On Mon, Jun 17, 2024 at 10:24:54PM -0400, Kent Overstreet wrote:
> On Mon, Jun 17, 2024 at 05:15:29PM -0700, Paul E. McKenney wrote:
> > If it ever becomes helpful to make this work for any of the RCU Tasks
> > flavors, the rcutorture_type enum from kernel/rcu/rcu.h might be of help.
> > Until then, NULL works just fine.
>
> Certainly
>
> > > - Tracks pending items in a radix tree; falls back to linked list if
> > > radix tree allocation fails
> >
> > I am still having trouble seeing what a radix tree brings to the table
> > in this case, but again, migtht as well dig in.
>
> It's simpler, and (marginally) more efficient than list-of-pages.
>
> My new implementation of kvfree_call_rcu() on top of this data structure
> is now looking faster than the existing version - and I'm using smaller
> nodes (512 bytes instead of full pages).
OK. We will need to look carefully at the performance comparisons.
Lots of places for overhead to hide in both schemes. ;-)
> > > +static inline unsigned long __start_poll_synchronize_rcu(struct srcu_struct *ssp)
> > > +{
> > > + return ssp
> > > + ? start_poll_synchronize_srcu(ssp)
> > > + : start_poll_synchronize_rcu();
> > > +}
> >
> > I don't get why you
> >
> > > +static inline void __rcu_barrier(struct srcu_struct *ssp)
> > > +{
> > > + return ssp
> > > + ? srcu_barrier(ssp)
> > > + : rcu_barrier();
> > > +}
> >
> > You will likely need something like these:
> >
> > static inline void __get_completed_synchronize_rcu(struct srcu_struct *ssp)
> > {
> > return ssp
> > ? get_completed_synchronize_srcu(ssp)
> > : get_completed_synchronize_rcu();
> > }
> >
> > static inline void __same_state_synchronize_rcu(struct srcu_struct *ssp)
> > {
> > return ssp
> > ? same_state_synchronize_srcu(ssp)
> > : same_state_synchronize_rcu();
> > }
> >
> > The point of these is to get around the earlier issue where three cookies
> > appeared to be valid due to time delays. But perhaps you found some
> > other way around that.
>
> Yes, that comes from a race between poll_state_synchronize() and
> start_poll_synchronize(), easily handled with a loop.
Plus call_srcu() can also get its oar in. But more later.
> To be honest, I'm still not getting your intended use of the other
> helpers; to be honest, just comparing sequence numbers with
> ULONG_CMP_GE/LT seems simplest to my brain.
