Re: [PATCH v2 net-next 5/5] mptcp: implement delegated actions

On Fri, 22 Jan 2021 09:25:07 +0100 Paolo Abeni wrote:

On Thu, 2021-01-21 at 17:34 -0800, Jakub Kicinski wrote:

quoted

On Wed, 20 Jan 2021 15:39:14 +0100 Paolo Abeni wrote:  

quoted

On MPTCP-level ack reception, the packet scheduler
may select a subflow other then the current one.

Prior to this commit we rely on the workqueue to trigger
action on such subflow.

This changeset introduces an infrastructure that allows
any MPTCP subflow to schedule actions (MPTCP xmit) on
others subflows without resorting to (multiple) process
reschedule.  

If your work doesn't reschedule there should not be multiple 
rescheds, no?  

Thank you for looking into this.

With the workqueue it would be:

<running process> -> BH -> (process scheduler) -> MPTCP workqueue ->
(process scheduler) -> <running process>

With this infra is:

<running process> -> BH -> BH -> <running process>

You assume your process is running when you got the Rx IRQ but also
is RPS enabled? For MPTCP there is no way to do reuseport to match Rx
CPU to processing thread, right?

quoted

quoted

A dummy NAPI instance is used instead. When MPTCP needs to
trigger action an a different subflow, it enqueues the target
subflow on the NAPI backlog and schedule such instance as needed.

The dummy NAPI poll method walks the sockets backlog and tries
to acquire the (BH) socket lock on each of them. If the socket
is owned by the user space, the action will be completed by
the sock release cb, otherwise push is started.

This change leverages the delegated action infrastructure
to avoid invoking the MPTCP worker to spool the pending data,
when the packet scheduler picks a subflow other then the one
currently processing the incoming MPTCP-level ack.

Additionally we further refine the subflow selection
invoking the packet scheduler for each chunk of data
even inside __mptcp_subflow_push_pending().  

Is there much precedence for this sort of hijacking of NAPI 
for protocol work?   

AFAICS, xfrm is using a similar trick in the receive path.

Note that we uses TX-only NAPIs, so this does not pollute the napi hash
table.

Ack, I think what you have should work with busy polling and threaded
IRQs as well. Just another piece of the puzzle to keep in one's head
when thinking about NAPI processing :)

quoted

Do you need it because of locking?  

This infrastructure is used to avoid the workqueue usage in the MPTCP
receive path (to push pending data). With many mptcp connections
established that would be very bad for tput and latency. This
infrastructure is not strictly needed from a functinal PoV, but I was
unable to find any other way to avoid the workqueue usage.

But it is due to locking or is it not? Because you're running the
callback in the same context, so otherwise why not just call the
function directly? Can't be batching, it's after GRO so we won't 
batch much more.

Please let me know if the above is clear enough!

Sure thing, it just looks like interesting work. I'll apply end 
of the day if nobody has any comments.