kernel/linux.git/tools/perf/util/evlist.h, branch v6.4

perf evlist: Remove nr_groups

2023-03-13T20:42:27+00:00

Maintaining the number of groups during event parsing is problematic and since changing to sort/regroup events can only be computed by a linear pass over the evlist. As the value is generally only used in tests, rather than hold it in a variable compute it by passing over the evlist when necessary. This change highlights that libpfm's counting of groups with a single entry disagreed with regular event parsing. The libpfm tests are updated accordingly. Signed-off-by: Ian Rogers Cc: Adrian Hunter Cc: Alexander Shishkin Cc: Florian Fischer Cc: Ingo Molnar Cc: James Clark Cc: Jiri Olsa Cc: John Garry Cc: Kajol Jain Cc: Kan Liang Cc: Kim Phillips Cc: Leo Yan Cc: Mark Rutland Cc: Namhyung Kim Cc: Peter Zijlstra Cc: Ravi Bangoria Cc: Sean Christopherson Cc: Steinar H. Gunderson Cc: Stephane Eranian Cc: Suzuki Poulouse Cc: Xing Zhengjun Link: https://lore.kernel.org/r/20230312021543.3060328-2-irogers@google.com Signed-off-by: Arnaldo Carvalho de Melo

perf parse-events: Sort and group parsed events

2023-03-13T20:42:26+00:00

This change is intended to be a no-op for most current cases, the default sort order is the order the events were parsed. Where it varies is in how groups are handled. Previously an uncore and core event that are grouped would most often cause the group to be removed: ``` $ perf stat -e '{instructions,uncore_imc_free_running_0/data_total/}' -a sleep 1 WARNING: grouped events cpus do not match, disabling group: anon group { instructions, uncore_imc_free_running_0/data_total/ } ... ``` However, when wildcards are used the events should be re-sorted and re-grouped in parse_events__set_leader, but this currently fails for simple examples: ``` $ perf stat -e '{uncore_imc_free_running/data_read/,uncore_imc_free_running/data_write/}' -a sleep 1 Performance counter stats for 'system wide': MiB uncore_imc_free_running/data_read/ MiB uncore_imc_free_running/data_write/ 1.000996992 seconds time elapsed ``` A futher failure mode, fixed in this patch, is to force topdown events into a group. This change moves sorting the evsels in the evlist after parsing. It requires parsing to set up groups. First the evsels are sorted respecting the existing groupings and parse order, but also reordering to ensure evsels of the same PMU and group appear together. So that software and aux events respect groups, their pmu_name is taken from the group leader. The sorting is done with list_sort removing a memory allocation. After sorting a pass is done to correct the group leaders and for topdown events ensuring they have a group leader. This fixes the problems seen before: ``` $ perf stat -e '{uncore_imc_free_running/data_read/,uncore_imc_free_running/data_write/}' -a sleep 1 Performance counter stats for 'system wide': 727.42 MiB uncore_imc_free_running/data_read/ 81.84 MiB uncore_imc_free_running/data_write/ 1.000948615 seconds time elapsed ``` As well as making groups not fail for cases like: ``` $ perf stat -e '{imc_free_running_0/data_total/,imc_free_running_1/data_total/}' -a sleep 1 Performance counter stats for 'system wide': 256.47 MiB imc_free_running_0/data_total/ 256.48 MiB imc_free_running_1/data_total/ 1.001165442 seconds time elapsed ``` Signed-off-by: Ian Rogers Cc: Adrian Hunter Cc: Alexander Shishkin Cc: Florian Fischer Cc: Ingo Molnar Cc: James Clark Cc: Jiri Olsa Cc: John Garry Cc: Kajol Jain Cc: Kan Liang Cc: Kim Phillips Cc: Leo Yan Cc: Mark Rutland Cc: Namhyung Kim Cc: Peter Zijlstra Cc: Ravi Bangoria Cc: Sean Christopherson Cc: Steinar H. Gunderson Cc: Stephane Eranian Cc: Suzuki Poulouse Cc: Xing Zhengjun Link: https://lore.kernel.org/r/20230312021543.3060328-2-irogers@google.com Signed-off-by: Arnaldo Carvalho de Melo

perf evlist: Remove group option.

2022-12-14T18:28:18+00:00

The group option predates grouping events using curly braces added in commit 89efb029502d7f2d ("perf tools: Add support to parse event group syntax"). The --group option was retained for legacy support (in August 2012) but keeping it adds complexity. Signed-off-by: Ian Rogers Cc: Adrian Hunter Cc: Alexander Shishkin Cc: Alexey Bayduraev Cc: Andi Kleen Cc: Anshuman Khandual Cc: Athira Jajeev Cc: Eelco Chaudron Cc: German Gomez Cc: Ingo Molnar Cc: James Clark Cc: Jiri Olsa Cc: Kan Liang Cc: Kim Phillips Cc: Leo Yan Cc: Mark Rutland Cc: Namhyung Kim Cc: Peter Zijlstra Cc: Riccardo Mancini Cc: Sandipan Das Cc: Sean Christopherson Cc: Shaomin Deng Cc: Stephane Eranian Cc: Timothy Hayes Cc: Xing Zhengjun Cc: bpf@vger.kernel.org Link: https://lore.kernel.org/r/20221213232651.1269909-6-irogers@google.com Signed-off-by: Arnaldo Carvalho de Melo

perf build: Use libtraceevent from the system

2022-12-14T14:16:12+00:00

Remove the LIBTRACEEVENT_DYNAMIC and LIBTRACEFS_DYNAMIC make command line variables. If libtraceevent isn't installed or NO_LIBTRACEEVENT=1 is passed to the build, don't compile in libtraceevent and libtracefs support. This also disables CONFIG_TRACE that controls "perf trace". CONFIG_LIBTRACEEVENT is used to control enablement in Build/Makefiles, HAVE_LIBTRACEEVENT is used in C code. Without HAVE_LIBTRACEEVENT tracepoints are disabled and as such the commands kmem, kwork, lock, sched and timechart are removed. The majority of commands continue to work including "perf test". Committer notes: Fixed up a tools/perf/util/Build reject and added: #include to tools/perf/util/scripting-engines/trace-event-perl.c. Committer testing: $ rpm -qi libtraceevent-devel Name : libtraceevent-devel Version : 1.5.3 Release : 2.fc36 Architecture: x86_64 Install Date: Mon 25 Jul 2022 03:20:19 PM -03 Group : Unspecified Size : 27728 License : LGPLv2+ and GPLv2+ Signature : RSA/SHA256, Fri 15 Apr 2022 02:11:58 PM -03, Key ID 999f7cbf38ab71f4 Source RPM : libtraceevent-1.5.3-2.fc36.src.rpm Build Date : Fri 15 Apr 2022 10:57:01 AM -03 Build Host : buildvm-x86-05.iad2.fedoraproject.org Packager : Fedora Project Vendor : Fedora Project URL : https://git.kernel.org/pub/scm/libs/libtrace/libtraceevent.git/ Bug URL : https://bugz.fedoraproject.org/libtraceevent Summary : Development headers of libtraceevent Description : Development headers of libtraceevent-libs $ Default build: $ ldd ~/bin/perf | grep tracee libtraceevent.so.1 => /lib64/libtraceevent.so.1 (0x00007f1dcaf8f000) $ # perf trace -e sched:* --max-events 10 0.000 migration/0/17 sched:sched_migrate_task(comm: "", pid: 1603763 (perf), prio: 120, dest_cpu: 1) 0.005 migration/0/17 sched:sched_wake_idle_without_ipi(cpu: 1) 0.011 migration/0/17 sched:sched_switch(prev_comm: "", prev_pid: 17 (migration/0), prev_state: 1, next_comm: "", next_prio: 120) 1.173 :0/0 sched:sched_wakeup(comm: "", pid: 3138 (gnome-terminal-), prio: 120) 1.180 :0/0 sched:sched_switch(prev_comm: "", prev_prio: 120, next_comm: "", next_pid: 3138 (gnome-terminal-), next_prio: 120) 0.156 migration/1/21 sched:sched_migrate_task(comm: "", pid: 1603763 (perf), prio: 120, orig_cpu: 1, dest_cpu: 2) 0.160 migration/1/21 sched:sched_wake_idle_without_ipi(cpu: 2) 0.166 migration/1/21 sched:sched_switch(prev_comm: "", prev_pid: 21 (migration/1), prev_state: 1, next_comm: "", next_prio: 120) 1.183 :0/0 sched:sched_wakeup(comm: "", pid: 1602985 (kworker/u16:0-f), prio: 120, target_cpu: 1) 1.186 :0/0 sched:sched_switch(prev_comm: "", prev_prio: 120, next_comm: "", next_pid: 1602985 (kworker/u16:0-f), next_prio: 120) # Had to tweak tools/perf/util/setup.py to make sure the python binding shared object links with libtraceevent if -DHAVE_LIBTRACEEVENT is present in CFLAGS. Building with NO_LIBTRACEEVENT=1 uncovered some more build failures: - Make building of data-convert-bt.c to CONFIG_LIBTRACEEVENT=y - perf-$(CONFIG_LIBTRACEEVENT) += scripts/ - bpf_kwork.o needs also to be dependent on CONFIG_LIBTRACEEVENT=y - The python binding needed some fixups and util/trace-event.c can't be built and linked with the python binding shared object, so remove it in tools/perf/util/setup.py and exclude it from the list of dependencies in the python/perf.so Makefile.perf target. Building without libtraceevent-devel installed uncovered more build failures: - The python binding tools/perf/util/python.c was assuming that traceevent/parse-events.h was always available, which was the case when we defaulted to using the in-kernel tools/lib/traceevent/ files, now we need to enclose it under ifdef HAVE_LIBTRACEEVENT, just like the other parts of it that deal with tracepoints. - We have to ifdef the rules in the Build files with CONFIG_LIBTRACEEVENT=y to build builtin-trace.c and tools/perf/trace/beauty/ as we only ifdef setting CONFIG_TRACE=y when setting NO_LIBTRACEEVENT=1 in the make command line, not when we don't detect libtraceevent-devel installed in the system. Simplification here to avoid these two ways of disabling builtin-trace.c and not having CONFIG_TRACE=y when libtraceevent-devel isn't installed is the clean way. From Athira: tools/perf/arch/powerpc/util/Build -perf-y += kvm-stat.o +perf-$(CONFIG_LIBTRACEEVENT) += kvm-stat.o Then, ditto for arm64 and s390, detected by container cross build tests. - s/390 uses test__checkevent_tracepoint() that is now only available if HAVE_LIBTRACEEVENT is defined, enclose the callsite with ifder HAVE_LIBTRACEEVENT. Also from Athira: With this change, I could successfully compile in these environment: - Without libtraceevent-devel installed - With libtraceevent-devel installed - With “make NO_LIBTRACEEVENT=1” Then, finally rename CONFIG_TRACEEVENT to CONFIG_LIBTRACEEVENT for consistency with other libraries detected in tools/perf/. Signed-off-by: Ian Rogers Tested-by: Arnaldo Carvalho de Melo Tested-by: Athira Rajeev Cc: Alexander Shishkin Cc: Jiri Olsa Cc: Mark Rutland Cc: Namhyung Kim Cc: Nick Desaulniers Cc: Peter Zijlstra Cc: Stephane Eranian Cc: bpf@vger.kernel.org Link: http://lore.kernel.org/lkml/20221205225940.3079667-3-irogers@google.com Signed-off-by: Arnaldo Carvalho de Melo

perf tools: Add evlist__add_sched_switch()

2022-10-06T11:03:53+00:00

Add a help to create a system-wide sched_switch event. One merit is that it sets the system-wide bit before adding it to evlist so that the libperf can handle the cpu and thread maps correctly. Reviewed-by: Adrian Hunter Signed-off-by: Namhyung Kim Cc: Ian Rogers Cc: Ingo Molnar Cc: Jiri Olsa Cc: Kan Liang Cc: Leo Yan Cc: Peter Zijlstra Link: https://lore.kernel.org/r/20221003204647.1481128-5-namhyung@kernel.org Signed-off-by: Arnaldo Carvalho de Melo

perf record: Allow multiple recording time ranges

2022-10-04T11:55:19+00:00

AUX area traces can produce too much data to record successfully or analyze subsequently. Add another means to reduce data collection by allowing multiple recording time ranges. This is useful, for instance, in cases where a workload produces predictably reproducible events in specific time ranges. Today we only have perf record -D to start at a specific region, or some complicated approach using snapshot mode and external scripts sending signals or using the fifos. But these approaches are difficult to set up compared with simply having perf do it. Extend perf record option -D/--delay option to specifying relative time stamps for start stop controlled by perf with the right time offset, for instance: perf record -e intel_pt// -D 10-20,30-40 to record 10ms to 20ms into the trace and 30ms to 40ms. Example: The example workload is: $ cat repeat-usleep.c int usleep(useconds_t usec); int usage(int ret, const char *msg) { if (msg) fprintf(stderr, "%s\n", msg); fprintf(stderr, "Usage is: repeat-usleep \n"); return ret; } int main(int argc, char *argv[]) { unsigned long usecs; char *end_ptr; if (argc != 2) return usage(1, "Error: Wrong number of arguments!"); errno = 0; usecs = strtoul(argv[1], &end_ptr, 0); if (errno || *end_ptr || usecs > UINT_MAX) return usage(1, "Error: Invalid argument!"); while (1) { int ret = usleep(usecs); if (ret & errno != EINTR) return usage(1, "Error: usleep() failed!"); } return 0; } $ perf record -e intel_pt//u --delay 10-20,40-70,110-160 -- ./repeat-usleep 500 Events disabled Events enabled Events disabled Events enabled Events disabled Events enabled Events disabled [ perf record: Woken up 5 times to write data ] [ perf record: Captured and wrote 0.204 MB perf.data ] Terminated A dlfilter is used to determine continuous data collection (timestamps less than 1ms apart): $ cat dlfilter-show-delays.c static __u64 start_time; static __u64 last_time; int start(void **data, void *ctx) { printf("%-17s\t%-9s\t%-6s\n", " Time", " Duration", " Delay"); return 0; } int filter_event_early(void *data, const struct perf_dlfilter_sample *sample, void *ctx) { __u64 delta; if (!sample->time) return 1; if (!last_time) goto out; delta = sample->time - last_time; if (delta < 1000000) goto out2;; printf("%17.9f\t%9.1f\t%6.1f\n", start_time / 1000000000.0, (last_time - start_time) / 1000000.0, delta / 1000000.0); out: start_time = sample->time; out2: last_time = sample->time; return 1; } int stop(void *data, void *ctx) { printf("%17.9f\t%9.1f\n", start_time / 1000000000.0, (last_time - start_time) / 1000000.0); return 0; } The result shows the times roughly match the --delay option: $ perf script --itrace=qb --dlfilter dlfilter-show-delays.so Time Duration Delay 39215.302317300 9.7 20.5 39215.332480217 30.4 40.9 39215.403837717 49.8 Signed-off-by: Adrian Hunter Cc: Alexey Bayduraev Cc: Andi Kleen Cc: Ian Rogers Cc: Jiri Olsa Cc: Namhyung Kim Link: https://lore.kernel.org/r/20220824072814.16422-6-adrian.hunter@intel.com Signed-off-by: Arnaldo Carvalho de Melo

perf evlist: Add evlist__{en/dis}able_non_dummy()

2022-10-04T11:55:19+00:00

Dummy events are used to provide sideband information like MMAP events that are always needed even when main events are disabled. Add functions that take that into account. Signed-off-by: Adrian Hunter Acked-by: Ian Rogers Cc: Alexey Bayduraev Cc: Andi Kleen Cc: Jiri Olsa Cc: Namhyung Kim Link: https://lore.kernel.org/r/20220824072814.16422-5-adrian.hunter@intel.com Signed-off-by: Arnaldo Carvalho de Melo

perf record: Change evlist->ctl_fd to use fdarray_flag__non_perf_event

2022-10-04T11:55:19+00:00

Patch "perf record: Fix way of handling non-perf-event pollfds" added a generic way to handle non-perf-event file descriptors like evlist->ctl_fd. Use it instead of handling evlist->ctl_fd separately. Signed-off-by: Adrian Hunter Acked-by: Ian Rogers Cc: Alexey Bayduraev Cc: Andi Kleen Cc: Jiri Olsa Cc: Namhyung Kim Link: https://lore.kernel.org/r/20220824072814.16422-4-adrian.hunter@intel.com Signed-off-by: Arnaldo Carvalho de Melo

perf x86 evlist: Add default hybrid events for perf stat

2022-07-29T16:42:35+00:00

Provide a new solution to replace the reverted commit ac2dc29edd21f9ec ("perf stat: Add default hybrid events") For the default software attrs, nothing is changed. For the default hardware attrs, create a new evsel for each hybrid pmu. With the new solution, adding a new default attr will not require the special support for the hybrid platform anymore. Also, the "--detailed" is supported on the hybrid platform With the patch, $ perf stat -a -ddd sleep 1 Performance counter stats for 'system wide': 32,231.06 msec cpu-clock # 32.056 CPUs utilized 529 context-switches # 16.413 /sec 32 cpu-migrations # 0.993 /sec 69 page-faults # 2.141 /sec 176,754,151 cpu_core/cycles/ # 5.484 M/sec (41.65%) 161,695,280 cpu_atom/cycles/ # 5.017 M/sec (49.92%) 48,595,992 cpu_core/instructions/ # 1.508 M/sec (49.98%) 32,363,337 cpu_atom/instructions/ # 1.004 M/sec (58.26%) 10,088,639 cpu_core/branches/ # 313.010 K/sec (58.31%) 6,390,582 cpu_atom/branches/ # 198.274 K/sec (58.26%) 846,201 cpu_core/branch-misses/ # 26.254 K/sec (66.65%) 676,477 cpu_atom/branch-misses/ # 20.988 K/sec (58.27%) 14,290,070 cpu_core/L1-dcache-loads/ # 443.363 K/sec (66.66%) 9,983,532 cpu_atom/L1-dcache-loads/ # 309.749 K/sec (58.27%) 740,725 cpu_core/L1-dcache-load-misses/ # 22.982 K/sec (66.66%) cpu_atom/L1-dcache-load-misses/ 480,441 cpu_core/LLC-loads/ # 14.906 K/sec (66.67%) 326,570 cpu_atom/LLC-loads/ # 10.132 K/sec (58.27%) 329 cpu_core/LLC-load-misses/ # 10.208 /sec (66.68%) 0 cpu_atom/LLC-load-misses/ # 0.000 /sec (58.32%) cpu_core/L1-icache-loads/ 21,982,491 cpu_atom/L1-icache-loads/ # 682.028 K/sec (58.43%) 4,493,189 cpu_core/L1-icache-load-misses/ # 139.406 K/sec (33.34%) 4,711,404 cpu_atom/L1-icache-load-misses/ # 146.176 K/sec (50.08%) 13,713,090 cpu_core/dTLB-loads/ # 425.462 K/sec (33.34%) 9,384,727 cpu_atom/dTLB-loads/ # 291.170 K/sec (50.08%) 157,387 cpu_core/dTLB-load-misses/ # 4.883 K/sec (33.33%) 108,328 cpu_atom/dTLB-load-misses/ # 3.361 K/sec (50.08%) cpu_core/iTLB-loads/ cpu_atom/iTLB-loads/ 37,655 cpu_core/iTLB-load-misses/ # 1.168 K/sec (33.32%) 61,661 cpu_atom/iTLB-load-misses/ # 1.913 K/sec (50.03%) cpu_core/L1-dcache-prefetches/ cpu_atom/L1-dcache-prefetches/ cpu_core/L1-dcache-prefetch-misses/ cpu_atom/L1-dcache-prefetch-misses/ 1.005466919 seconds time elapsed Signed-off-by: Kan Liang Acked-by: Ian Rogers Acked-by: Namhyung Kim Cc: Alexander Shishkin Cc: Andi Kleen Cc: Ingo Molnar Cc: Jiri Olsa Cc: Peter Zijlstra Link: https://lore.kernel.org/r/20220721065706.2886112-5-zhengjun.xing@linux.intel.com Signed-off-by: Xing Zhengjun Signed-off-by: Arnaldo Carvalho de Melo

perf evlist: Always use arch_evlist__add_default_attrs()

2022-07-29T16:41:59+00:00

Current perf stat uses the evlist__add_default_attrs() to add the generic default attrs, and uses arch_evlist__add_default_attrs() to add the Arch specific default attrs, e.g., Topdown for x86. It works well for the non-hybrid platforms. However, for a hybrid platform, the hard code generic default attrs don't work. Uses arch_evlist__add_default_attrs() to replace the evlist__add_default_attrs(). The arch_evlist__add_default_attrs() is modified to invoke the same __evlist__add_default_attrs() for the generic default attrs. No functional change. Add default_null_attrs[] to indicate the arch specific attrs. No functional change for the arch specific default attrs either. Signed-off-by: Kan Liang Acked-by: Ian Rogers Acked-by: Namhyung Kim Cc: Alexander Shishkin Cc: Andi Kleen Cc: Ingo Molnar Cc: Jiri Olsa Cc: Peter Zijlstra Link: https://lore.kernel.org/r/20220721065706.2886112-4-zhengjun.xing@linux.intel.com Signed-off-by: Xing Zhengjun Signed-off-by: Arnaldo Carvalho de Melo