In perf's 'mem-mode', one can get access to a whole bunch of details specific to a
particular sample instruction. A bunch of those details relate to the data
address.
One interesting thing you can do with data addresses is to convert them into a unique
cacheline they belong too. Organizing these data cachelines into similar groups and sorting
them can reveal cache contention.
This patch creates an alogorithm based on various sample details that can help group
entries together into data cachelines and allows 'perf report' to sort on it.
The algorithm relies on having proper mmap2 support in the kernel to help determine
if the memory map the data address belongs to is private to a pid or globally shared.
The alogortithm is as follows:
o group cpumodes together
o group entries with discovered maps together
o sort on major, minor, inode and inode generation numbers
o if userspace anon, then sort on pid
o sort on cachelines based on data addresses
The 'dcacheline' sort option in 'perf report' only works in 'mem-mode'.
Sample output:
#
# Samples: 206 of event 'cpu/mem-loads/pp'
# Total weight : 2534
# Sort order : dcacheline,pid
#
# Overhead Samples Data Cacheline Command: Pid
# ........ ............ ...................................................................... ..................
#
13.22% 1 [k] 0xffff88042f08ebc0 swapper: 0
9.27% 1 [k] 0xffff88082e8cea80 swapper: 0
3.59% 2 [k] 0xffffffff819ba180 swapper: 0
0.32% 1 [k] arch_trigger_all_cpu_backtrace_handler_na.23901+0xffffffffffffffe0 swapper: 0
0.32% 1 [k] timekeeper_seq+0xfffffffffffffff8 swapper: 0
Note: Added a '+1' to symlen size in hists__calc_col_len to prevent the next column
from prematurely tabbing over and mis-aligning. Not sure what the problem is.
Signed-off-by: Don Zickus <dzickus@redhat.com>
Link: http://lkml.kernel.org/r/1401208087-181977-8-git-send-email-dzickus@redhat.com
Signed-off-by: Jiri Olsa <jolsa@kernel.org>
After output/sort fields refactoring, it's expensive
to check the elide bool in its current location inside
the 'struct sort_entry'.
The perf_hpp__should_skip function gets highly noticable in
workloads with high number of output/sort fields, like for:
$ perf report -i perf-test.data -F overhead,sample,period,comm,pid,dso,symbol,cpu --stdio
Performance report:
9.70% perf [.] perf_hpp__should_skip
Moving the elide bool into the 'struct perf_hpp_fmt', which
makes the perf_hpp__should_skip just single struct read.
Got speedup of around 22% for my test perf.data workload.
The change should not harm any other workload types.
Performance counter stats for (10 runs):
before:
358,319,732,626 cycles ( +- 0.55% )
467,129,581,515 instructions # 1.30 insns per cycle ( +- 0.00% )
150.943975206 seconds time elapsed ( +- 0.62% )
now:
278,785,972,990 cycles ( +- 0.12% )
370,146,797,640 instructions # 1.33 insns per cycle ( +- 0.00% )
116.416670507 seconds time elapsed ( +- 0.31% )
Acked-by: Namhyung Kim <namhyung@kernel.org>
Cc: Arnaldo Carvalho de Melo <acme@kernel.org>
Cc: Corey Ashford <cjashfor@linux.vnet.ibm.com>
Cc: David Ahern <dsahern@gmail.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: Ingo Molnar <mingo@kernel.org>
Cc: Namhyung Kim <namhyung@kernel.org>
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/20140601142622.GA9131@krava.brq.redhat.com
Signed-off-by: Jiri Olsa <jolsa@kernel.org>
The new ->add_entry_cb() will be called after an entry was added to
the histogram. It's used for code sharing between perf report and
perf top. Note that ops->add_*_entry() should set iter->he properly
in order to call the ->add_entry_cb.
Also pass @arg to the callback function. It'll be used by perf top
later.
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
Tested-by: Arun Sharma <asharma@fb.com>
Tested-by: Rodrigo Campos <rodrigo@sdfg.com.ar>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Link: http://lkml.kernel.org/r/87k393g999.fsf@sejong.aot.lge.com
Signed-off-by: Jiri Olsa <jolsa@kernel.org>
Print accumulated stat of a hist entry if requested.
To do that, add new HPP_PERCENT_ACC_FNS macro and generate a
perf_hpp_fmt using it. The __hpp__sort_acc() function sorts entries
by accumulated period value. When accumulated periods of two entries
are same (i.e. single path callchain) put the caller above since
accumulation tends to put callers on higher position for obvious
reason.
Also add "overhead_children" output field to be selected by user.
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
Tested-by: Arun Sharma <asharma@fb.com>
Tested-by: Rodrigo Campos <rodrigo@sdfg.com.ar>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Link: http://lkml.kernel.org/r/1401335910-16832-11-git-send-email-namhyung@kernel.org
Signed-off-by: Jiri Olsa <jolsa@kernel.org>
There're some duplicate code when adding hist entries. They are
different in that some have branch info or mem info but generally do
same thing. So introduce new struct hist_entry_iter and add callbacks
to customize each case in general way.
The new perf_evsel__add_entry() function will look like:
iter->prepare_entry();
iter->add_single_entry();
while (iter->next_entry())
iter->add_next_entry();
iter->finish_entry();
This will help further work like the cumulative callchain patchset.
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
Tested-by: Arun Sharma <asharma@fb.com>
Tested-by: Rodrigo Campos <rodrigo@sdfg.com.ar>
Cc: David Ahern <dsahern@gmail.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: Stephane Eranian <eranian@google.com>
Link: http://lkml.kernel.org/r/1401335910-16832-3-git-send-email-namhyung@kernel.org
Signed-off-by: Jiri Olsa <jolsa@kernel.org>
The -F/--fields option is to allow user setup output field in any
order. It can receive any sort keys and following (hpp) fields:
overhead, overhead_sys, overhead_us, sample and period
If guest profiling is enabled, overhead_guest_{sys,us} will be
available too.
The output fields also affect sort order unless you give -s/--sort
option. And any keys specified on -s option, will also be added to
the output field list automatically.
$ perf report -F sym,sample,overhead
...
# Symbol Samples Overhead
# .......................... ............ ........
#
[.] __cxa_atexit 2 2.50%
[.] __libc_csu_init 4 5.00%
[.] __new_exitfn 3 3.75%
[.] _dl_check_map_versions 1 1.25%
[.] _dl_name_match_p 4 5.00%
[.] _dl_setup_hash 1 1.25%
[.] _dl_sysdep_start 1 1.25%
[.] _init 5 6.25%
[.] _setjmp 6 7.50%
[.] a 8 10.00%
[.] b 8 10.00%
[.] brk 1 1.25%
[.] c 8 10.00%
Note that, the example output above is captured after applying next
patch which fixes sort/comparing behavior.
Requested-by: Ingo Molnar <mingo@kernel.org>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
Acked-by: Ingo Molnar <mingo@kernel.org>
Link: http://lkml.kernel.org/r/1400480762-22852-12-git-send-email-namhyung@kernel.org
Signed-off-by: Jiri Olsa <jolsa@kernel.org>
Until now the hpp and sort functions do similar jobs different ways.
Since the sort functions converted/wrapped to hpp formats it can do
the job in a uniform way.
The perf_hpp__sort_list has a list of hpp formats to sort entries and
the perf_hpp__list has a list of hpp formats to print output result.
To have a backward compatibility, it automatically adds 'overhead'
field in front of sort list. And then all of fields in sort list
added to the output list (if it's not already there).
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
Acked-by: Ingo Molnar <mingo@kernel.org>
Link: http://lkml.kernel.org/n/tip-7g3h86woz2sckg3h1lj42ygj@git.kernel.org
Signed-off-by: Jiri Olsa <jolsa@kernel.org>
The hist_browser__reset() is only called right after a filter is
applied so it needs to udpate browser->nr_entries properly. We cannot
use hists->nr_non_filtered_entreis directly since it's possible that
such entries are also filtered out by minimum percentage limit.
In addition when a filter is used for perf top, hist browser's
nr_entries field was not updated after applying the filter. But it
needs to be updated as new samples are coming.
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
Link: http://lkml.kernel.org/r/1398327843-31845-11-git-send-email-namhyung@kernel.org
Signed-off-by: Jiri Olsa <jolsa@kernel.org>
Add hist.percentage option for setting default value of the
symbol_conf.filter_relative. It affects the output of various perf
commands (like perf report, top and diff) only if filter(s) applied.
An user can write .perfconfig file like below to show absolute
percentage of filtered entries by default:
$ cat ~/.perfconfig
[hist]
percentage = absolute
And it can be changed through command line:
$ perf report --percentage relative
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
Link: http://lkml.kernel.org/r/1397145720-8063-6-git-send-email-namhyung@kernel.org
Signed-off-by: Jiri Olsa <jolsa@redhat.com>
The --percentage option is for controlling overhead percentage
displayed. It can only receive either of "relative" or "absolute".
Move the parser callback function into a common location since it's
used by multiple commands now.
For more information, please see previous commit same thing done to
"perf report".
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
Link: http://lkml.kernel.org/r/1397145720-8063-4-git-send-email-namhyung@kernel.org
Signed-off-by: Jiri Olsa <jolsa@redhat.com>
The --percentage option is for controlling overhead percentage
displayed. It can only receive either of "relative" or "absolute".
"relative" means it's relative to filtered entries only so that the
sum of shown entries will be always 100%. "absolute" means it retains
the original value before and after the filter is applied.
$ perf report -s comm
# Overhead Command
# ........ ............
#
74.19% cc1
7.61% gcc
6.11% as
4.35% sh
4.14% make
1.13% fixdep
...
$ perf report -s comm -c cc1,gcc --percentage absolute
# Overhead Command
# ........ ............
#
74.19% cc1
7.61% gcc
$ perf report -s comm -c cc1,gcc --percentage relative
# Overhead Command
# ........ ............
#
90.69% cc1
9.31% gcc
Note that it has zero effect if no filter was applied.
Suggested-by: Arnaldo Carvalho de Melo <acme@ghostprotocols.net>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
Link: http://lkml.kernel.org/r/1397145720-8063-3-git-send-email-namhyung@kernel.org
Signed-off-by: Jiri Olsa <jolsa@redhat.com>
When filtering by thread, dso or symbol on TUI it also update total
period so that the output shows different result than no filter - the
percentage changed to relative to filtered entries only. Sometimes
this is not desired since users might expect same results with filter.
So new filtered_* fields to hists->stats to count them separately.
They'll be controlled/used by user later.
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
Link: http://lkml.kernel.org/r/1397145720-8063-2-git-send-email-namhyung@kernel.org
Signed-off-by: Jiri Olsa <jolsa@redhat.com>
