[LLVMdev] DataFlowSanitizer design discussion

[Peter Collingbourne]

On Thu, Jun 13, 2013 at 03:00:46PM -0700, Peter Collingbourne wrote:
> Hi,
> 
> I am starting a thread to discuss the design of DataFlowSanitizer,
> a compiler instrumentation based analysis tool which I am hoping to
> bring into LLVM.  As a starting point, I have included the current
> version of the design document below.  Comments are appreciated.

Any further comments on the below?  I've updated the design document
to add a use case at Kostya's request, but I'd appreciate any further
review of the design.

Thanks,
Peter


DataFlowSanitizer Design Document
*********************************

This document sets out the design for DataFlowSanitizer, a general
dynamic data flow analysis.  Unlike other Sanitizer tools, this tool
is not designed to detect a specific class of bugs on its own.
Instead, it provides a generic dynamic data flow analysis framework to
be used by clients to help detect application-specific issues within
their own code.

DataFlowSanitizer is a program instrumentation which can associate a
number of taint labels with any data stored in any memory region
accessible by the program. The analysis is dynamic, which means that
it operates on a running program, and tracks how the labels propagate
through that program. The tool shall support a large (>100) number of
labels, such that programs which operate on large numbers of data
items may be analysed with each data item being tracked separately.


Use Cases
=========

This instrumentation can be used as a tool to help monitor how data
flows from a program's inputs (sources) to its outputs (sinks). This
has applications from a privacy/security perspective in that one can
audit how a sensitive data item is used within a program and ensure it
isn't exiting the program anywhere it shouldn't be.


Interface
=========

A number of functions are provided which will create taint labels,
attach labels to memory regions and extract the set of labels
associated with a specific memory region. These functions are declared
in the header file "sanitizer/dfsan_interface.h".

   /// Creates and returns a base label with the given description and user data.
   dfsan_label dfsan_create_label(const char *desc, void *userdata);

   /// Sets the label for each address in [addr,addr+size) to \c label.
   void dfsan_set_label(dfsan_label label, void *addr, size_t size);

   /// Sets the label for each address in [addr,addr+size) to the union of the
   /// current label for that address and \c label.
   void dfsan_add_label(dfsan_label label, void *addr, size_t size);

   /// Retrieves the label associated with the given data.
   ///
   /// The type of 'data' is arbitrary.  The function accepts a value of any type,
   /// which can be truncated or extended (implicitly or explicitly) as necessary.
   /// The truncation/extension operations will preserve the label of the original
   /// value.
   dfsan_label dfsan_get_label(long data);

   /// Retrieves a pointer to the dfsan_label_info struct for the given label.
   const struct dfsan_label_info *dfsan_get_label_info(dfsan_label label);

   /// Returns whether the given label label contains the label elem.
   int dfsan_has_label(dfsan_label label, dfsan_label elem);

   /// If the given label label contains a label with the description desc, returns
   /// that label, else returns 0.
   dfsan_label dfsan_has_label_with_desc(dfsan_label label, const char *desc);


Taint label representation
==========================

As stated above, the tool must track a large number of taint labels.
This poses an implementation challenge, as most multiple-label
tainting systems assign one label per bit to shadow storage, and union
taint labels using a bitwise or operation. This will not scale to
clients which use hundreds or thousands of taint labels, as the label
union operation becomes O(n) in the number of supported labels, and
data associated with it will quickly dominate the live variable set,
causing register spills and hampering performance.

Instead, a low overhead approach is proposed which is best-case
O(log_2 n) during execution. The underlying assumption is that the
required space of label unions is sparse, which is a reasonable
assumption to make given that we are optimizing for the case where
applications mostly copy data from one place to another, without often
invoking the need for an actual union operation. The representation of
a taint label is a 16-bit integer, and new labels are allocated
sequentially from a pool. The label identifier 0 is special, and means
that the data item is unlabelled.

When a label union operation is requested at a join point (any
arithmetic or logical operation with two or more operands, such as
addition), the code checks whether a union is required, whether the
same union has been requested before, and whether one union label
subsumes the other. If so, it returns the previously allocated union
label. If not, it allocates a new union label from the same pool used
for new labels.

Specifically, the instrumentation pass will insert code like this to
decide the union label "lu" for a pair of labels "l1" and "l2":

   if (l1 == l2)
     lu = l1;
   else
     lu = __dfsan_union(l1, l2);

The equality comparison is outlined, to provide an early exit in the
common cases where the program is processing unlabelled data, or where
the two data items have the same label.  "__dfsan_union" is a runtime
library function which performs all other union computation.

Further optimizations are possible, for example if "l1" is known at
compile time to be zero (e.g. it is derived from a constant), "l2" can
be used for "lu", and vice versa.


Memory layout and label management
==================================

The following is the current memory layout for Linux/x86_64:

+-----------------+-----------------+----------------------+
| Start           | End             | Use                  |
+=================+=================+======================+
| 0x700000008000  | 0x800000000000  | application memory   |
+-----------------+-----------------+----------------------+
| 0x200200000000  | 0x700000008000  | unused               |
+-----------------+-----------------+----------------------+
| 0x200000000000  | 0x200200000000  | union table          |
+-----------------+-----------------+----------------------+
| 0x000000010000  | 0x200000000000  | shadow memory        |
+-----------------+-----------------+----------------------+
| 0x000000000000  | 0x000000010000  | reserved by kernel   |
+-----------------+-----------------+----------------------+

Each byte of application memory corresponds to two bytes of shadow
memory, which are used to store its taint label. As for LLVM SSA
registers, we have not found it necessary to associate a label with
each byte or bit of data, as some other tools do. Instead, labels are
associated directly with registers.  Loads will result in a union of
all shadow labels corresponding to bytes loaded (which most of the
time will be short circuited by the initial comparison) and stores
will result in a copy of the label to the shadow of all bytes stored
to.

-- 
Peter