The ADD path holds the manager 'device' mutex, and invalidate_cached_device
takes it. Its own doc says it must not be called from inside that mutex. Calling
it there deadlocked the session: on .173 the ADD stopped after
SET_RENDER_ADAPTER, no monitor was created, and the client reported 'no frames
received'.
The mechanism itself is sound and measured — recycling the driver's WUDFHost
clears the declare (pid 3872 -> 19932, adapter_luid 0x8ed607 -> 0x1a8f6ca,
cursor_excluded true -> false, next session streamed normally). What is missing
is a call site that runs OUTSIDE that mutex and is still on every session's
path; the handshake site tried earlier is not reached on this host.
Off unless PUNKTFUNK_CURSOR_RECYCLE=1. A session that self-composites is the
old behaviour; a deadlocked one is a regression, and the tree should not carry
that while the call site is unsolved.
enricobuehler
c5eea458b4