srvxdp.h File Reference

AF_XDP driver integration for the HTTP/3 server: opens one AF_XDP socket plus its BPF redirect filter, and drives the resulting rings with the app-facing mmsg_buf shape. More...

#include "app/http3/server/srvxdpbpf/srvxdpbpf.h"
#include "common/bytes/span/span.h"
#include "transport/io/socket/io/udp.h"
#include "transport/io/xdp/xdpmac/xdpmac.h"
#include "transport/io/xdp/xsksetup/xsksetup.h"
#include "transport/io/xdp/xskumem/xskumem.h"

Go to the source code of this file.

Data Structures
struct	wired_srvxdp_cfg
	Caller-supplied configuration for one AF_XDP-backed server socket. More...
struct	wired_srvxdp
	One open AF_XDP driver instance: the socket/rings, the TX frame pool, the learned peer MAC cache, and — when this instance opened its own BPF redirect filter rather than joining a shared one — those BPF objects. More...

Functions
i64	wired_srvxdp_open (wired_srvxdp *x, const wired_srvxdp_cfg *cfg)
	Open one AF_XDP-backed server socket with its own BPF redirect filter: wired_srvxdpbpf_open on cfg->ifindex, then xsksetup_open, then point the map's queue_id slot at the new socket.
i64	wired_srvxdp_open_shared (wired_srvxdp *x, const wired_srvxdp_cfg *cfg, wired_srvxdpbpf *bpf)
	Open one AF_XDP-backed server socket against an already-open shared per-interface BPF object: xsksetup_open, then register the new socket's fd in bpf's cfg->queue_id slot.
i64	wired_srvxdp_rx_burst (wired_srvxdp *x, mmsg_buf *bufs, usz nbufs)
	Drain up to nbufs received datagrams into bufs.
i64	wired_srvxdp_send (wired_srvxdp *x, const sockaddr *dst, wired_span pkt)
	Send one datagram to dst.
void	wired_srvxdp_close (wired_srvxdp *x)
	Tear down x: close the owned BPF link/prog/map fds (skipped when x was opened against a shared wired_srvxdpbpf, which the caller owns), then xsksetup_close the socket/rings/UMEM.
void	wired_srvxdp_print_stats (i64 fd)
	Read XDP_STATISTICS for fd (xsksetup_stats) and log one line per counter (name + value) via wired_log_str.

Detailed Description

AF_XDP driver integration for the HTTP/3 server: opens one AF_XDP socket plus its BPF redirect filter, and drives the resulting rings with the app-facing mmsg_buf shape.

Delegates every actual mechanism to its lower-layer domain (xsksetup for the syscall plumbing, xdpbpf for the filter program, xdpframe for the wire codec, xdpmac for the reflect cache, xskumem for the TX frame pool): this file is only the driving glue between them.

Function Documentation

wired_srvxdp_close()

void wired_srvxdp_close

(

wired_srvxdp *

x

)

Tear down x: close the owned BPF link/prog/map fds (skipped when x was opened against a shared wired_srvxdpbpf, which the caller owns), then xsksetup_close the socket/rings/UMEM.

Parameters

x	driver instance to close

wired_srvxdp_open()

i64 wired_srvxdp_open	(	wired_srvxdp *	x,
		const wired_srvxdp_cfg *	cfg )

Open one AF_XDP-backed server socket with its own BPF redirect filter: wired_srvxdpbpf_open on cfg->ifindex, then xsksetup_open, then point the map's queue_id slot at the new socket.

The TX frame pool is initialized over UMEM frames 64..127 (the RX pool, frames 0..63, is left to the kernel's fill/rx rings). On any failure, everything already built is torn down and a negative errno is returned. For multiple queues on one interface, use wired_srvxdp_open_shared instead: the kernel rejects a second XDP link on the same ifindex with -EBUSY.

Parameters

x	zero-initialized output; filled in on success
cfg	interface/queue/address/port and bind/attach flags

Returns

0 on success, or a negative errno

wired_srvxdp_open_shared()

i64 wired_srvxdp_open_shared	(	wired_srvxdp *	x,
		const wired_srvxdp_cfg *	cfg,
		wired_srvxdpbpf *	bpf )

Open one AF_XDP-backed server socket against an already-open shared per-interface BPF object: xsksetup_open, then register the new socket's fd in bpf's cfg->queue_id slot.

x keeps no reference to bpf afterwards (x->bpf stays -1/-1/-1, unowned): the caller owns bpf's lifetime and must keep it open while any registered socket is in use. On failure the socket is closed and a negative errno is returned.

Parameters

x	zero-initialized output; filled in on success
cfg	interface/queue/address/port and bind flags (attach_flags is unused: attachment happened in wired_srvxdpbpf_open)
bpf	shared BPF objects for cfg->ifindex (wired_srvxdpbpf_open)

Returns

0 on success, or a negative errno

wired_srvxdp_print_stats()

void wired_srvxdp_print_stats

(

i64

fd

)

Read XDP_STATISTICS for fd (xsksetup_stats) and log one line per counter (name + value) via wired_log_str.

A no-op (silently returns) if the stats read fails.

Parameters

fd	the XDP socket file descriptor

wired_srvxdp_rx_burst()

i64 wired_srvxdp_rx_burst	(	wired_srvxdp *	x,
		mmsg_buf *	bufs,
		usz	nbufs )

Drain up to nbufs received datagrams into bufs.

First reaps the completion ring (returning any finished TX frames to txpool), then consumes up to nbufs entries from the RX ring: each frame is parsed (xdpframe_parse), its payload copied into bufs[i].buf, its source into bufs[i].src, and the peer's MAC learned into the reflect cache (the first RX frame also seeds x->our_mac). Every consumed RX frame is returned to the fill ring before this call returns.

Parameters

x	open driver instance
bufs	array of nbufs receive slots
nbufs	number of slots in bufs

Returns

number of datagrams written into bufs (0..nbufs)

wired_srvxdp_send()

i64 wired_srvxdp_send	(	wired_srvxdp *	x,
		const sockaddr *	dst,
		wired_span	pkt )

Send one datagram to dst.

Drops (returns 0, letting QUIC retransmit) when dst's MAC is not yet learned, or when the TX frame pool is exhausted even after one completion-ring reap. On success, builds a complete eth+IPv4+UDP frame around pkt into a TX-pool frame, submits it to the TX ring, and kicks the kernel to service it (a kick failure is ignored: tests exercising this without a real fd expect it to fail).

Parameters

x	open driver instance
dst	destination address
pkt	the QUIC datagram to send

Returns

1 sent, 0 dropped