Skip to contentSkip to conversation

The host->h_inuse NLM's nlm_gc_hosts()

DISCUSSION Wendy Cheng 2 participants 3 messages Jun 5, 2006 – Jun 13, 2006

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

I'm chasing a bug in our distribution - when signalling lockd, the lock references do not seem to get released. I just spot the "nlm_gc_hosts" explicitly set

host->h_inuse = 0

then call nlmsvc_mark_resources() to remark h_inuse to 1 if there is a lock reference via:

nlmsvc_mark_resources ->
nlm_traverse_files ->
nlm_inspect_file->
nlm_traverse_locks->

    183 again:
    184         file->f_locks = 0;
    185         for (fl = inode->i_flock; fl; fl = fl->fl_next) {
    186                 if (fl->fl_lmops != &nlmsvc_lock_operations)
    187                         continue;
    188
    189                 /* update current lock count */
    190                 file->f_locks++;
    191                 lockhost = (struct nlm_host *) fl->fl_owner;
    192                 if (action == NLM_ACT_MARK)
    193                         lockhost->h_inuse = 1;
    194                 else if (action == NLM_ACT_CHECK)
    195                         return 1;
    196                 else if (action == NLM_ACT_UNLOCK) {
    197                         struct file_lock lock = *fl;
    198
    199                         if (host && lockhost != host)
    200                                 continue;

So in theory, from #193, the lockhost->h_inuse is supposed to get set back to 1. However, I'm wondering whether the fl->fl_owner would get overwritten anywhere by (task_struct *)current->files in various other code paths so we could end up updating different structures ? This could allow the host gets freed such that nlmsvc_invalidate_all() will no longer find this host ?

Will keep doing this bug hunt tomorrow but wondering someone knows this code well can assure me that this won't happen ? Many thanks in advance.

Also the assumption here is that by sending lockd a (any) signal would kickoff nlmsvc_invalidate_all() to remove the locks. Is this a correct assumption ?

-- Wendy

_______________________________________________
NFS maillist - [email protected]
https://lists.sourceforge.net/lists/listinfo/nfs

#1 lore raw
Quoted 2 lines · Wendy Cheng

I'm chasing a bug in our distribution - when signalling lockd, the lock references do not seem to get released.

It turns out to be a false alarm. What happened was that we had reports saying RHCS (Red Hat cluster suite) in some cases would not be able to umount the filesystem during active-active failover whenever nfs client was holding posix locks.

Extensive tracing shows the umount failure is actually caused by export entry's vfsmount reference due to the way unexport was handled. Sending a signal to lockd *did* release the posix locks as expected.

On the other hand, we do have some other NLM issues (with patches) that we would like to discuss with the list. More on that later.

-- Wendy

_______________________________________________
NFS maillist - [email protected]
https://lists.sourceforge.net/lists/listinfo/nfs

#2 lore raw
In reply to Wendy Cheng
Quoted 11 lines · On Monday June 5, [email protected] wrote:

So in theory, from #193, the lockhost->h_inuse is supposed to get set back to 1. However, I'm wondering whether the fl->fl_owner would get overwritten anywhere by (task_struct *)current->files in various other code paths so we could end up updating different structures ? This could allow the host gets freed such that nlmsvc_invalidate_all() will no longer find this host ?

Will keep doing this bug hunt tomorrow but wondering someone knows this code well can assure me that this won't happen ? Many thanks in advance.

It certainly shouldn't happen, but I cannot promise it won't - if there is a bug somewhere.

Quoted 3 lines

Also the assumption here is that by sending lockd a (any) signal would kickoff nlmsvc_invalidate_all() to remove the locks. Is this a correct assumption ?

Not 'any' signal - only SIGKILL. All other signals are blocked for lockd.

NeilBrown

_______________________________________________
NFS maillist - [email protected]
https://lists.sourceforge.net/lists/listinfo/nfs