Re: [PATCH] virtio_console: allocate inbufs in add_port() only if it is needed

[Michael S. Tsirkin]

On Wed, Nov 06, 2019 at 03:02:25PM +0100, Laurent Vivier wrote:
> On 06/11/2019 14:56, Michael S. Tsirkin wrote:
> > On Fri, Oct 18, 2019 at 06:47:18PM +0200, Laurent Vivier wrote:
> >> When we hot unplug a virtserialport and then try to hot plug again,
> >> it fails:
> >>
> >> (qemu) chardev-add socket,id=serial0,path=/tmp/serial0,server,nowait
> >> (qemu) device_add virtserialport,bus=virtio-serial0.0,nr=2,\
> >>                   chardev=serial0,id=serial0,name=serial0
> >> (qemu) device_del serial0
> >> (qemu) device_add virtserialport,bus=virtio-serial0.0,nr=2,\
> >>                   chardev=serial0,id=serial0,name=serial0
> >> kernel error:
> >>   virtio-ports vport2p2: Error allocating inbufs
> >> qemu error:
> >>   virtio-serial-bus: Guest failure in adding port 2 for device \
> >>                      virtio-serial0.0
> >>
> >> This happens because buffers for the in_vq are allocated when the port is
> >> added but are not released when the port is unplugged.
> >>
> >> They are only released when virtconsole is removed (see a7a69ec0d8e4)
> >>
> >> To avoid the problem and to be symmetric, we could allocate all the buffers
> >> in init_vqs() as they are released in remove_vqs(), but it sounds like
> >> a waste of memory.
> >>
> >> Rather than that, this patch changes add_port() logic to only allocate the
> >> buffers if the in_vq has available free slots.
> >>
> >> Fixes: a7a69ec0d8e4 ("virtio_console: free buffers after reset")
> >> Cc: mst@xxxxxxxxxx
> >> Signed-off-by: Laurent Vivier <lvivier@xxxxxxxxxx>
> >> ---
> >>  drivers/char/virtio_console.c | 17 +++++++++++------
> >>  1 file changed, 11 insertions(+), 6 deletions(-)
> >>
> >> diff --git a/drivers/char/virtio_console.c b/drivers/char/virtio_console.c
> >> index 7270e7b69262..77105166fe01 100644
> >> --- a/drivers/char/virtio_console.c
> >> +++ b/drivers/char/virtio_console.c
> >> @@ -1421,12 +1421,17 @@ static int add_port(struct ports_device *portdev, u32 id)
> >>  	spin_lock_init(&port->outvq_lock);
> >>  	init_waitqueue_head(&port->waitqueue);
> >>  
> >> -	/* Fill the in_vq with buffers so the host can send us data. */
> >> -	nr_added_bufs = fill_queue(port->in_vq, &port->inbuf_lock);
> >> -	if (!nr_added_bufs) {
> >> -		dev_err(port->dev, "Error allocating inbufs\n");
> >> -		err = -ENOMEM;
> >> -		goto free_device;
> >> +	/* if the in_vq has not already been filled (the port has already been
> >> +	 * used and unplugged), fill the in_vq with buffers so the host can
> >> +	 * send us data.
> >> +	 */
> >> +	if (port->in_vq->num_free != 0) {
> >> +		nr_added_bufs = fill_queue(port->in_vq, &port->inbuf_lock);
> >> +		if (!nr_added_bufs) {
> >> +			dev_err(port->dev, "Error allocating inbufs\n");
> >> +			err = -ENOMEM;
> >> +			goto free_device;
> >> +		}
> >>  	}
> >>  
> >>  	if (is_rproc_serial(port->portdev->vdev))
> > 
> > Well fill_queue will just add slots as long as it can.
> > So on a full queue it does nothing. How does this patch help?
> 
> Yes, but in this case it returns 0 and so add_port() fails and exits
> with -ENOMEM and the device is freed. It's what this patch tries to avoid.
> 
> Thanks,
> Laurent

Oh I see. However it's a bit asymmetrical to special case ring full.
How about making fill_queue return int and testing return code for
-ENOSPC instead? Will also help propagate errors correctly.

And I guess CC stable?

-- 
MST