[Question] Batched Tx in MCTP USB Binding

[santoshpuranik]

Hi Jeremy,

Starting this to discuss possible implementations that I am looking at to enable packing multiple MCTP packets into a single USB OUT transfer in the MCTP USB binding.

What are your thoughts on:

• The packing strategy: Should the binding layer that reads from the netdev queue "wait" a certain (a few microseconds?) time to possibly pool multiple packets in a transfer? I am not sure if that is a good idea given we don't know that there are more packets in the queue.
• The actual implementation: This kind of depends on the above question, but if we choose not to implement the "wait for more packets" strategy, does it make sense for the route.c implementation (do_fragment_route) that pushes packets to the netdev queue to set a hint via netdev_xmit_more() flag to indicate to the binding that more packets are on the way? The binding may then use this hint to defer the URB submission until netdev_xmit_more() is set OR we would exceed the max USB bulk xfer limit.
• The binding may just look at the EOM field in the header to know that there must be more packets on the way and wait like above until EOM is received OR max packet size would exceed the USB transfer limit.

Any quick thoughts on which of these (or something else entirely!) might be the right approach?

Thanks,
Santosh

[santoshpuranik]

Another (hackier) approach may be to form a large enough SKB with space reserved for the USB packets headers right in route.c and have the binding layer fill in the binding headers.

[jk-ozlabs]

I already have an implementation for this, using the xmit_more flag, and changing the dst output API to now accept a list of skbs. Seems to be working well in testing, I'll get a branch out shortly.

[santoshpuranik]

I already have an implementation for this, using the xmit_more flag, and changing the dst output API to now accept a list of skbs. Seems to be working well in testing, I'll get a branch out shortly.

Oh nice! I can help test that with multiple USB devices too. What is the strategy used there? We only batch when we have enough skbs to send as a list, right? That is, say when an application makes a large transfer that needs multiple packets and not wait till we get a certain number of packets (or timeout) before pooling.

[jk-ozlabs]

Test implementation is in our test/mctp-usb branch: https://github.com/CodeConstruct/linux/commits/test/mctp-usb/

[santoshpuranik]

Thank you! I'll test this out.