fix: ensure Python backend shm cleanup on forced shutdown

[Vinya567]

What does the PR do?

Runtime fix for orphaned triton_python_backend_shm_region_* files when Python backend stub shutdown exceeds the server exit timeout. Pairs with the QA-side change in server#8881.

Ticket

• Linear: TRI-1330 — Phase 2 (runtime cure)
• Server-side (Phase 1, merged): test: tolerate non-zero Triton exit on Jetson graceful shutdown server#8881

Root cause

• stub-timeout-seconds was parsed at backend init but not applied inside TerminateStub — the wait on the stub was effectively unbounded.
• Parent-owned shm regions were only removed via the SharedMemoryManager destructor, so a forced/abnormal exit could skip cleanup and leak /dev/shm/triton_python_backend_shm_region_*.

Fix

• Wire stub_timeout_seconds_ through StubLauncher::TerminateStub with a bounded pop + WaitForStubProcessWithTimeout; fall back to KillStubProcess on timeout (src/stub_launcher.{cc,h}).
• Add explicit SharedMemoryManager::RemoveShmRegion() and call it from TerminateStub (src/shm_manager.{cc,h}).
• Register parent-owned shm regions for std::atexit cleanup as a safety net for forced exits.

Test plan

• IGX-Orin GitLab job 366443271: L0_backend_python--IGX-Orin lifecycle PASSED with this PR + server#8881; /dev/shm shm-region count unchanged before/after run.
• x86: no regression on L0_backend_python/lifecycle.

Checklist

• PR title reflects the change and is of format <type>: <description>
• Changes are described in the pull request.
• Related PRs are referenced.
• Added test plan and verified test passes.

Commit Type

• fix

[greptile-apps]

Greptile Summary

This PR fixes two root causes of orphaned /dev/shm/triton_python_backend_shm_region_* files after forced/timeout-triggered server shutdown: an effectively unbounded wait in TerminateStub (the stub timeout was parsed but never applied) and shm regions that were only removed via the SharedMemoryManager destructor (skipped on abnormal exits).

• Bounded teardown (stub_launcher.cc): TerminateStub now shares a single stub_timeout_seconds_ budget between the Pop (finalize response) and WaitForStubProcessWithTimeout (process exit polling), falling back to KillStubProcess on timeout. All three issues flagged in the first review round (double timeout, final-sleep off-by-one, int64_t→int narrowing) are addressed.
• Explicit shm removal (shm_manager.cc/.h): RemoveShmRegion() is called unconditionally at the end of TerminateStub, and parent-owned regions are tracked in a global set with an atexit handler as a last-resort safety net for exits that bypass normal cleanup.

Confidence Score: 5/5

Safe to merge; the core teardown logic is correct and all three issues from the first review round are addressed.

The bounded teardown path (single shared budget, correct remaining-seconds accounting, final waitpid recheck) and the idempotent RemoveShmRegion flow are logically sound. The two remaining comments are about the atexit handler behaviour in a shared-library unload scenario and a missing flag-reset on an exceedingly rare atexit failure — neither affects normal operation or the primary fix.

src/shm_manager.cc — the atexit registration block warrants a second look if Triton ever dynamically unloads backends via dlclose mid-session.

Important Files Changed

Filename	Overview
src/shm_manager.cc	Adds anonymous-namespace globals for atexit-based shm cleanup: RegisterParentShmRegion, UnregisterParentShmRegion, CleanupParentShmRegions. Splits the destructor into destructor + RemoveShmRegion() (idempotent). Logic is sound; atexit handler has a theoretical safety concern in shared-library contexts.
src/shm_manager.h	Adds RemoveShmRegion() public declaration. Minimal, clean change.
src/stub_launcher.cc	Wires stub_timeout_seconds_ through TerminateStub, adds WaitForStubProcessWithTimeout (bounded poll-sleep loop with a final recheck to avoid killing an already-exiting stub), and calls RemoveShmRegion unconditionally. All three reviewer-raised issues from the previous round are addressed.
src/stub_launcher.h	Adds WaitForStubProcessWithTimeout declaration and stub_timeout_seconds_ field. Clean, consistent with implementation.

Sequence Diagram

sequenceDiagram
    participant Server
    participant StubLauncher
    participant Stub
    participant ShmManager

    Server->>StubLauncher: TerminateStub()
    alt is_healthy
        StubLauncher->>Stub: Push(PYTHONSTUB_FinalizeRequest)
        StubLauncher->>StubLauncher: Pop(pop_timeout_ms, success)
        alt success within timeout
            StubLauncher->>StubLauncher: compute remaining_seconds
            StubLauncher->>StubLauncher: WaitForStubProcessWithTimeout(remaining_seconds)
            alt stub exited in time
                StubLauncher-->>StubLauncher: OK
            else timeout
                StubLauncher->>Stub: KillStubProcess() SIGKILL
            end
        else timeout / no response
            StubLauncher->>Stub: KillStubProcess() SIGKILL
        end
    else not healthy
        StubLauncher->>Stub: KillStubProcess() SIGKILL
    end
    StubLauncher->>ShmManager: RemoveShmRegion() idempotent
    ShmManager->>ShmManager: shm_unlink + UnregisterParentShmRegion
    Note over ShmManager: atexit(CleanupParentShmRegions) fires at exit for any regions not cleaned up by TerminateStub

Loading

_{Reviews (3): Last reviewed commit: "polish: log atexit registration failure;..." | Re-trigger Greptile}