Re: [PATCH] perf/amd/ibs: Use interrupt regs ip for stack unwinding

From: Namhyung Kim
Date: Mon May 02 2022 - 02:08:24 EST


Hello,

On Thu, Apr 28, 2022 at 10:15 PM Ravi Bangoria <ravi.bangoria@xxxxxxx> wrote:
>
> IbsOpRip is recorded when IBS interrupt is triggered. But there is
> a skid from the time IBS interrupt gets triggered to the time the
> interrupt is presented to the core. Meanwhile processor would have
> moved ahead and thus IbsOpRip will be inconsistent with rsp and rbp
> recorded as part of the interrupt regs. This causes issues while
> unwinding stack using the ORC unwinder as it needs consistent rip,
> rsp and rbp. Fix this by using rip from interrupt regs instead of
> IbsOpRip for stack unwinding.
>
> Fixes: ee9f8fce99640 ("x86/unwind: Add the ORC unwinder")
> Reported-by: Dmitry Monakhov <dmtrmonakhov@xxxxxxxxxxxxxx>
> Suggested-by: Peter Zijlstra <peterz@xxxxxxxxxxxxx>
> Signed-off-by: Ravi Bangoria <ravi.bangoria@xxxxxxx>

Acked-by: Namhyung Kim <namhyung@xxxxxxxxxx>

Thanks,
Namhyung


> ---
> arch/x86/events/amd/ibs.c | 18 ++++++++++++++++++
> 1 file changed, 18 insertions(+)
>
> diff --git a/arch/x86/events/amd/ibs.c b/arch/x86/events/amd/ibs.c
> index 9739019d4b67..171941043f53 100644
> --- a/arch/x86/events/amd/ibs.c
> +++ b/arch/x86/events/amd/ibs.c
> @@ -304,6 +304,16 @@ static int perf_ibs_init(struct perf_event *event)
> hwc->config_base = perf_ibs->msr;
> hwc->config = config;
>
> + /*
> + * rip recorded by IbsOpRip will not be consistent with rsp and rbp
> + * recorded as part of interrupt regs. Thus we need to use rip from
> + * interrupt regs while unwinding call stack. Setting _EARLY flag
> + * makes sure we unwind call-stack before perf sample rip is set to
> + * IbsOpRip.
> + */
> + if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN)
> + event->attr.sample_type |= __PERF_SAMPLE_CALLCHAIN_EARLY;
> +
> return 0;
> }
>
> @@ -687,6 +697,14 @@ static int perf_ibs_handle_irq(struct perf_ibs *perf_ibs, struct pt_regs *iregs)
> data.raw = &raw;
> }
>
> + /*
> + * rip recorded by IbsOpRip will not be consistent with rsp and rbp
> + * recorded as part of interrupt regs. Thus we need to use rip from
> + * interrupt regs while unwinding call stack.
> + */
> + if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN)
> + data.callchain = perf_callchain(event, iregs);
> +
> throttle = perf_event_overflow(event, &data, &regs);
> out:
> if (throttle) {
> --
> 2.27.0
>