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>

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>
Paul Mackerras · Ingo Molnar
1 parent 220b140b52
Showing 6 changed files with 81 additions and 13 deletions Side-by-side Diff
tools/perf/Makefile
tools/perf/builtin-record.c
tools/perf/builtin-stat.c
tools/perf/builtin-top.c
tools/perf/util/cpumap.c
tools/perf/util/cpumap.h
@@ -387,6 +387,7 @@
 LIB_H += util/trace-event.h
 LIB_H += util/probe-finder.h
 LIB_H += util/probe-event.h
+LIB_H += util/cpumap.h
  
 LIB_OBJS += util/abspath.o
 LIB_OBJS += util/alias.o
@@ -433,6 +434,7 @@
 LIB_OBJS += util/hist.o
 LIB_OBJS += util/probe-event.o
 LIB_OBJS += util/util.o
+LIB_OBJS += util/cpumap.o
  
 BUILTIN_OBJS += builtin-annotate.o
  
@@ -22,6 +22,7 @@
 #include "util/debug.h"
 #include "util/session.h"
 #include "util/symbol.h"
+#include "util/cpumap.h"
  
 #include <unistd.h>
 #include <sched.h>
@@ -421,9 +422,6 @@
 	char buf;
  
 	page_size = sysconf(_SC_PAGE_SIZE);
-	nr_cpus = sysconf(_SC_NPROCESSORS_ONLN);
-	assert(nr_cpus <= MAX_NR_CPUS);
-	assert(nr_cpus >= 0);
  
 	atexit(sig_atexit);
 	signal(SIGCHLD, sig_handler);
  
@@ -547,8 +545,9 @@
 	if ((!system_wide && !inherit) || profile_cpu != -1) {
 		open_counters(profile_cpu, target_pid);
 	} else {
+		nr_cpus = read_cpu_map();
 		for (i = 0; i < nr_cpus; i++)
-			open_counters(i, target_pid);
+			open_counters(cpumap[i], target_pid);
 	}
  
 	if (file_new) {
@@ -45,6 +45,7 @@
 #include "util/event.h"
 #include "util/debug.h"
 #include "util/header.h"
+#include "util/cpumap.h"
  
 #include <sys/prctl.h>
 #include <math.h>
@@ -151,7 +152,7 @@
 		unsigned int cpu;
  
 		for (cpu = 0; cpu < nr_cpus; cpu++) {
-			fd[cpu][counter] = sys_perf_event_open(attr, -1, cpu, -1, 0);
+			fd[cpu][counter] = sys_perf_event_open(attr, -1, cpumap[cpu], -1, 0);
 			if (fd[cpu][counter] < 0 && verbose)
 				fprintf(stderr, ERR_PERF_OPEN, counter,
 					fd[cpu][counter], strerror(errno));
@@ -519,9 +520,10 @@
 		nr_counters = ARRAY_SIZE(default_attrs);
 	}
  
-	nr_cpus = sysconf(_SC_NPROCESSORS_ONLN);
-	assert(nr_cpus <= MAX_NR_CPUS);
-	assert((int)nr_cpus >= 0);
+	if (system_wide)
+		nr_cpus = read_cpu_map();
+	else
+		nr_cpus = 1;
  
 	/*
 	 * We dont want to block the signals - that would cause
@@ -28,6 +28,7 @@
 #include <linux/rbtree.h>
 #include "util/parse-options.h"
 #include "util/parse-events.h"
+#include "util/cpumap.h"
  
 #include "util/debug.h"
  
@@ -1123,7 +1124,7 @@
  
 	cpu = profile_cpu;
 	if (target_pid == -1 && profile_cpu == -1)
-		cpu = i;
+		cpu = cpumap[i];
  
 	attr = attrs + counter;
  
  
@@ -1347,12 +1348,10 @@
 		attrs[counter].sample_period = default_interval;
 	}
  
-	nr_cpus = sysconf(_SC_NPROCESSORS_ONLN);
-	assert(nr_cpus <= MAX_NR_CPUS);
-	assert(nr_cpus >= 0);
-
 	if (target_pid != -1 || profile_cpu != -1)
 		nr_cpus = 1;
+	else
+		nr_cpus = read_cpu_map();
  
 	get_term_dimensions(&winsize);
 	if (print_entries == 0) {
+#include "util.h"
+#include "../perf.h"
+#include "cpumap.h"
+#include <assert.h>
+#include <stdio.h>
+
+int cpumap[MAX_NR_CPUS];
+
+static int default_cpu_map(void)
+{
+	int nr_cpus, i;
+
+	nr_cpus = sysconf(_SC_NPROCESSORS_ONLN);
+	assert(nr_cpus <= MAX_NR_CPUS);
+	assert((int)nr_cpus >= 0);
+
+	for (i = 0; i < nr_cpus; ++i)
+		cpumap[i] = i;
+
+	return nr_cpus;
+}
+
+int read_cpu_map(void)
+{
+	FILE *onlnf;
+	int nr_cpus = 0;
+	int n, cpu, prev;
+	char sep;
+
+	onlnf = fopen("/sys/devices/system/cpu/online", "r");
+	if (!onlnf)
+		return default_cpu_map();
+
+	sep = 0;
+	prev = -1;
+	for (;;) {
+		n = fscanf(onlnf, "%u%c", &cpu, &sep);
+		if (n <= 0)
+			break;
+		if (prev >= 0) {
+			assert(nr_cpus + cpu - prev - 1 < MAX_NR_CPUS);
+			while (++prev < cpu)
+				cpumap[nr_cpus++] = prev;
+		}
+		assert (nr_cpus < MAX_NR_CPUS);
+		cpumap[nr_cpus++] = cpu;
+		if (n == 2 && sep == '-')
+			prev = cpu;
+		else
+			prev = -1;
+		if (n == 1 || sep == '\n')
+			break;
+	}
+	fclose(onlnf);
+	if (nr_cpus > 0)
+		return nr_cpus;
+
+	return default_cpu_map();
+}
+#ifndef __PERF_CPUMAP_H
+#define __PERF_CPUMAP_H
+
+extern int read_cpu_map(void);
+extern int cpumap[];
+
+#endif /* __PERF_CPUMAP_H */
...	...	@@ -387,6 +387,7 @@
387	387	LIB_H += util/trace-event.h
388	388	LIB_H += util/probe-finder.h
389	389	LIB_H += util/probe-event.h
	390	+LIB_H += util/cpumap.h
390	391
391	392	LIB_OBJS += util/abspath.o
392	393	LIB_OBJS += util/alias.o
...	...	@@ -433,6 +434,7 @@
433	434	LIB_OBJS += util/hist.o
434	435	LIB_OBJS += util/probe-event.o
435	436	LIB_OBJS += util/util.o
	437	+LIB_OBJS += util/cpumap.o
436	438
437	439	BUILTIN_OBJS += builtin-annotate.o
438	440
...	...	@@ -22,6 +22,7 @@
22	22	#include "util/debug.h"
23	23	#include "util/session.h"
24	24	#include "util/symbol.h"
	25	+#include "util/cpumap.h"
25	26
26	27	#include <unistd.h>
27	28	#include <sched.h>
...	...	@@ -421,9 +422,6 @@
421	422	char buf;
422	423
423	424	page_size = sysconf(_SC_PAGE_SIZE);
424		- nr_cpus = sysconf(_SC_NPROCESSORS_ONLN);
425		- assert(nr_cpus <= MAX_NR_CPUS);
426		- assert(nr_cpus >= 0);
427	425
428	426	atexit(sig_atexit);
429	427	signal(SIGCHLD, sig_handler);
430	428
...	...	@@ -547,8 +545,9 @@
547	545	if ((!system_wide && !inherit) \|\| profile_cpu != -1) {
548	546	open_counters(profile_cpu, target_pid);
549	547	} else {
	548	+ nr_cpus = read_cpu_map();
550	549	for (i = 0; i < nr_cpus; i++)
551		- open_counters(i, target_pid);
	550	+ open_counters(cpumap[i], target_pid);
552	551	}
553	552
554	553	if (file_new) {
...	...	@@ -45,6 +45,7 @@
45	45	#include "util/event.h"
46	46	#include "util/debug.h"
47	47	#include "util/header.h"
	48	+#include "util/cpumap.h"
48	49
49	50	#include <sys/prctl.h>
50	51	#include <math.h>
...	...	@@ -151,7 +152,7 @@
151	152	unsigned int cpu;
152	153
153	154	for (cpu = 0; cpu < nr_cpus; cpu++) {
154		- fd[cpu][counter] = sys_perf_event_open(attr, -1, cpu, -1, 0);
	155	+ fd[cpu][counter] = sys_perf_event_open(attr, -1, cpumap[cpu], -1, 0);
155	156	if (fd[cpu][counter] < 0 && verbose)
156	157	fprintf(stderr, ERR_PERF_OPEN, counter,
157	158	fd[cpu][counter], strerror(errno));
...	...	@@ -519,9 +520,10 @@
519	520	nr_counters = ARRAY_SIZE(default_attrs);
520	521	}
521	522
522		- nr_cpus = sysconf(_SC_NPROCESSORS_ONLN);
523		- assert(nr_cpus <= MAX_NR_CPUS);
524		- assert((int)nr_cpus >= 0);
	523	+ if (system_wide)
	524	+ nr_cpus = read_cpu_map();
	525	+ else
	526	+ nr_cpus = 1;
525	527
526	528	/*
527	529	* We dont want to block the signals - that would cause
...	...	@@ -28,6 +28,7 @@
28	28	#include <linux/rbtree.h>
29	29	#include "util/parse-options.h"
30	30	#include "util/parse-events.h"
	31	+#include "util/cpumap.h"
31	32
32	33	#include "util/debug.h"
33	34
...	...	@@ -1123,7 +1124,7 @@
1123	1124
1124	1125	cpu = profile_cpu;
1125	1126	if (target_pid == -1 && profile_cpu == -1)
1126		- cpu = i;
	1127	+ cpu = cpumap[i];
1127	1128
1128	1129	attr = attrs + counter;
1129	1130
1130	1131
...	...	@@ -1347,12 +1348,10 @@
1347	1348	attrs[counter].sample_period = default_interval;
1348	1349	}
1349	1350
1350		- nr_cpus = sysconf(_SC_NPROCESSORS_ONLN);
1351		- assert(nr_cpus <= MAX_NR_CPUS);
1352		- assert(nr_cpus >= 0);
1353		-
1354	1351	if (target_pid != -1 \|\| profile_cpu != -1)
1355	1352	nr_cpus = 1;
	1353	+ else
	1354	+ nr_cpus = read_cpu_map();
1356	1355
1357	1356	get_term_dimensions(&winsize);
1358	1357	if (print_entries == 0) {
	1	+#include "util.h"
	2	+#include "../perf.h"
	3	+#include "cpumap.h"
	4	+#include <assert.h>
	5	+#include <stdio.h>
	6	+
	7	+int cpumap[MAX_NR_CPUS];
	8	+
	9	+static int default_cpu_map(void)
	10	+{
	11	+ int nr_cpus, i;
	12	+
	13	+ nr_cpus = sysconf(_SC_NPROCESSORS_ONLN);
	14	+ assert(nr_cpus <= MAX_NR_CPUS);
	15	+ assert((int)nr_cpus >= 0);
	16	+
	17	+ for (i = 0; i < nr_cpus; ++i)
	18	+ cpumap[i] = i;
	19	+
	20	+ return nr_cpus;
	21	+}
	22	+
	23	+int read_cpu_map(void)
	24	+{
	25	+ FILE *onlnf;
	26	+ int nr_cpus = 0;
	27	+ int n, cpu, prev;
	28	+ char sep;
	29	+
	30	+ onlnf = fopen("/sys/devices/system/cpu/online", "r");
	31	+ if (!onlnf)
	32	+ return default_cpu_map();
	33	+
	34	+ sep = 0;
	35	+ prev = -1;
	36	+ for (;;) {
	37	+ n = fscanf(onlnf, "%u%c", &cpu, &sep);
	38	+ if (n <= 0)
	39	+ break;
	40	+ if (prev >= 0) {
	41	+ assert(nr_cpus + cpu - prev - 1 < MAX_NR_CPUS);
	42	+ while (++prev < cpu)
	43	+ cpumap[nr_cpus++] = prev;
	44	+ }
	45	+ assert (nr_cpus < MAX_NR_CPUS);
	46	+ cpumap[nr_cpus++] = cpu;
	47	+ if (n == 2 && sep == '-')
	48	+ prev = cpu;
	49	+ else
	50	+ prev = -1;
	51	+ if (n == 1 \|\| sep == '\n')
	52	+ break;
	53	+ }
	54	+ fclose(onlnf);
	55	+ if (nr_cpus > 0)
	56	+ return nr_cpus;
	57	+
	58	+ return default_cpu_map();
	59	+}
	1	+#ifndef __PERF_CPUMAP_H
	2	+#define __PERF_CPUMAP_H
	3	+
	4	+extern int read_cpu_map(void);
	5	+extern int cpumap[];
	6	+
	7	+#endif /* __PERF_CPUMAP_H */