On 27.8.2024 15.18, FPS wrote:
On 8/27/24 1:31 PM, Mathias Nyman wrote:
Should be harmless, the cycle bit and capital 'C' changes each time the
ringbuffer wraps around.
This is how TRB consumers/producers keep track of where we are in the ring.
segs 1 vs 2 just tells that we have allocated 2 segments for *Intel host
event ringbuffer while only one for Renesas.
OK, thanks for that explanation. I uploaded the full Intel controller
trace here (curl'able link):
https://uni-bielefeld.sciebo.de/s/0O4XIzW529sKYQM/download
And here is the Renesas trace:
https://uni-bielefeld.sciebo.de/s/jB4qqFL0okPlYwN/download
Another difference which I find maybe more interesting then. If you
scroll down to where the steady state has been reached, e.g. timestamp
119173.008004 for the Intel trace and timestamp 564371.959089 for the
Renesas trace, then there are always 8 xhci_handle_transfer calls for
TDs of size 48 and 8 queue_trb calls between doorbell rings for the
Renesas controller, but for the Intel controller it looks different:
There are also always 8 xhci_queue_trb calls, but either 7 or 9
xhci_handle_transfer calls. This is quite odd, no?
The 9th event means that xHC hardware managed to handle one more isoc transfer
before the xhci interrupt handler was called and started handling events.
xhci interrupt handler handles all pending events on the event ring, but an
actual interrupt is generated only on last (8th) TD.
xHC completes 1st TD of URB X, writes event on event ring
125us passes
xHC completes 2nd TD of URB X, writes event on event ring
... (same for all TDs 3-7)
xHC completes 8th TD (last in URB X), writes event on event ring, triggers interrupt.
125us passes
xHC completes 1st TD of URB Y, writes event on event ring.
xhci interrupt handler is finally running, processing all 9 pending transfers,
it will give back URB X after processing its last (8th) event, but continue
processing the first TD of URB Y
So there is an at least 125us delay between xHC completing 8th TD, triggering
an interrupt and interrupt handler actually running, finishing processing events.
Thanks
Mathias
