RE: [PATCH] sample/hw_breakpoint: avoid sample hw_breakpoint recursion for arm/arm64

[wangxu (AE)]

-----Original Message-----
From: Peter Zijlstra [mailto:peterz@xxxxxxxxxxxxx] 
Sent: Thursday, September 26, 2019 5:14 PM
To: wangxu (AE) <wangxu72@xxxxxxxxxx>
Cc: mingo@xxxxxxxxxx; acme@xxxxxxxxxx; mark.rutland@xxxxxxx; alexander.shishkin@xxxxxxxxxxxxxxx; namhyung@xxxxxxxxxx; gregkh@xxxxxxxxxxxxxxxxxxx; tglx@xxxxxxxxxxxxx; rfontana@xxxxxxxxxx; allison@xxxxxxxxxxx; linux-kernel@xxxxxxxxxxxxxxx
Subject: Re: [PATCH] sample/hw_breakpoint: avoid sample hw_breakpoint recursion for arm/arm64

On Mon, Sep 23, 2019 at 04:09:35PM +0800, wangxu wrote:
> From: Wang Xu <wangxu72@xxxxxxxxxx>
> 
> For x86/ppc, hw_breakpoint is triggered after the instruction is 
> executed.
> 
> For arm/arm64, which is triggered before the instruction executed.
> Arm/arm64 skips the instruction by using single step. But it only 
> supports default overflow_handler.

Where is the recusion.. ?

For arm/arm64, hw_breakpoint is triggered before the instruction executed.
When instruction_A is triggered, watchpoint_handler() will deal with this exception, and after return instruction_A will be triggerd ...

One using samples/hw_breakpoint/data_breakpoint.c in arm/arm64 will meet this problem.


> This patch provides a chance to avoid sample hw_breakpoint recursion 
> for arm/arm64 by adding 'struct perf_event_attr.bp_step'.

This patch also lacks justification for why this needs to come with ABI changes. There is also a distinct lack of comments.

I agree too. but have no better idea... 
This problem is really a big pit, especially for one not familiar with implementation differences in hw breakpoint for different architectures.



> Signed-off-by: Wang Xu <wangxu72@xxxxxxxxxx>
> ---
>  include/linux/perf_event.h              | 3 +++
>  include/uapi/linux/perf_event.h         | 3 ++-
>  samples/hw_breakpoint/data_breakpoint.c | 1 +
>  3 files changed, 6 insertions(+), 1 deletion(-)
> 
> diff --git a/include/linux/perf_event.h b/include/linux/perf_event.h 
> index 61448c1..f270eb7 100644
> --- a/include/linux/perf_event.h
> +++ b/include/linux/perf_event.h
> @@ -1024,6 +1024,9 @@ extern int perf_event_output(struct perf_event *event,
>  		return true;
>  	if (unlikely(event->overflow_handler == perf_event_output_backward))
>  		return true;
> +	/* avoid sample hw_breakpoint recursion */
> +	if (unlikely(event->attr.bp_step))
> +		return true;

This is just _wrong_.. it says that every event with bp_step set always is a 'default overflow handler', irrespective of what the overflow handler actually is.

Thanks for comments.

Function is_default_overflow_handler() was introduced in 1879445dfa7bbd6fe21b09c5cc72f4934798afed , which is only be called in arch/arm[64]/kernel/hw_breakpoint.c, and will never be used in other arch/ (I think). 

But keeping is_default_overflow_handler() unchanged, changing ' if (is_default_overflow_handler(bp)) ' to ' if (is_default_overflow_handler(bp) || unlikely(event->attr.bp_step) ) ' will be better in arch/arm[64]/kernel/hw_breakpoint.c.

>  	return false;
>  }
>