Merge branch 'unstable' into redis_6_merge

Former-commit-id: 32f34f63fd1025ae8337e8f30e9622abf08ee014

src/fastlock.cpp

@@ -74,6 +74,10 @@ extern int g_fInCrash;
     #define __has_feature(x) 0
 #endif
 
+#ifdef __linux__
+extern "C" void unlock_futex(struct fastlock *lock, uint16_t ifutex);
+#endif
+
 #if __has_feature(thread_sanitizer)
 
     /* Report that a lock has been created at address "lock". */
@@ -206,6 +210,11 @@ DeadlockDetector g_dlock;
 static_assert(sizeof(pid_t) <= sizeof(fastlock::m_pidOwner), "fastlock::m_pidOwner not large enough");
 uint64_t g_longwaits = 0;
 
+extern "C" void fastlock_panic(struct fastlock *lock)
+{
+    _serverPanic(__FILE__, __LINE__, "fastlock lock/unlock mismatch for: %s", lock->szName);
+}
+
 uint64_t fastlock_getlongwaitcount()
 {
     uint64_t rval;
@@ -337,31 +346,6 @@ extern "C" int fastlock_trylock(struct fastlock *lock, int fWeak)
     return false;
 }
 
-#ifdef __linux__
-#define ROL32(v, shift) ((v << shift) | (v >> (32-shift)))
-void unlock_futex(struct fastlock *lock, uint16_t ifutex)
-{
-    unsigned mask = (1U << (ifutex % 32));
-    unsigned futexT;
-    __atomic_load(&lock->futex, &futexT, __ATOMIC_RELAXED);
-    futexT &= mask;
-    
-    if (futexT == 0)
-        return;
-    
-    for (;;)
-    {
-        __atomic_load(&lock->futex, &futexT, __ATOMIC_ACQUIRE);
-        futexT &= mask;
-        if (!futexT)
-            break;
-
-        if (futex(&lock->m_ticket.u, FUTEX_WAKE_BITSET_PRIVATE, INT_MAX, nullptr, mask) == 1)
-            break;
-    }
-}
-#endif
-
 extern "C" void fastlock_unlock(struct fastlock *lock)
 {
     --lock->m_depth;
@@ -384,6 +368,26 @@ extern "C" void fastlock_unlock(struct fastlock *lock)
 }
 #endif
 
+#ifdef __linux__
+#define ROL32(v, shift) ((v << shift) | (v >> (32-shift)))
+extern "C" void unlock_futex(struct fastlock *lock, uint16_t ifutex)
+{
+    unsigned mask = (1U << (ifutex % 32));
+    unsigned futexT;
+    
+    for (;;)
+    {
+        __atomic_load(&lock->futex, &futexT, __ATOMIC_ACQUIRE);
+        futexT &= mask;
+        if (!futexT)
+            break;
+
+        if (futex(&lock->m_ticket.u, FUTEX_WAKE_BITSET_PRIVATE, INT_MAX, nullptr, mask) == 1)
+            break;
+    }
+}
+#endif
+
 extern "C" void fastlock_free(struct fastlock *lock)
 {
     // NOP

src/fastlock_x64.asm

@@ -126,6 +126,7 @@ fastlock_trylock:
 
 .ALIGN 16
 .global fastlock_unlock
+.type   fastlock_unlock,@function
 fastlock_unlock:
 	# RDI points to the struct:
 	#	int32_t m_pidOwner
@@ -133,34 +134,19 @@ fastlock_unlock:
 	# [rdi+64] ...
 	#	uint16_t active
 	#	uint16_t avail
-	push r11
 	sub dword ptr [rdi+4], 1     # decrement m_depth, don't use dec because it partially writes the flag register and we don't know its state
 	jnz .LDone                   # if depth is non-zero this is a recursive unlock, and we still hold it
 	mov dword ptr [rdi], -1      # pidOwner = -1 (we don't own it anymore)
-	mov ecx, [rdi+64]            # get current active (this one)
-	inc ecx                      # bump it to the next thread
-	mov [rdi+64], cx             # give up our ticket (note: lock is not required here because the spinlock itself guards this variable)
+	mov esi, [rdi+64]            # get current active (this one)
+	inc esi                      # bump it to the next thread
+	mov word ptr [rdi+64], si    # give up our ticket (note: lock is not required here because the spinlock itself guards this variable)
 	mfence                       # sync other threads
 	# At this point the lock is removed, however we must wake up any pending futexs
-	mov r9d, 1                   # eax is the bitmask for 2 threads
-	rol r9d, cl                  # place the mask in the right spot for the next 2 threads
-	add rdi, 64                  # rdi now points to the token
+	mov edx, [rdi+64+4]          # load the futex mask
+	bt edx, esi                  # is the next thread waiting on a futex?
+	jc unlock_futex              # unlock the futex if necessary
+	ret                          # if not we're done.
 .ALIGN 16
-.LRetryWake:
-	mov r11d, [rdi+4]            # load the futex mask
-	and r11d, r9d                # are any threads waiting on a futex?
-	jz .LDone                    # if not we're done.
-	# we have to wake the futexs
-                                 # rdi ARG1 futex (already in rdi)
-	mov esi, (10 | 128)          # rsi ARG2 FUTEX_WAKE_BITSET_PRIVATE
-	mov edx, 0x7fffffff          # rdx ARG3 INT_MAX (number of threads to wake)
-	xor r10d, r10d               # r10 ARG4 NULL
-	mov r8, rdi                  # r8 ARG5 dup rdi
-                                 # r9 ARG6 mask (already set above)
-	mov eax, 202                 # sys_futex
-	syscall
-	cmp eax, 1                   # did we wake as many as we expected?
-	jnz .LRetryWake
 .LDone:
-	pop r11
+	js fastlock_panic            # panic if we made m_depth negative
 	ret

src/networking.cpp

@@ -3193,6 +3193,7 @@ int processEventsWhileBlocked(int iel) {
         serverAssert(c->flags & CLIENT_PROTECTED);
         c->lock.unlock();
     }
+
     aeReleaseLock();
     serverAssertDebug(!GlobalLocksAcquired());
     try
@@ -3215,6 +3216,7 @@ int processEventsWhileBlocked(int iel) {
         locker.release();
         throw;
     }
+    
     AeLocker locker;
     if (c != nullptr)
         c->lock.lock();