Re: [PATCH] x86, aperture: Check for GART before accessing GART registers

[Aravind Gopalakrishnan <aravind.gopalakrishnan@amd.com>]

On 4/7/2015 9:57 AM, Borislav Petkov wrote:
> On Tue, Apr 07, 2015 at 09:46:26AM -0500, Aravind Gopalakrishnan wrote:
>> Okay. I'll do that and correct the typos Ingo pointed out earlier and
>> resend.
> Btw, I think you should do the same in early_gart_iommu_check() too.
>
> Doing the testing this way would mean that we first are testing for GART
> hw presence and then do the rest of checks. If no GART hw, the rest of
> the checks are meaningless.
>
> Also, when testing do a "pci=noearly" boot which should make
>
> 	!early_pci_allowed()
>
> true and thus test that path too.
>

Here are results from further testing:
1. on platforms with both iommu and gart
     - with pci=noearly, we break out of init routines in aperture_64.c 
early. amd_iommu_init() will run through it's init routine.
         - if amd_iommu_init() fails somewhere, we fall back to 
gart_iommu_init()
         - gart_iommu_init() fails since gart_iommu_aperture is not set.
         - fall back to swiotlb.
     - with amd_iommu=off
         - init routines in aperture_64.c run fine as both 
amd_gart_present() and early_pci_allowed() are true
         -  amd_iommu_detect() fails due to command line arg.
         - fall back to gart iommu
     - with pci=noearly and amd_iommu=off
         - break out of aperture_64.c init routines, and 
amd_iommu_detect() fails.
         - fall back to swiotlb

2. on platform with no gart but iommu present,
     - pci=noearly option is not relevant as we break before that due to 
!amd_gart_present()
     - if amd_iommu_init() fails somewhere, we fall back to swiotlb, 
else use iommu

3. on platforms with no gart and no iommu
     - use swiotlb regardless of any command line options passed

Thanks,
-Aravind.