Hello, Peter. On Thu, Oct 01, 2026 at 12:59:09PM +0200, Peter Zijlstra wrote: > Ideally RX would be split for each network queue, rather than lumped > into the one giant softirq that nobody owns.
Rx can be split per queue but the granularity doesn't match. There can be way more cgroups than the hardware can split queues for. > Furthermore, without ingress throttling, your RX back charge could > completely deplete the actual cgroup time quota. If rx processing slows down, the rx traffic slows down with it. We'd likely still want to allot some minimum amount of time to non-rx work while making sure the debt doesn't spiral out of control. If a dedicated kthread were handling rx for the cgroup, rx would be sharing CPU time with the rest of the cgroup without choking it completely. The back charging can follow the same corollary. > Anyway, if you get per queue RX processing threads, then you can move > them into cgroups where so desired. Same limit here. There are setups with a lot more cgroups than CPUs. Thanks. -- tejun
