From 40fa6e2023626c750ba8d87b41e5d0f0e3f43ce7 Mon Sep 17 00:00:00 2001
From: Robin Gareus <robin@gareus.org>
Date: Thu, 12 May 2022 04:49:11 +0200
Subject: [PATCH] Improve PBD::Semaphore performance

Skip syscall if no threads are waiting and decouple atomic
value from futex address so that FUTEX_WAIT does not fail.
---
 libs/pbd/pbd/semutils.h |  3 ++-
 libs/pbd/semutils.cc    | 34 ++++++++++++----------------------
 2 files changed, 14 insertions(+), 23 deletions(-)

diff --git a/libs/pbd/pbd/semutils.h b/libs/pbd/pbd/semutils.h
index b1b6bf41f4..bdc47072d2 100644
--- a/libs/pbd/pbd/semutils.h
+++ b/libs/pbd/pbd/semutils.h
@@ -47,7 +47,8 @@ class LIBPBD_API Semaphore {
 	sem_t* _sem;
 	sem_t* ptr_to_sem() const { return _sem; }
 #elif defined USE_FUTEX_SEMAPHORE
-	std::atomic<unsigned int> _value;
+	int              _futex;
+	std::atomic<int> _value;
 #else
 	mutable sem_t _sem;
 	sem_t* ptr_to_sem() const { return &_sem; }
diff --git a/libs/pbd/semutils.cc b/libs/pbd/semutils.cc
index ba7b4e2bcb..0d8c603aa2 100644
--- a/libs/pbd/semutils.cc
+++ b/libs/pbd/semutils.cc
@@ -23,6 +23,7 @@
 #ifdef USE_FUTEX_SEMAPHORE
 #include <errno.h>
 #include <linux/futex.h>
+#include <sched.h>
 #include <syscall.h>
 #include <unistd.h>
 #endif
@@ -103,23 +104,19 @@ Semaphore::reset ()
 int
 Semaphore::signal ()
 {
-	std::atomic_fetch_add_explicit (&_value, 1, std::memory_order_release);
-	return syscall (__NR_futex, &_value, FUTEX_WAKE_PRIVATE, 1, NULL, NULL, 0);
+	if (std::atomic_fetch_add_explicit (&_value, 1, std::memory_order_relaxed) < 0) {
+		while (syscall (__NR_futex, &_futex, FUTEX_WAKE_PRIVATE, 1, NULL, NULL, 0) < 1) {
+			sched_yield();
+		}
+	}
+	return 0;
 }
 
 int
 Semaphore::wait ()
 {
-	unsigned int value = 1;
-	while (!std::atomic_compare_exchange_weak_explicit (&_value, &value, value - 1, std::memory_order_acquire, std::memory_order_relaxed)) {
-		if (value == 0) {
-			if (syscall (__NR_futex, &_value, FUTEX_WAIT_PRIVATE, 0, NULL, NULL, 0)) {
-				if (errno != EAGAIN && errno != EINTR) {
-					return 1;
-				}
-			}
-			value = 1;
-		}
+	if (std::atomic_fetch_sub_explicit (&_value, 1, std::memory_order_relaxed) <= 0) {
+		syscall(__NR_futex, &_futex, FUTEX_WAIT_PRIVATE, _futex, NULL, 0, 0);
 	}
 	return 0;
 }
@@ -127,16 +124,9 @@ Semaphore::wait ()
 int
 Semaphore::reset ()
 {
-	int rv = 0;
-	unsigned int value = 1;
-
-	while (!std::atomic_compare_exchange_weak_explicit (&_value, &value, value - 1, std::memory_order_acquire, std::memory_order_relaxed)) {
-		if (value == 0) {
-			break;
-		}
-		++rv;
-	}
-	return rv;
+	int value = _value;
+	_value = 0;
+	return value;
 }
 
 #endif