Re: [PATCH V4 00/17] Stitch LBR call stack (Perf Tools)

From: Arnaldo Carvalho de Melo
Date: Fri Apr 17 2020 - 13:48:11 EST


Em Thu, Mar 19, 2020 at 01:25:00PM -0700, kan.liang@xxxxxxxxxxxxxxx escreveu:
> From: Kan Liang <kan.liang@xxxxxxxxxxxxxxx>
>
> Changes since V3:
> - There is no dependency among the 'capabilities'. If perf fails to read
> one, it should not impact others. Continue to parse the rest of caps.
> (Patch 1)
> - Use list_for_each_entry() to replace perf_pmu__scan_caps() (Patch 1 &
> 2)
> - Combine the declaration plus assignment when possible (Patch 1 & 2)
> - Add check for script/report/c2c.. (Patch 13, 14 & 16)
>
> Changes since V2:
> - Check strdup() in Patch 1
> - Split several patches into smaller patches
>
> Changes since V1:
> - Rebase on top of commit 5100c2b77049 ("perf header: Add check for
> unexpected use of reserved membrs in event attr")
> - Fix compling error with GCC9 in patch 1.
>
>
> 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")
>
> 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 maximum number of LBRs is
> required. Patch 1 - 4 retrieve the capabilities information from sysfs
> and save them in perf header.
>
> Patch 5 - 12 implements the LBR stitching approach.
>
> Users can use the options introduced in patch 13-16 to enable the LBR
> stitching approach for perf report, script, top and c2c.
>
> Patch 17 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

Next time provide the full test proggie, I had to expand those ... to
reproduce your results, all I have is in perf/core, some patches are
still to be processed, will continue later, have to stop now, see:

https://git.kernel.org/pub/scm/linux/kernel/git/acme/linux.git/commit/?id=13cfba6b741ff

For my testing, looks really great!

- Arnaldo

> 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 (17):
> perf pmu: Add support for PMU capabilities
> perf header: Support CPU PMU capabilities
> perf record: Clear HEADER_CPU_PMU_CAPS for non LBR call stack mode
> perf stat: Clear HEADER_CPU_PMU_CAPS
> perf machine: Remove the indent in resolve_lbr_callchain_sample
> perf machine: Refine the function for LBR call stack reconstruction
> perf machine: Factor out lbr_callchain_add_kernel_ip()
> perf machine: Factor out lbr_callchain_add_lbr_ip()
> perf thread: Add a knob for LBR stitch approach
> perf tools: Save previous sample for LBR stitching approach
> perf tools: Save previous cursor nodes for LBR stitching approach
> 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
>
> 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 | 12 +
> tools/perf/builtin-record.c | 3 +
> tools/perf/builtin-report.c | 12 +
> tools/perf/builtin-script.c | 12 +
> tools/perf/builtin-stat.c | 1 +
> tools/perf/builtin-top.c | 11 +
> tools/perf/util/branch.h | 19 +-
> tools/perf/util/callchain.h | 8 +
> tools/perf/util/env.h | 3 +
> tools/perf/util/header.c | 108 +++++
> tools/perf/util/header.h | 1 +
> tools/perf/util/hist.c | 23 +
> tools/perf/util/machine.c | 423 +++++++++++++++---
> tools/perf/util/pmu.c | 82 ++++
> tools/perf/util/pmu.h | 9 +
> tools/perf/util/sort.c | 2 +-
> tools/perf/util/sort.h | 2 +
> tools/perf/util/thread.c | 2 +
> tools/perf/util/thread.h | 35 ++
> tools/perf/util/top.h | 1 +
> 25 files changed, 757 insertions(+), 70 deletions(-)
>
> --
> 2.17.1
>

--

- Arnaldo