Hi Mathias, > Code on top of 6.11-rc4 can be found in my > feature_transfer_event_rework branch: > > git://git.kernel.org/pub/scm/linux/kernel/git/mnyman/xhci.git Thanks for pointing me to this. It does look like a very promising approach. This is the principle of "matching before handling" taken to the extreme, as nothing is changed during the first pass through the TD list. It seems to work well. I feel like the counting of distances from ring->dequeue is a little, I don't know, unexpected maybe, but it does indeed produce surprisingly compact and simple code thanks to those segment numbers. That being said, I've always been a strong case of NIH syndrome, so I implemented a similar search for passed TDs which relies solely on walking the ring in order from dequeue to the event DMA pointer. So I guess we will have two implementations to compare and choose from. I will try to develop this approach into a complete implementation that actually works and can be tested. I want to maximally preserve original behavior, except for obvious bugs (I found an obscure one already) to simplify validation and review. So, for example, no removing the skip flag for now. I'm not sure if it's a good idea at all, as it frankly is sweeping bugs under the rug. I have one HC which sometimes silently misses TDs for no known reason and it misses so many of them that auto-skipping wouldn't help anyway. OTOH, when I get the TRB errors, at least I know the HC is screwed. Regards, Michal
