Skip to contentSkip to conversation

arm-smmu: check for generic bindings first

PATCH In review Stefano Stabellini 2 participants 3 messages Aug 28, 2019 IOMMU

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 arm-smmu: check for generic bindings first

From: Stefano Stabellini <[email protected]>

Today, the arm-smmu driver checks for mmu-masters on device tree, the legacy property, if it is absent it assumes that we are using the new bindings. If it is present it assumes that we are using the legacy bindings. All arm-smmus need to use the same bindings: legacy or new.

There are two issues with this:

- we are not actually checking for the new bindings explicitly It would be better to have an explicit check for the new bindings rather than assuming we must be using the new if the old are not there.

- old and new bindings cannot coexist
It would be nice to be able to provide both old and new bindings so that software that hasn't been updated yet is still able to get IOMMU information from the legacy bindings while at the same time newer software can get the latest information via the new bindings. (Xen has not been updated to use the new binding yet for instance.) The current code breaks under these circumstances because if the old bindings are present, the new are not even checked.

This patch changes the scheme by checking for #iommu-cells, which is only present with the new bindings, before checking for mmu-masters. The new bindings are always favored when present. All SMMUs still need to use the same bindings: mix-and-matching new and old bindings between different SMMUs is not allowed.

---

Let me know if you'd like me to turn the two using_*_binding variables into a single one.

Also, please note that this is not meant as an excuse not to get Xen updated to use the new binding.

 drivers/iommu/arm-smmu.c | 22 +++++++++++++---------
 1 file changed, 13 insertions(+), 9 deletions(-)
drivers/iommu/arm-smmu.c +13 -9
index 64977c131ee6..79b518ff215c 100644
@@ -2118,7 +2118,7 @@ static int arm_smmu_device_dt_probe(struct platform_device *pdev,
 {
 	const struct arm_smmu_match_data *data;
 	struct device *dev = &pdev->dev;
-	bool legacy_binding;
+	bool legacy_binding, generic_binding;
 
 	if (of_property_read_u32(dev->of_node, "#global-interrupts",
 				 &smmu->num_global_irqs)) {
@@ -2132,16 +2132,20 @@ static int arm_smmu_device_dt_probe(struct platform_device *pdev,
 
 	parse_driver_options(smmu);
 
-	legacy_binding = of_find_property(dev->of_node, "mmu-masters", NULL);
-	if (legacy_binding && !using_generic_binding) {
-		if (!using_legacy_binding)
-			pr_notice("deprecated \"mmu-masters\" DT property in use; DMA API support unavailable\n");
-		using_legacy_binding = true;
-	} else if (!legacy_binding && !using_legacy_binding) {
+	generic_binding = of_find_property(dev->of_node, "#iommu-cells", NULL);
+	if (generic_binding && !using_legacy_binding) {
 		using_generic_binding = true;
 	} else {
-		dev_err(dev, "not probing due to mismatched DT properties\n");
-		return -ENODEV;
+		legacy_binding = of_find_property(dev->of_node, "mmu-masters",
+						  NULL);
+		if (legacy_binding && !using_generic_binding) {
+			if (!using_legacy_binding)
+				pr_notice("deprecated \"mmu-masters\" DT property in use; DMA API support unavailable\n");
+			using_legacy_binding = true;
+		} else {
+			dev_err(dev, "not probing due to mismatched DT properties\n");
+			return -ENODEV;
+		}
 	}
 
 	if (of_dma_is_coherent(dev->of_node))
Signature · 1 line
2.17.1
reviewer
#1 lore raw
Quoted 7 lines · Stefano Stabellini

From: Stefano Stabellini <[email protected]>

Today, the arm-smmu driver checks for mmu-masters on device tree, the legacy property, if it is absent it assumes that we are using the new bindings. If it is present it assumes that we are using the legacy bindings. All arm-smmus need to use the same bindings: legacy or new.

Quoted 5 lines

There are two issues with this:

- we are not actually checking for the new bindings explicitly It would be better to have an explicit check for the new bindings rather than assuming we must be using the new if the old are not there.

The legacy binding is the special case, though. Isn't it perfectly logical to check for the special case, and assume the normal case otherwise? ;)

Quoted 8 lines

- old and new bindings cannot coexist
It would be nice to be able to provide both old and new bindings so that software that hasn't been updated yet is still able to get IOMMU information from the legacy bindings while at the same time newer software can get the latest information via the new bindings. (Xen has not been updated to use the new binding yet for instance.) The current code breaks under these circumstances because if the old bindings are present, the new are not even checked.

FWIW that was a deliberate design decision. We didn't want to see DTs with both bindings at once - the legacy binding support remains only for the sake of DTBs deployed in fossilised firmware that can't/won't ever be updated. And I'm pretty sure the intervening 3 years have only reinforced that position. Only a couple more months now until it will have been deprecated for longer than it wasn't :)

Do note that "mmu-masters" isn't just deprecated for giggles either - the binding made some flawed assumptions, and there are Stream ID topologies that it cannot possibly describe, which do exist in the real world.

Quoted 6 lines

This patch changes the scheme by checking for #iommu-cells, which is only present with the new bindings, before checking for mmu-masters. The new bindings are always favored when present. All SMMUs still need to use the same bindings: mix-and-matching new and old bindings between different SMMUs is not allowed.

Quoted 8 lines

---

Let me know if you'd like me to turn the two using_*_binding variables into a single one.

Also, please note that this is not meant as an excuse not to get Xen updated to use the new binding.

Fair enough, but conversely, what makes it Linux's responsibility to handle? If Xen wants to deal with funky hybrid DTs then why shouldn't it be Xen's job to just filter deprecated properties out of whatever it presents to Linux?

Robin.

Quoted 46 lines
  drivers/iommu/arm-smmu.c | 22 +++++++++++++---------
  1 file changed, 13 insertions(+), 9 deletions(-)
drivers/iommu/arm-smmu.c +13 -9
index 64977c131ee6..79b518ff215c 100644
@@ -2118,7 +2118,7 @@ static int arm_smmu_device_dt_probe(struct platform_device *pdev,
  {
  	const struct arm_smmu_match_data *data;
  	struct device *dev = &pdev->dev;
-	bool legacy_binding;
+	bool legacy_binding, generic_binding;
 
  	if (of_property_read_u32(dev->of_node, "#global-interrupts",
  				 &smmu->num_global_irqs)) {
@@ -2132,16 +2132,20 @@ static int arm_smmu_device_dt_probe(struct platform_device *pdev,
 
  	parse_driver_options(smmu);
 
-	legacy_binding = of_find_property(dev->of_node, "mmu-masters", NULL);
-	if (legacy_binding && !using_generic_binding) {
-		if (!using_legacy_binding)
-			pr_notice("deprecated \"mmu-masters\" DT property in use; DMA API support unavailable\n");
-		using_legacy_binding = true;
-	} else if (!legacy_binding && !using_legacy_binding) {
+	generic_binding = of_find_property(dev->of_node, "#iommu-cells", NULL);
+	if (generic_binding && !using_legacy_binding) {
  		using_generic_binding = true;
  	} else {
-		dev_err(dev, "not probing due to mismatched DT properties\n");
-		return -ENODEV;
+		legacy_binding = of_find_property(dev->of_node, "mmu-masters",
+						  NULL);
+		if (legacy_binding && !using_generic_binding) {
+			if (!using_legacy_binding)
+				pr_notice("deprecated \"mmu-masters\" DT property in use; DMA API support unavailable\n");
+			using_legacy_binding = true;
+		} else {
+			dev_err(dev, "not probing due to mismatched DT properties\n");
+			return -ENODEV;
+		}
  	}
 
  	if (of_dma_is_coherent(dev->of_node))
#2 lore raw
Quoted 8 lines · Robin Murphy
Quoted 7 lines · Stefano Stabellini

From: Stefano Stabellini <[email protected]>

Today, the arm-smmu driver checks for mmu-masters on device tree, the legacy property, if it is absent it assumes that we are using the new bindings. If it is present it assumes that we are using the legacy bindings. All arm-smmus need to use the same bindings: legacy or new.

Quoted 8 lines
Quoted 5 lines

There are two issues with this:

- we are not actually checking for the new bindings explicitly It would be better to have an explicit check for the new bindings rather than assuming we must be using the new if the old are not there.

The legacy binding is the special case, though. Isn't it perfectly logical to check for the special case, and assume the normal case otherwise? ;)

I don't know ... all cases are special and all cases are normal depending on who you ask or when. The normal case of today can be the special case of tomorrow and vice versa.

Quoted 9 lines
Quoted 8 lines

- old and new bindings cannot coexist
It would be nice to be able to provide both old and new bindings so that software that hasn't been updated yet is still able to get IOMMU information from the legacy bindings while at the same time newer software can get the latest information via the new bindings. (Xen has not been updated to use the new binding yet for instance.) The current code breaks under these circumstances because if the old bindings are present, the new are not even checked.

Quoted 6 lines

FWIW that was a deliberate design decision. We didn't want to see DTs with both bindings at once - the legacy binding support remains only for the sake of DTBs deployed in fossilised firmware that can't/won't ever be updated. And I'm pretty sure the intervening 3 years have only reinforced that position. Only a couple more months now until it will have been deprecated for longer than it wasn't :)

This way the people that would be willing to put in extra effort to maintain compatibility wouldn't be able to achieve the goal.

Quoted 16 lines

Do note that "mmu-masters" isn't just deprecated for giggles either - the binding made some flawed assumptions, and there are Stream ID topologies that it cannot possibly describe, which do exist in the real world.

Quoted 12 lines

This patch changes the scheme by checking for #iommu-cells, which is only present with the new bindings, before checking for mmu-masters. The new bindings are always favored when present. All SMMUs still need to use the same bindings: mix-and-matching new and old bindings between different SMMUs is not allowed.

---

Let me know if you'd like me to turn the two using_*_binding variables into a single one.

Quoted 6 lines
Quoted 2 lines

Also, please note that this is not meant as an excuse not to get Xen updated to use the new binding.

Fair enough, but conversely, what makes it Linux's responsibility to handle? If Xen wants to deal with funky hybrid DTs then why shouldn't it be Xen's job to just filter deprecated properties out of whatever it presents to Linux?

Yes, Xen does it just fine. But if you boot Linux with the same DT but without Xen, problems occur.

The code looked suboptimal to me because of the reasons I listed in the commit message. I thought it was done this way by "mistake". However, I now realize that if it is written this way on purpose, then I have a very little chance of convincing you to change it.

Quoted 49 lines
Quoted 48 lines
  drivers/iommu/arm-smmu.c | 22 +++++++++++++---------
  1 file changed, 13 insertions(+), 9 deletions(-)
drivers/iommu/arm-smmu.c +13 -8
index 64977c131ee6..79b518ff215c 100644
@@ -2118,7 +2118,7 @@ static int arm_smmu_device_dt_probe(struct
 platform_device *pdev,
  {
  	const struct arm_smmu_match_data *data;
  	struct device *dev = &pdev->dev;
-	bool legacy_binding;
+	bool legacy_binding, generic_binding;
    	if (of_property_read_u32(dev->of_node, "#global-interrupts",
  				 &smmu->num_global_irqs)) {
@@ -2132,16 +2132,20 @@ static int arm_smmu_device_dt_probe(struct
 platform_device *pdev,
    	parse_driver_options(smmu);
  -	legacy_binding = of_find_property(dev->of_node, "mmu-masters", NULL);
-	if (legacy_binding && !using_generic_binding) {
-		if (!using_legacy_binding)
-			pr_notice("deprecated \"mmu-masters\" DT property in
 use; DMA API support unavailable\n");
-		using_legacy_binding = true;
-	} else if (!legacy_binding && !using_legacy_binding) {
+	generic_binding = of_find_property(dev->of_node, "#iommu-cells",
 NULL);
+	if (generic_binding && !using_legacy_binding) {
  		using_generic_binding = true;
  	} else {
-		dev_err(dev, "not probing due to mismatched DT properties\n");
-		return -ENODEV;
+		legacy_binding = of_find_property(dev->of_node, "mmu-masters",
+						  NULL);
+		if (legacy_binding && !using_generic_binding) {
+			if (!using_legacy_binding)
+				pr_notice("deprecated \"mmu-masters\" DT
 property in use; DMA API support unavailable\n");
+			using_legacy_binding = true;
+		} else {
+			dev_err(dev, "not probing due to mismatched DT
 properties\n");
+			return -ENODEV;
+		}
  	}
    	if (of_dma_is_coherent(dev->of_node))