kernelNFS(lockd) problem and patch suggestion

[Juan Gomez <juang@us.ibm.com>]

[-- Attachment #1: Type: text/plain, Size: 1064 bytes --]





Hi all,

I noticed that after starting a linux NFS server, the first lock request
gets delayed about 30 seconds even after waiting the 45 seconds
corresponding to the grace period.
At first I thought this was due to the fact that being the first time a
lock is acquired through the server lockd/statd interaction was required
and the creation of a new file was involved
(i.e. statd/sm/client_IP) but still 30 seconds was way too long...

After taking a look at the code I realized that the lockd thread sets grace
period and then goes to sleep for a long time waiting for messages and the
first message always gets processed
before checking if the grace period has completed (which it might after
sleeping for a long time).

I think this can be easily fixed by clearing the grace period after waiting
for messages and before procesing them for correct operation as described
in the following patch.

I will appreciate if this patch can be considered for inclusion in the
general distribution of the kernel.

(See attached file: lockd_delay_patch.2.4.19)


Juan

[-- Attachment #2: lockd_delay_patch.2.4.19 --]
[-- Type: application/octet-stream, Size: 1398 bytes --]

*** linux-2.4.19/fs/lockd/svc.c	Sun Oct 21 10:32:33 2001
--- linux-2.4.19.lockd_delay_patch/fs/lockd/svc.c	Mon Oct  7 22:16:20 2002
***************
*** 144,151 ****
  		 */
  		if (!nlmsvc_grace_period) {
  			timeout = nlmsvc_retry_blocked();
! 		} else if (time_before(grace_period_expire, jiffies))
! 			clear_grace_period();
  
  		/*
  		 * Find a socket with data available and call its
--- 144,150 ----
  		 */
  		if (!nlmsvc_grace_period) {
  			timeout = nlmsvc_retry_blocked();
! 		} 
  
  		/*
  		 * Find a socket with data available and call its
***************
*** 163,168 ****
--- 162,177 ----
  
  		dprintk("lockd: request from %08x\n",
  			(unsigned)ntohl(rqstp->rq_addr.sin_addr.s_addr));
+ 
+                 /* We need to do the clear/grace period here and not before svc_recv()
+                  * because svc_recv may sleep longer than the grace period and the 
+                  * first request may be falsely processed as if the server was in the 
+                  * grace period when it is not causing unnecessary delays.
+                  * Juan C. Gomez j_carlos_gome@yahoo.com
+                  */
+                 if (nlmsvc_grace_period && time_before(grace_period_expire, jiffies)){
+                   clear_grace_period();
+                 }
  
  		/*
  		 * Look up the NFS client handle. The handle is needed for