On Fri, Jul 31, 2020 at 09:19:11AM +0100, Quentin Perret wrote:
> On Friday 31 Jul 2020 at 09:14:36 (+0100), Quentin Perret wrote:
> > On Thursday 30 Jul 2020 at 16:33:48 (+0100), Will Deacon wrote:
> > > +void *kvm_pgtable_hyp_alloc_cookie(u32 va_bits)
> > > +{
> > > + struct kvm_pgtable *pgt = kzalloc(sizeof(*pgt), GFP_KERNEL);
> > > +
> > > + if (!pgt)
> > > + return NULL;
> > > +
> > > + pgt->ia_bits = va_bits;
> > > + pgt->start_level = kvm_start_level(va_bits);
> > > +
> > > + pgt->pgd = (kvm_pte_t *)get_zeroed_page(GFP_KERNEL);
> > > + if (!pgt->pgd) {
> > > + kfree(pgt);
> > > + pgt = NULL;
> > > + }
> > > +
> > > + return pgt;
> > > +}
> > > +
> > > +static int hyp_free_walker(u64 addr, u64 end, u32 level, kvm_pte_t *ptep,
> > > + enum kvm_pgtable_walk_flags flag, void * const arg)
> > > +{
> > > + free_page((unsigned long)kvm_pte_follow(*ptep));
> > > + return 0;
> > > +}
> > > +
> > > +void kvm_pgtable_hyp_free_cookie(void *cookie)
> > > +{
> > > + size_t pgd_sz;
> > > + struct kvm_pgtable *pgt = cookie;
> > > + struct kvm_pgtable_walker walker = {
> > > + .cb = hyp_free_walker,
> > > + .flags = KVM_PGTABLE_WALK_TABLE_POST,
> > > + };
> > > +
> > > + if (kvm_pgtable_walk(cookie, 0, BIT(pgt->ia_bits), &walker))
> > > + kvm_err("Failed to free page-table pages\n");
> > > +
> > > + pgd_sz = kvm_pgd_pages(pgt) * PAGE_SIZE;
> > > + free_pages_exact(pgt->pgd, pgd_sz);
> >
> > Given that the pgd is unconditionally a single page for the stage 1 case
> > (as per kvm_pgtable_hyp_alloc_cookie() above), should we simplify this
> > to a simple free_page()? Or did you want to factorize this with the
> > stage 2 free path?
>
> Hmm, or maybe it's the alloc() path that needs fixing actually ...
No, I think you're right. The hyp stage-1 PGD is always a single page, so I
can simplify the free() path. At one point, I had a common free() path for
stage-1 and stage-2, but that didn't last very long, as the stage-2
page-table pages are refcounted so you end up having to deal with them
differently anyway.
Do you think there is a functional issue here? I don't immediately see it.
Will
_______________________________________________
kvmarm mailing list
kvmarm@xxxxxxxxxxxxxxxxxxxxx
https://lists.cs.columbia.edu/mailman/listinfo/kvmarm
