Skip to contentSkip to conversation

iommu/dma: Use NUMA aware memory allocations in __iommu_dma_alloc_pages()

PATCH Superseded John Garry 2 participants 3 messages Nov 20, 2018

Subsystems (2)

v2 of 4 Superseded Latest is v4 →

lore
Share & tools
lore

John Garry

author
#0 lore raw

PATCH v2 1/1 iommu/dma: Use NUMA aware memory allocations in __iommu_dma_alloc_pages()

From: Ganapatrao Kulkarni <[email protected]>

Change function __iommu_dma_alloc_pages() to allocate memory/pages for DMA from respective device NUMA node.

[JPG: Modifed to use kvzalloc() and fixed indentation]

---
Difference v1->v2:
- Add Ganapatrao's tag and change author

This patch was originally posted by Ganapatrao in [1].

However, after initial review, it was never reposted (due to lack of cycles, I think). In addition, the functionality in its sibling patches were merged through patches, as mentioned in [2]; this also refers to a discussion on device local allocations vs CPU local allocations for DMA pool, and which is better [3].

However, as mentioned in [3], dma_alloc_coherent() uses the locality information from the device - as in direct DMA - so this patch is just applying this same policy.

[1] https://lore.kernel.org/patchwork/patch/833004/
[2] https://lkml.org/lkml/2018/8/22/391
[3] https://www.mail-archive.com/[email protected]/msg1692998.html

drivers/iommu/dma-iommu.c +10 -10
index d1b0475..ada00bc 100644
@@ -449,20 +449,17 @@ static void __iommu_dma_free_pages(struct page **pages, int count)
 	kvfree(pages);
 }
 
-static struct page **__iommu_dma_alloc_pages(unsigned int count,
-		unsigned long order_mask, gfp_t gfp)
+static struct page **__iommu_dma_alloc_pages(struct device *dev,
+		unsigned int count, unsigned long order_mask, gfp_t gfp)
 {
 	struct page **pages;
-	unsigned int i = 0, array_size = count * sizeof(*pages);
+	unsigned int i = 0, nid = dev_to_node(dev);
 
 	order_mask &= (2U << MAX_ORDER) - 1;
 	if (!order_mask)
 		return NULL;
 
-	if (array_size <= PAGE_SIZE)
-		pages = kzalloc(array_size, GFP_KERNEL);
-	else
-		pages = vzalloc(array_size);
+	pages = kvzalloc_node(count * sizeof(*pages), GFP_KERNEL, nid);
 	if (!pages)
 		return NULL;
 
@@ -483,8 +480,10 @@ static struct page **__iommu_dma_alloc_pages(unsigned int count,
 			unsigned int order = __fls(order_mask);
 
 			order_size = 1U << order;
-			page = alloc_pages((order_mask - order_size) ?
-					   gfp | __GFP_NORETRY : gfp, order);
+			page = alloc_pages_node(nid,
+						(order_mask - order_size) ?
+						gfp | __GFP_NORETRY : gfp,
+						order);
 			if (!page)
 				continue;
 			if (!order)
@@ -569,7 +568,8 @@ struct page **iommu_dma_alloc(struct device *dev, size_t size, gfp_t gfp,
 		alloc_sizes = min_size;
 
 	count = PAGE_ALIGN(size) >> PAGE_SHIFT;
-	pages = __iommu_dma_alloc_pages(count, alloc_sizes >> PAGE_SHIFT, gfp);
+	pages = __iommu_dma_alloc_pages(dev, count, alloc_sizes >> PAGE_SHIFT,
+					gfp);
 	if (!pages)
 		return NULL;
 
Signature · 1 line
1.9.1
maintainer
#1 lore raw
Quoted 45 lines · John Garry

From: Ganapatrao Kulkarni <[email protected]>

Change function __iommu_dma_alloc_pages() to allocate memory/pages for DMA from respective device NUMA node.

[JPG: Modifed to use kvzalloc() and fixed indentation]

---
Difference v1->v2:
- Add Ganapatrao's tag and change author

This patch was originally posted by Ganapatrao in [1].

However, after initial review, it was never reposted (due to lack of cycles, I think). In addition, the functionality in its sibling patches were merged through patches, as mentioned in [2]; this also refers to a discussion on device local allocations vs CPU local allocations for DMA pool, and which is better [3].

However, as mentioned in [3], dma_alloc_coherent() uses the locality information from the device - as in direct DMA - so this patch is just applying this same policy.

[1] https://lore.kernel.org/patchwork/patch/833004/
[2] https://lkml.org/lkml/2018/8/22/391
[3] https://www.mail-archive.com/[email protected]/msg1692998.html

drivers/iommu/dma-iommu.c +3 -3
index d1b0475..ada00bc 100644
@@ -449,20 +449,17 @@ static void __iommu_dma_free_pages(struct page **pages, int count)
  	kvfree(pages);
  }
 
-static struct page **__iommu_dma_alloc_pages(unsigned int count,
-		unsigned long order_mask, gfp_t gfp)
+static struct page **__iommu_dma_alloc_pages(struct device *dev,
+		unsigned int count, unsigned long order_mask, gfp_t gfp)
  {
  	struct page **pages;
-	unsigned int i = 0, array_size = count * sizeof(*pages);
+	unsigned int i = 0, nid = dev_to_node(dev);

drivers/iommu/dma-iommu.c static void __iommu_dma_free_pages(struct page **pages, int count) · line 462

Quoted 10 lines
drivers/iommu/dma-iommu.c +1 -4
 
  	order_mask &= (2U << MAX_ORDER) - 1;
  	if (!order_mask)
  		return NULL;
 
-	if (array_size <= PAGE_SIZE)
-		pages = kzalloc(array_size, GFP_KERNEL);
-	else
-		pages = vzalloc(array_size);
+	pages = kvzalloc_node(count * sizeof(*pages), GFP_KERNEL, nid);

The pages array is only accessed by the CPU servicing the iommu_dma_alloc() call, and is usually freed again before that call even returns. It's certainly never touched by the device, so forcing it to a potentially non-local node doesn't make a great deal of sense.

drivers/iommu/dma-iommu.c static struct page **__iommu_dma_alloc_pages(unsigned int count, · line 486

Quoted 13 lines
drivers/iommu/dma-iommu.c +4 -2
  	if (!pages)
  		return NULL;
 
@@ -483,8 +480,10 @@ static struct page **__iommu_dma_alloc_pages(unsigned int count,
  			unsigned int order = __fls(order_mask);
 
  			order_size = 1U << order;
-			page = alloc_pages((order_mask - order_size) ?
-					   gfp | __GFP_NORETRY : gfp, order);
+			page = alloc_pages_node(nid,
+						(order_mask - order_size) ?
+						gfp | __GFP_NORETRY : gfp,
+						order);

If we're touching this, can we sort out that horrendous ternary? FWIW I found I have a local version of the original patch which I tweaked at the time, and apparently I reworked this hunk as below, which does seem somewhat nicer for the same diffstat.

Robin.

@@ -446,10 +443,12 @@ static struct page **__iommu_dma_alloc_pages(unsigned int count,

                for (order_mask &= (2U << __fls(count)) - 1;
                     order_mask; order_mask &= ~order_size) {
                        unsigned int order = __fls(order_mask);

+ gfp_t alloc_flags = gfp;

                        order_size = 1U << order;

- page = alloc_pages((order_mask - order_size) ?
- gfp | __GFP_NORETRY : gfp, order);
+ if (order_size < order_mask)
+ alloc_flags |= __GFP_NORETRY;

+                       page = alloc_pages_node(nid, alloc_flags, order);
                        if (!page)
                                continue;
                        if (!order)
author
#2 lore raw
Quoted 52 lines · Robin Murphy
Quoted 51 lines · John Garry

From: Ganapatrao Kulkarni <ganapatrao.kulkarni-YGCgFSpz5w/[email protected]>

Change function __iommu_dma_alloc_pages() to allocate memory/pages for DMA from respective device NUMA node.

[JPG: Modifed to use kvzalloc() and fixed indentation]

---
Difference v1->v2:
- Add Ganapatrao's tag and change author

This patch was originally posted by Ganapatrao in [1].

However, after initial review, it was never reposted (due to lack of cycles, I think). In addition, the functionality in its sibling patches were merged through patches, as mentioned in [2]; this also refers to a discussion on device local allocations vs CPU local allocations for DMA pool, and which is better [3].

However, as mentioned in [3], dma_alloc_coherent() uses the locality information from the device - as in direct DMA - so this patch is just applying this same policy.

[1] https://lore.kernel.org/patchwork/patch/833004/
[2] https://lkml.org/lkml/2018/8/22/391
[3]
https://www.mail-archive.com/[email protected]/msg1692998.html

drivers/iommu/dma-iommu.c +3 -3
index d1b0475..ada00bc 100644
@@ -449,20 +449,17 @@ static void __iommu_dma_free_pages(struct page
 **pages, int count)
      kvfree(pages);
  }
  -static struct page **__iommu_dma_alloc_pages(unsigned int count,
-        unsigned long order_mask, gfp_t gfp)
+static struct page **__iommu_dma_alloc_pages(struct device *dev,
+        unsigned int count, unsigned long order_mask, gfp_t gfp)
  {
      struct page **pages;
-    unsigned int i = 0, array_size = count * sizeof(*pages);
+    unsigned int i = 0, nid = dev_to_node(dev);
        order_mask &= (2U << MAX_ORDER) - 1;
      if (!order_mask)
          return NULL;
  -    if (array_size <= PAGE_SIZE)
-        pages = kzalloc(array_size, GFP_KERNEL);
Quoted 8 lines

drivers/iommu/dma-iommu.c static void __iommu_dma_free_pages(struct page **pages, int count) · line 462

Quoted 3 lines
drivers/iommu/dma-iommu.c +1 -2
-    else
-        pages = vzalloc(array_size);
+    pages = kvzalloc_node(count * sizeof(*pages), GFP_KERNEL, nid);

The pages array is only accessed by the CPU servicing the iommu_dma_alloc() call, and is usually freed again before that call even returns. It's certainly never touched by the device, so forcing it to a potentially non-local node doesn't make a great deal of sense.

Right, it seems sensible to not make this allocation include the device-locality requirement, so can leave as is. However modifying to use kvzalloc() would seem ok.

Quoted 25 lines

drivers/iommu/dma-iommu.c static struct page **__iommu_dma_alloc_pages(unsigned int count, · line 486

Quoted 13 lines
drivers/iommu/dma-iommu.c +0 -0
      if (!pages)
          return NULL;
  @@ -483,8 +480,10 @@ static struct page

**__iommu_dma_alloc_pages(unsigned int count,

              unsigned int order = __fls(order_mask);
                order_size = 1U << order;
drivers/iommu/dma-iommu.c +4 -2
-            page = alloc_pages((order_mask - order_size) ?
-                       gfp | __GFP_NORETRY : gfp, order);
+            page = alloc_pages_node(nid,
+                        (order_mask - order_size) ?
+                        gfp | __GFP_NORETRY : gfp,
+                        order);

If we're touching this, can we sort out that horrendous ternary? FWIW I found I have a local version of the original patch which I tweaked at the time, and apparently I reworked this hunk as below, which does seem somewhat nicer for the same diffstat.

Robin.

(quoted diff) +0 -0
@@ -446,10 +443,12 @@ static struct page
 **__iommu_dma_alloc_pages(unsigned int count,
                for (order_mask &= (2U << __fls(count)) - 1;

Review static struct page · line 452

Quoted 10 lines
(quoted diff) +3 -2
                     order_mask; order_mask &= ~order_size) {
                        unsigned int order = __fls(order_mask);
+                       gfp_t alloc_flags = gfp;
 
                        order_size = 1U << order;
-                       page = alloc_pages((order_mask - order_size) ?
-                                          gfp | __GFP_NORETRY : gfp,
 order);
+                       if (order_size < order_mask)
+                               alloc_flags |= __GFP_NORETRY;

Sure, this can be included

Quoted 7 lines
(quoted diff) +1 -0
+                       page = alloc_pages_node(nid, alloc_flags, order);
                        if (!page)
                                continue;
                        if (!order)

.

Cheers,
John