Messages in this thread | | | Date | Thu, 15 Feb 2024 15:01:55 +0100 | From | Lorenzo Bianconi <> | Subject | Re: [PATCH net-next] page_pool: disable direct recycling based on pool->cpuid on destroy |
| |
> From: Lorenzo Bianconi <lorenzo@kernel.org> > Date: Thu, 15 Feb 2024 14:37:10 +0100 > > >> From: Toke Høiland-Jørgensen <toke@redhat.com> > >> Date: Thu, 15 Feb 2024 13:05:30 +0100 > >> > >>> Alexander Lobakin <aleksander.lobakin@intel.com> writes: > >>> > >>>> Now that direct recycling is performed basing on pool->cpuid when set, > >>>> memory leaks are possible: > >>>> > >>>> 1. A pool is destroyed. > >>>> 2. Alloc cache is emptied (it's done only once). > >>>> 3. pool->cpuid is still set. > >>>> 4. napi_pp_put_page() does direct recycling basing on pool->cpuid. > >>>> 5. Now alloc cache is not empty, but it won't ever be freed. > >>> > >>> Did you actually manage to trigger this? pool->cpuid is only set for the > >>> system page pool instance which is never destroyed; so this seems a very > >>> theoretical concern? > >> > >> To both Lorenzo and Toke: > >> > >> Yes, system page pools are never destroyed, but we might latter use > >> cpuid in non-persistent PPs. Then there will be memory leaks. > >> I was able to trigger this by creating bpf/test_run page_pools with the > >> cpuid set to test direct recycling of live frames. > > > > what about avoiding the page to be destroyed int this case? I do not like the > > I think I didn't get what you wanted to say here :s
My assumption here was cpuid will be set just system page_pool so it is just a matter of not running page_pool_destroy for them. Anyway in the future we could allow to set cpuid even for non-system page_pool if the pool is linked to a given rx-queue and the queue is pinned to a given cpu.
Regards, Lorenzo
> > Rewriting cpuid doesn't introduce any new checks on hotpath. Destroying > the pool is slowpath and we shouldn't hurt hotpath to handle it. > > > idea of overwriting the cpuid field for it. > > We also overwrite pp->p.napi field a couple lines below. It happens only > when destroying the pool, we don't care about the fields at this point. > > > > > Regards, > > Lorenzo > > > >> > >>> > >>> I guess we could still do this in case we find other uses for setting > >>> the cpuid; I don't think the addition of the READ_ONCE() will have any > >>> measurable overhead on the common arches? > >> > >> READ_ONCE() is cheap, but I thought it's worth mentioning in the > >> commitmsg anyway :) > >> > >>> > >>> -Toke > > Thanks, > Olek [unhandled content-type:application/pgp-signature] | |