[llvm-dev] Contributing a new sanitizer for pointer casts

[Stephen Kell via llvm-dev]

> What it would take to get to a point where there are no crunch "FPs"
> by design? E.g given the following code:
> 
>   A *a = malloc(sizeof(A));
>   UnrelatedToA *uta = (UnrelatedToA *)a;
>   LikeA *la = (LikeA *)uta;
>   deref(la);
> 
> Could the diagnostic for the assignment to "uta" be suppressed, since
> "uta" isn't dereferenced?

Yes, it could. I'm sticking my neck out slightly since this isn't
currently implemented. But what should happen is

- uta is trapped as "invalid / do-not-use" in the normal __is_a() cast
  check

- on the cast to LikeA*, the (de-trapped) pointer is re-checked against
  the underlying object. Assuming "LikeA" is in the list of types for
  which __like_a() has been turned on, this would pass and yield a
  valid pointer.

How to get LikeA into that list automatically is a matter for thought,
and probably need a fuller example to work through.

If we're talking about structs, code that needs __like_a() is almost
always code that requires -fno-strict-aliasing.  (An exception is
sockaddr-like types with padding in.)

The other relaxation of __like_a() is for signed/unsigned sloppiness,
which is more common. That could probably be turned on by default; I
just personally like being stricter.

For completeness, I should say the other thing that messes with
libcrunch is casting to a completely unrelated aggregate type (i.e.
wouldn't satisfy __like_a()) just for the sake of using some particular
member to do an access. So if I have 

struct S { int x; ... };

some code might cast any old int* to S*, if it only wants to do ->x on
the pointer. With a lot of gymnastics I could support this, either by
rewriting certain field accesses as casts or by some trap-side logic. I
don't know any code that does this (though I'd love to be pointed at
some) so it hasn't seemed worth supporting so far.

> > I take your point about the danger that "people would turn it off".
> > As an aside -- more below -- one thing I've worked quite hard on is
> > to ensure that you don't need to rebuild to turn it off, which at
> > least prevents a hard exit.
> 
> How does the ability to disable sanitization without rebuilding help
> prevent hard exits?

My poor wording; I meant "hard exit" of the user, i.e. it's easy for
the user to turn it back on later, without rebuilding. There's not much
performance drawback to building with the checks in, if you don't
preload the runtime.

Though also, I forgot to say that failed checks in libcrunch don't
(unless you want them to) cause termination of the program. That is
another thing that makes occasional false positives easier to deal with
(-- not that I believe there need to be any false positives, but
still...).

> > Beyond that, there is quite a long tail of mitigation measures
> > against false positives, including some not mentioned in the paper
> > (and, er, some not implemented). The big one, and one that isn't in
> > the paper, is "trap pointers" -- I outlined this in response to Hal
> > in the TBAA sanitizer thread. In short, it delays warnings from the
> > time of cast to the time of use.
> 
> Oh, neat! Would it handle the example above?

Yes, modulo the complexity of turning on __like_a() automagically.

> > I only implemented this later, to support bounds
> > checking rather than cast checking, but it applies to the latter
> > too. (Sneak preview: I'm working on bounds checker I believe will
> > offer a different, possibly better, precision/performance trade-off
> > than ASan, but that's still in progress.) 
> > 
> > Another possible mitigation is to use "__like_a" by default in some
> > circumstances, perhaps helped by an up-front static analysis pass to
> > find things like structure prefixing. That's pure idea-stuff right
> > now.
> 
> I'm not sure what it would take to prototype and evaluate this, but
> it sounds like a promising mitigation.

Indeed. From my perspective, the blocker to developing and evaluating
this stuff is it's already assuming uncommon code, and often the
properties are not syntactically obvious... so getting a good set of
evaluation cases together is like gathering needles from a haystack.
Maybe it could be useful to collect a corpus of "hard to sanitize" code
(e.g. perl would count :-). Alternatively (and less convincing but...)
there are already synthetic tests, like the Cerberus de-facto tests,
and maybe older things like C-torture. I have run libcrunch on the
Cerberus tests before and from an initial scan did not see any
surprises (i.e. it complains at the things I expected) although it has
acquired more tests since then, and I should reserve final judgement
til I go through more carefully.

> > Doing a build-world of FreeBSD or similar, as you suggest, is
> > exactly the sort of experience that is required, and inevitably, in
> > doing so there will be a lot of handle-turning to fix bugs and the
> > like. I'd rather go through this on an implementation that has a
> > chance of getting adopted... so I think the LLVM port has to happen
> > first, though I understand that the contribution per se may have to
> > wait.
> > 
> >> I've certainly seen code at Apple which would fail
> >> libcrunch's "is-a" checks, but which "works fine" according to the
> >> code owners. Would you be open to reducing the number of cases
> >> libcrunch is able to diagnose in order to reduce FP rates?
> > 
> > I'm curious -- can you give me a sketch of what this code is doing?
> 
> It's the example I provided above. It might also take some work to
> teach crunch about type relationships in ObjC, or at least to make
> crunch ignore ObjC types.

Ah yes. I don't know Objective-C well enough to say what's best there.

> > If it's just a case of "creating wrong-typed pointers that aren't
> > deref'd", then I think we're good. In general I'd rather fix the FPs
> > precisely, rather than introduce underapproximation, and I've yet to
> > see cases where that really can't be done.
> 
> Makes sense.
> 
> 
> >> Second, do you see any way to simplify the wrappers required to
> >> invoke libcrunch? E.g is the runtime preload step necessary? Apart
> >> from the runtime shared object, can the other tools required to use
> >> libcrunch be consolidated into the clang binary?
> > 
> > Great question. The plan is to consolidate the tools, yes -- but
> > probably mostly into the gold plugin.
> 
> Would it be feasible to introduce a clang tool that runs after -cc1
> to do some of the heavy lifting? This might be more portable than a
> linker plugin.

There is quite a bit of stuff that is movable, so yes, such a tool
could take over some of the work. I suppose you could even generate
type information on a per-CU basis, at the cost of generating more
comdat linker input... now I think about it, doing the dedup there
may be faster. 

The main things that have to happen at link time are wrapping of
allocator functions (each is basically linked with --wrap, together
with some generated code) and emitting DSOs with the type metadata. I
suppose the "separate metadata objects" model could be ditched
entirely, though it is quite useful to be able to generate metadata for
binaries you didn't build yourself (e.g. libc).

> > Only a couple of steps work on the individual compilation units,
> > and they can be deferred until link time anyhow. It does mean that
> > each compilation unit's allocation site information has to get
> > dumped somewhere, to be picked up at link time.
> 
> Sorry for the naive question, but could you explain how the
> allocation site information is generated? Is this something that
> could be deferred until link time?

No problem. Logically, the allocation site analysis is a data flow
analysis finding where "sizeof"-derived values reach the "size" argument
positions of allocation calls. That's why Chris (Diamand) had to
come up with the hack that preserves sizeof into the LLVM IR (via dummy
function calls).

This information comes out in a simple bunch of tuples (file, line,
type). At link time, the debugging information is used to extend these
tuples with the actual address of the allocation call. This goes into
the loadable metadata object and is used by the runtime.

So you can defer it only as long as you have "sizeof" around... of
course if you don't mind re-reading the source, you can even do it
way after build time, and that is sometimes useful (again, thinking
libc).

A radical approach might be to give up classifying allocation sites
altogether, and just say "the first cast determines the type". I am
uncomfortable with this, but it would simplify things.

(I'm uncomfortable because I feel that allocation sites have meaning,
from a debugging and comprehension point of view -- not limited to
their type info, but thinking e.g. about their use as alias analysis
classes -- so there is value in reifying them at debug/sanitize time
and having allocation site info available for diagnostics/inspection.)

> > This is currently in a .i.allocs text file, but that is a bit
> > untidy... I've thought about a DWARF extension
> > (DW_TAG_allocation_site or similar) instead, which would be cleaner.
> > 
> > About preloading -- I happen to like this design, modulo the
> > clunky invocation, because it means you can easily turn the
> > checking on or off. For executables, you could simply link the
> > preload stuff in, but that would lose the on/off switch, and
> > doesn't work for shared libraries. Another idea I've had is to use
> > a customised loader (ld.so) instead of a preload library. So you
> > would do
> > 
> > $ ./myprogram   # runs normally
> > 
> > or 
> > 
> > $ /path/to/crunchld.so ./myprogram    # runs +crunch
> > 
> > This would more-or-less replicate the usage model of Valgrind tools.
> > (In general, my feeling is that extending the ld.so is more sane
> > than using LD_PRELOAD, although I realise it will seem a bit
> > off-the-wall.)
> 
> I have a preference for building separately when enabling sanitizers.
> I feel this way because I don't have reasons to disable sanitization
> once I've enabled it in one build tree, and because the performance
> of the sanitized binary is impacted. Having this as the supported
> workflow may be more convenient if it makes it possible to skip the
> preload step.

That's fair -- I can see both ways. My preference has been to minimise
the "reasons to disable" in the first place. In this case, the slowdown
is usually minimal if you don't load the library, so it's job mostly
done. That model doesn't work for all sanitizers, of course.

(Still, long-term, I feel my job is only really done when deployed
binaries have the checks compiled in -- at least outside of HPC-style
workloads.) 

> >> Third, how do you plan on keeping the libcrunch integration into
> >> clang in-sync with the runtime, if the runtime is to live outside
> >> of the llvm tree? 
> > 
> > I think that part is no problem. The interface between instrumented
> > code and the runtime is pretty narrow (basically just the __is_a(),
> > __like_a() and similar functions) and does not embed any LLVM
> > stuff, so it would be stable under LLVM churn. The same would hold
> > for the gold plugin... LLVM/clang's involvement is mostly just to
> > do the instrumentation, which is a pretty well-isolated task. The
> > allocation site metadata I just mentioned is probably the only
> > change-prone thing.
> > 
> > Of course, there is a problem of breakage/churn in the
> > libcrunch/liballocs stuff... in particular, things like new DWARF
> > features can break things. I have no good answer to that, except
> > that more hands make light(er) work.
> > 
> >> How will the feature be tested? 
> > 
> > On that, I have no idea, and I'd have to take your lead on that. To
> > make an analogy with another out-of-tree dependency, how do you test
> > LTO against GNU binutils? I'm hoping, perhaps naively, that it could
> > work similarly.
> 
> There are some in-tree tests which use libLTO.dylib and ld64. The
> tests depend on those tools being available. For end-to-end testing
> of crunch, you could do something similar. The instrumentation and
> driver changes could be tested in isolation.
> 
> My preference would be to have the sanitizer runtime live in
> compiler-rt, which is where end-to-end testing of the other
> sanitizers currently happens.

I can see why that's desirable. And it could ultimately be achieved, no
doubt, although it feels like more work than I can reasonably get done
myself, at least on this occasion.

I should add that the libcrunch runtime is not really just that; it's
mostly liballocs, which is designed as a general-purpose way to 
add run-time type information to a Unix-like process -- with
applications beyond sanitizer-like tools. I see it as a run-time
service in its own right, not "just" an adjunct to a compiler-based
tool. Some of the funkier experimental applications of liballocs have
been on avoiding FFIs (in a V8 extension), adding type info to
memory-mapped files, and (internally) doing JIT-like dynamic code
generation in a way that is transparent to the dynamic linker.

Conversely, if one focuses only on this one use case, it would
probably be straightforward-ish to reimplement a stripped-down
replacement specialised to type checking -- especially if you don't
mind forgoing things like nested/custom allocators, resuming after
using a trap pointer (some hairy instruction-decoding gymnastics using
a lot of borrowed GPL'd code) or really high coverage of the address
space (e.g. liballocs can answer queries on pointers pointing into the
auxiliary vector, or sometimes into ld.so/libc-internal structures,
etc.). This replacement would be a better fit for compiler-rt. I'm
still wary of saying I would have the resource to implement it....

> >> Currently, the
> >> sanitizer runtimes ship with official LLVM downloads: what would
> >> the deployment story for libcrunch be (how would users get it)?
> > 
> > Initially they would build the libcrunch tree, which would build for
> > them the gold plugin and runtime. Again, I think it would be rather
> > like the LTO story (a web page with "follow these instructions").
> > And of course distributors could build binary packages... again, it
> > helps that the interface between compiler and plugin/runtime is
> > fairly simple.
> > 
> > (Full disclosure: currently you have to build both the liballocs
> > tree and the libcrunch tree, separately. Eventually the former will
> > be a submodule of the latter. I have too many levels of submodule
> > going on....)
> 
> IMO a model where all of the sanitizer components live in-tree would
> make testing and deploying the sanitizer easier. It may also make
> maintenance easier, e.g if it's ever necessary to change the
> libcrunch <-> clang interface.
> 
> I know you mentioned up front that you don't intend to port the
> runtime to compiler-rt, as this would require a major rewrite.
> However, I think that would be the right way to go, and I'd be
> willing to help out with reviews etc.

I can understand this preference. Thanks for being willing to help on
your side. 

> I'm curious to hear what others think about having an out-of-tree
> sanitizer runtime. I'm not sure that the situation with LTO and gold
> is precedent for this, but others may feel differently.

(Of course it's different, but I'd argue in the same spirit as e.g.
using the host system linker until the LLVM project found the
time/resources to implement a replacement.)

Stephen