[PATCH v2 0/5] Optimize iommu_map_sg() performance

[Robin Murphy]

On 2021-01-21 21:30, isaacm at codeaurora.org wrote:
> On 2021-01-12 08:00, Robin Murphy wrote:
>> On 2021-01-11 14:54, Isaac J. Manjarres wrote:
>>> The iommu_map_sg() code currently iterates through the given
>>> scatter-gather list, and in the worst case, invokes iommu_map()
>>> for each element in the scatter-gather list, which calls into
>>> the IOMMU driver through an indirect call. For an IOMMU driver
>>> that uses a format supported by the io-pgtable code, the IOMMU
>>> driver will then call into the io-pgtable code to map the chunk.
>>>
>>> Jumping between the IOMMU core code, the IOMMU driver, and the
>>> io-pgtable code and back for each element in a scatter-gather list
>>> is not efficient.
>>>
>>> Instead, add a map_sg() hook in both the IOMMU driver ops and the
>>> io-pgtable ops. iommu_map_sg() can then call into the IOMMU driver's
>>> map_sg() hook with the entire scatter-gather list, which can call
>>> into the io-pgtable map_sg() hook, which can process the entire
>>> scatter-gather list, signficantly reducing the number of indirect
>>> calls, and jumps between these layers, boosting performance.
>>
>> Out of curiosity, how much of the difference is attributable to actual
>> indirect call overhead vs. the additional massive reduction in visits
>> to arm_smmu_rpm_{get,put} that you fail to mention?There are ways to
> I did an experiment where I compared the two approaches without any calls
> to arm_smmu_rpm_[get/put]. There's still a large amount of difference
> without the overhead incurred by power management calls. Here are the 
> results:
> 
> no optimizations and no power management calls:
>   size        iommu_map_sg
>        4K            0.609 us
>       64K            8.583 us
>        1M          136.083 us
>        2M          273.145 us
>       12M         1442.119 us
>       24M         2876.078 us
>       32M         3832.041 us
> 
> iommu_map_sg optimizations and no power management calls:
> size        iommu_map_sg
>        4K            0.645 us
>       64K            1.229 us
>        1M            9.531 us
>        2M           23.198 us
>       12M           99.250 us
>       24M          185.713 us
>       32M          248.781 us
> 
>  From here, we can see that the amount of latency incurred by the indirect
> calls is fairly large.

OK, that's pretty much in line with what I was imagining, just wanted to 
confirm (if you ended up actually changing the power state around each 
page then the caller would likely be doing something very stupid).

I'm guessing the optimised numbers above looking ~20% slower than the 
ones below is just indicative of a high variance between runs, or maybe 
there's some funky cache interaction that really does make the RPM 
checks have effectively negative overhead.

>> optimise indirect calling that would benefit *all* cases, rather than
>> just one operation for one particular driver.
> Do you mind sharing some more information on how to optimize the existing
> approach further, such that it benefits other drivers as well?

This article touches on some of the possible techniques:

https://lwn.net/Articles/774743/

>>> On a system that uses the ARM SMMU driver, and the ARM LPAE format,
>>> the current implementation of iommu_map_sg() yields the following
>>> latencies for mapping scatter-gather lists of various sizes. These
>>> latencies are calculated by repeating the mapping operation 10 times:
>>>
>>>      size        iommu_map_sg latency
>>>        4K            0.624 us
>>>       64K            9.468 us
>>>        1M          122.557 us
>>>        2M          239.807 us
>>>       12M         1435.979 us
>>>       24M         2884.968 us
>>>       32M         3832.979 us
>>>
>>> On the same system, the proposed modifications yield the following
>>> results:
>>>
>>>      size        iommu_map_sg latency
>>>        4K            3.645 us
>>>       64K            4.198 us
>>>        1M           11.010 us
>>>        2M           17.125 us
>>>       12M           82.416 us
>>>       24M          158.677 us
>>>       32M          210.468 us
>>>
>>> The procedure for collecting the iommu_map_sg latencies is
>>> the same in both experiments. Clearly, reducing the jumps
>>> between the different layers in the IOMMU code offers a
>>> signficant performance boost in iommu_map_sg() latency.
>>
>> Presumably those are deliberately worst-case numbers? After all, a
>> 32MB scatterlist *could* incur less overhead than a 64KB one if things
>> line up just right (still 16 ->map calls, but each with one fewer
> Yes, these are worst case numbers (i.e. a buffer is composed entirely
> of 4 KB pages, so higher order mappings don't get used).
>> level of pagetable to traverse). TBH I find the significant regression
>> of the 4KB case the most interesting - what's going on there?
> That was an error on my part. After fixing my error, I observed that the
> time spent mapping the 4 KB buffer is comparable with and without 
> optimizations,
> which is expected.
>>
>> My main reservation here is that we get an explosion of duplicate
>> copies of almost the same code, and it's code that's just non-trivial
>> enough to start being bug-prone. And it's all still only for one
>> specific operation - your argument about calling through multiple
>> layers for each element applies just as much to iommu_map() itself, so
>> why aren't we trying to make more fundamental improvements with wider
>> benefits? Indeed I can't imagine the existing iommu_map_sg() loop
>> really adds significant overhead compared to a single iommu_map() call
>> that results in the equivalent set of ->map calls to the driver.
>>
>> At a glance, I reckon that simply extending the internal ->map and
>> ->unmap interfaces to encode a number of consecutive identical pages
>> would already get us a large chunk of the way there; then we'd be in a
>> better place to consider options for the io-pgtable interface.
>>
> Do you mean physically contiguous pages? If so, that still wouldn't help 
> the
> case where a buffer is composed entirely of 4 KB pages, correct?

Indeed, simply reducing the number of internal calls will be a fairly 
cheap win for most typical cases - both dma_map_page() for more than one 
page, and dma_map_sg() from users like the block layer rather than 
gigantic pathological dma-buf imports - but we still want to work on 
getting the individual call overhead down to a reasonable level as well.

Thanks,
Robin.

>>> Changes since v1:
>>>
>>> -Fixed an off by one error in arm_[lpae/v7s]_map_by_pgsize
>>> when checking if the IOVA and physical address ranges being
>>> mapped are within the appropriate limits.
>>> -Added Sai Prakash Ranjan's "Tested-by" tag.
>>>
>>> Thanks,
>>> Isaac
>>>
>>> Isaac J. Manjarres (5):
>>>    iommu/io-pgtable: Introduce map_sg() as a page table op
>>>    iommu/io-pgtable-arm: Hook up map_sg()
>>>    iommu/io-pgtable-arm-v7s: Hook up map_sg()
>>>    iommu: Introduce map_sg() as an IOMMU op for IOMMU drivers
>>>    iommu/arm-smmu: Hook up map_sg()
>>>
>>>   drivers/iommu/arm/arm-smmu/arm-smmu.c | 19 ++++++++
>>>   drivers/iommu/io-pgtable-arm-v7s.c    | 90 
>>> +++++++++++++++++++++++++++++++++++
>>>   drivers/iommu/io-pgtable-arm.c        | 86 
>>> +++++++++++++++++++++++++++++++++
>>>   drivers/iommu/iommu.c                 | 25 ++++++++--
>>>   include/linux/io-pgtable.h            |  6 +++
>>>   include/linux/iommu.h                 | 13 +++++
>>>   6 files changed, 234 insertions(+), 5 deletions(-)
>>>
>>
>> _______________________________________________
>> linux-arm-kernel mailing list
>> linux-arm-kernel at lists.infradead.org
>> http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
> _______________________________________________
> iommu mailing list
> iommu at lists.linux-foundation.org
> https://lists.linuxfoundation.org/mailman/listinfo/iommu