[Question] vmalloc latency in RT-Linux

[Question] vmalloc latency in RT-Linux

[Zhipeng Shi]

I noticed in rt-linux, vmalloc has a large latency. This is because the 
free_vmap_area_lock is held for a long time in the function  
__purge_vmap_area_lazy. 

In non-RT-Linux, because the function spin_is_contended is well 
implemented, so there will be no such problem.

But in RT-Linux, spin_is_contended simply returns 0. I don't understand 
why this function was implemented like this before, but in order to
solve this problem, I thought of two ways.

The first is to modify the spin_is_contended definition in spinlock_rt.h 
as shown below, but I'm not sure if the change has side-effects:

-#define spin_is_contended(lock)    (((void)(lock), 0))
+static inline int spin_is_contended(spinlock_t *lock)
+{
+ unsigned long *p = (unsigned long *) &lock->lock.owner;
+
+ return (READ_ONCE(*p) & RT_MUTEX_HAS_WAITERS);
+}

The second is by reducing the number of lazy_max_pages, but it will lead 
to lower performance of vmalloc.

Guys, Do you have any good ideas?

Best regards,
Zhipeng

Re: [Question] vmalloc latency in RT-Linux

[Baoquan He]

On 06/21/22 at 08:15pm, Zhipeng Shi wrote:
> I noticed in rt-linux, vmalloc has a large latency. This is because the 
> free_vmap_area_lock is held for a long time in the function  
> __purge_vmap_area_lazy. 
> 
> In non-RT-Linux, because the function spin_is_contended is well 
> implemented, so there will be no such problem.
> 
> But in RT-Linux, spin_is_contended simply returns 0. I don't understand 
> why this function was implemented like this before, but in order to
> solve this problem, I thought of two ways.
> 
> The first is to modify the spin_is_contended definition in spinlock_rt.h 
> as shown below, but I'm not sure if the change has side-effects:
> 
> -#define spin_is_contended(lock)    (((void)(lock), 0))
> +static inline int spin_is_contended(spinlock_t *lock)
> +{
> + unsigned long *p = (unsigned long *) &lock->lock.owner;
> +
> + return (READ_ONCE(*p) & RT_MUTEX_HAS_WAITERS);
> +}
> 
> The second is by reducing the number of lazy_max_pages, but it will lead 
> to lower performance of vmalloc.

__purge_vmap_area_lazy() has cond_resched_lock() to reschedule and drop
the lock. From your saying, it's spin_is_contended() which is not
working well to make rescheduling happen during __purge_vmap_area_lazy()
handling. Then the fixing should be done in lock side.

> 
> Guys, Do you have any good ideas?
> 
> Best regards,
> Zhipeng
> 

Re: [Question] vmalloc latency in RT-Linux

[Waiman Long]

On 6/23/22 06:51, Baoquan He wrote:
> On 06/21/22 at 08:15pm, Zhipeng Shi wrote:
>> I noticed in rt-linux, vmalloc has a large latency. This is because the
>> free_vmap_area_lock is held for a long time in the function
>> __purge_vmap_area_lazy.
>>
>> In non-RT-Linux, because the function spin_is_contended is well
>> implemented, so there will be no such problem.
>>
>> But in RT-Linux, spin_is_contended simply returns 0. I don't understand
>> why this function was implemented like this before, but in order to
>> solve this problem, I thought of two ways.
>>
>> The first is to modify the spin_is_contended definition in spinlock_rt.h
>> as shown below, but I'm not sure if the change has side-effects:
>>
>> -#define spin_is_contended(lock)    (((void)(lock), 0))
>> +static inline int spin_is_contended(spinlock_t *lock)
>> +{
>> + unsigned long *p = (unsigned long *) &lock->lock.owner;
>> +
>> + return (READ_ONCE(*p) & RT_MUTEX_HAS_WAITERS);
>> +}
>>
>> The second is by reducing the number of lazy_max_pages, but it will lead
>> to lower performance of vmalloc.
> __purge_vmap_area_lazy() has cond_resched_lock() to reschedule and drop
> the lock. From your saying, it's spin_is_contended() which is not
> working well to make rescheduling happen during __purge_vmap_area_lazy()
> handling. Then the fixing should be done in lock side.

Sebastian had sent out patch last year to fix spin_is_contended().

https://lore.kernel.org/lkml/20210906143004.2259141-3-bigeasy@linutronix.de/

However, there is no follow-up after some discussion and the patch 
wasn't merged.

Cheers,
Longman

Re: [Question] vmalloc latency in RT-Linux

[Baoquan He]

On 06/23/22 at 02:04pm, Waiman Long wrote:
> On 6/23/22 06:51, Baoquan He wrote:
> > On 06/21/22 at 08:15pm, Zhipeng Shi wrote:
> > > I noticed in rt-linux, vmalloc has a large latency. This is because the
> > > free_vmap_area_lock is held for a long time in the function
> > > __purge_vmap_area_lazy.
> > > 
> > > In non-RT-Linux, because the function spin_is_contended is well
> > > implemented, so there will be no such problem.
> > > 
> > > But in RT-Linux, spin_is_contended simply returns 0. I don't understand
> > > why this function was implemented like this before, but in order to
> > > solve this problem, I thought of two ways.
> > > 
> > > The first is to modify the spin_is_contended definition in spinlock_rt.h
> > > as shown below, but I'm not sure if the change has side-effects:
> > > 
> > > -#define spin_is_contended(lock)    (((void)(lock), 0))
> > > +static inline int spin_is_contended(spinlock_t *lock)
> > > +{
> > > + unsigned long *p = (unsigned long *) &lock->lock.owner;
> > > +
> > > + return (READ_ONCE(*p) & RT_MUTEX_HAS_WAITERS);
> > > +}
> > > 
> > > The second is by reducing the number of lazy_max_pages, but it will lead
> > > to lower performance of vmalloc.
> > __purge_vmap_area_lazy() has cond_resched_lock() to reschedule and drop
> > the lock. From your saying, it's spin_is_contended() which is not
> > working well to make rescheduling happen during __purge_vmap_area_lazy()
> > handling. Then the fixing should be done in lock side.
> 
> Sebastian had sent out patch last year to fix spin_is_contended().
> 
> https://lore.kernel.org/lkml/20210906143004.2259141-3-bigeasy@linutronix.de/
> 
> However, there is no follow-up after some discussion and the patch wasn't
> merged.

That's great. Thanks, Longman.

Then this is a good chance to reconsider it, maybe with a test from Zhipeng.

Re: [Question] vmalloc latency in RT-Linux

[Zhipeng Shi]

On Fri, Jun 24, 2022 at 10:39:43AM +0800, Baoquan He wrote:
> On 06/23/22 at 02:04pm, Waiman Long wrote:
> > On 6/23/22 06:51, Baoquan He wrote:
> > > On 06/21/22 at 08:15pm, Zhipeng Shi wrote:
> > > > I noticed in rt-linux, vmalloc has a large latency. This is because the
> > > > free_vmap_area_lock is held for a long time in the function
> > > > __purge_vmap_area_lazy.
> > > > 
> > > > In non-RT-Linux, because the function spin_is_contended is well
> > > > implemented, so there will be no such problem.
> > > > 
> > > > But in RT-Linux, spin_is_contended simply returns 0. I don't understand
> > > > why this function was implemented like this before, but in order to
> > > > solve this problem, I thought of two ways.
> > > > 
> > > > The first is to modify the spin_is_contended definition in spinlock_rt.h
> > > > as shown below, but I'm not sure if the change has side-effects:
> > > > 
> > > > -#define spin_is_contended(lock)    (((void)(lock), 0))
> > > > +static inline int spin_is_contended(spinlock_t *lock)
> > > > +{
> > > > + unsigned long *p = (unsigned long *) &lock->lock.owner;
> > > > +
> > > > + return (READ_ONCE(*p) & RT_MUTEX_HAS_WAITERS);
> > > > +}
> > > > 
> > > > The second is by reducing the number of lazy_max_pages, but it will lead
> > > > to lower performance of vmalloc.
> > > __purge_vmap_area_lazy() has cond_resched_lock() to reschedule and drop
> > > the lock. From your saying, it's spin_is_contended() which is not
> > > working well to make rescheduling happen during __purge_vmap_area_lazy()
> > > handling. Then the fixing should be done in lock side.
> > 
> > Sebastian had sent out patch last year to fix spin_is_contended().
> > 
> > https://lore.kernel.org/lkml/20210906143004.2259141-3-bigeasy@linutronix.de/
> > 
> > However, there is no follow-up after some discussion and the patch wasn't
> > merged.
> 
> That's great. Thanks, Longman.
> 
> Then this is a good chance to reconsider it, maybe with a test from Zhipeng.

Before that, since I didn't find the patch that Sebastian sent before, 
I sent relevant patch for this problem (now it seems that Sebastian's 
changes are better than mine) and test scripts. please refer to the 
following links:

https://lore.kernel.org/lkml/20220608142457.GA2400218@ubuntu20/

With this patch, max-latency of vmalloc reduce from 10+ msec to
200+ usec, this because spin_lock is released halfway through 
__purge_vmap_area_lazy.

Best regards,
Zhipeng

Re: [Question] vmalloc latency in RT-Linux

[Sebastian Andrzej Siewior]

On 2022-06-24 10:39:43 [+0800], Baoquan He wrote:
> Then this is a good chance to reconsider it, maybe with a test from Zhipeng.

I reconsidered and it was dropped purpose, see
	https://lore.kernel.org/lkml/YT80AB8%2FG59QBSVq@hirez.programming.kicks-ass.net/

Sebastian

Re: [Question] vmalloc latency in RT-Linux

[Waiman Long]

On 6/24/22 02:46, Sebastian Andrzej Siewior wrote:
> On 2022-06-24 10:39:43 [+0800], Baoquan He wrote:
>> Then this is a good chance to reconsider it, maybe with a test from Zhipeng.
> I reconsidered and it was dropped purpose, see
> 	https://lore.kernel.org/lkml/YT80AB8%2FG59QBSVq@hirez.programming.kicks-ass.net/

I do agree that is_contended may not that useful for rwlock, but it can 
be useful for spinlock. Will you consider a version just for rt_spinlock?

Cheers,
Longman