linux-stable.git/tools/perf, branch v5.17.2

perf vendor events: Update metrics for SkyLake Server

2022-04-08T11:59:02+00:00

commit 3bad20d7d129c3b3063658a0f83974dfe6dac5c4 upstream.

Based on TMA_metrics-full.csv version 4.3 at 01.org:
    https://download.01.org/perfmon/
Events are updated to version 1.26:
    https://download.01.org/perfmon/SKX
Json files generated by:
    https://github.com/intel/event-converter-for-linux-perf

Fixes were made that allow the skx-metrics.json to successfully
generate, bringing back TopdownL1 metrics.

Tested:

  $ perf test
  ...
    6: Parse event definition strings                                  : Ok
    7: Simple expression parser                                        : Ok
  ...
    9: Parse perf pmu format                                           : Ok
   10: PMU events                                                      :
   10.1: PMU event table sanity                                        : Ok
   10.2: PMU event map aliases                                         : Ok
   10.3: Parsing of PMU event table metrics                            : Ok
   10.4: Parsing of PMU event table metrics with fake PMUs             : Ok
  ...
   68: Parse and process metrics                                       : Ok
  ...
   88: perf stat metrics (shadow stat) test                            : Ok
   89: perf all metricgroups test                                      : Ok
   90: perf all metrics test                                           : Skip
   91: perf all PMU test                                               : Ok
  ...

90 skips due to a lack of floating point samples, which is
understandable.

Fixes: c4ad8fabd03f76ed ("perf vendor events: Update metrics for SkyLake Server")
Reviewed-by: Kan Liang 
Signed-off-by: Ian Rogers 
Cc: Alexander Shishkin 
Cc: Alexandre Torgue 
Cc: Andi Kleen 
Cc: Ingo Molnar 
Cc: James Clark 
Cc: Jin Yao 
Cc: Jiri Olsa 
Cc: John Garry 
Cc: Mark Rutland 
Cc: Maxime Coquelin 
Cc: Namhyung Kim 
Cc: Peter Zijlstra 
Cc: Stephane Eranian 
Cc: Zhengjun Xing 
Link: https://lore.kernel.org/r/20220201015858.1226914-3-irogers@google.com
Signed-off-by: Arnaldo Carvalho de Melo 
Signed-off-by: Greg Kroah-Hartman

perf stat: Fix forked applications enablement of counters

2022-04-08T11:58:32+00:00

[ Upstream commit d0a0a511493d269514fcbd852481cdca32c95350 ]

I have run into the following issue:

 # perf stat -a -e new_pmu/INSTRUCTION_7/ --  mytest -c1 7

 Performance counter stats for 'system wide':

                 0      new_pmu/INSTRUCTION_7/

       0.000366428 seconds time elapsed
 #

The new PMU for s390 counts the execution of certain CPU instructions.
The root cause is the extremely small run time of the mytest program. It
just executes some assembly instructions and then exits.

In above invocation the instruction is executed exactly one time (-c1
option). The PMU is expected to report this one time execution by a
counter value of one, but fails to do so in some cases, not all.

Debugging reveals the invocation of the child process is done
*before* the counter events are installed and enabled.

Tracing reveals that sometimes the child process starts and exits before
the event is installed on all CPUs. The more CPUs the machine has, the
more often this miscount happens.

Fix this by reversing the start of the work load after the events have
been installed on the specified CPUs. Now the comment also matches the
code.

Output after:

 # perf stat -a -e new_pmu/INSTRUCTION_7/ --  mytest -c1 7

 Performance counter stats for 'system wide':

                 1      new_pmu/INSTRUCTION_7/

       0.000366428 seconds time elapsed
 #

Now the correct result is reported rock solid all the time regardless
how many CPUs are online.

Reviewers notes:

Jiri:

Right, without -a the event has enable_on_exec so the race does not
matter, but it's a problem for system wide with fork.

Namhyung:

Agreed. Also we may move the enable_counters() and the clock code out of
the if block to be shared with the else block.

Fixes: acf2892270dcc428 ("perf stat: Use perf_evlist__prepare/start_workload()")
Signed-off-by: Thomas Richter 
Acked-by: Jiri Olsa 
Acked-by: Namhyung Kim 
Acked-by: Sumanth Korikkar 
Cc: Heiko Carstens 
Cc: Sven Schnelle 
Cc: Vasily Gorbik 
Link: https://lore.kernel.org/r/20220317155346.577384-1-tmricht@linux.ibm.com
Signed-off-by: Arnaldo Carvalho de Melo 
Signed-off-by: Sasha Levin

perf parse-events: Move slots only with topdown

2022-04-08T11:58:31+00:00

[ Upstream commit bc355822f0d9623b632069105d425c822d124cc8 ]

If slots isn't with a topdown event then moving it is unnecessary. For
example {instructions, slots} is re-ordered:

  $ perf stat -e '{instructions,slots}' -a sleep 1

   Performance counter stats for 'system wide':

         936,600,825      slots
         144,440,968      instructions

         1.006061423 seconds time elapsed

Which can break tools expecting the command line order to match the
printed order. It is necessary to move the slots event first when it
appears with topdown events. Add extra checking so that the slots event
is only moved in the case of there being a topdown event like:

  $ perf stat -e '{instructions,slots,topdown-fe-bound}' -a sleep 1

   Performance counter stats for 'system wide':

          2427568570      slots
           300927614      instructions
           551021649      topdown-fe-bound

         1.001771803 seconds time elapsed

Fixes: 94dbfd6781a0e87b ("perf parse-events: Architecture specific leader override")
Reported-by: Kan Liang 
Signed-off-by: Ian Rogers 
Tested-by: Kan Liang 
Cc: Alexander Shishkin 
Cc: Alexandre Torgue 
Cc: Andi Kleen 
Cc: Ingo Molnar 
Cc: James Clark 
Cc: Jiri Olsa 
Cc: John Garry 
Cc: Mark Rutland 
Cc: Maxime Coquelin 
Cc: Namhyung Kim 
Cc: Peter Zijlstra 
Cc: Stephane Eranian 
Cc: Xing Zhengjun 
Link: https://lore.kernel.org/r/20220321223344.1034479-1-irogers@google.com
Signed-off-by: Arnaldo Carvalho de Melo 
Signed-off-by: Sasha Levin

perf test arm64: Test unwinding using fame-pointer (fp) mode

2022-04-08T11:58:31+00:00

[ Upstream commit cd6382d82752737e43ef3617bb9e72913d2b1d47 ]

Add a shell script to check that the call-graphs generated using frame
pointers (--call-graph fp) are complete and not missing leaf functions:

  | $ perf test 88 -v
  |  88: Check Arm64 callgraphs are complete in fp mode                  :
  | --- start ---
  | test child forked, pid 8734
  |  + Compiling test program (/tmp/test_program.Cz3yL)...
  |  + Recording (PID=8749)...
  |  + Stopping perf-record...
  | test_program.Cz
  |                  728 leaf
  |                  753 parent
  |                  76c main
  | test child finished with 0
  | ---- end ----
  | Check Arm SPE callgraphs are complete in fp mode: Ok

It's supposed to work with both unwinders:

  | $ make                # for libunwind (default)
  | $ make NO_LIBUNWIND=1 # for libdw

Tester notes:

Ran it on N1SDP and it passes, and it fails if b9f6fbb3b2c29736 ("perf
arm64: Inject missing frames when using 'perf record --call-graph=fp'")
isn't applied.

Fixes: b9f6fbb3b2c29736 ("perf arm64: Inject missing frames when using 'perf record --call-graph=fp'")
Suggested-by: Jiri Olsa 
Reviewed-by: James Clark 
Tested-by: James Clark 
Signed-off-by: German Gomez 
Cc: Alexandre Truong 
Cc: Leo Yan 
Cc: Mark Rutland 
Cc: Namhyung Kim 
Link: https://lore.kernel.org/r/20220316172015.98000-1-german.gomez@arm.com
Signed-off-by: Arnaldo Carvalho de Melo 
Signed-off-by: Sasha Levin

perf parse-events: Ignore case in topdown.slots check

2022-03-18T21:39:09+00:00

An issue with icelakex metrics:

  https://git.kernel.org/pub/scm/linux/kernel/git/acme/linux.git/tree/tools/perf/pmu-events/arch/x86/icelakex/icx-metrics.json?h=perf/core&id=65eab2bc7dab326ee892ec5a4c749470b368b51a#n48

That causes the slots not to be first.

Fixes: 94dbfd6781a0e87b ("perf parse-events: Architecture specific leader override")
Reported-by: Caleb Biggers 
Signed-off-by: Ian Rogers 
Cc: Alexander Shishkin 
Cc: Alexandre Torgue 
Cc: Andi Kleen 
Cc: Ingo Molnar 
Cc: James Clark 
Cc: Jiri Olsa 
Cc: John Garry 
Cc: Kan Liang 
Cc: Mark Rutland 
Cc: Maxime Coquelin 
Cc: Namhyung Kim 
Cc: Peter Zijlstra 
Cc: Stephane Eranian 
Cc: Zhengjun Xing 
Link: https://lore.kernel.org/r/20220317224309.543736-1-irogers@google.com
Signed-off-by: Arnaldo Carvalho de Melo

perf evlist: Avoid iteration for empty evlist.

2022-03-18T21:39:09+00:00

As seen with 'perf stat --null ..' and reported in:
https://lore.kernel.org/lkml/YjCLcpcX2peeQVCH@kernel.org/

v2. Avoids setting evsel in the empty list case as suggested by Jiri Olsa.

    Committer testing:

Before:

  $  perf stat --null sleep 1
  Segmentation fault (core dumped)
  $

After:

  $  perf stat --null sleep 1

   Performance counter stats for 'sleep 1':

         1.010340646 seconds time elapsed

         0.001420000 seconds user
         0.000000000 seconds sys
  $

Fixes: 472832d2c000b961 ("perf evlist: Refactor evlist__for_each_cpu()")
Reported-by: Arnaldo Carvalho de Melo 
Signed-off-by: Ian Rogers 
Tested-by: Arnaldo Carvalho de Melo 
Acked-by: Jiri Olsa 
Cc: Adrian Hunter 
Cc: Ian Rogers 
Cc: Namhyung Kim 
Link: https://lore.kernel.org/r/20220317231643.550902-1-irogers@google.com
Signed-off-by: Arnaldo Carvalho de Melo

perf symbols: Fix symbol size calculation condition

2022-03-18T21:39:09+00:00

Before this patch, the symbol end address fixup to be called, needed two
conditions being met:

  if (prev->end == prev->start && prev->end != curr->start)

Where
  "prev->end == prev->start" means that prev is zero-long
                             (and thus needs a fixup)
and
  "prev->end != curr->start" means that fixup hasn't been applied yet

However, this logic is incorrect in the following situation:

*curr  = {rb_node = {__rb_parent_color = 278218928,
  rb_right = 0x0, rb_left = 0x0},
  start = 0xc000000000062354,
  end = 0xc000000000062354, namelen = 40, type = 2 '\002',
  binding = 0 '\000', idle = 0 '\000', ignore = 0 '\000',
  inlined = 0 '\000', arch_sym = 0 '\000', annotate2 = false,
  name = 0x1159739e "kprobe_optinsn_page\t[__builtin__kprobes]"}

*prev = {rb_node = {__rb_parent_color = 278219041,
  rb_right = 0x109548b0, rb_left = 0x109547c0},
  start = 0xc000000000062354,
  end = 0xc000000000062354, namelen = 12, type = 2 '\002',
  binding = 1 '\001', idle = 0 '\000', ignore = 0 '\000',
  inlined = 0 '\000', arch_sym = 0 '\000', annotate2 = false,
  name = 0x1095486e "optinsn_slot"}

In this case, prev->start == prev->end == curr->start == curr->end,
thus the condition above thinks that "we need a fixup due to zero
length of prev symbol, but it has been probably done, since the
prev->end == curr->start", which is wrong.

After the patch, the execution path proceeds to arch__symbols__fixup_end
function which fixes up the size of prev symbol by adding page_size to
its end offset.

Fixes: 3b01a413c196c910 ("perf symbols: Improve kallsyms symbol end addr calculation")
Signed-off-by: Michael Petlan 
Cc: Athira Jajeev 
Cc: Jiri Olsa 
Cc: Kajol Jain 
Cc: Madhavan Srinivasan 
Link: http://lore.kernel.org/lkml/20220317135536.805-1-mpetlan@redhat.com
Signed-off-by: Arnaldo Carvalho de Melo

perf parse: Fix event parser error for hybrid systems

2022-03-12T14:04:55+00:00

This bug happened on hybrid systems when both cpu_core and cpu_atom
have the same event name such as "UOPS_RETIRED.MS" while their event
terms are different, then during perf stat, the event for cpu_atom
will parse fail and then no output for cpu_atom.

UOPS_RETIRED.MS -> cpu_core/period=0x1e8483,umask=0x4,event=0xc2,frontend=0x8/
UOPS_RETIRED.MS -> cpu_atom/period=0x1e8483,umask=0x1,event=0xc2/

It is because event terms in the "head" of parse_events_multi_pmu_add
will be changed to event terms for cpu_core after parsing UOPS_RETIRED.MS
for cpu_core, then when parsing the same event for cpu_atom, it still
uses the event terms for cpu_core, but event terms for cpu_atom are
different with cpu_core, the event parses for cpu_atom will fail. This
patch fixes it, the event terms should be parsed from the original
event.

This patch can work for the hybrid systems that have the same event
in more than 2 PMUs. It also can work in non-hybrid systems.

Before:

  # perf stat -v  -e  UOPS_RETIRED.MS  -a sleep 1

  Using CPUID GenuineIntel-6-97-1
  UOPS_RETIRED.MS -> cpu_core/period=0x1e8483,umask=0x4,event=0xc2,frontend=0x8/
  Control descriptor is not initialized
  UOPS_RETIRED.MS: 2737845 16068518485 16068518485

 Performance counter stats for 'system wide':

         2,737,845      cpu_core/UOPS_RETIRED.MS/

       1.002553850 seconds time elapsed

After:

  # perf stat -v  -e  UOPS_RETIRED.MS  -a sleep 1

  Using CPUID GenuineIntel-6-97-1
  UOPS_RETIRED.MS -> cpu_core/period=0x1e8483,umask=0x4,event=0xc2,frontend=0x8/
  UOPS_RETIRED.MS -> cpu_atom/period=0x1e8483,umask=0x1,event=0xc2/
  Control descriptor is not initialized
  UOPS_RETIRED.MS: 1977555 16076950711 16076950711
  UOPS_RETIRED.MS: 568684 8038694234 8038694234

 Performance counter stats for 'system wide':

         1,977,555      cpu_core/UOPS_RETIRED.MS/
           568,684      cpu_atom/UOPS_RETIRED.MS/

       1.004758259 seconds time elapsed

Fixes: fb0811535e92c6c1 ("perf parse-events: Allow config on kernel PMU events")
Reviewed-by: Kan Liang 
Signed-off-by: Zhengjun Xing 
Cc: Adrian Hunter 
Cc: Alexander Shishkin 
Cc: Andi Kleen 
Cc: Ian Rogers 
Cc: Ingo Molnar 
Cc: Jiri Olsa 
Cc: Peter Zijlstra 
Link: https://lore.kernel.org/r/20220307151627.30049-1-zhengjun.xing@linux.intel.com
Signed-off-by: Arnaldo Carvalho de Melo

perf bench: Fix NULL check against wrong variable

2022-03-12T13:49:13+00:00

We did a NULL check after "epollfdp = calloc(...)", but we checked
"epollfd" instead of "epollfdp".

Signed-off-by: Weiguo Li 
Acked-by: Davidlohr Bueso 
Cc: Alexander Shishkin 
Cc: Davidlohr Bueso 
Cc: Ingo Molnar 
Cc: Jiri Olsa 
Cc: Mark Rutland 
Cc: Namhyung Kim 
Cc: Peter Zijlstra 
Link: https://lore.kernel.org/r/tencent_B5D64530EB9C7DBB8D2C88A0C790F1489D0A@qq.com
Signed-off-by: Arnaldo Carvalho de Melo

perf parse-events: Fix NULL check against wrong variable

2022-03-12T13:47:52+00:00

We did a null check after "tmp->symbol = strdup(...)", but we checked
"list->symbol" other than "tmp->symbol".

Reviewed-by: John Garry 
Signed-off-by: Weiguo Li 
Cc: Alexander Shishkin 
Cc: Ingo Molnar 
Cc: Jiri Olsa 
Cc: Mark Rutland 
Cc: Namhyung Kim 
Cc: Peter Zijlstra 
Link: https://lore.kernel.org/r/tencent_DF39269807EC9425E24787E6DB632441A405@qq.com
Signed-off-by: Arnaldo Carvalho de Melo