perf_counter/x86: Always use NMI for performance-monitoring interrupt

Always use NMI for performance-monitoring interrupt as there could be
racy situations if we switch between irq and nmi mode frequently.

Signed-off-by: Yong Wang <yong.y.wang@intel.com>
LKML-Reference: <20090529052835.GA13657@ywang-moblin2.bj.intel.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

perf_counter tools: Fix top symbol table max_ip typo

Signed-off-by: Mike Galbraith <efault@gmx.de>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
LKML-Reference: <new-submission>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

perf_counter tools: Fix top symbol table dump typo

Signed-off-by: Mike Galbraith <efault@gmx.de>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
LKML-Reference: <new-submission>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

perf_counter tools: Document '--' option parsing terminator

Signed-off-by: Mike Galbraith <efault@gmx.de>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
LKML-Reference: <new-submission>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

perf_counter tools: Convert builtin-top to use libperf symbol routines

Now both perf top and report use the same routines.

Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
LKML-Reference: <20090528175541.GG4747@ghostprotocols.net>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

perf_counter tools: Optionally pass a symbol filter to the dso load routines

Will be used by perf top.

Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
LKML-Reference: <20090528175526.GF4747@ghostprotocols.net>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

perf_counter tools: Consolidate dso methods to load kernel symbols

Now one has just to use dso__load_kernel() optionally passing a vmlinux
filename.

Will make things easier for perf top that will want to pass a callback
to filter some symbols.

Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

perf_counter tools: struct symbol priv area

When creating a dso instance allow asking that all symbols in this dso
have a private area just before the symbol.

perf top will use this for its counters, etc.

Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
LKML-Reference: <20090528175513.GD4747@ghostprotocols.net>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

perf_counter tools: Move symbol resolution classes from report to libperf

Will be used by perf top as well.

Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
LKML-Reference: <20090528175504.GC4747@ghostprotocols.net>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

perf_counter: Fix race in attaching counters to tasks and exiting

Commit 564c2b21 ("perf_counter: Optimize context switch between
identical inherited contexts") introduced a race where it is possible
that a counter being attached to a task could get attached to the
wrong task, if the task is one that has inherited its context from
another task via fork.  This happens because the optimized context
switch could switch the context to another task after find_get_context
has read task->perf_counter_ctxp.  In fact, it's possible that the
context could then get freed, if the other task then exits.

This fixes the problem by protecting both the context switch and the
critical code in find_get_context with spinlocks.  The context switch
locks the cxt->lock of both the outgoing and incoming contexts before
swapping them.  That means that once code such as find_get_context
has obtained the spinlock for the context associated with a task,
the context can't get swapped to another task.  However, the context
may have been swapped in the interval between reading
task->perf_counter_ctxp and getting the lock, so it is necessary to
check and retry.

To make sure that none of the contexts being looked at in
find_get_context can get freed, this changes the context freeing code
to use RCU.  Thus an rcu_read_lock() is sufficient to ensure that no
contexts can get freed.  This part of the patch is lifted from a patch
posted by Peter Zijlstra.

This also adds a check to make sure that we can't add a counter to a
task that is exiting.

There is also a race between perf_counter_exit_task and
find_get_context; this solves the race by moving the get_ctx that
was in perf_counter_alloc into the locked region in find_get_context,
so that once find_get_context has got the context for a task, it
won't get freed even if the task calls perf_counter_exit_task.  It
doesn't matter if new top-level (non-inherited) counters get attached
to the context after perf_counter_exit_task has detached the context
from the task.  They will just stay there and never get scheduled in
until the counters' fds get closed, and then perf_release will remove
them from the context and eventually free the context.

With this, we are now doing the unclone in find_get_context rather
than when a counter was added to or removed from a context (actually,
we were missing the unclone_ctx() call when adding a counter to a
context).  We don't need to unclone when removing a counter from a
context because we have no way to remove a counter from a cloned
context.

This also takes out the smp_wmb() in find_get_context, which Peter
Zijlstra pointed out was unnecessary because the cmpxchg implies a
full barrier anyway.

Signed-off-by: Paul Mackerras <paulus@samba.org>
Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Corey Ashford <cjashfor@linux.vnet.ibm.com>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Marcelo Tosatti <mtosatti@redhat.com>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: John Kacur <jkacur@redhat.com>
LKML-Reference: <18974.33033.667187.273886@cargo.ozlabs.ibm.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

perf_counter tools: report: Add help text for --sort

Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Corey Ashford <cjashfor@linux.vnet.ibm.com>
Cc: Marcelo Tosatti <mtosatti@redhat.com>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: John Kacur <jkacur@redhat.com>
LKML-Reference: <new-submission>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

perf_counter tools: report: Implement header output for --sort variants

Implement this style of header:

 #
 # Overhead          Command       File: Symbol
 # ........          .......       ............
 #

for the various --sort variants as well.

Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Corey Ashford <cjashfor@linux.vnet.ibm.com>
Cc: Marcelo Tosatti <mtosatti@redhat.com>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: John Kacur <jkacur@redhat.com>
LKML-Reference: <new-submission>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

perf_counter: Fix perf_counter_init_task() on !CONFIG_PERF_COUNTERS

Pointed out by compiler warnings:

   tip/include/linux/perf_counter.h:644: warning: no return statement in function returning non-void

Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Corey Ashford <cjashfor@linux.vnet.ibm.com>
Cc: Marcelo Tosatti <mtosatti@redhat.com>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: John Kacur <jkacur@redhat.com>
LKML-Reference: <new-submission>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

pref_counter: tools: report: Robustify in case of weird events

This error condition:

  aldebaran:~/linux/linux/Documentation/perf_counter> perf report
  dso__load_sym: cannot get elf header.
  failed to open: /etc/ld.so.cache
  problem processing PERF_EVENT_MMAP, bailing out

caused the profile to be very short - as the error was at the beginning
of the file and we bailed out completely.

Be more permissive and consider the event broken instead.

Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Corey Ashford <cjashfor@linux.vnet.ibm.com>
Cc: Marcelo Tosatti <mtosatti@redhat.com>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: John Kacur <jkacur@redhat.com>
LKML-Reference: <new-submission>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

pref_counter: tools: report: Add header printout & prettify

Old default output:

 3.12%    perf-report [.] ./perf-report:       dsos__find
 2.44%    perf-report [k] kernel:              kallsyms_expand_symbol
 2.28%          :4483 [.] <unknown>:           <unknown>
 2.05%          :4174 [k] kernel:              _spin_lock_irqsave
 2.01%    perf-report [k] kernel:              vsnprintf
 1.92%    perf-report [k] kernel:              format_decode
 1.92%          :4438 [k] kernel:              _spin_lock

New default output:

 #
 # Overhead          Command       File: Symbol
 # ........          .......       ............
 #
      6.54%             perf  [k]  kernel: kallsyms_expand_symbol
      6.26%             perf  [.]  /home/mingo/tip/Documentation/perf_counter/perf: dso__insert_symbol
      4.76%             perf  [.]  /home/mingo/tip/Documentation/perf_counter/perf: hex2long
      4.55%             perf  [k]  kernel: number
      4.48%             perf  [k]  kernel: format_decode
      4.09%             perf  [k]  kernel: vsnprintf

Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Corey Ashford <cjashfor@linux.vnet.ibm.com>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: John Kacur <jkacur@redhat.com>
Cc: Mike Galbraith <efault@gmx.de>
LKML-Reference: <20090527182101.229504802@chello.nl>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

pref_counter: tools: report: Add dso sorting

Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Corey Ashford <cjashfor@linux.vnet.ibm.com>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: John Kacur <jkacur@redhat.com>
Cc: Mike Galbraith <efault@gmx.de>
LKML-Reference: <20090527182101.229504802@chello.nl>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

perf_counter: tools: report: Add comm sorting

Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Corey Ashford <cjashfor@linux.vnet.ibm.com>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: John Kacur <jkacur@redhat.com>
Cc: Mike Galbraith <efault@gmx.de>
LKML-Reference: <20090527182101.129302022@chello.nl>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

pref_counter: tools: report: Add --sort option

option parsing for dynamic sorting.

Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Corey Ashford <cjashfor@linux.vnet.ibm.com>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: John Kacur <jkacur@redhat.com>
Cc: Mike Galbraith <efault@gmx.de>
LKML-Reference: <20090527182101.041817692@chello.nl>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

perf_counter: tools: report: Dynamic sort/print bits

Make the sorting and printing dynamic.

Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Corey Ashford <cjashfor@linux.vnet.ibm.com>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: John Kacur <jkacur@redhat.com>
Cc: Mike Galbraith <efault@gmx.de>
LKML-Reference: <20090527182100.921953817@chello.nl>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

perf_counter: tools: report: Rework histogram code

In preparation for configurable sorting, rework the histgram code a bit.

Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Corey Ashford <cjashfor@linux.vnet.ibm.com>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: John Kacur <jkacur@redhat.com>
Cc: Mike Galbraith <efault@gmx.de>
LKML-Reference: <20090527182100.796410098@chello.nl>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

perf_counter: tools: report: Add vmlinux support

Allow to use vmlinux instead of kallsyms.

Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Corey Ashford <cjashfor@linux.vnet.ibm.com>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: John Kacur <jkacur@redhat.com>
Cc: Mike Galbraith <efault@gmx.de>
LKML-Reference: <20090527182100.740018486@chello.nl>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

perf_counter: tools: /usr/lib/debug%s.debug support

Some distros seem to store debuginfo in weird places.

Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Corey Ashford <cjashfor@linux.vnet.ibm.com>
Cc: Marcelo Tosatti <mtosatti@redhat.com>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: John Kacur <jkacur@redhat.com>
LKML-Reference: <new-submission>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

perf report: Remove <ctype.h> include

Pekka reported build failure in builtin-report.c:

    CC builtin-report.o
    In file included from builtin-report.c:7:
    /usr/include/ctype.h:102: error: expected expression before token

And observed:

| Removing #include <ctype.h> from builtin-report.c makes the problem
| go away. I am running Ubuntu 9.04 that has gcc 4.3.3 and libc 2.9.

Reported-by: Pekka J Enberg <penberg@cs.helsinki.fi>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Corey Ashford <cjashfor@linux.vnet.ibm.com>
Cc: Marcelo Tosatti <mtosatti@redhat.com>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: John Kacur <jkacur@redhat.com>
LKML-Reference: <new-submission>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

perf record: Fix the profiling of existing pid or whole box

Perf record bails if no command argument is provided, so you can't use
naked -a or -p to profile a running task or the whole box.

Allow foreground profiling of an existing pid or the entire system.

[ Impact: fix command option handling bug ]

Signed-off-by: Mike Galbraith <efault@gmx.de>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Corey Ashford <cjashfor@linux.vnet.ibm.com>
Cc: Marcelo Tosatti <mtosatti@redhat.com>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: John Kacur <jkacur@redhat.com>
LKML-Reference: <new-submission>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

perf_counter tools: Add built-in pager support

Add Git's pager.c (and sigchain) code. A command only
has to call setup_pager() to get paged interactive
output.

Non-interactive (redirected, command-piped, etc.) uses
are not affected.

Update perf-report to make use of this.

[ Impact: new feature ]

Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Corey Ashford <cjashfor@linux.vnet.ibm.com>
Cc: Marcelo Tosatti <mtosatti@redhat.com>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: John Kacur <jkacur@redhat.com>
LKML-Reference: <new-submission>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

perf_counter tools: Rename output.perf to perf.data

output.perf is only output to perf-record - it's input to
perf-report. So change it to a more direction-neutral name.

Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Corey Ashford <cjashfor@linux.vnet.ibm.com>
Cc: Marcelo Tosatti <mtosatti@redhat.com>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: John Kacur <jkacur@redhat.com>
LKML-Reference: <new-submission>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

perf_counter tools: Introduce stricter C code checking

Tighten up our C code requirements:

 - disallow warnings
 - disallow declarations-mixed-with-statements
 - require proper prototypes
 - require C99 (with gcc extensions)

Fix up a ton of problems these measures unearth:

 - unused functions
 - needlessly global functions
 - missing prototypes
 - code mixed with declarations

Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Corey Ashford <cjashfor@linux.vnet.ibm.com>
Cc: Marcelo Tosatti <mtosatti@redhat.com>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: John Kacur <jkacur@redhat.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
LKML-Reference: <20090526222155.GJ4424@ghostprotocols.net>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

perf report: Show the IP only in --verbose mode

perf: report should show the IP only in --verbose mode

[acme@emilia ~]$ perf report | head
 4.95          find [k] _spin_lock
 2.19          find [k] ext3fs_dirhash [ext3]
 1.87          find [k] __rcu_read_lock
 1.86          find [k] _atomic_dec_and_lock
 1.86          find [.] /lib64/libc-2.5.so: __GI_strlen
 1.85          find [k] __kmalloc
 1.62          find [.] /lib64/libc-2.5.so: vfprintf
 1.59          find [k] __rcu_read_unlock
 1.55          find [k] __d_lookup

Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Corey Ashford <cjashfor@linux.vnet.ibm.com>
Cc: Marcelo Tosatti <mtosatti@redhat.com>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: John Kacur <jkacur@redhat.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
LKML-Reference: <20090526224614.GK4424@ghostprotocols.net>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

perf report: Only load text symbols from kallsyms, fix

- allow 'W' symbols too
 - Convert initializations to C99 style
 - whitespace cleanups

Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Corey Ashford <cjashfor@linux.vnet.ibm.com>
Cc: Marcelo Tosatti <mtosatti@redhat.com>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: John Kacur <jkacur@redhat.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
LKML-Reference: <20090526222155.GJ4424@ghostprotocols.net>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

perf report: Only load text symbols from kallsyms

Just like we do for userspace when reading the symtab, reducing the
number of entries we insert on the symbols rbtree.

Before:

[acme@emilia ~]$ rm -f perf_report.perf ; perf record -o perf_report.perf perf stat perf report > /dev/null

 Performance counter stats for 'perf':

     218.138382  task clock ticks     (msecs)
              4  context switches     (events)
              8  CPU migrations       (events)
           2136  pagefaults           (events)
       32746212  CPU cycles           (events)  (scaled from 67.04%)
       11961102  instructions         (events)  (scaled from 66.19%)
          49841  cache references     (events)  (scaled from 21.96%)
          13777  cache misses         (events)  (scaled from 21.98%)

 Wall-clock time elapsed:   218.702477 msecs

[acme@emilia ~]$ perf report -i perf_report.perf | head
11.06          perf [.] 0x00000000000057cb /home/acme/git/linux-2.6-tip/Documentation/perf_counter/perf: dso__find_symbol
 9.15          perf [.] 0x00000000000056a0 /home/acme/git/linux-2.6-tip/Documentation/perf_counter/perf: dso__insert_symbol
 8.72          perf [k] 0xffffffff8101b1d2 intel_pmu_enable_all
 8.51          perf [.] 0x0000000000006672 /home/acme/git/linux-2.6-tip/Documentation/perf_counter/perf: thread__symbol_incnew
 3.83          perf [k] 0xffffffff811cfc5a vsnprintf
 3.40          perf [.] 0x0000000000005e33 /home/acme/git/linux-2.6-tip/Documentation/perf_counter/perf: hex
 3.40          perf [.] 0x0000000000005ec7 /home/acme/git/linux-2.6-tip/Documentation/perf_counter/perf: hex2long
 3.19          perf [k] 0xffffffff811ce1c1 number
 2.77          perf [.] 0x0000000000006869 /home/acme/git/linux-2.6-tip/Documentation/perf_counter/perf: threads__findnew
 2.77          perf [.] 0x000000000000fde3 /home/acme/git/linux-2.6-tip/Documentation/perf_counter/perf: rb_insert_color
[acme@emilia ~]$

After:

acme@emilia ~]$ rm -f perf_report.perf ; perf record -o perf_report.perf perf stat perf report > /dev/null

 Performance counter stats for 'perf':

     190.228511  task clock ticks     (msecs)
              4  context switches     (events)
              7  CPU migrations       (events)
           1625  pagefaults           (events)
       29578745  CPU cycles           (events)  (scaled from 66.92%)
       10516914  instructions         (events)  (scaled from 66.47%)
          44015  cache references     (events)  (scaled from 22.04%)
           8248  cache misses         (events)  (scaled from 22.07%)

 Wall-clock time elapsed:   190.816096 msecs

[acme@emilia ~]$ perf report -i perf_report.perf | head
15.99          perf [.] 0x00000000000057a9 /home/acme/git/linux-2.6-tip/Documentation/perf_counter/perf: dso__find_symbol
10.87          perf [.] 0x000000000000674d /home/acme/git/linux-2.6-tip/Documentation/perf_counter/perf: thread__symbol_incnew
 8.74          perf [k] 0xffffffff8101b1d2 intel_pmu_enable_all
 5.54          perf [.] 0x0000000000005e42 /home/acme/git/linux-2.6-tip/Documentation/perf_counter/perf: hex
 4.48          perf [.] 0x0000000000005ebe /home/acme/git/linux-2.6-tip/Documentation/perf_counter/perf: hex2long
 4.48          perf [k] 0xffffffff811cfba0 vsnprintf
 3.84          perf [.] 0x00000000000056b4 /home/acme/git/linux-2.6-tip/Documentation/perf_counter/perf: dso__insert_symbol
 3.62          perf [.] 0x00000000000068d0 /home/acme/git/linux-2.6-tip/Documentation/perf_counter/perf: threads__findnew
 3.20          perf [k] 0xffffffff811ce0b3 number
 2.56          perf [.] 0x0000000000006d78 /home/acme/git/linux-2.6-tip/Documentation/perf_counter/perf: __cmd_report
[acme@emilia ~]$

[ Impact: optimization ]

Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Corey Ashford <cjashfor@linux.vnet.ibm.com>
Cc: Marcelo Tosatti <mtosatti@redhat.com>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: John Kacur <jkacur@redhat.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
LKML-Reference: <20090526222155.GJ4424@ghostprotocols.net>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

perf report: Use hex2long instead of sscanf

Before:

[acme@emilia ~]$ perf record -o perf_report.perf perf stat perf report > /dev/null

 Performance counter stats for 'perf':

     245.414985  task clock ticks     (msecs)
              6  context switches     (events)
              6  CPU migrations       (events)
           2108  pagefaults           (events)
       37493013  CPU cycles           (events)  (scaled from 67.04%)
       13576789  instructions         (events)  (scaled from 66.76%)
          57931  cache references     (events)  (scaled from 21.96%)
          12263  cache misses         (events)  (scaled from 21.98%)

 Wall-clock time elapsed:   246.575587 msecs

[acme@emilia ~]$ perf report -i perf_report.perf | head
12.15          perf [.] 0x000000000005432a /lib64/libc-2.5.so: _IO_vfscanf_internal
 9.38          perf [k] 0xffffffff8101b1d2 intel_pmu_enable_all
 8.53          perf [.] 0x00000000000056b8 /home/acme/git/linux-2.6-tip/Documentation/perf_counter/perf: dso__insert_symbol
 6.61          perf [.] 0x00000000000057cb /home/acme/git/linux-2.6-tip/Documentation/perf_counter/perf: dso__find_symbol
 5.33          perf [k] 0xffffffff811ce082 number
 4.69          perf [.] 0x0000000000034829 /lib64/libc-2.5.so: ____strtoull_l_internal
 4.48          perf [.] 0x0000000000006505 /home/acme/git/linux-2.6-tip/Documentation/perf_counter/perf: thread__symbol_incnew
 3.41          perf [.] 0x000000000000fce6 /home/acme/git/linux-2.6-tip/Documentation/perf_counter/perf: rb_insert_color
 3.20          perf [k] 0xffffffff811cfc01 vsnprintf
 2.99          perf [k] 0xffffffff811ce5e8 format_decode

After:

[acme@emilia ~]$ perf record -o perf_report.perf perf stat perf report > /dev/null

 Performance counter stats for 'perf':

     218.186805  task clock ticks     (msecs)
              4  context switches     (events)
              7  CPU migrations       (events)
           2133  pagefaults           (events)
       32735365  CPU cycles           (events)  (scaled from 67.04%)
       11952309  instructions         (events)  (scaled from 66.26%)
          50314  cache references     (events)  (scaled from 21.96%)
          13228  cache misses         (events)  (scaled from 21.98%)

 Wall-clock time elapsed:   218.810451 msecs

[acme@emilia ~]$ perf report -i perf_report.perf | head
10.68          perf [.] 0x000000000000578d /home/acme/git/linux-2.6-tip/Documentation/perf_counter/perf: dso__find_symbol
 9.62          perf [.] 0x00000000000065f7 /home/acme/git/linux-2.6-tip/Documentation/perf_counter/perf: thread__symbol_incnew
 9.40          perf [.] 0x00000000000056b4 /home/acme/git/linux-2.6-tip/Documentation/perf_counter/perf: dso__insert_symbol
 9.19          perf [k] 0xffffffff8101b1d2 intel_pmu_enable_all
 5.13          perf [.] 0x0000000000005ec7 /home/acme/git/linux-2.6-tip/Documentation/perf_counter/perf: hex2long
 4.49          perf [k] 0xffffffff81083808 kallsyms_expand_symbol
 3.85          perf [k] 0xffffffff811ce2c1 number
 3.63          perf [.] 0x0000000000005e81 /home/acme/git/linux-2.6-tip/Documentation/perf_counter/perf: hex
 2.99          perf [.] 0x000000000000fd5b /home/acme/git/linux-2.6-tip/Documentation/perf_counter/perf: rb_insert_color
 2.99          perf [k] 0xffffffff811cf251 string
[acme@emilia ~]$

[ Impact: optimization ]

Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Corey Ashford <cjashfor@linux.vnet.ibm.com>
Cc: Marcelo Tosatti <mtosatti@redhat.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: John Kacur <jkacur@redhat.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
LKML-Reference: <20090526222057.GI4424@ghostprotocols.net>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

perf report: Sort output by symbol usage

[acme@emilia ~]$ perf record find / > /dev/null 2>&1
[acme@emilia ~]$ perf stat perf report | head -20
 4.95          find [k] 0xffffffff81393d65 _spin_lock
 3.89          find [.] 0x000000000000af89 /usr/bin/find: <unknown>
 2.19          find [k] 0xffffffffa00518e0 ext3fs_dirhash
 1.87          find [k] 0xffffffff810a6cea __rcu_read_lock
 1.86          find [k] 0xffffffff811c7312 _atomic_dec_and_lock
 1.86          find [.] 0x00000000000782ab /lib64/libc-2.5.so: __GI_strlen
 1.85          find [k] 0xffffffff810fedfb __kmalloc
 1.62          find [.] 0x00000000000430ff /lib64/libc-2.5.so: vfprintf
 1.59          find [k] 0xffffffff810a6d6d __rcu_read_unlock
 1.55          find [k] 0xffffffff81119395 __d_lookup
 1.39          find [.] 0x0000000000071b40 /lib64/libc-2.5.so: _int_malloc
 1.30          find [k] 0xffffffffa031c4fc nfs_do_filldir
 1.21          find [k] 0xffffffff811876a5 avc_has_perm_noaudit
 1.15          find [k] 0xffffffff810fef62 kmem_cache_alloc
 1.07          find [k] 0xffffffff811d03fb copy_user_generic_string
 1.03          find [k] 0xffffffffa0043882 ext3_htree_store_dirent
 0.99          find [k] 0xffffffff81393ebb _spin_lock_bh
 0.98          find [k] 0xffffffffa03319a2 nfs3_decode_dirent
 0.97          find [k] 0xffffffff8100bf20 system_call
 0.92          find [k] 0xffffffff8139437e _spin_unlock

 Performance counter stats for 'perf':

     244.278972  task clock ticks     (msecs)
              8  context switches     (events)
              9  CPU migrations       (events)
           2104  pagefaults           (events)
       35329669  CPU cycles           (events)  (scaled from 75.40%)
       13740366  instructions         (events)  (scaled from 75.49%)
          59073  cache references     (events)  (scaled from 24.60%)
            196  cache misses         (events)  (scaled from 24.51%)

 Wall-clock time elapsed:   246.060717 msecs

[acme@emilia ~]$
[acme@emilia ~]$ grep "model name" /proc/cpuinfo | head -1
model name : Intel(R) Xeon(R) CPU           E5405  @ 2.00GHz
[acme@emilia ~]$ grep "model name" /proc/cpuinfo | wc -l
8
[acme@emilia ~]$

Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
LKML-Reference: <20090526191904.GH4424@ghostprotocols.net>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

perf report: More robust error handling

Don't let funny events confuse us, stick to what we know and
try to find sensible data again.

If we find an unknown event, check we're still u64 aligned, and
increment by one u64. This ensures we're bound to happen upon a
valid event soon.

Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Corey Ashford <cjashfor@linux.vnet.ibm.com>
Cc: Marcelo Tosatti <mtosatti@redhat.com>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: John Kacur <jkacur@redhat.com>
LKML-Reference: <new-submission>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

perf report: add more debugging

Add the offset of the file we are analyzing, and the size of the record.

In case of problems it's easier to see where the parser lost track.

Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Corey Ashford <cjashfor@linux.vnet.ibm.com>
Cc: Marcelo Tosatti <mtosatti@redhat.com>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: John Kacur <jkacur@redhat.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

perf report: add counter for unknown events

Add a counter for unknown event records.

[ Impact: improve debugging ]

Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Corey Ashford <cjashfor@linux.vnet.ibm.com>
Cc: Marcelo Tosatti <mtosatti@redhat.com>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: John Kacur <jkacur@redhat.com>
LKML-Reference: <new-submission>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

perf report: add --dump-raw-trace option

To help the inspection of various data files, implement an ASCII dump
method that just dumps the records as they are read in - then we exit.

[ Impact: new feature ]

Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Corey Ashford <cjashfor@linux.vnet.ibm.com>
Cc: Marcelo Tosatti <mtosatti@redhat.com>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: John Kacur <jkacur@redhat.com>
LKML-Reference: <new-submission>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

perf: Don't assume /proc/kallsyms is ordered

perf: Don't assume /proc/kallsyms is ordered

Since we _are_ ordering it by the symbol start, just traverse the
freshly built rbtree setting the prev->end members to curr->start - 1.

Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Corey Ashford <cjashfor@linux.vnet.ibm.com>
Cc: Marcelo Tosatti <mtosatti@redhat.com>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: John Kacur <jkacur@redhat.com>
LKML-Reference: <20090526152134.GF4424@ghostprotocols.net>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

perf report: Fix kernel symbol resolution

kallsyms have just the symbol start, so we need to read two lines
to get the len.

[ Impact: fix incorrect kernel symbol display in perf report ]

Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Corey Ashford <cjashfor@linux.vnet.ibm.com>
Cc: Marcelo Tosatti <mtosatti@redhat.com>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: John Kacur <jkacur@redhat.com>
LKML-Reference: <new-submission>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

perf report: Fix ELF symbol parsing

[ Impact: fix DSO symbol output in perf report ]

Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Corey Ashford <cjashfor@linux.vnet.ibm.com>
Cc: Marcelo Tosatti <mtosatti@redhat.com>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: John Kacur <jkacur@redhat.com>
LKML-Reference: <new-submission>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

perf top: fix typo in -d option

Clean up copy/paste options parsing conversion error.

[ Impact: reactivate -d option ]

Signed-off-by: Mike Galbraith <efault@gmx.de>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Corey Ashford <cjashfor@linux.vnet.ibm.com>
Cc: Marcelo Tosatti <mtosatti@redhat.com>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: John Kacur <jkacur@redhat.com>
LKML-Reference: <new-submission>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

perf top: Remove leftover NMI/IRQ bits

79202b removed IRQ/NMI mode selection, so remove it from
perf top as well.

[ Impact: cleanup ]

Signed-off-by: Mike Galbraith <efault@gmx.de>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Corey Ashford <cjashfor@linux.vnet.ibm.com>
Cc: Marcelo Tosatti <mtosatti@redhat.com>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: John Kacur <jkacur@redhat.com>
LKML-Reference: <new-submission>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

perf report: Add help/manpage

Add a (minimal) manpage for perf report.

Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Corey Ashford <cjashfor@linux.vnet.ibm.com>
Cc: Marcelo Tosatti <mtosatti@redhat.com>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: John Kacur <jkacur@redhat.com>
LKML-Reference: <new-submission>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

perf record: Convert to Git option parsing

Remove getopt usage and use Git's much more advanced and more compact
command option library.

Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Corey Ashford <cjashfor@linux.vnet.ibm.com>
Cc: Marcelo Tosatti <mtosatti@redhat.com>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: John Kacur <jkacur@redhat.com>
LKML-Reference: <new-submission>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

perf report: Fix segfault on unknown symbols

Ingo reported:

> Program received signal SIGSEGV, Segmentation fault.
> 0x0000003e25080f80 in strlen () from /lib64/libc.so.6
> Missing separate debuginfos, use: debuginfo-install elfutils.x86_64
> glibc.x86_64 zlib.x86_64
> (gdb) bt
> #0  0x0000003e25080f80 in strlen () from /lib64/libc.so.6
> #1  0x0000003e2506954e in fputs () from /lib64/libc.so.6
> #2  0x00000000004059e8 in cmd_report (argc=<value optimized out>,
>     argv=<value optimized out>) at builtin-report.c:521
> #3  0x0000000000402dad in handle_internal_command (argc=1, argv=0x7fffe1218e30)
>     at perf.c:226
> #4  0x0000000000402f6d in main (argc=1, argv=0x7fffe1218e30) at perf.c:324
> (gdb)

Signed-off-by Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Corey Ashford <cjashfor@linux.vnet.ibm.com>
Cc: Marcelo Tosatti <mtosatti@redhat.com>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

perf_counter: Use rb_tree for symhists and threads in report

Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Corey Ashford <cjashfor@linux.vnet.ibm.com>
Cc: Marcelo Tosatti <mtosatti@redhat.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
LKML-Reference: <new-submission>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

perf_counter: Add our private copy of list.h

Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Corey Ashford <cjashfor@linux.vnet.ibm.com>
Cc: Marcelo Tosatti <mtosatti@redhat.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
LKML-Reference: <new-submission>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

perf_counter: Use rb_trees in perf report

Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Corey Ashford <cjashfor@linux.vnet.ibm.com>
Cc: Marcelo Tosatti <mtosatti@redhat.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
LKML-Reference: <new-submission>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

perf_counter: Implement dso__load using libelf

Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Corey Ashford <cjashfor@linux.vnet.ibm.com>
Cc: Marcelo Tosatti <mtosatti@redhat.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
LKML-Reference: <new-submission>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

perf_counter tools: remove the standalone perf-report utility

With a built-in 'perf report' command now available, remove the
standalone implementation for good.

Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Corey Ashford <cjashfor@linux.vnet.ibm.com>
Cc: Marcelo Tosatti <mtosatti@redhat.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
LKML-Reference: <new-submission>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

perf_counter: First part of 'perf report' conversion to C + elfutils

Integrate perf-report into 'perf', as builtin-report.c.

Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Corey Ashford <cjashfor@linux.vnet.ibm.com>
Cc: Marcelo Tosatti <mtosatti@redhat.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
LKML-Reference: <new-submission>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

perf top: Convert to Git option parsing

Remove getopt usage and use Git's much more advanced and more compact
command option library.

Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Corey Ashford <cjashfor@linux.vnet.ibm.com>
Cc: Marcelo Tosatti <mtosatti@redhat.com>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: John Kacur <jkacur@redhat.com>
LKML-Reference: <new-submission>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

perf stat: Convert to Git option parsing

Remove getopt usage and use Git's much more advanced and more compact
command option library.

Extend the event parser library with the extensions that were in
perf-stat before.

Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Corey Ashford <cjashfor@linux.vnet.ibm.com>
Cc: Marcelo Tosatti <mtosatti@redhat.com>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: John Kacur <jkacur@redhat.com>
LKML-Reference: <new-submission>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

perf_counter tools: Librarize event string parsing

Extract the event string parser from builtin-record.c, and
librarize it - to be reused in other commands.

Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Corey Ashford <cjashfor@linux.vnet.ibm.com>
Cc: Marcelo Tosatti <mtosatti@redhat.com>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: John Kacur <jkacur@redhat.com>
LKML-Reference: <new-submission>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

perf record: Convert to Git option parsing

Remove getopt usage and use Git's much more advanced and more compact
command option library.

Git's library (util/parse-options.[ch]) constructs help texts and
error messages automatically, and has a number of other convenience
features as well.

Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Corey Ashford <cjashfor@linux.vnet.ibm.com>
Cc: Marcelo Tosatti <mtosatti@redhat.com>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: John Kacur <jkacur@redhat.com>
LKML-Reference: <new-submission>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

perf stat: Remove unused variable

[ Impact: cleanup ]

Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Corey Ashford <cjashfor@linux.vnet.ibm.com>
Cc: Marcelo Tosatti <mtosatti@redhat.com>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: John Kacur <jkacur@redhat.com>
LKML-Reference: <new-submission>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

perf record: Straighten out argv types

[ Impact: cleanup ]

Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Corey Ashford <cjashfor@linux.vnet.ibm.com>
Cc: Marcelo Tosatti <mtosatti@redhat.com>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: John Kacur <jkacur@redhat.com>
LKML-Reference: <new-submission>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

perf_counter: Initialize ->oncpu properly

This shouldnt matter normally (and i have not seen any
misbehavior), because active counters always have a
proper ->oncpu value - but nevertheless initialize the
field properly to -1.

[ Impact: cleanup ]

Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Corey Ashford <cjashfor@linux.vnet.ibm.com>
Cc: Marcelo Tosatti <mtosatti@redhat.com>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: John Kacur <jkacur@redhat.com>
LKML-Reference: <new-submission>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

perf_counter, x86: Make NMI lockups more robust

We have a debug check that detects stuck NMIs and returns with
the PMU disabled in the global ctrl MSR - but i managed to trigger
a situation where this was not enough to deassert the NMI.

So clear/reset the full PMU and keep the disable count balanced when
exiting from here. This way the box produces a debug warning but
stays up and is more debuggable.

[ Impact: in case of PMU related bugs, recover more gracefully ]

Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Corey Ashford <cjashfor@linux.vnet.ibm.com>
Cc: Marcelo Tosatti <mtosatti@redhat.com>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: John Kacur <jkacur@redhat.com>
LKML-Reference: <new-submission>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

perf_counter, x86: Fix APIC NMI programming

My Nehalem box locks up in certain situations (with an
always-asserted NMI causing a lockup) if the PMU LVT
entry is programmed between NMI and IRQ mode with a
high frequency.

Standardize exlusively on NMIs instead.

[ Impact: fix lockup ]

Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Corey Ashford <cjashfor@linux.vnet.ibm.com>
Cc: Marcelo Tosatti <mtosatti@redhat.com>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: John Kacur <jkacur@redhat.com>
LKML-Reference: <new-submission>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

perf_counter: powerpc: Implement interrupt throttling

This implements interrupt throttling on powerpc.  Since we don't have
individual count enable/disable or interrupt enable/disable controls
per counter, this simply sets the hardware counter to 0, meaning that
it will not interrupt again until it has counted 2^31 counts, which
will take at least 2^30 cycles assuming a maximum of 2 counts per
cycle.  Also, we set counter->hw.period_left to the maximum possible
value (2^63 - 1), so we won't report overflows for this counter for
the forseeable future.

The unthrottle operation restores counter->hw.period_left and the
hardware counter so that we will once again report a counter overflow
after counter->hw.irq_period counts.

[ Impact: new perfcounters robustness feature on PowerPC ]

Signed-off-by: Paul Mackerras <paulus@samba.org>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Corey Ashford <cjashfor@linux.vnet.ibm.com>
LKML-Reference: <18971.35823.643362.446774@cargo.ozlabs.ibm.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

perf_counter: fix warning & lockup

 - remove bogus warning
 - fix wakeup from NMI path lockup
 - also fix up whitespace noise in perf_counter.h

Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Corey Ashford <cjashfor@linux.vnet.ibm.com>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: John Kacur <jkacur@redhat.com>
LKML-Reference: <20090525153931.703093461@chello.nl>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

Revert "perf_counter, x86: speed up the scheduling fast-path"

This reverts commit b68f1d2e7aa21029d73c7d453a8046e95d351740.

It is causing problems (stuck/stuttering profiling) - when mixed
NMI and non-NMI counters are used.

Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Corey Ashford <cjashfor@linux.vnet.ibm.com>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: John Kacur <jkacur@redhat.com>
LKML-Reference: <20090525153931.703093461@chello.nl>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

perf_counter: Generic per counter interrupt throttle

Introduce a generic per counter interrupt throttle.

This uses the perf_counter_overflow() quick disable to throttle a specific
counter when its going too fast when a pmu->unthrottle() method is provided
which can undo the quick disable.

Power needs to implement both the quick disable and the unthrottle method.

Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Corey Ashford <cjashfor@linux.vnet.ibm.com>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: John Kacur <jkacur@redhat.com>
LKML-Reference: <20090525153931.703093461@chello.nl>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

perf_counter: x86: Remove interrupt throttle

remove the x86 specific interrupt throttle

Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Corey Ashford <cjashfor@linux.vnet.ibm.com>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: John Kacur <jkacur@redhat.com>
LKML-Reference: <20090525153931.616671838@chello.nl>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

perf_counter: x86: Expose INV and EDGE bits

Expose the INV and EDGE bits of the PMU to raw configs.

Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Corey Ashford <cjashfor@linux.vnet.ibm.com>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: John Kacur <jkacur@redhat.com>
LKML-Reference: <20090525153931.494709027@chello.nl>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

perf_counter: Fix PERF_COUNTER_CONTEXT_SWITCHES for cpu counters

Ingo noticed that cpu counters had 0 context switches, even though
there was plenty scheduling on the cpu.

Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Corey Ashford <cjashfor@linux.vnet.ibm.com>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: John Kacur <jkacur@redhat.com>
LKML-Reference: <20090525124600.419025548@chello.nl>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

perf_counter: Propagate inheritance failures down the fork() path

Fail fork() when we fail inheritance for some reason (-ENOMEM most likely).

Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Corey Ashford <cjashfor@linux.vnet.ibm.com>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: John Kacur <jkacur@redhat.com>
LKML-Reference: <20090525124600.324656474@chello.nl>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

perf_counter: Make pctrl() affect inherited counters too

Paul noted that the new ptcrl() didn't work on child counters.

Reported-by: Paul Mackerras <paulus@samba.org>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Corey Ashford <cjashfor@linux.vnet.ibm.com>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: John Kacur <jkacur@redhat.com>
LKML-Reference: <20090525124600.203151469@chello.nl>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

perf_counter: Remove unused ABI bits

extra_config_len isn't used for anything, remove it.

Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Corey Ashford <cjashfor@linux.vnet.ibm.com>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: John Kacur <jkacur@redhat.com>
LKML-Reference: <20090525124600.116035832@chello.nl>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

perf_counter: Fix perf-$cmd invokation

Fix:

  $ perf-top
  fatal: cannot handle -top internally

Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Corey Ashford <cjashfor@linux.vnet.ibm.com>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: John Kacur <jkacur@redhat.com>
LKML-Reference: <20090525124559.995591577@chello.nl>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

perf stat: flip around ':k' and ':u' flags

This output:

 $ perf stat -e 0:1:k -e 0:1:u ./hello
  Performance counter stats for './hello':
          140131  instructions         (events)
         1906968  instructions         (events)

Is quite confusing - as :k means "user instructions", :u means
"kernel instructions".

Flip them around - as the 'exclude' property is not intuitive in
the flag naming.

Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Corey Ashford <cjashfor@linux.vnet.ibm.com>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: John Kacur <jkacur@redhat.com>
Cc: Mike Galbraith <efault@gmx.de>
LKML-Reference: <new-submission>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

perf_counter: Move child perfcounter init to after scheduler init

Initialize a task's perfcounters (inherit from parent, etc.) after
the child task's scheduler fields have been initialized already.

[ Impact: cleanup ]

Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Corey Ashford <cjashfor@linux.vnet.ibm.com>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: John Kacur <jkacur@redhat.com>
Cc: Mike Galbraith <efault@gmx.de>
LKML-Reference: <new-submission>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

perf top: Reduce display overhead

Iterate over the symbol table once per display interval, and
copy/sort/tally/decay only those symbols which are active.

Before:

 top - 10:14:53 up  4:08, 17 users,  load average: 1.17, 1.53, 1.49
 Tasks: 273 total,   5 running, 268 sleeping,   0 stopped,   0 zombie
 Cpu(s):  6.9%us, 38.2%sy,  0.0%ni, 19.9%id,  0.0%wa,  0.0%hi, 35.0%si,  0.0%st

   PID USER      PR  NI  VIRT  RES  SHR S %CPU %MEM    TIME+  P COMMAND
 28504 root      20   0  1044  260  164 S   58  0.0   0:04.19 2 netserver
 28499 root      20   0  1040  412  316 R   51  0.0   0:04.15 0 netperf
 28500 root      20   0  1040  408  316 R   50  0.0   0:04.14 1 netperf
 28503 root      20   0  1044  260  164 S   50  0.0   0:04.01 1 netserver
 28501 root      20   0  1044  260  164 S   49  0.0   0:03.99 0 netserver
 28502 root      20   0  1040  412  316 S   43  0.0   0:03.96 2 netperf
 28468 root      20   0 1892m 325m  972 S   16 10.8   0:10.50 3 perf
 28467 root      20   0 1892m 325m  972 R    2 10.8   0:00.72 3 perf

After:

 top - 10:16:30 up  4:10, 17 users,  load average: 2.27, 1.88, 1.62
 Tasks: 273 total,   6 running, 267 sleeping,   0 stopped,   0 zombie
 Cpu(s):  2.5%us, 39.7%sy,  0.0%ni, 24.6%id,  0.0%wa,  0.0%hi, 33.3%si,  0.0%st

   PID USER      PR  NI  VIRT  RES  SHR S %CPU %MEM    TIME+  P COMMAND
 28590 root      20   0  1040  412  316 S   54  0.0   0:07.85 2 netperf
 28589 root      20   0  1044  260  164 R   54  0.0   0:07.84 0 netserver
 28588 root      20   0  1040  412  316 R   50  0.0   0:07.89 1 netperf
 28591 root      20   0  1044  256  164 S   50  0.0   0:07.82 1 netserver
 28587 root      20   0  1040  408  316 R   47  0.0   0:07.61 0 netperf
 28592 root      20   0  1044  260  164 R   47  0.0   0:07.85 2 netserver
 28378 root      20   0  8732 1300  860 R    2  0.0   0:01.81 3 top
 28577 root      20   0 1892m 165m  972 R    2  5.5   0:00.48 3 perf
 28578 root      20   0 1892m 165m  972 S    2  5.5   0:00.04 3 perf

[ Impact: optimization ]

Signed-off-by: Mike Galbraith <efault@gmx.de>
Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
LKML-Reference: <new-submission>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

perf_counter tools: increase limits, fix

NR_CPUS and NR_COUNTERS goes up quadratic ... 1024x4096 was far
too ambitious upper limit - go for 256x256 which is still plenty.

[ Impact: reduce perf tool memory consumption ]

Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Corey Ashford <cjashfor@linux.vnet.ibm.com>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: John Kacur <jkacur@redhat.com>
Cc: Mike Galbraith <efault@gmx.de>
LKML-Reference: <new-submission>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

perf_counter: Increase mmap limit

In a default 'perf top' run the tool will create a counter for
each online CPU. With enough CPUs this will eventually exhaust
the default limit.

So scale it up with the number of online CPUs.

Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Corey Ashford <cjashfor@linux.vnet.ibm.com>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: John Kacur <jkacur@redhat.com>
Cc: Mike Galbraith <efault@gmx.de>
LKML-Reference: <new-submission>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

perf top: fix segfault

c6eb13 increased stack usage such that perf-top now croaks on startup.

Take event_array and mmap_array off the stack to prevent segfault on boxen
with smallish ulimit -s setting.

Signed-off-by: Mike Galbraith <efault@gmx.de>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Corey Ashford <cjashfor@linux.vnet.ibm.com>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: John Kacur <jkacur@redhat.com>
LKML-Reference: <new-submission>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

perf_counter: Remove perf_counter_context::nr_enabled

now that pctrl() no longer disables other people's counters,
remove the PMU cache code that deals with that.

Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Corey Ashford <cjashfor@linux.vnet.ibm.com>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: John Kacur <jkacur@redhat.com>
LKML-Reference: <20090523163013.032998331@chello.nl>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

perf_counter: Change pctrl() behaviour

Instead of en/dis-abling all counters acting on a particular
task, en/dis- able all counters we created.

[ v2: fix crash on first counter enable ]

Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Corey Ashford <cjashfor@linux.vnet.ibm.com>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: John Kacur <jkacur@redhat.com>
LKML-Reference: <20090523163012.916937244@chello.nl>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

perf_counter: Simplify context cleanup

Use perf_counter_remove_from_context() to remove counters from
the context.

Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Corey Ashford <cjashfor@linux.vnet.ibm.com>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: John Kacur <jkacur@redhat.com>
LKML-Reference: <20090523163012.796275849@chello.nl>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

perf_counter: Fix userspace build

recent userspace (F11) seems to already include the
linux/unistd.h bits which means we cannot include the version
in the kernel sources due to the header guards being the same.

Ensure we include the kernel version first.

Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Corey Ashford <cjashfor@linux.vnet.ibm.com>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: John Kacur <jkacur@redhat.com>
LKML-Reference: <20090523163012.739756497@chello.nl>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

perf_counter: Sanitize context locking

Ensure we're consistent with the context locks.

 context->mutex
   context->lock
     list_{add,del}_counter();

so that either lock is sufficient to stabilize the context.

Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Corey Ashford <cjashfor@linux.vnet.ibm.com>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: John Kacur <jkacur@redhat.com>
LKML-Reference: <20090523163012.618790733@chello.nl>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

perf_counter: Sanitize counter->mutex

s/counter->mutex/counter->child_mutex/ and make sure its only
used to protect child_list.

The usage in __perf_counter_exit_task() doesn't appear to be
problematic since ctx->mutex also covers anything related to fd
tear-down.

Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Corey Ashford <cjashfor@linux.vnet.ibm.com>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: John Kacur <jkacur@redhat.com>
LKML-Reference: <20090523163012.533186528@chello.nl>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

perf_counter: Fix dynamic irq_period logging

We call perf_adjust_freq() from perf_counter_task_tick() which
is is called under the rq->lock causing lock recursion.
However, it's no longer required to be called under the
rq->lock, so remove it from under it.

Also, fix up some related comments.

Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Corey Ashford <cjashfor@linux.vnet.ibm.com>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: John Kacur <jkacur@redhat.com>
LKML-Reference: <20090523163012.476197912@chello.nl>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

perf_counter tools: increase limits

I tried to run with 300 active counters and the tools bailed out
because our limit was at 64. So increase the counter limit to 1024
and the CPU limit to 4096.

Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Corey Ashford <cjashfor@linux.vnet.ibm.com>
Cc: Marcelo Tosatti <mtosatti@redhat.com>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
LKML-Reference: <new-submission>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

perf_counter: fix !PERF_COUNTERS build failure

Update the !CONFIG_PERF_COUNTERS prototype too, for
perf_counter_task_sched_out().

[ Impact: build fix ]

Signed-off-by: Paul Mackerras <paulus@samba.org>
Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Corey Ashford <cjashfor@linux.vnet.ibm.com>
Cc: Marcelo Tosatti <mtosatti@redhat.com>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
LKML-Reference: <18966.10666.517218.332164@cargo.ozlabs.ibm.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

perf_counter: Optimize context switch between identical inherited contexts

When monitoring a process and its descendants with a set of inherited
counters, we can often get the situation in a context switch where
both the old (outgoing) and new (incoming) process have the same set
of counters, and their values are ultimately going to be added together.
In that situation it doesn't matter which set of counters are used to
count the activity for the new process, so there is really no need to
go through the process of reading the hardware counters and updating
the old task's counters and then setting up the PMU for the new task.

This optimizes the context switch in this situation.  Instead of
scheduling out the perf_counter_context for the old task and
scheduling in the new context, we simply transfer the old context
to the new task and keep using it without interruption.  The new
context gets transferred to the old task.  This means that both
tasks still have a valid perf_counter_context, so no special case
is introduced when the old task gets scheduled in again, either on
this CPU or another CPU.

The equivalence of contexts is detected by keeping a pointer in
each cloned context pointing to the context it was cloned from.
To cope with the situation where a context is changed by adding
or removing counters after it has been cloned, we also keep a
generation number on each context which is incremented every time
a context is changed.  When a context is cloned we take a copy
of the parent's generation number, and two cloned contexts are
equivalent only if they have the same parent and the same
generation number.  In order that the parent context pointer
remains valid (and is not reused), we increment the parent
context's reference count for each context cloned from it.

Since we don't have individual fds for the counters in a cloned
context, the only thing that can make two clones of a given parent
different after they have been cloned is enabling or disabling all
counters with prctl.  To account for this, we keep a count of the
number of enabled counters in each context.  Two contexts must have
the same number of enabled counters to be considered equivalent.

Here are some measurements of the context switch time as measured with
the lat_ctx benchmark from lmbench, comparing the times obtained with
and without this patch series:

-----Unmodified----- With this patch series
Counters: none 2 HW 4H+4S none 2 HW 4H+4S

2 processes:
Average 3.44 6.45 11.24 3.12 3.39 3.60
St dev 0.04 0.04 0.13 0.05 0.17 0.19

8 processes:
Average 6.45 8.79 14.00 5.57 6.23 7.57
St dev 1.27 1.04 0.88 1.42 1.46 1.42

32 processes:
Average 5.56 8.43 13.78 5.28 5.55 7.15
St dev 0.41 0.47 0.53 0.54 0.57 0.81

The numbers are the mean and standard deviation of 20 runs of
lat_ctx.  The "none" columns are lat_ctx run directly without any
counters.  The "2 HW" columns are with lat_ctx run under perfstat,
counting cycles and instructions.  The "4H+4S" columns are lat_ctx run
under perfstat with 4 hardware counters and 4 software counters
(cycles, instructions, cache references, cache misses, task
clock, context switch, cpu migrations, and page faults).

[ Impact: performance optimization of counter context-switches ]

Signed-off-by: Paul Mackerras <paulus@samba.org>
Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Corey Ashford <cjashfor@linux.vnet.ibm.com>
Cc: Marcelo Tosatti <mtosatti@redhat.com>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
LKML-Reference: <18966.10666.517218.332164@cargo.ozlabs.ibm.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

perf_counter: Dynamically allocate tasks' perf_counter_context struct

This replaces the struct perf_counter_context in the task_struct with
a pointer to a dynamically allocated perf_counter_context struct.  The
main reason for doing is this is to allow us to transfer a
perf_counter_context from one task to another when we do lazy PMU
switching in a later patch.

This has a few side-benefits: the task_struct becomes a little smaller,
we save some memory because only tasks that have perf_counters attached
get a perf_counter_context allocated for them, and we can remove the
inclusion of <linux/perf_counter.h> in sched.h, meaning that we don't
end up recompiling nearly everything whenever perf_counter.h changes.

The perf_counter_context structures are reference-counted and freed
when the last reference is dropped.  A context can have references
from its task and the counters on its task.  Counters can outlive the
task so it is possible that a context will be freed well after its
task has exited.

Contexts are allocated on fork if the parent had a context, or
otherwise the first time that a per-task counter is created on a task.
In the latter case, we set the context pointer in the task struct
locklessly using an atomic compare-and-exchange operation in case we
raced with some other task in creating a context for the subject task.

This also removes the task pointer from the perf_counter struct.  The
task pointer was not used anywhere and would make it harder to move a
context from one task to another.  Anything that needed to know which
task a counter was attached to was already using counter->ctx->task.

The __perf_counter_init_context function moves up in perf_counter.c
so that it can be called from find_get_context, and now initializes
the refcount, but is otherwise unchanged.

We were potentially calling list_del_counter twice: once from
__perf_counter_exit_task when the task exits and once from
__perf_counter_remove_from_context when the counter's fd gets closed.
This adds a check in list_del_counter so it doesn't do anything if
the counter has already been removed from the lists.

Since perf_counter_task_sched_in doesn't do anything if the task doesn't
have a context, and leaves cpuctx->task_ctx = NULL, this adds code to
__perf_install_in_context to set cpuctx->task_ctx if necessary, i.e. in
the case where the current task adds the first counter to itself and
thus creates a context for itself.

This also adds similar code to __perf_counter_enable to handle a
similar situation which can arise when the counters have been disabled
using prctl; that also leaves cpuctx->task_ctx = NULL.

[ Impact: refactor counter context management to prepare for new feature ]

Signed-off-by: Paul Mackerras <paulus@samba.org>
Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Corey Ashford <cjashfor@linux.vnet.ibm.com>
Cc: Marcelo Tosatti <mtosatti@redhat.com>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
LKML-Reference: <18966.10075.781053.231153@cargo.ozlabs.ibm.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

perf_counter: Fix context removal deadlock

Disable the PMU globally before removing a counter from a
context. This fixes the following lockup:

[22081.741922] ------------[ cut here ]------------
[22081.746668] WARNING: at arch/x86/kernel/cpu/perf_counter.c:803 intel_pmu_handle_irq+0x9b/0x24e()
[22081.755624] Hardware name: X8DTN
[22081.758903] perfcounters: irq loop stuck!
[22081.762985] Modules linked in:
[22081.766136] Pid: 11082, comm: perf Not tainted 2.6.30-rc6-tip #226
[22081.772432] Call Trace:
[22081.774940]  <NMI>  [<ffffffff81019aed>] ? intel_pmu_handle_irq+0x9b/0x24e
[22081.781993]  [<ffffffff81019aed>] ? intel_pmu_handle_irq+0x9b/0x24e
[22081.788368]  [<ffffffff8104505c>] ? warn_slowpath_common+0x77/0xa3
[22081.794649]  [<ffffffff810450d3>] ? warn_slowpath_fmt+0x40/0x45
[22081.800696]  [<ffffffff81019aed>] ? intel_pmu_handle_irq+0x9b/0x24e
[22081.807080]  [<ffffffff814d1a72>] ? perf_counter_nmi_handler+0x3f/0x4a
[22081.813751]  [<ffffffff814d2d09>] ? notifier_call_chain+0x58/0x86
[22081.819951]  [<ffffffff8105b250>] ? notify_die+0x2d/0x32
[22081.825392]  [<ffffffff814d1414>] ? do_nmi+0x8e/0x242
[22081.830538]  [<ffffffff814d0f0a>] ? nmi+0x1a/0x20
[22081.835342]  [<ffffffff8117e102>] ? selinux_file_free_security+0x0/0x1a
[22081.842105]  [<ffffffff81018793>] ? x86_pmu_disable_counter+0x15/0x41
[22081.848673]  <<EOE>>  [<ffffffff81018f3d>] ? x86_pmu_disable+0x86/0x103
[22081.855512]  [<ffffffff8108fedd>] ? __perf_counter_remove_from_context+0x0/0xfe
[22081.862926]  [<ffffffff8108fcbc>] ? counter_sched_out+0x30/0xce
[22081.868909]  [<ffffffff8108ff36>] ? __perf_counter_remove_from_context+0x59/0xfe
[22081.876382]  [<ffffffff8106808a>] ? smp_call_function_single+0x6c/0xe6
[22081.882955]  [<ffffffff81091b96>] ? perf_release+0x86/0x14c
[22081.888600]  [<ffffffff810c4c84>] ? __fput+0xe7/0x195
[22081.893718]  [<ffffffff810c213e>] ? filp_close+0x5b/0x62
[22081.899107]  [<ffffffff81046a70>] ? put_files_struct+0x64/0xc2
[22081.905031]  [<ffffffff8104841a>] ? do_exit+0x1e2/0x6ef
[22081.910360]  [<ffffffff814d0a60>] ? _spin_lock_irqsave+0x9/0xe
[22081.916292]  [<ffffffff8104898e>] ? do_group_exit+0x67/0x93
[22081.921953]  [<ffffffff810489cc>] ? sys_exit_group+0x12/0x16
[22081.927759]  [<ffffffff8100baab>] ? system_call_fastpath+0x16/0x1b
[22081.934076] ---[ end trace 3a3936ce3e1b4505 ]---

And could potentially also fix the lockup reported by Marcelo Tosatti.

Also, print more debug info in case of a detected lockup.

[ Impact: fix lockup ]

Reported-by: Marcelo Tosatti <mtosatti@redhat.com>
Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Corey Ashford <cjashfor@linux.vnet.ibm.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
LKML-Reference: <new-submission>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

perf_counter: Optimize sched in/out of counters

Avoid a function call for !group counters by directly calling the counter
function.

[ Impact: micro-optimize the code ]

Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Corey Ashford <cjashfor@linux.vnet.ibm.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Marcelo Tosatti <mtosatti@redhat.com>
Cc: John Kacur <jkacur@redhat.com>
LKML-Reference: <20090520102553.511933670@chello.nl>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

perf_counter: Optimize disable of time based sw counters

Currently we call hrtimer_cancel() unconditionally on disable of time based
software counters. Avoid when possible.

[ Impact: micro-optimize the code ]

Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Corey Ashford <cjashfor@linux.vnet.ibm.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Marcelo Tosatti <mtosatti@redhat.com>
Cc: John Kacur <jkacur@redhat.com>
LKML-Reference: <20090520102553.388185031@chello.nl>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

perf_counter: Log irq_period changes

For the dynamic irq_period code, log whenever we change the period so that
analyzing code can normalize the event flow.

[ Impact: add new feature to allow more precise profiling ]

Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Corey Ashford <cjashfor@linux.vnet.ibm.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Marcelo Tosatti <mtosatti@redhat.com>
Cc: John Kacur <jkacur@redhat.com>
LKML-Reference: <20090520102553.298769743@chello.nl>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

perf_counter: Solve the rotate_ctx vs inherit race differently

Instead of disabling RR scheduling of the counters, use a different list
that does not get rotated to iterate the counters on inheritance.

[ Impact: cleanup, optimization ]

Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Corey Ashford <cjashfor@linux.vnet.ibm.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Marcelo Tosatti <mtosatti@redhat.com>
Cc: John Kacur <jkacur@redhat.com>
LKML-Reference: <20090520102553.237504544@chello.nl>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

perf_counter: fix counter inheritance race

Context rotation should not occur when we are in the middle of
walking the counter list when inheriting counters ...

[ Impact: fix occasionally incorrect perf stat results ]

Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Corey Ashford <cjashfor@linux.vnet.ibm.com>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Marcelo Tosatti <mtosatti@redhat.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

perf_counter: fix counter freeing logic

Fix counter lifetime bugs which explain the crashes reported by
Marcelo Tosatti and Arnaldo Carvalho de Melo.

The new rule is: flushing + freeing is only done for a task's
own counters, never for other tasks.

[ Impact: fix crashes/lockups with inherited counters ]

Reported-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Reported-by: Marcelo Tosatti <mtosatti@redhat.com>
Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Corey Ashford <cjashfor@linux.vnet.ibm.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

perf_counter, x86: speed up the scheduling fast-path

We have to set up the LVT entry only at counter init time, not at
every switch-in time.

There's friction between NMI and non-NMI use here - we'll probably
remove the per counter configurability of it - but until then, dont
slow down things ...

[ Impact: micro-optimization ]

Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Srivatsa Vaddagiri <vatsa@in.ibm.com>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Corey Ashford <cjashfor@linux.vnet.ibm.com>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Marcelo Tosatti <mtosatti@redhat.com>
LKML-Reference: <new-submission>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

perf_counter: powerpc: initialize cpuhw pointer before use

Commit 9e35ad38 ("perf_counter: Rework the perf counter
disable/enable") added code to the powerpc hw_perf_enable (renamed
from hw_perf_restore) to test cpuhw->disabled and return immediately
if it is not set (i.e. if the PMU is already enabled).

Unfortunately the test got added before cpuhw was initialized,
resulting in an oops the first time hw_perf_enable got called.
This fixes it by moving the initialization of cpuhw to before
cpuhw->disabled is tested.

[ Impact: fix oops-causing bug on powerpc ]

Signed-off-by: Paul Mackerras <paulus@samba.org>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Corey Ashford <cjashfor@linux.vnet.ibm.com>
LKML-Reference: <18960.56772.869734.304631@drongo.ozlabs.ibm.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

Merge commit 'v2.6.30-rc6' into perfcounters/core

Merge reason: this branch was on an -rc4 base, merge it up to -rc6
              to get the latest upstream fixes.

Signed-off-by: Ingo Molnar <mingo@elte.hu>

perf_counter, x86: fix zero irq_period counters

The quirk to irq_period unearthed an unrobustness we had in the
hw_counter initialization sequence: we left irq_period at 0, which
was then quirked up to 2 ... which then generated a _lot_ of
interrupts during 'perf stat' runs, slowed them down and skewed
the counter results in general.

Initialize irq_period to the maximum instead.

[ Impact: fix perf stat results ]

Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Srivatsa Vaddagiri <vatsa@in.ibm.com>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Corey Ashford <cjashfor@linux.vnet.ibm.com>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Marcelo Tosatti <mtosatti@redhat.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

perf_counter: fix threaded task exit

Flushing counters in __exit_signal() with irqs disabled is not
a good idea as perf_counter_exit_task() acquires mutexes. So
flush it before acquiring the tasklist lock.

(Note, we still need a fix for when the PID has been unhashed.)

[ Impact: fix crash with inherited counters ]

Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Srivatsa Vaddagiri <vatsa@in.ibm.com>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Corey Ashford <cjashfor@linux.vnet.ibm.com>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Marcelo Tosatti <mtosatti@redhat.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

perf_counter: Fix counter inheritance

Srivatsa Vaddagiri reported that a Java workload triggers this
warning in kernel/exit.c:

   WARN_ON_ONCE(!list_empty(&tsk->perf_counter_ctx.counter_list));

Add the inherited counter propagation on self-detach, this could
cause counter leaks and incomplete stats in threaded code like
the below:

  #include <pthread.h>
  #include <unistd.h>

  void *thread(void *arg)
  {
          sleep(5);
          return NULL;
  }

  void main(void)
  {
          pthread_t thr;
          pthread_create(&thr, NULL, thread, NULL);
  }

Reported-by: Srivatsa Vaddagiri <vatsa@in.ibm.com>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Corey Ashford <cjashfor@linux.vnet.ibm.com>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Marcelo Tosatti <mtosatti@redhat.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>