On 27.03.2018 12:18, Vladimir Davydov wrote: > On Mon, Mar 26, 2018 at 06:20:55PM +0300, Kirill Tkhai wrote: >> On 24.03.2018 21:45, Vladimir Davydov wrote: >>> On Wed, Mar 21, 2018 at 04:21:29PM +0300, Kirill Tkhai wrote: >>>> The patch introduces mcg_shrinkers array to keep memcg-aware >>>> shrinkers in order of their shrinker::id. >>>> >>>> This allows to access the shrinkers dirrectly by the id, >>>> without iteration over shrinker_list list. >>> >>> Why don't you simply use idr instead of ida? With idr you wouldn't need >>> the array mapping shrinker id to shrinker ptr. AFAIU you need this >>> mapping to look up the shrinker by id in shrink_slab. The latter doesn't >>> seem to be a hot path so using idr there should be acceptable. Since we >>> already have shrinker_rwsem, which is taken for reading by shrink_slab, >>> we wouldn't even need any additional locking for it. >> >> The reason is ida may allocate memory, and since list_lru_add() can't fail, >> we can't do that there. If we allocate all the ida memory at the time of >> memcg creation (i.e., preallocate it), this is not different to the way >> the bitmap makes. >> >> While bitmap has the agvantage, since it's simplest data structure (while >> ida has some radix tree overhead). >> >> Also, bitmap does not require a lock, there is single atomic operation >> to set or clear a bit, and it scales better, when anything. > > I didn't mean the per-memcg bitmaps - I think it's OK to use plain > arrays for them and reallocate them with the aid of RCU. > > What I actually mean is the mapping shrink_id => shrinker. AFAIU it > isn't accessed from list_lru, it is only needed to look up a shrinker > by id from shrink_slab(). The latter is rather a slow path so I think > we can use an IDR for this mapping instead of IDA + plain array. This is good idea. Thanks, Kirill
