Linux Virtualization - Feb 2019 - [PATCH RFC] virtio: hint if callbacks surprisingly might sleep

On Thu, 31 Jan 2019 10:27:53 -0500
"Michael S. Tsirkin" <mst at redhat.com> wrote:
> On Thu, Jan 31, 2019 at 01:53:14PM +0100, Cornelia Huck wrote:
> > A virtio transport is free to implement some of the callbacks in
> > virtio_config_ops in a matter that they cannot be called from
> > atomic context (e.g. virtio-ccw, which maps a lot of the callbacks
> > to channel I/O, which is an inherently asynchronous mechanism).
> > This can be very surprising for developers using the much more
> > common virtio-pci transport, just to find out that things break
> > when used on s390.
> > 
> > The documentation for virtio_config_ops now contains a comment
> > explaining this, but it makes sense to add a might_sleep() annotation
> > to various wrapper functions in the virtio core to avoid surprises
> > later.
> > 
> > Note that annotations are NOT added to two classes of calls:
> > - direct calls from device drivers (all current callers should be
> >   fine, however)
> > - calls which clearly won't be made from atomic context (such as
> >   those ultimately coming in via the driver core)
> > 
> > Signed-off-by: Cornelia Huck <cohuck at redhat.com>  
> 
> 
> Makes sense to me. I don't think we should push our luck in
> this release though, better defer until the merge window.
Friendly ping, as we're quite close to the release of 5.0 now.

On Wed, Feb 13, 2019 at 02:44:14PM +0100, Cornelia Huck
wrote:
> On Thu, 31 Jan 2019 10:27:53 -0500
> "Michael S. Tsirkin" <mst at redhat.com> wrote:
> 
> > On Thu, Jan 31, 2019 at 01:53:14PM +0100, Cornelia Huck wrote:
> > > A virtio transport is free to implement some of the callbacks in
> > > virtio_config_ops in a matter that they cannot be called from
> > > atomic context (e.g. virtio-ccw, which maps a lot of the
callbacks
> > > to channel I/O, which is an inherently asynchronous mechanism).
> > > This can be very surprising for developers using the much more
> > > common virtio-pci transport, just to find out that things break
> > > when used on s390.
> > > 
> > > The documentation for virtio_config_ops now contains a comment
> > > explaining this, but it makes sense to add a might_sleep()
annotation
> > > to various wrapper functions in the virtio core to avoid
surprises
> > > later.
> > > 
> > > Note that annotations are NOT added to two classes of calls:
> > > - direct calls from device drivers (all current callers should be
> > >   fine, however)
> > > - calls which clearly won't be made from atomic context (such
as
> > >   those ultimately coming in via the driver core)
> > > 
> > > Signed-off-by: Cornelia Huck <cohuck at redhat.com>  
> > 
> > 
> > Makes sense to me. I don't think we should push our luck in
> > this release though, better defer until the merge window.
> 
> Friendly ping, as we're quite close to the release of 5.0 now.
Queued now, thanks!