Description of problem:

Currently, it's possible for a write to occur while a read-ahead is in progress, and nothing is done to interrupt or cancel that read-ahead.  As a result, a user read which is fully subsequent to that write (even if it's O_SYNC) could still get the read-ahead data which is now stale.

How reproducible:

It's a race, found by inspection, so it's probably pretty hard to hit (hasn't been observed in practice except by stopping things in a debugger).  It might be possible to construct a log-tailing test case, as that would leave one process reading just behind where another was writing on the same client.  Meanwhile, since read-ahead is just a performance optimization, it should be safe to disable it when the risk can be anticipated.