Re: [PATCH 1/3] KVM: arm64: Don't defer TLB invalidation when zapping table entries

From: Oliver Upton <oliver.upton@linux.dev>
To: Will Deacon <will@kernel.org>
Cc: kvmarm@lists.linux.dev, linux-arm-kernel@lists.infradead.org,
	Catalin Marinas <catalin.marinas@arm.com>,
	Gavin Shan <gshan@redhat.com>, Marc Zyngier <maz@kernel.org>,
	Mostafa Saleh <smostafa@google.com>,
	Quentin Perret <qperret@google.com>,
	Raghavendra Rao Ananta <rananta@google.com>,
	Ryan Roberts <ryan.roberts@arm.com>,
	Shaoqin Huang <shahuang@redhat.com>,
	Suzuki K Poulose <suzuki.poulose@arm.com>,
	Zenghui Yu <yuzenghui@huawei.com>
Subject: Re: [PATCH 1/3] KVM: arm64: Don't defer TLB invalidation when zapping table entries
Date: Tue, 26 Mar 2024 01:34:17 -0700	[thread overview]
Message-ID: <ZgKIides0YAA2j5Y@linux.dev> (raw)
In-Reply-To: <20240325185158.8565-2-will@kernel.org>

Hey Will,

On Mon, Mar 25, 2024 at 06:51:56PM +0000, Will Deacon wrote:
> Commit 7657ea920c54 ("KVM: arm64: Use TLBI range-based instructions for
> unmap") introduced deferred TLB invalidation for the stage-2 page-table
> so that range-based invalidation can be used for the accumulated
> addresses. This works fine if the structure of the page-tables remains
> unchanged, but if entire tables are zapped and subsequently freed then
> we transiently leave the hardware page-table walker with a reference
> to freed memory thanks to the translation walk caches.

Yikes! Well spotted. This is rather unfortunate because the unmap path
has been found to be a massive pain w/o aggregating invalidations. But
sacrificing correctness in the name of performance... No thanks :)

> diff --git a/arch/arm64/kvm/hyp/pgtable.c b/arch/arm64/kvm/hyp/pgtable.c
> index 3fae5830f8d2..de0b667ba296 100644
> --- a/arch/arm64/kvm/hyp/pgtable.c
> +++ b/arch/arm64/kvm/hyp/pgtable.c
> @@ -896,9 +896,11 @@ static void stage2_unmap_put_pte(const struct kvm_pgtable_visit_ctx *ctx,
>  	if (kvm_pte_valid(ctx->old)) {
>  		kvm_clear_pte(ctx->ptep);
>  
> -		if (!stage2_unmap_defer_tlb_flush(pgt))
> +		if (!stage2_unmap_defer_tlb_flush(pgt) ||
> +		    kvm_pte_table(ctx->old, ctx->level)) {
>  			kvm_call_hyp(__kvm_tlb_flush_vmid_ipa, mmu,
>  					ctx->addr, ctx->level);
> +		}

I'm not sure this is correct, though. My understanding of TTL is that
we're telling hardware where the *leaf* entry we're invalidating is
found, however here we know that the addressed PTE is a table entry.

So maybe in the case of a table PTE this invalidation should
TLBI_TTL_UNKNOWN.

>  	}
>  
>  	mm_ops->put_page(ctx->ptep);

At least for the 'normal' MMU where we use RCU, this could be changed to
->free_unlinked_table() which would defer the freeing of memory til
after the invalidation completes. But that still hoses pKVM's stage-2
MMU freeing in-place.

-- 
Thanks,
Oliver