Delay in WRITE delegation recall during COMMIT

[Ajay Gulati]

Hi All,
I am testing write delegations on nfsv4 and it seems that there is some 
sort of loop between COMMIT and DELEGRETURN, that causes commit to take 90 
seconds on closing of a file.
The scenario to duplicate is as follows:

f1 = open(file1, RDONLY)
close(f1)
f2 = open(file2, RDWR)	// you will get write delegation
write(f2, 75KB data)
f3= open(file2,RDWR)
close(f2)
close(f3)

Here the close takes 90 seconds to finish. tcpdump shows that 
commit fails 4-5 times with NFS4ERR_DELAY, delegation 
is being recalled as part of nfsd_commit()->nfsd_open()->breaklease() 
execution path. The server ultimately times out on delegation recall.

I think, the client is waiting for commit to succeed for returning 
delegation and the server is doing the recall as part of commit. I am not 
quite sure why this is taking so long, when delegation recall is spawned 
as a separate thread.

Same problem happens if some other client opens the file2 with RDWR 
permissions. The open(file2,RDWR) from other client takes 90 sec to 
return.

I can post exact program and tcpdump if needed. It should be fairly easy 
to reproduce though.
I am using kernel: 2.6.17-rc6 and CITI patch for that kernel.

If you have seen this before or know the reason why this is happening, 
please let me know.

Thanks
Ajay