0.13.1 - 2025-02-01

Release Notes

This release adds Windows support. It uses ETW via xperf to record the system activity to an ETL file. Then samply converts the ETL file.

Samply asks for Adminstrator privileges during profiling. This is necessary for ETW to work.

Thanks to @jrmuizel for getting this off the ground. Most of the Windows implementation was initially written by him. ETW is rather lightly documented, so this required a lot of research.

Also thanks to @vvuk, who integrated Jeff's code into samply and contributed hugely to getting this ready for production!

And thanks to the authors of the https://github.com/n4r1b/ferrisetw crate; samply uses etw-reader which started out as a fork of ferrisetw.

Known issues:

• By default, you won't get Windows symbols, but you can use samply record --windows-symbol-server https://msdl.microsoft.com/download/symbols to fix this - this will download symbols for Windows system libraries and kernel stacks from Microsoft's server. I'm planning to add a config file for samply so that symbol servers can be configured more permanently, but it doesn't exist yet.
• Missing symbols for precompiled .NET code: This is getsentry/pdb#153, which has a potential patch in getsentry/pdb#154.
• CoreCLR support could be better - some of it isn't working correctly any more (see #483)

Breaking changes

• The minimum supported Rust version is now 1.77.

Features

• Windows: Initial support.
• macOS: Support attaching to running processes and their subprocesses (#190, by @vvuk, and #425, by @tmm1)
• macOS: Add samply setup to code-sign samply so that attaching to running processes can work (#217 + #353, by @vvuk)
• All platforms: samply import has much better support for Android simpleperf now
• All platforms: Add --main-thread-only flag
• All platforms: Add --include-args argument
• Windows, Linux: Add --per-cpu-threads flag
• All platforms: Add --symbol-dir, --windows-symbol-server, --windows-symbol-cache, --breakpad-symbol-server, --breakpad-symbol-dir, --breakpad-symbol-cache, and --simpleperf-binary-cache arguments (various PRs, including some by @ishitatsuyuki)
• All platforms: Add --address option to specify the IP address at which the local server is listening (#234, by @Rjected)
• All platforms: Add --unstable-presymbolicate flag (#202, by @vvuk)

Fixes

• Fix build errors related to zerocopy and zerocopy_derive (#356, by @mox692)
• macOS: Fix library enumeration on macOS 15 Sequoia (#403, by @Maaarcocr)

Install samply 0.13.1

Install prebuilt binaries via shell script

curl --proto '=https' --tlsv1.2 -LsSf https://github.com/mstange/samply/releases/download/samply-v0.13.1/samply-installer.sh | sh

Install prebuilt binaries via powershell script

powershell -ExecutionPolicy Bypass -c "irm https://github.com/mstange/samply/releases/download/samply-v0.13.1/samply-installer.ps1 | iex"

Download samply 0.13.1

File	Platform	Checksum
samply-aarch64-apple-darwin.tar.xz	Apple Silicon macOS	checksum
samply-x86_64-apple-darwin.tar.xz	Intel macOS	checksum
samply-x86_64-pc-windows-msvc.zip	x64 Windows	checksum
samply-aarch64-unknown-linux-gnu.tar.xz	ARM64 Linux	checksum
samply-x86_64-unknown-linux-gnu.tar.xz	x64 Linux	checksum
samply-x86_64-unknown-linux-musl.tar.xz	x64 MUSL Linux	checksum

0.12.0 - 2024-04-16

Release Notes

Breaking changes

• The minimum supported Rust version is now 1.74.
• samply load perf.data is now called samply import perf.data.
• The --port alias has changed from -p to -P.

Features

• Linux: Allow attaching to running processes with samply record -p [pid] (#18, by @ishitatsuyuki)
• Linux, macOS: Support Jitdump in samply record.
• Linux: Support Jitdump in samply import perf.data without perf inject --jit.
• Linux, macOS: Support /tmp/perf-[pid].map(#34 + #36, by @bnjbvr)
• Linux, macOS: Support specifying environment variables after samply record.
• Linux, macOS: Add --iteration-count and--reuse-threads flags to samply record.
• Linux: Support symbolication with .dwo and .dwp files.
• Linux: Support unwinding and symbolicating VDSO frames.
• Linux, macOS: Support overwriting the launched browser with $BROWSER (#50, by @ishitatsuyuki)
• Linux, macOS: Add --profile-name argument to samply record and samply import to allow overriding the profile name (#68, by @rukai)
• Linux, macOS: Support Scala Native demangling (#109, by @keynmol)
• macOS: Support --main-thread-only in samply record, for lower-overhead sampling
• macOS, Linux: Unstable support for adding markers from marker-[pid].txt files which are opened (and, on Linux, mmap'ed) during profiling.
• Linux: Support kernel symbols when importing perf.data files with kernel stacks, if /proc/sys/kernel/kptr_restrict is 0.
• Android: Support importing perf.data files recorded with simpleperf's --trace-offcpu flag.

In progress

• Linux: Groundwork to support profiling Wine apps (by @ishitatsuyuki)

Fixes

• Linux, macOS: Don't discard information from processes with reused process IDs (e.g. due to exec).
• Linux: Support recording on more types of machines, by falling back to software perf events in more cases. (#70, by @rkd-msw)
• Linux: Fix out-of-order samples. (#30 + #62, by @ishitatsuyuki)
• Linux: Fix unwinding and symbolicating in processes which have forked without exec.
• Linux: Capture startup work of launched processes more reliably.
• Linux: Fix debuglink symbolication in certain cases. (#38, by @zecakeh)
• Linux: Fix stackwalking if unwinding information is stored in compressed .debug_frame sections. (#10, by @bobrik)
• macOS: Fix symbolication of system libraries on x86_64 macOS 13+.
• Android: Allow building samply for Android. (#76, by @flxo)
• macOS: Fix Jitdump symbolication for functions which were JITted just before the sample was taken (#128, by @vvuk)
• macOS, Linux: More reliable handling of Ctrl+C during profiling.
• macOS: Support recording workloads with deep recursion by eliding the middle of long stacks and not running out of memory.
• x86_64: Improve disassembly of relative jumps by displaying the absolute target address (#54, by @jrmuizel)
• macOS: Use yellow instead of blue, for consistency with Linux which uses yellow for user stacks and orange for kernel stacks.

Other

• Improve build times by using the separate serde-derive crate (#65, by @CryZe)

Install samply 0.12.0

Install prebuilt binaries via shell script

curl --proto '=https' --tlsv1.2 -LsSf https://github.com/mstange/samply/releases/download/samply-v0.12.0/samply-installer.sh | sh

Install prebuilt binaries via powershell script

powershell -c "irm https://github.com/mstange/samply/releases/download/samply-v0.12.0/samply-installer.ps1 | iex"

Download samply 0.12.0

File	Platform	Checksum
samply-aarch64-apple-darwin.tar.xz	Apple Silicon macOS	checksum
samply-x86_64-apple-darwin.tar.xz	Intel macOS	checksum
samply-x86_64-pc-windows-msvc.zip	x64 Windows	checksum
samply-x86_64-unknown-linux-gnu.tar.xz	x64 Linux	checksum
samply-x86_64-unknown-linux-musl.tar.xz	x64 MUSL Linux	checksum

v0.11.0

This release comes with the following fixes:

• Fixed a panic when closing the profiler tab during loading. (#11)
• On Linux, samply load perf.data will now include kernel symbols from /proc/kallsyms if run with root privileges.
• On Linux, if the DEBUGINFOD_URLS environment variable is set, samply will fetch symbols from the listed debuginfod servers.
• In the profile JSON, additional properties are set to hide some unnecessary UI elements.
• In the profile JSON, macOS library information now has the right "arch" values. Furthermore, symbolication of macOS system libraries from the dyld shared cache will only check the dyld cache files for that architecture.

v0.10.1

This release raises the minimum supported Rust version to 1.61.

It comes with the following fixes compared to 0.9:

• On macOS 13, system libraries will have symbols again. (ajdusted dyld shared cache paths)
• On Linux, there will be fewer panics during recording.
• On Linux, there will be a useful error message if perf event paranoid settings are inadequate.
• CLI argument parsing is improved when recording executables with arguments. 0.9 sometimes required the use of --.

v0.8.0

This release rewrites stackwalking to make use of various types of unwinding info. Stacks should now be higher quality, and frame pointers are no longer required.