Message-Id: <20191119143411.3482-1-kan.liang@linux.intel.com>
Date:   Tue, 19 Nov 2019 06:33:58 -0800
From:   kan.liang@...ux.intel.com
To:     peterz@...radead.org, acme@...hat.com, mingo@...nel.org,
        linux-kernel@...r.kernel.org
Cc:     jolsa@...nel.org, namhyung@...nel.org, vitaly.slobodskoy@...el.com,
        pavel.gerasimov@...el.com, ak@...ux.intel.com, eranian@...gle.com,
        mpe@...erman.id.au, Kan Liang <kan.liang@...ux.intel.com>
Subject: [PATCH V4 00/13] Stitch LBR call stack

From: Kan Liang <kan.liang@...ux.intel.com>


Changes since V3
- Add the new branch sample type at the end of enum
  perf_branch_sample_type.
- Rebase the user space patch on top of acme's perf/core branch

Changes since V2
- Move tos into struct perf_branch_stack

Changes since V1
- Add a new branch sample type for LBR TOS. Drop the sample type in V1.
- Add check in perf header to detect unknown input bits in event attr
- Save and use the LBR cursor nodes from previous sample to avoid
  duplicate calculation of cursor nodes.
- Add fast path for duplicate entries check. It benefits all call stack
  parsing, not just for stitch LBR call stack. It can be merged
  independetely.

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 in
patch 1 to 3 to dump the LBR Top-of-Stack (TOS) information for perf
tools.

Only when the new branch sample type is set, the TOS information is
dumped into the PERF_SAMPLE_BRANCH_STACK output. Perf tool should check
the attr.branch_sample_type, and apply the corresponding format for
PERF_SAMPLE_BRANCH_STACK samples. The check is introduced in Patch 4.

Besides, the maximum number of LBRs is required as well. Patch 5 & 6
retrieve the capabilities information from sysfs and save them in perf
header.

Patch 7 & 8 implements the LBR stitching approach.

Users can use the options introduced in patch 9-12 to enable the LBR
stitching approach for perf report, script, top and c2c.

Patch 13 adds fast path for duplicate entries check. It benefits all
call stack parsing, not just for stitch LBR call stack. It can be
merged independetely.


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 (13):
  perf/core: Add new branch sample type for LBR TOS
  perf/x86/intel: Output LBR TOS information
  perf tools: Support new branch sample type for LBR TOS
  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

 arch/x86/events/intel/lbr.c                   |   9 +
 include/linux/perf_event.h                    |   2 +
 include/uapi/linux/perf_event.h               |  16 +-
 kernel/events/core.c                          |  13 +-
 tools/include/uapi/linux/perf_event.h         |  16 +-
 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                   |   6 +
 tools/perf/builtin-stat.c                     |   1 +
 tools/perf/builtin-top.c                      |  11 +
 tools/perf/util/branch.h                      |   5 +-
 tools/perf/util/callchain.h                   |  12 +-
 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                      | 148 +++++++
 tools/perf/util/header.h                      |   1 +
 tools/perf/util/hist.c                        |  23 +
 tools/perf/util/machine.c                     | 409 +++++++++++++++---
 tools/perf/util/parse-branch-options.c        |   3 +-
 tools/perf/util/perf_event_attr_fprintf.c     |   3 +-
 tools/perf/util/pmu.c                         |  87 ++++
 tools/perf/util/pmu.h                         |  12 +
 tools/perf/util/sort.c                        |   2 +-
 tools/perf/util/sort.h                        |   2 +
 tools/perf/util/thread.c                      |   2 +
 tools/perf/util/thread.h                      |  34 ++
 tools/perf/util/top.h                         |   1 +
 35 files changed, 843 insertions(+), 78 deletions(-)

-- 
2.17.1

Powered by blists - more mailing lists