Replace the original PERF_FUNCTION_BEGIN and PERF_FUNCTION_END pair
with a single PERF macro. This design is also more flexible, removing
the restriction of have only one measurement per function.
The next patch will make use of this macro.
Signed-off-by: Andy Zhou <az...@nicira.com>
---
lib/perf-counter.c | 13 ++++++----
lib/perf-counter.h | 75 +++++++++++++++++++++---------------------------------
2 files changed, 37 insertions(+), 51 deletions(-)
diff --git a/lib/perf-counter.c b/lib/perf-counter.c
index 8c859cc..d8c0532 100644
--- a/lib/perf-counter.c
+++ b/lib/perf-counter.c
@@ -31,7 +31,7 @@
#include "shash.h"
#include "util.h"
-static struct shash perf_counters;
+static struct shash perf_counters = SHASH_INITIALIZER(&perf_counters);
static int fd__ = 0;
uint64_t
@@ -59,10 +59,11 @@ perf_event_open(struct perf_event_attr *hw_event, pid_t pid,
/* Set up perf event counters to read user space instruction counters
* only for this process, on all cpus. */
-static void
+static int
perf_event_setup(void)
{
struct perf_event_attr pe;
+ int fd;
memset(&pe, 0, sizeof(struct perf_event_attr));
pe.type = PERF_TYPE_HARDWARE;
@@ -72,11 +73,13 @@ perf_event_setup(void)
pe.exclude_kernel = 1;
pe.exclude_hv = 1;
- fd__ = perf_event_open(&pe, 0, -1, -1, 0);
- if (fd__ > 0) {
+ fd = perf_event_open(&pe, 0, -1, -1, 0);
+ if (fd > 0) {
ioctl(fd__, PERF_EVENT_IOC_RESET, 0);
ioctl(fd__, PERF_EVENT_IOC_ENABLE, 0);
}
+
+ return fd;
}
static void
@@ -159,7 +162,7 @@ void
perf_counters_init(void)
{
shash_init(&perf_counters);
- perf_event_setup();
+ fd__ = perf_event_setup();
}
void
diff --git a/lib/perf-counter.h b/lib/perf-counter.h
index 23365de..ca3aaba 100644
--- a/lib/perf-counter.h
+++ b/lib/perf-counter.h
@@ -44,54 +44,37 @@
* Those are not fundamental limits, but only limited by current
* implementation.
*
- * Function instruction counter sample point Usage
- * ================================================
+ * Usage:
+ * =======
*
- * There are two macros provided:
+ * Adding performance counter is easy. Simply use the following macro to
+ * wrap around the expression you are interested in measuring.
*
- * Macro 'PERF_FUNCTON_COUNT_BEGIN' needs to be inserted towards the
- * beginning of the function where local variables are declared.
+ * PERF(name, expr).
*
- * Macro 'PERF_FUNCTON_COUNT_END' needs to appear in the same function,
- * some where below 'PERF_FUNCTION_COUNT_BEGIN', usually towards of
- * a function.
+ * The 'expr' is a set of C expressions you are interested in measuring.
+ * 'name' is the counter name.
*
- * For example:
+ * For example, if we are interested in performance of perf_func():
*
- * void my_func() {
- * int some_local_variable;
- *
- * PERF_FUNCTION_COUNT_BEGIN;
- *
- * < implementation >
- *
- * PERF_FUNCTION_COUNT_END
+ * int perf_func() {
+ * <implemenation>
* }
*
- * This will maintain the number of times 'my_func()' is called, total
- * number of instructions '<implementation>' executed during all those calls.
+ * void func() {
+ * int rt;
*
- * Currently there are two limitation:
- * 1). At most one pair can appear in the same variable scope.
- * 2). The Macros use function name as the counter name for display.
- * Thus, all functions in one annotation session are required to
- * have unique names.
+ * ...
+ * PERF("perf_func", rt = perf_func());
*
- * Note, there is no requirement for those macros to be balanced.
- * For example:
+ * return rt;
+ * }
*
- * void my_func(int i){
*
- * PERF_FUNCTION_COUNT_BEGIN;
+ * This will maintain the number of times 'perf_func()' is called, total
+ * number of instructions '<implementation>' plus function call overhead
+ * executed.
*
- * if (i == 300) {
- * PERF_FUNCTION_COUNT_END;
- * return;
- * } else {
- * <some code>
- * }
- * }
- * will work just fine.
*/
#if defined(__linux__) && defined(HAVE_LINUX_PERF_EVENT_H)
@@ -120,17 +103,17 @@ void perf_counter_accumulate(struct perf_counter *counter,
char *perf_counters_to_string(void);
/* User access macros. */
-#define PERF_FUNCTION_BEGIN \
- static struct perf_counter x__ = PERF_COUNTER_ONCE_INITIALIZER(__func__); \
- uint64_t start_count__ = perf_counter_read(&start_count__); \
-
-#define PERF_FUNCTION_END \
- perf_counter_accumulate(&x__, start_count__);
-
+#define PERF(name, expr) \
+ { \
+ static struct perf_counter c = PERF_COUNTER_ONCE_INITIALIZER(name);\
+ uint64_t start_count = perf_counter_read(&start_count); \
+ \
+ expr; \
+ \
+ perf_counter_accumulate(&c, start_count); \
+ }
#else
-
-#define PERF_FUNCTON_BEGIN
-#define PERF_FUNCTON_END
+#define PERF(name, expr)
static inline void perf_counters_init(void) {}
static inline void perf_counters_destroy(void) {}
--
1.9.1
_______________________________________________
dev mailing list
dev@openvswitch.org
http://openvswitch.org/mailman/listinfo/dev
