On Fri, Nov 17, 2006 at 02:27:15PM -0500, Alan Stern wrote:
> On Fri, 17 Nov 2006, Paul E. McKenney wrote:
>
> > > It works for me, but the overhead is still large. Before it would take
> > > 8-12 jiffies for a synchronize_srcu() to complete without there actually
> > > being any reader locks active, now it takes 2-3 jiffies. So it's
> > > definitely faster, and as suspected the loss of two of three
> > > synchronize_sched() cut down the overhead to a third.
> >
> > Good to hear, thank you for trying it out!
> >
> > > It's still too heavy for me, by far the most calls I do to
> > > synchronize_srcu() doesn't have any reader locks pending. I'm still a
> > > big advocate of the fastpath srcu_readers_active() check. I can
> > > understand the reluctance to make it the default, but for my case it's
> > > "safe enough", so if we could either export srcu_readers_active() or
> > > export a synchronize_srcu_fast() (or something like that), then SRCU
> > > would be a good fit for barrier vs plug rework.
> >
> > OK, will export the interface. Do your queues have associated locking?
> >
> > > > Attached is a patch that compiles, but probably goes down in flames
> > > > otherwise.
> > >
> > > Works here :-)
> >
> > I have at least a couple bugs that would show up under low-memory
> > situations, will fix and post an update.
>
> Perhaps a better approach to the initialization problem would be to assume
> that either:
>
> 1. The srcu_struct will be initialized before it is used, or
>
> 2. When it is used before initialization, the system is running
> only one thread.
Are these assumptions valid? If so, they would indeed simplify things
a bit.
> In other words, statically allocated SRCU strucures that get used during
> system startup must be initialized before the system starts multitasking.
> That seems like a reasonable requirement.
>
> This eliminates worries about readers holding mutexes. It doesn't
> solve the issues surrounding your hardluckref, but maybe it makes them
> easier to think about.
For the moment, I cheaped out and used a mutex_trylock. If this can block,
I will need to add a separate spinlock to guard per_cpu_ref allocation.
Hmmm... How to test this? Time for the wrapper around alloc_percpu()
that randomly fails, I guess. ;-)
Thanx, Paul
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to [EMAIL PROTECTED]
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
