Re: Interesting qemu/virt-manager bug about the "rotational" attribute on virtio-blk disks

From: "Richard W.M. Jones" <rjones@redhat.com>
To: Stefan Hajnoczi <stefanha@redhat.com>
Cc: qemu-discuss@nongnu.org, "Michael S. Tsirkin" <mst@redhat.com>,
	virtualization@lists.linux-foundation.org
Subject: Re: Interesting qemu/virt-manager bug about the "rotational" attribute on virtio-blk disks
Date: Thu, 23 Jul 2020 11:40:56 +0100	[thread overview]
Message-ID: <20200723104056.GN3888@redhat.com> (raw)
In-Reply-To: <20200723103239.GD186372@stefanha-x1.localdomain>

On Thu, Jul 23, 2020 at 11:32:39AM +0100, Stefan Hajnoczi wrote:
> On Thu, Jul 16, 2020 at 11:33:44AM +0200, Stefano Garzarella wrote:
> > +Cc Michael, Stefan, virtualization@lists.linux-foundation.org
> > 
> > On Thu, Jul 16, 2020 at 09:06:14AM +0100, Richard W.M. Jones wrote:
> > > 
> > > https://bugzilla.redhat.com/show_bug.cgi?id=1857515
> > > 
> > > A virtio-blk disk which is backed by a raw file on an SSD,
> > > inside the guest shows rotational = 1.
> > > 
> > > I assumed that qemu must have a "rotational" property for disks and
> > > this would be communicated by virtio to the guest, but qemu and virtio
> > > don't seem to have this.  Pretty surprising!  Is it called something
> > > other than "rotational"?
> > > 
> > 
> > I'm not sure if we need to add this property in QEMU, but in Linux
> > I found these flags (include/linux/blkdev.h) for the block queues:
> > 
> >     #define QUEUE_FLAG_NONROT	6	/* non-rotational device (SSD) */
> >     #define QUEUE_FLAG_VIRT		QUEUE_FLAG_NONROT /* paravirt device */
> > 
> > xen-blkfront driver is the only one that sets the QUEUE_FLAG_VIRT,
> > should we do the same in the virtio-blk driver regardless of the backend?
> 
> The ability to control this flag would be interesting for performance
> experiments.

I expect there's a pretty strong interaction between this feature and
the shared guest io_uring stuff isn't there?  I'm not sure if it'll be
a positive or negative effect though.

Rich.

> The problem with changing the default is that regressions can be
> expected. Certain workloads benefit while others regress.
> 
> I suggest:
> 1. Make it controllable so that QUEUE_FLAG_NONROT can be set or clear
>    (not hardcoded to a single value).
> 2. The device can communicate the optimal setting from the host. The
>    SCSI protocol already conveys this information. Virtio-blk needs a
>    feature bit and possibly config space field.
> 3. Make it migration-safe. It needs to be configured explicitly so the
>    value doesn't change suddenly across migration.
> 
> Stefan

-- 
Richard Jones, Virtualization Group, Red Hat http://people.redhat.com/~rjones
Read my programming and virtualization blog: http://rwmj.wordpress.com
Fedora Windows cross-compiler. Compile Windows programs, test, and
build Windows installers. Over 100 libraries supported.
http://fedoraproject.org/wiki/MinGW