Re: [PATCH v1 3/3] perf tests: Disable bp_signal testing for arm64

[Arnaldo Carvalho de Melo <arnaldo.melo@gmail.com>]

Em Fri, Oct 18, 2019 at 04:55:31PM +0800, Leo Yan escreveu:
> As there have several discussions for enabling Perf breakpoint signal
> testing on arm64 platform; arm64 needs to rely on single-step to execute
> the breakpointed instruction and then reinstall the breakpoint exception
> handler.  But if hook the breakpoint with a signal, the signal handler
> will do the stepping rather than the breakpointed instruction, this
> causes infinite loops as below:
> 
>          Kernel space              |            Userspace
> -----------------------------------|--------------------------------
>                                    |  __test_function() -> hit
> 				   |                       breakpoint
>   breakpoint_handler()             |
>     `-> user_enable_single_step()  |
>   do_signal()                      |
>                                    |  sig_handler() -> Step one
> 				   |                instruction and
> 				   |                trap to kernel
>   single_step_handler()            |
>     `-> reinstall_suspended_bps()  |
>                                    |  __test_function() -> hit
> 				   |     breakpoint again and
> 				   |     repeat up flow infinitely
> 
> As Will Deacon mentioned [1]: "that we require the overflow handler to
> do the stepping on arm/arm64, which is relied upon by GDB/ptrace. The
> hw_breakpoint code is a complete disaster so my preference would be to
> rip out the perf part and just implement something directly in ptrace,
> but it's a pretty horrible job".  Though Will commented this on arm
> architecture, but the comment also can apply on arm64 architecture.
> 
> For complete information, I searched online and found a few years back,
> Wang Nan sent one patch 'arm64: Store breakpoint single step state into
> pstate' [2]; the patch tried to resolve this issue by avoiding single
> stepping in signal handler and defer to enable the signal stepping when
> return to __test_function().  The fixing was not merged due to the
> concern for missing to handle different usage cases.
> 
> Based on the info, the most feasible way is to skip Perf breakpoint
> signal testing for arm64 and this could avoid the duplicate
> investigation efforts when people see the failure.  This patch skips
> this case on arm64 platform, which is same with arm architecture.

Ok, applying,

- Arnaldo
 
> [1] https://lkml.org/lkml/2018/11/15/205
> [2] https://lkml.org/lkml/2015/12/23/477
> 
> Signed-off-by: Leo Yan <leo.yan@linaro.org>
> ---
>  tools/perf/tests/bp_signal.c | 15 ++++++---------
>  1 file changed, 6 insertions(+), 9 deletions(-)
> 
> diff --git a/tools/perf/tests/bp_signal.c b/tools/perf/tests/bp_signal.c
> index c1c2c13de254..166f411568a5 100644
> --- a/tools/perf/tests/bp_signal.c
> +++ b/tools/perf/tests/bp_signal.c
> @@ -49,14 +49,6 @@ asm (
>  	"__test_function:\n"
>  	"incq (%rdi)\n"
>  	"ret\n");
> -#elif defined (__aarch64__)
> -extern void __test_function(volatile long *ptr);
> -asm (
> -	".globl __test_function\n"
> -	"__test_function:\n"
> -	"str x30, [x0]\n"
> -	"ret\n");
> -
>  #else
>  static void __test_function(volatile long *ptr)
>  {
> @@ -302,10 +294,15 @@ bool test__bp_signal_is_supported(void)
>  	 * stepping into the SIGIO handler and getting stuck on the
>  	 * breakpointed instruction.
>  	 *
> +	 * Since arm64 has the same issue with arm for the single-step
> +	 * handling, this case also gets suck on the breakpointed
> +	 * instruction.
> +	 *
>  	 * Just disable the test for these architectures until these
>  	 * issues are resolved.
>  	 */
> -#if defined(__powerpc__) || defined(__s390x__) || defined(__arm__)
> +#if defined(__powerpc__) || defined(__s390x__) || defined(__arm__) || \
> +    defined(__aarch64__)
>  	return false;
>  #else
>  	return true;
> -- 
> 2.17.1

-- 

- Arnaldo