- fix readdir vs fragment race by keeping a separate frag pos, and ignoring dentries below it
mds
+- fix client op replay
+ - ordering
+ - should to set caps/lock states so that ops can proceed without revoking from client
+ - need to make sure replayed ops finish before new ops proceed. globally across the cluster. ick.
- fix up *_RDCACHE vs FILE_RDCACHE semantics
- ability to read attribute value is distinct from being able to hold cached pages?
- file recovery maybe needs to scan entire file range for a truncation event? (and object attr maybe needs file offset, not object offset, or original truncation?)