Add ucontext coroutine implementation for generic fallback.

common.mk

@@ -917,8 +917,10 @@ strlcat.$(OBJEXT): {$(VPATH)}strlcat.c
 strlcpy.$(OBJEXT): {$(VPATH)}strlcpy.c
 strstr.$(OBJEXT): {$(VPATH)}strstr.c
 nt.$(OBJEXT): {$(VPATH)}nt.c
+
+COROUTINE_SRC = $(COROUTINE_H:.h=).$(ASMEXT)
 .coroutine_obj $(COROUTINE_OBJ): \
-	{$(VPATH)}$(COROUTINE_H:.h=).$(ASMEXT) \
+	{$(VPATH)}$(COROUTINE_SRC:/ucontext/Context.S=/ucontext/Context.c) \
 	$(COROUTINE_H:/Context.h=/.time)
 $(COROUTINE_H:/Context.h=/.time):
 	$(Q) $(MAKEDIRS) $(@D)

configure.ac

@@ -2263,7 +2263,7 @@ AS_CASE(["$rb_cv_fiber_coroutine"], [yes|''], [
             AS_CASE(["$ac_cv_sizeof_voidp"],
                 [8], [ rb_cv_fiber_coroutine=amd64 ],
                 [4], [ rb_cv_fiber_coroutine=x86 ],
-        	[*], [ rb_cv_fiber_coroutine= ]
+                [*], [ rb_cv_fiber_coroutine= ]
             )
         ],
         [*86-linux], [
@@ -2282,7 +2282,7 @@ AS_CASE(["$rb_cv_fiber_coroutine"], [yes|''], [
             rb_cv_fiber_coroutine=ppc64le
         ],
         [*], [
-            rb_cv_fiber_coroutine=
+            rb_cv_fiber_coroutine=ucontext
         ]
     )
     AC_MSG_RESULT(${rb_cv_fiber_coroutine:-no})

cont.c

@@ -187,7 +187,7 @@ struct rb_fiber_struct {
 #if FIBER_USE_NATIVE
 #if defined(FIBER_USE_COROUTINE)
 #define FIBER_ALLOCATE_STACK
-    coroutine_context context;
+    struct coroutine_context context;
     void *ss_sp;
     size_t ss_size;
 #elif defined(_WIN32)
@@ -802,7 +802,7 @@ cont_restore_thread(rb_context_t *cont)
 #if FIBER_USE_NATIVE
 #if defined(FIBER_USE_COROUTINE)
 static COROUTINE
-fiber_entry(coroutine_context * from, coroutine_context * to)
+fiber_entry(struct coroutine_context * from, struct coroutine_context * to)
 {
     rb_fiber_start();
 }
@@ -914,7 +914,7 @@ fiber_initialize_machine_stack_context(rb_fiber_t *fib, size_t size)
     ptr = fiber_machine_stack_alloc(size);
     fib->ss_sp = ptr;
     fib->ss_size = size;
-    coroutine_initialize(&fib->context, fiber_entry, ptr+size, size);
+    coroutine_initialize(&fib->context, fiber_entry, ptr, size);
     sec->machine.stack_start = (VALUE*)(ptr + STACK_DIR_UPPER(0, size));
     sec->machine.stack_maxsize = size - RB_PAGE_SIZE;
 #elif defined(_WIN32)
@@ -1530,7 +1530,7 @@ root_fiber_alloc(rb_thread_t *th)
 
 #if FIBER_USE_NATIVE
 #if defined(FIBER_USE_COROUTINE)
-    coroutine_initialize(&fib->context, NULL, NULL, 0);
+    coroutine_initialize_main(&fib->context);
 #elif defined(_WIN32)
     /* setup fib_handle for root Fiber */
     if (fib->fib_handle == 0) {

coroutine/amd64/Context.S

@@ -27,7 +27,7 @@ PREFIXED_SYMBOL(SYMBOL_PREFIX,coroutine_transfer):
 	# Restore callee stack pointer
 	movq (%rsi), %rsp
 
-	# Restore callee stack
+	# Restore callee state
 	popq %r15
 	popq %r14
 	popq %r13

coroutine/amd64/Context.h

@@ -10,35 +10,32 @@
 #include <assert.h>
 #include <string.h>
 
-#if __cplusplus
-extern "C" {
-#endif
-
 #define COROUTINE __attribute__((noreturn)) void
 
 enum {COROUTINE_REGISTERS = 6};
 
-typedef struct
+struct coroutine_context
 {
     void **stack_pointer;
-} coroutine_context;
+};
 
-typedef COROUTINE(* coroutine_start)(coroutine_context *from, coroutine_context *self);
+typedef COROUTINE(* coroutine_start)(struct coroutine_context *from, struct coroutine_context *self);
+
+static inline void coroutine_initialize_main(struct coroutine_context * context) {
+    context->stack_pointer = NULL;
+}
 
 static inline void coroutine_initialize(
-    coroutine_context *context,
+    struct coroutine_context *context,
     coroutine_start start,
-    void *stack_pointer,
+    void *stack,
-    size_t stack_size
+    size_t size
 ) {
-    /* Force 16-byte alignment */
+    assert(start && stack && size >= 1024);
-    context->stack_pointer = (void**)((uintptr_t)stack_pointer & ~0xF);
 
-    if (!start) {
+    // Stack grows down. Force 16-byte alignment.
-        assert(!context->stack_pointer);
+    char * top = (char*)stack + size;
-        /* We are main coroutine for this thread */
+    context->stack_pointer = (void**)((uintptr_t)top & ~0xF);
-        return;
-    }
 
     *--context->stack_pointer = NULL;
     *--context->stack_pointer = (void*)start;
@@ -47,13 +44,9 @@ static inline void coroutine_initialize(
     memset(context->stack_pointer, 0, sizeof(void*) * COROUTINE_REGISTERS);
 }
 
-coroutine_context * coroutine_transfer(coroutine_context * current, coroutine_context * target);
+struct coroutine_context * coroutine_transfer(struct coroutine_context * current, struct coroutine_context * target);
 
-static inline void coroutine_destroy(coroutine_context * context)
+static inline void coroutine_destroy(struct coroutine_context * context)
 {
     context->stack_pointer = NULL;
 }
-
-#if __cplusplus
-}
-#endif

coroutine/arm32/Context.h

@@ -10,34 +10,32 @@
 #include <assert.h>
 #include <string.h>
 
-#if __cplusplus
-extern "C" {
-#endif
-
 #define COROUTINE __attribute__((noreturn)) void
 
 enum {COROUTINE_REGISTERS = 8};
 
-typedef struct
+struct coroutine_context
 {
     void **stack_pointer;
-} coroutine_context;
+};
 
-typedef COROUTINE(* coroutine_start)(coroutine_context *from, coroutine_context *self);
+typedef COROUTINE(* coroutine_start)(struct coroutine_context *from, struct coroutine_context *self);
+
+static inline void coroutine_initialize_main(struct coroutine_context * context) {
+    context->stack_pointer = NULL;
+}
 
 static inline void coroutine_initialize(
-    coroutine_context *context,
+    struct coroutine_context *context,
     coroutine_start start,
-    void *stack_pointer,
+    void *stack,
-    size_t stack_size
+    size_t size
 ) {
-    context->stack_pointer = (void**)stack_pointer;
+    assert(start && stack && size >= 1024);
 
-    if (!start) {
+    // Stack grows down. Force 16-byte alignment.
-        assert(!context->stack_pointer);
+    char * top = (char*)stack + size;
-        /* We are main coroutine for this thread */
+    context->stack_pointer = (void**)((uintptr_t)top & ~0xF);
-        return;
-    }
 
     *--context->stack_pointer = (void*)start;
 
@@ -45,12 +43,8 @@ static inline void coroutine_initialize(
     memset(context->stack_pointer, 0, sizeof(void*) * COROUTINE_REGISTERS);
 }
 
-coroutine_context * coroutine_transfer(coroutine_context * current, coroutine_context * target);
+struct coroutine_context * coroutine_transfer(struct coroutine_context * current, struct coroutine_context * target);
 
-static inline void coroutine_destroy(coroutine_context * context)
+static inline void coroutine_destroy(struct coroutine_context * context)
 {
 }
-
-#if __cplusplus
-}
-#endif

coroutine/arm64/Context.h

@@ -10,35 +10,32 @@
 #include <assert.h>
 #include <string.h>
 
-#if __cplusplus
-extern "C" {
-#endif
-
 #define COROUTINE __attribute__((noreturn)) void
 
 enum {COROUTINE_REGISTERS = 0xb0 / 8};
 
-typedef struct
+struct coroutine_context
 {
     void **stack_pointer;
-} coroutine_context;
+};
 
-typedef COROUTINE(* coroutine_start)(coroutine_context *from, coroutine_context *self);
+typedef COROUTINE(* coroutine_start)(struct coroutine_context *from, struct coroutine_context *self);
+
+static inline void coroutine_initialize_main(struct coroutine_context * context) {
+    context->stack_pointer = NULL;
+}
 
 static inline void coroutine_initialize(
-    coroutine_context *context,
+    struct coroutine_context *context,
     coroutine_start start,
-    void *stack_pointer,
+    void *stack,
-    size_t stack_size
+    size_t size
 ) {
-    /* Force 16-byte alignment */
+    assert(start && stack && size >= 1024);
-    context->stack_pointer = (void**)((uintptr_t)stack_pointer & ~0xF);
 
-    if (!start) {
+    // Stack grows down. Force 16-byte alignment.
-        assert(!context->stack_pointer);
+    char * top = (char*)stack + size;
-        /* We are main coroutine for this thread */
+    context->stack_pointer = (void**)((uintptr_t)top & ~0xF);
-        return;
-    }
 
     context->stack_pointer -= COROUTINE_REGISTERS;
     memset(context->stack_pointer, 0, sizeof(void*) * COROUTINE_REGISTERS);
@@ -46,12 +43,8 @@ static inline void coroutine_initialize(
     context->stack_pointer[0xa0 / 8] = (void*)start;
 }
 
-coroutine_context * coroutine_transfer(coroutine_context * current, coroutine_context * target);
+struct coroutine_context * coroutine_transfer(struct coroutine_context * current, struct coroutine_context * target);
 
-static inline void coroutine_destroy(coroutine_context * context)
+static inline void coroutine_destroy(struct coroutine_context * context)
 {
 }
-
-#if __cplusplus
-}
-#endif

coroutine/ppc64le/Context.h

@@ -3,10 +3,6 @@
 #include <assert.h>
 #include <string.h>
 
-#if __cplusplus
-extern "C" {
-#endif
-
 #define COROUTINE __attribute__((noreturn)) void
 
 enum {
@@ -15,27 +11,28 @@ enum {
   + 4  /* space for fiber_entry() to store the link register */
 };
 
-typedef struct
+struct coroutine_context
 {
     void **stack_pointer;
-} coroutine_context;
+};
 
-typedef COROUTINE(* coroutine_start)(coroutine_context *from, coroutine_context *self);
+typedef COROUTINE(* coroutine_start)(struct coroutine_context *from, struct coroutine_context *self);
+
+static inline void coroutine_initialize_main(struct coroutine_context * context) {
+    context->stack_pointer = NULL;
+}
 
 static inline void coroutine_initialize(
-    coroutine_context *context,
+    struct coroutine_context *context,
     coroutine_start start,
-    void *stack_pointer,
+    void *stack,
-    size_t stack_size
+    size_t size
 ) {
-    /* Force 16-byte alignment */
+    assert(start && stack && size >= 1024);
-    context->stack_pointer = (void**)((uintptr_t)stack_pointer & ~0xF);
 
-    if (!start) {
+    // Stack grows down. Force 16-byte alignment.
-        assert(!context->stack_pointer);
+    char * top = (char*)stack + size;
-        /* We are main coroutine for this thread */
+    context->stack_pointer = (void**)((uintptr_t)top & ~0xF);
-        return;
-    }
 
     context->stack_pointer -= COROUTINE_REGISTERS;
     memset(context->stack_pointer, 0, sizeof(void*) * COROUTINE_REGISTERS);
@@ -44,13 +41,9 @@ static inline void coroutine_initialize(
     context->stack_pointer[18] = ((char*)start) + 8;
 }
 
-coroutine_context * coroutine_transfer(coroutine_context * current, coroutine_context * target);
+struct coroutine_context * coroutine_transfer(struct coroutine_context * current, struct coroutine_context * target);
 
-static inline void coroutine_destroy(coroutine_context * context)
+static inline void coroutine_destroy(struct coroutine_context * context)
 {
     context->stack_pointer = NULL;
 }
-
-#if __cplusplus
-}
-#endif

coroutine/ucontext/Context.c

@@ -0,0 +1,16 @@
+/*
+ *  This file is part of the "Coroutine" project and released under the MIT License.
+ *
+ *  Created by Samuel Williams on 24/6/2019.
+ *  Copyright, 2019, by Samuel Williams. All rights reserved.
+*/
+
+#include "Context.h"
+
+void coroutine_trampoline(void * _start, void * _context)
+{
+    coroutine_start start = _start;
+    struct coroutine_context * context = _context;
+
+    start(context->from, context);
+}

coroutine/ucontext/Context.h

@@ -0,0 +1,66 @@
+/*
+ *  This file is part of the "Coroutine" project and released under the MIT License.
+ *
+ *  Created by Samuel Williams on 24/6/2019.
+ *  Copyright, 2019, by Samuel Williams. All rights reserved.
+*/
+
+#pragma once
+
+#include <assert.h>
+#include <stddef.h>
+#include <ucontext.h>
+
+#define COROUTINE __attribute__((noreturn)) void
+
+struct coroutine_context
+{
+    ucontext_t state;
+    struct coroutine_context * from;
+};
+
+typedef COROUTINE(* coroutine_start)(struct coroutine_context *from, struct coroutine_context *self);
+
+COROUTINE coroutine_trampoline(void * _start, void * _context);
+
+static inline void coroutine_initialize_main(struct coroutine_context * context) {
+    context->from = NULL;
+    getcontext(&context->state);
+}
+
+static inline void coroutine_initialize(
+    struct coroutine_context *context,
+    coroutine_start start,
+    void *stack,
+    size_t size
+) {
+    assert(start && stack && size >= 1024);
+
+    coroutine_initialize_main(context);
+
+    context->state.uc_stack.ss_size = size;
+    // Despite what it's called, this is not actually a stack pointer. It points to the address of the stack allocation (the lowest address).
+    context->state.uc_stack.ss_sp = (char*)stack;
+    context->state.uc_stack.ss_flags = 0;
+    context->state.uc_link = NULL;
+
+    makecontext(&context->state, (void(*)(void))coroutine_trampoline, 2, (void*)start, (void*)context);
+}
+
+static inline struct coroutine_context * coroutine_transfer(struct coroutine_context * current, struct coroutine_context * target)
+{
+    struct coroutine_context * previous = target->from;
+
+    target->from = current;
+    swapcontext(&current->state, &target->state);
+    target->from = previous;
+
+    return target;
+}
+
+static inline void coroutine_destroy(struct coroutine_context * context)
+{
+    context->state.uc_stack.ss_sp = NULL;
+    context->state.uc_stack.ss_size = 0;
+    context->from = NULL;
+}

coroutine/win32/Context.h

@@ -10,53 +10,47 @@
 #include <assert.h>
 #include <string.h>
 
-#if __cplusplus
-extern "C" {
-#endif
-
 #define COROUTINE __declspec(noreturn) void __fastcall
 
 /* This doesn't include thread information block */
 enum {COROUTINE_REGISTERS = 4};
 
-typedef struct
+struct coroutine_context
 {
     void **stack_pointer;
-} coroutine_context;
+};
 
-typedef void(__fastcall * coroutine_start)(coroutine_context *from, coroutine_context *self);
+typedef void(__fastcall * coroutine_start)(struct coroutine_context *from, struct coroutine_context *self);
+
+static inline void coroutine_initialize_main(struct coroutine_context * context) {
+    context->stack_pointer = NULL;
+}
 
 static inline void coroutine_initialize(
-    coroutine_context *context,
+    struct coroutine_context *context,
     coroutine_start start,
-    void *stack_pointer,
+    void *stack,
-    size_t stack_size
+    size_t size
 ) {
-    context->stack_pointer = (void**)stack_pointer;
+    assert(start && stack && size >= 1024);
 
-    if (!start) {
+    // Stack grows down. Force 16-byte alignment.
-        assert(!context->stack_pointer);
+    char * top = (char*)stack + size;
-        /* We are main coroutine for this thread */
+    context->stack_pointer = (void**)((uintptr_t)top & ~0xF);
-        return;
-    }
 
     *--context->stack_pointer = (void*)start;
 
     /* Windows Thread Information Block */
     *--context->stack_pointer = 0; /* fs:[0] */
-    *--context->stack_pointer = (void*)stack_pointer; /* fs:[4] */
+    *--context->stack_pointer = (void*)top; /* fs:[4] */
-    *--context->stack_pointer = (void*)((char *)stack_pointer - stack_size);  /* fs:[8] */
+    *--context->stack_pointer = (void*)stack;  /* fs:[8] */
 
     context->stack_pointer -= COROUTINE_REGISTERS;
     memset(context->stack_pointer, 0, sizeof(void*) * COROUTINE_REGISTERS);
 }
 
-coroutine_context * __fastcall coroutine_transfer(coroutine_context * current, coroutine_context * target);
+struct coroutine_context * __fastcall coroutine_transfer(struct coroutine_context * current, struct coroutine_context * target);
 
-static inline void coroutine_destroy(coroutine_context * context)
+static inline void coroutine_destroy(struct coroutine_context * context)
 {
 }
-
-#if __cplusplus
-}
-#endif

coroutine/win64/Context.h

@@ -10,10 +10,6 @@
 #include <assert.h>
 #include <string.h>
 
-#if __cplusplus
-extern "C" {
-#endif
-
 #define COROUTINE __declspec(noreturn) void
 
 enum {
@@ -21,29 +17,30 @@ enum {
     COROUTINE_XMM_REGISTERS = 1+10*2,
 };
 
-typedef struct
+struct coroutine_context
 {
     void **stack_pointer;
-} coroutine_context;
+};
 
-typedef void(* coroutine_start)(coroutine_context *from, coroutine_context *self);
+typedef void(* coroutine_start)(struct coroutine_context *from, struct coroutine_context *self);
 
 void coroutine_trampoline();
 
-static inline void coroutine_initialize(
+static inline void coroutine_initialize_main(struct coroutine_context * context) {
-    coroutine_context *context,
+    context->stack_pointer = NULL;
-    coroutine_start start,
+}
-    void *stack_pointer,
-    size_t stack_size
-) {
-    /* Force 16-byte alignment */
-    context->stack_pointer = (void**)((uintptr_t)stack_pointer & ~0xF);
 
-    if (!start) {
+static inline void coroutine_initialize(
-        assert(!context->stack_pointer);
+    struct coroutine_context *context,
-        /* We are main coroutine for this thread */
+    coroutine_start start,
-        return;
+    void *stack,
-    }
+    size_t size
+) {
+    assert(start && stack && size >= 1024);
+
+    // Stack grows down. Force 16-byte alignment.
+    char * top = (char*)stack + size;
+    context->stack_pointer = (void**)((uintptr_t)top & ~0xF);
 
     /* Win64 ABI requires space for arguments */
     context->stack_pointer -= 4;
@@ -55,20 +52,16 @@ static inline void coroutine_initialize(
 
     /* Windows Thread Information Block */
     /* *--context->stack_pointer = 0; */ /* gs:[0x00] is not used */
-    *--context->stack_pointer = (void*)stack_pointer; /* gs:[0x08] */
+    *--context->stack_pointer = (void*)top; /* gs:[0x08] */
-    *--context->stack_pointer = (void*)((char *)stack_pointer - stack_size);  /* gs:[0x10] */
+    *--context->stack_pointer = (void*)stack;  /* gs:[0x10] */
 
     context->stack_pointer -= COROUTINE_REGISTERS;
     memset(context->stack_pointer, 0, sizeof(void*) * COROUTINE_REGISTERS);
     memset(context->stack_pointer - COROUTINE_XMM_REGISTERS, 0, sizeof(void*) * COROUTINE_XMM_REGISTERS);
 }
 
-coroutine_context * coroutine_transfer(coroutine_context * current, coroutine_context * target);
+struct coroutine_context * coroutine_transfer(struct coroutine_context * current, struct coroutine_context * target);
 
-static inline void coroutine_destroy(coroutine_context * context)
+static inline void coroutine_destroy(struct coroutine_context * context)
 {
 }
-
-#if __cplusplus
-}
-#endif

coroutine/x86/Context.h

@@ -10,35 +10,32 @@
 #include <assert.h>
 #include <string.h>
 
-#if __cplusplus
-extern "C" {
-#endif
-
 #define COROUTINE __attribute__((noreturn, fastcall)) void
 
 enum {COROUTINE_REGISTERS = 4};
 
-typedef struct
+struct coroutine_context
 {
     void **stack_pointer;
-} coroutine_context;
+};
 
-typedef COROUTINE(* coroutine_start)(coroutine_context *from, coroutine_context *self) __attribute__((fastcall));
+typedef COROUTINE(* coroutine_start)(struct coroutine_context *from, struct coroutine_context *self) __attribute__((fastcall));
+
+static inline void coroutine_initialize_main(struct coroutine_context * context) {
+    context->stack_pointer = NULL;
+}
 
 static inline void coroutine_initialize(
-    coroutine_context *context,
+    struct coroutine_context *context,
     coroutine_start start,
-    void *stack_pointer,
+    void *stack,
-    size_t stack_size
+    size_t size
 ) {
-    /* Force 16-byte alignment */
+    assert(start && stack && size >= 1024);
-    context->stack_pointer = (void**)((uintptr_t)stack_pointer & ~0xF);
 
-    if (!start) {
+    // Stack grows down. Force 16-byte alignment.
-        assert(!context->stack_pointer);
+    char * top = (char*)stack + size;
-        /* We are main coroutine for this thread */
+    context->stack_pointer = (void**)((uintptr_t)top & ~0xF);
-        return;
-    }
 
     *--context->stack_pointer = NULL;
     *--context->stack_pointer = (void*)start;
@@ -47,13 +44,9 @@ static inline void coroutine_initialize(
     memset(context->stack_pointer, 0, sizeof(void*) * COROUTINE_REGISTERS);
 }
 
-coroutine_context * coroutine_transfer(coroutine_context * current, coroutine_context * target) __attribute__((fastcall));
+struct coroutine_context * coroutine_transfer(struct coroutine_context * current, struct coroutine_context * target) __attribute__((fastcall));
 
-static inline void coroutine_destroy(coroutine_context * context)
+static inline void coroutine_destroy(struct coroutine_context * context)
 {
     context->stack_pointer = NULL;
 }
-
-#if __cplusplus
-}
-#endif