ocaml-virtio/lib/mmio/ at main · gazagnaire.org/ocaml-virtio · Tangled

gazagnaire.org / ocaml-virtio

0

Virtio-serial port IO for guest VMs

0

ocaml-virtio / lib / mmio /

at main 15 files

Thomas Gazagnaire virtio: state the descriptor address range wire 1.2.0 no longer masks 10d ago

virtio: publish ring ownership across hardware fences Weakly ordered ARM could expose split avail.idx or packed ownership flags before the descriptor fields they protect. Drivers could likewise consume packed fields or reuse transmit buffers without acquiring the device's completion. Separate field preparation from ownership publication in the net, block, and console frontends, and fence both sides of each handoff.

1 month ago

virtio: publish ring ownership across hardware fences Weakly ordered ARM could expose split avail.idx or packed ownership flags before the descriptor fields they protect. Drivers could likewise consume packed fields or reuse transmit buffers without acquiring the device's completion. Separate field preparation from ownership publication in the net, block, and console frontends, and fence both sides of each handoff.

1 month ago

virtio: Preserve full-width descriptor addresses

1 month ago

virtio: drive the mmio device model allocation-free per frame The device model walked each descriptor chain into a list of buffer records and read every field through a freshly allocated sub-slice. Replace chain/read/write with fold_chain and staged field getters/setters over a reused buffer: fold_chain visits a chain's descriptors with no list and no record, first_writable locates the RX buffer directly, and the MEM signature narrows to the two bulk copies the hot path needs. desc.addr becomes a uint63 int and inject_rx returns -1 rather than None, so a delivered frame allocates nothing.

2 months ago

virtio: drive VIRTIO_F_EVENT_IDX and harden PCI bring-up for vz Apple's Virtualization.framework exposed three protocol gaps that QEMU tolerated: - The framework clears FEATURES_OK unless the driver accepts VIRTIO_F_INDIRECT_DESC and VIRTIO_F_EVENT_IDX. Both are now accepted by the net and console drivers; indirect descriptors are never posted (legal), and EVENT_IDX is driven properly: the driver publishes used_event (the trailing le16 of the available ring, spec 2.7.5) after every reap, since a device honoring the feature holds back used-ring publication until used_idx passes it. - The framework acknowledges device_status writes asynchronously on its own dispatch queue. The FEATURES_OK re-read now polls bounded instead of declaring a spurious refusal, and set_driver_ok writes the full accumulated status byte instead of read-modify-write, which a stale read would corrupt. - The console gains the VIRTIO_CONSOLE_F_MULTIPORT control-queue handshake (spec 5.3.6.2) as a pure Control state machine driven from connect: DEVICE_READY, PORT_READY per DEVICE_ADD, PORT_OPEN on CONSOLE_PORT. A non-multiport device (Apple's serial-port console, QEMU without virtio-serial) falls back to the plain single-port path. New multiport/opened introspection on the console.

1 month ago

virtio: drive VIRTIO_F_EVENT_IDX and harden PCI bring-up for vz Apple's Virtualization.framework exposed three protocol gaps that QEMU tolerated: - The framework clears FEATURES_OK unless the driver accepts VIRTIO_F_INDIRECT_DESC and VIRTIO_F_EVENT_IDX. Both are now accepted by the net and console drivers; indirect descriptors are never posted (legal), and EVENT_IDX is driven properly: the driver publishes used_event (the trailing le16 of the available ring, spec 2.7.5) after every reap, since a device honoring the feature holds back used-ring publication until used_idx passes it. - The framework acknowledges device_status writes asynchronously on its own dispatch queue. The FEATURES_OK re-read now polls bounded instead of declaring a spurious refusal, and set_driver_ok writes the full accumulated status byte instead of read-modify-write, which a stale read would corrupt. - The console gains the VIRTIO_CONSOLE_F_MULTIPORT control-queue handshake (spec 5.3.6.2) as a pure Control state machine driven from connect: DEVICE_READY, PORT_READY per DEVICE_ADD, PORT_OPEN on CONSOLE_PORT. A non-multiport device (Apple's serial-port console, QEMU without virtio-serial) falls back to the plain single-port path. New multiport/opened introspection on the console.

1 month ago

virtio: add the virtio-mmio net-device driver Add Virtio_mmio.Net_device, the guest virtio-net driver that assembles the register Transport, the split-ring Driver, and the Net_driver frontend over a DMA arena. It was the Xen backend's virtio_net.ml; it is platform-neutral -- the caller injects the candidate register windows, the MMIO accessors, a DMA allocator, and a memory barrier -- so any backend (Xen dom0, Unikraft) drives the same QEMU virtio-mmio transport through it. Device-tree discovery stays with the platform; this probes the candidate addresses it is given.

1 month ago

virtio-net: a multi-queue guest driver, one queue pair per shard Net.Front now brings up N receive/transmit pairs, not one. connect takes ?queue_pairs: more than one negotiates VIRTIO_NET_F_MQ and the control virtqueue, clamps the count to the device's advertised max_virtqueue_pairs, programs virtqueues 2p/2p+1 per pair plus the control queue, and posts VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET (Virtio_net.ctrl_mq_set) to bring the extra pairs online. The device is a set of independent queue handles -- send/receive/failed act on a queue, mac/mtu/queues/queue_count on the device -- so one share-nothing stack shard drives each pair: this is the guest side of scaling the sharded TCP stack past a single queue's ceiling. Net.Back gains service_ctrl, the device mirror that applies the pairs-set command and acknowledges it, so the loopback oracle is faithful. The Mmio model now serves a config word that runs off the end of the config by zero-padding its tail, as a real device does, so a two-byte field like max_virtqueue_pairs at the config tail reads correctly. test_net_device drives three pairs end to end: each carries its own frame both ways over its own virtqueues, and the device services the pairs-set command.

1 month ago

virtio: give the mmio and vring headings their own paragraph odoc requires a paragraph to start on its own line, so the parenthetical trailing each heading was a parse error and the heading did not render.

1 month ago

virtio: publish ring ownership across hardware fences Weakly ordered ARM could expose split avail.idx or packed ownership flags before the descriptor fields they protect. Drivers could likewise consume packed fields or reuse transmit buffers without acquiring the device's completion. Separate field preparation from ownership publication in the net, block, and console frontends, and fence both sides of each handoff.

1 month ago

virtio: publish ring ownership across hardware fences Weakly ordered ARM could expose split avail.idx or packed ownership flags before the descriptor fields they protect. Drivers could likewise consume packed fields or reuse transmit buffers without acquiring the device's completion. Separate field preparation from ownership publication in the net, block, and console frontends, and fence both sides of each handoff.

1 month ago

virtio: add a QueueNumMax reader to the mmio transport

1 month ago

virtio: add a QueueNumMax reader to the mmio transport

1 month ago

virtio: state the descriptor address range wire 1.2.0 no longer masks wire 1.2.0 removed uint63, because an 8-byte field spans more values than an Optint.Int63.t holds and decoding masked the difference away on both the OCaml and the generated C path. The descriptor address keeps that carrier -- it is immediate on a 64-bit host and a descriptor is read and written on the ring's hot path -- and the description now says which addresses it accepts: a constraint refuses one with bit 63 set, in the OCaml decoder and in the EverParse validator alike. The ring is written by the guest, so dropping the top bit of an address it chose would be a device-model defect rather than a rounding of it; every address this stack puts on a ring comes from an OCaml int, which cannot reach bit 63. The lengths, indices, flags and the virtio-net header fields are small numbers and stay plain ints, where UInt16.v and SInt32.of_int refuse a value the field cannot hold.

1 week ago

virtio: give the mmio and vring headings their own paragraph odoc requires a paragraph to start on its own line, so the parenthetical trailing each heading was a parse error and the heading did not render.

1 month ago