We are using the Python scripting interface in perf to extract kernel
events relevant for performance analysis of HPC codes. We noticed that
the "perf script" call allocates a significant amount of memory (in the
order of several 100 MiB) during it's run, e.g. 125 MiB for a 25 MiB
input file:
$> perf record -o perf.data -a -R -g fp \
-e power:cpu_frequency -e sched:sched_switch \
-e sched:sched_migrate_task -e sched:sched_process_exit \
-e sched:sched_process_fork -e sched:sched_process_exec \
-e cycles -m 4096 --freq 4000
$> /usr/bin/time perf script -i perf.data -s dummy_script.py
0.84user 0.13system 0:01.92elapsed 51%CPU (0avgtext+0avgdata
125532maxresident)k
73072inputs+0outputs (57major+33086minor)pagefaults 0swaps
Upon further investigation using the valgrind massif tool, we noticed
that Python objects that are created in trace-event-python.c via
PyString_FromString*() (and their Integer and Long counterparts) are
never free'd.
The reason for this seem to be missing Py_DECREF calls on the objects
that are returned by these functions and stored in the Python
dictionaries. The Python dictionaries do not steal references (as
opposed to Python tuples and lists) but instead add their own reference.
Hence, the reference that is returned by these object creation functions
is never released and the memory is leaked. (see [1,2])
The attached patch fixes this by wrapping all relevant calls to
PyDict_SetItemString() and decrementing the reference counter
immediately after the Python function call.
This reduces the allocated memory to a reasonable amount:
$> /usr/bin/time perf script -i perf.data -s dummy_script.py
0.73user 0.05system 0:00.79elapsed 99%CPU (0avgtext+0avgdata
49132maxresident)k
0inputs+0outputs (0major+14045minor)pagefaults 0swaps
For comparison, with a 120 MiB input file the memory consumption
reported by time drops from almost 600 MiB to 146 MiB.
The patch has been tested using Linux 3.8.2 with Python 2.7.4 and Linux
3.11.6 with Python 2.7.5.
Please let me know if you need any further information.
[1] http://docs.python.org/2/c-api/tuple.html#PyTuple_SetItem
[2] http://docs.python.org/2/c-api/dict.html#PyDict_SetItemString
Signed-off-by: Joseph Schuchart <joseph.schuchart@tu-dresden.de>
Reviewed-by: Tom Zanussi <tom.zanussi@linux.intel.com>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Tom Zanussi <tom.zanussi@linux.intel.com>
Link: http://lkml.kernel.org/r/1381468543-25334-4-git-send-email-namhyung@kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
At this point, --fentry (mcount function entry) option for gcc fuzzes
the debuginfo variable locations by skipping the mcount instruction
offset (on x86, this is a 5 byte call instruction).
This makes variable searching fail at the entry of functions which
are mcount'ed.
e.g.)
Available variables at vfs_read
@<vfs_read+0>
(No matched variables)
This patch adds additional location search at the function entry point
to solve this issue, which tries to find the earliest address for the
variable location.
Note that this only works with function parameters (formal parameters)
because any local variables should not exist on the function entry
address (those are not initialized yet).
With this patch, perf probe shows correct parameters if possible;
# perf probe --vars vfs_read
Available variables at vfs_read
@<vfs_read+0>
char* buf
loff_t* pos
size_t count
struct file* file
Signed-off-by: Masami Hiramatsu <masami.hiramatsu.pt@hitachi.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Link: http://lkml.kernel.org/r/20131011071025.15557.13275.stgit@udc4-manage.rcp.hitachi.co.jp
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
When callgraph data was included in the perf data file, it may take a
long time to scan all those data and merge them together especially if
the stored callchains are long and the perf data file itself is large,
like a Gbyte or so.
The callchain stack is currently limited to PERF_MAX_STACK_DEPTH (127).
This is a large value. Usually the callgraph data that developers are
most interested in are the first few levels, the rests are usually not
looked at.
This patch adds a new --max-stack option to perf-report to limit the
depth of callchain stack data to look at to reduce the time it takes for
perf-report to finish its processing. It trades the presence of trailing
stack information with faster speed.
The following table shows the elapsed time of doing perf-report on a
perf.data file of size 985,531,828 bytes.
--max_stack Elapsed Time Output data size
----------- ------------ ----------------
not set 88.0s 124,422,651
64 87.5s 116,303,213
32 87.2s 112,023,804
16 86.6s 94,326,380
8 59.9s 33,697,248
4 40.7s 10,116,637
-g none 27.1s 2,555,810
Signed-off-by: Waiman Long <Waiman.Long@hp.com>
Acked-by: David Ahern <dsahern@gmail.com>
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: Aswin Chandramouleeswaran <aswin@hp.com>
Cc: David Ahern <dsahern@gmail.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Jiri Olsa <jolsa@redhat.com>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Scott J Norton <scott.norton@hp.com>
Cc: Stephane Eranian <eranian@google.com>
Link: http://lkml.kernel.org/r/1382107129-2010-4-git-send-email-Waiman.Long@hp.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Linus reported that sometimes 'perf report -s symbol' exits without any
message on TUI. David and Jiri found that it's because it failed to add
a hist entry due to an invalid symbol length.
It turns out that sorting by symbol (address) was broken since it only
compares symbol addresses. The symbol address is a relative address
within a dso thus just checking its address can result in merging
unrelated symbols together. Fix it by checking dso before comparing
symbol address.
Reported-by: Linus Torvalds <torvalds@linux-foundation.org>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
Cc: David Ahern <dsahern@gmail.com>
Cc: Ingo Molnar <mingo@kernel.org>
Cc: Jiri Olsa <jolsa@redhat.com>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Link: http://lkml.kernel.org/r/1381802517-18812-1-git-send-email-namhyung@kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Current collapse stage has a scalability problem which can be reproduced
easily with a parallel kernel build.
This is because it needs to traverse every children of callchains
linearly during the collapse/merge stage.
Converting it to a rbtree reduced the overhead significantly.
On my 400MB perf.data file which recorded with make -j32 kernel build:
$ time perf --no-pager report --stdio > /dev/null
before:
real 6m22.073s
user 6m18.683s
sys 0m0.706s
after:
real 0m20.780s
user 0m19.962s
sys 0m0.689s
During the perf report the overhead on append_chain_children went down
from 96.69% to 18.16%:
- 18.16% perf perf [.] append_chain_children
- append_chain_children
- 77.48% append_chain_children
+ 69.79% merge_chain_branch
- 22.96% append_chain_children
+ 67.44% merge_chain_branch
+ 30.15% append_chain_children
+ 2.41% callchain_append
+ 7.25% callchain_append
+ 12.26% callchain_append
+ 10.22% merge_chain_branch
+ 11.58% perf perf [.] dso__find_symbol
+ 8.02% perf perf [.] sort__comm_cmp
+ 5.48% perf libc-2.17.so [.] malloc_consolidate
Reported-by: Linus Torvalds <torvalds@linux-foundation.org>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: Ingo Molnar <mingo@kernel.org>
Cc: Jiri Olsa <jolsa@redhat.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Link: http://lkml.kernel.org/r/1381468543-25334-2-git-send-email-namhyung@kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
For now, we disable the extended MMAP record support (MMAP2).
We have identified cases where it would not report the correct mapping
information, clone(VM_CLONE) but with separate pids. We will revisit
the support once we find a solution for this case.
The patch changes the kernel to return EINVAL if attr->mmap2 is set. The
patch also modifies the perf tool to use regular PERF_RECORD_MMAP for
synthetic events and it also prevents the tool from requesting
attr->mmap2 mode because the kernel would reject it.
The support will be revisited once the kenrel interface is updated.
In V2, we reduce the patch to the strict minimum.
In V3, we avoid calling perf_event_open() with mmap2 set because we know
it will fail and require fallback retry.
Signed-off-by: Stephane Eranian <eranian@google.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: David Ahern <dsahern@gmail.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Jiri Olsa <jolsa@redhat.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Link: http://lkml.kernel.org/r/20131017173215.GA8820@quad
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
kcore can be used to view the running kernel object code. However,
kcore changes as modules are loaded and unloaded, and when the kernel
decides to modify its own code. Consequently it is useful to create a
copy of kcore at a particular time. Unlike vmlinux, kcore is not unique
for a given build-id. And in addition, the kallsyms and modules files
are also needed. The tool therefore creates a directory:
~/.debug/[kernel.kcore]/<build-id>/<YYYYmmddHHMMSShh>
which contains: kcore, kallsyms and modules.
Note that the copied kcore contains only code sections. See the
kcore_copy() function for how that is determined.
The tool will not make additional copies of kcore if there is already
one with the same modules at the same addresses.
Currently, perf tools will not look for kcore in the cache. That is
addressed in another patch.
Signed-off-by: Adrian Hunter <adrian.hunter@intel.com>
Cc: David Ahern <dsahern@gmail.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: Jiri Olsa <jolsa@redhat.com>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Namhyung Kim <namhyung@gmail.com>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Stephane Eranian <eranian@google.com>
Link: http://lkml.kernel.org/r/525BF849.5030405@intel.com
[ renamed 'index' to 'idx' to avoid shadowing string.h symbol in f12,
use at least one member initializer when initializing a struct to
zeros, also to fix the build on f12 ]
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
trace-event-parse.c:parse_proc_kallsyms()
Old GCC (4.4.2) does not see through the code flow of get_srcline() and
gets confused about the status of 'file' and 'line':
CC /tmp/build/perf/util/srcline.o
cc1: warnings being treated as errors
util/srcline.c: In function ¿get_srcline¿:
util/srcline.c:226: error: ¿file¿ may be used uninitialized in this function
util/srcline.c:227: error: ¿line¿ may be used uninitialized in this function
make[1]: *** [/tmp/build/perf/util/srcline.o] Error 1
make: *** [install] Error 2
make: Leaving directory `/home/acme/git/linux/tools/perf'
[acme@fedora12 linux]$
Help out GCC by initializing 'file' and 'line'.
Signed-off-by: David Ahern <dsahern@gmail.com>
Cc: Ingo Molnar <mingo@kernel.org>
Cc: Jiri Olsa <jolsa@redhat.com>
Cc: Namhyung Kim <namhyung@gmail.com>
Link: http://lkml.kernel.org/n/tip-h8k7h49z3cndqgjdftkmm9f8@git.kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Task comm's are getting lost when processing events from a file. The
problem is that the trace struct used by the live processing has its
host machine and the perf-session used for file based processing has its
host machine. Fix by having both references point to the same machine.
Before:
0.030 ( 0.001 ms): :27743/27743 brk( ...
0.057 ( 0.004 ms): :27743/27743 mmap(len: 4096, prot: READ|WRITE, flags: ...
0.075 ( 0.006 ms): :27743/27743 access(filename: 0x7f3809fbce00, mode: R ...
0.091 ( 0.005 ms): :27743/27743 open(filename: 0x7f3809fba14c, flags: CLOEXEC ...
...
After:
0.030 ( 0.001 ms): make/27743 brk( ...
0.057 ( 0.004 ms): make/27743 mmap(len: 4096, prot: READ|WRITE, flags: ...
0.075 ( 0.006 ms): make/27743 access(filename: 0x7f3809fbce00, mode: R ...
0.091 ( 0.005 ms): make/27743 open(filename: 0x7f3809fba14c, flags: CLOEXEC ...
...
Signed-off-by: David Ahern <dsahern@gmail.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: Ingo Molnar <mingo@kernel.org>
Cc: Jiri Olsa <jolsa@redhat.com>
Cc: Namhyung Kim <namhyung.kim@lge.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Stephane Eranian <eranian@google.com>
Link: http://lkml.kernel.org/r/1380395584-9025-4-git-send-email-dsahern@gmail.com
[ Moved creation of new host machine to a separate constructor: machine__new_host() ]
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
The perf_evlist__mmap_read used 'union perf_event' as a placeholder for
event crossing the mmap boundary.
This is ok for sample shorter than ~PATH_MAX. However we could grow up
to the maximum sample size which is 16 bits max.
I hit this overflow issue when using 'perf top -G dwarf' which produces
sample with the size around 8192 bytes. We could configure any valid
sample size here using: '-G dwarf,size'.
Using array with sample max size instead for the event placeholder. Also
adding another safe check for the dynamic size of the user stack.
TODO: The 'struct perf_mmap' is quite big now, maybe we could use some
lazy allocation for event_copy size.
Signed-off-by: Jiri Olsa <jolsa@redhat.com>
Acked-by: David Ahern <dsahern@gmail.com>
Cc: Corey Ashford <cjashfor@linux.vnet.ibm.com>
Cc: David Ahern <dsahern@gmail.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Link: http://lkml.kernel.org/r/1380721599-24285-1-git-send-email-jolsa@redhat.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
