On Wed, Nov 09, 2022 at 02:26:36PM -0800, Ben Gardon wrote:
> On Mon, Nov 7, 2022 at 1:59 PM Oliver Upton <oliver.upton@xxxxxxxxx> wrote:
> >
> > In order to service stage-2 faults in parallel, stage-2 table walkers
> > must take exclusive ownership of the PTE being worked on. An additional
> > requirement of the architecture is that software must perform a
> > 'break-before-make' operation when changing the block size used for
> > mapping memory.
> >
> > Roll these two concepts together into helpers for performing a
> > 'break-before-make' sequence. Use a special PTE value to indicate a PTE
> > has been locked by a software walker. Additionally, use an atomic
> > compare-exchange to 'break' the PTE when the stage-2 page tables are
> > possibly shared with another software walker. Elide the DSB + TLBI if
> > the evicted PTE was invalid (and thus not subject to break-before-make).
> >
> > All of the atomics do nothing for now, as the stage-2 walker isn't fully
> > ready to perform parallel walks.
> >
> > Signed-off-by: Oliver Upton <oliver.upton@xxxxxxxxx>
> > ---
> > arch/arm64/kvm/hyp/pgtable.c | 80 +++++++++++++++++++++++++++++++++---
> > 1 file changed, 75 insertions(+), 5 deletions(-)
> >
> > diff --git a/arch/arm64/kvm/hyp/pgtable.c b/arch/arm64/kvm/hyp/pgtable.c
> > index f4dd77c6c97d..b9f0d792b8d9 100644
> > --- a/arch/arm64/kvm/hyp/pgtable.c
> > +++ b/arch/arm64/kvm/hyp/pgtable.c
> > @@ -49,6 +49,12 @@
> > #define KVM_INVALID_PTE_OWNER_MASK GENMASK(9, 2)
> > #define KVM_MAX_OWNER_ID 1
> >
> > +/*
> > + * Used to indicate a pte for which a 'break-before-make' sequence is in
> > + * progress.
> > + */
> > +#define KVM_INVALID_PTE_LOCKED BIT(10)
> > +
> > struct kvm_pgtable_walk_data {
> > struct kvm_pgtable_walker *walker;
> >
> > @@ -674,6 +680,11 @@ static bool stage2_pte_is_counted(kvm_pte_t pte)
> > return !!pte;
> > }
> >
> > +static bool stage2_pte_is_locked(kvm_pte_t pte)
> > +{
> > + return !kvm_pte_valid(pte) && (pte & KVM_INVALID_PTE_LOCKED);
> > +}
> > +
> > static bool stage2_try_set_pte(const struct kvm_pgtable_visit_ctx *ctx, kvm_pte_t new)
> > {
> > if (!kvm_pgtable_walk_shared(ctx)) {
> > @@ -684,6 +695,64 @@ static bool stage2_try_set_pte(const struct kvm_pgtable_visit_ctx *ctx, kvm_pte_
> > return cmpxchg(ctx->ptep, ctx->old, new) == ctx->old;
> > }
> >
> > +/**
> > + * stage2_try_break_pte() - Invalidates a pte according to the
> > + * 'break-before-make' requirements of the
> > + * architecture.
> > + *
> > + * @ctx: context of the visited pte.
> > + * @mmu: stage-2 mmu
> > + *
> > + * Returns: true if the pte was successfully broken.
> > + *
> > + * If the removed pte was valid, performs the necessary serialization and TLB
> > + * invalidation for the old value. For counted ptes, drops the reference count
> > + * on the containing table page.
> > + */
> > +static bool stage2_try_break_pte(const struct kvm_pgtable_visit_ctx *ctx,
> > + struct kvm_s2_mmu *mmu)
> > +{
> > + struct kvm_pgtable_mm_ops *mm_ops = ctx->mm_ops;
> > +
> > + if (stage2_pte_is_locked(ctx->old)) {
> > + /*
> > + * Should never occur if this walker has exclusive access to the
> > + * page tables.
> > + */
> > + WARN_ON(!kvm_pgtable_walk_shared(ctx));
> > + return false;
> > + }
> > +
> > + if (!stage2_try_set_pte(ctx, KVM_INVALID_PTE_LOCKED))
> > + return false;
> > +
> > + /*
> > + * Perform the appropriate TLB invalidation based on the evicted pte
> > + * value (if any).
> > + */
> > + if (kvm_pte_table(ctx->old, ctx->level))
> > + kvm_call_hyp(__kvm_tlb_flush_vmid, mmu);
> > + else if (kvm_pte_valid(ctx->old))
> > + kvm_call_hyp(__kvm_tlb_flush_vmid_ipa, mmu, ctx->addr, ctx->level);
> > +
> > + if (stage2_pte_is_counted(ctx->old))
> > + mm_ops->put_page(ctx->ptep);
> > +
> > + return true;
> > +}
> > +
> > +static void stage2_make_pte(const struct kvm_pgtable_visit_ctx *ctx, kvm_pte_t new)
> > +{
> > + struct kvm_pgtable_mm_ops *mm_ops = ctx->mm_ops;
> > +
> > + WARN_ON(!stage2_pte_is_locked(*ctx->ptep));
> > +
> > + if (stage2_pte_is_counted(new))
> > + mm_ops->get_page(ctx->ptep);
> > +
> > + smp_store_release(ctx->ptep, new);
> > +}
> > +
> > static void stage2_put_pte(const struct kvm_pgtable_visit_ctx *ctx, struct kvm_s2_mmu *mmu,
> > struct kvm_pgtable_mm_ops *mm_ops)
> > {
> > @@ -812,17 +881,18 @@ static int stage2_map_walk_leaf(const struct kvm_pgtable_visit_ctx *ctx,
> > if (!childp)
> > return -ENOMEM;
> >
> > + if (!stage2_try_break_pte(ctx, data->mmu)) {
> > + mm_ops->put_page(childp);
> > + return -EAGAIN;
> > + }
> > +
> > /*
> > * If we've run into an existing block mapping then replace it with
> > * a table. Accesses beyond 'end' that fall within the new table
> > * will be mapped lazily.
> > */
> > - if (stage2_pte_is_counted(ctx->old))
> > - stage2_put_pte(ctx, data->mmu, mm_ops);
> > -
> > new = kvm_init_table_pte(childp, mm_ops);
>
> Does it make any sense to move this before the "break" to minimize the
> critical section in which the PTE is locked?
I had rationalized this before as doing less work in the threads that
would lose a race, but the critical section is very likely to be
performance sensitive as we're unmapping memory after all.
Thanks for the suggestion, I'll fold it in the next spin.
--
Best,
Oliver
