Skip to contentSkip to conversation

iommu/amd: flush not present cache in iommu_map_page

PATCH In review Tom Murphy 3 participants 6 messages Apr 24, 2019 IOMMU

lore
Share & tools
lore

Tom Murphy

author
#0 lore raw

PATCH 1/1 iommu/amd: flush not present cache in iommu_map_page

check if there is a not-present cache present and flush it if there is.

---
 drivers/iommu/amd_iommu.c | 6 ++++++
 1 file changed, 6 insertions(+)
drivers/iommu/amd_iommu.c +6 -0
index f7cdd2ab7f11..8ef43224aae0 100644
@@ -1636,6 +1636,12 @@ static int iommu_map_page(struct protection_domain *dom,
 		pte[i] = __pte;
 
 	update_domain(dom);
+	if (!dom->updated) {
+		if (unlikely(amd_iommu_np_cache)) {
+			domain_flush_pages(dom, bus_addr, page_size);
+			domain_flush_complete(dom);
+		}
+	}
 
 	/* Everything flushed out, free pages now */
 	free_page_list(freelist);
Signature · 1 line
2.17.1
#1 lore raw
Quoted 12 lines · Tom Murphy via iommu

check if there is a not-present cache present and flush it if there is.

---
 drivers/iommu/amd_iommu.c | 6 ++++++
 1 file changed, 6 insertions(+)
drivers/iommu/amd_iommu.c +0 -0
index f7cdd2ab7f11..8ef43224aae0 100644

drivers/iommu/amd_iommu.c static int iommu_map_page(struct protection_domain *dom, · line 1644

Quoted 9 lines
drivers/iommu/amd_iommu.c +6 -0
@@ -1636,6 +1636,12 @@ static int iommu_map_page(struct protection_domain *dom,
 		pte[i] = __pte;
 
 	update_domain(dom);
+	if (!dom->updated) {
+		if (unlikely(amd_iommu_np_cache)) {
+			domain_flush_pages(dom, bus_addr, page_size);
+			domain_flush_complete(dom);
+		}
+	}

The two conditions can go into one if statement to make this a little more clear.

And I'd really like to understand the unlikely - amd_iommu_np_cache is set based on a hardware capability, so it seems rather odd to mark it unlikely. Dynamic branch prediction really should do the right thing here usually.

author
#2 lore raw
Quoted 2 lines

The two conditions can go into one if statement to make this a little more clear.

Ah, yeah of course

Quoted 4 lines

And I'd really like to understand the unlikely - amd_iommu_np_cache is set based on a hardware capability, so it seems rather odd to mark it unlikely. Dynamic branch prediction really should do the right thing here usually.

Here is the commit which added it without any explanation:
https://github.com/torvalds/linux/commit/270cab2426cdc6307725e4f1f46ecf8ab8e69193

should we remove it seen as there's no explanation given ?

Quoted 31 lines · Christoph Hellwig
Quoted 12 lines · Tom Murphy via iommu

check if there is a not-present cache present and flush it if there is.

---
 drivers/iommu/amd_iommu.c | 6 ++++++
 1 file changed, 6 insertions(+)
drivers/iommu/amd_iommu.c +0 -0
index f7cdd2ab7f11..8ef43224aae0 100644

drivers/iommu/amd_iommu.c static int iommu_map_page(struct protection_domain *dom, · line 1644

Quoted 9 lines
drivers/iommu/amd_iommu.c +6 -0
@@ -1636,6 +1636,12 @@ static int iommu_map_page(struct protection_domain *dom,
              pte[i] = __pte;
 
      update_domain(dom);
+     if (!dom->updated) {
+             if (unlikely(amd_iommu_np_cache)) {
+                     domain_flush_pages(dom, bus_addr, page_size);
+                     domain_flush_complete(dom);
+             }
+     }

The two conditions can go into one if statement to make this a little more clear.

And I'd really like to understand the unlikely - amd_iommu_np_cache is set based on a hardware capability, so it seems rather odd to mark it unlikely. Dynamic branch prediction really should do the right thing here usually.

#3 lore raw
Quoted 9 lines · Tom Murphy
Quoted 4 lines

And I'd really like to understand the unlikely - amd_iommu_np_cache is set based on a hardware capability, so it seems rather odd to mark it unlikely. Dynamic branch prediction really should do the right thing here usually.

Here is the commit which added it without any explanation:
https://github.com/torvalds/linux/commit/270cab2426cdc6307725e4f1f46ecf8ab8e69193

should we remove it seen as there's no explanation given ?

I'd be tempted to do that. But lets just ask Joerg if he has any opinion..

maintainer
#4 lore raw
Quoted 2 lines · Christoph Hellwig

I'd be tempted to do that. But lets just ask Joerg if he has any opinion..

The reason was that it is an unlikely path, as hardware implementations are not allowed to set this bit. It is purely for emulated AMD IOMMUs. I have not measured whether this annotation has any performance benefit, but I find it more readable at least.

Regards,

Joerg

PS: Why did you drop me from the Cc list of the previous replies?

author
#5 lore raw
Quoted 9 lines · Joerg Roedel
Quoted 2 lines · Christoph Hellwig

I'd be tempted to do that. But lets just ask Joerg if he has any opinion..

The reason was that it is an unlikely path, as hardware implementations are not allowed to set this bit. It is purely for emulated AMD IOMMUs. I have not measured whether this annotation has any performance benefit, but I find it more readable at least.

In that case I will keep it in.

Quoted 6 lines

Regards,

Joerg

PS: Why did you drop me from the Cc list of the previous replies?