Skip to contentSkip to conversation

perf trace: Add filter Suppport, V2

PATCH In review Li Zefan 3 participants 8 messages Oct 13, 2009

Subsystems (3)
lore
Share & tools
lore

Li Zefan

author
#0 lore raw

cover perf trace: Add filter Suppport, V2

This patchset adds filter support for perf counter, so not all profile events are recorded but only those match the filters we set.

An example:

#./perf record -R -f -e irq:irq_handler_entry --filter irq==18 ^C

 # ./perf trace
            perf-4303  ... irq_handler_entry: irq=18 handler=eth0
            init-0     ... irq_handler_entry: irq=18 handler=eth0
            init-0     ... irq_handler_entry: irq=18 handler=eth0
            init-0     ... irq_handler_entry: irq=18 handler=eth0
            init-0     ... irq_handler_entry: irq=18 handler=eth0

The syntax of filter expressions can be found in
Documentation/trace/events.txt.

Major changes from v1:

- Make filters per perf event, instead of per tracepoint.

- Use "--filter" to set a filter for a trace event, instead of mixing it into "-e" option.

---
 include/linux/ftrace_event.h       |   11 +-
 include/linux/perf_counter.h       |    1 +
 include/linux/perf_event.h         |    7 +
 kernel/perf_event.c                |   81 +++++++++-
 kernel/trace/trace.h               |    6 +-
 kernel/trace/trace_events_filter.c |  320 +++++++++++++++++++++---------------
 tools/perf/builtin-record.c        |   15 ++-
 tools/perf/util/parse-events.c     |   26 +++-
 tools/perf/util/parse-events.h     |    2 +
 9 files changed, 322 insertions(+), 147 deletions(-)
author
#1 lore raw

PATCH 1/4 tracing/filters: Refactor subsystem filter code

Change:
for_each_pred
for_each_subsystem
To:
for_each_subsystem
for_each_pred

This change also prepares for later patches.

---
 kernel/trace/trace.h               |    1 -
 kernel/trace/trace_events_filter.c |  124 +++++++++++++-----------------------
 2 files changed, 45 insertions(+), 80 deletions(-)
kernel/trace/trace.h +0 -1
index b17e0ea..300cdae 100644
@@ -705,7 +705,6 @@ struct event_filter {
 	int			n_preds;
 	struct filter_pred	**preds;
 	char			*filter_string;
-	bool			no_reset;
 };
 
 struct event_subsystem {
kernel/trace/trace_events_filter.c +45 -79

179 lines not on this page · raw message

Signature · 1 line
1.6.3
author
#2 lore raw

PATCH 2/4 tracing/filters: Use a different op for glob match

In reply to Li Zefan

"==" will always do a full match, and "~" will do a glob match.

In the future, we may add "=~" for regex match.

---
 kernel/trace/trace.h               |    2 +-
 kernel/trace/trace_events_filter.c |   59 +++++++++++++++++-------------------
 2 files changed, 29 insertions(+), 32 deletions(-)
kernel/trace/trace.h +1 -1
index 300cdae..b5f30d5 100644
@@ -724,7 +724,7 @@ typedef int (*filter_pred_fn_t) (struct filter_pred *pred, void *event,
 typedef int (*regex_match_func)(char *str, struct regex *r, int len);
 
 enum regex_type {
-	MATCH_FULL,
+	MATCH_FULL = 0,
 	MATCH_FRONT_ONLY,
 	MATCH_MIDDLE_ONLY,
 	MATCH_END_ONLY,
kernel/trace/trace_events_filter.c +28 -31

102 lines not on this page · raw message

Signature · 1 line
1.6.3
author
#3 lore raw

PATCH 3/4 tracing/profile: Add filter support

In reply to Li Zefan

- Add an ioctl to allocate a filter for a perf event.

- Free the filter when the associated perf event is to be freed.

- Do the filtering in perf_swevent_match().

---
 include/linux/ftrace_event.h       |   11 +++-
 include/linux/perf_counter.h       |    1 +
 include/linux/perf_event.h         |    7 ++
 kernel/perf_event.c                |   81 ++++++++++++++++++--
 kernel/trace/trace.h               |    3 +-
 kernel/trace/trace_events_filter.c |  149 +++++++++++++++++++++++++++++-------
 6 files changed, 214 insertions(+), 38 deletions(-)
include/linux/ftrace_event.h +10 -1
index 4ec5e67..d117704 100644
@@ -144,7 +144,7 @@ extern char			*trace_profile_buf_nmi;
 #define MAX_FILTER_STR_VAL	256	/* Should handle KSYM_SYMBOL_LEN */
 
 extern void destroy_preds(struct ftrace_event_call *call);
-extern int filter_match_preds(struct ftrace_event_call *call, void *rec);
+extern int filter_match_preds(struct event_filter *filter, void *rec);
 extern int filter_current_check_discard(struct ring_buffer *buffer,
 					struct ftrace_event_call *call,
 					void *rec,
@@ -186,4 +186,13 @@ do {									\
 		__trace_printk(ip, fmt, ##args);			\
 } while (0)
 
+#ifdef CONFIG_EVENT_PROFILE
+struct perf_event;
+extern int ftrace_profile_enable(int event_id);
+extern void ftrace_profile_disable(int event_id);
+extern int ftrace_profile_set_filter(struct perf_event *event, int event_id,
+				     char *filter_str);
+extern void ftrace_profile_free_filter(struct perf_event *event);
+#endif
+
 #endif /* _LINUX_FTRACE_EVENT_H */
include/linux/perf_counter.h +1 -0
index 7b7fbf4..91a2b43 100644
@@ -225,6 +225,7 @@ struct perf_counter_attr {
 #define PERF_COUNTER_IOC_RESET		_IO ('$', 3)
 #define PERF_COUNTER_IOC_PERIOD		_IOW('$', 4, u64)
 #define PERF_COUNTER_IOC_SET_OUTPUT	_IO ('$', 5)
+#define PERF_COUNTER_IOC_SET_FILTER	_IOW('$', 6, char *)
 
 enum perf_counter_ioc_flags {
 	PERF_IOC_FLAG_GROUP		= 1U << 0,
include/linux/perf_event.h +7 -0
index 2e6d95f..b7ac0f2 100644
@@ -221,6 +221,7 @@ struct perf_event_attr {
 #define PERF_EVENT_IOC_RESET		_IO ('$', 3)
 #define PERF_EVENT_IOC_PERIOD		_IOW('$', 4, u64)
 #define PERF_EVENT_IOC_SET_OUTPUT	_IO ('$', 5)
+#define PERF_EVENT_IOC_SET_FILTER	_IOW('$', 6, char *)
 
 enum perf_event_ioc_flags {
 	PERF_IOC_FLAG_GROUP		= 1U << 0,
@@ -633,7 +634,13 @@ struct perf_event {
 
 	struct pid_namespace		*ns;
 	u64				id;
+
+#ifdef CONFIG_EVENT_PROFILE
+	struct event_filter		*filter;
+	int				filter_active;
 #endif
+
+#endif /* CONFIG_PERF_EVENTS */
 };
 
 /**
kernel/perf_event.c +74 -7

160 lines not on this page · raw message

kernel/trace/trace.h +2 -1
index b5f30d5..671bedb 100644
@@ -765,7 +765,8 @@ filter_check_discard(struct ftrace_event_call *call, void *rec,
 		     struct ring_buffer *buffer,
 		     struct ring_buffer_event *event)
 {
-	if (unlikely(call->filter_active) && !filter_match_preds(call, rec)) {
+	if (unlikely(call->filter_active) &&
+	    !filter_match_preds(call->filter, rec)) {
 		ring_buffer_discard_commit(buffer, event);
 		return 1;
 	}
kernel/trace/trace_events_filter.c +120 -29

275 lines not on this page · raw message

Signature · 1 line
1.6.3
maintainer
#4 lore raw

Re: [PATCH 3/4] tracing/profile: Add filter support

include/linux/perf_event.h struct perf_event { · line 641

Quoted 4 lines · Li Zefan
include/linux/perf_event.h +3 -0
+#ifdef CONFIG_EVENT_PROFILE
+       struct event_filter             *filter;
+       int                             filter_active;
 #endif

Why do you need filter_active, isn't !filter equivalent?

author
#5 lore raw

Re: [PATCH 3/4] tracing/profile: Add filter support

Quoted 8 lines · Peter Zijlstra

include/linux/perf_event.h struct perf_event { · line 641

Quoted 4 lines · Li Zefan
include/linux/perf_event.h +3 -0
+#ifdef CONFIG_EVENT_PROFILE
+       struct event_filter             *filter;
+       int                             filter_active;
 #endif

Why do you need filter_active, isn't !filter equivalent?

Indeed. I realized this right after I sent out this patchset. I'll fix it. Thanks.

#6 lore raw

Re: [PATCH 3/4] tracing/profile: Add filter support

In reply to Li Zefan

include/linux/ftrace_event.h do { \ · line 196

Quoted 8 lines · Li Zefan
include/linux/ftrace_event.h +8 -0
+#ifdef CONFIG_EVENT_PROFILE
+struct perf_event;
+extern int ftrace_profile_enable(int event_id);
+extern void ftrace_profile_disable(int event_id);
+extern int ftrace_profile_set_filter(struct perf_event *event, int event_id,
+				     char *filter_str);
+extern void ftrace_profile_free_filter(struct perf_event *event);
+#endif

This reminds me - i think we should eliminate CONFIG_EVENT_PROFILE - it's an unnecessary Kconfig complication. If both PERF_EVENTS and EVENT_TRACING is enabled we should expose generic tracepoints.

Nor is it limited to event 'profiling', so it has become a misnomer as well.

Ingo

author
#7 lore raw

PATCH 4/4 perf trace: Add filter Suppport

In reply to Li Zefan

Add a new option "--filter <filter_str>" to perf record, and it should be right after "-e trace_point":

#./perf record -R -f -e irq:irq_handler_entry --filter irq==18 ^C

 # ./perf trace
            perf-4303  ... irq_handler_entry: irq=18 handler=eth0
            init-0     ... irq_handler_entry: irq=18 handler=eth0
            init-0     ... irq_handler_entry: irq=18 handler=eth0
            init-0     ... irq_handler_entry: irq=18 handler=eth0
            init-0     ... irq_handler_entry: irq=18 handler=eth0

See Documentation/trace/events.txt for the syntax of filter expressions.

---
 tools/perf/builtin-record.c    |   15 ++++++++++++++-
 tools/perf/util/parse-events.c |   26 ++++++++++++++++++++++++--
 tools/perf/util/parse-events.h |    2 ++
 3 files changed, 40 insertions(+), 3 deletions(-)
tools/perf/builtin-record.c +14 -1
index 59af03d..2f83c8f 100644
@@ -374,9 +374,11 @@ static struct perf_header_attr *get_header_attr(struct perf_event_attr *a, int n
 
 static void create_counter(int counter, int cpu, pid_t pid)
 {
+	char *filter = filters[counter];
 	struct perf_event_attr *attr = attrs + counter;
 	struct perf_header_attr *h_attr;
 	int track = !counter; /* only the first counter needs these */
+	int ret;
 	struct {
 		u64 count;
 		u64 time_enabled;
@@ -479,7 +481,6 @@ try_again:
 		multiplex_fd = fd[nr_cpu][counter];
 
 	if (multiplex && fd[nr_cpu][counter] != multiplex_fd) {
-		int ret;
 
 		ret = ioctl(fd[nr_cpu][counter], PERF_EVENT_IOC_SET_OUTPUT, multiplex_fd);
 		assert(ret != -1);
@@ -499,6 +500,16 @@ try_again:
 		}
 	}
 
+	if (filter != NULL) {
+		ret = ioctl(fd[nr_cpu][counter],
+			    PERF_EVENT_IOC_SET_FILTER, filter);
+		if (ret) {
+			error("failed to set filter with %d (%s)\n", errno,
+			      strerror(errno));
+			exit(-1);
+		}
+	}
+
 	ioctl(fd[nr_cpu][counter], PERF_EVENT_IOC_ENABLE);
 }
 
@@ -676,6 +687,8 @@ static const struct option options[] = {
 	OPT_CALLBACK('e', "event", NULL, "event",
 		     "event selector. use 'perf list' to list available events",
 		     parse_events),
+	OPT_CALLBACK(0, "filter", NULL, "filter",
+		     "event filter", parse_filter),
 	OPT_INTEGER('p', "pid", &target_pid,
 		    "record events on existing pid"),
 	OPT_INTEGER('r', "realtime", &realtime_prio,
tools/perf/util/parse-events.c +24 -2
index 87c424d..5573290 100644
@@ -8,9 +8,10 @@
 #include "cache.h"
 #include "header.h"
 
-int					nr_counters;
+int				nr_counters;
 
 struct perf_event_attr		attrs[MAX_COUNTERS];
+char				*filters[MAX_COUNTERS];
 
 struct event_symbol {
 	u8		type;
@@ -705,7 +706,6 @@ static void store_event_type(const char *orgname)
 	perf_header__push_event(id, orgname);
 }
 
-
 int parse_events(const struct option *opt __used, const char *str, int unset __used)
 {
 	struct perf_event_attr attr;
@@ -742,6 +742,28 @@ int parse_events(const struct option *opt __used, const char *str, int unset __u
 	return 0;
 }
 
+int parse_filter(const struct option *opt __used, const char *str,
+		 int unset __used)
+{
+	int i = nr_counters - 1;
+	int len = strlen(str);
+
+	if (i < 0 || attrs[i].type != PERF_TYPE_TRACEPOINT) {
+		fprintf(stderr,
+			"-F option should follow a -e tracepoint option\n");
+		return -1;
+	}
+
+	filters[i] = malloc(len + 1);
+	if (!filters[i]) {
+		fprintf(stderr, "not enough memory to hold filter string\n");
+		return -1;
+	}
+	strcpy(filters[i], str);
+
+	return 0;
+}
+
 static const char * const event_type_descriptors[] = {
 	"",
 	"Hardware event",
tools/perf/util/parse-events.h +2 -0
index 8626a43..b8c1f64 100644
@@ -17,11 +17,13 @@ extern struct tracepoint_path *tracepoint_id_to_path(u64 config);
 extern int			nr_counters;
 
 extern struct perf_event_attr attrs[MAX_COUNTERS];
+extern char *filters[MAX_COUNTERS];
 
 extern const char *event_name(int ctr);
 extern const char *__event_name(int type, u64 config);
 
 extern int parse_events(const struct option *opt, const char *str, int unset);
+extern int parse_filter(const struct option *opt, const char *str, int unset);
 
 #define EVENTS_HELP_MAX (128*1024)
 
Signature · 1 line
1.6.3