It would be useful for perf to support a mode to query the inline stack
for a given callgraph address. This would simplify finding the right
code in code that does a lot of inlining.
The srcline.c has contained the code which supports to translate the
address to filename:line_nr. This patch just extends the function to let
it support getting the inline stacks.
It introduces the inline_list which will store the inline function
result (filename:line_nr and funcname).
If BFD lib is not supported, the result is only filename:line_nr.
Signed-off-by: Yao Jin <yao.jin@linux.intel.com>
Tested-by: Milian Wolff <milian.wolff@kdab.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Jiri Olsa <jolsa@kernel.org>
Cc: Kan Liang <kan.liang@intel.com>
Link: http://lkml.kernel.org/r/1490474069-15823-3-git-send-email-yao.jin@linux.intel.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Introduce dso__name() and filename_split() out of existing code because
these codes will be used in several places in next patch.
For filename_split(), it may also solve a potential memory leak in
existing code. In existing addr2line(),
sep = strchr(filename, ':');
if (sep) {
*sep++ = '\0';
*file = filename;
*line_nr = strtoul(sep, NULL, 0);
ret = 1;
}
out:
pclose(fp);
return ret;
If sep is NULL, filename is not freed or returned via file.
Signed-off-by: Yao Jin <yao.jin@linux.intel.com>
Tested-by: Milian Wolff <milian.wolff@kdab.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Jiri Olsa <jolsa@kernel.org>
Cc: Kan Liang <kan.liang@intel.com>
Link: http://lkml.kernel.org/r/1490474069-15823-2-git-send-email-yao.jin@linux.intel.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
This patch fix some spelling typos found in printk.
[jkosina@suse.cz: drop arch/arm64/kernel/hibernate.c that was already
in place]
Signed-off-by: Masanari Iida <standby24x7@gmail.com>
Acked-by: Randy Dunlap <rdunlap@infradead.org>
Signed-off-by: Jiri Kosina <jkosina@suse.cz>
Move the printing of perf expressions and internal events to a new
clearer --details flag, instead of lumping it together with other debug
options in --debug. This makes it clearer to use.
Before
perf list --debug
...
unc_m_power_critical_throttle_cycles
[Cycles all ranks are in critical thermal throttle. Unit: uncore_imc]
uncore_imc_2/event=0x86/ MetricName: power_critical_throttle_cycles % MetricExpr: (unc_m_power_critical_throttle_cycles / unc_m_clockticks) * 100.
after
perf list --details
...
unc_m_power_critical_throttle_cycles
[Cycles all ranks are in critical thermal throttle. Unit: uncore_imc]
uncore_imc_2/event=0x86/ MetricName: power_critical_throttle_cycles % MetricExpr: (unc_m_power_critical_throttle_cycles / unc_m_clockticks) * 100.
Signed-off-by: Andi Kleen <ak@linux.intel.com>
Acked-by: Jiri Olsa <jolsa@kernel.org>
Tested-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Link: http://lkml.kernel.org/r/20170320201711.14142-14-andi@firstfloor.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Add support for a new JSON event attribute to name MetricExpr for better
output in perf stat.
If the event has no MetricName it uses the normal event name instead to
describe the metric.
Before
% perf stat -a -I 1000 -e '{unc_p_clockticks,unc_p_freq_max_os_cycles}' --metric-only
time unc_p_freq_max_os_cycles
1.000149775 15.7
2.000344807 19.3
3.000502544 16.7
4.000640656 6.6
5.000779955 9.9
After
% perf stat -a -I 1000 -e '{unc_p_clockticks,unc_p_freq_max_os_cycles}' --metric-only
time freq_max_os_cycles %
1.000149775 15.7
2.000344807 19.3
3.000502544 16.7
4.000640656 6.6
5.000779955 9.9
Signed-off-by: Andi Kleen <ak@linux.intel.com>
Acked-by: Jiri Olsa <jolsa@kernel.org>
Link: http://lkml.kernel.org/r/20170320201711.14142-13-andi@firstfloor.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Output the metric expr in perf list when --debug is specified, so that
the user can check the formula.
Before:
% perf list
...
unc_m_power_channel_ppd
[Cycles where DRAM ranks are in power down (CKE) mode. Derived from unc_m_power_channel_ppd. Unit:
uncore_imc]
uncore_imc_2/event=0x85/
After:
% perf list --debug
...
unc_m_power_channel_ppd
[Cycles where DRAM ranks are in power down (CKE) mode. Derived from unc_m_power_channel_ppd. Unit:
uncore_imc]
Perf: uncore_imc_2/event=0x85/ MetricExpr: (unc_m_power_channel_ppd / unc_m_clockticks) * 100.
Signed-off-by: Andi Kleen <ak@linux.intel.com>
Acked-by: Jiri Olsa <jolsa@kernel.org>
Link: http://lkml.kernel.org/r/20170320201711.14142-12-andi@firstfloor.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Add generic infrastructure to perf stat to output ratios for
"MetricExpr" entries in the event lists. Many events are more useful as
ratios than in raw form, typically some count in relation to total
ticks.
Transfer the MetricExpr information from the alias to the evsel.
We mark the events that need to be collected for MetricExpr, and also
link the events using them with a pointer. The code is careful to always
prefer the right event in the same group to minimize multiplexing
errors. At the moment only a single relation is supported.
Then add a rblist to the stat shadow code that remembers stats based on
the cpu and context.
Then finally update and retrieve and print these values similarly to the
existing hardcoded perf metrics. We use the simple expression parser
added earlier to evaluate the expression.
Normally we just output the result without further commentary, but for
--metric-only this would lead to empty columns. So for this case use the
original event as description.
There is no attempt to automatically add the MetricExpr event, if it is
missing, however we suggest it to the user, because the user tool
doesn't have enough information to reliably construct a group that is
guaranteed to schedule. So we leave that to the user.
% perf stat -a -I 1000 -e '{unc_p_clockticks,unc_p_freq_max_os_cycles}'
1.000147889 800,085,181 unc_p_clockticks
1.000147889 93,126,241 unc_p_freq_max_os_cycles # 11.6
2.000448381 800,218,217 unc_p_clockticks
2.000448381 142,516,095 unc_p_freq_max_os_cycles # 17.8
3.000639852 800,243,057 unc_p_clockticks
3.000639852 162,292,689 unc_p_freq_max_os_cycles # 20.3
% perf stat -a -I 1000 -e '{unc_p_clockticks,unc_p_freq_max_os_cycles}' --metric-only
# time freq_max_os_cycles %
1.000127077 0.9
2.000301436 0.7
3.000456379 0.0
v2: Change from DivideBy to MetricExpr
v3: Use expr__ prefix. Support more than one other event.
v4: Update description
v5: Only print warning message once for multiple PMUs.
Signed-off-by: Andi Kleen <ak@linux.intel.com>
Acked-by: Jiri Olsa <jolsa@kernel.org>
Link: http://lkml.kernel.org/r/20170320201711.14142-11-andi@firstfloor.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Add a simple expression parser good enough to parse JSON relation
expressions. The parser is implemented using bison.
This is just intended as an simple parser for internal usage in the
event lists, not the beginning of a "perf scripting language"
v2: Use expr__ prefix instead of expr_
Support multiple free variables for parser
Committer note:
The v2 patch had:
%define api.pure full
In expr.y, that is a feature introduced in bison 2.7, to have reentrant
parsers, not using global variables, which would make tools/perf stop
building with the bison version shipped in older distros, so Andi
realised that the other parsers (e.g. parse-events.y) were using:
%pure-parser
Which is present in older versions of bison and fits the bill.
I added:
CFLAGS_expr-bison.o += -DYYENABLE_NLS=0 -DYYLTYPE_IS_TRIVIAL=0 -w
To finally make it build, copying what was there for pmu-bison.o,
another parser.
Signed-off-by: Andi Kleen <ak@linux.intel.com>
Acked-by: Jiri Olsa <jolsa@kernel.org>
Link: http://lkml.kernel.org/r/20170320201711.14142-8-andi@firstfloor.org
[ stdlib.h is needed in tests/expr.c for free() fixing build in systems such as ubuntu:16.04-x-s390 ]
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Special case uncore_ prefix in PMU match, to allow for shorter event
uncore specifications.
Before:
perf stat -a -e uncore_cbox/event=0x35,umask=0x1,filter_opc=0x19C/ sleep 1
After
perf stat -a -e cbox/event=0x35,umask=0x1,filter_opc=0x19C/ sleep 1
Committer tests:
# perf list uncore
List of pre-defined events (to be used in -e):
uncore_cbox_0/clockticks/ [Kernel PMU event]
uncore_cbox_1/clockticks/ [Kernel PMU event]
uncore_imc/data_reads/ [Kernel PMU event]
uncore_imc/data_writes/ [Kernel PMU event]
# perf stat -a -e cbox_0/clockticks/ sleep 1
Performance counter stats for 'system wide':
281,474,976,653,084 cbox_0/clockticks/
1.000870129 seconds time elapsed
#
Signed-off-by: Andi Kleen <ak@linux.intel.com>
Acked-by: Jiri Olsa <jolsa@kernel.org>
Tested-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Link: http://lkml.kernel.org/r/20170320201711.14142-7-andi@firstfloor.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
When the user specifies a pmu directly, expand it automatically with a
prefix match for all available PMUs, similar as we do for the normal
aliases now.
This allows to specify attributes for duplicated boxes quickly. For
example uncore_cbox_{0,6}/.../ can be now specified as uncore_cbox/.../
and it gets automatically expanded for all boxes.
This generally makes it more concise to write uncore specifications, and
also avoids the need to know the exact topology of the system.
Before:
% perf stat -a -e uncore_cbox_0/event=0x35,umask=0x1,filter_opc=0x19C/,\
uncore_cbox_1/event=0x35,umask=0x1,filter_opc=0x19C/,\
uncore_cbox_2/event=0x35,umask=0x1,filter_opc=0x19C/,\
uncore_cbox_3/event=0x35,umask=0x1,filter_opc=0x19C/,\
uncore_cbox_4/event=0x35,umask=0x1,filter_opc=0x19C/,\
uncore_cbox_5/event=0x35,umask=0x1,filter_opc=0x19C/ sleep 1
After:
% perf stat -a -e uncore_cbox/event=0x35,umask=0x1,filter_opc=0x19C/ sleep 1
v2: Handle all bison rules. Move multi add code to separate function.
Handle uncore_ prefix correctly.
v3: Move parse_events_multi_pmu_add to separate patch. Move uncore
prefix check to separate patch.
Signed-off-by: Andi Kleen <ak@linux.intel.com>
Acked-by: Jiri Olsa <jolsa@kernel.org>
Link: http://lkml.kernel.org/r/20170320201711.14142-6-andi@firstfloor.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
The uncore PMU has a lot of duplicated PMUs for different subsystems.
When expanding an uncore alias we usually end up with a large
number of identically named aliases, which makes perf stat
output difficult to read.
Automatically sum them up in perf stat, unless --no-merge is specified.
This can be default because only the uncores generally have duplicated
aliases. Other PMUs have unique names.
Before:
% perf stat --no-merge -a -e unc_c_llc_lookup.any sleep 1
Performance counter stats for 'system wide':
694,976 Bytes unc_c_llc_lookup.any
706,304 Bytes unc_c_llc_lookup.any
956,608 Bytes unc_c_llc_lookup.any
782,720 Bytes unc_c_llc_lookup.any
605,696 Bytes unc_c_llc_lookup.any
442,816 Bytes unc_c_llc_lookup.any
659,328 Bytes unc_c_llc_lookup.any
509,312 Bytes unc_c_llc_lookup.any
263,936 Bytes unc_c_llc_lookup.any
592,448 Bytes unc_c_llc_lookup.any
672,448 Bytes unc_c_llc_lookup.any
608,640 Bytes unc_c_llc_lookup.any
641,024 Bytes unc_c_llc_lookup.any
856,896 Bytes unc_c_llc_lookup.any
808,832 Bytes unc_c_llc_lookup.any
684,864 Bytes unc_c_llc_lookup.any
710,464 Bytes unc_c_llc_lookup.any
538,304 Bytes unc_c_llc_lookup.any
1.002577660 seconds time elapsed
After:
% perf stat -a -e unc_c_llc_lookup.any sleep 1
Performance counter stats for 'system wide':
2,685,120 Bytes unc_c_llc_lookup.any
1.002648032 seconds time elapsed
v2: Split collect_aliases. Rename alias flag.
v3: Make sure unsupported/not counted is always printed.
v4: Factor out callback change into separate patch.
v5: Move check for bad results here
Move merged check into collect_data
Signed-off-by: Andi Kleen <ak@linux.intel.com>
Acked-by: Jiri Olsa <jolsa@kernel.org>
Link: http://lkml.kernel.org/r/20170320201711.14142-3-andi@firstfloor.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
The 'grep -v "filename"' applied to the objdump command output cause a
side effect eliminating filename:linenr of output of 'objdump -l' if the
object file name and source file name are the same, fix it.
E.g. the output of the following objdump command in symbol__disassemble():
$ objdump -l -d -S -C /home/taeung/hello --start-address=...
/home/taeung/hello: file format elf64-x86-64
Disassembly of section .text:
0000000000400526 <main>:
main():
/home/taeung/hello.c:4
void main()
{
400526: 55 push %rbp
400527: 48 89 e5 mov %rsp,%rbp
/home/taeung/hello.c:5
...
But it uses grep -v "filename" e.g. "/home/taeung/hello" in the objdump
command to remove the first line containing file name and file format
("/home/taeung/hello: file format elf64-x86-64"):
Before:
$ objdump -l -d -S -C /home/taeung/hello | grep /home/taeung/hello
But this causes a side effect, removing filename:linenr too, because the
object file and source file have the same name e.g. "/home/taueng/hello",
"/home/taeung/hello.c"
So more do a better match by using grep -v as below to correctly remove
that first line:
"/home/taeung/hello: file format elf64-x86-64"
After:
$ objdump -l -d -S -C /home/taeung/hello | grep /home/taeung/hello:
Signed-off-by: Taeung Song <treeze.taeung@gmail.com>
Tested-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Jiri Olsa <jolsa@kernel.org>
Cc: Masami Hiramatsu <mhiramat@kernel.org>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Wang Nan <wangnan0@huawei.com>
Link: http://lkml.kernel.org/r/1489978617-31396-5-git-send-email-treeze.taeung@gmail.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
An sdt probe can be associated with arguments but they were not passed
to the user probe tracing interface (uprobe_events); this patch adapts
the sdt argument descriptors according to the uprobe input format.
As the uprobe parser does not support scaled address mode, perf will
skip arguments which cannot be adapted to the uprobe format.
Here are the results:
$ perf buildid-cache -v --add test_sdt
$ perf probe -x test_sdt sdt_libfoo:table_frob
$ perf probe -x test_sdt sdt_libfoo:table_diddle
$ perf record -e sdt_libfoo:table_frob -e sdt_libfoo:table_diddle test_sdt
$ perf script
test_sdt ... 666.255678: sdt_libfoo:table_frob: (4004d7) arg0=0 arg1=0
test_sdt ... 666.255683: sdt_libfoo:table_diddle: (40051a) arg0=0 arg1=0
test_sdt ... 666.255686: sdt_libfoo:table_frob: (4004d7) arg0=1 arg1=2
test_sdt ... 666.255689: sdt_libfoo:table_diddle: (40051a) arg0=3 arg1=4
test_sdt ... 666.255692: sdt_libfoo:table_frob: (4004d7) arg0=2 arg1=4
test_sdt ... 666.255694: sdt_libfoo:table_diddle: (40051a) arg0=6 arg1=8
Signed-off-by: Alexis Berlemont <alexis.berlemont@gmail.com>
Acked-by: Masami Hiramatsu <mhiramat@kernel.org>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Hemant Kumar <hemant@linux.vnet.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Ravi Bangoria <ravi.bangoria@linux.vnet.ibm.com>
Link: http://lkml.kernel.org/r/20161214000732.1710-3-alexis.berlemont@gmail.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
On old perf, when using 'perf probe -d' to delete an inexistent event,
it returns errno, eg,
-bash-4.3# perf probe -d xxx || echo $?
Info: Event "*:xxx" does not exist.
Error: Failed to delete events.
255
But now perf_del_probe_events() will always set ret = 0, different from
previous del_perf_probe_events(). After this, it returns errno again,
eg,
-bash-4.3# ./perf probe -d xxx || echo $?
"xxx" does not hit any event.
Error: Failed to delete events.
254
And it is more appropriate to return -ENOENT instead of -EPERM.
Signed-off-by: Kefeng Wang <wangkefeng.wang@huawei.com>
Acked-by: Masami Hiramatsu <mhiramat@kernel.org>
Cc: Hanjun Guo <guohanjun@huawei.com>
Cc: Jiri Olsa <jolsa@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Wang Nan <wangnan0@huawei.com>
Fixes: dddc7ee32f ("perf probe: Fix an error when deleting probes successfully")
Link: http://lkml.kernel.org/r/1489738592-61011-1-git-send-email-wangkefeng.wang@huawei.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
There are many SDT markers in powerpc whose uprobe definition goes
beyond current MAX_CMDLEN, especially when target filename is long and
sdt marker has long list of arguments. For example, definition of sdt
marker
method__compile__end: 8@17 8@9 8@10 -4@8 8@7 -4@6 8@5 -4@4 1@37(28)
from file
/usr/lib/jvm/java-1.8.0-openjdk-1.8.0.91-2.b14.fc22.ppc64/jre/lib/ppc64/server/libjvm.so
is
p:sdt_hotspot/method__compile__end /usr/lib/jvm/java-1.8.0-openjdk-\
1.8.0.91-2.b14.fc22.ppc64/jre/lib/ppc64/server/libjvm.so:0x4c4e00\
arg1=%gpr17:u64 arg2=%gpr9:u64 arg3=%gpr10:u64 arg4=%gpr8:s32\
arg5=%gpr7:u64 arg6=%gpr6:s32 arg7=%gpr5:u64 arg8=%gpr4:s32\
arg9=+37(%gpr28):u8
'perf probe' fails with segfault for such markers. As the uprobe_events
file accepts definitions up to 4094 characters(4096 - 2 (\n\0)),
increase value of MAX_CMDLEN match that.
Signed-off-by: Ravi Bangoria <ravi.bangoria@linux.vnet.ibm.com>
Acked-by: Masami Hiramatsu <mhiramat@kernel.org>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Alexis Berlemont <alexis.berlemont@gmail.com>
Cc: Madhavan Srinivasan <maddy@linux.vnet.ibm.com>
Cc: Michael Ellerman <mpe@ellerman.id.au>
Cc: Naveen N. Rao <naveen.n.rao@linux.vnet.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Link: http://lkml.kernel.org/r/20170207054547.3690-1-ravi.bangoria@linux.vnet.ibm.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Pull perf fixes from Thomas Gleixner:
"A set of perf related fixes:
- fix a CR4.PCE propagation issue caused by usage of mm instead of
active_mm and therefore propagated the wrong value.
- perf core fixes, which plug a use-after-free issue and make the
event inheritance on fork more robust.
- a tooling fix for symbol handling"
* 'perf-urgent-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip:
perf symbols: Fix symbols__fixup_end heuristic for corner cases
x86/perf: Clarify why x86_pmu_event_mapped() isn't racy
x86/perf: Fix CR4.PCE propagation to use active_mm instead of mm
perf/core: Better explain the inherit magic
perf/core: Simplify perf_event_free_task()
perf/core: Fix event inheritance on fork()
perf/core: Fix use-after-free in perf_release()
This patch decodes the 'partial' flag in AUX records and prints
a warning to the user, so that they don't have to guess why their
PT traces contain gaps (or missing altogether):
Warning:
AUX data had gaps in it 8 times out of 8!
Are you running a KVM guest in the background?
Trying to be even more helpful, we will detect if the user's kvm driver sets up
exclusive VMX root mode for the entire lifespan of the kvm process:
Reloading kvm_intel module with vmm_exclusive=0
will reduce the gaps to only guest's timeslices.
Note however, that you'll still have gaps in cpu-wide traces even with
vmm_exclusive=0, but the number of gaps will be below 100% (as opposed to the
above example).
Currently this is the only reason for partial records.
Signed-off-by: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Stephane Eranian <eranian@google.com>
Cc: Vince Weaver <vince@deater.net>
Link: http://lkml.kernel.org/r/8760j941ig.fsf@ashishki-desk.ger.corp.intel.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
When I use -i option for report subcommand, it doesn't accept it. We
need add common options using OPT_PARENT macro.
perf lock report -i lock_perf.data
Error: unknown switch `i'
Usage: perf lock report [<options>]
-f, --force don't complain, do it
-k, --key <acquired> key for sorting ...
Signed-off-by: Changbin Du <changbin.du@intel.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Link: http://lkml.kernel.org/r/20170317055342.8284-1-changbin.du@intel.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
The current symbols__fixup_end() heuristic for the last entry in the rb
tree is suboptimal as it leads to not being able to recognize the symbol
in the call graph in a couple of corner cases, for example:
i) If the symbol has a start address (f.e. exposed via kallsyms)
that is at a page boundary, then the roundup(curr->start, 4096)
for the last entry will result in curr->start == curr->end with
a symbol length of zero.
ii) If the symbol has a start address that is shortly before a page
boundary, then also here, curr->end - curr->start will just be
very few bytes, where it's unrealistic that we could perform a
match against.
Instead, change the heuristic to roundup(curr->start, 4096) + 4096, so
that we can catch such corner cases and have a better chance to find
that specific symbol. It's still just best effort as the real end of the
symbol is unknown to us (and could even be at a larger offset than the
current range), but better than the current situation.
Alexei reported that he recently run into case i) with a JITed eBPF
program (these are all page aligned) as the last symbol which wasn't
properly shown in the call graph (while other eBPF program symbols in
the rb tree were displayed correctly). Since this is a generic issue,
lets try to improve the heuristic a bit.
Reported-and-Tested-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Fixes: 2e538c4a18 ("perf tools: Improve kernel/modules symbol lookup")
Link: http://lkml.kernel.org/r/bb5c80d27743be6f12afc68405f1956a330e1bc9.1489614365.git.daniel@iogearbox.net
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Pull perf/core improvements and fixes from Arnaldo Carvalho de Melo:
New features:
- Add 'brstackinsn' field in 'perf script' to reuse the x86 instruction
decoder used in the Intel PT code to study hot paths to samples (Andi Kleen)
Kernel changes:
- Default UPROBES_EVENTS to Y (Alexei Starovoitov)
- Fix check for kretprobe offset within function entry (Naveen N. Rao)
Infrastructure changes:
- Introduce util func is_sdt_event() (Ravi Bangoria)
- Make perf_event__synthesize_mmap_events() scale on older kernels where
reading /proc/pid/maps is way slower than reading /proc/pid/task/pid/maps (Stephane Eranian)
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Implement printing instruction sequences as hex dump for branch stacks.
This relies on the x86 instruction decoder used by the PT decoder to
find the lengths of instructions to dump them individually.
This is good enough for pattern matching.
This allows to study hot paths for individual samples, together with
branch misprediction and cycle count / IPC information if available (on
Skylake systems).
% perf record -b ...
% perf script -F brstackinsn
...
read_hpet+67:
ffffffff9905b843 insn: 74 ea # PRED
ffffffff9905b82f insn: 85 c9
ffffffff9905b831 insn: 74 12
ffffffff9905b833 insn: f3 90
ffffffff9905b835 insn: 48 8b 0f
ffffffff9905b838 insn: 48 89 ca
ffffffff9905b83b insn: 48 c1 ea 20
ffffffff9905b83f insn: 39 f2
ffffffff9905b841 insn: 89 d0
ffffffff9905b843 insn: 74 ea # PRED
Only works when no special branch filters are specified.
Occasionally the path does not reach up to the sample IP, as the LBRs
may be frozen before executing a final jump. In this case we print a
special message.
The instruction dumper piggy backs on the existing infrastructure from
the IP PT decoder.
An earlier iteration of this patch relied on a disassembler, but this
version only uses the existing instruction decoder.
Committer note:
Added hint about how to get suitable perf.data files for use with
'-F brstackinsm':
$ perf record usleep 1
[ perf record: Woken up 1 times to write data ]
[ perf record: Captured and wrote 0.018 MB perf.data (8 samples) ]
$
$ perf script -F brstackinsn
Display of branch stack assembler requested, but non all-branch filter set
Hint: run 'perf record -b ...'
$
Signed-off-by: Andi Kleen <ak@linux.intel.com>
Tested-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Jiri Olsa <jolsa@kernel.org>
Cc: Michael Ellerman <mpe@ellerman.id.au>
Link: http://lkml.kernel.org/r/20170223234634.583-1-andi@firstfloor.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
This patch significantly improves the execution time of
perf_event__synthesize_mmap_events() when running perf record on systems
where processes have lots of threads.
It just happens that cat /proc/pid/maps support uses a O(N^2) algorithm to
generate each map line in the maps file. If you have 1000 threads, then you
have necessarily 1000 stacks. For each vma, you need to check if it
corresponds to a thread's stack. With a large number of threads, this can take
a very long time. I have seen latencies >> 10mn.
As of today, perf does not use the fact that a mapping is a stack, therefore we
can work around the issue by using /proc/pid/tasks/pid/maps. This entry does
not try to map a vma to stack and is thus much faster with no loss of
functonality.
The proc-map-timeout logic is kept in case users still want some upper limit.
In V2, we fix the file path from /proc/pid/tasks/pid/maps to actual
/proc/pid/task/pid/maps, tasks -> task. Thanks Arnaldo for catching this.
Committer note:
This problem seems to have been elliminated in the kernel since commit :
b18cb64ead ("fs/proc: Stop trying to report thread stacks").
Signed-off-by: Stephane Eranian <eranian@google.com>
Acked-by: Jiri Olsa <jolsa@redhat.com>
Cc: Andy Lutomirski <luto@kernel.org>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Link: http://lkml.kernel.org/r/20170315135059.GC2177@redhat.com
Link: http://lkml.kernel.org/r/1489598233-25586-1-git-send-email-eranian@google.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
We indicate support for accepting sym+offset with kretprobes through a
line in ftrace README. Parse the same to identify support and choose the
appropriate format for kprobe_events.
As an example, without this perf patch, but with the ftrace changes:
naveen@ubuntu:~/linux/tools/perf$ sudo cat /sys/kernel/debug/tracing/README | grep kretprobe
place (kretprobe): [<module>:]<symbol>[+<offset>]|<memaddr>
naveen@ubuntu:~/linux/tools/perf$
naveen@ubuntu:~/linux/tools/perf$ sudo ./perf probe -v do_open%return
probe-definition(0): do_open%return
symbol:do_open file:(null) line:0 offset:0 return:1 lazy:(null)
0 arguments
Looking at the vmlinux_path (8 entries long)
Using /boot/vmlinux for symbols
Open Debuginfo file: /boot/vmlinux
Try to find probe point from debuginfo.
Matched function: do_open [2d0c7d8]
Probe point found: do_open+0
Matched function: do_open [35d76b5]
found inline addr: 0xc0000000004ba984
Failed to find "do_open%return",
because do_open is an inlined function and has no return point.
An error occurred in debuginfo analysis (-22).
Trying to use symbols.
Opening /sys/kernel/debug/tracing//kprobe_events write=1
Writing event: r:probe/do_open do_open+0
Writing event: r:probe/do_open_1 do_open+0
Added new events:
probe:do_open (on do_open%return)
probe:do_open_1 (on do_open%return)
You can now use it in all perf tools, such as:
perf record -e probe:do_open_1 -aR sleep 1
naveen@ubuntu:~/linux/tools/perf$ sudo cat /sys/kernel/debug/kprobes/list
c000000000041370 k kretprobe_trampoline+0x0 [OPTIMIZED]
c0000000004433d0 r do_open+0x0 [DISABLED]
c0000000004433d0 r do_open+0x0 [DISABLED]
And after this patch (and the subsequent powerpc patch):
naveen@ubuntu:~/linux/tools/perf$ sudo ./perf probe -v do_open%return
probe-definition(0): do_open%return
symbol:do_open file:(null) line:0 offset:0 return:1 lazy:(null)
0 arguments
Looking at the vmlinux_path (8 entries long)
Using /boot/vmlinux for symbols
Open Debuginfo file: /boot/vmlinux
Try to find probe point from debuginfo.
Matched function: do_open [2d0c7d8]
Probe point found: do_open+0
Matched function: do_open [35d76b5]
found inline addr: 0xc0000000004ba984
Failed to find "do_open%return",
because do_open is an inlined function and has no return point.
An error occurred in debuginfo analysis (-22).
Trying to use symbols.
Opening /sys/kernel/debug/tracing//README write=0
Opening /sys/kernel/debug/tracing//kprobe_events write=1
Writing event: r:probe/do_open _text+4469712
Writing event: r:probe/do_open_1 _text+4956248
Added new events:
probe:do_open (on do_open%return)
probe:do_open_1 (on do_open%return)
You can now use it in all perf tools, such as:
perf record -e probe:do_open_1 -aR sleep 1
naveen@ubuntu:~/linux/tools/perf$ sudo cat /sys/kernel/debug/kprobes/list
c000000000041370 k kretprobe_trampoline+0x0 [OPTIMIZED]
c0000000004433d0 r do_open+0x0 [DISABLED]
c0000000004ba058 r do_open+0x8 [DISABLED]
Signed-off-by: Naveen N. Rao <naveen.n.rao@linux.vnet.ibm.com>
Acked-by: Masami Hiramatsu <mhiramat@kernel.org>
Cc: Ananth N Mavinakayanahalli <ananth@linux.vnet.ibm.com>
Cc: Michael Ellerman <mpe@ellerman.id.au>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: linuxppc-dev@lists.ozlabs.org
Link: http://lkml.kernel.org/r/496ef9f33c1ab16286ece9dd62aa672807aef91c.1488961018.git.naveen.n.rao@linux.vnet.ibm.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
This patch introduces a cgroup identifier entry field in perf report to
identify or distinguish data of different cgroups. It uses the device
number and inode number of cgroup namespace, included in perf data with
the new PERF_RECORD_NAMESPACES event, as cgroup identifier.
With the assumption that each container is created with it's own cgroup
namespace, this allows assessment/analysis of multiple containers at
once.
A simple test for this would be to clone a few processes passing
SIGCHILD & CLONE_NEWCROUP flags to each of them, execute shell and run
different workloads on each of those contexts, while running perf
record command with --namespaces option.
Shown below is the output of perf report, sorted with cgroup identifier,
on perf.data generated with the above test scenario, clearly indicating
one context's considerable use of kernel memory in comparison with
others:
$ perf report -s cgroup_id,sample --stdio
#
# Total Lost Samples: 0
#
# Samples: 5K of event 'kmem:kmalloc'
# Event count (approx.): 5965
#
# Overhead cgroup id (dev/inode) Samples
# ........ ..................... ............
#
81.27% 3/0xeffffffb 4848
16.24% 3/0xf00000d0 969
1.16% 3/0xf00000ce 69
0.82% 3/0xf00000cf 49
0.50% 0/0x0 30
While this is a start, there is further scope of improving this. For
example, instead of cgroup namespace's device and inode numbers, dev
and inode numbers of some or all namespaces may be used to distinguish
which processes are running in a given container context.
Also, scripts to map device and inode info to containers sounds
plausible for better tracing of containers.
Signed-off-by: Hari Bathini <hbathini@linux.vnet.ibm.com>
Tested-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Alexei Starovoitov <ast@fb.com>
Cc: Ananth N Mavinakayanahalli <ananth@linux.vnet.ibm.com>
Cc: Aravinda Prasad <aravinda@linux.vnet.ibm.com>
Cc: Brendan Gregg <brendan.d.gregg@gmail.com>
Cc: Daniel Borkmann <daniel@iogearbox.net>
Cc: Eric Biederman <ebiederm@xmission.com>
Cc: Jiri Olsa <jolsa@redhat.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Sargun Dhillon <sargun@sargun.me>
Cc: Steven Rostedt <rostedt@goodmis.org>
Link: http://lkml.kernel.org/r/148891933338.25309.756882900782042645.stgit@hbathini.in.ibm.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
