On 21.03.2018 19:20, Matthew Wilcox wrote: > On Wed, Mar 21, 2018 at 06:43:01PM +0300, Kirill Tkhai wrote: >> On 21.03.2018 18:26, Matthew Wilcox wrote: >>> On Wed, Mar 21, 2018 at 06:12:17PM +0300, Kirill Tkhai wrote: >>>> On 21.03.2018 17:56, Matthew Wilcox wrote: >>>>> Why use your own bitmap here? Why not use an IDA which can grow and >>>>> shrink automatically without you needing to play fun games with RCU? >>>> >>>> Bitmap allows to use unlocked set_bit()/clear_bit() to maintain the map >>>> of not empty shrinkers. >>>> >>>> So, the reason to use IDR here is to save bitmap memory? Does this mean >>>> IDA works fast with sparse identifiers? It seems they require per-memcg >>>> lock to call IDR primitives. I just don't have information about this. >>>> >>>> If so, which IDA primitive can be used to set particular id in bitmap? >>>> There is idr_alloc_cyclic(idr, NULL, id, id+1, GFP_KERNEL) only I see >>>> to do that. >>> >>> You're confusing IDR and IDA in your email, which is unfortunate. >>> >>> You can set a bit in an IDA by calling ida_simple_get(ida, n, n, GFP_FOO); >>> You clear it by calling ida_simple_remove(ida, n); >> >> I moved to IDR in the message, since IDA uses global spinlock. It will be >> taken every time a first object is added to list_lru, or last is removed. >> These may be frequently called operations, and they may scale not good >> on big machines. > > I'm fixing the global spinlock issue with the IDA. Not going to be ready > for 4.17, but hopefully for 4.18. It will be nice to see that in kernel. >> Using IDR will allow us to introduce memcg-related locks, but I'm still not >> sure it's easy to introduce them in scalable-way. Simple set_bit()/clear_bit() >> do not require locks at all. > > They're locked operations ... they may not have an explicit spinlock > associated with them, but the locking still happens. Yes, they are not ideal in this way. >>> The identifiers aren't going to be all that sparse; after all you're >>> allocating them from a global IDA. Up to 62 identifiers will allocate >>> no memory; 63-1024 identifiers will allocate a single 128 byte chunk. >>> Between 1025 and 65536 identifiers, you'll allocate a 576-byte chunk >>> and then 128-byte chunks for each block of 1024 identifiers (*). One of >>> the big wins with the IDA is that it will shrink again after being used. >>> I didn't read all the way through your patchset to see if you bother to >>> shrink your bitmap after it's no longer used, but most resizing bitmaps >>> we have in the kernel don't bother with that part. >>> >>> (*) Actually it's more complex than that... between 1025 and 1086, >>> you'll have a 576 byte chunk, a 128-byte chunk and then use 62 bits of >>> the next pointer before allocating a 128 byte chunk when reaching ID >>> 1087. Similar things happen for the 62 bits after 2048, 3076 and so on. >>> The individual chunks aren't shrunk until they're empty so if you set ID >>> 1025 and then ID 1100, then clear ID 1100, the 128-byte chunk will remain >>> allocated until ID 1025 is cleared. This probably doesn't matter to you. >> >> Sound great, thanks for explaining this. The big problem I see is >> that IDA/IDR add primitives allocate memory, while they will be used >> in the places, where they mustn't fail. There is list_lru_add(), and >> it's called unconditionally in current kernel code. The patchset makes >> the bitmap be populated in this function. So, we can't use IDR there. > > Maybe we can use GFP_NOFAIL here. They're small allocations, so we're > only asking for single-page allocations to not fail, which shouldn't > put too much strain on the VM. Oh. I'm not sure about this. Even if each allocation is small, there is theoretically possible a situation, when many lists will want to add first element. list_lru_add() is called from iput() for example. Kirill
