Re: [PATCH] perf/x86/intel/lbr: fix branch type encoding

From: Stephane Eranian
Date: Thu Aug 11 2022 - 11:28:51 EST


On Thu, Aug 11, 2022 at 5:42 PM Liang, Kan <kan.liang@xxxxxxxxxxxxxxx> wrote:
>
>
>
> On 2022-08-11 10:17 a.m., Stephane Eranian wrote:
> > On Thu, Aug 11, 2022 at 3:23 PM Liang, Kan <kan.liang@xxxxxxxxxxxxxxx> wrote:
> >>
> >>
> >>
> >> On 2022-08-10 5:06 p.m., Stephane Eranian wrote:
> >>> With architected LBR, the procesosr can record the type of each sampled taken
> >>> branch. The type is encoded in 4-bit field in the LBR_INFO MSR of each entry.
> >>>
> >>> The branch type must then extracted and saved in the perf_branch_entry in the
> >>> perf_events sampling buffer. With the current code, the raw Intel encoding of
> >>> the branch is exported to user tools.
> >>
> >> In the intel_pmu_lbr_filter(), the raw encoding will be converted into
> >> the X86_BR_* format via arch_lbr_br_type_map[]. Then the
> >> common_branch_type() will convert the X86_BR_* format to the generic
> >> PERF_BR_* type and expose to user tools.
> >>
> >> I double check the existing arch_lbr_br_type_map[] and branch_map[].
> >> They should generate the same PERF_BR_* type as your arch_lbr_type_map[].
> >>
> >> Is there a test case which I can use to reproduce the problem?
> >>
> > I was doing a simple:
> > $ perf record -b -e cpu/event=0xc4/ ....
> > $ perf report -D
> > Looking at the LBR information and the BR type, many entries has no branch type.
> > What I see is a function where you do: e->type = get_lbr_br_type() and
> > that is what
> > is then saved in the buffer. Unless I am missing a later patch.
> >
>
> To get the LBR type, the save_type filter option must be applied. See
> 60f83fa6341d ("perf record: Create a new option save_type in
> --branch-filter").
>
That seems overly complicated. I don't recall having to pass a new option
to get the LBR latency. It showed up automatically. So why for branch_type?

> The -b only include the ANY option. Maybe we should extend the -b option
> to ANY|SAVE_TYPE.
>
Ok, that explains it then. I think we need to simplify.


> Thanks,
> Kan
>
> >
> >> Thanks,
> >> Kan
> >>
> >>> Yet tools, such as perf, expected the
> >>> branch type to be encoded using perf_events branch type enum
> >>> (see tools/perf/util/branch.c). As a result of the discrepancy, the output of
> >>> perf report -D shows bogus branch types.
> >>>
> >>> Fix the problem by converting the Intel raw encoding into the perf_events
> >>> branch type enum values. With that in place and with no changes to the tools,
> >>> the branch types are now reported properly.
> >>>
> >>> Signed-off-by: Stephane Eranian <eranian@xxxxxxxxxx>
> >>> ---
> >>> arch/x86/events/intel/lbr.c | 35 ++++++++++++++++++++++++++++++++---
> >>> 1 file changed, 32 insertions(+), 3 deletions(-)
> >>>
> >>> diff --git a/arch/x86/events/intel/lbr.c b/arch/x86/events/intel/lbr.c
> >>> index 4f70fb6c2c1e..ef63d4d46b50 100644
> >>> --- a/arch/x86/events/intel/lbr.c
> >>> +++ b/arch/x86/events/intel/lbr.c
> >>> @@ -894,9 +894,23 @@ static DEFINE_STATIC_KEY_FALSE(x86_lbr_mispred);
> >>> static DEFINE_STATIC_KEY_FALSE(x86_lbr_cycles);
> >>> static DEFINE_STATIC_KEY_FALSE(x86_lbr_type);
> >>>
> >>> -static __always_inline int get_lbr_br_type(u64 info)
> >>> +/*
> >>> + * Array index encodes IA32_LBR_x_INFO Branch Type Encodings
> >>> + * as per Intel SDM Vol3b Branch Types section
> >>> + */
> >>> +static const int arch_lbr_type_map[]={
> >>> + [0] = PERF_BR_COND,
> >>> + [1] = PERF_BR_IND,
> >>> + [2] = PERF_BR_UNCOND,
> >>> + [3] = PERF_BR_IND_CALL,
> >>> + [4] = PERF_BR_CALL,
> >>> + [5] = PERF_BR_RET,
> >>> +};
> >>> +#define ARCH_LBR_TYPE_COUNT ARRAY_SIZE(arch_lbr_type_map)
> >>> +
> >>> +static __always_inline u16 get_lbr_br_type(u64 info)
> >>> {
> >>> - int type = 0;
> >>> + u16 type = 0;
> >>>
> >>> if (static_branch_likely(&x86_lbr_type))
> >>> type = (info & LBR_INFO_BR_TYPE) >> LBR_INFO_BR_TYPE_OFFSET;
> >>> @@ -904,6 +918,21 @@ static __always_inline int get_lbr_br_type(u64 info)
> >>> return type;
> >>> }
> >>>
> >>> +/*
> >>> + * The kernel cannot expose raw Intel branch type encodings because they are
> >>> + * not generic. Instead, the function below maps the encoding to the
> >>> + * perf_events user visible branch types.
> >>> + */
> >>> +static __always_inline int get_lbr_br_type_mapping(u64 info)
> >>> +{
> >>> + if (static_branch_likely(&x86_lbr_type)) {
> >>> + u16 raw_type = get_lbr_br_type(info);
> >>> + if (raw_type < ARCH_LBR_TYPE_COUNT)
> >>> + return arch_lbr_type_map[raw_type];
> >>> + }
> >>> + return PERF_BR_UNKNOWN;
> >>> +}
> >>> +
> >>> static __always_inline bool get_lbr_mispred(u64 info)
> >>> {
> >>> bool mispred = 0;
> >>> @@ -957,7 +986,7 @@ static void intel_pmu_store_lbr(struct cpu_hw_events *cpuc,
> >>> e->in_tx = !!(info & LBR_INFO_IN_TX);
> >>> e->abort = !!(info & LBR_INFO_ABORT);
> >>> e->cycles = get_lbr_cycles(info);
> >>> - e->type = get_lbr_br_type(info);
> >>> + e->type = get_lbr_br_type_mapping(info);
> >>> }
> >>>
> >>> cpuc->lbr_stack.nr = i;