Re: [Qemu-devel] [PATCH v2] pci: relax pci_msi_get_message()

[Michael S. Tsirkin]

On Tue, Nov 22, 2016 at 04:08:50PM +0800, Peter Xu wrote:
> We are very strict in the past getting MSIs from commit
> d1f6af6a1 ("kvm-irqchip: simplify kvm_irqchip_add_msi_route"), assuming
> that MSI should be configured before hand when fetching. When we have
> unrecognized configurations, we panic the system. However looks like
> this is too strict to be working on some platform, and issues occured.
> Firstly it's found on a ppc case and fixed by David in:
> 
>   6d17a01 vfio/pci: Fix regression in MSI routing configuration
> 
> However we encountered another case now with windows virtio driver and
> reported (and possibly more):
> 
>   http://bugs.debian.org/844361
> 
> To make every driver/hardware happy, let's loosen the rule and go back
> to the original behavior - instead of panic the system, when we try to
> fetch MSI without configured MSI/MSI-X system, we just provide an empty
> message to make drivers happy.
> 
> Reported-by: Maciej Kotliński <address@hidden>
> Signed-off-by: Peter Xu <address@hidden>
> ---
>  hw/pci/pci.c | 8 +++++---
>  1 file changed, 5 insertions(+), 3 deletions(-)
> 
> diff --git a/hw/pci/pci.c b/hw/pci/pci.c
> index 24fae16..0cd2bb0 100644
> --- a/hw/pci/pci.c
> +++ b/hw/pci/pci.c
> @@ -2606,9 +2606,11 @@ MSIMessage pci_get_msi_message(PCIDevice *dev, int 
> vector)
>      } else if (msi_enabled(dev)) {
>          msg = msi_get_message(dev, vector);
>      } else {
> -        /* Should never happen */
> -        error_report("%s: unknown interrupt type", __func__);
> -        abort();
> +        /*
> +         * Device is not configured with MSI/MSI-X yet, let's provide
> +         * an empty message to make all device drivers happy.
> +         */
> +        memset(&msg, 0, sizeof(msg));
>      }
>      return msg;
>  }

Looks like a hack to me. Even if callers happen to treat
0 message as a nop, there's no guarantee that's true for
all platforms. Pls change pci_get_msi_message to
return an error code, and fix callers to check that.

> -- 
> 2.7.4