On Fri, Apr 08, 2016 at 04:35:05PM +0100, Ian Campbell wrote: > On Fri, 2016-04-01 at 15:23 +0100, Stefan Hajnoczi wrote: > > This series is based on Michael Tsirkin's vhost branch (v4.5-rc6). > > > > I'm about to process Claudio Imbrenda's locking fixes for virtio-vsock but > > first I want to share the latest version of the code. Several people are > > playing with vsock now so sharing the latest code should avoid duplicate work. > > Thanks for this, I've been using it in my project and it mostly seems > fine. > > One wrinkle I came across, which I'm not sure if it is by design or a > problem is that I can see this sequence coming from the guest (with > other activity in between): > > 1) OP_SHUTDOWN w/ flags == SHUTDOWN_RX > 2) OP_SHUTDOWN w/ flags == SHUTDOWN_TX > 3) OP_SHUTDOWN w/ flags == SHUTDOWN_TX|SHUTDOWN_RX > > I orignally had my backend close things down at #2, however this meant > that when #3 arrived it was for a non-existent socket (or, worse, an > active one if the ports got reused). I checked v5 of the spec > proposal[0] which says: > If these bits are set and there are no more virtqueue buffers > pending the socket is disconnected. > > but I'm not entirely sure if this behaviour contradicts this or not > (the bits have both been set at #2, but not at the same time). > > BTW, how does one tell if there are no more virtqueue buffers pending > or not while processing the op? #2 is odd. The shutdown bits are sticky so they cannot be cleared once set. I would have expected just #1 and #3. The behavior you observe look like a bug. The spec text does not convey the meaning of OP_SHUTDOWN well. OP_SHUTDOWN SHUTDOWN_TX|SHUTDOWN_RX means no further rx/tx is possible for this connection. "there are no more virtqueue buffers pending the socket" really means that this isn't an immediate close from the perspective of the application. If the application still has unread rx buffers then the socket stays readable until the rx data has been fully read. > Another thing I noticed, which is really more to do with the generic > AF_VSOCK bits than anything to do with your patches is that there is no > limitations on which vsock ports a non-privileged user can bind to and > relatedly that there is no netns support so e.g. users in unproivileged > containers can bind to any vsock port and talk to the host, which might > be undesirable. For my use for now I just went with the big hammer > approach of denying access from anything other than init_net > namespace[1] while I consider what the right answer is. From the vhost point of view each netns should have its own AF_VSOCK namespace. This way two containers could act as "the host" (CID 2) for their respective guests.
Attachment:
signature.asc
Description: PGP signature
