USB: EHCI: fix performance regression

This patch (as1099) fixes a performance regression in ehci-hcd.  The
fundamental problem is that queue headers get removed from the
schedule too quickly, since the code checks for a counter advancing
rather than making an actual time-based check.  The latency involved
in removing the queue header and then relinking it can severely
degrade certain kinds of workloads.

The patch replaces a simple counter with a timestamp derived from the
controller's uframe value.  In addition, the delay for unlinking an
idle queue header is increased from 5 ms to 10 ms; since some
controllers (nVidia) have a latency of up to 1 ms for unlinking, this
reduces the relative impact from 20% to 10%.

Finally, a logical error left over from the IAA watchdog-timer
conversion is corrected.  Now the driver will always either unlink an
idle queue header or set up a timer to unlink it later.  The old code
would sometimes fail to do either.

Signed-off-by: Alan Stern <stern@rowland.harvard.edu>
Cc: David Brownell <david-b@pacbell.net>
Cc: Leonid <leonidv11@gmail.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@suse.de>

drivers/usb/host/ehci-hcd.c

@@ -84,7 +84,8 @@ static const char	hcd_name [] = "ehci_hcd";
 #define EHCI_IAA_MSECS		10		/* arbitrary */
 #define EHCI_IO_JIFFIES		(HZ/10)		/* io watchdog > irq_thresh */
 #define EHCI_ASYNC_JIFFIES	(HZ/20)		/* async idle timeout */
-#define EHCI_SHRINK_JIFFIES	(HZ/200)	/* async qh unlink delay */
+#define EHCI_SHRINK_JIFFIES	(HZ/100)	/* async qh unlink delay */
+#define EHCI_SHRINK_UFRAMES	(10*8)		/* same value in uframes */
 
 /* Initial IRQ latency:  faster than hw default */
 static int log2_irq_thresh = 0;		// 0 to 6

drivers/usb/host/ehci-q.c

@@ -1116,8 +1116,7 @@ static void scan_async (struct ehci_hcd *ehci)
 	struct ehci_qh		*qh;
 	enum ehci_timer_action	action = TIMER_IO_WATCHDOG;
 
-	if (!++(ehci->stamp))
-		ehci->stamp++;
+	ehci->stamp = ehci_readl(ehci, &ehci->regs->frame_index);
 	timer_action_done (ehci, TIMER_ASYNC_SHRINK);
 rescan:
 	qh = ehci->async->qh_next.qh;
@@ -1148,12 +1147,14 @@ static void scan_async (struct ehci_hcd *ehci)
 			 * doesn't stay idle for long.
 			 * (plus, avoids some kind of re-activation race.)
 			 */
-			if (list_empty (&qh->qtd_list)) {
-				if (qh->stamp == ehci->stamp)
+			if (list_empty(&qh->qtd_list) &&
+					qh->qh_state == QH_STATE_LINKED) {
+				if (!ehci->reclaim &&
+				    ((ehci->stamp - qh->stamp) & 8191) >=
+						EHCI_SHRINK_UFRAMES)
+					start_unlink_async(ehci, qh);
+				else
 					action = TIMER_ASYNC_SHRINK;
-				else if (!ehci->reclaim
-					    && qh->qh_state == QH_STATE_LINKED)
-					start_unlink_async (ehci, qh);
 			}
 
 			qh = qh->qh_next.qh;