registry_cleanup_stale runs on every graph query, costing one stat() per live entry #52

query_all() calls registry_cleanup_stale() before every graph query (rmw_graph.cpp#L46-L58). It has 13 call sites, so this covers rmw_get_node_names, rmw_count_publishers, rmw_count_subscribers, rmw_get_topic_names_and_types, rmw_get_service_names_and_types and the rest.

registry_cleanup_stale() walks every non-empty slot, snapshots 1164 bytes from it, and calls stat("/proc/<pid>") to check whether the owner is still alive (registry.cpp#L359-L371).

Reclaiming slots owned by dead processes is garbage collection. Running it on a read path makes every graph question pay for it.

Measured sweep cost:

rmw_get_topic_names_and_types calls query_all twice, once for publishers and once for subscriptions (rmw_graph.cpp#L261-L262), so it pays the full cost twice.

Anything that polls the graph turns this into a continuous load. A recorder polling the topic list at the common 10 Hz default costs about 0.5 s of CPU per second against a registry with ~9000 entries, and it pays that whether or not the graph changed.

Suggested fix

• Rate limit registry_cleanup_stale to at most once per interval per process. One second is plenty. Small change, large effect.
• Make rmw_get_topic_names_and_types sweep once instead of twice.
• Longer term, drop the per-slot stat() entirely. A dead peer already makes the doorbell sendto fail with ECONNREFUSED, so liveness can be learned without a syscall per entity.

Worth noting separately: stat("/proc/<pid>") also succeeds for a zombie, since the entry survives until the parent reaps it. A slot whose owner has exited but not been reaped is therefore treated as alive and never reclaimed by this path.