Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

[Feature Request] Add resolved callstacks to Execution Events #118

Open
ddeadguyy opened this issue Feb 16, 2024 · 4 comments
Open

[Feature Request] Add resolved callstacks to Execution Events #118

ddeadguyy opened this issue Feb 16, 2024 · 4 comments
Labels
enhancement New feature or request

Comments

@ddeadguyy
Copy link

I'm hoping for the same New Thread Stack and Ready Thread Stack options that Windows CPU Usage (Precise) has.

@ivberg
Copy link
Contributor

ivberg commented Feb 21, 2024

agreed that would be nice

Is this for LTTng or for Perfetto events?

while stackwalk is theoretically possible to obtain on newest tracing, decoding stack and symbols remain a challenge.
There is a lack of a nice symbol decoding methods and format like on Windows. This is necessary to decode the stackwalk both in general for Linux/Android as well as in these trace formats.

I think we have experimental barely working support for LTTng for just KM (kernel-mode) stacks if a special file is provided similar to what Trace Compass does.

@ddeadguyy
Copy link
Author

This would be for LTTng Execution Events.

Is "experimental barely working support for LTTng for just KM (kernel-mode) stacks" documented anywhere?

We're also evaluating VTune in our search for a WPA-like Linux profiler. It captures Transitions that involve callstacks with our resolved symbols in them. Not quite as comprehensive as Windows CPU Usage (Precise) would be, though.

@ivberg
Copy link
Contributor

ivberg commented Feb 22, 2024

We're also evaluating VTune in our search for a WPA-like Linux profiler.

So since you used precise language about CPU Usage Precise and scheduling events, I know you are interested in them. However, just to be clear and if others can make use of this, I did want to say we do support Linux profiling with stacks (KM/UM). This is equivalent to Windows CPU Usage (Sampled).

You used the term "Linux profiler" here probably in the generic sense of the word, not the specific sense of an actual profiler which samples the CPU at a specific interval, determining where CPU time is spent on which functions and the stack that led there (profiling). Last I checked LTTng did not support profiling in this sense. Instead, these tools rely on Linux kernel perf tool cpu-clock events where the stack is decoded by perf/Linux on the box before reading into our tool. All this is documented here - https://github.com/microsoft/Microsoft-Performance-Tools-Linux-Android/blob/develop/LinuxTraceLogCapture.md#perf

Is "experimental barely working support for LTTng for just KM (kernel-mode) stacks" documented anywhere?

With that bit on profiling support out of the way let me move to your specific follow-up question. No it has not been documented before the experimental support we have, but probably you could get it to work with a few bugfixes if you want to look. AFAIK symbol info has not been added to Trace Compass traces, although it could technically be done. Therefore, there has to be some manual way to resolve symbols. LTTng grabs the undecoded callstack, but something still needs to resolve the symbols.
I will explain / document here where we are:

  1. Context / Inspiration - Trace Compass (OSS) is a popular GUI for reading LTTng traces, although it has different features than this toolkit and WPA. Trace compass supports providing the kallsyms (KM only) and loading it in the GUI. See https://archive.eclipse.org/tracecompass.incubator/doc/org.eclipse.tracecompass.incubator.kernel.doc.user/User-Guide.html
  2. This is where we attempt to read kallsyms for Perf cpu-clock events converted to LTTng CTF format. The conversion is a bit of a PITA to do - not really recommended. Anyways, the current checked in algo I think is wrong or sucked here, but it gives you the idea of what to do and where to do it.
  3. Anyways, this could be fixed and ported to LTTng scheduling events such that at least the KM stacks could be decoded similar to Trace Compass

Maybe try the kallsyms symbol support in Trace Compass if you can, and see if it works well enough for you to want to use it. Then we would be open to a contribution here in these tools to get similar support

Long term here is what I would suggest for the LTTNg and Linux folks to better support call stacks and symbols

  1. LTTng needs to support profiling (cpu-clock) events
  2. Embed binary id / signature info into the LTTng trace necessary for symbol decode (similar to what Windows does)
  3. Have some sort of symbol server or on-demand method to pull down symbols to decode (similar to what Windows does)

@ivberg ivberg added the enhancement New feature or request label Jun 4, 2024
@ivberg
Copy link
Contributor

ivberg commented Jun 4, 2024

FYI This (2) is now deprecated - kallsyms for Perf cpu-clock events converted to LTTng CTF format

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
enhancement New feature or request
Projects
None yet
Development

No branches or pull requests

2 participants