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

From: Arnaldo Carvalho de Melo
Date: Fri Aug 12 2022 - 15:35:50 EST


Em Thu, Aug 11, 2022 at 11:56:56AM -0400, Liang, Kan escreveu:
>
>
> On 2022-08-11 11:33 a.m., Stephane Eranian wrote:
> > On Thu, Aug 11, 2022 at 6:28 PM Stephane Eranian <eranian@xxxxxxxxxx> wrote:
> >>
> >> 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.
> >>
> > In fact, I don't see a case where you would not benefit from the branch type.
> > Furthermore, not having the branch type DOES NOT save any space in the
> > branch record (given we have a reserved field). So I think I prefer not having
> > to specify yet another cmdline option to get the branch type.
>
>
> I think the option is to avoid the overhead of disassembling of branch
> instruction. See eb0baf8a0d92 ("perf/core: Define the common branch type
> classification")
> "Since the disassembling of branch instruction needs some overhead,
> a new PERF_SAMPLE_BRANCH_TYPE_SAVE is introduced to indicate if it
> needs to disassemble the branch instruction and record the branch
> type."
>
> I have no idea how big the overhead is. If we can always be benefit from
> the branch type. I guess we can make it default on.

Would you be so nice as to add a paragraph with such explanation to the
man page so that what happened to Stephane doesn't trip other people?

- Arnaldo

> > In fact, if you do
> > not pass the option, then perf report -D reports some bogus branch types, i.e.,
> > not all entries have empty types.
>
> Yes, that's an issue for Arch LBR. If the overhead is not a problem, it
> should be gone after we make the SAVE_TYPE default. Otherwise, I think
> we need a patch to clear the fields if the SAVE_TYPE is not applied.
>
> 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;

--

- Arnaldo