He Fengqing wrote:
> It seems inc_misses_counter() suffers from same issue fixed in
> the commit d979617aa84d ("bpf: Fixes possible race in update_prog_stats()
> for 32bit arches"):
> As it can run while interrupts are enabled, it could
> be re-entered and the u64_stats syncp could be mangled.
>
> Fixes: 9ed9e9ba2337 ("bpf: Count the number of times recursion was prevented")
> Signed-off-by: He Fengqing <hefengqing@xxxxxxxxxx>
> ---
> kernel/bpf/trampoline.c | 5 +++--
> 1 file changed, 3 insertions(+), 2 deletions(-)
Appears possible through sleepable progs.
Acked-by: John Fastabend <john.fastabend@xxxxxxxxx>
> diff --git a/kernel/bpf/trampoline.c b/kernel/bpf/trampoline.c
> index 4b6974a195c1..5e7edf913060 100644
> --- a/kernel/bpf/trampoline.c
> +++ b/kernel/bpf/trampoline.c
> @@ -550,11 +550,12 @@ static __always_inline u64 notrace bpf_prog_start_time(void)
> static void notrace inc_misses_counter(struct bpf_prog *prog)
> {
> struct bpf_prog_stats *stats;
> + unsigned int flags;
>
> stats = this_cpu_ptr(prog->stats);
> - u64_stats_update_begin(&stats->syncp);
> + flags = u64_stats_update_begin_irqsave(&stats->syncp);
> u64_stats_inc(&stats->misses);
> - u64_stats_update_end(&stats->syncp);
> + u64_stats_update_end_irqrestore(&stats->syncp, flags);
> }
>
> /* The logic is similar to bpf_prog_run(), but with an explicit
> --
> 2.25.1
>
