Merge tag 'perf-tools-for-v5.10-2020-10-15' of git://git.kernel.org/pub/scm/linux/kernel/git/acme/linux

Pull perf tools updates from Arnaldo Carvalho de Melo:

 - cgroup improvements for 'perf stat', allowing for compact
   specification of events and cgroups in the command line.

 - Support per thread topdown metrics in 'perf stat'.

 - Support sample-read topdown metric group in 'perf record'

 - Show start of latency in addition to its start in 'perf sched
   latency'.

 - Add min, max to 'perf script' futex-contention output, in addition to
   avg.

 - Allow usage of 'perf_event_attr->exclusive' attribute via the new
   ':e' event modifier.

 - Add 'snapshot' command to 'perf record --control', using it with
   Intel PT.

 - Support FIFO file names as alternative options to 'perf record
   --control'.

 - Introduce branch history "streams", to compare 'perf record' runs
   with 'perf diff' based on branch records and report hot streams.

 - Support PE executable symbol tables using libbfd, to profile, for
   instance, wine binaries.

 - Add filter support for option 'perf ftrace -F/--funcs'.

 - Allow configuring the 'disassembler_style' 'perf annotate' knob via
   'perf config'

 - Update CascadelakeX and SkylakeX JSON vendor events files.

 - Add support for parsing perchip/percore JSON vendor events.

 - Add power9 hv_24x7 core level metric events.

 - Add L2 prefetch, ITLB instruction fetch hits JSON events for AMD
   zen1.

 - Enable Family 19h users by matching Zen2 AMD vendor events.

 - Use debuginfod in 'perf probe' when required debug files not found
   locally.

 - Display negative tid in non-sample events in 'perf script'.

 - Make GTK2 support opt-in

 - Add build test with GTK+

 - Add missing -lzstd to the fast path feature detection

 - Add scripts to auto generate 'mmap', 'mremap' string<->id tables for
   use in 'perf trace'.

 - Show python test script in verbose mode.

 - Fix uncore metric expressions

 - Msan uninitialized use fixes.

 - Use condition variables in 'perf bench numa'

 - Autodetect python3 binary in systems without python2.

 - Support md5 build ids in addition to sha1.

 - Add build id 'perf test' regression test.

 - Fix printable strings in python3 scripts.

 - Fix off by ones in 'perf trace' in arches using libaudit.

 - Fix JSON event code for events referencing std arch events.

 - Introduce 'perf test' shell script for Arm CoreSight testing.

 - Add rdtsc() for Arm64 for used in the PERF_RECORD_TIME_CONV metadata
   event and in 'perf test tsc'.

 - 'perf c2c' improvements: Add "RMT Load Hit" metric, "Total Stores",
   fixes and documentation update.

 - Fix usage of reloc_sym in 'perf probe' when using both kallsyms and
   debuginfo files.

 - Do not print 'Metric Groups:' unnecessarily in 'perf list'

 - Refcounting fixes in the event parsing code.

 - Add expand cgroup event 'perf test' entry.

 - Fix out of bounds CPU map access when handling armv8_pmu events in
   'perf stat'.

 - Add build-id injection 'perf bench' benchmark.

 - Enter namespace when reading build-id in 'perf inject'.

 - Do not load map/dso when injecting build-id speeding up the 'perf
   inject' process.

 - Add --buildid-all option to avoid processing all samples, just the
   mmap metadata events.

 - Add feature test to check if libbfd has buildid support

 - Add 'perf test' entry for PE binary format support.

 - Fix typos in power8 PMU vendor events JSON files.

 - Hide libtraceevent non API functions.

* tag 'perf-tools-for-v5.10-2020-10-15' of git://git.kernel.org/pub/scm/linux/kernel/git/acme/linux: (113 commits)
  perf c2c: Update documentation for metrics reorganization
  perf c2c: Add metrics "RMT Load Hit"
  perf c2c: Correct LLC load hit metrics
  perf c2c: Change header for LLC local hit
  perf c2c: Use more explicit headers for HITM
  perf c2c: Change header from "LLC Load Hitm" to "Load Hitm"
  perf c2c: Organize metrics based on memory hierarchy
  perf c2c: Display "Total Stores" as a standalone metrics
  perf c2c: Display the total numbers continuously
  perf bench: Use condition variables in numa.
  perf jevents: Fix event code for events referencing std arch events
  perf diff: Support hot streams comparison
  perf streams: Report hot streams
  perf streams: Calculate the sum of total streams hits
  perf streams: Link stream pair
  perf streams: Compare two streams
  perf streams: Get the evsel_streams by evsel_idx
  perf streams: Introduce branch history "streams"
  perf intel-pt: Improve PT documentation slightly
  perf tools: Add support for exclusive groups/events
  ...

4 files changed, 524 insertions, 21 deletions

diff --git a/tools/perf/bench/Build b/tools/perf/bench/Build
index 878db6a59a41..e43f46931b41 100644
--- a/tools/perf/bench/Build
+++ b/tools/perf/bench/Build
@@ -12,6 +12,7 @@ perf-y += epoll-ctl.o
 perf-y += synthesize.o
 perf-y += kallsyms-parse.o
 perf-y += find-bit-bench.o
+perf-y += inject-buildid.o
 
 perf-$(CONFIG_X86_64) += mem-memcpy-x86-64-asm.o
 perf-$(CONFIG_X86_64) += mem-memset-x86-64-asm.o
diff --git a/tools/perf/bench/bench.h b/tools/perf/bench/bench.h
index 2804812d4154..eac36afab2b3 100644
--- a/tools/perf/bench/bench.h
+++ b/tools/perf/bench/bench.h
@@ -47,6 +47,7 @@ int bench_epoll_wait(int argc, const char **argv);
 int bench_epoll_ctl(int argc, const char **argv);
 int bench_synthesize(int argc, const char **argv);
 int bench_kallsyms_parse(int argc, const char **argv);
+int bench_inject_build_id(int argc, const char **argv);
 
 #define BENCH_FORMAT_DEFAULT_STR	"default"
 #define BENCH_FORMAT_DEFAULT		0
diff --git a/tools/perf/bench/inject-buildid.c b/tools/perf/bench/inject-buildid.c
new file mode 100644
index 000000000000..280227e3ffd7
--- /dev/null
+++ b/tools/perf/bench/inject-buildid.c
@@ -0,0 +1,476 @@
+// SPDX-License-Identifier: GPL-2.0
+#include <stdlib.h>
+#include <stddef.h>
+#include <ftw.h>
+#include <fcntl.h>
+#include <errno.h>
+#include <unistd.h>
+#include <pthread.h>
+#include <sys/mman.h>
+#include <sys/wait.h>
+#include <linux/kernel.h>
+#include <linux/time64.h>
+#include <linux/list.h>
+#include <linux/err.h>
+#include <internal/lib.h>
+#include <subcmd/parse-options.h>
+
+#include "bench.h"
+#include "util/data.h"
+#include "util/stat.h"
+#include "util/debug.h"
+#include "util/event.h"
+#include "util/symbol.h"
+#include "util/session.h"
+#include "util/build-id.h"
+#include "util/synthetic-events.h"
+
+#define MMAP_DEV_MAJOR  8
+#define DSO_MMAP_RATIO  4
+
+static unsigned int iterations = 100;
+static unsigned int nr_mmaps   = 100;
+static unsigned int nr_samples = 100;  /* samples per mmap */
+
+static u64 bench_sample_type;
+static u16 bench_id_hdr_size;
+
+struct bench_data {
+	int			pid;
+	int			input_pipe[2];
+	int			output_pipe[2];
+	pthread_t		th;
+};
+
+struct bench_dso {
+	struct list_head	list;
+	char			*name;
+	int			ino;
+};
+
+static int nr_dsos;
+static struct bench_dso *dsos;
+
+extern int cmd_inject(int argc, const char *argv[]);
+
+static const struct option options[] = {
+	OPT_UINTEGER('i', "iterations", &iterations,
+		     "Number of iterations used to compute average (default: 100)"),
+	OPT_UINTEGER('m', "nr-mmaps", &nr_mmaps,
+		     "Number of mmap events for each iteration (default: 100)"),
+	OPT_UINTEGER('n', "nr-samples", &nr_samples,
+		     "Number of sample events per mmap event (default: 100)"),
+	OPT_INCR('v', "verbose", &verbose,
+		 "be more verbose (show iteration count, DSO name, etc)"),
+	OPT_END()
+};
+
+static const char *const bench_usage[] = {
+	"perf bench internals inject-build-id <options>",
+	NULL
+};
+
+/*
+ * Helper for collect_dso that adds the given file as a dso to dso_list
+ * if it contains a build-id.  Stops after collecting 4 times more than
+ * we need (for MMAP2 events).
+ */
+static int add_dso(const char *fpath, const struct stat *sb __maybe_unused,
+		   int typeflag, struct FTW *ftwbuf __maybe_unused)
+{
+	struct bench_dso *dso = &dsos[nr_dsos];
+	struct build_id bid;
+
+	if (typeflag == FTW_D || typeflag == FTW_SL)
+		return 0;
+
+	if (filename__read_build_id(fpath, &bid) < 0)
+		return 0;
+
+	dso->name = realpath(fpath, NULL);
+	if (dso->name == NULL)
+		return -1;
+
+	dso->ino = nr_dsos++;
+	pr_debug2("  Adding DSO: %s\n", fpath);
+
+	/* stop if we collected enough DSOs */
+	if ((unsigned int)nr_dsos == DSO_MMAP_RATIO * nr_mmaps)
+		return 1;
+
+	return 0;
+}
+
+static void collect_dso(void)
+{
+	dsos = calloc(nr_mmaps * DSO_MMAP_RATIO, sizeof(*dsos));
+	if (dsos == NULL) {
+		printf("  Memory allocation failed\n");
+		exit(1);
+	}
+
+	if (nftw("/usr/lib/", add_dso, 10, FTW_PHYS) < 0)
+		return;
+
+	pr_debug("  Collected %d DSOs\n", nr_dsos);
+}
+
+static void release_dso(void)
+{
+	int i;
+
+	for (i = 0; i < nr_dsos; i++) {
+		struct bench_dso *dso = &dsos[i];
+
+		free(dso->name);
+	}
+	free(dsos);
+}
+
+/* Fake address used by mmap and sample events */
+static u64 dso_map_addr(struct bench_dso *dso)
+{
+	return 0x400000ULL + dso->ino * 8192ULL;
+}
+
+static u32 synthesize_attr(struct bench_data *data)
+{
+	union perf_event event;
+
+	memset(&event, 0, sizeof(event.attr) + sizeof(u64));
+
+	event.header.type = PERF_RECORD_HEADER_ATTR;
+	event.header.size = sizeof(event.attr) + sizeof(u64);
+
+	event.attr.attr.type = PERF_TYPE_SOFTWARE;
+	event.attr.attr.config = PERF_COUNT_SW_TASK_CLOCK;
+	event.attr.attr.exclude_kernel = 1;
+	event.attr.attr.sample_id_all = 1;
+	event.attr.attr.sample_type = bench_sample_type;
+
+	return writen(data->input_pipe[1], &event, event.header.size);
+}
+
+static u32 synthesize_fork(struct bench_data *data)
+{
+	union perf_event event;
+
+	memset(&event, 0, sizeof(event.fork) + bench_id_hdr_size);
+
+	event.header.type = PERF_RECORD_FORK;
+	event.header.misc = PERF_RECORD_MISC_FORK_EXEC;
+	event.header.size = sizeof(event.fork) + bench_id_hdr_size;
+
+	event.fork.ppid = 1;
+	event.fork.ptid = 1;
+	event.fork.pid = data->pid;
+	event.fork.tid = data->pid;
+
+	return writen(data->input_pipe[1], &event, event.header.size);
+}
+
+static u32 synthesize_mmap(struct bench_data *data, struct bench_dso *dso,
+			   u64 timestamp)
+{
+	union perf_event event;
+	size_t len = offsetof(struct perf_record_mmap2, filename);
+	u64 *id_hdr_ptr = (void *)&event;
+	int ts_idx;
+
+	len += roundup(strlen(dso->name) + 1, 8) + bench_id_hdr_size;
+
+	memset(&event, 0, min(len, sizeof(event.mmap2)));
+
+	event.header.type = PERF_RECORD_MMAP2;
+	event.header.misc = PERF_RECORD_MISC_USER;
+	event.header.size = len;
+
+	event.mmap2.pid = data->pid;
+	event.mmap2.tid = data->pid;
+	event.mmap2.maj = MMAP_DEV_MAJOR;
+	event.mmap2.ino = dso->ino;
+
+	strcpy(event.mmap2.filename, dso->name);
+
+	event.mmap2.start = dso_map_addr(dso);
+	event.mmap2.len = 4096;
+	event.mmap2.prot = PROT_EXEC;
+
+	if (len > sizeof(event.mmap2)) {
+		/* write mmap2 event first */
+		writen(data->input_pipe[1], &event, len - bench_id_hdr_size);
+		/* zero-fill sample id header */
+		memset(id_hdr_ptr, 0, bench_id_hdr_size);
+		/* put timestamp in the right position */
+		ts_idx = (bench_id_hdr_size / sizeof(u64)) - 2;
+		id_hdr_ptr[ts_idx] = timestamp;
+		writen(data->input_pipe[1], id_hdr_ptr, bench_id_hdr_size);
+	} else {
+		ts_idx = (len / sizeof(u64)) - 2;
+		id_hdr_ptr[ts_idx] = timestamp;
+		writen(data->input_pipe[1], &event, len);
+	}
+	return len;
+}
+
+static u32 synthesize_sample(struct bench_data *data, struct bench_dso *dso,
+			     u64 timestamp)
+{
+	union perf_event event;
+	struct perf_sample sample = {
+		.tid = data->pid,
+		.pid = data->pid,
+		.ip = dso_map_addr(dso),
+		.time = timestamp,
+	};
+
+	event.header.type = PERF_RECORD_SAMPLE;
+	event.header.misc = PERF_RECORD_MISC_USER;
+	event.header.size = perf_event__sample_event_size(&sample, bench_sample_type, 0);
+
+	perf_event__synthesize_sample(&event, bench_sample_type, 0, &sample);
+
+	return writen(data->input_pipe[1], &event, event.header.size);
+}
+
+static u32 synthesize_flush(struct bench_data *data)
+{
+	struct perf_event_header header = {
+		.size = sizeof(header),
+		.type = PERF_RECORD_FINISHED_ROUND,
+	};
+
+	return writen(data->input_pipe[1], &header, header.size);
+}
+
+static void *data_reader(void *arg)
+{
+	struct bench_data *data = arg;
+	char buf[8192];
+	int flag;
+	int n;
+
+	flag = fcntl(data->output_pipe[0], F_GETFL);
+	fcntl(data->output_pipe[0], F_SETFL, flag | O_NONBLOCK);
+
+	/* read out data from child */
+	while (true) {
+		n = read(data->output_pipe[0], buf, sizeof(buf));
+		if (n > 0)
+			continue;
+		if (n == 0)
+			break;
+
+		if (errno != EINTR && errno != EAGAIN)
+			break;
+
+		usleep(100);
+	}
+
+	close(data->output_pipe[0]);
+	return NULL;
+}
+
+static int setup_injection(struct bench_data *data, bool build_id_all)
+{
+	int ready_pipe[2];
+	int dev_null_fd;
+	char buf;
+
+	if (pipe(ready_pipe) < 0)
+		return -1;
+
+	if (pipe(data->input_pipe) < 0)
+		return -1;
+
+	if (pipe(data->output_pipe) < 0)
+		return -1;
+
+	data->pid = fork();
+	if (data->pid < 0)
+		return -1;
+
+	if (data->pid == 0) {
+		const char **inject_argv;
+		int inject_argc = 2;
+
+		close(data->input_pipe[1]);
+		close(data->output_pipe[0]);
+		close(ready_pipe[0]);
+
+		dup2(data->input_pipe[0], STDIN_FILENO);
+		close(data->input_pipe[0]);
+		dup2(data->output_pipe[1], STDOUT_FILENO);
+		close(data->output_pipe[1]);
+
+		dev_null_fd = open("/dev/null", O_WRONLY);
+		if (dev_null_fd < 0)
+			exit(1);
+
+		dup2(dev_null_fd, STDERR_FILENO);
+
+		if (build_id_all)
+			inject_argc++;
+
+		inject_argv = calloc(inject_argc + 1, sizeof(*inject_argv));
+		if (inject_argv == NULL)
+			exit(1);
+
+		inject_argv[0] = strdup("inject");
+		inject_argv[1] = strdup("-b");
+		if (build_id_all)
+			inject_argv[2] = strdup("--buildid-all");
+
+		/* signal that we're ready to go */
+		close(ready_pipe[1]);
+
+		cmd_inject(inject_argc, inject_argv);
+
+		exit(0);
+	}
+
+	pthread_create(&data->th, NULL, data_reader, data);
+
+	close(ready_pipe[1]);
+	close(data->input_pipe[0]);
+	close(data->output_pipe[1]);
+
+	/* wait for child ready */
+	if (read(ready_pipe[0], &buf, 1) < 0)
+		return -1;
+	close(ready_pipe[0]);
+
+	return 0;
+}
+
+static int inject_build_id(struct bench_data *data, u64 *max_rss)
+{
+	int status;
+	unsigned int i, k;
+	struct rusage rusage;
+	u64 len = 0;
+
+	/* this makes the child to run */
+	if (perf_header__write_pipe(data->input_pipe[1]) < 0)
+		return -1;
+
+	len += synthesize_attr(data);
+	len += synthesize_fork(data);
+
+	for (i = 0; i < nr_mmaps; i++) {
+		int idx = rand() % (nr_dsos - 1);
+		struct bench_dso *dso = &dsos[idx];
+		u64 timestamp = rand() % 1000000;
+
+		pr_debug2("   [%d] injecting: %s\n", i+1, dso->name);
+		len += synthesize_mmap(data, dso, timestamp);
+
+		for (k = 0; k < nr_samples; k++)
+			len += synthesize_sample(data, dso, timestamp + k * 1000);
+
+		if ((i + 1) % 10 == 0)
+			len += synthesize_flush(data);
+	}
+
+	/* tihs makes the child to finish */
+	close(data->input_pipe[1]);
+
+	wait4(data->pid, &status, 0, &rusage);
+	*max_rss = rusage.ru_maxrss;
+
+	pr_debug("   Child %d exited with %d\n", data->pid, status);
+
+	return 0;
+}
+
+static void do_inject_loop(struct bench_data *data, bool build_id_all)
+{
+	unsigned int i;
+	struct stats time_stats, mem_stats;
+	double time_average, time_stddev;
+	double mem_average, mem_stddev;
+
+	init_stats(&time_stats);
+	init_stats(&mem_stats);
+
+	pr_debug("  Build-id%s injection benchmark\n", build_id_all ? "-all" : "");
+
+	for (i = 0; i < iterations; i++) {
+		struct timeval start, end, diff;
+		u64 runtime_us, max_rss;
+
+		pr_debug("  Iteration #%d\n", i+1);
+
+		if (setup_injection(data, build_id_all) < 0) {
+			printf("  Build-id injection setup failed\n");
+			break;
+		}
+
+		gettimeofday(&start, NULL);
+		if (inject_build_id(data, &max_rss) < 0) {
+			printf("  Build-id injection failed\n");
+			break;
+		}
+
+		gettimeofday(&end, NULL);
+		timersub(&end, &start, &diff);
+		runtime_us = diff.tv_sec * USEC_PER_SEC + diff.tv_usec;
+		update_stats(&time_stats, runtime_us);
+		update_stats(&mem_stats, max_rss);
+
+		pthread_join(data->th, NULL);
+	}
+
+	time_average = avg_stats(&time_stats) / USEC_PER_MSEC;
+	time_stddev = stddev_stats(&time_stats) / USEC_PER_MSEC;
+	printf("  Average build-id%s injection took: %.3f msec (+- %.3f msec)\n",
+	       build_id_all ? "-all" : "", time_average, time_stddev);
+
+	/* each iteration, it processes MMAP2 + BUILD_ID + nr_samples * SAMPLE */
+	time_average = avg_stats(&time_stats) / (nr_mmaps * (nr_samples + 2));
+	time_stddev = stddev_stats(&time_stats) / (nr_mmaps * (nr_samples + 2));
+	printf("  Average time per event: %.3f usec (+- %.3f usec)\n",
+		time_average, time_stddev);
+
+	mem_average = avg_stats(&mem_stats);
+	mem_stddev = stddev_stats(&mem_stats);
+	printf("  Average memory usage: %.0f KB (+- %.0f KB)\n",
+		mem_average, mem_stddev);
+}
+
+static int do_inject_loops(struct bench_data *data)
+{
+
+	srand(time(NULL));
+	symbol__init(NULL);
+
+	bench_sample_type  = PERF_SAMPLE_IDENTIFIER | PERF_SAMPLE_IP;
+	bench_sample_type |= PERF_SAMPLE_TID | PERF_SAMPLE_TIME;
+	bench_id_hdr_size  = 32;
+
+	collect_dso();
+	if (nr_dsos == 0) {
+		printf("  Cannot collect DSOs for injection\n");
+		return -1;
+	}
+
+	do_inject_loop(data, false);
+	do_inject_loop(data, true);
+
+	release_dso();
+	return 0;
+}
+
+int bench_inject_build_id(int argc, const char **argv)
+{
+	struct bench_data data;
+
+	argc = parse_options(argc, argv, options, bench_usage, 0);
+	if (argc) {
+		usage_with_options(bench_usage, options);
+		exit(EXIT_FAILURE);
+	}
+
+	return do_inject_loops(&data);
+}
+
diff --git a/tools/perf/bench/numa.c b/tools/perf/bench/numa.c
index f85bceccc459..11726ec6285f 100644
--- a/tools/perf/bench/numa.c
+++ b/tools/perf/bench/numa.c
@@ -137,12 +137,13 @@ struct global_info {
 	u8			*data;
 
 	pthread_mutex_t		startup_mutex;
+	pthread_cond_t		startup_cond;
 	int			nr_tasks_started;
 
-	pthread_mutex_t		startup_done_mutex;
-
 	pthread_mutex_t		start_work_mutex;
+	pthread_cond_t		start_work_cond;
 	int			nr_tasks_working;
+	bool			start_work;
 
 	pthread_mutex_t		stop_work_mutex;
 	u64			bytes_done;
@@ -483,6 +484,18 @@ static void init_global_mutex(pthread_mutex_t *mutex)
 	pthread_mutex_init(mutex, &attr);
 }
 
+/*
+ * Return a process-shared (global) condition variable:
+ */
+static void init_global_cond(pthread_cond_t *cond)
+{
+	pthread_condattr_t attr;
+
+	pthread_condattr_init(&attr);
+	pthread_condattr_setpshared(&attr, PTHREAD_PROCESS_SHARED);
+	pthread_cond_init(cond, &attr);
+}
+
 static int parse_cpu_list(const char *arg)
 {
 	p0.cpu_list_str = strdup(arg);
@@ -1136,15 +1149,18 @@ static void *worker_thread(void *__tdata)
 	if (g->p.serialize_startup) {
 		pthread_mutex_lock(&g->startup_mutex);
 		g->nr_tasks_started++;
+		/* The last thread wakes the main process. */
+		if (g->nr_tasks_started == g->p.nr_tasks)
+			pthread_cond_signal(&g->startup_cond);
+
 		pthread_mutex_unlock(&g->startup_mutex);
 
 		/* Here we will wait for the main process to start us all at once: */
 		pthread_mutex_lock(&g->start_work_mutex);
+		g->start_work = false;
 		g->nr_tasks_working++;
-
-		/* Last one wake the main process: */
-		if (g->nr_tasks_working == g->p.nr_tasks)
-			pthread_mutex_unlock(&g->startup_done_mutex);
+		while (!g->start_work)
+			pthread_cond_wait(&g->start_work_cond, &g->start_work_mutex);
 
 		pthread_mutex_unlock(&g->start_work_mutex);
 	}
@@ -1441,8 +1457,9 @@ static int init(void)
 
 	/* Startup serialization: */
 	init_global_mutex(&g->start_work_mutex);
+	init_global_cond(&g->start_work_cond);
 	init_global_mutex(&g->startup_mutex);
-	init_global_mutex(&g->startup_done_mutex);
+	init_global_cond(&g->startup_cond);
 	init_global_mutex(&g->stop_work_mutex);
 
 	init_thread_data();
@@ -1502,9 +1519,6 @@ static int __bench_numa(const char *name)
 	pids = zalloc(g->p.nr_proc * sizeof(*pids));
 	pid = -1;
 
-	/* All threads try to acquire it, this way we can wait for them to start up: */
-	pthread_mutex_lock(&g->start_work_mutex);
-
 	if (g->p.serialize_startup) {
 		tprintf(" #\n");
 		tprintf(" # Startup synchronization: ..."); fflush(stdout);
@@ -1526,22 +1540,29 @@ static int __bench_numa(const char *name)
 		pids[i] = pid;
 
 	}
-	/* Wait for all the threads to start up: */
-	while (g->nr_tasks_started != g->p.nr_tasks)
-		usleep(USEC_PER_MSEC);
-
-	BUG_ON(g->nr_tasks_started != g->p.nr_tasks);
 
 	if (g->p.serialize_startup) {
+		bool threads_ready = false;
 		double startup_sec;
 
-		pthread_mutex_lock(&g->startup_done_mutex);
+		/*
+		 * Wait for all the threads to start up. The last thread will
+		 * signal this process.
+		 */
+		pthread_mutex_lock(&g->startup_mutex);
+		while (g->nr_tasks_started != g->p.nr_tasks)
+			pthread_cond_wait(&g->startup_cond, &g->startup_mutex);
 
-		/* This will start all threads: */
-		pthread_mutex_unlock(&g->start_work_mutex);
+		pthread_mutex_unlock(&g->startup_mutex);
 
-		/* This mutex is locked - the last started thread will wake us: */
-		pthread_mutex_lock(&g->startup_done_mutex);
+		/* Wait for all threads to be at the start_work_cond. */
+		while (!threads_ready) {
+			pthread_mutex_lock(&g->start_work_mutex);
+			threads_ready = (g->nr_tasks_working == g->p.nr_tasks);
+			pthread_mutex_unlock(&g->start_work_mutex);
+			if (!threads_ready)
+				usleep(1);
+		}
 
 		gettimeofday(&stop, NULL);
 
@@ -1555,7 +1576,11 @@ static int __bench_numa(const char *name)
 		tprintf(" #\n");
 
 		start = stop;
-		pthread_mutex_unlock(&g->startup_done_mutex);
+		/* Start all threads running. */
+		pthread_mutex_lock(&g->start_work_mutex);
+		g->start_work = true;
+		pthread_mutex_unlock(&g->start_work_mutex);
+		pthread_cond_broadcast(&g->start_work_cond);
 	} else {
 		gettimeofday(&start, NULL);
 	}