On Tue, Mar 12, 2024 at 10:22:54AM +0100, Vlastimil Babka wrote: > On 3/1/24 19:53, Roman Gushchin wrote: > > On Fri, Mar 01, 2024 at 09:51:18AM -0800, Linus Torvalds wrote: > >> What I *think* I'd want for this case is > >> > >> (a) allow the accounting to go over by a bit > >> > >> (b) make sure there's a cheap way to ask (before) about "did we go > >> over the limit" > >> > >> IOW, the accounting never needed to be byte-accurate to begin with, > >> and making it fail (cheaply and early) on the next file allocation is > >> fine. > >> > >> Just make it really cheap. Can we do that? > >> > >> For example, maybe don't bother with the whole "bytes and pages" > >> stuff. Just a simple "are we more than one page over?" kind of > >> question. Without the 'stock_lock' mess for sub-page bytes etc > >> > >> How would that look? Would it result in something that can be done > >> cheaply without locking and atomics and without excessive pointer > >> indirection through many levels of memcg data structures? > > > > I think it's possible and I'm currently looking into batching charge, > > objcg refcnt management and vmstats using per-task caching. It should > > speed up things for the majority of allocations. > > For allocations from an irq context and targeted allocations > > (where the target memcg != memcg of the current task) we'd probably need to > > keep the old scheme. I hope to post some patches relatively soon. > > Do you think this will work on top of this series, i.e. patches 1+2 could be > eventually put to slab/for-next after the merge window, or would it > interfere with your changes? Please, go on and merge them, I'll rebase on top of it, it will be even better for my work. I made a couple of comments there, but overall they look very good to me, thank you for doing this work! > > > I tried to optimize the current implementation but failed to get any > > significant gains. It seems that the overhead is very evenly spread across > > objcg pointer access, charge management, objcg refcnt management and vmstats. I started working on the thing, but it's a bit more complicated than I initially thought because: 1) there are allocations made from a !in_task() context, so we need to handle this correctly 2) tasks can be moved between cgroups concurrently to memory allocations. fortunately my recent changes provide a path here, but it adds to the complexity. In alternative world where tasks can't move between cgroups the life would be so much better (and faster too, we could remove a ton of synchronization). 3) we do have per-numa-node per-memcg stats, which are less trivial to cache on struct task I hope to resolve these issues somehow and post patches, but probably will need a bit more time. Thanks!
