Re: INFO: HARDIRQ-safe -> HARDIRQ-unsafe lock order detected for kmemleak_lock

[Eric Paris <eparis@redhat.com>]

On Wed, 2009-09-02 at 12:44 +0100, Catalin Marinas wrote:
> On Wed, 2009-09-02 at 11:54 +0100, Catalin Marinas wrote:
> > On Tue, 2009-09-01 at 16:55 -0400, Eric Paris wrote:
> > > I wrote a multithreaded inotify syscall pounder intended to create
> > > files, destroy files, create watches, and destroy watches with the
> > > maximum number of races possible.  Instead after letting it run a while
> > > I came upon this!  And then my system started to crash in all sorts of
> > > fun and glorious ways (kmem_cache_alloc bugs/panics/whatever)
> > > 
> > > -Eric
> > > 
> > > [ 2235.913737] ======================================================
> > > [ 2235.914084] [ INFO: HARDIRQ-safe -> HARDIRQ-unsafe lock order detected ]
> > > [ 2235.914084] 2.6.31-rc8-next-20090901 #64
> > > [ 2235.914084] ------------------------------------------------------
> > > [ 2235.914084] syscall_thrash/2516 [HC0[0]:SC0[0]:HE0:SE1] is trying to acquire:
> > > [ 2235.914084]  (kthread_create_lock){+.+...}, at: [<ffffffff81091543>] kthread_create+0x73/0x180
> > > [ 2235.914084] 
> > > [ 2235.914084] and this task is already holding:
> > > [ 2235.914084]  (kmemleak_lock){..----}, at: [<ffffffff81152611>] create_object+0x161/0x2e0
> > > [ 2235.914084] which would create a new lock dependency:
> > > [ 2235.914084]  (kmemleak_lock){..----} -> (kthread_create_lock){+.+...}
> > 
> > Are there other messages from kmemleak printed before that? It looks to
> > me like kmemleak got an exceptional situation (not being able to
> > allocate memory or inserting a pointer into the prio search tree) and it
> > disabled itself. When disabling, it starts a clean-up thread and AFAICT
> > that's the only condition when kmemleak_lock -> kthread_create_lock
> > dependency would be created.
> > 
> > I'm not sure whether disabling interrupts around kthread_run in
> > kmemleak_cleanup() would solve the problem. Otherwise, maybe the
> > kmemleak clean-up thread should take a different form or just a thread
> > waiting for a clean-up event (it currently acquires a mutex and cannot
> > be used in interrupt context).
> 
> It looks like the kthread_create_lock cannot be acquired in interrupt
> context anyway, so the patch below changes this to a workqueue.
> 
> 
> kmemleak: Do no create the clean-up thread during kmemleak_disable()

I'm not going to be able to test as it hasn't happened again.  I do
remember seeing messages about kmem leaks being found around the time of
the issue.  Although I was working the box as hard as I could, so maybe
I accidentally OOM'd it.  Wish there were more I could do to test or
confirm!

-Eric