Re: [PATCH rcu] srcu: Guarantee non-negative return value from srcu_read_lock()

[Paul E. McKenney]

On Wed, Oct 23, 2024 at 02:58:07PM +0800, Alan Huang wrote:
> On Oct 22, 2024, at 22:26, Paul E. McKenney <paulmck@xxxxxxxxxx> wrote:
> > 
> > On Tue, Oct 22, 2024 at 12:13:12AM -0700, Christoph Hellwig wrote:
> >> On Tue, Oct 22, 2024 at 09:10:18AM +0200, Peter Zijlstra wrote:
> >>> Ah, well, the thing that got us here is that we (Andrii and me) wanted
> >>> to use -1 as an 'invalid' value to indicate SRCU is not currently in
> >>> use.
> >>> 
> >>> So it all being int is really rather convenient :-)
> >> 
> >> Then please document that use.  Maybe even with a symolic name for
> >> -1 that clearly describes these uses.
> > 
> > Would this work?
> > 
> > #define SRCU_INVALID_INDEX -1
> 
> Is there any similar guarantee of the return value of get_state_synchronize_rcu
> or start_poll_synchronize_rcu, like invalid value?

Yes, there is a get_completed_synchronize_rcu() function that returns a
value that causes poll_state_synchronize_rcu() to always return true.
There is also a get_completed_synchronize_rcu_full() function that
returns a value that causes poll_state_synchronize_rcu_full() to always
return true.

There has been some discussion of another set of values that cause
poll_state_synchronize_rcu() and poll_state_synchronize_rcu_full() to
always return false, but there is not yet a use case for this.  Easy to
provide if required, but why further explode the RCU API unless it really
is required?

							Thanx, Paul

> > Whatever the name, maybe Peter and Andrii define this under #ifndef
> > right now, and we get it into include/linux/srcu.h over time.
> > 
> > Or is there a better way?  Or name, for that matter.
> > 
> > Thanx, Paul
> > 
>