On Thu, Sep 17, 2020 at 8:46 AM Yonghong Song <yhs@xxxxxx> wrote:
>
> If a bucket contains a lot of sockets, during bpf_iter traversing
> a bucket, concurrent userspace bpf_map_update_elem() and
> bpf program bpf_sk_storage_{get,delete}() may experience
> some undesirable delays as they will compete with bpf_iter
> for bucket lock.
>
> Note that the number of buckets for bpf_sk_storage_map
> is roughly the same as the number of cpus. So if there
> are lots of sockets in the system, each bucket could
> contain lots of sockets.
>
> Different actual use cases may experience different delays.
> Here, using selftest bpf_iter subtest bpf_sk_storage_map,
> I hacked the kernel with ktime_get_mono_fast_ns()
> to collect the time when a bucket was locked
> during bpf_iter prog traversing that bucket. This way,
> the maximum incurred delay was measured w.r.t. the
> number of elements in a bucket.
> # elems in each bucket delay(ns)
> 64 17000
> 256 72512
> 2048 875246
>
> The potential delays will be further increased if
> we have even more elemnts in a bucket. Using rcu_read_lock()
> is a reasonable compromise here. It may lose some precision, e.g.,
> access stale sockets, but it will not hurt performance of
> bpf program or user space application which also tries
> to get/delete or update map elements.
>
> Cc: Martin KaFai Lau <kafai@xxxxxx>
> Acked-by: Song Liu <songliubraving@xxxxxx>
> Signed-off-by: Yonghong Song <yhs@xxxxxx>
> ---
> net/core/bpf_sk_storage.c | 31 +++++++++++++------------------
> 1 file changed, 13 insertions(+), 18 deletions(-)
>
> Changelog:
> v3 -> v4:
> - use rcu_dereference/hlist_next_rcu for hlist_entry_safe. (Martin)
Applied. Thanks
