intel-iommu: Fix leaks in pagetable freeing

At best the current code only seems to free the leaf pagetables and the root. If you're unlucky enough to have a large gap (like any QEMU guest with more than 3G of memory), only the first chunk of leaf pagetables are freed (plus the root). This is a massive memory leak. This patch re-writes the pagetable freeing function to use a recursive algorithm and manages to not only free all the pagetables, but does it without any apparent performance loss versus the current broken version.

Signed-off-by Alex Williamson <[email protected]>

Cc [email protected]

---

Suggesting for stable, would like to see some soak time, but it's hard to imagine this being any worse than the current code.

This likely also affects device domains, but the current code does ok at freeing individual leaf pagetables and driver domains would only get a full pruning if the driver or device is removed.

Some test programs:
https://github.com/awilliam/tests/blob/master/kvm-huge-guest-test.c
https://github.com/awilliam/tests/blob/master/vfio-huge-guest-test.c

Both of these simulate a large guest on a small host system. They mmap 4G of memory and map it across a large address space just like QEMU would (aside from re-using the same mmap across multiple IOVAs). On existing code the vfio version (w/o a KVM memory slot limit) will leak over 1G of pagetables per run.

 drivers/iommu/intel-iommu.c |   72 +++++++++++++++++++++----------------------
 1 file changed, 35 insertions(+), 37 deletions(-)

drivers/iommu/intel-iommu.c +35 -37

index eec0d3e..15e9b57 100644
@@ -890,56 +890,54 @@ static int dma_pte_clear_range(struct dmar_domain *domain,
 	return order;
 }
 
+static void dma_pte_free_level(struct dmar_domain *domain, int level,
+			       struct dma_pte *pte, unsigned long pfn,
+			       unsigned long start_pfn, unsigned long last_pfn)
+{
+	pfn = max(start_pfn, pfn);
+	pte = &pte[pfn_level_offset(pfn, level)];
+
+	do {
+		unsigned long level_pfn;
+		struct dma_pte *level_pte;
+
+		if (!dma_pte_present(pte) || dma_pte_superpage(pte))
+			goto next;
+
+		level_pfn = pfn & level_mask(level - 1);
+		level_pte = phys_to_virt(dma_pte_addr(pte));
+
+		if (level > 2)
+			dma_pte_free_level(domain, level - 1, level_pte,
+					   level_pfn, start_pfn, last_pfn);
+
+		/* If range covers entire pagetable, free it */
+		if (!(start_pfn > level_pfn ||
+		      last_pfn < level_pfn + level_size(level))) {
+			dma_clear_pte(pte);
+			domain_flush_cache(domain, pte, sizeof(*pte));
+			free_pgtable_page(level_pte);
+		}
+next:
+		pfn += level_size(level);
+	} while (!first_pte_in_page(++pte) && pfn <= last_pfn);
+}
+
 /* free page table pages. last level pte should already be cleared */
 static void dma_pte_free_pagetable(struct dmar_domain *domain,
 				   unsigned long start_pfn,
 				   unsigned long last_pfn)
 {
 	int addr_width = agaw_to_width(domain->agaw) - VTD_PAGE_SHIFT;
-	struct dma_pte *first_pte, *pte;
-	int total = agaw_to_level(domain->agaw);
-	int level;
-	unsigned long tmp;
-	int large_page = 2;
 
 	BUG_ON(addr_width < BITS_PER_LONG && start_pfn >> addr_width);
 	BUG_ON(addr_width < BITS_PER_LONG && last_pfn >> addr_width);
 	BUG_ON(start_pfn > last_pfn);
 
 	/* We don't need lock here; nobody else touches the iova range */
-	level = 2;
-	while (level <= total) {
-		tmp = align_to_level(start_pfn, level);
-
-		/* If we can't even clear one PTE at this level, we're done */
-		if (tmp + level_size(level) - 1 > last_pfn)
-			return;
-
-		do {
-			large_page = level;
-			first_pte = pte = dma_pfn_level_pte(domain, tmp, level, &large_page);
-			if (large_page > level)
-				level = large_page + 1;
-			if (!pte) {
-				tmp = align_to_level(tmp + 1, level + 1);
-				continue;
-			}
-			do {
-				if (dma_pte_present(pte)) {
-					free_pgtable_page(phys_to_virt(dma_pte_addr(pte)));
-					dma_clear_pte(pte);
-				}
-				pte++;
-				tmp += level_size(level);
-			} while (!first_pte_in_page(pte) &&
-				 tmp + level_size(level) - 1 <= last_pfn);
+	dma_pte_free_level(domain, agaw_to_level(domain->agaw),
+			   domain->pgd, 0, start_pfn, last_pfn);
 
-			domain_flush_cache(domain, first_pte,
-					   (void *)pte - (void *)first_pte);
-			
-		} while (tmp && tmp + level_size(level) - 1 <= last_pfn);
-		level++;
-	}
 	/* free pgd */
 	if (start_pfn == 0 && last_pfn == DOMAIN_MAX_PFN(domain->gaw)) {
 		free_pgtable_page(domain->pgd);

This is a pretty massive memory leak, anyone @Intel care? Thanks,

Alex

Quoted 120 lines · Alex Williamson

At best the current code only seems to free the leaf pagetables and the root. If you're unlucky enough to have a large gap (like any QEMU guest with more than 3G of memory), only the first chunk of leaf pagetables are freed (plus the root). This is a massive memory leak. This patch re-writes the pagetable freeing function to use a recursive algorithm and manages to not only free all the pagetables, but does it without any apparent performance loss versus the current broken version.

Signed-off-by Alex Williamson <[email protected]>

Cc [email protected]

---

Suggesting for stable, would like to see some soak time, but it's hard to imagine this being any worse than the current code.

This likely also affects device domains, but the current code does ok at freeing individual leaf pagetables and driver domains would only get a full pruning if the driver or device is removed.

Some test programs:
https://github.com/awilliam/tests/blob/master/kvm-huge-guest-test.c
https://github.com/awilliam/tests/blob/master/vfio-huge-guest-test.c

Both of these simulate a large guest on a small host system. They mmap 4G of memory and map it across a large address space just like QEMU would (aside from re-using the same mmap across multiple IOVAs). On existing code the vfio version (w/o a KVM memory slot limit) will leak over 1G of pagetables per run.

 drivers/iommu/intel-iommu.c |   72 +++++++++++++++++++++----------------------
 1 file changed, 35 insertions(+), 37 deletions(-)

drivers/iommu/intel-iommu.c +35 -31

index eec0d3e..15e9b57 100644
@@ -890,56 +890,54 @@ static int dma_pte_clear_range(struct dmar_domain *domain,
 	return order;
 }
 
+static void dma_pte_free_level(struct dmar_domain *domain, int level,
+			       struct dma_pte *pte, unsigned long pfn,
+			       unsigned long start_pfn, unsigned long last_pfn)
+{
+	pfn = max(start_pfn, pfn);
+	pte = &pte[pfn_level_offset(pfn, level)];
+
+	do {
+		unsigned long level_pfn;
+		struct dma_pte *level_pte;
+
+		if (!dma_pte_present(pte) || dma_pte_superpage(pte))
+			goto next;
+
+		level_pfn = pfn & level_mask(level - 1);
+		level_pte = phys_to_virt(dma_pte_addr(pte));
+
+		if (level > 2)
+			dma_pte_free_level(domain, level - 1, level_pte,
+					   level_pfn, start_pfn, last_pfn);
+
+		/* If range covers entire pagetable, free it */
+		if (!(start_pfn > level_pfn ||
+		      last_pfn < level_pfn + level_size(level))) {
+			dma_clear_pte(pte);
+			domain_flush_cache(domain, pte, sizeof(*pte));
+			free_pgtable_page(level_pte);
+		}
+next:
+		pfn += level_size(level);
+	} while (!first_pte_in_page(++pte) && pfn <= last_pfn);
+}
+
 /* free page table pages. last level pte should already be cleared */
 static void dma_pte_free_pagetable(struct dmar_domain *domain,
 				   unsigned long start_pfn,
 				   unsigned long last_pfn)
 {
 	int addr_width = agaw_to_width(domain->agaw) - VTD_PAGE_SHIFT;
-	struct dma_pte *first_pte, *pte;
-	int total = agaw_to_level(domain->agaw);
-	int level;
-	unsigned long tmp;
-	int large_page = 2;
 
 	BUG_ON(addr_width < BITS_PER_LONG && start_pfn >> addr_width);
 	BUG_ON(addr_width < BITS_PER_LONG && last_pfn >> addr_width);
 	BUG_ON(start_pfn > last_pfn);
 
 	/* We don't need lock here; nobody else touches the iova range */
-	level = 2;
-	while (level <= total) {
-		tmp = align_to_level(start_pfn, level);
-
-		/* If we can't even clear one PTE at this level, we're done */
-		if (tmp + level_size(level) - 1 > last_pfn)
-			return;
-
-		do {
-			large_page = level;
-			first_pte = pte = dma_pfn_level_pte(domain, tmp, level, &large_page);
-			if (large_page > level)
-				level = large_page + 1;
-			if (!pte) {
-				tmp = align_to_level(tmp + 1, level + 1);
-				continue;
-			}
-			do {
-				if (dma_pte_present(pte)) {
-					free_pgtable_page(phys_to_virt(dma_pte_addr(pte)));
-					dma_clear_pte(pte);
-				}
-				pte++;
-				tmp += level_size(level);
-			} while (!first_pte_in_page(pte) &&
-				 tmp + level_size(level) - 1 <= last_pfn);
+	dma_pte_free_level(domain, agaw_to_level(domain->agaw),
+			   domain->pgd, 0, start_pfn, last_pfn);
 

drivers/iommu/intel-iommu.c static int dma_pte_clear_range(struct dmar_domain *domain, · line 943

Quoted 10 lines

drivers/iommu/intel-iommu.c +0 -6

-			domain_flush_cache(domain, first_pte,
-					   (void *)pte - (void *)first_pte);
-			
-		} while (tmp && tmp + level_size(level) - 1 <= last_pfn);
-		level++;
-	}
 	/* free pgd */
 	if (start_pfn == 0 && last_pfn == DOMAIN_MAX_PFN(domain->gaw)) {
 		free_pgtable_page(domain->pgd);

At best the current code only seems to free the leaf pagetables and the root. If you're unlucky enough to have a large gap (like any QEMU guest with more than 3G of memory), only the first chunk of leaf pagetables are freed (plus the root). This is a massive memory leak. This patch re-writes the pagetable freeing function to use a recursive algorithm and manages to not only free all the pagetables, but does it without any apparent performance loss versus the current broken version.

Signed-off-by Alex Williamson <[email protected]>

Cc [email protected]

At best the current code only seems to free the leaf pagetables and the root. If you're unlucky enough to have a large gap (like any QEMU guest with more than 3G of memory), only the first chunk of leaf pagetables are freed (plus the root). This is a massive memory leak. This patch re-writes the pagetable freeing function to use a recursive algorithm and manages to not only free all the pagetables, but does it without any apparent performance loss versus the current broken version.

Signed-off-by Alex Williamson <[email protected]>

Cc [email protected]

---

Suggesting for stable, would like to see some soak time, but it's hard to imagine this being any worse than the current code.

Quoted 13 lines · Alex Williamson

At best the current code only seems to free the leaf pagetables and the root. If you're unlucky enough to have a large gap (like any QEMU guest with more than 3G of memory), only the first chunk of leaf pagetables are freed (plus the root). This is a massive memory leak. This patch re-writes the pagetable freeing function to use a recursive algorithm and manages to not only free all the pagetables, but does it without any apparent performance loss versus the current broken version.

Signed-off-by Alex Williamson <[email protected]>

Cc [email protected]

---

Quoted 2 lines

Suggesting for stable, would like to see some soak time, but it's hard to imagine this being any worse than the current code.

Btw, I have a backport for the 3.0.x series which builds fine here, in case you guys are interested :)