Skip to content

attachplane: a displacement notice lost to its own write deadline leaves a client believing it controls #89

Description

@jiashuoz

Found by the seventh review of #84, pre-existing on main's attach path and untouched by that PR.

Interleaving

Peer A has been told something (so displace owes it a notice). A's client stops draining. The notice takes A's announce hold uncontended and dies inside its own write when the handoff step's deadline expires. The stream deliberately does not close a socket whose caller's deadline expired (so one take-over cannot disconnect a client reading scrollback), and the post-write correction loop in announceAs is gated on the caller's context, so it never turns. Result: the plane has A at view@N, A's client still believes attached control N-1, its stdin is dropped at the plane, its heartbeat is inert for a viewer, and its take-control key sends no claim because the client believes it controls. Only detaching recovers it.

Reproduced by a probe on the review's throwaway copy; no shipped test pins it.

Fix direction

Record that a peer is owed a notice when announceAs returns without landing one (send error, or the early return when the hold could not be taken in time), and retry it on the existing heartbeat tick, extended to cover a viewer whose last announce was dropped. Add the probe as the regression test.

Acceptance

The probe passes; the stress set and the mutation battery from the #84 reviews stay clean; no new unbounded wait on any handoff path.

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    bugSomething isn't working

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions