USB: EHCI: go back to using the system clock for QH unlinks
authorAlan Stern <stern@rowland.harvard.edu>
Tue, 5 Jul 2011 16:34:05 +0000 (12:34 -0400)
committerGreg Kroah-Hartman <gregkh@suse.de>
Fri, 8 Jul 2011 21:55:08 +0000 (14:55 -0700)
This patch (as1477) fixes a problem affecting a few types of EHCI
controller.  Contrary to what one might expect, these controllers
automatically stop their internal frame counter when no ports are
enabled.  Since ehci-hcd currently relies on the frame counter for
determining when it should unlink QHs from the async schedule, those
controllers run into trouble: The frame counter stops and the QHs
never get unlinked.

Some systems have also experienced other problems traced back to
commit b963801164618e25fbdc0cd452ce49c3628b46c8 (USB: ehci-hcd unlink
speedups), which made the original switch from using the system clock
to using the frame counter.  It never became clear what the reason was
for these problems, but evidently it is related to use of the frame
counter.

To fix all these problems, this patch more or less reverts that commit
and goes back to using the system clock.  But this can't be done
cleanly because other changes have since been made to the scan_async()
subroutine.  One of these changes involved the tricky logic that tries
to avoid rescanning QHs that have already been seen when the scanning
loop is restarted, which happens whenever an URB is given back.
Switching back to clock-based unlinks would make this logic even more
complicated.

Therefore the new code doesn't rescan the entire async list whenever a
giveback occurs.  Instead it rescans only the current QH and continues
on from there.  This requires the use of a separate pointer to keep
track of the next QH to scan, since the current QH may be unlinked
while the scanning is in progress.  That new pointer must be global,
so that it can be adjusted forward whenever the _next_ QH gets
unlinked.  (uhci-hcd uses this same trick.)

Simplification of the scanning loop removes a level of indentation,
which accounts for the size of the patch.  The amount of code changed
is relatively small, and it isn't exactly a reversion of the
b963801164 commit.

This fixes Bugzilla #32432.

Signed-off-by: Alan Stern <stern@rowland.harvard.edu>
CC: <stable@kernel.org>
Tested-by: Matej Kenda <matejken@gmail.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@suse.de>
drivers/usb/host/ehci-hcd.c
drivers/usb/host/ehci-q.c
drivers/usb/host/ehci.h

index 4ee62bef372375bd4088fd4592912d232c64229a..2902199fa8ff910ee9f3bba997078a9908feabe4 100644 (file)
@@ -90,7 +90,8 @@ 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_FRAMES     5               /* async qh unlink delay */
+#define EHCI_SHRINK_JIFFIES    (DIV_ROUND_UP(HZ, 200) + 1)
+                                               /* 200-ms async qh unlink delay */
 
 /* Initial IRQ latency:  faster than hw default */
 static int log2_irq_thresh = 0;                // 0 to 6
@@ -148,10 +149,7 @@ timer_action(struct ehci_hcd *ehci, enum ehci_timer_action action)
                        break;
                /* case TIMER_ASYNC_SHRINK: */
                default:
-                       /* add a jiffie since we synch against the
-                        * 8 KHz uframe counter.
-                        */
-                       t = DIV_ROUND_UP(EHCI_SHRINK_FRAMES * HZ, 1000) + 1;
+                       t = EHCI_SHRINK_JIFFIES;
                        break;
                }
                mod_timer(&ehci->watchdog, t + jiffies);
index 5d6bc624c961ae96b6f76447d04b432a948e9de9..9bf3c0d983c494b4ae0cf00fee622eb1b6b72f24 100644 (file)
@@ -1231,6 +1231,8 @@ static void start_unlink_async (struct ehci_hcd *ehci, struct ehci_qh *qh)
 
        prev->hw->hw_next = qh->hw->hw_next;
        prev->qh_next = qh->qh_next;
+       if (ehci->qh_scan_next == qh)
+               ehci->qh_scan_next = qh->qh_next.qh;
        wmb ();
 
        /* If the controller isn't running, we don't have to wait for it */
@@ -1256,53 +1258,49 @@ static void scan_async (struct ehci_hcd *ehci)
        struct ehci_qh          *qh;
        enum ehci_timer_action  action = TIMER_IO_WATCHDOG;
 
-       ehci->stamp = ehci_readl(ehci, &ehci->regs->frame_index);
        timer_action_done (ehci, TIMER_ASYNC_SHRINK);
-rescan:
        stopped = !HC_IS_RUNNING(ehci_to_hcd(ehci)->state);
-       qh = ehci->async->qh_next.qh;
-       if (likely (qh != NULL)) {
-               do {
-                       /* clean any finished work for this qh */
-                       if (!list_empty(&qh->qtd_list) && (stopped ||
-                                       qh->stamp != ehci->stamp)) {
-                               int temp;
-
-                               /* unlinks could happen here; completion
-                                * reporting drops the lock.  rescan using
-                                * the latest schedule, but don't rescan
-                                * qhs we already finished (no looping)
-                                * unless the controller is stopped.
-                                */
-                               qh = qh_get (qh);
-                               qh->stamp = ehci->stamp;
-                               temp = qh_completions (ehci, qh);
-                               if (qh->needs_rescan)
-                                       unlink_async(ehci, qh);
-                               qh_put (qh);
-                               if (temp != 0) {
-                                       goto rescan;
-                               }
-                       }
 
-                       /* unlink idle entries, reducing DMA usage as well
-                        * as HCD schedule-scanning costs.  delay for any qh
-                        * we just scanned, there's a not-unusual case that it
-                        * doesn't stay idle for long.
-                        * (plus, avoids some kind of re-activation race.)
+       ehci->qh_scan_next = ehci->async->qh_next.qh;
+       while (ehci->qh_scan_next) {
+               qh = ehci->qh_scan_next;
+               ehci->qh_scan_next = qh->qh_next.qh;
+ rescan:
+               /* clean any finished work for this qh */
+               if (!list_empty(&qh->qtd_list)) {
+                       int temp;
+
+                       /*
+                        * Unlinks could happen here; completion reporting
+                        * drops the lock.  That's why ehci->qh_scan_next
+                        * always holds the next qh to scan; if the next qh
+                        * gets unlinked then ehci->qh_scan_next is adjusted
+                        * in start_unlink_async().
                         */
-                       if (list_empty(&qh->qtd_list)
-                                       && qh->qh_state == QH_STATE_LINKED) {
-                               if (!ehci->reclaim && (stopped ||
-                                       ((ehci->stamp - qh->stamp) & 0x1fff)
-                                               >= EHCI_SHRINK_FRAMES * 8))
-                                       start_unlink_async(ehci, qh);
-                               else
-                                       action = TIMER_ASYNC_SHRINK;
-                       }
+                       qh = qh_get(qh);
+                       temp = qh_completions(ehci, qh);
+                       if (qh->needs_rescan)
+                               unlink_async(ehci, qh);
+                       qh->unlink_time = jiffies + EHCI_SHRINK_JIFFIES;
+                       qh_put(qh);
+                       if (temp != 0)
+                               goto rescan;
+               }
 
-                       qh = qh->qh_next.qh;
-               } while (qh);
+               /* unlink idle entries, reducing DMA usage as well
+                * as HCD schedule-scanning costs.  delay for any qh
+                * we just scanned, there's a not-unusual case that it
+                * doesn't stay idle for long.
+                * (plus, avoids some kind of re-activation race.)
+                */
+               if (list_empty(&qh->qtd_list)
+                               && qh->qh_state == QH_STATE_LINKED) {
+                       if (!ehci->reclaim && (stopped ||
+                                       time_after_eq(jiffies, qh->unlink_time)))
+                               start_unlink_async(ehci, qh);
+                       else
+                               action = TIMER_ASYNC_SHRINK;
+               }
        }
        if (action == TIMER_ASYNC_SHRINK)
                timer_action (ehci, TIMER_ASYNC_SHRINK);
index fa3129fe1ee08383cc3ed1cdce73161f2ee8e7fe..e4feec3457fb3b4300ad27164dffcca469f53438 100644 (file)
@@ -75,6 +75,7 @@ struct ehci_hcd {                     /* one per controller */
        struct ehci_qh          *async;
        struct ehci_qh          *dummy;         /* For AMD quirk use */
        struct ehci_qh          *reclaim;
+       struct ehci_qh          *qh_scan_next;
        unsigned                scanning : 1;
 
        /* periodic schedule support */
@@ -119,7 +120,6 @@ struct ehci_hcd {                   /* one per controller */
        struct timer_list       iaa_watchdog;
        struct timer_list       watchdog;
        unsigned long           actions;
-       unsigned                stamp;
        unsigned                periodic_stamp;
        unsigned                random_frame;
        unsigned long           next_statechange;
@@ -345,6 +345,7 @@ struct ehci_qh {
        struct ehci_qh          *reclaim;       /* next to reclaim */
 
        struct ehci_hcd         *ehci;
+       unsigned long           unlink_time;
 
        /*
         * Do NOT use atomic operations for QH refcounting. On some CPUs