Retbleed: Exploiting retpolines

Syonyk · July 12, 2022, 5:04pm

Retbleed (CVE-2022-29900 and CVE-2022-29901) is the new addition to the family of speculative execution attacks that exploit branch target injection to leak information, which we call Spectre-BTI. Unlike its siblings, who trigger harmful branch target speculation by exploiting indirect jumps or calls, Retbleed exploits return instructions. This means a great deal, since it undermines some of our current Spectre-BTI defenses.

One such defense that many of our operating systems use today is called retpoline. Retpolines work by replacing indirect jumps and calls with returns. Back in 2018, retpolines were invented to prevent unauthorized attackers from stealing information from the memory of the system using Spectre-BTI. While some were concerned about whether returns could be susceptible to Spectre-BTI attacks too, returns were deemed impractical to exploit, and concerns were brushed off the shoulder. The primary reason being that return target predictions are, under normal microarchitectural conditions, not predicted as indirect branches.

As it turns out however, Retbleed is indeed practical to exploit, thanks to the following two insights:

We found that we can trigger the microarchitectural conditions, on both AMD and Intel CPUs, that forces returns to be predicted like indirect branches. We also built the necessary tools to discover locations in the Linux kernel where these conditions are met.

We found that we can inject branch targets that reside inside the kernel address-space, even as an unprivileged user. Even though we cannot access branch targets inside the kernel address-space — branching to such a target results in a page fault — the Branch Prediction Unit will update itself upon observing a branch and assume that it was legally executed, even if it’s to a kernel address.

Man, we are so good at computers, even our mitigations have vulnerabilities!