linux.git/tools/perf/util/stat.h, branch v5.8

perf metricgroup: Add options to not group or merge

2020-05-28T13:03:28+00:00

Add --metric-no-group that causes all events within metrics to not be
grouped. This can allow the event to get more time when multiplexed, but
may also lower accuracy.
Add --metric-no-merge option. By default events in different metrics may
be shared if the group of events for one metric is the same or larger
than that of the second. Sharing may increase or lower accuracy and so
is now configurable.

Signed-off-by: Ian Rogers 
Acked-by: Jiri Olsa 
Cc: Alexander Shishkin 
Cc: Andi Kleen 
Cc: Andrii Nakryiko 
Cc: Cong Wang 
Cc: Jin Yao 
Cc: John Garry 
Cc: Kajol Jain 
Cc: Kan Liang 
Cc: Kim Phillips 
Cc: Mark Rutland 
Cc: Namhyung Kim 
Cc: Paul Clarke 
Cc: Peter Zijlstra 
Cc: Song Liu 
Cc: Srikar Dronamraju 
Cc: Stephane Eranian 
Cc: Vince Weaver 
Cc: bpf@vger.kernel.org
Cc: netdev@vger.kernel.org
Link: http://lore.kernel.org/lkml/20200520182011.32236-7-irogers@google.com
Signed-off-by: Arnaldo Carvalho de Melo

perf config: Add stat.big-num support

2020-05-28T13:03:27+00:00

Add support for new "stat.big-num" boolean option.

This allows a user to set a default for "--no-big-num" for "perf stat"
commands.

--
  $ perf config stat.big-num
  $ perf stat --event cycles /bin/true

   Performance counter stats for '/bin/true':

             778,849      cycles
  [...]
  $ perf config stat.big-num=false
  $ perf config stat.big-num
  stat.big-num=false
  $ perf stat --event cycles /bin/true

   Performance counter stats for '/bin/true':

              769622      cycles
  [...]
--

There is an interaction with "--field-separator" that must be
accommodated, such that specifying "--big-num --field-separator={x}"
still reports an invalid combination of options.

Documentation for perf-config and perf-stat updated.

Signed-off-by: Paul Clarke 
Tested-by: Arnaldo Carvalho de Melo 
Link: http://lore.kernel.org/lkml/1589991815-17951-1-git-send-email-pc@us.ibm.com
Signed-off-by: Arnaldo Carvalho de Melo

perf stat: Report summary for interval mode

2020-05-28T13:03:27+00:00

Currently 'perf stat' supports to print counts at regular interval (-I),
but it's not very easy for user to get the overall statistics.

The patch uses 'evsel->prev_raw_counts' to get counts for summary.  Copy
the counts to 'evsel->counts' after printing the interval results.
Next, we just follow the non-interval processing.

Let's see some examples,

 root@kbl-ppc:~# perf stat -e cycles -I1000 --interval-count 2
 #           time             counts unit events
      1.000412064          2,281,114      cycles
      2.001383658          2,547,880      cycles

  Performance counter stats for 'system wide':

          4,828,994      cycles

        2.002860349 seconds time elapsed

 root@kbl-ppc:~# perf stat -e cycles,instructions -I1000 --interval-count 2
 #           time             counts unit events
      1.000389902          1,536,093      cycles
      1.000389902            420,226      instructions              #    0.27  insn per cycle
      2.001433453          2,213,952      cycles
      2.001433453            735,465      instructions              #    0.33  insn per cycle

  Performance counter stats for 'system wide':

          3,750,045      cycles
          1,155,691      instructions              #    0.31  insn per cycle

        2.003023361 seconds time elapsed

 root@kbl-ppc:~# perf stat -M CPI,IPC -I1000 --interval-count 2
 #           time             counts unit events
      1.000435121            905,303      inst_retired.any          #      2.9 CPI
      1.000435121          2,663,333      cycles
      1.000435121            914,702      inst_retired.any          #      0.3 IPC
      1.000435121          2,676,559      cpu_clk_unhalted.thread
      2.001615941          1,951,092      inst_retired.any          #      1.8 CPI
      2.001615941          3,551,357      cycles
      2.001615941          1,950,837      inst_retired.any          #      0.5 IPC
      2.001615941          3,551,044      cpu_clk_unhalted.thread

  Performance counter stats for 'system wide':

          2,856,395      inst_retired.any          #      2.2 CPI
          6,214,690      cycles
          2,865,539      inst_retired.any          #      0.5 IPC
          6,227,603      cpu_clk_unhalted.thread

        2.003403078 seconds time elapsed

Committer testing:

Before:

  # perf stat -e cycles -I1000 --interval-count 2
  #           time             counts unit events
       1.000618627         26,877,408      cycles
       2.001417968        233,672,829      cycles
  #

After:

  # perf stat -e cycles -I1000 --interval-count 2
  #           time             counts unit events
       1.001531815      5,341,388,792      cycles
       2.002936530        100,073,912      cycles

   Performance counter stats for 'system wide':

       5,441,462,704      cycles

         2.004893794 seconds time elapsed

  #

Signed-off-by: Jin Yao 
Reviewed-by: Jiri Olsa 
Tested-by: Arnaldo Carvalho de Melo 
Cc: Alexander Shishkin 
Cc: Andi Kleen 
Cc: Kan Liang 
Cc: Peter Zijlstra 
Link: http://lore.kernel.org/lkml/20200520042737.24160-6-yao.jin@linux.intel.com
Signed-off-by: Arnaldo Carvalho de Melo

perf stat: Save aggr value to first member of prev_raw_counts

2020-05-28T13:03:27+00:00

To collect the overall statistics for interval mode, we copy the counts
from evsel->prev_raw_counts to evsel->counts.

For AGGR_GLOBAL mode, because the perf_stat_process_counter creates aggr
values from per cpu values, but the per cpu values are 0, so the
calculated aggr values will be always 0.

This patch uses a trick that saves the previous aggr value to the first
member of perf_counts, then aggr calculation in process_counter_values
can work correctly for AGGR_GLOBAL.

 v6:
 ---
 Add comments in perf_evlist__save_aggr_prev_raw_counts.

Signed-off-by: Jin Yao 
Reviewed-by: Jiri Olsa 
Cc: Alexander Shishkin 
Cc: Andi Kleen 
Cc: Jin Yao 
Cc: Kan Liang 
Cc: Peter Zijlstra 
Link: http://lore.kernel.org/lkml/20200520042737.24160-5-yao.jin@linux.intel.com
Signed-off-by: Arnaldo Carvalho de Melo

perf stat: Copy counts from prev_raw_counts to evsel->counts

2020-05-28T13:03:27+00:00

It would be useful to support the overall statistics for perf-stat
interval mode. For example, report the summary at the end of "perf-stat
-I" output.

But since perf-stat can support many aggregation modes, such as
--per-thread, --per-socket, -M and etc, we need a solution which doesn't
bring much complexity.

The idea is to use 'evsel->prev_raw_counts' which is updated in each
interval and it's saved with the latest counts. Before reporting the
summary, we copy the counts from evsel->prev_raw_counts to
evsel->counts, and next we just follow non-interval processing.

 v5:
 ---
 Don't save the previous aggr value to the member of [cpu0,thread0]
 in perf_counts. Originally that was a trick because the
 perf_stat_process_counter would create aggr values from per cpu
 values. But we don't need to do that all the time. We will
 handle it in next patch.

Signed-off-by: Jin Yao 
Reviewed-by: Jiri Olsa 
Cc: Alexander Shishkin 
Cc: Andi Kleen 
Cc: Jin Yao 
Cc: Kan Liang 
Cc: Peter Zijlstra 
Link: http://lore.kernel.org/lkml/20200520042737.24160-4-yao.jin@linux.intel.com
Signed-off-by: Arnaldo Carvalho de Melo

perf stat: Show percore counts in per CPU output

2020-03-04T13:34:09+00:00

We have supported the event modifier "percore" which sums up the event
counts for all hardware threads in a core and show the counts per core.

For example,

 # perf stat -e cpu/event=cpu-cycles,percore/ -a -A -- sleep 1

  Performance counter stats for 'system wide':

 S0-D0-C0                395,072      cpu/event=cpu-cycles,percore/
 S0-D0-C1                851,248      cpu/event=cpu-cycles,percore/
 S0-D0-C2                954,226      cpu/event=cpu-cycles,percore/
 S0-D0-C3              1,233,659      cpu/event=cpu-cycles,percore/

This patch provides a new option "--percore-show-thread". It is used
with event modifier "percore" together to sum up the event counts for
all hardware threads in a core but show the counts per hardware thread.

This is essentially a replacement for the any bit (which is gone in
Icelake). Per core counts are useful for some formulas, e.g. CoreIPC.
The original percore version was inconvenient to post process. This
variant matches the output of the any bit.

With this patch, for example,

 # perf stat -e cpu/event=cpu-cycles,percore/ -a -A --percore-show-thread  -- sleep 1

  Performance counter stats for 'system wide':

 CPU0               2,453,061      cpu/event=cpu-cycles,percore/
 CPU1               1,823,921      cpu/event=cpu-cycles,percore/
 CPU2               1,383,166      cpu/event=cpu-cycles,percore/
 CPU3               1,102,652      cpu/event=cpu-cycles,percore/
 CPU4               2,453,061      cpu/event=cpu-cycles,percore/
 CPU5               1,823,921      cpu/event=cpu-cycles,percore/
 CPU6               1,383,166      cpu/event=cpu-cycles,percore/
 CPU7               1,102,652      cpu/event=cpu-cycles,percore/

We can see counts are duplicated in CPU pairs (CPU0/CPU4, CPU1/CPU5,
CPU2/CPU6, CPU3/CPU7).

The interval mode also works. For example,

 # perf stat -e cpu/event=cpu-cycles,percore/ -a -A --percore-show-thread  -I 1000
 #           time CPU                    counts unit events
      1.000425421 CPU0                 925,032      cpu/event=cpu-cycles,percore/
      1.000425421 CPU1                 430,202      cpu/event=cpu-cycles,percore/
      1.000425421 CPU2                 436,843      cpu/event=cpu-cycles,percore/
      1.000425421 CPU3               1,192,504      cpu/event=cpu-cycles,percore/
      1.000425421 CPU4                 925,032      cpu/event=cpu-cycles,percore/
      1.000425421 CPU5                 430,202      cpu/event=cpu-cycles,percore/
      1.000425421 CPU6                 436,843      cpu/event=cpu-cycles,percore/
      1.000425421 CPU7               1,192,504      cpu/event=cpu-cycles,percore/

If we offline CPU5, the result is:

 # perf stat -e cpu/event=cpu-cycles,percore/ -a -A --percore-show-thread -- sleep 1

  Performance counter stats for 'system wide':

 CPU0               2,752,148      cpu/event=cpu-cycles,percore/
 CPU1               1,009,312      cpu/event=cpu-cycles,percore/
 CPU2               2,784,072      cpu/event=cpu-cycles,percore/
 CPU3               2,427,922      cpu/event=cpu-cycles,percore/
 CPU4               2,752,148      cpu/event=cpu-cycles,percore/
 CPU6               2,784,072      cpu/event=cpu-cycles,percore/
 CPU7               2,427,922      cpu/event=cpu-cycles,percore/

        1.001416041 seconds time elapsed

 v4:
 ---
 Ravi Bangoria reports an issue in v3. Once we offline a CPU,
 the output is not correct. The issue is we should use the cpu
 idx in print_percore_thread rather than using the cpu value.

 v3:
 ---
 1. Fix the interval mode output error
 2. Use cpu value (not cpu index) in config->aggr_get_id().
 3. Refine the code according to Jiri's comments.

 v2:
 ---
 Add the explanation in change log. This is essentially a replacement
 for the any bit. No code change.

Signed-off-by: Jin Yao 
Tested-by: Ravi Bangoria 
Acked-by: Jiri Olsa 
Cc: Alexander Shishkin 
Cc: Andi Kleen 
Cc: Kan Liang 
Cc: Peter Zijlstra 
Link: http://lore.kernel.org/lkml/20200214080452.26402-1-yao.jin@linux.intel.com
Signed-off-by: Arnaldo Carvalho de Melo

perf stat: Use affinity for opening events

2019-11-29T15:20:45+00:00

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 
Acked-by: Jiri Olsa 
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

perf stat: Add --per-node agregation support

2019-11-06T18:49:39+00:00

Adding new --per-node option to aggregate counts per NUMA
nodes for system-wide mode measurements.

You can specify --per-node in live mode:

  # perf stat  -a -I 1000 -e cycles --per-node
  #           time node   cpus             counts unit events
       1.000542550 N0       20          6,202,097      cycles
       1.000542550 N1       20            639,559      cycles
       2.002040063 N0       20          7,412,495      cycles
       2.002040063 N1       20          2,185,577      cycles
       3.003451699 N0       20          6,508,917      cycles
       3.003451699 N1       20            765,607      cycles
  ...

Or in the record/report stat session:

  # perf stat record -a -I 1000 -e cycles
  #           time             counts unit events
       1.000536937         10,008,468      cycles
       2.002090152          9,578,539      cycles
       3.003625233          7,647,869      cycles
       4.005135036          7,032,086      cycles
  ^C     4.340902364          3,923,893      cycles

  # perf stat report --per-node
  #           time node   cpus             counts unit events
       1.000536937 N0       20          9,355,086      cycles
       1.000536937 N1       20            653,382      cycles
       2.002090152 N0       20          7,712,838      cycles
       2.002090152 N1       20          1,865,701      cycles
       3.003625233 N0       20          6,604,441      cycles
       3.003625233 N1       20          1,043,428      cycles
       4.005135036 N0       20          6,350,522      cycles
       4.005135036 N1       20            681,564      cycles
       4.340902364 N0       20          3,403,188      cycles
       4.340902364 N1       20            520,705      cycles

Signed-off-by: Jiri Olsa 
Cc: Alexander Shishkin 
Cc: Alexey Budankov 
Cc: Andi Kleen 
Cc: Joe Mario 
Cc: Kan Liang 
Cc: Michael Petlan 
Cc: Namhyung Kim 
Cc: Peter Zijlstra 
Link: http://lkml.kernel.org/r/20190904073415.723-4-jolsa@kernel.org
Signed-off-by: Arnaldo Carvalho de Melo

perf stat: Support --all-kernel/--all-user

2019-10-15T11:39:42+00:00

'perf record' has supported --all-kernel / --all-user to configure all
used events to run in kernel space or run in user space. But 'perf stat'
doesn't support these options.

It would be useful to support these options in 'perf stat' too to keep
the same semantics available in both tools.

Signed-off-by: Jin Yao 
Acked-by: Jiri Olsa 
Cc: Alexander Shishkin 
Cc: Andi Kleen 
Cc: Kan Liang 
Cc: Peter Zijlstra 
Link: http://lore.kernel.org/lkml/20191011050545.3899-1-yao.jin@linux.intel.com
Signed-off-by: Arnaldo Carvalho de Melo

perf stat: Reset previous counts on repeat with interval

2019-09-20T13:28:26+00:00

When using 'perf stat' with repeat and interval option, it shows wrong
values for events.

The wrong values will be shown for the first interval on the second and
subsequent repetitions.

Without the fix:

  # perf stat -r 3 -I 2000 -e faults -e sched:sched_switch -a sleep 5

     2.000282489                 53      faults
     2.000282489                513      sched:sched_switch
     4.005478208              3,721      faults
     4.005478208              2,666      sched:sched_switch
     5.025470933                395      faults
     5.025470933              1,307      sched:sched_switch
     2.009602825 1,84,46,74,40,73,70,95,47,520      faults 		<------
     2.009602825 1,84,46,74,40,73,70,95,49,568      sched:sched_switch  <------
     4.019612206              4,730      faults
     4.019612206              2,746      sched:sched_switch
     5.039615484              3,953      faults
     5.039615484              1,496      sched:sched_switch
     2.000274620 1,84,46,74,40,73,70,95,47,520      faults		<------
     2.000274620 1,84,46,74,40,73,70,95,47,520      sched:sched_switch	<------
     4.000480342              4,282      faults
     4.000480342              2,303      sched:sched_switch
     5.000916811              1,322      faults
     5.000916811              1,064      sched:sched_switch
  #

prev_raw_counts is allocated when using intervals. This is used when
calculating the difference in the counts of events when using interval.

The current counts are stored in prev_raw_counts to calculate the
differences in the next iteration.

On the first interval of the second and subsequent repetitions,
prev_raw_counts would be the values stored in the last interval of the
previous repetitions, while the current counts will only be for the
first interval of the current repetition.

Hence there is a possibility of events showing up as big number.

Fix this by resetting prev_raw_counts whenever perf stat repeats the
command.

With the fix:

  # perf stat -r 3 -I 2000 -e faults -e sched:sched_switch -a sleep 5

     2.019349347              2,597      faults
     2.019349347              2,753      sched:sched_switch
     4.019577372              3,098      faults
     4.019577372              2,532      sched:sched_switch
     5.019415481              1,879      faults
     5.019415481              1,356      sched:sched_switch
     2.000178813              8,468      faults
     2.000178813              2,254      sched:sched_switch
     4.000404621              7,440      faults
     4.000404621              1,266      sched:sched_switch
     5.040196079              2,458      faults
     5.040196079                556      sched:sched_switch
     2.000191939              6,870      faults
     2.000191939              1,170      sched:sched_switch
     4.000414103                541      faults
     4.000414103                902      sched:sched_switch
     5.000809863                450      faults
     5.000809863                364      sched:sched_switch
  #

Committer notes:

This was broken since the cset introducing the --interval feature, i.e.
--repeat + --interval wasn't tested at that point, add the Fixes tag so
that automatic scripts can pick this up.

Fixes: 13370a9b5bb8 ("perf stat: Add interval printing")
Signed-off-by: Srikar Dronamraju 
Acked-by: Jiri Olsa 
Tested-by: Arnaldo Carvalho de Melo 
Tested-by: Ravi Bangoria 
Cc: Namhyung Kim 
Cc: Naveen N. Rao 
Cc: Stephane Eranian 
Cc: stable@vger.kernel.org # v3.9+
Link: http://lore.kernel.org/lkml/20190904094738.9558-2-srikar@linux.vnet.ibm.com
[ Fixed up conflicts with libperf, i.e. some perf_{evsel,evlist} lost the 'perf' prefix ]
Signed-off-by: Arnaldo Carvalho de Melo