DLPX-95965 _framestr() parsing is fragile - should use drgn StackFrame API properties

[dbjwhs-perforce]

DLPX-95965 Replace fragile string parsing in _framestr() with drgn StackFrame API

Problem

@mmaybee called out in #352 that the _framestr() function was using string parsing on str(frame) output to extract frame information. This approach was brittle and could break if drgn's string representation format changed.

Solution

This PR replaces the string parsing with direct use of drgn StackFrame API properties:

• frame.pc - Program counter (address)
• frame.function_name - Function name (or None if unavailable)
• frame.symbol() - Symbol object for fallback name lookup
• frame.source() - Returns (filename, line, column) tuple
• frame.is_inline - Boolean for inline frame detection

Changes

1. Updated _framestr() signature: Added required frame_index: int parameter.

2. Rewrote frame formatting logic: Uses API properties directly with error handling via try/except for LookupError

3. Updated call sites:

   • Changed for frame in stack_trace: to for frame_index, frame in enumerate(stack_trace):
   • Updated calls from _framestr(frame, False) to _framestr(frame, False, frame_index)
   • Updated calls from _framestr(frame, True) to _framestr(frame, True, self.frame_id)

Testing

Tested with crash dump from kernel 6.14.0-27-dx2025082915-8ba235c2d-generic:

sdb> trace 0xffff8f1a285e5400
TASK: 0xffff8f1a285e5400 INTERRUPTIBLE PID: 12546
#0  0xffffffff92718fd0 in context_switch() at core.c:5378:2 (inlined)
#1  0xffffffff92718fd0 in __schedule() at core.c:6765:8
#2  0xffffffff92719389 in __schedule_loop() at core.c:6842:3 (inlined)
#3  0xffffffff92719389 in schedule() at core.c:6857:2
#4  0xffffffffc06c8a40
#5  0xffffffff917ab35b in kthread() at kthread.c:464:9
#6  0xffffffff916d9dd4 in ret_from_fork() at process.c:153:3

All features working correctly:

• Frame indices from enumerate
• Addresses from frame.pc
• Function names from frame.function_name
• Source locations from frame.source() showing file:line:col
• Inline markers from frame.is_inline

Benefits

• More robust: No longer dependent on string format
• More maintainable: Uses documented API instead of parsing heuristics
• Better error handling: Explicit try/except for missing data
• Clearer code: Intent is obvious from property names

Ensuring no regressions

Verified Command Compatibility

All related stack inspection commands continue to work correctly with the updated _framestr() implementation:

trace command - Displays full stack trace with frame indices, addresses, functions, and source locations:

sdb> trace 0xffff8f1a285e5400
TASK: 0xffff8f1a285e5400 INTERRUPTIBLE PID: 12546
#0  0xffffffff92718fd0 in context_switch() at core.c:5378:2 (inlined)
#1  0xffffffff92718fd0 in __schedule() at core.c:6765:8
#2  0xffffffff92719389 in __schedule_loop() at core.c:6842:3 (inlined)
#3  0xffffffff92719389 in schedule() at core.c:6857:2
...

frame command - Shows detailed frame information with function arguments (uses _framestr() with args=True):

sdb> frame 0
#0  0xffffffff92718fd0 in context_switch (rq=0xffff8f566bbb6e80, prev=0xffff8f1a285e5400, 
    next=0xffff8f1840f32a00, rf=0xffffa36c7599bd20) at core.c:5378:2 (inlined)

locals command - Displays local variables in the current frame:

sdb> locals
rf = (struct rq_flags *)0xffffa36c7599bd20
next = (struct task_struct *)0xffff8f1840f32a00
prev = (struct task_struct *)0xffff8f1a285e5400
rq = (struct rq *)0xffff8f566bbb6e80

registers command - Shows CPU register values at the current frame:

sdb> registers
rbx = 18446620200070246016
rbp = 18446642284957646200
rsp = 18446642284957646096
r12 = 18446619941242033152
r13 = 18446619933064505856
r14 = 0
r15 = 18446619941242036240
rip = 18446744071871500240

These commands demonstrate that the refactored _framestr() function maintains full backward compatibility while providing more robust frame formatting through the drgn StackFrame API.

[lyriclake]

Thanks for taking this on!

[dbjwhs-perforce]

@mmaybee if you look back at this heads up I removed the <no address found> handling because @lyriclake pointed out that frame.pc is a Final[int] attribute that cannot fail, and we already access it without protection at line 268. So all that got removed.