Add --off-cpu option to enable the off-cpu profiling with BPF. It'd
use a bpf_output event and rename it to "offcpu-time". Samples will
be synthesized at the end of the record session using data from a BPF
map which contains the aggregated off-cpu time at context switches.
So it needs root privilege to get the off-cpu profiling.
Each sample will have a separate user stacktrace so it will skip
kernel threads. The sample ip will be set from the stacktrace and
other sample data will be updated accordingly. Currently it only
handles some basic sample types.
The sample timestamp is set to a dummy value just not to bother with
other events during the sorting. So it has a very big initial value
and increase it on processing each samples.
Good thing is that it can be used together with regular profiling like
cpu cycles. If you don't want to that, you can use a dummy event to
enable off-cpu profiling only.
Example output:
$ sudo perf record --off-cpu perf bench sched messaging -l 1000
$ sudo perf report --stdio --call-graph=no
# Total Lost Samples: 0
#
# Samples: 41K of event 'cycles'
# Event count (approx.): 42137343851
...
# Samples: 1K of event 'offcpu-time'
# Event count (approx.): 587990831640
#
# Children Self Command Shared Object Symbol
# ........ ........ ............... .................. .........................
#
81.66% 0.00% sched-messaging libc-2.33.so [.] __libc_start_main
81.66% 0.00% sched-messaging perf [.] cmd_bench
81.66% 0.00% sched-messaging perf [.] main
81.66% 0.00% sched-messaging perf [.] run_builtin
81.43% 0.00% sched-messaging perf [.] bench_sched_messaging
40.86% 40.86% sched-messaging libpthread-2.33.so [.] __read
37.66% 37.66% sched-messaging libpthread-2.33.so [.] __write
2.91% 2.91% sched-messaging libc-2.33.so [.] __poll
...
As you can see it spent most of off-cpu time in read and write in
bench_sched_messaging(). The --call-graph=no was added just to make
the output concise here.
It uses perf hooks facility to control BPF program during the record
session rather than adding new BPF/off-cpu specific calls.
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
Acked-by: Ian Rogers <irogers@google.com>
Tested-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Blake Jones <blakejones@google.com>
Cc: Hao Luo <haoluo@google.com>
Cc: Ingo Molnar <mingo@kernel.org>
Cc: Jiri Olsa <jolsa@kernel.org>
Cc: Milian Wolff <milian.wolff@kdab.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Song Liu <songliubraving@fb.com>
Cc: bpf@vger.kernel.org
Link: https://lore.kernel.org/r/20220518224725.742882-3-namhyung@kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
ptwrite is an Intel x86 instruction that writes arbitrary values into an
Intel PT trace. It is not supported on all hardware, so provide an
alternative that makes use of TNT packets to convey the payload data.
TNT packets encode Taken/Not-taken conditional branch information, so
taking branches based on the payload value will encode the value into
the TNT packet. Refer to the changes to the documentation file
perf-intel-pt.txt in this patch for an example.
Signed-off-by: Adrian Hunter <adrian.hunter@intel.com>
Cc: Jiri Olsa <jolsa@kernel.org>
Link: https://lore.kernel.org/r/20220509152400.376613-2-adrian.hunter@intel.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
When analyzing with 'perf script', it's useful to understand the
captured instruction and the next sequential instruction.
To calculate the address of the next sequential instruction, the length
of the captured instruction is required.
For example, you can’t know the next sequential instruction after an
unconditional branch unless you calculate that based on its length.
For branch stacks, 'perf script' only prints the instruction bytes with
'brstackinsn', but lacks the instruction length.
Add 'brstackinsnlen' to print the instruction length.
$ perf script -F ip,brstackinsn,brstackinsnlen --xed
7fa555be8f75
_start:
00007fa555be8090 mov %rsp, %rdi ilen: 3
00007fa555be8093 callq 0x7fa555be8ea0 ilen: 5 # PRED 102 cycles [102] 0.02 IPC
_dl_start+38:
00007fa555be8ec6 movq %rdx,0x227853(%rip) ilen: 7
00007fa555be8ecd leaq 0x227f94(%rip),%rdx ilen: 7
Signed-off-by: Kan Liang <kan.liang@linux.intel.com>
Cc: Ahmad Yasin <ahmad.yasin@intel.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Jiri Olsa <jolsa@kernel.org>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Xing Zhengjun <zhengjun.xing@linux.intel.com>
Link: https://lore.kernel.org/r/1647871212-184070-1-git-send-email-kan.liang@linux.intel.com
[ Added the new field to tools/perf/Documentation/perf-script.txt ]
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Extend --threads option in perf record command line interface.
The option can have a value in the form of masks that specify
CPUs to be monitored with data streaming threads and its layout
in system topology. The masks can be filtered using CPU mask
provided via -C option.
The specification value can be user defined list of masks. Masks
separated by colon define CPUs to be monitored by one thread and
affinity mask of that thread is separated by slash. For example:
<cpus mask 1>/<affinity mask 1>:<cpu mask 2>/<affinity mask 2>
specifies parallel threads layout that consists of two threads
with corresponding assigned CPUs to be monitored.
The specification value can be a string e.g. "cpu", "core" or
"package" meaning creation of data streaming thread for every
CPU or core or package to monitor distinct CPUs or CPUs grouped
by core or package.
The option provided with no or empty value defaults to per-cpu
parallel threads layout creating data streaming thread for every
CPU being monitored.
Document --threads option syntax and parallel data streaming modes
in Documentation/perf-record.txt.
Suggested-by: Jiri Olsa <jolsa@kernel.org>
Suggested-by: Namhyung Kim <namhyung@kernel.org>
Reviewed-by: Riccardo Mancini <rickyman7@gmail.com>
Signed-off-by: Alexey Bayduraev <alexey.v.bayduraev@linux.intel.com>
Tested-by: Jiri Olsa <jolsa@kernel.org>
Tested-by: Riccardo Mancini <rickyman7@gmail.com>
Acked-by: Andi Kleen <ak@linux.intel.com>
Acked-by: Namhyung Kim <namhyung@gmail.com>
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: Alexander Antonov <alexander.antonov@linux.intel.com>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Alexei Budankov <abudankov@huawei.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Link: https://lore.kernel.org/r/079e2619be70c465317cf7c9fdaf5fa069728c32.1642440724.git.alexey.v.bayduraev@linux.intel.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
In previous patch, we have supported the syntax which enables
the event on a specified pmu, such as:
cpu_core/<event>/
cpu_atom/<event>/
While this syntax is not very easy for applying on a set of
events or applying on a group. In following example, we have to
explicitly assign the pmu prefix.
# ./perf stat -e '{cpu_core/cycles/,cpu_core/instructions/}' -- sleep 1
Performance counter stats for 'sleep 1':
1,158,545 cpu_core/cycles/
1,003,113 cpu_core/instructions/
1.002428712 seconds time elapsed
A much easier way is:
# ./perf stat --cputype core -e '{cycles,instructions}' -- sleep 1
Performance counter stats for 'sleep 1':
1,101,071 cpu_core/cycles/
939,892 cpu_core/instructions/
1.002363142 seconds time elapsed
For this example, the '--cputype' enables the events from specified
pmu (cpu_core).
If '--cputype' conflicts with pmu prefix, '--cputype' is ignored.
# ./perf stat --cputype core -e cycles,cpu_atom/instructions/ -a -- sleep 1
Performance counter stats for 'system wide':
21,003,407 cpu_core/cycles/
367,886 cpu_atom/instructions/
1.002203520 seconds time elapsed
Signed-off-by: Jin Yao <yao.jin@linux.intel.com>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Jin Yao <yao.jin@intel.com>
Cc: Jiri Olsa <jolsa@kernel.org>
Cc: Kan Liang <kan.liang@linux.intel.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Link: http://lore.kernel.org/lkml/20210909062215.10278-1-yao.jin@linux.intel.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
AMD processors have events with event select codes and unit masks larger
than a byte. The core PMU, for example, uses 12-bit event select codes
split between bits 0-7 and 32-35 of the PERF_CTL MSRs as can be seen
from /sys/bus/event_sources/devices/cpu/format/*.
The Processor Programming Reference (PPR) lists the event codes as
unified 12-bit hexadecimal values instead and the split between the bits
is not apparent to someone who is not aware of the layout of the
PERF_CTL MSRs.
8-bit event select codes continue to work as the layout matches that of
the PERF_CTL MSRs i.e. bits 0-7 for event select and 8-15 for unit mask.
This adds more details in the perf man pages about using
/sys/bus/event_sources/devices/*/format/* for determining the correct
raw event encoding scheme.
E.g. the "op_cache_hit_miss.op_cache_hit" event with code 0x28f and
umask 0x03 can be programmed using its symbolic name as:
$ sudo perf --debug perf-event-open stat -e op_cache_hit_miss.op_cache_hit sleep 1
------------------------------------------------------------
perf_event_attr:
type 4
size 128
config 0x20000038f
sample_type IDENTIFIER
read_format TOTAL_TIME_ENABLED|TOTAL_TIME_RUNNING
disabled 1
inherit 1
enable_on_exec 1
exclude_guest 1
------------------------------------------------------------
[...]
One might use a simple eventsel+umask combination based on what the
current man pages say and incorrectly program the event as:
$ sudo perf --debug perf-event-open stat -e r0328f sleep 1
------------------------------------------------------------
perf_event_attr:
type 4
size 128
config 0x328f
sample_type IDENTIFIER
read_format TOTAL_TIME_ENABLED|TOTAL_TIME_RUNNING
disabled 1
inherit 1
enable_on_exec 1
exclude_guest 1
------------------------------------------------------------
[...]
When it should have been based on the format from sysfs:
$ cat /sys/bus/event_source/devices/cpu/format/event
config:0-7,32-35
$ sudo perf --debug perf-event-open stat -e r20000038f sleep 1
------------------------------------------------------------
perf_event_attr:
type 4
size 128
config 0x20000038f
sample_type IDENTIFIER
read_format TOTAL_TIME_ENABLED|TOTAL_TIME_RUNNING
disabled 1
inherit 1
enable_on_exec 1
exclude_guest 1
------------------------------------------------------------
[...]
Reviewed-by: Kajol Jain <kjain@linux.ibm.com>
Signed-off-by: Sandipan Das <sandipan.das@amd.com>
Acked-by: Jiri Olsa <jolsa@redhat.com>
Cc: Ananth Narayan <ananth.narayan@amd.com>
Cc: Kim Phillips <kim.phillips@amd.com>
Cc: Ravi Bangoria <ravi.bangoria@amd.com>
Cc: Robert Richter <rrichter@amd.com>
Cc: Santosh Shukla <santosh.shukla@amd.com>
Link: https://lore.kernel.org/r/20211123084613.243792-1-sandipan.das@amd.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
The instruction latency information can be recorded on
some platforms, e.g., the Intel Sapphire Rapids server. With both memory
latency (weight) and the new instruction latency information, users can
easily locate the expensive load instructions, and also understand the time
spent in different stages. The users can optimize their applications in
different pipeline stages.
Add a new field "ins_lat" to filter the instruction latency information,
which is available with sample type PERF_SAMPLE_WEIGHT_STRUCT.
Signed-off-by: Kan Liang <kan.liang@linux.intel.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Jiri Olsa <jolsa@redhat.com>
Cc: Joe Mario <jmario@redhat.com>
Link: https://lore.kernel.org/r/1632929894-102778-2-git-send-email-kan.liang@linux.intel.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Add a new dlfilter to show cycles.
Cycle counts are accumulated per CPU (or per thread if CPU is not recorded)
from IPC information, and printed together with the change since the last
print, at the start of each line. Separate counts are kept for branches,
instructions or other events.
Note also, the itrace A option can be useful to provide higher granularity
cycle information.
Example:
$ perf record -e intel_pt/cyc/u uname
Linux
[ perf record: Woken up 1 times to write data ]
[ perf record: Captured and wrote 0.044 MB perf.data ]
$ perf script --itrace=A --call-trace --dlfilter dlfilter-show-cycles.so --deltatime | head
0 perf-exec 8509 [001] 0.000000000: psb offs: 0
0 perf-exec 8509 [001] 0.000000000: cbr: 42 freq: 4219 MHz (156%)
833 833 uname 8509 [001] 0.000047689: (/usr/lib/x86_64-linux-gnu/ld-2.31.so ) _start
833 uname 8509 [001] 0.000003261: (/usr/lib/x86_64-linux-gnu/ld-2.31.so ) _dl_start
2015 1182 uname 8509 [001] 0.000000282: (/usr/lib/x86_64-linux-gnu/ld-2.31.so ) _dl_start
2676 661 uname 8509 [001] 0.000002629: (/usr/lib/x86_64-linux-gnu/ld-2.31.so ) _dl_start
3612 936 uname 8509 [001] 0.000001232: (/usr/lib/x86_64-linux-gnu/ld-2.31.so ) _dl_start
4579 967 uname 8509 [001] 0.000002519: (/usr/lib/x86_64-linux-gnu/ld-2.31.so ) _dl_start
6145 1566 uname 8509 [001] 0.000001050: (/usr/lib/x86_64-linux-gnu/ld-2.31.so ) _dl_setup_hash
6239 94 uname 8509 [001] 0.000000023: (/usr/lib/x86_64-linux-gnu/ld-2.31.so ) _dl_sysdep_start
Reviewed-by: Andi Kleen <ak@linux.intel.com>
Signed-off-by: Adrian Hunter <adrian.hunter@intel.com>
Cc: Jiri Olsa <jolsa@redhat.com>
Link: https://lore.kernel.org/r/20211027080334.365596-5-adrian.hunter@intel.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Normally, for cycle-acccurate mode, IPC values are an exact number of
instructions and cycles. Due to the granularity of timestamps, that happens
only when a CYC packet correlates to the event.
Support the itrace 'A' option, to use instead, the number of cycles
associated with the current timestamp. This provides IPC information for
every change of timestamp, but at the expense of accuracy. Due to the
granularity of timestamps, the actual number of cycles increases even
though the cycles reported does not. The number of instructions is known,
but if IPC is reported, cycles can be too low and so IPC is too high. Note
that inaccuracy decreases as the period of sampling increases i.e. if the
number of cycles is too low by a small amount, that becomes less
significant if the number of cycles is large.
Furthermore, it can be used in conjunction with dlfilter-show-cycles.so
to provide higher granularity cycle information.
Reviewed-by: Andi Kleen <ak@linux.intel.com>
Signed-off-by: Adrian Hunter <adrian.hunter@intel.com>
Cc: Jiri Olsa <jolsa@redhat.com>
Link: https://lore.kernel.org/r/20211027080334.365596-4-adrian.hunter@intel.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Like all locally-built programs, dlfilters may need to be re-built if
shared libraries they use change. Also there may be unexpected results
if the dfilter uses different versions of the shared libraries that perf
uses.
Note those things in the documentation.
Signed-off-by: Adrian Hunter <adrian.hunter@intel.com>
Cc: Jiri Olsa <jolsa@redhat.com>
Link: https //lore.kernel.org/r/20210811101036.17986-5-adrian.hunter@intel.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
perf_dlfilter_fns must not be const, because it is not.
Declaring it const can result in it being mapped read-only, causing a
segfaullt when it is written. Update documentation accordingly.
Signed-off-by: Adrian Hunter <adrian.hunter@intel.com>
Cc: Jiri Olsa <jolsa@redhat.com>
Fixes: 8defa7147d5572 ("perf script Add API for filtering via dynamically loaded shared object")
Link: https //lore.kernel.org/r/20210811101036.17986-2-adrian.hunter@intel.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
