Re: Kernel RCU: shrink the size of the struct rcu_head

[Mathieu Desnoyers <mathieu.desnoyers@polymtl.ca>]

* Paul E. McKenney (paulmck@linux.vnet.ibm.com) wrote:
> On Sun, Oct 18, 2009 at 07:29:18PM -0400, Mathieu Desnoyers wrote:
> > Hi Paul,
> > 
> > I noticed that you already discussed the possibility of shrinking the
> > struct rcu_head by removing the function pointer.
> > (http://kernel.org/pub/linux/kernel/people/paulmck/rcutodo.html)
> > 
> > The ideas brought in so far require having per-callback lists, which
> > involves a bit of management overhead and don't permit keeping the
> > call_rcu() in cpu order.
> 
> But please note that this is on the "Possibly Dubious Changes" list.  ;-)
> 
> > You might want to look into the Userspace RCU urcu-defer.c
> > implementation, where I perform pointer encoding to compact the usual
> > case, expected to be the same callback passed as parameter multiple
> > times in a row to call_rcu(). This is very typical with multiple free()
> > calls for different data structures next to each other.
> > 
> > This typically keeps the size of the information to encode per callback
> > down to a minimum: the size of a single pointer. It would be good to
> > trace the kernel usage of call_rcu() to see if my assumption holds.
> > 
> > I just thought I should tell you before you start looking at this
> > issue further.
> 
> So the idea is to maintain a per-CPU queue of function pointers, but
> with the pointers on this queue encoded to save space, correct?

Yes, exactly.

>  If I
> understand correctly, the user-level rcu-defer implementation relies on
> the following:
> 
> 1.	It is illegal to call _rcu_defer_queue() within an RCU read-side
> 	critical section (due to the call to rcu_defer_barrier_thread()
> 	which in turn calls synchronize_rcu().  This is necessary to
> 	handle queue overflow.  (Which appears to be why you introduce
> 	a new API, as it is legal to invoke call_rcu() from within an
> 	RCU read-side critical section.)

When dealing with queue overflow, I figured we have 4 alternatives.
Either:

1, 2, 3) We proceed to execution of {the single, all, thread local}
         callback(s) on the spot after a synchronize_rcu().

4) We expand the queue by allocating more memory.

The idea of pointer encoding to save space could be used with any of 1,
2, 3, or 4. As you say, call_rcu() requires (4), because it tolerates
being called from an rcu read-side C.S.. 1, 2, 3 are incompatible with
read-side C.S. context because they require to use synchronize_rcu()
within the C.S., which would deadlock on its calling context.

Now, there is a rationale for the choice of (3) in my urcu-defer
implementation:

* It's how I can deal with memory full (-ENOMEM) without letting the
  system die with exit(). How does the kernel call_rcu() deal with this
  currently ? BUG_ON, WARN_ON ?

* It acts as a rate limiter for urcu_defer_queue(). Basically, if a
  thread starts enqueuing callbacks too fast, it will eventually fill its
  queue and have to empty it itself. AFAIK, It's not possible to do that
  if you allow call_rcu() to be called from read-side C.S..

I could even extend rcu_defer_queue() to take a second rate-limiter
callback, which would check if the thread went over some threshold and
give a more precise limit (e.g. amount of memory to be freed) on the
rate than the "4096 callbacks in flight max", which have been chosen by
benchmarks, but is a bit arbitrary in terms of overall callback effect.

How important is it to permit enqueuing callbacks from within rcu
read-side C.S. in terms of real-life usage ? If it is really that
important to fill this use-case, then I could have a mode for call_rcu()
that expands the RCU callback queue upon overflow. But as I argue above,
I really prefer the control we have with a fixed-sized queue.

> 
> 2.	It is OK to wait for a grace period when a thread calls
> 	rcu_defer_unregister_thread() while exiting.  In the kernel,
> 	this is roughly equivalent to the CPU_DYING notifier, which
> 	cannot block, thus cannot wait for a grace period.
> 
> 	I could imagine copying the per-CPU buffer somewhere, though
> 	my experience with the RCU/CPU-hotplug interface does not
> 	encourage me in this direction.  ;-)

As you say, we don't _have_ to empty the queue before putting a
thread/cpu offline. We could simply copy the unplugged cpu queue to an
orphan queue, as you currently do in your implementation. I agree that
it would be more suitable to the cpu hotplug CPU_DYING execution
context, due to its inherent trickiness.

Thanks,

Mathieu

> 
> 							Thanx, Paul

-- 
Mathieu Desnoyers
OpenPGP key fingerprint: 8CD5 52C3 8E3C 4140 715F  BA06 3F25 A8FE 3BAE 9A68