| Message ID | 20240325074943.289909-1-bigeasy@linutronix.de (mailing list archive) |
|---|---|
| Headers | show |
| Series | net: Provide SMP threads for backlog NAPI | expand |
Hello: This series was applied to netdev/net-next.git (main) by Paolo Abeni <pabeni@redhat.com>: On Mon, 25 Mar 2024 08:40:27 +0100 you wrote: > The RPS code and "deferred skb free" both send IPI/ function call > to a remote CPU in which a softirq is raised. This leads to a warning on > PREEMPT_RT because raising softiqrs from function call led to undesired > behaviour in the past. I had duct tape in RT for the "deferred skb free" > and Wander Lairson Costa reported the RPS case. > > This series only provides support for SMP threads for backlog NAPI, I > did not attach a patch to make it default and remove the IPI related > code to avoid confusion. I can post it for reference it asked. > > [...] Here is the summary with links: - [v6,net-next,1/4] net: Remove conditional threaded-NAPI wakeup based on task state. https://git.kernel.org/netdev/net-next/c/56364c910691 - [v6,net-next,2/4] net: Allow to use SMP threads for backlog NAPI. https://git.kernel.org/netdev/net-next/c/dad6b9770263 - [v6,net-next,3/4] net: Use backlog-NAPI to clean up the defer_list. https://git.kernel.org/netdev/net-next/c/80d2eefcb4c8 - [v6,net-next,4/4] net: Rename rps_lock to backlog_lock. https://git.kernel.org/netdev/net-next/c/765b11f8f4e2 You are awesome, thank you!
The RPS code and "deferred skb free" both send IPI/ function call to a remote CPU in which a softirq is raised. This leads to a warning on PREEMPT_RT because raising softiqrs from function call led to undesired behaviour in the past. I had duct tape in RT for the "deferred skb free" and Wander Lairson Costa reported the RPS case. This series only provides support for SMP threads for backlog NAPI, I did not attach a patch to make it default and remove the IPI related code to avoid confusion. I can post it for reference it asked. The RedHat performance team was so kind to provide some testing here. The series (with the IPI code removed) has been tested and no regression vs without the series has been found. For testing iperf3 was used on 25G interface, provided by mlx5, ix40e or ice driver and RPS was enabled. I can provide the individual test results if needed. Changes: - v5…v6 https://lore.kernel.org/all/20240309090824.2956805-1-bigeasy@linutronix.de/ - Rebase on top of current net-next. - v4…v5 https://lore.kernel.org/all/20240305120002.1499223-1-bigeasy@linutronix.de/ - Rebase on top of current net-next. - v3…v4 https://lore.kernel.org/all/20240228121000.526645-1-bigeasy@linutronix.de/ - Rebase on top of current net-next, collect Acks. - Add struct softnet_data as an argument to kick_defer_list_purge(). - Add sd_has_rps_ipi_waiting() check to napi_threaded_poll_loop() which was accidentally removed. - v2…v3 https://lore.kernel.org/all/20240221172032.78737-1-bigeasy@linutronix.de/ - Move the "if use_backlog_threads()" case into the CONFIG_RPS block within napi_schedule_rps(). - Use __napi_schedule_irqoff() instead of napi_schedule_rps() in kick_defer_list_purge(). - v1…v2 https://lore.kernel.org/all/20230929162121.1822900-1-bigeasy@linutronix.de/ - Patch #1 is new. It ensures that NAPI_STATE_SCHED_THREADED is always set (instead conditional based on task state) and the smboot thread logic relies on this bit now. In v1 NAPI_STATE_SCHED was used but is racy. - The defer list clean up is split out and also relies on NAPI_STATE_SCHED_THREADED. This fixes a different race. - RFC…v1 https://lore.kernel.org/all/20230814093528.117342-1-bigeasy@linutronix.de/ - Patch #2 has been removed. Removing the warning is still an option. - There are two patches in the series: - Patch #1 always creates backlog threads - Patch #2 creates the backlog threads if requested at boot time, mandatory on PREEMPT_RT. So it is either or and I wanted to show how both look like. - The kernel test robot reported a performance regression with loopback (stress-ng --udp X --udp-ops Y) against the RFC version. The regression is now avoided by using local-NAPI if backlog processing is requested on the local CPU. Sebastian