Re: [PATCH V1 net-next 1/2] pgtable: Add API to query if write combining is available

[Moshe Lazer <moshel@dev.mellanox.co.il>]

On 10/8/2014 7:24 PM, David Miller wrote:
> From: Moshe Lazer <moshel@dev.mellanox.co.il>
> Date: Wed, 08 Oct 2014 11:44:57 +0300
>
>>> #if defined(__i386__) || defined(__x86_64__)
>>> 	if (map->type == _DRM_REGISTERS && !(map->flags & _DRM_WRITE_COMBINING))
>>> 		tmp = pgprot_noncached(tmp);
>>> 	else
>>> 		tmp = pgprot_writecombine(tmp);
>>> #elif defined(__powerpc__)
>>> 	pgprot_val(tmp) |= _PAGE_NO_CACHE;
>>> 	if (map->type == _DRM_REGISTERS)
>>> 		pgprot_val(tmp) |= _PAGE_GUARDED;
>>> #elif defined(__ia64__)
>>> 	if (efi_range_is_wc(vma->vm_start, vma->vm_end -
>>> 				    vma->vm_start))
>>> 		tmp = pgprot_writecombine(tmp);
>>> 	else
>>> 		tmp = pgprot_noncached(tmp);
>>> #elif defined(__sparc__) || defined(__arm__) || defined(__mips__)
>>> 	tmp = pgprot_noncached(tmp);
>>> #endif
>> The idea was to provide an indication as for whether the arch supports
>> write-combining in general.
>> If we want to benefit from blue flame operations, we need to map the
>> blue flame registers as write combining - otherwise there is no
>> benefit. So we would like to know if write combining is supported by
>> the system or not.
>>
> You completely miss my point.  On a given architectuire it might be
> _illegal_ to map certain address ranges as write-combining without
> checks like the ones above that ia64 needs.
>
> Therefore your proposed interface is by definition insufficient.
Thanks David, I'll try to clarify my point.
For me the writecombine_available() is a way to know if the 
pgprot_writecombine() is effective or just cover call to the 
pgprot_noncached().
I want to use the writecombine_available() regardless to the mapping 
address.
For example in mlx4 query_device I want to indicate that blue-flame is 
not supported if  `writecombine_available() ==  false`.
In this case we don't have the mapping address yet.

Later on if an arch has write-combining (writecombine_available() ==  
true) when we try to map the blue-flame registers (in mlx4_ib_mmap):

     vma->vm_page_prot = pgprot_writecombine(vma->vm_page_prot);

     if (io_remap_pfn_range(vma, vma->vm_start,
         to_mucontext(context)->uar.pfn +
         dev->dev->caps.num_uars,
         PAGE_SIZE, vma->vm_page_prot))
             return -EAGAIN;

I can be sure that pgprot_writecombine() is not a cover for 
pgprot_noncached().
The address checks that you mentioned should be part of 
io_remap_pfn_range, this function should fail if the vma->vm_start is 
not compatible to the vma->vm_page_prot.
Please let me know if I misunderstood something.