Network stack and VirtIO networking¶

Layered packet processing¶

Networking converts application data into packets transmitted through a network interface and performs the reverse operation on receive.

A useful separation is:

application / socket
        ↓
transport protocol
        ↓
network protocol
        ↓
link framing
        ↓
network interface driver
        ↓
device queues / wire or virtual host

Each layer owns a different addressing and reliability problem. Ethernet addresses a local link, IP addresses routed networks, and transport protocols identify endpoints and delivery semantics.

Network interface¶

A network driver moves packet buffers between host memory and a device. With VirtIO-net, the guest and virtual device exchange descriptor chains through virtqueues.

Receive buffers must remain allocated while the host can fill them. Transmit buffers/descriptors must not be recycled until completion establishes that the device no longer consumes them.

This is the same asynchronous ownership rule seen in storage and GPU devices.

Packets are untrusted input¶

Network packets originate outside the kernel's trust boundary. Every length, header offset and protocol field must be validated before use.

A malformed length can otherwise turn packet parsing into an out-of-bounds memory access. Checksums and protocol validity are not substitutes for memory-safety bounds checking.

Sockets¶

A socket provides a process/application-facing endpoint abstraction. The kernel maps application operations onto protocol state and packet queues.

Current ChrisOS socket state has ownership relationships to processes/slots. The locking audit notes that socket-table mutation is BSP/receive-path sensitive and does not yet have a general socket lock suitable for arbitrary AP receive concurrency.

That limitation must be preserved in documentation rather than generalized into "fully SMP network stack."

Transfer service¶

net_xfer.c provides project-specific host/guest transfer functionality. Application protocols built on top of TCP or another transport should be treated separately from the transport implementation itself.

Interrupt and polling boundaries¶

A device can notify receive completion asynchronously. The driver path must coordinate IRQ context with consumers that read or mutate packet queues.

If receive processing is deferred to a non-IRQ context, the handoff itself is a synchronization boundary.

Validation¶

A useful network test hierarchy includes:

parser/unit tests for malformed frames and lengths;
driver queue lifecycle tests;
deterministic virtual-device tests;
QEMU end-to-end packet exchange;
physical NIC tests for supported hardware profiles.

Success with VirtIO-net demonstrates the paravirtual path, not an arbitrary physical Ethernet controller.