[PATCH v7 6/7] KVM: guest_memfd: Handle invalidation of shared memory

[Fuad Tabba <tabba@xxxxxxxxxx>]

When guest_memfd backed memory is invalidated, e.g., on punching holes,
releasing, ensure that the sharing states are updated and that any
folios in a transient state are restored to an appropriate state.

Signed-off-by: Fuad Tabba <tabba@xxxxxxxxxx>
---
 virt/kvm/guest_memfd.c | 57 ++++++++++++++++++++++++++++++++++++++++++
 1 file changed, 57 insertions(+)

diff --git a/virt/kvm/guest_memfd.c b/virt/kvm/guest_memfd.c
index ce19bd6c2031..eec9d5e09f09 100644
--- a/virt/kvm/guest_memfd.c
+++ b/virt/kvm/guest_memfd.c
@@ -118,6 +118,16 @@ static struct folio *kvm_gmem_get_folio(struct inode *inode, pgoff_t index)
 	return filemap_grab_folio(inode->i_mapping, index);
 }
 
+#ifdef CONFIG_KVM_GMEM_SHARED_MEM
+static void kvm_gmem_offset_range_invalidate_shared(struct inode *inode,
+						    pgoff_t start, pgoff_t end);
+#else
+static inline void kvm_gmem_offset_range_invalidate_shared(struct inode *inode,
+							   pgoff_t start, pgoff_t end)
+{
+}
+#endif
+
 static void kvm_gmem_invalidate_begin(struct kvm_gmem *gmem, pgoff_t start,
 				      pgoff_t end)
 {
@@ -127,6 +137,7 @@ static void kvm_gmem_invalidate_begin(struct kvm_gmem *gmem, pgoff_t start,
 	unsigned long index;
 
 	xa_for_each_range(&gmem->bindings, index, slot, start, end - 1) {
+		struct file *file = READ_ONCE(slot->gmem.file);
 		pgoff_t pgoff = slot->gmem.pgoff;
 
 		struct kvm_gfn_range gfn_range = {
@@ -146,6 +157,16 @@ static void kvm_gmem_invalidate_begin(struct kvm_gmem *gmem, pgoff_t start,
 		}
 
 		flush |= kvm_mmu_unmap_gfn_range(kvm, &gfn_range);
+
+		/*
+		 * If this gets called after kvm_gmem_unbind() it means that all
+		 * in-flight operations are gone, and the file has been closed.
+		 */
+		if (file) {
+			kvm_gmem_offset_range_invalidate_shared(file_inode(file),
+								gfn_range.start,
+								gfn_range.end);
+		}
 	}
 
 	if (flush)
@@ -512,6 +533,42 @@ static int kvm_gmem_offset_clear_shared(struct inode *inode, pgoff_t index)
 	return r;
 }
 
+/*
+ * Callback when invalidating memory that is potentially shared.
+ *
+ * Must be called with the offsets_lock write lock held.
+ */
+static void kvm_gmem_offset_range_invalidate_shared(struct inode *inode,
+						    pgoff_t start, pgoff_t end)
+{
+	struct xarray *shared_offsets = &kvm_gmem_private(inode)->shared_offsets;
+	rwlock_t *offsets_lock = &kvm_gmem_private(inode)->offsets_lock;
+	pgoff_t i;
+
+	lockdep_assert_held_write(offsets_lock);
+
+	for (i = start; i < end; i++) {
+		/*
+		 * If the folio is NONE_SHARED, it indicates that it's
+		 * transitioning to private (GUEST_SHARED). Transition it to
+		 * shared (ALL_SHARED) and remove the callback.
+		 */
+		if (xa_to_value(xa_load(shared_offsets, i)) == KVM_GMEM_NONE_SHARED) {
+			struct folio *folio = filemap_lock_folio(inode->i_mapping, i);
+
+			if (!WARN_ON_ONCE(IS_ERR(folio))) {
+				if (folio_test_guestmem(folio))
+					kvm_gmem_restore_pending_folio(folio, inode);
+
+				folio_unlock(folio);
+				folio_put(folio);
+			}
+		}
+
+		xa_erase(shared_offsets, i);
+	}
+}
+
 /*
  * Marks the range [start, end) as not shared with the host. If the host doesn't
  * have any references to a particular folio, then that folio is marked as
-- 
2.49.0.472.ge94155a9ec-goog