Re: another problem with latest code drops

[Lachlan McIlroy <lachlan@sgi.com>]

Dave Chinner wrote:
> On Thu, Oct 16, 2008 at 06:35:03PM +1000, Lachlan McIlroy wrote:
>> Dave Chinner wrote:
>>> On Thu, Oct 16, 2008 at 05:38:39PM +1000, Lachlan McIlroy wrote:
>>>> Dave Chinner wrote:
>>>>> On Thu, Oct 16, 2008 at 12:06:21PM +1000, Lachlan McIlroy wrote:
>>>>>> fsstress started reporting these errors
>>>>>>
>>>>>> fsstress: check_cwd failure
>>>>>> fsstress: check_cwd failure
>>>>>> fsstress: check_cwd failure
>>>>>> fsstress: check_cwd failure
>>>>>> fsstress: check_cwd failure
>>>>>> ...
>>> ....
>>>>> Ah, yes. A shutdown in a directory transaction. Have you applied the
>>>>> fix to the directory block allocation transaction accounting that was one
>>>>> of the last patches I posted?
>>>> Yes, I checked that in yesterday and ran with it overnight.
>>> OK.
>>>
>>>>> If so, then there's some other problem in that code that we'll
>>>>> need a reproducable test case to be able to find....
>>>> I was running 8 copies of this command:
>>>> fsstress -p 64 -n 10000000 -d /mnt/data/fsstress.$i
>>>>
>>>> I tried it again but this time the system ran out of memory
>>>> and locked up hard.  I couldn't see why though - maybe a memory
>>>> leak.
>>> I just ran up the same load in a UML session. I'd say it's this
>>> slab:
>>>
>>>   2482   2481  99%    0.23K    146       17       584K xfs_btree_cur
>>>
>>> which is showing a leak. It is slowly growing on my system
>>> and dropping the caches doesn't reduce it's size. At least it's
>>> a place to start looking - somewhere in the new btree code we
>>> seem to be leaking a btree cursor....
>> I'm not seeing a leak in that slab - actually that slab doesn't even
>> show up.
> 
> Overnight the xfs_btree_cur slab made it up to about 7000 in use
> entries, so there is definitely a leak there, though it is a slow
> one.
> 
>> I am seeing a lot of memory used here though:
>>
>> 116605669 116605669  26%    0.23K 6859157       17  27436628K selinux_inode_security
> 
> Ah - I don't run selinux. Sounds like a bug that needs reporting
> to lkml...

I'm sure this is caused by your changes that introduced inode_init_always().
It re-initialises an existing inode without destroying it first so it calls
security_inode_alloc() without calling security_inode_free().