Skip to contentSkip to conversation

dma-direct: Skip cache prep for HighMem coherent allocations

PATCH In review Aneesh Kumar K.V (Arm) 4 participants 14 messages Jan 2 – Jan 20 DMA mapping helpers

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 dma-direct: Skip cache prep for HighMem coherent allocations

dma_direct_alloc() calls arch_dma_prep_coherent() to clean any dirty cache lines from the kernel linear alias before creating a coherent remapping.

HighMem pages have no kernel alias mapping, so there are no alias cache lines to clean. Skip arch_dma_prep_coherent() for HighMem allocations.

---
 kernel/dma/direct.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)
kernel/dma/direct.c +2 -1
index 50c3fe2a1d55..ffa267020a1e 100644
@@ -272,7 +272,8 @@ void *dma_direct_alloc(struct device *dev, size_t size,
 			prot = pgprot_decrypted(prot);
 
 		/* remove any dirty cache lines on the kernel alias */
-		arch_dma_prep_coherent(page, size);
+		if (!PageHighMem(page))
+			arch_dma_prep_coherent(page, size);
 
 		/* create a coherent mapping */
 		ret = dma_common_contiguous_remap(page, size, prot,
Signature · 1 line
2.43.0
maintainer
#1 lore raw
Quoted 8 lines · Aneesh Kumar K.V (Arm)

dma_direct_alloc() calls arch_dma_prep_coherent() to clean any dirty cache lines from the kernel linear alias before creating a coherent remapping.

HighMem pages have no kernel alias mapping, so there are no alias cache lines to clean. Skip arch_dma_prep_coherent() for HighMem allocations.

Indeed this is an overhead to call prep for HighMem pages, but on the other hand highmem support is being phased out according to
https://lwn.net/ml/all/[email protected]/ Does it make sense to apply this assuming that it will be removed soon?

Quoted 9 lines
---
  kernel/dma/direct.c | 3 ++-
  1 file changed, 2 insertions(+), 1 deletion(-)
kernel/dma/direct.c +0 -0
index 50c3fe2a1d55..ffa267020a1e 100644

kernel/dma/direct.c void *dma_direct_alloc(struct device *dev, size_t size, · line 279

Quoted 9 lines
kernel/dma/direct.c +2 -1
@@ -272,7 +272,8 @@ void *dma_direct_alloc(struct device *dev, size_t size,
  			prot = pgprot_decrypted(prot);
 
  		/* remove any dirty cache lines on the kernel alias */
-		arch_dma_prep_coherent(page, size);
+		if (!PageHighMem(page))
+			arch_dma_prep_coherent(page, size);
 
  		/* create a coherent mapping */
  		ret = dma_common_contiguous_remap(page, size, prot,

Best regards

Signature · 2 lines
Marek Szyprowski, PhD
Samsung R&D Institute Poland
#2 lore raw
Quoted 8 lines · Marek Szyprowski
Quoted 7 lines · Aneesh Kumar K.V (Arm)

dma_direct_alloc() calls arch_dma_prep_coherent() to clean any dirty cache lines from the kernel linear alias before creating a coherent remapping.

HighMem pages have no kernel alias mapping, so there are no alias cache lines to clean. Skip arch_dma_prep_coherent() for HighMem allocations.

Quoted 6 lines
Quoted 1 line

Indeed this is an overhead to call prep for HighMem pages, but on the other hand highmem support is being phased out according to
https://lwn.net/ml/all/[email protected]/ Does it make sense to apply this assuming that it will be removed soon?

I think "soon" is overstating what the plan is. With my proposed series, the majority of current highmem users are changed to no longer use it by default, but there are still three ways in which users will get highmem for a number of years:

- anything with more than 2GB RAM inevitably uses highmem for the top portion of physical memory. These are much less common than 2GB systems, but are not going away soon.
- systems with sparse physical memory where the first and last page are more than 2GB apart currently still rely on highmem even if the total RAM is 2GB or less. This happens e.g. on Tegra114 or RZ-G1H, IIRC. I have plans to address this in the future.
- Users that have applications using a lot of virtual memory still have the option to go back to old configurations with CONFIG_VMSPLIT_3G by selecting CONFIG_EXPERT.

On the other hand, my proposed change does mean that we have more freedom to optimize for the non-highmem case. I think we can require that the CMA area is in lowmem, and we can make the use of highmem computationally more expensive if it helps simplify code.

Arnd

maintainer
#3 lore raw
Quoted 30 lines · Arnd Bergmann
Quoted 5 lines · Marek Szyprowski
Quoted 4 lines · Aneesh Kumar K.V (Arm)

dma_direct_alloc() calls arch_dma_prep_coherent() to clean any dirty cache lines from the kernel linear alias before creating a coherent remapping.

Quoted 8 lines
Quoted 4 lines

HighMem pages have no kernel alias mapping, so there are no alias cache lines to clean. Skip arch_dma_prep_coherent() for HighMem allocations.

Indeed this is an overhead to call prep for HighMem pages, but on the other hand highmem support is being phased out according to
https://lwn.net/ml/all/[email protected]/ Does it make sense to apply this assuming that it will be removed soon?

I think "soon" is overstating what the plan is. With my proposed series, the majority of current highmem users are changed to no longer use it by default, but there are still three ways in which users will get highmem for a number of years:

- anything with more than 2GB RAM inevitably uses highmem for the top portion of physical memory. These are much less common than 2GB systems, but are not going away soon.
- systems with sparse physical memory where the first and last page are more than 2GB apart currently still rely on highmem even if the total RAM is 2GB or less. This happens e.g. on Tegra114 or RZ-G1H, IIRC. I have plans to address this in the future.
- Users that have applications using a lot of virtual memory still have the option to go back to old configurations with CONFIG_VMSPLIT_3G by selecting CONFIG_EXPERT.

Quoted 4 lines

On the other hand, my proposed change does mean that we have more freedom to optimize for the non-highmem case. I think we can require that the CMA area is in lowmem, and we can make the use of highmem computationally more expensive if it helps simplify code.

Thanks for the summary, I will apply then this patch as a fix for v6.19-rc.

Best regards

Signature · 2 lines
Marek Szyprowski, PhD
Samsung R&D Institute Poland
reviewer
#4 lore raw
In reply to Aneesh Kumar K.V (Arm)
Quoted 6 lines · Aneesh Kumar K.V (Arm)

dma_direct_alloc() calls arch_dma_prep_coherent() to clean any dirty cache lines from the kernel linear alias before creating a coherent remapping.

HighMem pages have no kernel alias mapping, so there are no alias cache lines to clean. Skip arch_dma_prep_coherent() for HighMem allocations.

This is assuming that caches are always cleaned when unmapping highmem, and no still-mapped highmem pages are dirty - how is that guaranteed? The fact that they're not in the linear map doesn't mean they don't necessarily have kernel aliases in either vmalloc pagetables or caches.

Thanks,
Robin.

Quoted 19 lines
---
  kernel/dma/direct.c | 3 ++-
  1 file changed, 2 insertions(+), 1 deletion(-)
kernel/dma/direct.c +2 -1
index 50c3fe2a1d55..ffa267020a1e 100644
@@ -272,7 +272,8 @@ void *dma_direct_alloc(struct device *dev, size_t size,
  			prot = pgprot_decrypted(prot);
 
  		/* remove any dirty cache lines on the kernel alias */
-		arch_dma_prep_coherent(page, size);
+		if (!PageHighMem(page))
+			arch_dma_prep_coherent(page, size);
 
  		/* create a coherent mapping */
  		ret = dma_common_contiguous_remap(page, size, prot,
maintainer
#5 lore raw
Quoted 5 lines · Robin Murphy
Quoted 4 lines · Aneesh Kumar K.V (Arm)

dma_direct_alloc() calls arch_dma_prep_coherent() to clean any dirty cache lines from the kernel linear alias before creating a coherent remapping.

Quoted 8 lines
Quoted 2 lines

HighMem pages have no kernel alias mapping, so there are no alias cache lines to clean. Skip arch_dma_prep_coherent() for HighMem allocations.

This is assuming that caches are always cleaned when unmapping highmem, and no still-mapped highmem pages are dirty - how is that guaranteed? The fact that they're not in the linear map doesn't mean they don't necessarily have kernel aliases in either vmalloc pagetables or caches.

Right, so it is better to keep this unconditional arch_dma_prep_coherent() call. I will drop it from dma-mapping-fixes then.

Best regards

Signature · 2 lines
Marek Szyprowski, PhD
Samsung R&D Institute Poland
reviewer
#6 lore raw
Quoted 16 lines · Marek Szyprowski
Quoted 5 lines · Robin Murphy
Quoted 4 lines · Aneesh Kumar K.V (Arm)

dma_direct_alloc() calls arch_dma_prep_coherent() to clean any dirty cache lines from the kernel linear alias before creating a coherent remapping.

Quoted 8 lines
Quoted 2 lines

HighMem pages have no kernel alias mapping, so there are no alias cache lines to clean. Skip arch_dma_prep_coherent() for HighMem allocations.

This is assuming that caches are always cleaned when unmapping highmem, and no still-mapped highmem pages are dirty - how is that guaranteed? The fact that they're not in the linear map doesn't mean they don't necessarily have kernel aliases in either vmalloc pagetables or caches.

Quoted 2 lines

Right, so it is better to keep this unconditional arch_dma_prep_coherent() call. I will drop it from dma-mapping-fixes then.

Yeah, I think the confusing thing here is that there are architectures with CONFIG_HIGHMEM that don't actually check for and handle it in their arch_dma_prep_coherent() as they seemingly should, however I'm not sure off-hand whether they also support/use highmem CMA in the manner that could end up being an issue in practice (the lack of any reports of crashes or DMA corruption over the last however many years suggests not...)

Thanks,
Robin.

#7 lore raw
Quoted 20 lines · Robin Murphy
Quoted 16 lines · Marek Szyprowski
Quoted 5 lines · Robin Murphy
Quoted 4 lines · Aneesh Kumar K.V (Arm)

dma_direct_alloc() calls arch_dma_prep_coherent() to clean any dirty cache lines from the kernel linear alias before creating a coherent remapping.

Quoted 8 lines
Quoted 2 lines

HighMem pages have no kernel alias mapping, so there are no alias cache lines to clean. Skip arch_dma_prep_coherent() for HighMem allocations.

This is assuming that caches are always cleaned when unmapping highmem, and no still-mapped highmem pages are dirty - how is that guaranteed? The fact that they're not in the linear map doesn't mean they don't necessarily have kernel aliases in either vmalloc pagetables or caches.

Quoted 2 lines

Right, so it is better to keep this unconditional arch_dma_prep_coherent() call. I will drop it from dma-mapping-fixes then.

Quoted 7 lines

Yeah, I think the confusing thing here is that there are architectures with CONFIG_HIGHMEM that don't actually check for and handle it in their arch_dma_prep_coherent() as they seemingly should, however I'm not sure off-hand whether they also support/use highmem CMA in the manner that could end up being an issue in practice (the lack of any reports of crashes or DMA corruption over the last however many years suggests not...)

Should we then remove the PageHighMem() check with DMA_ATTR_NO_KERNEL_MAPPING?

-aneesh

maintainer
#8 lore raw
Quoted 17 lines · Aneesh Kumar K.V
Quoted 16 lines · Robin Murphy
Quoted 15 lines · Marek Szyprowski
Quoted 5 lines · Robin Murphy
Quoted 4 lines · Aneesh Kumar K.V (Arm)

dma_direct_alloc() calls arch_dma_prep_coherent() to clean any dirty cache lines from the kernel linear alias before creating a coherent remapping.

Quoted 7 lines
Quoted 2 lines

HighMem pages have no kernel alias mapping, so there are no alias cache lines to clean. Skip arch_dma_prep_coherent() for HighMem allocations.

This is assuming that caches are always cleaned when unmapping highmem, and no still-mapped highmem pages are dirty - how is that guaranteed? The fact that they're not in the linear map doesn't mean they don't necessarily have kernel aliases in either vmalloc pagetables or caches.

Right, so it is better to keep this unconditional arch_dma_prep_coherent() call. I will drop it from dma-mapping-fixes then.

Quoted 8 lines
Quoted 7 lines

Yeah, I think the confusing thing here is that there are architectures with CONFIG_HIGHMEM that don't actually check for and handle it in their arch_dma_prep_coherent() as they seemingly should, however I'm not sure off-hand whether they also support/use highmem CMA in the manner that could end up being an issue in practice (the lack of any reports of crashes or DMA corruption over the last however many years suggests not...)

Should we then remove the PageHighMem() check with DMA_ATTR_NO_KERNEL_MAPPING?

Right, this has to be unified.

Best regards

Signature · 2 lines
Marek Szyprowski, PhD
Samsung R&D Institute Poland
#9 lore raw
Quoted 25 lines · Marek Szyprowski
Quoted 24 lines · Aneesh Kumar K.V
Quoted 23 lines · Robin Murphy
Quoted 8 lines · Marek Szyprowski
Quoted 7 lines · Robin Murphy
Quoted 6 lines · Aneesh Kumar K.V (Arm)

dma_direct_alloc() calls arch_dma_prep_coherent() to clean any dirty cache lines from the kernel linear alias before creating a coherent remapping.

HighMem pages have no kernel alias mapping, so there are no alias cache lines to clean. Skip arch_dma_prep_coherent() for HighMem allocations.

Quoted 7 lines
Quoted 5 lines

This is assuming that caches are always cleaned when unmapping highmem, and no still-mapped highmem pages are dirty - how is that guaranteed? The fact that they're not in the linear map doesn't mean they don't necessarily have kernel aliases in either vmalloc pagetables or caches.

Right, so it is better to keep this unconditional arch_dma_prep_coherent() call. I will drop it from dma-mapping-fixes then.

Yeah, I think the confusing thing here is that there are architectures with CONFIG_HIGHMEM that don't actually check for and handle it in their arch_dma_prep_coherent() as they seemingly should, however I'm not sure off-hand whether they also support/use highmem CMA in the manner that could end up being an issue in practice (the lack of any reports of crashes or DMA corruption over the last however many years suggests not...)

Quoted 3 lines
Quoted 1 line

Should we then remove the PageHighMem() check with DMA_ATTR_NO_KERNEL_MAPPING?

Right, this has to be unified.

I had a related question, how do we handle cache flushes required for architectures that don't implement CONFIG_ARCH_HAS_DMA_PREP_COHERENT (arch/arm)?

-aneesh

maintainer
#10 lore raw
Quoted 26 lines · Aneesh Kumar K.V
Quoted 25 lines · Marek Szyprowski
Quoted 24 lines · Aneesh Kumar K.V
Quoted 23 lines · Robin Murphy
Quoted 8 lines · Marek Szyprowski
Quoted 7 lines · Robin Murphy
Quoted 6 lines · Aneesh Kumar K.V (Arm)

dma_direct_alloc() calls arch_dma_prep_coherent() to clean any dirty cache lines from the kernel linear alias before creating a coherent remapping.

HighMem pages have no kernel alias mapping, so there are no alias cache lines to clean. Skip arch_dma_prep_coherent() for HighMem allocations.

Quoted 7 lines
Quoted 5 lines

This is assuming that caches are always cleaned when unmapping highmem, and no still-mapped highmem pages are dirty - how is that guaranteed? The fact that they're not in the linear map doesn't mean they don't necessarily have kernel aliases in either vmalloc pagetables or caches.

Right, so it is better to keep this unconditional arch_dma_prep_coherent() call. I will drop it from dma-mapping-fixes then.

Yeah, I think the confusing thing here is that there are architectures with CONFIG_HIGHMEM that don't actually check for and handle it in their arch_dma_prep_coherent() as they seemingly should, however I'm not sure off-hand whether they also support/use highmem CMA in the manner that could end up being an issue in practice (the lack of any reports of crashes or DMA corruption over the last however many years suggests not...)

Quoted 4 lines
Quoted 2 lines
Quoted 1 line

Should we then remove the PageHighMem() check with DMA_ATTR_NO_KERNEL_MAPPING?

Right, this has to be unified.

I had a related question, how do we handle cache flushes required for architectures that don't implement CONFIG_ARCH_HAS_DMA_PREP_COHERENT (arch/arm)?

ARM 32bit architecture provides arch_dma_alloc(), which handles cache management internally.

Best regards

Signature · 2 lines
Marek Szyprowski, PhD
Samsung R&D Institute Poland
#11 lore raw
Quoted 27 lines · Marek Szyprowski
Quoted 26 lines · Aneesh Kumar K.V
Quoted 25 lines · Marek Szyprowski
Quoted 24 lines · Aneesh Kumar K.V
Quoted 23 lines · Robin Murphy
Quoted 8 lines · Marek Szyprowski
Quoted 7 lines · Robin Murphy
Quoted 6 lines · Aneesh Kumar K.V (Arm)

dma_direct_alloc() calls arch_dma_prep_coherent() to clean any dirty cache lines from the kernel linear alias before creating a coherent remapping.

HighMem pages have no kernel alias mapping, so there are no alias cache lines to clean. Skip arch_dma_prep_coherent() for HighMem allocations.

Quoted 7 lines
Quoted 5 lines

This is assuming that caches are always cleaned when unmapping highmem, and no still-mapped highmem pages are dirty - how is that guaranteed? The fact that they're not in the linear map doesn't mean they don't necessarily have kernel aliases in either vmalloc pagetables or caches.

Right, so it is better to keep this unconditional arch_dma_prep_coherent() call. I will drop it from dma-mapping-fixes then.

Yeah, I think the confusing thing here is that there are architectures with CONFIG_HIGHMEM that don't actually check for and handle it in their arch_dma_prep_coherent() as they seemingly should, however I'm not sure off-hand whether they also support/use highmem CMA in the manner that could end up being an issue in practice (the lack of any reports of crashes or DMA corruption over the last however many years suggests not...)

Quoted 8 lines
Quoted 4 lines
Quoted 2 lines
Quoted 1 line

Should we then remove the PageHighMem() check with DMA_ATTR_NO_KERNEL_MAPPING?

Right, this has to be unified.

I had a related question, how do we handle cache flushes required for architectures that don't implement CONFIG_ARCH_HAS_DMA_PREP_COHERENT (arch/arm)?

ARM 32bit architecture provides arch_dma_alloc(), which handles cache management internally.

But dma_direct_alloc does the below

	if ((attrs & DMA_ATTR_NO_KERNEL_MAPPING) &&
	    !force_dma_unencrypted(dev) && !is_swiotlb_for_alloc(dev))
		return dma_direct_alloc_no_mapping(dev, size, dma_handle, gfp);

	if (!dev_is_dma_coherent(dev)) {
		if (IS_ENABLED(CONFIG_ARCH_HAS_DMA_ALLOC) &&
		    !is_swiotlb_for_alloc(dev))
			return arch_dma_alloc(dev, size, dma_handle, gfp,
					      attrs);

IIUC, this implies we won't call arch_dma_alloc for DMA_ATTR_NO_KERNEL_MAPPING ?

-aneesh

maintainer
#12 lore raw
Quoted 47 lines · Aneesh Kumar K.V
Quoted 27 lines · Marek Szyprowski
Quoted 26 lines · Aneesh Kumar K.V
Quoted 25 lines · Marek Szyprowski
Quoted 24 lines · Aneesh Kumar K.V
Quoted 23 lines · Robin Murphy
Quoted 8 lines · Marek Szyprowski
Quoted 7 lines · Robin Murphy
Quoted 6 lines · Aneesh Kumar K.V (Arm)

dma_direct_alloc() calls arch_dma_prep_coherent() to clean any dirty cache lines from the kernel linear alias before creating a coherent remapping.

HighMem pages have no kernel alias mapping, so there are no alias cache lines to clean. Skip arch_dma_prep_coherent() for HighMem allocations.

Quoted 7 lines
Quoted 5 lines

This is assuming that caches are always cleaned when unmapping highmem, and no still-mapped highmem pages are dirty - how is that guaranteed? The fact that they're not in the linear map doesn't mean they don't necessarily have kernel aliases in either vmalloc pagetables or caches.

Right, so it is better to keep this unconditional arch_dma_prep_coherent() call. I will drop it from dma-mapping-fixes then.

Yeah, I think the confusing thing here is that there are architectures with CONFIG_HIGHMEM that don't actually check for and handle it in their arch_dma_prep_coherent() as they seemingly should, however I'm not sure off-hand whether they also support/use highmem CMA in the manner that could end up being an issue in practice (the lack of any reports of crashes or DMA corruption over the last however many years suggests not...)

Quoted 7 lines
Quoted 4 lines
Quoted 2 lines
Quoted 1 line

Should we then remove the PageHighMem() check with DMA_ATTR_NO_KERNEL_MAPPING?

Right, this has to be unified.

I had a related question, how do we handle cache flushes required for architectures that don't implement CONFIG_ARCH_HAS_DMA_PREP_COHERENT (arch/arm)?

ARM 32bit architecture provides arch_dma_alloc(), which handles cache management internally.

But dma_direct_alloc does the below

	if ((attrs & DMA_ATTR_NO_KERNEL_MAPPING) &&
	    !force_dma_unencrypted(dev) && !is_swiotlb_for_alloc(dev))
		return dma_direct_alloc_no_mapping(dev, size, dma_handle, gfp);

	if (!dev_is_dma_coherent(dev)) {
		if (IS_ENABLED(CONFIG_ARCH_HAS_DMA_ALLOC) &&
		    !is_swiotlb_for_alloc(dev))
			return arch_dma_alloc(dev, size, dma_handle, gfp,
					      attrs);
Quoted 2 lines

IIUC, this implies we won't call arch_dma_alloc for DMA_ATTR_NO_KERNEL_MAPPING ?

It looks that commit 849facea92fa ("dma-direct: simplify the DMA_ATTR_NO_KERNEL_MAPPING handling") changed this. Before that arch_dma_alloc() was called regardless of the provided DMA attrs. Indeed, this should be fixed. Frankly speaking this probably means that there are no active users of this combination (arch/arm, DMA_ATTR_NO_KERNEL_MAPPING and dma-direct) or nobody cares.

The DMA_ATTR_NO_KERNEL_MAPPING attribute was initially introduced for the DMA-IOMMU based implementation, where it saves significant amount of resources.

Best regards

Signature · 2 lines
Marek Szyprowski, PhD
Samsung R&D Institute Poland
#13 lore raw
Quoted 51 lines · Marek Szyprowski
Quoted 47 lines · Aneesh Kumar K.V
Quoted 27 lines · Marek Szyprowski
Quoted 26 lines · Aneesh Kumar K.V
Quoted 25 lines · Marek Szyprowski
Quoted 24 lines · Aneesh Kumar K.V
Quoted 23 lines · Robin Murphy
Quoted 8 lines · Marek Szyprowski
Quoted 7 lines · Robin Murphy
Quoted 6 lines · Aneesh Kumar K.V (Arm)

dma_direct_alloc() calls arch_dma_prep_coherent() to clean any dirty cache lines from the kernel linear alias before creating a coherent remapping.

HighMem pages have no kernel alias mapping, so there are no alias cache lines to clean. Skip arch_dma_prep_coherent() for HighMem allocations.

Quoted 7 lines
Quoted 5 lines

This is assuming that caches are always cleaned when unmapping highmem, and no still-mapped highmem pages are dirty - how is that guaranteed? The fact that they're not in the linear map doesn't mean they don't necessarily have kernel aliases in either vmalloc pagetables or caches.

Right, so it is better to keep this unconditional arch_dma_prep_coherent() call. I will drop it from dma-mapping-fixes then.

Yeah, I think the confusing thing here is that there are architectures with CONFIG_HIGHMEM that don't actually check for and handle it in their arch_dma_prep_coherent() as they seemingly should, however I'm not sure off-hand whether they also support/use highmem CMA in the manner that could end up being an issue in practice (the lack of any reports of crashes or DMA corruption over the last however many years suggests not...)

Quoted 7 lines
Quoted 4 lines
Quoted 2 lines
Quoted 1 line

Should we then remove the PageHighMem() check with DMA_ATTR_NO_KERNEL_MAPPING?

Right, this has to be unified.

I had a related question, how do we handle cache flushes required for architectures that don't implement CONFIG_ARCH_HAS_DMA_PREP_COHERENT (arch/arm)?

ARM 32bit architecture provides arch_dma_alloc(), which handles cache management internally.

But dma_direct_alloc does the below

	if ((attrs & DMA_ATTR_NO_KERNEL_MAPPING) &&
	    !force_dma_unencrypted(dev) && !is_swiotlb_for_alloc(dev))
		return dma_direct_alloc_no_mapping(dev, size, dma_handle, gfp);

	if (!dev_is_dma_coherent(dev)) {
		if (IS_ENABLED(CONFIG_ARCH_HAS_DMA_ALLOC) &&
		    !is_swiotlb_for_alloc(dev))
			return arch_dma_alloc(dev, size, dma_handle, gfp,
					      attrs);
Quoted 2 lines

IIUC, this implies we won't call arch_dma_alloc for DMA_ATTR_NO_KERNEL_MAPPING ?

Quoted 7 lines

It looks that commit 849facea92fa ("dma-direct: simplify the DMA_ATTR_NO_KERNEL_MAPPING handling") changed this. Before that arch_dma_alloc() was called regardless of the provided DMA attrs. Indeed, this should be fixed. Frankly speaking this probably means that there are no active users of this combination (arch/arm, DMA_ATTR_NO_KERNEL_MAPPING and dma-direct) or nobody cares.

Before that commit we did

	if (!IS_ENABLED(CONFIG_ARCH_HAS_DMA_SET_UNCACHED) &&
	    !IS_ENABLED(CONFIG_DMA_DIRECT_REMAP) &&
	    dma_alloc_need_uncached(dev, attrs))
		return arch_dma_alloc(dev, size, dma_handle, gfp, attrs);

dma_alloc_need_uncached() returned false for DMA_ATTR_NO_KERNEL_MAPPING

static __always_inline bool dma_alloc_need_uncached(struct device *dev,
unsigned long attrs)

{
	if (dev_is_dma_coherent(dev))
		return false;
	if (attrs & DMA_ATTR_NO_KERNEL_MAPPING)
		return false;
	return true;
}

It looks like for DMA_ATTR_NO_KERNEL_MAPPING, we never called into arch_dma_alloc()?

Quoted 5 lines

The DMA_ATTR_NO_KERNEL_MAPPING attribute was initially introduced for the DMA-IOMMU based implementation, where it saves significant amount of resources.

-aneesh