On Thu, Aug 01, 2024, Alex Bennée wrote:
> Sean Christopherson <seanjc@xxxxxxxxxx> writes:
>
> > Add an API to release an unused page, i.e. to put a page without marking
> > it accessed or dirty. The API will be used when KVM faults-in a page but
> > bails before installing the guest mapping (and other similar flows).
> >
> > Signed-off-by: Sean Christopherson <seanjc@xxxxxxxxxx>
> > ---
> > include/linux/kvm_host.h | 9 +++++++++
> > 1 file changed, 9 insertions(+)
> >
> > diff --git a/include/linux/kvm_host.h b/include/linux/kvm_host.h
> > index 3d9617d1de41..c5d39a337aa3 100644
> > --- a/include/linux/kvm_host.h
> > +++ b/include/linux/kvm_host.h
> > @@ -1201,6 +1201,15 @@ unsigned long gfn_to_hva_prot(struct kvm *kvm, gfn_t gfn, bool *writable);
> > unsigned long gfn_to_hva_memslot(struct kvm_memory_slot *slot, gfn_t gfn);
> > unsigned long gfn_to_hva_memslot_prot(struct kvm_memory_slot *slot, gfn_t gfn,
> > bool *writable);
> > +
> > +static inline void kvm_release_page_unused(struct page *page)
> > +{
> > + if (!page)
> > + return;
> > +
> > + put_page(page);
> > +}
>
> I guess it's unfamiliarity with the mm layout but I was trying to find
> where the get_pages come from to see the full pattern of allocate and
> return. I guess somewhere in the depths of hva_to_pfn() from
> hva_to_pfn_retry()?
If successful, get_user_page_fast_only() and get_user_pages_unlocked() grab a
reference on behalf of the caller.
As of this patch, hva_to_pfn_remapped() also grabs a reference to pages that
appear to be refcounted, which is the underlying wart this series aims to fix.
In KVM's early days, it _only_ supported GUP, i.e. if KVM got a pfn, that pfn
was (a) backed by struct page and (b) KVM had a reference to said page. That
led to the current mess, as KVM didn't get reworked to properly track pages vs.
pfns when support for VM_MIXEDMAP was added.
/*
* Get a reference here because callers of *hva_to_pfn* and
* *gfn_to_pfn* ultimately call kvm_release_pfn_clean on the
* returned pfn. This is only needed if the VMA has VM_MIXEDMAP
* set, but the kvm_try_get_pfn/kvm_release_pfn_clean pair will
* simply do nothing for reserved pfns.
*
* Whoever called remap_pfn_range is also going to call e.g.
* unmap_mapping_range before the underlying pages are freed,
* causing a call to our MMU notifier.
*
* Certain IO or PFNMAP mappings can be backed with valid
* struct pages, but be allocated without refcounting e.g.,
* tail pages of non-compound higher order allocations, which
* would then underflow the refcount when the caller does the
* required put_page. Don't allow those pages here.
*/
if (!kvm_try_get_pfn(pfn))
r = -EFAULT;
