Re: [PATCH 00/13] Rename k[v]free_rcu() single argument to k[v]free_rcu_mightsleep()

From: Paul E. McKenney
Date: Thu Feb 23 2023 - 13:31:29 EST


On Thu, Feb 23, 2023 at 07:57:13AM -0700, Jens Axboe wrote:
> On 2/1/23 8:08 AM, Uladzislau Rezki (Sony) wrote:
> > This small series is based on Paul's "dev" branch. Head is 6002817348a1c610dc1b1c01ff81654cdec12be4
> > it renames a single argument of k[v]free_rcu() to its new k[v]free_rcu_mightsleep() name.
> >
> > 1.
> > The problem is that, recently we have run into a precedent when
> > a user intended to give a second argument to kfree_rcu() API but
> > forgot to do it in a code so a call became as a single argument
> > of kfree_rcu() API.
> >
> > 2.
> > Such mistyping can lead to hidden bags where sleeping is forbidden.
> >
> > 3.
> > _mightsleep() prefix gives much more information for which contexts
> > it can be used for.
>
> This patchset seems weird to me. We have a LOT of calls that might
> sleep, yet we don't suffix them all with _mightsleep(). Why is this
> any different? Why isn't this just a might_sleep() call in the
> actual helper, which will suffice for checkers and catch it at
> runtime as well.

Fair enough, and the situation that this patchset is addressing is also a
bit unusual. This change was requested by Eric Dumazet due to a situation
where someone forgot the optional second argument to kfree_rcu(). Now,
you are right that this would be caught if invoked from a non-sleepable
context, but there are also cases where sleeping is legal, but where the
occasional wait for an RCU grace period would be a problem. The checkers
cannot easily catch this sort of thing, and hence the change in name.

Hey, the combined one/two-argument form seemed like a good idea at
the time! ;-)

Thanx, Paul