On Wed, Jan 17, 2024 at 2:20 PM Roman Gushchin <roman.gushchin@xxxxxxxxx> wrote: > > On Wed, Jan 17, 2024 at 01:02:19PM -0800, Shakeel Butt wrote: > > On Wed, Jan 17, 2024 at 12:21 PM Linus Torvalds > > <torvalds@xxxxxxxxxxxxxxxxxxxx> wrote: > > > > > > On Wed, 17 Jan 2024 at 11:39, Josh Poimboeuf <jpoimboe@xxxxxxxxxx> wrote: > > > > > > > > That's a good point. If the microbenchmark isn't likely to be even > > > > remotely realistic, maybe we should just revert the revert until if/when > > > > somebody shows a real world impact. > > > > > > > > Linus, any objections to that? > > > > > > We use SLAB_ACCOUNT for much more common allocations like queued > > > signals, so I would tend to agree with Jeff that it's probably just > > > some not very interesting microbenchmark that shows any file locking > > > effects from SLAB_ALLOC, not any real use. > > > > > > That said, those benchmarks do matter. It's very easy to say "not > > > relevant in the big picture" and then the end result is that > > > everything is a bit of a pig. > > > > > > And the regression was absolutely *ENORMOUS*. We're not talking "a few > > > percent". We're talking a 33% regression that caused the revert: > > > > > > https://lore.kernel.org/lkml/20210907150757.GE17617@xsang-OptiPlex-9020/ > > > > > > I wish our SLAB_ACCOUNT wasn't such a pig. Rather than account every > > > single allocation, it would be much nicer to account at a bigger > > > granularity, possibly by having per-thread counters first before > > > falling back to the obj_cgroup_charge. Whatever. > > > > > > It's kind of stupid to have a benchmark that just allocates and > > > deallocates a file lock in quick succession spend lots of time > > > incrementing and decrementing cgroup charges for that repeated > > > alloc/free. > > > > > > However, that problem with SLAB_ACCOUNT is not the fault of file > > > locking, but more of a slab issue. > > > > > > End result: I think we should bring in Vlastimil and whoever else is > > > doing SLAB_ACCOUNT things, and have them look at that side. > > > > > > And then just enable SLAB_ACCOUNT for file locks. But very much look > > > at silly costs in SLAB_ACCOUNT first, at least for trivial > > > "alloc/free" patterns.. > > > > > > Vlastimil? Who would be the best person to look at that SLAB_ACCOUNT > > > thing? See commit 3754707bcc3e (Revert "memcg: enable accounting for > > > file lock caches") for the history here. > > > > > > > Roman last looked into optimizing this code path. I suspect > > mod_objcg_state() to be more costly than obj_cgroup_charge(). I will > > try to measure this path and see if I can improve it. > > It's roughly an equal split between mod_objcg_state() and obj_cgroup_charge(). > And each is comparable (by order of magnitude) to the slab allocation cost > itself. On the free() path a significant cost comes simple from reading > the objcg pointer (it's usually a cache miss). > > So I don't see how we can make it really cheap (say, less than 5% overhead) > without caching pre-accounted objects. Maybe this is what we want. Now we are down to just SLUB, maybe such caching of pre-accounted objects can be in SLUB layer and we can decide to keep this caching per-kmem-cache opt-in or always on. > > I thought about merging of charge and stats handling paths, which _maybe_ can > shave off another 20-30%, but there still will be a double-digit% accounting > overhead. > > I'm curious to hear other ideas and suggestions. > > Thanks!
