rcu: a per-table choice between RCU and SRCU for the walk, decided by a number #1184

rcu.map walks inside an SRCU read-side critical section since #1179, so its callback may sleep and every entry a writer unlinks goes through call_srcu before kfree_rcu. Measured on the shared host (aarch64, three runs of a million operations each, minimum ns per operation, with a spread between runs of 110 to 172 on master alone): a replace costs 110 on master and 113 on #1179, an insert and remove pair 134 and 145, and a walk 164 and 146 per key visited. The difference between call_srcu and kfree_rcu on the writer sits inside the noise, the walk got cheaper, and lookups are untouched. Not measured: memory in flight, since an unlinked entry now lives an SRCU grace period, tens of ms, plus an RCU one, instead of one.

A table walked only from interrupt-context runtimes, whose callback cannot sleep anyway, could keep the walk under rcu_read_lock and free with kfree_rcu alone. The choice is per table and not per call, since the writer decides how it frees, and it costs a flag at creation, two walk paths, a refusal of the sleeping walker and its documentation; the tables written hot in the tree, systrack's, are walked from process context and would stay SRCU. Not done now: open it when a consumer shows a hot-writer table walked only from hooks, with that table's number in hand.