Created attachment 21380 [details] NFSv4 BUG ON log My NFS server rebooted. The machine with the kernel in question, one of many clients, spit out the attached error in dmesg, and all NFS activity on the machine blocked forever, necessitating a reboot. This is not true on any of the other NFS clients on the network, which vary between 2.6.18 and 2.6.27, so it may be A) 64-bit specific somehow (the rest are 32-bit), B) recently introduced, or C) recently exposed by some existing bad behavior in NFS recovery being removed. Machine was "vanilla" 2.6.30-rc4 (with commits b827e496c893de0c0f142abfaeb8730a2fd6b37f and 7fdf523067666b0eaff330f362401ee50ce187c4 added), 64-bit. NFSv4 mounted with rw,nosuid,nodev,noatime,hard,intr,nolock,sloppy,rsize=8192,wsize=8192,tcp,timeo=600. I'll try reproducing this on latest GIT shortly, but it's hard to reproduce (since it only occurs after the NFS server reboots, and not even consistently then), so I don't know when I'll be able to report back that it occurs or not.
http://article.gmane.org/gmane.linux.nfs/26502 is a previous trace/occurrence of this, which should demonstrate exactly how hard it is to reproduce.
On Sun, 2009-05-17 at 04:44 +0000, bugzilla-daemon@bugzilla.kernel.org wrote: > http://bugzilla.kernel.org/show_bug.cgi?id=13330 > > Summary: nfs4 NULL pointer dereference in _nfs4_do_setlk > Product: File System > Version: 2.5 > Kernel Version: 2.6.30-rc4 > Platform: All > OS/Version: Linux > Tree: Mainline > Status: NEW > Severity: high > Priority: P1 > Component: NFS > AssignedTo: trond.myklebust@fys.uio.no > ReportedBy: rercola@acm.jhu.edu > Regression: No > > > Created an attachment (id=21380) > --> (http://bugzilla.kernel.org/attachment.cgi?id=21380) > NFSv4 BUG ON log > > My NFS server rebooted. > > The machine with the kernel in question, one of many clients, spit out the > attached error in dmesg, and all NFS activity on the machine blocked forever, > necessitating a reboot. > > This is not true on any of the other NFS clients on the network, which vary > between 2.6.18 and 2.6.27, so it may be A) 64-bit specific somehow (the rest > are 32-bit), B) recently introduced, or C) recently exposed by some existing > bad behavior in NFS recovery being removed. > > Machine was "vanilla" 2.6.30-rc4 (with commits > b827e496c893de0c0f142abfaeb8730a2fd6b37f and > 7fdf523067666b0eaff330f362401ee50ce187c4 added), 64-bit. NFSv4 mounted with > > rw,nosuid,nodev,noatime,hard,intr,nolock,sloppy,rsize=8192,wsize=8192,tcp,timeo=600. > > I'll try reproducing this on latest GIT shortly, but it's hard to reproduce > (since it only occurs after the NFS server reboots, and not even consistently > then), so I don't know when I'll be able to report back that it occurs or > not. Switching to email... I'm having trouble reproducing this, and staring at the code itself isn't helping (as far as I can see, the locking using nfsi->rwsem should work). Could you therefore please try the attached patch? Cheers Trond
Created attachment 21474 [details] NFSv4: Fix an Oops in the lock recovery code
On Thu, 2009-05-21 at 17:03 -0400, Trond Myklebust wrote: > On Sun, 2009-05-17 at 04:44 +0000, bugzilla-daemon@bugzilla.kernel.org > wrote: > > http://bugzilla.kernel.org/show_bug.cgi?id=13330 > > > > Summary: nfs4 NULL pointer dereference in _nfs4_do_setlk > > Product: File System > > Version: 2.5 > > Kernel Version: 2.6.30-rc4 > > Platform: All > > OS/Version: Linux > > Tree: Mainline > > Status: NEW > > Severity: high > > Priority: P1 > > Component: NFS > > AssignedTo: trond.myklebust@fys.uio.no > > ReportedBy: rercola@acm.jhu.edu > > Regression: No > > > > > > Created an attachment (id=21380) > > --> (http://bugzilla.kernel.org/attachment.cgi?id=21380) > > NFSv4 BUG ON log > > > > My NFS server rebooted. > > > > The machine with the kernel in question, one of many clients, spit out the > > attached error in dmesg, and all NFS activity on the machine blocked > forever, > > necessitating a reboot. > > > > This is not true on any of the other NFS clients on the network, which vary > > between 2.6.18 and 2.6.27, so it may be A) 64-bit specific somehow (the > rest > > are 32-bit), B) recently introduced, or C) recently exposed by some > existing > > bad behavior in NFS recovery being removed. > > > > Machine was "vanilla" 2.6.30-rc4 (with commits > > b827e496c893de0c0f142abfaeb8730a2fd6b37f and > > 7fdf523067666b0eaff330f362401ee50ce187c4 added), 64-bit. NFSv4 mounted with > > > rw,nosuid,nodev,noatime,hard,intr,nolock,sloppy,rsize=8192,wsize=8192,tcp,timeo=600. > > > > I'll try reproducing this on latest GIT shortly, but it's hard to reproduce > > (since it only occurs after the NFS server reboots, and not even > consistently > > then), so I don't know when I'll be able to report back that it occurs or > not. > > Switching to email... > > I'm having trouble reproducing this, and staring at the code itself > isn't helping (as far as I can see, the locking using nfsi->rwsem should > work). > > Could you therefore please try the attached patch? You might also want to apply the attached fix to the RENEW function. It corrects a bug that can cause NFSv4 clients to fail to recover state during the grace period... Cheers Trond
Created attachment 21475 [details] NFSv4: Fix NFSv4 async renewal
Created attachment 21600 [details] A new, more different trace with RIP at _setlk ...well that's different. NFS behavior appears unaffected - no stuck threads, nothing. Just...the BUG ON in the attachment. Kernel git 0f6f49a8cd0163fdb1723ed29f01fc65177108dc with the patches in this bug applied.
Uuh... You appear to have a lock that contains a pointer to an invalid struct file. That is indeed a novelty.
Created attachment 21610 [details] Debugging patch.
Could you try adding this debugging patch in order to see what is going on with the open context here? It should be impossible to trigger the BUG_ON(nfs_file_open_context(request->fl_file)->state == NULL) from nfs4_lock_delegation_recall(), since nfs_delegation_claim_opens() explicitly tests for that condition. The only explanation I can see, therefore, is a reference leak of some sort.
Created attachment 21628 [details] A BUG ON with trond's debugging patch Oh look, a BUG ON for you. With your patch applied.
Sorry. That's a red herring... That's just the 'hung task' watchdog complaining because some process was holding the directory mutex for more than 120 seconds. It is quite normal for that to happen if the server was down for any extended period of time. Please just ignore those particular warnings, or turn off the hung task watchdog altogether...
I'm aware. I've seen them before, I was just including it to show that there was no other interesting printout before the BUG ON at the end. [And there is, indeed, a BUG ON at the end.]
Oh... I missed the one at the very end. Same apparently impossible state as before the last debugging patch. We pass the if (nfs_file_open_context(fl->fl_file)->state != state) test in nfs4_reclaim_locks, then apparently hit the BUG_ON(nfs_file_open_context(fl->fl_file)->state == NULL); in _nfs4_do_setlk. Could you please check that line 3472 in fs/nfs/nfs4proc.c matches mine, and is indeed that particular BUG_ON?
line 3472: BUG_ON(fl->fl_ops == NULL); in _do_setlk.
OK. That explains it. Your nfs4proc.c is shifted by 2 lines...
I'm reverting your patch, incidentally - the machine no longer stays up for longer than 30 minutes with it, and BUG ONs within 10 of boot.
Handled-By : Trond Myklebust <trond.myklebust@fys.uio.no>
Created attachment 21710 [details] NFS: Ensure we always hold the BKL when dereferencing inode->i_flock This patch shouldn't really make a difference if you're just doing NFSv4, but it doesn't hurt to try it. If you are running a samba server on top of NFS (or some other program that uses leases), then it might explain what is going on...
Created attachment 21751 [details] BUG ON with that patch applied "Great."
Oh for crying out loud... I missed your 'nolock' mount option. Of course that's going to fail...
Created attachment 21753 [details] NFSv4: Fix the 'nolock' option regression NFSv4 should just ignore the 'nolock' option. It is an NFSv2/v3 thing... This fixes the Oops in http://bugzilla.kernel.org/show_bug.cgi?id=13330
Does the above patch indeed fix the Oops? It should turn off the 'nolock' option and thus ensure that locks are managed by the NFSv4 code. The alternative, if we want to support 'nolock' on NFSv4 (we haven't earlier) is simply to turn off lock recovery for those mounts.
On Sunday 07 June 2009, Trond Myklebust wrote: > On Sun, 2009-06-07 at 11:52 +0200, Rafael J. Wysocki wrote: > > This message has been generated automatically as a part of a report > > of recent regressions. > > > > The following bug entry is on the current list of known regressions > > from 2.6.29. Please verify if it still should be listed and let me know > > (either way). > > > > > > Bug-Entry : http://bugzilla.kernel.org/show_bug.cgi?id=13330 > > Subject : nfs4 NULL pointer dereference in _nfs4_do_setlk > > Submitter : Rich Ercolani <rercola@acm.jhu.edu> > > Date : 2009-05-17 04:44 (22 days old) > > Handled-By : Trond Myklebust <trond.myklebust@fys.uio.no> > > It should still be listed. I do believe I've finally identified the > cause, and have provided a patch, however that fix has not yet been > verified by Rich...
Confirm. I've tried hard to trigger this again, to no effect. :)
Thanks. I'll send it to Linus tomorrow...