perf top: Add processing thread

Add a new thread that takes care of the hist creating to alleviate the
main reader thread so it can keep perf mmaps served in time so that we
reduce the possibility of losing events.

The 'perf top' command now spawns 2 extra threads, the data processing
is the following:

  1) The main thread reads the data from mmaps and queues them to
     ordered events object;

  2) The processing threads takes the data from the ordered events
     object and create initial histogram;

  3) The GUI thread periodically sorts the initial histogram and
     presents it.

Passing the data between threads 1 and 2 is done by having 2 ordered
events queues. One is always being stored by thread 1 while the other is
flushed out in thread 2.

Passing the data between threads 2 and 3 stays the same as was initially
for threads 1 and 3.

Signed-off-by: Jiri Olsa <jolsa@kernel.org>
Acked-by: David S. Miller <davem@davemloft.net>
Acked-by: Namhyung Kim <namhyung@kernel.org>
Tested-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Link: http://lkml.kernel.org/n/tip-hhf4hllgkmle9wl1aly1jli0@git.kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>

tools/perf/builtin-top.c

@@ -46,6 +46,7 @@
 #include "arch/common.h"
 
 #include "util/debug.h"
+#include "util/ordered-events.h"
 
 #include <assert.h>
 #include <elf.h>
@@ -830,78 +831,28 @@ static void perf_top__mmap_read_idx(struct perf_top *top, int idx)
 {
 	struct record_opts *opts = &top->record_opts;
 	struct perf_evlist *evlist = top->evlist;
-	struct perf_sample sample;
-	struct perf_evsel *evsel;
 	struct perf_mmap *md;
-	struct perf_session *session = top->session;
 	union perf_event *event;
-	struct machine *machine;
-	int ret;
 
 	md = opts->overwrite ? &evlist->overwrite_mmap[idx] : &evlist->mmap[idx];
 	if (perf_mmap__read_init(md) < 0)
 		return;
 
 	while ((event = perf_mmap__read_event(md)) != NULL) {
-		ret = perf_evlist__parse_sample(evlist, event, &sample);
-		if (ret) {
-			pr_err("Can't parse sample, err = %d\n", ret);
-			goto next_event;
-		}
+		u64 timestamp = -1ULL;
+		int ret;
 
-		evsel = perf_evlist__id2evsel(session->evlist, sample.id);
-		assert(evsel != NULL);
-
-		if (event->header.type == PERF_RECORD_SAMPLE)
-			++top->samples;
-
-		switch (sample.cpumode) {
-		case PERF_RECORD_MISC_USER:
-			++top->us_samples;
-			if (top->hide_user_symbols)
-				goto next_event;
-			machine = &session->machines.host;
+		ret = perf_evlist__parse_sample_timestamp(evlist, event, &timestamp);
+		if (ret && ret != -1)
 			break;
-		case PERF_RECORD_MISC_KERNEL:
-			++top->kernel_samples;
-			if (top->hide_kernel_symbols)
-				goto next_event;
-			machine = &session->machines.host;
-			break;
-		case PERF_RECORD_MISC_GUEST_KERNEL:
-			++top->guest_kernel_samples;
-			machine = perf_session__find_machine(session,
-							     sample.pid);
-			break;
-		case PERF_RECORD_MISC_GUEST_USER:
-			++top->guest_us_samples;
-			/*
-			 * TODO: we don't process guest user from host side
-			 * except simple counting.
-			 */
-			goto next_event;
-		default:
-			if (event->header.type == PERF_RECORD_SAMPLE)
-				goto next_event;
-			machine = &session->machines.host;
-			break;
-		}
 
+		pthread_mutex_lock(&top->qe.lock);
+		ret = ordered_events__queue(top->qe.in, event, timestamp, 0);
+		pthread_mutex_unlock(&top->qe.lock);
 
-		if (event->header.type == PERF_RECORD_SAMPLE) {
-			perf_event__process_sample(&top->tool, event, evsel,
-						   &sample, machine);
-		} else if (event->header.type == PERF_RECORD_LOST) {
-			perf_top__process_lost(top, event, evsel);
-		} else if (event->header.type == PERF_RECORD_LOST_SAMPLES) {
-			perf_top__process_lost_samples(top, event, evsel);
-		} else if (event->header.type < PERF_RECORD_MAX) {
-			hists__inc_nr_events(evsel__hists(evsel), event->header.type);
-			machine__process_event(machine, event, &sample);
-		} else
-			++session->evlist->stats.nr_unknown_events;
-next_event:
 		perf_mmap__consume(md);
+		if (ret)
+			break;
 	}
 
 	perf_mmap__read_done(md);
@@ -1084,6 +1035,125 @@ static int callchain_param__setup_sample_type(struct callchain_param *callchain)
 	return 0;
 }
 
+static struct ordered_events *rotate_queues(struct perf_top *top)
+{
+	struct ordered_events *in = top->qe.in;
+
+	if (top->qe.in == &top->qe.data[1])
+		top->qe.in = &top->qe.data[0];
+	else
+		top->qe.in = &top->qe.data[1];
+
+	return in;
+}
+
+static void *process_thread(void *arg)
+{
+	struct perf_top *top = arg;
+
+	while (!done) {
+		struct ordered_events *out, *in = top->qe.in;
+
+		if (!in->nr_events) {
+			usleep(100);
+			continue;
+		}
+
+		pthread_mutex_lock(&top->qe.lock);
+		out = rotate_queues(top);
+		pthread_mutex_unlock(&top->qe.lock);
+
+		if (ordered_events__flush(out, OE_FLUSH__TOP))
+			pr_err("failed to process events\n");
+	}
+
+	return NULL;
+}
+
+static int deliver_event(struct ordered_events *qe,
+			 struct ordered_event *qevent)
+{
+	struct perf_top *top = qe->data;
+	struct perf_evlist *evlist = top->evlist;
+	struct perf_session *session = top->session;
+	union perf_event *event = qevent->event;
+	struct perf_sample sample;
+	struct perf_evsel *evsel;
+	struct machine *machine;
+	int ret = -1;
+
+	ret = perf_evlist__parse_sample(evlist, event, &sample);
+	if (ret) {
+		pr_err("Can't parse sample, err = %d\n", ret);
+		goto next_event;
+	}
+
+	evsel = perf_evlist__id2evsel(session->evlist, sample.id);
+	assert(evsel != NULL);
+
+	if (event->header.type == PERF_RECORD_SAMPLE)
+		++top->samples;
+
+	switch (sample.cpumode) {
+	case PERF_RECORD_MISC_USER:
+		++top->us_samples;
+		if (top->hide_user_symbols)
+			goto next_event;
+		machine = &session->machines.host;
+		break;
+	case PERF_RECORD_MISC_KERNEL:
+		++top->kernel_samples;
+		if (top->hide_kernel_symbols)
+			goto next_event;
+		machine = &session->machines.host;
+		break;
+	case PERF_RECORD_MISC_GUEST_KERNEL:
+		++top->guest_kernel_samples;
+		machine = perf_session__find_machine(session,
+						     sample.pid);
+		break;
+	case PERF_RECORD_MISC_GUEST_USER:
+		++top->guest_us_samples;
+		/*
+		 * TODO: we don't process guest user from host side
+		 * except simple counting.
+		 */
+		goto next_event;
+	default:
+		if (event->header.type == PERF_RECORD_SAMPLE)
+			goto next_event;
+		machine = &session->machines.host;
+		break;
+	}
+
+	if (event->header.type == PERF_RECORD_SAMPLE) {
+		perf_event__process_sample(&top->tool, event, evsel,
+					   &sample, machine);
+	} else if (event->header.type == PERF_RECORD_LOST) {
+		perf_top__process_lost(top, event, evsel);
+	} else if (event->header.type == PERF_RECORD_LOST_SAMPLES) {
+		perf_top__process_lost_samples(top, event, evsel);
+	} else if (event->header.type < PERF_RECORD_MAX) {
+		hists__inc_nr_events(evsel__hists(evsel), event->header.type);
+		machine__process_event(machine, event, &sample);
+	} else
+		++session->evlist->stats.nr_unknown_events;
+
+	ret = 0;
+next_event:
+	return ret;
+}
+
+static void init_process_thread(struct perf_top *top)
+{
+	ordered_events__init(&top->qe.data[0], deliver_event, top);
+	ordered_events__init(&top->qe.data[1], deliver_event, top);
+	ordered_events__set_copy_on_queue(&top->qe.data[0], true);
+	ordered_events__set_copy_on_queue(&top->qe.data[1], true);
+	top->qe.in = &top->qe.data[0];
+	pthread_mutex_init(&top->qe.lock, NULL);
+}
+
 static int __cmd_top(struct perf_top *top)
 {
 	char msg[512];
@@ -1091,7 +1161,7 @@ static int __cmd_top(struct perf_top *top)
 	struct perf_evsel_config_term *err_term;
 	struct perf_evlist *evlist = top->evlist;
 	struct record_opts *opts = &top->record_opts;
-	pthread_t thread;
+	pthread_t thread, thread_process;
 	int ret;
 
 	top->session = perf_session__new(NULL, false, NULL);
@@ -1115,6 +1185,8 @@ static int __cmd_top(struct perf_top *top)
 	if (top->nr_threads_synthesize > 1)
 		perf_set_multithreaded();
 
+	init_process_thread(top);
+
 	machine__synthesize_threads(&top->session->machines.host, &opts->target,
 				    top->evlist->threads, false,
 				    top->nr_threads_synthesize);
@@ -1155,10 +1227,15 @@ static int __cmd_top(struct perf_top *top)
                 perf_evlist__enable(top->evlist);
 
 	ret = -1;
+	if (pthread_create(&thread_process, NULL, process_thread, top)) {
+		ui__error("Could not create process thread.\n");
+		goto out_delete;
+	}
+
 	if (pthread_create(&thread, NULL, (use_browser > 0 ? display_thread_tui :
 							    display_thread), top)) {
 		ui__error("Could not create display thread.\n");
-		goto out_delete;
+		goto out_join_thread;
 	}
 
 	if (top->realtime_prio) {
@@ -1193,6 +1270,8 @@ static int __cmd_top(struct perf_top *top)
 	ret = 0;
 out_join:
 	pthread_join(thread, NULL);
+out_join_thread:
+	pthread_join(thread_process, NULL);
 out_delete:
 	perf_session__delete(top->session);
 	top->session = NULL;
@@ -1284,6 +1363,7 @@ int cmd_top(int argc, const char **argv)
 			 * stays in overwrite mode. -acme
 			 * */
 			.overwrite	= 0,
+			.sample_time	= true,
 		},
 		.max_stack	     = sysctl__max_stack(),
 		.annotation_opts     = annotation__default_options,

tools/perf/util/ordered-events.c

@@ -279,8 +279,10 @@ int ordered_events__flush(struct ordered_events *oe, enum oe_flush how)
 
 	switch (how) {
 	case OE_FLUSH__FINAL:
-		oe->next_flush = ULLONG_MAX;
 		show_progress = true;
+		__fallthrough;
+	case OE_FLUSH__TOP:
+		oe->next_flush = ULLONG_MAX;
 		break;
 
 	case OE_FLUSH__HALF:

tools/perf/util/ordered-events.h

@@ -18,6 +18,7 @@ enum oe_flush {
 	OE_FLUSH__FINAL,
 	OE_FLUSH__ROUND,
 	OE_FLUSH__HALF,
+	OE_FLUSH__TOP,
 };
 
 struct ordered_events;

tools/perf/util/top.h

@@ -40,6 +40,12 @@ struct perf_top {
 	const char	   *sym_filter;
 	float		   min_percent;
 	unsigned int	   nr_threads_synthesize;
+
+	struct {
+		struct ordered_events	*in;
+		struct ordered_events	 data[2];
+		pthread_mutex_t		 lock;
+	} qe;
 };
 
 #define CONSOLE_CLEAR ""