On Wed, Sep 18, 2019 at 05:51:10AM +0000, Yonghong Song wrote: > > Adding cc to bpf@xxxxxxxxxxxxxxx mailing list since this is really > bpf related. > > On 9/17/19 10:24 PM, jinshan.xiong@xxxxxxxxx wrote: > > From: Jinshan Xiong <jinshan.xiong@xxxxxxxxx> > > > > Invoking bpf program only if kprobe based perf_event has been added into > > the percpu list. This is essential to make event tracing for cgroup to work > > properly. > > The issue actually exists for bpf programs with kprobe, uprobe, > tracepoint and trace_syscall_enter/exit. > > In all these places, bpf program is called regardless of > whether there are perf events or not on this cpu. > This provides bpf programs more opportunities to see > the events. I guess this is by design. > Alexei/Daniel, could you clarify? Yes. It is by design. When bpf is attached to kprobe it will fire on all cpus. Per-cpu or per-task or per-cgroup filtering is already done inside bpf programs. We cannot make this change now it will break all users.
