Re: [PATCH 06/12] vfio/ap_ops: Convert to use vfio_register_group_dev()

From: Jason Gunthorpe <jgg@nvidia.com>
To: Tony Krowiak <akrowiak@linux.ibm.com>
Cc: Cornelia Huck <cohuck@redhat.com>,
	Christian Borntraeger <borntraeger@de.ibm.com>,
	Harald Freudenberger <freude@linux.ibm.com>,
	Vasily Gorbik <gor@linux.ibm.com>,
	Heiko Carstens <hca@linux.ibm.com>,
	linux-s390@vger.kernel.org, Halil Pasic <pasic@linux.ibm.com>,
	Pierre Morel <pmorel@linux.ibm.com>,
	"Raj, Ashok" <ashok.raj@intel.com>,
	Dan Williams <dan.j.williams@intel.com>,
	Daniel Vetter <daniel@ffwll.ch>, Christoph Hellwig <hch@lst.de>,
	Leon Romanovsky <leonro@nvidia.com>,
	Max Gurtovoy <mgurtovoy@nvidia.com>,
	Tarun Gupta <targupta@nvidia.com>,
	"Jason J . Herne" <jjherne@linux.ibm.com>
Subject: Re: [PATCH 06/12] vfio/ap_ops: Convert to use vfio_register_group_dev()
Date: Tue, 4 May 2021 13:04:32 -0300	[thread overview]
Message-ID: <20210504160432.GB94750@nvidia.com> (raw)
In-Reply-To: <01ada05b-ecfb-491f-9a34-6d6de54262a0@linux.ibm.com>

On Tue, May 04, 2021 at 09:58:45AM -0400, Tony Krowiak wrote:
> 
> 
> On 5/3/21 4:33 PM, Jason Gunthorpe wrote:
> > On Mon, May 03, 2021 at 04:14:43PM -0400, Tony Krowiak wrote:
> > 
> > > This case will occur whenever a user removes the mdev
> > > by echoing a '1' into the mdev's sysfs 'remove' attribute
> > > file. I'm not sure it can be considered graceful to take away
> > > all of the crypto devices from a guest while it is running,
> > > but there is a way to process the remove callback without
> > > leaving things in a "weird, half-dead state".
> > It is acceptable to just sleep here until whatever user controlled
> > condition is resolved.
> > 
> > Jason
> 
> I suppose we could do that, but the user that tried to remove
> the mdev via its sysfs 'remove' attribute will be left sitting
> there wondering why the operation didn't complete. That
> could result in leaving the user hanging in perpetuity.

Yes.

If the driver can't implement a disconnection then that is
unavoidable. What it does today by leaking memory under user control
is not acceptable.

> IMHO, the callback should continue to return an int and
> the caller should display the error if a non-zero rc is
> returned.

Nope, there is a reason removal is not allowed to fail.. sysfs remove
isn't the only reason the mdev driver could be destroyed, the
underlying physical device could be unplugged or other things.

Drivers need to implement a proper remove.

Jason