On 12/15, Pavel Begunkov wrote:
On 12/15/21 16:51, sdf@xxxxxxxxxx wrote:
> On 12/15, Pavel Begunkov wrote:
> > Add per socket fast path for not enabled BPF skb filtering, which
sheds
> > a nice chunk of send/recv overhead when affected. Testing udp with 128
> > byte payload and/or zerocopy with any payload size showed 2-3%
> > improvement in requests/s on the tx side using fast NICs across
network,
> > and around 4% for dummy device. Same goes for rx, not measured, but
> > numbers should be relatable.
> > In my understanding, this should affect a good share of machines, and
at
> > least it includes my laptops and some checked servers.
>
> > The core of the problem is that even though there is
> > cgroup_bpf_enabled_key guarding from __cgroup_bpf_run_filter_skb()
> > overhead, there are cases where we have several cgroups and loading a
> > BPF program to one also makes all others to go through the slow path
> > even when they don't have any BPF attached. It's even worse, because
> > apparently systemd or some other early init loads some BPF and so
> > triggers exactly this situation for normal networking.
>
> > Signed-off-by: Pavel Begunkov <asml.silence@xxxxxxxxx>
> > ---
>
> > v2: replace bitmask appoach with empty_prog_array (suggested by
Martin)
> > v3: add "bpf_" prefix to empty_prog_array (Martin)
>
> > � include/linux/bpf-cgroup.h | 24 +++++++++++++++++++++---
> > � include/linux/bpf.h������� | 13 +++++++++++++
> > � kernel/bpf/cgroup.c������� | 18 ++----------------
> > � kernel/bpf/core.c��������� | 16 ++++------------
> > � 4 files changed, 40 insertions(+), 31 deletions(-)
>
> > diff --git a/include/linux/bpf-cgroup.h b/include/linux/bpf-cgroup.h
> > index 11820a430d6c..c6dacdbdf565 100644
> > --- a/include/linux/bpf-cgroup.h
> > +++ b/include/linux/bpf-cgroup.h
> > @@ -219,11 +219,28 @@ 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_prog_array_is_empty(struct cgroup_bpf *cgrp_bpf,
> > +���������������� enum cgroup_bpf_attach_type type)
> > +{
> > +��� struct bpf_prog_array *array =
rcu_access_pointer(cgrp_bpf->effective[type]);
> > +
> > +��� return array == &bpf_empty_prog_array.hdr;
> > +}
> > +
> > +#define CGROUP_BPF_TYPE_ENABLED(sk, atype)���������������������� \
> > +({������������������������������������������ \
> > +��� struct cgroup *__cgrp =
sock_cgroup_ptr(&(sk)->sk_cgrp_data);���������� \
> > +������������������������������������������ \
> > +��� !__cgroup_bpf_prog_array_is_empty(&__cgrp->bpf,
(atype));���������� \
> > +})
> > +
> > � /* Wrappers for __cgroup_bpf_run_filter_skb() guarded by
cgroup_bpf_enabled. */
> > � #define BPF_CGROUP_RUN_PROG_INET_INGRESS(sk, skb)����������������� \
> > � ({����������������������������������������� \
> > ����� int __ret = 0;��������������������������������� \
> > -��� if (cgroup_bpf_enabled(CGROUP_INET_INGRESS))������������� \
> > +��� if (cgroup_bpf_enabled(CGROUP_INET_INGRESS) && sk
&&������������� \
> > +������� CGROUP_BPF_TYPE_ENABLED((sk),
CGROUP_INET_INGRESS))���������� \
>
> Why not add this __cgroup_bpf_run_filter_skb check to
> __cgroup_bpf_run_filter_skb? Result of sock_cgroup_ptr() is already
there
> and you can use it. Maybe move the things around if you want
> it to happen earlier.
For inlining. Just wanted to get it done right, otherwise I'll likely be
returning to it back in a few months complaining that I see measurable
overhead from the function call :)
Do you expect that direct call to bring any visible overhead?
Would be nice to compare that inlined case vs
__cgroup_bpf_prog_array_is_empty inside of __cgroup_bpf_run_filter_skb
while you're at it (plus move offset initialization down?).
