Skip to content

L
linux
Switch branch/tag
  1. 06 Jan, 2016 8 commits
    • Andi Kleen's avatar
      perf/x86: Use INST_RETIRED.PREC_DIST for cycles: ppp · 72469764
      Andi Kleen authored 
      Add a new 'three-p' precise level, that uses INST_RETIRED.PREC_DIST as
base. The basic mechanism of abusing the inverse cmask to get all
cycles works the same as before.

PREC_DIST is available on Sandy Bridge or later. It had some problems
on Sandy Bridge, so we only use it on IvyBridge and later. I tested it
on Broadwell and Skylake.

PREC_DIST has special support for avoiding shadow effects, which can
give better results compare to UOPS_RETIRED. The drawback is that
PREC_DIST can only schedule on counter 1, but that is ok for cycle
sampling, as there is normally no need to do multiple cycle sampling
runs in parallel. It is still possible to run perf top in parallel, as
that doesn't use precise mode. Also of course the multiplexing can
still allow parallel operation.

:pp stays with the previous event.

Example:

Sample a loop with 10 sqrt with old cycles:pp

	  0.14 │10:   sqrtps %xmm1,%xmm0     <--------------
	  9.13 │      sqrtps %xmm1,%xmm0
	 11.58 │      sqrtps %xmm1,%xmm0
	 11.51 │      sqrtps %xmm1,%xmm0
	  6.27 │      sqrtps %xmm1,%xmm0
	 10.38 │      sqrtps %xmm1,%xmm0
	 12.20 │      sqrtps %xmm1,%xmm0
	 12.74 │      sqrtps %xmm1,%xmm0
	  5.40 │      sqrtps %xmm1,%xmm0
	 10.14 │      sqrtps %xmm1,%xmm0
	 10.51 │    ↑ jmp    10

We expect all 10 sqrt to get roughly the sample number of samples.

But you can see that the instruction directly after the JMP is
systematically underestimated in the result, due to sampling shadow
effects.

With the new PREC_DIST based sampling this problem is gone and all
instructions show up roughly evenly:

	  9.51 │10:   sqrtps %xmm1,%xmm0
	 11.74 │      sqrtps %xmm1,%xmm0
	 11.84 │      sqrtps %xmm1,%xmm0
	  6.05 │      sqrtps %xmm1,%xmm0
	 10.46 │      sqrtps %xmm1,%xmm0
	 12.25 │      sqrtps %xmm1,%xmm0
	 12.18 │      sqrtps %xmm1,%xmm0
	  5.26 │      sqrtps %xmm1,%xmm0
	 10.13 │      sqrtps %xmm1,%xmm0
	 10.43 │      sqrtps %xmm1,%xmm0
	  0.16 │    ↑ jmp    10

Even with PREC_DIST there is still sampling skid and the result is not
completely even, but systematic shadow effects are significantly
reduced.

The improvements are mainly expected to make a difference in high IPC
code. With low IPC it should be similar.
Signed-off-by: default avatarAndi Kleen <ak@linux.intel.com>
Signed-off-by: default avatarPeter Zijlstra (Intel) <peterz@infradead.org>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Jiri Olsa <jolsa@redhat.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Stephane Eranian <eranian@google.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Vince Weaver <vincent.weaver@maine.edu>
Cc: hpa@zytor.com
Link: http://lkml.kernel.org/r/1448929689-13771-2-git-send-email-andi@firstfloor.orgSigned-off-by: default avatarIngo Molnar <mingo@kernel.org>
      72469764
    • Andi Kleen's avatar
      perf/x86: Use INST_RETIRED.TOTAL_CYCLES_PS for cycles:pp for Skylake · 442f5c74
      Andi Kleen authored 
      I added UOPS_RETIRED.ALL by mistake to the Skylake PEBS event list for
cycles:pp. But the event is not documented for Skylake, and has some
issues.

The recommended replacement for cycles:pp is to use
INST_RETIRED.ANY+pebs as a base, similar to what CPUs before Sandy
Bridge did. This new event is called INST_RETIRED.TOTAL_CYCLES_PS. The
event is not really new, but has been already used by perf before
Sandy Bridge for the original cycles:p

Note the SDM doesn't document that event either, but it's being
documented in the latest version of the event list on:

  https://download.01.org/perfmon/SKL

This patch does:

 - Remove UOPS_RETIRED.ALL from the Skylake PEBS event list

 - Add INST_RETIRED.ANY to the Skylake PEBS event list, and an table entry to
   allow cmask=16,inv=1 for cycles:pp

 - We don't need an extra entry for the base INST_RETIRED event,
   because it is already covered by the catch-all PEBS table entry.

 - Switch Skylake to use the Core2 PEBS alias (which is
   INST_RETIRED.TOTAL_CYCLES_PS)
Signed-off-by: default avatarAndi Kleen <ak@linux.intel.com>
Signed-off-by: default avatarPeter Zijlstra (Intel) <peterz@infradead.org>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Jiri Olsa <jolsa@redhat.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Stephane Eranian <eranian@google.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Vince Weaver <vincent.weaver@maine.edu>
Cc: hpa@zytor.com
Link: http://lkml.kernel.org/r/1448929689-13771-1-git-send-email-andi@firstfloor.orgSigned-off-by: default avatarIngo Molnar <mingo@kernel.org>
      442f5c74
    • Andi Kleen's avatar
      perf/x86: Allow zero PEBS status with only single active event · 01330d72
      Andi Kleen authored 
      Normally we drop PEBS events with a zero status field. But when
there is only a single PEBS event active we can assume the
PEBS record is for that event. The PEBS buffer is always flushed
when PEBS events are disabled, so there is no risk of mishandling
state PEBS records this way.
Signed-off-by: default avatarAndi Kleen <ak@linux.intel.com>
Signed-off-by: default avatarPeter Zijlstra (Intel) <peterz@infradead.org>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Jiri Olsa <jolsa@redhat.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Stephane Eranian <eranian@google.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Vince Weaver <vincent.weaver@maine.edu>
Link: http://lkml.kernel.org/r/1449177740-5422-2-git-send-email-andi@firstfloor.orgSigned-off-by: default avatarIngo Molnar <mingo@kernel.org>
      01330d72
    • Andi Kleen's avatar
      perf/x86: Remove warning for zero PEBS status · 957ea1fd
      Andi Kleen authored 
      The recent commit:

  75f80859 ("perf/x86/intel/pebs: Robustify PEBS buffer drain")

causes lots of warnings on different CPUs before Skylake
when running PEBS intensive workloads.

They can have a zero status field in the PEBS record when
PEBS is racing with clearing of GLOBAl_STATUS.

This also can cause hangs (it seems there are still
problems with printk in NMI).

Disable the warning, but still ignore the record.
Signed-off-by: default avatarAndi Kleen <ak@linux.intel.com>
Signed-off-by: default avatarPeter Zijlstra (Intel) <peterz@infradead.org>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Jiri Olsa <jolsa@redhat.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Stephane Eranian <eranian@google.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Vince Weaver <vincent.weaver@maine.edu>
Link: http://lkml.kernel.org/r/1449177740-5422-1-git-send-email-andi@firstfloor.orgSigned-off-by: default avatarIngo Molnar <mingo@kernel.org>
      957ea1fd
    • Peter Zijlstra's avatar
      perf/core: Collapse more IPI loops · 7b648018
      Peter Zijlstra authored 
      This patch collapses the two 'hard' cases, which are
perf_event_{dis,en}able().

I cannot seem to convince myself the current code is correct.

So starting with perf_event_disable(); we don't strictly need to test
for event->state == ACTIVE, ctx->is_active is enough. If the event is
not scheduled while the ctx is, __perf_event_disable() still does the
right thing.  Its a little less efficient to IPI in that case,
over-all simpler.

For perf_event_enable(); the same goes, but I think that's actually
broken in its current form. The current condition is: ctx->is_active
&& event->state == OFF, that means it doesn't do anything when
!ctx->active && event->state == OFF. This is wrong, it should still
mark the event INACTIVE in that case, otherwise we'll still not try
and schedule the event once the context becomes active again.

This patch implements the two function using the new
event_function_call() and does away with the tricky event->state
tests.
Signed-off-by: default avatarPeter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: default avatarAlexander Shishkin <alexander.shishkin@intel.com>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Jiri Olsa <jolsa@redhat.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Stephane Eranian <eranian@google.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Vince Weaver <vincent.weaver@maine.edu>
Signed-off-by: default avatarIngo Molnar <mingo@kernel.org>
      7b648018
    • Ingo Molnar's avatar
      Merge branch 'perf/urgent' into perf/core, to pick up fixes before applying new changes · 9cc96b0a
      Ingo Molnar authored 
      Signed-off-by: default avatarIngo Molnar <mingo@kernel.org>
      9cc96b0a
    • Peter Zijlstra's avatar
      perf: Fix race in swevent hash · 12ca6ad2
      Peter Zijlstra authored 
      There's a race on CPU unplug where we free the swevent hash array
while it can still have events on. This will result in a
use-after-free which is BAD.

Simply do not free the hash array on unplug. This leaves the thing
around and no use-after-free takes place.

When the last swevent dies, we do a for_each_possible_cpu() iteration
anyway to clean these up, at which time we'll free it, so no leakage
will occur.
Reported-by: default avatarSasha Levin <sasha.levin@oracle.com>
Tested-by: default avatarSasha Levin <sasha.levin@oracle.com>
Signed-off-by: default avatarPeter Zijlstra (Intel) <peterz@infradead.org>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: Jiri Olsa <jolsa@redhat.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Stephane Eranian <eranian@google.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Vince Weaver <vincent.weaver@maine.edu>
Signed-off-by: default avatarIngo Molnar <mingo@kernel.org>
      12ca6ad2
    • Peter Zijlstra's avatar
      perf: Fix race in perf_event_exec() · c1274499
      Peter Zijlstra authored 
      I managed to tickle this warning:

  [ 2338.884942] ------------[ cut here ]------------
  [ 2338.890112] WARNING: CPU: 13 PID: 35162 at ../kernel/events/core.c:2702 task_ctx_sched_out+0x6b/0x80()
  [ 2338.900504] Modules linked in:
  [ 2338.903933] CPU: 13 PID: 35162 Comm: bash Not tainted 4.4.0-rc4-dirty #244
  [ 2338.911610] Hardware name: Intel Corporation S2600GZ/S2600GZ, BIOS SE5C600.86B.02.02.0002.122320131210 12/23/2013
  [ 2338.923071]  ffffffff81f1468e ffff8807c6457cb8 ffffffff815c680c 0000000000000000
  [ 2338.931382]  ffff8807c6457cf0 ffffffff810c8a56 ffffe8ffff8c1bd0 ffff8808132ed400
  [ 2338.939678]  0000000000000286 ffff880813170380 ffff8808132ed400 ffff8807c6457d00
  [ 2338.947987] Call Trace:
  [ 2338.950726]  [<ffffffff815c680c>] dump_stack+0x4e/0x82
  [ 2338.956474]  [<ffffffff810c8a56>] warn_slowpath_common+0x86/0xc0
  [ 2338.963195]  [<ffffffff810c8b4a>] warn_slowpath_null+0x1a/0x20
  [ 2338.969720]  [<ffffffff811a49cb>] task_ctx_sched_out+0x6b/0x80
  [ 2338.976244]  [<ffffffff811a62d2>] perf_event_exec+0xe2/0x180
  [ 2338.982575]  [<ffffffff8121fb6f>] setup_new_exec+0x6f/0x1b0
  [ 2338.988810]  [<ffffffff8126de83>] load_elf_binary+0x393/0x1660
  [ 2338.995339]  [<ffffffff811dc772>] ? get_user_pages+0x52/0x60
  [ 2339.001669]  [<ffffffff8121e297>] search_binary_handler+0x97/0x200
  [ 2339.008581]  [<ffffffff8121f8b3>] do_execveat_common.isra.33+0x543/0x6e0
  [ 2339.016072]  [<ffffffff8121fcea>] SyS_execve+0x3a/0x50
  [ 2339.021819]  [<ffffffff819fc165>] stub_execve+0x5/0x5
  [ 2339.027469]  [<ffffffff819fbeb2>] ? entry_SYSCALL_64_fastpath+0x12/0x71
  [ 2339.034860] ---[ end trace ee1337c59a0ddeac ]---

Which is a WARN_ON_ONCE() indicating that cpuctx->task_ctx is not
what we expected it to be.

This is because context switches can swap the task_struct::perf_event_ctxp[]
pointer around. Therefore you have to either disable preemption when looking
at current, or hold ctx->lock.

Fix perf_event_enable_on_exec(), it loads current->perf_event_ctxp[]
before disabling interrupts, therefore a preemption in the right place
can swap contexts around and we're using the wrong one.
Signed-off-by: default avatarPeter Zijlstra (Intel) <peterz@infradead.org>
Cc: Alexander Potapenko <glider@google.com>
Cc: Arnaldo Carvalho de Melo <acme@kernel.org>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Dmitry Vyukov <dvyukov@google.com>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Jiri Olsa <jolsa@redhat.com>
Cc: Kostya Serebryany <kcc@google.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Sasha Levin <sasha.levin@oracle.com>
Cc: Stephane Eranian <eranian@google.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Vince Weaver <vincent.weaver@maine.edu>
Cc: syzkaller <syzkaller@googlegroups.com>
Link: http://lkml.kernel.org/r/20151210195740.GG6357@twins.programming.kicks-ass.netSigned-off-by: default avatarIngo Molnar <mingo@kernel.org>
      c1274499
  3. 18 Dec, 2015 4 commits
    • Ingo Molnar's avatar
      Merge tag 'perf-core-for-mingo-3' of... · d64fe8e6
      Ingo Molnar authored 
      Merge tag 'perf-core-for-mingo-3' of git://git.kernel.org/pub/scm/linux/kernel/git/acme/linux into perf/core

Pull new perf tool feature from Arnaldo Carvalho de Melo:

" User visible changes:

  - Generate perf.data files from 'perf stat', to tap into the scripting
    capabilities perf has instead of defining a 'perf stat' specific scripting
    support to calculate event ratios, etc. Simple example:

    $ perf stat record -e cycles usleep 1

     Performance counter stats for 'usleep 1':

           1,134,996      cycles

         0.000670644 seconds time elapsed

    $ perf stat report

     Performance counter stats for '/home/acme/bin/perf stat record -e cycles usleep 1':

           1,134,996      cycles

         0.000670644 seconds time elapsed

    $

    It generates PERF_RECORD_ userspace records to store the details:

    $ perf report -D | grep PERF_RECORD
    0xf0 [0x28]: PERF_RECORD_THREAD_MAP nr: 1 thread: 27637
    0x118 [0x12]: PERF_RECORD_CPU_MAP nr: 1 cpu: 65535
    0x12a [0x40]: PERF_RECORD_STAT_CONFIG
    0x16a [0x30]: PERF_RECORD_STAT
    -1 -1 0x19a [0x40]: PERF_RECORD_MMAP -1/0: [0xffffffff81000000(0x1f000000) @ 0xffffffff81000000]: x [kernel.kallsyms]_text
    0x1da [0x18]: PERF_RECORD_STAT_ROUND
    [acme@ssdandy linux]$

    An effort was made to make perf.data files generated like this to not
    generate cryptic messages when processed by older tools.

    The 'perf script' bits need rebasing, will go up later.

  Jiri's cover letter for this series:

  The initial attempt defined its own formula lang and allowed triggering user's
  script on the end of the stat command:

    http://marc.info/?l=linux-kernel&m=136742146322273&w=2

  This patchset abandons the idea of new formula language and rather adds support
  to:

    - store stat data into perf.data file
    - add python support to process stat events

  Basically it allows to store stat data into perf.data and post process it with
  python scripts in a similar way we do for sampling data.

  The stat data are stored in new stat, stat-round, stat-config user events.
    stat        - stored for each read syscall of the counter
    stat round  - stored for each interval or end of the command invocation
    stat config - stores all the config information needed to process data
                  so report tool could restore the same output as record

  The python script can now define 'stat__<eventname>_<modifier>' functions
  to get stat events data and 'stat__interval' to get stat-round data.

  See CPI script example in scripts/python/stat-cpi.py."

Also a few other changes:

User visible changes:

  - Make command line options always available, even when they
    depend on some feature being enabled, warning the user about
    use of such options (Wang Nan)

  - Support --vmlinux in perf record, useful, so far, for eBPF,
    where we will set up events that will be used in the record
    session (He Kuang)

  - Automatically disable collecting branch flags and cycles with
    --call-graph lbr. This allows avoiding a bunch of extra MSR
    reads in the PMI on Skylake.  (Andi Kleen)

Infrastructure changes:

  - Dump the stack when a 'perf test -v ' entry segfaults, so far we
    would have to run it under gdb with 'set follow-fork-mode child'
    set to get a proper backtrace (Arnaldo Carvalho de Melo)

  - Initialize the refcnt in 'struct thread' to 1 and fixup its
    users accordingly, so that we try to have the same refcount
    model accross the perf codebase (Arnaldo Carvalho de Melo)

  - More prep work for moving the subcmd infrastructure out of
    tools/perf/ and into tools/lib/subcmd/ to be used by other
    tools/ living utilities (Josh Poimboeuf)

  - Fix 'perf test' hist testcases when kptr_restrict is on (Namhyung Kim)
Signed-off-by: default avatarArnaldo Carvalho de Melo <acme@redhat.com>
Signed-off-by: default avatarIngo Molnar <mingo@kernel.org>
      d64fe8e6
    • Ingo Molnar's avatar
      Merge branch 'perf/urgent' into perf/core, to make sure a cherry-picked commit... · 141a361e
      Ingo Molnar authored 
      Merge branch 'perf/urgent' into perf/core, to make sure a cherry-picked commit does not create conflicts
Signed-off-by: default avatarIngo Molnar <mingo@kernel.org>
      141a361e
    • Ingo Molnar's avatar
      Merge tag 'perf-urgent-for-mingo' of... · 2d2e7ac1
      Ingo Molnar authored 
      Merge tag 'perf-urgent-for-mingo' of git://git.kernel.org/pub/scm/linux/kernel/git/acme/linux into perf/urgent

Pull perf/urgent tooling fix from Arnaldo Carvalho de Melo:

  User visible changes:

    - Fix 'perf list' segfault due to lack of support for PERF_CONF_SW_BPF_OUTPUT
      in an array used just for printing available events, robustify the code
      involved (Arnaldo Carvalho de Melo)
Signed-off-by: default avatarArnaldo Carvalho de Melo <acme@redhat.com>
Signed-off-by: default avatarIngo Molnar <mingo@kernel.org>
      2d2e7ac1
    • Ingo Molnar's avatar
      Merge tag 'perf-core-for-mingo-2.1' of... · b21daaed
      Ingo Molnar authored 
      Merge tag 'perf-core-for-mingo-2.1' of git://git.kernel.org/pub/scm/linux/kernel/git/acme/linux into perf/core

Pull perf/core improvements from Arnaldo Carvalho de Melo:

User visible changes:

  - Add record.build-id config option to 'perf record', to allow configuring
    in the ~/.perfconfig file if and how build-ids should be processed, allowing
    a permanent setting for options such as -B and -N: (Namhyung Kim)

    $ perf record -h -B -N

     Usage: perf record [<options>] [<command>]
        or: perf record [<options>] -- <command> [<options>]

        -B, --no-buildid       do not collect buildids in perf.data
        -N, --no-buildid-cache do not update the buildid cache

    $

Infrastructure changes:

  - Move code for options parsing and subcommand handling from tools/perf/
    to tools/lib/subcmd/, so that it can be used by other tools/ living
    utilities (Josh Poimboeuf)
Signed-off-by: default avatarArnaldo Carvalho de Melo <acme@redhat.com>
Signed-off-by: default avatarIngo Molnar <mingo@kernel.org>
      b21daaed
  5. 17 Dec, 2015 28 commits
GitLab Nexedi Edition | About GitLab | About Nexedi | 沪ICP备2021021310号-2 | 沪ICP备2021021310号-7