Hi Aaron,
On Thu, Jul 11, 2019 at 12:50 PM Aaron Williams awilliams@marvell.com wrote:
Hi All,
I'm running into a problem with the XHCI driver on our SOC where I am getting spurious success TRBs after a short packet response. In this particular case, the driver for a USB network controller has a receive buffer that crosses a 64K boundary. This results in two TRBs being created to handle this with the first TRB having the chain and ISP bits set and the second having the IOC and ISP bits set as expected in this case.
The network controller has a short response (as is expected) followed by a success response event (which is also valid) for the second TRB. The U-Boot XHCI driver, however, does not handle this properly. The xhci_bulk_tx function only looks at the first response and does not dequeue the second success response. The problem is when the next transaction occurs it dequeues the success response from the previous transaction so the responses events no longer match the TRBs. At some point the network driver goes from receive to transmit and at this point the driver panics because the event is still a response from a previous receive event and not the transmit event causing the endpoints to mismatch. This only happens in the two TRB case due to the receive buffer crossing a 64K boundary. The short packet response points to the first TRB and the success points to the second TRB.
In the Linux driver I see the XHCI_SPURIOUS_SUCCESS flag is available. I'm not certain if this applies to our hardware or not. In any case, this could be solved in one of two ways. One way to solve this is to make sure that any driver where a short response is possible to not have the buffer cross a 64K boundary. The other (and correct) solution is to have the XHCI driver handle this case. If a short packet response is received and there are multiple TRBs involved, check if there is an additional success response that points to the last TRB.
To me it seems valid that the two response events are received. The specification only says that there will be one response if the TRB has both the ISP and IOC bits set, but this is not the case. I came across this commit in the Linux kernel which seems to address this case: https://kernel.opensuse.org/cgit/kernel/commit/?id=ad808333d8201d53075a11bc8...
Are you using an Intel SoC? From the commit you gave, it looks it only applies to a specific Intel PCH chipset.
Regards, Bin