Allan Zhang wrote:
> BPF_PROG_TYPE_SOCK_OPS program can reenter bpf_event_output because it can
> be called from atomic and non-atomic contexts since we don't have
> bpf_prog_active to prevent it happen.
>
> This patch enables 3 level of nesting to support normal, irq and nmi
> context.
>
> We can easily reproduce the issue by running neper crr mode with 100 flows
> and 10 threads from neper client side.
>
> Here is the whole stack dump:
>
> [ 515.228898] WARNING: CPU: 20 PID: 14686 at kernel/trace/bpf_trace.c:549
> bpf_event_output+0x1f9/0x220
> [ 515.228903] CPU: 20 PID: 14686 Comm: tcp_crr Tainted: G W
> 4.15.0-smp-fixpanic #44
> [ 515.228904] Hardware name: Intel TBG,ICH10/Ikaria_QC_1b, BIOS 1.22.0
> 06/04/2018
> [ 515.228905] RIP: 0010:bpf_event_output+0x1f9/0x220
[...]
> Fixes: a5a3a828cd00 ("bpf: add perf event notificaton support for sock_ops")
>
> Effort: BPF
> Signed-off-by: Allan Zhang <[email protected]>
> Reviewed-by: Stanislav Fomichev <[email protected]>
> Reviewed-by: Eric Dumazet <[email protected]>
> ---
LGTM thanks.
Acked-by: John Fastabend <[email protected]>
