xhci: Fix failed enqueue in the middle of isoch TD.

xhci: Fix failed enqueue in the middle of isoch TD.

When an isochronous transfer is enqueued, xhci_queue_isoc_tx_prepare()
will ensure that there is enough room on the transfer rings for all of the
isochronous TDs for that URB.  However, when xhci_queue_isoc_tx() is
enqueueing individual isoc TDs, the prepare_transfer() function can fail
if the endpoint state has changed to disabled, error, or some other
unknown state.

With the current code, if Nth TD (not the first TD) fails, the ring is
left in a sorry state.  The partially enqueued TDs are left on the ring,
and the first TRB of the TD is not given back to the hardware.  The
enqueue pointer is left on the TRB after the last successfully enqueued
TD.  This means the ring is basically useless.  Any new transfers will be
enqueued after the failed TDs, which the hardware will never read because
the cycle bit indicates it does not own them.  The ring will fill up with
untransferred TDs, and the endpoint will be basically unusable.

The untransferred TDs will also remain on the TD list.  Since the td_list
is a FIFO, this basically means the ring handler will be waiting on TDs
that will never be completed (or worse, dereference memory that doesn't
exist any more).

Change the code to clean up the isochronous ring after a failed transfer.
If the first TD failed, simply return and allow the xhci_urb_enqueue
function to free the urb_priv.  If the Nth TD failed, first remove the TDs
from the td_list.  Then convert the TRBs that were enqueued into No-op
TRBs.  Make sure to flip the cycle bit on all enqueued TRBs (including any
link TRBs in the middle or between TDs), but leave the cycle bit of the
first TRB (which will show software-owned) intact.  Then move the ring
enqueue pointer back to the first TRB and make sure to change the
xhci_ring's cycle state to what is appropriate for that ring segment.

This ensures that the No-op TRBs will be overwritten by subsequent TDs,
and the hardware will not start executing random TRBs because the cycle
bit was left as hardware-owned.

This bug is unlikely to be hit, but it was something I noticed while
tracking down the watchdog timer issue.  I verified that the fix works by
injecting some errors on the 250th isochronous URB queued, although I
could not verify that the ring is in the correct state because uvcvideo
refused to talk to the device after the first usb_submit_urb() failed.
Ring debugging shows that the ring looks correct, however.

This patch should be backported to kernels as old as 2.6.36.

Signed-off-by: Sarah Sharp <sarah.a.sharp@linux.intel.com>
Cc: Andiry Xu <andiry.xu@amd.com>
Cc: stable@kernel.org