Re: [Qemu-devel] [PATCH RFC 0/3] vfio: allow to notify unmap for very big region

[Peter Xu]

On Thu, Jan 19, 2017 at 09:21:10PM -0700, Alex Williamson wrote:
> On Fri, 20 Jan 2017 11:43:28 +0800
> Peter Xu <address@hidden> wrote:
> 
> > On Thu, Jan 19, 2017 at 10:54:37AM -0700, Alex Williamson wrote:
> > > On Thu, 19 Jan 2017 17:25:29 +0800
> > > Peter Xu <address@hidden> wrote:
> > >   
> > > > This requirement originates from the VT-d vfio series:
> > > > 
> > > >   https://lists.nongnu.org/archive/html/qemu-devel/2017-01/msg03495.html
> > > > 
> > > > The goal of this series is to allow IOMMU to notify unmap with very
> > > > big IOTLB range, for example, with base=0 and size=2^63-1 (to unmap
> > > > the whole address space).
> > > > 
> > > > The first patch is a good to have, for traces.
> > > > 
> > > > The second one is a cleanup of existing code, only.  
> > > 
> > > Sort of, it does add some overhead to the unmap path, but you remove
> > > that and more in the third patch.  
> > 
> > Hmm, yes, I tried to get the ram pointer even for unmap. I should
> > remove the ending "only".
> > 
> > >    
> > > > The third one moves the further RAM translation and check into map
> > > > operation logic, so that it'll free unmap operations.
> > > > 
> > > > The series is marked as RFC since I am not sure whether this is a
> > > > workable way. Anyway, please review to help confirm it.  
> > > 
> > > It seems reasonable to me,  
> > 
> > Good to know that you didn't disagree on this. :) Then let me take
> > these patches along with that series in the next post (since that
> > series will depend on this one, so I'll keep them together, though
> > please maintainers decide on how you'll merge them when you want to),
> > 
> > > except for the example here of using 2^63-1,
> > > which I expect is to work around the vfio {un}map API bug as we
> > > discussed on irc.  
> > 
> > Actually not only for that one, I don't know whether we have problem
> > here:
> > 
> > (I mentioned this in IRC, in case you missed that, I paste it here as
> >  well)
> > 
> > static struct vfio_dma *vfio_find_dma(struct vfio_iommu *iommu,
> >                                   dma_addr_t start, size_t size)
> > {
> >     struct rb_node *node = iommu->dma_list.rb_node;
> > 
> >     while (node) {
> >             struct vfio_dma *dma = rb_entry(node, struct vfio_dma, node);
> > 
> >             if (start + size <= dma->iova)
> >                     node = node->rb_left;
> >             else if (start >= dma->iova + dma->size)
> >                           ^^^^^^^^^^^^^^^^^^^^^
> >                           Could it overflow here? <--------------
> >                     node = node->rb_right;
> >             else
> >                     return dma;
> >     }
> > 
> >     return NULL;
> > }
> > 
> > I used 2^63-1 to try to avoid both cases.
> 
> 
> Perhaps I'm not seeing the overflow you're trying to identify,
> but dma->iova + dma->size cannot overflow, these are existing DMA
> mappings and the DMA mapping path does test for overflow.  I think we
> can consider those sanitized.  I am wondering if the test above it is
> suspect though, start + size doesn't seem to be checked for overflow on
> the unmap path.  It seems pretty easy to avoid from the user side
> though, but maybe we should add a test for it in the kernel.

Ah, yes. :-)

> 
>  
> > > For everyone, the root of the problem is that the
> > > ioctls use:
> > > 
> > >         __u64   iova;                           /* IO virtual address */
> > >         __u64   size;                           /* Size of mapping 
> > > (bytes) */
> > > 
> > > So we can only express a size of 2^64-1 and we have an off-by-one error
> > > trying to perform a map/unmap of an entire 64-bit space.  Note when
> > > designing an API, use start/end rather than base/size to avoid this.  
> > 
> > Lesson learned.
> > 
> > > 
> > > What I don't want to see is for this API bug to leak out into the rest
> > > of the QEMU code such that intel_iommu code, or iommu code in general
> > > subtly avoids it by artificially using a smaller range.  VT-d hardware
> > > has an actual physical address space of either 2^39 or 2^48 bits, so if
> > > you want to make the iommu address space match the device we're trying
> > > to emulate, that's perfectly fine.  AIUI, AMD-Vi does actually have a
> > > 64-bit address space on the IOMMU, so to handle that case I'd expect
> > > the simplest solution would be to track the and mapped iova high water
> > > mark per container in vfio and truncate unmaps to that high water end
> > > address.  Realistically we're probably not going to see iovas at the end
> > > of the 64-bit address space, but we can come up with some other
> > > workaround in the vfio code or update the kernel API if we do.  Thanks,  
> > 
> > Agree that high watermark can be a good solution for VT-d. I'll use
> > that instead of 2^63-1. Though for AMD, if it supports 64bits, we may
> > still need to solve existing issues in the future, since in that case
> > the high watermark can be (hwaddr)-1 as well if guest specifies it.
> > 
> > (Though anyway I'm not sure when AMD vIOMMUs will be ready for device
> >  assignment, so I don't think that's a big issue at least for now)
> 
> Even with a true 64bit address space, it seems very unlikely we'd have
> iovas at the very top of that address space.

Yes, especially if the IOVA is allocated inside kernel. However, since
we are with DPDK these days, that's still possible?

I remembered that Jason has mentioned about OOM attacker for a tree -
since we have a per-domain tree in kernel vfio, it's also possible
that aggresive guest (DPDK user) doing special pattern of mapping that
can let the tree grow into a very big one, finally exaust the host
memory. Is this a problem as well?

> If we do, another idea
> would be that iommus often have a reserved iova range for MSI mapping,
> both VT-d and AMD-Vi do.  We know there would be no user iovas within
> those ranges, so it would make a clean point to split a full 64-bit
> unmap in two.  Hopefully we'll have reporting of that reserved range up
> through vfio should/when we need something like that.

Since we are at here... IIUC VT-d has reserved 0xfeeXXXXX for MSI, do
we have protection in intel iommu driver that user should not use IOVA
inside this range (I guess not?)? Though this is totally out of topic
since even this is a problem, it'll be for intel-iommu driver's.

> We could
> also simply track both the start and end of that mapping that sets the
> high water mark, unmapping all except that, then that alone as two
> separate ioctls.  I'm not too concerned that we'll be able to work
> around it if we need to.  Thanks,

Sure. Thanks,

-- peterx