add zcontext code

example/echo_client/echo_client_stackfull.cpp

@@ -36,14 +36,25 @@ static void echo_client(HANDLE iocp_handle, const char* lp_server_addr)
 
 	FiberOVERLAPPED ov;
 
+
 	auto result = WSAConnectEx(sock, (const SOCKADDR*) &server_addr, INET_ADDRSTRLEN, 0, 0, 0, &ov.ov);
+
+	float c = ov.last_error + 1;
+
+	c*=99;
+	printf("c = %f before\n", c);
+
 	ov.last_error = WSAGetLastError();
 
 	if (!(!result && ov.last_error != WSA_IO_PENDING))
 	{
 		get_overlapped_result(&ov);
 	}
 
+	c /= 99;
+
+	printf("c = %f\n", c);
+
 	if (ov.last_error)
 	{
 		printf("connection failed\n");
@@ -113,7 +124,13 @@ int main(int argc, char* argv[])
 
 #endif
 
+	float cc = 8;
+	cc *= argc;
+	printf("cc = %f before\n", cc);
 	create_detached_coroutine(echo_client, iocp_handle, (const char*)(argc == 2 ? argv[1] : "127.0.0.1"));
+	cc *= argc;
+
+	printf("cc = %f\n", cc);
 
 	run_event_loop(iocp_handle);
 	CloseHandle(iocp_handle);

µasync/CMakeLists.txt

@@ -2,6 +2,7 @@
 
 option(DISABLE_BOOST_CONTEXT "use ucontext even if boost is found" OFF)
 option(DISABLE_UCONTEXT "disable ucontext and use setjmp/longjmp" OFF)
+option(USE_ZCONTEXT "use zcontext api to do context switch" OFF)
 
 cmake_policy(SET CMP0167 NEW)
 find_package(Boost 1.86 COMPONENTS context)
@@ -24,4 +25,11 @@ endif()
 
 if (DISABLE_UCONTEXT)
 target_compile_definitions(uasync INTERFACE -DDISABLE_UCONTEXT=1)
-endif()
+endif()
+
+if (USE_ZCONTEXT)
+enable_language(ASM)
+add_library(zcontext asm/zcontext_x86_64.S)
+target_compile_definitions(uasync INTERFACE -DUSE_ZCONTEXT=1)
+target_link_libraries(uasync INTERFACE zcontext)
+endif()

µasync/asm/zcontext_x86_64.S

@@ -0,0 +1,65 @@
+
+.align 16
+.global zcontext_swap
+// RDI = to save sp to
+// RSI = to load sp from
+// RDX = argument
+zcontext_swap: // void* zcontext_swap(zcontext_t* from, zcontext_t* to, void* argument)
+
+// 保存非易失性寄存器
+    push %rbp
+    push %rbx
+    push %r10
+    push %r12
+    push %r13
+    push %r14
+    push %r15
+//  保存浮点上下文
+    sub $16, %rsp
+    fnstcw  0(%rsp)
+    stmxcsr 8(%rsp)
+
+//  切换栈指针
+    mov %rsp, (%rdi)
+    mov (%rsi), %rsp
+
+//  恢复浮点上下文
+    fldcw   (%rsp)
+    ldmxcsr 8(%rsp)
+    add $16, %rsp
+//  恢复非易失性寄存器
+    pop %r15
+    pop %r14
+    pop %r13
+    pop %r12
+    pop %r10
+    pop %rbx
+    pop %rbp
+
+    // 返回 argument
+    mov %rdx, %rax
+    mov (%rsp), %rcx
+    retq
+.align 16
+
+
+.align 16
+zcontext_entry_point:
+    mov %r15, %rdi
+    call *%r14
+    hlt
+
+.align 16
+.global zcontext_setup
+// RDI = target
+// RSI = func
+// RDX = argument
+zcontext_setup: // void zcontext_setup(zcontext_t* target, void (*func)(void*arg), void* argument)
+    movq    (%rdi), %rax
+    movq    %rdx, -64(%rax)
+    subq    $80, %rax
+    movq    %rsi, 24(%rax)
+    movq    zcontext_entry_point@GOTPCREL(%rip), %rcx;
+    movq    %rcx, 72(%rax)
+    movq    %rax, (%rdi)
+    ret

µasync/include/universal_fiber.hpp

@@ -26,7 +26,7 @@
 #define USE_SETJMP
 #elif defined(DISABLE_UCONTEXT)
 #define USE_SETJMP
-#else
+#elif !defined(USE_ZCONTEXT)
 #define USE_UCONTEXT
 #endif
 
@@ -46,6 +46,8 @@
 #endif
 #elif defined(USE_SETJMP)
 #include <setjmp.h>
+#elif defined(USE_ZCONTEXT)
+#include "zcontext.h"
 #endif
 
 #ifdef __linux__
@@ -72,6 +74,8 @@ typedef struct FiberOVERLAPPED
 	ucontext_t target;
 #elif defined (USE_SETJMP)
 	jmp_buf target_jmp;
+#elif defined (USE_ZCONTEXT)
+	zcontext_t target;
 #endif
 
 	DWORD byte_transfered;
@@ -80,6 +84,12 @@ typedef struct FiberOVERLAPPED
 	BOOL resultOk;
 } FiberOVERLAPPED;
 
+struct HelperStack
+{
+	alignas(64) char sp[448];
+	alignas(64) char sp_top[1];
+};
+
 struct FiberContext
 {
 	unsigned long long sp[
@@ -94,6 +104,8 @@ struct FiberContext
 	alignas(64) void* func_ptr; // NOTE： &func_ptr 相当于栈顶
 #if defined (USE_UCONTEXT)
 	ucontext_t ctx;
+#elif defined (USE_ZCONTEXT)
+	zcontext_t ctx;
 #endif
 };
 
@@ -145,15 +157,10 @@ inline thread_local LPVOID __please_delete_me = NULL;
 #elif defined (USE_UCONTEXT)
 inline thread_local ucontext_t* __current_yield_ctx = NULL;
 #elif defined (USE_SETJMP)
-
-struct HelperStack
-{
-	alignas(64) char sp[448];
-	alignas(64) char sp_top[1];
-};
-
 inline thread_local jmp_buf __current_jump_buf;
 inline thread_local FiberContext* __please_delete_me;
+#elif  defined (USE_ZCONTEXT)
+inline thread_local zcontext_t* __current_yield_zctx = NULL;
 #endif
 // <<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<
 
@@ -213,6 +220,17 @@ inline DWORD get_overlapped_result(FiberOVERLAPPED* ov)
 	WSASetLastError(ov->last_error);
 	return ov->byte_transfered;
 }
+#elif defined (USE_ZCONTEXT)
+
+inline DWORD get_overlapped_result(FiberOVERLAPPED* ov)
+{
+	assert(__current_yield_zctx && "get_overlapped_result should be called by a ucontext based coroutine!");
+
+	zcontext_swap(&ov->target, __current_yield_zctx, 0);
+
+	WSASetLastError(ov->last_error);
+	return ov->byte_transfered;
+}
 
 #endif
 
@@ -263,6 +281,17 @@ inline void process_overlapped_event(OVERLAPPED* _ov,
 		FiberContextAlloctor{}.deallocate(__please_delete_me);
 		__please_delete_me = NULL;
 	}
+#elif defined (USE_ZCONTEXT)
+	zcontext_t self;
+	zcontext_t* old = __current_yield_zctx;
+	__current_yield_zctx = &self;
+	auto __please_delete_me = zcontext_swap(&self, &ovl_res->target, 0);
+	__current_yield_zctx = old;
+	if (__please_delete_me)
+	{
+		FiberContextAlloctor{}.deallocate((FiberContext*)__please_delete_me);
+		__please_delete_me = NULL;
+	}
 #endif
 }
 
@@ -321,7 +350,7 @@ inline void exit_event_loop_when_empty(HANDLE iocp_handle)
 // different stackfull implementations
 // >>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>
 
-#ifdef USE_SETJMP
+#if defined (USE_SETJMP) || defined (USE_ZCONTEXT)
 inline void execute_on_new_stack(void* new_sp, void (*jump_target)(void*), void * param)
 {
 #if defined (__x86_64__)
@@ -396,12 +425,12 @@ static inline void WINAPI __coroutine_entry_point(LPVOID param)
 #endif // defined(USE_WINFIBER)
 
 
-#if defined (USE_UCONTEXT) || defined (USE_SETJMP) || defined (USE_FCONTEXT)
+#if defined (USE_UCONTEXT) || defined (USE_SETJMP) || defined (USE_FCONTEXT) || defined (USE_ZCONTEXT)
 
 template<typename... Args>
 #if defined (USE_FCONTEXT)
 inline void __coroutine_entry_point(boost::context::detail::transfer_t arg)
-#else //if defined (USE_UCONTEXT) || defined (USE_SETJMP)
+#else //if defined (USE_UCONTEXT) || defined (USE_SETJMP) || defined (USE_ZCONTEXT)
 static inline void __coroutine_entry_point(FiberContext* ctx)
 #endif
 {
@@ -418,7 +447,6 @@ static inline void __coroutine_entry_point(FiberContext* ctx)
 	{
 		typedef void (* real_func_type)(Args...);
 		auto  real_func_ptr = reinterpret_cast<real_func_type>(ctx->func_ptr);
-
 		std::apply(real_func_ptr, std::move(*fiber_args));
 	}
 
@@ -451,12 +479,15 @@ static inline void __coroutine_entry_point(FiberContext* ctx)
 #elif defined(USE_SETJMP)
 	__please_delete_me = ctx;
 	longjmp(__current_jump_buf, 1);
+#elif defined (USE_ZCONTEXT)
+	zcontext_t self;
+	zcontext_swap(&self, __current_yield_zctx, ctx);
 #endif
 
 }
 #endif
 
-#if defined (USE_FCONTEXT) || defined (USE_UCONTEXT) || defined (USE_SETJMP)
+#if defined (USE_FCONTEXT) || defined (USE_UCONTEXT) || defined (USE_ZCONTEXT)
 
 template<typename... Args>
 inline void create_detached_coroutine(void (*func_ptr)(Args...), Args... args)
@@ -482,39 +513,76 @@ inline void create_detached_coroutine(void (*func_ptr)(Args...), Args... args)
 	__current_yield_ctx = &self;
 
 	ucontext_t* new_ctx = & new_fiber_ctx->ctx;
+
+	auto stack_size = (new_fiber_ctx->sp_top - new_fiber_ctx->sp) * sizeof (unsigned long long);
+
 	getcontext(new_ctx);
 	new_ctx->uc_stack.ss_sp = new_fiber_ctx->sp;
 	new_ctx->uc_stack.ss_flags = 0;
-	new_ctx->uc_stack.ss_size = sizeof(new_fiber_ctx->sp);
+	new_ctx->uc_stack.ss_size = stack_size;
 	new_ctx->uc_link = __current_yield_ctx;
 
 	typedef void (*__func)(void);
 	makecontext(new_ctx, (__func)__coroutine_entry_point<Args...>, 1, new_fiber_ctx);
 
 	swapcontext(&self, new_ctx);
 	__current_yield_ctx = old;
-#	elif defined (USE_SETJMP)
+#elif defined (USE_ZCONTEXT)
+
+	// setup a new stack, and jump to __coroutine_entry_point
+	typedef void(*entry_point_type)(FiberContext*);
+
+	entry_point_type entry_func = & __coroutine_entry_point<Args...>;
+	unsigned long long * new_sp = new_fiber_ctx->sp_top;
+
+	zcontext_t self;
+	zcontext_t* old = __current_yield_zctx;
+	__current_yield_zctx = &self;
+
+	new_fiber_ctx->ctx.sp = new_sp;
+
+	zcontext_setup(&new_fiber_ctx->ctx, reinterpret_cast<void (*)(void*)>(entry_func), new_fiber_ctx);
+	auto __please_delete_me = zcontext_swap(&self, &new_fiber_ctx->ctx, 0);
+	__current_yield_zctx = old;
+
+	if (__please_delete_me)
+	{
+		FiberContextAlloctor{}.deallocate((FiberContext*) __please_delete_me);
+		__please_delete_me = NULL;
+	}
+#	endif
+}
+
+#endif  //defined (USE_SETJMP) || defined (USE_ZCONTEXT) || defined (USE_ZCONTEXT)
+
+#if defined (USE_SETJMP)
+
+template<typename... Args>
+inline void create_detached_coroutine(void (*func_ptr)(Args...), Args... args)
+{
+	using arg_tuple = std::tuple<Args...>;
+
+	++ out_standing_coroutines;
+
+	FiberContext* new_fiber_ctx = FiberContextAlloctor{}.allocate();
+	new_fiber_ctx->func_ptr = reinterpret_cast<void*>(func_ptr);
+	// placement new argument passed to function
+	new (new_fiber_ctx->sp) arg_tuple{std::forward<Args>(args)...};
+
 	// jmp_buf_fiber_entry(&fiber_param);
 	if (!setjmp(__current_jump_buf))
 	{
 		// setup a new stack, and jump to __coroutine_entry_point
 		typedef void(*entry_point_type)(FiberContext*);
 
 		entry_point_type entry_func = & __coroutine_entry_point<Args...>;
-		void * new_sp = &new_fiber_ctx->sp_top;
+		void * new_sp = new_fiber_ctx->sp_top;
 
 		execute_on_new_stack(new_sp, reinterpret_cast<void (*)(void*)>(entry_func), new_fiber_ctx);
 	}
-	if (__please_delete_me)
-	{
-		free(__please_delete_me);
-		__please_delete_me = NULL;
-	}
-
-#	endif
 }
 
-#endif
+#endif // defined (USE_SETJMP)
 
 #if defined(USE_WINFIBER)
 

µasync/include/zcontext.h

@@ -0,0 +1,11 @@
+
+#pragma once
+
+struct zcontext_t
+{
+    void* sp;// pointer to active stack buttom
+};
+
+void* zcontext_swap(zcontext_t* from, zcontext_t* to, void* argument) asm("zcontext_swap");
+
+void zcontext_setup(zcontext_t* target, void (*func)(void*arg), void* argument) asm("zcontext_setup");