Indistinguishable from Jesse

Today I'm releasing two fuzzers: jsfunfuzz, which tests JavaScript engines, and DOMFuzz, which tests layout and DOM APIs.

Over the last 11 years, these fuzzers have found 6450 Firefox bugs, including 790 bugs that were rated as security-critical.

I had to keep these fuzzers private for a long time because of the frequency with which they found security holes in Firefox. But three things have changed that have tipped the balance toward openness.

First, each area of Firefox has been through many fuzz-fix cycles. So now I'm mostly finding regressions in the Nightly channel, and the severe ones are fixed well before they reach most Firefox users. Second, modern Firefox is much less fragile, thanks to architectural changes to areas that once oozed with fuzz bugs. Third, other security researchers have noticed my success and demonstrated that they can write similarly powerful fuzzers.

My fuzzers are no longer unique in their ability to find security bugs, but they are unusual in their ability to churn out reliable, reduced testcases. Each fuzzer alternates between randomly building a JS string and then evaling it. This construction makes it possible to make a reproduction file from the same generated strings. Furthermore, most DOMFuzz modules are designed so their functions will have the same effect even if other parts of the testcase are removed. As a result, a simple testcase reduction tool can reduce most testcases from 3000 lines to 3-10 lines, and I can usually finish reducing testcases in less than 15 minutes.

The ease of getting reduced testcases lets me afford to report less severe bugs. Occasionally, one of these turns out to be a security bug in disguise. But most importantly, these bug reports help me establish positive relationships with Firefox developers, by frequently saving them time.

A JavaScript engine developer can easily spend a day trying to figure out why a web site doesn't work in Firefox. If instead I can give them a simple testcase that shows an incorrect result with a new JS optimization enabled, they can quickly find the source of the bug and fix it. Similarly, they much prefer reliable assertion testcases over bug reports saying "sometimes, Google Maps crashes after a while".

As a result, instead of being hostile to fuzzing, Firefox developers actively help me fuzz their code. They've added numerous assertions to their code, allowing fuzzers to notice as soon as the smallest thing goes wrong. They've fixed most of the bugs that impede fuzzing progress. And several have suggested new ways to test their code, even (especially) ways that scare them.

Developers working on the JavaScript engine have been especially helpful. First, they ensured I could test their code directly, apart from the rest of the browser. They already had a JavaScript shell for running regression tests, and they added a --fuzzing-safe option to disable the more dangerous testing functions.

The JS team also created a large set of testing functions to let me control things that would normally be based on heuristics. Fuzzers can now choose when garbage collection happens and even how much. They can make expensive JITs kick in after 2 loop iterations rather than 100. Fuzzers can even simulate out-of-memory conditions. All of these things make it possible to create small, reliable testcases for nasty classes of bugs.

Finally, the JS team has supported differential testing, a form of fuzzing where output is checked for correctness against some oracle. In this case, the oracle is the same JavaScript engine with most of its optimizations disabled. By fixing inconsistencies quickly and supporting --enable-more-deterministic, they've ensured that differential testing doesn't get stuck finding the same problems repeatedly.

Please join us on IRC, or just dive in and contribute! Your suggestions and patches can have a large impact: fuzzer modules often act together to find complex interactions within the browser. For example, bug 893333 was found by my designMode module interacting with a <table> module contributed by a Firefox developer, Mats Palmgren. Likewise, bug 1158427 was found by Christoph Diehl's WebAudio module combined with my reflection-based API-discovery modules.

If your contributions result in me finding a security bug, and I think I wouldn't have found it otherwise, I'll make sure you get a bug bounty as if you had reported it yourself.

To the next 6450 browser bug fixes!

Fuzzers make things go wrong.
Assertions make sure we find out.

Assertions can improve code quality in many ways, but they truly shine when combined with fuzzing. Fuzzing is normally limited to finding obvious symptoms like crashes, because it's rare to be able to tell correct behavior from incorrect behavior when the input is generated randomly. Assertions expand the scope of fuzzing to include everything they check.

Assertions can even help find crash bugs: some bugs are relatively easy for fuzzers to trigger, but only lead to crashes when additional conditions are met. A well-placed assertion can let us know every time we trigger the bug.

Fuzzing JS and DOM has found about 4000 assertion bugs, including about 300 security bugs.

Asserting safe use of generic data structures

Assertions in widely-used data structures can find bugs in many callers.

• Array indices must be within bounds. This simple precondition assert in nsTArray has caught about 90 bugs.
• Hash tables must not be modified during enumeration. If the modification happened to resize the hash table, it would leave stack pointers dangling. This PLDHashTable assertion has caught over 50 bugs.
• Cached values should not be out of date. When a cache's get method takes a key and a closure for computing values in the case of a cache miss, debug builds can check whether the cached values are still correct. This is effectively a form of differential testing that notices bugs in cache-invalidation logic.

Asserting module invariants

When an entire module must maintain an invariant, a single assertion can catch dozens of bugs.

• Compartment mismatches. When a JS object in one page's compartment references an object in another, it must do so through a wrappers that enforces security policies. Without these assertions, we would have missed over 25 violations of Firefox's script security model.
• Phases of layout. These assertions have strings like "Should be in an update while creating frames" and "reflowing in the middle of frame construction". More phase and nesting assertions are wanted, but sometimes special cases like plugins get in the way.

Making the frame arena safer

Gecko's CSS box objects, called "frames", are created and destroyed manually. They are allocated within an arena to reduce malloc overhead and fragmentation. The arena also made it possible to reduce the risk associated with manual memory management. A combination of assertions (in debug builds) and runtime mitigations (in all builds) mitigates dangling pointer bugs that involve frames.

• When the arena is destroyed, debug builds assert that all objects in the arena were also destroyed. Over 60 bugs have been caught by the assertion. About half of the bugs that trigger the assertion can lead to exploitable crashes, but without a specially crafted testcase, they will not crash at all.
• While the arena is still alive, deleted frames are overwritten with a special poison pattern. If any code uses a pointer from a deleted frame, the browser will segfault safely. This mitigation, called frame poisoning, has prevented dozens of bugs from being exploitable.
• Writing over the poison trips another assertion. This assertion is actually more prone to catch hardware errors than software bugs, so it has been modified to help distinguish between the two.

Requests for Gecko developers

Please add assertions, especially when:

• A bug would be a security hole
• Crashing is not guaranteed
• Many callers must fulfill a precondition
• Complex, extensive code must maintain an invariant

Also consider:

• Ensure assertions in third-party libraries are enabled in debug builds of Firefox.
• Fix bugs requesting new assertions.
• Fix my assertion bugs to allow my fuzzers to find more.

I have mixed feelings about requiring Mozillians to “agree” to the Mozilla Manifesto. I get the impression that many volunteers aren’t fond of “commercial involvement” (9). Firefox development often does not live up to the ideals of absolute security (4) or transparency (8), so we’d be asking new contributors to commit to behavior for which they may have little support.

Meanwhile, the manifesto is oddly silent on two issues that many Mozillians care about deeply. First, it says little about privacy. “Shaping your own experience on the Internet” (5) suggests control over customized ads, but not control over tracking by advertisers or governments.

Second, the manifesto does not adequately address removing barriers to contribution or promoting inclusiveness in community processes. The relevant principles (6, 8) are worded as vague beliefs rather than strong values. Compare with my favorite part of the Ada Initiative FAQ:

“Open technology and culture are shaping the future of global society. If we want that society to be socially just and to serve the interests of all people, [all kinds of people] must be involved in its creation and organization.”

Rather than asking each Mozillian to agree to the entire manifesto, let’s instead encourage everyone to Likert the 10 existing principles and add a few of their own.

Indicating how you feel about each principle is more memorable than clicking “Agree” once. Each Mozillian would have a personal version of the Manifesto to remind them what drives them to contribute. Such a survey could also lead to better understanding of the community and suggest improvements to the Manifesto.

My DOM fuzzer can now find bugs where the layout of a DOM tree depends on its history.

In this example, forcing a re-layout swapped a “1” and “3” on the screen. My fuzzer didn’t know which rendering was correct, but it could tell that Firefox was being inconsistent.

Gecko developer Simon Montagu quickly determined that 13ت is the correct rendering and attached a patch. Later, when a user reported that the bug affected Persian comments on Facebook, we were able to backport Simon’s fix to Firefox 11.

How it works

The fuzzer starts by making random dynamic changes to a page. Then it compares two snapshots: one taken immediately after the dynamic changes, and another taken after also forcing a relayout.

To force a relayout, it removes the root from the document and then adds it back:

  var r = document.documentElement; 
  document.removeChild(r);
  document.appendChild(r);

Like reftest, it uses drawWindow() to take snapshots and compareCanvases() to compare them.

In theory, I could also look for bugs where dynamic changes do not repaint enough of the window. But I've been told that testing for painting invalidation bugs is tricky, so I'll wait until most of the layout bugs are fixed.

Exceptions

Since the testcases are random, I have to be heavy-handed in ignoring known bugs. If I file a rendering bug where the weirdest part of the testcase is floats, I'll have the fuzzer ignore inconsistent rendering in testcases with floats until the bug is fixed.

The current list of exceptions is fairly large and includes key web technologies:

• CSS border/padding (bug 718452)
• CSS position: relative/absolute (bug 728100)
• CSS float (bug 725928)
• Non-ASCII text (bug 726460)
• Right-to-left text (bug 730562)
• <table> (bug 467723)
• MathML (bug 522393)
• SVG (bug 723376, bug 475216)
• Anything that causes coordinate overflow
• Anything that causes assertion failures (which are tracked separately)

Last week, I asked Luke Wagner to explain some security bugs that he fixed in the past. I hoped to learn from each bug at multiple levels, in ways that could help prevent future security bugs from arising and persisting.

Luke is one of the developers working on Firefox's JavaScript engine, which is currently our largest source of critical security bugs.

Method

I imagined we would recurse in exhaustive breadth and exhausting depth. Instead, we recursed only on the most interesting items, and refined a checklist of starting points:

• What was the bug?
• What went wrong in the developer's thinking that caused the bug to be introduced?
• What made the bug exploitable?
• What caused us to use especially dangerous features of C++?
• Could a new abstraction make it possible to do this both fast and safe?
• What caused the bug to persist? Could we have caught this earlier with improved regression tests, fuzz testing, dynamic analysis, or static analysis?

Luke and I made trees for all ten bugs, at first on paper and later using EtherPad. Then I extracted and categorized what I thought were the most useful lessons and recommendations.

Recommendations for introducing fewer bugs

Casts

• Create centralized, type-restricted cast functions. This protects you when you change the representation of one of the types. It also protects against mistakes that cause the input type to be incorrect.

Sentinel values

• Use tagged unions instead.
• Use a typed wrapper (a struct containing a single value). When assigning from the underlying numeric type, convert using one of two functions: one that checks for special values, and one that explicitly does not.
• Audit existing code paths to ensure they cannot generate the special value.

Clarity of invariants

• Increase use of methods named AssertInvariants
• Create an alias for JS_ASSERTION called JS_INVARIANT.

Interacting with other developers

• If you're about to do something gross because someone else doesn't expose the right API/helper, maybe you should get it exposed.

JS Engine specific

• Any patch that touches rooting should be reviewed by Igor.
• Interpreter could have better abstraction and encapsulation for its stack.

Recommendations for catching bugs earlier

Static analysis

• Find all casts (C-style casts, the reinterpret_cast keyword, and casts through unions) for a given type. Could be used to enforce centralization or to find things that should be centralized.
• Be suspicious of a function with multiple return statements, all of which return the same primitive value.
• Be suspicious of a function returning true/success in an OOM path.

Dynamic analysis

• Ask Valgrind developers what they think of providing (in valgrind.h) a way to tie the addressability of "stacklike memory" to a variable that represents the end of the stack.

Fuzzing

• We should fuzz worker threads somehow.
  • In browser (slow and messy, but it's what users are running).
  • In thread-safe shell (--enable-threadsafe?), which has "toy workers".
• We should fuzz compartments better.
  • I should ask Blake and Andreas for help with testing compartments and wrappers.
  • I should ask Gary to run jsfunfuzz in xpcshell, where I can test both same-origin and different-origin compartments, and thus get more interesting wrappers.
• We should give JS OOM fuzzing another shot.

Next steps

I'm curious if others have additional ideas for what could have prevented the ten bugs we looked at. For example, someone like Jeff Walden, who loves to write exhaustive regression tests, might have ideas that Luke and I did not consider.

I'd also like to do this kind of analysis with a other developers on bugs they have fixed.

Not all of our neighbors followed us. Some asked — demanded? — that we send back supplies.

We acknowledged their request, but our immediate task was to explore this Isle of Rapidity. What surprises would we discover? What surprises would discover us?

To survive in this strange land, we would have to befriend new neighbors. Living for so long atop Mount Annum, we had almost forgotten how to introduce ourselves.

But we had brought much to share. We had barely opened our packs when the wind seemed to whisper:

Here, gifts arrive almost before you send them.

Maybe it wouldn’t be so hard to make friends here.

And there was something inexplicably familiar about this island. Was it the scent of the flowers? The rhythmic waves in the distance? The chattering of wildlife, almost a chorus?

Here, a gift to your neighbor is equally a gift to yourself.

We felt a sudden shift in perception: the Isle of Rapidity was home.

Our friends thought us mad.

We had thrived atop Mount Annum. It was the highest peak as far as the eye could see.

Once, we had sprained an ankle on the Foothills of Many Betas in the west. We remembered miserable visits to the Bog of Eternal Driver Approval in the east. Every time, we had quickly returned home.

But there we were, on the summit of Mount Annum, climbing into a cannon aimed at 4°N, 6°W.

So far away, and yet oddly specific. We lit the fuse and plugged our ears.

We flew over the stifling Bog of Eternal Driver Approval. We flew over the perilous Sea of Recklessness.

We landed, as we hoped, on the uncharted Isle of Rapidity. The landing was unexpectedly soft.

We’ve shaken off the dust and gunpowder. We’ve begun to tend our wounds. We’re excited about the upcoming climb.

In the months before Firefox's first rapid release, one concern echoed throughout engineering: crashes.

We had always relied on long stabilization periods to get crash rates down. Firefox 4 would be our last high-stability release. We hoped improvements on other aspects of quality would outweigh the decreased stability.

But then something surprising happened. We released Firefox 5, and Firefox didn’t get crashier.

KaiRo’s explanation parallels what I’ve seen helping with MemShrink:

• The channel cascade gives each release 12 weeks of pure stabilization.
• The channel audiences help by comparing alphas to alphas.
• The short cycles enable backouts and reduce the desire to land half-baked features.

“Rapid release” doesn't mean building Firefox the way we always have, x times faster. It’s a new process that fits together in beautiful yet fragile ways.