SFrame V3: Towards efficient stack-tracing in GNU/Linux
CEME 1212 | Fri 07 Aug 2 p.m.–2:45 p.m.
Presented by
-
Indu Bhagat has been a regular contributor to the various GNU Toolchain components over the last few years. She also helps maintain SFrame, CTF/BTF formats in the GNU Toolchain.
Indu Bhagat has been a regular contributor to the various GNU Toolchain components over the last few years. She also helps maintain SFrame, CTF/BTF formats in the GNU Toolchain.
Abstract
First released in early 2023, SFrame (Simple Frame) offers a fast and efficient method for stack tracing without relying on frame pointers. A new version of the format, SFrame V3, was released with Binutils 2.46 (February 2026). Compared to V2, it features support for text sizes greater than 2GiB, more flexibility in representing topmost frames, support for marking outermost frames, and support for marking signal trampolines, among other improvements.
In this talk, we will introduce the format and explore how it interfaces with the Linux kernel’s deferred unwinder infrastructure (allowing kernel tools like perf and other tracing mechanisms to safely read user-space stacks and get reliable user-space stack traces quickly). We will also examine the ongoing integration efforts to bring SFrame-based stack tracing holistically to user space.
Finally, we will map out the missing pieces and call upon the community to help drive the seamless, distro-wide adoption we need to finally move away from Frame Pointers and reclaim CPU performance on some targets.
First released in early 2023, SFrame (Simple Frame) offers a fast and efficient method for stack tracing without relying on frame pointers. A new version of the format, SFrame V3, was released with Binutils 2.46 (February 2026). Compared to V2, it features support for text sizes greater than 2GiB, more flexibility in representing topmost frames, support for marking outermost frames, and support for marking signal trampolines, among other improvements.
In this talk, we will introduce the format and explore how it interfaces with the Linux kernel’s deferred unwinder infrastructure (allowing kernel tools like perf and other tracing mechanisms to safely read user-space stacks and get reliable user-space stack traces quickly). We will also examine the ongoing integration efforts to bring SFrame-based stack tracing holistically to user space.
Finally, we will map out the missing pieces and call upon the community to help drive the seamless, distro-wide adoption we need to finally move away from Frame Pointers and reclaim CPU performance on some targets.