perf tool: Add cgroup support

This patch adds the ability to filter monitoring based on container groups
(cgroups) for both perf stat and perf record. It is possible to monitor
multiple cgroup in parallel. There is one cgroup per event. The cgroups to
monitor are passed via a new -G option followed by a comma separated list of
cgroup names.

The cgroup filesystem has to be mounted. Given a cgroup name, the perf tool
finds the corresponding directory in the cgroup filesystem and opens it. It
then passes that file descriptor to the kernel.

Example:

$ perf stat -B -a -e cycles:u,cycles:u,cycles:u -G test1,,test2 -- sleep 1
 Performance counter stats for 'sleep 1':

      2,368,667,414  cycles                   test1
      2,369,661,459  cycles
      <not counted>  cycles                   test2

        1.001856890  seconds time elapsed

Signed-off-by: Stephane Eranian <eranian@google.com>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
LKML-Reference: <4d590290.825bdf0a.7d0a.4890@mx.google.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

4 files changed, 211 insertions, 2 deletions

diff --git a/tools/perf/util/cgroup.c b/tools/perf/util/cgroup.c
new file mode 100644
index 0000000..9fea755
--- /dev/null
+++ b/tools/perf/util/cgroup.c
@@ -0,0 +1,178 @@
+#include "util.h"
+#include "../perf.h"
+#include "parse-options.h"
+#include "evsel.h"
+#include "cgroup.h"
+#include "debugfs.h" /* MAX_PATH, STR() */
+#include "evlist.h"
+
+int nr_cgroups;
+
+static int
+cgroupfs_find_mountpoint(char *buf, size_t maxlen)
+{
+	FILE *fp;
+	char mountpoint[MAX_PATH+1], tokens[MAX_PATH+1], type[MAX_PATH+1];
+	char *token, *saved_ptr;
+	int found = 0;
+
+	fp = fopen("/proc/mounts", "r");
+	if (!fp)
+		return -1;
+
+	/*
+	 * in order to handle split hierarchy, we need to scan /proc/mounts
+	 * and inspect every cgroupfs mount point to find one that has
+	 * perf_event subsystem
+	 */
+	while (fscanf(fp, "%*s %"STR(MAX_PATH)"s %"STR(MAX_PATH)"s %"
+				STR(MAX_PATH)"s %*d %*d\n",
+				mountpoint, type, tokens) == 3) {
+
+		if (!strcmp(type, "cgroup")) {
+
+			token = strtok_r(tokens, ",", &saved_ptr);
+
+			while (token != NULL) {
+				if (!strcmp(token, "perf_event")) {
+					found = 1;
+					break;
+				}
+				token = strtok_r(NULL, ",", &saved_ptr);
+			}
+		}
+		if (found)
+			break;
+	}
+	fclose(fp);
+	if (!found)
+		return -1;
+
+	if (strlen(mountpoint) < maxlen) {
+		strcpy(buf, mountpoint);
+		return 0;
+	}
+	return -1;
+}
+
+static int open_cgroup(char *name)
+{
+	char path[MAX_PATH+1];
+	char mnt[MAX_PATH+1];
+	int fd;
+
+
+	if (cgroupfs_find_mountpoint(mnt, MAX_PATH+1))
+		return -1;
+
+	snprintf(path, MAX_PATH, "%s/%s", mnt, name);
+
+	fd = open(path, O_RDONLY);
+	if (fd == -1)
+		fprintf(stderr, "no access to cgroup %s\n", path);
+
+	return fd;
+}
+
+static int add_cgroup(struct perf_evlist *evlist, char *str)
+{
+	struct perf_evsel *counter;
+	struct cgroup_sel *cgrp = NULL;
+	int n;
+	/*
+	 * check if cgrp is already defined, if so we reuse it
+	 */
+	list_for_each_entry(counter, &evlist->entries, node) {
+		cgrp = counter->cgrp;
+		if (!cgrp)
+			continue;
+		if (!strcmp(cgrp->name, str))
+			break;
+
+		cgrp = NULL;
+	}
+
+	if (!cgrp) {
+		cgrp = zalloc(sizeof(*cgrp));
+		if (!cgrp)
+			return -1;
+
+		cgrp->name = str;
+
+		cgrp->fd = open_cgroup(str);
+		if (cgrp->fd == -1) {
+			free(cgrp);
+			return -1;
+		}
+	}
+
+	/*
+	 * find corresponding event
+	 * if add cgroup N, then need to find event N
+	 */
+	n = 0;
+	list_for_each_entry(counter, &evlist->entries, node) {
+		if (n == nr_cgroups)
+			goto found;
+		n++;
+	}
+	if (cgrp->refcnt == 0)
+		free(cgrp);
+
+	return -1;
+found:
+	cgrp->refcnt++;
+	counter->cgrp = cgrp;
+	return 0;
+}
+
+void close_cgroup(struct cgroup_sel *cgrp)
+{
+	if (!cgrp)
+		return;
+
+	/* XXX: not reentrant */
+	if (--cgrp->refcnt == 0) {
+		close(cgrp->fd);
+		free(cgrp->name);
+		free(cgrp);
+	}
+}
+
+int parse_cgroups(const struct option *opt __used, const char *str,
+		  int unset __used)
+{
+	struct perf_evlist *evlist = *(struct perf_evlist **)opt->value;
+	const char *p, *e, *eos = str + strlen(str);
+	char *s;
+	int ret;
+
+	if (list_empty(&evlist->entries)) {
+		fprintf(stderr, "must define events before cgroups\n");
+		return -1;
+	}
+
+	for (;;) {
+		p = strchr(str, ',');
+		e = p ? p : eos;
+
+		/* allow empty cgroups, i.e., skip */
+		if (e - str) {
+			/* termination added */
+			s = strndup(str, e - str);
+			if (!s)
+				return -1;
+			ret = add_cgroup(evlist, s);
+			if (ret) {
+				free(s);
+				return -1;
+			}
+		}
+		/* nr_cgroups is increased een for empty cgroups */
+		nr_cgroups++;
+		if (!p)
+			break;
+		str = p+1;
+	}
+	return 0;
+}
diff --git a/tools/perf/util/cgroup.h b/tools/perf/util/cgroup.h
new file mode 100644
index 0000000..89acd6d
--- /dev/null
+++ b/tools/perf/util/cgroup.h
@@ -0,0 +1,17 @@
+#ifndef __CGROUP_H__
+#define __CGROUP_H__
+
+struct option;
+
+struct cgroup_sel {
+	char *name;
+	int fd;
+	int refcnt;
+};
+
+
+extern int nr_cgroups; /* number of explicit cgroups defined */
+extern void close_cgroup(struct cgroup_sel *cgrp);
+extern int parse_cgroups(const struct option *opt, const char *str, int unset);
+
+#endif /* __CGROUP_H__ */
diff --git a/tools/perf/util/evsel.c b/tools/perf/util/evsel.c
index 211063e..c974e08 100644
--- a/tools/perf/util/evsel.c
+++ b/tools/perf/util/evsel.c
@@ -85,6 +85,7 @@ void perf_evsel__exit(struct perf_evsel *evsel)
 void perf_evsel__delete(struct perf_evsel *evsel)
 {
 	perf_evsel__exit(evsel);
+	close_cgroup(evsel->cgrp);
 	free(evsel);
 }
 
@@ -163,21 +164,32 @@ static int __perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
 			      struct thread_map *threads, bool group, bool inherit)
 {
 	int cpu, thread;
+	unsigned long flags = 0;
+	int pid = -1;
 
 	if (evsel->fd == NULL &&
 	    perf_evsel__alloc_fd(evsel, cpus->nr, threads->nr) < 0)
 		return -1;
 
+	if (evsel->cgrp) {
+		flags = PERF_FLAG_PID_CGROUP;
+		pid = evsel->cgrp->fd;
+	}
+
 	for (cpu = 0; cpu < cpus->nr; cpu++) {
 		int group_fd = -1;
 
 		evsel->attr.inherit = (cpus->map[cpu] < 0) && inherit;
 
 		for (thread = 0; thread < threads->nr; thread++) {
+
+			if (!evsel->cgrp)
+				pid = threads->map[thread];
+
 			FD(evsel, cpu, thread) = sys_perf_event_open(&evsel->attr,
-								     threads->map[thread],
+								     pid,
 								     cpus->map[cpu],
-								     group_fd, 0);
+								     group_fd, flags);
 			if (FD(evsel, cpu, thread) < 0)
 				goto out_close;
 
diff --git a/tools/perf/util/evsel.h b/tools/perf/util/evsel.h
index eecdc3a..1d3d5a3 100644
--- a/tools/perf/util/evsel.h
+++ b/tools/perf/util/evsel.h
@@ -6,6 +6,7 @@
 #include "../../../include/linux/perf_event.h"
 #include "types.h"
 #include "xyarray.h"
+#include "cgroup.h"
  
 struct perf_counts_values {
 	union {
@@ -45,6 +46,7 @@ struct perf_evsel {
 	struct perf_counts	*counts;
 	int			idx;
 	void			*priv;
+	struct cgroup_sel	*cgrp;
 };
 
 struct cpu_map;