Re: [PATCH 04/13] tracing: Rename kvfree_rcu() to kvfree_rcu_mightsleep()

[Joel Fernandes]

On Wed, Mar 15, 2023 at 11:44 PM Paul E. McKenney <paulmck@xxxxxxxxxx> wrote:
>
> On Wed, Mar 15, 2023 at 10:23:23PM -0400, Steven Rostedt wrote:
> > On Wed, 15 Mar 2023 17:37:30 -0700
> > "Paul E. McKenney" <paulmck@xxxxxxxxxx> wrote:
> >
> > >
> > > That does work, and I am guessing that the size increase is not a big
> > > problem for you there.
> >
> > Well, I was fine with it as long as it stayed in the headers, where
> > ugliness is warmly welcomed. Just ask all the #ifdefs.
> >
> > >
> > > > That's a cop out, just removing the one case you care about. Fact is
> > > > the naming is awful, and the 1/2 argument thing is making it worse.
> > > > If a big change is warranted, why not do it right and ACTUALLY
> > > > get it right?
> > >
> > > You both do realize that the kvfree_rcu_mightsleep() definition is
> > > already in mainline, right?
> > >
> > > Anyway, to sum up, kvfree_rcu_mightsleep()--or whatever the entire
> > > community eventually decides to name it--can do any of the following:
> > >
> > > 1.  Put the pointer into an already allocated array of pointers.
> > >
> > > 2.  Allocate a new array of pointers, have the allocation succeed
> > >     without sleeping, then put the pointer into an already allocated
> > >     array of pointers.
> > >
> > > 3.  Allocate a new array of pointers, have the allocation succeed
> > >     after sleeping, then put the pointer into an already allocated
> > >     array of pointers.
> > >
> > > 4.  Attempt to allocate a new array of pointers, have the allocation
> > >     fail (presumably after sleeping), then invoke synchronize_rcu()
> > >     directly.
> > >
> > > Too much fun!  ;-)
> > >
> >
> >   kvfree_rcu_kitchen_sink() ?
> >
> >   kvfree_rcu_goldie_locks()?
> >
> > I honestly like the name "headless" as that perfectly describes the
> > difference between kvfree_rcu(arg1, arg2) and kvfree_rcu(arg1).
> >
> > Whereas mightsleep() is confusing to me because it doesn't tell me why
> > kvfree_rcu() has two args and kvfree_rcu_mightsleep() has only one.
> > Usually, code that has two sleep variants is about limiting the
> > functionality of the atomic friendly one.
>
>         kvfree_rcu_alloc_head()?
>         kvfree_rcu_dynhead()?
>         kvfree_rcu_gearhead()?
>         kvfree_rcu_radiohead()?
>         kvfree_rcu_getahead()?
>
> I don't know about you guys, but to me, kvfree_rcu_mightsleep() is
> sounding better and better by comparison...

Indeed, and one could argue that "headless" sounds like something out
of a horror movie ;-). Which of course does match the situation when
the API is applied incorrectly.

 - Joel