2.4.20 NFS / EJUKEBOX problem

2.4.20 NFS / EJUKEBOX problem

[Paul N]

Hi,
I'm seeing a weird problem with NFSv3's EJUKEBOX when
mounting a irix machine running DMF (software for managing offline/tape
copies of files).    When an NFS read request is made for an 'offline' file
it causes other write requests to stall until the 'offline'  file is 
unmigrated to
disk and the read can resume. 
 From looking at tcpdump it seems that the write requests proceed until
an EJUKEBOX msg from the read is returned from the server.  At that point
all write requests are stuck behind the  jukeboxed read. 

Could someone please give me some information on this?

Thanks
Paul

Re: 2.4.20 NFS / EJUKEBOX problem

[Trond Myklebust]

>>>>> " " == Paul N <Paul> writes:

     > Hi, I'm seeing a weird problem with NFSv3's EJUKEBOX when
     > mounting a irix machine running DMF (software for managing
     > offline/tape copies of files).  When an NFS read request is
     > made for an 'offline' file it causes other write requests to
     > stall until the 'offline' file is unmigrated to

     > disk and the read can resume. From looking at tcpdump it seems
     > that the write requests proceed until

     > an EJUKEBOX msg from the read is returned from the server.  At
     > that point all write requests are stuck behind the jukeboxed
     > read. Could someone please give me some information on this?

The writes are probably waiting in 'nfs_try_to_free_pages()' because
you are hitting the hard limit on outstanding read/write requests.
The NFS code is not allowed to allocate more than MAX_REQUEST_HARD
requests per mountpoint at any time since there is otherwise no way
for the VM to notify us in case of resource starvation.

Cheers,
  Trond