Re: [PATCH] vfio/type1: optimize vfio_pin_pages_remote() for hugetlbfs folio

[Alex Williamson <alex.williamson@redhat.com>]

On Fri, 16 May 2025 11:18:16 -0300
Jason Gunthorpe <jgg@ziepe.ca> wrote:

> On Thu, May 15, 2025 at 03:19:46PM -0600, Alex Williamson wrote:
> > On Tue, 13 May 2025 11:57:30 +0800
> > lizhe.67@bytedance.com wrote:
> >   
> > > From: Li Zhe <lizhe.67@bytedance.com>
> > > 
> > > When vfio_pin_pages_remote() is called with a range of addresses that
> > > includes hugetlbfs folios, the function currently performs individual
> > > statistics counting operations for each page. This can lead to significant
> > > performance overheads, especially when dealing with large ranges of pages.
> > > 
> > > This patch optimize this process by batching the statistics counting
> > > operations.
> > > 
> > > The performance test results for completing the 8G VFIO IOMMU DMA mapping,
> > > obtained through trace-cmd, are as follows. In this case, the 8G virtual
> > > address space has been mapped to physical memory using hugetlbfs with
> > > pagesize=2M.
> > > 
> > > Before this patch:
> > > funcgraph_entry:      # 33813.703 us |  vfio_pin_map_dma();
> > > 
> > > After this patch:
> > > funcgraph_entry:      # 15635.055 us |  vfio_pin_map_dma();
> > > 
> > > Signed-off-by: Li Zhe <lizhe.67@bytedance.com>
> > > ---
> > >  drivers/vfio/vfio_iommu_type1.c | 49 +++++++++++++++++++++++++++++++++
> > >  1 file changed, 49 insertions(+)  
> > 
> > Hi,
> > 
> > Thanks for looking at improvements in this area...  
> 
> Why not just use iommufd? Doesn't it already does all these
> optimizations?

We don't have feature parity yet (P2P DMA), we don't have libvirt
support, and many users are on kernels or product stacks where iommufd
isn't available yet.

> Indeed today you can use iommufd with a memfd handle which should
> return the huge folios directly from the hugetlbfs and we never
> iterate with 4K pages.

Good to know we won't need to revisit this, and maybe "good enough"
here, without tackling the batch size or gup page pointers is enough to
hold users over until iommufd.  Thanks,

Alex