@verbatim
Copyright (C) 2013
Free Software Foundation, Inc.
This file is part of GCC.
GCC is free software; you can redistribute it and/or modify it under the terms of the GNU General Public License as published by the Free Software Foundation; either version 3, or (at your option) any later version.
GCC is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for more details.
You should have received a copy of the GNU General Public License along with GCC; see the file COPYING3. If not see http://www.gnu.org/licenses/.
Virtual Table Pointer Security Pass - Detect corruption of vtable pointers
before using them for virtual method dispatches.
This file is part of the vtable security feature implementation.
The vtable security feature is designed to detect when a virtual
call is about to be made through an invalid vtable pointer
(possibly due to data corruption or malicious attacks). The
compiler finds every virtual call, and inserts a verification call
before the virtual call. The verification call takes the actual
vtable pointer value in the object through which the virtual call
is being made, and compares the vtable pointer against a set of all
valid vtable pointers that the object could contain (this set is
based on the declared type of the object). If the pointer is in
the valid set, execution is allowed to continue; otherwise the
program is halted.
There are several pieces needed in order to make this work: 1. For
every virtual class in the program (i.e. a class that contains
virtual methods), we need to build the set of all possible valid
vtables that an object of that class could point to. This includes
vtables for any class(es) that inherit from the class under
consideration. 2. For every such data set we build up, we need a
way to find and reference the data set. This is complicated by the
fact that the real vtable addresses are not known until runtime,
when the program is loaded into memory, but we need to reference the
sets at compile time when we are inserting verification calls into
the program. 3. We need to find every virtual call in the program,
and insert the verification call (with the appropriate arguments)
before the virtual call. 4. We need some runtime library pieces:
the code to build up the data sets at runtime; the code to actually
perform the verification using the data sets; and some code to set
protections on the data sets, so they themselves do not become
hacker targets.
To find and reference the set of valid vtable pointers for any given
virtual class, we create a special global variable for each virtual
class. We refer to this as the "vtable map variable" for that
class. The vtable map variable has the type "void *", and is
initialized by the compiler to NULL. At runtime when the set of
valid vtable pointers for a virtual class, e.g. class Foo, is built,
the vtable map variable for class Foo is made to point to the set.
During compile time, when the compiler is inserting verification
calls into the program, it passes the vtable map variable for the
appropriate class to the verification call, so that at runtime the
verification call can find the appropriate data set.
The actual set of valid vtable pointers for a virtual class,
e.g. class Foo, cannot be built until runtime, when the vtables get
loaded into memory and their addresses are known. But the knowledge
about which vtables belong in which class' hierarchy is only known
at compile time. Therefore at compile time we collect class
hierarchy and vtable information about every virtual class, and we
generate calls to build up the data sets at runtime. To build the
data sets, we call one of the functions we add to the runtime
library, __VLTRegisterPair. __VLTRegisterPair takes two arguments,
a vtable map variable and the address of a vtable. If the vtable
map variable is currently NULL, it creates a new data set (hash
table), makes the vtable map variable point to the new data set, and
inserts the vtable address into the data set. If the vtable map
variable is not NULL, it just inserts the vtable address into the
data set. In order to make sure that our data sets are built before
any verification calls happen, we create a special constructor
initialization function for each compilation unit, give it a very
high initialization priority, and insert all of our calls to
__VLTRegisterPair into our special constructor initialization
function.
The vtable verification feature is controlled by the flag
'-fvtable-verify='. There are three flavors of this:
'-fvtable-verify=std', '-fvtable-verify=preinit', and
'-fvtable-verify=none'. If the option '-fvtable-verfy=preinit' is
used, then our constructor initialization function gets put into the
preinit array. This is necessary if there are data sets that need
to be built very early in execution. If the constructor
initialization function gets put into the preinit array, the we also
add calls to __VLTChangePermission at the beginning and end of the
function. The call at the beginning sets the permissions on the
data sets and vtable map variables to read/write, and the one at the
end makes them read-only. If the '-fvtable-verify=std' option is
used, the constructor initialization functions are executed at their
normal time, and the __VLTChangePermission calls are handled
differently (see the comments in libstdc++-v3/libsupc++/vtv_rts.cc).
The option '-fvtable-verify=none' turns off vtable verification.
This file contains code for the tree pass that goes through all the
statements in each basic block, looking for virtual calls, and
inserting a call to __VLTVerifyVtablePointer (with appropriate
arguments) before each one. It also contains the hash table
functions for the data structures used for collecting the class
hierarchy data and building/maintaining the vtable map variable data
are defined in gcc/vtable-verify.h. These data structures are
shared with the code in the C++ front end that collects the class
hierarchy & vtable information and generates the vtable map
variables (see cp/vtable-class-hierarchy.c). This tree pass should
run just before the gimple is converted to RTL.
Some implementation details for this pass:
To find all of the virtual calls, we iterate through all the
gimple statements in each basic block, looking for any call
statement with the code "OBJ_TYPE_REF". Once we have found the
virtual call, we need to find the vtable pointer through which the
call is being made, and the type of the object containing the
pointer (to find the appropriate vtable map variable). We then use
these to build a call to __VLTVerifyVtablePointer, passing the
vtable map variable, and the vtable pointer. We insert the
verification call just after the gimple statement that gets the
vtable pointer out of the object, and we update the next
statement to depend on the result returned from
__VLTVerifyVtablePointer (the vtable pointer value), to ensure
subsequent compiler phases don't remove or reorder the call (it's no
good to have the verification occur after the virtual call, for
example). To find the vtable pointer being used (and the type of
the object) we search backwards through the def_stmts chain from the
virtual call (see verify_bb_vtables for more details).
1.8.1.1