The objdump utility has useful --prefix / --prefix-strip options to
allow changing source code file names hardcoded into executables' debug
info. Add options to 'perf report', 'perf top' and 'perf annotate',
which are then passed to objdump.
$ mkdir foo
$ echo 'main() { for (;;); }' > foo/foo.c
$ gcc -g foo/foo.c
foo/foo.c:1:1: warning: return type defaults to ‘int’ [-Wimplicit-int]
1 | main() { for (;;); }
| ^~~~
$ perf record ./a.out
^C[ perf record: Woken up 1 times to write data ]
[ perf record: Captured and wrote 0.230 MB perf.data (5721 samples) ]
$ mv foo bar
$ perf annotate
<does not show source code>
$ perf annotate --prefix=/home/ak/lsrc/git/bar --prefix-strip=5
<does show source code>
Signed-off-by: Andi Kleen <ak@linux.intel.com>
Tested-by: Jiri Olsa <jolsa@redhat.com>
LPU-Reference: 20200107210444.214071-1-andi@firstfloor.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
clock_nanosleep() accepts absolute values of expiration time when
TIMER_ABSTIME flag is set. This absolute value is inside the task's
time namespace, and has to be converted to the host's time.
There is timens_ktime_to_host() helper for converting time, but
it accepts ktime argument.
As a preparation, make hrtimer_nanosleep() accept a clock value in ktime
instead of timespec64.
Co-developed-by: Dmitry Safonov <dima@arista.com>
Signed-off-by: Andrei Vagin <avagin@openvz.org>
Signed-off-by: Dmitry Safonov <dima@arista.com>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Link: https://lore.kernel.org/r/20191112012724.250792-17-dima@arista.com
Allow user to limit output to one or more CPUs. Really helpful on
systems with a large number of cpus.
Committer testing:
# perf sched record -a sleep 1
[ perf record: Woken up 1 times to write data ]
[ perf record: Captured and wrote 1.765 MB perf.data (1412 samples) ]
[root@quaco ~]# perf sched timehist | head
Samples do not have callchains.
time cpu task name wait time sch delay run time
[tid/pid] (msec) (msec) (msec)
--------------- ------ ------------------------------ --------- --------- ---------
66307.802686 [0000] perf[13086] 0.000 0.000 0.000
66307.802700 [0000] migration/0[12] 0.000 0.001 0.014
66307.802766 [0001] perf[13086] 0.000 0.000 0.000
66307.802774 [0001] migration/1[15] 0.000 0.001 0.007
66307.802841 [0002] perf[13086] 0.000 0.000 0.000
66307.802849 [0002] migration/2[20] 0.000 0.001 0.008
66307.802913 [0003] perf[13086] 0.000 0.000 0.000
#
# perf sched timehist --cpu 2 | head
Samples do not have callchains.
time cpu task name wait time sch delay run time
[tid/pid] (msec) (msec) (msec)
--------------- ------ ------------------------------ --------- --------- ---------
66307.802841 [0002] perf[13086] 0.000 0.000 0.000
66307.802849 [0002] migration/2[20] 0.000 0.001 0.008
66307.964485 [0002] <idle> 0.000 0.000 161.635
66307.964811 [0002] CPU 0/KVM[3589/3561] 0.000 0.056 0.325
66307.965477 [0002] <idle> 0.325 0.000 0.666
66307.965553 [0002] CPU 0/KVM[3589/3561] 0.666 0.024 0.076
66307.966456 [0002] <idle> 0.076 0.000 0.903
#
Signed-off-by: David Ahern <dsahern@gmail.com>
Tested-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Jiri Olsa <jolsa@kernel.org>
Cc: Namhyung Kim <namhyung@kernel.org>
Link: http://lore.kernel.org/lkml/20191204173925.66976-1-dsahern@kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Use struct mmap_cpu_mask type for the tool's thread and mmap data
buffers to overcome current 1024 CPUs mask size limitation of cpu_set_t
type.
Currently glibc's cpu_set_t type has an internal mask size limit of 1024
CPUs.
Moving to the 'struct mmap_cpu_mask' type allows overcoming that limit.
The tools bitmap API is used to manipulate objects of 'struct mmap_cpu_mask'
type.
Committer notes:
To print the 'nbits' struct member we must use %zd, since it is a
size_t, this fixes the build in some toolchains/arches.
Reported-by: Andi Kleen <ak@linux.intel.com>
Signed-off-by: Alexey Budankov <alexey.budankov@linux.intel.com>
Acked-by: Jiri Olsa <jolsa@redhat.com>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Link: http://lore.kernel.org/lkml/96d7e2ff-ce8b-c1e0-d52c-aa59ea96f0ea@linux.intel.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
When a map is create to represent the main kernel area (vmlinux) with
map__new2() we allocate an extra area to store a pointer to the 'struct
maps' for the kernel maps, so that we can access that struct when
loading ELF files or kallsyms, as we will need to split it in multiple
maps, one per kernel module or ELF section (such as ".init.text").
So when map->dso->kernel is non-zero, it is expected that
map__kmap(map)->kmaps to be set to the tree of kernel maps (modules,
chunks of the main kernel, bpf progs put in place via
PERF_RECORD_KSYMBOL, the main kernel).
This was not the case when we were splitting the main kernel into chunks
for its ELF sections, which ended up making 'perf report --children'
processing a perf.data file with callchains to trip on
__map__is_kernel(), when we press ENTER to see the popup menu for main
histogram entries that starts at a symbol in the ".init.text" ELF
section, e.g.:
- 8.83% 0.00% swapper [kernel.vmlinux].init.text [k] start_kernel
start_kernel
cpu_startup_entry
do_idle
cpuidle_enter
cpuidle_enter_state
intel_idle
Fix it.
Reviewed-by: Jiri Olsa <jolsa@kernel.org>
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: Namhyung Kim <namhyung@kernel.org>
Link: https://lore.kernel.org/lkml/20191218190120.GB13282@kernel.org/
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
We observed an issue that was some extra columns displayed after switching
perf data file in browser. The steps to reproduce:
1. perf record -a -e cycles,instructions -- sleep 3
2. perf report --group
3. In browser, we use hotkey 's' to switch to another perf.data
4. Now in browser, the extra columns 'Self' and 'Children' are displayed.
The issue is setup_sorting() executed again after repeat path, so dimensions
are added again.
This patch checks the last key returned from __cmd_report(). If it's
K_SWITCH_INPUT_DATA, skips the setup_sorting().
Fixes: ad0de0971b ("perf report: Enable the runtime switching of perf data file")
Signed-off-by: Jin Yao <yao.jin@linux.intel.com>
Tested-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Acked-by: Jiri Olsa <jolsa@redhat.com>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Feng Tang <feng.tang@intel.com>
Cc: Jin Yao <yao.jin@intel.com>
Cc: Kan Liang <kan.liang@linux.intel.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Link: http://lore.kernel.org/lkml/20191220013722.20592-1-yao.jin@linux.intel.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Before this patch, perf expected that there might be NPROC*4 unique
cache entries at max, however, it also expected that some of them would
be shared and/or of the same size, thus the final number of entries
would be reduced to be lower than NPROC*4. In case the number of entries
hadn't been reduced (was NPROC*4), the warning was printed.
However, some systems might have unusual cache topology, such as the
following two-processor KVM guest:
cpu level shared_cpu_list size
0 1 0 32K
0 1 0 64K
0 2 0 512K
0 3 0 8192K
1 1 1 32K
1 1 1 64K
1 2 1 512K
1 3 1 8192K
This KVM guest has 8 (NPROC*4) unique cache entries, which used to make
perf printing the message, although there actually aren't "way too many
cpu caches".
v2: Removing unused argument.
v3: Unifying the way we obtain number of cpus.
v4: Removed '& UINT_MAX' construct which is redundant.
Signed-off-by: Michael Petlan <mpetlan@redhat.com>
Acked-by: Jiri Olsa <jolsa@redhat.com>
LPU-Reference: 20191208162056.20772-1-mpetlan@redhat.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Commit f01642e491 ("perf metricgroup: Support multiple events for
metricgroup") introduced support for multiple events in a metric group.
But with the current upstream, metric events names are not printed
properly
In power9 platform:
command:# ./perf stat --metric-only -M translation -C 0 -I 1000 sleep 2
1.000208486
2.000368863
2.001400558
Similarly in skylake platform:
command:./perf stat --metric-only -M Power -I 1000
1.000579994
2.002189493
With current upstream version, issue is with event name comparison logic
in find_evsel_group(). Current logic is to compare events belonging to a
metric group to the events in perf_evlist. Since the break statement is
missing in the loop used for comparison between metric group and
perf_evlist events, the loop continues to execute even after getting a
pattern match, and end up in discarding the matches.
Incase of single metric event belongs to metric group, its working fine,
because in case of single event once it compare all events it reaches to
end of perf_evlist.
Example for single metric event in power9 platform:
command:# ./perf stat --metric-only -M branches_per_inst -I 1000 sleep 1
1.000094653 0.2
1.001337059 0.0
This patch fixes the issue by making sure once we found all events
belongs to that metric event matched in find_evsel_group(), we
successfully break from that loop by adding corresponding condition.
With this patch:
In power9 platform:
command:# ./perf stat --metric-only -M translation -C 0 -I 1000 sleep 2
result:#
time derat_4k_miss_rate_percent derat_4k_miss_ratio derat_miss_ratio derat_64k_miss_rate_percent derat_64k_miss_ratio dslb_miss_rate_percent islb_miss_rate_percent
1.000135672 0.0 0.3 1.0 0.0 0.2 0.0 0.0
2.000380617 0.0 0.0 0.0 0.0 0.0 0.0 0.0
command:# ./perf stat --metric-only -M Power -I 1000
Similarly in skylake platform:
result:#
time Turbo_Utilization C3_Core_Residency C6_Core_Residency C7_Core_Residency C2_Pkg_Residency C3_Pkg_Residency C6_Pkg_Residency C7_Pkg_Residency
1.000563580 0.3 0.0 2.6 44.2 21.9 0.0 0.0 0.0
2.002235027 0.4 0.0 2.7 43.0 20.7 0.0 0.0 0.0
Committer testing:
Before:
[root@seventh ~]# perf stat --metric-only -M Power -I 1000
# time
1.000383223
2.001168182
3.001968545
4.002741200
5.003442022
^C 5.777687244
[root@seventh ~]#
After the patch:
[root@seventh ~]# perf stat --metric-only -M Power -I 1000
# time Turbo_Utilization C3_Core_Residency C6_Core_Residency C7_Core_Residency C2_Pkg_Residency C3_Pkg_Residency C6_Pkg_Residency C7_Pkg_Residency
1.000406577 0.4 0.1 1.4 97.0 0.0 0.0 0.0 0.0
2.001481572 0.3 0.0 0.6 97.9 0.0 0.0 0.0 0.0
3.002332585 0.2 0.0 1.0 97.5 0.0 0.0 0.0 0.0
4.003196624 0.2 0.0 0.3 98.6 0.0 0.0 0.0 0.0
5.004063851 0.3 0.0 0.7 97.7 0.0 0.0 0.0 0.0
^C 5.471260276 0.2 0.0 0.5 49.3 0.0 0.0 0.0 0.0
[root@seventh ~]#
[root@seventh ~]# dmesg | grep -i skylake
[ 0.187807] Performance Events: PEBS fmt3+, Skylake events, 32-deep LBR, full-width counters, Intel PMU driver.
[root@seventh ~]#
Fixes: f01642e491 ("perf metricgroup: Support multiple events for metricgroup")
Signed-off-by: Kajol Jain <kjain@linux.ibm.com>
Reviewed-by: Ravi Bangoria <ravi.bangoria@linux.ibm.com>
Tested-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Anju T Sudhakar <anju@linux.vnet.ibm.com>
Cc: Jin Yao <yao.jin@linux.intel.com>
Cc: Jiri Olsa <jolsa@kernel.org>
Cc: Kan Liang <kan.liang@linux.intel.com>
Cc: Madhavan Srinivasan <maddy@linux.vnet.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Link: http://lore.kernel.org/lkml/20191120084059.24458-1-kjain@linux.ibm.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
'perf top' stopped working on hw architectures that do not provide a
get_cpuid() implementation and thus fallback to the weak get_cpuid()
default function.
This is done because at annotation time we may need it in the arch
specific annotation init routine, but that is only being used by arches
that do provide a get_cpuid() implementation:
$ find tools/ -name "*.[ch]" | xargs grep 'evlist->env'
tools/perf/builtin-top.c: top.evlist->env = &perf_env;
tools/perf/util/evsel.c: return evsel->evlist->env;
tools/perf/util/s390-cpumsf.c: sf->machine_type = s390_cpumsf_get_type(session->evlist->env->cpuid);
tools/perf/util/header.c: session->evlist->env = &header->env;
tools/perf/util/sample-raw.c: const char *arch_pf = perf_env__arch(evlist->env);
$
$ find tools/perf/arch -name "*.[ch]" | xargs grep -w get_cpuid
tools/perf/arch/x86/util/auxtrace.c: ret = get_cpuid(buffer, sizeof(buffer));
tools/perf/arch/x86/util/header.c:get_cpuid(char *buffer, size_t sz)
tools/perf/arch/powerpc/util/header.c:get_cpuid(char *buffer, size_t sz)
tools/perf/arch/s390/util/header.c: * Implementation of get_cpuid().
tools/perf/arch/s390/util/header.c:int get_cpuid(char *buffer, size_t sz)
tools/perf/arch/s390/util/header.c: if (buf && get_cpuid(buf, 128))
$
For 'report' or 'script', i.e. tools working on perf.data files, that is
setup while reading the header, its just top that needs to explicitely
read it at tool start.
Fixes: 608127f737 ("perf top: Initialize perf_env->cpuid, needed by the per arch annotation init routine")
Reported-by: John Garry <john.garry@huawei.com>
Analysed-by: Jiri Olsa <jolsa@kernel.org>
Reviewed-by: Mark Rutland <mark.rutland@arm.com>
Tested-by: Mark Rutland <mark.rutland@arm.com>
Tested-by: John Garry <john.garry@huawei.com> # arm64
Acked-by: Jiri Olsa <jolsa@redhat.com>
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Will Deacon <will@kernel.org>
Link: https://lkml.kernel.org/n/tip-lxwjr0cd2eggzx04a780ffrv@git.kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
To pick up BPF fixes to allow a clean 'make -C tools/perf build-test':
7c3977d1e8 libbpf: Fix sym->st_value print on 32-bit arches
1fd450f992 libbpf: Fix up generation of bpf_helper_defs.h
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
The ID index event is used when decoding, but can result in the
following error:
$ perf record --aux-sample -e '{intel_pt//,branch-misses}:u' ls
$ perf inject -i perf.data -o perf.data.inj --itrace=be
$ perf script -i perf.data.inj
0x1020 [0x410]: failed to process type: 69 [No such file or directory]
Fix by having 'perf inject' drop the ID index event.
Fixes: c0a6de06c4 ("perf record: Add support for AUX area sampling")
Signed-off-by: Adrian Hunter <adrian.hunter@intel.com>
Cc: Jiri Olsa <jolsa@redhat.com>
Link: http://lore.kernel.org/lkml/20191204120800.8138-1-adrian.hunter@intel.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
If perf.data is recorded without -d, don't allow user to use --mem-mode
with 'perf report'. symbol_daddr and phys_daddr can be recorded
separately and may be present in the perf.data but at the report time
they are associated with mem-mode fields and thus this restriction
applies to them as well.
Before:
$ perf record ls
$ perf report --mem-mode --stdio
# Overhead Local Weight Memory access Symbol
# ........ ............ ............. .......................
55.56% 0 N/A [k] 0xffffffff81a00ae7
After:
$ perf report --mem-mode --stdio
Error:
Selected --mem-mode but no mem data. Did you call perf record without -d?
Suggested-by: Arnaldo Carvalho de Melo <acme@kernel.org>
Signed-off-by: Ravi Bangoria <ravi.bangoria@linux.ibm.com>
Acked-by: Jiri Olsa <jolsa@kernel.org>
Tested-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Jin Yao <yao.jin@linux.intel.com>
Cc: Kan Liang <kan.liang@linux.intel.com>
Cc: Mark Rutland <mark.rutland@arm.com>
Cc: Namhyung Kim <namhyung@kernel.org>
Link: http://lore.kernel.org/lkml/20191114132213.5419-4-ravi.bangoria@linux.ibm.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Daniel Borkmann says:
====================
pull-request: bpf 2019-12-02
The following pull-request contains BPF updates for your *net* tree.
We've added 10 non-merge commits during the last 6 day(s) which contain
a total of 10 files changed, 60 insertions(+), 51 deletions(-).
The main changes are:
1) Fix vmlinux BTF generation for binutils pre v2.25, from Stanislav Fomichev.
2) Fix libbpf global variable relocation to take symbol's st_value offset
into account, from Andrii Nakryiko.
3) Fix libbpf build on powerpc where check_abi target fails due to different
readelf output format, from Aurelien Jarno.
4) Don't set BPF insns RO for the case when they are JITed in order to avoid
fragmenting the direct map, from Daniel Borkmann.
5) Fix static checker warning in btf_distill_func_proto() as well as a build
error due to empty enum when BPF is compiled out, from Alexei Starovoitov.
6) Fix up generation of bpf_helper_defs.h for perf, from Arnaldo Carvalho de Melo.
====================
Signed-off-by: David S. Miller <davem@davemloft.net>
The 'perf kvm' subcommand has options that it in turn passes to other
perf subcommands such as 'report' and 'record', particularly -i and -o
end up setting the same variable that will then be used for 'record's -o
and report '-i', which ends up being confusing, leading some to think
that both -i and -o can be used with 'report'.
Improve the man page to state that -i is used with the post-processing
subcommands while -o is used just with 'record' and that to save the
output of 'report' one should simply redirect its output to a file.
Noticed while reading the https://www.linux-kvm.org/page/Perf_events
page.
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: Jiri Olsa <jolsa@kernel.org>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Steve Dickson <steved@redhat.com>
Cc: William Cohen <wcohen@redhat.com>
Link: https://lkml.kernel.org/n/tip-tclbttvmgtm525fvmh85f7d9@git.kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
And update linux/linkage.h, which requires in turn that we make these
files switch from ENTRY()/ENDPROC() to SYM_FUNC_START()/SYM_FUNC_END():
tools/perf/arch/arm64/tests/regs_load.S
tools/perf/arch/arm/tests/regs_load.S
tools/perf/arch/powerpc/tests/regs_load.S
tools/perf/arch/x86/tests/regs_load.S
We also need to switch SYM_FUNC_START_LOCAL() to SYM_FUNC_START() for
the functions used directly by 'perf bench', and update
tools/perf/check_headers.sh to ignore those changes when checking if the
kernel original files drifted from the copies we carry.
This is to get the changes from:
6dcc5627f6 ("x86/asm: Change all ENTRY+ENDPROC to SYM_FUNC_*")
ef1e03152c ("x86/asm: Make some functions local")
e9b9d020c4 ("x86/asm: Annotate aliases")
And address these tools/perf build warnings:
Warning: Kernel ABI header at 'tools/arch/x86/lib/memcpy_64.S' differs from latest version at 'arch/x86/lib/memcpy_64.S'
diff -u tools/arch/x86/lib/memcpy_64.S arch/x86/lib/memcpy_64.S
Warning: Kernel ABI header at 'tools/arch/x86/lib/memset_64.S' differs from latest version at 'arch/x86/lib/memset_64.S'
diff -u tools/arch/x86/lib/memset_64.S arch/x86/lib/memset_64.S
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: Borislav Petkov <bp@suse.de>
Cc: Jiri Olsa <jolsa@kernel.org>
Cc: Jiri Slaby <jslaby@suse.cz>
Cc: Namhyung Kim <namhyung@kernel.org>
Link: https://lkml.kernel.org/n/tip-tay3l8x8k11p7y3qcpqh9qh5@git.kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
I forgot to fill in the map_symbol->maps field in append_inlines() which
then makes code down the line segfault when trying to deref it.
It doesn't make any sense to have an addr_location with its 'map' member
not NULL while its 'maps' is NULL, after all al->maps is where al->map
is in.
It is done that way so that we don't have to have in each 'struct map' a
pointer to the 'struct maps' it is in, as we had in the past when we
would have 'map->mg', before 'struct maps' was combined with 'struct
map_groups', because there was always a one-to-one relationship for
these structs.
This fixes a segfault when processing DWARF callgraphs in 'perf report'.
Reported-by: Jiri Olsa <jolsa@kernel.org>
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: Namhyung Kim <namhyung@kernel.org>
Fixes: 08f6680e62 ("perf tools: Add a 'struct map_groups' pointer to 'struct map_symbol'")
Link: http://lore.kernel.org/lkml/20191129160631.GD26963@kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Restructure event enabling/disabling to use affinity, which
minimizes the number of IPIs needed.
Before on a large test case with 94 CPUs:
% time seconds usecs/call calls errors syscall
------ ----------- ----------- --------- --------- ----------------
54.65 1.899986 22 84812 660 ioctl
after:
39.21 0.930451 10 84796 644 ioctl
Signed-off-by: Andi Kleen <ak@linux.intel.com>
Acked-by: Jiri Olsa <jolsa@kernel.org>
Link: http://lore.kernel.org/lkml/20191121001522.180827-13-andi@firstfloor.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Restructure event reading to use affinity to minimize the number of IPIs
needed.
Before on a large test case with 94 CPUs:
% time seconds usecs/call calls errors syscall
------ ----------- ----------- --------- --------- ----------------
3.16 0.106079 4 22082 read
After:
3.43 0.081295 3 22082 read
Signed-off-by: Andi Kleen <ak@linux.intel.com>
Acked-by: Jiri Olsa <jolsa@kernel.org>
Link: http://lore.kernel.org/lkml/20191121001522.180827-11-andi@firstfloor.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Restructure the event opening in perf stat to cycle through the events
by CPU after setting affinity to that CPU.
This eliminates IPI overhead in the perf API.
We have to loop through the CPU in the outter builtin-stat code instead
of leaving that to low level functions.
It has to change the weak group fallback strategy slightly. Since we
cannot easily undo the opens for other CPUs move the weak group retry to
a separate loop.
Before with a large test case with 94 CPUs:
% time seconds usecs/call calls errors syscall
------ ----------- ----------- --------- --------- ----------------
42.75 4.050910 67 60046 110 perf_event_open
After:
26.86 0.944396 16 58069 110 perf_event_open
(the number changes slightly because the weak group retries
work differently and the test case relies on weak groups)
Committer notes:
Added one of the hunks in a patch provided by Andi after I noticed that
the "event times" 'perf test' entry was segfaulting.
Signed-off-by: Andi Kleen <ak@linux.intel.com>
Acked-by: Jiri Olsa <jolsa@kernel.org>
Link: http://lore.kernel.org/lkml/20191121001522.180827-10-andi@firstfloor.org
Link: http://lore.kernel.org/lkml/20191127232657.GL84886@tassilo.jf.intel.com # Fix
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Closing a perf fd can also trigger an IPI to the target CPU.
Use the same affinity technique as we use for reading/enabling events to
closing to optimize the CPU transitions.
Before on a large test case with 94 CPUs:
% time seconds usecs/call calls errors syscall
------ ----------- ----------- --------- --------- ----------------
32.56 3.085463 50 61483 close
After:
10.54 0.735704 11 61485 close
Signed-off-by: Andi Kleen <ak@linux.intel.com>
Acked-by: Jiri Olsa <jolsa@kernel.org>
Link: http://lore.kernel.org/lkml/20191121001522.180827-8-andi@firstfloor.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Maintain a cpumap in the evlist that is the union of all the cpus of the
events.
This needs a cpumap merge operation, which is added together with tests.
v2:
Add tests for cpu map merge
Fix handling of duplicates
Rename _update to _merge
Factor out sorting.
Fix handling of NULL maps in merge
v3:
Add comments and empty lines to _merge
Committer testing:
# perf test "Merge cpu map"
52: Merge cpu map : Ok
#
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://lore.kernel.org/lkml/20191121001522.180827-5-andi@firstfloor.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
