iommu/dma: Zero pages manually in a length of scatterlist

The __GFP_ZERO will be passed down to the generic page allocation routine which zeros everything page by page. This is safe to be a generic way but not efficient for iommu allocation that organizes contiguous pages using scatterlist.

So this changes drops __GFP_ZERO from the flag, and adds a manual memset after page/sg allocations, using the length of scatterlist.

My test result of a 2.5MB size allocation shows iommu_dma_alloc() takes 46% less time, reduced from averagely 925 usec to 500 usec.

Signed-off-by Nicolin Chen <[email protected]>

---
  drivers/iommu/dma-iommu.c | 18 ++++++++++++++++++
  1 file changed, 18 insertions(+)

drivers/iommu/dma-iommu.c +3 -0

index d1b04753b204..e48d995e65c5 100644
@@ -551,10 +551,13 @@ struct page **iommu_dma_alloc(struct device *dev, size_t size, gfp_t gfp,
  	struct iommu_domain *domain = iommu_get_dma_domain(dev);
  	struct iommu_dma_cookie *cookie = domain->iova_cookie;
  	struct iova_domain *iovad = &cookie->iovad;
+	struct scatterlist *s;
  	struct page **pages;
  	struct sg_table sgt;
  	dma_addr_t iova;
  	unsigned int count, min_size, alloc_sizes = domain->pgsize_bitmap;
+	bool gfp_zero = false;
+	int i;
 
  	*handle = IOMMU_MAPPING_ERROR;
 

@@ -568,6 +571,15 @@ struct page **iommu_dma_alloc(struct device *dev, size_t size, gfp_t gfp,
  	if (attrs & DMA_ATTR_ALLOC_SINGLE_PAGES)

drivers/iommu/dma-iommu.c +2 -0

+	}
+
  	if (!(prot & IOMMU_CACHE)) {
  		struct sg_mapping_iter miter;
  		/*

Quoted 8 lines · Nicolin Chen

The __GFP_ZERO will be passed down to the generic page allocation routine which zeros everything page by page. This is safe to be a generic way but not efficient for iommu allocation that organizes contiguous pages using scatterlist.

So this changes drops __GFP_ZERO from the flag, and adds a manual memset after page/sg allocations, using the length of scatterlist.

Quoted 2 lines

My test result of a 2.5MB size allocation shows iommu_dma_alloc() takes 46% less time, reduced from averagely 925 usec to 500 usec.

Assuming this is for arm64, I'm somewhat surprised that memset() could be that much faster than clear_page(), since they should effectively amount to the same thing (a DC ZVA loop). What hardware is this on? Profiling to try

and see exactly where the extra time goes would be interesting too.

Quoted 5 lines

drivers/iommu/dma-iommu.c +2 -0

@@ -568,6 +571,15 @@ struct page **iommu_dma_alloc(struct device *dev, size_t size, gfp_t gfp,
  	if (attrs & DMA_ATTR_ALLOC_SINGLE_PAGES)
  		alloc_sizes = min_size;
+	/*
+	 * The generic zeroing in a length of one page size is slow,

drivers/iommu/dma-iommu.c struct page **iommu_dma_alloc(struct device *dev, size_t size, gfp_t gfp, · line 581

Quoted 6 lines

drivers/iommu/dma-iommu.c +6 -0

+	 * so do it manually in a length of scatterlist size instead
+	 */
+	if (gfp & __GFP_ZERO) {
+		gfp &= ~__GFP_ZERO;
+		gfp_zero = true;
+	}

Or just mask it out in __iommu_dma_alloc_pages()?

drivers/iommu/dma-iommu.c struct page **iommu_dma_alloc(struct device *dev, size_t size, gfp_t gfp, · line 598

Quoted 7 lines

drivers/iommu/dma-iommu.c +4 -0

@@ -581,6 +593,12 @@ struct page **iommu_dma_alloc(struct device *dev, size_t size, gfp_t gfp,
  	if (sg_alloc_table_from_pages(&sgt, pages, count, 0, size, GFP_KERNEL))
  		goto out_free_iova;
+	if (gfp_zero) {
+		/* Now zero all the pages in the scatterlist */
+		for_each_sg(sgt.sgl, s, sgt.orig_nents, i)
+			memset(sg_virt(s), 0, s->length);

What if the pages came from highmem? I know that doesn't happen on arm64 today, but the point of this code *is* to be generic, and other users will arrive eventually.

Quoted 3 lines

What if the pages came from highmem? I know that doesn't happen on arm64 today, but the point of this code *is* to be generic, and other users will arrive eventually.

Hmm, so it probably should use sg_miter_start/stop() too? Looking at the flush routine doing in PAGE_SIZE for each iteration, would be possible to map and memset contiguous pages together? Actually the flush routine might be also optimized if we can map contiguous pages.

Quoted 4 lines · Nicolin Chen

Quoted 3 lines

What if the pages came from highmem? I know that doesn't happen on arm64 today, but the point of this code *is* to be generic, and other users will arrive eventually.

Quoted 5 lines

Hmm, so it probably should use sg_miter_start/stop() too? Looking at the flush routine doing in PAGE_SIZE for each iteration, would be possible to map and memset contiguous pages together? Actually the flush routine might be also optimized if we can map contiguous pages.

FYI, I have patches I plan to submit soon that gets rid of the struct scatterlist use in this code to simplify it:

...and I have some significant objections to that simplification which I plan to respond with ;)

(namely that it defaults the whole higher-order page allocation business which will have varying degrees of performance impact on certain cases)

Quoted 9 lines · Robin Murphy

Quoted 8 lines · Nicolin Chen

The __GFP_ZERO will be passed down to the generic page allocation routine which zeros everything page by page. This is safe to be a generic way but not efficient for iommu allocation that organizes contiguous pages using scatterlist.

So this changes drops __GFP_ZERO from the flag, and adds a manual memset after page/sg allocations, using the length of scatterlist.

Quoted 6 lines

Quoted 2 lines

My test result of a 2.5MB size allocation shows iommu_dma_alloc() takes 46% less time, reduced from averagely 925 usec to 500 usec.

Assuming this is for arm64, I'm somewhat surprised that memset() could be that much faster than clear_page(), since they should effectively amount to the same thing (a DC ZVA loop). What hardware is this on? Profiling to try

I am running with tegra186-p2771-0000.dtb so it's arm64 yes.

Quoted 1 line

and see exactly where the extra time goes would be interesting too.

I re-ran the test to get some accuracy within the function and got:

1) pages = __iommu_dma_alloc_pages(count, alloc_sizes >> PAGE_SHIFT, gfp);
    // reduced from 422 usec to 56 usec == 366 usec less
2) if (!(prot & IOMMU_CACHE)) {...}	//flush routine
    // reduced from 439 usec to 236 usec == 203 usec less

Note: new memset takes about 164 usec, resulting in 400 usec diff
for the entire iommu_dma_alloc() function call.

It looks like this might be more than the diff between clear_page and memset, and might be related to mapping and cache. Any idea?

Quoted 5 lines

Quoted 5 lines

(quoted diff) +2 -0

@@ -568,6 +571,15 @@ struct page **iommu_dma_alloc(struct device *dev, size_t size, gfp_t gfp,
   	if (attrs & DMA_ATTR_ALLOC_SINGLE_PAGES)
   		alloc_sizes = min_size;
+	/*
+	 * The generic zeroing in a length of one page size is slow,

Quoted 8 lines

Review struct page **iommu_dma_alloc(struct device *dev, size_t size, gfp_t gfp, · line 580

Quoted 6 lines

(quoted diff) +6 -0

+	 * so do it manually in a length of scatterlist size instead
+	 */
+	if (gfp & __GFP_ZERO) {
+		gfp &= ~__GFP_ZERO;
+		gfp_zero = true;
+	}

Or just mask it out in __iommu_dma_alloc_pages()?

Yea, the change here would be neater then.

Quoted 11 lines

Review struct page **iommu_dma_alloc(struct device *dev, size_t size, gfp_t gfp, · line 598

Quoted 7 lines

(quoted diff) +4 -0

@@ -581,6 +593,12 @@ struct page **iommu_dma_alloc(struct device *dev, size_t size, gfp_t gfp,
   	if (sg_alloc_table_from_pages(&sgt, pages, count, 0, size, GFP_KERNEL))
   		goto out_free_iova;
+	if (gfp_zero) {
+		/* Now zero all the pages in the scatterlist */
+		for_each_sg(sgt.sgl, s, sgt.orig_nents, i)
+			memset(sg_virt(s), 0, s->length);

What if the pages came from highmem? I know that doesn't happen on arm64 today, but the point of this code *is* to be generic, and other users will arrive eventually.

Hmm, so it probably should use sg_miter_start/stop() too? Looking at the flush routine doing in PAGE_SIZE for each iteration, would be possible to map and memset contiguous pages together? Actually the flush routine might be also optimized if we can map contiguous pages.

Quoted 8 lines

I re-ran the test to get some accuracy within the function and got:

1) pages = __iommu_dma_alloc_pages(count, alloc_sizes >> PAGE_SHIFT, gfp);
    // reduced from 422 usec to 56 usec == 366 usec less
2) if (!(prot & IOMMU_CACHE)) {...}	//flush routine
    // reduced from 439 usec to 236 usec == 203 usec less

Note: new memset takes about 164 usec, resulting in 400 usec diff
for the entire iommu_dma_alloc() function call.

Quoted 2 lines

It looks like this might be more than the diff between clear_page and memset, and might be related to mapping and cache. Any idea?

Hmm, I guess it might not be so much clear_page() itself as all the gubbins involved in getting there from prep_new_page(). I could perhaps make some vague guesses about how the A57 cores might get tickled by the different code patterns, but the Denver cores are well beyond my ability to reason about. Out of even further curiosity, how does the quick hack below compare?

Quoted 12 lines

Quoted 11 lines

Review struct page **iommu_dma_alloc(struct device *dev, size_t size, gfp_t gfp, · line 598

Quoted 7 lines

(quoted diff) +4 -0

@@ -581,6 +593,12 @@ struct page **iommu_dma_alloc(struct device *dev, size_t size, gfp_t gfp,
   	if (sg_alloc_table_from_pages(&sgt, pages, count, 0, size, GFP_KERNEL))
   		goto out_free_iova;
+	if (gfp_zero) {
+		/* Now zero all the pages in the scatterlist */
+		for_each_sg(sgt.sgl, s, sgt.orig_nents, i)
+			memset(sg_virt(s), 0, s->length);

What if the pages came from highmem? I know that doesn't happen on arm64 today, but the point of this code *is* to be generic, and other users will arrive eventually.

Quoted 5 lines

Hmm, so it probably should use sg_miter_start/stop() too? Looking at the flush routine doing in PAGE_SIZE for each iteration, would be possible to map and memset contiguous pages together? Actually the flush routine might be also optimized if we can map contiguous pages.

I suppose the ideal point at which to do it would be after the remapping when we have the entire buffer contiguous in vmalloc space and can make best use of prefetchers etc. - DMA_ATTR_NO_KERNEL_MAPPING is a bit of a spanner in the works, but we could probably accommodate a special case for that. As Christoph points out, this isn't really the place to be looking for performance anyway (unless it's pathologically bad as per the

DMA_ATTR_ALLOC_SINGLE_PAGES fun), but if we're looking at pulling the remapping out of the arch code, maybe we could aim to rework the zeroing completely as part of that.

The __GFP_ZERO will be passed down to the generic page allocation routine which zeros everything page by page. This is safe to be a generic way but not efficient for iommu allocation that organizes contiguous pages using scatterlist.

So this changes drops __GFP_ZERO from the flag, and adds a manual memset after page/sg allocations, using the length of scatterlist.

My test result of a 2.5MB size allocation shows iommu_dma_alloc() takes 46% less time, reduced from averagely 925 usec to 500 usec.

Quoted 5 lines · Nicolin Chen

The __GFP_ZERO will be passed down to the generic page allocation routine which zeros everything page by page. This is safe to be a generic way but not efficient for iommu allocation that organizes contiguous pages using scatterlist.

Quoted 5 lines

So this changes drops __GFP_ZERO from the flag, and adds a manual memset after page/sg allocations, using the length of scatterlist.

My test result of a 2.5MB size allocation shows iommu_dma_alloc() takes 46% less time, reduced from averagely 925 usec to 500 usec.

And in what case does dma_alloc_* performance even matter?