On Tue, Jun 27, 2023 at 10:56:21AM -0700, Linus Torvalds wrote:
> On Sun, 25 Jun 2023 at 08:35, Paul E. McKenney <paulmck@xxxxxxxxxx> wrote:
> >
> > git://git.kernel.org/pub/scm/linux/kernel/git/paulmck/linux-rcu.git tags/rcu.2023.06.22a
> >
> > o Eliminate the single-argument variant of k[v]free_rcu() now
> > that all uses have been converted to k[v]free_rcu_mightsleep().
>
> Well, clearly not all users had been.
>
> The base of this RCU was v6.4-rc1, and when that commit was done, we
> still had a single-argument variant:
>
> 7e3f926bf453 ("rcu/kvfree: Eliminate k[v]free_rcu() single argument macro")
>
> but look here:
>
> git grep 'kfree_rcu([^,()][^,()]*)' 7e3f926bf453
>
> results in
>
> 7e3f926bf453:drivers/infiniband/sw/rxe/rxe_verbs.c: kfree_rcu(mr);
>
> so the RCU tree itself can not possibly have built cleanly.
>
> How the heck did this pass testing in linux-next? Did linux-next just
> assume that it was a merge error, and fix it up?
Because idiot here failed to notice that the needed change was only
in -next, and not yet in mainline.
What I needed to have done instead was to keep this commit in -next,
but not send it to mainline until the v6.6 merge window. Or maybe to
send it as a separate pull request once the rdma commit hit mainline.
> Anyway, I *did* fix it up, changing the 'kfree_rcu()' to
> 'kfree_rcu_mightsleep()', but no, this was not a merge artifact. This
> was purely "the RCU tree did not build on its own", and as a result
> the tree does not bisect cleanly if you have rdma enabled.
>
> Adding rdma people to the participants just to let them know that this
> happened, but it's not their fault. This is on the RCU tree, and lack
> of proper coverage testing.
Apologies to all for my confusion, and thank you Linus for cleaning up
my mess!
Thanx, Paul
