Re: [PATCH RFC v3 0/8] vsock: add datagram support to virtio and vhost
From: Bobby Eshleman
Date: Thu Nov 19 2026 - 16:52:53 EST
On Thu, Nov 19, 2026 at 08:06:44AM -0800, Sergio Lopez Pascual wrote:
> Bobby Eshleman <bobbyeshleman@xxxxxxxx> writes:
>
> > On Fri, Nov 13, 2026 at 06:15:40PM +0100, Sergio Lopez wrote:
> >> This RFC adds SOCK_DGRAM support to virtio-vsock and vhost-vsock through
> >> the existing AF_VSOCK API. It is intended for guest and host services
> >> that exchange independent requests, responses, or events and need message
> >> boundaries and sender addresses without a stream connection or application
> >> framing.
> >>
> >> VIRTIO_VSOCK_F_DGRAM gates virtio traffic in both directions. Each virtio
> >> send carries zero to 64 KiB of payload in one VIRTIO_VSOCK_TYPE_DGRAM
> >> packet. On the RX virtqueue, the device may split a datagram across
> >> page-sized guest buffers.
> >> START and EOM delimit the message; the guest reassembles at most 64
> >> fragments and 64 KiB before delivering it to a socket. Invalid or
> >> incomplete messages are discarded. Datagram delivery is best effort:
> >> there are no peer credits, acknowledgements, or retransmissions. Send
> >> admission uses the socket send-memory limit, and SIOCOUTQ reports charged
> >> send memory.
> >>
> >> Datagram sockets have a separate port space from stream and seqpacket
> >> sockets. The latter two continue to share their port space. Datagram
> >> routes are chosen per destination among guest-to-host, host-to-guest, and
> >> loopback transports, with VMCI as a fallback. VMCI still owns its socket
> >> handle when it is present; loopback works without a directional transport.
> >> Hyper-V continues to reject datagram sockets.
> >>
> >> The Sashiko review of v2 also reported that a VMCI datagram bind in a
> >> local-mode network namespace creates a receive handle in VMCI's global
> >> port space, so a sender outside the namespace can deliver datagrams
> >> across the namespace isolation boundary. The issue predates this series
> >> and it's getting addressed in a separate patch:
> >>
> >> https://lore.kernel.org/all/20261113-vsock-vmci-netns-isolation-v1-1-ced02054ea61@xxxxxxxxxx
> >>
> >> Within this series, local-mode namespaces bind datagrams only in the
> >> core table and never create a VMCI handle.
> >>
> >> Bobby Eshleman and Amery Hung explored this feature in their v6 series
> >> and v7 work-in-progress branch. This series is a fresh implementation
> >> against the current vsock code. It follows the separate datagram port
> >> space, per-destination routing, and VMCI fallback requirements discussed
> >> there. For guest receive buffers, it uses bounded fragmentation and
> >> reassembly rather than full-message buffers or the earlier drop/error
> >> queue approach:
> >>
> >> https://lore.kernel.org/r/20240710212555.1617795-1-amery.hung@xxxxxxxxxxxxx
> >> https://github.com/ameryhung/linux/tree/vsock-dgram-v7
> >>
> >> A corresponding virtio specification update has been drafted separately.
> >> The wire format and feature negotiation still need specification review
> >> before this series can be considered for merging.
> >>
> >> The first patches expose the core port allocator and add datagram queue
> >> reporting, then implement the common data path, guest and host feature
> >> negotiation, routing, and RX fragmentation. The last patch extends
> >> vsock_test with local and peer tests for message boundaries, routing,
> >> connected peers, port sharing, queue accounting, and delayed reads.
> >>
> >> Testing used QEMU with the F_DGRAM property patch below, an x86-64
> >> KASAN/lockdep kernel, paired guest tests, and a nested Linux vhost-vsock
> >> backend. The tested paths include 64 KiB messages, delayed-read bursts,
> >> feature-disabled ingress, stream and seqpacket controls, and header-only
> >> guest RX buffers. A temporary userspace backend still fails stream test
> >> 21 on both the series base and tip; that case passes with Linux vhost.
> >> VMCI's host personality was exercised in a QEMU guest, including
> >> coexistence with virtio and self-send without loopback. A VMware VMCI guest
> >> device and Hyper-V hardware were not exercised.
> >>
> >> https://lists.gnu.org/archive/html/qemu-devel/2026-10/msg08074.html
> >>
> >> Code was written and reviewed both manually and assisted by Codex
> >> (GPT-6.1-Sol) and Claude (Opus 4.6).
> >>
> >> Signed-off-by: Sergio Lopez <slp@xxxxxxxxxx>
> >> ---
> >> Changes in v3:
> >> - Preserve Hyper-V stream routing and prefer virtio for guest CID_HOST
> >> datagrams (thanks to Sashiko AI bot).
> >> - Bind explicit non-VMCI CIDs and local-mode sockets without a VMCI
> >> handle, resolve wildcard destinations, and retain VMCI self routes
> >> (thanks to Sashiko AI bot).
> >> - Recheck bound addresses after receive takes the socket lock and set
> >> TO_HOST only for forwarded guest sources (thanks to Sashiko AI bot).
> >> - Probe negotiated virtio datagrams and numeric local CIDs before running
> >> dependent tests; cover wildcard and VMCI coexistence (thanks to Sashiko
> >> AI bot).
> >> - Link to v2: https://lore.kernel.org/r/20261113-vsock-dgram-support-v2-0-bacfc85b71ce@xxxxxxxxxx
> >>
> >> Changes in v2:
> >> - Keep guest RX buffers page-sized; add START/EOM framing and bounded
> >> datagram reassembly (thanks to Alexei Starovoitov and Bobby Eshleman).
> >> - Preserve stream and seqpacket data when an RX buffer has no payload
> >> room.
> >> - Separate datagram and connection-oriented port spaces, and share
> >> receive metadata and queueing (thanks to Alexei Starovoitov).
> >> - Check negotiated support and destination CID on ingress, and require
> >> an exact connected peer CID and port (thanks to Sashiko AI bot).
> >> - Select a capable route for each destination, retain VMCI as fallback,
> >> and support loopback-only sockets (thanks to Alexei Starovoitov).
> >> - Preserve the connected peer address across temporary parent fallback
> >> (thanks to Sashiko AI bot).
> >> - Use socket write-memory accounting for datagram SIOCOUTQ instead of
> >> separate transport counters (thanks to Alexei Starovoitov).
> >> - Extend selftests for fragmentation, delayed reads, port sharing,
> >> connected peers, and capability-based skips; fix the test build and
> >> README entries (thanks to Alexei Starovoitov, Bobby Eshleman,
> >> and Sashiko AI bot).
> >> - Link to v1: https://lore.kernel.org/r/20261028-vsock-dgram-support-v1-0-e457cf7cee78@xxxxxxxxxxx
> >>
> >> ---
> >> Sergio Lopez (8):
> >> vsock: expose the core port allocator to transports
> >> vsock: report datagram receive and send queue state
> >> vsock/virtio: implement the datagram data path
> >> vsock/virtio: negotiate datagram support on the guest
> >> vhost/vsock: negotiate datagram support with guests
> >> vsock: allow datagrams to use directional transports
> >> vsock: fragment datagrams across receive buffers
> >> vsock: add datagram tests to vsock_test
> >>
> >> Documentation/admin-guide/sysctl/net.rst | 8 +-
> >> drivers/vhost/vsock.c | 51 +-
> >> include/linux/virtio_vsock.h | 8 +-
> >> include/net/af_vsock.h | 31 +-
> >> include/uapi/linux/virtio_vsock.h | 10 +
> >> net/vmw_vsock/af_vsock.c | 526 +++++++++++----
> >> net/vmw_vsock/hyperv_transport.c | 11 +-
> >> net/vmw_vsock/virtio_transport.c | 167 ++++-
> >> net/vmw_vsock/virtio_transport_common.c | 172 ++++-
> >> net/vmw_vsock/vmci_transport.c | 124 ++--
> >> net/vmw_vsock/vsock_bpf.c | 3 +-
> >> net/vmw_vsock/vsock_loopback.c | 2 +-
> >> tools/testing/vsock/Makefile | 3 +-
> >> tools/testing/vsock/README | 33 +-
> >> tools/testing/vsock/util.c | 62 +-
> >> tools/testing/vsock/util.h | 10 +
> >> tools/testing/vsock/vsock_test.c | 175 ++++-
> >> tools/testing/vsock/vsock_test_dgram.c | 1061 ++++++++++++++++++++++++++++++
> >> tools/testing/vsock/vsock_test_dgram.h | 30 +
> >> 19 files changed, 2195 insertions(+), 292 deletions(-)
> >> ---
> >> base-commit: 65ba8f34ec886ec72b2c07ddc8cf3d815c32869d
> >> change-id: 20261027-vsock-dgram-support-683a4e0aebe7
> >>
> >> Best regards,
> >> --
> >> Sergio Lopez <slp@xxxxxxxxxx>
> >>
> >
> > Hey Sergio,
> >
> > Just a gentle heads up that net-next asks for 24hrs before reposting.
>
> Whoops, sorry!
>
> > I'll start working my way through the series, sorry if my reviews come
> > back spread out across a few days, not sure I'll have enough time in a
> > single sitting.
>
> I was planning on sending a v4 fixing Sashiko's issues. Should I wait
> for your review first?
>
> Thanks!
> Sergio.
>
Feel free to send without my review, I'm lagging behind!
Best,
Bobby