[PATCH 0/4] perf lock contention: Improve call stack handling (v1)

Namhyung Kim posted 4 patches 3 years, 7 months ago
tools/perf/Documentation/perf-lock.txt        |  6 ++
tools/perf/builtin-lock.c                     | 89 ++++++++++++++-----
tools/perf/util/bpf_lock_contention.c         | 21 +++--
.../perf/util/bpf_skel/lock_contention.bpf.c  |  3 +-
tools/perf/util/lock-contention.h             |  3 +
5 files changed, 96 insertions(+), 26 deletions(-)
[PATCH 0/4] perf lock contention: Improve call stack handling (v1)
Posted by Namhyung Kim 3 years, 7 months ago
Hello,

I found that call stack from the lock tracepoint (using bpf_get_stackid)
can be different on each configuration.  For example it's very different
when I run it on a VM than on a real machine.

The perf lock contention relies on the stack trace to get the lock
caller names, this kind of difference can be annoying.  Ideally we could
skip stack trace entries for internal BPF or lock functions and get the
correct caller, but it's not the case as of today.  Currently it's hard
coded to control the behavior of stack traces for the lock contention
tracepoints.

To handle those differences, add two new options to control the number of
stack entries and how many it skips.  The default value worked well on
my VM setup, but I had to use --stack-skip=5 on real machines.

You can get it from 'perf/lock-stack-v1' branch in

  git://git.kernel.org/pub/scm/linux/kernel/git/namhyung/linux-perf.git

Thanks,
Namhyung


Namhyung Kim (4):
  perf lock contention: Factor out get_symbol_name_offset()
  perf lock contention: Show full callstack with -v option
  perf lock contention: Allow to change stack depth and skip
  perf lock contention: Skip stack trace from BPF

 tools/perf/Documentation/perf-lock.txt        |  6 ++
 tools/perf/builtin-lock.c                     | 89 ++++++++++++++-----
 tools/perf/util/bpf_lock_contention.c         | 21 +++--
 .../perf/util/bpf_skel/lock_contention.bpf.c  |  3 +-
 tools/perf/util/lock-contention.h             |  3 +
 5 files changed, 96 insertions(+), 26 deletions(-)


base-commit: 6c3bd8d3e01d9014312caa52e4ef1c29d5249648
-- 
2.37.2.789.g6183377224-goog
Re: [PATCH 0/4] perf lock contention: Improve call stack handling (v1)
Posted by Arnaldo Carvalho de Melo 3 years, 6 months ago
Em Wed, Sep 07, 2022 at 11:37:50PM -0700, Namhyung Kim escreveu:
> Hello,
> 
> I found that call stack from the lock tracepoint (using bpf_get_stackid)
> can be different on each configuration.  For example it's very different
> when I run it on a VM than on a real machine.
> 
> The perf lock contention relies on the stack trace to get the lock
> caller names, this kind of difference can be annoying.  Ideally we could
> skip stack trace entries for internal BPF or lock functions and get the
> correct caller, but it's not the case as of today.  Currently it's hard
> coded to control the behavior of stack traces for the lock contention
> tracepoints.
> 
> To handle those differences, add two new options to control the number of
> stack entries and how many it skips.  The default value worked well on
> my VM setup, but I had to use --stack-skip=5 on real machines.
> 
> You can get it from 'perf/lock-stack-v1' branch in
> 
>   git://git.kernel.org/pub/scm/linux/kernel/git/namhyung/linux-perf.git

This clashed with a patch you Acked earlier, so lets see if someone has
extra review comments and a v2 become needed for other reason, when you
can refresh it, ok?

- Arnaldo
 
> Thanks,
> Namhyung
> 
> 
> Namhyung Kim (4):
>   perf lock contention: Factor out get_symbol_name_offset()
>   perf lock contention: Show full callstack with -v option
>   perf lock contention: Allow to change stack depth and skip
>   perf lock contention: Skip stack trace from BPF
> 
>  tools/perf/Documentation/perf-lock.txt        |  6 ++
>  tools/perf/builtin-lock.c                     | 89 ++++++++++++++-----
>  tools/perf/util/bpf_lock_contention.c         | 21 +++--
>  .../perf/util/bpf_skel/lock_contention.bpf.c  |  3 +-
>  tools/perf/util/lock-contention.h             |  3 +
>  5 files changed, 96 insertions(+), 26 deletions(-)
> 
> 
> base-commit: 6c3bd8d3e01d9014312caa52e4ef1c29d5249648
> -- 
> 2.37.2.789.g6183377224-goog

-- 

- Arnaldo
Re: [PATCH 0/4] perf lock contention: Improve call stack handling (v1)
Posted by Namhyung Kim 3 years, 6 months ago
Hi Arnaldo,

On Thu, Sep 8, 2022 at 11:43 AM Arnaldo Carvalho de Melo
<acme@kernel.org> wrote:
>
> Em Wed, Sep 07, 2022 at 11:37:50PM -0700, Namhyung Kim escreveu:
> > Hello,
> >
> > I found that call stack from the lock tracepoint (using bpf_get_stackid)
> > can be different on each configuration.  For example it's very different
> > when I run it on a VM than on a real machine.
> >
> > The perf lock contention relies on the stack trace to get the lock
> > caller names, this kind of difference can be annoying.  Ideally we could
> > skip stack trace entries for internal BPF or lock functions and get the
> > correct caller, but it's not the case as of today.  Currently it's hard
> > coded to control the behavior of stack traces for the lock contention
> > tracepoints.
> >
> > To handle those differences, add two new options to control the number of
> > stack entries and how many it skips.  The default value worked well on
> > my VM setup, but I had to use --stack-skip=5 on real machines.
> >
> > You can get it from 'perf/lock-stack-v1' branch in
> >
> >   git://git.kernel.org/pub/scm/linux/kernel/git/namhyung/linux-perf.git
>
> This clashed with a patch you Acked earlier, so lets see if someone has
> extra review comments and a v2 become needed for other reason, when you
> can refresh it, ok?

Sounds good!

Thanks,
Namhyung
Re: [PATCH 0/4] perf lock contention: Improve call stack handling (v1)
Posted by Arnaldo Carvalho de Melo 3 years, 6 months ago
Em Thu, Sep 08, 2022 at 04:44:15PM -0700, Namhyung Kim escreveu:
> Hi Arnaldo,
> 
> On Thu, Sep 8, 2022 at 11:43 AM Arnaldo Carvalho de Melo
> <acme@kernel.org> wrote:
> >
> > Em Wed, Sep 07, 2022 at 11:37:50PM -0700, Namhyung Kim escreveu:
> > > Hello,
> > >
> > > I found that call stack from the lock tracepoint (using bpf_get_stackid)
> > > can be different on each configuration.  For example it's very different
> > > when I run it on a VM than on a real machine.
> > >
> > > The perf lock contention relies on the stack trace to get the lock
> > > caller names, this kind of difference can be annoying.  Ideally we could
> > > skip stack trace entries for internal BPF or lock functions and get the
> > > correct caller, but it's not the case as of today.  Currently it's hard
> > > coded to control the behavior of stack traces for the lock contention
> > > tracepoints.
> > >
> > > To handle those differences, add two new options to control the number of
> > > stack entries and how many it skips.  The default value worked well on
> > > my VM setup, but I had to use --stack-skip=5 on real machines.
> > >
> > > You can get it from 'perf/lock-stack-v1' branch in
> > >
> > >   git://git.kernel.org/pub/scm/linux/kernel/git/namhyung/linux-perf.git
> >
> > This clashed with a patch you Acked earlier, so lets see if someone has
> > extra review comments and a v2 become needed for other reason, when you
> > can refresh it, ok?
> 
> Sounds good!

Have you resubmitted this? /me goes on the backlog...

- Arnaldo
Re: [PATCH 0/4] perf lock contention: Improve call stack handling (v1)
Posted by Namhyung Kim 3 years, 6 months ago
On Tue, Sep 20, 2022 at 1:22 PM Arnaldo Carvalho de Melo
<acme@kernel.org> wrote:
>
> Em Thu, Sep 08, 2022 at 04:44:15PM -0700, Namhyung Kim escreveu:
> > Hi Arnaldo,
> >
> > On Thu, Sep 8, 2022 at 11:43 AM Arnaldo Carvalho de Melo
> > <acme@kernel.org> wrote:
> > >
> > > Em Wed, Sep 07, 2022 at 11:37:50PM -0700, Namhyung Kim escreveu:
> > > > Hello,
> > > >
> > > > I found that call stack from the lock tracepoint (using bpf_get_stackid)
> > > > can be different on each configuration.  For example it's very different
> > > > when I run it on a VM than on a real machine.
> > > >
> > > > The perf lock contention relies on the stack trace to get the lock
> > > > caller names, this kind of difference can be annoying.  Ideally we could
> > > > skip stack trace entries for internal BPF or lock functions and get the
> > > > correct caller, but it's not the case as of today.  Currently it's hard
> > > > coded to control the behavior of stack traces for the lock contention
> > > > tracepoints.
> > > >
> > > > To handle those differences, add two new options to control the number of
> > > > stack entries and how many it skips.  The default value worked well on
> > > > my VM setup, but I had to use --stack-skip=5 on real machines.
> > > >
> > > > You can get it from 'perf/lock-stack-v1' branch in
> > > >
> > > >   git://git.kernel.org/pub/scm/linux/kernel/git/namhyung/linux-perf.git
> > >
> > > This clashed with a patch you Acked earlier, so lets see if someone has
> > > extra review comments and a v2 become needed for other reason, when you
> > > can refresh it, ok?
> >
> > Sounds good!
>
> Have you resubmitted this? /me goes on the backlog...

Yep :)

https://lore.kernel.org/r/20220912055314.744552-1-namhyung@kernel.org
Re: [PATCH 0/4] perf lock contention: Improve call stack handling (v1)
Posted by Arnaldo Carvalho de Melo 3 years, 6 months ago
Em Tue, Sep 20, 2022 at 02:04:47PM -0700, Namhyung Kim escreveu:
> On Tue, Sep 20, 2022 at 1:22 PM Arnaldo Carvalho de Melo
> <acme@kernel.org> wrote:
> >
> > Em Thu, Sep 08, 2022 at 04:44:15PM -0700, Namhyung Kim escreveu:
> > > Hi Arnaldo,
> > >
> > > On Thu, Sep 8, 2022 at 11:43 AM Arnaldo Carvalho de Melo
> > > <acme@kernel.org> wrote:
> > > >
> > > > Em Wed, Sep 07, 2022 at 11:37:50PM -0700, Namhyung Kim escreveu:
> > > > > Hello,
> > > > >
> > > > > I found that call stack from the lock tracepoint (using bpf_get_stackid)
> > > > > can be different on each configuration.  For example it's very different
> > > > > when I run it on a VM than on a real machine.
> > > > >
> > > > > The perf lock contention relies on the stack trace to get the lock
> > > > > caller names, this kind of difference can be annoying.  Ideally we could
> > > > > skip stack trace entries for internal BPF or lock functions and get the
> > > > > correct caller, but it's not the case as of today.  Currently it's hard
> > > > > coded to control the behavior of stack traces for the lock contention
> > > > > tracepoints.
> > > > >
> > > > > To handle those differences, add two new options to control the number of
> > > > > stack entries and how many it skips.  The default value worked well on
> > > > > my VM setup, but I had to use --stack-skip=5 on real machines.
> > > > >
> > > > > You can get it from 'perf/lock-stack-v1' branch in
> > > > >
> > > > >   git://git.kernel.org/pub/scm/linux/kernel/git/namhyung/linux-perf.git
> > > >
> > > > This clashed with a patch you Acked earlier, so lets see if someone has
> > > > extra review comments and a v2 become needed for other reason, when you
> > > > can refresh it, ok?
> > >
> > > Sounds good!
> >
> > Have you resubmitted this? /me goes on the backlog...
> 
> Yep :)
> 
> https://lore.kernel.org/r/20220912055314.744552-1-namhyung@kernel.org

It applies now, testing :-)

- Arnaldo