Re: [Qemu-devel] [PATCH] Asynchronous reliable and configurable cache flush

[Jamie Lokier <jamie@shareable.org>]

Paul Brook wrote:
> > @@ -1021,6 +1031,13 @@ static void ide_write_dma_cb(void *opaque, int ret)
> >  
> >      /* end of transfer ? */
> >      if (s->nsector == 0) {
> > +       if (!s->write_cache) {
> > +           ret = bdrv_flush(s->bs);
> > +           if (ret != 0) {
> > +               ide_dma_error(s);
> > +               return;
> > +           }
> > +       }
> 
> By my reading this is adding a synchronous flush to the end of an
> async write operation, which in practice makes the whole operation
> synchronous.

Looks that way to me too.  It might be simplest to open the device
with O_DSYNC when !s->write_cache and user actually wants fdatasync,
so that async write can be used.  I suspect every platform with useful
Posix AIO has O_DSYNC.

Otherwise, chaining where the completion of aio_write triggers
aio_fsync instead of reporting completion to the guest?

-- Jamie