Initial community commit

Src/replicant/nu/win-x86/LockFreeLIFO.asm

@@ -0,0 +1,50 @@
+.686
+.model FLAT
+
+PUBLIC _lifo_push
+_TEXT SEGMENT
+lifo = 4      ; size = 4
+entry = 8      ; size = 4
+_lifo_push PROC
+	mov ecx, DWORD PTR 4[esp] ; ecx holds lifo
+	mov edx, DWORD PTR 8[esp] ; edx holds the new entry
+again:
+	mov eax, DWORD PTR [ecx] ; eax holds the old head
+	mov DWORD PTR[edx], eax ; new node's 'next' is set to the old head
+	lock cmpxchg DWORD PTR [ecx], edx
+	jnz again
+	ret	0
+_lifo_push ENDP
+
+PUBLIC _lifo_pop
+_TEXT SEGMENT
+lifo = 4      ; size = 4
+_lifo_pop PROC
+	push esi
+	push ebx
+	mov	esi, DWORD PTR 12[esp] ; esi holds lifo
+again:
+    ; if re-ordered loads become an issue, we could use cmpxchg8b to read in (after zeroing ebx/ecx) or maybe use movq
+	mov edx, DWORD PTR [esi+4] ; counter
+	; or we could put an LFENCE here
+	mov eax, DWORD PTR [esi] ; pointer
+	test eax, eax
+	jz bail
+
+	mov ecx, edx ; counter
+	mov ebx, DWORD PTR [eax] ; pointer->next
+	inc ecx
+
+	lock cmpxchg8b QWORD PTR [esi]
+	jnz again
+
+bail:
+	pop ebx
+	pop esi
+	ret 0
+_lifo_pop ENDP
+
+_TEXT ENDS
+
+END
+

Src/replicant/nu/win-x86/LockFreeLIFO.c

@@ -0,0 +1,48 @@
+#include "LockFreeLIFO.h"
+#include "foundation/atomics.h"
+
+/* win32 implementation */
+
+void lifo_init(lifo_t *lifo)
+{
+	lifo->head = 0;
+	lifo->aba = 0;
+}
+
+#if 0 // defined in LockFreeLIFO.asm
+void lifo_push(lifo_t *lifo, queue_node_t *cl)
+{
+	queue_node_t *new_head = cl;
+	queue_node_t *old_head = 0;
+	do
+	{
+		old_head = (queue_node_t *)lifo->head;
+		new_head->Next = old_head;
+	} while (!nx_atomic_cmpxchg_pointer(old_head, new_head, (void * volatile *)&lifo->head));
+}
+
+queue_node_t *lifo_pop(lifo_t *lifo)
+{
+	lifo_t old_head, new_head;
+	do
+	{
+		old_head = *lifo;
+		if (!old_head.head)
+			return 0;
+
+		new_head.head = old_head.head->Next;
+		new_head.aba = old_head.aba+1;
+	} while (!nx_atomic_cmpxchg2(*(int64_t *)&old_head, *(int64_t *)&new_head, (volatile int64_t *)&lifo->head));
+	return (queue_node_t *)old_head.head;
+}
+#endif
+
+queue_node_t *lifo_malloc(size_t bytes)
+{
+	return _aligned_malloc(bytes, MEMORY_ALLOCATION_ALIGNMENT);
+}
+
+void lifo_free(queue_node_t *ptr)
+{
+	_aligned_free(ptr);
+}

Src/replicant/nu/win-x86/LockFreeLIFO.h

@@ -0,0 +1,31 @@
+#pragma once
+#include "foundation/types.h"
+#include "nu/queue_node.h"
+#include "foundation/align.h"
+/* lock free stack object 
+multiple threads can push and pop without locking
+note that order is not guaranteed.  that is, if Thread 1 calls Insert before Thread 2, Thread 2's item might still make it in before.
+*/
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+	typedef NALIGN(8) struct lifo_struct_t
+	{
+		volatile queue_node_t *head;
+		uint32_t aba;
+	} lifo_t;
+
+	/* use this to allocate an object that will go into this */
+	queue_node_t *lifo_malloc(size_t bytes); 
+	void lifo_free(queue_node_t *ptr);
+
+	void lifo_init(lifo_t *lifo);
+	void lifo_push(lifo_t *lifo, queue_node_t *cl);
+	queue_node_t *lifo_pop(lifo_t *lifo);
+
+	
+
+#ifdef __cplusplus
+}
+#endif

Src/replicant/nu/win-x86/ThreadLoop.cpp

@@ -0,0 +1,84 @@
+#include "ThreadLoop.h"
+
+lifo_t ThreadLoop::procedure_cache = {0,};
+lifo_t ThreadLoop::cache_bases= {0,};
+
+#define PROCEDURE_CACHE_SEED 64
+ThreadLoop::ThreadLoop()
+{
+	mpscq_init(&procedure_queue);
+	procedure_notification = CreateSemaphore(0, 0, LONG_MAX, 0);
+	kill_switch = CreateEvent(0, TRUE, FALSE, 0);
+}
+
+void ThreadLoop::RefillCache()
+{
+	threadloop_node_t *cache_seed = (threadloop_node_t *)malloc(PROCEDURE_CACHE_SEED*sizeof(threadloop_node_t));
+	
+	if (cache_seed)
+	{
+		int i=PROCEDURE_CACHE_SEED;
+		while (--i)
+		{
+			lifo_push(&procedure_cache, (queue_node_t *)&cache_seed[i]);
+		}
+		lifo_push(&cache_bases, (queue_node_t *)cache_seed);
+	}
+	else
+	{
+		Sleep(0); // yield and hope that someone else pops something off soon		
+	}
+	
+}
+
+void ThreadLoop::Run()
+{
+	HANDLE events[] = {kill_switch, procedure_notification};
+	while (WaitForMultipleObjects(2, events, FALSE, INFINITE) == WAIT_OBJECT_0 + 1)
+	{
+		for (;;)
+		{
+			threadloop_node_t *apc = (threadloop_node_t *)mpscq_pop(&procedure_queue);
+			if (apc == (threadloop_node_t *)1) /* special return value that indicates a busy list */
+			{
+				Sleep(0); // yield so that the thread that got pre-empted during push can finish
+			}
+			else
+			{
+				if (apc)
+				{
+					apc->func(apc->param1, apc->param2, apc->real_value);
+					lifo_push(&procedure_cache, apc);
+				}
+				else
+				{
+					break;	
+				}				
+			}
+		}
+	}
+}
+
+threadloop_node_t *ThreadLoop::GetAPC()
+{
+	threadloop_node_t *apc = 0;
+
+	do 
+	{
+		apc = (threadloop_node_t *)lifo_pop(&procedure_cache);
+		if (!apc)
+			RefillCache();
+	} while (!apc);
+	return apc;
+}
+
+void ThreadLoop::Schedule(threadloop_node_t *apc)
+{
+	if (mpscq_push(&procedure_queue, apc) == 0)
+		ReleaseSemaphore(procedure_notification, 1, 0);
+}
+
+void ThreadLoop::Kill()
+{
+	SetEvent(kill_switch);
+}

Src/replicant/nu/win-x86/ThreadLoop.h

@@ -0,0 +1,37 @@
+#pragma once
+#include "nu/lfmpscq.h"
+#include "nu/LockFreeLIFO.h"
+#include <windows.h>
+
+struct threadloop_node_t : public queue_node_t
+{
+	void (*func)(void *param1, void *param2, double real_value);
+
+	void *param1;
+	void *param2;
+	double real_value;
+};
+
+class ThreadLoop
+{
+public:
+	ThreadLoop();
+	threadloop_node_t *GetAPC(); // returns a node for you to fill out
+	void Schedule(threadloop_node_t *apc);
+	void Run();
+	void Kill();
+private:
+	void RefillCache();
+
+	HANDLE procedure_notification;
+	HANDLE kill_switch;
+	mpscq_t procedure_queue;
+
+	/* Memory cache to be able to run APCs without having the memory manager lock 
+	we'll allocate 100 at a time (#defined by PROCEDURE_CACHE_SEED)
+	and allocate new ones only if the cache is empty (which unfortunately will lock)
+	cache_bases holds the pointers we've allocated (to free on destruction of this object)
+	and procedure_cache holds the individual pointers */
+	static lifo_t procedure_cache;
+	static lifo_t cache_bases;
+};