Re: [PATCH V3 3/3] virtio-blk: Add bio-based IO path for virtio-blk

[Asias He]

On 07/29/2012 08:59 PM, Michael S. Tsirkin wrote:
> On Sat, Jul 28, 2012 at 10:38:41AM +0800, Asias He wrote:
> > On 07/27/2012 08:33 AM, Rusty Russell wrote:
> > > On Fri, 13 Jul 2012 16:38:51 +0800, Asias He <asias@xxxxxxxxxx> wrote:
> > > > Add 'virtio_blk.use_bio=1' to kernel cmdline or 'modprobe virtio_blk
> > > > use_bio=1' to enable ->make_request_fn() based I/O path.
> > >
> > > This patch conflicts with Paolo's Bonzini's 'virtio-blk: allow toggling
> > > host cache between writeback and writethrough' which is also queued (see
> > > linux-next).
> >
> > Rebased against Paolo's patch in V4.
> >
> > > I'm not sure what the correct behavior for bio & cacheflush is, if any.
> >
> > REQ_FLUSH is not supported in the bio path.
> >
> > > But as to the patch itself: it's a hack.
> > >
> > > 1) Leaving the guest's admin to turn on the switch is a terrible choice.
> > > 2) The block layer should stop merging and sorting when a device is
> > >     fast, not the driver.
> > > 3) I pointed out that slow disks have low IOPS, so why is this
> > >     conditional?  Sure, more guest exits, but it's still a small number
> > >     for a slow device.
> > > 4) The only case where we want merging is on a slow device when the host
> > >     isn't doing it.
> > >
> > > Now, despite this, I'm prepared to commit it.  But in my mind it's a
> > > hack: we should aim for use_bio to be based on a feature bit fed from
> > > the host, and use the module parameter only if we want to override it.
> >
> > OK. A feature bit from host sound like a choice but a switch is also
> > needed on host side.
>
> qemu automatically gives you the ability to control
> any feature bit.

Automatically?

> > And for other OS, e.g. Windows, the bio thing
> > does not apply at all.
>
> Let's try to define when it's a good idea. Is it a hint to guest that
> backend handles small accesses efficiently so ok to disable batching?

Yes. It's also a hint for latency reduction.

> > Anyway, I have to admit that adding a module parameter here is not
> > the best choice. Let's think more.
> >
> > --
> > Asias


--
Asias
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/