin(2). 526 */ 527 void 528 sysev_fini(fmd_hdl_t *hdl) 529 { 530 if (strcmp(sysev_channel, FM_ERROR_CHAN) != 0) { 531 sysevent_evc_unsubscribe(sysev_evc, sysev_sid); 532 sysevent_evc_unbind(sysev_evc); 533 } 534 535 if (fmd.d_sysev_hdl != NULL) 536 sysevent_unbind_handle(fmd.d_sysev_hdl); 537 538 if (sysev_xprt != NULL) { 539 /* 540 * Wait callback returns before destroy the transport. 541 */ 542 (void) pthread_mutex_lock(&sysev_mutex); 543 sysev_exiting = 1; 544 while (sysev_xprt_refcnt > 0) 545 (void) pthread_cond_wait(&sysev_cv, &sysev_mutex); 546 (void) pthread_mutex_unlock(&sysev_mutex); 547 fmd_xprt_close(hdl, sysev_xprt); 548 } 549 550 fmd_prop_free_string(hdl, sysev_class); 551 fmd_prop_free_string(hdl, sysev_channel); 552 fmd_prop_free_string(hdl, sysev_device); 553 fmd_prop_free_string(hdl, sysev_sid); 554 } 555