Skip to contentSkip to conversation

iommu: Don't reserve IOVA when address and size are zero

PATCH In review Ashish Mhetre 2 participants 3 messages Nov 23, 2023 – Nov 28, 2023

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 iommu: Don't reserve IOVA when address and size are zero

When the bootloader/firmware doesn't setup the framebuffers, their address and size are zero in "iommu-addresses" property. If we intend to use display driver in kernel without framebuffer then it's causing the display IOMMU mappings to fail as IOVA is reserved with size and address as zero.
An ideal solution would be firmware removing the "iommu-addresses" property and corresponding "memory-region" if display is not present. But the kernel should be able to handle this by checking for size and address of IOVA and skipping the IOVA reservation if both are 0.

Fixes a5bf3cfce8cb ("iommu: Implement of_iommu_get_resv_regions()")
---
 drivers/iommu/of_iommu.c | 4 ++++
 1 file changed, 4 insertions(+)
drivers/iommu/of_iommu.c +4 -0
index 157b286e36bf..150ef65d357a 100644
@@ -255,6 +255,10 @@ void of_iommu_get_resv_regions(struct device *dev, struct list_head *list)
 				size_t length;
 
 				maps = of_translate_dma_region(np, maps, &iova, &length);
+				if (iova == 0 && length == 0) {
+					dev_dbg(dev, "Skipping IOVA reservation as address and size are zero\n");
+					continue;
+				}
 				type = iommu_resv_region_get_type(dev, &phys, iova, length);
 
 				region = iommu_alloc_resv_region(iova, length, prot, type,
Signature · 1 line
2.17.1
reviewer
#1 lore raw
Quoted 5 lines · Ashish Mhetre

When the bootloader/firmware doesn't setup the framebuffers, their address and size are zero in "iommu-addresses" property. If we intend to use display driver in kernel without framebuffer then it's causing the display IOMMU mappings to fail as IOVA is reserved with size and address as zero.

Can you clarify the problem there? Looking at the code in iova_reserve_iommu_regions() I'm guessing it's that "region->start + region->length - 1" underflows so reserve_iova() actually ends up reserving the entire valid IOVA space?

Quoted 4 lines

An ideal solution would be firmware removing the "iommu-addresses" property and corresponding "memory-region" if display is not present. But the kernel should be able to handle this by checking for size and address of IOVA and skipping the IOVA reservation if both are 0.

Surely it doesn't make sense to reserve a 0-length region at *any* base address? The symptom above wouldn't be quite the same if the base was nonzero, but corrupting the rbtree with an entry where pfn_hi < pfn_lo would definitely not be good either.

Quoted 11 lines
Fixes a5bf3cfce8cb ("iommu: Implement of_iommu_get_resv_regions()")
---
  drivers/iommu/of_iommu.c | 4 ++++
  1 file changed, 4 insertions(+)
drivers/iommu/of_iommu.c +0 -0
index 157b286e36bf..150ef65d357a 100644

drivers/iommu/of_iommu.c void of_iommu_get_resv_regions(struct device *dev, struct list_head *list) · line 259

Quoted 5 lines
drivers/iommu/of_iommu.c +2 -0
@@ -255,6 +255,10 @@ void of_iommu_get_resv_regions(struct device *dev, struct list_head *list)
  				size_t length;
 
  				maps = of_translate_dma_region(np, maps, &iova, &length);
+				if (iova == 0 && length == 0) {
+					dev_dbg(dev, "Skipping IOVA reservation as address and size are zero\n");

FWIW I'd be inclined to log a visible warning that firmware is giving us nonsense.

Thanks,
Robin.

Quoted 5 lines
drivers/iommu/of_iommu.c +2 -0
+					continue;
+				}
  				type = iommu_resv_region_get_type(dev, &phys, iova, length);
 
  				region = iommu_alloc_resv_region(iova, length, prot, type,
#2 lore raw
Quoted 10 lines · Robin Murphy

External email: Use caution opening links or attachments

Quoted 5 lines · Ashish Mhetre

When the bootloader/firmware doesn't setup the framebuffers, their address and size are zero in "iommu-addresses" property. If we intend to use display driver in kernel without framebuffer then it's causing the display IOMMU mappings to fail as IOVA is reserved with size and address as zero.

Quoted 4 lines

Can you clarify the problem there? Looking at the code in iova_reserve_iommu_regions() I'm guessing it's that "region->start + region->length - 1" underflows so reserve_iova() actually ends up reserving the entire valid IOVA space?

Yes, that's the problem which lead to dma_map call failures from display driver. I don't have the logs handy to pin-point the exact function which failed as this issue was seen before few months.

Quoted 11 lines
Quoted 5 lines

An ideal solution would be firmware removing the "iommu-addresses" property and corresponding "memory-region" if display is not present. But the kernel should be able to handle this by checking for size and address of IOVA and skipping the IOVA reservation if both are 0.

Surely it doesn't make sense to reserve a 0-length region at *any* base address? The symptom above wouldn't be quite the same if the base was nonzero, but corrupting the rbtree with an entry where pfn_hi < pfn_lo would definitely not be good either.

Agreed, we should restrict reservation for 0-length region at any base. I will update the condition in next version.

Quoted 14 lines
Quoted 14 lines
Fixes a5bf3cfce8cb ("iommu: Implement of_iommu_get_resv_regions()")

---
  drivers/iommu/of_iommu.c | 4 ++++

  1 file changed, 4 insertions(+)
drivers/iommu/of_iommu.c +0 -0
index 157b286e36bf..150ef65d357a 100644
@@ -255,6 +255,10 @@ void of_iommu_get_resv_regions(struct device 
 *dev, struct list_head *list)
                               size_t length;
 
Quoted 8 lines

drivers/iommu/of_iommu.c void of_iommu_get_resv_regions(struct device *dev, struct list_head *list)

Quoted 4 lines
drivers/iommu/of_iommu.c +2 -0
                               maps = of_translate_dma_region(np, maps, &iova, &length);
+                             if (iova == 0 && length == 0) {
+                                     dev_dbg(dev, "Skipping IOVA 
 reservation as address and size are zero\n");

FWIW I'd be inclined to log a visible warning that firmware is giving us nonsense.

Okay, I'll replace dev_dbg() with dev_warn() in next version.

Quoted 8 lines

Thanks,
Robin.

Quoted 5 lines
drivers/iommu/of_iommu.c +2 -0
+ continue;
+                             }
                               type = iommu_resv_region_get_type(dev, &phys, iova, length);

                              region = iommu_alloc_resv_region(iova, length, prot, type,