Commit 0a892c1c94 ("perf record: Add dummy event during system wide
synthesis") reveals an issue with Intel PT system wide tracing.
Specifically that Intel PT already adds a dummy tracking event, and it
is not the first event. Adding another dummy tracking event causes
duplicated sideband events. Fix by checking for an existing dummy
tracking event first.
Example showing duplicated switch events:
Before:
# perf record -a -e intel_pt//u uname
Linux
[ perf record: Woken up 1 times to write data ]
[ perf record: Captured and wrote 0.895 MB perf.data ]
# perf script --no-itrace --show-switch-events | head
swapper 0 [007] 6390.516222: PERF_RECORD_SWITCH_CPU_WIDE OUT preempt next pid/tid: 11/11
swapper 0 [007] 6390.516222: PERF_RECORD_SWITCH_CPU_WIDE OUT preempt next pid/tid: 11/11
rcu_sched 11 [007] 6390.516223: PERF_RECORD_SWITCH_CPU_WIDE IN prev pid/tid: 0/0
rcu_sched 11 [007] 6390.516224: PERF_RECORD_SWITCH_CPU_WIDE IN prev pid/tid: 0/0
rcu_sched 11 [007] 6390.516227: PERF_RECORD_SWITCH_CPU_WIDE OUT next pid/tid: 0/0
rcu_sched 11 [007] 6390.516227: PERF_RECORD_SWITCH_CPU_WIDE OUT next pid/tid: 0/0
swapper 0 [007] 6390.516228: PERF_RECORD_SWITCH_CPU_WIDE IN prev pid/tid: 11/11
swapper 0 [007] 6390.516228: PERF_RECORD_SWITCH_CPU_WIDE IN prev pid/tid: 11/11
swapper 0 [002] 6390.516415: PERF_RECORD_SWITCH_CPU_WIDE OUT preempt next pid/tid: 5556/5559
swapper 0 [002] 6390.516416: PERF_RECORD_SWITCH_CPU_WIDE OUT preempt next pid/tid: 5556/5559
After:
# perf record -a -e intel_pt//u uname
Linux
[ perf record: Woken up 1 times to write data ]
[ perf record: Captured and wrote 0.868 MB perf.data ]
# perf script --no-itrace --show-switch-events | head
swapper 0 [005] 6450.567013: PERF_RECORD_SWITCH_CPU_WIDE OUT preempt next pid/tid: 7179/7181
perf 7181 [005] 6450.567014: PERF_RECORD_SWITCH_CPU_WIDE IN prev pid/tid: 0/0
perf 7181 [005] 6450.567028: PERF_RECORD_SWITCH_CPU_WIDE OUT next pid/tid: 0/0
swapper 0 [005] 6450.567029: PERF_RECORD_SWITCH_CPU_WIDE IN prev pid/tid: 7179/7181
swapper 0 [005] 6450.571699: PERF_RECORD_SWITCH_CPU_WIDE OUT preempt next pid/tid: 11/11
rcu_sched 11 [005] 6450.571700: PERF_RECORD_SWITCH_CPU_WIDE IN prev pid/tid: 0/0
rcu_sched 11 [005] 6450.571702: PERF_RECORD_SWITCH_CPU_WIDE OUT next pid/tid: 0/0
swapper 0 [005] 6450.571703: PERF_RECORD_SWITCH_CPU_WIDE IN prev pid/tid: 11/11
swapper 0 [005] 6450.579703: PERF_RECORD_SWITCH_CPU_WIDE OUT preempt next pid/tid: 11/11
rcu_sched 11 [005] 6450.579704: PERF_RECORD_SWITCH_CPU_WIDE IN prev pid/tid: 0/0
Signed-off-by: Adrian Hunter <adrian.hunter@intel.com>
Tested-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Jiri Olsa <jolsa@redhat.com>
Link: http://lore.kernel.org/lkml/20200629091955.17090-3-adrian.hunter@intel.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
As they are 'struct evsel' methods or related routines, not part of
tools/lib/perf/, aka libperf, to whom the perf_ prefix belongs.
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: Jiri Olsa <jolsa@kernel.org>
Cc: Namhyung Kim <namhyung@kernel.org>
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
As the "perf_" prefix should be restricted to functions and types in
tools/lib/perf/, aka libperf, this way we reduce a bit the confusion for
types only in libperf or the ones in the more contained tools/perf/
project.
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: Jiri Olsa <jolsa@kernel.org>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Song Liu <songliubraving@fb.com>
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
A new branch sample type PERF_SAMPLE_BRANCH_HW_INDEX has been introduced
in latest kernel.
Enable HW_INDEX by default in LBR call stack mode.
If kernel doesn't support the sample type, switching it off.
Add HW_INDEX in attr_fprintf as well. User can check whether the branch
sample type is set via debug information or header.
Committer testing:
First collect some samples with LBR callchains, system wide, for a few
seconds:
# perf record --call-graph lbr -a sleep 5
[ perf record: Woken up 1 times to write data ]
[ perf record: Captured and wrote 0.625 MB perf.data (224 samples) ]
#
Now lets use 'perf evlist -v' to look at the branch_sample_type:
# perf evlist -v
cycles: size: 120, { sample_period, sample_freq }: 4000, sample_type: IP|TID|TIME|CALLCHAIN|CPU|PERIOD|BRANCH_STACK, read_format: ID, disabled: 1, inherit: 1, mmap: 1, comm: 1, freq: 1, task: 1, precise_ip: 3, sample_id_all: 1, exclude_guest: 1, mmap2: 1, comm_exec: 1, ksymbol: 1, bpf_event: 1, branch_sample_type: USER|CALL_STACK|NO_FLAGS|NO_CYCLES|HW_INDEX
#
So the machine has the kernel feature, and it was correctly added to
perf_event_attr.branch_sample_type, for the default 'cycles' event.
If we do it in another machine, where the kernel lacks the HW_INDEX
feature, we get:
# perf record --call-graph lbr -a sleep 2s
[ perf record: Woken up 1 times to write data ]
[ perf record: Captured and wrote 1.690 MB perf.data (499 samples) ]
# perf evlist -v
cycles: size: 120, { sample_period, sample_freq }: 4000, sample_type: IP|TID|TIME|CALLCHAIN|CPU|PERIOD|BRANCH_STACK, read_format: ID, disabled: 1, inherit: 1, mmap: 1, comm: 1, freq: 1, task: 1, precise_ip: 3, sample_id_all: 1, exclude_guest: 1, mmap2: 1, comm_exec: 1, ksymbol: 1, bpf_event: 1, branch_sample_type: USER|CALL_STACK|NO_FLAGS|NO_CYCLES
#
No HW_INDEX in attr.branch_sample_type.
Signed-off-by: Kan Liang <kan.liang@linux.intel.com>
Tested-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: Alexey Budankov <alexey.budankov@linux.intel.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Jiri Olsa <jolsa@redhat.com>
Cc: Mathieu Poirier <mathieu.poirier@linaro.org>
Cc: Michael Ellerman <mpe@ellerman.id.au>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Pavel Gerasimov <pavel.gerasimov@intel.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Ravi Bangoria <ravi.bangoria@linux.ibm.com>
Cc: Stephane Eranian <eranian@google.com>
Cc: Vitaly Slobodskoy <vitaly.slobodskoy@intel.com>
Link: http://lore.kernel.org/lkml/20200228163011.19358-3-kan.liang@linux.intel.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
The low level index of raw branch records for the most recent branch can
be recorded in a sample with PERF_SAMPLE_BRANCH_HW_INDEX
branch_sample_type. Extend struct branch_stack to support it.
However, if the PERF_SAMPLE_BRANCH_HW_INDEX is not applied, only nr and
entries[] will be output by kernel. The pointer of entries[] could be
wrong, since the output format is different with new struct
branch_stack. Add a variable no_hw_idx in struct perf_sample to
indicate whether the hw_idx is output. Add get_branch_entry() to return
corresponding pointer of entries[0].
To make dummy branch sample consistent as new branch sample, add hw_idx
in struct dummy_branch_stack for cs-etm and intel-pt.
Apply the new struct branch_stack for synthetic events as well.
Extend test case sample-parsing to support new struct branch_stack.
Committer notes:
Renamed get_branch_entries() to perf_sample__branch_entries() to have
proper namespacing and pave the way for this to be moved to libperf,
eventually.
Add 'static' to that inline as it is in a header.
Add 'hw_idx' to 'struct dummy_branch_stack' in cs-etm.c to fix the build
on arm64.
Signed-off-by: Kan Liang <kan.liang@linux.intel.com>
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: Alexey Budankov <alexey.budankov@linux.intel.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Jiri Olsa <jolsa@redhat.com>
Cc: Mathieu Poirier <mathieu.poirier@linaro.org>
Cc: Michael Ellerman <mpe@ellerman.id.au>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Pavel Gerasimov <pavel.gerasimov@intel.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Ravi Bangoria <ravi.bangoria@linux.ibm.com>
Cc: Stephane Eranian <eranian@google.com>
Cc: Vitaly Slobodskoy <vitaly.slobodskoy@intel.com>
Link: http://lore.kernel.org/lkml/20200228163011.19358-2-kan.liang@linux.intel.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Restructure event reading to use affinity to minimize the number of IPIs
needed.
Before on a large test case with 94 CPUs:
% time seconds usecs/call calls errors syscall
------ ----------- ----------- --------- --------- ----------------
3.16 0.106079 4 22082 read
After:
3.43 0.081295 3 22082 read
Signed-off-by: Andi Kleen <ak@linux.intel.com>
Acked-by: Jiri Olsa <jolsa@kernel.org>
Link: http://lore.kernel.org/lkml/20191121001522.180827-11-andi@firstfloor.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Restructure the event opening in perf stat to cycle through the events
by CPU after setting affinity to that CPU.
This eliminates IPI overhead in the perf API.
We have to loop through the CPU in the outter builtin-stat code instead
of leaving that to low level functions.
It has to change the weak group fallback strategy slightly. Since we
cannot easily undo the opens for other CPUs move the weak group retry to
a separate loop.
Before with a large test case with 94 CPUs:
% time seconds usecs/call calls errors syscall
------ ----------- ----------- --------- --------- ----------------
42.75 4.050910 67 60046 110 perf_event_open
After:
26.86 0.944396 16 58069 110 perf_event_open
(the number changes slightly because the weak group retries
work differently and the test case relies on weak groups)
Committer notes:
Added one of the hunks in a patch provided by Andi after I noticed that
the "event times" 'perf test' entry was segfaulting.
Signed-off-by: Andi Kleen <ak@linux.intel.com>
Acked-by: Jiri Olsa <jolsa@kernel.org>
Link: http://lore.kernel.org/lkml/20191121001522.180827-10-andi@firstfloor.org
Link: http://lore.kernel.org/lkml/20191127232657.GL84886@tassilo.jf.intel.com # Fix
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Some uncore metrics don't work as expected. For example, on
cascadelakex:
root@lkp-csl-2sp2:~# perf stat -M UNC_M_PMM_BANDWIDTH.TOTAL -a -- sleep 1
Performance counter stats for 'system wide':
1841092 unc_m_pmm_rpq_inserts
3680816 unc_m_pmm_wpq_inserts
1.001775055 seconds time elapsed
root@lkp-csl-2sp2:~# perf stat -M UNC_M_PMM_READ_LATENCY -a -- sleep 1
Performance counter stats for 'system wide':
860649746 unc_m_pmm_rpq_occupancy.all
1840557 unc_m_pmm_rpq_inserts
12790627455 unc_m_clockticks
1.001773348 seconds time elapsed
No metrics 'UNC_M_PMM_BANDWIDTH.TOTAL' or 'UNC_M_PMM_READ_LATENCY' are
reported.
The issue is, the case of an alias expanding to mulitple events is not
supported, typically the uncore events. (see comments in
find_evsel_group()).
For UNC_M_PMM_BANDWIDTH.TOTAL in above example, the expanded event group
is '{unc_m_pmm_rpq_inserts,unc_m_pmm_wpq_inserts}:W', but the actual
events passed to find_evsel_group are:
unc_m_pmm_rpq_inserts
unc_m_pmm_rpq_inserts
unc_m_pmm_rpq_inserts
unc_m_pmm_rpq_inserts
unc_m_pmm_rpq_inserts
unc_m_pmm_rpq_inserts
unc_m_pmm_wpq_inserts
unc_m_pmm_wpq_inserts
unc_m_pmm_wpq_inserts
unc_m_pmm_wpq_inserts
unc_m_pmm_wpq_inserts
unc_m_pmm_wpq_inserts
For this multiple events case, it's not supported well.
This patch introduces a new field 'metric_leader' in struct evsel. The
first event is considered as a metric leader. For the rest of same
events, they point to the first event via it's metric_leader field in
struct evsel.
This design is for adding the counting results of all same events to the
first event in group (the metric_leader).
With this patch,
root@lkp-csl-2sp2:~# perf stat -M UNC_M_PMM_BANDWIDTH.TOTAL -a -- sleep 1
Performance counter stats for 'system wide':
1842108 unc_m_pmm_rpq_inserts # 337.2 MB/sec UNC_M_PMM_BANDWIDTH.TOTAL
3682209 unc_m_pmm_wpq_inserts
1.001819706 seconds time elapsed
root@lkp-csl-2sp2:~# perf stat -M UNC_M_PMM_READ_LATENCY -a -- sleep 1
Performance counter stats for 'system wide':
861970685 unc_m_pmm_rpq_occupancy.all # 219.4 ns UNC_M_PMM_READ_LATENCY
1842772 unc_m_pmm_rpq_inserts
12790196356 unc_m_clockticks
1.001749103 seconds time elapsed
Now we can see the correct metrics 'UNC_M_PMM_BANDWIDTH.TOTAL' and
'UNC_M_PMM_READ_LATENCY'.
Signed-off-by: Jin Yao <yao.jin@linux.intel.com>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Jiri Olsa <jolsa@kernel.org>
Cc: Kan Liang <kan.liang@linux.intel.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Link: http://lore.kernel.org/lkml/20190828055932.8269-5-yao.jin@linux.intel.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
