Re: inconsistent {IN-HARDIRQ-W} -> {HARDIRQ-ON-W} usage withhcd_urb_list_lock

[Alan Stern]

On Thu, 11 Jul 2013, Sarah Sharp wrote:

> On Sat, Jul 06, 2013 at 04:48:47PM -0400, Alan Stern wrote:
> > On Sat, 6 Jul 2013, Maarten Lankhorst wrote:
> > 
> > > I didn't even know I still had lockdep on.
> > > The following lockdep splat happened when I plugged in a usb bluetooth dongle, using 
> > > the pre-rc1 3.11 kernel at HEAD b2c311075db
> > > 
> > > =================================
> > > [ INFO: inconsistent lock state ]
> > > 3.10.0+ #106 Not tainted
> > > ---------------------------------
> > > inconsistent {IN-HARDIRQ-W} -> {HARDIRQ-ON-W} usage.
> > > irq/42-xhci_hcd/97 [HC0[0]:SC0[2]:HE1:SE0] takes:
> > >  (hcd_urb_list_lock){?.....}, at: [<ffffffff8149440e>] usb_hcd_unlink_urb_from_ep+0x28/0x4e
> > 
> > > stack backtrace:
> > > CPU: 1 PID: 97 Comm: irq/42-xhci_hcd Not tainted 3.10.0+ #106
> > > Hardware name: Acer Aspire M3985/Aspire M3985, BIOS P01-A1 03/12/2012
> > >  ffffffff8210c150 ffff88040834da48 ffffffff81691af4 0000000000000007
> > >  ffff8804082e20b0 ffff88040834daa8 ffffffff8168cb10 0000000000000002
> > >  ffff880400000001 ffff880400000000 ffffffff8100f4f7 ffff88040834dac4
> > > Call Trace:
> > >  [<ffffffff81691af4>] dump_stack+0x4f/0x84
> > >  [<ffffffff8168cb10>] print_usage_bug+0x1f5/0x206
> > >  [<ffffffff8100f4f7>] ? save_stack_trace+0x2f/0x50
> > >  [<ffffffff810af30c>] mark_lock+0x276/0x2cf
> > >  [<ffffffff810ae8cc>] ? check_usage_forwards+0x12f/0x12f
> > >  [<ffffffff810af925>] __lock_acquire+0x5c0/0x1c2e
> > >  [<ffffffff810b1e04>] ? mark_held_locks+0x6d/0x117
> > >  [<ffffffff8168e8d4>] ? __slab_free+0x1c7/0x2ed
> > >  [<ffffffff810b1f5a>] ? trace_hardirqs_on_caller+0xac/0x1bb
> > >  [<ffffffff810b2076>] ? trace_hardirqs_on+0xd/0xf
> > >  [<ffffffff8149440e>] ? usb_hcd_unlink_urb_from_ep+0x28/0x4e
> > >  [<ffffffff810b1556>] lock_acquire+0x87/0x139
> > >  [<ffffffff8149440e>] ? usb_hcd_unlink_urb_from_ep+0x28/0x4e
> > >  [<ffffffff81698a1a>] _raw_spin_lock+0x3b/0x4a
> > >  [<ffffffff8149440e>] ? usb_hcd_unlink_urb_from_ep+0x28/0x4e
> > >  [<ffffffff8149440e>] usb_hcd_unlink_urb_from_ep+0x28/0x4e
> > >  [<ffffffff814bf55a>] xhci_irq+0x5ac/0x143d
> > >  [<ffffffff81699171>] ? _raw_spin_unlock_irq+0x3b/0x5d
> > >  [<ffffffff8108386d>] ? finish_task_switch+0x7c/0x101
> > >  [<ffffffff81083830>] ? finish_task_switch+0x3f/0x101
> > >  [<ffffffff81697060>] ? __schedule+0x42a/0x885
> > >  [<ffffffff810d7fdb>] ? irq_thread_fn+0x48/0x48
> > >  [<ffffffff814c03fc>] xhci_msi_irq+0x11/0x15
> > 
> > It looks like xhci_msi_irq() needs to call local_irq_save() and 
> > local_irq_restore().
> 
> Alan, does this only need to be called when MSI is enabled?
> xhci_msi_irq() basically calls xhci_irq, which is called if MSI is
> disabled and legacy PCI interrupts are used.
> 
> I'm trying to figure out whether to put the calls to local_irq_save()
> and local_irq_restore() in xhci_msi_irq() or xhci_irq().

Ming Lei pointed out that my comment was wrong.  The reason for the
lockdep violation was because Maarten had set up xhci-hcd to use
threaded interrupts.  You can see this in the current process name in
the lockdep warning: "irq/42-xhci_hcd" is the name of an
interrupt-handler thread.

Without threaded interrupts, it seems likely that no changes are
needed.  Certainly no changes are needed in xhci_irq(), because usbcore
disables local interrupts explicitly before invoking legacy interrupt
handlers (see usb_hcd_irq() in hcd.c).  In the absense of threaded 
interrupts, I assume xhci_msi_irq() will be invoked with interrupts 
disabled -- if not then it will need to disable them before calling 
xhci_irq().

Alan Stern

--
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/