12 Sep, 2009
2 commits
-
Conflicts:
arch/sparc/Kconfig -
…el/git/tip/linux-2.6-tip
* 'tracing-core-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/linux-2.6-tip: (105 commits)
ring-buffer: only enable ring_buffer_swap_cpu when needed
ring-buffer: check for swapped buffers in start of committing
tracing: report error in trace if we fail to swap latency buffer
tracing: add trace_array_printk for internal tracers to use
tracing: pass around ring buffer instead of tracer
tracing: make tracing_reset safe for external use
tracing: use timestamp to determine start of latency traces
tracing: Remove mentioning of legacy latency_trace file from documentation
tracing/filters: Defer pred allocation, fix memory leak
tracing: remove users of tracing_reset
tracing: disable buffers and synchronize_sched before resetting
tracing: disable update max tracer while reading trace
tracing: print out start and stop in latency traces
ring-buffer: disable all cpu buffers when one finds a problem
ring-buffer: do not count discarded events
ring-buffer: remove ring_buffer_event_discard
ring-buffer: fix ring_buffer_read crossing pages
ring-buffer: remove unnecessary cpu_relax
ring-buffer: do not swap buffers during a commit
ring-buffer: do not reset while in a commit
...
06 Sep, 2009
1 commit
-
Merge reason: move from -rc5 to -rc9.
Signed-off-by: Ingo Molnar
05 Sep, 2009
2 commits
-
This patch improves some (common) inefficiencies in the
handling of directory lookups:- not using the d_type information returned by the kernel
- constructing (absolute) paths for file operation even though
directory-relative operations using the *at functions is
possibleThere are more places to fix but this is a start.
Signed-off-by: Ulrich Drepper
Signed-off-by: Arnaldo Carvalho de Melo
Cc: Frederic Weisbecker
Cc: Peter Zijlstra
Cc: Mike Galbraith
LKML-Reference:
Signed-off-by: Ingo Molnar -
Remove some, now useless, global storage.
Don't calculate the stddev when not needed.Signed-off-by: Peter Zijlstra
LKML-Reference:
Signed-off-by: Ingo Molnar
04 Sep, 2009
5 commits
-
Use the more advanced single pass variance algorithm outlined
on the wikipedia page. This is numerically more stable for
larger sample sets.Signed-off-by: Peter Zijlstra
LKML-Reference:
Signed-off-by: Ingo Molnar -
When we're computing the mean by sampling the distribution,
then the std dev of the mean is related to the std dev of the
sample set by:stddev_mean = std_dev / sqrt(N)
Which is exactly what we want.
This results in the error on the mean decreasing with
increasing number of samples.Also fix the scaled == -1, aka not counted case.
Signed-off-by: Peter Zijlstra
LKML-Reference:
Signed-off-by: Ingo Molnar -
Since we don't need all the individual samples to calculate the
error remove both the limit and the storage overhead associated
with that.Signed-off-by: Peter Zijlstra
LKML-Reference:
Signed-off-by: Ingo Molnar -
The current noise computation does:
\Sum abs(n_i - avg(n)) * N^-1.5
Which is (afaik) not a regular noise function, and needs the
complete sample set available to post-process.Change this to use a regular stddev computation which can be
done by keeping a two sums:stddev = sqrt( 1/N (\Sum n_i^2) - avg(n)^2 )
For which we only need to keep \Sum n_i and \Sum n_i^2.
Signed-off-by: Peter Zijlstra
Cc:
LKML-Reference:
Signed-off-by: Ingo Molnar -
This wires up the perf_counter_open() syscall so that basic
software support for perf is working.Signed-off-by: Jens Axboe
Signed-off-by: David S. Miller
03 Sep, 2009
6 commits
-
We did not account for the enclosing \0. Depending on what malloc()
gave us this resulted in corrupted version string printouts.Cc: Peter Zijlstra
Cc: Mike Galbraith
Cc: Paul Mackerras
Cc: Arnaldo Carvalho de Melo
Cc: Frederic Weisbecker
LKML-Reference:
Signed-off-by: Ingo Molnar -
Print out more accurate timestamps - usecs does not cut it
anymore on fast enough boxes ;-)Cc: Peter Zijlstra
Cc: Mike Galbraith
Cc: Paul Mackerras
Cc: Arnaldo Carvalho de Melo
Cc: Frederic Weisbecker
LKML-Reference:
Signed-off-by: Ingo Molnar -
Leave the input fd at the data area.
It does not matter right now - but seeking at the end of it
certainly did not make sense.Cc: Peter Zijlstra
Cc: Mike Galbraith
Cc: Paul Mackerras
Cc: Arnaldo Carvalho de Melo
Cc: Frederic Weisbecker
LKML-Reference:
Signed-off-by: Ingo Molnar -
We started parsing perf.data at head 0. This caused -D to
segfault and it could possibly also case incorrect trace
entries to be displayed.Parse it at data_offset instead.
Cc: Peter Zijlstra
Cc: Mike Galbraith
Cc: Paul Mackerras
Cc: Arnaldo Carvalho de Melo
Cc: Frederic Weisbecker
LKML-Reference:
Signed-off-by: Ingo Molnar -
Before:
perf-21082 [013] 0.000000: sched_wakeup_new: task perf:21083 [120] success=1 [015]
perf-21082 [013] 0.000000: sched_migrate_task: task perf:21082 [120] from: 13 to: 15
perf-21082 [013] 0.000000: sched_process_fork: parent perf:21082 child perf:21083
true-21083 [015] 0.000000: sched_wakeup: task migration/15:33 [0] success=1 [015]
perf-21082 [013] 0.000000: sched_switch: task perf:21082 [120] (S) ==> swapper:0 [140]
true-21083 [015] 0.000000: sched_switch: task perf:21083 [120] (R) ==> migration/15:33 [0]
true-21083 [011] 0.000000: sched_process_exit: task true:21083 [120]After:
perf-21082 [013] 14674.797613: sched_wakeup_new: task perf:21083 [120] success=1 [015]
perf-21082 [013] 14674.797506: sched_migrate_task: task perf:21082 [120] from: 13 to: 15
perf-21082 [013] 14674.797610: sched_process_fork: parent perf:21082 child perf:21083
true-21083 [015] 14674.797725: sched_wakeup: task migration/15:33 [0] success=1 [015]
perf-21082 [013] 14674.797722: sched_switch: task perf:21082 [120] (S) ==> swapper:0 [140]
true-21083 [015] 14674.797729: sched_switch: task perf:21083 [120] (R) ==> migration/15:33 [0]
true-21083 [011] 14674.798159: sched_process_exit: task true:21083 [120]Cc: Peter Zijlstra
Cc: Mike Galbraith
Cc: Paul Mackerras
Cc: Arnaldo Carvalho de Melo
Cc: Frederic Weisbecker
LKML-Reference:
Signed-off-by: Ingo Molnar -
Sample, record, parse and print the CPU field - it had all zeroes before.
Before (watch the second column, the CPU values):
perf-32685 [000] 0.000000: sched_wakeup_new: task perf:32686 [120] success=1 [011]
perf-32685 [000] 0.000000: sched_migrate_task: task perf:32685 [120] from: 1 to: 11
perf-32685 [000] 0.000000: sched_process_fork: parent perf:32685 child perf:32686
true-32686 [000] 0.000000: sched_wakeup: task migration/11:25 [0] success=1 [011]
true-32686 [000] 0.000000: sched_wakeup: task distccd:12793 [125] success=1 [015]
true-32686 [000] 0.000000: sched_wakeup: task distccd:12793 [125] success=1 [015]
perf-32685 [000] 0.000000: sched_switch: task perf:32685 [120] (S) ==> swapper:0 [140]
true-32686 [000] 0.000000: sched_switch: task perf:32686 [120] (R) ==> migration/11:25 [0]
true-32686 [000] 0.000000: sched_switch: task perf:32686 [120] (R) ==> distccd:12793 [125]
true-32686 [000] 0.000000: sched_switch: task true:32686 [120] (R) ==> distccd:12793 [125]
true-32686 [000] 0.000000: sched_process_exit: task true:32686 [120]
true-32686 [000] 0.000000: sched_stat_wait: task: distccd:12793 wait: 6767985949080 [ns]
true-32686 [000] 0.000000: sched_stat_wait: task: distccd:12793 wait: 6767986139446 [ns]
true-32686 [000] 0.000000: sched_stat_sleep: task: distccd:12793 sleep: 132844 [ns]
true-32686 [000] 0.000000: sched_stat_sleep: task: distccd:12793 sleep: 131724 [ns]After:
perf-32685 [001] 0.000000: sched_wakeup_new: task perf:32686 [120] success=1 [011]
perf-32685 [001] 0.000000: sched_migrate_task: task perf:32685 [120] from: 1 to: 11
perf-32685 [001] 0.000000: sched_process_fork: parent perf:32685 child perf:32686
true-32686 [011] 0.000000: sched_wakeup: task migration/11:25 [0] success=1 [011]
true-32686 [015] 0.000000: sched_wakeup: task distccd:12793 [125] success=1 [015]
true-32686 [015] 0.000000: sched_wakeup: task distccd:12793 [125] success=1 [015]
perf-32685 [001] 0.000000: sched_switch: task perf:32685 [120] (S) ==> swapper:0 [140]
true-32686 [011] 0.000000: sched_switch: task perf:32686 [120] (R) ==> migration/11:25 [0]
true-32686 [015] 0.000000: sched_switch: task perf:32686 [120] (R) ==> distccd:12793 [125]
true-32686 [015] 0.000000: sched_switch: task true:32686 [120] (R) ==> distccd:12793 [125]
true-32686 [015] 0.000000: sched_process_exit: task true:32686 [120]
true-32686 [015] 0.000000: sched_stat_wait: task: distccd:12793 wait: 6767985949080 [ns]
true-32686 [015] 0.000000: sched_stat_wait: task: distccd:12793 wait: 6767986139446 [ns]
true-32686 [015] 0.000000: sched_stat_sleep: task: distccd:12793 sleep: 132844 [ns]
true-32686 [015] 0.000000: sched_stat_sleep: task: distccd:12793 sleep: 131724 [ns]So we can now see how this workload migrated between CPUs.
Cc: Peter Zijlstra
Cc: Paul Mackerras
Cc: Frederic Weisbecker
Cc: Li Zefan
Cc: Mike Galbraith
Cc: Arnaldo Carvalho de Melo
LKML-Reference:
Signed-off-by: Ingo Molnar
02 Sep, 2009
2 commits
-
Older versions of GCC are rather stupid about strict aliasing:
util/trace-event-parse.c: In function 'parse_cmdlines':
util/trace-event-parse.c:93: warning: dereferencing type-punned pointer will break strict-aliasing rules
util/trace-event-parse.c: In function 'parse_proc_kallsyms':
util/trace-event-parse.c:155: warning: dereferencing type-punned pointer will break strict-aliasing rules
util/trace-event-parse.c:157: warning: dereferencing type-punned pointer will break strict-aliasing rules
util/trace-event-parse.c:158: warning: dereferencing type-punned pointer will break strict-aliasing rules
util/trace-event-parse.c: In function 'parse_ftrace_printk':
util/trace-event-parse.c:294: warning: dereferencing type-punned pointer will break strict-aliasing rules
util/trace-event-parse.c:295: warning: dereferencing type-punned pointer will break strict-aliasing rules
make: *** [util/trace-event-parse.o] Error 1Make it clear to GCC that we intend with those pointers, by passing
them through via an explicit (void *) cast.We might want to add -fno-strict-aliasing as well, like the kernel
itself does.Cc: Peter Zijlstra
Cc: Mike Galbraith
Cc: Paul Mackerras
Cc: Arnaldo Carvalho de Melo
Cc: Frederic Weisbecker
LKML-Reference:
Signed-off-by: Ingo Molnar -
Make it easier to turn warnings on/off by using a separate
line for each warning added.Some of the warnings have too much of a nuisance factor and
we might want to turn them off in the future.Cc: Arjan van de Ven
Cc: Peter Zijlstra
Cc: Mike Galbraith
Cc: Paul Mackerras
Cc: Arnaldo Carvalho de Melo
Cc: Frederic Weisbecker
LKML-Reference:
Signed-off-by: Ingo Molnar
31 Aug, 2009
6 commits
-
Complete support for __str_loc type strings of ftrace events
which have dynamic offsets values set for each of them inside
their sammples.Before:
geany-5759 [000] 0.000000: lock_release: name
geany-5759 [000] 0.000000: lock_release: name
geany-5759 [000] 0.000000: lock_release: name
kondemand/0-362 [000] 0.000000: lock_release: name
pdflush-421 [000] 0.000000: lock_release: nameAfter:
geany-5759 [000] 0.000000: lock_release: &u->lock
geany-5759 [000] 0.000000: lock_release: key
geany-5759 [000] 0.000000: lock_release: &group->notification_mutex
kondemand/0-362 [000] 0.000000: lock_release: &rq->lock
pdflush-421 [000] 0.000000: lock_release: &rq->lockSigned-off-by: Frederic Weisbecker
Cc: Frederic Weisbecker
Cc: Peter Zijlstra
Cc: Arnaldo Carvalho de Melo
Cc: Steven Rostedt
LKML-Reference:
Signed-off-by: Ingo Molnar
Cc: Peter Zijlstra
Cc: Arnaldo Carvalho de Melo
Cc: Steven Rostedt -
In perf tools, we hardcode the pid 0 cmdline resolving to
"idle" because the init task is not included in the COMM
events.But the idle tasks secondary cpus are resolved into their
"init" name through the COMM events.We have then such strange result in perf report (ditto with
trace):19.66% init [kernel] [k] acpi_idle_enter_c1
17.32% [idle] [kernel] [k] acpi_idle_enter_c1It's then better to unify the swapper tasks into a single init
name.Signed-off-by: Frederic Weisbecker
Cc: Frederic Weisbecker
Cc: Peter Zijlstra
Cc: Arnaldo Carvalho de Melo
LKML-Reference:
Signed-off-by: Ingo Molnar
Cc: Peter Zijlstra
Cc: Arnaldo Carvalho de Melo -
The cmd-trace tool used the cmdline file and resolved the idle
thread using a hardcoded check for the 0 task pid.Now we have a centralized way to do that from perf using
register_idle_thread() API.Before:
:0-0 [000] 0.000000: irq_handler_entry: irq=0 handler=name
:0-0 [000] 0.000000: irq_handler_entry: irq=0 handler=nameAfter:
[idle]-0 [000] 0.000000: irq_handler_entry: irq=0 handler=name
[idle]-0 [000] 0.000000: irq_handler_entry: irq=0 handler=nameSigned-off-by: Frederic Weisbecker
Cc: Peter Zijlstra
Cc: Arnaldo Carvalho de Melo
LKML-Reference:
Signed-off-by: Ingo Molnar -
Librarize register_idle_thread() used by annotate and report.
Signed-off-by: Frederic Weisbecker
Cc: Peter Zijlstra
Cc: Arnaldo Carvalho de Melo
LKML-Reference:
Signed-off-by: Ingo Molnar -
Add missing documentation for the following parameters:
- perf record -R
- perf report -gSigned-off-by: Frederic Weisbecker
Cc: Peter Zijlstra
Cc: Arnaldo Carvalho de Melo
LKML-Reference:
Signed-off-by: Ingo Molnar -
Merge reason: this topic is ready now to merge into the main
development branch for .32, with functional
perf trace output.Signed-off-by: Ingo Molnar
28 Aug, 2009
3 commits
-
This improves patch fa6963b24 so that perf.data stuff that has
been dumped as root can be read (annotate/report) by a user
without the use of the --force.Rationale is that root has plenty of ways to screw us (usually)
that do not require twisted schemes involving specially
crafting a perf.data.Signed-off-by: Pierre Habouzit
Cc: Paul Mackerras
Cc: Peter Zijlstra
Cc:
LKML-Reference:
Signed-off-by: Ingo Molnar -
Some string fields are not printed because of a missing printf
in the post-processing.Before:
perf-10070 [000] 0.000000: sched_switch: task :10070 [120] (R) ==> :5720 [120]
geany-5720 [000] 0.000000: sched_switch: task :5720 [120] (S) ==> :10070 [120]
perf-10070 [000] 0.000000: sched_switch: task :10070 [120] (R) ==> :5720 [120]
geany-5720 [000] 0.000000: sched_switch: task :5720 [120] (S) ==> :10070 [120]
-0 [000] 0.000000: sched_switch: task :0 [140] (R) ==> :361 [115]After:
perf-10070 [000] 0.000000: sched_switch: task perf:10070 [120] (R) ==> geany:5720 [120]
geany-5720 [000] 0.000000: sched_switch: task geany:5720 [120] (S) ==> perf:10070 [120]
perf-10070 [000] 0.000000: sched_switch: task perf:10070 [120] (R) ==> geany:5720 [120]
geany-5720 [000] 0.000000: sched_switch: task geany:5720 [120] (S) ==> perf:10070 [120]
-0 [000] 0.000000: sched_switch: task swapper:0 [140] (R) ==> kondemand/1:361 [115]Signed-off-by: Frederic Weisbecker
Cc: Peter Zijlstra
Cc: Arnaldo Carvalho de Melo
Cc: Steven Rostedt
LKML-Reference:
Signed-off-by: Ingo Molnar -
While opening a trace event counter, every events are saved in
the trace.info file. But we only want to save the
specifications of the events we are using.Signed-off-by: Frederic Weisbecker
Cc: Peter Zijlstra
Cc: Arnaldo Carvalho de Melo
Cc: Steven Rostedt
LKML-Reference:
Signed-off-by: Ingo Molnar
27 Aug, 2009
1 commit
-
Signed-off-by: Arnaldo Carvalho de Melo
Cc: Peter Zijlstra
Cc: Mike Galbraith
Cc: Paul Mackerras
Cc: Frederic Weisbecker
LKML-Reference:
Signed-off-by: Ingo Molnar
24 Aug, 2009
1 commit
-
Conflicts:
tools/perf/builtin-annotate.c
tools/perf/builtin-report.cMerge reason: resolve these conflicts.
Signed-off-by: Ingo Molnar
22 Aug, 2009
1 commit
-
Add OPT_END to option array of perf-trace for fixing a SEGV bug when
showing perf-trace help message.Without this patch;
./perf trace -husage: perf trace []
-D, --dump-raw-trace dump raw trace in ASCII
-v, --verbose be more verbose (show symbol address, etc)
-f, Segmentation faultWith this patch:
./perf trace -husage: perf trace []
-D, --dump-raw-trace dump raw trace in ASCII
-v, --verbose be more verbose (show symbol address, etc)Signed-off-by: Masami Hiramatsu
Cc: systemtap
Cc: DLE
Cc: Masami Hiramatsu
Cc: Steven Rostedt
Cc: Frederic Weisbecker
Cc: Peter Zijlstra
Cc: Arnaldo Carvalho de Melo
Cc: Mike Galbraith
Cc: Xiao Guangrong
Cc: Zhaolei
Cc: Li Zefan
Cc: Lai Jiangshan
Cc: Tom Zanussi
LKML-Reference:
Signed-off-by: Ingo Molnar
19 Aug, 2009
2 commits
-
Add an owner check to opening perf.data files and a switch to
silence it.Because perf-report/perf-annotate are binary parsers reading
another users' perf.data file could be a security risk if the
file were explicitly engineered to trigger bugs in the parser
(we hope of course there are non such bugs, but you never
know).Signed-off-by: Peter Zijlstra
Cc: Arnaldo Carvalho de Melo
Cc: Frederic Weisbecker
Cc: Mike Galbraith
Cc: Peter Zijlstra
Cc: Paul Mackerras
LKML-Reference:
Signed-off-by: Ingo Molnar -
pushd tools/perf/Documentation
make html
popdis failing for me...
ASCIIDOC perf-annotate.html
ERROR: unsafe: include file: /etc/asciidoc/./stylesheets/xhtml11.css
ERROR: unsafe: include file:
/etc/asciidoc/./stylesheets/xhtml11-manpage.css
ERROR: unsafe: include file:
/etc/asciidoc/./stylesheets/xhtml11-quirks.css
make: *** [perf-annotate.html] Error 1Apparently asciidoc "unsafe" is the default mode of operation
in practice.https://bugzilla.redhat.com/show_bug.cgi?id=506953
Works tidily now.
Signed-off-by: Kyle McMartin
Cc: Peter Zijlstra
Cc: Mike Galbraith
Cc: Paul Mackerras
Cc: Arnaldo Carvalho de Melo
Cc: Frederic Weisbecker
LKML-Reference:
Signed-off-by: Ingo Molnar
18 Aug, 2009
8 commits
-
The librarization of the thread helpers between annotate and
report lost some perf report specifics.thread__insert_map() had its most uptodate version in perf
report which cared about partial map overlapping. In case of
overlap between two maps, perf annotate's version removes the
whole old map without considering if it partially or
absolutely overlaps the new map.We exported the odd version, change it by using the perf
report version.Signed-off-by: Frederic Weisbecker
Cc: Frederic Weisbecker
Cc: Peter Zijlstra
Cc: Arnaldo Carvalho de Melo
Cc: Mike Galbraith
LKML-Reference:
Signed-off-by: Ingo Molnar -
The librarization of the thread helpers between annotate and
report lost some perf report specifics.This patch fixes the thread comm column adjusting that has
been omitted during this export.Signed-off-by: Frederic Weisbecker
Cc: Frederic Weisbecker
Cc: Peter Zijlstra
Cc: Arnaldo Carvalho de Melo
Cc: Mike Galbraith
LKML-Reference:
Signed-off-by: Ingo Molnar -
Linus reported this perf annotate segfault:
[torvalds@nehalem git]$ perf annotate unmap_vmas
Segmentation fault#0 map__clone (self=) at builtin-annotate.c:236
#1 thread__fork (self=) at builtin-annotate.c:372The bug here was that builtin-annotate.c was a copy of
builtin-report.c and a threading related fix to builtin-report.c
didnt get propagated to builtin-annotate.c ...Reported-by: Linus Torvalds
Cc: Peter Zijlstra
Cc: Mike Galbraith
Cc: Paul Mackerras
Cc: Arnaldo Carvalho de Melo
Cc: Frederic Weisbecker
LKML-Reference:
Signed-off-by: Ingo Molnar -
The _XOPEN_SOURCE* defines are not really needed on Linux and
it's not like we'll port this to AIX ;-)The define also broke the build with gcc 4.4.1:
CC util/trace-event-parse.o
In file included from util/trace-event-parse.c:32:
util/util.h:43:1: error: "_XOPEN_SOURCE" redefinedSo remove them.
Cc: Peter Zijlstra
Cc: Mike Galbraith
Cc: Paul Mackerras
Cc: Arnaldo Carvalho de Melo
Cc: Frederic Weisbecker
LKML-Reference:
Signed-off-by: Ingo Molnar -
The ftrace event format parser handles the usual casts but not
the cast to pointers. Such casts have been introduced recently
with the module trace events and raise the following parsing
error:Fatal: bad op token )
This is because it considers the "*" character as a binary
operator. Make it then aware of casts to pointers.Signed-off-by: Frederic Weisbecker
Cc: Peter Zijlstra
Cc: Arnaldo Carvalho de Melo
Cc: Mike Galbraith
Cc: Steven Rostedt
LKML-Reference:
Signed-off-by: Ingo Molnar -
You can enable a counter's PERF_SAMPLE_RAW attribute in two
fashions:- using the -R option (every counters get PERF_SAMPLE_RAW)
- using the :record suffix in a trace event counter nameCurrently we record the events info in a trace.info file from
perf record when the former method is used but we omit it with
the latter.Check both situations.
Signed-off-by: Frederic Weisbecker
Cc: Peter Zijlstra
Cc: Arnaldo Carvalho de Melo
Cc: Mike Galbraith
Cc: Steven Rostedt
LKML-Reference:
Signed-off-by: Ingo Molnar -
When a user runs perf trace using an input with logged
counters without PERF_SAMPLE_RAW attribute, warn by giving a
nice tip.Signed-off-by: Frederic Weisbecker
Cc: Peter Zijlstra
Cc: Arnaldo Carvalho de Melo
Cc: Mike Galbraith
Cc: Steven Rostedt
LKML-Reference:
Signed-off-by: Ingo Molnar -
While running perf report -g in a perf.data file that hasn't
been recorded in callchain mode, the error reported has a
spelling issue:./perf report -g
selected -c but no callchain data. Did you call perf record without -g?Fix it.
Signed-off-by: Frederic Weisbecker
Cc: Peter Zijlstra
Cc: Arnaldo Carvalho de Melo
Cc: Mike Galbraith
Cc: Steven Rostedt
LKML-Reference:
Signed-off-by: Ingo Molnar