Re: [PATCH RESEND v2 0/8] Cache-coherent DMA access using UIO

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

On Thu, 11 Aug 2016 09:30:19 +0530
Anup Patel <anup.patel@broadcom.com> wrote:

> Hi Arnd,
> 
> On Wed, Aug 10, 2016 at 9:25 PM, Arnd Bergmann <arnd@arndb.de> wrote:
> > On Monday, August 8, 2016 11:22:29 AM CEST Anup Patel wrote:  
> >> The goal of this patchset is to improve UIO framework and UIO dmem
> >> driver to allow cache-coherent DMA accesses from user-space.
> >>
> >> This patchset is based on two previous patchsets:
> >> 1) [PATCH v5 0/6] UIO driver for APM X-Gene QMTM
> >> (Refer, http://www.spinics.net/lists/devicetree/msg58244.html)
> >> 2) [PATCH 0/4] Fix and extend uio_dmem_genirq
> >> (Refer, https://lkml.org/lkml/2016/5/17/141)
> >>
> >> We have adopted only patch0-3 of patchset1 which was abandoned
> >> long time back. We have taken care of last few unaddressed comments
> >> on these patches.
> >>
> >> The patchset2 is quite recent has been adopted entirely. We have
> >> taken care review comments on these patches too.
> >>
> >> This patchset is based on v4.7-rc7 tag and it is available in uio-v2
> >> branch of https://github.com/Broadcom/arm64-linux.git  
> >
> >
> > UIO devices are generally meant to be things that do not
> > perform DMA and that don't screw up the rest of the system
> > when misused. A device that is able to access any physical
> > memory doesn't belong into this category. The way that
> > uio_dmem_genirq.c gets around this is by requiring the device
> > to be created by some code that sets up a separate IOMMU
> > domain first, but the DT probing here doesn't do that.
> > Note that IOMMU domains typically use 32-bit addressing,
> > so the entire "dma_mask from property" dance isn't even
> > required.  
> 
> IMHO, UIO devices are meant for things that are not behind
> any IOMMU hardware.
> 
> Yes, any mis-programming in user space using UIO can
> potentially screw-up the rest of the system but this is
> generally a known/assumed fact for people who are using UIO.
> 
> >
> > Also, this seems to duplicate a lot of the work that
> > went into "vfio". Can you explain why we need another way
> > of doing the same thing here?  
> 
> We can only use "vfio" for devices that are behind some
> kind of IOMMU (Right??). For devices not having IOMMU
> support will have to use UIO for user space access.
> 
> Particularly, there are lot of FPGA-based solutions and legacy
> hardware which do not have IOMMU support (devices on
> FPGA or specific devices).
> 
> In our use case, we have some FPGA-based device which
> does not have IOMMU support and we are accessing this
> FPGA-based device from user-space.
> 
> This patchset only tries to extend "uio" and "uio_dmem_genirq".
> There is no intention of duplicating what has been already
> done for "vfio".
> 
> I do agree that "vfio" should eventually become defacto method
> of accessing devices in user space but that requires devices to
> always have IOMMU support.

A vfio no-iommu mode exists since v4.5:

http://git.kernel.org/cgit/linux/kernel/git/torvalds/linux.git/commit/?id=03a76b60f8ba27974e2d252bc555d2c103420e15

Thanks,
Alex