On Fri, Jun 14, 2024 at 10:28 AM Sebastian Andrzej Siewior
[off-list ref] wrote:

On 2024-06-12 13:18:29 [-0400], Steven Rostedt wrote:

quoted

On Wed, 12 Jun 2024 18:44:34 +0200
Sebastian Andrzej Siewior [off-list ref] wrote:

quoted

Softirq is preemptible on PREEMPT_RT. Without a per-CPU lock in
local_bh_disable() there is no guarantee that only one device is
transmitting at a time.
With preemption and multiple senders it is possible that the per-CPU
recursion counter gets incremented by different threads and exceeds
XMIT_RECURSION_LIMIT leading to a false positive recursion alert.

Instead of adding a lock to protect the per-CPU variable it is simpler
to make the counter per-task. Sending and receiving skbs happens always
in thread context anyway.

Having a lock to protected the per-CPU counter would block/ serialize two
sending threads needlessly. It would also require a recursive lock to
ensure that the owner can increment the counter further.

Make the recursion counter a task_struct member on PREEMPT_RT.

I'm curious to what would be the harm to using a per_task counter
instead of per_cpu outside of PREEMPT_RT. That way, we wouldn't have to
have the #ifdef.

There should be a hole on !RT, too so we shouldn't gain weight. The
limit is set to 8 so an u8 would be enough. The counter is only accessed
with BH-disabled so it will be used only in one context since it can't
schedule().

I think it should work fine. netdev folks, you want me to remove that
ifdef and use a per-Task counter unconditionally?

It depends if this adds another cache line miss/dirtying or not.

What about other fields from softnet_data.xmit ?