perf tools: Fix sparse CPU numbering related bugs
At present, the perf subcommands that do system-wide monitoring
(perf stat, perf record and perf top) don't work properly unless
the online cpus are numbered 0, 1, ..., N-1. These tools ask
for the number of online cpus with sysconf(_SC_NPROCESSORS_ONLN)
and then try to create events for cpus 0, 1, ..., N-1.
This creates problems for systems where the online cpus are
numbered sparsely. For example, a POWER6 system in
single-threaded mode (i.e. only running 1 hardware thread per
core) will have only even-numbered cpus online.
This fixes the problem by reading the /sys/devices/system/cpu/online
file to find out which cpus are online. The code that does that is in
tools/perf/util/cpumap.[ch], and consists of a read_cpu_map()
function that sets up a cpumap[] array and returns the number of
online cpus. If /sys/devices/system/cpu/online can't be read or
can't be parsed successfully, it falls back to using sysconf to
ask how many cpus are online and sets up an identity map in cpumap[].
The perf record, perf stat and perf top code then calls
read_cpu_map() in the system-wide monitoring case (instead of
sysconf) and uses cpumap[] to get the cpu numbers to pass to
perf_event_open.
Signed-off-by: Paul Mackerras <paulus@samba.org>
Cc: Anton Blanchard <anton@samba.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Arnaldo Carvalho de Melo <acme@infradead.org>
LKML-Reference: <20100310093609.GA3959@brick.ozlabs.ibm.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
2010-03-10 16:36:09 +07:00
|
|
|
#include "util.h"
|
2013-12-09 23:14:24 +07:00
|
|
|
#include <api/fs/fs.h>
|
perf tools: Fix sparse CPU numbering related bugs
At present, the perf subcommands that do system-wide monitoring
(perf stat, perf record and perf top) don't work properly unless
the online cpus are numbered 0, 1, ..., N-1. These tools ask
for the number of online cpus with sysconf(_SC_NPROCESSORS_ONLN)
and then try to create events for cpus 0, 1, ..., N-1.
This creates problems for systems where the online cpus are
numbered sparsely. For example, a POWER6 system in
single-threaded mode (i.e. only running 1 hardware thread per
core) will have only even-numbered cpus online.
This fixes the problem by reading the /sys/devices/system/cpu/online
file to find out which cpus are online. The code that does that is in
tools/perf/util/cpumap.[ch], and consists of a read_cpu_map()
function that sets up a cpumap[] array and returns the number of
online cpus. If /sys/devices/system/cpu/online can't be read or
can't be parsed successfully, it falls back to using sysconf to
ask how many cpus are online and sets up an identity map in cpumap[].
The perf record, perf stat and perf top code then calls
read_cpu_map() in the system-wide monitoring case (instead of
sysconf) and uses cpumap[] to get the cpu numbers to pass to
perf_event_open.
Signed-off-by: Paul Mackerras <paulus@samba.org>
Cc: Anton Blanchard <anton@samba.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Arnaldo Carvalho de Melo <acme@infradead.org>
LKML-Reference: <20100310093609.GA3959@brick.ozlabs.ibm.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
2010-03-10 16:36:09 +07:00
|
|
|
#include "../perf.h"
|
|
|
|
#include "cpumap.h"
|
|
|
|
#include <assert.h>
|
|
|
|
#include <stdio.h>
|
2013-02-14 19:57:27 +07:00
|
|
|
#include <stdlib.h>
|
perf tools: Fix sparse CPU numbering related bugs
At present, the perf subcommands that do system-wide monitoring
(perf stat, perf record and perf top) don't work properly unless
the online cpus are numbered 0, 1, ..., N-1. These tools ask
for the number of online cpus with sysconf(_SC_NPROCESSORS_ONLN)
and then try to create events for cpus 0, 1, ..., N-1.
This creates problems for systems where the online cpus are
numbered sparsely. For example, a POWER6 system in
single-threaded mode (i.e. only running 1 hardware thread per
core) will have only even-numbered cpus online.
This fixes the problem by reading the /sys/devices/system/cpu/online
file to find out which cpus are online. The code that does that is in
tools/perf/util/cpumap.[ch], and consists of a read_cpu_map()
function that sets up a cpumap[] array and returns the number of
online cpus. If /sys/devices/system/cpu/online can't be read or
can't be parsed successfully, it falls back to using sysconf to
ask how many cpus are online and sets up an identity map in cpumap[].
The perf record, perf stat and perf top code then calls
read_cpu_map() in the system-wide monitoring case (instead of
sysconf) and uses cpumap[] to get the cpu numbers to pass to
perf_event_open.
Signed-off-by: Paul Mackerras <paulus@samba.org>
Cc: Anton Blanchard <anton@samba.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Arnaldo Carvalho de Melo <acme@infradead.org>
LKML-Reference: <20100310093609.GA3959@brick.ozlabs.ibm.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
2010-03-10 16:36:09 +07:00
|
|
|
|
2011-01-04 02:49:48 +07:00
|
|
|
static struct cpu_map *cpu_map__default_new(void)
|
perf tools: Fix sparse CPU numbering related bugs
At present, the perf subcommands that do system-wide monitoring
(perf stat, perf record and perf top) don't work properly unless
the online cpus are numbered 0, 1, ..., N-1. These tools ask
for the number of online cpus with sysconf(_SC_NPROCESSORS_ONLN)
and then try to create events for cpus 0, 1, ..., N-1.
This creates problems for systems where the online cpus are
numbered sparsely. For example, a POWER6 system in
single-threaded mode (i.e. only running 1 hardware thread per
core) will have only even-numbered cpus online.
This fixes the problem by reading the /sys/devices/system/cpu/online
file to find out which cpus are online. The code that does that is in
tools/perf/util/cpumap.[ch], and consists of a read_cpu_map()
function that sets up a cpumap[] array and returns the number of
online cpus. If /sys/devices/system/cpu/online can't be read or
can't be parsed successfully, it falls back to using sysconf to
ask how many cpus are online and sets up an identity map in cpumap[].
The perf record, perf stat and perf top code then calls
read_cpu_map() in the system-wide monitoring case (instead of
sysconf) and uses cpumap[] to get the cpu numbers to pass to
perf_event_open.
Signed-off-by: Paul Mackerras <paulus@samba.org>
Cc: Anton Blanchard <anton@samba.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Arnaldo Carvalho de Melo <acme@infradead.org>
LKML-Reference: <20100310093609.GA3959@brick.ozlabs.ibm.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
2010-03-10 16:36:09 +07:00
|
|
|
{
|
2011-01-04 02:49:48 +07:00
|
|
|
struct cpu_map *cpus;
|
|
|
|
int nr_cpus;
|
perf tools: Fix sparse CPU numbering related bugs
At present, the perf subcommands that do system-wide monitoring
(perf stat, perf record and perf top) don't work properly unless
the online cpus are numbered 0, 1, ..., N-1. These tools ask
for the number of online cpus with sysconf(_SC_NPROCESSORS_ONLN)
and then try to create events for cpus 0, 1, ..., N-1.
This creates problems for systems where the online cpus are
numbered sparsely. For example, a POWER6 system in
single-threaded mode (i.e. only running 1 hardware thread per
core) will have only even-numbered cpus online.
This fixes the problem by reading the /sys/devices/system/cpu/online
file to find out which cpus are online. The code that does that is in
tools/perf/util/cpumap.[ch], and consists of a read_cpu_map()
function that sets up a cpumap[] array and returns the number of
online cpus. If /sys/devices/system/cpu/online can't be read or
can't be parsed successfully, it falls back to using sysconf to
ask how many cpus are online and sets up an identity map in cpumap[].
The perf record, perf stat and perf top code then calls
read_cpu_map() in the system-wide monitoring case (instead of
sysconf) and uses cpumap[] to get the cpu numbers to pass to
perf_event_open.
Signed-off-by: Paul Mackerras <paulus@samba.org>
Cc: Anton Blanchard <anton@samba.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Arnaldo Carvalho de Melo <acme@infradead.org>
LKML-Reference: <20100310093609.GA3959@brick.ozlabs.ibm.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
2010-03-10 16:36:09 +07:00
|
|
|
|
|
|
|
nr_cpus = sysconf(_SC_NPROCESSORS_ONLN);
|
2011-01-04 02:49:48 +07:00
|
|
|
if (nr_cpus < 0)
|
|
|
|
return NULL;
|
|
|
|
|
|
|
|
cpus = malloc(sizeof(*cpus) + nr_cpus * sizeof(int));
|
|
|
|
if (cpus != NULL) {
|
|
|
|
int i;
|
|
|
|
for (i = 0; i < nr_cpus; ++i)
|
|
|
|
cpus->map[i] = i;
|
perf tools: Fix sparse CPU numbering related bugs
At present, the perf subcommands that do system-wide monitoring
(perf stat, perf record and perf top) don't work properly unless
the online cpus are numbered 0, 1, ..., N-1. These tools ask
for the number of online cpus with sysconf(_SC_NPROCESSORS_ONLN)
and then try to create events for cpus 0, 1, ..., N-1.
This creates problems for systems where the online cpus are
numbered sparsely. For example, a POWER6 system in
single-threaded mode (i.e. only running 1 hardware thread per
core) will have only even-numbered cpus online.
This fixes the problem by reading the /sys/devices/system/cpu/online
file to find out which cpus are online. The code that does that is in
tools/perf/util/cpumap.[ch], and consists of a read_cpu_map()
function that sets up a cpumap[] array and returns the number of
online cpus. If /sys/devices/system/cpu/online can't be read or
can't be parsed successfully, it falls back to using sysconf to
ask how many cpus are online and sets up an identity map in cpumap[].
The perf record, perf stat and perf top code then calls
read_cpu_map() in the system-wide monitoring case (instead of
sysconf) and uses cpumap[] to get the cpu numbers to pass to
perf_event_open.
Signed-off-by: Paul Mackerras <paulus@samba.org>
Cc: Anton Blanchard <anton@samba.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Arnaldo Carvalho de Melo <acme@infradead.org>
LKML-Reference: <20100310093609.GA3959@brick.ozlabs.ibm.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
2010-03-10 16:36:09 +07:00
|
|
|
|
2011-01-04 02:49:48 +07:00
|
|
|
cpus->nr = nr_cpus;
|
|
|
|
}
|
perf tools: Fix sparse CPU numbering related bugs
At present, the perf subcommands that do system-wide monitoring
(perf stat, perf record and perf top) don't work properly unless
the online cpus are numbered 0, 1, ..., N-1. These tools ask
for the number of online cpus with sysconf(_SC_NPROCESSORS_ONLN)
and then try to create events for cpus 0, 1, ..., N-1.
This creates problems for systems where the online cpus are
numbered sparsely. For example, a POWER6 system in
single-threaded mode (i.e. only running 1 hardware thread per
core) will have only even-numbered cpus online.
This fixes the problem by reading the /sys/devices/system/cpu/online
file to find out which cpus are online. The code that does that is in
tools/perf/util/cpumap.[ch], and consists of a read_cpu_map()
function that sets up a cpumap[] array and returns the number of
online cpus. If /sys/devices/system/cpu/online can't be read or
can't be parsed successfully, it falls back to using sysconf to
ask how many cpus are online and sets up an identity map in cpumap[].
The perf record, perf stat and perf top code then calls
read_cpu_map() in the system-wide monitoring case (instead of
sysconf) and uses cpumap[] to get the cpu numbers to pass to
perf_event_open.
Signed-off-by: Paul Mackerras <paulus@samba.org>
Cc: Anton Blanchard <anton@samba.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Arnaldo Carvalho de Melo <acme@infradead.org>
LKML-Reference: <20100310093609.GA3959@brick.ozlabs.ibm.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
2010-03-10 16:36:09 +07:00
|
|
|
|
2011-01-04 02:49:48 +07:00
|
|
|
return cpus;
|
perf tools: Fix sparse CPU numbering related bugs
At present, the perf subcommands that do system-wide monitoring
(perf stat, perf record and perf top) don't work properly unless
the online cpus are numbered 0, 1, ..., N-1. These tools ask
for the number of online cpus with sysconf(_SC_NPROCESSORS_ONLN)
and then try to create events for cpus 0, 1, ..., N-1.
This creates problems for systems where the online cpus are
numbered sparsely. For example, a POWER6 system in
single-threaded mode (i.e. only running 1 hardware thread per
core) will have only even-numbered cpus online.
This fixes the problem by reading the /sys/devices/system/cpu/online
file to find out which cpus are online. The code that does that is in
tools/perf/util/cpumap.[ch], and consists of a read_cpu_map()
function that sets up a cpumap[] array and returns the number of
online cpus. If /sys/devices/system/cpu/online can't be read or
can't be parsed successfully, it falls back to using sysconf to
ask how many cpus are online and sets up an identity map in cpumap[].
The perf record, perf stat and perf top code then calls
read_cpu_map() in the system-wide monitoring case (instead of
sysconf) and uses cpumap[] to get the cpu numbers to pass to
perf_event_open.
Signed-off-by: Paul Mackerras <paulus@samba.org>
Cc: Anton Blanchard <anton@samba.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Arnaldo Carvalho de Melo <acme@infradead.org>
LKML-Reference: <20100310093609.GA3959@brick.ozlabs.ibm.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
2010-03-10 16:36:09 +07:00
|
|
|
}
|
|
|
|
|
2011-01-04 02:49:48 +07:00
|
|
|
static struct cpu_map *cpu_map__trim_new(int nr_cpus, int *tmp_cpus)
|
perf tools: Fix sparse CPU numbering related bugs
At present, the perf subcommands that do system-wide monitoring
(perf stat, perf record and perf top) don't work properly unless
the online cpus are numbered 0, 1, ..., N-1. These tools ask
for the number of online cpus with sysconf(_SC_NPROCESSORS_ONLN)
and then try to create events for cpus 0, 1, ..., N-1.
This creates problems for systems where the online cpus are
numbered sparsely. For example, a POWER6 system in
single-threaded mode (i.e. only running 1 hardware thread per
core) will have only even-numbered cpus online.
This fixes the problem by reading the /sys/devices/system/cpu/online
file to find out which cpus are online. The code that does that is in
tools/perf/util/cpumap.[ch], and consists of a read_cpu_map()
function that sets up a cpumap[] array and returns the number of
online cpus. If /sys/devices/system/cpu/online can't be read or
can't be parsed successfully, it falls back to using sysconf to
ask how many cpus are online and sets up an identity map in cpumap[].
The perf record, perf stat and perf top code then calls
read_cpu_map() in the system-wide monitoring case (instead of
sysconf) and uses cpumap[] to get the cpu numbers to pass to
perf_event_open.
Signed-off-by: Paul Mackerras <paulus@samba.org>
Cc: Anton Blanchard <anton@samba.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Arnaldo Carvalho de Melo <acme@infradead.org>
LKML-Reference: <20100310093609.GA3959@brick.ozlabs.ibm.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
2010-03-10 16:36:09 +07:00
|
|
|
{
|
2011-01-04 02:49:48 +07:00
|
|
|
size_t payload_size = nr_cpus * sizeof(int);
|
|
|
|
struct cpu_map *cpus = malloc(sizeof(*cpus) + payload_size);
|
|
|
|
|
|
|
|
if (cpus != NULL) {
|
|
|
|
cpus->nr = nr_cpus;
|
|
|
|
memcpy(cpus->map, tmp_cpus, payload_size);
|
|
|
|
}
|
|
|
|
|
|
|
|
return cpus;
|
|
|
|
}
|
|
|
|
|
2012-09-10 14:53:50 +07:00
|
|
|
struct cpu_map *cpu_map__read(FILE *file)
|
2011-01-04 02:49:48 +07:00
|
|
|
{
|
|
|
|
struct cpu_map *cpus = NULL;
|
perf tools: Fix sparse CPU numbering related bugs
At present, the perf subcommands that do system-wide monitoring
(perf stat, perf record and perf top) don't work properly unless
the online cpus are numbered 0, 1, ..., N-1. These tools ask
for the number of online cpus with sysconf(_SC_NPROCESSORS_ONLN)
and then try to create events for cpus 0, 1, ..., N-1.
This creates problems for systems where the online cpus are
numbered sparsely. For example, a POWER6 system in
single-threaded mode (i.e. only running 1 hardware thread per
core) will have only even-numbered cpus online.
This fixes the problem by reading the /sys/devices/system/cpu/online
file to find out which cpus are online. The code that does that is in
tools/perf/util/cpumap.[ch], and consists of a read_cpu_map()
function that sets up a cpumap[] array and returns the number of
online cpus. If /sys/devices/system/cpu/online can't be read or
can't be parsed successfully, it falls back to using sysconf to
ask how many cpus are online and sets up an identity map in cpumap[].
The perf record, perf stat and perf top code then calls
read_cpu_map() in the system-wide monitoring case (instead of
sysconf) and uses cpumap[] to get the cpu numbers to pass to
perf_event_open.
Signed-off-by: Paul Mackerras <paulus@samba.org>
Cc: Anton Blanchard <anton@samba.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Arnaldo Carvalho de Melo <acme@infradead.org>
LKML-Reference: <20100310093609.GA3959@brick.ozlabs.ibm.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
2010-03-10 16:36:09 +07:00
|
|
|
int nr_cpus = 0;
|
2011-01-04 02:49:48 +07:00
|
|
|
int *tmp_cpus = NULL, *tmp;
|
|
|
|
int max_entries = 0;
|
perf tools: Fix sparse CPU numbering related bugs
At present, the perf subcommands that do system-wide monitoring
(perf stat, perf record and perf top) don't work properly unless
the online cpus are numbered 0, 1, ..., N-1. These tools ask
for the number of online cpus with sysconf(_SC_NPROCESSORS_ONLN)
and then try to create events for cpus 0, 1, ..., N-1.
This creates problems for systems where the online cpus are
numbered sparsely. For example, a POWER6 system in
single-threaded mode (i.e. only running 1 hardware thread per
core) will have only even-numbered cpus online.
This fixes the problem by reading the /sys/devices/system/cpu/online
file to find out which cpus are online. The code that does that is in
tools/perf/util/cpumap.[ch], and consists of a read_cpu_map()
function that sets up a cpumap[] array and returns the number of
online cpus. If /sys/devices/system/cpu/online can't be read or
can't be parsed successfully, it falls back to using sysconf to
ask how many cpus are online and sets up an identity map in cpumap[].
The perf record, perf stat and perf top code then calls
read_cpu_map() in the system-wide monitoring case (instead of
sysconf) and uses cpumap[] to get the cpu numbers to pass to
perf_event_open.
Signed-off-by: Paul Mackerras <paulus@samba.org>
Cc: Anton Blanchard <anton@samba.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Arnaldo Carvalho de Melo <acme@infradead.org>
LKML-Reference: <20100310093609.GA3959@brick.ozlabs.ibm.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
2010-03-10 16:36:09 +07:00
|
|
|
int n, cpu, prev;
|
|
|
|
char sep;
|
|
|
|
|
|
|
|
sep = 0;
|
|
|
|
prev = -1;
|
|
|
|
for (;;) {
|
2012-09-10 14:53:50 +07:00
|
|
|
n = fscanf(file, "%u%c", &cpu, &sep);
|
perf tools: Fix sparse CPU numbering related bugs
At present, the perf subcommands that do system-wide monitoring
(perf stat, perf record and perf top) don't work properly unless
the online cpus are numbered 0, 1, ..., N-1. These tools ask
for the number of online cpus with sysconf(_SC_NPROCESSORS_ONLN)
and then try to create events for cpus 0, 1, ..., N-1.
This creates problems for systems where the online cpus are
numbered sparsely. For example, a POWER6 system in
single-threaded mode (i.e. only running 1 hardware thread per
core) will have only even-numbered cpus online.
This fixes the problem by reading the /sys/devices/system/cpu/online
file to find out which cpus are online. The code that does that is in
tools/perf/util/cpumap.[ch], and consists of a read_cpu_map()
function that sets up a cpumap[] array and returns the number of
online cpus. If /sys/devices/system/cpu/online can't be read or
can't be parsed successfully, it falls back to using sysconf to
ask how many cpus are online and sets up an identity map in cpumap[].
The perf record, perf stat and perf top code then calls
read_cpu_map() in the system-wide monitoring case (instead of
sysconf) and uses cpumap[] to get the cpu numbers to pass to
perf_event_open.
Signed-off-by: Paul Mackerras <paulus@samba.org>
Cc: Anton Blanchard <anton@samba.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Arnaldo Carvalho de Melo <acme@infradead.org>
LKML-Reference: <20100310093609.GA3959@brick.ozlabs.ibm.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
2010-03-10 16:36:09 +07:00
|
|
|
if (n <= 0)
|
|
|
|
break;
|
|
|
|
if (prev >= 0) {
|
2011-01-04 02:49:48 +07:00
|
|
|
int new_max = nr_cpus + cpu - prev - 1;
|
|
|
|
|
|
|
|
if (new_max >= max_entries) {
|
|
|
|
max_entries = new_max + MAX_NR_CPUS / 2;
|
|
|
|
tmp = realloc(tmp_cpus, max_entries * sizeof(int));
|
|
|
|
if (tmp == NULL)
|
|
|
|
goto out_free_tmp;
|
|
|
|
tmp_cpus = tmp;
|
|
|
|
}
|
|
|
|
|
perf tools: Fix sparse CPU numbering related bugs
At present, the perf subcommands that do system-wide monitoring
(perf stat, perf record and perf top) don't work properly unless
the online cpus are numbered 0, 1, ..., N-1. These tools ask
for the number of online cpus with sysconf(_SC_NPROCESSORS_ONLN)
and then try to create events for cpus 0, 1, ..., N-1.
This creates problems for systems where the online cpus are
numbered sparsely. For example, a POWER6 system in
single-threaded mode (i.e. only running 1 hardware thread per
core) will have only even-numbered cpus online.
This fixes the problem by reading the /sys/devices/system/cpu/online
file to find out which cpus are online. The code that does that is in
tools/perf/util/cpumap.[ch], and consists of a read_cpu_map()
function that sets up a cpumap[] array and returns the number of
online cpus. If /sys/devices/system/cpu/online can't be read or
can't be parsed successfully, it falls back to using sysconf to
ask how many cpus are online and sets up an identity map in cpumap[].
The perf record, perf stat and perf top code then calls
read_cpu_map() in the system-wide monitoring case (instead of
sysconf) and uses cpumap[] to get the cpu numbers to pass to
perf_event_open.
Signed-off-by: Paul Mackerras <paulus@samba.org>
Cc: Anton Blanchard <anton@samba.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Arnaldo Carvalho de Melo <acme@infradead.org>
LKML-Reference: <20100310093609.GA3959@brick.ozlabs.ibm.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
2010-03-10 16:36:09 +07:00
|
|
|
while (++prev < cpu)
|
2011-01-04 02:49:48 +07:00
|
|
|
tmp_cpus[nr_cpus++] = prev;
|
|
|
|
}
|
|
|
|
if (nr_cpus == max_entries) {
|
|
|
|
max_entries += MAX_NR_CPUS;
|
|
|
|
tmp = realloc(tmp_cpus, max_entries * sizeof(int));
|
|
|
|
if (tmp == NULL)
|
|
|
|
goto out_free_tmp;
|
|
|
|
tmp_cpus = tmp;
|
perf tools: Fix sparse CPU numbering related bugs
At present, the perf subcommands that do system-wide monitoring
(perf stat, perf record and perf top) don't work properly unless
the online cpus are numbered 0, 1, ..., N-1. These tools ask
for the number of online cpus with sysconf(_SC_NPROCESSORS_ONLN)
and then try to create events for cpus 0, 1, ..., N-1.
This creates problems for systems where the online cpus are
numbered sparsely. For example, a POWER6 system in
single-threaded mode (i.e. only running 1 hardware thread per
core) will have only even-numbered cpus online.
This fixes the problem by reading the /sys/devices/system/cpu/online
file to find out which cpus are online. The code that does that is in
tools/perf/util/cpumap.[ch], and consists of a read_cpu_map()
function that sets up a cpumap[] array and returns the number of
online cpus. If /sys/devices/system/cpu/online can't be read or
can't be parsed successfully, it falls back to using sysconf to
ask how many cpus are online and sets up an identity map in cpumap[].
The perf record, perf stat and perf top code then calls
read_cpu_map() in the system-wide monitoring case (instead of
sysconf) and uses cpumap[] to get the cpu numbers to pass to
perf_event_open.
Signed-off-by: Paul Mackerras <paulus@samba.org>
Cc: Anton Blanchard <anton@samba.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Arnaldo Carvalho de Melo <acme@infradead.org>
LKML-Reference: <20100310093609.GA3959@brick.ozlabs.ibm.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
2010-03-10 16:36:09 +07:00
|
|
|
}
|
2011-01-04 02:49:48 +07:00
|
|
|
|
|
|
|
tmp_cpus[nr_cpus++] = cpu;
|
perf tools: Fix sparse CPU numbering related bugs
At present, the perf subcommands that do system-wide monitoring
(perf stat, perf record and perf top) don't work properly unless
the online cpus are numbered 0, 1, ..., N-1. These tools ask
for the number of online cpus with sysconf(_SC_NPROCESSORS_ONLN)
and then try to create events for cpus 0, 1, ..., N-1.
This creates problems for systems where the online cpus are
numbered sparsely. For example, a POWER6 system in
single-threaded mode (i.e. only running 1 hardware thread per
core) will have only even-numbered cpus online.
This fixes the problem by reading the /sys/devices/system/cpu/online
file to find out which cpus are online. The code that does that is in
tools/perf/util/cpumap.[ch], and consists of a read_cpu_map()
function that sets up a cpumap[] array and returns the number of
online cpus. If /sys/devices/system/cpu/online can't be read or
can't be parsed successfully, it falls back to using sysconf to
ask how many cpus are online and sets up an identity map in cpumap[].
The perf record, perf stat and perf top code then calls
read_cpu_map() in the system-wide monitoring case (instead of
sysconf) and uses cpumap[] to get the cpu numbers to pass to
perf_event_open.
Signed-off-by: Paul Mackerras <paulus@samba.org>
Cc: Anton Blanchard <anton@samba.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Arnaldo Carvalho de Melo <acme@infradead.org>
LKML-Reference: <20100310093609.GA3959@brick.ozlabs.ibm.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
2010-03-10 16:36:09 +07:00
|
|
|
if (n == 2 && sep == '-')
|
|
|
|
prev = cpu;
|
|
|
|
else
|
|
|
|
prev = -1;
|
|
|
|
if (n == 1 || sep == '\n')
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
|
2011-01-04 02:49:48 +07:00
|
|
|
if (nr_cpus > 0)
|
|
|
|
cpus = cpu_map__trim_new(nr_cpus, tmp_cpus);
|
|
|
|
else
|
|
|
|
cpus = cpu_map__default_new();
|
|
|
|
out_free_tmp:
|
|
|
|
free(tmp_cpus);
|
2012-09-10 14:53:50 +07:00
|
|
|
return cpus;
|
|
|
|
}
|
|
|
|
|
|
|
|
static struct cpu_map *cpu_map__read_all_cpu_map(void)
|
|
|
|
{
|
|
|
|
struct cpu_map *cpus = NULL;
|
|
|
|
FILE *onlnf;
|
|
|
|
|
|
|
|
onlnf = fopen("/sys/devices/system/cpu/online", "r");
|
|
|
|
if (!onlnf)
|
|
|
|
return cpu_map__default_new();
|
|
|
|
|
|
|
|
cpus = cpu_map__read(onlnf);
|
2011-01-04 02:49:48 +07:00
|
|
|
fclose(onlnf);
|
|
|
|
return cpus;
|
perf tools: Fix sparse CPU numbering related bugs
At present, the perf subcommands that do system-wide monitoring
(perf stat, perf record and perf top) don't work properly unless
the online cpus are numbered 0, 1, ..., N-1. These tools ask
for the number of online cpus with sysconf(_SC_NPROCESSORS_ONLN)
and then try to create events for cpus 0, 1, ..., N-1.
This creates problems for systems where the online cpus are
numbered sparsely. For example, a POWER6 system in
single-threaded mode (i.e. only running 1 hardware thread per
core) will have only even-numbered cpus online.
This fixes the problem by reading the /sys/devices/system/cpu/online
file to find out which cpus are online. The code that does that is in
tools/perf/util/cpumap.[ch], and consists of a read_cpu_map()
function that sets up a cpumap[] array and returns the number of
online cpus. If /sys/devices/system/cpu/online can't be read or
can't be parsed successfully, it falls back to using sysconf to
ask how many cpus are online and sets up an identity map in cpumap[].
The perf record, perf stat and perf top code then calls
read_cpu_map() in the system-wide monitoring case (instead of
sysconf) and uses cpumap[] to get the cpu numbers to pass to
perf_event_open.
Signed-off-by: Paul Mackerras <paulus@samba.org>
Cc: Anton Blanchard <anton@samba.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Arnaldo Carvalho de Melo <acme@infradead.org>
LKML-Reference: <20100310093609.GA3959@brick.ozlabs.ibm.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
2010-03-10 16:36:09 +07:00
|
|
|
}
|
2010-05-28 17:00:01 +07:00
|
|
|
|
2011-01-04 02:49:48 +07:00
|
|
|
struct cpu_map *cpu_map__new(const char *cpu_list)
|
2010-05-28 17:00:01 +07:00
|
|
|
{
|
2011-01-04 02:49:48 +07:00
|
|
|
struct cpu_map *cpus = NULL;
|
2010-05-28 17:00:01 +07:00
|
|
|
unsigned long start_cpu, end_cpu = 0;
|
|
|
|
char *p = NULL;
|
|
|
|
int i, nr_cpus = 0;
|
2011-01-04 02:49:48 +07:00
|
|
|
int *tmp_cpus = NULL, *tmp;
|
|
|
|
int max_entries = 0;
|
2010-05-28 17:00:01 +07:00
|
|
|
|
|
|
|
if (!cpu_list)
|
2011-01-04 02:49:48 +07:00
|
|
|
return cpu_map__read_all_cpu_map();
|
2010-05-28 17:00:01 +07:00
|
|
|
|
|
|
|
if (!isdigit(*cpu_list))
|
2011-01-04 02:49:48 +07:00
|
|
|
goto out;
|
2010-05-28 17:00:01 +07:00
|
|
|
|
|
|
|
while (isdigit(*cpu_list)) {
|
|
|
|
p = NULL;
|
|
|
|
start_cpu = strtoul(cpu_list, &p, 0);
|
|
|
|
if (start_cpu >= INT_MAX
|
|
|
|
|| (*p != '\0' && *p != ',' && *p != '-'))
|
|
|
|
goto invalid;
|
|
|
|
|
|
|
|
if (*p == '-') {
|
|
|
|
cpu_list = ++p;
|
|
|
|
p = NULL;
|
|
|
|
end_cpu = strtoul(cpu_list, &p, 0);
|
|
|
|
|
|
|
|
if (end_cpu >= INT_MAX || (*p != '\0' && *p != ','))
|
|
|
|
goto invalid;
|
|
|
|
|
|
|
|
if (end_cpu < start_cpu)
|
|
|
|
goto invalid;
|
|
|
|
} else {
|
|
|
|
end_cpu = start_cpu;
|
|
|
|
}
|
|
|
|
|
|
|
|
for (; start_cpu <= end_cpu; start_cpu++) {
|
|
|
|
/* check for duplicates */
|
|
|
|
for (i = 0; i < nr_cpus; i++)
|
2011-01-04 02:49:48 +07:00
|
|
|
if (tmp_cpus[i] == (int)start_cpu)
|
2010-05-28 17:00:01 +07:00
|
|
|
goto invalid;
|
|
|
|
|
2011-01-04 02:49:48 +07:00
|
|
|
if (nr_cpus == max_entries) {
|
|
|
|
max_entries += MAX_NR_CPUS;
|
|
|
|
tmp = realloc(tmp_cpus, max_entries * sizeof(int));
|
|
|
|
if (tmp == NULL)
|
|
|
|
goto invalid;
|
|
|
|
tmp_cpus = tmp;
|
|
|
|
}
|
|
|
|
tmp_cpus[nr_cpus++] = (int)start_cpu;
|
2010-05-28 17:00:01 +07:00
|
|
|
}
|
|
|
|
if (*p)
|
|
|
|
++p;
|
|
|
|
|
|
|
|
cpu_list = p;
|
|
|
|
}
|
|
|
|
|
2011-01-04 02:49:48 +07:00
|
|
|
if (nr_cpus > 0)
|
|
|
|
cpus = cpu_map__trim_new(nr_cpus, tmp_cpus);
|
|
|
|
else
|
|
|
|
cpus = cpu_map__default_new();
|
2010-05-28 17:00:01 +07:00
|
|
|
invalid:
|
2011-01-04 02:49:48 +07:00
|
|
|
free(tmp_cpus);
|
|
|
|
out:
|
|
|
|
return cpus;
|
|
|
|
}
|
|
|
|
|
2012-01-19 23:07:23 +07:00
|
|
|
size_t cpu_map__fprintf(struct cpu_map *map, FILE *fp)
|
|
|
|
{
|
|
|
|
int i;
|
|
|
|
size_t printed = fprintf(fp, "%d cpu%s: ",
|
|
|
|
map->nr, map->nr > 1 ? "s" : "");
|
|
|
|
for (i = 0; i < map->nr; ++i)
|
|
|
|
printed += fprintf(fp, "%s%d", i ? ", " : "", map->map[i]);
|
|
|
|
|
|
|
|
return printed + fprintf(fp, "\n");
|
|
|
|
}
|
|
|
|
|
2011-01-04 02:49:48 +07:00
|
|
|
struct cpu_map *cpu_map__dummy_new(void)
|
|
|
|
{
|
|
|
|
struct cpu_map *cpus = malloc(sizeof(*cpus) + sizeof(int));
|
|
|
|
|
|
|
|
if (cpus != NULL) {
|
|
|
|
cpus->nr = 1;
|
|
|
|
cpus->map[0] = -1;
|
|
|
|
}
|
|
|
|
|
|
|
|
return cpus;
|
2010-05-28 17:00:01 +07:00
|
|
|
}
|
2011-01-15 01:19:12 +07:00
|
|
|
|
|
|
|
void cpu_map__delete(struct cpu_map *map)
|
|
|
|
{
|
|
|
|
free(map);
|
|
|
|
}
|
2013-02-06 21:46:01 +07:00
|
|
|
|
|
|
|
int cpu_map__get_socket(struct cpu_map *map, int idx)
|
|
|
|
{
|
|
|
|
FILE *fp;
|
|
|
|
const char *mnt;
|
|
|
|
char path[PATH_MAX];
|
|
|
|
int cpu, ret;
|
|
|
|
|
|
|
|
if (idx > map->nr)
|
|
|
|
return -1;
|
|
|
|
|
|
|
|
cpu = map->map[idx];
|
|
|
|
|
2013-11-06 00:48:50 +07:00
|
|
|
mnt = sysfs__mountpoint();
|
2013-02-06 21:46:01 +07:00
|
|
|
if (!mnt)
|
|
|
|
return -1;
|
|
|
|
|
2013-02-14 19:57:27 +07:00
|
|
|
snprintf(path, PATH_MAX,
|
2013-02-06 21:46:01 +07:00
|
|
|
"%s/devices/system/cpu/cpu%d/topology/physical_package_id",
|
|
|
|
mnt, cpu);
|
|
|
|
|
|
|
|
fp = fopen(path, "r");
|
|
|
|
if (!fp)
|
|
|
|
return -1;
|
|
|
|
ret = fscanf(fp, "%d", &cpu);
|
|
|
|
fclose(fp);
|
|
|
|
return ret == 1 ? cpu : -1;
|
|
|
|
}
|
|
|
|
|
2013-02-14 19:57:27 +07:00
|
|
|
static int cmp_ids(const void *a, const void *b)
|
2013-02-06 21:46:01 +07:00
|
|
|
{
|
2013-02-14 19:57:27 +07:00
|
|
|
return *(int *)a - *(int *)b;
|
|
|
|
}
|
|
|
|
|
|
|
|
static int cpu_map__build_map(struct cpu_map *cpus, struct cpu_map **res,
|
|
|
|
int (*f)(struct cpu_map *map, int cpu))
|
|
|
|
{
|
|
|
|
struct cpu_map *c;
|
2013-02-06 21:46:01 +07:00
|
|
|
int nr = cpus->nr;
|
|
|
|
int cpu, s1, s2;
|
|
|
|
|
2013-02-14 19:57:27 +07:00
|
|
|
/* allocate as much as possible */
|
|
|
|
c = calloc(1, sizeof(*c) + nr * sizeof(int));
|
|
|
|
if (!c)
|
2013-02-06 21:46:01 +07:00
|
|
|
return -1;
|
|
|
|
|
|
|
|
for (cpu = 0; cpu < nr; cpu++) {
|
2013-02-14 19:57:27 +07:00
|
|
|
s1 = f(cpus, cpu);
|
|
|
|
for (s2 = 0; s2 < c->nr; s2++) {
|
|
|
|
if (s1 == c->map[s2])
|
2013-02-06 21:46:01 +07:00
|
|
|
break;
|
|
|
|
}
|
2013-02-14 19:57:27 +07:00
|
|
|
if (s2 == c->nr) {
|
|
|
|
c->map[c->nr] = s1;
|
|
|
|
c->nr++;
|
2013-02-06 21:46:01 +07:00
|
|
|
}
|
|
|
|
}
|
2013-02-14 19:57:27 +07:00
|
|
|
/* ensure we process id in increasing order */
|
|
|
|
qsort(c->map, c->nr, sizeof(int), cmp_ids);
|
|
|
|
|
|
|
|
*res = c;
|
2013-02-06 21:46:01 +07:00
|
|
|
return 0;
|
|
|
|
}
|
2013-02-14 19:57:27 +07:00
|
|
|
|
2013-02-14 19:57:29 +07:00
|
|
|
int cpu_map__get_core(struct cpu_map *map, int idx)
|
|
|
|
{
|
|
|
|
FILE *fp;
|
|
|
|
const char *mnt;
|
|
|
|
char path[PATH_MAX];
|
|
|
|
int cpu, ret, s;
|
|
|
|
|
|
|
|
if (idx > map->nr)
|
|
|
|
return -1;
|
|
|
|
|
|
|
|
cpu = map->map[idx];
|
|
|
|
|
2013-11-06 00:48:50 +07:00
|
|
|
mnt = sysfs__mountpoint();
|
2013-02-14 19:57:29 +07:00
|
|
|
if (!mnt)
|
|
|
|
return -1;
|
|
|
|
|
|
|
|
snprintf(path, PATH_MAX,
|
|
|
|
"%s/devices/system/cpu/cpu%d/topology/core_id",
|
|
|
|
mnt, cpu);
|
|
|
|
|
|
|
|
fp = fopen(path, "r");
|
|
|
|
if (!fp)
|
|
|
|
return -1;
|
|
|
|
ret = fscanf(fp, "%d", &cpu);
|
|
|
|
fclose(fp);
|
|
|
|
if (ret != 1)
|
|
|
|
return -1;
|
|
|
|
|
|
|
|
s = cpu_map__get_socket(map, idx);
|
|
|
|
if (s == -1)
|
|
|
|
return -1;
|
|
|
|
|
|
|
|
/*
|
|
|
|
* encode socket in upper 16 bits
|
|
|
|
* core_id is relative to socket, and
|
|
|
|
* we need a global id. So we combine
|
|
|
|
* socket+ core id
|
|
|
|
*/
|
|
|
|
return (s << 16) | (cpu & 0xffff);
|
|
|
|
}
|
|
|
|
|
2013-02-14 19:57:27 +07:00
|
|
|
int cpu_map__build_socket_map(struct cpu_map *cpus, struct cpu_map **sockp)
|
|
|
|
{
|
|
|
|
return cpu_map__build_map(cpus, sockp, cpu_map__get_socket);
|
|
|
|
}
|
2013-02-14 19:57:29 +07:00
|
|
|
|
|
|
|
int cpu_map__build_core_map(struct cpu_map *cpus, struct cpu_map **corep)
|
|
|
|
{
|
|
|
|
return cpu_map__build_map(cpus, corep, cpu_map__get_core);
|
|
|
|
}
|