On Wed, Feb 19, 2025 at 10:57:29AM -0500, Joel Fernandes wrote:
> On 2/19/2025 8:54 AM, Paul E. McKenney wrote:
> >>> the system, by scaling the default by the number of CPUs. After all,
> >>> we want to test both the ->gpwrap path *and* the non-->gpwrap path.
> >> Agreed. But do we want to test both these paths only under rcutorture
> >> testing or
> >> were you considering we do it for regular operation as well? i.e. wrap in
> >> 10
> >> minutes in regular operation. I am guessing you want it only for rcutorture
> >> _and_ have a 10 minute default.
> >>
> >> If that is the case, we would then need rcu_gpnum_ovf() to consult
> >> rcutorture,
> >> if it is enabled, about what what the threshold is.
> > The rcu_gp_slow() function does something similar for the grace-period
> > kthread's injected delays, so that can work. Probably better than putting
> > the module parameter in tree.c, because putting it in rcutorture makes
> > in more clear that it is for debugging.
>
> Ok thanks, just to clarify:
>
> Normal operation is still ULONG/2 threshold.
Yes, please keep this be the default.
> rcutorture operation will be 10 minutes (overridable by rcutorture module
> option, not module option in non-test kernel code itself).
That sounds good also.
> If we are in disagreement about this, please let me know. Otherwise I will
> work
> on a patch as a next step on this.
Very good, looking forward to seeing what you come up with!
Thanx, Paul
