On Fri, 7 Apr 2023 07:55:48 +0200
Christoph Hellwig <hch@xxxxxx> wrote:
> On Tue, Mar 28, 2023 at 09:54:35AM +0200, Petr Tesarik wrote:
> > I tend to agree here. However, it's the DMABUF design itself that causes
> > some trouble. The buffer is allocated by the v3d driver, which does not
> > have the restriction, so the DMA API typically allocates an address
> > somewhere near the 4G boundary. Userspace then exports the buffer, sends
> > it to another process as a file descriptor and imports it into the vc4
> > driver, which requires DMA below 1G. In the beginning, v3d had no idea
> > that the buffer would be exported to userspace, much less that it would
> > be later imported into vc4.
>
> Then we need to either:
>
> a) figure out a way to communicate these addressing limitations
AFAICS this would require a complete overhaul of the dma-buf userspace
API so that intended imports are communicated at export time. In other
words, it would be quite intrusive. Not my preferrence.
> b) find a way to migrate a buffer into other memory, similar to
> how page migration works for page cache
Let me express the idea in my own words to make sure I get it right.
When a DMA buffer is imported, but before it is ultimately pinned in
memory, the importing device driver checks whether the buffer meets its
DMA constraints. If not, it calls a function provided by the exporting
device driver to migrate the buffer. This makes sense, but:
1) The operation must be implemented in the exporting driver; this
will take some time.
2) In theory, there may be no overlap between the exporting device
and the importing device. OTOH I'm not aware of any real-world
example, so we can probably return a suitable error code, and
that's it.
Anyway, I have already written in another reply that my original use
case is moot, because a more recent distribution can do the job without
using dma-buf, so it has been fixed in user space, be it in GNOME,
pipewire, or Mesa (I don't really have to know).
At this point I would go with the assumption that large buffers
allocated by media subsystems will not hit swiotlb. Consequently, I
don't plan to spend more time on this branch of the story.
> > BTW my testing also suggests that the streaming DMA API is quite
> > inefficient, because UAS performance _improved_ with swiotlb=force.
> > Sure, this should probably be addressed in the UAS and/or xHCI driver,
> > but what I mean is that moving away from swiotlb may even cause
> > performance regressions, which is counter-intuitive. At least I would
> > _not_ have expected it.
>
> That is indeed very odd. Are you running with a very slow iommu
> driver there? Or what is the actual use case there in general?
This was on a Raspberry Pi 4, which does not have any IOMMU. IOW it
looks like copying data around can be faster than sending it straight
to the device. When I have some more time, I must investigate what is
really happening there, because it does not make any sense to me.
Petr T
