- 16 Nov, 2022 30 commits
-
-
Ian Rogers authored
hashmap.h comes from libbpf but isn't installed with its headers. Always use the header file of the code in util. Change the hashmap.h dependency in expr.h to a forward declaration, add the necessary header file includes in the C files. Signed-off-by:
Ian Rogers <irogers@google.com> Acked-by:
Namhyung Kim <namhyung@kernel.org> Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com> Cc: Andrii Nakryiko <andrii.nakryiko@gmail.com> Cc: Jiri Olsa <jolsa@kernel.org> Cc: Mark Rutland <mark.rutland@arm.com> Cc: Masahiro Yamada <masahiroy@kernel.org> Cc: Nick Desaulniers <ndesaulniers@google.com> Cc: Nicolas Schier <nicolas@fjasle.eu> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Stephane Eranian <eranian@google.com> Cc: bpf@vger.kernel.org Link: http://lore.kernel.org/lkml/20221109184914.1357295-12-irogers@google.comSigned-off-by:
Arnaldo Carvalho de Melo <acme@redhat.com>
-
Ian Rogers authored
The perf build currently has a '-Itools/lib' on the CC command line. This causes issues as the libapi, libsubcmd, libtraceevent, libbpf and libsymbol headers are all found via this path, making it impossible to override include behavior. Change the libsymbol build mirroring the libbpf, libsubcmd, libapi, libperf and libtraceevent build, so that it is installed in a directory along with its headers. A later change will modify the include behavior. Don't build kallsyms.o as part of util as this will lead to duplicate definitions. Add kallsym's directory to the MANIFEST rather than individual files, so that the Build and Makefile are added to a source tar ball. Signed-off-by:
-
Ian Rogers authored
Add sufficient Makefile for libsymbol to be built as a dependency and header files installed. Signed-off-by:
-
Ian Rogers authored
Headers necessary for the perf build. Note, internal headers are also installed as these are necessary for the build. Signed-off-by:
-
Ian Rogers authored
Headers necessary for the perf build. Signed-off-by:
-
Ian Rogers authored
The perf build currently has a '-Itools/lib' on the CC command line. This causes issues as the libapi, libsubcmd, libtraceevent, libbpf headers are all found via this path, making it impossible to override include behavior. Change the libtraceevent build mirroring the libbpf, libsubcmd, libapi and libperf build, so that it is installed in a directory along with its headers. A later change will modify the include behavior. Similarly, the plugins are now installed into libtraceevent_plugins except they have no header files. Signed-off-by:
-
Ian Rogers authored
The perf build currently has a '-Itools/lib' on the CC command line. This causes issues as the libapi, libsubcmd, libtraceevent, libbpf headers are all found via this path, making it impossible to override include behavior. Change the libperf build mirroring the libbpf, libsubcmd and libapi build, so that it is installed in a directory along with its headers. A later change will modify the include behavior. Signed-off-by:
-
Ian Rogers authored
The perf build currently has a '-Itools/lib' on the CC command line. This causes issues as the libapi, libsubcmd, libtraceevent, libbpf headers are all found via this path, making it impossible to override include behavior. Change the libapi build mirroring the libbpf and libsubcmd build, so that it is installed in a directory along with its headers. A later change will modify the include behavior. Signed-off-by:
-
Ian Rogers authored
The perf build currently has a '-Itools/lib' on the CC command line. This causes issues as the libapi, libsubcmd, libtraceevent, libbpf headers are all found via this path, making it impossible to override include behavior. Change the libsubcmd build mirroring the libbpf build, so that it is installed in a directory along with its headers. A later change will modify the include behavior. Signed-off-by:
-
Ian Rogers authored
This allows libsubcmd to be installed as a dependency. Signed-off-by:
-
Ian Rogers authored
This allows libapi to be installed as a dependency. Signed-off-by:
-
Namhyung Kim authored
Normally, --for-each-cgroup only works with AGGR_GLOBAL. However the --topdown on some cpu (e.g. Intel Skylake) converts it to the AGGR_CORE internally. To support those machines, add print_aggr_cgroup and handle the events like in print_cgroup_events(). $ perf stat -a --for-each-cgroup system.slice,user.slice --topdown sleep 1 nmi_watchdog enabled with topdown. May give wrong results. Disable with echo 0 > /proc/sys/kernel/nmi_watchdog Performance counter stats for 'system wide': retiring bad speculation frontend bound backend bound S0-D0-C0 2 system.slice 49.0% -46.6% 31.4% S0-D0-C1 2 system.slice 55.5% 8.0% 45.5% -9.0% S0-D0-C2 2 system.slice 87.8% 22.1% 30.3% -40.3% S0-D0-C3 2 system.slice 53.3% -11.9% 45.2% 13.4% S0-D0-C0 2 user.slice 123.5% 4.0% 48.5% -75.9% S0-D0-C1 2 user.slice 19.9% 6.5% 89.9% -16.3% S0-D0-C2 2 user.slice 29.9% 7.9% 71.3% -9.1 S0-D0-C3 2 user.slice 28.0% 7.2% 43.3% 21.5% 1.004136937 seconds time elapsed Signed-off-by: -
Namhyung Kim authored
When we have events for each cgroup, the metric should be printed for each cgroup separately. Add print_cgroup_counter() to handle that situation properly. Also change print_metric_headers() not to print duplicate headers by checking cgroups. $ perf stat -a --for-each-cgroup system.slice,user.slice --metric-only sleep 1 Performance counter stats for 'system wide': GHz insn per cycle branch-misses of all branches system.slice 3.792 0.61 3.24% user.slice 3.661 2.32 0.37% 1.016111516 seconds time elapsed Signed-off-by: -
Namhyung Kim authored
For the metric-only case, add new functions to handle the start and the end of each metric display. Signed-off-by:
-
Namhyung Kim authored
The prefix is needed for interval mode to print timestamp at the beginning of each line. But the it's tricky for the metric only mode since it doesn't print every evsel and combines the metrics into a single line. So it needed to pass 'first' argument to print_counter_aggrdata() to determine if the current event is being printed at first. This makes the code hard to read. Let's move the logic out of the function and do it in the outer print loop. This would enable further cleanups later. Signed-off-by:
-
Namhyung Kim authored
Likewise, I think it'd better to have the control inside the function, and keep the higher level function clearer. Signed-off-by:
-
Namhyung Kim authored
There are print_header() and print_interval() to print header lines before actual counter values. Also print_metric_headers() needs to be called for the metric-only case. Let's move all these logics to a single place including num_print_iv to refresh the headers for interval mode. Signed-off-by:
-
Namhyung Kim authored
The print would run only if metric_only is not set, but it's already in a block that says it's in metric_only case. And there's no place to change the setting. Signed-off-by:
-
Namhyung Kim authored
Instead of using magic values, define symbolic constants and use them. Also add aggr_header_std[] array to simplify aggr_mode handling. Signed-off-by:
-
Namhyung Kim authored
This logic does not print the time directly, but it just puts the timestamp in the buffer as a prefix. To reduce the confusion, factor out the code into a separate function. Signed-off-by:
-
Namhyung Kim authored
The print_metric_headers() shows metric headers a little bit for each mode. Split it out to make the code clearer. Signed-off-by:
-
Namhyung Kim authored
We don't know how long cgroup name is, but at least we can align short ones like below. $ perf stat -a --for-each-cgroup system.slice,user.slice true Performance counter stats for 'system wide': 0.13 msec cpu-clock system.slice # 0.010 CPUs utilized 4 context-switches system.slice # 31.989 K/sec 1 cpu-migrations system.slice # 7.997 K/sec 0 page-faults system.slice # 0.000 /sec 450,673 cycles system.slice # 3.604 GHz (92.41%) 161,216 instructions system.slice # 0.36 insn per cycle (92.41%) 32,678 branches system.slice # 261.332 M/sec (92.41%) 2,628 branch-misses system.slice # 8.04% of all branches (92.41%) 14.29 msec cpu-clock user.slice # 1.163 CPUs utilized 35 context-switches user.slice # 2.449 K/sec 12 cpu-migrations user.slice # 839.691 /sec 57 page-faults user.slice # 3.989 K/sec 49,683,026 cycles user.slice # 3.477 GHz (99.38%) 110,790,266 instructions user.slice # 2.23 insn per cycle (99.38%) 24,552,255 branches user.slice # 1.718 G/sec (99.38%) 127,779 branch-misses user.slice # 0.52% of all branches (99.38%) 0.012289431 seconds time elapsed Signed-off-by: -
Namhyung Kim authored
Unfortunately, event running time, percentage and noise data are printed in different positions in normal output than CSV/JSON. I think it's better to put such details in where it actually prints. So add before_metric argument to print_noise() and print_running() and call them twice before and after the metric. Signed-off-by:
-
Namhyung Kim authored
In the printout() function, it checks if the event is bad (i.e. not counted or not supported) and print the result. But it does the same what abs_printout() is doing. So add an argument to indicate the value is ok or not and use the same function in both cases. Signed-off-by:
-
Namhyung Kim authored
And split it for each output mode like others. I believe it makes the code simpler and more intuitive. Now abs_printout() becomes just to call sub-functions. Signed-off-by:
-
Namhyung Kim authored
The aggr_printout() function is to print aggr_id and count (nr). Split it for each output mode to simplify the code. Signed-off-by:
-
Namhyung Kim authored
Likewise, split print_cgroup() for each output mode. Signed-off-by:
-
Namhyung Kim authored
Likewise, split print_noise_pct() for each output mode. Although it's a tiny function, more logic will be added soon so it'd be better split it and treat it in the same way. Signed-off-by:
-
Namhyung Kim authored
To make the code more obvious and hopefully simpler, factor out the code for each output mode - stdio, CSV, JSON. Signed-off-by:
-
Namhyung Kim authored
The --interval-clear option makes perf stat to clear the terminal at each interval. But it doesn't need to clear the screen when it saves to a file. Make it fail when it's enabled with the output options. $ perf stat -I 1 --interval-clear -o myfile true --interval-clear does not work with output Usage: perf stat [<options>] [<command>] -o, --output <file> output file name --log-fd <n> log output to fd, instead of stderr --interval-clear clear screen in between new interval Signed-off-by:
-
- 15 Nov, 2022 8 commits
-
-
Ian Rogers authored
Previously print_pmu_events() would compute the values to be printed, place them in struct sevent, sort them and then print them. Modify the code so that struct sevent holds just the PMU and event, sort these and then in the main print loop calculate aliases for names, etc. This avoids memory allocations for copied values as they are computed then printed. Signed-off-by:
-
Ian Rogers authored
The current code computes an array of symbol names then sorts and prints them. Use a strlist to create a list of names that is sorted and then print it. Signed-off-by:
-
Ian Rogers authored
The current code computes an array of cache names then sorts and prints them. Use a strlist to create a list of names that is sorted. Keep the hybrid names, it is unclear how to generalize it, but drop the computation of evt_pmus that is never used. Signed-off-by:
-
Ian Rogers authored
Deprecate the --cputype option and add a --unit option where '--unit cpu_atom' behaves like '--cputype atom'. The --unit option can be used with arbitrary PMUs, for example: ``` $ perf list --unit msr pmu List of pre-defined events (to be used in -e or -M): msr/aperf/ [Kernel PMU event] msr/cpu_thermal_margin/ [Kernel PMU event] msr/mperf/ [Kernel PMU event] msr/pperf/ [Kernel PMU event] msr/smi/ [Kernel PMU event] msr/tsc/ [Kernel PMU event] ``` Signed-off-by:
-
Ian Rogers authored
In print_tracepoint_events() use tracing_events__scandir_alphasort() and scandir alphasort so that the subsystem and events are sorted and don't need a secondary qsort. Locally this results in the following change: ... ext4:ext4_zero_range [Tracepoint event] - fib6:fib6_table_lookup [Tracepoint event] fib:fib_table_lookup [Tracepoint event] + fib6:fib6_table_lookup [Tracepoint event] filelock:break_lease_block [Tracepoint event] ... ie fib6 now is after fib and not before it. This is more consistent with how numbers are more generally sorted, such as: ... syscalls:sys_enter_renameat [Tracepoint event] syscalls:sys_enter_renameat2 [Tracepoint event] ... and so an improvement over the qsort approach. Signed-off-by:
-
Ian Rogers authored
tracing_events__opendir() allows iteration over files in <debugfs>/tracing/events but with an arbitrary sort order. Add a scandir alternative where the results are alphabetically sorted. Signed-off-by:
-
Ian Rogers authored
Add documentation to 'struct perf_pmu' and the associated structs of 'perf_pmu_alias' and 'perf_pmu_format'. Signed-off-by:
-
Ian Rogers authored
Replace usage with perf_pmu__is_hybrid(). Suggested-by:
Kan Liang <kan.liang@linux.intel.com> Signed-off-by:
-
- 14 Nov, 2022 2 commits
-
-
Namhyung Kim authored
It should have a comma after 'cpus' for socket and die aggregation mode. The output of the following command shows the issue. $ sudo perf stat -a --per-socket -x, --metric-only -I1 true Before: +--- here V time,socket,cpusGhz,insn per cycle,branch-misses of all branches, 0.000908461,S0,8,0.950,1.65,1.21, After: time,socket,cpus,GHz,insn per cycle,branch-misses of all branches, 0.000683094,S0,8,0.593,2.00,0.60, Signed-off-by: -
Namhyung Kim authored
It should not print "summary" for each event when --metric-only is set. Before: $ sudo perf stat -a --per-socket --summary -x, --metric-only true time,socket,cpusGhz,insn per cycle,branch-misses of all branches, 0.000709079,S0,8,0.893,2.40,0.45, S0,8, summary, summary, summary, summary, summary,0.893, summary,2.40, summary, summary,0.45, After: $ sudo perf stat -a --per-socket --summary -x, --metric-only true time,socket,cpusGHz,insn per cycle,branch-misses of all branches, 0.000882297,S0,8,0.598,1.64,0.64, summary,S0,8,0.598,1.64,0.64, Signed-off-by:
-
