Re: [PATCH v4 46/48] mm: shrinker: make memcg slab shrink lockless

[Qi Zheng <zhengqi.arch@bytedance.com>]

Hi Dave,

On 2023/8/8 10:44, Dave Chinner wrote:
> On Mon, Aug 07, 2023 at 07:09:34PM +0800, Qi Zheng wrote:
>> Like global slab shrink, this commit also uses refcount+RCU method to make
>> memcg slab shrink lockless.
> 
> This patch does random code cleanups amongst the actual RCU changes.
> Can you please move the cleanups to a spearate patch to reduce the
> noise in this one?

Sure, will do.

> 
>> diff --git a/mm/shrinker.c b/mm/shrinker.c
>> index d318f5621862..fee6f62904fb 100644
>> --- a/mm/shrinker.c
>> +++ b/mm/shrinker.c
>> @@ -107,6 +107,12 @@ static struct shrinker_info *shrinker_info_protected(struct mem_cgroup *memcg,
>>   					 lockdep_is_held(&shrinker_rwsem));
>>   }
>>   
>> +static struct shrinker_info *shrinker_info_rcu(struct mem_cgroup *memcg,
>> +					       int nid)
>> +{
>> +	return rcu_dereference(memcg->nodeinfo[nid]->shrinker_info);
>> +}
> 
> This helper doesn't add value. It doesn't tell me that
> rcu_read_lock() needs to be held when it is called, for one....

How about adding a comment or an assertion here?

> 
>>   static int expand_one_shrinker_info(struct mem_cgroup *memcg, int new_size,
>>   				    int old_size, int new_nr_max)
>>   {
>> @@ -198,7 +204,7 @@ void set_shrinker_bit(struct mem_cgroup *memcg, int nid, int shrinker_id)
>>   		struct shrinker_info_unit *unit;
>>   
>>   		rcu_read_lock();
>> -		info = rcu_dereference(memcg->nodeinfo[nid]->shrinker_info);
>> +		info = shrinker_info_rcu(memcg, nid);
> 
> ... whilst the original code here was obviously correct.
> 
>>   		unit = info->unit[shriner_id_to_index(shrinker_id)];
>>   		if (!WARN_ON_ONCE(shrinker_id >= info->map_nr_max)) {
>>   			/* Pairs with smp mb in shrink_slab() */
>> @@ -211,7 +217,7 @@ void set_shrinker_bit(struct mem_cgroup *memcg, int nid, int shrinker_id)
>>   
>>   static DEFINE_IDR(shrinker_idr);
>>   
>> -static int prealloc_memcg_shrinker(struct shrinker *shrinker)
>> +static int shrinker_memcg_alloc(struct shrinker *shrinker)
> 
> Cleanups in a separate patch.

OK.

> 
>> @@ -253,10 +258,15 @@ static long xchg_nr_deferred_memcg(int nid, struct shrinker *shrinker,
>>   {
>>   	struct shrinker_info *info;
>>   	struct shrinker_info_unit *unit;
>> +	long nr_deferred;
>>   
>> -	info = shrinker_info_protected(memcg, nid);
>> +	rcu_read_lock();
>> +	info = shrinker_info_rcu(memcg, nid);
>>   	unit = info->unit[shriner_id_to_index(shrinker->id)];
>> -	return atomic_long_xchg(&unit->nr_deferred[shriner_id_to_offset(shrinker->id)], 0);
>> +	nr_deferred = atomic_long_xchg(&unit->nr_deferred[shriner_id_to_offset(shrinker->id)], 0);
>> +	rcu_read_unlock();
>> +
>> +	return nr_deferred;
>>   }
> 
> This adds two rcu_read_lock() sections to every call to
> do_shrink_slab(). It's not at all clear ifrom any of the other code
> that do_shrink_slab() now has internal rcu_read_lock() sections....

The xchg_nr_deferred_memcg() will only be called in shrink_slab_memcg(),
so other code doesn't need to know that information?

> 
>> @@ -464,18 +480,23 @@ static unsigned long shrink_slab_memcg(gfp_t gfp_mask, int nid,
>>   	if (!mem_cgroup_online(memcg))
>>   		return 0;
>>   
>> -	if (!down_read_trylock(&shrinker_rwsem))
>> -		return 0;
>> -
>> -	info = shrinker_info_protected(memcg, nid);
>> +again:
>> +	rcu_read_lock();
>> +	info = shrinker_info_rcu(memcg, nid);
>>   	if (unlikely(!info))
>>   		goto unlock;
>>   
>> -	for (; index < shriner_id_to_index(info->map_nr_max); index++) {
>> +	if (index < shriner_id_to_index(info->map_nr_max)) {
>>   		struct shrinker_info_unit *unit;
>>   
>>   		unit = info->unit[index];
>>   
>> +		/*
>> +		 * The shrinker_info_unit will not be freed, so we can
>> +		 * safely release the RCU lock here.
>> +		 */
>> +		rcu_read_unlock();
> 
> Why - what guarantees that the shrinker_info_unit exists at this
> point? We hold no reference to it, we hold no reference to any
> shrinker, etc. What provides this existence guarantee?

The shrinker_info_unit is never freed unless the memcg is destroyed.
Here we hold the refcount of this memcg (mem_cgroup_iter() -->
css_tryget()), so the shrinker_info_unit will not be freed.

> 
>> +
>>   		for_each_set_bit(offset, unit->map, SHRINKER_UNIT_BITS) {
>>   			struct shrink_control sc = {
>>   				.gfp_mask = gfp_mask,
>> @@ -485,12 +506,14 @@ static unsigned long shrink_slab_memcg(gfp_t gfp_mask, int nid,
>>   			struct shrinker *shrinker;
>>   			int shrinker_id = calc_shrinker_id(index, offset);
>>   
>> +			rcu_read_lock();
>>   			shrinker = idr_find(&shrinker_idr, shrinker_id);
>> -			if (unlikely(!shrinker || !(shrinker->flags & SHRINKER_REGISTERED))) {
>> -				if (!shrinker)
>> -					clear_bit(offset, unit->map);
>> +			if (unlikely(!shrinker || !shrinker_try_get(shrinker))) {
>> +				clear_bit(offset, unit->map);
>> +				rcu_read_unlock();
>>   				continue;
>>   			}
>> +			rcu_read_unlock();
>>   
>>   			/* Call non-slab shrinkers even though kmem is disabled */
>>   			if (!memcg_kmem_online() &&
>> @@ -523,15 +546,20 @@ static unsigned long shrink_slab_memcg(gfp_t gfp_mask, int nid,
>>   					set_shrinker_bit(memcg, nid, shrinker_id);
>>   			}
>>   			freed += ret;
>> -
>> -			if (rwsem_is_contended(&shrinker_rwsem)) {
>> -				freed = freed ? : 1;
>> -				goto unlock;
>> -			}
>> +			shrinker_put(shrinker);
> 
> Ok, so why is this safe to call without holding the rcu read lock?
> The global shrinker has to hold the rcu_read_lock() whilst calling
> shrinker_put() to guarantee the validity of the list next pointer,
> but we don't hold off RCU here so what guarantees a racing global
> shrinker walk doesn't trip over this shrinker_put() call dropping
> the refcount to zero and freeing occuring in a different context...

This will not be a problem, even if shrinker::refcount is reduced to
0 here, the racing global shrinker walk already holds the rcu lock.

         shrink_slab            shrink_slab_memcg
         ===========            =================

         rcu_read_lock()
         shrinker_put()
                                shrinker_put()

And in shrink_slab_memcg(), the shrinker is not required to traverse the
next bit in the shrinker_info_unit::map, so there is no need to hold the
rcu lock to ensure the existence of this shrinker.

> 
> 
>> +		/*
>> +		 * We have already exited the read-side of rcu critical section
>> +		 * before calling do_shrink_slab(), the shrinker_info may be
>> +		 * released in expand_one_shrinker_info(), so reacquire the
>> +		 * shrinker_info.
>> +		 */
>> +		index++;
>> +		goto again;
> 
> With that, what makes the use of shrinker_info in
> xchg_nr_deferred_memcg() in do_shrink_slab() coherent and valid?

Holding rcu lock can ensure that the old shrinker_info will not be
freed, and the shrinker_info_unit::nr_deferred can also be indexed from
the old shrinker_info::unit[x], so the updated nr_deferred will not be
lost.

Thanks,
Qi

> 
> -Dave.