Re: [PATCH v4 00/44] Fix perf on Intel hybrid CPUs
From: Liang, Kan
Date: Sun May 14 2023 - 08:03:55 EST
On 2023-05-12 2:33 p.m., Arnaldo Carvalho de Melo wrote:
> Em Wed, May 03, 2023 at 04:56:36PM -0400, Liang, Kan escreveu:
>>
>>
>> On 2023-05-02 6:38 p.m., Ian Rogers wrote:
>>> TL;DR: hybrid doesn't crash, json metrics work on hybrid on both PMUs
>>> or individually, event parsing doesn't always scan all PMUs, more and
>>> new tests that also run without hybrid, less code.
>>>
>>> The first 4 patches are aimed at Linux 6.4 to address issues raised,
>>> in particular by Kan, on the existing perf stat behavior with json
>>> metrics. They avoid duplicated events by removing groups. They don't
>>> hide events and metrics to make event multiplexing obvious. They avoid
>>> terminating perf when paranoia is higher due to certain events that
>>> always fail. They avoid rearranging events by PMUs when the events
>>> aren't in a group.
>>>
>>> The next 5 patches avoid grouping events for metrics where they could
>>> never succeed and were previously posted as:
>>> "perf vendor events intel: Add xxx metric constraints"
>>> https://lore.kernel.org/all/20230419005423.343862-1-irogers@xxxxxxxxxx/
>>> In general the generated json is coming from:
>>> https://github.com/intel/perfmon/pull/73
>>>
>>> Next are some general and test improvements.
>>>
>>> Next event parsing is rewritten to not scan all PMUs for the benefit
>>> of raw and legacy cache parsing, instead these are handled by the
>>> lexer and a new term type. This ultimately removes the need for the
>>> event parser for hybrid to be recursive as legacy cache can be just a
>>> term. Tests are re-enabled for events with hyphens, so AMD's
>>> branch-brs event is now parsable.
>>>
>>> The cputype option is made a generic pmu filter flag and is tested
>>> even on non-hybrid systems.
>>>
>>> The final patches address specific json metric issues on hybrid, in
>>> both the json metrics and the metric code.
>>>
>>> The patches add slightly more code than they remove, in areas like
>>> better json metric constraints and tests, but in the core util code,
>>> the removal of hybrid is a net reduction:
>>> 22 files changed, 711 insertions(+), 1016 deletions(-)
>>>
>>> Sample output is contained in the v1 patch set:
>>> https://lore.kernel.org/lkml/bff481ba-e60a-763f-0aa0-3ee53302c480@xxxxxxxxxxxxxxx/
>>>
>>> Tested on Tigerlake, Skylake and Alderlake CPUs.
>>>
>>> The v4 patch set:
>>> - rebase, 1 of the Linux 6.4 recommended patches are merged leaving:
>>> 1) perf metric: Change divide by zero and !support events behavior
>>> 2) perf stat: Introduce skippable evsels
>>> 3) perf metric: Json flag to not group events if gathering a metric group
>>> 4) perf parse-events: Don't reorder ungrouped events by pmu
>>> whose diffstat is:
>>> 30 files changed, 326 insertions(+), 33 deletions(-)
>>> but without the vendor event updates (the tend to be large as they
>>> repeat something per architecture per metric) is just:
>>> 10 files changed, 90 insertions(+), 32 deletions(-)
>>
>> I have tested the 4 patches on top of the perf-tools-next branch on both
>> Cascade Lake and Raptor Lake. The result looks good to me.
>>
>> They address the permission error found in the default mode of perf stat
>> on the Cascade Lake. Thanks Ian for the fix.
>>
>> Arnaldo, could you please consider to back port them for the 6.4?
>
> Yes, its in perf-tools now, will go to Linus next week.
Thanks Arnaldo!
>
> What about the other patches? I saw some you provided your review, what
> about the others, are you ok with them?
>
Yes, I'm OK with the patch set. It fixes many issues. Thanks Ian.
(My tests mainly focus on the area in which the patch set may touch. I
did the tests on various platforms, ADL (hybrid), Cascade Lake, SPR.)
Tested-by: Kan Liang <kan.liang@xxxxxxxxxxxxxxx>
But there are still some issues. I don't think they are introduced by
this patch set. We may fix them later separately.
- Segmentation fault with perf stat --topdown on ADL (hybrid) and
Cascade Lake.
It looks like a legacy issue, may not be introduced by this patch set.
Here is the backtrace. It looks like there is a NULL metric_group.
(gdb) backtrace
#0 0x00007ffff73035d1 in __strstr_sse2_unaligned () from /lib64/libc.so.6
#1 0x00000000004f9019 in metricgroup__topdown_max_level_callback
(pm=<optimized out>, table=<optimized out>,
data=0x7fffffff92f4) at util/metricgroup.c:1722
#2 0x00000000005e8a31 in pmu_metrics_table_for_each_metric
(table=0xcb74d0 <pmu_events_map+368>,
fn=fn@entry=0x4f8ff0 <metricgroup__topdown_max_level_callback>,
data=data@entry=0x7fffffff92f4)
at pmu-events/pmu-events.c:61123
#3 0x00000000004fbc3b in metricgroups__topdown_max_level () at
util/metricgroup.c:1742
#4 0x000000000042c135 in add_default_attributes () at builtin-stat.c:1845
#5 cmd_stat (argc=0, argv=0x7fffffffe3e0) at builtin-stat.c:2446
#6 0x00000000004b922b in run_builtin (p=p@entry=0xd5c530
<commands+336>, argc=argc@entry=2,
argv=argv@entry=0x7fffffffe3e0) at perf.c:323
#7 0x000000000040e373 in handle_internal_command (argv=0x7fffffffe3e0,
argc=2) at perf.c:377
#8 run_argv (argv=<synthetic pointer>, argcp=<synthetic pointer>) at
perf.c:421
#9 main (argc=2, argv=0x7fffffffe3e0) at perf.c:537
(gdb)
Also, the return type is unsigned int, but a bool is given.
unsigned int metricgroups__topdown_max_level(void)
{
unsigned int max_level = 0;
const struct pmu_metrics_table *table = pmu_metrics_table__find();
if (!table)
return false;
- The perf metric and metricgroups fail on different platforms.
Ian and I have discussed it. We agree to address it later separately.
102: perf all metricgroups test
ADL (hybrid)
103: perf all metrics test
ADL (hybrid), Cascade Lake, SPR
- perf list: The [Kernel PMU event] is missed for all the hardware cache
events.
It impacts both hybrid and non-hybrid platforms.
It's a user-visible change introduced by the patch set.
I don't know if anyone cares whether it's a kernel event or a regular
event. It doesn't bother me. So I'm OK with it.
cpu:
L1-dcache-loads OR cpu/L1-dcache-loads/
L1-dcache-load-misses OR cpu/L1-dcache-load-misses/
L1-dcache-stores OR cpu/L1-dcache-stores/
L1-icache-load-misses OR cpu/L1-icache-load-misses/
LLC-loads OR cpu/LLC-loads/
LLC-load-misses OR cpu/LLC-load-misses/
LLC-stores OR cpu/LLC-stores/
LLC-store-misses OR cpu/LLC-store-misses/
dTLB-loads OR cpu/dTLB-loads/
dTLB-load-misses OR cpu/dTLB-load-misses/
dTLB-stores OR cpu/dTLB-stores/
dTLB-store-misses OR cpu/dTLB-store-misses/
iTLB-load-misses OR cpu/iTLB-load-misses/
branch-loads OR cpu/branch-loads/
branch-load-misses OR cpu/branch-load-misses/
node-loads OR cpu/node-loads/
node-load-misses OR cpu/node-load-misses/
branch-instructions OR cpu/branch-instructions/ [Kernel PMU event]
branch-misses OR cpu/branch-misses/ [Kernel PMU event]
bus-cycles OR cpu/bus-cycles/ [Kernel PMU event]
cache-misses OR cpu/cache-misses/ [Kernel PMU event]
cache-references OR cpu/cache-references/ [Kernel PMU event]
cpu-cycles OR cpu/cpu-cycles/ [Kernel PMU event]
instructions OR cpu/instructions/ [Kernel PMU event]
- The --cputype only works for the metric in the default mode.
I can still see the cpu_atom events with --cputype core
It may be something we can improve later.
# ./perf stat --cputype core sleep 2
Performance counter stats for 'sleep 2':
0.52 msec task-clock # 0.000 CPUs
utilized
1 context-switches # 1.939 K/sec
0 cpu-migrations # 0.000 /sec
69 page-faults # 133.770 K/sec
2,569,423 cpu_core/cycles/ # 4.981 G/sec
<not counted> cpu_atom/cycles/
(0.00%)
3,287,691 cpu_core/instructions/ # 6.374 G/sec
<not counted> cpu_atom/instructions/
(0.00%)
555,848 cpu_core/branches/ # 1.078 G/sec
<not counted> cpu_atom/branches/
(0.00%)
8,398 cpu_core/branch-misses/ # 16.281 M/sec
<not counted> cpu_atom/branch-misses/
(0.00%)
15,416,538 cpu_core/TOPDOWN.SLOTS/ # 36.1 %
tma_backend_bound
# 23.9 % tma_retiring
# 5.6 %
tma_bad_speculation
# 34.4 %
tma_frontend_bound
3,687,877 cpu_core/topdown-retiring/
846,398 cpu_core/topdown-bad-spec/
5,320,217 cpu_core/topdown-fe-bound/
5,562,045 cpu_core/topdown-be-bound/
14,149 cpu_core/INT_MISC.UOP_DROPPING/ # 27.431 M/sec
Thanks,
Kan