On Mon, 4 Sep 2017 10:44:02 +0000 Andrew Bainbridge <[email protected]> wrote:
> I'm a newbie. I want to learn more about to use mbufs to achieve the best > throughput. My application is something like a VPN server. In pseudo code: > > while 1: > pkt = recv() > if pkt.ip.daddr == CLIENT: > new_pkt = encap(pkt) > else: > new_pkt = decap(pkt) > send(new_pkt) > > Where encap() prepends an IP and UDP header, and decap() does the opposite. > > Most of each packet I send is the same as one I just received. Is it possible > to do the send without having to allocate a new mbuf and memcpy into it? > > I want to learn more about how the system works at the low level. > > My guess of how it works is that the NIC reads in a packet from the Ethernet > cable and writes it into its on-chip SRAM. Once it has enough data buffered, > or enough time has elapsed it does a PCIe write request to copy the data into > system RAM. The simplest scheme would be to have a single large circular > buffer in system RAM and for the packets to be written nose-to-tail into that > buffer. > Does DPDK do that? I guess not. I guess the supported cards all support > scatter/gather, which AFAICT means the NICs are smart enough to understand an > array of pointers to buffers. > > So what then? I have many 1500 byte buffers allocated, and I give the NIC an > array of pointers to those buffers. The NIC then "scatters" the input stream > into these buffers, one packet per buffer. > > I guess the best scheme for my application would be if I could tell the NIC > to always leave 30 bytes or so of headroom on each packet, so that I can > prepend the extra headers in the encap case. Can I request that when I > configure the mbufs? > > If you can point me to some kind of tutorial or blog post that covers this > area, that would also be helpful. > > Thanks, > Andrew You are heading off on a tangent. That is not how DPDK works. DPDK works like Linux and FreeBSD kernel. Either read the documentation or look at how examples work.
