Date:	Fri, 18 Dec 2015 12:41:20 +0000
From:	Robin Murphy <robin.murphy@....com>
To:	Doug Anderson <dianders@...omium.org>,
	Russell King <linux@....linux.org.uk>
Cc:	Laurent Pinchart <laurent.pinchart+renesas@...asonboard.com>,
	"linux-kernel@...r.kernel.org" <linux-kernel@...r.kernel.org>,
	Pawel Osciak <pawel@...iak.com>, mike.looijmans@...ic.nl,
	lorenx4@...il.com, Dmitry Torokhov <dmitry.torokhov@...il.com>,
	Will Deacon <will.deacon@....com>,
	Tomasz Figa <tfiga@...omium.org>,
	David Rientjes <rientjes@...gle.com>,
	Carlo Caione <carlo@...one.org>,
	Andrew Morton <akpm@...ux-foundation.org>,
	"linux-arm-kernel@...ts.infradead.org" 
	<linux-arm-kernel@...ts.infradead.org>,
	Marek Szyprowski <m.szyprowski@...sung.com>
Subject: Re: [PATCH] ARM: dma-mapping: Just allocate one chunk at a time

Hi Doug,

On 17/12/15 22:31, Doug Anderson wrote:
> Hi,
>
> On Thu, Dec 17, 2015 at 12:30 PM, Douglas Anderson
> <dianders@...omium.org> 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 be efficient about this
>> by starting out allocating large chunks and then moving to smaller and
>> smaller chunk sizes until it succeeds.
>>
>> 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 24 seconds to allocate 4 megs of memory (!!).
>>
>> A second problem (and maybe even more important) is that allocating big
>> chunks when we don't need them is just not a good idea anyway.  The
>> first thing we do with these big chunks is break them into smaller
>> chunks!  If we allocate small chunks:
>> - The memory manager doesn't need to work so hard to give us big chunks.
>> - We can save the big chunks for those that really need them and this
>>    code can make great use of all the small chunks sitting around.
>>
>> Let's simplify by just allocating one page at a time.  We may make more
>> total allocate calls but it works way better.  In real world tests that
>> used to sometimes see a 24 second allocation call I can now see at most
>> 250 ms.
>
> Off-list I talked to Dmitry about this a little bit and he pointed out
> that contiguous chunks actually give a benefit to the IOMMU.  I don't
> think the benefit outweighs the cost in this case, but I'm happy to
> hear what others have to say.  I did some quick printouts and it turns
> out that even when requesting page at a time the memory manager
> (unsurprisingly) can in many cases still give us pages that are
> contiguous.
>
> Also I'm happy to post up
> <https://chromium-review.googlesource.com/#/c/319210/> which sorts the
> array and could possibly give us larger chunks of contiguous memory.

I think sorting individually-allocated pages really isn't worth the 
effort - I'm not aware of anything that's going to be capable of using 
larger page/section mappings without also having the necessary physical 
alignment, and if you _can_ cobble together, say, 2MB worth of 
contiguous pages *at 2MB alignment*, then you would have been far better 
off just asking the slab allocator for that in the first place.

That's the key point of the higher-order allocation - not that you get 
some contiguous pages, but that the region you get is also naturally 
aligned to its size physically. That we break up the CPU page tables for 
that region into individual pages is just an inconsequential 
implementation detail from the IOMMU side. When you _do_ have plenty of 
unfragmented free memory it can really be a big win - here's an 
instrumented example of what happens on my Juno with the ARM HDLCD/SMMU 
combo setting up a framebuffer at boot time:


   iommu_dma_alloc: alloc size 0x753000, 1875 pages
   __iommu_dma_alloc_pages: allocated at order 10
   __iommu_dma_alloc_pages: allocated at order 9
   __iommu_dma_alloc_pages: allocated at order 8
   __iommu_dma_alloc_pages: allocated at order 6
   __iommu_dma_alloc_pages: allocated at order 4
   __iommu_dma_alloc_pages: allocated at order 1
   __iommu_dma_alloc_pages: allocated at order 0
   iommu: map: iova 0xff800000 pa 0x00000009f5400000 size 0x400000
   iommu: mapping: iova 0xff800000 pa 0x00000009f5400000 pgsize 0x200000
   iommu: mapping: iova 0xffa00000 pa 0x00000009f5600000 pgsize 0x200000
   iommu: map: iova 0xffc00000 pa 0x00000000fa200000 size 0x200000
   iommu: mapping: iova 0xffc00000 pa 0x00000000fa200000 pgsize 0x200000
   iommu: map: iova 0xffe00000 pa 0x00000009f5a00000 size 0x100000
   iommu: mapping: iova 0xffe00000 pa 0x00000009f5a00000 pgsize 0x1000
   iommu: mapping: iova 0xffe01000 pa 0x00000009f5a01000 pgsize 0x1000
   iommu: mapping: iova 0xffe02000 pa 0x00000009f5a02000 pgsize 0x1000
   iommu: mapping: iova 0xffe03000 pa 0x00000009f5a03000 pgsize 0x1000
   ...

Since the IOVA region itself is aligned to 8MB (for the total size) and 
the physical regions come out in optimal decreasing order, we're able to 
map over 80% of the whole buffer with just 3 section entries, with a 
corresponding saving on TLB pressure, page table maintenance (cache 
flushing), etc.

That said, unless you're in the middle of some crazy allocator-thrashing 
race, then it's probably safe to assume that once allocation fails at a 
given order that's going to remain the case in the near future - would 
you mind taking the following diff for a spin under your test conditions 
to see how it compares?

Robin.

----->8-----
diff --git a/arch/arm/mm/dma-mapping.c b/arch/arm/mm/dma-mapping.c
index dfb5001..95e75c4 100644
--- a/arch/arm/mm/dma-mapping.c
+++ b/arch/arm/mm/dma-mapping.c
@@ -1129,6 +1129,7 @@ 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;
+       unsigned int order = MAX_ORDER;

         if (array_size <= PAGE_SIZE)
                 pages = kzalloc(array_size, GFP_KERNEL);
@@ -1160,9 +1161,10 @@ static struct page **__iommu_alloc_buffer(struct 
device *dev, size_t size,
         gfp |= __GFP_NOWARN | __GFP_HIGHMEM;

         while (count) {
-               int j, order;
+               int j;

-               for (order = __fls(count); order > 0; --order) {
+               for (order = min_t(unsigned int, 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

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

Powered by blists - more mailing lists