Re: [BPF PATCH for-next] cgroup/bpf: fast path for not loaded skb BPF filtering

[Martin KaFai Lau <kafai@fb.com>]

On Sat, Dec 11, 2021 at 01:15:05AM +0000, Pavel Begunkov wrote:
> On 12/11/21 00:38, Martin KaFai Lau wrote:
> > On Fri, Dec 10, 2021 at 02:23:34AM +0000, Pavel Begunkov wrote:
> > > cgroup_bpf_enabled_key static key guards from overhead in cases where
> > > no cgroup bpf program of a specific type is loaded in any cgroup. Turn
> > > out that's not always good enough, e.g. when there are many cgroups but
> > > ones that we're interesting in are without bpf. It's seen in server
> > > environments, but the problem seems to be even wider as apparently
> > > systemd loads some BPF affecting my laptop.
> > > 
> > > Profiles for small packet or zerocopy transmissions over fast network
> > > show __cgroup_bpf_run_filter_skb() taking 2-3%, 1% of which is from
> > > migrate_disable/enable(), and similarly on the receiving side. Also
> > > got +4-5% of t-put for local testing.
> > > 
> > > Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
> > > ---
> > >   include/linux/bpf-cgroup.h | 24 +++++++++++++++++++++---
> > >   kernel/bpf/cgroup.c        | 23 +++++++----------------
> > >   2 files changed, 28 insertions(+), 19 deletions(-)
> > > 
> > > diff --git a/include/linux/bpf-cgroup.h b/include/linux/bpf-cgroup.h
> > > index 11820a430d6c..99b01201d7db 100644
> > > --- a/include/linux/bpf-cgroup.h
> > > +++ b/include/linux/bpf-cgroup.h
> > > @@ -141,6 +141,9 @@ struct cgroup_bpf {
> > >   	struct list_head progs[MAX_CGROUP_BPF_ATTACH_TYPE];
> > >   	u32 flags[MAX_CGROUP_BPF_ATTACH_TYPE];
> > > +	/* for each type tracks whether effective prog array is not empty */
> > > +	unsigned long enabled_mask;
> > > +
> > >   	/* list of cgroup shared storages */
> > >   	struct list_head storages;
> > > @@ -219,11 +222,25 @@ int bpf_percpu_cgroup_storage_copy(struct bpf_map *map, void *key, void *value);
> > >   int bpf_percpu_cgroup_storage_update(struct bpf_map *map, void *key,
> > >   				     void *value, u64 flags);
> > > +static inline bool __cgroup_bpf_type_enabled(struct cgroup_bpf *cgrp_bpf,
> > > +					     enum cgroup_bpf_attach_type atype)
> > > +{
> > > +	return test_bit(atype, &cgrp_bpf->enabled_mask);
> > > +}
> > > +
> > > +#define CGROUP_BPF_TYPE_ENABLED(sk, atype)				       \
> > > +({									       \
> > > +	struct cgroup *__cgrp = sock_cgroup_ptr(&(sk)->sk_cgrp_data);	       \
> > > +									       \
> > > +	__cgroup_bpf_type_enabled(&__cgrp->bpf, (atype));		       \
> > > +})
> > I think it should directly test if the array is empty or not instead of
> > adding another bit.
> > 
> > Can the existing __cgroup_bpf_prog_array_is_empty(cgrp, ...) test be used instead?
> 
> That was the first idea, but it's still heavier than I'd wish. 0.3%-0.7%
> in profiles, something similar in reqs/s. rcu_read_lock/unlock() pair is
> cheap but anyway adds 2 barrier()s, and with bitmasks we can inline
> the check.
It sounds like there is opportunity to optimize
__cgroup_bpf_prog_array_is_empty().

How about using rcu_access_pointer(), testing with &empty_prog_array.hdr,
and then inline it?  The cgroup prog array cannot be all
dummy_bpf_prog.prog.  If that could be the case, it should be replaced
with &empty_prog_array.hdr earlier, so please check.