From bf008669f26f70426ca30ee8cdcbd48e41a74718 Mon Sep 17 00:00:00 2001
From: Philip Withnall <pwithnall@endlessos.org>
Date: Thu, 15 Dec 2022 13:00:39 +0000
Subject: [PATCH 1/2] =?UTF-8?q?gvariant:=20Check=20offset=20table=20doesn?=
 =?UTF-8?q?=E2=80=99t=20fall=20outside=20variant=20bounds?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

When dereferencing the first entry in the offset table for a tuple,
check that it doesn’t fall outside the bounds of the variant first.

This prevents an out-of-bounds read from some non-normal tuples.

This bug was introduced in commit 73d0aa81c2575a5c9ae77d.

Includes a unit test, although the test will likely only catch the
original bug if run with asan enabled.

Signed-off-by: Philip Withnall <pwithnall@endlessos.org>

Fixes: #2840
oss-fuzz#54302
---
 glib/gvariant-serialiser.c | 12 ++++++--
 glib/tests/gvariant.c      | 63 ++++++++++++++++++++++++++++++++++++++
 2 files changed, 72 insertions(+), 3 deletions(-)

diff --git a/glib/gvariant-serialiser.c b/glib/gvariant-serialiser.c
index e9b0eab2b4..af863cc526 100644
--- a/glib/gvariant-serialiser.c
+++ b/glib/gvariant-serialiser.c
@@ -984,7 +984,8 @@ gvs_tuple_get_member_bounds (GVariantSerialised  value,
 
   member_info = g_variant_type_info_member_info (value.type_info, index_);
 
-  if (member_info->i + 1)
+  if (member_info->i + 1 &&
+      offset_size * (member_info->i + 1) <= value.size)
     member_start = gvs_read_unaligned_le (value.data + value.size -
                                           offset_size * (member_info->i + 1),
                                           offset_size);
@@ -995,7 +996,8 @@ gvs_tuple_get_member_bounds (GVariantSerialised  value,
   member_start &= member_info->b;
   member_start |= member_info->c;
 
-  if (member_info->ending_type == G_VARIANT_MEMBER_ENDING_LAST)
+  if (member_info->ending_type == G_VARIANT_MEMBER_ENDING_LAST &&
+      offset_size * (member_info->i + 1) <= value.size)
     member_end = value.size - offset_size * (member_info->i + 1);
 
   else if (member_info->ending_type == G_VARIANT_MEMBER_ENDING_FIXED)
@@ -1006,11 +1008,15 @@ gvs_tuple_get_member_bounds (GVariantSerialised  value,
       member_end = member_start + fixed_size;
     }
 
-  else /* G_VARIANT_MEMBER_ENDING_OFFSET */
+  else if (member_info->ending_type == G_VARIANT_MEMBER_ENDING_OFFSET &&
+           offset_size * (member_info->i + 2) <= value.size)
     member_end = gvs_read_unaligned_le (value.data + value.size -
                                         offset_size * (member_info->i + 2),
                                         offset_size);
 
+  else  /* invalid */
+    member_end = G_MAXSIZE;
+
   if (out_member_start != NULL)
     *out_member_start = member_start;
   if (out_member_end != NULL)
diff --git a/glib/tests/gvariant.c b/glib/tests/gvariant.c
index ba837af48d..e4a2518aa6 100644
--- a/glib/tests/gvariant.c
+++ b/glib/tests/gvariant.c
@@ -5562,6 +5562,67 @@ test_normal_checking_tuple_offsets4 (void)
   g_variant_unref (variant);
 }
 
+/* This is a regression test that dereferencing the first element in the offset
+ * table doesn’t dereference memory before the start of the GVariant. The first
+ * element in the offset table gives the offset of the final member in the
+ * tuple (the offset table is stored in reverse), and the position of this final
+ * member is needed to check that none of the tuple members overlap with the
+ * offset table
+ *
+ * See https://gitlab.gnome.org/GNOME/glib/-/issues/2840 */
+static void
+test_normal_checking_tuple_offsets5 (void)
+{
+  /* A tuple of type (sss) in normal form would have an offset table with two
+   * entries:
+   *  - The first entry (lowest index in the table) gives the offset of the
+   *    third `s` in the tuple, as the offset table is reversed compared to the
+   *    tuple members.
+   *  - The second entry (highest index in the table) gives the offset of the
+   *    second `s` in the tuple.
+   *  - The offset of the first `s` in the tuple is always 0.
+   *
+   * See §2.5.4 (Structures) of the GVariant specification for details, noting
+   * that the table is only layed out this way because all three members of the
+   * tuple have non-fixed sizes.
+   *
+   * It’s not clear whether the 0xaa data of this variant is part of the strings
+   * in the tuple, or part of the offset table. It doesn’t really matter. This
+   * is a regression test to check that the code to validate the offset table
+   * doesn’t unconditionally try to access the first entry in the offset table
+   * by subtracting the table size from the end of the GVariant data.
+   *
+   * In this non-normal case, that would result in an address off the start of
+   * the GVariant data, and an out-of-bounds read, because the GVariant is one
+   * byte long, but the offset table is calculated as two bytes long (with 1B
+   * sized entries) from the tuple’s type.
+   */
+  const GVariantType *data_type = G_VARIANT_TYPE ("(sss)");
+  const guint8 data[] = { 0xaa };
+  gsize size = sizeof (data);
+  GVariant *variant = NULL;
+  GVariant *normal_variant = NULL;
+  GVariant *expected = NULL;
+
+  g_test_bug ("https://gitlab.gnome.org/GNOME/glib/-/issues/2840");
+
+  variant = g_variant_new_from_data (data_type, data, size, FALSE, NULL, NULL);
+  g_assert_nonnull (variant);
+
+  g_assert_false (g_variant_is_normal_form (variant));
+
+  normal_variant = g_variant_get_normal_form (variant);
+  g_assert_nonnull (normal_variant);
+
+  expected = g_variant_new_parsed ("('', '', '')");
+  g_assert_cmpvariant (expected, variant);
+  g_assert_cmpvariant (expected, normal_variant);
+
+  g_variant_unref (expected);
+  g_variant_unref (normal_variant);
+  g_variant_unref (variant);
+}
+
 /* Test that an otherwise-valid serialised GVariant is considered non-normal if
  * its offset table entries are too wide.
  *
@@ -5813,6 +5874,8 @@ main (int argc, char **argv)
                    test_normal_checking_tuple_offsets3);
   g_test_add_func ("/gvariant/normal-checking/tuple-offsets4",
                    test_normal_checking_tuple_offsets4);
+  g_test_add_func ("/gvariant/normal-checking/tuple-offsets5",
+                   test_normal_checking_tuple_offsets5);
   g_test_add_func ("/gvariant/normal-checking/tuple-offsets/minimal-sized",
                    test_normal_checking_tuple_offsets_minimal_sized);
   g_test_add_func ("/gvariant/normal-checking/empty-object-path",
-- 
GitLab


From 4d0bed8c4690f7a2692474ef6a570bd99ef45ef1 Mon Sep 17 00:00:00 2001
From: Philip Withnall <pwithnall@endlessos.org>
Date: Thu, 15 Dec 2022 16:49:28 +0000
Subject: [PATCH 2/2] gvariant: Propagate trust when getting a child of a
 serialised variant

If a variant is trusted, that means all its children are trusted, so
ensure that their checked offsets are set as such.

This allows a lot of the offset table checks to be avoided when getting
children from trusted serialised tuples, which speeds things up.

No unit test is included because this is just a performance fix. If
there are other slownesses, or regressions, in serialised `GVariant`
performance, the fuzzing setup will catch them like it did this one.

This change does reduce the time to run the oss-fuzz reproducer from 80s
to about 0.7s on my machine.

Signed-off-by: Philip Withnall <pwithnall@endlessos.org>

Fixes: #2841
oss-fuzz#54314
---
 glib/gvariant-core.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/glib/gvariant-core.c b/glib/gvariant-core.c
index 75d23f39af..57bbf88cc9 100644
--- a/glib/gvariant-core.c
+++ b/glib/gvariant-core.c
@@ -1198,8 +1198,8 @@ g_variant_get_child_value (GVariant *value,
     child->contents.serialised.bytes =
       g_bytes_ref (value->contents.serialised.bytes);
     child->contents.serialised.data = s_child.data;
-    child->contents.serialised.ordered_offsets_up_to = s_child.ordered_offsets_up_to;
-    child->contents.serialised.checked_offsets_up_to = s_child.checked_offsets_up_to;
+    child->contents.serialised.ordered_offsets_up_to = (value->state & STATE_TRUSTED) ? G_MAXSIZE : s_child.ordered_offsets_up_to;
+    child->contents.serialised.checked_offsets_up_to = (value->state & STATE_TRUSTED) ? G_MAXSIZE : s_child.checked_offsets_up_to;
 
     return child;
   }
-- 
GitLab