Re: [RFC][PATCH 03/11] x86/mm: Page size aware flush_tlb_mm_range()

[Dave Hansen <dave.hansen@linux.intel.com>]

> +static inline void tlb_flush(struct mmu_gather *tlb)
> +{
> +	unsigned long start = 0UL, end = TLB_FLUSH_ALL;
> +	unsigned int invl_shift = tlb_get_unmap_shift(tlb);

I had to go back and look at

	https://patchwork.kernel.org/patch/10587207/

to figure out what was going on.  I wonder if we could make the code a
bit more standalone.

This at least needs a comment about what it's getting from 'tlb'.  Maybe
just:

	/* Find the smallest page size that we unmapped: */

> --- a/arch/x86/include/asm/tlbflush.h
> +++ b/arch/x86/include/asm/tlbflush.h
> @@ -507,23 +507,25 @@ struct flush_tlb_info {
>  	unsigned long		start;
>  	unsigned long		end;
>  	u64			new_tlb_gen;
> +	unsigned int		invl_shift;
>  };

Maybe we really should just call this flush_stride or something.

>  #define local_flush_tlb() __flush_tlb()
>  
>  #define flush_tlb_mm(mm)	flush_tlb_mm_range(mm, 0UL, TLB_FLUSH_ALL, 0UL)
>  
> -#define flush_tlb_range(vma, start, end)	\
> -		flush_tlb_mm_range(vma->vm_mm, start, end, vma->vm_flags)
> +#define flush_tlb_range(vma, start, end)			\
> +		flush_tlb_mm_range((vma)->vm_mm, start, end,	\
> +				(vma)->vm_flags & VM_HUGETLB ? PMD_SHIFT : PAGE_SHIFT)

This is safe.  But, Couldn't this PMD_SHIFT also be PUD_SHIFT for a 1G
hugetlb page?

>  void native_flush_tlb_others(const struct cpumask *cpumask,
> --- a/arch/x86/mm/tlb.c
> +++ b/arch/x86/mm/tlb.c
> @@ -522,12 +522,12 @@ static void flush_tlb_func_common(const
>  	    f->new_tlb_gen == mm_tlb_gen) {
>  		/* Partial flush */
>  		unsigned long addr;
> -		unsigned long nr_pages = (f->end - f->start) >> PAGE_SHIFT;
> +		unsigned long nr_pages = (f->end - f->start) >> f->invl_shift;

We might want to make this nr_invalidations or nr_flushes now so we
don't get it confused with PAGE_SIZE stuff.

Otherwise, this makes me a *tiny* bit nervous.  I think we're good about
ensuring that we fully flush 4k mappings from the TLB before going up to
a 2MB mapping because of all the errata we've had there over the years.
But, had we left 4k mappings around, the old flushing code would have
cleaned them up for us.

This certainly tightly ties the invalidations to what was in the page
tables.  If that diverged from the TLB at some point, there's certainly
more exposure here.

Looks fun, though. :)