From ad5bbee6b2c63c0718263f3929180f93b646d77c Mon Sep 17 00:00:00 2001
From: J Reece Wilson <me@reece.sx>
Date: Fri, 18 Aug 2023 22:51:47 +0100
Subject: [PATCH] [*] Linux: optimize futex wrappers

---
 Source/AuProcAddresses.Linux.cpp        | 63 ++++++++++++++++---------
 Source/AuProcAddresses.Linux.hpp        | 19 +++++---
 Source/IO/IPC/AuIPCMutexFutex.Linux.cpp |  4 +-
 3 files changed, 54 insertions(+), 32 deletions(-)

diff --git a/Source/AuProcAddresses.Linux.cpp b/Source/AuProcAddresses.Linux.cpp
index ef17f951..a2e0a969 100644
--- a/Source/AuProcAddresses.Linux.cpp
+++ b/Source/AuProcAddresses.Linux.cpp
@@ -49,34 +49,24 @@ namespace Aurora
         return syscall(SYS_futex, uaddr, futex_op, val, timeout, uaddr2, val3);
     }
 
-    int futex_wait(volatile uint32_t *addr, uint32_t expected)
-    {
-        return futex((uint32_t *)addr, FUTEX_WAIT, expected, 0, 0, 0);
-    }
-
-    int futex_wait(volatile uint32_t *addr, uint32_t expected, const struct timespec *timeout)
-    {
-        if (timeout)
-        {
-            return futex((uint32_t *)addr, FUTEX_WAIT_BITSET, expected, timeout, 0, FUTEX_BITSET_MATCH_ANY);
-        }
-        else
-        {
-            return futex((uint32_t *)addr, FUTEX_WAIT, expected, timeout, 0, 0);
-        }
-    }
-
-    int futex_wake(volatile uint32_t *addr, uint32_t nthreads)
-    {
-        return futex((uint32_t *)addr, FUTEX_WAKE, nthreads, 0, 0, 0);
-    }
-
     int futex_wait(uint32_t *addr, uint32_t expected)
     {
-        return futex(addr, FUTEX_WAIT, expected, 0, 0, 0);
+        return futex(addr, FUTEX_WAIT | FUTEX_PRIVATE_FLAG, expected, 0, 0, 0);
     }
 
     int futex_wait(uint32_t *addr, uint32_t expected, const struct timespec *timeout)
+    {
+        if (timeout)
+        {
+            return futex(addr, FUTEX_WAIT_BITSET | FUTEX_PRIVATE_FLAG, expected, timeout, 0, FUTEX_BITSET_MATCH_ANY);
+        }
+        else
+        {
+            return futex(addr, FUTEX_WAIT | FUTEX_PRIVATE_FLAG, expected, timeout, 0, 0);
+        }
+    }
+
+    int futex_wait_shared(uint32_t *addr, uint32_t expected, const struct timespec *timeout)
     {
         if (timeout)
         {
@@ -89,10 +79,37 @@ namespace Aurora
     }
 
     int futex_wake(uint32_t *addr, uint32_t nthreads)
+    {
+        return futex(addr, FUTEX_WAKE | FUTEX_PRIVATE_FLAG, nthreads, 0, 0, 0);
+    }
+
+    int futex_wake_shared(uint32_t *addr, uint32_t nthreads)
     {
         return futex(addr, FUTEX_WAKE, nthreads, 0, 0, 0);
     }
 
+    int futex_wait(volatile uint32_t *addr, uint32_t expected)
+    {
+        return futex((uint32_t *)addr, FUTEX_WAIT | FUTEX_PRIVATE_FLAG, expected, 0, 0, 0);
+    }
+
+    int futex_wait(volatile uint32_t *addr, uint32_t expected, const struct timespec *timeout)
+    {
+        if (timeout)
+        {
+            return futex((uint32_t *)addr, FUTEX_WAIT_BITSET | FUTEX_PRIVATE_FLAG, expected, timeout, 0, FUTEX_BITSET_MATCH_ANY);
+        }
+        else
+        {
+            return futex((uint32_t *)addr, FUTEX_WAIT | FUTEX_PRIVATE_FLAG, expected, timeout, 0, 0);
+        }
+    }
+
+    int futex_wake(volatile uint32_t *addr, uint32_t nthreads)
+    {
+        return futex((uint32_t *)addr, FUTEX_WAKE | FUTEX_PRIVATE_FLAG, nthreads, 0, 0, 0);
+    }
+
     int io_setup(unsigned nr, aio_context_t *ctxp)
     {
         return syscall(__NR_io_setup, nr, ctxp);
diff --git a/Source/AuProcAddresses.Linux.hpp b/Source/AuProcAddresses.Linux.hpp
index e47f6ffa..420571db 100644
--- a/Source/AuProcAddresses.Linux.hpp
+++ b/Source/AuProcAddresses.Linux.hpp
@@ -24,13 +24,6 @@ namespace Aurora
 
     long set_robust_list(struct robust_list_head *head, size_t len);
 
-    int futex_wake(volatile uint32_t *addr, uint32_t nthreads);
-
-    int futex_wait(volatile uint32_t *addr, uint32_t expected,
-                   const struct timespec *timeout);
-    
-    int futex_wait(volatile uint32_t *addr, uint32_t expected);
-
     int futex_wake(uint32_t *addr, uint32_t nthreads);
 
     int futex_wait(uint32_t *addr, uint32_t expected,
@@ -38,6 +31,18 @@ namespace Aurora
     
     int futex_wait(uint32_t *addr, uint32_t expected);
 
+    int futex_wake_shared(uint32_t *addr, uint32_t nthreads);
+
+    int futex_wait_shared(uint32_t *addr, uint32_t expected,
+                          const struct timespec *timeout);
+    
+    int futex_wake(volatile uint32_t *addr, uint32_t nthreads);
+
+    int futex_wait(volatile uint32_t *addr, uint32_t expected,
+                   const struct timespec *timeout);
+    
+    int futex_wait(volatile uint32_t *addr, uint32_t expected);
+
     int io_submit(aio_context_t ctx, long nr, struct iocb **iocbpp);
     
     int io_cancel(aio_context_t ctx_id, struct iocb *iocb,
diff --git a/Source/IO/IPC/AuIPCMutexFutex.Linux.cpp b/Source/IO/IPC/AuIPCMutexFutex.Linux.cpp
index 69aac763..0be2a5ce 100644
--- a/Source/IO/IPC/AuIPCMutexFutex.Linux.cpp
+++ b/Source/IO/IPC/AuIPCMutexFutex.Linux.cpp
@@ -322,7 +322,7 @@ static bool LinuxLockFutex(AuUInt32 *futex, AuUInt32 timeout)
         bContended = old != kFutexValueUnlocked;
         if (bContended)
         {
-            int res = Aurora::futex_wait(futex, old, timeout ? &tspec : nullptr);
+            int res = Aurora::futex_wait_shared(futex, old, timeout ? &tspec : nullptr);
             if (res < 0)
             {
                 if (res == ETIMEDOUT || errno == ETIMEDOUT)
@@ -375,7 +375,7 @@ static bool LinuxUnlockFutex(AuUInt32 *futex)
         return false;
     }
 
-    Aurora::futex_wake(futex, 1);
+    Aurora::futex_wake_shared(futex, 1);
 
     return true;
 }