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

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

On Tue May 21, 2024 at 2:43 AM AEST, Christophe Leroy wrote:
>
>
> Le 20/05/2024 à 14:54, Nicholas Piggin a écrit :
> > 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.
>
> I don't know how book3s/64 works exactly, but on nohash, when you get a 
> TLB miss exception the only thing you have is the address and you don't 
> know yes it is a hugepage so you get the PTE as if it was a 4k page and 
> it is only when you read that PTE that you know it is a hugepage.
>
> Ok, on book3s/64 the page size seems to be encoded inside the segment so 
> maybe it is a bit different but anyway the TLB miss exception (or DSI ?) 
> can happen at any address.

Right.

If you think of the hash page table as a software loaded TLB (which
is how Linux kind of thinks of it), then DSI is a TLB miss. hash_page_x
calls find the Linux pte and load that translation into hash page table.

One of the hard parts is keeping them coherent with low overhead. This
requires pte bits H_PAGE_BUSY as a lock and H_PAGE_HASHPTE which means
it might be in the hash table. So Linux PTE and hash PTE have to be
1:1 in general.

There are probably cases where we could get away from 1:1, but I would
much prefer not to. Maybe read-only access would be okay though. But
the hash_page will have to always operate on the 0th pte, which I think
we get via segment size masking, same for any set / update / clear of
the pte.

> > 
> > What paths look at the N > 0 PTEs of a contiguous page entry?
> > 
>
> pte_offset_kernel() or pte_offset_map_lock() will land on any contiguous 
> PTE based on the address handed to pte_index(), as if it was a standard 
> (4k or 64k) page.
>
> pte_index() doesn't know it is a hugepage, that's the reason why we need 
> to duplicate the entry.

From the mm/ side of things, hugetlb page tables are always walked via
the huge vma which knows the page size and could align address... I
guess except for fast gup? Which should be read-only. So okay you do
need to replicate huge ptes for fast gup at least. Any others?

There's going to need to be a little more to it. __hash_page_huge sets
PTE accessed and dirty for example, so if we allow any PTE readers to
check the non-0th pte we would have to do something about that.

How do you deal with dirty/accessed bits for other subarchs?

We could just remove the hash_page setting of those bits and just cause
a fault and require Linux mm to set them. At least for hugepages we
could do that probably without any real performance worry.

Thanks,
Nick