| Message ID | 20230522064330.189127-1-irogers@google.com (mailing list archive) |
|---|---|
| Headers | show |
| Series | PMU refactoring and improvements | expand |
On 2023-05-22 2:43 a.m., Ian Rogers wrote: > Separate the code in pmu.[ch] into the set/list of PMUs and the code > for a particular PMU. Move the set/list of PMUs code into > pmus.[ch]. Clean up hybrid code and remove hybrid PMU list, it is > sufficient to scan PMUs looking for core ones. Add core PMU list and > perf_pmus__scan_core that just reads core PMUs. Switch code that skips > non-core PMUs during a perf_pmus__scan, to use the > perf_pmus__scan_core variant. Don't scan sysfs for PMUs if all such > PMUs have been previously scanned/loaded. Scanning just core PMUs, for > the cases it is applicable, can improve the sysfs reading time by more > than 4 fold on my laptop, as servers generally have many more uncore > PMUs the improvement there should be larger: > > ``` > $ perf bench internals pmu-scan -i 1000 > Computing performance of sysfs PMU event scan for 1000 times > Average core PMU scanning took: 989.231 usec (+- 1.535 usec) > Average PMU scanning took: 4309.425 usec (+- 74.322 usec) > ``` > > The patch "perf pmu: Separate pmu and pmus" moves and renames a lot of > functions, and is consequently large. The changes are trivial, but > kept together to keep the overall number of patches more reasonable. > > v2. Address Kan's review comments wrt "cycles" -> "cycles:P" and > "uncore_pmus" -> "other_pmus". > > Ian Rogers (23): > perf tools: Warn if no user requested CPUs match PMU's CPUs > perf evlist: Remove evlist__warn_hybrid_group > perf evlist: Remove __evlist__add_default > perf evlist: Reduce scope of evlist__has_hybrid > perf pmu: Remove perf_pmu__hybrid_mounted > perf pmu: Detect ARM and hybrid PMUs with sysfs > perf pmu: Add is_core to pmu > perf pmu: Rewrite perf_pmu__has_hybrid to avoid list > perf x86: Iterate hybrid PMUs as core PMUs > perf topology: Avoid hybrid list for hybrid topology > perf evsel: Compute is_hybrid from PMU being core > perf header: Avoid hybrid PMU list in write_pmu_caps > perf metrics: Remove perf_pmu__is_hybrid use > perf stat: Avoid hybrid PMU list > perf mem: Avoid hybrid PMU list > perf pmu: Remove perf_pmu__hybrid_pmus list > perf pmus: Prefer perf_pmu__scan over perf_pmus__for_each_pmu > perf x86 mem: minor refactor to is_mem_loads_aux_event > perf pmu: Separate pmu and pmus > perf pmus: Split pmus list into core and other > perf pmus: Allow just core PMU scanning > perf pmus: Avoid repeated sysfs scanning > perf pmus: Ensure all PMUs are read for find_by_type The patch set also triggers Segmentation fault with default mode on my hybrid machine. # ./perf stat sleep 1 Performance counter stats for 'sleep 1': 0.53 msec task-clock # 0.001 CPUs utilized 1 context-switches # 1.875 K/sec 0 cpu-migrations # 0.000 /sec 68 page-faults # 127.476 K/sec Segmentation fault (core dumped) Program received signal SIGSEGV, Segmentation fault. evsel__is_hybrid (evsel=0x55555609a1a0) at util/evsel.c:3143 3143 return pmu->is_core; (gdb) backtrace #0 evsel__is_hybrid (evsel=0x55555609a1a0) at util/evsel.c:3143 #1 evsel__is_hybrid (evsel=evsel@entry=0x55555609a1a0) at util/evsel.c:3135 #2 0x0000555555759468 in hybrid_uniquify (config=0x555555f931e0 <stat_config>, evsel=0x55555609a1a0) at util/stat-display.c:813 #3 uniquify_counter (counter=0x55555609a1a0, config=0x555555f931e0 <stat_config>) at util/stat-display.c:818 #4 print_counter_aggrdata (config=config@entry=0x555555f931e0 <stat_config>, counter=counter@entry=0x55555609a1a0, aggr_idx=aggr_idx@entry=0, os=os@entry=0x7fffffff8fe0) at util/stat-display.c:888 #5 0x000055555575b119 in print_counter (os=<optimized out>, counter=<optimized out>, config=<optimized out>) at util/stat-display.c:1019 #6 print_counter (os=0x7fffffff8fe0, counter=0x55555609a1a0, config=0x555555f931e0 <stat_config>) at util/stat-display.c:1009 #7 evlist__print_counters (evlist=0x555556029da0, config=config@entry=0x555555f931e0 <stat_config>, _target=_target@entry=0x555555f42de0 <target>, ts=ts@entry=0x0, argc=argc@entry=2, argv=argv@entry=0x7fffffffe1d0) at util/stat-display.c:1480 #8 0x000055555562009c in print_counters (argv=0x7fffffffe1d0, argc=2, ts=0x0) at builtin-stat.c:979 #9 print_counters (argv=0x7fffffffe1d0, argc=2, ts=0x0) at builtin-stat.c:971 #10 cmd_stat (argc=2, argv=0x7fffffffe1d0) at builtin-stat.c:2832 #11 0x00005555556b6670 in run_builtin (p=p@entry=0x555555f9c590 <commands+336>, argc=argc@entry=3, argv=argv@entry=0x7fffffffe1d0) at perf.c:323 #12 0x00005555555ff2d9 in handle_internal_command (argv=0x7fffffffe1d0, argc=3) at perf.c:377 #13 run_argv (argv=<synthetic pointer>, argcp=<synthetic pointer>) at perf.c:421 #14 main (argc=3, argv=0x7fffffffe1d0) at perf.c:537 Thanks, Kan > > tools/perf/arch/arm/util/auxtrace.c | 7 +- > tools/perf/arch/arm/util/cs-etm.c | 4 +- > tools/perf/arch/arm64/util/pmu.c | 6 +- > tools/perf/arch/x86/tests/hybrid.c | 7 +- > tools/perf/arch/x86/util/auxtrace.c | 5 +- > tools/perf/arch/x86/util/evlist.c | 25 +- > tools/perf/arch/x86/util/evsel.c | 27 +- > tools/perf/arch/x86/util/intel-bts.c | 4 +- > tools/perf/arch/x86/util/intel-pt.c | 4 +- > tools/perf/arch/x86/util/mem-events.c | 17 +- > tools/perf/arch/x86/util/perf_regs.c | 15 +- > tools/perf/arch/x86/util/topdown.c | 5 +- > tools/perf/bench/pmu-scan.c | 60 ++-- > tools/perf/builtin-c2c.c | 9 +- > tools/perf/builtin-list.c | 4 +- > tools/perf/builtin-mem.c | 9 +- > tools/perf/builtin-record.c | 29 +- > tools/perf/builtin-stat.c | 15 +- > tools/perf/builtin-top.c | 10 +- > tools/perf/tests/attr.c | 4 +- > tools/perf/tests/event_groups.c | 7 +- > tools/perf/tests/parse-events.c | 15 +- > tools/perf/tests/parse-metric.c | 4 +- > tools/perf/tests/pmu-events.c | 6 +- > tools/perf/tests/switch-tracking.c | 4 +- > tools/perf/tests/topology.c | 4 +- > tools/perf/util/Build | 2 - > tools/perf/util/cpumap.h | 2 +- > tools/perf/util/cputopo.c | 16 +- > tools/perf/util/env.c | 5 +- > tools/perf/util/evlist-hybrid.c | 162 --------- > tools/perf/util/evlist-hybrid.h | 15 - > tools/perf/util/evlist.c | 67 +++- > tools/perf/util/evlist.h | 9 +- > tools/perf/util/evsel.c | 57 +-- > tools/perf/util/evsel.h | 3 - > tools/perf/util/header.c | 27 +- > tools/perf/util/mem-events.c | 17 +- > tools/perf/util/metricgroup.c | 9 +- > tools/perf/util/parse-events.c | 24 +- > tools/perf/util/parse-events.y | 3 +- > tools/perf/util/pfm.c | 6 +- > tools/perf/util/pmu-hybrid.c | 52 --- > tools/perf/util/pmu-hybrid.h | 32 -- > tools/perf/util/pmu.c | 482 ++------------------------ > tools/perf/util/pmu.h | 26 +- > tools/perf/util/pmus.c | 477 ++++++++++++++++++++++++- > tools/perf/util/pmus.h | 15 +- > tools/perf/util/print-events.c | 15 +- > tools/perf/util/python-ext-sources | 1 - > tools/perf/util/stat-display.c | 21 +- > 51 files changed, 819 insertions(+), 1032 deletions(-) > delete mode 100644 tools/perf/util/evlist-hybrid.c > delete mode 100644 tools/perf/util/evlist-hybrid.h > delete mode 100644 tools/perf/util/pmu-hybrid.c > delete mode 100644 tools/perf/util/pmu-hybrid.h >
Separate the code in pmu.[ch] into the set/list of PMUs and the code for a particular PMU. Move the set/list of PMUs code into pmus.[ch]. Clean up hybrid code and remove hybrid PMU list, it is sufficient to scan PMUs looking for core ones. Add core PMU list and perf_pmus__scan_core that just reads core PMUs. Switch code that skips non-core PMUs during a perf_pmus__scan, to use the perf_pmus__scan_core variant. Don't scan sysfs for PMUs if all such PMUs have been previously scanned/loaded. Scanning just core PMUs, for the cases it is applicable, can improve the sysfs reading time by more than 4 fold on my laptop, as servers generally have many more uncore PMUs the improvement there should be larger: ``` $ perf bench internals pmu-scan -i 1000 Computing performance of sysfs PMU event scan for 1000 times Average core PMU scanning took: 989.231 usec (+- 1.535 usec) Average PMU scanning took: 4309.425 usec (+- 74.322 usec) ``` The patch "perf pmu: Separate pmu and pmus" moves and renames a lot of functions, and is consequently large. The changes are trivial, but kept together to keep the overall number of patches more reasonable. v2. Address Kan's review comments wrt "cycles" -> "cycles:P" and "uncore_pmus" -> "other_pmus". Ian Rogers (23): perf tools: Warn if no user requested CPUs match PMU's CPUs perf evlist: Remove evlist__warn_hybrid_group perf evlist: Remove __evlist__add_default perf evlist: Reduce scope of evlist__has_hybrid perf pmu: Remove perf_pmu__hybrid_mounted perf pmu: Detect ARM and hybrid PMUs with sysfs perf pmu: Add is_core to pmu perf pmu: Rewrite perf_pmu__has_hybrid to avoid list perf x86: Iterate hybrid PMUs as core PMUs perf topology: Avoid hybrid list for hybrid topology perf evsel: Compute is_hybrid from PMU being core perf header: Avoid hybrid PMU list in write_pmu_caps perf metrics: Remove perf_pmu__is_hybrid use perf stat: Avoid hybrid PMU list perf mem: Avoid hybrid PMU list perf pmu: Remove perf_pmu__hybrid_pmus list perf pmus: Prefer perf_pmu__scan over perf_pmus__for_each_pmu perf x86 mem: minor refactor to is_mem_loads_aux_event perf pmu: Separate pmu and pmus perf pmus: Split pmus list into core and other perf pmus: Allow just core PMU scanning perf pmus: Avoid repeated sysfs scanning perf pmus: Ensure all PMUs are read for find_by_type tools/perf/arch/arm/util/auxtrace.c | 7 +- tools/perf/arch/arm/util/cs-etm.c | 4 +- tools/perf/arch/arm64/util/pmu.c | 6 +- tools/perf/arch/x86/tests/hybrid.c | 7 +- tools/perf/arch/x86/util/auxtrace.c | 5 +- tools/perf/arch/x86/util/evlist.c | 25 +- tools/perf/arch/x86/util/evsel.c | 27 +- tools/perf/arch/x86/util/intel-bts.c | 4 +- tools/perf/arch/x86/util/intel-pt.c | 4 +- tools/perf/arch/x86/util/mem-events.c | 17 +- tools/perf/arch/x86/util/perf_regs.c | 15 +- tools/perf/arch/x86/util/topdown.c | 5 +- tools/perf/bench/pmu-scan.c | 60 ++-- tools/perf/builtin-c2c.c | 9 +- tools/perf/builtin-list.c | 4 +- tools/perf/builtin-mem.c | 9 +- tools/perf/builtin-record.c | 29 +- tools/perf/builtin-stat.c | 15 +- tools/perf/builtin-top.c | 10 +- tools/perf/tests/attr.c | 4 +- tools/perf/tests/event_groups.c | 7 +- tools/perf/tests/parse-events.c | 15 +- tools/perf/tests/parse-metric.c | 4 +- tools/perf/tests/pmu-events.c | 6 +- tools/perf/tests/switch-tracking.c | 4 +- tools/perf/tests/topology.c | 4 +- tools/perf/util/Build | 2 - tools/perf/util/cpumap.h | 2 +- tools/perf/util/cputopo.c | 16 +- tools/perf/util/env.c | 5 +- tools/perf/util/evlist-hybrid.c | 162 --------- tools/perf/util/evlist-hybrid.h | 15 - tools/perf/util/evlist.c | 67 +++- tools/perf/util/evlist.h | 9 +- tools/perf/util/evsel.c | 57 +-- tools/perf/util/evsel.h | 3 - tools/perf/util/header.c | 27 +- tools/perf/util/mem-events.c | 17 +- tools/perf/util/metricgroup.c | 9 +- tools/perf/util/parse-events.c | 24 +- tools/perf/util/parse-events.y | 3 +- tools/perf/util/pfm.c | 6 +- tools/perf/util/pmu-hybrid.c | 52 --- tools/perf/util/pmu-hybrid.h | 32 -- tools/perf/util/pmu.c | 482 ++------------------------ tools/perf/util/pmu.h | 26 +- tools/perf/util/pmus.c | 477 ++++++++++++++++++++++++- tools/perf/util/pmus.h | 15 +- tools/perf/util/print-events.c | 15 +- tools/perf/util/python-ext-sources | 1 - tools/perf/util/stat-display.c | 21 +- 51 files changed, 819 insertions(+), 1032 deletions(-) delete mode 100644 tools/perf/util/evlist-hybrid.c delete mode 100644 tools/perf/util/evlist-hybrid.h delete mode 100644 tools/perf/util/pmu-hybrid.c delete mode 100644 tools/perf/util/pmu-hybrid.h