Re: [tip: perf/core] perf tests: Disable bp_signal testing for arm64

From: Will Deacon
Date: Tue Oct 22 2019 - 09:14:34 EST


On Mon, Oct 21, 2019 at 11:18:54PM -0000, tip-bot2 for Leo Yan wrote:
> The following commit has been merged into the perf/core branch of tip:
>
> Commit-ID: 6a5f3d94cb69a185b921cb92c39888dc31009acb
> Gitweb: https://git.kernel.org/tip/6a5f3d94cb69a185b921cb92c39888dc31009acb
> Author: Leo Yan <leo.yan@xxxxxxxxxx>
> AuthorDate: Fri, 18 Oct 2019 16:55:31 +08:00
> Committer: Arnaldo Carvalho de Melo <acme@xxxxxxxxxx>
> CommitterDate: Sat, 19 Oct 2019 15:35:01 -03:00
>
> perf tests: Disable bp_signal testing for arm64
>
> As there are 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 we 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.
>
> [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@xxxxxxxxxx>
> Cc: Adrian Hunter <adrian.hunter@xxxxxxxxx>
> Cc: Alexander Shishkin <alexander.shishkin@xxxxxxxxxxxxxxx>
> Cc: Brajeswar Ghosh <brajeswar.linux@xxxxxxxxx>
> Cc: Florian Fainelli <f.fainelli@xxxxxxxxx>
> Cc: Jiri Olsa <jolsa@xxxxxxxxxx>
> Cc: Mark Rutland <mark.rutland@xxxxxxx>
> Cc: Michael Petlan <mpetlan@xxxxxxxxxx>
> Cc: Namhyung Kim <namhyung@xxxxxxxxxx>
> Cc: Peter Zijlstra <peterz@xxxxxxxxxxxxx>
> Cc: Song Liu <songliubraving@xxxxxx>
> Cc: Souptick Joarder <jrdr.linux@xxxxxxxxx>
> Cc: Will Deacon <will@xxxxxxxxxx>
> Link: http://lore.kernel.org/lkml/20191018085531.6348-3-leo.yan@xxxxxxxxxx
> Signed-off-by: Arnaldo Carvalho de Melo <acme@xxxxxxxxxx>
> ---
> 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 c1c2c13..166f411 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.

Freudian slip?

Will