Re: [PATCH 00/12] Stitch LBR call stack (Perf Tools)

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

Em Fri, Feb 28, 2020 at 08:29:59AM -0800, kan.liang@linux.intel.com escreveu:
> From: Kan Liang <kan.liang@linux.intel.com>
> 
> The kernel patches have been merged into linux-next.
>   commit bbfd5e4fab63 ("perf/core: Add new branch sample type for HW
> index of raw branch records")
>   commit db278b90c326 ("perf/x86/intel: Output LBR TOS information
> correctly")

I saw it landed in tip/perf/core, going thru this patchset now.

Thanks,

- Arnaldo
 
> Start from Haswell, Linux perf can utilize the existing Last Branch
> Record (LBR) facility to record call stack. However, the depth of the
> reconstructed LBR call stack limits to the number of LBR registers.
> E.g. on skylake, the depth of reconstructed LBR call stack is <= 32
> That's because HW will overwrite the oldest LBR registers when it's
> full.
> 
> However, the overwritten LBRs may still be retrieved from previous
> sample. At that moment, HW hasn't overwritten the LBR registers yet.
> Perf tools can stitch those overwritten LBRs on current call stacks to
> get a more complete call stack.
> 
> To determine if LBRs can be stitched, the physical index of LBR
> registers is required. A new branch sample type is introduced to dump
> the physical index of the most recent LBR aka Top-of-Stack (TOS)
> information for perf tools.
> Patch 1 & 2 extend struct branch_stack to support the new branch sample
> type, PERF_SAMPLE_BRANCH_HW_INDEX.
> 
> Since the output format of PERF_SAMPLE_BRANCH_STACK will be changed
> when the new branch sample type is set, an older version of perf tool
> may parse the perf.data incorrectly. Furthermore, there is no warning
> if this case happens. Because current perf header never check for
> unknown input bits in attr. Patch 3 adds check for event attr. (Can be
> merged independently.)
> 
> Besides the physical index, the maximum number of LBRs is required as
> well. Patch 4 & 5 retrieve the capabilities information from sysfs
> and save them in perf header.
> 
> Patch 6 & 7 implements the LBR stitching approach.
> 
> Users can use the options introduced in patch 8-11 to enable the LBR
> stitching approach for perf report, script, top and c2c.
> 
> Patch 12 adds a fast path for duplicate entries check. It benefits all
> call stack parsing, not just for stitch LBR call stack. It can be
> merged independently.
> 
> 
> The stitching approach base on LBR call stack technology. The known
> limitations of LBR call stack technology still apply to the approach,
> e.g. Exception handing such as setjmp/longjmp will have calls/returns
> not match.
> This approach is not full proof. There can be cases where it creates
> incorrect call stacks from incorrect matches. There is no attempt
> to validate any matches in another way. So it is not enabled by default.
> However in many common cases with call stack overflows it can recreate
> better call stacks than the default lbr call stack output. So if there
> are problems with LBR overflows this is a possible workaround.
> 
> Regression:
> Users may collect LBR call stack on a machine with new perf tool and
> new kernel (support LBR TOS). However, they may parse the perf.data with
> old perf tool (not support LBR TOS). The old tool doesn't check
> attr.branch_sample_type. Users probably get incorrect information
> without any warning.
> 
> Performance impact:
> The processing time may increase with the LBR stitching approach
> enabled. The impact depends on the increased depth of call stacks.
> 
> For a simple test case tchain_edit with 43 depth of call stacks.
> perf record --call-graph lbr -- ./tchain_edit
> perf report --stitch-lbr
> 
> Without --stitch-lbr, perf report only display 32 depth of call stacks.
> With --stitch-lbr, perf report can display all 43 depth of call stacks.
> The depth of call stacks increase 34.3%.
> 
> Correspondingly, the processing time of perf report increases 39%,
> Without --stitch-lbr:                           11.0 sec
> With --stitch-lbr:                              15.3 sec
> 
> The source code of tchain_edit.c is something similar as below.
> noinline void f43(void)
> {
>         int i;
>         for (i = 0; i < 10000;) {
> 
>                 if(i%2)
>                         i++;
>                 else
>                         i++;
>         }
> }
> 
> noinline void f42(void)
> {
>         int i;
>         for (i = 0; i < 100; i++) {
>                 f43();
>                 f43();
>                 f43();
>         }
> }
> 
> noinline void f41(void)
> {
>         int i;
>         for (i = 0; i < 100; i++) {
>                 f42();
>                 f42();
>                 f42();
>         }
> }
> noinline void f40(void)
> {
>         f41();
> }
> 
> ... ...
> 
> noinline void f32(void)
> {
>         f33();
> }
> 
> noinline void f31(void)
> {
>         int i;
> 
>         for (i = 0; i < 10000; i++) {
>                 if(i%2)
>                         i++;
>                 else
>                         i++;
>         }
> 
>         f32();
> }
> 
> noinline void f30(void)
> {
>         f31();
> }
> 
> ... ...
> 
> noinline void f1(void)
> {
>         f2();
> }
> 
> int main()
> {
>         f1();
> }
> 
> Kan Liang (12):
>   perf tools: Add hw_idx in struct branch_stack
>   perf tools: Support PERF_SAMPLE_BRANCH_HW_INDEX
>   perf header: Add check for event attr
>   perf pmu: Add support for PMU capabilities
>   perf header: Support CPU PMU capabilities
>   perf machine: Refine the function for LBR call stack reconstruction
>   perf tools: Stitch LBR call stack
>   perf report: Add option to enable the LBR stitching approach
>   perf script: Add option to enable the LBR stitching approach
>   perf top: Add option to enable the LBR stitching approach
>   perf c2c: Add option to enable the LBR stitching approach
>   perf hist: Add fast path for duplicate entries check approach
> 
>  tools/include/uapi/linux/perf_event.h         |   8 +-
>  tools/perf/Documentation/perf-c2c.txt         |  11 +
>  tools/perf/Documentation/perf-report.txt      |  11 +
>  tools/perf/Documentation/perf-script.txt      |  11 +
>  tools/perf/Documentation/perf-top.txt         |   9 +
>  .../Documentation/perf.data-file-format.txt   |  16 +
>  tools/perf/builtin-c2c.c                      |   6 +
>  tools/perf/builtin-record.c                   |   3 +
>  tools/perf/builtin-report.c                   |   6 +
>  tools/perf/builtin-script.c                   |  76 ++--
>  tools/perf/builtin-stat.c                     |   1 +
>  tools/perf/builtin-top.c                      |  11 +
>  tools/perf/tests/sample-parsing.c             |   7 +-
>  tools/perf/util/branch.h                      |  27 +-
>  tools/perf/util/callchain.h                   |  12 +-
>  tools/perf/util/cs-etm.c                      |   1 +
>  tools/perf/util/env.h                         |   3 +
>  tools/perf/util/event.h                       |   1 +
>  tools/perf/util/evsel.c                       |  20 +-
>  tools/perf/util/evsel.h                       |   6 +
>  tools/perf/util/header.c                      | 147 ++++++
>  tools/perf/util/header.h                      |   1 +
>  tools/perf/util/hist.c                        |  26 +-
>  tools/perf/util/intel-pt.c                    |   2 +
>  tools/perf/util/machine.c                     | 424 +++++++++++++++---
>  tools/perf/util/perf_event_attr_fprintf.c     |   1 +
>  tools/perf/util/pmu.c                         |  87 ++++
>  tools/perf/util/pmu.h                         |  12 +
>  .../scripting-engines/trace-event-python.c    |  30 +-
>  tools/perf/util/session.c                     |   8 +-
>  tools/perf/util/sort.c                        |   2 +-
>  tools/perf/util/sort.h                        |   2 +
>  tools/perf/util/synthetic-events.c            |   6 +-
>  tools/perf/util/thread.c                      |   2 +
>  tools/perf/util/thread.h                      |  34 ++
>  tools/perf/util/top.h                         |   1 +
>  36 files changed, 900 insertions(+), 131 deletions(-)
> 
> -- 
> 2.17.1
> 

-- 

- Arnaldo