@@ -509,29 +509,35 @@ static u32 *as_get_pte(struct tegra_smmu_as *as, dma_addr_t iova,
return &pt[pte];
}
-static void as_put_pte(struct tegra_smmu_as *as, dma_addr_t iova)
+static void tegra_smmu_pte_put_use(struct tegra_smmu_as *as, unsigned long iova)
{
+ struct tegra_smmu *smmu = as->smmu;
u32 pde = (iova >> SMMU_PDE_SHIFT) & 0x3ff;
- u32 pte = (iova >> SMMU_PTE_SHIFT) & 0x3ff;
u32 *count = page_address(as->count);
- u32 *pd = page_address(as->pd), *pt;
+ u32 *pd = page_address(as->pd);
struct page *page;
- page = pfn_to_page(pd[pde] & as->smmu->pfn_mask);
- pt = page_address(page);
+ page = pfn_to_page(pd[pde] & smmu->pfn_mask);
/*
* When no entries in this page table are used anymore, return the
* memory page to the system.
*/
- if (pt[pte] != 0) {
- if (--count[pde] == 0) {
- ClearPageReserved(page);
- __free_page(page);
- pd[pde] = 0;
- }
+ if (--count[pde] == 0) {
+ unsigned int offset = pde * sizeof(*pd);
- pt[pte] = 0;
+ /* Clear the page directory entry first */
+ pd[pde] = 0;
+
+ /* Flush the page directory entry */
+ smmu->soc->ops->flush_dcache(as->pd, offset, sizeof(*pd));
+ smmu_flush_ptc(smmu, as->pd, offset);
+ smmu_flush_tlb_section(smmu, as->id, iova);
+ smmu_flush(smmu);
+
+ /* Finally, free the page */
+ ClearPageReserved(page);
+ __free_page(page);
}
}
@@ -569,17 +575,20 @@ static size_t tegra_smmu_unmap(struct iommu_domain *domain, unsigned long iova,
u32 *pte;
pte = as_get_pte(as, iova, &page);
- if (!pte)
+ if (!pte || !*pte)
return 0;
+ *pte = 0;
+
offset = offset_in_page(pte);
- as_put_pte(as, iova);
smmu->soc->ops->flush_dcache(page, offset, 4);
smmu_flush_ptc(smmu, page, offset);
smmu_flush_tlb_group(smmu, as->id, iova);
smmu_flush(smmu);
+ tegra_smmu_pte_put_use(as, iova);
+
return size;
}
The Tegra SMMU unmap path has several problems: 1. as_pte_put() can perform a write-after-free 2. tegra_smmu_unmap() can perform cache maintanence on a page we have just freed. 3. when a page table is unmapped, there is no CPU cache maintanence of the write clearing the page directory entry, nor is there any maintanence of the IOMMU to ensure that it sees the page table has gone. Fix this by getting rid of as_pte_put(), and instead coding the PTE unmap separately from the PDE unmap, placing the PDE unmap after the PTE unmap has been completed. Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk> --- drivers/iommu/tegra-smmu.c | 37 +++++++++++++++++++++++-------------- 1 file changed, 23 insertions(+), 14 deletions(-)