Re: [PATCH RFC 00/21] perf intel-pt: Better support for perf record --cpu

[Ian Rogers]

On Fri, Apr 22, 2022 at 9:24 AM Adrian Hunter <adrian.hunter@xxxxxxxxx> wrote:
>
> Hi
>
> Here are patches to support capturing Intel PT sideband events such as
> mmap, task, context switch, text poke etc, on every CPU even when tracing
> selected user_requested_cpus.  That is, when using the perf record -C or
>  --cpu option.
>
> This is needed for:
> 1. text poke: a text poke on any CPU affects all CPUs
> 2. tracing user space: a user space process can migrate between CPUs so
> mmap events that happen on a different CPU can be needed to decode a
> user_requested_cpus CPU.
>
> For example:
>
>         Trace on CPU 1:
>
>         perf record --kcore -C 1 -e intel_pt// &
>
>         Start a task on CPU 0:
>
>         taskset 0x1 testprog &
>
>         Migrate it to CPU 1:
>
>         taskset -p 0x2 <testprog pid>
>
>         Stop tracing:
>
>         kill %1
>
>         Prior to these changes there will be errors decoding testprog
>         in userspace because the comm and mmap events for testprog will not
>         have been captured.
>
> There is quite a bit of preparation:
>
> The first 5 patches stop auxtrace mixing up mmap idx between evlist and
> evsel.  That is going to matter when
> evlist->all_cpus != evlist->user_requested_cpus != evsel->cpus:
>
>       libperf evsel: Factor out perf_evsel__ioctl()
>       libperf evsel: Add perf_evsel__enable_thread()
>       perf evlist: Use libperf functions in evlist__enable_event_idx()
>       perf auxtrace: Move evlist__enable_event_idx() to auxtrace.c
>       perf auxtrace: Do not mix up mmap idx
>
> The next 6 patches stop attempts to auxtrace mmap when it is not an
> auxtrace event e.g. when mmapping the CPUs on which only sideband is
> captured:
>
>       libperf evlist: Remove ->idx() per_cpu parameter
>       libperf evlist: Move ->idx() into mmap_per_evsel()
>       libperf evlist: Add evsel as a parameter to ->idx()
>       perf auxtrace: Record whether an auxtrace mmap is needed
>       perf auxctrace: Add mmap_needed to auxtrace_mmap_params
>       perf auxtrace: Remove auxtrace_mmap_params__set_idx() per_cpu parameter
>
> The next 5 patches switch to setting up dummy event maps before adding the
> evsel so that the evsel is subject to map propagation, primarily to cause
> addition of the evsel's CPUs to all_cpus.
>
>       perf evlist: Factor out evlist__dummy_event()
>       perf evlist: Add evlist__add_system_wide_dummy()
>       perf record: Use evlist__add_system_wide_dummy() in record__config_text_poke()
>       perf intel-pt: Use evlist__add_system_wide_dummy() for switch tracking
>       perf intel-pt: Track sideband system-wide when needed
>
> The remaining 5 patches make more significant changes.
>
> First change from using user_requested_cpus to using all_cpus where necessary:
>
>       perf tools: Allow all_cpus to be a superset of user_requested_cpus
>
> Secondly, mmap all per-thread and all per-cpu events:
>
>       libperf evlist: Allow mixing per-thread and per-cpu mmaps
>
> Stop using system_wide flag for uncore because it will not work anymore:
>
>       perf stat: Add per_cpu_only flag for uncore
>
> Finally change map propagation so that system-wide events retain their cpus and
> (dummy) threads:
>
>       perf tools: Allow system-wide events to keep their own CPUs
>       perf tools: Allow system-wide events to keep their own threads
>
>
> Adrian Hunter (21):
>       libperf evsel: Factor out perf_evsel__ioctl()
>       libperf evsel: Add perf_evsel__enable_thread()
>       perf evlist: Use libperf functions in evlist__enable_event_idx()
>       perf auxtrace: Move evlist__enable_event_idx() to auxtrace.c
>       perf auxtrace: Do not mix up mmap idx
>       libperf evlist: Remove ->idx() per_cpu parameter
>       libperf evlist: Move ->idx() into mmap_per_evsel()
>       libperf evlist: Add evsel as a parameter to ->idx()
>       perf auxtrace: Record whether an auxtrace mmap is needed
>       perf auxctrace: Add mmap_needed to auxtrace_mmap_params
>       perf auxtrace: Remove auxtrace_mmap_params__set_idx() per_cpu parameter
>       perf evlist: Factor out evlist__dummy_event()
>       perf evlist: Add evlist__add_system_wide_dummy()
>       perf record: Use evlist__add_system_wide_dummy() in record__config_text_poke()
>       perf intel-pt: Use evlist__add_system_wide_dummy() for switch tracking
>       perf intel-pt: Track sideband system-wide when needed
>       perf tools: Allow all_cpus to be a superset of user_requested_cpus
>       libperf evlist: Allow mixing per-thread and per-cpu mmaps
>       perf stat: Add per_cpu_only flag for uncore
>       perf tools: Allow system-wide events to keep their own CPUs
>       perf tools: Allow system-wide events to keep their own threads
>
>  tools/lib/perf/evlist.c                  |  67 +++++++------------
>  tools/lib/perf/evsel.c                   |  29 +++++++--
>  tools/lib/perf/include/internal/evlist.h |   3 +-
>  tools/lib/perf/include/internal/evsel.h  |   1 +
>  tools/lib/perf/include/perf/evsel.h      |   1 +
>  tools/perf/arch/arm/util/cs-etm.c        |   1 +
>  tools/perf/arch/arm64/util/arm-spe.c     |   1 +
>  tools/perf/arch/s390/util/auxtrace.c     |   1 +
>  tools/perf/arch/x86/util/intel-bts.c     |   1 +
>  tools/perf/arch/x86/util/intel-pt.c      |  32 ++++------
>  tools/perf/builtin-record.c              |  39 +++++-------
>  tools/perf/builtin-stat.c                |   5 +-
>  tools/perf/util/auxtrace.c               |  31 +++++++--
>  tools/perf/util/auxtrace.h               |   8 ++-
>  tools/perf/util/evlist.c                 | 106 +++++++++++++++----------------
>  tools/perf/util/evlist.h                 |   7 +-
>  tools/perf/util/evsel.c                  |   1 +
>  tools/perf/util/evsel.h                  |   1 +
>  tools/perf/util/mmap.c                   |   4 +-
>  tools/perf/util/parse-events.c           |   2 +-
>  20 files changed, 176 insertions(+), 165 deletions(-)
>
>
> Regards
> Adrian

Thanks Adrian, I'm very much in favor of this patch set. Can we add
some tests for intel-pt? They could be part of the 'perf record' shell
test:
https://git.kernel.org/pub/scm/linux/kernel/git/acme/linux.git/tree/tools/perf/tests/shell/record.sh?h=perf/core
but probably better as there own thing. A command line broken prior to
this change would be great!

Aside from testing, we should clean up what dummy CPUs in the cpu maps
means. As per:
https://lore.kernel.org/linux-perf-users/CAP-5=fWfs2td9nZLGdEBD+C5s=upa_7SORab8tQ7qH=jX--F7w@xxxxxxxxxxxxxx/

I also think landing:
https://lore.kernel.org/linux-perf-users/20220503041757.2365696-3-irogers@xxxxxxxxxx/
will help as it avoids the all_cpus map containing references to CPUs
from PMU sysfs that had been overridden.

Thanks,
Ian