static void
fmd_ckpt_restore_events(fmd_ckpt_t *ckp, fcf_secidx_t sid,
void (*func)(void *, fmd_event_t *), void *arg)
{
const fcf_event_t *fcfe;
const fcf_sec_t *sp;
fmd_timeval_t ftv;
fmd_log_t *lp, *errlp;
uint_t i, n;
uint32_t e_maj, e_min;
uint64_t e_ino;
if ((sp = fmd_ckpt_secptr(ckp, sid, FCF_SECT_EVENTS)) == NULL) {
fmd_ckpt_error(ckp, EFMD_CKPT_INVAL,
"invalid link to section %u: expected events\n", sid);
}
if (sp->fcfs_size == 0)
return; /* empty events section or type none */
fcfe = fmd_ckpt_dataptr(ckp, sp);
n = sp->fcfs_size / sp->fcfs_entsize;
/*
* Hold the reader lock on log pointers to block log rotation during
* the section restore so that we can safely insert refs to d_errlog.
*/
(void) pthread_rwlock_rdlock(&fmd.d_log_lock);
errlp = fmd.d_errlog;
e_maj = major(errlp->log_stat.st_dev);
e_min = minor(errlp->log_stat.st_dev);
e_ino = errlp->log_stat.st_ino;
for (i = 0; i < n; i++) {
fmd_event_t *ep;
ftv.ftv_sec = fcfe->fcfe_todsec;
ftv.ftv_nsec = fcfe->fcfe_todnsec;
if (e_ino == fcfe->fcfe_inode &&
e_maj == fcfe->fcfe_major &&
e_min == fcfe->fcfe_minor)
lp = errlp;
else
lp = NULL;
ep = fmd_event_recreate(FMD_EVT_PROTOCOL,
&ftv, NULL, NULL, lp, fcfe->fcfe_offset, 0);
fmd_event_hold(ep);
func(arg, ep);
fmd_event_rele(ep);
fcfe = (fcf_event_t *)((uintptr_t)fcfe + sp->fcfs_entsize);
}
(void) pthread_rwlock_unlock(&fmd.d_log_lock);
}
void
fmd_run(fmd_t *dp, int pfd)
{
char *nodc_key[] = { FMD_FLT_NODC, NULL };
char nodc_str[128];
struct sigaction act;
int status = FMD_EXIT_SUCCESS;
const char *name;
fmd_conf_path_t *pap;
fmd_event_t *e;
int dbout, err;
/*
* Cache all the current debug property settings in d_fmd_debug,
* d_fmd_dbout, d_hdl_debug, and d_hdl_dbout. If a given debug mask
* is non-zero and the corresponding dbout mask is zero, set dbout
* to a sensible default value based on whether we have daemonized.
*/
(void) fmd_conf_getprop(dp->d_conf, "dbout", &dbout);
if (dp->d_fmd_debug != 0 && dbout == 0)
dp->d_fmd_dbout = dp->d_fg? FMD_DBOUT_STDERR : FMD_DBOUT_SYSLOG;
else
dp->d_fmd_dbout = dbout;
(void) fmd_conf_getprop(dp->d_conf, "client.debug", &dp->d_hdl_debug);
(void) fmd_conf_getprop(dp->d_conf, "client.dbout", &dbout);
if (dp->d_hdl_debug != 0 && dbout == 0)
dp->d_hdl_dbout = dp->d_fg? FMD_DBOUT_STDERR : FMD_DBOUT_SYSLOG;
else
dp->d_hdl_dbout = dbout;
/*
* Initialize remaining major program data structures such as the
* clock, dispatch queues, log files, module hash collections, etc.
* This work is done here rather than in fmd_create() to permit the -o
* command-line option to modify properties after fmd_create() is done.
*/
name = dp->d_rootdir != NULL &&
*dp->d_rootdir != '\0' ? dp->d_rootdir : NULL;
if ((dp->d_topo = topo_open(TOPO_VERSION, name, &err)) == NULL) {
fmd_error(EFMD_EXIT, "failed to initialize "
"topology library: %s\n", topo_strerror(err));
}
dp->d_clockptr = dp->d_clockops->fto_init();
dp->d_xprt_ids = fmd_idspace_create("xprt_ids", 1, INT_MAX);
fmd_xprt_suspend_all();
(void) door_server_create(fmd_door);
fmd_dr_init();
dp->d_rmod->mod_timerids = fmd_idspace_create(dp->d_pname, 1, 16);
dp->d_timers = fmd_timerq_create();
dp->d_disp = fmd_dispq_create();
dp->d_cases = fmd_case_hash_create();
/*
* The root module's mod_queue is created with limit zero, making it
* act like /dev/null; anything inserted here is simply ignored.
*/
dp->d_rmod->mod_queue = fmd_eventq_create(dp->d_rmod,
&dp->d_rmod->mod_stats->ms_evqstat, &dp->d_rmod->mod_stats_lock, 0);
/*
* Once our subsystems that use signals have been set up, install the
* signal handler for the fmd_thr_signal() API. Verify that the signal
* being used for this purpose doesn't conflict with something else.
*/
(void) fmd_conf_getprop(dp->d_conf, "client.thrsig", &dp->d_thr_sig);
if (sigaction(dp->d_thr_sig, NULL, &act) != 0) {
fmd_error(EFMD_EXIT, "invalid signal selected for "
"client.thrsig property: %d\n", dp->d_thr_sig);
}
if (act.sa_handler != SIG_IGN && act.sa_handler != SIG_DFL) {
fmd_error(EFMD_EXIT, "signal selected for client.thrsig "
"property is already in use: %d\n", dp->d_thr_sig);
}
act.sa_handler = fmd_signal;
act.sa_flags = 0;
(void) sigemptyset(&act.sa_mask);
(void) sigaction(dp->d_thr_sig, &act, NULL);
(void) fmd_conf_getprop(dp->d_conf, "schemedir", &name);
dp->d_schemes = fmd_scheme_hash_create(dp->d_rootdir, name);
(void) fmd_conf_getprop(dp->d_conf, "log.rsrc", &name);
dp->d_asrus = fmd_asru_hash_create(dp->d_rootdir, name);
(void) fmd_conf_getprop(dp->d_conf, "log.error", &name);
dp->d_errlog = fmd_log_open(dp->d_rootdir, name, FMD_LOG_ERROR);
(void) fmd_conf_getprop(dp->d_conf, "log.fault", &name);
dp->d_fltlog = fmd_log_open(dp->d_rootdir, name, FMD_LOG_FAULT);
if (dp->d_asrus == NULL || dp->d_errlog == NULL || dp->d_fltlog == NULL)
fmd_error(EFMD_EXIT, "failed to initialize log files\n");
/*
* Before loading modules, create an empty control event which will act
* as a global barrier for module event processing. Each module we
* load successfully will insert it at their head of their event queue,
* and then pause inside of fmd_ctl_rele() after dequeuing the event.
* This module barrier is required for two reasons:
*
* (a) During module loading, the restoration of case checkpoints may
* result in a list.* event being recreated for which the intended
* subscriber has not yet loaded depending on the load order. Such
* events could then result in spurious "no subscriber" errors.
*
* (b) During errlog replay, a sequence of errors from a long time ago
* may be replayed, and the module may attempt to install relative
* timers associated with one or more of these events. If errlog
* replay were "racing" with active module threads, an event E1
* that resulted in a relative timer T at time E1 + N nsec could
* fire prior to an event E2 being enqueued, even if the relative
* time ordering was E1 < E2 < E1 + N, causing mis-diagnosis.
*/
dp->d_mod_event = e = fmd_event_create(FMD_EVT_CTL,
FMD_HRT_NOW, NULL, fmd_ctl_init(NULL));
fmd_event_hold(e);
/*
* Once all data structures are initialized, we load all of our modules
* in order according to class in order to load up any subscriptions.
* Once built-in modules are loaded, we detach from our waiting parent.
*/
dp->d_mod_hash = fmd_modhash_create();
if (fmd_builtin_loadall(dp->d_mod_hash) != 0 && !dp->d_fg)
fmd_error(EFMD_EXIT, "failed to initialize fault manager\n");
(void) fmd_conf_getprop(dp->d_conf, "self.name", &name);
dp->d_self = fmd_modhash_lookup(dp->d_mod_hash, name);
if (dp->d_self != NULL && fmd_module_dc_key2code(dp->d_self,
nodc_key, nodc_str, sizeof (nodc_str)) == 0)
(void) fmd_conf_setprop(dp->d_conf, "nodiagcode", nodc_str);
fmd_rpc_init();
dp->d_running = 1; /* we are now officially an active fmd */
/*
* Now that we're running, if a pipe fd was specified, write an exit
* status to it to indicate that our parent process can safely detach.
* Then proceed to loading the remaining non-built-in modules.
*/
if (pfd >= 0)
(void) write(pfd, &status, sizeof (status));
/*
* Before loading all modules, repopulate the ASRU cache from its
* persistent repository on disk. Then during module loading, the
* restoration of checkpoint files will reparent any active cases.
*/
fmd_asru_hash_refresh(dp->d_asrus);
(void) fmd_conf_getprop(dp->d_conf, "plugin.path", &pap);
fmd_modhash_loadall(dp->d_mod_hash, pap, &fmd_rtld_ops, ".so");
(void) fmd_conf_getprop(dp->d_conf, "agent.path", &pap);
fmd_modhash_loadall(dp->d_mod_hash, pap, &fmd_proc_ops, NULL);
/*
* With all modules loaded, replay fault events from the ASRU cache for
* any ASRUs that must be retired, replay error events from the errlog
* that did not finish processing the last time ran, and then release
* the global module barrier by executing a final rele on d_mod_event.
*/
fmd_asru_hash_replay(dp->d_asrus);
(void) pthread_rwlock_rdlock(&dp->d_log_lock);
fmd_log_replay(dp->d_errlog, (fmd_log_f *)fmd_err_replay, dp);
fmd_log_update(dp->d_errlog);
(void) pthread_rwlock_unlock(&dp->d_log_lock);
dp->d_mod_event = NULL;
fmd_event_rele(e);
/*
* Finally, awaken any threads associated with receiving events from
* open transports and tell them to proceed with fmd_xprt_recv().
*/
fmd_xprt_resume_all();
fmd_gc(dp, 0, 0);
dp->d_booted = 1;
}
