On 5/9/19 7:20 PM, katcherw(a)gmail.com wrote: For NFSv3, there isn't an explicit close in the protocol. This leaves servers with two choices: close the file every time, and incur the overhead of opening it every time, or cache the opens for a while, and close them later. This is what Ganesha does. MDCACHE has a 2 layer LRU that keeps track entries. When an entry transitions from L1 to L2 of the LRU, it's open file handle is closed. This is done in the background by a scrub thread, and is governed by a high-water and low-water mark on the number of FDs, configured by FD_HWMark_Percent and FD_LWMark_Percent. When lwmark is passed, the thread wakes up; when hwmark is passed, it runs really hard. There's a hard limit, detected at startup from the system rlimit. Percentages are relative to this. Long story short: NFSv4 files will generally be closed when the client is done; NFSv3 files will hang around Daniel