We're working on a high availability experiment
project for our linux-based server and the nfsds hang
from times to times. It is observed that the version 3
procedure 12/13 (rmdir and remove) obtains an inode
semaphore within fh_lock()/nfsd_unlink() but never
releases it. Theoretically this inode is going to get
released but what can prevent the following two cases
that would end up hanging the nfsd ?
1. Due to slow network and/or system, the client re-
sends the same request but using a different xid. The
nfsd that accepts the new request would be stuck in
nfsd_unlink() waiting for the inode semaphore.
2. A "create" immediately follows the "remove" for the
same file.
There is a flag called fh_locked but it is passed in
from each individual nfs request (a local copy, not
from dcache) that doesn't help. I also browse thru
some file system's (say ext3) inode free code but
can't find anywhere this semaphore gets released.
We're still debugging our failover logic but if
someone could confirm the above are indeed a bug (or
explain why it will not happen) would be of great
helps.
Or should this semaphore get "zero"ed out by file
system (say, ext3) ?
We're on 2.4.19 kernel.
Thanks for the help.
Wendy
Adding fh_put() to release the semaphore after
nfsd_unlink() within nfs3proc.c does make the
hang disappear. Hope we can see this get fixed
in the future code.
wendy
----- Original Message -----
From: Wendy Cheng
To: [email protected]
Sent: Tuesday, February 11, 2003 12:17 PM
Subject: the inode semaphore
We're working on a high availability experiment
project for our linux-based server and the nfsds hang
from times to times. It is observed that the version 3
procedure 12/13 (rmdir and remove) obtains an inode
semaphore within fh_lock()/nfsd_unlink() but never
releases it. Theoretically this inode is going to get
released but what can prevent the following two cases
that would end up hanging the nfsd ?
1. Due to slow network and/or system, the client re-
sends the same request but using a different xid. The
nfsd that accepts the new request would be stuck in
nfsd_unlink() waiting for the inode semaphore.
2. A "create" immediately follows the "remove" for the
same file.
There is a flag called fh_locked but it is passed in
from each individual nfs request (a local copy, not
from dcache) that doesn't help. I also browse thru
some file system's (say ext3) inode free code but
can't find anywhere this semaphore gets released.
We're still debugging our failover logic but if
someone could confirm the above are indeed a bug (or
explain why it will not happen) would be of great
helps.
Or should this semaphore get "zero"ed out by file
system (say, ext3) ?
We're on 2.4.19 kernel.
Thanks for the help.
Wendy