[PATCH 3/3] kasan: don't assume percpu shadow allocations will succeed

[Daniel Axtens]

syzkaller and the fault injector showed that I was wrong to assume
that we could ignore percpu shadow allocation failures.

Handle failures properly. Merge all the allocated areas back into the free
list and release the shadow, then clean up and return NULL. The shadow
is released unconditionally, which relies upon the fact that the release
function is able to tolerate pages not being present.

Also clean up shadows in the recovery path - currently they are not
released, which leaks a bit of memory.

Fixes: 3c5c3cfb9ef4 ("kasan: support backing vmalloc space with real shadow memory")
Reported-by: syzbot+82e323920b78d54aaed5@xxxxxxxxxxxxxxxxxxxxxxxxx
Reported-by: syzbot+59b7daa4315e07a994f1@xxxxxxxxxxxxxxxxxxxxxxxxx
Cc: Dmitry Vyukov <dvyukov@xxxxxxxxxx>
Cc: Andrey Ryabinin <aryabinin@xxxxxxxxxxxxx>
Signed-off-by: Daniel Axtens <dja@xxxxxxxxxx>
---
 mm/vmalloc.c | 48 ++++++++++++++++++++++++++++++++++++++----------
 1 file changed, 38 insertions(+), 10 deletions(-)

diff --git a/mm/vmalloc.c b/mm/vmalloc.c
index 37af94b6cf30..fa5688093a88 100644
--- a/mm/vmalloc.c
+++ b/mm/vmalloc.c
@@ -3291,7 +3291,7 @@ struct vm_struct **pcpu_get_vm_areas(const unsigned long *offsets,
 	struct vmap_area **vas, *va;
 	struct vm_struct **vms;
 	int area, area2, last_area, term_area;
-	unsigned long base, start, size, end, last_end;
+	unsigned long base, start, size, end, last_end, orig_start, orig_end;
 	bool purged = false;
 	enum fit_type type;
 
@@ -3421,6 +3421,15 @@ struct vm_struct **pcpu_get_vm_areas(const unsigned long *offsets,
 
 	spin_unlock(&free_vmap_area_lock);
 
+	/* populate the kasan shadow space */
+	for (area = 0; area < nr_vms; area++) {
+		if (kasan_populate_vmalloc(vas[area]->va_start, sizes[area]))
+			goto err_free_shadow;
+
+		kasan_unpoison_vmalloc((void *)vas[area]->va_start,
+				       sizes[area]);
+	}
+
 	/* insert all vm's */
 	spin_lock(&vmap_area_lock);
 	for (area = 0; area < nr_vms; area++) {
@@ -3431,13 +3440,6 @@ struct vm_struct **pcpu_get_vm_areas(const unsigned long *offsets,
 	}
 	spin_unlock(&vmap_area_lock);
 
-	/* populate the shadow space outside of the lock */
-	for (area = 0; area < nr_vms; area++) {
-		/* assume success here */
-		kasan_populate_vmalloc(vas[area]->va_start, sizes[area]);
-		kasan_unpoison_vmalloc((void *)vms[area]->addr, sizes[area]);
-	}
-
 	kfree(vas);
 	return vms;
 
@@ -3449,8 +3451,12 @@ struct vm_struct **pcpu_get_vm_areas(const unsigned long *offsets,
 	 * and when pcpu_get_vm_areas() is success.
 	 */
 	while (area--) {
-		merge_or_add_vmap_area(vas[area], &free_vmap_area_root,
-				       &free_vmap_area_list);
+		orig_start = vas[area]->va_start;
+		orig_end = vas[area]->va_end;
+		va = merge_or_add_vmap_area(vas[area], &free_vmap_area_root,
+					    &free_vmap_area_list);
+		kasan_release_vmalloc(orig_start, orig_end,
+				      va->va_start, va->va_end);
 		vas[area] = NULL;
 	}
 
@@ -3485,6 +3491,28 @@ struct vm_struct **pcpu_get_vm_areas(const unsigned long *offsets,
 	kfree(vas);
 	kfree(vms);
 	return NULL;
+
+err_free_shadow:
+	spin_lock(&free_vmap_area_lock);
+	/*
+	 * We release all the vmalloc shadows, even the ones for regions that
+	 * hadn't been successfully added. This relies on kasan_release_vmalloc
+	 * being able to tolerate this case.
+	 */
+	for (area = 0; area < nr_vms; area++) {
+		orig_start = vas[area]->va_start;
+		orig_end = vas[area]->va_end;
+		va = merge_or_add_vmap_area(vas[area], &free_vmap_area_root,
+					    &free_vmap_area_list);
+		kasan_release_vmalloc(orig_start, orig_end,
+				      va->va_start, va->va_end);
+		vas[area] = NULL;
+		kfree(vms[area]);
+	}
+	spin_unlock(&free_vmap_area_lock);
+	kfree(vas);
+	kfree(vms);
+	return NULL;
 }
 
 /**
-- 
2.20.1