Skip to content

Latest commit

 

History

History
37 lines (28 loc) · 1.54 KB

README.md

File metadata and controls

37 lines (28 loc) · 1.54 KB
Raw

libsinet - S is for "Speed"

This is libsinet, a lightweight userland implementation of the UDP/IPv4 stack designed to plug into the netmap framework.

Design Objectives

  • multithread-safe, multithread-optimized
  • provably small CPU run-time overhead
  • high-PPS optimized
  • as close as socket(2)/kqueue(2) API (aka "netold") as possible
  • UDP/IPv4/layerII initially, IPv6 / TCP in mind

Differences between libsinet and netold

  • Void pointer is used as an object descriptor instead of int
  • Error reporting is performed into user supplied int buffer user has an option to ignore that by providing NULL pointer instead
  • to facilitate zero-copying, all i/o is performed by library-managed buffers linked directly into kernel rings. Fixed and limited number of those buffers are allocated during the sin_init(). It is the applications duty to use those wisely, making sure that each each buffer produced by the sin_recv() call is either going to be consumed by the sin_send() method, or recycled by invoking sin_pkt_release(). If application needs more packets for sending data out than it gets fromsin_recv() then additional packets may be produced by sin_pkg_alloc()

Caveats

  • In order for pass-through mode to work between hardware and software rings, hardware checksum off-loading must be disabled. As an example, checksum off-loading can be disabled for the em(4) driver as follows:
    # ifconfig emX -rxcsum
    # ifconfig emX -txcsum