On Tue, Nov 16, 2021 at 07:37:34AM +0000, Ciara Loftus wrote: > The common case for AF_XDP sockets (xsks) is creating a single xsk on a queue for sending and > receiving frames as this is analogous to HW packet steering through RSS and other classification > methods in the NIC. AF_XDP uses the xdp redirect infrastructure to direct packets to the socket. It > was designed for the much more complicated case of DEVMAP xdp_redirects which directs traffic to > another netdev and thus potentially another driver. In the xsk redirect case, by skipping the > unnecessary parts of this common code we can significantly improve performance and pave the way > for batching in the driver. This RFC proposes one such way to simplify the infrastructure which > yields a 27% increase in throughput and a decrease in cycles per packet of 24 cycles [1]. The goal > of this RFC is to start a discussion on how best to simplify the single-socket datapath while > providing one method as an example. > > Current approach: > 1. XSK pointer: an xsk is created and a handle to the xsk is stored in the XSKMAP. > 2. XDP program: bpf_redirect_map helper triggers the XSKMAP lookup which stores the result (handle > to the xsk) and the map type (XSKMAP) in the percpu bpf_redirect_info struct. The XDP_REDIRECT > action is returned. > 3. XDP_REDIRECT handling called by the driver: the map type (XSKMAP) is read from the > bpf_redirect_info which selects the xsk_map_redirect path. The xsk pointer is retrieved from the > bpf_redirect_info and the XDP descriptor is pushed to the xsk's Rx ring. The socket is added to a > list for flushing later. > 4. xdp_do_flush: iterate through the lists of all maps that can be used for redirect (CPUMAP, > DEVMAP and XSKMAP). When XSKMAP is flushed, go through all xsks that had any traffic redirected to > them and bump the Rx ring head pointer(s). > > For the end goal of submitting the descriptor to the Rx ring and bumping the head pointer of that > ring, only some of these steps are needed. The rest is overhead. The bpf_redirect_map > infrastructure is needed for all other redirect operations, but is not necessary when redirecting > to a single AF_XDP socket. And similarly, flushing the list for every map type in step 4 is not > necessary when only one socket needs to be flushed. > > Proposed approach: > 1. XSK pointer: an xsk is created and a handle to the xsk is stored both in the XSKMAP and also the > netdev_rx_queue struct. > 2. XDP program: new bpf_redirect_xsk helper returns XDP_REDIRECT_XSK. > 3. XDP_REDIRECT_XSK handling called by the driver: the xsk pointer is retrieved from the > netdev_rx_queue struct and the XDP descriptor is pushed to the xsk's Rx ring. > 4. xsk_flush: fetch the handle from the netdev_rx_queue and flush the xsk. > > This fast path is triggered on XDP_REDIRECT_XSK if: > (i) AF_XDP socket SW Rx ring configured > (ii) Exactly one xsk attached to the queue > If any of these conditions are not met, fall back to the same behavior as the original approach: > xdp_redirect_map. This is handled under-the-hood in the new bpf_xdp_redirect_xsk helper so the user > does not need to be aware of these conditions. I don't think the micro optimization for specific use case warrants addition of new apis. Please optimize it without adding new actions and new helpers.
