Re: [PATCH 4/6] mm: add a gup_fixup_start_addr hook
From: Catalin Marinas
Date: Wed May 29 2019 - 04:23:35 EST
Hi Christoph,
On Sat, 25 May 2019 at 14:33, Christoph Hellwig <hch@xxxxxx> wrote:
> diff --git a/mm/gup.c b/mm/gup.c
> index f173fcbaf1b2..1c21ecfbf38b 100644
> --- a/mm/gup.c
> +++ b/mm/gup.c
> @@ -2117,6 +2117,10 @@ static void gup_pgd_range(unsigned long addr, unsigned long end,
> } while (pgdp++, addr = next, addr != end);
> }
>
> +#ifndef gup_fixup_start_addr
> +#define gup_fixup_start_addr(start) (start)
> +#endif
As you pointed out in a subsequent reply, we could use the
untagged_addr() macro from Andrey (or a shorter "untag_addr" if you
want it to look like a verb).
> #ifndef gup_fast_permitted
> /*
> * Check if it's allowed to use __get_user_pages_fast() for the range, or
> @@ -2145,7 +2149,7 @@ int __get_user_pages_fast(unsigned long start, int nr_pages, int write,
> unsigned long flags;
> int nr = 0;
>
> - start &= PAGE_MASK;
> + start = gup_fixup_start_addr(start) & PAGE_MASK;
> len = (unsigned long) nr_pages << PAGE_SHIFT;
> end = start + len;
>
> @@ -2218,7 +2222,7 @@ int get_user_pages_fast(unsigned long start, int nr_pages,
> unsigned long addr, len, end;
> int nr = 0, ret = 0;
>
> - start &= PAGE_MASK;
> + start = gup_fixup_start_addr(start) & PAGE_MASK;
> addr = start;
> len = (unsigned long) nr_pages << PAGE_SHIFT;
> end = start + len;
In Andrey's patch [1] we don't fix __get_user_pages_fast(), only
__get_user_pages() as it needs to do a find_vma() search. I wonder
whether this is actually necessary for the *_fast() versions. If the
top byte is non-zero (i.e. tagged address), 'end' would also have the
same tag. The page table macros like pgd_index() and pgd_addr_end()
already take care of masking out the top bits (at least for arm64)
since they need to work on kernel address with the top bits all 1. So
gup_pgd_range() should cope with tagged addresses already.
[1] https://lore.kernel.org/lkml/d234cd71774f35229bdfc0a793c34d6712b73093.1557160186.git.andreyknvl@xxxxxxxxxx/
--
Catalin