[PATCH v5 0/5] dma-mapping: Patches for speeding up allocation

[PATCH v5 0/5] dma-mapping: Patches for speeding up allocation

[Douglas Anderson]

This series of patches will speed up memory allocation in dma-mapping
quite a bit.

The first patch ("ARM: dma-mapping: Optimize allocation") is hopefully
not terribly controversial: it merely doesn't try as hard to allocate
big chunks once it gets the first failure.  Since it's unlikely that
further big chunks will help (they're not likely to be virtually aligned
anyway), this should give a big speedup with no real regression to speak
of.  Yes, things could be made better, but this seems like a sane start.

The second patch ("common: DMA-mapping: add DMA_ATTR_NO_HUGE_PAGE
attribute") models MADV_NOHUGEPAGE as I understand it.  Hopefully folks
are happy with following that lead.  It does nothing by itself.

The third patch ("ARM: dma-mapping: Use DMA_ATTR_NO_HUGE_PAGE hint to
optimize allocation") simply applies the 2nd patch.  Again it's pretty
simple.  ...and again it does nothing by itself.

Thue fourth patch ("[media] videobuf2-dc: Let drivers specify DMA
attrs") comes from the ChromeOS tree (authored by Tomasz Figa) and
allows the fifth patch.

The fifth patch ("[media] s5p-mfc: Set DMA_ATTR_NO_HUGE_PAGE") uses the
new attribute.  For a second user, you can see the out of tree patch for
rk3288 at <https://chromium-review.googlesource.com/#/c/320498/>.

All testing was done on the chromeos kernel-3.8 and kernel-3.14.
Sanity (compile / boot) testing was done on a v4.4-rc6-based kernel on
rk3288, though the video codec isn't there.  I don't have graphics / MFC
working well on exynos, so the MFC change was only compile-tested
upstream.  Hopefully someone upstream whose setup for MFC can give a
Tested-by for these?

Also note that v2 of this series had an extra patch
<https://patchwork.kernel.org/patch/7888861/> that would attempt to sort
the allocation results to opportunistically get some extra alignment.  I
dropped that, but it could be re-introduced if there was interest.  I
found that it did give a little extra alignment sometimes, but maybe not
enough to justify the extra complexity.  It also was a bit half-baked
since it really should have tried harder to ensure alignment.

Changes in v5:
- renamed DMA_ATTR_NOHUGEPAGE to DMA_ATTR_NO_HUGE_PAGE
- s/ping ping/ping pong/
- Let drivers specify DMA attrs new for v5
- s5p-mfc patch new for v5

Changes in v4:
- renamed DMA_ATTR_SEQUENTIAL to DMA_ATTR_NOHUGEPAGE
- added Marek's ack

Changes in v3:
- add DMA_ATTR_SEQUENTIAL attribute new for v3
- Use DMA_ATTR_SEQUENTIAL hint patch new for v3.

Changes in v2:
- No longer just 1 page at a time, but gives up higher order quickly.
- Only tries important higher order allocations that might help us.

Douglas Anderson (4):
  ARM: dma-mapping: Optimize allocation
  common: DMA-mapping: add DMA_ATTR_NO_HUGE_PAGE attribute
  ARM: dma-mapping: Use DMA_ATTR_NO_HUGE_PAGE hint to optimize
    allocation
  [media] s5p-mfc: Set DMA_ATTR_NO_HUGE_PAGE

Tomasz Figa (1):
  [media] videobuf2-dc: Let drivers specify DMA attrs

 Documentation/DMA-attributes.txt               | 23 ++++++++++++++++
 arch/arm/mm/dma-mapping.c                      | 38 ++++++++++++++++----------
 drivers/media/platform/s5p-mfc/s5p_mfc.c       | 13 +++++++--
 drivers/media/v4l2-core/videobuf2-dma-contig.c | 33 ++++++++++++++--------
 include/linux/dma-attrs.h                      |  1 +
 include/media/videobuf2-dma-contig.h           | 11 +++++++-
 6 files changed, 91 insertions(+), 28 deletions(-)

-- 
2.6.0.rc2.230.g3dd15c0

[PATCH v5 1/5] ARM: dma-mapping: Optimize allocation

[Douglas Anderson]

The __iommu_alloc_buffer() is expected to be called to allocate pretty
sizeable buffers.  Upon simple tests of video I saw it trying to
allocate 4,194,304 bytes.  The function tries to allocate large chunks
in order to optimize IOMMU TLB usage.

The current function is very, very slow.

One problem is the way it keeps trying and trying to allocate big
chunks.  Imagine a very fragmented memory that has 4M free but no
contiguous pages at all.  Further imagine allocating 4M (1024 pages).
We'll do the following memory allocations:
- For page 1:
  - Try to allocate order 10 (no retry)
  - Try to allocate order 9 (no retry)
  - ...
  - Try to allocate order 0 (with retry, but not needed)
- For page 2:
  - Try to allocate order 9 (no retry)
  - Try to allocate order 8 (no retry)
  - ...
  - Try to allocate order 0 (with retry, but not needed)
- ...
- ...

Total number of calls to alloc() calls for this case is:
  sum(int(math.log(i, 2)) + 1 for i in range(1, 1025))
  => 9228

The above is obviously worse case, but given how slow alloc can be we
really want to try to avoid even somewhat bad cases.  I timed the old
code with a device under memory pressure and it wasn't hard to see it
take more than 120 seconds to allocate 4 megs of memory! (NOTE: testing
was done on kernel 3.14, so possibly mainline would behave
differently).

A second problem is that allocating big chunks under memory pressure
when we don't need them is just not a great idea anyway unless we really
need them.  We can make due pretty well with smaller chunks so it's
probably wise to leave bigger chunks for other users once memory
pressure is on.

Let's adjust the allocation like this:

1. If a big chunk fails, stop trying to hard and bump down to lower
   order allocations.
2. Don't try useless orders.  The whole point of big chunks is to
   optimize the TLB and it can really only make use of 2M, 1M, 64K and
   4K sizes.

We'll still tend to eat up a bunch of big chunks, but that might be the
right answer for some users.  A future patch could possibly add a new
DMA_ATTR that would let the caller decide that TLB optimization isn't
important and that we should use smaller chunks.  Presumably this would
be a sane strategy for some callers.

Signed-off-by: Douglas Anderson <dianders@chromium.org>
Acked-by: Marek Szyprowski <m.szyprowski@samsung.com>
---
Changes in v5: None
Changes in v4:
- Added Marek's ack

Changes in v3: None
Changes in v2:
- No longer just 1 page at a time, but gives up higher order quickly.
- Only tries important higher order allocations that might help us.

 arch/arm/mm/dma-mapping.c | 34 ++++++++++++++++++++--------------
 1 file changed, 20 insertions(+), 14 deletions(-)

diff --git a/arch/arm/mm/dma-mapping.c b/arch/arm/mm/dma-mapping.c
index 0eca3812527e..bc9cebfa0891 100644
--- a/arch/arm/mm/dma-mapping.c
+++ b/arch/arm/mm/dma-mapping.c
@@ -1122,6 +1122,9 @@ static inline void __free_iova(struct dma_iommu_mapping *mapping,
 	spin_unlock_irqrestore(&mapping->lock, flags);
 }
 
+/* We'll try 2M, 1M, 64K, and finally 4K; array must end with 0! */
+static const int iommu_order_array[] = { 9, 8, 4, 0 };
+
 static struct page **__iommu_alloc_buffer(struct device *dev, size_t size,
 					  gfp_t gfp, struct dma_attrs *attrs)
 {
@@ -1129,6 +1132,7 @@ static struct page **__iommu_alloc_buffer(struct device *dev, size_t size,
 	int count = size >> PAGE_SHIFT;
 	int array_size = count * sizeof(struct page *);
 	int i = 0;
+	int order_idx = 0;
 
 	if (array_size <= PAGE_SIZE)
 		pages = kzalloc(array_size, GFP_KERNEL);
@@ -1162,22 +1166,24 @@ static struct page **__iommu_alloc_buffer(struct device *dev, size_t size,
 	while (count) {
 		int j, order;
 
-		for (order = __fls(count); order > 0; --order) {
-			/*
-			 * We do not want OOM killer to be invoked as long
-			 * as we can fall back to single pages, so we force
-			 * __GFP_NORETRY for orders higher than zero.
-			 */
-			pages[i] = alloc_pages(gfp | __GFP_NORETRY, order);
-			if (pages[i])
-				break;
+		order = iommu_order_array[order_idx];
+
+		/* Drop down when we get small */
+		if (__fls(count) < order) {
+			order_idx++;
+			continue;
 		}
 
-		if (!pages[i]) {
-			/*
-			 * Fall back to single page allocation.
-			 * Might invoke OOM killer as last resort.
-			 */
+		if (order) {
+			/* See if it's easy to allocate a high-order chunk */
+			pages[i] = alloc_pages(gfp | __GFP_NORETRY, order);
+
+			/* Go down a notch@first sign of pressure */
+			if (!pages[i]) {
+				order_idx++;
+				continue;
+			}
+		} else {
 			pages[i] = alloc_pages(gfp, 0);
 			if (!pages[i])
 				goto error;
-- 
2.6.0.rc2.230.g3dd15c0

[PATCH v5 3/5] ARM: dma-mapping: Use DMA_ATTR_NO_HUGE_PAGE hint to optimize allocation

[Douglas Anderson]

If we know that TLB efficiency will not be an issue when memory is
accessed then it's not terribly important to allocate big chunks of
memory.  The whole point of allocating the big chunks was that it would
make TLB usage efficient.

As Marek Szyprowski indicated:
    Please note that mapping memory with larger pages significantly
    improves performance, especially when IOMMU has a little TLB
    cache. This can be easily observed when multimedia devices do
    processing of RGB data with 90/270 degree rotation
Image rotation is distinctly an operation that needs to bounce around
through memory, so it makes sense that TLB efficiency is important
there.

Video decoding, on the other hand, is a fairly sequential operation.
During video decoding it's not expected that we'll be jumping all over
memory.  Decoding video is also pretty heavy and the TLB misses aren't a
huge deal.  Presumably most HW video acceleration users of dma-mapping
will not care about huge pages and will set DMA_ATTR_NO_HUGE_PAGE.

Allocating big chunks of memory is quite expensive, especially if we're
doing it repeadly and memory is full.  In one (out of tree) usage model
it is common that arm_iommu_alloc_attrs() is called 16 times in a row,
each one trying to allocate 4 MB of memory.  This is called whenever the
system encounters a new video, which could easily happen while the
memory system is stressed out.  In fact, on certain social media
websites that auto-play video and have infinite scrolling, it's quite
common to see not just one of these 16x4MB allocations but 2 or 3 right
after another.  Asking the system even to do a small amount of extra
work to give us big chunks in this case is just not a good use of time.

Allocating big chunks of memory is also expensive indirectly.  Even if
we ask the system not to do ANY extra work to allocate _our_ memory,
we're still potentially eating up all big chunks in the system.
Presumably there are other users in the system that aren't quite as
flexible and that actually need these big chunks.  By eating all the big
chunks we're causing extra work for the rest of the system.  We also may
start making other memory allocations fail.  While the system may be
robust to such failures (as is the case with dwc2 USB trying to allocate
buffers for Ethernet data and with WiFi trying to allocate buffers for
WiFi data), it is yet another big performance hit.

Signed-off-by: Douglas Anderson <dianders@chromium.org>
Acked-by: Marek Szyprowski <m.szyprowski@samsung.com>
---
Changes in v5:
- renamed DMA_ATTR_NOHUGEPAGE to DMA_ATTR_NO_HUGE_PAGE

Changes in v4:
- renamed DMA_ATTR_SEQUENTIAL to DMA_ATTR_NOHUGEPAGE
- added Marek's ack

Changes in v3:
- Use DMA_ATTR_SEQUENTIAL hint patch new for v3.

Changes in v2: None

 arch/arm/mm/dma-mapping.c | 4 ++++
 1 file changed, 4 insertions(+)

diff --git a/arch/arm/mm/dma-mapping.c b/arch/arm/mm/dma-mapping.c
index bc9cebfa0891..e9fb2929cb7b 100644
--- a/arch/arm/mm/dma-mapping.c
+++ b/arch/arm/mm/dma-mapping.c
@@ -1158,6 +1158,10 @@ static struct page **__iommu_alloc_buffer(struct device *dev, size_t size,
 		return pages;
 	}
 
+	/* Go straight to 4K chunks if caller says it's OK. */
+	if (dma_get_attr(DMA_ATTR_NO_HUGE_PAGE, attrs))
+		order_idx = ARRAY_SIZE(iommu_order_array) - 1;
+
 	/*
 	 * IOMMU can map any pages, so himem can also be used here
 	 */
-- 
2.6.0.rc2.230.g3dd15c0

[PATCH v5 5/5] [media] s5p-mfc: Set DMA_ATTR_NO_HUGE_PAGE

[Douglas Anderson]

We do video allocation all the time and we need it to be fast.  Plus TLB
efficiency isn't terribly important for video.

That means we want to set DMA_ATTR_NO_HUGE_PAGE.

See also the previous change ("ARM: dma-mapping: Use
DMA_ATTR_NO_HUGE_PAGE hint to optimize allocation").

Signed-off-by: Douglas Anderson <dianders@chromium.org>
---
Changes in v5:
- s5p-mfc patch new for v5

Changes in v4: None
Changes in v3: None
Changes in v2: None

 drivers/media/platform/s5p-mfc/s5p_mfc.c | 13 +++++++++++--
 1 file changed, 11 insertions(+), 2 deletions(-)

diff --git a/drivers/media/platform/s5p-mfc/s5p_mfc.c b/drivers/media/platform/s5p-mfc/s5p_mfc.c
index 927ab4928779..7ea5d0d262bb 100644
--- a/drivers/media/platform/s5p-mfc/s5p_mfc.c
+++ b/drivers/media/platform/s5p-mfc/s5p_mfc.c
@@ -1095,6 +1095,7 @@ static int s5p_mfc_alloc_memdevs(struct s5p_mfc_dev *dev)
 /* MFC probe function */
 static int s5p_mfc_probe(struct platform_device *pdev)
 {
+	DEFINE_DMA_ATTRS(attrs);
 	struct s5p_mfc_dev *dev;
 	struct video_device *vfd;
 	struct resource *res;
@@ -1164,12 +1165,20 @@ static int s5p_mfc_probe(struct platform_device *pdev)
 		}
 	}
 
-	dev->alloc_ctx[0] = vb2_dma_contig_init_ctx(dev->mem_dev_l);
+	/*
+	 * We'll do mostly sequential access, so sacrifice TLB efficiency for
+	 * faster allocation.
+	 */
+	dma_set_attr(DMA_ATTR_NO_HUGE_PAGE, &attrs);
+
+	dev->alloc_ctx[0] = vb2_dma_contig_init_ctx_attrs(dev->mem_dev_l,
+							  &attrs);
 	if (IS_ERR(dev->alloc_ctx[0])) {
 		ret = PTR_ERR(dev->alloc_ctx[0]);
 		goto err_res;
 	}
-	dev->alloc_ctx[1] = vb2_dma_contig_init_ctx(dev->mem_dev_r);
+	dev->alloc_ctx[1] = vb2_dma_contig_init_ctx_attrs(dev->mem_dev_r,
+							  &attrs);
 	if (IS_ERR(dev->alloc_ctx[1])) {
 		ret = PTR_ERR(dev->alloc_ctx[1]);
 		goto err_mem_init_ctx_1;
-- 
2.6.0.rc2.230.g3dd15c0

[PATCH v5 1/5] ARM: dma-mapping: Optimize allocation

[Robin Murphy]

Hi Doug,

On 08/01/16 23:05, Douglas Anderson wrote:
> The __iommu_alloc_buffer() is expected to be called to allocate pretty
> sizeable buffers.  Upon simple tests of video I saw it trying to
> allocate 4,194,304 bytes.  The function tries to allocate large chunks
> in order to optimize IOMMU TLB usage.
>
> The current function is very, very slow.
>
> One problem is the way it keeps trying and trying to allocate big
> chunks.  Imagine a very fragmented memory that has 4M free but no
> contiguous pages at all.  Further imagine allocating 4M (1024 pages).
> We'll do the following memory allocations:
> - For page 1:
>    - Try to allocate order 10 (no retry)
>    - Try to allocate order 9 (no retry)
>    - ...
>    - Try to allocate order 0 (with retry, but not needed)
> - For page 2:
>    - Try to allocate order 9 (no retry)
>    - Try to allocate order 8 (no retry)
>    - ...
>    - Try to allocate order 0 (with retry, but not needed)
> - ...
> - ...
>
> Total number of calls to alloc() calls for this case is:
>    sum(int(math.log(i, 2)) + 1 for i in range(1, 1025))
>    => 9228
>
> The above is obviously worse case, but given how slow alloc can be we
> really want to try to avoid even somewhat bad cases.  I timed the old
> code with a device under memory pressure and it wasn't hard to see it
> take more than 120 seconds to allocate 4 megs of memory! (NOTE: testing
> was done on kernel 3.14, so possibly mainline would behave
> differently).
>
> A second problem is that allocating big chunks under memory pressure
> when we don't need them is just not a great idea anyway unless we really
> need them.  We can make due pretty well with smaller chunks so it's
> probably wise to leave bigger chunks for other users once memory
> pressure is on.
>
> Let's adjust the allocation like this:
>
> 1. If a big chunk fails, stop trying to hard and bump down to lower
>     order allocations.
> 2. Don't try useless orders.  The whole point of big chunks is to
>     optimize the TLB and it can really only make use of 2M, 1M, 64K and
>     4K sizes.
>
> We'll still tend to eat up a bunch of big chunks, but that might be the
> right answer for some users.  A future patch could possibly add a new
> DMA_ATTR that would let the caller decide that TLB optimization isn't
> important and that we should use smaller chunks.  Presumably this would
> be a sane strategy for some callers.

Now that I've had time to think about it properly:

Reviewed-by: Robin Murphy <robin.murphy@arm.com>

I just had an absolutely disgusting idea of how to get the same 
progression with just a single variable and no static array, but I'll 
keep that firmly to myself as it's almost IOCCC-grade WTF :D

Thanks,
Robin.

> Signed-off-by: Douglas Anderson <dianders@chromium.org>
> Acked-by: Marek Szyprowski <m.szyprowski@samsung.com>
> ---
> Changes in v5: None
> Changes in v4:
> - Added Marek's ack
>
> Changes in v3: None
> Changes in v2:
> - No longer just 1 page at a time, but gives up higher order quickly.
> - Only tries important higher order allocations that might help us.
>
>   arch/arm/mm/dma-mapping.c | 34 ++++++++++++++++++++--------------
>   1 file changed, 20 insertions(+), 14 deletions(-)
>
> diff --git a/arch/arm/mm/dma-mapping.c b/arch/arm/mm/dma-mapping.c
> index 0eca3812527e..bc9cebfa0891 100644
> --- a/arch/arm/mm/dma-mapping.c
> +++ b/arch/arm/mm/dma-mapping.c
> @@ -1122,6 +1122,9 @@ static inline void __free_iova(struct dma_iommu_mapping *mapping,
>   	spin_unlock_irqrestore(&mapping->lock, flags);
>   }
>
> +/* We'll try 2M, 1M, 64K, and finally 4K; array must end with 0! */
> +static const int iommu_order_array[] = { 9, 8, 4, 0 };
> +
>   static struct page **__iommu_alloc_buffer(struct device *dev, size_t size,
>   					  gfp_t gfp, struct dma_attrs *attrs)
>   {
> @@ -1129,6 +1132,7 @@ static struct page **__iommu_alloc_buffer(struct device *dev, size_t size,
>   	int count = size >> PAGE_SHIFT;
>   	int array_size = count * sizeof(struct page *);
>   	int i = 0;
> +	int order_idx = 0;
>
>   	if (array_size <= PAGE_SIZE)
>   		pages = kzalloc(array_size, GFP_KERNEL);
> @@ -1162,22 +1166,24 @@ static struct page **__iommu_alloc_buffer(struct device *dev, size_t size,
>   	while (count) {
>   		int j, order;
>
> -		for (order = __fls(count); order > 0; --order) {
> -			/*
> -			 * We do not want OOM killer to be invoked as long
> -			 * as we can fall back to single pages, so we force
> -			 * __GFP_NORETRY for orders higher than zero.
> -			 */
> -			pages[i] = alloc_pages(gfp | __GFP_NORETRY, order);
> -			if (pages[i])
> -				break;
> +		order = iommu_order_array[order_idx];
> +
> +		/* Drop down when we get small */
> +		if (__fls(count) < order) {
> +			order_idx++;
> +			continue;
>   		}
>
> -		if (!pages[i]) {
> -			/*
> -			 * Fall back to single page allocation.
> -			 * Might invoke OOM killer as last resort.
> -			 */
> +		if (order) {
> +			/* See if it's easy to allocate a high-order chunk */
> +			pages[i] = alloc_pages(gfp | __GFP_NORETRY, order);
> +
> +			/* Go down a notch at first sign of pressure */
> +			if (!pages[i]) {
> +				order_idx++;
> +				continue;
> +			}
> +		} else {
>   			pages[i] = alloc_pages(gfp, 0);
>   			if (!pages[i])
>   				goto error;
>

[PATCH v5 1/5] ARM: dma-mapping: Optimize allocation

[Tomasz Figa]

On Wed, Jan 13, 2016 at 9:17 PM, Robin Murphy <robin.murphy@arm.com> wrote:
> Hi Doug,
>
>
> On 08/01/16 23:05, Douglas Anderson wrote:
>>
>> The __iommu_alloc_buffer() is expected to be called to allocate pretty
>> sizeable buffers.  Upon simple tests of video I saw it trying to
>> allocate 4,194,304 bytes.  The function tries to allocate large chunks
>> in order to optimize IOMMU TLB usage.
>>
>> The current function is very, very slow.
>>
>> One problem is the way it keeps trying and trying to allocate big
>> chunks.  Imagine a very fragmented memory that has 4M free but no
>> contiguous pages at all.  Further imagine allocating 4M (1024 pages).
>> We'll do the following memory allocations:
>> - For page 1:
>>    - Try to allocate order 10 (no retry)
>>    - Try to allocate order 9 (no retry)
>>    - ...
>>    - Try to allocate order 0 (with retry, but not needed)
>> - For page 2:
>>    - Try to allocate order 9 (no retry)
>>    - Try to allocate order 8 (no retry)
>>    - ...
>>    - Try to allocate order 0 (with retry, but not needed)
>> - ...
>> - ...
>>
>> Total number of calls to alloc() calls for this case is:
>>    sum(int(math.log(i, 2)) + 1 for i in range(1, 1025))
>>    => 9228
>>
>> The above is obviously worse case, but given how slow alloc can be we
>> really want to try to avoid even somewhat bad cases.  I timed the old
>> code with a device under memory pressure and it wasn't hard to see it
>> take more than 120 seconds to allocate 4 megs of memory! (NOTE: testing
>> was done on kernel 3.14, so possibly mainline would behave
>> differently).
>>
>> A second problem is that allocating big chunks under memory pressure
>> when we don't need them is just not a great idea anyway unless we really
>> need them.  We can make due pretty well with smaller chunks so it's
>> probably wise to leave bigger chunks for other users once memory
>> pressure is on.
>>
>> Let's adjust the allocation like this:
>>
>> 1. If a big chunk fails, stop trying to hard and bump down to lower
>>     order allocations.
>> 2. Don't try useless orders.  The whole point of big chunks is to
>>     optimize the TLB and it can really only make use of 2M, 1M, 64K and
>>     4K sizes.
>>
>> We'll still tend to eat up a bunch of big chunks, but that might be the
>> right answer for some users.  A future patch could possibly add a new
>> DMA_ATTR that would let the caller decide that TLB optimization isn't
>> important and that we should use smaller chunks.  Presumably this would
>> be a sane strategy for some callers.
>
>
> Now that I've had time to think about it properly:
>
> Reviewed-by: Robin Murphy <robin.murphy@arm.com>
>
> I just had an absolutely disgusting idea of how to get the same progression
> with just a single variable and no static array, but I'll keep that firmly
> to myself as it's almost IOCCC-grade WTF :D

Just out of curiosity, a bitmap and loop with fls() and clearing bit
on failure or something more freaky? :)

Anyway:

Reviewed-by: Tomasz Figa <tfiga@chromium.org>

Best regards,
Tomasz

[PATCH v5 1/5] ARM: dma-mapping: Optimize allocation

[Robin Murphy]

On 13/01/16 17:33, Tomasz Figa wrote:
> On Wed, Jan 13, 2016 at 9:17 PM, Robin Murphy <robin.murphy@arm.com> wrote:
>> Hi Doug,
>>
>>
>> On 08/01/16 23:05, Douglas Anderson wrote:
>>>
>>> The __iommu_alloc_buffer() is expected to be called to allocate pretty
>>> sizeable buffers.  Upon simple tests of video I saw it trying to
>>> allocate 4,194,304 bytes.  The function tries to allocate large chunks
>>> in order to optimize IOMMU TLB usage.
>>>
>>> The current function is very, very slow.
>>>
>>> One problem is the way it keeps trying and trying to allocate big
>>> chunks.  Imagine a very fragmented memory that has 4M free but no
>>> contiguous pages at all.  Further imagine allocating 4M (1024 pages).
>>> We'll do the following memory allocations:
>>> - For page 1:
>>>     - Try to allocate order 10 (no retry)
>>>     - Try to allocate order 9 (no retry)
>>>     - ...
>>>     - Try to allocate order 0 (with retry, but not needed)
>>> - For page 2:
>>>     - Try to allocate order 9 (no retry)
>>>     - Try to allocate order 8 (no retry)
>>>     - ...
>>>     - Try to allocate order 0 (with retry, but not needed)
>>> - ...
>>> - ...
>>>
>>> Total number of calls to alloc() calls for this case is:
>>>     sum(int(math.log(i, 2)) + 1 for i in range(1, 1025))
>>>     => 9228
>>>
>>> The above is obviously worse case, but given how slow alloc can be we
>>> really want to try to avoid even somewhat bad cases.  I timed the old
>>> code with a device under memory pressure and it wasn't hard to see it
>>> take more than 120 seconds to allocate 4 megs of memory! (NOTE: testing
>>> was done on kernel 3.14, so possibly mainline would behave
>>> differently).
>>>
>>> A second problem is that allocating big chunks under memory pressure
>>> when we don't need them is just not a great idea anyway unless we really
>>> need them.  We can make due pretty well with smaller chunks so it's
>>> probably wise to leave bigger chunks for other users once memory
>>> pressure is on.
>>>
>>> Let's adjust the allocation like this:
>>>
>>> 1. If a big chunk fails, stop trying to hard and bump down to lower
>>>      order allocations.
>>> 2. Don't try useless orders.  The whole point of big chunks is to
>>>      optimize the TLB and it can really only make use of 2M, 1M, 64K and
>>>      4K sizes.
>>>
>>> We'll still tend to eat up a bunch of big chunks, but that might be the
>>> right answer for some users.  A future patch could possibly add a new
>>> DMA_ATTR that would let the caller decide that TLB optimization isn't
>>> important and that we should use smaller chunks.  Presumably this would
>>> be a sane strategy for some callers.
>>
>>
>> Now that I've had time to think about it properly:
>>
>> Reviewed-by: Robin Murphy <robin.murphy@arm.com>
>>
>> I just had an absolutely disgusting idea of how to get the same progression
>> with just a single variable and no static array, but I'll keep that firmly
>> to myself as it's almost IOCCC-grade WTF :D
>
> Just out of curiosity, a bitmap and loop with fls() and clearing bit
> on failure or something more freaky? :)

Got a Python interpreter handy?

order = 9
for i in range(4):
     print order
     order = (order - 1) & 0xc

Like I said, disgusting :D

Robin.

>
> Anyway:
>
> Reviewed-by: Tomasz Figa <tfiga@chromium.org>
>
> Best regards,
> Tomasz
>