Re: [PATCH v2] usbip: give back URBs for unsent unlink requests during cleanup

[Shuah Khan]

On 8/11/21 7:58 AM, Anirudh Rayabharam wrote:
> On Tue, Aug 10, 2021 at 05:25:51PM -0600, Shuah Khan wrote:
> > On 8/6/21 12:13 PM, Anirudh Rayabharam wrote:
> > > In vhci_device_unlink_cleanup(), the URBs for unsent unlink requests are
> > > not given back. This sometimes causes usb_kill_urb to wait indefinitely
> > > for that urb to be given back. syzbot has reported a hung task issue [1]
> > > for this.
> > >
> > > To fix this, give back the urbs corresponding to unsent unlink requests
> > > (unlink_tx list) similar to how urbs corresponding to unanswered unlink
> > > requests (unlink_rx list) are given back. Since the code is almost the
> > > same, extract it into a new function and call it for both unlink_rx and
> > > unlink_tx lists.
> >
> > Let's not do the refactor - let's first fix the problem and then the refactor.
>
> Sure, I will make it a two patch series where the first one fixes the
> problem and the second one does the refactor.
>
> > > [1]: https://syzkaller.appspot.com/bug?id=08f12df95ae7da69814e64eb5515d5a85ed06b76
> > >
> > > Reported-by: syzbot+74d6ef051d3d2eacf428@xxxxxxxxxxxxxxxxxxxxxxxxx
> > > Tested-by: syzbot+74d6ef051d3d2eacf428@xxxxxxxxxxxxxxxxxxxxxxxxx
> > > Signed-off-by: Anirudh Rayabharam <mail@xxxxxxxxxxxxx>
> > > ---
> > >
> > > Changes in v2:
> > > Use WARN_ON() instead of BUG() when unlink_list is neither unlink_tx nor
> > > unlink_rx.
> > >
> > > v1: https://lore.kernel.org/lkml/20210806164015.25263-1-mail@xxxxxxxxxxxxx/
> > >
> > > ---
> > >    drivers/usb/usbip/vhci_hcd.c | 45 +++++++++++++++++++++++++-----------
> > >    1 file changed, 32 insertions(+), 13 deletions(-)
> > >
> > > diff --git a/drivers/usb/usbip/vhci_hcd.c b/drivers/usb/usbip/vhci_hcd.c
> > > index 4ba6bcdaa8e9..67e638f4c455 100644
> > > --- a/drivers/usb/usbip/vhci_hcd.c
> > > +++ b/drivers/usb/usbip/vhci_hcd.c
> > > @@ -945,7 +945,8 @@ static int vhci_urb_dequeue(struct usb_hcd *hcd, struct urb *urb, int status)
> > >    	return 0;
> > >    }
> > > -static void vhci_device_unlink_cleanup(struct vhci_device *vdev)
> > > +static void __vhci_cleanup_unlink_list(struct vhci_device *vdev,
> > > +		struct list_head *unlink_list)
> > >    {
> > >    	struct vhci_hcd *vhci_hcd = vdev_to_vhci_hcd(vdev);
> > >    	struct usb_hcd *hcd = vhci_hcd_to_hcd(vhci_hcd);
> > > @@ -953,23 +954,23 @@ static void vhci_device_unlink_cleanup(struct vhci_device *vdev)
> > >    	struct vhci_unlink *unlink, *tmp;
> > >    	unsigned long flags;
> > > +	if (WARN(unlink_list != &vdev->unlink_tx
> > > +				&& unlink_list != &vdev->unlink_rx,
> > > +			"Invalid list passed to __vhci_cleanup_unlink_list\n"))
> > > +		return;
> > > +
> >
> > With this change, this will be only place unlink_rx is used without
> > vdev->priv_lock hold? Please explain why this is safe.
>
> Well, this doesn't read or modify the contents of unlink_rx and unlink_tx.
> So, it looks safe to me. Let me know if I'm missing something here.
>
> > >    	spin_lock_irqsave(&vhci->lock, flags);
> > >    	spin_lock(&vdev->priv_lock);
> > > -	list_for_each_entry_safe(unlink, tmp, &vdev->unlink_tx, list) {
> > > -		pr_info("unlink cleanup tx %lu\n", unlink->unlink_seqnum);
> > > -		list_del(&unlink->list);
> > > -		kfree(unlink);
> > > -	}
> > > -
> > > -	while (!list_empty(&vdev->unlink_rx)) {
> > > +	list_for_each_entry_safe(unlink, tmp, unlink_list, list) {
> > >    		struct urb *urb;
> > > -		unlink = list_first_entry(&vdev->unlink_rx, struct vhci_unlink,
> > > -			list);
> > > -
> > > -		/* give back URB of unanswered unlink request */
> > > -		pr_info("unlink cleanup rx %lu\n", unlink->unlink_seqnum);
> > > +		if (unlink_list == &vdev->unlink_tx)
> > > +			pr_info("unlink cleanup tx %lu\n",
> > > +					unlink->unlink_seqnum);
> > > +		else
> > > +			pr_info("unlink cleanup rx %lu\n",
> > > +					unlink->unlink_seqnum);
> > >    		urb = pickup_urb_and_free_priv(vdev, unlink->unlink_seqnum);
> > >    		if (!urb) {
> > > @@ -1001,6 +1002,24 @@ static void vhci_device_unlink_cleanup(struct vhci_device *vdev)
> > >    	spin_unlock_irqrestore(&vhci->lock, flags);
> > >    }
> > > +static inline void vhci_cleanup_unlink_tx(struct vhci_device *vdev)
> > > +{
> > > +	__vhci_cleanup_unlink_list(vdev, &vdev->unlink_tx);
> >
> > With this change, this will be only place unlink_rx is used without
> > vdev->priv_lock hold? Please explain why this is safe.
> >
> > > +}
> > > +
> >
> > Is there a need for this layer?
> >
> > > +static inline void vhci_cleanup_unlink_rx(struct vhci_device *vdev)
> > > +{
> > > +	__vhci_cleanup_unlink_list(vdev, &vdev->unlink_rx);
> >
> > With this change, this will be only place unlink_rx is used without
> > vdev->priv_lock hold? Please explain why this is safe.
> >
> > > +}
> > > +
> > Is there a need for this layer?
>
> I added these wrappers purely for convenience. There is no other purpose.
> Would you prefer this patch without the wrappers?

Yes. Prefer it without the wrappers. When you take the wrappers
out, I think the unlink_rx could be within spinlock hold easily.

thanks,
-- Shuah