On Fri, Apr 04, 2025 at 10:30:43AM +0200, Stefano Garzarella wrote: > On Fri, Apr 04, 2025 at 04:14:51AM -0400, Michael S. Tsirkin wrote: > > On Fri, Apr 04, 2025 at 10:04:38AM +0200, Alexander Graf wrote: > > > > > > On 03.04.25 14:21, Michael S. Tsirkin wrote: > > > > On Wed, Apr 02, 2025 at 12:14:24PM -0400, Stefan Hajnoczi wrote: > > > > > On Tue, Apr 01, 2025 at 08:13:49PM +0000, Alexander Graf wrote: > > > > > > Ever since the introduction of the virtio vsock driver, it included > > > > > > pushback logic that blocks it from taking any new RX packets until > > > > > > the > > > > > > TX queue backlog becomes shallower than the virtqueue size. > > > > > > > > > > > > This logic works fine when you connect a user space application on > > > > > > the > > > > > > hypervisor with a virtio-vsock target, because the guest will stop > > > > > > receiving data until the host pulled all outstanding data from the > > > > > > VM. > > > > > > > > > > > > With Nitro Enclaves however, we connect 2 VMs directly via vsock: > > > > > > > > > > > > Parent Enclave > > > > > > > > > > > > RX -------- TX > > > > > > TX -------- RX > > > > > > > > > > > > This means we now have 2 virtio-vsock backends that both have the > > > > > > pushback > > > > > > logic. If the parent's TX queue runs full at the same time as the > > > > > > Enclave's, both virtio-vsock drivers fall into the pushback path and > > > > > > no longer accept RX traffic. However, that RX traffic is TX traffic > > > > > > on > > > > > > the other side which blocks that driver from making any forward > > > > > > progress. We're now in a deadlock. > > > > > > > > > > > > To resolve this, let's remove that pushback logic altogether and > > > > > > rely on > > > > > > higher levels (like credits) to ensure we do not consume unbounded > > > > > > memory. > > > > > The reason for queued_replies is that rx packet processing may emit tx > > > > > packets. Therefore tx virtqueue space is required in order to process > > > > > the rx virtqueue. > > > > > > > > > > queued_replies puts a bound on the amount of tx packets that can be > > > > > queued in memory so the other side cannot consume unlimited memory. > > > > > Once > > > > > that bound has been reached, rx processing stops until the other side > > > > > frees up tx virtqueue space. > > > > > > > > > > It's been a while since I looked at this problem, so I don't have a > > > > > solution ready. In fact, last time I thought about it I wondered if > > > > > the > > > > > design of virtio-vsock fundamentally suffers from deadlocks. > > > > > > > > > > I don't think removing queued_replies is possible without a > > > > > replacement > > > > > for the bounded memory and virtqueue exhaustion issue though. Credits > > > > > are not a solution - they are about socket buffer space, not about > > > > > virtqueue space, which includes control packets that are not accounted > > > > > by socket buffer space. > > > > > > > > Hmm. > > > > Actually, let's think which packets require a response. > > > > > > > > VIRTIO_VSOCK_OP_REQUEST > > > > VIRTIO_VSOCK_OP_SHUTDOWN > > > > VIRTIO_VSOCK_OP_CREDIT_REQUEST > > > > > > > > > > > > the response to these always reports a state of an existing socket. > > > > and, only one type of response is relevant for each socket. > > > > > > > > So here's my suggestion: > > > > stop queueing replies on the vsock device, instead, > > > > simply store the response on the socket, and create a list of sockets > > > > that have replies to be transmitted > > > > > > > > > > > > WDYT? > > > > > > > > > Wouldn't that create the same problem again? The socket will eventually > > > push > > > back any new data that it can take because its FIFO is full. At that > > > point, > > > the "other side" could still have a queue full of requests on exactly that > > > socket that need to get processed. We can now not pull those packets off > > > the > > > virtio queue, because we can not enqueue responses. > > > > Either I don't understand what you wrote or I did not explain myself > > clearly. > > I didn't fully understand either, but with this last message of yours it's > clear to me and I like the idea! > > > > > In this idea there needs to be a single response enqueued > > like this in the socket, because, no more than one ever needs to > > be outstanding per socket. > > > > For example, until VIRTIO_VSOCK_OP_REQUEST > > is responded to, the socket is not active and does not need to > > send anything. > > One case I see is responding when we don't have the socket listening (e.g. > the port is not open), so if before the user had a message that the port was > not open, now instead connect() will timeout. So we could respond if we have > space in the virtqueue, otherwise discard it without losing any important > information or guarantee of a lossless channel. > > So in summary: > > - if we have an associated socket, then always respond (possibly > allocating memory in the intermediate queue if the virtqueue is full > as we already do). We need to figure out if a flood of > VIRTIO_VSOCK_OP_CREDIT_REQUEST would cause problems, but we can always > decide not to respond if we have sent this identical information > before.
If taking this path, need to consider not responding is within spec or not. But again, credit update needed is just a single flag we need to set on a socket. If we have anything we need to send, it can also update the credits. > - if there is no associated socket, we only respond if virtqueue has > space. > > I like it and it seems feasible without changing anything in the > specification. > > Did I get it right? > > Thanks, > Stefano That was the idea, yes. -- MST
