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 2008 Sun Microsystems, Inc. All rights reserved. 24 * Use is subject to license terms. 25 */ 26 27 #pragma ident "%Z%%M% %I% %E% SMI" 28 29 #include <sys/sysevent/eventdefs.h> 30 #include <sys/sysevent.h> 31 #include <sys/sysevent_impl.h> 32 #include <sys/fm/protocol.h> 33 #include <sys/sysmacros.h> 34 #include <sys/dumphdr.h> 35 #include <sys/dumpadm.h> 36 #include <sys/fm/util.h> 37 38 #include <libsysevent.h> 39 #include <libnvpair.h> 40 #include <alloca.h> 41 #include <limits.h> 42 #include <strings.h> 43 #include <unistd.h> 44 #include <fcntl.h> 45 #include <errno.h> 46 47 #undef MUTEX_HELD 48 #undef RW_READ_HELD 49 #undef RW_WRITE_HELD 50 51 #include <fmd_api.h> 52 #include <fmd_log.h> 53 #include <fmd_subr.h> 54 #include <fmd_dispq.h> 55 #include <fmd_dr.h> 56 #include <fmd_module.h> 57 #include <fmd_protocol.h> 58 #include <fmd_scheme.h> 59 #include <fmd_error.h> 60 61 #include <fmd.h> 62 63 static char *sysev_channel; /* event channel to which we are subscribed */ 64 static char *sysev_class; /* event class to which we are subscribed */ 65 static char *sysev_device; /* device path to use for replaying events */ 66 static char *sysev_sid; /* event channel subscriber identifier */ 67 static void *sysev_evc; /* event channel cookie from evc_bind */ 68 69 static fmd_xprt_t *sysev_xprt; 70 static int sysev_xprt_refcnt; 71 static fmd_hdl_t *sysev_hdl; 72 73 static struct sysev_stats { 74 fmd_stat_t dump_replay; 75 fmd_stat_t dump_lost; 76 fmd_stat_t bad_class; 77 fmd_stat_t bad_attr; 78 fmd_stat_t eagain; 79 } sysev_stats = { 80 { "dump_replay", FMD_TYPE_UINT64, "events replayed from dump device" }, 81 { "dump_lost", FMD_TYPE_UINT64, "events lost from dump device" }, 82 { "bad_class", FMD_TYPE_UINT64, "events dropped due to invalid class" }, 83 { "bad_attr", FMD_TYPE_UINT64, "events dropped due to invalid nvlist" }, 84 { "eagain", FMD_TYPE_UINT64, "events retried due to low memory" }, 85 }; 86 87 static pthread_cond_t sysev_cv = PTHREAD_COND_INITIALIZER; 88 static pthread_mutex_t sysev_mutex = PTHREAD_MUTEX_INITIALIZER; 89 static int sysev_replay_wait = 1; 90 static int sysev_exiting; 91 92 /* 93 * Entry point for legacy sysevents. This function is responsible for two 94 * things: passing off interesting events to the DR handler, and converting 95 * sysevents into resource events that modules can then subscribe to. 96 */ 97 static void 98 sysev_legacy(sysevent_t *sep) 99 { 100 const char *class = sysevent_get_class_name(sep); 101 const char *subclass = sysevent_get_subclass_name(sep); 102 char *fullclass; 103 size_t len; 104 nvlist_t *attr, *nvl; 105 fmd_event_t *e; 106 hrtime_t hrt; 107 108 /* notify the DR subsystem of the event */ 109 fmd_dr_event(sep); 110 111 /* get the matching sysevent name */ 112 len = snprintf(NULL, 0, "%s%s.%s", SYSEVENT_RSRC_CLASS, 113 class, subclass); 114 fullclass = alloca(len + 1); 115 (void) snprintf(fullclass, len + 1, "%s%s.%s", 116 SYSEVENT_RSRC_CLASS, class, subclass); 117 118 /* construct the event payload */ 119 (void) nvlist_xalloc(&nvl, NV_UNIQUE_NAME, &fmd.d_nva); 120 (void) nvlist_add_string(nvl, FM_CLASS, fullclass); 121 (void) nvlist_add_uint8(nvl, FM_VERSION, FM_RSRC_VERSION); 122 if (sysevent_get_attr_list(sep, &attr) == 0) { 123 (void) nvlist_merge(nvl, attr, 0); 124 nvlist_free(attr); 125 } 126 127 /* 128 * Dispatch the event. Ideally, we'd like to use the same transport 129 * interface as sysev_recv(), but because the legacy sysevent mechanism 130 * puts in a thread outside fmd's control, using the module APIs is 131 * impossible. 132 */ 133 sysevent_get_time(sep, &hrt); 134 (void) nvlist_lookup_string(nvl, FM_CLASS, &fullclass); 135 e = fmd_event_create(FMD_EVT_PROTOCOL, hrt, nvl, fullclass); 136 fmd_dispq_dispatch(fmd.d_disp, e, fullclass); 137 } 138 139 /* 140 * Receive an event from the SysEvent channel and post it to our transport. 141 * Under extreme low-memory situations where we cannot event unpack the event, 142 * we can request that SysEvent redeliver the event later by returning EAGAIN. 143 * If we do this too many times, the kernel will drop the event. Rather than 144 * keeping state per-event, we simply attempt a garbage-collect, hoping that 145 * enough free memory will be available by the time the event is redelivered. 146 */ 147 static int 148 sysev_recv(sysevent_t *sep, void *arg) 149 { 150 uint64_t seq = sysevent_get_seq(sep); 151 fmd_xprt_t *xp = arg; 152 nvlist_t *nvl; 153 hrtime_t hrt; 154 int rc = 0; 155 156 (void) pthread_mutex_lock(&sysev_mutex); 157 if (sysev_exiting == 1) { 158 while (sysev_xprt_refcnt > 0) 159 (void) pthread_cond_wait(&sysev_cv, &sysev_mutex); 160 (void) pthread_mutex_unlock(&sysev_mutex); 161 return (EAGAIN); 162 } 163 sysev_xprt_refcnt++; 164 while (sysev_replay_wait) 165 (void) pthread_cond_wait(&sysev_cv, &sysev_mutex); 166 (void) pthread_mutex_unlock(&sysev_mutex); 167 168 if (strcmp(sysevent_get_class_name(sep), EC_FM) != 0) { 169 fmd_hdl_error(sysev_hdl, "discarding event 0x%llx: unexpected" 170 " transport class %s\n", seq, sysevent_get_class_name(sep)); 171 sysev_stats.bad_class.fmds_value.ui64++; 172 } else if (sysevent_get_attr_list(sep, &nvl) != 0) { 173 if (errno == EAGAIN || errno == ENOMEM) { 174 fmd_modhash_tryapply(fmd.d_mod_hash, fmd_module_trygc); 175 fmd_scheme_hash_trygc(fmd.d_schemes); 176 sysev_stats.eagain.fmds_value.ui64++; 177 rc = EAGAIN; 178 } else { 179 fmd_hdl_error(sysev_hdl, "discarding event 0x%llx: " 180 "missing or invalid payload", seq); 181 sysev_stats.bad_attr.fmds_value.ui64++; 182 } 183 } else { 184 sysevent_get_time(sep, &hrt); 185 fmd_xprt_post(sysev_hdl, xp, nvl, hrt); 186 } 187 188 (void) pthread_mutex_lock(&sysev_mutex); 189 if (--sysev_xprt_refcnt == 0 && sysev_exiting == 1) 190 (void) pthread_cond_broadcast(&sysev_cv); 191 (void) pthread_mutex_unlock(&sysev_mutex); 192 193 return (rc); 194 } 195 196 /* 197 * Checksum algorithm used by the dump transport for verifying the content of 198 * error reports saved on the dump device (copy of the kernel's checksum32()). 199 */ 200 static uint32_t 201 sysev_checksum(void *cp_arg, size_t length) 202 { 203 uchar_t *cp, *ep; 204 uint32_t sum = 0; 205 206 for (cp = cp_arg, ep = cp + length; cp < ep; cp++) 207 sum = ((sum >> 1) | (sum << 31)) + *cp; 208 209 return (sum); 210 } 211 212 /* 213 * Replay saved events from the dump transport. This function is installed as 214 * the timer callback and is called only once during the module's lifetime. 215 */ 216 /*ARGSUSED*/ 217 static void 218 sysev_replay(fmd_hdl_t *hdl, id_t id, void *arg) 219 { 220 char *dumpdev; 221 off64_t off, off0; 222 int fd, err; 223 224 /* 225 * Determine the appropriate dump device to use for replaying pending 226 * error reports. If the device property is NULL (default), we 227 * open and query /dev/dump to determine the current dump device. 228 */ 229 if ((dumpdev = sysev_device) == NULL) { 230 if ((fd = open("/dev/dump", O_RDONLY)) == -1) { 231 fmd_hdl_error(hdl, "failed to open /dev/dump " 232 "to locate dump device for event replay"); 233 goto done; 234 } 235 236 dumpdev = alloca(PATH_MAX); 237 err = ioctl(fd, DIOCGETDEV, dumpdev); 238 (void) close(fd); 239 240 if (err == -1) { 241 if (errno != ENODEV) { 242 fmd_hdl_error(hdl, "failed to obtain " 243 "path to dump device for event replay"); 244 } 245 goto done; 246 } 247 } 248 249 if (strcmp(dumpdev, "/dev/null") == 0) 250 goto done; /* return silently and skip replay for /dev/null */ 251 252 /* 253 * Open the appropriate device and then determine the offset of the 254 * start of the ereport dump region located at the end of the device. 255 */ 256 if ((fd = open64(dumpdev, O_RDWR | O_DSYNC)) == -1) { 257 fmd_hdl_error(hdl, "failed to open dump transport %s " 258 "(pending events will not be replayed)", dumpdev); 259 goto done; 260 } 261 262 off = DUMP_OFFSET + DUMP_LOGSIZE + DUMP_ERPTSIZE; 263 off = off0 = lseek64(fd, -off, SEEK_END) & -DUMP_OFFSET; 264 265 if (off == (off64_t)-1LL) { 266 fmd_hdl_error(hdl, "failed to seek dump transport %s " 267 "(pending events will not be replayed)", dumpdev); 268 (void) close(fd); 269 goto done; 270 } 271 272 /* 273 * The ereport dump region is a sequence of erpt_dump_t headers each of 274 * which is followed by packed nvlist data. We iterate over them in 275 * order, unpacking and dispatching each one to our dispatch queue. 276 */ 277 for (;;) { 278 char nvbuf[ERPT_DATA_SZ]; 279 uint32_t chksum; 280 erpt_dump_t ed; 281 nvlist_t *nvl; 282 283 fmd_timeval_t ftv, tod; 284 hrtime_t hrt; 285 uint64_t ena; 286 287 if (pread64(fd, &ed, sizeof (ed), off) != sizeof (ed)) { 288 fmd_hdl_error(hdl, "failed to read from dump " 289 "transport %s (pending events lost)", dumpdev); 290 break; 291 } 292 293 if (ed.ed_magic == 0 && ed.ed_size == 0) 294 break; /* end of list: all zero */ 295 296 if (ed.ed_magic == 0) { 297 off += sizeof (ed) + ed.ed_size; 298 continue; /* continue searching */ 299 } 300 301 if (ed.ed_magic != ERPT_MAGIC) { 302 /* 303 * Stop reading silently if the first record has the 304 * wrong magic number; this likely indicates that we 305 * rebooted from non-FMA bits or paged over the dump. 306 */ 307 if (off == off0) 308 break; 309 310 fmd_hdl_error(hdl, "invalid dump transport " 311 "record at %llx (magic number %x, expected %x)\n", 312 (u_longlong_t)off, ed.ed_magic, ERPT_MAGIC); 313 break; 314 } 315 316 if (ed.ed_size > ERPT_DATA_SZ) { 317 fmd_hdl_error(hdl, "invalid dump transport " 318 "record at %llx size (%u exceeds limit)\n", 319 (u_longlong_t)off, ed.ed_size); 320 break; 321 } 322 323 if (pread64(fd, nvbuf, ed.ed_size, 324 off + sizeof (ed)) != ed.ed_size) { 325 fmd_hdl_error(hdl, "failed to read dump " 326 "transport event (offset %llx)", (u_longlong_t)off); 327 328 sysev_stats.dump_lost.fmds_value.ui64++; 329 goto next; 330 } 331 332 if ((chksum = sysev_checksum(nvbuf, 333 ed.ed_size)) != ed.ed_chksum) { 334 fmd_hdl_error(hdl, "dump transport event at " 335 "offset %llx is corrupt (checksum %x != %x)\n", 336 (u_longlong_t)off, chksum, ed.ed_chksum); 337 338 sysev_stats.dump_lost.fmds_value.ui64++; 339 goto next; 340 } 341 342 if ((err = nvlist_xunpack(nvbuf, 343 ed.ed_size, &nvl, &fmd.d_nva)) != 0) { 344 fmd_hdl_error(hdl, "failed to unpack dump " 345 "transport event at offset %llx: %s\n", 346 (u_longlong_t)off, fmd_strerror(err)); 347 348 sysev_stats.dump_lost.fmds_value.ui64++; 349 goto next; 350 } 351 352 /* 353 * If ed_hrt_nsec is set it contains the gethrtime() value from 354 * when the event was originally enqueued for the transport. 355 * If it is zero, we use the weaker bound ed_hrt_base instead. 356 */ 357 if (ed.ed_hrt_nsec != 0) 358 hrt = ed.ed_hrt_nsec; 359 else 360 hrt = ed.ed_hrt_base; 361 362 /* 363 * If this is an FMA protocol event of class "ereport.*" that 364 * contains valid ENA, we can improve the precision of 'hrt'. 365 */ 366 if (nvlist_lookup_uint64(nvl, FM_EREPORT_ENA, &ena) == 0) 367 hrt = fmd_time_ena2hrt(hrt, ena); 368 369 /* 370 * Now convert 'hrt' to an adjustable TOD based on the values 371 * in ed_tod_base which correspond to one another and are 372 * sampled before reboot using the old gethrtime() clock. 373 * fmd_event_recreate() will use this TOD value to re-assign 374 * the event an updated gethrtime() value based on the current 375 * value of the non-adjustable gethrtime() clock. Phew. 376 */ 377 tod.ftv_sec = ed.ed_tod_base.sec; 378 tod.ftv_nsec = ed.ed_tod_base.nsec; 379 fmd_time_hrt2tod(ed.ed_hrt_base, &tod, hrt, &ftv); 380 381 (void) nvlist_remove_all(nvl, FMD_EVN_TOD); 382 (void) nvlist_add_uint64_array(nvl, 383 FMD_EVN_TOD, (uint64_t *)&ftv, 2); 384 385 fmd_xprt_post(hdl, sysev_xprt, nvl, 0); 386 sysev_stats.dump_replay.fmds_value.ui64++; 387 388 next: 389 /* 390 * Reset the magic number for the event record to zero so that 391 * we do not replay the same event multiple times. 392 */ 393 ed.ed_magic = 0; 394 395 if (pwrite64(fd, &ed, sizeof (ed), off) != sizeof (ed)) { 396 fmd_hdl_error(hdl, "failed to mark dump " 397 "transport event (offset %llx)", (u_longlong_t)off); 398 } 399 400 off += sizeof (ed) + ed.ed_size; 401 } 402 403 (void) close(fd); 404 done: 405 (void) pthread_mutex_lock(&sysev_mutex); 406 sysev_replay_wait = 0; 407 (void) pthread_cond_broadcast(&sysev_cv); 408 (void) pthread_mutex_unlock(&sysev_mutex); 409 } 410 411 static const fmd_prop_t sysev_props[] = { 412 { "class", FMD_TYPE_STRING, EC_ALL }, /* event class */ 413 { "device", FMD_TYPE_STRING, NULL }, /* replay device */ 414 { "channel", FMD_TYPE_STRING, FM_ERROR_CHAN }, /* channel name */ 415 { "sid", FMD_TYPE_STRING, "fmd" }, /* subscriber id */ 416 { NULL, 0, NULL } 417 }; 418 419 static const fmd_hdl_ops_t sysev_ops = { 420 NULL, /* fmdo_recv */ 421 sysev_replay, /* fmdo_timeout */ 422 NULL, /* fmdo_close */ 423 NULL, /* fmdo_stats */ 424 NULL, /* fmdo_gc */ 425 NULL, /* fmdo_send */ 426 }; 427 428 static const fmd_hdl_info_t sysev_info = { 429 "SysEvent Transport Agent", "1.0", &sysev_ops, sysev_props 430 }; 431 432 /* 433 * Bind to the sysevent channel we use for listening for error events and then 434 * subscribe to appropriate events received over this channel. Setup the 435 * legacy sysevent handler for creating sysevent resources and forwarding DR 436 * events. 437 */ 438 void 439 sysev_init(fmd_hdl_t *hdl) 440 { 441 uint_t flags; 442 const char *subclasses[] = { EC_SUB_ALL }; 443 444 if (fmd_hdl_register(hdl, FMD_API_VERSION, &sysev_info) != 0) 445 return; /* invalid property settings */ 446 447 (void) fmd_stat_create(hdl, FMD_STAT_NOALLOC, sizeof (sysev_stats) / 448 sizeof (fmd_stat_t), (fmd_stat_t *)&sysev_stats); 449 450 sysev_channel = fmd_prop_get_string(hdl, "channel"); 451 sysev_class = fmd_prop_get_string(hdl, "class"); 452 sysev_device = fmd_prop_get_string(hdl, "device"); 453 sysev_sid = fmd_prop_get_string(hdl, "sid"); 454 455 if (sysev_channel == NULL) 456 fmd_hdl_abort(hdl, "channel property must be defined\n"); 457 458 if (sysev_sid == NULL) 459 fmd_hdl_abort(hdl, "sid property must be defined\n"); 460 461 if ((errno = sysevent_evc_bind(sysev_channel, &sysev_evc, 462 EVCH_CREAT | EVCH_HOLD_PEND)) != 0) { 463 fmd_hdl_abort(hdl, "failed to bind to event transport " 464 "channel %s", sysev_channel); 465 } 466 467 sysev_xprt = fmd_xprt_open(hdl, FMD_XPRT_RDONLY, NULL, NULL); 468 sysev_hdl = hdl; 469 470 /* 471 * If we're subscribing to the default channel, keep our subscription 472 * active even if we die unexpectedly so we continue queuing events. 473 * If we're not (e.g. running under fmsim), do not specify SUB_KEEP so 474 * that our event channel will be destroyed if we die unpleasantly. 475 */ 476 if (strcmp(sysev_channel, FM_ERROR_CHAN) == 0) 477 flags = EVCH_SUB_KEEP | EVCH_SUB_DUMP; 478 else 479 flags = EVCH_SUB_DUMP; 480 481 errno = sysevent_evc_subscribe(sysev_evc, 482 sysev_sid, sysev_class, sysev_recv, sysev_xprt, flags); 483 484 if (errno != 0) { 485 if (errno == EEXIST) { 486 fmd_hdl_abort(hdl, "another fault management daemon is " 487 "active on transport channel %s\n", sysev_channel); 488 } else { 489 fmd_hdl_abort(hdl, "failed to subscribe to %s on " 490 "transport channel %s", sysev_class, sysev_channel); 491 } 492 } 493 494 /* 495 * Once the transport is open, install a single timer to fire at once 496 * in the context of the module's thread to run sysev_replay(). This 497 * thread will block in its first fmd_xprt_post() until fmd is ready. 498 */ 499 fmd_hdl_debug(hdl, "transport '%s' open\n", sysev_channel); 500 (void) fmd_timer_install(hdl, NULL, NULL, 0); 501 502 /* 503 * Open the legacy sysevent handle and subscribe to all events. These 504 * are automatically converted to "resource.sysevent.*" events so that 505 * modules can manage these events without additional infrastructure. 506 */ 507 if (geteuid() != 0) 508 return; 509 510 if ((fmd.d_sysev_hdl = 511 sysevent_bind_handle(sysev_legacy)) == NULL) 512 fmd_hdl_abort(hdl, "failed to bind to legacy sysevent channel"); 513 514 if (sysevent_subscribe_event(fmd.d_sysev_hdl, EC_ALL, 515 subclasses, 1) != 0) 516 fmd_hdl_abort(hdl, "failed to subscribe to legacy sysevents"); 517 } 518 519 /* 520 * Close the channel by unsubscribing and unbinding. We only do this when a 521 * a non-default channel has been selected. If we're using FM_ERROR_CHAN, 522 * the system default, we do *not* want to unsubscribe because the kernel will 523 * remove the subscriber queue and any events published in our absence will 524 * therefore be lost. This scenario may occur when, for example, fmd is sent 525 * a SIGTERM by init(1M) during reboot but an error is detected and makes it 526 * into the sysevent channel queue before init(1M) manages to call uadmin(2). 527 */ 528 void 529 sysev_fini(fmd_hdl_t *hdl) 530 { 531 if (strcmp(sysev_channel, FM_ERROR_CHAN) != 0) { 532 sysevent_evc_unsubscribe(sysev_evc, sysev_sid); 533 sysevent_evc_unbind(sysev_evc); 534 } 535 536 if (fmd.d_sysev_hdl != NULL) 537 sysevent_unbind_handle(fmd.d_sysev_hdl); 538 539 if (sysev_xprt != NULL) { 540 /* 541 * Wait callback returns before destroy the transport. 542 */ 543 (void) pthread_mutex_lock(&sysev_mutex); 544 sysev_exiting = 1; 545 while (sysev_xprt_refcnt > 0) 546 (void) pthread_cond_wait(&sysev_cv, &sysev_mutex); 547 (void) pthread_mutex_unlock(&sysev_mutex); 548 fmd_xprt_close(hdl, sysev_xprt); 549 } 550 551 fmd_prop_free_string(hdl, sysev_class); 552 fmd_prop_free_string(hdl, sysev_channel); 553 fmd_prop_free_string(hdl, sysev_device); 554 fmd_prop_free_string(hdl, sysev_sid); 555 } 556