oprofile Log


Commit Date  
[635313] by Maynard Johnson Maynard Johnson

Plug timing hole between JIT agent and opjitconv that can corrupt dump file

There's a window of time where oprofile's Java JIT agent library may be
writing to a JIT dump file at a time when the opjitconv is simultaneously
making a copy of it for purposes of processing its records to create an
ELF file. It's possible that opjitconv may make a copy of the file when
it's not in a coherent state -- i.e., the JIT agent may be in the process
of performing multiple writes for a record. There is no OS-level locking
of the file, so neither of the processes involved (the JIT agent and
opjitconv) are aware the other one is accessing the file.

No actual bugs have been reported due to this problem, so this patch is
a pre-emptive fix. Symptoms of this problem occurring would include
opjitconv not processing all JIT dump files, and operf displaying the message
JIT dump processing exited abnormally: 1
Passing "--verbose=debug" to operf to collect detailed information on the JIT
conversion process would show the following error message:

record past end of file
JIT convert error -1

Signed-off-by: Maynard Johnson <maynardj@us.ibm.com>

2014-02-12 14:41:40 Tree
[a5eec4] by Maynard Johnson Maynard Johnson

Minor fixup for previous commit

The previous commit for the new APM X-Gene (AaArch64 ARMv8)
processor went through a number of iterations before acceptance.
I missed changing one of the references to the new CPU type
from CPU_ARM_V8_XGENE to CPU_ARM_V8_APM_XGENE when I committed it.
This patch fixes that.

Signed-off-by: Maynard Johnson <maynardj@us.ibm.com>

2014-02-12 14:29:15 Tree
[34d006] by William Cohen William Cohen , pushed by Maynard Johnson Maynard Johnson

Provide basic AArch64 (ARMv8) support

The AArch64 (ARMv8) support is provided as an ARM variant to allow use
in both 32-bit and 64-bit ARM environments. The support in this patch
is just the basic events described in the AArch64 documentation.
AArch64 processor implementation may provide additional implementation
specific events. One could add code to recognize those processor
specific implementations and include the armv8-pmuv3-common base
events into the event sets for the processor implementations.
The APM X-Gene processor type is included in this patch as an
implementation, although there are no known processor-specific events
to add at this time.

Below is example run on the ARM Foundation simulator collecting data
on a build of OProfile.

$ cd oprofile
$ operf make
...
$ opreport -t 5
Using /home/wcohen/oprofile/oprofile/oprofile_data/samples/ for samples directory.

WARNING: Lost samples detected! See /home/wcohen/oprofile/oprofile/oprofile_data/samples/operf.log for details.
CPU: ARM AArch64
Counted CPU_CYCLES events (Cycle) with a unit mask of 0x00 (No unit mask) count 100000
CPU_CYCLES:100000|
samples| %|
------------------
10943 90.5877 make
CPU_CYCLES:100000|
samples| %|
------------------
5281 48.2592 make
4543 41.5151 libc-2.17.so
1079 9.8602 kallsyms
40 0.3655 ld-2.17.so
735 6.0844 sh
CPU_CYCLES:100000|
samples| %|
------------------
321 43.6735 kallsyms
298 40.5442 libc-2.17.so
94 12.7891 bash
22 2.9932 ld-2.17.so

Signed-off-by: William Cohen <wcohen@redhat.com>

2014-02-12 14:05:38 Tree
[492e52] by Maynard Johnson Maynard Johnson

Enhance ocount to support millisecond time intervals

The ocount tool currently has a --time-interval option that
allows users to print a snapshot of event counts every n seconds,
where 'n' is an integer value equal to or greater than 1.
A user requested a change to allow time intervals of 100 ms.
The user stated that he uses an event counting tool on a
non-Linux operating system, and that tool had the option of
collecting counts every 100 ms. Now, he wanted the same ability
on Linux so he could do direct comparison of data. The
original choice of one second granularity for the time interval
option was arbitrary, so the request (based on a real world
use case) seemed reasonable.

This patch changes ocount's '--time-interval' option to accept
values in units of milliseconds instead of seconds. At this
time, the implementation supports only 100 ms granularity, so
time intervals passed will be rounded to the nearest 100 ms.

NOTE: Current users of ocount who make use of the '--time-interval'
option will see a change in behavior. I don't believe this is
a serious compatibility issue since the timestamp printed along
with the event counts makes it very clear what the actual
time interval is, and the user can check the ocount man page
for clarification.

Signed-off-by: Maynard Johnson <maynardj@us.ibm.com>

2014-02-07 23:02:41 Tree
[31389d] by Maynard Johnson Maynard Johnson

Fix PM_RUN_CYC and PM_RUN_INST_CMPL event codes broken by previous commit

Signed-off-by: Maynard Johnson <maynardj@us.ibm.com>

2014-02-07 16:27:46 Tree
[029735] by Maynard Johnson Maynard Johnson

Fix various event names and codes for IBM architected and POWER8 events

Signed-off-by: Maynard Johnson <maynardj@us.ibm.com>

2014-02-07 14:58:28 Tree
[7243fa] by Maynard Johnson Maynard Johnson

Make cpu type POWER8E equivalent to POWER8

Recent mainline kernel changes resulted in a cpu type of
"POWER8E" being displayed in /proc/cpuinfo for certain revisions
of the IBM POWER8 processor model. But for profiling and
counting of native events, we can ignore the differences between
POWER8 and POWER8E. This patch addresses that issue.

Signed-off-by: Maynard Johnson <maynardj@us.ibm.com>

2014-02-04 14:27:10 Tree
[717d45] by Maynard Johnson Maynard Johnson

Fix up event codes for marked architected events

Fourteen events in the set of architected events had the wrong
event encoding. All 14 were "marked" events, used in random
sampling.

Signed-off-by: Maynard Johnson <maynardj@us.ibm.com>

2014-02-03 23:50:54 Tree
[fd05da] by Maynard Johnson Maynard Johnson

Remove 'extra' attribute from ophelp XML output; bump schema version

As discussed on the oprofile mailing list on Sep 24, 2013, there is
no value add in keeping the 'extra' attribute in ophelp's XML output.
The previous commit added the 'name' field to the XML output, and
that is actual valuable information that consumers of the XML output
should use when coding event specifications to pass to operf or
ocount.

This patch removes the 'extra' attribute and also bumps the schema
version (both in the ophelp.xsd and the XML instance documents).
The schema bump is needed mostly due to removing the 'extra' attribute;
but another reason for it is to draw attention to the new 'name'
attribute, which consumers really must use (when present) in order
to be sure they can properly specify the unitmask that the user
requests.

Signed-off-by: Maynard Johnson <maynardj@us.ibm.com>

2014-02-03 14:47:30 Tree
[c0b7ac] by Maynard Johnson Maynard Johnson

Make operf/ocount detect invalid timer mode from opcontrol

Certain architecture/processor models are not supported by the
legacy oprofile kernel driver, so the cpu type detected (when
opcontrol is run) is "timer". If a user runs opcontrol on such
a system and does not unload the oprofile kernel module (using
'opcontrol --deinit') prior to using operf or ocount, the operf
and ocount tools will fail; for example, running 'operf <cmd>'
(with no event specified) will fail with the following unhelpful
message:

Unable to find default event

This failure is due to how operf and ocount ascertain the cpu
type by calling libop/op_events.c:op_get_cpu_type(). That function
will look first in the oprofilefs (/dev/oprofile/cpu_type, in particular)
to try to determine the running cpu type. In the case described above,
the /dev/oprofile/cpu_type file contains 'timer', so operf and ocount
then try to find the default event for cpu type 'timer', but since timer
mode is not supported on either operf or ocount, there is default event,
so the tool fails.

Additionally, the ophelp command, when run in the same situation (where
the /dev/oprofile/cpu_type file contains 'timer') will not display the
expected list of native events; instead it simply says:
Using timer interrupt.

This patch makes operf and ocount exit immediately with a helpful message
if cpu type 'timer' is detected. A helpful message is also added to ophelp
for the same circumstances.

Signed-off-by: Maynard Johnson <maynardj@us.ibm.com>

2014-01-31 15:05:18 Tree
[ed40d8] by William Cohen William Cohen , pushed by Maynard Johnson Maynard Johnson

Print unit mask name where applicable in ophelp XML output

Some Intel architectures have named unit masks and it would be useful
to include the unit mask name in the XML output. This patch also
updates the ophelp.xsd schema file to include the optional unit
mask 'name' field.

Signed-off-by: William Cohen <wcohen@redhat.com>

2014-01-28 17:05:46 Tree
[be6d22] by Maynard Johnson Maynard Johnson

Fix issues detected by Coverity

Will Cohen ran Coverity against oprofile and reported some issues
on Nov 20, 2013. I submitted the current oprofile source to the
Coverity webpage, and a couple new issues were detected. This
patch addresses most of these issues. Some issues are either
false positives from Coverity's analysis or have been marked
as "Intentional" so as to have Coverity ignore them.

Signed-off-by: Maynard Johnson <maynardj@us.ibm.com>

2014-01-27 21:44:18 Tree
[d6ca28] by Maynard Johnson Maynard Johnson

Reduce overhead of operf waiting for profiled app to end

The original implementation was very inefficient in how operf
waited for the profiled app to end. We reduce much of the overhead
by doing a nanosleep for 100 ms, waking up, checking the status
of the app (via waitpid), and going back to sleep again if the app
is still running.

Signed-off-by: Maynard Johnson <maynardj@us.ibm.com>

2014-01-24 23:07:10 Tree
[65176c] by Maynard Johnson Maynard Johnson

Fix regression in IBM POWER8 running in POWER7 compat mode

A commit made on Dec 17, 2013 ("Allow all native events for IBM POWER8
in POWER7 compat mode) broke support for POWER8 in POWER7 compat mode.
Instead, oprofile attempts to treat it as a normal POWER7 processor,
which is not correct. A user reported the following error when
running operf with the default CYCLES event:

terminate called after throwing an instance of 'std::runtime_error'
what(): libpfm cannot find event code for CYCLES; cannot continue
Aborted

This patch fixes this problem.

Signed-off-by: Maynard Johnson <maynardj@us.ibm.com>

2014-01-21 20:43:02 Tree
[1c1636] by Maynard Johnson Maynard Johnson

Minor man page cleanups for the ocount command

Signed-off-by: Maynard Johnson <maynardj@us.ibm.com>

2014-01-17 22:52:45 Tree
[d0eb98] by Leonid Moiseichuk Leonid Moiseichuk , pushed by Maynard Johnson Maynard Johnson

Add minimal (armv7-common) support for ARMv7 Krait

Just common ARM PMU events are supported with this patch due
to lack of available documentation.

Other developers reports that operf also works nice.
Tested on device and using "make distcheck".
Signed-off-by: Leonid Moiseichuk <l.moiseichuk@samsung.com>

2014-01-10 15:13:28 Tree
[e3edba] by Maynard Johnson Maynard Johnson

Whitespace fix in configure.ac from previous commit

Signed-off-by: Maynard Johnson <maynardj@us.ibm.com>

2014-01-09 21:53:08 Tree
[a265c5] by Maynard Johnson Maynard Johnson

Enable oprofile for new ppc64le architecture

Signed-off-by: Maynard Johnson <maynardj@us.ibm.com>

2014-01-09 21:47:09 Tree
[88ed74] by Maynard Johnson Maynard Johnson

Fix "Unable to open cpu_type file for reading" for IBM POWER7+

Using operf to do profiling on an IBM POWER7+ may result in
the following error message:

Unable to open cpu_type file for reading

This patch fixes the problem. There is also a simple workaround of
running 'opcontrol --init'.

Signed-off-by: Maynard Johnson <maynardj@us.ibm.com>

2014-01-09 21:07:21 Tree
[286f23] by Maynard Johnson Maynard Johnson

Fix compile errors occurring with gcc 4.8.x

Signed-off-by: Maynard Johnson <maynardj@us.ibm.com>

2014-01-08 18:31:24 Tree
[bb56ad] by Maynard Johnson Maynard Johnson

Fix ocount man page and usage regarding counting modes

The ocount tool must be run with one and only one of the following
counting modes:
o system-wide
o process-list
o thread-list
o cpu-list
o command [args]

The ocount man page and usage printout was missing the logical OR
separator ('|') between cpu-list and command modes. This patch
fixes that.

Signed-off-by: Maynard Johnson <maynardj@us.ibm.com>

2014-01-03 15:49:29 Tree
[344bac] by Maynard Johnson Maynard Johnson

Remove unused variable 'tab' from previous commit

Signed-off-by: Maynard Johnson <maynardj@us.ibm.com>

2014-01-02 16:23:45 Tree
[51e615] by Maynard Johnson Maynard Johnson

opreport from 'operf --callgraph' profile shows incorrect recursive calls

When you collect a callgraph profile with operf, the opreport output
incorrectly implies recursive calls. For example, a simple memcpy
testcase that has the following true callchain:
main -> do_my_memcpy -> memcpy (libc)

appears as follows with 'opreport --callgraph' (focusing here just
on the do_my_memcpy callers and callees):

4757 50.0000 memcpyt do_my_memcpy
4757 50.0000 memcpyt main
4757 6.3185 memcpyt do_my_memcpy
4757 49.9842 memcpyt do_my_memcpy
4757 49.9842 memcpyt do_my_memcpy [self]
3 0.0315 no-vmlinux /no-vmlinux

NOTE: Lines above the non-indented line show the callers of do_my_memcpy;
lines below the non-indented line show the callees of do_my_memcpy.
So it appears that do_my_memcpy calls itself, which it does not do.

If I use 'perf record' to get a callgraph profile, the 'perf report'
looks like the following:

6.88% memcpyt memcpyt [.] do_my_memcpy
|
--- do_my_memcpy
main
__libc_start_main

So here, too, it seems to me that do_my_memcpy calls do_my_memcpy.
When I reported this issue to perf/perf_events kernel developers,
I was basically told that this behavior was "by design".

This patch makes an effort to handle this issue by having operf drop
the first address in the callchain if and only if it is the same
address as the second address in the callchain.

Signed-off-by: Maynard Johnson <maynardj@us.ibm.com>

2014-01-02 14:35:17 Tree
[746b5c] by Maynard Johnson Maynard Johnson

Update TODO list

Signed-off-by: Maynard Johnson <maynardj@us.ibm.com>

2013-12-20 15:26:28 Tree
[46a673] by Maynard Johnson Maynard Johnson

Fix sample attribution problem when using multiple events

A serious bug was found that affects operf profiling with
multiple events. Samples for an event may be incorrectly
attributed to another event. For example, profiling on
a Sandybridge laptop with CPU_CLK_UNHALTED and INST_RETIRED
events produces the following summary counts from opreport:

CPU_CLK_UNHALTED |INST_RETIRED |
samples| %| samples| %|
------------------------------------
32412 100.000 20104 100.000 foo

Using operf to produce separate profiles for these two events
results in these sample counts:

CPU_CLK_UNHALTED |
samples| %|
------------------
18962 100.000 foo

INST_RETIRED |
samples| %|
------------------
33464 100.000 foo

This patch fixes the problem.

Signed-off-by: Maynard Johnson <maynardj@us.ibm.com>

2013-12-20 15:10:46 Tree
Older >