Repository navigation
fix(event): reset the process-global reader across replaced terminals #80
Description
Problem
Issue #77 makes EOF/HUP terminal input a stable error in one Unix event source. Crossterm's process-global INTERNAL_EVENT_READER, however, owns that source for the process lifetime. After an embedded/public tut::run returns from terminal EOF, replacing fd0 with a new tty and calling tut::run again still reaches the old source's terminal_closed latch.
Blindly dropping or rebuilding the source is unsafe when event-stream is enabled: EventStream caches the old source's poll waker, so a worker blocked on a new generation may no longer be woken on Drop.
The CLI normally exits after #77 and is not affected. This issue tracks the reusable-library/process-global lifecycle separately rather than expanding the bounded EOF fix.
Required design
- Give the global reader an explicit source generation and a stable terminal identity.
- Preserve a stable EOF for the failed generation while the underlying tty identity is unchanged.
- When a later session presents a different valid tty, initialize exactly one new source generation before reading it.
- Make
EventStreamobtain/wake the active generation; no cached stale waker may strand a worker. - Deliver source errors deterministically—do not lose an EOF between the worker and executor, duplicate it into an unrelated generation, or spin while retrying.
- Preserve the Bound Unix terminal input memory and sequence lifetime #74 byte/time/event bounds and fix(input): terminate cleanly when the controlling terminal closes #77 HUP ordering.
Acceptance
- Public same-process test: run against tty A, close A, observe the exact EOF error, bind fd0 to tty B, and successfully run/read from B.
- An unchanged closed tty continues to return the same EOF without source recreation or busy looping.
- Concurrent
poll/readremains serialized by the existing reader lock; generation changes are atomic. -
EventStreamreceives the EOF once, then can use a later generation; dropping it always wakes the currently blocked worker. - Source initialization failure is typed and does not publish a half-initialized generation.
- Current/MSRV strict Clippy/tests, Linux/macOS PTY tests, and event-stream configurations pass.
Activity
m1ngsama commentedon Aug 10, 2026
Design audit after 0.0.4 confirms that this must be a generation-aware lifecycle change, not
source = UnixInternalEventSource::new()after EOF.Additional constraints discovered while tracing both Unix backends:
use-dev-ttybackend currently relies on process-lifetime registration of its SIGWINCH pipe. Rebuilding sources without an RAII registration/unregistration owner would leak registrations and descriptors across generations.EventStreamcannot cache one source waker for its whole lifetime. The reader should expose an atomic active-generation snapshot, and worker results/errors must carry that generation so EOF cannot be lost, duplicated into a replacement terminal, or delivered to the wrong executor turn.The safest implementation order is therefore: (1) source-owned tty identity and registration RAII, (2) generation-tagged global reader state and transactional replacement, (3) generation-aware EventStream worker/result/waker handoff, then (4) the public tty-A/EOF/tty-B end-to-end test. A narrower singleton reset would satisfy the synchronous happy path while leaving stream hangs and
use-dev-ttyresource leaks, so it should not be merged independently.m1ngsama commentedon Aug 11, 2026
Completed by signed crossterm fork commit
e3888223b5064543484cc08c3966cc05d711003aand TUT mergefe91fc4e3b1f0dbe6d13e566d187a3eab285fd89(PR #106).Final evidence:
tut::runafter reading B input;