Re: [RFC V2 1/2] arm64/mm: Change THP helpers per generic memory semantics

[Catalin Marinas]

On Mon, Jul 06, 2020 at 09:27:04AM +0530, Anshuman Khandual wrote:
> On 07/02/2020 05:41 PM, Catalin Marinas wrote:
> > On Mon, Jun 15, 2020 at 06:45:17PM +0530, Anshuman Khandual wrote:
> >> --- a/arch/arm64/include/asm/pgtable.h
> >> +++ b/arch/arm64/include/asm/pgtable.h
> >> @@ -353,15 +353,92 @@ static inline int pmd_protnone(pmd_t pmd)
> >>  }
> >>  #endif
> >>  
> >> +#define pmd_table(pmd)	((pmd_val(pmd) & PMD_TYPE_MASK) ==  PMD_TYPE_TABLE)
> >> +#define pmd_sect(pmd)	((pmd_val(pmd) & PMD_TYPE_MASK) ==  PMD_TYPE_SECT)
> >> +
> >> +#ifdef CONFIG_TRANSPARENT_HUGEPAGE
> >>  /*
> >> - * THP definitions.
> >> + * PMD Level Encoding (THP Enabled)
> >> + *
> >> + * 0b00 - Not valid	Not present	NA
> >> + * 0b10 - Not valid	Present		Huge  (Splitting)
> >> + * 0b01 - Valid		Present		Huge  (Mapped)
> >> + * 0b11 - Valid		Present		Table (Mapped)
> >>   */
> > 
> > I wonder whether it would be easier to read if we add a dedicated
> > PMD_SPLITTING bit, only when bit 0 is cleared. This bit can be high (say
> > 59), it doesn't really matter as the entry is not valid.
> 
> Could make (PMD[0b00] = 0b10) be represented as PMD_SPLITTING just for
> better reading purpose. But if possible, IMHO it is efficient and less
> vulnerable to use HW defined PTE attribute bit positions including SW
> usable ones than the reserved bits, for a PMD state representation.
> 
> Earlier proposal used PTE_SPECIAL (bit 56) instead. Using PMD_TABLE_BIT
> helps save bit 56 for later. Thinking about it again, would not these
> unused higher bits [59..63] create any problem ? For example while
> enabling THP swapping without split via ARCH_WANTS_THP_SWAP or something
> else later when these higher bits might be required. I am not sure, just
> speculating.

The swap encoding goes to bit 57, so going higher shouldn't be an issue.

> But, do you see any particular problem with PMD_TABLE_BIT ?

No. Only that we have some precedent like PTE_PROT_NONE (bit 58) and
wondering whether we could use a high bit as well here. If we can get
them to overlap, it simplifies this patch further.

> > The only doubt I have is that pmd_mkinvalid() is used in other contexts
> > when it's not necessarily splitting a pmd (search for the
> > pmdp_invalidate() calls). So maybe a better name like PMD_PRESENT with a
> > comment that pmd_to_page() is valid (i.e. no migration or swap entry).
> > Feel free to suggest a better name.
> 
> PMD_INVALID_PRESENT sounds better ?

No strong opinion either way. Yours is clearer.

> >> +static inline pmd_t pmd_mksplitting(pmd_t pmd)
> >> +{
> >> +	unsigned long val = pmd_val(pmd);
> >>  
> >> -#ifdef CONFIG_TRANSPARENT_HUGEPAGE
> >> -#define pmd_trans_huge(pmd)	(pmd_val(pmd) && !(pmd_val(pmd) & PMD_TABLE_BIT))
> >> +	return __pmd((val & ~PMD_TYPE_MASK) | PMD_TABLE_BIT);
> >> +}
> >> +
> >> +static inline pmd_t pmd_clrsplitting(pmd_t pmd)
> >> +{
> >> +	unsigned long val = pmd_val(pmd);
> >> +
> >> +	return __pmd((val & ~PMD_TYPE_MASK) | PMD_TYPE_SECT);
> >> +}
> >> +
> >> +static inline bool pmd_splitting(pmd_t pmd)
> >> +{
> >> +	unsigned long val = pmd_val(pmd);
> >> +
> >> +	if ((val & PMD_TYPE_MASK) == PMD_TABLE_BIT)
> >> +		return true;
> >> +	return false;
> >> +}
> >> +
> >> +static inline bool pmd_mapped(pmd_t pmd)
> >> +{
> >> +	return pmd_sect(pmd);
> >> +}
> >> +
> >> +static inline pmd_t pmd_mkinvalid(pmd_t pmd)
> >> +{
> >> +	/*
> >> +	 * Invalidation should not have been invoked on
> >> +	 * a PMD table entry. Just warn here otherwise.
> >> +	 */
> >> +	WARN_ON(pmd_table(pmd));
> >> +	return pmd_mksplitting(pmd);
> >> +}
> > 
> > And here we wouldn't need t worry about table checks.
> 
> This is just a temporary sanity check validating the assumption
> that a table entry would never be called with pmdp_invalidate().
> This can be dropped later on if required.

You could use a VM_WARN_ON.

> >> +static inline int pmd_present(pmd_t pmd);
> >> +
> >> +static inline int pmd_trans_huge(pmd_t pmd)
> >> +{
> >> +	if (!pmd_present(pmd))
> >> +		return 0;
> >> +
> >> +	if (!pmd_val(pmd))
> >> +		return 0;
> >> +
> >> +	if (pmd_mapped(pmd))
> >> +		return 1;
> >> +
> >> +	if (pmd_splitting(pmd))
> >> +		return 1;
> >> +	return 0;
> > 
> > Doesn't your new pmd_present() already check for splitting? I think
> 
> I actually meant pte_present() here instead, my bad.
> 
> > checking for bit 0 and the new PMD_PRESENT. That would be similar to
> > what we do with PTE_PROT_NONE. Actually, you could use the same bit for
> > both.
> 
> IIUC PROT NONE is supported at PMD level as well. Hence with valid bit
> cleared, there is a chance for misinterpretation between pmd_protnone()
> and pmd_splitting() if the same bit (PTE_PROT_NONE) is used.

We can indeed have a PROT_NONE pmd but does it matter? All you need is
that pmdp_invalidate() returns the original (present pmd) and writes a
value that is still pmd_present() while invalid. You never modify the
new value again AFAICT (only the old one to rebuild the pmd).

It is indeed a problem if set_pmd_at() clears the new
PMD_INVALID_PRESENT bit but my understanding is that it doesn't need to
(see below).

> >> diff --git a/arch/arm64/mm/mmu.c b/arch/arm64/mm/mmu.c
> >> index 990929c8837e..337519031115 100644
> >> --- a/arch/arm64/mm/mmu.c
> >> +++ b/arch/arm64/mm/mmu.c
> >> @@ -22,6 +22,8 @@
> >>  #include <linux/io.h>
> >>  #include <linux/mm.h>
> >>  #include <linux/vmalloc.h>
> >> +#include <linux/swap.h>
> >> +#include <linux/swapops.h>
> >>  
> >>  #include <asm/barrier.h>
> >>  #include <asm/cputype.h>
> >> @@ -1483,3 +1485,21 @@ static int __init prevent_bootmem_remove_init(void)
> >>  }
> >>  device_initcall(prevent_bootmem_remove_init);
> >>  #endif
> >> +
> >> +#ifdef CONFIG_TRANSPARENT_HUGEPAGE
> >> +void set_pmd_at(struct mm_struct *mm, unsigned long addr,
> >> +		pmd_t *pmdp, pmd_t pmd)
> >> +{
> >> +	/*
> >> +	 * PMD migration entries need to retain splitting PMD
> >> +	 * representation created with pmdp_invalidate(). But
> >> +	 * any non-migration entry which just might have been
> >> +	 * invalidated previously, still need be a normal huge
> >> +	 * page. Hence selectively clear splitting entries.
> >> +	 */
> >> +	if (!is_migration_entry(pmd_to_swp_entry(pmd)))
> >> +		pmd = pmd_clrsplitting(pmd);
> >> +
> >> +	set_pte_at(mm, addr, (pte_t *)pmdp, pmd_pte(pmd));
> >> +}
> >> +#endif
> > 
> > So a pmdp_invalidate() returns the old pmd. Do we ever need to rebuild a
> > pmd based on the actual bits in the new invalidated pmdp? Wondering how
> > the table bit ends up here that we need to pmd_clrsplitting().
> 
> Yes, a pmd is always rebuilt via set_pmd_at() with the old value as
> returned from an earlier pmdp_invalidate() but which may have been
> changed with standard page table entry transformations. Basically,
> it will not be created afresh from the pfn and VMA flags.

My point is that pmdp_invalidate() is never called on an already invalid
pmd. A valid pmd should never have the PMD_INVALID_PRESENT bit set.
Therefore, set_pmd_at() does not need to clear any such bit as it wasn't
in the old value returned by pmdp_invalidate().

> Any additional bit set in PMD via pmdp_invalidate() needs to be
> cleared off in set_pmd_at(), unless it is a migration entry.

I'm not convinced we need to unless we nest pmdp_invalidate() calls
(have you seen any evidence of this?).

-- 
Catalin