| Message ID | 20170822015439.11263-1-npiggin@gmail.com (mailing list archive) |
|---|---|
| State | Not Applicable |
| Headers | show |
On Tue, 22 Aug 2017, Nicholas Piggin wrote: > I would have preferred to get comments from the timer maintainers, but > they've been busy or away for the past copule of weeks. Perhaps you > would consider carrying it until then? Yes, I was on vacation, but that patch never hit my LKML inbox .... > diff --git a/kernel/time/timer.c b/kernel/time/timer.c > index 8f5d1bf18854..dd7be9fe6839 100644 > --- a/kernel/time/timer.c > +++ b/kernel/time/timer.c > @@ -203,6 +203,7 @@ struct timer_base { > bool migration_enabled; > bool nohz_active; > bool is_idle; > + bool was_idle; /* was it idle since last run/fwded */ Please do not use tail comments. They are a horrible habit and disturb the reading flow. I'm not fond of that name either. Something like forward_clk or a similar self explaining name would be appreciated. > /* > @@ -938,6 +945,13 @@ __mod_timer(struct timer_list *timer, unsigned long expires, bool pending_only) > * same array bucket then just return: > */ > if (timer_pending(timer)) { > + /* > + * The downside of this optimization is that it can result in > + * larger granularity than you would get from adding a new > + * timer with this expiry. Would a timer flag for networking > + * be appropriate, then we can try to keep expiry of general > + * timers within ~1/8th of their interval? This really depends on the timer usage type. If you use it merily for TCP timeouts or similar things, then this does not matter at all because then the timer is the safety net in case something goes wrong. If you lose packets on the transport the larger granularity is the least of your worries. From earlier discussions with the networking folks these timeouts are the reason for this optimization because you avoid the expensive dequeue/requeue operation if the successful communication is fast. If the timer has a short relative expiry and is used for sending packets or similar purposes, then it usually sits in the first bucket and has no granularity issues anyway. But from staring at trace data provided by google and facebook these timer are not rearmed while pending, they fire, do networking stuff and then get rearmed. So I rather avoid that kind of misleading comment. The first sentence surely can stay. Other than that, nice detective work! Thanks, tglx
On Tue, 22 Aug 2017 09:45:46 +0200 (CEST) Thomas Gleixner <tglx@linutronix.de> wrote: > On Tue, 22 Aug 2017, Nicholas Piggin wrote: > > I would have preferred to get comments from the timer maintainers, but > > they've been busy or away for the past copule of weeks. Perhaps you > > would consider carrying it until then? > > Yes, I was on vacation, but that patch never hit my LKML inbox .... Hmm okay. Well that's fine, we're just getting done testing it here. > > > diff --git a/kernel/time/timer.c b/kernel/time/timer.c > > index 8f5d1bf18854..dd7be9fe6839 100644 > > --- a/kernel/time/timer.c > > +++ b/kernel/time/timer.c > > @@ -203,6 +203,7 @@ struct timer_base { > > bool migration_enabled; > > bool nohz_active; > > bool is_idle; > > + bool was_idle; /* was it idle since last run/fwded */ > > Please do not use tail comments. They are a horrible habit and disturb the > reading flow. Sure. > I'm not fond of that name either. Something like forward_clk > or a similar self explaining name would be appreciated. Well I actually had that initially (must_forward). But on the other hand that's less explanatory about the state of the timer base I thought. Anyway I could go either way and you're going to be looking at this code more than me in future :) > > > /* > > @@ -938,6 +945,13 @@ __mod_timer(struct timer_list *timer, unsigned long expires, bool pending_only) > > * same array bucket then just return: > > */ > > if (timer_pending(timer)) { > > + /* > > + * The downside of this optimization is that it can result in > > + * larger granularity than you would get from adding a new > > + * timer with this expiry. Would a timer flag for networking > > + * be appropriate, then we can try to keep expiry of general > > + * timers within ~1/8th of their interval? > > This really depends on the timer usage type. > > If you use it merily for TCP timeouts or similar things, then this does not > matter at all because then the timer is the safety net in case something > goes wrong. If you lose packets on the transport the larger granularity is > the least of your worries. > > From earlier discussions with the networking folks these timeouts are the > reason for this optimization because you avoid the expensive > dequeue/requeue operation if the successful communication is fast. > > If the timer has a short relative expiry and is used for sending packets or > similar purposes, then it usually sits in the first bucket and has no > granularity issues anyway. But from staring at trace data provided by > google and facebook these timer are not rearmed while pending, they fire, > do networking stuff and then get rearmed. > > So I rather avoid that kind of misleading comment. The first sentence > surely can stay. Right. The question was actually there for you to answer, so thanks. I can certainly understand the use-case and importance of keeping it light. > Other than that, nice detective work! Thanks for taking a look (and thanks everyone who's been testing and hacking on it). Do you want me to re-send with the changes? Thanks, Nick
diff --git a/kernel/time/timer.c b/kernel/time/timer.c index 8f5d1bf18854..dd7be9fe6839 100644 --- a/kernel/time/timer.c +++ b/kernel/time/timer.c @@ -203,6 +203,7 @@ struct timer_base { bool migration_enabled; bool nohz_active; bool is_idle; + bool was_idle; /* was it idle since last run/fwded */ DECLARE_BITMAP(pending_map, WHEEL_SIZE); struct hlist_head vectors[WHEEL_SIZE]; } ____cacheline_aligned; @@ -856,13 +857,19 @@ get_target_base(struct timer_base *base, unsigned tflags) static inline void forward_timer_base(struct timer_base *base) { - unsigned long jnow = READ_ONCE(jiffies); + unsigned long jnow; /* - * We only forward the base when it's idle and we have a delta between - * base clock and jiffies. + * We only forward the base when we are idle or have just come out + * of idle (was_idle logic), and have a delta between base clock + * and jiffies. In the common case, run_timers will take care of it. */ - if (!base->is_idle || (long) (jnow - base->clk) < 2) + if (likely(!base->was_idle)) + return; + + jnow = READ_ONCE(jiffies); + base->was_idle = base->is_idle; + if ((long)(jnow - base->clk) < 2) return; /* @@ -938,6 +945,13 @@ __mod_timer(struct timer_list *timer, unsigned long expires, bool pending_only) * same array bucket then just return: */ if (timer_pending(timer)) { + /* + * The downside of this optimization is that it can result in + * larger granularity than you would get from adding a new + * timer with this expiry. Would a timer flag for networking + * be appropriate, then we can try to keep expiry of general + * timers within ~1/8th of their interval? + */ if (timer->expires == expires) return 1; @@ -948,6 +962,7 @@ __mod_timer(struct timer_list *timer, unsigned long expires, bool pending_only) * dequeue/enqueue dance. */ base = lock_timer_base(timer, &flags); + forward_timer_base(base); clk = base->clk; idx = calc_wheel_index(expires, clk); @@ -964,6 +979,7 @@ __mod_timer(struct timer_list *timer, unsigned long expires, bool pending_only) } } else { base = lock_timer_base(timer, &flags); + forward_timer_base(base); } ret = detach_if_pending(timer, base, false); @@ -991,12 +1007,10 @@ __mod_timer(struct timer_list *timer, unsigned long expires, bool pending_only) raw_spin_lock(&base->lock); WRITE_ONCE(timer->flags, (timer->flags & ~TIMER_BASEMASK) | base->cpu); + forward_timer_base(base); } } - /* Try to forward a stale timer base clock */ - forward_timer_base(base); - timer->expires = expires; /* * If 'idx' was calculated above and the base time did not advance @@ -1112,6 +1126,7 @@ void add_timer_on(struct timer_list *timer, int cpu) WRITE_ONCE(timer->flags, (timer->flags & ~TIMER_BASEMASK) | cpu); } + forward_timer_base(base); debug_activate(timer, timer->expires); internal_add_timer(base, timer); @@ -1499,8 +1514,10 @@ u64 get_next_timer_interrupt(unsigned long basej, u64 basem) /* * If we expect to sleep more than a tick, mark the base idle: */ - if ((expires - basem) > TICK_NSEC) + if ((expires - basem) > TICK_NSEC) { + base->was_idle = true; base->is_idle = true; + } } raw_spin_unlock(&base->lock); @@ -1611,6 +1628,17 @@ static __latent_entropy void run_timer_softirq(struct softirq_action *h) { struct timer_base *base = this_cpu_ptr(&timer_bases[BASE_STD]); + /* + * was_idle must be cleared before running timers so that any timer + * functions that call mod_timer will not try to forward the base. + * + * The deferrable base does not do idle tracking at all, so we do + * not forward it. This can result in very large variations in + * granularity for deferrable timers, but they can be deferred for + * long periods due to idle. + */ + base->was_idle = false; + __run_timers(base); if (IS_ENABLED(CONFIG_NO_HZ_COMMON) && base->nohz_active) __run_timers(this_cpu_ptr(&timer_bases[BASE_DEF]));
When a timer base is idle, it is forwarded when a new timer is added to ensure that granularity does not become excessive. When not idle, the timer tick is expected to increment the base. However there are several problems: - If an existing timer is modified, the base is forwarded only after the index is calculated. - The base is not forwarded by add_timer_on. - There is a window after a timer is restarted from a nohz ide, after it is marked not-idle and before the timer tick on this CPU, where a timer may be added but the ancient base does not get forwarded. These result in excessive granularity (a 1 jiffy timeout can blow out to 100s of jiffies), which cause the rcu lockup detector to trigger, among other things. Fix this by keeping track of whether the timer base has been idle since it was last run or forwarded, and if so then forward it before adding a new timer. There is still a problem where the mod_timer optimization where it's modified with the same expiry time can result in excessive granularity relative to the new shorter interval. That is not addressed by this patch because checking base->was_idle would increase overhead and it's a rather special case (you could reason that the caller should not expect change in absolute expiry time due to such an operation). So that is noted as a comment. As well as fixing the visible RCU softlockup failures, I tested an idle system (with no lockup watchdogs running) and traced all non-deferrable timer expiries for 1000s, and analysed wakeup latency relative to requested latency. 1.0 means we slept for as many jiffies as requested, 2.0 means we slept 2x the time (this suffers jiffies round-up skew at low absolute times): max avg std upstream 506.0 1.20 4.68 patched 2.0 1.08 0.15 This was noticed due to the lockup detector Kconfig changes dropping it out of people's .configs. When the lockup detectors are enabled, no CPU can go idle for longer than 4 seconds, which limits the granularity errors. Sub-optimal timer behaviour is observable on a smaller scale: max avg std upstream 9.0 1.05 0.19 patched 2.0 1.04 0.11 Tested-by: David Miller <davem@davemloft.net> Signed-off-by: Nicholas Piggin <npiggin@gmail.com> --- Hi Andrew, I would have preferred to get comments from the timer maintainers, but they've been busy or away for the past copule of weeks. Perhaps you would consider carrying it until then? Thanks, Nick kernel/time/timer.c | 44 ++++++++++++++++++++++++++++++++++++-------- 1 file changed, 36 insertions(+), 8 deletions(-)