[PATCH 14/30] KVM: arm64: Hook up reclaim hypercall to pkvm_pgtable_stage2_destroy()

[Quentin Perret]

On Friday 09 Jan 2026 at 14:35:57 (+0000), Will Deacon wrote:
> On Tue, Jan 06, 2026 at 02:59:19PM +0000, Quentin Perret wrote:
> > On Monday 05 Jan 2026 at 15:49:22 (+0000), Will Deacon wrote:
> > > During teardown of a protected guest, its memory pages must be reclaimed
> > > from the hypervisor by issuing the '__pkvm_reclaim_dying_guest_page'
> > > hypercall.
> > > 
> > > Add a new helper, __pkvm_pgtable_stage2_reclaim(), which is called
> > > during the VM teardown operation to reclaim pages from the hypervisor
> > > and drop the GUP pin on the host.
> > > 
> > > Signed-off-by: Will Deacon <will at kernel.org>
> > > ---
> > >  arch/arm64/kvm/pkvm.c | 31 ++++++++++++++++++++++++++++++-
> > >  1 file changed, 30 insertions(+), 1 deletion(-)
> > > 
> > > diff --git a/arch/arm64/kvm/pkvm.c b/arch/arm64/kvm/pkvm.c
> > > index 1814e17d600e..8be91051699e 100644
> > > --- a/arch/arm64/kvm/pkvm.c
> > > +++ b/arch/arm64/kvm/pkvm.c
> > > @@ -322,6 +322,32 @@ int pkvm_pgtable_stage2_init(struct kvm_pgtable *pgt, struct kvm_s2_mmu *mmu,
> > >  	return 0;
> > >  }
> > >  
> > > +static int __pkvm_pgtable_stage2_reclaim(struct kvm_pgtable *pgt, u64 start, u64 end)
> > > +{
> > > +	struct kvm *kvm = kvm_s2_mmu_to_kvm(pgt->mmu);
> > > +	pkvm_handle_t handle = kvm->arch.pkvm.handle;
> > > +	struct pkvm_mapping *mapping;
> > > +	int ret;
> > > +
> > > +	for_each_mapping_in_range_safe(pgt, start, end, mapping) {
> > > +		struct page *page;
> > > +
> > > +		ret = kvm_call_hyp_nvhe(__pkvm_reclaim_dying_guest_page,
> > > +					handle, mapping->gfn);
> > > +		if (WARN_ON(ret))
> > > +			return ret;
> > > +
> > > +		page = pfn_to_page(mapping->pfn);
> > > +		WARN_ON_ONCE(mapping->nr_pages != 1);
> > > +		unpin_user_pages_dirty_lock(&page, 1, true);
> > > +		account_locked_vm(current->mm, 1, false);
> > > +		pkvm_mapping_remove(mapping, &pgt->pkvm_mappings);
> > > +		kfree(mapping);
> > 
> > Nit: this might take a while, worth adding a cond_resched() in here?
> 
> Commit 4ddfab5436b6 ("KVM: arm64: Reschedule as needed when destroying
> the stage-2 page-tables") breaks the destruction operation up into
> chunks so we now reschedule from the caller in stage2_destroy_range().
> 
> Do you think we need an additional reschedule?

I would intuitively think so, but I've also just realized we already
don't have that resched for np guests, so this patch is absolutely fine
as-is.

It's just that knocking down a whole 1G (assuming 4K pages) will be
significantly faster for normal KVM than for pKVM. It's one call in
pgtable.c with possibly a single deferred TLBI vs 1 hcall and 1 TLBI per
4K page. But I guess this is really performance tuning, so this can
always come later. Feel free to ignore :)