Re: [Patch] jbd commit code deadloop when installing Linux

[Andrew Morton <akpm@osdl.org>]

On 28 Jun 2006 14:50:29 +0800
Zou Nan hai <nanhai.zou@intel.com> wrote:

> On Wed, 2006-06-28 at 16:04, Andrew Morton wrote:
> > On 28 Jun 2006 14:02:57 +0800
> > Zou Nan hai <nanhai.zou@intel.com> wrote:
> > 
> > > > > However I think cond_resched_lock and cond_resched_softirq also need fix
> > > > > to make the semantic consistent.
> > > > > 
> > > > > Please check the following patch.
> > > > > 
> > > > 
> > > > Ah.  I think the return value from these functions should mean "something
> > > > disruptive happened", if you like.
> > > > 
> > > > See, the callers of cond_resched_lock() aren't interested in whether
> > > > cond_resched_lock() actually called schedule().  They want to know whether
> > > > cond_resched_lock() dropped the lock.  Because if the lock was dropped, the
> > > > caller needs to take some special action, regardless of whether schedule()
> > > > was finally called.
> > > > 
> > > > So I think the patch I queued is OK, agree?
> > > 
> > >   I am afraid the code like cond_resched_lock check in
> > > fs/jbd/checkpoint.c log_do_checkpoint may fall into endless retry in
> > > some condition, will it?
> > 
> > Oh crap, yes.  If need_resched() and system_state==SYSTEM_BOOTING then
> > cond_resched_lock() will drop the lock but won't schedule.  So it'll return
> > true but won't clear need_resched() and the caller will lock up.
> > 
> > So if cond_resched_foo() ends up dropping the lock it _must_ call
> > schedule() to clear need_resched().
> > 
> > So, how about this (it needs some code comments!)
> > 
> > 
> 
>  The patch works for the install test env.

Thanks.

>  However I still have some concern on cond_resched_lock(), on an UP 
> kernel it will return 1 if schedule happen, but actually it does not
> drop any lock, that semantic seems to be different to SMP kernel. 

That's OK (I think - I don't have a good track record in this thread).

If the kernel is non-preemptible and UP, we want to return true from
cond_resched_foo() if we called schedule().  Because schedule() might allow
a different thread into the kernel which might modify the locked data.

And if the kernel is preemptible and UP, we want to return true from
cond_resched_foo() if we dropped the lock, because that internally does a
preempt_enable().

And the patch (hopefully) satisfies those requirements.  Does that all
sound solid?