del_timer_sync scalability patch
Hello.
I am not sure, but I think this patch incorrect.
Review repeat: · line 484
Quoted 4 lines
(quoted diff) +0
-0@@ -466,6 +482,7 @@ repeat: set_running_timer(base, timer); smp_wmb(); timer->base = NULL;
------> WINDOW <------
Review repeat: · line 486
Quoted 2 lines
(quoted diff) +1
-0+ set_last_running(timer, base); spin_unlock_irq(&base->lock);
Suppose it is the first invocation of timer, sotimer->last_running == NULL.
What if del_timer_sync() happens in that window?
del_timer_sync:
del_timer(); // timer->base == NULL, returns base = timer->last_running; if (base) // no, it is still NULL ... if (timer->base != NULL || timer->last_running != base) goto del_again; // not taken return;
I think it is not enough to exchange these 2 lines in __run_timers, we also need barriers.
Oleg.
Quoted 2 lines · Oleg Nesterov
I think it is not enough to exchange these 2 lines in
__run_timers, we also need barriers.
Maybe its best to drop last_running_timer as Ingo suggested.
Replace the magic with a flag that can be set to stop scheduling a timer again.
Then del_timer_sync may
1. Set the flag not to reschedule
2. delete the timer
3. wait till the timer function is complete
4. (eventually verify that the timer is really gone)
5. reset the no reschedule flag
cover del_timer_sync: proof of concept
Quoted 6 lines · Andrew Morton
If we're prepared to rule that a timer handler is not allowed to do
add_timer_on()then a recurring timer is permanently pinned to a CPU, isn't it?That should make things simpler?
In that case I think that both problems (race and scalability) can be solved without increasing sizeof(timer_list).
What if timer_list had ->pending field? Then we can do:
timer_pending:
return timer->pending;
__mod_timer:
internal_add_timer(new_base, timer); timer->base = new_base; timer->pending = 1;
__run_timers:
list_del(&timer->entry); set_running_timer(base, timer); /* do not change timer->base */ timer->pending = 0; spin_unlock(base->lock); timer->function();
del_timer:
if (!timer->pending) return 0; base = timer->base; ...
del_timer_sync:
base = timer->base; if (!base) return 0; spin_lock(base->lock); if (base != timer->base) goto del_again; if (base->running_timer == timer) goto del_again; if (timer->pending) list_del(&timer->entry); timer->pending = 0; timer->base = NULL;
The ->pending flag could live in the least significant bit of timer->base, this way we:
1. do not waste the space
2. can read/write base+pending atomically
These patches are incomplete/suboptimal, just a proof of concept.
Oleg.
Re: [PATCH 0/2] del_timer_sync: proof of concept
I like the idea and it would solve the concerns that we had. The encoding of a bit in a pointer is weird but we have done that before in the page_struct.
However, this also means that __run_timers will not free up the timer and it has to be explicitly freed with del_timer_??. There may be code that relies on not having to delete a single shot timer after it has been fired.
Re: [PATCH 0/2] del_timer_sync: proof of concept
Quoted 3 lines · Christoph Lameter
However, this also means that
__run_timerswill not free up the timer and it has to be explicitly freed withdel_timer_??.
I am not sure I understand you but no, del_timer{,_sync} is not needed.
__run_timer deletes timer from base->tv? list and clears 'pending flag'.
__del_timer_sync sets ->_base = NULL, but it is merely optimization. It could set ->_base = base, but in that case next del_timer_sync() call will need spin_lock(base->lock) again.
Oleg.
Re: [PATCH 0/2] del_timer_sync: proof of concept
Quoted 5 lines · Oleg Nesterov
Quoted 3 lines · Christoph Lameter
However, this also means that
__run_timerswill not free up the timer and it has to be explicitly freed withdel_timer_??.
Quoted 7 lines
I am not sure I understand you but no, del_timer{,_sync} is not needed.
__run_timerdeletes timer frombase->tv? list and clears 'pending flag'.
__del_timer_syncsets ->_base = NULL, but it is merely optimization. It could set ->_base = base, but in that case nextdel_timer_sync()call will needspin_lock(base->lock)again.
For some reason I thought that ->base == NULL would have special significance outside of the function you discussed. Looks fine to me now.
Re: [PATCH 0/2] del_timer_sync: proof of concept
In reply to Oleg NesterovQuoted 6 lines · Andrew Morton
If we're prepared to rule that a timer handler is not allowed to do
add_timer_on()then a recurring timer is permanently pinned to a CPU, isn't it?That should make things simpler?
I think that current inplementation of del_timer_sync() don't like add_timer_on() too.
Consider the timer running on CPU_0. It sets timer->expires = jiffies, and calls add_timer_on(1). Now it is possible that local timer interrupt on CPU_1 happens and starts that timer before timer->function returns on CPU_0.
del_timer_sync() detects that timer is running on CPU_0, waits while ->running_timer == timer, and returns. The timer still runs on CPU_1.
Oleg.
PATCH 1/2 del_timer_sync: proof of concept
In reply to Christoph LameterThis patch renames ->base to _base. Now this field is used only in __get_base/__set_base helpers.
It is ugly, just to reduce the size of patch.
No changes in kernel/timer.o, so it is correct.
include/linux/timer.h renamed from include/linux/timer.h~1_BASE +9 -4
@@ -18,16 +18,21 @@ struct timer_list {void (*function)(unsigned long );unsigned long data; -struct tvec_t_base_s *base; +struct tvec_t_base_s *_base; }; +static inline int __timer_pending (struct tvec_t_base_s *base) +{ +return base !=NULL ; +} +#define TIMER_MAGIC 0x4b87ad6e#define TIMER_INITIALIZER (_function, _expires, _data) { \ .function = (_function), \ .expires = (_expires), \ .data = (_data), \ - .base = NULL, \ + ._base = NULL, \ .magic = TIMER_MAGIC, \ .lock = SPIN_LOCK_UNLOCKED, \ }
@@ -41,7 +46,7 @@ struct timer_list {*/ static inline void init_timer (struct timer_list * timer) { - timer->base =NULL ; + timer->_base =NULL ; timer->magic =TIMER_MAGIC ;spin_lock_init (&timer->lock); }
@@ -58,7 +63,7 @@ static inline void init_timer(struct tim*/ static inline int timer_pending (const struct timer_list * timer) { -return timer->base !=NULL ; +return __timer_pending (timer->_base); }extern void add_timer_on (struct timer_list *timer,int cpu);
kernel/timer.c renamed from kernel/timer.c~1_BASE +25 -11
@@ -84,6 +84,20 @@ static inline void set_running_timer(tve#endif } +static inline tvec_base_t *__get_base (struct timer_list *timer) +{ +return timer->_base; +} + +static inline void __set_base (struct timer_list *timer, +tvec_base_t *base,int pending) +{ +if (pending) + timer->_base = base; +else + timer->_base =NULL ; +} +/* Fake initialization */ static DEFINE_PER_CPU(tvec_base_t, tvec_bases) = { SPIN_LOCK_UNLOCKED };
@@ -167,7 +181,7 @@ int __mod_timer(struct timer_list *timerspin_lock_irqsave (&timer->lock, flags); new_base = &__get_cpu_var (tvec_bases); repeat: - old_base = timer->base; + old_base =__get_base (timer);/* * Prevent deadlocks via ordering by old_base < new_base.
@@ -184,14 +198,14 @@ repeat:* The timer base might have been cancelled while we were * trying to take the lock(s): */ -if (timer->base != old_base) { +if (__get_base (timer) != old_base) {spin_unlock (&new_base->lock);spin_unlock (&old_base->lock); goto repeat; } }else {spin_lock (&new_base->lock); -if (timer->base != old_base) { +if (__get_base (timer) != old_base) {spin_unlock (&new_base->lock); goto repeat; }
@@ -207,7 +221,7 @@ repeat: } timer->expires = expires;internal_add_timer (new_base, timer); - timer->base = new_base; +__set_base (timer, new_base,1 );if (old_base && (new_base != old_base))spin_unlock (&old_base->lock);
@@ -237,7 +251,7 @@ void add_timer_on(struct timer_list *timspin_lock_irqsave (&base->lock, flags);internal_add_timer (base, timer); - timer->base = base; +__set_base (timer, base,1 );spin_unlock_irqrestore (&base->lock, flags); }
@@ -299,18 +313,18 @@ int del_timer(struct timer_list *timer)check_timer (timer); repeat: - base = timer->base; + base =__get_base (timer);if (!base)return 0 ;spin_lock_irqsave (&base->lock, flags); -if (base != timer->base) { +if (base !=__get_base (timer)) {spin_unlock_irqrestore (&base->lock, flags); goto repeat; }list_del (&timer->entry);/* Need to make sure that anybody who sees a NULL base also sees the list ops */ smp_wmb (); - timer->base =NULL ; +__set_base (timer, base,0 );spin_unlock_irqrestore (&base->lock, flags);return 1 ;
@@ -413,7 +427,7 @@ static int cascade(tvec_base_t *base, tvstruct timer_list *tmp; tmp = list_entry(curr, struct timer_list, entry); -BUG_ON (tmp->base != base); +BUG_ON (__get_base (tmp) != base); curr = curr->next;internal_add_timer (base, tmp); }
@@ -463,7 +477,7 @@ repeat:list_del (&timer->entry);set_running_timer (base, timer);smp_wmb (); - timer->base =NULL ; +__set_base (timer, base,0 );spin_unlock_irq (&base->lock); {u32 preempt_count =preempt_count ();
@@ -1309,7 +1323,7 @@ static int migrate_timer_list(tvec_base_return 0 ;list_del (&timer->entry);internal_add_timer (new_base, timer); - timer->base = new_base; +__set_base (timer, new_base,1 );spin_unlock (&timer->lock); }return 1 ;
PATCH 2/2 del_timer_sync: proof of concept
In reply to Christoph LameterNew rules:
->_base & 1 : is timer pending ->_base & ~1 : timer's base
If ->_base == NULL, this timer is not running on any cpu. Cleared only in del_timer_sync().
Note that now we don't need del_singleshot_timer_sync().
Here is the code of del_timer_sync:
int del_timer_sync(struct timer_list *timer)
{
int ret = 0;
for (;;) {
unsigned long flags;
tvec_base_t *_base, *base;
_base = timer->_base;
if (!_base)
return ret;
base = __timer_base(_base);
spin_lock_irqsave(&base->lock, flags);
if (timer->_base != _base)
goto unlock;
if (base->running_timer == timer)
goto unlock;
if (__timer_pending(_base)) {
list_del(&timer->entry);
ret = 1;
}
smp_wmb();
timer->_base = NULL;
unlock:
spin_unlock_irqrestore(&base->lock, flags);
}
}include/linux/timer.h renamed from include/linux/timer.h~2_PEND +3 -1
@@ -21,9 +21,11 @@ struct timer_list {struct tvec_t_base_s *_base; }; +#define __TIMER_PENDING 1 +static inline int __timer_pending (struct tvec_t_base_s *base) { -return base !=NULL ; +return ((unsigned long )base & __TIMER_PENDING) !=0 ; }#define TIMER_MAGIC 0x4b87ad6e
kernel/timer.c renamed from kernel/timer.c~2_PEND +40 -20
84 lines not on this page · raw message
Re: [PATCH 2/2] del_timer_sync: proof of concept
Quoted 3 lines · Oleg Nesterov
New rules:
->_base & 1 : is timer pending ->_base & ~1 : timer's base
how would it look like if we had a separate timer->pending field after all? Would it be faster/cleaner?
(we dont need to keep them small _that_ bad - if there's a good reason we should rather add a clean new field than to encode two fields into one field and complicate the code.)
Ingo
Re: [PATCH 2/2] del_timer_sync: proof of concept
Quoted 9 lines · Ingo Molnar
Quoted 3 lines · Oleg Nesterov
New rules:
->_base & 1 : is timer pending ->_base & ~1 : timer's basehow would it look like if we had a separate
timer->pendingfield after all? Would it be faster/cleaner?
The only change visible outside kernel/timer.c is:
static inline int timer_pending(const struct timer_list * timer) {
- return timer->base != NULL;
+ return timer->base & 1;
}
Currently __get_base() usage in the kernel/time.c suboptimal and should be cleanuped, I see no other problems with performance.
Quoted 3 lines
(we dont need to keep them small _that_ bad - if there's a good reason we should rather add a clean new field than to encode two fields into one field and complicate the code.)
I think that separate timer->pending field will require more changes, because we can't read/write base+pending atomically.
int del_timer()
{
again:
if (!timer->pending) // not strictly necessary, but it is
return 0; // nice to avoid locking
base = timer->base;
if (!base)
return 0;
spin_lock(base->lock);
if (!timer->pending) {
spin_unlock();
goto again;
}
if (timer->base != base) {
spin_unlock();
goto again;
}
....
}Note also, that we have to audit every timer->base usage anyway, because currently it mix base and pending.
But may be you are right, the encoding of a bit in a pointer is indeed weird.
Oleg.
Re: [PATCH 2/2] del_timer_sync: proof of concept
Quoted 2 lines · Oleg Nesterov
I think that separate
timer->pendingfield will require more changes, because we can't read/write base+pending atomically.
i think that's the killer argument in favor of the bit-trick. Being able to read/write base+pending atomically is a good excuse. Using less RAM alone is not necessarily a good excuse.
Andrew, could we give Oleg's 2 patches a whirl in -mm?
Ingo
I suspect that del_timer_sync() in its current form is racy.
CPU 0 CPU 1
__run_timers() sets timer->base = NULL
del_timer_sync() starts, calls del_timer(), it returns because timer->base == NULL. waits until the run is complete: while (base->running_timer == timer) preempt_check_resched(); calls schedule(), or long interrupt happens.
timer reschedules itself, __run_timers()
exits.
base->running_timer == NULL, end of loop.
next timer interrupt, __run_timers() picks
this timer again, sets timer->base = NULL
if (timer_pending(timer)) // no, timer->base == NULL goto del_again; // not taken del_timer_sync() returns
timer runs.
No?
Oleg.
Quoted 18 lines · Oleg Nesterov
I suspect that
del_timer_sync()in its current form is racy.CPU 0 CPU 1
__run_timers()setstimer->base= NULLdel_timer_sync() starts, calls del_timer(), it returns because timer->base == NULL. waits until the run is complete: while (base->running_timer == timer) preempt_check_resched(); calls schedule(), or long interrupt happens.timer reschedules itself,
__run_timers()
exits.
Quoted 6 lines
base->running_timer == NULL, end of loop.next timer interrupt,
__run_timers()picks
this timer again, setstimer->base= NULLif (timer_pending(timer)) // no, timer->base == NULL
timer->base is != NULL because the timer has rescheduled itself. __mod_timer sets timer->bBase
Quoted 1 line
goto
del_again; // not taken
Yes it is taken!
Quoted 4 lines
del_timer_sync()returnstimer runs.
Nope.
Quoted 5 lines · Christoph Lameter
Quoted 2 lines · Oleg Nesterov
I suspect that
del_timer_sync()in its current form is racy.
...snip...
Quoted 2 lines
Quoted 2 lines
next timer interrupt,
__run_timers()picks
this timer again, setstimer->base= NULL
^^^^^^^^^^^^^^^^^^^^^^^
Quoted 5 lines
Quoted 2 lines
if (timer_pending(timer)) // no, timer->base == NULL
timer->baseis != NULL because the timer has rescheduled itself.__mod_timersetstimer->bBase
Christoph, please look again. Yes, __mod_timer sets timer->base, but it is cleared in the _next_ timer interrupt on CPU 0.
Andrew, Ingo, what do you think?
Oleg.
How about this take on the problem?
When a potential periodic timer is deleted through timer_del_sync, all cpus are scanned to determine if the timer is running on that cpu. In a NUMA configuration doing so will cause NUMA interlink traffic which limits the scalability of timers.
The following patch removes the magic in the timer_list structure (Andrew suggested that we may not need it anymore) and replaces it with two u8 variables that give us some additional state of the timer
running -> Set if the timer function is currently executing
shutdown -> Set if del_timer_sync is running and the timer
should not be rescheduled.
include/linux/timer.h +7 -6
@@ -13,22 +13,22 @@ struct timer_list {unsigned long expires;spinlock_t lock; -unsigned long magic;void (*function)(unsigned long );unsigned long data; -struct tvec_t_base_s *base; +struct tvec_t_base_s *base;/* ==NULL if timer not scheduled */ +u8 running;/* function is currently executing */ +u8 shutdown;/* do not schedule this timer */ }; -#define TIMER_MAGIC 0x4b87ad6e -#define TIMER_INITIALIZER (_function, _expires, _data) { \ .function = (_function), \ .expires = (_expires), \ .data = (_data), \ .base = NULL, \ - .magic = TIMER_MAGIC, \ + .running = 0, \ + .shutdown = 0, \ .lock = SPIN_LOCK_UNLOCKED, \ }
@@ -42,7 +42,8 @@ struct timer_list {static inline void init_timer (struct timer_list * timer) { timer->base =NULL ; - timer->magic =TIMER_MAGIC ; + timer->running =0 ; + timer->shutdown =0 ;spin_lock_init (&timer->lock); }
kernel/timer.c +30 -51
154 lines not on this page · raw message
kernel/timer.c repeat: · line 457
Quoted 6 lines · Christoph Lameter
kernel/timer.c +1
-0@@ -476,6 +454,7 @@ repeat: } }spin_lock_irq (&base->lock); + timer->running =0 ;
^^^^^^^^^^^^^^^^^^
kernel/timer.c repeat: · line 460
Quoted 3 lines
kernel/timer.c +0
-0goto repeat; } }
This is imho wrong. The timer probably don't exist when timer_list->function returns.
I mean, timer_list->function could deallocate timer's memory.
Oleg.
Quoted 3 lines · Christoph Lameter
The following patch removes the magic in the
timer_liststructure (Andrew suggested that we may not need it anymore) and replaces it with two u8 variables that give us some additional state of the timer
The 'remove the magic' observation is not a 'backdoor' to introduce new fields 'for free'. So please dont mix the two things into one patch.
include/linux/timer.h struct timer_list { · line 22
Quoted 2 lines
include/linux/timer.h +2
-0+u8 running;/* function is currently executing */ +u8 shutdown;/* do not schedule this timer */
it may as well be cleaner to do the timer->base_running thing after all, and to do this in __run_timers():
timer->base = NULL; timer->base_running = base;
note that we cannot clear ->base_running in __run_timers() [we dont own any access to the timer anymore, it may be kfree()d, etc.], so del_timer_sync() has to make sure that the timer->base_running info is not stale - it is a hint only. But it looks doable, and it should solve the NUMA scalability problem.
Ingo