In the existing implementation, multiple requests queued up on an
endpoint are subject to getting evicted without transmission.
For both control and bulk endpoints, their respective logic found in
usba_control_irq()/usba_ep_irq() guarantees that TX FIFO is empty before
data is sent out, and that request_complete() gets called once the
transaction has been finished. At this point however, if any additional
requests are found on the endpoint queue, they will be processed by
submit_next_request(), which makes no checks against the above
conditions, trashing data on a busy FIFO and neglecting completion
handlers.
Fix the above issues by removing the calls to submit_next_request(),
and thus forcing the pending requests to be processed on the next pass
of the respective endpoint logic. While at it, remove a DBG message, as
that branch becomes part of regular flow.
This restores mass storage mode operation on Microchip ATSAMA5D27 SoC.
Signed-off-by: Artur Rojek <artur@conclusive.pl>
bc8e94c902