Nasty non-recursive lockmgr panic on softdep only enabled UFS partition when filesystem full

[Garrett Cooper]

On May 4, 2011, at 2:07 AM, Kostik Belousov wrote:

> On Tue, May 03, 2011 at 11:58:49PM -0700, Garrett Cooper wrote:
>> On Tue, May 3, 2011 at 11:42 PM, Garrett Cooper <yanegomi at gmail.com> wrote:
>>> On Tue, May 3, 2011 at 10:59 PM, Kirk McKusick <mckusick at mckusick.com> wrote:
>>>>> Date: Tue, 3 May 2011 22:40:26 -0700
>>>>> Subject: Nasty non-recursive lockmgr panic on softdep only enabled UFS
>>>>>  partition when filesystem full
>>>>> From: Garrett Cooper <yanegomi at gmail.com>
>>>>> To: Jeff Roberson <jeff at freebsd.org>,
>>>>>         Marshall Kirk McKusick <mckusick at mckusick.com>
>>>>> Cc: FreeBSD Current <freebsd-current at freebsd.org>
>>>>> 
>>>>> Hi Jeff and Dr. McKusick,
>>>>>     Ran into this panic when /usr ran out of space doing a make
>>>>> universe on amd64/r221219 (it took ~15 minutes for the panic to occur
>>>>> after the filesystem ran out of space -- wasn't quite sure what it was
>>>>> doing at the time):
>>>>> 
>>>>> ...
>>>>> 
>>>>>     Let me know what other commands you would like for me to run in kgdb.
>>>>> Thanks,
>>>>> -Garrett
>>>> 
>>>> You did not indicate whether you are running an 8.X system or a 9-current
>>>> system. It would be helpful to know that.
>>> 
>>> I've actually been running CURRENT for a few years now, but you're right --
>>> I didn't mention that part.
>>> 
>>>> Jeff thinks that there may be a potential race in the locking code for
>>>> softdep_request_cleanup. If so, this patch for 9-current should fix it:
>>>> 
>>>> Index: ffs_softdep.c
>>>> ===================================================================
>>>> --- ffs_softdep.c       (revision 221385)
>>>> +++ ffs_softdep.c       (working copy)
>>>> @@ -11380,7 +11380,8 @@
>>>>                                continue;
>>>>                        }
>>>>                        MNT_IUNLOCK(mp);
>>>> -                       if (vget(lvp, LK_EXCLUSIVE | LK_INTERLOCK, curthread)) {
>>>> +                       if (vget(lvp, LK_EXCLUSIVE | LK_NOWAIT | LK_INTERLOCK,
>>>> +                           curthread)) {
>>>>                                MNT_ILOCK(mp);
>>>>                                continue;
>>>>                        }
>>>> 
>>>> If you are running an 8.X system, hopefully you will be able to apply it.
>>> 
>>>    I've applied it, rebuilt and installed the kernel, and trying to
>>> repro the case again. Will let you know how things go!
>> 
>>    Happened again with the change. It's really easy to repro:
>> 
>> 1. Get a filesystem with UFS+SU
>> 2. Execute something that does a large number of small writes to a partition.
>> 3. 'dd if=/dev/zero of=FOO bs=10m' on the same partition
>> 
>>    The kernel will panic with the issue I discussed above.
>> Thanks!
> 
> Jeff' change is required to avoid LORs, but it is not sufficient to
> prevent recursion. We must skip the vnode supplied as a parameter to
> softdep_request_cleanup(). Theoretically, other vnodes might be also
> locked by curthread, thus I think the change below is needed. Try this.
> 
> diff --git a/sys/ufs/ffs/ffs_softdep.c b/sys/ufs/ffs/ffs_softdep.c
> index a6d4441..25fa5d6 100644
> --- a/sys/ufs/ffs/ffs_softdep.c
> +++ b/sys/ufs/ffs/ffs_softdep.c
> @@ -11380,7 +11380,9 @@ retry:
> 				continue;
> 			}
> 			MNT_IUNLOCK(mp);
> -			if (vget(lvp, LK_EXCLUSIVE | LK_INTERLOCK, curthread)) {
> +			if (VOP_ISLOCKED(lvp) ||
> +			    vget(lvp, LK_EXCLUSIVE | LK_INTERLOCK | LK_NOWAIT,
> +			    curthread)) {
> 				MNT_ILOCK(mp);
> 				continue;
> 			}

	Ran into the same panic after I applied the patch above with the repro steps I described before. One thing that I noticed is that the issue isn't as easy to reproduce unless you add the dd in parallel with the make operation.
Thanks,
-Garrett