[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

RE: [PATCH] xen: introduce xen_vring_use_dma



> Subject: Re: [PATCH] xen: introduce xen_vring_use_dma
> 
> On Thu, Jun 25, 2020 at 10:31:27AM -0700, Stefano Stabellini wrote:
> > On Wed, 24 Jun 2020, Michael S. Tsirkin wrote:
> > > On Wed, Jun 24, 2020 at 02:53:54PM -0700, Stefano Stabellini wrote:
> > > > On Wed, 24 Jun 2020, Michael S. Tsirkin wrote:
> > > > > On Wed, Jun 24, 2020 at 10:59:47AM -0700, Stefano Stabellini wrote:
> > > > > > On Wed, 24 Jun 2020, Michael S. Tsirkin wrote:
> > > > > > > On Wed, Jun 24, 2020 at 05:17:32PM +0800, Peng Fan wrote:
> > > > > > > > Export xen_swiotlb for all platforms using xen swiotlb
> > > > > > > >
> > > > > > > > Use xen_swiotlb to determine when vring should use dma
> > > > > > > > APIs to map the
> > > > > > > > ring: when xen_swiotlb is enabled the dma API is required.
> > > > > > > > When it is disabled, it is not required.
> > > > > > > >
> > > > > > > > Signed-off-by: Peng Fan <peng.fan@xxxxxxx>
> > > > > > >
> > > > > > > Isn't there some way to use VIRTIO_F_IOMMU_PLATFORM for
> this?
> > > > > > > Xen was there first, but everyone else is using that now.
> > > > > >
> > > > > > Unfortunately it is complicated and it is not related to
> > > > > > VIRTIO_F_IOMMU_PLATFORM :-(
> > > > > >
> > > > > >
> > > > > > The Xen subsystem in Linux uses dma_ops via swiotlb_xen to
> > > > > > translate foreign mappings (memory coming from other VMs) to
> physical addresses.
> > > > > > On x86, it also uses dma_ops to translate Linux's idea of a
> > > > > > physical address into a real physical address (this is
> > > > > > unneeded on ARM.)
> > > > > >
> > > > > >
> > > > > > So regardless of VIRTIO_F_IOMMU_PLATFORM, dma_ops should be
> > > > > > used on Xen/x86 always and on Xen/ARM if Linux is Dom0
> > > > > > (because it has foreign
> > > > > > mappings.) That is why we have the if (xen_domain) return
> > > > > > true; in vring_use_dma_api.
> > > > >
> > > > > VIRTIO_F_IOMMU_PLATFORM makes guest always use DMA ops.
> > > > >
> > > > > Xen hack predates VIRTIO_F_IOMMU_PLATFORM so it *also* forces
> > > > > DMA ops even if VIRTIO_F_IOMMU_PLATFORM is clear.
> > > > >
> > > > > Unfortunately as a result Xen never got around to properly
> > > > > setting VIRTIO_F_IOMMU_PLATFORM.
> > > >
> > > > I don't think VIRTIO_F_IOMMU_PLATFORM would be correct for this
> > > > because the usage of swiotlb_xen is not a property of virtio,
> > >
> > >
> > > Basically any device without VIRTIO_F_ACCESS_PLATFORM (that is it's
> > > name in latest virtio spec, VIRTIO_F_IOMMU_PLATFORM is what linux
> > > calls it) is declared as "special, don't follow normal rules for
> > > access".
> > >
> > > So yes swiotlb_xen is not a property of virtio, but what *is* a
> > > property of virtio is that it's not special, just a regular device from 
> > > DMA
> POV.
> >
> > I am trying to understand what you meant but I think I am missing
> > something.
> >
> > Are you saying that modern virtio should always have
> > VIRTIO_F_ACCESS_PLATFORM, hence use normal dma_ops as any other
> devices?
> 
> I am saying it's a safe default. Clear VIRTIO_F_ACCESS_PLATFORM if you have
> some special needs e.g. you are very sure it's ok to bypass DMA ops, or you
> need to support a legacy guest (produced in the window between virtio 1
> support in 2014 and support for VIRTIO_F_ACCESS_PLATFORM in 2016).
> 
> 
> > If that is the case, how is it possible that virtio breaks on ARM
> > using the default dma_ops? The breakage is not Xen related (except
> > that Xen turns dma_ops on). The original message from Peng was:
> >
> >   vring_map_one_sg -> vring_use_dma_api
> >                    -> dma_map_page
> >                    -> __swiotlb_map_page
> >                             ->swiotlb_map_page
> >                             ->__dma_map_area(phys_to_virt(dma_to_phys(dev,
> dev_addr)), size, dir);
> >   However we are using per device dma area for rpmsg, phys_to_virt
> >   could not return a correct virtual address for virtual address in
> >   vmalloc area. Then kernel panic.
> >
> > I must be missing something. Maybe it is because it has to do with RPMesg?
> 
> I think it's an RPMesg bug, yes

rpmsg bug is another issue, it should not use dma_alloc_coherent for reserved 
area,
and use vmalloc_to_page.

Anyway here using dma api will also trigger issue.

> 
> >
> > > > > > You might have noticed that I missed one possible case above:
> > > > > > Xen/ARM DomU :-)
> > > > > >
> > > > > > Xen/ARM domUs don't need swiotlb_xen, it is not even
> > > > > > initialized. So if
> > > > > > (xen_domain) return true; would give the wrong answer in that case.
> > > > > > Linux would end up calling the "normal" dma_ops, not
> > > > > > swiotlb-xen, and the "normal" dma_ops fail.
> > > > > >
> > > > > >
> > > > > > The solution I suggested was to make the check in
> > > > > > vring_use_dma_api more flexible by returning true if the
> > > > > > swiotlb_xen is supposed to be used, not in general for all Xen
> > > > > > domains, because that is what the check was really meant to do.
> > > > >
> > > > > Why not fix DMA ops so they DTRT (nop) on Xen/ARM DomU? What is
> wrong with that?
> > > >
> > > > swiotlb-xen is not used on Xen/ARM DomU, the default dma_ops are
> > > > the ones that are used. So you are saying, why don't we fix the
> > > > default dma_ops to work with virtio?
> > > >
> > > > It is bad that the default dma_ops crash with virtio, so yes I
> > > > think it would be good to fix that. However, even if we fixed
> > > > that, the if
> > > > (xen_domain()) check in vring_use_dma_api is still a problem.
> > >
> > > Why is it a problem? It just makes virtio use DMA API.
> > > If that in turn works, problem solved.
> >
> > You are correct in the sense that it would work. However I do think it
> > is wrong for vring_use_dma_api to enable dma_ops/swiotlb-xen for
> > Xen/ARM DomUs that don't need it. There are many different types of
> > Xen guests, Xen x86 is drastically different from Xen ARM, it seems
> > wrong to treat them the same way.
> 
> I could imagine some future Xen hosts setting a flag somewhere in the
> platform capability saying "no xen specific flag, rely on
> "VIRTIO_F_ACCESS_PLATFORM". Then you set that accordingly in QEMU.
> How about that?
> 

Michael, Stefano,

So what's your suggestion here, that we could avoid similar issue
for virtio drivers in ARM DomU?

Thanks,
Peng.

> 
> >
> >
> > Anyway, re-reading the last messages of the original thread [1], it
> > looks like Peng had a clear idea on how to fix the general issue.
> > Peng, what happened with that?

We shrinked the rpmsg reserved area to workaround the issue.
So still use the dma apis in rpmsg.

But here I am going to address domu android trusty issue using
virtio.

> >
> >
> > [1]
> > https://eur01.safelinks.protection.outlook.com/?url=https%3A%2F%2Flore
> > .kernel.org%2Fpatchwork%2Fpatch%2F1033801%2F%231222404&amp;dat
> a=02%7C0
> >
> 1%7Cpeng.fan%40nxp.com%7C08ba48d3b3d54e775a8108d819e62fd0%7C68
> 6ea1d3bc
> >
> 2b4c6fa92cd99c5c301635%7C0%7C0%7C637287823721994475&amp;sdata
> =Cw4FHWrH
> > uVKBCn3%2BKS2VM7cWuGoTI6R7SHJrJSLY5Io%3D&amp;reserved=0




 


Rackspace

Lists.xenproject.org is hosted with RackSpace, monitoring our
servers 24x7x365 and backed by RackSpace's Fanatical Support®.