Re: [PATCH v2] vfio iommu type1: Improve vfio_iommu_type1_pin_pages performance

From: Cornelia Huck
Date: Wed Dec 09 2020 - 06:57:14 EST


On Tue, 8 Dec 2020 21:55:53 +0800
"xuxiaoyang (C)" <xuxiaoyang2@xxxxxxxxxx> wrote:

> On 2020/11/21 15:58, xuxiaoyang (C) wrote:
> > vfio_pin_pages() accepts an array of unrelated iova pfns and processes
> > each to return the physical pfn. When dealing with large arrays of
> > contiguous iovas, vfio_iommu_type1_pin_pages is very inefficient because
> > it is processed page by page.In this case, we can divide the iova pfn
> > array into multiple continuous ranges and optimize them. For example,
> > when the iova pfn array is {1,5,6,7,9}, it will be divided into three
> > groups {1}, {5,6,7}, {9} for processing. When processing {5,6,7}, the
> > number of calls to pin_user_pages_remote is reduced from 3 times to once.
> > For single page or large array of discontinuous iovas, we still use
> > vfio_pin_page_external to deal with it to reduce the performance loss
> > caused by refactoring.
> >
> > Signed-off-by: Xiaoyang Xu <xuxiaoyang2@xxxxxxxxxx>

(...)

>
> hi Cornelia Huck, Eric Farman, Zhenyu Wang, Zhi Wang
>
> vfio_pin_pages() accepts an array of unrelated iova pfns and processes
> each to return the physical pfn. When dealing with large arrays of
> contiguous iovas, vfio_iommu_type1_pin_pages is very inefficient because
> it is processed page by page. In this case, we can divide the iova pfn
> array into multiple continuous ranges and optimize them. I have a set
> of performance test data for reference.
>
> The patch was not applied
> 1 page 512 pages
> no huge pages: 1638ns 223651ns
> THP: 1668ns 222330ns
> HugeTLB: 1526ns 208151ns
>
> The patch was applied
> 1 page 512 pages
> no huge pages 1735ns 167286ns
> THP: 1934ns 126900ns
> HugeTLB: 1713ns 102188ns
>
> As Alex Williamson said, this patch lacks proof that it works in the
> real world. I think you will have some valuable opinions.

Looking at this from the vfio-ccw angle, I'm not sure how much this
would buy us, as we deal with IDAWs, which are designed so that they
can be non-contiguous. I guess this depends a lot on what the guest
does.

Eric, any opinion? Do you maybe also happen to have a test setup that
mimics workloads actually seen in the real world?