Three bugs in the xHCI endpoint-restart path used by all error recovery
(stall hard reset, transaction-error soft retry, resource retry,
split/babble hard reset):
1. Latent deadlock: restart_endpoint held the port_states write guard
across set_tr_deque_ptr(), which internally re-acquires a read guard
on the same key (std RwLock read-while-write on one thread).
Unobserved because error injection is not yet exercised at runtime
(P8-C). Fixed by scoping phase-1 ring priming so the guard drops
before the async command.
2. Doorbell ordering violated xHCI spec 4.6.8/4.6.10: after Reset
Endpoint the TR Dequeue Pointer is undefined, so Set TR Dequeue
Pointer must complete BEFORE the doorbell transitions the endpoint
Stopped->Running. The old order (doorbell first) ran the endpoint
with an undefined dequeue pointer — undefined xHC behavior on real
hardware. Linux rings the doorbell from the Set TR Dequeue command
completion path (xhci_handle_cmd_set_deq, ring.c:1416-1554); xhcid
now issues Set TR Dequeue, awaits completion, then rings.
3. Priming NoOp never executed: ring.register() was captured after
ring.next() advanced the enqueue index, so the dequeue pointed past
the NoOp (dead TRB). Now captured before next(), priming the
hardware dequeue AT the NoOp so the xHC executes it on restart —
same semantics as Linux xhci_move_dequeue_past_td pointing at the
first valid TRB.
Verified: cargo check clean (138 warnings, unchanged), 43/43 tests pass.
Red Bear OS
260003331e