bpf: fix corruption on concurrent perf_event_output calls
When tracing and networking programs are both attached in the system and both use event-output helpers that eventually call into perf_event_output(), then we could end up in a situation where the tracing attached program runs in user context while a cls_bpf program is triggered on that same CPU out of softirq context. Since both rely on the same per-cpu perf_sample_data, we could potentially corrupt it. This can only ever happen in a combination of the two types; all tracing programs use a bpf_prog_active counter to bail out in case a program is already running on that CPU out of a different context. XDP and cls_bpf programs by themselves don't have this issue as they run in the same context only. Therefore, split both perf_sample_data so they cannot be accessed from each other. Fixes: 20b9d7ac ("bpf: avoid excessive stack usage for perf_sample_data") Reported-by: NAlexei Starovoitov <ast@fb.com> Signed-off-by: NDaniel Borkmann <daniel@iogearbox.net> Tested-by: NSong Liu <songliubraving@fb.com> Acked-by: NAlexei Starovoitov <ast@kernel.org> Signed-off-by: NAlexei Starovoitov <ast@kernel.org>
Showing
想要评论请 注册 或 登录