09:04 <mrvn> anyone awake?

12:40 <tsuyoshi> mrvn: you only need to register the values you're protecting

12:41 <tsuyoshi> like if you store a function in some c struct, you only need to register the function really, and not the c struct

12:45 <mrvn> Yes.

12:46 <mrvn> Should I regider/deregister the closures every time I get one or isn't it better to keep a registered array of values and store the closures there whenever I have one?

12:46 <mrvn> i.e. modify the array

12:47 <mrvn> Actually forget that.

12:48 <mrvn> The closure is copied into a struct iocb and passed to the kernel. Later then the kernel returns it. SO I have no way to tell the Gc where the closure will be in memory. So I do have to keep a copy in a fixed location and register that.

12:56 <mrvn> Something like this: CAMLprim void read(value fn) { CAMLparam1(fn); value *p = malloc(sizeof(value)); *p = fn; caml_register_global_root(p); struct iocb io; io.data = p; do_io(&io); CAMLreturn0; } CAMLprim void callback(struct iocb *io) { caml_callback(*(io->data), Val_unit); caml_remove_global_root(io->data); }

12:56 <mrvn> +free(io->data);

13:09 <tsuyoshi> the kernel returns it?

13:10 <mrvn> The libaio returns a struct io_event { void *data; ... }

13:11 <mrvn> I have no control over where the "void *data" gets stored or copied around between when I give it to libaio (and thus the kernel) and when it comes back.

13:14 <mrvn> The io_event also contains the struct iocb *io. I guess I could just use that. Not sure why the void *data is copyied explicity into the io_event. Seems wastefull.

13:15 <mrvn> Then I would use caml_register_global_root(&io->data);, right?

13:51 <mrvn> it segfaults. :(

13:58 <mrvn> doh, I forgot to CAMLreturn(ml_str). I had a dummy CAMLreturn(Val_int(0)) there.

13:59 <mrvn> Now I need one more custom block for the aio context and it should be working.

14:00 <mrvn> Only problem I've got left is that at the end of the program the GC doesn't run so *_finalized() may never be called. How do I makes sure that it does?

14:04 <tsuyoshi> atexit() will work unless the process aborts

14:10 <mrvn> 'at_exit Gc.full_major' it is.

14:11 <mrvn> Or should I keep track of all custom blocks allocated and manually free them?

14:44 <mrvn> Are custom blocks moved during compaction?

19:11 <gildor> mrvn: just see your bug on debian-ocaml

19:12 <gildor> mrvn: I am really not sure this is bug

19:13 <gildor> mrvn: in OCaml case, finalizer is about memory, not tracking panding aio

19:13 <gildor> just as close_out