USB: EHCI: add a delay when unlinking an active QH

Michael Reutman reports that an AMD/ATI EHCI host controller on one of
his computers does not stop transferring data when an active bulk QH
is unlinked from the async schedule.  Apparently that host controller
fails to implement the IAA mechanism correctly when an active QH is
unlinked.  This leads to data corruption, because the controller
continues to update the QH in memory when the driver doesn't expect
it.  As a result, the next URB submitted for that QH can hang, because
the link pointers for the TD queue have been messed up.  This
misbehavior is observed quite regularly.

To be fair, the EHCI spec (section 4.8.2) says that active QHs should
not be unlinked.  It goes on to recommend a procedure that involves
waiting for the QH to go inactive before unlinking it.  In the real
world this is impractical, not least because the QH may _never_ go
inactive.  (What were they thinking?)  Sometimes we have no choice but
to unlink an active QH.

In an attempt to avoid the problems that can ensue, this patch changes
how the driver decides when the unlink is complete.  In addition to
waiting through two IAA cycles, in cases where the QH was not known to
be inactive beforehand we now wait until a 2-ms period has elapsed
with the host controller making no change to the QH data structure
(the hw_current and hw_token fields in particular).  The intuition
here is that after such a long period, the endpoint must be NAKing and
hopefully the QH has been dropped from the host controller's internal
cache.  There's no way to know if this reasoning is really valid --
the spec is no help in this regard -- but at least this approach fixes
Michael's problem.

The test for whether the QH is already known to be inactive involves
the reason for unlinking the QH originally.  If it was unlinked
because it had halted, or it stopped in response to a short read, or
it overlaid a dummy TD (a silicon bug), then it certainly is inactive.
If it was unlinked because the TD queue was empty and no TDs have been
added to the queue in the meantime, then it must be inactive.  Or if
the hardware status indicates that the QH is currently halted (even if
that wasn't the reason for unlinking it), then it is inactive.
Otherwise, if none of those checks apply, we go through the 2-ms
delay.

Signed-off-by: Alan Stern <stern@rowland.harvard.edu>
Reported-by: Michael Reutman <mreutman@epiqsolutions.com>
Tested-by: Michael Reutman <mreutman@epiqsolutions.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>

1 files changed, 51 insertions, 5 deletions

diff --git a/drivers/usb/host/ehci-q.c b/drivers/usb/host/ehci-q.c
index 15dcae8..a24341e 100644
--- a/drivers/usb/host/ehci-q.c
+++ b/drivers/usb/host/ehci-q.c
@@ -1341,14 +1341,60 @@ static void end_unlink_async(struct ehci_hcd *ehci)
 	 * after the IAA interrupt occurs.  In self-defense, always go
 	 * through two IAA cycles for each QH.
 	 */
-	else if (qh->qh_state == QH_STATE_UNLINK_WAIT) {
+	else if (qh->qh_state == QH_STATE_UNLINK) {
+		/*
+		 * Second IAA cycle has finished.  Process only the first
+		 * waiting QH (NVIDIA (?) bug).
+		 */
+		list_move_tail(&qh->unlink_node, &ehci->async_idle);
+	}
+
+	/*
+	 * AMD/ATI (?) bug: The HC can continue to use an active QH long
+	 * after the IAA interrupt occurs.  To prevent problems, QHs that
+	 * may still be active will wait until 2 ms have passed with no
+	 * change to the hw_current and hw_token fields (this delay occurs
+	 * between the two IAA cycles).
+	 *
+	 * The EHCI spec (4.8.2) says that active QHs must not be removed
+	 * from the async schedule and recommends waiting until the QH
+	 * goes inactive.  This is ridiculous because the QH will _never_
+	 * become inactive if the endpoint NAKs indefinitely.
+	 */
+
+	/* Some reasons for unlinking guarantee the QH can't be active */
+	else if (qh->unlink_reason & (QH_UNLINK_HALTED |
+			QH_UNLINK_SHORT_READ | QH_UNLINK_DUMMY_OVERLAY))
+		goto DelayDone;
+
+	/* The QH can't be active if the queue was and still is empty... */
+	else if	((qh->unlink_reason & QH_UNLINK_QUEUE_EMPTY) &&
+			list_empty(&qh->qtd_list))
+		goto DelayDone;
+
+	/* ... or if the QH has halted */
+	else if	(qh->hw->hw_token & cpu_to_hc32(ehci, QTD_STS_HALT))
+		goto DelayDone;
+
+	/* Otherwise we have to wait until the QH stops changing */
+	else {
+		__hc32		qh_current, qh_token;
+
+		qh_current = qh->hw->hw_current;
+		qh_token = qh->hw->hw_token;
+		if (qh_current != ehci->old_current ||
+				qh_token != ehci->old_token) {
+			ehci->old_current = qh_current;
+			ehci->old_token = qh_token;
+			ehci_enable_event(ehci,
+					EHCI_HRTIMER_ACTIVE_UNLINK, true);
+			return;
+		}
+ DelayDone:
 		qh->qh_state = QH_STATE_UNLINK;
 		early_exit = true;
 	}
-
-	/* Otherwise process only the first waiting QH (NVIDIA bug?) */
-	else
-		list_move_tail(&qh->unlink_node, &ehci->async_idle);
+	ehci->old_current = ~0;		/* Prepare for next QH */
 
 	/* Start a new IAA cycle if any QHs are waiting for it */
 	if (!list_empty(&ehci->async_unlink))