An event applied while the open's lookup was in flight could already be
reflected in the returned entry while sitting above the pre-lookup
cursor, so the fence under-covered and the queued invalidation replaced
the fresher entry. The open now re-reads until the cursor is stable
across the lookup; on sustained churn the last pre-lookup value stands,
which can only under-fence, never block a newer event.
A pre-RPC ping likewise cannot vouch for events committed during the
RPC itself. CacheRemoteObjectToLocalCluster now returns a log timestamp
stamped before the filer reads the entry — events are logged after
their store write on that same clock, so everything at or below it is
reflected in the returned entry, making the fence causal with the
response. The mount prefers the response event's timestamp, then this
log timestamp, and keeps the pre-RPC ping only for filers that return
neither.
Chris Lu
ab490013e0