memory barrier in counters_zero

[Alexander Bluhm]

Hi,

Inspired by Taylor's talk at EuroBSDCon I think a memory barrier
in counters_zero() is missing.  Reading uses two consumer barriers,
so writing should also have two.

Following code would have no barrier between writing generation
number and writing counters.

counters_leave();
counters_zero();

counters_leave() writes to generation number at the end, so
counters_zero() needs a barrier at the start.

ok?

bluhm

Index: kern/subr_percpu.c
===================================================================
RCS file: /data/mirror/openbsd/cvs/src/sys/kern/subr_percpu.c,v
retrieving revision 1.9
diff -u -p -r1.9 subr_percpu.c
--- kern/subr_percpu.c  10 Mar 2021 10:21:47 -0000      1.9
+++ kern/subr_percpu.c  17 Sep 2022 14:17:34 -0000
@@ -213,6 +213,7 @@ counters_zero(struct cpumem *cm, unsigne
        unsigned int i;
 
        counters = cpumem_first(&cmi, cm);
+       membar_producer();
        do {
                for (i = 0; i < n; i++)
                        counters[i] = 0;