+ }
+ }
+
+ timer_queue_unlock(timer_master_queue);
+}
+
+void
+timer_longterm_callout(timer_call_param_t p0, __unused timer_call_param_t p1)
+{
+ timer_longterm_t *tlp = (timer_longterm_t *) p0;
+
+ timer_longterm_update(tlp);
+}
+
+void
+timer_longterm_update_locked(timer_longterm_t *tlp)
+{
+ uint64_t latency;
+
+ TIMER_KDEBUG_TRACE(KDEBUG_TRACE,
+ DECR_TIMER_UPDATE | DBG_FUNC_START,
+ VM_KERNEL_UNSLIDE_OR_PERM(&tlp->queue),
+ tlp->threshold.deadline,
+ tlp->threshold.preempted,
+ tlp->queue.count, 0);
+
+ tlp->scan_time = mach_absolute_time();
+ if (tlp->threshold.preempted != TIMER_LONGTERM_NONE) {
+ tlp->threshold.preempts++;
+ tlp->threshold.deadline = tlp->threshold.preempted;
+ tlp->threshold.preempted = TIMER_LONGTERM_NONE;
+ /*
+ * Note: in the unlikely event that a pre-empted timer has
+ * itself been cancelled, we'll simply re-scan later at the
+ * time of the preempted/cancelled timer.
+ */
+ } else {
+ tlp->threshold.scans++;
+
+ /*
+ * Maintain a moving average of our wakeup latency.
+ * Clamp latency to 0 and ignore above threshold interval.
+ */
+ if (tlp->scan_time > tlp->threshold.deadline_set)
+ latency = tlp->scan_time - tlp->threshold.deadline_set;
+ else
+ latency = 0;
+ if (latency < tlp->threshold.interval) {
+ tlp->threshold.latency_min =
+ MIN(tlp->threshold.latency_min, latency);
+ tlp->threshold.latency_max =
+ MAX(tlp->threshold.latency_max, latency);
+ tlp->threshold.latency =
+ (tlp->threshold.latency*99 + latency) / 100;
+ }
+
+ timer_longterm_scan(tlp, tlp->scan_time);
+ }
+
+ tlp->threshold.deadline_set = tlp->threshold.deadline;
+ /* The next deadline timer to be set is adjusted */
+ if (tlp->threshold.deadline != TIMER_LONGTERM_NONE &&
+ tlp->threshold.deadline != TIMER_LONGTERM_SCAN_AGAIN) {
+ tlp->threshold.deadline_set -= tlp->threshold.margin;
+ tlp->threshold.deadline_set -= tlp->threshold.latency;
+ }
+
+ /* Throttle next scan time */
+ uint64_t scan_clamp = mach_absolute_time() + tlp->scan_limit;
+ if (tlp->threshold.deadline_set < scan_clamp)
+ tlp->threshold.deadline_set = scan_clamp;
+
+ TIMER_KDEBUG_TRACE(KDEBUG_TRACE,
+ DECR_TIMER_UPDATE | DBG_FUNC_END,
+ VM_KERNEL_UNSLIDE_OR_PERM(&tlp->queue),
+ tlp->threshold.deadline,
+ tlp->threshold.scans,
+ tlp->queue.count, 0);
+}
+
+void
+timer_longterm_update(timer_longterm_t *tlp)
+{
+ spl_t s = splclock();
+
+ timer_queue_lock_spin(timer_longterm_queue);
+
+ if (cpu_number() != master_cpu)
+ panic("timer_longterm_update_master() on non-boot cpu");
+
+ timer_longterm_update_locked(tlp);
+
+ if (tlp->threshold.deadline != TIMER_LONGTERM_NONE)
+ timer_call_enter(
+ &tlp->threshold.timer,
+ tlp->threshold.deadline_set,
+ TIMER_CALL_LOCAL | TIMER_CALL_SYS_CRITICAL);
+
+ timer_queue_unlock(timer_longterm_queue);
+ splx(s);
+}
+
+void
+timer_longterm_init(void)
+{
+ uint32_t longterm;
+ timer_longterm_t *tlp = &timer_longterm;
+
+ DBG("timer_longterm_init() tlp: %p, queue: %p\n", tlp, &tlp->queue);
+
+ /*
+ * Set the longterm timer threshold. Defaults to TIMER_LONGTERM_THRESHOLD
+ * or TIMER_LONGTERM_NONE (disabled) for server;
+ * overridden longterm boot-arg
+ */
+ tlp->threshold.interval = serverperfmode ? TIMER_LONGTERM_NONE
+ : TIMER_LONGTERM_THRESHOLD;
+ if (PE_parse_boot_argn("longterm", &longterm, sizeof (longterm))) {
+ tlp->threshold.interval = (longterm == 0) ?
+ TIMER_LONGTERM_NONE :
+ longterm * NSEC_PER_MSEC;
+ }
+ if (tlp->threshold.interval != TIMER_LONGTERM_NONE) {
+ printf("Longterm timer threshold: %llu ms\n",
+ tlp->threshold.interval / NSEC_PER_MSEC);
+ kprintf("Longterm timer threshold: %llu ms\n",
+ tlp->threshold.interval / NSEC_PER_MSEC);
+ nanoseconds_to_absolutetime(tlp->threshold.interval,
+ &tlp->threshold.interval);
+ tlp->threshold.margin = tlp->threshold.interval / 10;
+ tlp->threshold.latency_min = EndOfAllTime;
+ tlp->threshold.latency_max = 0;
+ }
+
+ tlp->threshold.preempted = TIMER_LONGTERM_NONE;
+ tlp->threshold.deadline = TIMER_LONGTERM_NONE;
+
+ lck_attr_setdefault(&timer_longterm_lck_attr);
+ lck_grp_attr_setdefault(&timer_longterm_lck_grp_attr);
+ lck_grp_init(&timer_longterm_lck_grp,
+ "timer_longterm", &timer_longterm_lck_grp_attr);
+ mpqueue_init(&tlp->queue,
+ &timer_longterm_lck_grp, &timer_longterm_lck_attr);
+
+ timer_call_setup(&tlp->threshold.timer,
+ timer_longterm_callout, (timer_call_param_t) tlp);
+
+ timer_longterm_queue = &tlp->queue;
+}
+
+enum {
+ THRESHOLD, QCOUNT,
+ ENQUEUES, DEQUEUES, ESCALATES, SCANS, PREEMPTS,
+ LATENCY, LATENCY_MIN, LATENCY_MAX, SCAN_LIMIT, PAUSES
+};
+uint64_t
+timer_sysctl_get(int oid)
+{
+ timer_longterm_t *tlp = &timer_longterm;
+
+ switch (oid) {
+ case THRESHOLD:
+ return (tlp->threshold.interval == TIMER_LONGTERM_NONE) ?
+ 0 : tlp->threshold.interval / NSEC_PER_MSEC;
+ case QCOUNT:
+ return tlp->queue.count;
+ case ENQUEUES:
+ return tlp->enqueues;
+ case DEQUEUES:
+ return tlp->dequeues;
+ case ESCALATES:
+ return tlp->escalates;
+ case SCANS:
+ return tlp->threshold.scans;
+ case PREEMPTS:
+ return tlp->threshold.preempts;
+ case LATENCY:
+ return tlp->threshold.latency;
+ case LATENCY_MIN:
+ return tlp->threshold.latency_min;
+ case LATENCY_MAX:
+ return tlp->threshold.latency_max;
+ case SCAN_LIMIT:
+ return tlp->scan_limit;
+ case PAUSES:
+ return tlp->scan_pauses;
+ default:
+ return 0;
+ }
+}
+
+/*
+ * timer_master_scan() is the inverse of timer_longterm_scan()
+ * since it un-escalates timers to the longterm queue.
+ */
+static void
+timer_master_scan(timer_longterm_t *tlp,
+ uint64_t now)
+{
+ queue_entry_t qe;
+ timer_call_t call;
+ uint64_t threshold;
+ uint64_t deadline;
+ mpqueue_head_t *timer_master_queue;
+
+ if (tlp->threshold.interval != TIMER_LONGTERM_NONE)
+ threshold = now + tlp->threshold.interval;
+ else
+ threshold = TIMER_LONGTERM_NONE;
+
+ timer_master_queue = timer_queue_cpu(master_cpu);
+ timer_queue_lock_spin(timer_master_queue);
+
+ qe = queue_first(&timer_master_queue->head);
+ while (!queue_end(&timer_master_queue->head, qe)) {
+ call = TIMER_CALL(qe);
+ deadline = TCE(call)->deadline;
+ qe = queue_next(qe);
+ if ((call->flags & TIMER_CALL_LOCAL) != 0)
+ continue;
+ if (!simple_lock_try(&call->lock)) {
+ /* case (2c) lock order inversion, dequeue only */
+ timer_call_entry_dequeue_async(call);
+ continue;
+ }
+ if (deadline > threshold) {
+ /* move from master to longterm */
+ timer_call_entry_dequeue(call);
+ timer_call_entry_enqueue_tail(call, timer_longterm_queue);
+ if (deadline < tlp->threshold.deadline) {
+ tlp->threshold.deadline = deadline;
+ tlp->threshold.call = call;
+ }
+ }
+ simple_unlock(&call->lock);
+ }
+ timer_queue_unlock(timer_master_queue);
+}
+
+static void
+timer_sysctl_set_threshold(uint64_t value)
+{
+ timer_longterm_t *tlp = &timer_longterm;
+ spl_t s = splclock();
+ boolean_t threshold_increase;
+
+ timer_queue_lock_spin(timer_longterm_queue);
+
+ timer_call_cancel(&tlp->threshold.timer);
+
+ /*
+ * Set the new threshold and note whther it's increasing.
+ */
+ if (value == 0) {
+ tlp->threshold.interval = TIMER_LONGTERM_NONE;
+ threshold_increase = TRUE;
+ timer_call_cancel(&tlp->threshold.timer);
+ } else {
+ uint64_t old_interval = tlp->threshold.interval;
+ tlp->threshold.interval = value * NSEC_PER_MSEC;
+ nanoseconds_to_absolutetime(tlp->threshold.interval,
+ &tlp->threshold.interval);
+ tlp->threshold.margin = tlp->threshold.interval / 10;
+ if (old_interval == TIMER_LONGTERM_NONE)
+ threshold_increase = FALSE;
+ else
+ threshold_increase = (tlp->threshold.interval > old_interval);
+ }
+
+ if (threshold_increase /* or removal */) {
+ /* Escalate timers from the longterm queue */
+ timer_longterm_scan(tlp, mach_absolute_time());
+ } else /* decrease or addition */ {
+ /*
+ * We scan the local/master queue for timers now longterm.
+ * To be strictly correct, we should scan all processor queues
+ * but timer migration results in most timers gravitating to the
+ * master processor in any case.
+ */
+ timer_master_scan(tlp, mach_absolute_time());
+ }
+
+ /* Set new timer accordingly */
+ tlp->threshold.deadline_set = tlp->threshold.deadline;
+ if (tlp->threshold.deadline != TIMER_LONGTERM_NONE) {
+ tlp->threshold.deadline_set -= tlp->threshold.margin;
+ tlp->threshold.deadline_set -= tlp->threshold.latency;
+ timer_call_enter(
+ &tlp->threshold.timer,
+ tlp->threshold.deadline_set,
+ TIMER_CALL_LOCAL | TIMER_CALL_SYS_CRITICAL);
+ }
+
+ /* Reset stats */
+ tlp->enqueues = 0;
+ tlp->dequeues = 0;
+ tlp->escalates = 0;
+ tlp->scan_pauses = 0;
+ tlp->threshold.scans = 0;
+ tlp->threshold.preempts = 0;
+ tlp->threshold.latency = 0;
+ tlp->threshold.latency_min = EndOfAllTime;
+ tlp->threshold.latency_max = 0;
+
+ timer_queue_unlock(timer_longterm_queue);
+ splx(s);
+}
projects / apple / xnu.git / blobdiff