Re: [Qemu-devel] 2.0 regression: loadvm assertion with ehci + tablet

[Paolo Bonzini]

Il 31/03/2014 20:14, Cole Robinson ha scritto:
> irq_state is cleared before pci_device_deassert_intx. But tries to clear all
> irqs via pci_irq_handler, but that function will exit without taking any
> action if the requested irq level matches what we already track in irq_state.
> Since irq_state is 0, pci_device_deassert_intx is basically a no-op. Any
> interrupts with level=1 will not be cleared, which is the case with the usb
> tablet after usb_detach.

Thanks for the analysis.  It's my bug indeed.

> This fixes things for me, but I have no idea if it's the proper fix:
>
> diff --git a/hw/pci/pci.c b/hw/pci/pci.c
> index 8f722dd..1912dfb 100644
> --- a/hw/pci/pci.c
> +++ b/hw/pci/pci.c
> @@ -189,9 +189,9 @@ static void pci_do_device_reset(PCIDevice *dev)
>  {
>      int r;
>
> +    pci_device_deassert_intx(dev);
>      dev->irq_state = 0;
>      pci_update_irq_status(dev);
> -    pci_device_deassert_intx(dev);
>      /* Clear all writable bits */
>      pci_word_test_and_clear_mask(dev->config + PCI_COMMAND,
>                                   pci_get_word(dev->wmask + PCI_COMMAND) |

Yes, the patch is fine.  Even better, dev->irq_state = 0 could become an 
assertion too, and the call to pci_update_irq_status is not necessary at 
all.

Paolo