On Mon, May 29, 2023 at 7:18 AM Marc Zyngier <maz@xxxxxxxxxx> wrote:
>
> On Fri, 19 May 2023 01:52:31 +0100,
> Raghavendra Rao Ananta <rananta@xxxxxxxxxx> wrote:
> >
> > The current implementation of the stage-2 unmap walker traverses
> > the given range and, as a part of break-before-make, performs
> > TLB invalidations with a DSB for every PTE. A multitude of this
> > combination could cause a performance bottleneck.
> >
> > Hence, if the system supports FEAT_TLBIRANGE, defer the TLB
> > invalidations until the entire walk is finished, and then
> > use range-based instructions to invalidate the TLBs in one go.
> > Condition this upon S2FWB in order to avoid walking the page-table
> > again to perform the CMOs after issuing the TLBI.
>
> But that's the real bottleneck. TLBIs are cheap compared to CMOs, even
> on remarkably bad implementations. What is your plan to fix this?
>
Correct me if I'm wrong, but my understanding was that a multiple
issuance of TLBI + DSB was the bottleneck, and this patch tries to
avoid this by issuing only one TLBI + DSB at the end.
> >
> > Rename stage2_put_pte() to stage2_unmap_put_pte() as the function
> > now serves the stage-2 unmap walker specifically, rather than
> > acting generic.
> >
> > Signed-off-by: Raghavendra Rao Ananta <rananta@xxxxxxxxxx>
> > ---
> > arch/arm64/kvm/hyp/pgtable.c | 35 ++++++++++++++++++++++++++++++-----
> > 1 file changed, 30 insertions(+), 5 deletions(-)
> >
> > diff --git a/arch/arm64/kvm/hyp/pgtable.c b/arch/arm64/kvm/hyp/pgtable.c
> > index b8f0dbd12f773..5832ee3418fb0 100644
> > --- a/arch/arm64/kvm/hyp/pgtable.c
> > +++ b/arch/arm64/kvm/hyp/pgtable.c
> > @@ -771,16 +771,34 @@ static void stage2_make_pte(const struct kvm_pgtable_visit_ctx *ctx, kvm_pte_t n
> > 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)
> > +static bool stage2_unmap_defer_tlb_flush(struct kvm_pgtable *pgt)
> > {
> > + /*
> > + * If FEAT_TLBIRANGE is implemented, defer the individial PTE
> > + * TLB invalidations until the entire walk is finished, and
> > + * then use the range-based TLBI instructions to do the
> > + * invalidations. Condition this upon S2FWB in order to avoid
> > + * a page-table walk again to perform the CMOs after TLBI.
> > + */
> > + return system_supports_tlb_range() && stage2_has_fwb(pgt);
> > +}
> > +
> > +static void stage2_unmap_put_pte(const struct kvm_pgtable_visit_ctx *ctx,
> > + struct kvm_s2_mmu *mmu,
> > + struct kvm_pgtable_mm_ops *mm_ops)
> > +{
> > + struct kvm_pgtable *pgt = ctx->arg;
> > +
> > /*
> > * Clear the existing PTE, and perform break-before-make with
> > * TLB maintenance if it was valid.
> > */
> > if (kvm_pte_valid(ctx->old)) {
> > kvm_clear_pte(ctx->ptep);
> > - kvm_call_hyp(__kvm_tlb_flush_vmid_ipa, mmu, ctx->addr, ctx->level);
> > +
> > + if (!stage2_unmap_defer_tlb_flush(pgt))
> > + kvm_call_hyp(__kvm_tlb_flush_vmid_ipa, mmu,
> > + ctx->addr, ctx->level);
>
> This really doesn't match the comment anymore.
>
Right, I can re-write this in the next spin.
> Overall, I'm very concerned that we lose the consistency property that
> the current code has: once called, the TLBs and the page tables are
> synchronised.
>
> Yes, this patch looks correct. But it is also really fragile.
>
Yeah, we were a little skeptical about this too. Till v2, we had a
different implementation in which we had an independent fast unmap
path that disconnects the PTE hierarchy if the unmap range was exactly
KVM_PGTABLE_MIN_BLOCK_LEVEL [1]. But this had some problems, and we
pivoted to the current implementation.
> > }
> >
> > mm_ops->put_page(ctx->ptep);
> > @@ -1015,7 +1033,7 @@ static int stage2_unmap_walker(const struct kvm_pgtable_visit_ctx *ctx,
> > * block entry and rely on the remaining portions being faulted
> > * back lazily.
> > */
> > - stage2_put_pte(ctx, mmu, mm_ops);
> > + stage2_unmap_put_pte(ctx, mmu, mm_ops);
> >
> > if (need_flush && mm_ops->dcache_clean_inval_poc)
> > mm_ops->dcache_clean_inval_poc(kvm_pte_follow(ctx->old, mm_ops),
> > @@ -1029,13 +1047,20 @@ static int stage2_unmap_walker(const struct kvm_pgtable_visit_ctx *ctx,
> >
> > int kvm_pgtable_stage2_unmap(struct kvm_pgtable *pgt, u64 addr, u64 size)
> > {
> > + int ret;
> > struct kvm_pgtable_walker walker = {
> > .cb = stage2_unmap_walker,
> > .arg = pgt,
> > .flags = KVM_PGTABLE_WALK_LEAF | KVM_PGTABLE_WALK_TABLE_POST,
> > };
> >
> > - return kvm_pgtable_walk(pgt, addr, size, &walker);
> > + ret = kvm_pgtable_walk(pgt, addr, size, &walker);
> > + if (stage2_unmap_defer_tlb_flush(pgt))
> > + /* Perform the deferred TLB invalidations */
> > + kvm_call_hyp(__kvm_tlb_flush_vmid_range, pgt->mmu,
> > + addr, addr + size);
>
> This "kvm_call_hyp(__kvm_tlb_flush_vmid_range,...)" could do with a
> wrapper from the point where you introduce it.
>
Sorry, I didn't get this comment. Do you mind elaborating on it?
Thank you.
Raghavendra
[1]: https://lore.kernel.org/all/20230206172340.2639971-8-rananta@xxxxxxxxxx/
> > +
> > + return ret;
> > }
> >
>
> Thanks,
>
> M.
>
> --
> Without deviation from the norm, progress is not possible.
