On Mon, 6 Mar 2023 17:06:44 +0000
Catalin Marinas <catalin.marinas@xxxxxxx> wrote:
> On Mon, Mar 06, 2023 at 05:15:48PM +0100, Gerald Schaefer wrote:
> > diff --git a/arch/arm64/include/asm/pgtable.h b/arch/arm64/include/asm/pgtable.h
> > index b6ba466e2e8a..0bd18de9fd97 100644
> > --- a/arch/arm64/include/asm/pgtable.h
> > +++ b/arch/arm64/include/asm/pgtable.h
> > @@ -57,7 +57,7 @@ static inline bool arch_thp_swp_supported(void)
> > * fault on one CPU which has been handled concurrently by another CPU
> > * does not need to perform additional invalidation.
> > */
> > -#define flush_tlb_fix_spurious_fault(vma, address) do { } while (0)
> > +#define flush_tlb_fix_spurious_fault(vma, address, ptep) do { } while (0)
>
> For arm64:
>
> Acked-by: Catalin Marinas <catalin.marinas@xxxxxxx>
>
> > diff --git a/arch/s390/include/asm/pgtable.h b/arch/s390/include/asm/pgtable.h
> > index 2c70b4d1263d..c1f6b46ec555 100644
> > --- a/arch/s390/include/asm/pgtable.h
> > +++ b/arch/s390/include/asm/pgtable.h
> > @@ -1239,7 +1239,8 @@ static inline int pte_allow_rdp(pte_t old, pte_t new)
> > }
> >
> > static inline void flush_tlb_fix_spurious_fault(struct vm_area_struct *vma,
> > - unsigned long address)
> > + unsigned long address,
> > + pte_t *ptep)
> > {
> > /*
> > * RDP might not have propagated the PTE protection reset to all CPUs,
> > @@ -1247,11 +1248,12 @@ static inline void flush_tlb_fix_spurious_fault(struct vm_area_struct *vma,
> > * NOTE: This will also be called when a racing pagetable update on
> > * another thread already installed the correct PTE. Both cases cannot
> > * really be distinguished.
> > - * Therefore, only do the local TLB flush when RDP can be used, to avoid
> > - * unnecessary overhead.
> > + * Therefore, only do the local TLB flush when RDP can be used, and the
> > + * PTE does not have _PAGE_PROTECT set, to avoid unnecessary overhead.
> > + * A local RDP can be used to do the flush.
> > */
> > - if (MACHINE_HAS_RDP)
> > - asm volatile("ptlb" : : : "memory");
> > + if (MACHINE_HAS_RDP && !(pte_val(*ptep) & _PAGE_PROTECT))
> > + __ptep_rdp(address, ptep, 0, 0, 1);
>
> I wonder whether passing the actual entry is somewhat quicker as it
> avoids another memory access (though it might already be in the cache).
The RDP instruction itself only requires the PTE pointer as input, or more
precisely a pointer to the pagetable origin. We calculate that from the PTE
pointer, by masking out some bits, w/o actual memory access to the PTE entry
value.
Of course, there is the pte_val(*ptep) & _PAGE_PROTECT check here, with
memory access, but this might get removed in the future. TBH, I simply
wasn't sure (enough) yet, if we could technically ever end up here with
_PAGE_PROTECT set at all. For "real" spurious protection faults, it should
never be set, not so sure about racing pagetable updates though.
So this might actually be an unnecessary / overly cautious check, that
gets removed in the future, and not worth passing along the PTE value
in addition to the pointer.
