Skip to contentSkip to conversation

CMA for arm64

PATCH Superseded Laura Abbott 2 participants 11 messages Dec 10, 2013 – Dec 12, 2013 arm64

v3 of 4 Superseded Latest is v4 →

lore
Share & tools
lore
#0 lore raw

v3 cover CMA for arm64

Hi,

This adds support for CMA in arm64

v3: Fix a missed line to actually set the dma_handle correctly

v2: Move the dma contiguous functions from swiotlb to arm64 directly

 arch/arm64/Kconfig                      |    1 +
 arch/arm64/include/asm/dma-contiguous.h |   29 +++++++++++++++++++++++++++++
 arch/arm64/mm/dma-mapping.c             |   27 ++++++++++++++++++++++++---
 arch/arm64/mm/init.c                    |    3 +++
 4 files changed, 57 insertions(+), 3 deletions(-)

Thanks,
Laura

#1 lore raw

PATCH v3 1/2 arm64: Check for NULL device before getting the coherent_dma_mask

The device passed in to dma_alloc may be NULL. Check for this before trying to get the coherent_dma_mask.

Cc Will Deacon <[email protected]>
Cc Catalin Marinas <[email protected]>
Cc Marek Szyprowski <[email protected]>
---
 arch/arm64/mm/dma-mapping.c |    2 +-
 1 files changed, 1 insertions(+), 1 deletions(-)
arch/arm64/mm/dma-mapping.c +1 -1
index 4bd7579..4134212 100644
@@ -33,7 +33,7 @@ static void *arm64_swiotlb_alloc_coherent(struct device *dev, size_t size,
 					  dma_addr_t *dma_handle, gfp_t flags,
 					  struct dma_attrs *attrs)
 {
-	if (IS_ENABLED(CONFIG_ZONE_DMA32) &&
+	if (dev && IS_ENABLED(CONFIG_ZONE_DMA32) &&
 	    dev->coherent_dma_mask <= DMA_BIT_MASK(32))
 		flags |= GFP_DMA32;
 	return swiotlb_alloc_coherent(dev, size, dma_handle, flags);
Signature · 2 lines
The Qualcomm Innovation Center, Inc. is a member of the Code Aurora Forum,
hosted by The Linux Foundation
#2 lore raw

PATCH v3 1/2 arm64: Check for NULL device before getting the coherent_dma_mask

Quoted 16 lines · Laura Abbott

The device passed in to dma_alloc may be NULL. Check for this before trying to get the coherent_dma_mask.

Cc Will Deacon <[email protected]>
Cc Catalin Marinas <[email protected]>
Cc Marek Szyprowski <[email protected]>
---
 arch/arm64/mm/dma-mapping.c |    2 +-
 1 files changed, 1 insertions(+), 1 deletions(-)
arch/arm64/mm/dma-mapping.c +0 -0
index 4bd7579..4134212 100644

arch/arm64/mm/dma-mapping.c static void *arm64_swiotlb_alloc_coherent(struct device *dev, size_t size, · line 39

Quoted 8 lines
arch/arm64/mm/dma-mapping.c +1 -1
@@ -33,7 +33,7 @@ static void *arm64_swiotlb_alloc_coherent(struct device *dev, size_t size,
 					  dma_addr_t *dma_handle, gfp_t flags,
 					  struct dma_attrs *attrs)
 {
-	if (IS_ENABLED(CONFIG_ZONE_DMA32) &&
+	if (dev && IS_ENABLED(CONFIG_ZONE_DMA32) &&
 	    dev->coherent_dma_mask <= DMA_BIT_MASK(32))
 		flags |= GFP_DMA32;
 	return swiotlb_alloc_coherent(dev, size, dma_handle, flags);

Unless I'm misreading the code, it looks like there are paths through swiotlb_alloc_coherent that will dereference the dev parameter without a NULL check. Are you sure we should allow for NULL devices here?

Will

#3 lore raw

PATCH v3 1/2 arm64: Check for NULL device before getting the coherent_dma_mask

Quoted 21 lines · Will Deacon
Quoted 20 lines · Laura Abbott

The device passed in to dma_alloc may be NULL. Check for this before trying to get the coherent_dma_mask.

Cc Will Deacon <[email protected]>
Cc Catalin Marinas <[email protected]>
Cc Marek Szyprowski <[email protected]>
---
  arch/arm64/mm/dma-mapping.c |    2 +-
  1 files changed, 1 insertions(+), 1 deletions(-)
arch/arm64/mm/dma-mapping.c +0 -1
index 4bd7579..4134212 100644
@@ -33,7 +33,7 @@ static void *arm64_swiotlb_alloc_coherent(struct device *dev, size_t size,
  					  dma_addr_t *dma_handle, gfp_t flags,
  					  struct dma_attrs *attrs)
  {
-	if (IS_ENABLED(CONFIG_ZONE_DMA32) &&
Quoted 9 lines

arch/arm64/mm/dma-mapping.c static void *arm64_swiotlb_alloc_coherent(struct device *dev, size_t size, · line 39

Quoted 4 lines
arch/arm64/mm/dma-mapping.c +1 -0
+	if (dev && IS_ENABLED(CONFIG_ZONE_DMA32) &&
  	    dev->coherent_dma_mask <= DMA_BIT_MASK(32))
  		flags |= GFP_DMA32;
  	return swiotlb_alloc_coherent(dev, size, dma_handle, flags);

Unless I'm misreading the code, it looks like there are paths through swiotlb_alloc_coherent that will dereference the dev parameter without a NULL check. Are you sure we should allow for NULL devices here?

The current ARM code allows for NULL devices so that would be a difference in behavior between arm and arm64. We're also relying on this behavior in some code. Where exactly in swiotlb_alloc_coherent does this dereference happen? The only one I see is checked with 'if (hwdev && hwdev->coherent_dma_mask)'

Quoted 2 lines

Will

Thanks,
Laura

Signature · 2 lines
Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum,
hosted by The Linux Foundation
#4 lore raw

PATCH v3 1/2 arm64: Check for NULL device before getting the coherent_dma_mask

Quoted 21 lines · Laura Abbott
Quoted 10 lines · Will Deacon
Quoted 9 lines · Laura Abbott
arch/arm64/mm/dma-mapping.c +0 -1
index 4bd7579..4134212 100644
@@ -33,7 +33,7 @@ static void *arm64_swiotlb_alloc_coherent(struct device *dev, size_t size,
  					  dma_addr_t *dma_handle, gfp_t flags,
  					  struct dma_attrs *attrs)
  {
-	if (IS_ENABLED(CONFIG_ZONE_DMA32) &&
Quoted 9 lines

arch/arm64/mm/dma-mapping.c static void *arm64_swiotlb_alloc_coherent(struct device *dev, size_t size, · line 39

Quoted 4 lines
arch/arm64/mm/dma-mapping.c +1 -0
+	if (dev && IS_ENABLED(CONFIG_ZONE_DMA32) &&
  	    dev->coherent_dma_mask <= DMA_BIT_MASK(32))
  		flags |= GFP_DMA32;
  	return swiotlb_alloc_coherent(dev, size, dma_handle, flags);

Unless I'm misreading the code, it looks like there are paths through swiotlb_alloc_coherent that will dereference the dev parameter without a NULL check. Are you sure we should allow for NULL devices here?

Quoted 5 lines

The current ARM code allows for NULL devices so that would be a difference in behavior between arm and arm64. We're also relying on this behavior in some code. Where exactly in swiotlb_alloc_coherent does this dereference happen? The only one I see is checked with 'if (hwdev && hwdev->coherent_dma_mask)'

phys_to_dma could, but doesn't. The one I spotted was buried down in:

map_single -> swiotlb_tbl_map_single -> dma_get_seg_boundary

Will

#5 lore raw

PATCH v3 1/2 arm64: Check for NULL device before getting the coherent_dma_mask

Quoted 28 lines · Will Deacon
Quoted 21 lines · Laura Abbott
Quoted 10 lines · Will Deacon
Quoted 9 lines · Laura Abbott
arch/arm64/mm/dma-mapping.c +0 -1
index 4bd7579..4134212 100644
@@ -33,7 +33,7 @@ static void *arm64_swiotlb_alloc_coherent(struct device *dev, size_t size,
   					  dma_addr_t *dma_handle, gfp_t flags,
   					  struct dma_attrs *attrs)
   {
-	if (IS_ENABLED(CONFIG_ZONE_DMA32) &&
Quoted 9 lines

arch/arm64/mm/dma-mapping.c static void *arm64_swiotlb_alloc_coherent(struct device *dev, size_t size, · line 39

Quoted 4 lines
arch/arm64/mm/dma-mapping.c +1 -0
+	if (dev && IS_ENABLED(CONFIG_ZONE_DMA32) &&
   	    dev->coherent_dma_mask <= DMA_BIT_MASK(32))
   		flags |= GFP_DMA32;
   	return swiotlb_alloc_coherent(dev, size, dma_handle, flags);

Unless I'm misreading the code, it looks like there are paths through swiotlb_alloc_coherent that will dereference the dev parameter without a NULL check. Are you sure we should allow for NULL devices here?

Quoted 5 lines

The current ARM code allows for NULL devices so that would be a difference in behavior between arm and arm64. We're also relying on this behavior in some code. Where exactly in swiotlb_alloc_coherent does this dereference happen? The only one I see is checked with 'if (hwdev && hwdev->coherent_dma_mask)'

Quoted 4 lines

phys_to_dma could, but doesn't. The one I spotted was buried down in:

   map_single -> swiotlb_tbl_map_single -> dma_get_seg_boundary

Ah yes I see that now.

I guess the question still stands though, should we fixup the parts of the code to fully support the NULL devices or do we tell clients to fix their code and fail the allocation with a warning not to pass NULL?

Quoted 2 lines

Will

Laura

Signature · 2 lines
Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum,
hosted by The Linux Foundation
#6 lore raw

PATCH v3 1/2 arm64: Check for NULL device before getting the coherent_dma_mask

Quoted 31 lines · Laura Abbott
Quoted 30 lines · Will Deacon
Quoted 21 lines · Laura Abbott
Quoted 10 lines · Will Deacon
Quoted 9 lines · Laura Abbott
arch/arm64/mm/dma-mapping.c +0 -1
index 4bd7579..4134212 100644
@@ -33,7 +33,7 @@ static void *arm64_swiotlb_alloc_coherent(struct device *dev, size_t size,
   					  dma_addr_t *dma_handle, gfp_t flags,
   					  struct dma_attrs *attrs)
   {
-	if (IS_ENABLED(CONFIG_ZONE_DMA32) &&
Quoted 9 lines

arch/arm64/mm/dma-mapping.c static void *arm64_swiotlb_alloc_coherent(struct device *dev, size_t size, · line 39

Quoted 4 lines
arch/arm64/mm/dma-mapping.c +1 -0
+	if (dev && IS_ENABLED(CONFIG_ZONE_DMA32) &&
   	    dev->coherent_dma_mask <= DMA_BIT_MASK(32))
   		flags |= GFP_DMA32;
   	return swiotlb_alloc_coherent(dev, size, dma_handle, flags);

Unless I'm misreading the code, it looks like there are paths through swiotlb_alloc_coherent that will dereference the dev parameter without a NULL check. Are you sure we should allow for NULL devices here?

Quoted 5 lines

The current ARM code allows for NULL devices so that would be a difference in behavior between arm and arm64. We're also relying on this behavior in some code. Where exactly in swiotlb_alloc_coherent does this dereference happen? The only one I see is checked with 'if (hwdev && hwdev->coherent_dma_mask)'

phys_to_dma could, but doesn't. The one I spotted was buried down in:

Quoted 8 lines
Quoted 2 lines
   map_single -> swiotlb_tbl_map_single -> dma_get_seg_boundary

Ah yes I see that now.

I guess the question still stands though, should we fixup the parts of the code to fully support the NULL devices or do we tell clients to fix their code and fail the allocation with a warning not to pass NULL?

Assuming it's always possible to pass in a non-NULL device, I'd be tempted to fix the callers. Can you think of a valid use-case where the caller doesn't have a device handle to pass in?

Will

#7 lore raw

PATCH v3 1/2 arm64: Check for NULL device before getting the coherent_dma_mask

Quoted 37 lines · Will Deacon
Quoted 36 lines · Laura Abbott
Quoted 28 lines · Will Deacon
Quoted 21 lines · Laura Abbott
Quoted 10 lines · Will Deacon
Quoted 9 lines · Laura Abbott
arch/arm64/mm/dma-mapping.c +0 -1
index 4bd7579..4134212 100644
@@ -33,7 +33,7 @@ static void *arm64_swiotlb_alloc_coherent(struct device *dev, size_t size,
    					  dma_addr_t *dma_handle, gfp_t flags,
    					  struct dma_attrs *attrs)
    {
-	if (IS_ENABLED(CONFIG_ZONE_DMA32) &&
Quoted 9 lines

arch/arm64/mm/dma-mapping.c static void *arm64_swiotlb_alloc_coherent(struct device *dev, size_t size, · line 39

Quoted 4 lines
arch/arm64/mm/dma-mapping.c +1 -0
+	if (dev && IS_ENABLED(CONFIG_ZONE_DMA32) &&
    	    dev->coherent_dma_mask <= DMA_BIT_MASK(32))
    		flags |= GFP_DMA32;
    	return swiotlb_alloc_coherent(dev, size, dma_handle, flags);

Unless I'm misreading the code, it looks like there are paths through swiotlb_alloc_coherent that will dereference the dev parameter without a NULL check. Are you sure we should allow for NULL devices here?

Quoted 5 lines

The current ARM code allows for NULL devices so that would be a difference in behavior between arm and arm64. We're also relying on this behavior in some code. Where exactly in swiotlb_alloc_coherent does this dereference happen? The only one I see is checked with 'if (hwdev && hwdev->coherent_dma_mask)'

Quoted 4 lines

phys_to_dma could, but doesn't. The one I spotted was buried down in:

    map_single -> swiotlb_tbl_map_single -> dma_get_seg_boundary

Ah yes I see that now.

Quoted 8 lines
Quoted 3 lines

I guess the question still stands though, should we fixup the parts of the code to fully support the NULL devices or do we tell clients to fix their code and fail the allocation with a warning not to pass NULL?

Assuming it's always possible to pass in a non-NULL device, I'd be tempted to fix the callers. Can you think of a valid use-case where the caller doesn't have a device handle to pass in?

Depends on your definition of 'valid' ;)

The only cases we've been able to come up with are places where we've been abusing the DMA APIs slightly to get chunks of contiguous memory through CMA for miscellaneous debugging/testing purposes. I only stumbled across this problem at all while writing a quick test case for CMA. I'd argue that if there is a real use case for non-device large contiguous allocations there should be a new dedicated API that doesn't break the dma abstraction.

I suggest for now we add a big WARN when a NULL device is passed in and see how many people complain.

Laura

Signature · 2 lines
Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum,
hosted by The Linux Foundation
#8 lore raw

PATCH v3 2/2 arm64: Enable CMA

In reply to Laura Abbott

arm64 bit targets need the features CMA provides. Add the appropriate hooks, header files, and Kconfig to allow this to happen.

Cc Will Deacon <[email protected]>
Cc Catalin Marinas <[email protected]>
Cc Marek Szyprowski <[email protected]>
---
 arch/arm64/Kconfig                      |    1 +
 arch/arm64/include/asm/dma-contiguous.h |   29 +++++++++++++++++++++++++++++
 arch/arm64/mm/dma-mapping.c             |   25 +++++++++++++++++++++++--
 arch/arm64/mm/init.c                    |    3 +++
 4 files changed, 56 insertions(+), 2 deletions(-)
 create mode 100644 arch/arm64/include/asm/dma-contiguous.h
arch/arm64/Kconfig +1 -0
index 9737e97..26e4bef 100644
@@ -22,6 +22,7 @@ config ARM64
 	select HAVE_DEBUG_KMEMLEAK
 	select HAVE_DMA_API_DEBUG
 	select HAVE_DMA_ATTRS
+	select HAVE_DMA_CONTIGUOUS
 	select HAVE_GENERIC_DMA_COHERENT
 	select HAVE_GENERIC_HARDIRQS
 	select HAVE_HW_BREAKPOINT if PERF_EVENTS
arch/arm64/include/asm/dma-contiguous.h new file +29 -0
new file mode 100644
index 0000000..bc32516
@@ -0,0 +1,29 @@
+/*
+ * Copyright (c) 2013, The Linux Foundation. All rights reserved.
+ *
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License version 2 and
+ * only version 2 as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ * GNU General Public License for more details.
+ */
+
+#ifndef ASMARM64_DMA_CONTIGUOUS_H
+#define ASMARM64_DMA_CONTIGUOUS_H
+
+#ifdef __KERNEL__
+#ifdef CONFIG_CMA
+
+#include <linux/types.h>
+#include <asm-generic/dma-contiguous.h>
+
+static inline void
+dma_contiguous_early_fixup(phys_addr_t base, unsigned long size) { }
+
+#endif
+#endif
+
+#endif
arch/arm64/mm/dma-mapping.c +23 -2
index 4134212..29d10b9 100644
@@ -21,6 +21,7 @@
 #include <linux/export.h>
 #include <linux/slab.h>
 #include <linux/dma-mapping.h>
+#include <linux/dma-contiguous.h>
 #include <linux/vmalloc.h>
 #include <linux/swiotlb.h>
 
@@ -36,14 +37,34 @@ static void *arm64_swiotlb_alloc_coherent(struct device *dev, size_t size,
 	if (dev && IS_ENABLED(CONFIG_ZONE_DMA32) &&
 	    dev->coherent_dma_mask <= DMA_BIT_MASK(32))
 		flags |= GFP_DMA32;
-	return swiotlb_alloc_coherent(dev, size, dma_handle, flags);
+	if (IS_ENABLED(CONFIG_DMA_CMA)) {
+		struct page *page;
+
+		page = dma_alloc_from_contiguous(dev, size >> PAGE_SHIFT,
+							get_order(size));
+		if (!page)
+			return NULL;
+
+		*dma_handle = phys_to_dma(dev, page_to_phys(page));
+		return page_address(page);
+	} else {
+		return swiotlb_alloc_coherent(dev, size, dma_handle, flags);
+	}
 }
 
 static void arm64_swiotlb_free_coherent(struct device *dev, size_t size,
 					void *vaddr, dma_addr_t dma_handle,
 					struct dma_attrs *attrs)
 {
-	swiotlb_free_coherent(dev, size, vaddr, dma_handle);
+	if (IS_ENABLED(CONFIG_DMA_CMA)) {
+		phys_addr_t paddr = dma_to_phys(dev, dma_handle);
+
+		dma_release_from_contiguous(dev,
+					phys_to_page(paddr),
+					size >> PAGE_SHIFT);
+	} else {
+		swiotlb_free_coherent(dev, size, vaddr, dma_handle);
+	}
 }
 
 static struct dma_map_ops arm64_swiotlb_dma_ops = {
arch/arm64/mm/init.c +3 -0
index 67e8d7c..74b7da1 100644
@@ -30,6 +30,7 @@
 #include <linux/memblock.h>
 #include <linux/sort.h>
 #include <linux/of_fdt.h>
+#include <linux/dma-contiguous.h>
 
 #include <asm/prom.h>
 #include <asm/sections.h>
@@ -173,6 +174,8 @@ void __init arm64_memblock_init(void)
 		memblock_reserve(base, size);
 	}
 
+	dma_contiguous_reserve(0);
+
 	memblock_allow_resize();
 	memblock_dump_all();
 }
Signature · 2 lines
The Qualcomm Innovation Center, Inc. is a member of the Code Aurora Forum,
hosted by The Linux Foundation
#9 lore raw

PATCH v3 2/2 arm64: Enable CMA

Hi Laura,

Couple of really minor comments...

Quoted 38 lines · Laura Abbott

arm64 bit targets need the features CMA provides. Add the appropriate hooks, header files, and Kconfig to allow this to happen.

Cc Will Deacon <[email protected]>
Cc Catalin Marinas <[email protected]>
Cc Marek Szyprowski <[email protected]>
---
 arch/arm64/Kconfig                      |    1 +
 arch/arm64/include/asm/dma-contiguous.h |   29 +++++++++++++++++++++++++++++
 arch/arm64/mm/dma-mapping.c             |   25 +++++++++++++++++++++++--
 arch/arm64/mm/init.c                    |    3 +++
 4 files changed, 56 insertions(+), 2 deletions(-)
 create mode 100644 arch/arm64/include/asm/dma-contiguous.h
arch/arm64/Kconfig +1 -0
index 9737e97..26e4bef 100644
@@ -22,6 +22,7 @@ config ARM64
 	select HAVE_DEBUG_KMEMLEAK
 	select HAVE_DMA_API_DEBUG
 	select HAVE_DMA_ATTRS
+	select HAVE_DMA_CONTIGUOUS
 	select HAVE_GENERIC_DMA_COHERENT
 	select HAVE_GENERIC_HARDIRQS
 	select HAVE_HW_BREAKPOINT if PERF_EVENTS
arch/arm64/include/asm/dma-contiguous.h new file +5 -0
new file mode 100644
index 0000000..bc32516
@@ -0,0 +1,29 @@
+/*
+ * Copyright (c) 2013, The Linux Foundation. All rights reserved.
+ *
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License version 2 and

arch/arm64/include/asm/dma-contiguous.h line 15

Quoted 10 lines
arch/arm64/include/asm/dma-contiguous.h new file +10 -0
+ * only version 2 as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ * GNU General Public License for more details.
+ */
+
+#ifndef ASMARM64_DMA_CONTIGUOUS_H
+#define ASMARM64_DMA_CONTIGUOUS_H

We've tried to keep these guards consistent for the arm64 headers, so this would be: __ASM_DMA_CONTIGUOUS_H.

arch/arm64/include/asm/dma-contiguous.h line 18

Quoted 3 lines
arch/arm64/include/asm/dma-contiguous.h new file +3 -0
+
+#ifdef __KERNEL__
+#ifdef CONFIG_CMA

Why is this not CONFIG_DMA_CMA?

Quoted 24 lines
arch/arm64/include/asm/dma-contiguous.h new file +10 -0
+#include <linux/types.h>
+#include <asm-generic/dma-contiguous.h>
+
+static inline void
+dma_contiguous_early_fixup(phys_addr_t base, unsigned long size) { }
+
+#endif
+#endif
+
+#endif
arch/arm64/mm/dma-mapping.c +1 -0
index 4134212..29d10b9 100644
@@ -21,6 +21,7 @@
 #include <linux/export.h>
 #include <linux/slab.h>
 #include <linux/dma-mapping.h>
+#include <linux/dma-contiguous.h>
 #include <linux/vmalloc.h>
 #include <linux/swiotlb.h>
 
@@ -36,14 +37,34 @@ static void *arm64_swiotlb_alloc_coherent(struct device *dev, size_t size,
 	if (dev && IS_ENABLED(CONFIG_ZONE_DMA32) &&

arch/arm64/mm/dma-mapping.c static void *arm64_swiotlb_alloc_coherent(struct device *dev, size_t size, · line 46

Quoted 10 lines
arch/arm64/mm/dma-mapping.c +7 -1
 	    dev->coherent_dma_mask <= DMA_BIT_MASK(32))
 		flags |= GFP_DMA32;
-	return swiotlb_alloc_coherent(dev, size, dma_handle, flags);
+	if (IS_ENABLED(CONFIG_DMA_CMA)) {
+		struct page *page;
+
+		page = dma_alloc_from_contiguous(dev, size >> PAGE_SHIFT,
+							get_order(size));
+		if (!page)
+			return NULL;

Seems a shame to fail the allocation if CMA can't manage it. Is there a good reason not to fall back to swiotlb (other than complicating the freeing paths)?

Will

#10 lore raw

PATCH v3 2/2 arm64: Enable CMA

Quoted 48 lines · Will Deacon

Hi Laura,

Couple of really minor comments...

Quoted 43 lines · Laura Abbott

arm64 bit targets need the features CMA provides. Add the appropriate hooks, header files, and Kconfig to allow this to happen.

Cc Will Deacon <[email protected]>
Cc Catalin Marinas <[email protected]>
Cc Marek Szyprowski <[email protected]>
---
  arch/arm64/Kconfig                      |    1 +
  arch/arm64/include/asm/dma-contiguous.h |   29 +++++++++++++++++++++++++++++
  arch/arm64/mm/dma-mapping.c             |   25 +++++++++++++++++++++++--
  arch/arm64/mm/init.c                    |    3 +++
  4 files changed, 56 insertions(+), 2 deletions(-)

create mode 100644 arch/arm64/include/asm/dma-contiguous.h

arch/arm64/Kconfig +1 -0
index 9737e97..26e4bef 100644
@@ -22,6 +22,7 @@ config ARM64
  	select HAVE_DEBUG_KMEMLEAK
  	select HAVE_DMA_API_DEBUG
  	select HAVE_DMA_ATTRS
+	select HAVE_DMA_CONTIGUOUS
  	select HAVE_GENERIC_DMA_COHERENT
  	select HAVE_GENERIC_HARDIRQS
  	select HAVE_HW_BREAKPOINT if PERF_EVENTS
arch/arm64/include/asm/dma-contiguous.h new file +10 -0
new file mode 100644
index 0000000..bc32516
@@ -0,0 +1,29 @@
+/*
+ * Copyright (c) 2013, The Linux Foundation. All rights reserved.
+ *
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License version 2 and
+ * only version 2 as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
Quoted 9 lines

arch/arm64/include/asm/dma-contiguous.h line 15

Quoted 5 lines
arch/arm64/include/asm/dma-contiguous.h new file +5 -0
+ * GNU General Public License for more details.
+ */
+
+#ifndef ASMARM64_DMA_CONTIGUOUS_H
+#define ASMARM64_DMA_CONTIGUOUS_H

We've tried to keep these guards consistent for the arm64 headers, so this would be: __ASM_DMA_CONTIGUOUS_H.

Fine.

Quoted 6 lines

arch/arm64/include/asm/dma-contiguous.h line 18

Quoted 3 lines
arch/arm64/include/asm/dma-contiguous.h new file +3 -0
+
+#ifdef __KERNEL__
+#ifdef CONFIG_CMA

Why is this not CONFIG_DMA_CMA?

Whoops.

Quoted 22 lines
Quoted 22 lines
arch/arm64/include/asm/dma-contiguous.h new file +10 -0
+#include <linux/types.h>
+#include <asm-generic/dma-contiguous.h>
+
+static inline void
+dma_contiguous_early_fixup(phys_addr_t base, unsigned long size) { }
+
+#endif
+#endif
+
+#endif
arch/arm64/mm/dma-mapping.c +1 -0
index 4134212..29d10b9 100644
@@ -21,6 +21,7 @@
  #include <linux/export.h>
  #include <linux/slab.h>
  #include <linux/dma-mapping.h>
+#include <linux/dma-contiguous.h>
  #include <linux/vmalloc.h>
  #include <linux/swiotlb.h>
 
Quoted 17 lines

arch/arm64/mm/dma-mapping.c static void *arm64_swiotlb_alloc_coherent(struct device *dev, size_t size, · line 46

Quoted 12 lines
arch/arm64/mm/dma-mapping.c +7 -1
@@ -36,14 +37,34 @@ static void *arm64_swiotlb_alloc_coherent(struct device *dev, size_t size,
  	if (dev && IS_ENABLED(CONFIG_ZONE_DMA32) &&
  	    dev->coherent_dma_mask <= DMA_BIT_MASK(32))
  		flags |= GFP_DMA32;
-	return swiotlb_alloc_coherent(dev, size, dma_handle, flags);
+	if (IS_ENABLED(CONFIG_DMA_CMA)) {
+		struct page *page;
+
+		page = dma_alloc_from_contiguous(dev, size >> PAGE_SHIFT,
+							get_order(size));
+		if (!page)
+			return NULL;

Seems a shame to fail the allocation if CMA can't manage it. Is there a good reason not to fall back to swiotlb (other than complicating the freeing paths)?

The current ARM code doesn't currently fall back but perhaps more importantly, the entire point of CMA is to be able to get allocations that are larger than the buddy allocator can handle. I'm not sure how likely it would be that if a CMA allocation failed the buddy allocator would be able to handle it successfully.

Quoted 2 lines

Will

Thanks,
Laura

Signature · 2 lines
Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum,
hosted by The Linux Foundation