Re: [Xen-devel] [Minios-devel] [PATCH MINI-OS v2] xenbus: notify the other end when necessary

[Wei Liu <wei.liu2@xxxxxxxxxx>]

On Mon, Oct 26, 2015 at 06:11:04PM +0100, Samuel Thibault wrote:
> Wei Liu, le Mon 26 Oct 2015 16:41:15 +0000, a écrit :
> > The oxenstored in staging.
> 
> Ok.  For a first, one fishy thing at quick sight is that the only
> occurence of rsp_cons (except at closure) is when writing, and not when
> going to sleep.  One issue there is that ml_interface_write only writes
> a contiguous piece of data, so when crossing the ring bound, it will
> return a short write while there *is* room!
> 
> One quick test you could do is calling Xs_ring.write a second time in
> tools/ocaml/libs/xb/xb.ml's write_mmap, something like below (untested,
> and my caml is old, so it's ugly, but you get the idea).
> 
> Samuel
> 
> diff --git a/tools/ocaml/libs/xb/xb.ml b/tools/ocaml/libs/xb/xb.ml
> index 50944b5..33298d2 100644
> --- a/tools/ocaml/libs/xb/xb.ml
> +++ b/tools/ocaml/libs/xb/xb.ml
> @@ -91,10 +91,12 @@ let write_fd back con s len =
>       Unix.write back.fd s 0 len
>  
>  let write_mmap back con s len =
> -     let ws = Xs_ring.write back.mmap s len in
> -     if ws > 0 then
> +     let ws = ref (Xs_ring.write back.mmap s len) in
> +     if !ws < len then
> +             ws := !ws + Xs_ring.write back.mmap (String.sub s !ws 
> (len-!ws)) (len-!ws);
> +     if !ws > 0 then
>               back.eventchn_notify ();
> -     ws
> +     !ws
>  
>  let write con s len =
>       match con.backend with

I'm now running a slightly modified version of this diff (to fix
compilation) and a mini-os only notifies the other end when it consumes
message.

The result is looking good, but I will let it run overnight to be sure.

If the backend is to blame, I think we need to accept the fact it is now
the de facto behaviour of xenstore ring and work around it in the
frontend...

Wei.

_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxx
http://lists.xen.org/xen-devel