[llvm-dev] llvm-dev Digest, Vol 136, Issue 22
Adve, Vikram Sadanand via llvm-dev
llvm-dev at lists.llvm.org
Fri Oct 9 09:45:10 PDT 2015
(Note to self: learn to scan the full digest for later messages in a thread before replying to an earlier message.)
Your reply to John answered some of my questions, but not all, and raised a new one:
> Maybe I should have been a bit clearer; we're really interested in full
> memory and type safety. We want to harden the system against memory
> corruption vulnerabilities. Process isolation isn't an issue, as we are
> in an embedded context where we don't have processes.
Do you also care about use-after-free and free-after-free errors? Handling free-after-free is not too hard, but general use-after-free can be expensive, unless you’re willing to spend a fair amount of extra memory. If dynamic allocation isn’t an issue for your (embedded) applications, that makes this whole problem go away, of course.
This brings me to the new question: exactly what unsafe features do you need to be concerned about? Our first paper in the SAFECode project focused on identifying a subset of C for which type and memory safety could be enforced without *any* run-time checks or GC. We found that array bounds checks were the biggest problem; ignoring these, we can ensure the safety of pointer and dynamic memory usage in all the embedded benchmarks we tried without any run-time checks. The paper appeared at LCTES:
Like the later work, this did not eliminate use-after-free but used APA to ensure that any use-after-free errors are “harmless” in that they did not violate all the other type- and memory-safety guarantees.
Let me know if you have any questions about this.
// Vikram S. Adve
// Professor, Department of Computer Science
// University of Illinois at Urbana-Champaign
// vadve at illinois.edu
// http://llvm.org <http://llvm.org/>
// "A ‘No’ uttered from the deepest conviction is better than
// a ‘Yes’ merely uttered to please, or worse, to avoid trouble."
// --Mahatma Gandhi
On 10/9/15, 1:05 AM, "llvm-dev on behalf of via llvm-dev" <llvm-dev-bounces at lists.llvm.org on behalf of llvm-dev at lists.llvm.org> wrote:
>Date: Thu, 08 Oct 2015 23:50:33 +0100
>From: Ed Robbins via llvm-dev <llvm-dev at lists.llvm.org>
>To: John Criswell <jtcriswel at gmail.com>, llvm-dev at lists.llvm.org
>Subject: Re: [llvm-dev] Pool allocator + safecode
> <1444344633.2719095.405296865.10B20F64 at webmail.messagingengine.com>
>Thanks again for the really nice response John.
>On Thu, Oct 8, 2015, at 10:14 PM, John Criswell wrote:
>>The original llvm-gcc (in which we modified gcc to produce LLVM bitcode)
>>no longer exists. When GCC was extended to use plugins, someone created
>>the DragonEgg plugin. I am not sure if that still exists.
>>Is there a reason you need a GCC-based front-end? Clang is intended to
>>be a drop-in replacement for GCC and supports many of the GCC extensions.
>The only reason was to make it easier to drop into the toolchain. I
>don't anticipate that being a huge amount of work to switch to clang
>anyway, but I saw that SafeCode was originally using llvm-gcc so I
>thought that might make life easy.
>> Integrating into Clang simply makes it easier to run
>>the SAFECode transformation passes after the LLVM passes have been run.
>So basically it means you don't have to run safecode manually?
>>DSA is a points-to and call graph analysis. As such, it is used for
>>many things in addition to APA and SAFECode. For example, the SMACK
>>verifier uses DSA.
>>The APA code has been bit rotting for awhile. While you can compile and
>>run it with LLVM 3.2, it won't work as well as it did for the paper, and
>>it probably needs some work to make it robust. Please remember that APA
>>has always been a research prototype.
>I really want to bring all this into trunk. I've spent this afternoon
>chugging through minor API incompatibilities trying to get APA to build.
>Not there yet!
>>> We really want the all singing all dancing safecode framework with APA
>>> as detailed in the 2005 TECS SafeCode paper (Memory Safety Without
>>> Garbage Collection for Embedded Applications). We are trying to build a
>>> C based embedded system that is type safe at the lowest possible run
>>> time cost. So I am also going to modify the uninitialized pointer MMU
>>> based stuff to work with the ARM Cortex M3 MPU. I don't think there are
>>> any shortcuts here (I'd be happy to be proved wrong though) - we need
>>If you just need to isolate processes on your system so that they don't
>>overwrite each other, it looks like the ARM Cortex M3 MPU can do that.
>>Is there a reason why that won't work?
>Maybe I should have been a bit clearer; we're really interested in full
>memory and type safety. We want to harden the system against memory
>corruption vulnerabilities. Process isolation isn't an issue, as we are
>in an embedded context where we don't have processes. I was really
>talking about setting uninitialised pointers to a value that is
>configured in the MPU to be inaccessible, and using a handler to abort
>in the case where an attempt is made to read/write to any of these
>>However, if you want to isolate processes using inferred type safety,
>>you will need to address three significant challenges:
>>1) Due to changes in LLVM, DSA has a difficult time inferring types. In
>>a nutshell, some LLVM optimizations turned typed-getelementptr (GEP)
>>instructions into GEPs that use byte-level indexing. DSA was written
>>with the assumption that GEPs carried high-level type information.
>>I think this is fixable by refactoring DSA's type inference code to act
>>more like Value Set Analysis: it would use a map between a 4-tuple and a
>>type. The 4-tuple (a, b, c, d) describes a formula ax + b in which a
>>and b denote offset and stride and c and d place a limit on the lower
>>and upper bounds of x. c and d can be +- infinity, allowing the 4-tuple
>>to denote a type that occurs in an unbounded array.
>This sounds like a good idea. I'm willing to give this a go, time
>>As an FYI, there are multiple research groups interested in accurate
>>points-to analysis results (mine included). I'll be holding a BoF at
>>the LLVM Developer's meeting this year to discuss who needs what and if
>>there's a way to develop it.
>That doesn't surprise me. If someone would like to do this it would be
>very handy from my point of view too!
>>2) Making APA robust. In my personal opinion, the original APA may be
>>more complicated than necessary for your application. APA currently
>>creates context sensitive pools, which means that it needs to pass pool
>>handles around. This requires transforming function signatures which
>>makes inter-operating with external code a real pain. It also makes the
>>A simpler design would be to create one pool for each type that DSA
>>infers. The points-to analysis would not be sound (unless it unified
>>all points-to sets that have the same type), but it would allow all
>>pools (or nearly all pools) to be global variables, greatly simplifying
>>the transformation. It would also continue to prevent the application
>>from accessing data outside its allocated memory.
>We aren't going to need to inter-operate with external code all that
>much (if at all), so perhaps this will not be an issue. DSA and the
>context sensitive pools of APA (despite the ugly transform) were
>originally what attracted us to SafeCode; we would like to remain sound.
>>3) The version of SAFECode in that paper used the Omega constraint
>>solver to prove that array accesses are safe. That code has long
>>bit-rotted away, and its implementation was not the most efficient (it
>>exec()'ed the Omega solver for every query). A better approach today
>>would be to integrate the constraint solver into the compiler proper.
>>Additionally, you now have other tools available, such as CVC4, Z3, and
>>SMACK/Boogie, for building and solving the constraints.
>We had already noticed that the limit to linear array access was way
>behind state of the art (we have a lot of experience with constraint
>solvers). Using an SMT solver would make sense, as you can use the
>theory that suits the particular problem - linear isn't good enough?
>You've got IDL, octagons etc.
>To start off with we'd like to get something running, even if it has
>limitations. Unless you think it's a really bad idea I will for the
>moment try to get the old APA working to some extent again. It isn't
>really an issue if I have to run APA and safecode as separate parts of
>the build process for the moment, or if parts of it are clunky, from our
>point of view it would just be good just to get some toy examples
>working relatively quickly.
>>As an FYI, later versions of SAFECode use run-time checks for
>>type-unsafe pools and array accesses so that it can support the full
>>generality of C (see Dhurjati's PLDI 2007 paper and my SVA
>>publications). You could probably do something simple in which
>>type-safe data goes into pools and type-unsafe data goes into a large
>>area for which loads and stores are subjected to simple SFI-like
>>instrumentation (e.g., Google Native Client). There may be solutions in
>>the middle of the spectrum between fast-but-difficult-to-implement and
>We want speed speed speed! If we only support a (preferably large)
>subset of C it doesn't matter - this is better than having additional
>run time checks. Is there an option to warn whenever a run-time check is
>inserted? I'll check out some of the other papers. I read the TECS paper
>and gave a few of the others a skim, but looks like I missed some stuff.
More information about the llvm-dev