NN_C99_INLINE void thread_state_awake (struct thread_state1 *ts1)
{
vtime_t vt = ts1->vtime;
vtime_t wd = ts1->watchdog;
if (vtime_asleep_p (vt))
ts1->vtime = vt + 1;
else
{
pa_membar_exit ();
ts1->vtime = vt + 2;
}
pa_membar_enter ();
if ( wd % 2 ){
ts1->watchdog = wd + 1;
} else {
ts1->watchdog = wd + 2;
}
}
static void *lease_renewal_thread (struct nn_servicelease *sl)
{
/* Do not check more often than once every 100ms (no particular
reason why it has to be 100ms), regardless of the lease settings.
Note: can't trust sl->self, may have been scheduled before the
assignment. */
const os_int64 min_progress_check_intv = 100 * T_MILLISECOND;
struct thread_state1 *self = lookup_thread_state ();
nn_mtime_t next_thread_cputime = { 0 };
nn_mtime_t tlast = { 0 };
int was_alive = 1;
unsigned i;
for (i = 0; i < thread_states.nthreads; i++)
{
sl->av_ary[i].alive = 1;
sl->av_ary[i].wd = thread_states.ts[i].watchdog - 1;
}
os_mutexLock (&sl->lock);
while (sl->keepgoing)
{
unsigned n_alive = 0;
nn_mtime_t tnow = now_mt ();
LOG_THREAD_CPUTIME (next_thread_cputime);
TRACE (("servicelease: tnow %"PA_PRId64":", tnow.v));
/* Check progress only if enough time has passed: there is no
guarantee that os_cond_timedwait wont ever return early, and we
do want to avoid spurious warnings. */
if (tnow.v < tlast.v + min_progress_check_intv)
{
n_alive = thread_states.nthreads;
}
else
{
tlast = tnow;
for (i = 0; i < thread_states.nthreads; i++)
{
if (thread_states.ts[i].state != THREAD_STATE_ALIVE)
n_alive++;
else
{
vtime_t vt = thread_states.ts[i].vtime;
vtime_t wd = thread_states.ts[i].watchdog;
int alive = vtime_asleep_p (vt) || vtime_asleep_p (wd) || vtime_gt (wd, sl->av_ary[i].wd);
n_alive += (unsigned) alive;
TRACE ((" %d(%s):%c:%u:%u->%u:", i, thread_states.ts[i].name, alive ? 'a' : 'd', vt, sl->av_ary[i].wd, wd));
sl->av_ary[i].wd = wd;
if (sl->av_ary[i].alive != alive)
{
const char *name = thread_states.ts[i].name;
const char *msg;
if (!alive)
msg = "failed to make progress";
else
msg = "once again made progress";
NN_WARNING2 ("thread %s %s\n", name ? name : "(anon)", msg);
sl->av_ary[i].alive = (char) alive;
}
}
}
}
/* Only renew the lease if all threads are alive, so that one
thread blocking for a while but not too extremely long will
cause warnings for that thread in the log file, but won't cause
the DDSI2 service to be marked as dead. */
if (n_alive == thread_states.nthreads)
{
TRACE ((": [%d] renewing\n", n_alive));
/* FIXME: perhaps it would be nice to control automatic
liveliness updates from here.
FIXME: should terminate failure of renew_cb() */
sl->renew_cb (sl->renew_arg);
was_alive = 1;
}
else
{
TRACE ((": [%d] NOT renewing\n", n_alive));
if (was_alive)
log_stack_traces ();
was_alive = 0;
}
#if SYSDEPS_HAVE_GETRUSAGE
/* If getrusage() is available, use it to log CPU and memory
statistics to the trace. Getrusage() can't fail if the
parameters are valid, and these are by the book. Still we
check. */
if (config.enabled_logcats & LC_TIMING)
{
struct rusage u;
if (getrusage (RUSAGE_SELF, &u) == 0)
{
nn_log (LC_TIMING,
"rusage: utime %d.%06d stime %d.%06d maxrss %ld data %ld vcsw %ld ivcsw %ld\n",
(int) u.ru_utime.tv_sec, (int) u.ru_utime.tv_usec,
(int) u.ru_stime.tv_sec, (int) u.ru_stime.tv_usec,
u.ru_maxrss, u.ru_idrss, u.ru_nvcsw, u.ru_nivcsw);
}
}
#endif
os_condTimedWait (&sl->cond, &sl->lock, sl->sleepTime);
/* We are never active in a way that matters for the garbage
collection of old writers, &c. */
thread_state_asleep (self);
}
os_mutexUnlock (&sl->lock);
return NULL;
}
