Re: [PATCH] mm, vmalloc: use __GFP_HIGHMEM implicitly

[Vlastimil Babka]

On 03/07/2017 07:57 PM, Matthew Wilcox wrote:
> On Tue, Mar 07, 2017 at 10:28:41AM -0800, Matthew Wilcox wrote:
>> On Tue, Mar 07, 2017 at 03:10:20PM +0100, Michal Hocko wrote:
>>> This patch simply uses __GFP_HIGHMEM implicitly when allocating pages to
>>> be mapped to the vmalloc space. Current users which add __GFP_HIGHMEM
>>> are simplified and drop the flag.
> 
> btw, I had another idea for GFP_HIGHMEM -- remove it when CONFIG_HIGHMEM
> isn't enabled.  Saves 26 bytes of .text and 64 bytes of .data on my
> laptop's kernel build.  What do you think?
> 
> Also, I suspect the layout of bits is suboptimal from an assembly
> language perspective.  I still mostly care about x86 which doesn't
> benefit, so I'm not inclined to do the work, but certainly ARM, PA-RISC,
> SPARC and Itanium would all benefit from having frequently-used bits
> (ie those used in GFP_KERNEL and GFP_ATOMIC) placed in the low 8 bits.
> 
> diff --git a/include/linux/gfp.h b/include/linux/gfp.h
> index 0fe0b6295ab5..d88cb532d7c8 100644
> --- a/include/linux/gfp.h
> +++ b/include/linux/gfp.h
> @@ -16,7 +16,11 @@ struct vm_area_struct;
>  
>  /* Plain integer GFP bitmasks. Do not use this directly. */
>  #define ___GFP_DMA		0x01u
> +#ifdef CONFIG_HIGHMEM
>  #define ___GFP_HIGHMEM		0x02u
> +#else
> +#define ___GFP_HIGHMEM		0x0u

Make sure you don't break the users of __def_gfpflag_names e.g.
format_flags(). IIRC zero is a terminator in the table.

But the savings don't seem to be worth the trouble.

> +#endif
>  #define ___GFP_DMA32		0x04u
>  #define ___GFP_MOVABLE		0x08u
>  #define ___GFP_RECLAIMABLE	0x10u
>