1307194 - Stack walking seems to miss frames sometimes

Status: RESOLVED WORKSFORME

(Core :: Gecko Profiler, defect)

Description

[Nick Fitzgerald [:fitzgen] [⏰PST; UTC-8]]

When digging through profiles with jimb and sfink, we found stack traces where every frame looked plausible, but there were too few frames and there *had* to be more in between them. Whether this is bad stack walking or not using DWARF info to get inlined frames (it seemed like too many missing frames to just be inlining TBH), I'm not 100% sure.

Comment 1

[Steve Fink [:sfink] [:s:]]

Example: https://cleopatra.io/#report=c9ad18b88e6461fb7848626e224409e93fc70031&invertCallback=true&selection=%22(total)%22,158,1

(most of __poll's time is when it is called by XRE_Main.)

Comment 2

[Markus Stange [:mstange]]

(In reply to Steve Fink [:sfink] [:s:] from comment #1)
> Example:
> https://cleopatra.io/
> #report=c9ad18b88e6461fb7848626e224409e93fc70031&invertCallback=true&selectio
> n=%22(total)%22,158,1

What you see here is just the "pseudostack". The profile doesn't have C++ callstacks because it was gathered on a build that didn't include stackwalking support or because the "stackwalk" profiler feature wasn't enabled. In the sidebar it says "Stackwalk: False".

> (most of __poll's time is when it is called by XRE_Main.)

This is probably while the process is waiting inside the native event loop, which is not instrumented with pseudostack labels, so the only pseudostack label on the stack is XRE_Main.

Comment 3

[Markus Stange [:mstange]]

I'm closing this bug because it's not actionable. Please file new bugs with example profiles when you find occurrences of bad stack walking in Nightly.