Re: [RFC PATCH v1] perf parse-events: Make legacy events lower priority than sysfs/json

[Hector Martin <marcan@marcan.st>]

On 2023/11/23 13:29, Ian Rogers wrote:
> The perf tool has previously made legacy events the priority so with
> or without a PMU the legacy event would be opened:
> 
> ```
> $ perf stat -e cpu-cycles,cpu/cpu-cycles/ true
> Using CPUID GenuineIntel-6-8D-1
> intel_pt default config: tsc,mtc,mtc_period=3,psb_period=3,pt,branch
> Attempting to add event pmu 'cpu' with 'cpu-cycles,' that may result in non-fatal errors
> After aliases, add event pmu 'cpu' with 'cpu-cycles,' that may result in non-fatal errors
> Control descriptor is not initialized
> ------------------------------------------------------------
> perf_event_attr:
>   type                             0 (PERF_TYPE_HARDWARE)
>   size                             136
>   config                           0 (PERF_COUNT_HW_CPU_CYCLES)
>   sample_type                      IDENTIFIER
>   read_format                      TOTAL_TIME_ENABLED|TOTAL_TIME_RUNNING
>   disabled                         1
>   inherit                          1
>   enable_on_exec                   1
>   exclude_guest                    1
> ------------------------------------------------------------
> sys_perf_event_open: pid 833967  cpu -1  group_fd -1  flags 0x8 = 3
> ------------------------------------------------------------
> perf_event_attr:
>   type                             0 (PERF_TYPE_HARDWARE)
>   size                             136
>   config                           0 (PERF_COUNT_HW_CPU_CYCLES)
>   sample_type                      IDENTIFIER
>   read_format                      TOTAL_TIME_ENABLED|TOTAL_TIME_RUNNING
>   disabled                         1
>   inherit                          1
>   enable_on_exec                   1
>   exclude_guest                    1
> ------------------------------------------------------------
> ...
> ```
> 
> Fixes to make hybrid/BIG.little PMUs behave correctly, ie as core PMUs
> capable of opening legacy events on each, removing hard coded
> "cpu_core" and "cpu_atom" Intel PMU names, etc. caused a behavioral
> difference on Apple/ARM due to latent issues in the PMU driver
> reported in:
> https://lore.kernel.org/lkml/08f1f185-e259-4014-9ca4-6411d5c1bc65@marcan.st/
> 
> As part of that report Mark Rutland <mark.rutland@arm.com> requested
> that legacy events not be higher in priority when a PMU is specified
> reversing what has until this change been perf's default
> behavior. With this change the above becomes:
> 
> ```
> $ perf stat -e cpu-cycles,cpu/cpu-cycles/ true
> Using CPUID GenuineIntel-6-8D-1
> Attempt to add: cpu/cpu-cycles=0/
> ..after resolving event: cpu/event=0x3c/
> Control descriptor is not initialized
> ------------------------------------------------------------
> perf_event_attr:
>   type                             0 (PERF_TYPE_HARDWARE)
>   size                             136
>   config                           0 (PERF_COUNT_HW_CPU_CYCLES)
>   sample_type                      IDENTIFIER
>   read_format                      TOTAL_TIME_ENABLED|TOTAL_TIME_RUNNING
>   disabled                         1
>   inherit                          1
>   enable_on_exec                   1
>   exclude_guest                    1
> ------------------------------------------------------------
> sys_perf_event_open: pid 827628  cpu -1  group_fd -1  flags 0x8 = 3
> ------------------------------------------------------------
> perf_event_attr:
>   type                             4 (PERF_TYPE_RAW)
>   size                             136
>   config                           0x3c
>   sample_type                      IDENTIFIER
>   read_format                      TOTAL_TIME_ENABLED|TOTAL_TIME_RUNNING
>   disabled                         1
>   inherit                          1
>   enable_on_exec                   1
>   exclude_guest                    1
> ------------------------------------------------------------
> ...
> ```
> 
> So the second event has become a raw event as
> /sys/devices/cpu/events/cpu-cycles exists.
> 
> A fix was necessary to config_term_pmu in parse-events.c as
> check_alias expansion needs to happen after config_term_pmu, and
> config_term_pmu may need calling a second time because of this.
> 
> config_term_pmu is updated to not use the legacy event when the PMU
> has such a named event (either from json or sysfs).
> 
> The bulk of this change is updating all of the parse-events test
> expectations so that if a sysfs/json event exists for a PMU the test
> doesn't fail - a further sign, if it were needed, that the legacy
> event priority was a known and tested behavior of the perf tool.
> 
> Signed-off-by: Ian Rogers <irogers@google.com>
> ---
> This is a large behavioral change:
> 1) the scope of the change means it should bake on linux-next and I
> don't believe should be a 6.7-rc fix.
> 2) a fixes tag and stable backport I don't think are appropriate. The
> real reported issue is with the PMU driver. A backport would bring the
> risk that later fixes, due to the large behavior change, wouldn't be
> backported and past releases get regressed in scenarios like
> hybrid. Backports for the perf tool are also less necessary than say a
> buggy PMU driver, as distributions should be updating to the latest
> perf tool regardless of what Linux kernel is being run (the perf tool
> is backward compatible).
> ---
>  tools/perf/tests/parse-events.c | 256 +++++++++++++++++++++++---------
>  tools/perf/util/parse-events.c  |  52 +++++--
>  tools/perf/util/pmu.c           |   8 +-
>  tools/perf/util/pmu.h           |   3 +-
>  4 files changed, 231 insertions(+), 88 deletions(-)
> 

Tested-by: Hector Martin <marcan@marcan.st>

$ sudo taskset -c 2 ./perf stat -e apple_icestorm_pmu/cycles/ -e
apple_firestorm_pmu/cycles/ -e cycles echo


 Performance counter stats for 'echo':

     <not counted>      apple_icestorm_pmu/cycles/
                       (0.00%)
            34,622      apple_firestorm_pmu/cycles/

            30,751      cycles


       0.000429625 seconds time elapsed

       0.000000000 seconds user
       0.000443000 seconds sys


$ sudo taskset -c 0 ./perf stat -e apple_icestorm_pmu/cycles/ -e
apple_firestorm_pmu/cycles/ -e cycles echo


 Performance counter stats for 'echo':

            13,413      apple_icestorm_pmu/cycles/

     <not counted>      apple_firestorm_pmu/cycles/
                       (0.00%)
     <not counted>      cycles
                       (0.00%)

       0.000898458 seconds time elapsed

       0.000908000 seconds user
       0.000000000 seconds sys

(It would be nice to have "cycles" match/aggregate both PMUs, but that's
a story for another day. The behavior above is what was there in 6.4 and
earlier.)

- Hector