Re: [PATCH v1 4/4] iommu/tegra: gart: Optimize map/unmap

[Dmitry Osipenko]

On 07.05.2018 11:04, Joerg Roedel wrote:
> On Mon, May 07, 2018 at 12:19:01AM +0300, Dmitry Osipenko wrote:
>> Probably the best variant would be to give an explicit control over syncing to a
>> user of the IOMMU API, like for example device driver may perform multiple
>> mappings / unmappings and then sync/flush in the end. I'm not sure that it's
>> really worth the hassle to shuffle the API right now, maybe we can implement it
>> later if needed. Joerg, do you have objections to a 'compound page' approach?
> 
> Have you measured the performance difference on both variants? The
> compound-page approach only works for cases when the physical memory you
> map contiguous and correctly aligned.

Yes, previously I actually only tested mapping of the contiguous allocations
(used for memory isolation purposes). But now I've re-tested all variants and
got somewhat interesting results.

Firstly it is not that easy to test a really sparse mapping simply because
memory allocator produces sparse allocation only when memory is _really_
fragmented. Pretty much all of the time the sparse allocations are contiguous or
they consist of a very few chunks that do not impose any noticeable performance
impact.

Secondly, the interesting part is that mapping / unmapping of a contiguous
allocation (CMA using DMA API) is slower by ~50% then doing it for a sparse
allocation (get_pages using bare IOMMU API). /I think/ it's a shortcoming of the
arch/arm/mm/dma-mapping.c, which also suffers from other inflexibilities that
Thierry faced recently. Though I haven't really tried to figure out what is the
bottleneck yet and Thierry was going to re-write ARM's dma-mapping
implementation anyway, I'll take a closer look at this issue a bit later.

I've implemented the iotlb_sync_map() and tested things with it. The end result
is the same as for the compound page approach, simply because actual allocations
are pretty much always contiguous.

> If it is really needed I would prefer a separate iotlb_sync_map()
> call-back that is just NULL when not needed. This way all users that
> don't need it only get a minimal penalty in the mapping path and you
> don't have any requirements on the physical memory you map to get good
> performance.
Summarizing, the iotlb_sync_map() is indeed better way. As you rightly noticed,
that approach is also optimal for the non-contiguous cases as we won't have to
flush on mapping of each contiguous chunk of the sparse allocation, but after
the whole mapping is done.

Thierry, Robin and Joerg - thanks for your input, I'll prepare patches
implementing the iotlb_sync_map.