The ibmvscsi driver maps pages via map_sg and frees the sg list by calling
unmap_single. This exposes a problem where the code had been using the
length field in the scatterlist to account for the allocation when the
mapping was created but the dma_length would be used when the mapping
was destroyed. The iommu code does coalesce the pages so the dma_length
can be smaller and would result in an entitlement accounting leak.
This patch frees up the entitlement after calling the iommu map_sg code
when pages have been coalesced so that freeing entitlement based on
dma_length is correct.
Signed-off-by: Robert Jennings <[EMAIL PROTECTED]>
---
This applies on top of the existing CMO patchset.
---
arch/powerpc/kernel/vio.c | 7 ++++++-
1 file changed, 6 insertions(+), 1 deletion(-)
Index: b/arch/powerpc/kernel/vio.c
===================================================================
--- a/arch/powerpc/kernel/vio.c
+++ b/arch/powerpc/kernel/vio.c
@@ -571,6 +571,11 @@ static int vio_dma_iommu_map_sg(struct d
atomic_inc(&viodev->cmo.allocs_failed);
}
+ for (sgl = sglist, count = 0; count < ret; count++, sgl++)
+ alloc_size -= roundup(sgl->dma_length, PAGE_SIZE);
+ if (alloc_size)
+ vio_cmo_dealloc(viodev, alloc_size);
+
return ret;
}
@@ -584,7 +589,7 @@ static void vio_dma_iommu_unmap_sg(struc
int count = 0;
for (sgl = sglist; count < nelems; count++, sgl++)
- alloc_size += roundup(sgl->length, PAGE_SIZE);
+ alloc_size += roundup(sgl->dma_length, PAGE_SIZE);
dma_iommu_ops.unmap_sg(dev, sglist, nelems, direction);
_______________________________________________
Linuxppc-dev mailing list
Linuxppc-dev@ozlabs.org
https://ozlabs.org/mailman/listinfo/linuxppc-dev
