This would be rare with HCDs that maintain chains of DMA
transfers, except if the HC dies in the middle of an I/O
request; so no rush to merge this.  It'd happen in a PIO
based HCD though ... :)


Async unlink of an URB from an endpoint's I/O queue _normally_ involves a
delay from handshaking with the host controller, to be sure the DMA queue
is inactive.  So urb->complete() runs after usb_unlink_urb() returns, and
from a different context.  But not always...

The completion may run immediately whenever the HCD knows that HC isn't
busy with the URB.  Maybe that HCD is in a HALT state, or the endpoint
queue is is temporarily off-schedule (halted, or dead after PM resume
from D3cold, etc) ... or maybe the HCD doesn't use DMA, so most unlinks
just list_del_init() and return.

This makes usb_sg_cancel() and sg_complete() drop the io->lock when they
cancel active urbs, preventing potential self-deadlock when that completion
handler runs immediately.
Signed-off-by: David Brownell <dbrownell@users.sourceforge.net>
Signed-off-by: Greg Kroah-Hartman <greg@kroah.com>