Re: Using RCU with rcu_read_lock()?

["Paul E. McKenney" <paulmck@linux.vnet.ibm.com>]

On Fri, Jun 15, 2007 at 04:25:02PM -0400, Dmitry Torokhov wrote:
> On 6/15/07, Paul E. McKenney <paulmck@linux.vnet.ibm.com> wrote:
> >On Sat, Jun 16, 2007 at 12:59:40AM +0530, Dipankar Sarma wrote:
> >> On Fri, Jun 15, 2007 at 09:04:19PM +0200, Peter Zijlstra wrote:
> >> > On Fri, 2007-06-15 at 15:00 -0400, Dmitry Torokhov wrote:
> >> > > Hi,
> >> > >
> >> > > I have a piece of code that is always called under a spinlock with
> >> > > interrups disabled. Within that piece of code I iterate through a
> >> > > list. I have another piece of code that wants to modify that list. I
> >> > > have 2 options:
> >> > >
> >> > > I don't want to do 1) because the otheir piece of code does not 
> >really
> >> > > care about object owning the spinlock and so acquiring the spinlock 
> >is
> >> > > "not nice". However it is guaranteed that the piece of code that
> >> > > accesses lock runs atomically with interrupts disabled. So
> >> > > rcu_read_lock() would be superfluos there.
> >> > >
> >> > > Is it possible to still use list_for_each_rcu() and friends to access
> >> > > that list without rcu_read_lock()? Or it is betteruse complete RCU
> >> > > interface and eat cost of couple of extra instrctions?
> >> >
> >> > Yes, preemptible rcu requires that you use the full interface, also, it
> >> > more clearly documents the code. Trying to find code that breaks these
> >> > assumptions is very tedious work after the fact.
> >> >
> >> > Please do use the RCU interface in full.
> >>
> >> As Peter said, you should use the strict RCU APIs and not rely
> >> on the current implementation of RCU to optimize. Things change.
> >> Plus static/dynamic checking becomes easier that way.
> >
> >What they said!!!
> >
> >There are a couple of other options, however:
> >
> >1.      Use preempt_disable() and preempt_enable() on the read side,
> >       and synchronize_sched() on the update side.
> >
> >2.      Use local_irq_save() and local_irq_restore() on the read side,
> >       and synchronize_sched() on the update side.  Usually not
> >       competitive -- unless interrupts needed to be disabled for some
> >       other reason anyway.  Which you in fact say that you do.
> 
> Right. The callsite that iterates through the list is essentially
> protected by spin_lock_irqsave()/spin_unlock_irqrestore() - needed for
> other reasons (such as updating internal state of a device - and that
> can happen from different contexts).

That will work!

> >I believe that #2 might do what you want.  But please, PLEASE carefully
> >comment this usage!!!
> 
> Would there be a reson not to use #2 but rather full RCU with
> rcu_read_lock()/synchronize_rcu()?

Probably not, but here are a couple of situations where the full RCU
might be preferred:

1.	If you were relying on interrupts being disabled within an
	interrupt handler (which they are -not- in -rt), then you would
	either need to add some form of local_irq_save() or, as you say,
	go to the rcu_read_lock() and synchronize_rcu() interfaces.

2.	If updates needed to use callbacks rather than synchronous waits
	for grace periods, in other words, if you needed call_rcu()
	instead of synchronize_rcu().  Of course, a callback API for
	_sched (call_rcu_sched() or some such) could be added if needed,
	though it would be better to avoid the API proliferation unless
	really badly needed.

						Thanx, Paul