Re: [PATCH -tip v2] kprobes: Use synchronize_rcu_tasks() for optprobe with CONFIG_PREEMPT

[Paul E. McKenney]

On Tue, Sep 12, 2017 at 10:10:51AM +0900, Masami Hiramatsu wrote:
> To enable jump optimized probe with CONFIG_PREEMPT, use
> synchronize_rcu_tasks() to wait for all tasks preempted
> on trampoline code back on track.
> 
> Since the jump optimized kprobes can replace multiple
> instructions, there can be tasks which are interrupted
> on the 2nd (or 3rd) instructions. If the kprobe
> replaces those instructions by a jump instruction,
> when those tasks back to the interrupted place, it is
> a middle of the jump instruction and causes a kernel
> panic.
> To avoid such tragedies in advance, kprobe optimizer
> prepare a detour route using normal kprobe (e.g.
> int3 breakpoint on x86), and wait for the tasks which
> is interrrupted on such place by synchronize_sched()
> when CONFIG_PREEMPT=n.
> If CONFIG_PREEMPT=y, things be more complicated, because
> such interrupted thread can be preempted (other thread
> can be scheduled in interrupt handler.) So, kprobes
> optimizer has to wait for those tasks scheduled normally.
> In this case we can use synchronize_rcu_tasks() which
> ensures that all preempted tasks back on track and
> schedule it.
> 
> Signed-off-by: Masami Hiramatsu <mhiramat@xxxxxxxxxx>

Reviewed-by: Paul E. McKenney <paulmck@xxxxxxxxxxxxxxxxxx>

> ---
>  arch/Kconfig     |    2 +-
>  kernel/kprobes.c |   18 +++++++++++++-----
>  2 files changed, 14 insertions(+), 6 deletions(-)
> 
> diff --git a/arch/Kconfig b/arch/Kconfig
> index 2520ca5b42eb..d495c06ae961 100644
> --- a/arch/Kconfig
> +++ b/arch/Kconfig
> @@ -90,7 +90,7 @@ config STATIC_KEYS_SELFTEST
>  config OPTPROBES
>  	def_bool y
>  	depends on KPROBES && HAVE_OPTPROBES
> -	depends on !PREEMPT
> +	select TASKS_RCU if PREEMPT
> 
>  config KPROBES_ON_FTRACE
>  	def_bool y
> diff --git a/kernel/kprobes.c b/kernel/kprobes.c
> index a1606a4224e1..6243b8b02511 100644
> --- a/kernel/kprobes.c
> +++ b/kernel/kprobes.c
> @@ -574,12 +574,20 @@ static void kprobe_optimizer(struct work_struct *work)
> 
>  	/*
>  	 * Step 2: Wait for quiesence period to ensure all running interrupts
> -	 * are done. Because optprobe may modify multiple instructions
> -	 * there is a chance that Nth instruction is interrupted. In that
> -	 * case, running interrupt can return to 2nd-Nth byte of jump
> -	 * instruction. This wait is for avoiding it.
> +	 * are done. Because optprobe may modify multiple instructions,
> +	 * there is a chance that the Nth instruction is interrupted. In that
> +	 * case, running interrupt can return to the Nth byte of jump
> +	 * instruction. This can be avoided by waiting for returning of
> +	 * such interrupts, since (until here) the first byte of the optimized
> +	 * probe is already replaced with normal kprobe (sw breakpoint) and
> +	 * all threads which reach to the probed address will hit it and
> +	 * bypass the copied instructions instead of executing the original.
> +	 * With CONFIG_PREEMPT, such interrupts can be preepmted. To wait
> +	 * for such thread, we will use synchronize_rcu_tasks() which ensures
> +	 * all preeempted tasks are scheduled normally (= not preempted.)
> +	 * So we can ensure there is no threads preempted at probed address.
>  	 */
> -	synchronize_sched();
> +	synchronize_rcu_tasks();
> 
>  	/* Step 3: Optimize kprobes after quiesence period */
>  	do_optimize_kprobes();
>