Version française
Home     About     Download     Resources     Contact us    
Browse thread
memory allocation when ... calling Ocaml from C called from multiple ocaml threads
[ Home ] [ Index: by date | by threads ]
[ Search: ]

[ Message by date: previous | next ] [ Message in thread: previous | next ] [ Thread: previous | next ]
Date: -- (:)
From: Vincenzo Ciancia <vincenzo_mlRE.MOVE@y...>
Subject: memory allocation when ... calling Ocaml from C called from multiple ocaml threads
Hi all, I am fighting again in my infinite quest for a working, 
multithreaded fuse binding for ocaml. It costed me a lot of free time 
to get where I am, I already have code for all the fuse callbacks, and 
single-threaded mainloop works well. When I use the multithreaded loop 
explained below, I can't find a stable way to allocate memory for read 
and write operations.

I have the following setup:

- ocaml starts
- ocaml calls fuse initialization
- ocaml starts an infinite loop where at each step
  - C is called to retrieve a fuse command from kernel

  - A ocaml thread is created, where

    - C is called again to process the command (it's not that simple and
      I would like to avoid implementing this in ocaml), and as a part
      of this processing:

      - (1) A C function for each filesystem operation is called, where

        - OCaml is called back to implement the operation

It's not too complicated: this way I reuse all fuse code for reading 
commands and send responses to the kernel, and process every fuse 
command in a separate thread (of course, I will put a limit but it's 
not so important right now). Threads are created in ocaml, since while 
googling for enter_blocking_section and friends I found in this list 
archives that it should be made that way. It would be much simpler if 
threads could be created from C, but it just does not work. I use 
leave_blocking_section when calling back ocaml.

I can't use CAMLparam, CAMLlocal and CAMLreturn when entering the C 
callback named (1), they just segfault, I think because these functions 
are called from C. Therefore I allocate needed temporaries needed to 
call back ocaml using C declarations, like 

 value vtmp=copy_string(path);
 value vres=callback(*closure,tmp);

Unexpectedly, this seems to work fine when hammering with parallel 
filesystem requests, but every attempt to allocate a buffer for read 
and write operations fails (in the following piece of code 
string_create is a callback to String.create):

int read (const char *path, char *buf, size_t size, off_t offset)
{
  [...]
  vtmp=callback(*ocaml_string_create,Val_int(size)); 
  vres=callback3(*read_closure,vpath,vtmp,copy_int64(offset));
  [...]
  res=Int_val(Field(vres,0)); /* vres has a sum type */
  memcpy(buf,String_val(vtmp),min(size,res));
  return res;
}

Obviously, vtmp will be garbage-collected before the callback will use 
it, but also alloc_string in place of the callback to 
ocaml_string_create will just segfault, and this is the whole point. 
How can I get out of this all?

Also, it would be more efficient to just pass the already allocated C 
buffer wrapped inside an ocaml string, even if it could be dangerous to 
pass around it. Is there a way to do so?

Thanks

Vincenzo