x86_64: make GART PTEs uncacheable
PATCH 1/1 x86_64: make GART PTEs uncacheable
This patches fixes the silent data corruption problems being seen using the GART iommu where 4kB of data where incorrect (seen mostly on Nvidia CK804 systems). This fix, to mark the memory regin the GART PTEs reside on as uncacheable, also brings the code in line with the AGP specification.
---
arch/x86_64/kernel/pci-gart.c +2 -1
@@ -519,7 +519,11 @@ static __init int init_k8_gatt(struct ag gatt_size = (aper_size >>PAGE_SHIFT ) *sizeof (u32); gatt = (void *)__get_free_pages (GFP_KERNEL ,get_order (gatt_size));if (!gatt) -panic ("Cannot allocate GATT table" ); +panic ("Cannot allocate GATT table" ); + if (change_page_attr_addr((unsigned long)gatt, gatt_size >> PAGE_SHIFT,
PAGE_KERNEL_NOCACHE))
+ panic("Could not set GART PTEs to uncacheable pages");
+ global_flush_tlb();
+
memset(gatt, 0, gatt_size);
agp_gatt_table = gatt;
Quoted 3 lines · On Monday 23 April 2007 11:14:10 Joachim Deguara wrote:
This patches fixes the silent data corruption problems being seen using the GART iommu where 4kB of data where incorrect (seen mostly on Nvidia CK804 systems).
Performance numbers? How much slower does this make this? Is it still faster than swiotlb?
Also this will always split up the direct memory mapping of the kernel, so you'll lose more TLB entries even for other data.
Quoted 2 lines
This fix, to mark the memory regin the GART PTEs reside on as uncacheable, also brings the code in line with the AGP specification.
Where in the AGP specification? I can't find any such requirement in AGPv3
-Andi
Quoted 7 lines · Andi Kleen
Quoted 3 lines · On Monday 23 April 2007 11:14:10 Joachim Deguara wrote:
This patches fixes the silent data corruption problems being seen using the GART iommu where 4kB of data where incorrect (seen mostly on Nvidia CK804 systems).
Performance numbers? How much slower does this make this? Is it still faster than swiotlb?
I can work on that as a side note, but while the GART IOMMU is still in the kernel then we need this fix.
Quoted 9 lines
Also this will always split up the direct memory mapping of the kernel, so you'll lose more TLB entries even for other data.
Quoted 2 lines
This fix, to mark the memory regin the GART PTEs reside on as uncacheable, also brings the code in line with the AGP specification.
Where in the AGP specification? I can't find any such requirement in AGPv3
Mark pointed this out and he can answer best. I believe he was referring to section 5.3.4 point 8:
"Core-logic accesses to the GART are not guaranteed to be coherent with host processor caches. In
order to avoid having to flush the cache after every GART update, portable system software should
place the GART in a range of the physical memory space that is considered un-cacheable by host
processors. (A good example is mapping the GART as UC in an IntelÒ PentiumÒ II processor).
However, the specification does not preclude the placement of the GART in cachable memory
space in cases where the coherency is guaranteed through some hardware or software
mechanism."
updated patch follows with corrected long lone.
-Joachim
---
This patches fixes the silent data corruption problems being seen using the GART iommu where 4kB of data where incorrect (seen mostly on Nvidia CK804 systems). This fix to mark the memory regin the GART PTEs reside on as uncacheable also brings the code in line with the AGP specification.
---
arch/x86_64/kernel/pci-gart.c +6 -1
@@ -519,7 +519,12 @@ static __init int init_k8_gatt(struct ag gatt_size = (aper_size >>PAGE_SHIFT ) *sizeof (u32); gatt = (void *)__get_free_pages (GFP_KERNEL ,get_order (gatt_size));if (!gatt) -panic ("Cannot allocate GATT table" ); +panic ("Cannot allocate GATT table" ); +if (change_page_attr_addr ((unsigned long )gatt, gatt_size >>PAGE_SHIFT , +PAGE_KERNEL_NOCACHE )) +panic ("Could not set GART PTEs to uncacheable pages" ); +global_flush_tlb (); +memset (gatt,0 , gatt_size); agp_gatt_table = gatt;
Quoted 2 lines · On Monday 23 April 2007 11:45:11 Joachim Deguara wrote:
I can work on that as a side note, but while the GART IOMMU is still in the kernel then we need this fix.
If it's too slow we can just use swiotlb instead. Probably while enlarging it.
-Andi