On Fri, 4 Sep 2020 17:39:17 +0200 Björn Töpel <bjorn.topel@xxxxxxxxx> wrote: > On 2020-09-04 17:11, Jesper Dangaard Brouer wrote: > > On Fri, 4 Sep 2020 15:53:28 +0200 Björn Töpel > > <bjorn.topel@xxxxxxxxx> wrote: > > > >> From: Björn Töpel <bjorn.topel@xxxxxxxxx> > >> > >> The xsk_do_redirect_rx_full() helper can be used to check if a > >> failure of xdp_do_redirect() was due to the AF_XDP socket had a > >> full Rx ring. > > > > This is very AF_XDP specific. I think that the cpumap could likely > > benefit from similar approach? e.g. if the cpumap kthread is > > scheduled on the same CPU. > > > > At least I thought this was *very* AF_XDP specific, since the kernel is > dependent of that userland runs. Allocation (source) and Rx ring (sink). > Maybe I was wrong! :-) > > The thing with AF_XDP zero-copy, is that we sort of assume that if a > user enabled that most packets will have XDP_REDIRECT to an AF_XDP socket. > > > > But for cpumap we only want this behavior if sched on the same CPU > > as RX-NAPI. This could be "seen" by the cpumap code itself in the > > case bq_flush_to_queue() drops packets, check if rcpu->cpu equal > > smp_processor_id(). Maybe I'm taking this too far? > > > > Interesting. So, if you're running on the same core, and redirect fail > for CPUMAP, you'd like to yield the NAPI loop? Is that really OK from a > fairness perspective? I mean, with AF_XDP zero-copy we pretty much know > that all actions will be redirect to socket. For CPUMAP type of > applications, can that assumption be made? Yes, you are right. The RX NAPI loop could be doing something else, and yielding the NAPI loop due to detecting same-CPU is stalling on cpumap delivery might not be correct action. I just tested the same-CPU processing case for cpumap (result below signature), and it doesn't exhibit the bad 'dropping-off-edge' performance slowdown. The cpumap code also already tries to mitigate this, by calling wake_up_process() for every 8 packets (CPU_MAP_BULK_SIZE). I find your patchset very interesting, as I believe we do need some kind of general push-back "flow-control" mechanism for XDP. Maybe I should solve this differently in our XDP-TX-QoS pipe dream ;-) -- Best regards, Jesper Dangaard Brouer MSc.CS, Principal Kernel Engineer at Red Hat LinkedIn: http://www.linkedin.com/in/brouer Quick benchmark of cpumap. Same CPU RX and cpumap processing: ---------------------------------- (Doing XDP_DROP on CPU) Running XDP/eBPF prog_name:xdp_cpu_map0 XDP-cpumap CPU:to pps drop-pps extra-info XDP-RX 4 9,189,700 0 0 XDP-RX total 9,189,700 0 cpumap-enqueue 4:4 9,189,696 0 8.00 bulk-average cpumap-enqueue sum:4 9,189,696 0 8.00 bulk-average cpumap_kthread 4 9,189,702 0 143,582 sched cpumap_kthread total 9,189,702 0 143,582 sched-sum redirect_err total 0 0 xdp_exception total 0 0 2nd remote XDP/eBPF prog_name: xdp1 XDP-cpumap CPU:to xdp-pass xdp-drop xdp-redir xdp-in-kthread 4 0 9,189,702 0 xdp-in-kthread total 0 9,189,702 0 %CPU 51,8 ksoftirqd/4 48,2 cpumap/4/map:17 (Doing XDP_PASS on CPU) Running XDP/eBPF prog_name:xdp_cpu_map0 XDP-cpumap CPU:to pps drop-pps extra-info XDP-RX 4 8,593,822 0 0 XDP-RX total 8,593,822 0 cpumap-enqueue 4:4 8,593,888 7,714,949 8.00 bulk-average cpumap-enqueue sum:4 8,593,888 7,714,949 8.00 bulk-average cpumap_kthread 4 878,930 0 13,732 sched cpumap_kthread total 878,930 0 13,732 sched-sum redirect_err total 0 0 xdp_exception total 0 0 2nd remote XDP/eBPF prog_name: xdp_redirect_dummy XDP-cpumap CPU:to xdp-pass xdp-drop xdp-redir xdp-in-kthread 4 878,931 0 0 xdp-in-kthread total 878,931 0 0 Another CPU getting cpumap redirected packets: ---------------------------------------------- (Doing XDP_DROP on CPU) Running XDP/eBPF prog_name:xdp_cpu_map0 XDP-cpumap CPU:to pps drop-pps extra-info XDP-RX 4 17,526,797 0 0 XDP-RX total 17,526,797 0 cpumap-enqueue 4:0 17,526,796 245,811 8.00 bulk-average cpumap-enqueue sum:0 17,526,796 245,811 8.00 bulk-average cpumap_kthread 0 17,281,001 0 16,351 sched cpumap_kthread total 17,281,001 0 16,351 sched-sum redirect_err total 0 0 xdp_exception total 0 0 2nd remote XDP/eBPF prog_name: xdp1 XDP-cpumap CPU:to xdp-pass xdp-drop xdp-redir xdp-in-kthread 0 0 17,281,001 0 xdp-in-kthread total 0 17,281,001 0 (Doing XDP_PASS on CPU) Running XDP/eBPF prog_name:xdp_cpu_map0 XDP-cpumap CPU:to pps drop-pps extra-info XDP-RX 4 14,603,587 0 0 XDP-RX total 14,603,587 0 cpumap-enqueue 4:0 14,603,582 12,999,248 8.00 bulk-average cpumap-enqueue sum:0 14,603,582 12,999,248 8.00 bulk-average cpumap_kthread 0 1,604,338 0 0 cpumap_kthread total 1,604,338 0 0 redirect_err total 0 0 xdp_exception total 0 0 2nd remote XDP/eBPF prog_name: xdp_redirect_dummy XDP-cpumap CPU:to xdp-pass xdp-drop xdp-redir xdp-in-kthread 0 1,604,338 0 0 xdp-in-kthread total 1,604,338 0 0