On Sat, Mar 07, 2020 at 12:02:12PM +0900, Masami Hiramatsu wrote:
> Prohibit probing on rcu_nmi_exit() and ist_exit() which
> are called from do_int3()'s kprobe path after kprobe_int3_handler().
>
> The commit c13324a505c7 ("x86/kprobes: Prohibit probing on
> functions before kprobe_int3_handler()") tried to fix similar
> issue, but it only marks the functions before kprobe_int3_handler()
> in do_int3().
>
> If we put a kprobe on rcu_nmi_exit() or ist_exit(), the kprobes
> will detect reentrance. However, it only skips the kprobe handler,
> exits from do_int3() and hits ist_exit() and rcu_nmi_exit() again.
> Thus, it causes another int3 exception and finally we will get
> the kernel panic with "Unrecoverable kprobe detected." error message.
>
> This is reproducible by the following commands.
>
> / # echo 0 > /proc/sys/debug/kprobes-optimization
> / # echo p vfs_read > /sys/kernel/debug/tracing/kprobe_events
> / # echo p rcu_nmi_exit >> /sys/kernel/debug/tracing/kprobe_events
> / # echo 1 > /sys/kernel/debug/tracing/events/kprobes/enable
>
> Fixes: c13324a505c7 ("x86/kprobes: Prohibit probing on functions before kprobe_int3_handler()")
> Signed-off-by: Masami Hiramatsu <mhiramat@xxxxxxxxxx>
> Cc: stable@xxxxxxxxxxxxxxx
>From an RCU perspective:
Reviewed-by: Paul E. McKenney <paulmck@xxxxxxxxxx>
> ---
> arch/x86/kernel/traps.c | 1 +
> kernel/rcu/tree.c | 1 +
> 2 files changed, 2 insertions(+)
>
> diff --git a/arch/x86/kernel/traps.c b/arch/x86/kernel/traps.c
> index 6ef00eb6fbb9..c63fb7697794 100644
> --- a/arch/x86/kernel/traps.c
> +++ b/arch/x86/kernel/traps.c
> @@ -115,6 +115,7 @@ void ist_exit(struct pt_regs *regs)
> if (!user_mode(regs))
> rcu_nmi_exit();
> }
> +NOKPROBE_SYMBOL(ist_exit);
>
> /**
> * ist_begin_non_atomic() - begin a non-atomic section in an IST exception
> diff --git a/kernel/rcu/tree.c b/kernel/rcu/tree.c
> index d91c9156fab2..c49ea0e919f9 100644
> --- a/kernel/rcu/tree.c
> +++ b/kernel/rcu/tree.c
> @@ -670,6 +670,7 @@ void rcu_nmi_exit(void)
> {
> rcu_nmi_exit_common(false);
> }
> +NOKPROBE_SYMBOL(rcu_nmi_exit);
>
> /**
> * rcu_irq_exit - inform RCU that current CPU is exiting irq towards idle
>
