passing function pointer to the C code using cgo

Iakov Davydov picture Iakov Davydov · May 11, 2016 · Viewed 9.2k times · Source

Starting from Go v1.6 cgo changed the rules of passing pointers to the C code golang/go#12416. The example of invoking a dynamic Go callback from C code from the wiki doesn't work anymore.

package main

import (
    "fmt"
    "unsafe"
)

/*
   extern void go_callback_int(void* foo, int p1);

   // normally you will have to define function or variables
   // in another separate C file to avoid the multiple definition
   // errors, however, using "static inline" is a nice workaround
   // for simple functions like this one.
   static inline void CallMyFunction(void* pfoo) {
       go_callback_int(pfoo, 5);
       }
*/
import "C"

//export go_callback_int
func go_callback_int(pfoo unsafe.Pointer, p1 C.int) {
    foo := *(*func(C.int))(pfoo)
    foo(p1)
}

func MyCallback(x C.int) {
    fmt.Println("callback with", x)
}

// we store it in a global variable so that the garbage collector
// doesn't clean up the memory for any temporary variables created.
var MyCallbackFunc = MyCallback

func Example() {
    C.CallMyFunction(unsafe.Pointer(&MyCallbackFunc))
}

func main() {
    Example()
}

The output looks like this:

panic: runtime error: cgo argument has Go pointer to Go pointer

What is the proper way to do this today? Preferably without hacks like hiding pointer from the language by converting it into uintptr_t.

Answer

Iakov Davydov picture Iakov Davydov · Jun 5, 2016

Starting from Go 1.6 cgo has new rules.

Go code may pass a Go pointer to C provided that the Go memory to which it points does not contain any Go pointers.

[source]

These rules are checked during the runtime, and if violated program crashes. At the moment it is possible to disable checks using GODEBUG=cgocheck=0 environment variable. But in the future, that might stop working.

So it is not possible anymore to pass a pointer to C code, if the memory to which it is pointing stores a Go function/method pointer. There are several ways to overcome this limitations, but I guess in most of them you should store a synchronized data structure which represents the correspondence between a certain id and the actual pointer. This way you can pass an id to the C code, not a pointer.

The code solving this problem might look like this:

package gocallback

import (
    "fmt"
    "sync"
)

/*
extern void go_callback_int(int foo, int p1);

// normally you will have to define function or variables
// in another separate C file to avoid the multiple definition
// errors, however, using "static inline" is a nice workaround
// for simple functions like this one.
static inline void CallMyFunction(int foo) {
    go_callback_int(foo, 5);
}
*/
import "C"

//export go_callback_int
func go_callback_int(foo C.int, p1 C.int) {
    fn := lookup(int(foo))
    fn(p1)
}

func MyCallback(x C.int) {
    fmt.Println("callback with", x)
}

func Example() {
    i := register(MyCallback)
    C.CallMyFunction(C.int(i))
    unregister(i)
}

var mu sync.Mutex
var index int
var fns = make(map[int]func(C.int))

func register(fn func(C.int)) int {
    mu.Lock()
    defer mu.Unlock()
    index++
    for fns[index] != nil {
        index++
    }
    fns[index] = fn
    return index
}

func lookup(i int) func(C.int) {
    mu.Lock()
    defer mu.Unlock()
    return fns[i]
}

func unregister(i int) {
    mu.Lock()
    defer mu.Unlock()
    delete(fns, i)
}

This code comes from the (updated) wiki page.