mirror of
https://github.com/permissionlesstech/bitchat.git
synced 2026-11-15 08:12:23 +00:00
On iOS every return to the foreground restarts Tor: ensureRunningOnForeground() calls restartArti(), which called the Rust arti_stop() inline. arti_stop() sleeps 200 ms while it holds Arti's state lock, and restartArti() runs on the main actor, so the main thread froze for about 200 ms on every resume (and on any other restart, such as recovery after a network path change). Nobody put this call on the main thread on purpose; it inherited it. TorManager is @MainActor, so restartArti() runs there even though it wraps its state changes in MainActor.run as if it ran somewhere else. shutdownCompletely() already makes the same call from inside Task.detached. Await arti_stop() on a detached task instead. The restart still waits for the stop to finish before it continues, so the sequence is the same: TorWillRestart, the generation bump and flag resets, then the stop, then the wait loop, then startIfNeeded(). The main actor is simply free while the stop runs. Deliberately unchanged: - The full restart on every foreground. It has been there since #564 ("iOS can suspend Tor harshly") and was device-verified in the v1.7.1 pass (#1580); attempts to avoid it never proved out. - The 200 ms sleep inside arti_stop(). It gives the old run_arti task time to clear IS_RUNNING and BOOTSTRAP_PROGRESS before arti_start sets them again, so it stays, and the stop still strictly precedes the start. No Rust or xcframework change. - Everything else around the restart: the TorWillRestart notification, the generation bump, the flag resets, and AppRuntime's URLSession rebuild and relay reset. Simulator A/B, same build with only the placement of this call toggled, measured by a background probe that timed main-queue round trips every 10 ms for 10 s after .active: - Before (7 resumes): arti_stop() took 202 ms on the main thread, and each resume had one 190-194 ms main-thread stall. Tor was ready 1.48-1.49 s after .active. - After (3 resumes): the worst main-thread stall per resume was 16.7-17.6 ms. Tor was ready 1.49-1.50 s after .active and the relays reconnected. TorManagerTests scans TorManager.swift and fails if arti_stop() is called anywhere outside a Task.detached closure. Run against the previous source, it flags the old call in restartArti(). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>