Re: [PATCH tip/core/rcu 03/17] rcu/tree: Skip entry into the page allocator for PREEMPT_RT

[Joel Fernandes <joel@joelfernandes.org>]

On Tue, Jul 07, 2020 at 07:34:41PM +0200, Uladzislau Rezki wrote:
> On Mon, Jul 06, 2020 at 02:06:45PM -0700, Paul E. McKenney wrote:
> > On Thu, Jul 02, 2020 at 10:19:08PM +0200, Sebastian Andrzej Siewior wrote:
> > > On 2020-07-02 09:48:26 [-0700], Paul E. McKenney wrote:
> > > > On Thu, Jul 02, 2020 at 04:12:16PM +0200, Sebastian Andrzej Siewior wrote:
> > > > > On 2020-06-30 11:35:34 [-0700], Paul E. McKenney wrote:
> > > > > > > This is not going to work together with the "wait context validator"
> > > > > > > (CONFIG_PROVE_RAW_LOCK_NESTING). As of -rc3 it should complain about
> > > > > > > printk() which is why it is still disabled by default.
> > > > > > 
> > > > > > Fixing that should be "interesting".  In particular, RCU CPU stall
> > > > > > warnings rely on the raw spin lock to reduce false positives due
> > > > > > to race conditions.  Some thought will be required here.
> > > > > 
> > > > > I don't get this part. Can you explain/give me an example where to look
> > > > > at?
> > > > 
> > > > Starting from the scheduler-clock interrupt's call into RCU,
> > > > we have rcu_sched_clock_irq() which calls rcu_pending() which
> > > > calls check_cpu_stall() which calls either print_cpu_stall() or
> > > > print_other_cpu_stall(), depending on whether the stall is happening on
> > > > the current CPU or on some other CPU, respectively.
> > > > 
> > > > Both of these last functions acquire the rcu_node structure's raw ->lock
> > > > and expect to do printk()s while holding it.
> > > 
> > > …
> > > > Thoughts?
> > > 
> > > Okay. So in the RT queue there is a printk() rewrite which fixes this
> > > kind of things. Upstream the printk() interface is still broken in this
> > > regard and therefore CONFIG_PROVE_RAW_LOCK_NESTING is disabled.
> > > [Earlier the workqueue would also trigger a warning but this has been
> > > fixed as of v5.8-rc1.]
> > > This was just me explaining why this bad, what debug function would
> > > report it and why it is not enabled by default.
> > 
> > Whew!!!  ;-)
> > 
> > > > > > > So assume that this is fixed and enabled then on !PREEMPT_RT it will
> > > > > > > complain that you have a raw_spinlock_t acquired (the one from patch
> > > > > > > 02/17) and attempt to acquire a spinlock_t in the memory allocator.
> > > > > > 
> > > > > > Given that the slab allocator doesn't acquire any locks until it gets
> > > > > > a fair way in, wouldn't it make sense to allow a "shallow" allocation
> > > > > > while a raw spinlock is held?  This would require yet another GFP_ flag,
> > > > > > but that won't make all that much of a difference in the total.  ;-)
> > > > > 
> > > > > That would be one way of dealing with. But we could go back to
> > > > > spinlock_t and keep the memory allocation even for RT as is. I don't see
> > > > > a downside of this. And we would worry about kfree_rcu() from real
> > > > > IRQ-off region once we get to it.
> > > > 
> > > > Once we get to it, your thought would be to do per-CPU queuing of
> > > > memory from IRQ-off kfree_rcu(), and have IRQ work or some such clean
> > > > up after it?  Or did you have some other trick in mind?
> > > 
> > > So for now I would very much like to revert the raw_spinlock_t back to
> > > the spinlock_t and add a migrate_disable() just avoid the tiny
> > > possible migration between obtaining the CPU-ptr and acquiring the lock
> > > (I think Joel was afraid of performance hit).
> > 
> > Performance is indeed a concern here.
> > 
> > > Should we get to a *real* use case where someone must invoke kfree_rcu()
> > > from a hard-IRQ-off region then we can think what makes sense. per-CPU
> > > queues and IRQ-work would be one way of dealing with it.
> > 
> > It looks like workqueues can also be used, at least in their current
> > form.  And timers.
> > 
> > Vlad, Joel, thoughts?
> >
> Some high level thoughts:
> 
> Currently everything is done in workqueue context, it means all freeing
> happens there. For RT kernel we can invoke a page allocator only for single
> kfree_rcu() argument(though we skip it). As for double one, it is impossible,
> that is why a simple path is used by linking rcu_head among each other for
> further reclaim in wq context. As of now, for RT, everything is already
> deferred.
> 
> If we revert to spinlock_t then calling of kfree_rcu() from hard IRQ
> context is broken, even though we think that for RT kernel it will
> never happen. Therefore i do not see a clear motivation and benefits
> why we should revert to spinlock_t.
> 
> IMHO, if we can avoid of such drawback i would go with that way, i.e.
> i would not like to think what to do with that when it becomes broken.

I am also of Vlad's opinion. It seems to me that doing the migrate_disable()
before dropping the raw spinlock should suffice.

Note that in the future, we may have other users of this path such as
a potential kmem_cache_free_rcu(). It seems burdensome to make these all
callable only from sleepable-context.

Is there a drawback of doing migrate_disable before dropping the raw internal
lock, that I'm missing?

thanks,

 - Joel