Re: [PATCH v10 01/17] iova: Export alloc_iova_fast() and free_iova_fast()

[Yongji Xie <xieyongji@bytedance.com>]

On Tue, Aug 3, 2021 at 6:54 PM Robin Murphy <robin.murphy@arm.com> wrote:
>
> On 2021-08-03 09:54, Yongji Xie wrote:
> > On Tue, Aug 3, 2021 at 3:41 PM Jason Wang <jasowang@redhat.com> wrote:
> >>
> >>
> >> 在 2021/7/29 下午3:34, Xie Yongji 写道:
> >>> Export alloc_iova_fast() and free_iova_fast() so that
> >>> some modules can use it to improve iova allocation efficiency.
> >>
> >>
> >> It's better to explain why alloc_iova() is not sufficient here.
> >>
> >
> > Fine.
>
> What I fail to understand from the later patches is what the IOVA domain
> actually represents. If the "device" is a userspace process then
> logically the "IOVA" would be the userspace address, so presumably
> somewhere you're having to translate between this arbitrary address
> space and actual usable addresses - if you're worried about efficiency
> surely it would be even better to not do that?
>

Yes, userspace daemon needs to translate the "IOVA" in a DMA
descriptor to the VA (from mmap(2)). But this actually doesn't affect
performance since it's an identical mapping in most cases.

> Presumably userspace doesn't have any concern about alignment and the
> things we have to worry about for the DMA API in general, so it's pretty
> much just allocating slots in a buffer, and there are far more effective
> ways to do that than a full-blown address space manager.

Considering iova allocation efficiency, I think the iova allocator is
better here. In most cases, we don't even need to hold a spin lock
during iova allocation.

> If you're going
> to reuse any infrastructure I'd have expected it to be SWIOTLB rather
> than the IOVA allocator. Because, y'know, you're *literally implementing
> a software I/O TLB* ;)
>

But actually what we can reuse in SWIOTLB is the IOVA allocator. And
the IOVA management in SWIOTLB is not what we want. For example,
SWIOTLB allocates and uses contiguous memory for bouncing, which is
not necessary in VDUSE case. And VDUSE needs coherent mapping which is
not supported by the SWIOTLB. Besides, the SWIOTLB works in singleton
mode (designed for platform IOMMU) , but VDUSE is based on on-chip
IOMMU (supports multiple instances). So I still prefer to reuse the
IOVA allocator to implement a MMU-based software IOTLB.

Thanks,
Yongji