[PATCH v4 0/3] dma-mapping: Patches for speeding up allocation

[PATCH v4 0/3] dma-mapping: Patches for speeding up allocation

[Douglas Anderson]

This series of 3 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_NOHUGEPAGE
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_NOHUGEPAGE hint to
optimize allocation") simply applies the 2nd patch.  Again it's pretty
simple.  ...and again it does nothing by itself.

Notably missing from this series is the fourth patch that adds teeth to
the second and third.  You can find that out of tree at
<https://chromium-review.googlesource.com/#/c/320498/>.  Unfortunately
the rk3288_vpu, which is what I'm working on, is out of tree.

All testing was done on the chromeos kernel-3.14.  Sanity (compile /
boot) testing was done on a v4.4-rc6-based kernel.

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 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 (3):
  ARM: dma-mapping: Optimize allocation
  common: DMA-mapping: add DMA_ATTR_NOHUGEPAGE attribute
  ARM: dma-mapping: Use DMA_ATTR_NOHUGEPAGE hint to optimize allocation

 Documentation/DMA-attributes.txt | 23 +++++++++++++++++++++++
 arch/arm/mm/dma-mapping.c        | 38 ++++++++++++++++++++++++--------------
 include/linux/dma-attrs.h        |  1 +
 3 files changed, 48 insertions(+), 14 deletions(-)

-- 
2.6.0.rc2.230.g3dd15c0

[PATCH v4 1/3] 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 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 v4 3/3] ARM: dma-mapping: Use DMA_ATTR_NOHUGEPAGE 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_NOHUGEPAGE.

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 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..96d71bcb4c3a 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_NOHUGEPAGE, 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