On 10/17/18 7:54 PM, Alexander Duyck wrote:
> This change makes it so that we use the same approach that was already in
> use on Sparc on all the archtectures that support a 64b long.
>
> This is mostly motivated by the fact that 7 to 10 store/move instructions
> are likely always going to be faster than having to call into a function
> that is not specialized for handling page init.
>
> An added advantage to doing it this way is that the compiler can get away
> with combining writes in the __init_single_page call. As a result the
> memset call will be reduced to only about 4 write operations, or at least
> that is what I am seeing with GCC 6.2 as the flags, LRU poitners, and
> count/mapcount seem to be cancelling out at least 4 of the 8 assignments on
> my system.
>
> One change I had to make to the function was to reduce the minimum page
> size to 56 to support some powerpc64 configurations.
>
> This change should introduce no change on SPARC since it already had this
> code. In the case of x86_64 I saw a reduction from 3.75s to 2.80s when
> initializing 384GB of RAM per node. Pavel Tatashin tested on a system with
> Broadcom's Stingray CPU and 48GB of RAM and found that __init_single_page()
> takes 19.30ns / 64-byte struct page before this patch and with this patch
> it takes 17.33ns / 64-byte struct page. Mike Rapoport ran a similar test on
> a OpenPower (S812LC 8348-21C) with Power8 processor and 128GB or RAM. His
> results per 64-byte struct page were 4.68ns before, and 4.59ns after this
> patch.
Reviewed-by: Pavel Tatashin <pavel.tatashin@xxxxxxxxxxxxx>
>
> Signed-off-by: Alexander Duyck <alexander.h.duyck@xxxxxxxxxxxxxxx>
> ---
> arch/sparc/include/asm/pgtable_64.h | 30 --------------------------
> include/linux/mm.h | 41 ++++++++++++++++++++++++++++++++---
> 2 files changed, 38 insertions(+), 33 deletions(-)
>
> diff --git a/arch/sparc/include/asm/pgtable_64.h b/arch/sparc/include/asm/pgtable_64.h
> index 1393a8ac596b..22500c3be7a9 100644
> --- a/arch/sparc/include/asm/pgtable_64.h
> +++ b/arch/sparc/include/asm/pgtable_64.h
> @@ -231,36 +231,6 @@
> extern struct page *mem_map_zero;
> #define ZERO_PAGE(vaddr) (mem_map_zero)
>
> -/* This macro must be updated when the size of struct page grows above 80
> - * or reduces below 64.
> - * The idea that compiler optimizes out switch() statement, and only
> - * leaves clrx instructions
> - */
> -#define mm_zero_struct_page(pp) do { \
> - unsigned long *_pp = (void *)(pp); \
> - \
> - /* Check that struct page is either 64, 72, or 80 bytes */ \
> - BUILD_BUG_ON(sizeof(struct page) & 7); \
> - BUILD_BUG_ON(sizeof(struct page) < 64); \
> - BUILD_BUG_ON(sizeof(struct page) > 80); \
> - \
> - switch (sizeof(struct page)) { \
> - case 80: \
> - _pp[9] = 0; /* fallthrough */ \
> - case 72: \
> - _pp[8] = 0; /* fallthrough */ \
> - default: \
> - _pp[7] = 0; \
> - _pp[6] = 0; \
> - _pp[5] = 0; \
> - _pp[4] = 0; \
> - _pp[3] = 0; \
> - _pp[2] = 0; \
> - _pp[1] = 0; \
> - _pp[0] = 0; \
> - } \
> -} while (0)
> -
> /* PFNs are real physical page numbers. However, mem_map only begins to record
> * per-page information starting at pfn_base. This is to handle systems where
> * the first physical page in the machine is at some huge physical address,
> diff --git a/include/linux/mm.h b/include/linux/mm.h
> index fcf9cc9d535f..6e2c9631af05 100644
> --- a/include/linux/mm.h
> +++ b/include/linux/mm.h
> @@ -98,10 +98,45 @@ static inline void set_max_mapnr(unsigned long limit) { }
>
> /*
> * On some architectures it is expensive to call memset() for small sizes.
> - * Those architectures should provide their own implementation of "struct page"
> - * zeroing by defining this macro in <asm/pgtable.h>.
> + * If an architecture decides to implement their own version of
> + * mm_zero_struct_page they should wrap the defines below in a #ifndef and
> + * define their own version of this macro in <asm/pgtable.h>
> */
> -#ifndef mm_zero_struct_page
> +#if BITS_PER_LONG == 64
> +/* This function must be updated when the size of struct page grows above 80
> + * or reduces below 56. The idea that compiler optimizes out switch()
> + * statement, and only leaves move/store instructions. Also the compiler can
> + * combine write statments if they are both assignments and can be reordered,
> + * this can result in several of the writes here being dropped.
> + */
> +#define mm_zero_struct_page(pp) __mm_zero_struct_page(pp)
> +static inline void __mm_zero_struct_page(struct page *page)
> +{
> + unsigned long *_pp = (void *)page;
> +
> + /* Check that struct page is either 56, 64, 72, or 80 bytes */
> + BUILD_BUG_ON(sizeof(struct page) & 7);
> + BUILD_BUG_ON(sizeof(struct page) < 56);
> + BUILD_BUG_ON(sizeof(struct page) > 80);
> +
> + switch (sizeof(struct page)) {
> + case 80:
> + _pp[9] = 0; /* fallthrough */
> + case 72:
> + _pp[8] = 0; /* fallthrough */
> + case 64:
> + _pp[7] = 0; /* fallthrough */
> + case 56:
> + _pp[6] = 0;
> + _pp[5] = 0;
> + _pp[4] = 0;
> + _pp[3] = 0;
> + _pp[2] = 0;
> + _pp[1] = 0;
> + _pp[0] = 0;
> + }
> +}
> +#else
> #define mm_zero_struct_page(pp) ((void)memset((pp), 0, sizeof(struct page)))
> #endif
>
>
