Re: [RFC PATCH v2 18/20] powerpc/64s: Use contiguous PMD/PUD instead of HUGEPD

["Nicholas Piggin" <npiggin@gmail.com>]

On Sat May 18, 2024 at 5:00 AM AEST, Christophe Leroy wrote:
> On book3s/64, the only user of hugepd is hash in 4k mode.
>
> All other setups (hash-64, radix-4, radix-64) use leaf PMD/PUD.
>
> Rework hash-4k to use contiguous PMD and PUD instead.
>
> In that setup there are only two huge page sizes: 16M and 16G.
>
> 16M sits at PMD level and 16G at PUD level.
>
> pte_update doesn't know page size, lets use the same trick as
> hpte_need_flush() to get page size from segment properties. That's
> not the most efficient way but let's do that until callers of
> pte_update() provide page size instead of just a huge flag.
>
> Signed-off-by: Christophe Leroy <christophe.leroy@csgroup.eu>
> ---
>  arch/powerpc/include/asm/book3s/64/hash-4k.h  | 15 --------
>  arch/powerpc/include/asm/book3s/64/hash.h     | 38 +++++++++++++++----
>  arch/powerpc/include/asm/book3s/64/hugetlb.h  | 38 -------------------
>  .../include/asm/book3s/64/pgtable-4k.h        | 34 -----------------
>  .../include/asm/book3s/64/pgtable-64k.h       | 20 ----------
>  arch/powerpc/include/asm/hugetlb.h            |  4 ++
>  .../include/asm/nohash/32/hugetlb-8xx.h       |  4 --
>  .../powerpc/include/asm/nohash/hugetlb-e500.h |  4 --
>  arch/powerpc/include/asm/page.h               |  8 ----
>  arch/powerpc/mm/book3s64/hash_utils.c         | 11 ++++--
>  arch/powerpc/mm/book3s64/pgtable.c            | 12 ------
>  arch/powerpc/mm/hugetlbpage.c                 | 19 ----------
>  arch/powerpc/mm/pgtable.c                     |  2 +-
>  arch/powerpc/platforms/Kconfig.cputype        |  1 -
>  14 files changed, 43 insertions(+), 167 deletions(-)
>
> diff --git a/arch/powerpc/include/asm/book3s/64/hash-4k.h b/arch/powerpc/include/asm/book3s/64/hash-4k.h
> index 6472b08fa1b0..c654c376ef8b 100644
> --- a/arch/powerpc/include/asm/book3s/64/hash-4k.h
> +++ b/arch/powerpc/include/asm/book3s/64/hash-4k.h
> @@ -74,21 +74,6 @@
>  #define remap_4k_pfn(vma, addr, pfn, prot)	\
>  	remap_pfn_range((vma), (addr), (pfn), PAGE_SIZE, (prot))
>  
> -#ifdef CONFIG_HUGETLB_PAGE
> -static inline int hash__hugepd_ok(hugepd_t hpd)
> -{
> -	unsigned long hpdval = hpd_val(hpd);
> -	/*
> -	 * if it is not a pte and have hugepd shift mask
> -	 * set, then it is a hugepd directory pointer
> -	 */
> -	if (!(hpdval & _PAGE_PTE) && (hpdval & _PAGE_PRESENT) &&
> -	    ((hpdval & HUGEPD_SHIFT_MASK) != 0))
> -		return true;
> -	return false;
> -}
> -#endif
> -
>  /*
>   * 4K PTE format is different from 64K PTE format. Saving the hash_slot is just
>   * a matter of returning the PTE bits that need to be modified. On 64K PTE,
> diff --git a/arch/powerpc/include/asm/book3s/64/hash.h b/arch/powerpc/include/asm/book3s/64/hash.h
> index faf3e3b4e4b2..509811ca7695 100644
> --- a/arch/powerpc/include/asm/book3s/64/hash.h
> +++ b/arch/powerpc/include/asm/book3s/64/hash.h
> @@ -4,6 +4,7 @@
>  #ifdef __KERNEL__
>  
>  #include <asm/asm-const.h>
> +#include <asm/book3s/64/slice.h>
>  
>  /*
>   * Common bits between 4K and 64K pages in a linux-style PTE.
> @@ -161,14 +162,10 @@ extern void hpte_need_flush(struct mm_struct *mm, unsigned long addr,
>  			    pte_t *ptep, unsigned long pte, int huge);
>  unsigned long htab_convert_pte_flags(unsigned long pteflags, unsigned long flags);
>  /* Atomic PTE updates */
> -static inline unsigned long hash__pte_update(struct mm_struct *mm,
> -					 unsigned long addr,
> -					 pte_t *ptep, unsigned long clr,
> -					 unsigned long set,
> -					 int huge)
> +static inline unsigned long hash__pte_update_one(pte_t *ptep, unsigned long clr,
> +						 unsigned long set)
>  {
>  	__be64 old_be, tmp_be;
> -	unsigned long old;
>  
>  	__asm__ __volatile__(
>  	"1:	ldarx	%0,0,%3		# pte_update\n\
> @@ -182,11 +179,38 @@ static inline unsigned long hash__pte_update(struct mm_struct *mm,
>  	: "r" (ptep), "r" (cpu_to_be64(clr)), "m" (*ptep),
>  	  "r" (cpu_to_be64(H_PAGE_BUSY)), "r" (cpu_to_be64(set))
>  	: "cc" );
> +
> +	return be64_to_cpu(old_be);
> +}
> +
> +static inline unsigned long hash__pte_update(struct mm_struct *mm,
> +					 unsigned long addr,
> +					 pte_t *ptep, unsigned long clr,
> +					 unsigned long set,
> +					 int huge)
> +{
> +	unsigned long old;
> +
> +	old = hash__pte_update_one(ptep, clr, set);
> +
> +	if (huge && IS_ENABLED(CONFIG_PPC_4K_PAGES)) {
> +		unsigned int psize = get_slice_psize(mm, addr);
> +		int nb, i;
> +
> +		if (psize == MMU_PAGE_16M)
> +			nb = SZ_16M / PMD_SIZE;
> +		else if (psize == MMU_PAGE_16G)
> +			nb = SZ_16G / PUD_SIZE;
> +		else
> +			nb = 1;
> +
> +		for (i = 1; i < nb; i++)
> +			hash__pte_update_one(ptep + i, clr, set);
> +	}
>  	/* huge pages use the old page table lock */
>  	if (!huge)
>  		assert_pte_locked(mm, addr);
>  
> -	old = be64_to_cpu(old_be);
>  	if (old & H_PAGE_HASHPTE)
>  		hpte_need_flush(mm, addr, ptep, old, huge);
>  

Nice series, I don't know this hugepd code very well but I'll try.
Why do you have to replicate the PTE entry here? The hash table refill
should always be working on the first PTE of the page otherwise we have
bigger problems.

What paths look at the N > 0 PTEs of a contiguous page entry?

Thanks,
Nick