perf trace: Don't relookup fields by name in each sample
authorArnaldo Carvalho de Melo <acme@redhat.com>
Wed, 6 Nov 2013 19:35:57 +0000 (16:35 -0300)
committerArnaldo Carvalho de Melo <acme@redhat.com>
Thu, 7 Nov 2013 13:40:47 +0000 (10:40 -0300)
Instead do the lookups just when creating the tracepoints, initially for
the most common, raw_syscalls:sys_{enter,exit}.

It works by having evsel->priv have a per tracepoint structure with
entries for the fields, for direct access, with the offset and a
function to get the value from the sample, doing the swap if needed.

Using a simple workload that does M millions write syscalls, we go from:

 # perf stat -i -e cycles /tmp/oldperf trace ./sc_hello 100 > /dev/null

 Performance counter stats for '/tmp/oldperf trace ./sc_hello 100':

     8,366,771,459 cycles

       2.668025928 seconds time elapsed

 # perf stat -i -e cycles perf trace ./sc_hello 100 > /dev/null

 Performance counter stats for 'perf trace ./sc_hello 100':

     8,345,187,650 cycles

       2.631748425 seconds time elapsed

Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: David Ahern <dsahern@gmail.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: Jiri Olsa <jolsa@redhat.com>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Stephane Eranian <eranian@google.com>
Link: http://lkml.kernel.org/n/tip-eyfhvoo510a5i10b27dnvm88@git.kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
tools/perf/builtin-trace.c

index ee59df3..329b783 100644 (file)
 # define MADV_UNMERGEABLE      13
 #endif
 
+struct tp_field {
+       int offset;
+       union {
+               u64 (*integer)(struct tp_field *field, struct perf_sample *sample);
+               void *(*pointer)(struct tp_field *field, struct perf_sample *sample);
+       };
+};
+
+#define TP_UINT_FIELD(bits) \
+static u64 tp_field__u##bits(struct tp_field *field, struct perf_sample *sample) \
+{ \
+       return *(u##bits *)(sample->raw_data + field->offset); \
+}
+
+TP_UINT_FIELD(8);
+TP_UINT_FIELD(16);
+TP_UINT_FIELD(32);
+TP_UINT_FIELD(64);
+
+#define TP_UINT_FIELD__SWAPPED(bits) \
+static u64 tp_field__swapped_u##bits(struct tp_field *field, struct perf_sample *sample) \
+{ \
+       u##bits value = *(u##bits *)(sample->raw_data + field->offset); \
+       return bswap_##bits(value);\
+}
+
+TP_UINT_FIELD__SWAPPED(16);
+TP_UINT_FIELD__SWAPPED(32);
+TP_UINT_FIELD__SWAPPED(64);
+
+static int tp_field__init_uint(struct tp_field *field,
+                              struct format_field *format_field,
+                              bool needs_swap)
+{
+       field->offset = format_field->offset;
+
+       switch (format_field->size) {
+       case 1:
+               field->integer = tp_field__u8;
+               break;
+       case 2:
+               field->integer = needs_swap ? tp_field__swapped_u16 : tp_field__u16;
+               break;
+       case 4:
+               field->integer = needs_swap ? tp_field__swapped_u32 : tp_field__u32;
+               break;
+       case 8:
+               field->integer = needs_swap ? tp_field__swapped_u64 : tp_field__u64;
+               break;
+       default:
+               return -1;
+       }
+
+       return 0;
+}
+
+static void *tp_field__ptr(struct tp_field *field, struct perf_sample *sample)
+{
+       return sample->raw_data + field->offset;
+}
+
+static int tp_field__init_ptr(struct tp_field *field, struct format_field *format_field)
+{
+       field->offset = format_field->offset;
+       field->pointer = tp_field__ptr;
+       return 0;
+}
+
+struct syscall_tp {
+       struct tp_field id;
+       union {
+               struct tp_field args, ret;
+       };
+};
+
+static int perf_evsel__init_tp_uint_field(struct perf_evsel *evsel,
+                                         struct tp_field *field,
+                                         const char *name)
+{
+       struct format_field *format_field = perf_evsel__field(evsel, name);
+
+       if (format_field == NULL)
+               return -1;
+
+       return tp_field__init_uint(field, format_field, evsel->needs_swap);
+}
+
+#define perf_evsel__init_sc_tp_uint_field(evsel, name) \
+       ({ struct syscall_tp *sc = evsel->priv;\
+          perf_evsel__init_tp_uint_field(evsel, &sc->name, #name); })
+
+static int perf_evsel__init_tp_ptr_field(struct perf_evsel *evsel,
+                                        struct tp_field *field,
+                                        const char *name)
+{
+       struct format_field *format_field = perf_evsel__field(evsel, name);
+
+       if (format_field == NULL)
+               return -1;
+
+       return tp_field__init_ptr(field, format_field);
+}
+
+#define perf_evsel__init_sc_tp_ptr_field(evsel, name) \
+       ({ struct syscall_tp *sc = evsel->priv;\
+          perf_evsel__init_tp_ptr_field(evsel, &sc->name, #name); })
+
+static void perf_evsel__delete_priv(struct perf_evsel *evsel)
+{
+       free(evsel->priv);
+       evsel->priv = NULL;
+       perf_evsel__delete(evsel);
+}
+
+static struct perf_evsel *perf_evsel__syscall_newtp(const char *direction,
+                                                   void *handler, int idx)
+{
+       struct perf_evsel *evsel = perf_evsel__newtp("raw_syscalls", direction, idx);
+
+       if (evsel) {
+               evsel->priv = malloc(sizeof(struct syscall_tp));
+
+               if (evsel->priv == NULL)
+                       goto out_delete;
+
+               if (perf_evsel__init_sc_tp_uint_field(evsel, id))
+                       goto out_delete;
+
+               evsel->handler = handler;
+       }
+
+       return evsel;
+
+out_delete:
+       perf_evsel__delete_priv(evsel);
+       return NULL;
+}
+
+#define perf_evsel__sc_tp_uint(evsel, name, sample) \
+       ({ struct syscall_tp *fields = evsel->priv; \
+          fields->name.integer(&fields->name, sample); })
+
+#define perf_evsel__sc_tp_ptr(evsel, name, sample) \
+       ({ struct syscall_tp *fields = evsel->priv; \
+          fields->name.pointer(&fields->name, sample); })
+
+static int perf_evlist__add_syscall_newtp(struct perf_evlist *evlist,
+                                         void *sys_enter_handler,
+                                         void *sys_exit_handler)
+{
+       int ret = -1;
+       int idx = evlist->nr_entries;
+       struct perf_evsel *sys_enter, *sys_exit;
+
+       sys_enter = perf_evsel__syscall_newtp("sys_enter", sys_enter_handler, idx++);
+       if (sys_enter == NULL)
+               goto out;
+
+       if (perf_evsel__init_sc_tp_ptr_field(sys_enter, args))
+               goto out_delete_sys_enter;
+
+       sys_exit = perf_evsel__syscall_newtp("sys_exit", sys_exit_handler, idx++);
+       if (sys_exit == NULL)
+               goto out_delete_sys_enter;
+
+       if (perf_evsel__init_sc_tp_uint_field(sys_exit, ret))
+               goto out_delete_sys_exit;
+
+       perf_evlist__add(evlist, sys_enter);
+       perf_evlist__add(evlist, sys_exit);
+
+       ret = 0;
+out:
+       return ret;
+
+out_delete_sys_exit:
+       perf_evsel__delete_priv(sys_exit);
+out_delete_sys_enter:
+       perf_evsel__delete_priv(sys_enter);
+       goto out;
+}
+
+
 struct syscall_arg {
        unsigned long val;
        struct thread *thread;
@@ -1392,7 +1575,7 @@ static int trace__sys_enter(struct trace *trace, struct perf_evsel *evsel,
        void *args;
        size_t printed = 0;
        struct thread *thread;
-       int id = perf_evsel__intval(evsel, sample, "id");
+       int id = perf_evsel__sc_tp_uint(evsel, id, sample);
        struct syscall *sc = trace__syscall_info(trace, evsel, id);
        struct thread_trace *ttrace;
 
@@ -1407,12 +1590,7 @@ static int trace__sys_enter(struct trace *trace, struct perf_evsel *evsel,
        if (ttrace == NULL)
                return -1;
 
-       args = perf_evsel__rawptr(evsel, sample, "args");
-       if (args == NULL) {
-               fprintf(trace->output, "Problems reading syscall arguments\n");
-               return -1;
-       }
-
+       args = perf_evsel__sc_tp_ptr(evsel, args, sample);
        ttrace = thread->priv;
 
        if (ttrace->entry_str == NULL) {
@@ -1445,7 +1623,7 @@ static int trace__sys_exit(struct trace *trace, struct perf_evsel *evsel,
        int ret;
        u64 duration = 0;
        struct thread *thread;
-       int id = perf_evsel__intval(evsel, sample, "id");
+       int id = perf_evsel__sc_tp_uint(evsel, id, sample);
        struct syscall *sc = trace__syscall_info(trace, evsel, id);
        struct thread_trace *ttrace;
 
@@ -1463,7 +1641,7 @@ static int trace__sys_exit(struct trace *trace, struct perf_evsel *evsel,
        if (trace->summary)
                thread__update_stats(ttrace, id, sample);
 
-       ret = perf_evsel__intval(evsel, sample, "ret");
+       ret = perf_evsel__sc_tp_uint(evsel, ret, sample);
 
        if (id == trace->audit.open_id && ret >= 0 && trace->last_vfs_getname) {
                trace__set_fd_pathname(thread, ret, trace->last_vfs_getname);
@@ -1675,8 +1853,7 @@ static int trace__run(struct trace *trace, int argc, const char **argv)
                goto out;
        }
 
-       if (perf_evlist__add_newtp(evlist, "raw_syscalls", "sys_enter", trace__sys_enter) ||
-               perf_evlist__add_newtp(evlist, "raw_syscalls", "sys_exit", trace__sys_exit))
+       if (perf_evlist__add_syscall_newtp(evlist, trace__sys_enter, trace__sys_exit))
                goto out_error_tp;
 
        perf_evlist__add_vfs_getname(evlist);