Re: data corruption with nvidia chipsets and IDE/SATA drives (k8 cpuerrata needed?)

From: Robert Hancock
Date: Mon Mar 05 2007 - 01:25:49 EST


Chip Coldwell wrote:
On Wed, 17 Jan 2007, Andi Kleen wrote:

On Wednesday 17 January 2007 07:31, Chris Wedgwood wrote:
On Tue, Jan 16, 2007 at 08:52:32PM +0100, Christoph Anton Mitterer wrote:
I agree,... it seems drastic, but this is the only really secure
solution.
I'd like to here from Andi how he feels about this? It seems like a
somewhat drastic solution in some ways given a lot of hardware doesn't
seem to be affected (or maybe in those cases it's just really hard to
hit, I don't know).
AMD is looking at the issue. Only Nvidia chipsets seem to be affected,
although there were similar problems on VIA in the past too.
Unless a good workaround comes around soon I'll probably default
to iommu=soft on Nvidia.

We (Sun, AMD, Nvidia and Red Hat) have been testing a patch that seems
to solve the problem. AMD and Nvidia analyzed an HDT trace that
seemed to indicate that CPU updates of the GATT were still in cache
when a subsequent table walk caused by a device load used a stale GATT
PTE. That analysis inspired this patch, submitted to this list as an
RFC. It is not obvious (to me, at least) why this problem has only
shown up on Nvidia SATA controllers.

We are continuing to investigate.

diff --git a/arch/x86_64/kernel/pci-gart.c b/arch/x86_64/kernel/pci-gart.c
index 030eb37..1dd461a 100644
--- a/arch/x86_64/kernel/pci-gart.c
+++ b/arch/x86_64/kernel/pci-gart.c
@@ -69,6 +69,8 @@ static u32 gart_unmapped_entry;
#define AGPEXTERN
#endif
+#define GATT_CLFLUSH(i) asm volatile ("clflush (%0)" :: "r" (iommu_gatt_base + (i)))
+
/* backdoor interface to AGP driver */
AGPEXTERN int agp_memory_reserved;
AGPEXTERN __u32 *agp_gatt_table;
@@ -221,6 +223,7 @@ static dma_addr_t dma_map_area(struct device *dev, dma_addr_t phys_mem,
for (i = 0; i < npages; i++) {
iommu_gatt_base[iommu_page + i] = GPTE_ENCODE(phys_mem);
SET_LEAK(iommu_page + i);
+ GATT_CLFLUSH(iommu_page + i);
phys_mem += PAGE_SIZE;
}
return iommu_bus_base + iommu_page*PAGE_SIZE + (phys_mem & ~PAGE_MASK);
@@ -348,6 +351,7 @@ static int __dma_map_cont(struct scatterlist *sg, int start, int stopat,
while (pages--) { iommu_gatt_base[iommu_page] = GPTE_ENCODE(addr); SET_LEAK(iommu_page);
+ GATT_CLFLUSH(iommu_page);
addr += PAGE_SIZE;
iommu_page++;
}



Andi, have you had a look at this? I'm a bit surprised at the lack of reaction to this find..

--
Robert Hancock Saskatoon, SK, Canada
To email, remove "nospam" from hancockr@xxxxxxxxxxxxx
Home Page: http://www.roberthancock.com/

-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/