Re: [Qemu-devel] [PATCH v1] pci: Don't register a specialized 'config_write' if default behavior is intended

[Marcel Apfelbaum <marcel.apfelbaum@gmail.com>]

On 06/17/2015 10:17 PM, Shmulik Ladkani wrote:
> Hi,
>
> On Wed, 17 Jun 2015 12:37:18 +0300, marcel.apfelbaum@gmail.com wrote:
>> BTW, did you notice a bug here?  If yes, can you elaborate?
>
> No, not a direct bug.
> We noticed this while working on related code areas.
>
> There's some history behind this.
>
> In 95d6580 'msi: Invoke msi/msix_write_config from PCI core', the calls
> to msi[x]_write_config have been added into pci_default_write_config,
> and many specialized 'config_write' methods have been eliminated.
>
> However there was a bug in 95d6580 - the values written to msi/msix
> were always 0.
> This was recently fixed in d7efb7e
>       'pci: avoid losing config updates to MSI/MSIX cap regs'
Got it.

>
> I assume that device authors were either (1) unware of the
> generalization, thus kept invoking msi[x]_write_config explicitly, or
> (2) trying to overcome the "lost writes".
>
> Anyway, I'm no PCI expert here, but I assume the side-effect invoking
> msi[x]_write_config twice (explicitly from the specialized config_write,
> then implicitly from pci_default_write_config) isn't desired.
Of course.

>
> Meaning, the suggested patch follows the spirit of 95d6580.
> Let me know if my analysis is flawed.
Thank you for the patch, you are completely right.
My 'Reviewed-by' tag is there, I think Michael, the PCI
maintainer, will take it shortly.

Thanks,
Marcel

>
> Regards,
> Shmulik
>