Re: Regression: USB is nfg after suspend/resume(RAM) cycle on Intelchipset

[Alan Stern]

On Tue, 29 May 2007, Mark Lord wrote:

> Mark Lord wrote:
> >..
> > 7ed92f1a149dddc3cb537ccd7441e98adac12c3e USB: make the autosuspend 
> > workqueue thread freezable
> > ef7f6c7084b333c7524dcd297e0578d43733a2a2 USB: more autosuspend timer stuff
> > 
> > I yanked them both, as they appeared to be releated based on the titles.
> > Reverting this pair of commits fixes the USB suspend/resume regression.
> 
> Okay, just to make it trivial,
> I've narrowed it down to only this commit from Alan Stern:
> 
> 7ed92f1a149dddc3cb537ccd7441e98adac12c3e USB: make the autosuspend workqueue thread freezable

I'm a little surprised; I would have expected the other patch to be the 
one responsible.  Are you sure you got the right one?

Your symptoms suggest that the ksuspend_usbd and khubd kernel threads
are in deadlock.  You could verify this by getting an Alt-SysRq-T stack 
trace of the two threads after resuming.  Andrew Morton saw this 
happen, under slightly different circumstances, on his system.

If this is indeed the case, the patch below ought to fix your problem.  
Essentially it replaces calls to flush_workqueue() with
cancel_sync_work().  I sent it to Andrew last week, but his problem
occurs only sporadically so it's hard to test.


On Tue, 29 May 2007, Linus Torvalds wrote:

> Heh. Have I mentioned how much I *hate* those kernel threads being frozen?
> 
> Just for fun, could you try if the patch that just rips out the freezer 
> calls from the STR code just fixes the problem too (instead of reverting 
> that commit?)

We'll know for certain after Mark tries this patch.  However at the 
moment I doubt that freezing the kernel threads has anything to do with 
the problem.

BTW, you mentioned in an earlier email that for suspend-to-RAM, instead
of freezing tasks we should freeze I/O.  It's a good idea, but have you
considered how much overhead it would add?  Even if the necessary
changes are confined to the subsystem level, just think about what
would happen for example with char device drivers.

The only common subsystem they share is VFS.  Every entry point into
VFS would therefore need to test whether a suspend-to-RAM was in
progress, and if it was, block until the suspend was over.  Each of
these tests would be on a hot path -- not something we really want to
do if it is at all avoidable.

Alan Stern



Index: 2.6.22-rc3/drivers/usb/core/hub.c
===================================================================
--- 2.6.22-rc3.orig/drivers/usb/core/hub.c
+++ 2.6.22-rc3/drivers/usb/core/hub.c
@@ -1158,6 +1158,30 @@ static void release_address(struct usb_d
 	}
 }
 
+#ifdef	CONFIG_USB_SUSPEND
+
+static void usb_stop_pm(struct usb_device *udev)
+{
+	/* Synchronize with the ksuspend thread to prevent any more
+	 * autosuspend requests from being submitted, and decrement
+	 * the parent's count of unsuspended children.
+	 */
+	usb_pm_lock(udev);
+	if (udev->parent && !udev->discon_suspended)
+		usb_autosuspend_device(udev->parent);
+	usb_pm_unlock(udev);
+
+	/* Stop any autosuspend requests already submitted */
+	cancel_rearming_delayed_work(&udev->autosuspend);
+}
+
+#else
+
+static inline void usb_stop_pm(struct usb_device *udev)
+{ }
+
+#endif
+
 /**
  * usb_disconnect - disconnect a device (usbcore-internal)
  * @pdev: pointer to device being disconnected
@@ -1224,13 +1248,7 @@ void usb_disconnect(struct usb_device **
 	*pdev = NULL;
 	spin_unlock_irq(&device_state_lock);
 
-	/* Decrement the parent's count of unsuspended children */
-	if (udev->parent) {
-		usb_pm_lock(udev);
-		if (!udev->discon_suspended)
-			usb_autosuspend_device(udev->parent);
-		usb_pm_unlock(udev);
-	}
+	usb_stop_pm(udev);
 
 	put_device(&udev->dev);
 }
Index: 2.6.22-rc3/drivers/usb/core/usb.c
===================================================================
--- 2.6.22-rc3.orig/drivers/usb/core/usb.c
+++ 2.6.22-rc3/drivers/usb/core/usb.c
@@ -184,10 +184,6 @@ static void usb_release_dev(struct devic
 
 	udev = to_usb_device(dev);
 
-#ifdef	CONFIG_USB_SUSPEND
-	cancel_delayed_work(&udev->autosuspend);
-	flush_workqueue(ksuspend_usb_wq);
-#endif
 	usb_destroy_configuration(udev);
 	usb_put_hcd(bus_to_hcd(udev->bus));
 	kfree(udev->product);
Index: 2.6.22-rc3/drivers/usb/core/hcd.c
===================================================================
--- 2.6.22-rc3.orig/drivers/usb/core/hcd.c
+++ 2.6.22-rc3/drivers/usb/core/hcd.c
@@ -1681,7 +1681,7 @@ void usb_remove_hcd(struct usb_hcd *hcd)
 	spin_unlock_irq (&hcd_root_hub_lock);
 
 #ifdef CONFIG_PM
-	flush_workqueue(ksuspend_usb_wq);
+	cancel_work_sync(&hcd->wakeup_work);
 #endif
 
 	mutex_lock(&usb_bus_list_lock);

-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/