Re: [PATCH bpf-next v3] docs/bpf: Add LRU internals description and graph

From: Joe Stringer
Date: Sat Apr 01 2023 - 14:57:49 EST


On Thu, Mar 16, 2023 at 11:05 PM Martin KaFai Lau <martin.lau@xxxxxxxxx> wrote:
>
> On 3/15/23 6:54 PM, Joe Stringer wrote:
> > On Tue, Mar 14, 2023 at 12:31 PM Martin KaFai Lau <martin.lau@xxxxxxxxx> wrote:
> >>
> >> Maybe a note somewhere to mention why it will still fail to
> >> shrink the list because the htab_lock_bucket() have detected potential
> >> deadlock/recursion which is a very unlikely case.
> >
> > I missed the "shrink the list" link here since it seems like this
> > could happen for any combination of update or delete elems for the
> > same bucket. But yeah given that also needs to happen on the same CPU,
> > it does seem very unlikely...
>
> shrink should try to shrink a couple of stale elems which are likely hashed to
> different buckets. If shrink hits htab_lock_bucket() EBUSY on all stale elems,
> shrink could fail but unlikely.

The failure case I had in mind for this is to assume that shrinking
succeeds and we find an LRU node during the htab_map_update_elem()
call through prealloc_lru_pop(), but then immediately afterwards it
makes a direct htab_lock_bucket() call which has just one chance to
succeed based on whether this CPU races against some other user of the
bucket lock. Still seems somewhat rare, but feasible to hit.

> > Could there be a case something like "userspace process is touching that bucket,
> > gets interrupted, then the same CPU runs an eBPF program that attempts to
> > update/delete elements in the same bucket"?
>
> raw_spin_lock_irqsave() is done after the percpu counter, so there is a gap but
> not sure if it matters though unless the production use case can really hit it.

Yeah unfortunately I'm going off an incident from last year and I
don't have this level of visibility into the failure scenario in a
prod-like environment today.