v1.48.0

libxtc (1.48.0-1) unstable; urgency=medium

• proc: fix a wrong-proc bug that stranded fibers on kqueue backends.
  xtc_proc_wait_fd took its proc from __current_proc, a thread-local the
  coroutine layer restores across every yield -- and under the multi-loop
  executor the restored value named a DIFFERENT proc than the running
  fiber 99.2% of the time (measured). It therefore registered another
  fiber's task as the readiness tag and displaced that fiber's live fd
  registration, so several loops ended up watching the same open file and
  raced to dispatch its readiness to their own tag; the real parker stayed
  blocked on a completion whose result was already published. struct
  xtc_coro now carries a proc back-pointer published by __proc_entry,
  wait_fd re-anchors from the running fiber, and __proc_free clears the
  link (the coro outlives the proc, so leaving it dangling was a
  use-after-free). Fixes the /m5/exec/Blk2_concurrent_commit hang that
  had burned the FreeBSD CI job's 6h ceiling for weeks; that job gates
  again. See docs/KNOWN_ISSUES.md.
• io: xtc_io_event_t carries the fd an event is for, populated by all nine
  backends, so the dispatcher unregisters the registration that actually
  fired instead of guessing from task->park_fd and the polling loop.
  task->park_fd is now atomic and every unregister CLAIMS it, replacing a
  torn test-then-use that a peer loop's dispatch could race.
• tools: xtc-stranded joins each parked fiber's loop against that loop's
  io_uring state and flags an fd park on a CQ-overflowed ring -- a park
  with a legitimate source is still doomed if the ring that owes it a
  completion cannot deliver one. Reported by a consumer whose two doomed
  fibers showed as "0 suspect".
• test: the doc-snippet gate probes for -lrt like it already probed for
  -ldl; without it every snippet failed to LINK on FreeBSD and the gate
  reported "10 of 10 broken" against a healthy library.

-- Greg Burd greg@burd.me Wed, 16 Sep 2026 00:00:00 -0400