[Qemu-devel] Re: [PATCH v7 4/6] char: Add qemu_chr_write_nb() for nonblocking writes

[Anthony Liguori]

On 05/05/2010 01:40 PM, Gerd Hoffmann wrote:
> > > + return s->chr_write_nb(s, buf, len);
> > > +}
> >
> > I really dislike the idea of adding another function for this.
>
> Having a explicit function for non-blocking mode IMHO is much better. 
> It makes things more clear when reading the code.
>
> Previous approach was to check for O_NONBLOCK (extra syscall) and 
> behave differently depending on whenever it is set or not (too much 
> hidden magic).
>
> > Can you
> > explain why you need this functionality for virtio-console and why this
> > functionality isn't needed for everything else?
>
> virtio-console is just the first user.  16550 emulation can use that 
> too.  With the blocking interface you'll stall the guest (sitting in 
> the blocking write syscall) in case the outgoing pipe is full.  With 
> the non-blocking interface you can simply delay the "you can send more 
> data" signal (update status register + raise IRQ) until you can 
> actually accept more data.  Meanwhile the guest can continue to run 
> and do something else.  Basically it allows to implement sane flow 
> control.

The use of char drivers is dominated by a small number of use cases 
which include serial ports, the monitor, and QMP.

In all of those cases, the devices can and should participate in control 
flow.  IMHO, we should make control flow part of the standard interface 
and fix all users appropriately.

I've seen blocking result in frozen guests many times in real 
deployments.  It's never the right thing to do as far as I can tell.

Regards,

Anthony Liguori

> cheers,
>   Gerd