in recent 7-STABLE: VOP_WRITE...is not exclusive locked but should be

[John Baldwin]

On Thursday 19 March 2009 12:02:51 pm Kostik Belousov wrote:
> On Thu, Mar 19, 2009 at 10:01:44AM -0400, John Baldwin wrote:
> > On Thursday 19 March 2009 8:05:34 am Tim Chase wrote:
> > > Hello,
> > > 
> > > I have a system that had been running quite well with an oldish 7-STABLE
> > > (from around August 7, 2008) but has started deadlocking within the past
> > > week or so.
> > > 
> > > I updated the kernel to a newer 7-STABLE (Mar 15, 2009) and enabled
> > > INVARIANTS, INVARIANT_SUPPORT, WITNESS, DEBUG_LOCKS DEBUG_VFS_LOCKS and
> > > DIAGNOSTIC and the message indicated in the subject line has now 
appeared
> > > 3 times as shown below.  Is this something to be terribly concerned 
about?
> > > Is there anything I can to to further track down the cause?  Since the
> > > system is a production mail server, I have it set to not drop into DDB
> > > when this happens.
> > > 
> > > The machine is a 4-core Xeon X5450 with 8G of RAM running FreeBSD
> > > amd64 and in userland it's pretty much just cyrus imapd and apache/php.
> > > The file systems are all ZFS on a bunch of SAS drives connected to a
> > > LSI Logic 1068 controller.
> > > 
> > > As to the deadlock that started this exercise, if the machine follows 
its
> > > recent pattern, that should happen within the next 2-4 hours.
> > 
> > Err, the vn_write() routine should be using an exclusive vnode lock:
> > 
> > vn_write()
> > {
> > 
> > 	...
> > 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, td);
> > 	if ((flags & FOF_OFFSET) == 0)
> > 		uio->uio_offset = fp->f_offset;
> > 	ioflag |= sequential_heuristic(uio, fp);
> > #ifdef MAC
> > 	error = mac_check_vnode_write(active_cred, fp->f_cred, vp);
> > 	if (error == 0)
> > #endif
> > 		error = VOP_WRITE(vp, uio, ioflag, fp->f_cred);
> > 	...
> > }
> > 
> > Can you check your /sys/kern/vfs_vnops.c and verify that LK_EXCLUSIVE is 
> > present in your vn_write() routine?  If so, then perhaps run memtest?
> 
> Note that this happens on the ZFS. Might be, ZFS unlocks the vnode and
> then relocks it with some invalid flags ?

Hmm, it depends on where the check is I guess (before or after zfs_write()).  
Tim, can you do 'l *VOP_WRITE_APV+0x155' on your kernel from gdb?

-- 
John Baldwin