coccinelle check scripts/coccinelle/locks/call_kern.cocci found that
in drivers/usb/host/xhci.c an allocation with GFP_KERNEL is done
with locks held:
xhci_resume
spin_lock_irq(xhci->lock)
xhci_setup_msix
kmalloc(GFP_KERNEL)
Change it to GFP_ATOMIC.
Signed-off-by: David Sterba <[email protected]>
CC: Sarah Sharp <[email protected]>
---
Exists in v2.6.37-rc1 and current linux-next
diff -u -p a/drivers/usb/host/xhci.c b/drivers/usb/host/xhci.c
--- a/drivers/usb/host/xhci.c 2010-10-27 11:27:23.534310054 +0200
+++ b/drivers/usb/host/xhci.c 2010-11-02 15:53:03.000000000 +0100
@@ -240,7 +240,7 @@ static int xhci_setup_msix(struct xhci_h
xhci->msix_entries =
kmalloc((sizeof(struct msix_entry))*xhci->msix_count,
- GFP_KERNEL);
+ GFP_ATOMIC);
if (!xhci->msix_entries) {
xhci_err(xhci, "Failed to allocate MSI-X entries\n");
return -ENOMEM;
On Wed, Nov 03, 2010 at 09:19:03AM +0100, Jiri Slaby wrote:
> On 11/02/2010 11:47 PM, David Sterba wrote:
> > coccinelle check scripts/coccinelle/locks/call_kern.cocci found that
> > in drivers/usb/host/xhci.c an allocation with GFP_KERNEL is done
> > with locks held:
> >
> > xhci_resume
> > spin_lock_irq(xhci->lock)
> > xhci_setup_msix
> > kmalloc(GFP_KERNEL)
> >
> > Change it to GFP_ATOMIC.
>
> Hi, I already reported that [1] and this is not enough. There are other
> sleepy calls like request_irq inside...
>
> [1] http://lkml.org/lkml/2010/10/23/17
Andiry is looking into this. Andiry, perhaps you don't need to take the
xHCI spinlock in xhci_resume()? If that function is being called
because the PCI device is being resumed, you know nothing else is going
to touch the xHCI host controller. (Except maybe the BIOS, but it isn't
going to respect xhci->lock at all.) The USB core certainly won't touch
the host controller until it's resumed. Maybe we could get an interrupt
with a port status change, but I think it's unlikely...
Alan, can you think of any reason the xHCI driver would need to grab its
host controller spinlock on PCI resume?
Sarah Sharp
On Thu, 4 Nov 2010, Sarah Sharp wrote:
> On Wed, Nov 03, 2010 at 09:19:03AM +0100, Jiri Slaby wrote:
> > On 11/02/2010 11:47 PM, David Sterba wrote:
> > > coccinelle check scripts/coccinelle/locks/call_kern.cocci found that
> > > in drivers/usb/host/xhci.c an allocation with GFP_KERNEL is done
> > > with locks held:
> > >
> > > xhci_resume
> > > spin_lock_irq(xhci->lock)
> > > xhci_setup_msix
> > > kmalloc(GFP_KERNEL)
> > >
> > > Change it to GFP_ATOMIC.
> >
> > Hi, I already reported that [1] and this is not enough. There are other
> > sleepy calls like request_irq inside...
> >
> > [1] http://lkml.org/lkml/2010/10/23/17
>
> Andiry is looking into this. Andiry, perhaps you don't need to take the
> xHCI spinlock in xhci_resume()? If that function is being called
> because the PCI device is being resumed, you know nothing else is going
> to touch the xHCI host controller. (Except maybe the BIOS, but it isn't
> going to respect xhci->lock at all.) The USB core certainly won't touch
> the host controller until it's resumed. Maybe we could get an interrupt
> with a port status change, but I think it's unlikely...
>
> Alan, can you think of any reason the xHCI driver would need to grab its
> host controller spinlock on PCI resume?
Maybe an interrupt isn't as unlikely as all that. To be safe you
should acquire the spinlock. The real question is why these other
routines are performing actions that could block in a path that holds
the lock. Maybe the spinlock should be acquired _after_ doing the
potentially-blocking operations.
Alan Stern