ACPI/IORT: handle potential overflows

The sum of dmaaddr and size may overflow, particularly considering there are cases where size will be U64_MAX.

Signed-off-by Andrew Jones <[email protected]>

---
  drivers/acpi/arm64/iort.c | 7 ++++++-
  1 file changed, 6 insertions(+), 1 deletion(-)

drivers/acpi/arm64/iort.c +6 -1

index 70f4e80b9246..a0f4c157ba5e 100644
@@ -1002,7 +1002,12 @@ void iort_dma_setup(struct device *dev, u64 *dma_addr, u64 *dma_size)
  	}
 
  	if (!ret) {
-		msb = fls64(dmaaddr + size - 1);
+		u64 dmaaddr_max = dmaaddr + size - 1;
+		if (dmaaddr_max >= dmaaddr)
+			msb = fls64(dmaaddr_max);
+		else
+			msb = 64;
+
  		/*
  		 * Round-up to the power-of-two mask or set
  		 * the mask to the whole 64-bit address space

Quoted 2 lines · Andrew Jones

The sum of dmaaddr and size may overflow, particularly considering there are cases where size will be U64_MAX.

Only if the firmware is broken in the first place, though. It would be weird to describe an explicit _DMA range of base=0 and size=U64_MAX, because it's effectively the same as just not having one at all, but it's not strictly illegal. However, since the ACPI System Memory address space is at most 64-bit, anything that would actually overflow here is already describing an impossibility - really, we should probably scream even louder about a firmware bug and reject it entirely, rather than quietly hiding it.

Robin.

Quoted 23 lines

Signed-off-by Andrew Jones <[email protected]>

---
  drivers/acpi/arm64/iort.c | 7 ++++++-
  1 file changed, 6 insertions(+), 1 deletion(-)

drivers/acpi/arm64/iort.c +6 -1

index 70f4e80b9246..a0f4c157ba5e 100644
@@ -1002,7 +1002,12 @@ void iort_dma_setup(struct device *dev, u64 *dma_addr, u64 *dma_size)
  	}
  	if (!ret) {
-		msb = fls64(dmaaddr + size - 1);
+		u64 dmaaddr_max = dmaaddr + size - 1;
+		if (dmaaddr_max >= dmaaddr)
+			msb = fls64(dmaaddr_max);
+		else
+			msb = 64;
+
  		/*
  		 * Round-up to the power-of-two mask or set
  		 * the mask to the whole 64-bit address space
 

Quoted 11 lines · Robin Murphy

Quoted 2 lines · Andrew Jones

The sum of dmaaddr and size may overflow, particularly considering there are cases where size will be U64_MAX.

Only if the firmware is broken in the first place, though. It would be weird to describe an explicit _DMA range of base=0 and size=U64_MAX, because it's effectively the same as just not having one at all, but it's not strictly illegal. However, since the ACPI System Memory address space is at most 64-bit, anything that would actually overflow here is already describing an impossibility - really, we should probably scream even louder about a firmware bug and reject it entirely, rather than quietly hiding it.

Ah, looking again I see the paths. Either acpi_dma_get_range() returns success, in which case base and size are fine, or it returns an EINVAL, in which case base=size=0, or it returns ENODEV in which case base is zero, so size may be set to U64_MAX by rc_dma_get_range() with no problem. The !dev_is_pci(dev) path is also fine since base=0.

While I agree that we should complain when firmware provides bad ACPI tables, my understanding of setting iort.memory_address_limit=64 was that it simply meant "no limit", regardless of the base. However I now see that it won't be used unless base=0. So I don't think we have a problem, and we don't even seem to be missing firmware sanity checks.

It might be nice to have a comment explaining this stuff somewhere, but otherwise sorry for the noise.

Thanks,
drew

Quoted 26 lines

Robin.

Quoted 23 lines

Signed-off-by Andrew Jones <[email protected]>

---
  drivers/acpi/arm64/iort.c | 7 ++++++-
  1 file changed, 6 insertions(+), 1 deletion(-)

drivers/acpi/arm64/iort.c +6 -1

index 70f4e80b9246..a0f4c157ba5e 100644
@@ -1002,7 +1002,12 @@ void iort_dma_setup(struct device *dev, u64 *dma_addr, u64 *dma_size)
  	}
  	if (!ret) {
-		msb = fls64(dmaaddr + size - 1);
+		u64 dmaaddr_max = dmaaddr + size - 1;
+		if (dmaaddr_max >= dmaaddr)
+			msb = fls64(dmaaddr_max);
+		else
+			msb = 64;
+
  		/*
  		 * Round-up to the power-of-two mask or set
  		 * the mask to the whole 64-bit address space
 

Hi Robin, Drew,

Quoted 13 lines · Andrew Jones

Quoted 11 lines · Robin Murphy

Quoted 2 lines · Andrew Jones

The sum of dmaaddr and size may overflow, particularly considering there are cases where size will be U64_MAX.

Only if the firmware is broken in the first place, though. It would be weird to describe an explicit _DMA range of base=0 and size=U64_MAX, because it's effectively the same as just not having one at all, but it's not strictly illegal. However, since the ACPI System Memory address space is at most 64-bit, anything that would actually overflow here is already describing an impossibility - really, we should probably scream even louder about a firmware bug and reject it entirely, rather than quietly hiding it.

Quoted 5 lines

Ah, looking again I see the paths. Either acpi_dma_get_range() returns success, in which case base and size are fine, or it returns an EINVAL, in which case base=size=0, or it returns ENODEV in which case base is zero, so size may be set to U64_MAX by rc_dma_get_range() with no problem. The !dev_is_pci(dev) path is also fine since base=0.

So practically putting an explicit memory_address_limit=64 is harmless as dmaaddr always is 0, right?

In QEMU I intended to update the ACPI code to comply with the rev D spec. in that case the RC table revision is 1 (rev D) and the memory_address_limit needs to be filled. If we don't want to restrict the limit, isn't it the right choice to set 64 here?

Quoted 27 lines · Auger Eric

Hi Robin, Drew,

Quoted 17 lines · Andrew Jones

Quoted 4 lines · Robin Murphy

Quoted 2 lines · Andrew Jones

The sum of dmaaddr and size may overflow, particularly considering there are cases where size will be U64_MAX.

Quoted 11 lines

Only if the firmware is broken in the first place, though. It would be weird
to describe an explicit _DMA range of base=0 and size=U64_MAX, because it's
effectively the same as just not having one at all, but it's not strictly
illegal. However, since the ACPI System Memory address space is at most 64-bit, anything that would actually overflow here is already describing an
impossibility - really, we should probably scream even louder about a firmware bug and reject it entirely, rather than quietly hiding it.

Quoted 6 lines

Ah, looking again I see the paths. Either acpi_dma_get_range() returns success, in which case base and size are fine, or it returns an EINVAL, in which case base=size=0, or it returns ENODEV in which case base is zero, so size may be set to U64_MAX by rc_dma_get_range() with no problem.
The !dev_is_pci(dev) path is also fine since base=0.

Quoted 7 lines

So practically putting an explicit memory_address_limit=64 is harmless as dmaaddr always is 0, right?

In QEMU I intended to update the ACPI code to comply with the rev D spec. in that case the RC table revision is 1 (rev D) and the memory_address_limit needs to be filled. If we don't want to restrict the limit, isn't it the right choice to set 64 here?

Indeed, the Memory Address Size Limit doesn't cater for offsets so can't run into this kind of overflow in the first place. For a fully-emulated PCI hierarchy I'd say 64 is not just harmless but in fact entirely correct - you're going to have more fun with VFIO passthrough if the host tables have more restrictive limits, but I guess that's a problem for the future ;)

Robin.

The sum of base and size may overflow, particularly considering there are cases where size will be U64_MAX. Also, end_pfn is unused, so we remove it. Finally, as size doesn't actually need to be IOMMU page aligned we remove it from the comment stating both it and base should be. I wonder if we shouldn't at least warn when base is not aligned?

Signed-off-by Andrew Jones <[email protected]>

---
  drivers/iommu/dma-iommu.c | 15 +++++++++------
  1 file changed, 9 insertions(+), 6 deletions(-)

drivers/iommu/dma-iommu.c +9 -6

index d1b04753b204..a0b01398b15c 100644
@@ -281,9 +281,9 @@ static void iommu_dma_flush_iotlb_all(struct iova_domain *iovad)
   * @size: Size of IOVA space
   * @dev: Device the domain is being initialised for
   *
- * @base and @size should be exact multiples of IOMMU page granularity to
- * avoid rounding surprises. If necessary, we reserve the page at address 0
- * to ensure it is an invalid IOVA. It is safe to reinitialise a domain, but
+ * @base should be an exact multiple of IOMMU page granularity to avoid
+ * rounding surprises. If necessary, we reserve the page at address 0 to
+ * ensure it is an invalid IOVA. It is safe to reinitialise a domain, but
   * any change which could make prior IOVAs invalid will fail.
   */
  int iommu_dma_init_domain(struct iommu_domain *domain, dma_addr_t base,

@@ -291,21 +291,24 @@ int iommu_dma_init_domain(struct iommu_domain *domain, dma_addr_t base,
  {
  	struct iommu_dma_cookie *cookie = domain->iova_cookie;
  	struct iova_domain *iovad = &cookie->iovad;
-	unsigned long order, base_pfn, end_pfn;
+	dma_addr_t max_addr = base + size - 1;
+	unsigned long order, base_pfn;
  	int attr;
 
  	if (!cookie || cookie->type != IOMMU_DMA_IOVA_COOKIE)
  		return -EINVAL;
 
+	if (max_addr < base)
+		max_addr = U64_MAX;
+
  	/* Use the smallest supported page size for IOVA granularity */
  	order = __ffs(domain->pgsize_bitmap);
  	base_pfn = max_t(unsigned long, 1, base >> order);
-	end_pfn = (base + size - 1) >> order;
 
  	/* Check the domain allows at least some access to the device... */
  	if (domain->geometry.force_aperture) {
  		if (base > domain->geometry.aperture_end ||
-		    base + size <= domain->geometry.aperture_start) {
+		    max_addr < domain->geometry.aperture_start) {
  			pr_warn("specified DMA range outside IOMMU capability\n");
  			return -EFAULT;
  		}