xdp_datagram_source.hpp File Reference

AF_XDP datagram source (ADR-0015, section 3): one XDP socket per (interface, RX queue), each with its own UMEM, fill and RX rings (the completion ring exists because bind requires it). More...

#include "fastmm/core/config_macros.hpp"
#include "fastmm/core/time.hpp"
#include "fastmm/net/datagram_source.hpp"
#include "fastmm/net/udp_frame.hpp"
#include "fastmm/net/xdp_program.hpp"
#include <sys/socket.h>
#include <atomic>
#include <cstddef>
#include <cstdint>
#include <span>
#include <string>
#include <string_view>
#include <utility>
#include <vector>

Go to the source code of this file.

Classes
struct	fastmm::net::XdpSubscription
struct	fastmm::net::XdpInterfaceQueues
struct	fastmm::net::XdpConfig
struct	fastmm::net::XdpStats
struct	fastmm::net::XdpSocketStats
	XDP_STATISTICS of one socket, as of the last refresh_stats(). More...
struct	fastmm::net::XdpInterfaceStatus
class	fastmm::net::XdpFilter
	The BPF objects of one interface: subscription hash map, XSKMAP, per-CPU fallback counter and the program, plus its XDP link once attached. More...
class	fastmm::net::XdpDatagramSource

Namespaces
namespace	fastmm
namespace	fastmm::net

Enumerations
enum class	fastmm::net::XdpMode : std::uint8_t { fastmm::net::Auto , fastmm::net::ZeroCopy , fastmm::net::NativeCopy , fastmm::net::Generic }
	Attach and bind mode. More...

Functions
constexpr std::string_view	fastmm::net::to_string (XdpMode m) noexcept
std::string	fastmm::net::xdp_capability_error ()
	"" when the process has CAP_NET_ADMIN, CAP_NET_RAW, CAP_BPF (or CAP_SYS_ADMIN) and CAP_IPC_LOCK in its effective set; otherwise names the missing ones and the setcap command.
std::string	fastmm::net::xdp_kernel_error ()
	"" on Linux 5.11 or later; otherwise the reason.

Detailed Description

AF_XDP datagram source (ADR-0015, section 3): one XDP socket per (interface, RX queue), each with its own UMEM, fill and RX rings (the completion ring exists because bind requires it).

A BPF program per interface, loaded and attached over raw bpf(2) calls, redirects UDP datagrams for the subscribed (group, port) pairs to the sockets and passes everything else to the kernel. A kernel UDP socket per multicast subscription joins the group so IGMP reports go out; it reads nothing. A unicast subscription is a local address (no join). Linux 5.11 or later; needs CAP_NET_ADMIN, CAP_NET_RAW, CAP_BPF and CAP_IPC_LOCK.

Single-threaded: open, poll and the stats calls belong to the net thread. poll() allocates nothing and returns every RX descriptor it took to the fill ring before it returns, so payload spans are valid only inside the handler call.