In [0] I introduced explicit locking for resources which are otherwise locked implicit locked by local_bh_disable() and this protections goes away if the lock in local_bh_disable() is removed on PREEMPT_RT. There was a big complained once it came to the locking of XDP related resources during XDP-redirect because it was mostly per-packet and the locking, even not contended, was seen as a problem [1]. Another XDP related problem was that I also touched every NIC-driver using XDP. This complicated the "how to write a NIC driver" further. To solve both problems I was thinking about an alternative and ended up with moving the data structure from per-CPU to per-task on PREEMPT_RT. Instead of adding it to task_struct, I added a pointer there and setup the struct on stack. In my debug build on x86-64 the struct bpf_xdp_storage has 112 bytes and collected the per-CPU ressources. I've been benchmark the outcome on !RT. I meassured 14,880,077 pkt/s with pktgen + "xdp-bench drop" on a ixgbe which is the interface limit. I then lowered the clockspeed of the CPU (on the RX side) and received approx between 12,156,279 and 12,476,759 pkt/s. The results were not consistent and varied between runs mostly in that range with and without the patches. v1…v2 https://lore.kernel.org/all/20240213145923.2552753-1-bigeasy@xxxxxxxxxxxxx - Renamed the container struct from xdp_storage to bpf_net_context. Suggested by Toke Høiland-Jørgensen. - Use the container struct also on !PREEMPT_RT builds. Store the pointer to the on-stack struct in a per-CPU variable. Suggested by Toke Høiland-Jørgensen. I'm posting here only two patches which replace the XDP redirect part (patches 14 to 24) from the original series. [0] https://lore.kernel.org/all/20231215171020.687342-1-bigeasy@xxxxxxxxxxxxx/ [1] https://lore.kernel.org/all/CAADnVQKJBpvfyvmgM29FLv+KpLwBBRggXWzwKzaCT9U-4bgxjA@xxxxxxxxxxxxxx/ [2] https://lore.kernel.org/all/20231215145059.3b42ee35@xxxxxxxxxx Sebastian
