On 09.08.2018 10:14, Michal Hocko wrote:
> On Wed 08-08-18 16:20:54, Kirill Tkhai wrote:
>> [Added two more places needed srcu_dereference(). All ->shrinker_map
>> dereferences must be under SRCU, and this v2 adds missed in previous]
>>
>> The patch makes shrinker list and shrinker_idr SRCU-safe
>> for readers. This requires synchronize_srcu() on finalize
>> stage unregistering stage, which waits till all parallel
>> shrink_slab() are finished
>>
>> Note, that patch removes rwsem_is_contended() checks from
>> the code, and this does not result in delays during
>> registration, since there is no waiting at all. Unregistration
>> case may be optimized by splitting unregister_shrinker()
>> in tho stages, and this is made in next patches.
>>
>> Also, keep in mind, that in case of SRCU is not allowed
>> to make unconditional (which is done in previous patch),
>> it is possible to use percpu_rw_semaphore instead of it.
>> percpu_down_read() will be used in shrink_slab_memcg()
>> and in shrink_slab(), and consecutive calls
>>
>> percpu_down_write(percpu_rwsem);
>> percpu_up_write(percpu_rwsem);
>>
>> will be used instead of synchronize_srcu().
>
> An obvious question. Why didn't you go that way? What are pros/cons of
> both approaches?
1)After percpu_rw_semaphore is introduced, shrink_slab() will be not able
to do successful percpu_down_read_trylock() for longer time in comparison
to current behavior:
[cpu0] [cpu1]
{un,}register_shrinker(); shrink_slab()
percpu_down_write(); percpu_down_read_trylock() -> fail
synchronize_rcu(); -> in some periods very slow on big SMP ...
shrink_slab()
percpu_down_read_trylock() -> fail
Also, register_shrinker() and unregister_shrinker() will become slower for the same reason.
Unlike unregister_shrinker(); register_shrinker() can't be made asynchronous/delayed, so
simple mount() performance will be worse.
It's possible, these both can be solved by using both percpu_rw_semaphore and rw_semaphore.
shrink_slab() may fall back to rw_semaphore in case of percpu_rw_semaphore can't be blocked:
shrink_slab()
{
bool percpu = true;
if (!percpu_down_read_try_lock()) {
if(!down_read_trylock())
return 0;
percpu = false;
}
shrinker = idr_find();
...
if (percpu)
percpu_up_read();
else
up_read();
}
register_shrinker()
{
down_write();
idr_alloc();
up_write();
}
unregister_shrinker()
{
percpu_down_write();
down_write();
idr_remove();
up_write();
percpu_up_write();
}
But a)On big machine this may turn in always down_read_trylock() like we have now;
b)I'm not sure, unlocked idr_find() is safe in parallel with idr_alloc(), maybe,
there is needed something else around it (I just haven't investigated this).
All the above are cons. Pros are not enabling SRCU.
2)SRCU. Pros are there are no the above problems; we will have completely unlocked and
scalable shrink_slab(). We will also have a possibility to avoid unregistering delays,
like I did for superblock shrinker. There will be full scalability.
Cons is enabling SRCU.
Kirill
