userfaultfd: shmem: allocate anonymous memory for MAP_PRIVATE shmem

Userfaultfd did not create private memory when UFFDIO_COPY was invoked
on a MAP_PRIVATE shmem mapping.  Instead it wrote to the shmem file,
even when that had not been opened for writing.  Though, fortunately,
that could only happen where there was a hole in the file.

Fix the shmem-backed implementation of UFFDIO_COPY to create private
memory for MAP_PRIVATE mappings.  The hugetlbfs-backed implementation
was already correct.

This change is visible to userland, if userfaultfd has been used in
unintended ways: so it introduces a small risk of incompatibility, but
is necessary in order to respect file permissions.

An app that uses UFFDIO_COPY for anything like postcopy live migration
won't notice the difference, and in fact it'll run faster because there
will be no copy-on-write and memory waste in the tmpfs pagecache
anymore.

Userfaults on MAP_PRIVATE shmem keep triggering only on file holes like
before.

The real zeropage can also be built on a MAP_PRIVATE shmem mapping
through UFFDIO_ZEROPAGE and that's safe because the zeropage pte is
never dirty, in turn even an mprotect upgrading the vma permission from
PROT_READ to PROT_READ|PROT_WRITE won't make the zeropage pte writable.

Link: http://lkml.kernel.org/r/[email protected]
Fixes: 4c27fe4 ("userfaultfd: shmem: add shmem_mcopy_atomic_pte for userfaultfd support")
Signed-off-by: Andrea Arcangeli <[email protected]>
Reported-by: Mike Rapoport <[email protected]>
Reviewed-by: Hugh Dickins <[email protected]>
Cc: <[email protected]>
Cc: "Dr. David Alan Gilbert" <[email protected]>
Cc: Jann Horn <[email protected]>
Cc: Mike Kravetz <[email protected]>
Cc: Peter Xu <[email protected]>
Cc: [email protected]
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>

Author: aagit

mm/userfaultfd.c

@@ -380,7 +380,17 @@ static __always_inline ssize_t mfill_atomic_pte(struct mm_struct *dst_mm,
 {
 	ssize_t err;
 
-	if (vma_is_anonymous(dst_vma)) {
+	/*
+	 * The normal page fault path for a shmem will invoke the
+	 * fault, fill the hole in the file and COW it right away. The
+	 * result generates plain anonymous memory. So when we are
+	 * asked to fill an hole in a MAP_PRIVATE shmem mapping, we'll
+	 * generate anonymous memory directly without actually filling
+	 * the hole. For the MAP_PRIVATE case the robustness check
+	 * only happens in the pagetable (to verify it's still none)
+	 * and not in the radix tree.
+	 */
+	if (!(dst_vma->vm_flags & VM_SHARED)) {
 		if (!zeropage)
 			err = mcopy_atomic_pte(dst_mm, dst_pmd, dst_vma,
 					       dst_addr, src_addr, page);
@@ -489,7 +499,8 @@ static __always_inline ssize_t __mcopy_atomic(struct mm_struct *dst_mm,
 	 * dst_vma.
 	 */
 	err = -ENOMEM;
-	if (vma_is_anonymous(dst_vma) && unlikely(anon_vma_prepare(dst_vma)))
+	if (!(dst_vma->vm_flags & VM_SHARED) &&
+	    unlikely(anon_vma_prepare(dst_vma)))
 		goto out_unlock;
 
 	while (src_addr < src_start + len) {