Message ID | 3a0ef8e5a838fcf1a3cdaecc029cc97dea5edf65.1384940804.git.viresh.kumar@linaro.org |
---|---|
State | New |
Headers | show |
Viresh, On Wed, 20 Nov 2013, Viresh Kumar wrote: > Migration of timers from idle cores to non-idle ones for power saving is very > well working and really saves a lot of power for us. What's currently not > working is the migration of running timers Or timers which re-arms themselves. > > There are complications with migrating timers which schedules themselves again > from their handler. del_timer_sync() can't detect that the timer's handler > yet has not finished. Because you try to violate the semantics of the existing code. There is a reason why we enforce that running timers must be requeued on the base they are running on. Now you try to duct tape it into submission. That's not going to fly. The timer wheel code is not designed to allow that and it has lots of other short comings and historic burdens. I'm not going to accept more duct tape and fragile hackery into that code. I'm already working on a complete replacement infrastructure, which avoids all the short comings of the current timer implementation including this one. It's going to be a massive effort to convert everything over to the new infrastructure so we can kill the timer wheel, but that's less scary and less risky than trying to add more fragility to the existing code. Thanks, tglx
On Wednesday 20 November 2013 05:12 PM, Thomas Gleixner wrote: > Viresh, > > On Wed, 20 Nov 2013, Viresh Kumar wrote: > >> Migration of timers from idle cores to non-idle ones for power saving is very >> well working and really saves a lot of power for us. What's currently not >> working is the migration of running timers Or timers which re-arms themselves. >> >> There are complications with migrating timers which schedules themselves again >> from their handler. del_timer_sync() can't detect that the timer's handler >> yet has not finished. > > Because you try to violate the semantics of the existing code. There > is a reason why we enforce that running timers must be requeued on the > base they are running on. > > Now you try to duct tape it into submission. That's not going to fly. > > The timer wheel code is not designed to allow that and it has lots of > other short comings and historic burdens. I'm not going to accept more > duct tape and fragile hackery into that code. > > I'm already working on a complete replacement infrastructure, which > avoids all the short comings of the current timer implementation > including this one. > > It's going to be a massive effort to convert everything over to the > new infrastructure so we can kill the timer wheel, but that's less > scary and less risky than trying to add more fragility to the existing > code. Thanks for the feedback. I Atleast know now that this patch doesn't have a future :)
diff --git a/include/linux/timer.h b/include/linux/timer.h index 8c5a197..ad00ebe 100644 --- a/include/linux/timer.h +++ b/include/linux/timer.h @@ -20,6 +20,7 @@ struct timer_list { void (*function)(unsigned long); unsigned long data; + int wait_for_migration_to_complete; int slack; diff --git a/kernel/timer.c b/kernel/timer.c index 6582b82..30a93e6 100644 --- a/kernel/timer.c +++ b/kernel/timer.c @@ -748,21 +748,15 @@ __mod_timer(struct timer_list *timer, unsigned long expires, new_base = per_cpu(tvec_bases, cpu); if (base != new_base) { - /* - * We are trying to schedule the timer on the local CPU. - * However we can't change timer's base while it is running, - * otherwise del_timer_sync() can't detect that the timer's - * handler yet has not finished. This also guarantees that - * the timer is serialized wrt itself. - */ - if (likely(base->running_timer != timer)) { - /* See the comment in lock_timer_base() */ - timer_set_base(timer, NULL); - spin_unlock(&base->lock); - base = new_base; - spin_lock(&base->lock); - timer_set_base(timer, base); - } + if (base->running_timer == timer) + timer->wait_for_migration_to_complete = 1; + + /* See the comment in lock_timer_base() */ + timer_set_base(timer, NULL); + spin_unlock(&base->lock); + base = new_base; + spin_lock(&base->lock); + timer_set_base(timer, base); } timer->expires = expires; @@ -992,7 +986,8 @@ int try_to_del_timer_sync(struct timer_list *timer) base = lock_timer_base(timer, &flags); - if (base->running_timer != timer) { + if ((base->running_timer != timer) && + !timer->wait_for_migration_to_complete) { timer_stats_timer_clear_start_info(timer); ret = detach_if_pending(timer, base, true); } @@ -1185,6 +1180,8 @@ static inline void __run_timers(struct tvec_base *base) call_timer_fn(timer, fn, data); spin_lock_irq(&base->lock); } + if (timer->wait_for_migration_to_complete) + timer->wait_for_migration_to_complete = 0; } } base->running_timer = NULL;
Hi Thomas, This was earlier discussed here (Well, Not much :)): https://lkml.org/lkml/2012/11/6/160 I am floating the idea again to get more attention on this patch. This is just an idea for now, haven't seen much testing.. Migration of timers from idle cores to non-idle ones for power saving is very well working and really saves a lot of power for us. What's currently not working is the migration of running timers Or timers which re-arms themselves. There are complications with migrating timers which schedules themselves again from their handler. del_timer_sync() can't detect that the timer's handler yet has not finished. __mod_timer migrates the timer with following code: ... spin_lock(&old_base->lock); ... timer_set_base(timer, NULL); spin_unlock(&old_base->lock); ->A spin_lock(&new_base->lock); ->B timer_set_base(timer, new_base); ... After the unlock at time A, old_base->running_timer may get updated to the next timer in queue. After lock at time B, lock_timer_base() will return new_base where another timer might be running timer at that point of time. Whereas, del_timer_sync() depends on below code to check if a timer's handler is currently running or not. if (base->running_timer != timer) Because there is window where timer's handler would be running and at the same time it wasn't marked as running timer for any of the bases (well, it matters only for its current base, i.e. new_base). del_timer_sync() wouldn't know this and will go on and delete the timer, whose handler is currently running. The new approach tries to mark such timers with wait_for_migration_to_complete flag (can be named better and some memory can be saved as PeterZ suggested), which will be used in del_timer_sync() to see if the timer is currently migrating and so isn't marked as running_timer of its base. Benefits: Well, obviously saving power for a core which is being interrupted again and again in its idle loop by this timer event. Which will also prevent the core to go in deeper idle states if possible. Please share your feedback on this approach. Signed-off-by: Viresh Kumar <viresh.kumar@linaro.org> --- include/linux/timer.h | 1 + kernel/timer.c | 29 +++++++++++++---------------- 2 files changed, 14 insertions(+), 16 deletions(-)