[Intel-gfx] git pull] drm for v4.1-rc1

[Bjorn Helgaas]

On Tue, Apr 21, 2015 at 11:07 AM, Linus Torvalds
<torvalds at linux-foundation.org> wrote:
> Hmm. The odd Intel PCI resource mess is back.
>
> Or maybe it never went away.
>
> I get these when suspending. Things *work*, but it's really spamming
> my logs a fair bit:
>
>   i915 0000:00:02.0: BAR 6: [??? 0x00000000 flags 0x2] has bogus alignment
>   pci_bus 0000:01: Allocating resources
>   pci_bus 0000:02: Allocating resources
>   i915 0000:00:02.0: BAR 6: [??? 0x00000000 flags 0x2] has bogus alignment
>   i915 0000:00:02.0: BAR 6: [??? 0x00000000 flags 0x2] has bogus alignment
>   i915 0000:00:02.0: BAR 6: [??? 0x00000000 flags 0x2] has bogus alignment
>   i915 0000:00:02.0: BAR 6: [??? 0x00000000 flags 0x2] has bogus alignment
>   i915 0000:00:02.0: BAR 6: [??? 0x00000000 flags 0x2] has bogus alignment
>   i915 0000:00:02.0: BAR 6: [??? 0x00000000 flags 0x2] has bogus alignment
>
> That resource is complete garbage. "flags 0x2" is not even a valid
> flag value. I'm *guessing* it might be IORESOURCE_ROM_SHADOW, but if
> that is valid, then it should also have have had the IORESOURCE_MEM
> bit, and it doesn't.
>
> (The low 8 bits of the resource flags depend on the high bits, which
> is why I say that I'm "guessing" at that ROM_SHADOW bit. It could be
> something else, like a IORESOURCE_MEM_CACHEABLE PnP bit - but that
> should not show up for PCI, and "BAR 6" is normally the ROM resource,
> so the ROM_SHADOW bit makes some sense.
>
> The only place that sets IORESOURCE_ROM_SHADOW that I find is the x86
> pci_fixup_video() function. That one checks for PCI_COMMAND_IO |
> PCI_COMMAND_MEMORY in the PCI command word, though. Why are the other
> bits not set?
>
> Both i915/dri people and PCI people on the Cc. This warning does *not*
> happen at bootup, but only at suspend time. So my suspicion is that
> somebody messes with the PCI ROM resource, and disables it or
> something, but the IORESOURCE_ROM_SHADOW never gets cleared. And then
> because res->flags is non-zero, the PCI scanning code doesn't ignore
> the resource.
>
> Just before the whole bogus alignment check, the PCI code does
>
>                 if (!(r->flags) || r->parent)
>                         continue;
>
> (don't ask me about the odd parenthesis) which *should* have
> triggered, but that IORESOURCE_ROM_SHADOW bit screws things up.
>
> Anybody?

I'll look into this.  It's been around a long time, but hasn't
percolated to the top of my list until now.

https://bugzilla.kernel.org/show_bug.cgi?id=16063 says "echo 1
>/sys/bus/pci/rescan" is also a reproducer.

Bjorn