Fixed terminate_mp_processes to cover failure workers.

[hexinw-nvidia]

Failure workers have orphan descendant processes left behind from SubprocessContext._close(). temrinate_mp_processes is the right place to catch all the orphan cleanup.

Note if the orphan processes are temprorily or permanently stuck in the kernel I/O wait. The user-level signl based kill would not help.

[greptile-apps]

Greptile Summary

This PR broadens terminate_mp_processes in utils.py to clean up all orphaned descendant processes in worker process groups (not just multiprocessing-specific ones), which fixes missed cleanup for failure workers whose descendants are left behind by SubprocessContext._close(). It also fixes two previously flagged inefficiencies: pre-fetched ppid is now consumed from process.info['ppid'] instead of re-reading via process.ppid(), and the unused name and cmdline attrs are removed from the psutil.process_iter call.

Key changes:

• Removed cmdline pattern matching (from multiprocessing.resource_tracker import main, from multiprocessing.spawn import spawn_main) — now any process with ppid=1 in a target pgid is terminated
• process.ppid() → process.info['ppid'] — correctly uses the pre-fetched cache, avoiding a redundant /proc read per process
• attrs trimmed from ['pid', 'ppid', 'name', 'cmdline'] → ['pid', 'ppid'] — removes two now-unnecessary pre-fetches
• Docstring updated to reflect the broadened scope and document the intentional single-pass (no recursive children walk) design decision
• launcher.py: Updated comments in _shutdown to clarify best-effort cleanup for orphan descendants

Confidence Score: 4/5

• Safe to merge; the broadened termination scope is intentional and well-guarded by the pgid filter.
• The logic change is focused and purposeful—broadening cleanup from multiprocessing-specific patterns to any process with ppid=1 in target pgids. This tighter filtering (ppid=1 AND pgid membership) actually improves safety compared to pattern matching. All previously flagged inefficiencies (ppid cache usage, unused attrs) have been fixed. The acknowledged limitation (D-state processes cannot be signaled) is a kernel constraint, not a code defect. No regressions are expected.
• No files require special attention.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A["_shutdown(death_sig, timeout)"] --> B["_pcontext.close(death_sig, timeout)"]
    B --> C{"Any orphan<br/>processes remain?"}
    C -- "Worker died normally<br/>children reparented to PID 1" --> D["terminate_mp_processes<br/>allowed_ppids={1}<br/>allowed_pgids=_children_pgids"]
    C -- "Worker stuck in D-state<br/>children still parented to worker" --> E["⚠ Orphans NOT caught<br/>(ppid ≠ 1 yet)"]
    D --> F["psutil.process_iter<br/>attrs=['pid','ppid']"]
    F --> G["For each process:<br/>check process.info['ppid'] in {1}"]
    G -- "ppid match" --> H["check os.getpgid(pid)<br/>in _children_pgids"]
    G -- "no match" --> F
    H -- "pgid match" --> I["process.terminate()<br/>(SIGTERM)"]
    H -- "no match" --> F
    I --> J["Exception swallowed<br/>(best-effort)"]
    E --> K["Known limitation:<br/>D-state processes<br/>cannot be signaled"]

Loading

_{Last reviewed commit: 5d7afa6}