Recovery refuses a starting owner that becomes ready after a failed re-attach; Linux zombies count as live #422

Closed
opened 2026-10-23 07:38:35 +02:00 by buildagent · 0 comments
Member

Related to #350, #382 and #416. The normal Linux run for PR #420 at a3c9f56b failed a_replaced_daemon_is_disclosed_in_both_envelopes (5043/5044 passed): re-attach received a connection reset, recovery found a live kernel lock and a starting registration, and returned immediately. A later RPC in the same tool call correctly refused a second heal, yielding internal_error. Windows, packages, pilots and corpus passed.

Two independently reproduced recovery defects:

  • The starting-owner guard never attempts an authenticated identity handshake after re-attach fails, even when the owner becomes ready between the two observations. A deterministic socket test reproduces this without sleeps; recovery must reuse the same-root authenticated owner without spawning, and reject a foreign root. Keep the existing reconnect timeout and enclosing tool budget.
  • Linux is_process_alive only checks /proc/<pid> existence, so an exited, unreaped owner is considered alive. A real child-process regression fails before the fix and must pass afterward. Unknown/unreadable process state should retain the conservative fallback.

The exact PID relationship in the CI failure was not recorded; these are demonstrated defects in its recovery path, not a claim that the zombie case caused that run. The end-to-end assertion now includes the killed PID. The unmodified end-to-end case passed 40 sequential and 80 concurrent local runs; these passes do not waive the red CI head.

Fix and negative controls are in progress for PR #420; release remains blocked until the new head passes normal CI.

Related to #350, #382 and #416. The normal Linux run for PR #420 at a3c9f56b failed `a_replaced_daemon_is_disclosed_in_both_envelopes` (5043/5044 passed): re-attach received a connection reset, recovery found a live kernel lock and a starting registration, and returned immediately. A later RPC in the same tool call correctly refused a second heal, yielding internal_error. Windows, packages, pilots and corpus passed. Two independently reproduced recovery defects: - The starting-owner guard never attempts an authenticated identity handshake after re-attach fails, even when the owner becomes ready between the two observations. A deterministic socket test reproduces this without sleeps; recovery must reuse the same-root authenticated owner without spawning, and reject a foreign root. Keep the existing reconnect timeout and enclosing tool budget. - Linux `is_process_alive` only checks `/proc/<pid>` existence, so an exited, unreaped owner is considered alive. A real child-process regression fails before the fix and must pass afterward. Unknown/unreadable process state should retain the conservative fallback. The exact PID relationship in the CI failure was not recorded; these are demonstrated defects in its recovery path, not a claim that the zombie case caused that run. The end-to-end assertion now includes the killed PID. The unmodified end-to-end case passed 40 sequential and 80 concurrent local runs; these passes do not waive the red CI head. Fix and negative controls are in progress for PR #420; release remains blocked until the new head passes normal CI.
Sign in to join this conversation.
No milestone
No project
No assignees
1 participant
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set

Reference
h-dv/code-index#422
No description provided.