Re: [Xen-devel] [PATCH 0/1] drm/xen-zcopy: Add Xen zero-copy helper DRM driver

["Roger Pau Monné" <roger.pau@citrix.com>]

On Wed, Apr 18, 2018 at 11:01:12AM +0300, Oleksandr Andrushchenko wrote:
> On 04/18/2018 10:35 AM, Roger Pau Monné wrote:
> > On Wed, Apr 18, 2018 at 09:38:39AM +0300, Oleksandr Andrushchenko wrote:
> > > On 04/17/2018 11:57 PM, Dongwon Kim wrote:
> > > > On Tue, Apr 17, 2018 at 09:59:28AM +0200, Daniel Vetter wrote:
> > > > > On Mon, Apr 16, 2018 at 12:29:05PM -0700, Dongwon Kim wrote:
> > > 3.2 Backend exports dma-buf to xen-front
> > > 
> > > In this case Dom0 pages are shared with DomU. As before, DomU can only write
> > > to these pages, not any other page from Dom0, so it can be still considered
> > > safe.
> > > But, the following must be considered (highlighted in xen-front's Kernel
> > > documentation):
> > >   - If guest domain dies then pages/grants received from the backend cannot
> > >     be claimed back - think of it as memory lost to Dom0 (won't be used for
> > > any
> > >     other guest)
> > >   - Misbehaving guest may send too many requests to the backend exhausting
> > >     its grant references and memory (consider this from security POV). As the
> > >     backend runs in the trusted domain we also assume that it is trusted as
> > > well,
> > >     e.g. must take measures to prevent DDoS attacks.
> > I cannot parse the above sentence:
> > 
> > "As the backend runs in the trusted domain we also assume that it is
> > trusted as well, e.g. must take measures to prevent DDoS attacks."
> > 
> > What's the relation between being trusted and protecting from DoS
> > attacks?
> I mean that we trust the backend that it can prevent Dom0
> from crashing in case DomU's frontend misbehaves, e.g.
> if the frontend sends too many memory requests etc.
> > In any case, all? PV protocols are implemented with the frontend
> > sharing pages to the backend, and I think there's a reason why this
> > model is used, and it should continue to be used.
> This is the first use-case above. But there are real-world
> use-cases (embedded in my case) when physically contiguous memory
> needs to be shared, one of the possible ways to achieve this is
> to share contiguous memory from Dom0 to DomU (the second use-case above)
> > Having to add logic in the backend to prevent such attacks means
> > that:
> > 
> >   - We need more code in the backend, which increases complexity and
> >     chances of bugs.
> >   - Such code/logic could be wrong, thus allowing DoS.
> You can live without this code at all, but this is then up to
> backend which may make Dom0 down because of DomU's frontend doing evil
> things

IMO we should design protocols that do not allow such attacks instead
of having to defend against them.

> > > 4. xen-front/backend/xen-zcopy synchronization
> > > 
> > > 4.1. As I already said in 2) all the inter VM communication happens between
> > > xen-front and the backend, xen-zcopy is NOT involved in that.
> > > When xen-front wants to destroy a display buffer (dumb/dma-buf) it issues a
> > > XENDISPL_OP_DBUF_DESTROY command (opposite to XENDISPL_OP_DBUF_CREATE).
> > > This call is synchronous, so xen-front expects that backend does free the
> > > buffer pages on return.
> > > 
> > > 4.2. Backend, on XENDISPL_OP_DBUF_DESTROY:
> > >    - closes all dumb handles/fd's of the buffer according to [3]
> > >    - issues DRM_IOCTL_XEN_ZCOPY_DUMB_WAIT_FREE IOCTL to xen-zcopy to make
> > > sure
> > >      the buffer is freed (think of it as it waits for dma-buf->release
> > > callback)
> > So this zcopy thing keeps some kind of track of the memory usage? Why
> > can't the user-space backend keep track of the buffer usage?
> Because there is no dma-buf UAPI which allows to track the buffer life cycle
> (e.g. wait until dma-buf's .release callback is called)
> > >    - replies to xen-front that the buffer can be destroyed.
> > > This way deletion of the buffer happens synchronously on both Dom0 and DomU
> > > sides. In case if DRM_IOCTL_XEN_ZCOPY_DUMB_WAIT_FREE returns with time-out
> > > error
> > > (BTW, wait time is a parameter of this IOCTL), Xen will defer grant
> > > reference
> > > removal and will retry later until those are free.
> > > 
> > > Hope this helps understand how buffers are synchronously deleted in case
> > > of xen-zcopy with a single protocol command.
> > > 
> > > I think the above logic can also be re-used by the hyper-dmabuf driver with
> > > some additional work:
> > > 
> > > 1. xen-zcopy can be split into 2 parts and extend:
> > > 1.1. Xen gntdev driver [4], [5] to allow creating dma-buf from grefs and
> > > vise versa,
> > I don't know much about the dma-buf implementation in Linux, but
> > gntdev is a user-space device, and AFAICT user-space applications
> > don't have any notion of dma buffers. How are such buffers useful for
> > user-space? Why can't this just be called memory?
> A dma-buf is seen by user-space as a file descriptor and you can
> pass it to different drivers then. For example, you can share a buffer
> used by a display driver for scanout with a GPU, to compose a picture
> into it:
> 1. User-space (US) allocates a display buffer from display driver
> 2. US asks display driver to export the dma-buf which backs up that buffer,
> US gets buffer's fd: dma_buf_fd
> 3. US asks GPU driver to import a buffer and provides it with dma_buf_fd
> 4. GPU renders contents into display buffer (dma_buf_fd)

After speaking with Oleksandr on IRC, I think the main usage of the
gntdev extension is to:

1. Create a dma-buf from a set of grant references.
2. Share dma-buf and get a list of grant references.

I think this set of operations could be broken into:

1.1 Map grant references into user-space using the gntdev.
1.2 Create a dma-buf out of a set of user-space virtual addresses.

2.1 Map a dma-buf into user-space.
2.2 Get grefs out of the user-space addresses where the dma-buf is
    mapped.

So it seems like what's actually missing is a way to:

 - Create a dma-buf from a list of user-space virtual addresses.
 - Allow to map a dma-buf into user-space, so it can then be used with
   the gntdev.

I think this is generic enough that it could be implemented by a
device not tied to Xen. AFAICT the hyper_dma guys also wanted
something similar to this.

> Finally, this is indeed some memory, but a bit more [1]
> > 
> > Also, (with my FreeBSD maintainer hat) how is this going to translate
> > to other OSes? So far the operations performed by the gntdev device
> > are mostly OS-agnostic because this just map/unmap memory, and in fact
> > they are implemented by Linux and FreeBSD.
> At the moment I can only see Linux implementation and it seems
> to be perfectly ok as we do not change Xen's APIs etc. and only
> use the existing ones (remember, we only extend gntdev/balloon
> drivers, all the changes in the Linux kernel)
> As the second note I can also think that we do not extend gntdev/balloon
> drivers and have re-worked xen-zcopy driver be a separate entity,
> say drivers/xen/dma-buf
> > > implement "wait" ioctl (wait for dma-buf->release): currently these are
> > > DRM_XEN_ZCOPY_DUMB_FROM_REFS, DRM_XEN_ZCOPY_DUMB_TO_REFS and
> > > DRM_XEN_ZCOPY_DUMB_WAIT_FREE
> > > 1.2. Xen balloon driver [6] to allow allocating contiguous buffers (not
> > > needed
> > > by current hyper-dmabuf, but is a must for xen-zcopy use-cases)
> > I think this needs clarifying. In which memory space do you need those
> > regions to be contiguous?
> Use-case: Dom0 has a HW driver which only works with contig memory
> and I want DomU to be able to directly write into that memory, thus
> implementing zero copying
> > 
> > Do they need to be contiguous in host physical memory, or guest
> > physical memory?
> Host
> > 
> > If it's in guest memory space, isn't there any generic interface that
> > you can use?
> > 
> > If it's in host physical memory space, why do you need this buffer to
> > be contiguous in host physical memory space? The IOMMU should hide all
> > this.
> There are drivers/HW which can only work with contig memory and
> if it is backed by an IOMMU then still it has to be contig in IPA
> space (real device doesn't know that it is actually IPA contig, not PA)

What's IPA contig?

Thanks, Roger.