Age | Commit message (Collapse) | Author | Files | Lines |
|
When perf data is in a pipe, it reads each event separately using
read(2) syscall. This is a huge performance bottleneck when
processing large data like in perf inject. Also perf inject needs to
use write(2) syscall for the output.
So convert it to use buffer I/O functions in stdio library for pipe
data. This makes inject-build-id bench time drops from 20ms to 8ms.
$ perf bench internals inject-build-id
# Running 'internals/inject-build-id' benchmark:
Average build-id injection took: 8.074 msec (+- 0.013 msec)
Average time per event: 0.792 usec (+- 0.001 usec)
Average memory usage: 8328 KB (+- 0 KB)
Average build-id-all injection took: 5.490 msec (+- 0.008 msec)
Average time per event: 0.538 usec (+- 0.001 usec)
Average memory usage: 7563 KB (+- 0 KB)
This patch enables it just for perf inject when used with pipe (it's a
default behavior). Maybe we could do it for perf record and/or report
later..
Committer testing:
Before:
$ perf stat -r 5 perf bench internals inject-build-id
# Running 'internals/inject-build-id' benchmark:
Average build-id injection took: 13.605 msec (+- 0.064 msec)
Average time per event: 1.334 usec (+- 0.006 usec)
Average memory usage: 12220 KB (+- 7 KB)
Average build-id-all injection took: 11.458 msec (+- 0.058 msec)
Average time per event: 1.123 usec (+- 0.006 usec)
Average memory usage: 11546 KB (+- 8 KB)
# Running 'internals/inject-build-id' benchmark:
Average build-id injection took: 13.673 msec (+- 0.057 msec)
Average time per event: 1.341 usec (+- 0.006 usec)
Average memory usage: 12508 KB (+- 8 KB)
Average build-id-all injection took: 11.437 msec (+- 0.046 msec)
Average time per event: 1.121 usec (+- 0.004 usec)
Average memory usage: 11812 KB (+- 7 KB)
# Running 'internals/inject-build-id' benchmark:
Average build-id injection took: 13.641 msec (+- 0.069 msec)
Average time per event: 1.337 usec (+- 0.007 usec)
Average memory usage: 12302 KB (+- 8 KB)
Average build-id-all injection took: 10.820 msec (+- 0.106 msec)
Average time per event: 1.061 usec (+- 0.010 usec)
Average memory usage: 11616 KB (+- 7 KB)
# Running 'internals/inject-build-id' benchmark:
Average build-id injection took: 13.379 msec (+- 0.074 msec)
Average time per event: 1.312 usec (+- 0.007 usec)
Average memory usage: 12334 KB (+- 8 KB)
Average build-id-all injection took: 11.288 msec (+- 0.071 msec)
Average time per event: 1.107 usec (+- 0.007 usec)
Average memory usage: 11657 KB (+- 8 KB)
# Running 'internals/inject-build-id' benchmark:
Average build-id injection took: 13.534 msec (+- 0.058 msec)
Average time per event: 1.327 usec (+- 0.006 usec)
Average memory usage: 12264 KB (+- 8 KB)
Average build-id-all injection took: 11.557 msec (+- 0.076 msec)
Average time per event: 1.133 usec (+- 0.007 usec)
Average memory usage: 11593 KB (+- 8 KB)
Performance counter stats for 'perf bench internals inject-build-id' (5 runs):
4,060.05 msec task-clock:u # 1.566 CPUs utilized ( +- 0.65% )
0 context-switches:u # 0.000 K/sec
0 cpu-migrations:u # 0.000 K/sec
101,888 page-faults:u # 0.025 M/sec ( +- 0.12% )
3,745,833,163 cycles:u # 0.923 GHz ( +- 0.10% ) (83.22%)
194,346,613 stalled-cycles-frontend:u # 5.19% frontend cycles idle ( +- 0.57% ) (83.30%)
708,495,034 stalled-cycles-backend:u # 18.91% backend cycles idle ( +- 0.48% ) (83.48%)
5,629,328,628 instructions:u # 1.50 insn per cycle
# 0.13 stalled cycles per insn ( +- 0.21% ) (83.57%)
1,236,697,927 branches:u # 304.602 M/sec ( +- 0.16% ) (83.44%)
17,564,877 branch-misses:u # 1.42% of all branches ( +- 0.23% ) (82.99%)
2.5934 +- 0.0128 seconds time elapsed ( +- 0.49% )
$
After:
$ perf stat -r 5 perf bench internals inject-build-id
# Running 'internals/inject-build-id' benchmark:
Average build-id injection took: 8.560 msec (+- 0.125 msec)
Average time per event: 0.839 usec (+- 0.012 usec)
Average memory usage: 12520 KB (+- 8 KB)
Average build-id-all injection took: 5.789 msec (+- 0.054 msec)
Average time per event: 0.568 usec (+- 0.005 usec)
Average memory usage: 11919 KB (+- 9 KB)
# Running 'internals/inject-build-id' benchmark:
Average build-id injection took: 8.639 msec (+- 0.111 msec)
Average time per event: 0.847 usec (+- 0.011 usec)
Average memory usage: 12732 KB (+- 8 KB)
Average build-id-all injection took: 5.647 msec (+- 0.069 msec)
Average time per event: 0.554 usec (+- 0.007 usec)
Average memory usage: 12093 KB (+- 7 KB)
# Running 'internals/inject-build-id' benchmark:
Average build-id injection took: 8.551 msec (+- 0.096 msec)
Average time per event: 0.838 usec (+- 0.009 usec)
Average memory usage: 12739 KB (+- 8 KB)
Average build-id-all injection took: 5.617 msec (+- 0.061 msec)
Average time per event: 0.551 usec (+- 0.006 usec)
Average memory usage: 12105 KB (+- 7 KB)
# Running 'internals/inject-build-id' benchmark:
Average build-id injection took: 8.403 msec (+- 0.097 msec)
Average time per event: 0.824 usec (+- 0.010 usec)
Average memory usage: 12770 KB (+- 8 KB)
Average build-id-all injection took: 5.611 msec (+- 0.085 msec)
Average time per event: 0.550 usec (+- 0.008 usec)
Average memory usage: 12134 KB (+- 8 KB)
# Running 'internals/inject-build-id' benchmark:
Average build-id injection took: 8.518 msec (+- 0.102 msec)
Average time per event: 0.835 usec (+- 0.010 usec)
Average memory usage: 12518 KB (+- 10 KB)
Average build-id-all injection took: 5.503 msec (+- 0.073 msec)
Average time per event: 0.540 usec (+- 0.007 usec)
Average memory usage: 11882 KB (+- 8 KB)
Performance counter stats for 'perf bench internals inject-build-id' (5 runs):
2,394.88 msec task-clock:u # 1.577 CPUs utilized ( +- 0.83% )
0 context-switches:u # 0.000 K/sec
0 cpu-migrations:u # 0.000 K/sec
103,181 page-faults:u # 0.043 M/sec ( +- 0.11% )
3,548,172,030 cycles:u # 1.482 GHz ( +- 0.30% ) (83.26%)
81,537,700 stalled-cycles-frontend:u # 2.30% frontend cycles idle ( +- 1.54% ) (83.24%)
876,631,544 stalled-cycles-backend:u # 24.71% backend cycles idle ( +- 1.14% ) (83.45%)
5,960,361,707 instructions:u # 1.68 insn per cycle
# 0.15 stalled cycles per insn ( +- 0.27% ) (83.26%)
1,269,413,491 branches:u # 530.054 M/sec ( +- 0.10% ) (83.48%)
11,372,453 branch-misses:u # 0.90% of all branches ( +- 0.52% ) (83.31%)
1.51874 +- 0.00642 seconds time elapsed ( +- 0.42% )
$
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
Acked-by: Jiri Olsa <jolsa@redhat.com>
Tested-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Ian Rogers <irogers@google.com>
Cc: Mark Rutland <mark.rutland@arm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Stephane Eranian <eranian@google.com>
Link: http://lore.kernel.org/lkml/20201030054742.87740-1-namhyung@kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
|
|
When processing address packet and counter packet, if the packet
contains extended header, it misses to account the extra one byte for
header length calculation, thus returns the wrong packet length.
To correct the packet length calculation, one possible fixing is simply
to plus extra 1 for extended header, but will spread some duplicate code
in the flows for processing address packet and counter packet.
Alternatively, we can refine the function arm_spe_get_payload() to not
only support short header and allow it to support extended header, and
rely on it for the packet length calculation.
So this patch refactors function arm_spe_get_payload() with a new
argument 'ext_hdr' for support extended header; the packet processing
flows can invoke this function to unify the packet length calculation.
Signed-off-by: Leo Yan <leo.yan@linaro.org>
Reviewed-by: Andre Przywara <andre.przywara@arm.com>
Link: https://lore.kernel.org/r/20201111071149.815-6-leo.yan@linaro.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
|
|
In function arm_spe_get_events(), the event packet's 'index' is assigned
as payload length, but the flow is not directive: it firstly gets the
packet length from the return value of arm_spe_get_payload(), the value
includes header length (1) and payload length:
int ret = arm_spe_get_payload(buf, len, packet);
and then reduces header length from packet length, so finally get the
payload length:
packet->index = ret - 1;
To simplify the code, this patch directly assigns payload length to
event packet's index; and at the end it calls arm_spe_get_payload() to
return the payload value.
Signed-off-by: Leo Yan <leo.yan@linaro.org>
Reviewed-by: Andre Przywara <andre.przywara@arm.com>
Link: https://lore.kernel.org/r/20201111071149.815-5-leo.yan@linaro.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
|
|
This patch defines macro to extract "sz" field from header, and renames
the function payloadlen() to arm_spe_payload_len().
Signed-off-by: Leo Yan <leo.yan@linaro.org>
Reviewed-by: Andre Przywara <andre.przywara@arm.com>
Link: https://lore.kernel.org/r/20201111071149.815-4-leo.yan@linaro.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
|
|
Fix a typo: s/iff/if.
Signed-off-by: Leo Yan <leo.yan@linaro.org>
Reviewed-by: Andre Przywara <andre.przywara@arm.com>
Link: https://lore.kernel.org/r/20201111071149.815-3-leo.yan@linaro.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
|
|
Include header linux/bitops.h, directly use its BIT() macro and remove
the self defined macros.
Committer notes:
Use BIT_ULL() instead of BIT to build on 32-bit arches as mentioned in
review by Andre Przywara <andre.przywara@arm.com>. I noticed the build
failure when crossbuilding to arm32 from x86_64.
Signed-off-by: Leo Yan <leo.yan@linaro.org>
Reviewed-by: Andre Przywara <andre.przywara@arm.com>
Link: https://lore.kernel.org/r/20201111071149.815-2-leo.yan@linaro.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
|
|
This patch is to add itrace option '-M' to synthesize memory event.
Signed-off-by: Leo Yan <leo.yan@linaro.org>
Acked-by: Jiri Olsa <jolsa@redhat.com>
Link: https://lore.kernel.org/r/20201106094853.21082-7-leo.yan@linaro.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
|
|
On the architectures with perf memory profiling, two types of hardware
events have been supported: load and store; if want to profile memory
for both load and store operations, the tool will use these two events
at the same time, the usage is:
# perf mem record -t load,store -- uname
But this cannot be applied for AUX tracing event, the same PMU event can
be used to only trace memory load, or only memory store, or trace for
both memory load and store.
This patch introduces a new event PERF_MEM_EVENTS__LOAD_STORE, which is
used to support the event which can record both memory load and store
operations.
When user specifies memory operation type as 'load,store', or doesn't
set type so use 'load,store' as default, if the arch supports the event
PERF_MEM_EVENTS__LOAD_STORE, the tool will convert the required
operations to this single event; otherwise, if the arch doesn't support
PERF_MEM_EVENTS__LOAD_STORE, the tool rolls back to enable both events
PERF_MEM_EVENTS__LOAD and PERF_MEM_EVENTS__STORE, which keeps the same
behaviour with before.
Signed-off-by: Leo Yan <leo.yan@linaro.org>
Acked-by: Jiri Olsa <jolsa@redhat.com>
Link: https://lore.kernel.org/r/20201106094853.21082-4-leo.yan@linaro.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
|
|
Different architectures might use different event or different event
parameters for memory profiling, this patch introduces a weak
perf_mem_events__ptr() function which allows to return back a
architecture specific memory event.
Since the variable 'perf_mem_events' can be only accessed by the
perf_mem_events__ptr() function, mark the variable as 'static', this
allows the architectures to define its own memory event array.
Signed-off-by: Leo Yan <leo.yan@linaro.org>
Acked-by: Jiri Olsa <jolsa@redhat.com>
Link: https://lore.kernel.org/r/20201106094853.21082-3-leo.yan@linaro.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
|
|
The perf tool searches a memory event name under the folder
'/sys/devices/cpu/events/', this leads to the limitation for the
selection of a memory profiling event which must be under this folder.
Thus it's impossible to use any other event as memory event which is not
under this specific folder, e.g. Arm SPE hardware event is not located
in '/sys/devices/cpu/events/' so it cannot be enabled for memory
profiling.
This patch changes to search folder from '/sys/devices/cpu/events/' to
'/sys/devices', so it give flexibility to find events which can be used
for memory profiling.
Signed-off-by: Leo Yan <leo.yan@linaro.org>
Acked-by: Jiri Olsa <jolsa@redhat.com>
Link: https://lore.kernel.org/r/20201106094853.21082-2-leo.yan@linaro.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
|
|
Add a new --quiet option to 'perf stat'. This is useful with 'perf stat
record' to write the data only to the perf.data file, which can lower
measurement overhead because the data doesn't need to be formatted.
On my 4C desktop:
% time ./perf stat record -e $(python -c 'print ",".join(["cycles"]*1000)') -a -I 1000 sleep 5
...
real 0m5.377s
user 0m0.238s
sys 0m0.452s
% time ./perf stat record --quiet -e $(python -c 'print ",".join(["cycles"]*1000)') -a -I 1000 sleep 5
real 0m5.452s
user 0m0.183s
sys 0m0.423s
In this example it cuts the user time by 20%. On systems with more cores
the savings are higher.
Signed-off-by: Andi Kleen <andi@firstfloor.org>
Acked-by: Jiri Olsa <jolsa@kernel.org>
Cc: Alexey Budankov <alexey.budankov@linux.intel.com>
Link: http://lore.kernel.org/lkml/20201027002737.30942-1-andi@firstfloor.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
|
|
To make the command line even more compact with cgroups, support regex
pattern matching in cgroup names.
$ perf stat -a -e cpu-clock,cycles --for-each-cgroup ^foo sleep 1
3,000.73 msec cpu-clock foo # 2.998 CPUs utilized
12,530,992,699 cycles foo # 7.517 GHz (100.00%)
1,000.61 msec cpu-clock foo/bar # 1.000 CPUs utilized
4,178,529,579 cycles foo/bar # 2.506 GHz (100.00%)
1,000.03 msec cpu-clock foo/baz # 0.999 CPUs utilized
4,176,104,315 cycles foo/baz # 2.505 GHz (100.00%)
1.000892614 seconds time elapsed
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
Acked-by: Jiri Olsa <jolsa@redhat.com>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Ian Rogers <irogers@google.com>
Cc: Mark Rutland <mark.rutland@arm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Stephane Eranian <eranian@google.com>
Link: http://lore.kernel.org/lkml/20201027072855.655449-2-namhyung@kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
|
|
If libbpf isn't selected, no need for a bunch of related code, that were
not even being used, as code using these perf_env methods was also
enclosed in HAVE_LIBBPF_SUPPORT.
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: Ian Rogers <irogers@google.com>
Cc: Jiri Olsa <jolsa@kernel.org>
Cc: Namhyung Kim <namhyung@kernel.org>
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
|
|
No need to include it otherwise.
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: Ian Rogers <irogers@google.com>
Cc: Jiri Olsa <jolsa@kernel.org>
Cc: Namhyung Kim <namhyung@kernel.org>
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
|
|
As it uses the 'deprecated' attribute in a way that breaks the build
with old gcc compilers, so to continue being able to build in such
systems where NO_LIBBPF=1 is being used, enclose it under
HAVE_LIBBPF_SUPPORT.
1 centos:6 : FAIL gcc (GCC) 4.4.7 20120313 (Red Hat 4.4.7-23)
2 oraclelinux:6 : FAIL gcc (GCC) 4.4.7 20120313 (Red Hat 4.4.7-23.0.1)
CC /tmp/build/perf/builtin-record.o
In file included from util/bpf-loader.h:11,
from builtin-record.c:39:
/git/linux/tools/lib/bpf/libbpf.h:203: error: wrong number of arguments specified for 'deprecated' attribute
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: Ian Rogers <irogers@google.com>
Cc: Jiri Olsa <jolsa@kernel.org>
Cc: Namhyung Kim <namhyung@kernel.org>
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
|
|
Some archs (e.g. x86 and Arm64) don't enable the configuration
CONFIG_MEMORY_HOTPLUG by default, if this configuration is not enabled
when build the kernel image, the SysFS for memory nodes will be missed.
This results in perf tool has no chance to catpure the memory nodes
information, when perf tool reports the result and detects no memory
nodes, it outputs "assertion failed at util/mem2node.c:99".
The output log doesn't give out reason for the failure and users have no
clue for how to fix it. This patch changes to use explicit way for
warning: it tells user that detected no memory nodes and suggests to
enable CONFIG_MEMORY_HOTPLUG for kernel building.
Signed-off-by: Leo Yan <leo.yan@linaro.org>
Acked-by: Jiri Olsa <jolsa@redhat.com>
Link: https://lore.kernel.org/r/20201019003613.8399-1-leo.yan@linaro.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
|
|
Support the MIPS architecture using the ins_ops association method. With
this patch, perf-annotate can work well on MIPS.
Testing it with a perf.data file collected on a mips machine:
$./perf annotate -i perf.data
: Disassembly of section .text:
:
: 00000000000be6a0 <get_next_seq>:
: get_next_seq():
0.00 : be6a0: lw v0,0(a0)
0.00 : be6a4: daddiu sp,sp,-128
0.00 : be6a8: ld a7,72(a0)
0.00 : be6ac: gssq s5,s4,80(sp)
0.00 : be6b0: gssq s1,s0,48(sp)
0.00 : be6b4: gssq s8,gp,112(sp)
0.00 : be6b8: gssq s7,s6,96(sp)
0.00 : be6bc: gssq s3,s2,64(sp)
0.00 : be6c0: sd a3,0(sp)
0.00 : be6c4: move s0,a0
0.00 : be6c8: sd v0,32(sp)
0.00 : be6cc: sd a5,8(sp)
0.00 : be6d0: sd zero,8(a0)
0.00 : be6d4: sd a6,16(sp)
0.00 : be6d8: ld s2,48(a0)
8.53 : be6dc: ld s1,40(a0)
9.42 : be6e0: ld v1,32(a0)
0.00 : be6e4: nop
0.00 : be6e8: ld s4,24(a0)
0.00 : be6ec: ld s5,16(a0)
0.00 : be6f0: sd a7,40(sp)
10.11 : be6f4: ld s6,64(a0)
...
The original patch link:
https://lore.kernel.org/patchwork/patch/1180480/
Signed-off-by: Dengcheng Zhu <dzhu@wavecomp.com>
Cc: Dengcheng Zhu <dzhu@wavecomp.com>
Cc: Jiaxun Yang <jiaxun.yang@flygoat.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Xuefeng Li <lixuefeng@loongson.cn>
Cc: linux-mips@vger.kernel.org
[ fanpeng@loongson.cn: Add missing "bgtzl", "bltzl", "bgezl", "blezl", "beql" and "bnel" for pre-R6processors ]
Signed-off-by: Peng Fan <fanpeng@loongson.cn>
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
|
|
It was missed to add a swap function for PERF_RECORD_CGROUP.
Fixes: ba78c1c5461c ("perf tools: Basic support for CGROUP event")
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
Acked-by: Jiri Olsa <jolsa@redhat.com>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Ian Rogers <irogers@google.com>
Cc: Mark Rutland <mark.rutland@arm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Stephane Eranian <eranian@google.com>
Link: http://lore.kernel.org/lkml/20201102140228.303657-1-namhyung@kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
|
|
We are missing swap for ino_generation field.
Fixes: 5c5e854bc760 ("perf tools: Add attr->mmap2 support")
Signed-off-by: Jiri Olsa <jolsa@kernel.org>
Acked-by: Namhyung Kim <namhyung@kernel.org>
Link: https://lore.kernel.org/r/20201101233103.3537427-2-jolsa@kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
|
|
We display garbage for undefined build_id objects, because we don't
initialize the output buffer.
Signed-off-by: Jiri Olsa <jolsa@kernel.org>
Acked-by: Namhyung Kim <namhyung@kernel.org>
Link: https://lore.kernel.org/r/20201101233103.3537427-1-jolsa@kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
|
|
To avoid this:
util/scripting-engines/trace-event-python.c: In function 'python_start_script':
util/scripting-engines/trace-event-python.c:1595:2: error: 'visibility' attribute ignored [-Werror=attributes]
1595 | PyMODINIT_FUNC (*initfunc)(void);
| ^~~~~~~~~~~~~~
That started breaking when building with PYTHON=python3 and these gcc
versions (I haven't checked with the clang ones, maybe it breaks there
as well):
# export PERF_TARBALL=http://192.168.86.5/perf/perf-5.9.0.tar.xz
# dm fedora:33 fedora:rawhide
1 107.80 fedora:33 : Ok gcc (GCC) 10.2.1 20201005 (Red Hat 10.2.1-5), clang version 11.0.0 (Fedora 11.0.0-1.fc33)
2 92.47 fedora:rawhide : Ok gcc (GCC) 10.2.1 20201016 (Red Hat 10.2.1-6), clang version 11.0.0 (Fedora 11.0.0-1.fc34)
#
Avoid that by ditching that 'initfunc' function pointer with its:
#define Py_EXPORTED_SYMBOL _attribute_ ((visibility ("default")))
#define PyMODINIT_FUNC Py_EXPORTED_SYMBOL PyObject*
And just call PyImport_AppendInittab() at the end of the ifdef python3
block with the functions that were being attributed to that initfunc.
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: Ian Rogers <irogers@google.com>
Cc: Jiri Olsa <jolsa@kernel.org>
Cc: Namhyung Kim <namhyung@kernel.org>
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
|
|
The addr in PERF_RECORD_KSYMBOL events for non-jited bpf progs points to
the bpf interpreter, ie. within kernel text section. When processing the
unregister event, this causes unexpected removal of vmlinux_map,
crashing perf later in cleanup:
# perf record -- timeout --signal=INT 2s /usr/share/bcc/tools/execsnoop
PCOMM PID PPID RET ARGS
[ perf record: Woken up 1 times to write data ]
[ perf record: Captured and wrote 0.208 MB perf.data (5155 samples) ]
perf: tools/include/linux/refcount.h:131: refcount_sub_and_test: Assertion `!(new > val)' failed.
Aborted (core dumped)
# perf script -D|grep KSYM
0 0xa40 [0x48]: PERF_RECORD_KSYMBOL addr ffffffffa9b6b530 len 0 type 1 flags 0x0 name bpf_prog_f958f6eb72ef5af6
0 0xab0 [0x48]: PERF_RECORD_KSYMBOL addr ffffffffa9b6b530 len 0 type 1 flags 0x0 name bpf_prog_8c42dee26e8cd4c2
0 0xb20 [0x48]: PERF_RECORD_KSYMBOL addr ffffffffa9b6b530 len 0 type 1 flags 0x0 name bpf_prog_f958f6eb72ef5af6
108563691893 0x33d98 [0x58]: PERF_RECORD_KSYMBOL addr ffffffffa9b6b3b0 len 0 type 1 flags 0x0 name bpf_prog_bc5697a410556fc2_syscall__execve
108568518458 0x34098 [0x58]: PERF_RECORD_KSYMBOL addr ffffffffa9b6b3f0 len 0 type 1 flags 0x0 name bpf_prog_45e2203c2928704d_do_ret_sys_execve
109301967895 0x34830 [0x58]: PERF_RECORD_KSYMBOL addr ffffffffa9b6b3b0 len 0 type 1 flags 0x1 name bpf_prog_bc5697a410556fc2_syscall__execve
109302007356 0x348b0 [0x58]: PERF_RECORD_KSYMBOL addr ffffffffa9b6b3f0 len 0 type 1 flags 0x1 name bpf_prog_45e2203c2928704d_do_ret_sys_execve
perf: tools/include/linux/refcount.h:131: refcount_sub_and_test: Assertion `!(new > val)' failed.
Here the addresses match the bpf interpreter:
# grep -e ffffffffa9b6b530 -e ffffffffa9b6b3b0 -e ffffffffa9b6b3f0 /proc/kallsyms
ffffffffa9b6b3b0 t __bpf_prog_run224
ffffffffa9b6b3f0 t __bpf_prog_run192
ffffffffa9b6b530 t __bpf_prog_run32
Fix by not allowing vmlinux_map to be removed by PERF_RECORD_KSYMBOL
unregister event.
Signed-off-by: Tommi Rantala <tommi.t.rantala@nokia.com>
Acked-by: Jiri Olsa <jolsa@redhat.com>
Tested-by: Jiri Olsa <jolsa@redhat.com>
Link: https://lore.kernel.org/r/20201016114718.54332-1-tommi.t.rantala@nokia.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
|
|
To pick the changes in:
85367030a6c7ef33 ("libbpf: Centralize poisoning and poison reallocarray()")
7d9c71e10baa3496 ("libbpf: Extract generic string hashing function for reuse")
That don't entail any changes in tools/perf.
This addresses this perf build warning:
Warning: Kernel ABI header at 'tools/perf/util/hashmap.h' differs from latest version at 'tools/lib/bpf/hashmap.h'
diff -u tools/perf/util/hashmap.h tools/lib/bpf/hashmap.h
Not a kernel ABI, its just that this uses the mechanism in place for
checking kernel ABI files drift.
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: Alexei Starovoitov <ast@kernel.org>
Cc: Andrii Nakryiko <andriin@fb.com>
Cc: Ian Rogers <irogers@google.com>
Cc: Jiri Olsa <jolsa@kernel.org>
Cc: Namhyung Kim <namhyung@kernel.org>
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
|
|
Pull perf tools updates from Arnaldo Carvalho de Melo:
- cgroup improvements for 'perf stat', allowing for compact
specification of events and cgroups in the command line.
- Support per thread topdown metrics in 'perf stat'.
- Support sample-read topdown metric group in 'perf record'
- Show start of latency in addition to its start in 'perf sched
latency'.
- Add min, max to 'perf script' futex-contention output, in addition to
avg.
- Allow usage of 'perf_event_attr->exclusive' attribute via the new
':e' event modifier.
- Add 'snapshot' command to 'perf record --control', using it with
Intel PT.
- Support FIFO file names as alternative options to 'perf record
--control'.
- Introduce branch history "streams", to compare 'perf record' runs
with 'perf diff' based on branch records and report hot streams.
- Support PE executable symbol tables using libbfd, to profile, for
instance, wine binaries.
- Add filter support for option 'perf ftrace -F/--funcs'.
- Allow configuring the 'disassembler_style' 'perf annotate' knob via
'perf config'
- Update CascadelakeX and SkylakeX JSON vendor events files.
- Add support for parsing perchip/percore JSON vendor events.
- Add power9 hv_24x7 core level metric events.
- Add L2 prefetch, ITLB instruction fetch hits JSON events for AMD
zen1.
- Enable Family 19h users by matching Zen2 AMD vendor events.
- Use debuginfod in 'perf probe' when required debug files not found
locally.
- Display negative tid in non-sample events in 'perf script'.
- Make GTK2 support opt-in
- Add build test with GTK+
- Add missing -lzstd to the fast path feature detection
- Add scripts to auto generate 'mmap', 'mremap' string<->id tables for
use in 'perf trace'.
- Show python test script in verbose mode.
- Fix uncore metric expressions
- Msan uninitialized use fixes.
- Use condition variables in 'perf bench numa'
- Autodetect python3 binary in systems without python2.
- Support md5 build ids in addition to sha1.
- Add build id 'perf test' regression test.
- Fix printable strings in python3 scripts.
- Fix off by ones in 'perf trace' in arches using libaudit.
- Fix JSON event code for events referencing std arch events.
- Introduce 'perf test' shell script for Arm CoreSight testing.
- Add rdtsc() for Arm64 for used in the PERF_RECORD_TIME_CONV metadata
event and in 'perf test tsc'.
- 'perf c2c' improvements: Add "RMT Load Hit" metric, "Total Stores",
fixes and documentation update.
- Fix usage of reloc_sym in 'perf probe' when using both kallsyms and
debuginfo files.
- Do not print 'Metric Groups:' unnecessarily in 'perf list'
- Refcounting fixes in the event parsing code.
- Add expand cgroup event 'perf test' entry.
- Fix out of bounds CPU map access when handling armv8_pmu events in
'perf stat'.
- Add build-id injection 'perf bench' benchmark.
- Enter namespace when reading build-id in 'perf inject'.
- Do not load map/dso when injecting build-id speeding up the 'perf
inject' process.
- Add --buildid-all option to avoid processing all samples, just the
mmap metadata events.
- Add feature test to check if libbfd has buildid support
- Add 'perf test' entry for PE binary format support.
- Fix typos in power8 PMU vendor events JSON files.
- Hide libtraceevent non API functions.
* tag 'perf-tools-for-v5.10-2020-10-15' of git://git.kernel.org/pub/scm/linux/kernel/git/acme/linux: (113 commits)
perf c2c: Update documentation for metrics reorganization
perf c2c: Add metrics "RMT Load Hit"
perf c2c: Correct LLC load hit metrics
perf c2c: Change header for LLC local hit
perf c2c: Use more explicit headers for HITM
perf c2c: Change header from "LLC Load Hitm" to "Load Hitm"
perf c2c: Organize metrics based on memory hierarchy
perf c2c: Display "Total Stores" as a standalone metrics
perf c2c: Display the total numbers continuously
perf bench: Use condition variables in numa.
perf jevents: Fix event code for events referencing std arch events
perf diff: Support hot streams comparison
perf streams: Report hot streams
perf streams: Calculate the sum of total streams hits
perf streams: Link stream pair
perf streams: Compare two streams
perf streams: Get the evsel_streams by evsel_idx
perf streams: Introduce branch history "streams"
perf intel-pt: Improve PT documentation slightly
perf tools: Add support for exclusive groups/events
...
|
|
We show the streams separately. They are divided into different sections.
1. "Matched hot streams"
2. "Hot streams in old perf data only"
3. "Hot streams in new perf data only".
For each stream, we report the cycles and hot percent (hits%).
For example,
cycles: 2, hits: 4.08%
--------------------------
main div.c:42
compute_flag div.c:28
Signed-off-by: Jin Yao <yao.jin@linux.intel.com>
Acked-by: Jiri Olsa <jolsa@kernel.org>
Link: https://lore.kernel.org/r/20201009022845.13141-7-yao.jin@linux.intel.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
|
|
We have used callchain_node->hit to measure the hot level of one stream.
This patch calculates the sum of hits of total streams.
Thus in next patch, we can use following formula to report hot percent
for one stream.
hot percent = callchain_node->hit / sum of total hits
Signed-off-by: Jin Yao <yao.jin@linux.intel.com>
Acked-by: Jiri Olsa <jolsa@kernel.org>
Link: https://lore.kernel.org/r/20201009022845.13141-6-yao.jin@linux.intel.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
|
|
In previous patch, we have created an evsel_streams for one event, and
top N hottest streams will be saved in a stream array in evsel_streams.
This patch compares total streams among two evsel_streams.
Once two streams are fully matched, they will be linked as a pair. From
the pair, we can know which streams are matched.
Signed-off-by: Jin Yao <yao.jin@linux.intel.com>
Acked-by: Jiri Olsa <jolsa@kernel.org>
Link: https://lore.kernel.org/r/20201009022845.13141-5-yao.jin@linux.intel.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
|
|
Stream is the branch history which is aggregated by the branch records
from perf samples. Now we support the callchain as stream.
If the callchain entries of one stream are fully matched with the
callchain entries of another stream, we think two streams are matched.
For example,
cycles: 1, hits: 26.80% cycles: 1, hits: 27.30%
----------------------- -----------------------
main div.c:39 main div.c:39
main div.c:44 main div.c:44
Above two streams are matched (we don't consider the case that source
code is changed).
The matching logic is, compare the chain string first. If it's not
matched, fallback to dso address comparison.
Signed-off-by: Jin Yao <yao.jin@linux.intel.com>
Acked-by: Jiri Olsa <jolsa@kernel.org>
Link: https://lore.kernel.org/r/20201009022845.13141-4-yao.jin@linux.intel.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
|
|
In previous patch, we have created evsel_streams array.
This patch returns the specified evsel_streams according to the
evsel_idx.
Signed-off-by: Jin Yao <yao.jin@linux.intel.com>
Acked-by: Jiri Olsa <jolsa@kernel.org>
Link: https://lore.kernel.org/r/20201009022845.13141-3-yao.jin@linux.intel.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
|
|
We define a stream as the branch history which is aggregated by the
branch records from perf samples. For example, the callchains aggregated
from the branch records are considered as streams. By browsing the hot
stream, we can understand the hot code path.
Now we only support the callchain for stream. For measuring the hot
level for a stream, we use the callchain_node->hit, higher is hotter.
There may be many callchains sampled so we only focus on the top N
hottest callchains. N is a user defined parameter or predefined default
value (nr_streams_max).
This patch creates an evsel_streams array per event, and saves the top N
hottest streams in a stream array.
So now we can get the per-event top N hottest streams.
Signed-off-by: Jin Yao <yao.jin@linux.intel.com>
Acked-by: Jiri Olsa <jolsa@kernel.org>
Link: https://lore.kernel.org/r/20201009022845.13141-2-yao.jin@linux.intel.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
|
|
Peter suggested that using the exclusive mode in perf could avoid some
problems with bad scheduling of groups. Exclusive is implemented in the
kernel, but wasn't exposed by the perf tool, so hard to use without
custom low level API users.
Add support for marking groups or events with :e for exclusive in the
perf tool. The implementation is basically the same as the existing
pinned attribute.
Committer testing:
# perf test "parse event"
6: Parse event definition strings : Ok
# perf test -v "parse event" |& grep :u*e
running test 56 'instructions:uep'
running test 57 '{cycles,cache-misses,branch-misses}:e'
#
#
# grep "model name" -m1 /proc/cpuinfo
model name : AMD Ryzen 9 3900X 12-Core Processor
#
# perf stat -a -e '{cycles,cache-misses,branch-misses}:e' sleep 1
Performance counter stats for 'system wide':
<not counted> cycles (0.00%)
<not counted> cache-misses (0.00%)
<not counted> branch-misses (0.00%)
1.001269893 seconds time elapsed
Some events weren't counted. Try disabling the NMI watchdog:
echo 0 > /proc/sys/kernel/nmi_watchdog
perf stat ...
echo 1 > /proc/sys/kernel/nmi_watchdog
# echo 0 > /proc/sys/kernel/nmi_watchdog
# perf stat -a -e '{cycles,cache-misses,branch-misses}:e' sleep 1
Performance counter stats for 'system wide':
1,298,663,141 cycles
30,962,215 cache-misses
5,325,150 branch-misses
1.001474934 seconds time elapsed
#
# The output for asking for precise events on AMD needs to improve, it
# supposedly works only for system wide or per CPU
#
# perf stat -a -e '{cycles,cache-misses,branch-misses}:uep' sleep 1
Error:
The sys_perf_event_open() syscall returned with 22 (Invalid argument) for event (cycles).
/bin/dmesg | grep -i perf may provide additional information.
# perf stat -a -e '{cycles,cache-misses,branch-misses}:ue' sleep 1
Performance counter stats for 'system wide':
746,363,126 cycles
16,881,611 cache-misses
2,871,259 branch-misses
1.001636066 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>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Link: http://lore.kernel.org/lkml/20201014144255.22699-1-andi@firstfloor.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
|
|
With shorter md5 build ids we need to align their paths properly with
other build ids:
$ perf buildid-list
17f4e448cc746582ea1881528deb549f7fdb3fd5 [kernel.kallsyms]
a50e350e97c43b4708d09bcd85ebfff7 .../tools/perf/buildid-ex-md5
1805c738c8f3ec0f47b7ea09080c28f34d18a82b /usr/lib64/ld-2.31.so
$
Signed-off-by: Jiri Olsa <jolsa@kernel.org>
Acked-by: Ian Rogers <irogers@google.com>
Link: https://lore.kernel.org/r/20201013192441.1299447-9-jolsa@kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
|
|
We do not store size with build ids in perf data, but there's enough
space to do it. Adding misc bit PERF_RECORD_MISC_BUILD_ID_SIZE to mark
build id event with size.
With this fix the dso with md5 build id will have correct build id data
and will be usable for debuginfod processing if needed (coming in
following patches).
Committer notes:
Use %zu with size_t to fix this error on 32-bit arches:
util/header.c: In function '__event_process_build_id':
util/header.c:2105:3: error: format '%lu' expects argument of type 'long unsigned int', but argument 6 has type 'size_t' [-Werror=format=]
pr_debug("build id event received for %s: %s [%lu]\n",
^
Signed-off-by: Jiri Olsa <jolsa@kernel.org>
Acked-by: Ian Rogers <irogers@google.com>
Link: https://lore.kernel.org/r/20201013192441.1299447-8-jolsa@kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
|
|
Passing build_id object to dso__build_id_equal(), so we can properly
check build id with different size than sha1.
Signed-off-by: Jiri Olsa <jolsa@kernel.org>
Acked-by: Ian Rogers <irogers@google.com>
Link: https://lore.kernel.org/r/20201013192441.1299447-7-jolsa@kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
|
|
Passing build_id object to dso__set_build_id(), so it's easier
to initialize dos's build id object.
Signed-off-by: Jiri Olsa <jolsa@kernel.org>
Acked-by: Ian Rogers <irogers@google.com>
Link: https://lore.kernel.org/r/20201013192441.1299447-6-jolsa@kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
|
|
Passing build_id object to build_id__sprintf function, so it can operate
with the proper size of build id.
This will create proper md5 build id readable names,
like following:
a50e350e97c43b4708d09bcd85ebfff7
instead of:
a50e350e97c43b4708d09bcd85ebfff700000000
Signed-off-by: Jiri Olsa <jolsa@kernel.org>
Acked-by: Ian Rogers <irogers@google.com>
Link: https://lore.kernel.org/r/20201013192441.1299447-5-jolsa@kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
|
|
Passing build id object to sysfs__read_build_id function, so it can
populate the size of the build_id object.
Signed-off-by: Jiri Olsa <jolsa@kernel.org>
Acked-by: Ian Rogers <irogers@google.com>
Link: https://lore.kernel.org/r/20201013192441.1299447-4-jolsa@kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
|
|
Pass a build_id object to filename__read_build_id function, so it can
populate the size of the build_id object.
Changing filename__read_build_id() code for both ELF/non-ELF code.
Signed-off-by: Jiri Olsa <jolsa@kernel.org>
Acked-by: Ian Rogers <irogers@google.com>
Link: https://lore.kernel.org/r/20201013192441.1299447-3-jolsa@kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
|
|
Replace build_id byte array with struct build_id object and all the code
that references it.
The objective is to carry size together with build id array, so it's
better to keep both together.
This is preparatory change for following patches, and there's no
functional change.
Signed-off-by: Jiri Olsa <jolsa@kernel.org>
Acked-by: Ian Rogers <irogers@google.com>
Link: https://lore.kernel.org/r/20201013192441.1299447-2-jolsa@kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
|
|
We'll use it to ask for extra config files to be loaded, profile like
stuff that will be used first to make 'perf trace' mimic 'strace' output
via a 'perf strace' command that just sets up 'perf trace' output.
At some point it'll be used for regression tests, where we'll run some
simple commands like:
perf strace ls > perf-strace.output
strace ls > strace.output
And then do some mutable syscall arg aware diff like tool to deal with
arguments for things like mmap, that change at each execution, to be
first ignored and then properly tracked when used accoss multiple
syscalls.
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: Ian Rogers <irogers@google.com>
Cc: Jiri Olsa <jolsa@kernel.org>
Cc: Namhyung Kim <namhyung@kernel.org>
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
|
|
To pick fixes that missed v5.9.
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
|
|
No need to load symbols in a DSO when injecting build-id. I guess the
reason was to check the DSO is a special file like anon files. Use some
helper functions in map.c to check them before reading build-id. Also
pass sample event's cpumode to a new build-id event.
It brought a speedup in the benchmark of 25 -> 21 msec on my laptop.
Also the memory usage (Max RSS) went down by ~200 KB.
# Running 'internals/inject-build-id' benchmark:
Average build-id injection took: 21.389 msec (+- 0.138 msec)
Average time per event: 2.097 usec (+- 0.014 usec)
Average memory usage: 8225 KB (+- 0 KB)
Committer notes:
Before:
$ perf stat -r5 perf bench internals inject-build-id > /dev/null
Performance counter stats for 'perf bench internals inject-build-id' (5 runs):
4,020.56 msec task-clock:u # 1.271 CPUs utilized ( +- 0.74% )
0 context-switches:u # 0.000 K/sec
0 cpu-migrations:u # 0.000 K/sec
123,354 page-faults:u # 0.031 M/sec ( +- 0.81% )
7,119,951,568 cycles:u # 1.771 GHz ( +- 1.74% ) (83.27%)
230,086,969 stalled-cycles-frontend:u # 3.23% frontend cycles idle ( +- 1.97% ) (83.41%)
1,168,298,765 stalled-cycles-backend:u # 16.41% backend cycles idle ( +- 1.13% ) (83.44%)
11,173,083,669 instructions:u # 1.57 insn per cycle
# 0.10 stalled cycles per insn ( +- 1.58% ) (83.31%)
2,413,908,936 branches:u # 600.392 M/sec ( +- 1.69% ) (83.26%)
46,576,289 branch-misses:u # 1.93% of all branches ( +- 2.20% ) (83.31%)
3.1638 +- 0.0309 seconds time elapsed ( +- 0.98% )
$
After:
$ perf stat -r5 perf bench internals inject-build-id > /dev/null
Performance counter stats for 'perf bench internals inject-build-id' (5 runs):
2,379.94 msec task-clock:u # 1.473 CPUs utilized ( +- 0.18% )
0 context-switches:u # 0.000 K/sec
0 cpu-migrations:u # 0.000 K/sec
62,584 page-faults:u # 0.026 M/sec ( +- 0.07% )
2,372,389,668 cycles:u # 0.997 GHz ( +- 0.29% ) (83.14%)
106,937,862 stalled-cycles-frontend:u # 4.51% frontend cycles idle ( +- 4.89% ) (83.20%)
581,697,915 stalled-cycles-backend:u # 24.52% backend cycles idle ( +- 0.71% ) (83.47%)
3,659,692,199 instructions:u # 1.54 insn per cycle
# 0.16 stalled cycles per insn ( +- 0.10% ) (83.63%)
791,372,961 branches:u # 332.518 M/sec ( +- 0.27% ) (83.39%)
10,648,083 branch-misses:u # 1.35% of all branches ( +- 0.22% ) (83.16%)
1.61570 +- 0.00172 seconds time elapsed ( +- 0.11% )
$
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
Original-patch-by: Stephane Eranian <eranian@google.com>
Tested-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Acked-by: Jiri Olsa <jolsa@redhat.com>
Link: https://lore.kernel.org/r/20201012070214.2074921-5-namhyung@kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
|
|
Sometimes I can see that 'perf record' piped with 'perf inject' take a
long time processing build-ids.
So introduce a inject-build-id benchmark to the internals benchmark
suite to measure its overhead regularly.
It runs the 'perf inject' command internally and feeds the given number
of synthesized events (MMAP2 + SAMPLE basically).
Usage: perf bench internals inject-build-id <options>
-i, --iterations <n> Number of iterations used to compute average (default: 100)
-m, --nr-mmaps <n> Number of mmap events for each iteration (default: 100)
-n, --nr-samples <n> Number of sample events per mmap event (default: 100)
-v, --verbose be more verbose (show iteration count, DSO name, etc)
By default, it measures average processing time of 100 MMAP2 events
and 10000 SAMPLE events. Below is a result on my laptop.
$ perf bench internals inject-build-id
# Running 'internals/inject-build-id' benchmark:
Average build-id injection took: 25.789 msec (+- 0.202 msec)
Average time per event: 2.528 usec (+- 0.020 usec)
Average memory usage: 8411 KB (+- 7 KB)
Committer testing:
$ perf bench
Usage:
perf bench [<common options>] <collection> <benchmark> [<options>]
# List of all available benchmark collections:
sched: Scheduler and IPC benchmarks
syscall: System call benchmarks
mem: Memory access benchmarks
numa: NUMA scheduling and MM benchmarks
futex: Futex stressing benchmarks
epoll: Epoll stressing benchmarks
internals: Perf-internals benchmarks
all: All benchmarks
$ perf bench internals
# List of available benchmarks for collection 'internals':
synthesize: Benchmark perf event synthesis
kallsyms-parse: Benchmark kallsyms parsing
inject-build-id: Benchmark build-id injection
$ perf bench internals inject-build-id
# Running 'internals/inject-build-id' benchmark:
Average build-id injection took: 14.202 msec (+- 0.059 msec)
Average time per event: 1.392 usec (+- 0.006 usec)
Average memory usage: 12650 KB (+- 10 KB)
Average build-id-all injection took: 12.831 msec (+- 0.071 msec)
Average time per event: 1.258 usec (+- 0.007 usec)
Average memory usage: 11895 KB (+- 10 KB)
$
$ perf stat -r5 perf bench internals inject-build-id
# Running 'internals/inject-build-id' benchmark:
Average build-id injection took: 14.380 msec (+- 0.056 msec)
Average time per event: 1.410 usec (+- 0.006 usec)
Average memory usage: 12608 KB (+- 11 KB)
Average build-id-all injection took: 11.889 msec (+- 0.064 msec)
Average time per event: 1.166 usec (+- 0.006 usec)
Average memory usage: 11838 KB (+- 10 KB)
# Running 'internals/inject-build-id' benchmark:
Average build-id injection took: 14.246 msec (+- 0.065 msec)
Average time per event: 1.397 usec (+- 0.006 usec)
Average memory usage: 12744 KB (+- 10 KB)
Average build-id-all injection took: 12.019 msec (+- 0.066 msec)
Average time per event: 1.178 usec (+- 0.006 usec)
Average memory usage: 11963 KB (+- 10 KB)
# Running 'internals/inject-build-id' benchmark:
Average build-id injection took: 14.321 msec (+- 0.067 msec)
Average time per event: 1.404 usec (+- 0.007 usec)
Average memory usage: 12690 KB (+- 10 KB)
Average build-id-all injection took: 11.909 msec (+- 0.041 msec)
Average time per event: 1.168 usec (+- 0.004 usec)
Average memory usage: 11938 KB (+- 10 KB)
# Running 'internals/inject-build-id' benchmark:
Average build-id injection took: 14.287 msec (+- 0.059 msec)
Average time per event: 1.401 usec (+- 0.006 usec)
Average memory usage: 12864 KB (+- 10 KB)
Average build-id-all injection took: 11.862 msec (+- 0.058 msec)
Average time per event: 1.163 usec (+- 0.006 usec)
Average memory usage: 12103 KB (+- 10 KB)
# Running 'internals/inject-build-id' benchmark:
Average build-id injection took: 14.402 msec (+- 0.053 msec)
Average time per event: 1.412 usec (+- 0.005 usec)
Average memory usage: 12876 KB (+- 10 KB)
Average build-id-all injection took: 11.826 msec (+- 0.061 msec)
Average time per event: 1.159 usec (+- 0.006 usec)
Average memory usage: 12111 KB (+- 10 KB)
Performance counter stats for 'perf bench internals inject-build-id' (5 runs):
4,267.48 msec task-clock:u # 1.502 CPUs utilized ( +- 0.14% )
0 context-switches:u # 0.000 K/sec
0 cpu-migrations:u # 0.000 K/sec
102,092 page-faults:u # 0.024 M/sec ( +- 0.08% )
3,894,589,578 cycles:u # 0.913 GHz ( +- 0.19% ) (83.49%)
140,078,421 stalled-cycles-frontend:u # 3.60% frontend cycles idle ( +- 0.77% ) (83.34%)
948,581,189 stalled-cycles-backend:u # 24.36% backend cycles idle ( +- 0.46% ) (83.25%)
5,835,587,719 instructions:u # 1.50 insn per cycle
# 0.16 stalled cycles per insn ( +- 0.21% ) (83.24%)
1,267,423,636 branches:u # 296.996 M/sec ( +- 0.22% ) (83.12%)
17,484,290 branch-misses:u # 1.38% of all branches ( +- 0.12% ) (83.55%)
2.84176 +- 0.00222 seconds time elapsed ( +- 0.08% )
$
Acked-by: Jiri Olsa <jolsa@redhat.com>
Tested-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
Link: https://lore.kernel.org/r/20201012070214.2074921-2-namhyung@kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
|
|
Hagen reported broken strings in python3 tracepoint scripts:
make PYTHON=python3
perf record -e sched:sched_switch -a -- sleep 5
perf script --gen-script py
perf script -s ./perf-script.py
[..]
sched__sched_switch 7 563231.759525792 0 swapper prev_comm=bytearray(b'swapper/7\x00\x00\x00\x00\x00\x00\x00'), prev_pid=0, prev_prio=120, prev_state=, next_comm=bytearray(b'mutex-thread-co\x00'),
The problem is in the is_printable_array function that does not take the
zero byte into account and claim such string as not printable, so the
code will create byte array instead of string.
Committer testing:
After this fix:
sched__sched_switch 3 484522.497072626 1158680 kworker/3:0-eve prev_comm=kworker/3:0, prev_pid=1158680, prev_prio=120, prev_state=I, next_comm=swapper/3, next_pid=0, next_prio=120
Sample: {addr=0, cpu=3, datasrc=84410401, datasrc_decode=N/A|SNP N/A|TLB N/A|LCK N/A, ip=18446744071841817196, period=1, phys_addr=0, pid=1158680, tid=1158680, time=484522497072626, transaction=0, values=[(0, 0)], weight=0}
sched__sched_switch 4 484522.497085610 1225814 perf prev_comm=perf, prev_pid=1225814, prev_prio=120, prev_state=, next_comm=migration/4, next_pid=30, next_prio=0
Sample: {addr=0, cpu=4, datasrc=84410401, datasrc_decode=N/A|SNP N/A|TLB N/A|LCK N/A, ip=18446744071841817196, period=1, phys_addr=0, pid=1225814, tid=1225814, time=484522497085610, transaction=0, values=[(0, 0)], weight=0}
Fixes: 249de6e07458 ("perf script python: Fix string vs byte array resolving")
Signed-off-by: Jiri Olsa <jolsa@kernel.org>
Tested-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Tested-by: Hagen Paul Pfeifer <hagen@jauu.net>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Mark Rutland <mark.rutland@arm.com>
Cc: Michael Petlan <mpetlan@redhat.com>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: stable@vger.kernel.org
Link: http://lore.kernel.org/lkml/20200928201135.3633850-1-jolsa@kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
|
|
perf_event_attr bp_addr is a u64. parse-events.y parses it as a u64, but
casts it to a void* and then parse-events.c casts it back to a u64.
Rather than all the casts, change the type of the address to be a u64.
This removes an issue noted in:
https://lore.kernel.org/lkml/20200903184359.GC3495158@kernel.org/
Signed-off-by: Ian Rogers <irogers@google.com>
Acked-by: Jiri Olsa <jolsa@redhat.com>
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Jin Yao <yao.jin@linux.intel.com>
Cc: Leo Yan <leo.yan@linaro.org>
Cc: Mark Rutland <mark.rutland@arm.com>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Stephane Eranian <eranian@google.com>
Link: http://lore.kernel.org/lkml/20200925003903.561568-1-irogers@google.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
|
|
This is a preparation for a test case of expanding events for multiple
cgroups. Instead of using real system cgroup, the test will use fake
cgroups so it needs a way to have them without a open file descriptor.
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
Acked-by: Jiri Olsa <jolsa@redhat.com>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Ian Rogers <irogers@google.com>
Cc: Mark Rutland <mark.rutland@arm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Stephane Eranian <eranian@google.com>
Link: http://lore.kernel.org/lkml/20200924124455.336326-5-namhyung@kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
|
|
The metricgroup__copy_metric_events() is to handle metrics events when
expanding event for cgroups. As the metric events keep pointers to
evsel, it should be refreshed when events are cloned during the
operation.
The perf_stat__collect_metric_expr() is also called in case an event has
a metric directly.
During the copy, it references evsel by index as the evlist now has
cloned evsels for the given cgroup.
Also kernel test robot found an issue in the python module import so add
empty implementations of those two functions to fix it.
Reported-by: kernel test robot <rong.a.chen@intel.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
Acked-by: Jiri Olsa <jolsa@redhat.com>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Ian Rogers <irogers@google.com>
Cc: John Garry <john.garry@huawei.com>
Cc: Kajol Jain <kjain@linux.ibm.com>
Cc: Mark Rutland <mark.rutland@arm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Stephane Eranian <eranian@google.com>
Link: http://lore.kernel.org/lkml/20200924124455.336326-4-namhyung@kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
|
|
The --for-each-cgroup option is a syntax sugar to monitor large number
of cgroups easily. Current command line requires to list all the events
and cgroups even if users want to monitor same events for each cgroup.
This patch addresses that usage by copying given events for each cgroup
on user's behalf.
For instance, if they want to monitor 6 events for 200 cgroups each they
should write 1200 event names (with -e) AND 1200 cgroup names (with -G)
on the command line. But with this change, they can just specify 6
events and 200 cgroups with a new option.
A simpler example below: It wants to measure 3 events for 2 cgroups ('A'
and 'B'). The result is that total 6 events are counted like below.
$ perf stat -a -e cpu-clock,cycles,instructions --for-each-cgroup A,B sleep 1
Performance counter stats for 'system wide':
988.18 msec cpu-clock A # 0.987 CPUs utilized
3,153,761,702 cycles A # 3.200 GHz (100.00%)
8,067,769,847 instructions A # 2.57 insn per cycle (100.00%)
982.71 msec cpu-clock B # 0.982 CPUs utilized
3,136,093,298 cycles B # 3.182 GHz (99.99%)
8,109,619,327 instructions B # 2.58 insn per cycle (99.99%)
1.001228054 seconds time elapsed
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
Acked-by: Jiri Olsa <jolsa@redhat.com>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Ian Rogers <irogers@google.com>
Cc: Mark Rutland <mark.rutland@arm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Stephane Eranian <eranian@google.com>
Link: http://lore.kernel.org/lkml/20200924124455.336326-3-namhyung@kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
|
|
The evsel__clone() is to create an exactly same evsel from same
attributes. The function assumes the given evsel is not configured
yet so it cares fields set during event parsing. Those fields are now
moved together as Jiri suggested. Note that metric events will be
handled by later patch.
It will be used by perf stat to generate separate events for each
cgroup.
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
Acked-by: Jiri Olsa <jolsa@redhat.com>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Ian Rogers <irogers@google.com>
Cc: Mark Rutland <mark.rutland@arm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Stephane Eranian <eranian@google.com>
Link: http://lore.kernel.org/lkml/20200924124455.336326-2-namhyung@kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
|
|
Alexei Starovoitov says:
====================
pull-request: bpf-next 2020-09-23
The following pull-request contains BPF updates for your *net-next* tree.
We've added 95 non-merge commits during the last 22 day(s) which contain
a total of 124 files changed, 4211 insertions(+), 2040 deletions(-).
The main changes are:
1) Full multi function support in libbpf, from Andrii.
2) Refactoring of function argument checks, from Lorenz.
3) Make bpf_tail_call compatible with functions (subprograms), from Maciej.
4) Program metadata support, from YiFei.
5) bpf iterator optimizations, from Yonghong.
====================
Signed-off-by: David S. Miller <davem@davemloft.net>
|