On Wed, 4 Nov 2015, Doug Anderson wrote: > In the ChromeOS gerrit > <https://chromium-review.googlesource.com/#/c/310583/> Julius Werner > points out that for EHCI it was good to take the optimization from > commit 9118f9eb4f1e ("USB: EHCI: improve interrupt qh unlink") before > this one. I'm still trying to learn USB / dwc2 so it's unclear to me > whether we also need a similar change before landing. > > I'll see if I can do some investigation about this and also some > benchmarking before and after. Certainly profiling the interrupt > handler itself showed a huge improvement, but I'd hate to see a > regression elsewhere. > > If anyone else knows better than I, please speak up! :) This is a matter of both efficiency and correctness. Giving back URBs in a tasklet is not a simple change. Have you read the kerneldoc for usb_submit_urb() in drivers/usb/core/urb.c? The portion about "Reserved Bandwidth Transfers" is highly relevant. I don't know how dwc2 goes about reserving bandwidth for periodic transfers, but if it relies on the endpoint queue being non-empty to maintain a reservation then it will be affected by this change. Alan Stern
