AF_XDP socket setup: socket(2)/setsockopt(2)/getsockopt(2)/mmap(2)/bind(2) sequence that turns a raw AF_XDP fd into four ready-to-poll SPSC rings (fill/comp/rx/tx) over one UMEM. More...

#include "common/platform/sys/syscall.h"
#include "transport/io/xdp/xskring/xskring.h"

Data Structures
struct	xsk_cfg
	Caller-supplied configuration for one AF_XDP socket. More...
struct	xsk
	One open AF_XDP socket: the fd, its UMEM, the four rings, and the mmap regions backing them (kept so xsksetup_close can munmap them). More...

Macros
#define	XSKSETUP_UMEM_FRAMES 1024u
	Fixed capacity: one UMEM frame per fill/comp/rx/tx ring entry slot count below, matching xskumem's XSKUMEM_FRAMES (1024 frames of 2048B = 2MiB).
#define	XSKSETUP_FRAME_SIZE 2048u
	Byte size of one UMEM frame, matching xskumem's XSKUMEM_FRAME_SIZE.
#define	XSKSETUP_UMEM_LEN ((u64)XSKSETUP_UMEM_FRAMES * XSKSETUP_FRAME_SIZE)
	Total UMEM byte length.
#define	XSKSETUP_RX_POOL_FRAMES 256u
	Number of UMEM frames (0..XSKSETUP_RX_POOL_FRAMES) reserved for the kernel's fill/RX side (xsksetup_prime_fill).
#define	XSKSETUP_RING_ENTRIES 64u
	Entries per ring (fill/comp/rx/tx), must be a power of two.

Functions
i64	xsksetup_open (xsk x, const xsk_cfg cfg)
	Open one AF_XDP socket bound to cfg->ifindex/queue_id: socket -> UMEM anonymous mmap -> XDP_UMEM_REG -> ring size setsockopts -> XDP_MMAP_OFFSETS -> mmap the four rings -> xskring_init each -> bind -> prime the fill ring with all RX-pool frames.
void	xsksetup_close (xsk *x)
	Tear down x: munmap every ring region and the UMEM, close the fd.
i64	xsksetup_kick_tx (i64 fd)
	Wake the kernel to service the TX ring (sendto with no data, matching the driver's XDP_COPY bind mode which needs an explicit kick).
i64	xsksetup_stats (i64 fd, u64 out[6])
	Read XDP_STATISTICS into out[6]: rx_dropped, rx_invalid_descs, tx_invalid_descs, rx_ring_full, rx_fill_ring_empty_descs, tx_ring_empty_descs (linux/if_xdp.h struct xdp_statistics order).

Detailed Description

AF_XDP socket setup: socket(2)/setsockopt(2)/getsockopt(2)/mmap(2)/bind(2) sequence that turns a raw AF_XDP fd into four ready-to-poll SPSC rings (fill/comp/rx/tx) over one UMEM.

Everything above this layer (ring op / frame parse / bpf) is pure and already tested; this file is only the syscall plumbing that wires those parts to the kernel.

Macro Definition Documentation

◆ XSKSETUP_RX_POOL_FRAMES

#define XSKSETUP_RX_POOL_FRAMES 256u

Number of UMEM frames (0..XSKSETUP_RX_POOL_FRAMES) reserved for the kernel's fill/RX side (xsksetup_prime_fill).

The remainder (XSKSETUP_UMEM_FRAMES - XSKSETUP_RX_POOL_FRAMES) is the server's TX pool (srvxdp's SRVXDP_TXPOOL_FRAMES). Skewed toward TX: an HTTP/3 server mostly transmits (response bodies) and rarely needs many RX frames in flight at once.

◆ XSKSETUP_UMEM_FRAMES

#define XSKSETUP_UMEM_FRAMES 1024u

Fixed capacity: one UMEM frame per fill/comp/rx/tx ring entry slot count below, matching xskumem's XSKUMEM_FRAMES (1024 frames of 2048B = 2MiB).

A 1MB file at a ~1400B QUIC packet size needs ~750 in-flight TX frames; the prior 128-frame UMEM (64 TX) stalled on txpool exhaustion well before that (observed as xdp_statistics.tx_ring_empty_descs = 14352 on a real VPS run). 1024 gives headroom above that without growing the mmap past a few MiB.

◆ XSKSETUP_UMEM_LEN

#define XSKSETUP_UMEM_LEN ((u64)XSKSETUP_UMEM_FRAMES * XSKSETUP_FRAME_SIZE)

Total UMEM byte length.

Cast to u64 before multiplying: both operands are u32, and a plain u32*u32 product then gets implicitly widened when used as a u64 length/offset (bugprone-implicit-widening-of-multiplication-result). The current constants fit in 32 bits either way, but computing the product in the target width is what the mmap length/UMEM offset math needs.

Function Documentation

◆ xsksetup_close()

void xsksetup_close ( xsk * x )

Tear down x: munmap every ring region and the UMEM, close the fd.

Safe to call twice (a second call is a no-op: fd is left at -1 after the first).

◆ xsksetup_open()

i64 xsksetup_open	(	xsk *	x,
		const xsk_cfg *	cfg )

Open one AF_XDP socket bound to cfg->ifindex/queue_id: socket -> UMEM anonymous mmap -> XDP_UMEM_REG -> ring size setsockopts -> XDP_MMAP_OFFSETS -> mmap the four rings -> xskring_init each -> bind -> prime the fill ring with all RX-pool frames.

On any failure, everything allocated so far is unwound and a negative errno is returned.

Parameters

x	zero-initialized output; filled in on success
cfg	ifindex/queue_id/bind_flags

Returns: 0 on success, or a negative errno

Data Structures

Macros

Functions