RE: [PATCH 1/3] mm: Export vmf_insert_mixed_mkwrite()

[Michael Kelley]

From: Christoph Hellwig <hch@xxxxxx> Sent: Thursday, April 10, 2025 12:42 AM
> 
> On Wed, Apr 09, 2025 at 02:10:26PM +0000, Michael Kelley wrote:
> > Hmmm. What's the reference to "as told last time"? I don't think I've had
> > this conversation before.
> 
> Hmm, there was a conversation about deferred I/O, and I remember the
> drm folks even defending their abuse of vmalloc_to_page on dma coherent
> memory against the documentation in the most silly way.  Maybe that was
> a different discussion of the same thing.

Yes, must have been a different discussion.

Turns out the hyperv_fb driver in a different configuration *is* using
dma_alloc_coherent() to get framebuffer memory, which is then managed
by deferred I/O. dma_alloc_coherent() is used as a wrapper around
cma_alloc() but only when the framebuffer size is > 4 MiB. The deferred
I/O code doesn't do vmalloc_to_page() since the CPU address from
dma_alloc_coherent() isn't a vmalloc addr.

That combo works OK, so wasn't something to be fixed in this patch set.
I'll see about a separate patch to just use cma_alloc() directly for that
case, though cma_alloc() and cma_release() would need to be EXPORTed.

> 
> >
> > For the hyperv_fb driver, the memory in question is allocated with a direct call
> > to alloc_pages(), not via dma_alloc_coherent(). There's no DMA in this scenario.
> > The memory is shared with the Hyper-V host and designated as the memory
> > for the virtual framebuffer device. It is then mapped into user space using the
> > mmap() system call against /dev/fb0. User space writes to the memory are
> > eventually (and I omit the details) picked up by the Hyper-V host and displayed.
> 
> Oh, great.
> 
> > Is your point that memory dma_alloc_coherent() memory must be treated as
> > a black box, and can't be deconstructed into individual pages? If so, that makes
> > sense to me.
> 
> Yes.

I'll add some comments to the fbdev deferred I/O code saying not to use it
with a framebuffer allocated with dma_alloc_coherent().

Thanks,

Michael

> 
> > But must the same treatment be applied to memory from
> > alloc_pages()? This is where I need some education.
> 
> No, that's just fine.