+Alex, who is looking at the huge-VM_PFNMAP angle in particular.
On Thu, Mar 14, 2024, Sean Christopherson wrote:
> -Christ{oph,ian} to avoid creating more noise...
>
> On Thu, Mar 14, 2024, David Stevens wrote:
> > Because of that, the specific type of pfns that don't work right now are
> > pfn_valid() && !PG_Reserved && !page_ref_count() - what I called the
> > non-refcounted pages in a bad choice of words. If that's correct, then
> > perhaps this series should go a little bit further in modifying
> > hva_to_pfn_remapped, but it isn't fundamentally wrong.
>
> Loosely related to all of this, I have a mildly ambitious idea. Well, one mildly
> ambitious idea, and one crazy ambitious idea. Crazy ambitious idea first...
>
> Something we (GCE side of Google) have been eyeballing is adding support for huge
> VM_PFNMAP memory, e.g. for mapping large amounts of device (a.k.a. GPU) memory
> into guests using hugepages. One of the hiccups is that follow_pte() doesn't play
> nice with hugepages, at all, e.g. even has a "VM_BUG_ON(pmd_trans_huge(*pmd))".
> Teaching follow_pte() to play nice with hugepage probably is doing, but making
> sure all existing users are aware, maybe not so much.
>
> My first (half baked, crazy ambitious) idea is to move away from follow_pte() and
> get_user_page_fast_only() for mmu_notifier-aware lookups, i.e. that don't need
> to grab references, and replace them with a new converged API that locklessly walks
> host userspace page tables, and grabs the hugepage size along the way, e.g. so that
> arch code wouldn't have to do a second walk of the page tables just to get the
> hugepage size.
>
> In other words, for the common case (mmu_notifier integration, no reference needed),
> route hva_to_pfn_fast() into the new API and walk the userspace page tables (probably
> only for write faults, to avoid CoW compliciations) before doing anything else.
>
> Uses of hva_to_pfn() that need to get a reference to the struct page couldn't be
> converted, e.g. when stuffing physical addresses into the VMCS for nested virtualization.
> But for everything else, grabbing a reference is a non-goal, i.e. actually "getting"
> a user page is wasted effort and actively gets in the way.
>
> I was initially hoping we could go super simple and use something like x86's
> host_pfn_mapping_level(), but there are too many edge cases in gup() that need to
> be respected, e.g. to avoid mapping memfd_secret pages into KVM guests. I.e. the
> API would need to be a formal mm-owned thing, not some homebrewed KVM implementation.
>
> I can't tell if the payoff would be big enough to justify the effort involved, i.e.
> having a single unified API for grabbing PFNs from the primary MMU might just be a
> pie-in-the-sky type idea.
>
> My second, less ambitious idea: the previously linked LWN[*] article about the
> writeback issues reminded me of something that has bugged me for a long time. IIUC,
> getting a writable mapping from the primary MMU marks the page/folio dirty, and that
> page/folio stays dirty until the data is written back and the mapping is made read-only.
> And because KVM is tapped into the mmu_notifiers, KVM will be notified *before* the
> RW=>RO conversion completes, i.e. before the page/folio is marked clean.
>
> I _think_ that means that calling kvm_set_page_dirty() when zapping a SPTE (or
> dropping any mmu_notifier-aware mapping) is completely unnecessary. If that is the
> case, _and_ we can weasel our way out of calling kvm_set_page_accessed() too, then
> with FOLL_GET plumbed into hva_to_pfn(), we can:
>
> - Drop kvm_{set,release}_pfn_{accessed,dirty}(), because all callers of hva_to_pfn()
> that aren't tied into mmu_notifiers, i.e. aren't guaranteed to drop mappings
> before the page/folio is cleaned, will *know* that they hold a refcounted struct
> page.
>
> - Skip "KVM: x86/mmu: Track if sptes refer to refcounted pages" entirely, because
> KVM never needs to know if a SPTE points at a refcounted page.
>
> In other words, double down on immediately doing put_page() after gup() if FOLL_GET
> isn't specified, and naturally make all KVM MMUs compatible with pfn_valid() PFNs
> that are acquired by follow_pte().
>
> I suspect we can simply mark pages as access when a page is retrieved from the primary
> MMU, as marking a page accessed when its *removed* from the guest is rather nonsensical.
> E.g. if a page is mapped into the guest for a long time and it gets swapped out, marking
> the page accessed when KVM drops its SPTEs in response to the swap adds no value. And
> through the mmu_notifiers, KVM already plays nice with setups that use idle page
> tracking to make reclaim decisions.
>
> [*] https://lwn.net/Articles/930667
