Google On-line Safety Weblog: Use-after-freedom: MiraclePtr

Reminiscence security bugs are essentially the most quite a few class of Chrome safety points and we’re persevering with to examine many options – each in C++ and in new programming languages. The most typical kind of reminiscence security bug is the “use-after-free”. We not too long ago posted about an thrilling sequence of applied sciences designed to forestall these. These applied sciences (collectively, *Scan, pronounced “star scan”) are very {powerful} however probably require {hardware} help for adequate efficiency.

Right now we’re going to speak a couple of totally different method to fixing the identical kind of bugs.

It’s arduous, if not unimaginable, to keep away from use-after-frees in a non-trivial codebase. It’s hardly ever a mistake by a single programmer. As a substitute, one programmer makes affordable assumptions about how a little bit of code will work, then a later change invalidates these assumptions. All of the sudden, the information isn’t legitimate so long as the unique programmer anticipated, and an exploitable bug outcomes.

These bugs have actual penalties. For instance, based on Google Menace Evaluation Group, a use-after-free within the ChromeHTML engine was exploited this 12 months by North Korea.

Half of the identified exploitable bugs in Chrome are use-after-frees:

Diving Deeper: Not All Use-After-Free Bugs Are Equal

Chrome has a multi-process structure, partly to make sure that internet content material is remoted right into a sandboxed “renderer” course of the place little hurt can happen. An attacker due to this fact normally wants to search out and exploit two vulnerabilities – one to attain code execution within the renderer course of, and one other bug to interrupt out of the sandbox.

The primary stage is commonly the simpler one. The attacker has numerous affect within the renderer course of. It’s straightforward to rearrange reminiscence in a selected approach, and the renderer course of acts upon many alternative sorts of internet content material, giving a big “assault floor” that might doubtlessly be exploited.

The second stage, escaping the renderer sandbox, is trickier. Attackers have two choices how to do that:

1. They’ll exploit a bug within the underlying working system (OS) via the restricted interfaces obtainable inside Chrome’s sandbox.
2. Or, they’ll exploit a bug in a extra {powerful}, privileged a part of Chrome – just like the “browser” course of. This course of coordinates all the opposite bits of Chrome, so essentially has to be omnipotent.

We think about the attackers squeezing via the slender a part of a funnel:

If we will scale back the scale of the slender a part of the funnel, we’ll make it as arduous as potential for attackers to assemble a full exploit chain. We are able to scale back the scale of the orange slice by eradicating entry to extra OS interfaces inside the renderer course of sandbox, and we’re repeatedly engaged on that. The MiraclePtr mission goals to cut back the scale of the blue slice.

Right here’s a pattern of 100 current excessive severity Chrome safety bugs that made it to the steady channel, divided by root trigger and by the method they have an effect on.

You would possibly discover:

• This doesn’t fairly add as much as 100 – that’s as a result of a couple of bugs had been in different processes past the renderer or browser.
• We claimed that the browser course of is the harder half to take advantage of, but there are extra potentially-exploitable bugs! Which may be so, however we imagine they’re usually more durable to take advantage of as a result of the attacker has much less management over reminiscence format.

As you may see, the largest class of bugs in every course of is: V8 within the renderer course of (JavaScript engine logic bugs – work in progress) and use-after-free bugs within the browser course of. If we will make that “skinny” bit thinner nonetheless by eradicating a few of these use-after-free bugs, we make the entire job of Chrome exploitation markedly more durable.

MiraclePtr: Stopping Exploitation of Use-After-Free Bugs

That is the place MiraclePtr is available in. It’s a expertise to forestall exploitation of use-after-free bugs. Not like aforementioned *Scan applied sciences that provide a non-invasive method to this downside, MiraclePtr depends on rewriting the codebase to make use of a brand new sensible pointer kind, raw_ptr<T>. There are a number of methods to implement MiraclePtr. We got here up with ~10 algorithms and in contrast the professionals and cons. After analyzing their efficiency overhead, reminiscence overhead, safety safety ensures, developer ergonomics, and so on., we concluded that BackupRefPtr was essentially the most promising resolution.

The BackupRefPtr algorithm is predicated on reference counting. It makes use of help of Chrome’s personal heap allocator, PartitionAlloc, which carves out just a little additional area for a hidden reference rely for every allocation. raw_ptr<T> increments or decrements the reference rely when it’s constructed, destroyed or modified. When the appliance calls free/delete and the reference rely is larger than 0, PartitionAlloc quarantines that reminiscence area as a substitute of instantly releasing it. The reminiscence area is then solely made obtainable for reuse as soon as the reference rely reaches 0. Quarantined reminiscence is poisoned to additional scale back the chance that use-after-free accesses will lead to exploitable circumstances, and in hope that future accesses result in an easy-to-debug crash, turning these safety points into less-dangerous ones.

class A { ... };
class B {
  B(A* a) : a_(a) {}
  void doSomething() { a_->doSomething(); }
  raw_ptr<A> a_;  // MiraclePtr
};

std::unique_ptr<A> a = std::make_unique<A>();
std::unique_ptr<B> b = std::make_unique<B>(a.get());
[…]
a = nullptr;  // The free is delayed as a result of the MiraclePtr continues to be pointing to the item.
b->doSomething();  // Use-after-free is neutralized.

We efficiently rewrote greater than 15,000 uncooked pointers within the Chrome codebase into raw_ptr<T>, then enabled BackupRefPtr for the browser course of on Home windows and Android (each 64 bit and 32 bit) in Chrome 102 Secure. We anticipate that MiraclePtr meaningfully reduces the browser course of assault floor of Chrome by defending ~50% of use-after-free points towards exploitation. We are actually engaged on enabling BackupRefPtr within the community, utility and GPU processes, and for different platforms. In the long run state, our aim is to allow BackupRefPtr on all platforms as a result of that ensures {that a} given pointer is protected for all customers of Chrome.

Balancing Safety and Efficiency

There is no such thing as a free lunch, nonetheless. This safety safety comes at a price, which we’ve rigorously weighed in our resolution making.

Unsurprisingly, the principle price is reminiscence. Fortunately, associated investments into PartitionAlloc over the previous 12 months led to 10-25% complete reminiscence financial savings, relying on utilization patterns and platforms. So we had been capable of spend a few of these financial savings on safety: MiraclePtr elevated the reminiscence utilization of the browser course of 4.5-6.5% on Home windows and three.5-5% on Android¹, nonetheless nicely under their earlier ranges. Whereas we had been apprehensive about quarantined reminiscence, in observe this can be a tiny fraction (0.01%) of the browser course of utilization. By far the larger wrongdoer is the extra reminiscence wanted to retailer the reference rely. One would possibly suppose that including 4 bytes to every allocation wouldn’t be an enormous deal. Nevertheless, there are a lot of small allocations in Chrome, so even the 4B overhead will not be negligible. PartitionAlloc additionally makes use of pre-defined bucket sizes, so this additional 4B pushes sure allocations (notably power-of-2 sized) into a bigger bucket, e.g. 4096B->5120B.

We additionally thought-about the efficiency price. Including an atomic increment/decrement on frequent operations reminiscent of pointer project has unavoidable overhead. Having excluded numerous performance-critical pointers, we drove this overhead down till we may achieve again the identical margin via different efficiency optimizations. On Home windows, no statistically important efficiency regressions had been noticed on most of our top-level efficiency metrics like Largest Contentful Paint, First Enter Delay, and so on. The one hostile change there¹ is a rise of the principle thread rivalry (~7%). On Android¹, along with the same improve in the principle thread rivalry (~6%), there have been small regressions in First Enter Delay (~1%), Enter Delay (~3%) and First Contentful Paint (~0.5%). We do not anticipate these regressions to have a noticeable impression on consumer expertise, and are assured that they’re strongly outweighed by the extra security for our customers.

We must always emphasize that MiraclePtr at the moment protects solely class/struct pointer fields, to reduce the overhead. As future work, we’re exploring choices to develop the pointer protection to on-stack pointers in order that we will defend towards extra use-after-free bugs.

Word that the first aim of MiraclePtr is to forestall exploitation of use-after-free bugs. Though it wasn’t designed for diagnosability, it already helped us discover and repair numerous bugs that had been beforehand undetected. We’ve got ongoing efforts to make MiraclePtr crash studies much more informative and actionable.

Proceed to Present Us Suggestions

Final however not least, we’d prefer to encourage safety researchers to proceed to report points via the Chrome Vulnerability Reward Program, even when these points are mitigated by MiraclePtr. We nonetheless have to make MiraclePtr obtainable to all customers, accumulate extra information on its impression via reported points, and additional refine our processes and tooling. Till that’s completed, we is not going to take into account MiraclePtr when figuring out the severity of a bug or the reward quantity.

¹ Measured in Chrome 99.