Re: [RFC PATCH] KVM: TDX: Decouple TDX init mem region from kvm_gmem_populate()

[Sean Christopherson <seanjc@google.com>]

On Mon, Jul 14, 2025, David Hildenbrand wrote:
> On 14.07.25 17:46, Sean Christopherson wrote:
> > On Mon, Jul 14, 2025, Yan Zhao wrote:
> > > On Fri, Jul 11, 2025 at 08:39:59AM -0700, Sean Christopherson wrote:
> > > > The below could be tweaked to batch get_user_pages() into an array of pointers,
> > > > but given that both SNP and TDX can only operate on one 4KiB page at a time, and
> > > > that hugepage support doesn't yet exist, trying to super optimize the hugepage
> > > > case straightaway doesn't seem like a pressing concern.
> > > 
> > > > static long __kvm_gmem_populate(struct kvm *kvm, struct kvm_memory_slot *slot,
> > > > 				struct file *file, gfn_t gfn, void __user *src,
> > > > 				kvm_gmem_populate_cb post_populate, void *opaque)
> > > > {
> > > > 	pgoff_t index = kvm_gmem_get_index(slot, gfn);
> > > > 	struct page *src_page = NULL;
> > > > 	bool is_prepared = false;
> > > > 	struct folio *folio;
> > > > 	int ret, max_order;
> > > > 	kvm_pfn_t pfn;
> > > > 
> > > > 	if (src) {
> > > > 		ret = get_user_pages((unsigned long)src, 1, 0, &src_page);
> > > get_user_pages_fast()?
> > > 
> > > get_user_pages() can't pass the assertion of mmap_assert_locked().
> > 
> > Oh, I forgot get_user_pages() requires mmap_lock to already be held.  I would
> > prefer to not use a fast variant, so that userspace isn't required to prefault
> > (and pin?) the source.
> > 
> > So get_user_pages_unlocked()?
> 
> Yes, but likely we really want get_user_pages_fast(), which will fallback to
> GUP-slow (+take the lock) in case it doesn't find what it needs in the page
> tables.
> 
> get_user_pages_fast_only() would be the variant that doesn't fallback to
> GUP-slow.

Doh, right, that's indeed what I want.