Re: RE: [PATCH v1 0/2] Virtio: fix some vq allocation issues

[Halil Pasic <pasic@linux.ibm.com>]

On Wed, 2 Jan 2019 14:23:38 +0100
Cornelia Huck <cohuck@redhat.com> wrote:

> On Wed, 2 Jan 2019 10:53:14 +0100
> Cornelia Huck <cohuck@redhat.com> wrote:
> 
> > On Tue, 1 Jan 2019 00:40:19 +0100
> > Halil Pasic <pasic@linux.ibm.com> wrote:
> 
> > > As I said, at the moment I don't have a preference regarding the fix,
> > > partly because I'm not sure if "reading config inside the handler" is OK
> > > or not. Maybe Connie or Michael can help us here. I'm however sure that
> > > commit 86a5597 "virtio-balloon: VIRTIO_BALLOON_F_FREE_PAGE_HINT"
> > > breaks virtio-balloon with the ccw transport (i.e. effectively breaks 
> > > virtio-balloon on s390): it used to work before and does not work
> > > after.  
> > 
> > Yes, that's unfortunate.
> > 
> > > 
> > > AFAICT tweaking the balloon code may be simpler than tweaking the
> > > virtio-ccw (transport code). ccw_io_helper() relies on getting
> > > an interrupt when the issued IO is done. If virtio-ccw is buggy, it
> > > needs to be fixed, but I'm not sure it is.  
> > 
> > I would not call virtio-ccw buggy, but it has some constraints that
> > virtio-pci apparently doesn't have (and which did not show up so far;
> > e.g. virtio-blk schedules a work item on config change, so there's no
> > deadlock there.)
> > 
> > One way to get out of that constraint (don't interact with the config
> > space directly in the config changed handler) would be to schedule a
> > work item in virtio-ccw that calls virtio_config_changed() for the
> > device. My understanding is that delaying the notification to a work
> > queue would be fine.
> 
> Unfortunately, calling virtio_config_changed() from a work item is not
> enough: That function takes the config_lock, and the virtio-ccw code to
> get the config both needs to allocate some memory and call schedule :/
> 
> The best option really seems to be
> - have virtio-balloon move the handling of the config change onto a
>   workqueue or something like that, and
> - document that you cannot read/write the virtio config space from an
>   atomic context
> 
> Unless someone has a better idea?
> 

I wonder, would making config_lock a mutex suffice?

I've already hinted that a virtio-balloon side fix is probably the
simpler variant. 

I agree, let's fix the regression first, and think about wether to teach
virtio-ccw to allow config manipulation from virtio_config_changed() or
not later.

Regards,
Halil