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>
This commit is contained in:
Alan Stern 2008-05-20 16:59:33 -04:00 committed by Greg Kroah-Hartman
parent b40e43fcc5
commit fa38dfcc56
2 changed files with 10 additions and 8 deletions

View file

@ -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

View file

@ -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 @@ rescan:
* 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;