Framebuffer mmap on PowerPC

[Arnd Bergmann]

On Thu, Aug 31, 2023, at 10:41, Thomas Zimmermann wrote:
> Hi,
>
> there's a per-architecture function called fb_pgprotect() that sets 
> VMA's vm_page_prot for mmaped framebuffers. Most architectures use a 
> simple implementation based on pgprot_writecomine() [1] or 
> pgprot_noncached(). [2]
>
> On PPC this function uses phys_mem_access_prot() and therefore requires 
> the mmap call's file struct. [3] Removing the file argument would help 
> with simplifying the caller of fb_pgprotect(). [4]
>
> Why is the file even required on PPC?
>
> Is it possible to replace phys_mem_access_prot() with something simpler 
> that does not use the file struct?

What what I can tell, the structure of the code is a result of
these constraints:

- some powerpc platforms use different page table flags for
  prefetchable vs nonprefetchable BARs on PCI memory.

- page table flags must match between all mappings, in particular
  here between /dev/fb0 and /dev/mem, as mismatched attributes
  cause a checkstop. On other architectures this may cause
  undefined behavior instead of a checkstop

It's unfortunate that we have multiple incompatible ways
to determine the page flags based on firmware (ia64),
pci (powerpc) or file->f_flags (arm, csky), when they all
try to solve the same problem here.

Christophe's suggested approach to simplify it is probably
fine, another way would be to pass the f_flags value instead
of the file pointer.

      Arnd