Re: [PATCH v3 3/7] dma-buf: Document RDMA non-ODP invalidate_mapping() special case

[Jason Gunthorpe <jgg@ziepe.ca>]

On Wed, Jan 21, 2026 at 03:15:46PM +0100, Christian König wrote:
> > And let's clarify what I said in my other email that this new revoke
> > semantic is not just a signal to maybe someday unmap but a hard
> > barrier that it must be done once the fences complete, similar to
> > non-pinned importers.
> 
> Well, I would avoid that semantics.
>
> Even when the exporter requests the mapping to be invalidated it
> does not mean that the mapping can go away immediately.
> 
> It's fine when accesses initiated after an invalidation and then
> waiting for fences go into nirvana and have undefined results, but
> they should not trigger PCI AER, warnings from the IOMMU or even
> worse end up in some MMIO BAR of a newly attached devices.

So what's the purpose of the fence if accesses can continue after
waiting for fences?

If we always have to wait for the unmap call, is the importer allowed
to call unmap while its own fences are outstanding?

> So if the exporter wants to be 100% sure that nobody is using the
> mapping any more then it needs to wait for the importer to call
> dma_buf_unmap_attachment().

We are trying to introduce this new idea called "revoke".

Revoke means the exporter does some defined sequence and after the end
of that sequence it knows there are no further DMA or CPU accesses to
its memory at all.

It has to happen in bounded time, so it can't get entangled with
waiting for userspace to do something (eg importer unmap via an ioctl)

It has to be an absolute statement because the VFIO and RDMA exporter
use cases can trigger UAFs and AERs if importers keep accessing.

So, what exactly should the export sequence be? We were proposing to
call invalidate_mapping() and when it returns there is no access.

The fence is missing, so now the sequences includes wait for the
fences.

And now you are saying we have to wait for all unmaps? Not only wait
for the unmaps, but the importers now also must call unmap as part of
their invalidate_mapping() callback.. Is that OK? Do existing
importers do that?

If all the above are yes, then lets document explicitly this is the
required sequence and we can try to make it work. Please say, because
we just don't know and keep getting surprised :)

Thanks,
Jason