From mboxrd@z Thu Jan  1 00:00:00 1970
From: will.deacon@arm.com (Will Deacon)
Date: Tue, 11 Jul 2017 18:52:57 +0100
Subject: prefetch inconsistency in copy_page()
In-Reply-To: <CAKv+Gu9hGwLJxBCMCTK2fsK31SgX7E=qc3PQcNXZA6OP2uUs8Q@mail.gmail.com>
References: <CAKv+Gu9hGwLJxBCMCTK2fsK31SgX7E=qc3PQcNXZA6OP2uUs8Q@mail.gmail.com>
Message-ID: <20170711175257.GF13977@arm.com>
To: linux-arm-kernel@lists.infradead.org
List-Id: linux-arm-kernel.lists.infradead.org

On Tue, Jul 11, 2017 at 10:42:19AM +0100, Ard Biesheuvel wrote:
> Hi all,
> 
> No big deal, but I spotted an inconsistency in how copy_page() does
> its prefetching. With the code as-is, it prefetches 2 lines at the
> start of the function, but it is off by one line in the loop, i.e., it
> prefetches 3 lines ahead (and skips a line at entry). So imo, we need
> either
> 
> """
> @@ -30,9 +30,10 @@
>   */
>  ENTRY(copy_page)
>  alternative_if ARM64_HAS_NO_HW_PREFETCH
> -       # Prefetch two cache lines ahead.
> +       # Prefetch three cache lines ahead.
>         prfm    pldl1strm, [x1, #128]
>         prfm    pldl1strm, [x1, #256]
> +       prfm    pldl1strm, [x1, #384]
>  alternative_else_nop_endif
> 
>         ldp     x2, x3, [x1]
> """
> 
> or
> 
> """
> @@ -50,7 +51,7 @@ alternative_else_nop_endif
>         subs    x18, x18, #128
> 
>  alternative_if ARM64_HAS_NO_HW_PREFETCH
> -       prfm    pldl1strm, [x1, #384]
> +       prfm    pldl1strm, [x1, #256]
>  alternative_else_nop_endif
> 
>         stnp    x2, x3, [x0]
> 
> """
> 
> to make things consistent again.
> 
> Thoughts?

Either of those look good to me. Andrew -- any preference? If not, I'll take
the second one.

Will