mm/hugetlb.c: fix UAF of vma in hugetlb fault pathway

Syzbot reports a UAF in hugetlb_fault().  This happens because
vmf_anon_prepare() could drop the per-VMA lock and allow the current VMA
to be freed before hugetlb_vma_unlock_read() is called.

We can fix this by using a modified version of vmf_anon_prepare() that
doesn't release the VMA lock on failure, and then release it ourselves
after hugetlb_vma_unlock_read().

Link: https://lkml.kernel.org/r/[email protected]
Fixes: 9acad7b ("hugetlb: use vmf_anon_prepare() instead of anon_vma_prepare()")
Reported-by: [email protected]
Closes: https://lore.kernel.org/linux-mm/[email protected]/
Signed-off-by: Vishal Moola (Oracle) <[email protected]>
Cc: Muchun Song <[email protected]>
Cc: <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>

Author: VMoola

mm/hugetlb.c

@@ -6048,7 +6048,7 @@ static vm_fault_t hugetlb_wp(struct folio *pagecache_folio,
 	 * When the original hugepage is shared one, it does not have
 	 * anon_vma prepared.
 	 */
-	ret = vmf_anon_prepare(vmf);
+	ret = __vmf_anon_prepare(vmf);
 	if (unlikely(ret))
 		goto out_release_all;
 
@@ -6247,7 +6247,7 @@ static vm_fault_t hugetlb_no_page(struct address_space *mapping,
 		}
 
 		if (!(vma->vm_flags & VM_MAYSHARE)) {
-			ret = vmf_anon_prepare(vmf);
+			ret = __vmf_anon_prepare(vmf);
 			if (unlikely(ret))
 				goto out;
 		}
@@ -6378,6 +6378,14 @@ static vm_fault_t hugetlb_no_page(struct address_space *mapping,
 	folio_unlock(folio);
 out:
 	hugetlb_vma_unlock_read(vma);
+
+	/*
+	 * We must check to release the per-VMA lock. __vmf_anon_prepare() is
+	 * the only way ret can be set to VM_FAULT_RETRY.
+	 */
+	if (unlikely(ret & VM_FAULT_RETRY))
+		vma_end_read(vma);
+
 	mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 	return ret;
 
@@ -6599,6 +6607,14 @@ vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
 	}
 out_mutex:
 	hugetlb_vma_unlock_read(vma);
+
+	/*
+	 * We must check to release the per-VMA lock. __vmf_anon_prepare() in
+	 * hugetlb_wp() is the only way ret can be set to VM_FAULT_RETRY.
+	 */
+	if (unlikely(ret & VM_FAULT_RETRY))
+		vma_end_read(vma);
+
 	mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 	/*
 	 * Generally it's safe to hold refcount during waiting page lock. But