fix(perf): purge inherited mappings on execve to fix unknown symbols

[GuillaumeLagrange]

Purge a process's inherited memory mappings when it execve's a new binary, so wall-time symbolication doesn't resolve frames against the wrong module.

A child forked from a parent (e.g. bash) inherits the parent's memory
mappings. When it then execve's its own binary, the inherited PIE
mappings remain and can collide with the new binary's mappings at the
same base address, leading to wrong symbolication and unknown symbols.

Handle COMM records flagged as execve and purge the exec'ing process's
inherited mappings so only its post-exec mappings are used.

Timing-aware module mapping would be a more general fix but is larger in scope; tracked separately in COD-1377.

Refs COD-2779

[codspeed-hq]

Merging this PR will not alter performance

⚠️ Unknown Walltime execution environment detected

Using the Walltime instrument on standard Hosted Runners will lead to inconsistent data.

For the most accurate results, we recommend using CodSpeed Macro Runners: bare-metal machines fine-tuned for performance measurement consistency.

✅ 7 untouched benchmarks

---

_{Comparing cod-2779-fix-unknown-symbols-in-walltime-and-add-e2e-gate (a54f0af) with main (bfff150)}

[not-matthias]

LGTM!

[greptile-apps]

Greptile Summary

This PR fixes incorrect symbolication when a child process (e.g. a benchmark forked from bash) exec's its own binary by purging stale inherited memory mappings on COMM records flagged as is_execve. A new purge_process_mappings helper iterates all loaded modules and removes the exec'ing PID's entries, ensuring that only post-exec MMAP2 records are used for symbolication.

• Adds a RecordType::COMM branch that fires purge_process_mappings for any tracked PID whose COMM record carries is_execve, clearing all module mappings inherited via inherit_parent_mappings on fork.
• Adds a chronological-ordering comment documenting the stream assumption.
• Adds a focused unit test (purge_process_mappings_removes_only_the_target_pid) that reproduces the fork+exec collision scenario and verifies the purge's selectivity.

Confidence Score: 5/5

Safe to merge; the change is narrowly scoped to purging stale per-PID mappings on execve and cannot affect PIDs that don't exec.

The fix correctly intercepts COMM records with is_execve, gates the purge behind pid_filter.should_include so only tracked processes are affected, and relies on the chronological ordering already required by the rest of the parsing loop. The new unit test covers the exact collision scenario and verifies isolation between PIDs. No correctness issues found.

No files require special attention.

Important Files Changed

Filename	Overview
src/executor/wall_time/profiler/perf/parse_perf_file.rs	Adds COMM/execve handling to purge inherited mappings for exec'ing PIDs; logic is correct, well-guarded by pid_filter, and covered by a new unit test.

Sequence Diagram

sequenceDiagram
    participant perf as perf stream
    participant parser as parse_for_memmap2
    participant modules as loaded_modules_by_path
    participant filter as pid_filter

    perf->>parser: "FORK(ppid=100, pid=200)"
    parser->>filter: add_child_if_parent_tracked(100, 200)
    filter-->>parser: true (child tracked)
    parser->>modules: inherit_parent_mappings(100 to 200)
    Note over modules: bash mappings copied to pid 200

    perf->>parser: "COMM(pid=200, is_execve=true)"
    parser->>filter: should_include(200)
    filter-->>parser: true
    parser->>modules: "purge_process_mappings(pid=200)"
    Note over modules: All pid 200 entries removed

    perf->>parser: "MMAP2(pid=200, path=/cpp/build/fractal_main)"
    parser->>modules: "process_mmap2_record(pid=200, new binary)"
    Note over modules: Clean post-exec mapping for pid 200

Loading

_{Reviews (4): Last reviewed commit: "fix(perf): purge inherited mappings on e..." | Re-trigger Greptile}