Re: [PATCH v3 0/5] enhance DMA CMA on x86

[konrad wilk <konrad.wilk-QHcLZuEGTsvQT0dZR+AlfA@public.gmane.org>]

On 10/3/2014 12:06 PM, Akinobu Mita wrote:
> 2014-10-03 23:27 GMT+09:00 Peter Hurley <peter-WaGBZJeGNqdsbIuE7sb01tBPR1lH4CV8@public.gmane.org>:
>> On 10/02/2014 07:08 PM, Akinobu Mita wrote:
>>> 2014-10-03 7:03 GMT+09:00 Peter Hurley <peter-WaGBZJeGNqdsbIuE7sb01tBPR1lH4CV8@public.gmane.org>:
>>>> On 10/02/2014 12:41 PM, Konrad Rzeszutek Wilk wrote:
>>>>> On Tue, Sep 30, 2014 at 09:49:54PM -0400, Peter Hurley wrote:
>>>>>> On 09/30/2014 07:45 PM, Thomas Gleixner wrote:
>>>
>>>>>> Which is different than if the plan is to ship production units for x86;
>>>>>> then a general purpose solution will be required.
>>>>>>
>>>>>> As to the good design of a general purpose solution for allocating and
>>>>>> mapping huge order pages, you are certainly more qualified to help Akinobu
>>>>>> than I am.
>>>>
>>>> What Akinobu's patches intend to support is:
>>>>
>>>>          phys_addr = dma_alloc_coherent(dev, 64 * 1024 * 1024, &bus_addr, GFP_KERNEL);
>>>>
>>>> which raises three issues:
>>>>
>>>> 1. Where do coherent blocks of this size come from?
>>>> 2. How to prevent fragmentation of these reserved blocks over time by
>>>>     existing DMA users?
>>>> 3. Is this support generically required across all iommu implementations on x86?
>>>>
>>>> Questions 1 and 2 are non-trivial, in the general case, otherwise the page
>>>> allocator would already do this. Simply dropping in the contiguous memory
>>>> allocator doesn't work because CMA does not have the same policy and performance
>>>> as the page allocator, and is already causing performance regressions even
>>>> in the absence of huge page allocations.
>>>
>>> Could you take a look at the patches I sent?  Can they fix these issues?
>>> https://lkml.org/lkml/2014/9/28/110
>>>
>>> With these patches, normal alloc_pages() is used for allocation first
>>> and dma_alloc_from_contiguous() is used as a fallback.
>>
>> Sure, I can test these patches this weekend.
>> Where are the unit tests?
>
> Thanks a lot.  I would like to know whether the performance regression
> you see will disappear or not with these patches as if CONFIG_DMA_CMA is
> disabled.
>
>>>> So that's why I raised question 3; is making the necessary compromises to support
>>>> 64MB coherent DMA allocations across all x86 iommu implementations actually
>>>> required?
>>>>
>>>> Prior to Akinobu's patches, the use of CMA by x86 iommu configurations was
>>>> designed to be limited to testing configurations, as the introductory
>>>> commit states:
>>>>
>>>> commit 0a2b9a6ea93650b8a00f9fd5ee8fdd25671e2df6
>>>> Author: Marek Szyprowski <m.szyprowski-Sze3O3UU22JBDgjK7y7TUQ@public.gmane.org>
>>>> Date:   Thu Dec 29 13:09:51 2011 +0100
>>>>
>>>>      X86: integrate CMA with DMA-mapping subsystem
>>>>
>>>>      This patch adds support for CMA to dma-mapping subsystem for x86
>>>>      architecture that uses common pci-dma/pci-nommu implementation. This
>>>>      allows to test CMA on KVM/QEMU and a lot of common x86 boxes.
>>>>
>>>>      Signed-off-by: Marek Szyprowski <m.szyprowski-Sze3O3UU22JBDgjK7y7TUQ@public.gmane.org>
>>>>      Signed-off-by: Kyungmin Park <kyungmin.park-Sze3O3UU22JBDgjK7y7TUQ@public.gmane.org>
>>>>      CC: Michal Nazarewicz <mina86-deATy8a+UHjQT0dZR+AlfA@public.gmane.org>
>>>>      Acked-by: Arnd Bergmann <arnd-r2nGTMty4D4@public.gmane.org>
>>>>
>>>>
>>>> Which brings me to my suggestion: if support for huge coherent DMA is
>>>> required only for a special test platform, then could not this support
>>>> be specific to a new iommu configuration, namely iommu=cma, which would
>>>> get initialized much the same way that iommu=calgary is now.
>>>>
>>>> The code for such a iommu configuration would mostly duplicate
>>>> arch/x86/kernel/pci-swiotlb.c and the CMA support would get removed from
>>>> the other x86 iommu implementations.
>>>
>>> I'm not sure I read correctly, though.  Can boot option 'cma=0' also
>>> help avoiding CMA from IOMMU implementation?
>>
>> Maybe, but that's not an appropriate solution for distro kernels.
>>
>> Nor does this address configurations that want a really large CMA so
>> 1GB huge pages can be allocated (not for DMA though).
>
> Now I see the point of iommu=cma you suggested.  But what should we do
> when CONFIG_SWIOTLB is disabled, especially for x86_32?
> Should we just introduce yet another flag to tell not using DMA_CMA
> instead of adding new swiotlb-like iommu implementation?
>

If you implement an DMA API producer - aka dma_ops (which is what Peter 
is thinking I believe) it won't matter which IOMMUs / DMA producers are 
selected right?

Or are you saying that CMA needs SWIOTLB to handle certain type of
pages as a fallback mechanism - and hence there needs to be a tight
relationship?

In which case I would look at making SWIOTLB be more library like - the 
Xen-SWIOTLB already does that by using certain parts of the SWIOTLB code
which are exposed to the rest of the kernel.