Framebuffer mmap on PowerPC

Framebuffer mmap on PowerPC

[Thomas Zimmermann]

[-- Attachment #1.1: Type: text/plain, Size: 1185 bytes --]

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?

Best regards
Thomas


[1] 
https://elixir.bootlin.com/linux/v6.5/source/include/asm-generic/fb.h#L19
[2] 
https://elixir.bootlin.com/linux/v6.5/source/arch/mips/include/asm/fb.h#L11
[3] 
https://elixir.bootlin.com/linux/v6.5/source/arch/powerpc/include/asm/fb.h#L12
[4] 
https://elixir.bootlin.com/linux/v6.5/source/drivers/video/fbdev/core/fbmem.c#L1299


-- 
Thomas Zimmermann
Graphics Driver Developer
SUSE Software Solutions Germany GmbH
Frankenstrasse 146, 90461 Nuernberg, Germany
GF: Ivo Totev, Andrew Myers, Andrew McDonald, Boudien Moerman
HRB 36809 (AG Nuernberg)

[-- Attachment #2: OpenPGP digital signature --]
[-- Type: application/pgp-signature, Size: 840 bytes --]

Re: Framebuffer mmap on PowerPC

[Christophe Leroy]

Le 31/08/2023 à 16:41, Thomas Zimmermann a écrit :
> 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?

Looks like phys_mem_access_prot() defaults to calling pgprot_noncached() 
see 
https://elixir.bootlin.com/linux/v6.5/source/arch/powerpc/mm/mem.c#L37 
but for a few platforms that's superseeded by 
pci_phys_mem_access_prot(), see 
https://elixir.bootlin.com/linux/v6.5/source/arch/powerpc/kernel/pci-common.c#L524

However, as far as I can see pci_phys_mem_access_prot() doesn't use file 
so you could likely drop that argument on phys_mem_access_prot() on 
powerpc. But when I for instance look at arm, I see that the file 
argument is used, see 
https://elixir.bootlin.com/linux/v6.5/source/arch/arm/mm/mmu.c#L713

So, the simplest is maybe the following, allthough that's probably worth 
a comment:

diff --git a/arch/powerpc/include/asm/fb.h b/arch/powerpc/include/asm/fb.h
index 5f1a2e5f7654..8b9b856f476e 100644
--- a/arch/powerpc/include/asm/fb.h
+++ b/arch/powerpc/include/asm/fb.h
@@ -6,10 +6,9 @@

  #include <asm/page.h>

-static inline void fb_pgprotect(struct file *file, struct 
vm_area_struct *vma,
-				unsigned long off)
+static inline void fb_pgprotect(struct vm_area_struct *vma, unsigned 
long off)
  {
-	vma->vm_page_prot = phys_mem_access_prot(file, off >> PAGE_SHIFT,
+	vma->vm_page_prot = phys_mem_access_prot(NULL, off >> PAGE_SHIFT,
  						 vma->vm_end - vma->vm_start,
  						 vma->vm_page_prot);
  }


Christophe


> 
> Best regards
> Thomas
> 
> 
> [1] 
> https://elixir.bootlin.com/linux/v6.5/source/include/asm-generic/fb.h#L19
> [2] 
> https://elixir.bootlin.com/linux/v6.5/source/arch/mips/include/asm/fb.h#L11
> [3] 
> https://elixir.bootlin.com/linux/v6.5/source/arch/powerpc/include/asm/fb.h#L12
> [4] 
> https://elixir.bootlin.com/linux/v6.5/source/drivers/video/fbdev/core/fbmem.c#L1299
> 
> 

Re: Framebuffer mmap on PowerPC

[Christophe Leroy]

Le 31/08/2023 à 19:38, Christophe Leroy a écrit :
> 
> 
> Le 31/08/2023 à 16:41, Thomas Zimmermann a écrit :
>> 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?
> 
> Looks like phys_mem_access_prot() defaults to calling pgprot_noncached() 
> see 
> https://elixir.bootlin.com/linux/v6.5/source/arch/powerpc/mm/mem.c#L37 
> but for a few platforms that's superseeded by 
> pci_phys_mem_access_prot(), see 
> https://elixir.bootlin.com/linux/v6.5/source/arch/powerpc/kernel/pci-common.c#L524
> 
> However, as far as I can see pci_phys_mem_access_prot() doesn't use file 
> so you could likely drop that argument on phys_mem_access_prot() on 
> powerpc. But when I for instance look at arm, I see that the file 
> argument is used, see 
> https://elixir.bootlin.com/linux/v6.5/source/arch/arm/mm/mmu.c#L713
> 
> So, the simplest is maybe the following, allthough that's probably worth 
> a comment:
> 
> diff --git a/arch/powerpc/include/asm/fb.h b/arch/powerpc/include/asm/fb.h
> index 5f1a2e5f7654..8b9b856f476e 100644
> --- a/arch/powerpc/include/asm/fb.h
> +++ b/arch/powerpc/include/asm/fb.h
> @@ -6,10 +6,9 @@
> 
>   #include <asm/page.h>
> 
> -static inline void fb_pgprotect(struct file *file, struct 
> vm_area_struct *vma,
> -                unsigned long off)
> +static inline void fb_pgprotect(struct vm_area_struct *vma, unsigned 
> long off)
>   {
> -    vma->vm_page_prot = phys_mem_access_prot(file, off >> PAGE_SHIFT,
> +    vma->vm_page_prot = phys_mem_access_prot(NULL, off >> PAGE_SHIFT,
>                            vma->vm_end - vma->vm_start,
>                            vma->vm_page_prot);
>   }

And while at it, maybe also replace off >> PAGE_SHIFT by PHYS_PFN(off)

Christophe

> 
> 
> Christophe
> 
> 
>>
>> Best regards
>> Thomas
>>
>>
>> [1] 
>> https://elixir.bootlin.com/linux/v6.5/source/include/asm-generic/fb.h#L19
>> [2] 
>> https://elixir.bootlin.com/linux/v6.5/source/arch/mips/include/asm/fb.h#L11
>> [3] 
>> https://elixir.bootlin.com/linux/v6.5/source/arch/powerpc/include/asm/fb.h#L12
>> [4] 
>> https://elixir.bootlin.com/linux/v6.5/source/drivers/video/fbdev/core/fbmem.c#L1299
>>
>>

Re: Framebuffer mmap on PowerPC

[Thomas Zimmermann]

[-- Attachment #1.1: Type: text/plain, Size: 3246 bytes --]

Hi

Am 31.08.23 um 19:38 schrieb Christophe Leroy:
> 
> 
> Le 31/08/2023 à 16:41, Thomas Zimmermann a écrit :
>> 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?
> 
> Looks like phys_mem_access_prot() defaults to calling pgprot_noncached()
> see
> https://elixir.bootlin.com/linux/v6.5/source/arch/powerpc/mm/mem.c#L37
> but for a few platforms that's superseeded by
> pci_phys_mem_access_prot(), see
> https://elixir.bootlin.com/linux/v6.5/source/arch/powerpc/kernel/pci-common.c#L524
> 
> However, as far as I can see pci_phys_mem_access_prot() doesn't use file
> so you could likely drop that argument on phys_mem_access_prot() on
> powerpc. But when I for instance look at arm, I see that the file
> argument is used, see
> https://elixir.bootlin.com/linux/v6.5/source/arch/arm/mm/mmu.c#L713

Right, I've seen these various implementations. Luckily, the ARM 
framebuffers use a plain pgprot_writecombine() without any references on 
to file.

> 
> So, the simplest is maybe the following, allthough that's probably worth
> a comment:

Could we drop the file argument from PPC's internal functions and 
provide this interface to fb_pgprotect()? phys_mem_access_prot() would 
be a trivial wrapper around that internal API. I'd provide a patch to do 
that.

Best regards
Thomas

> 
> diff --git a/arch/powerpc/include/asm/fb.h b/arch/powerpc/include/asm/fb.h
> index 5f1a2e5f7654..8b9b856f476e 100644
> --- a/arch/powerpc/include/asm/fb.h
> +++ b/arch/powerpc/include/asm/fb.h
> @@ -6,10 +6,9 @@
> 
>    #include <asm/page.h>
> 
> -static inline void fb_pgprotect(struct file *file, struct
> vm_area_struct *vma,
> -				unsigned long off)
> +static inline void fb_pgprotect(struct vm_area_struct *vma, unsigned
> long off)
>    {
> -	vma->vm_page_prot = phys_mem_access_prot(file, off >> PAGE_SHIFT,
> +	vma->vm_page_prot = phys_mem_access_prot(NULL, off >> PAGE_SHIFT,
>    						 vma->vm_end - vma->vm_start,
>    						 vma->vm_page_prot);
>    }
> 
> 
> Christophe
> 
> 
>>
>> Best regards
>> Thomas
>>
>>
>> [1]
>> https://elixir.bootlin.com/linux/v6.5/source/include/asm-generic/fb.h#L19
>> [2]
>> https://elixir.bootlin.com/linux/v6.5/source/arch/mips/include/asm/fb.h#L11
>> [3]
>> https://elixir.bootlin.com/linux/v6.5/source/arch/powerpc/include/asm/fb.h#L12
>> [4]
>> https://elixir.bootlin.com/linux/v6.5/source/drivers/video/fbdev/core/fbmem.c#L1299
>>
>>

-- 
Thomas Zimmermann
Graphics Driver Developer
SUSE Software Solutions Germany GmbH
Frankenstrasse 146, 90461 Nuernberg, Germany
GF: Ivo Totev, Andrew Myers, Andrew McDonald, Boudien Moerman
HRB 36809 (AG Nuernberg)

[-- Attachment #2: OpenPGP digital signature --]
[-- Type: application/pgp-signature, Size: 840 bytes --]

Re: 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