On Fri, May 29, 2020 at 08:50:25AM -0700, Andi Kleen wrote:
> >
> > ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
> > - if (ret == 0 && write)
> > + if (ret == 0 && write) {
> > + if (sysctl_overcommit_memory == OVERCOMMIT_NEVER)
> > + schedule_on_each_cpu(sync_overcommit_as);
>
> The schedule is not atomic.
> There's still a race window here over all the CPUs where the WARN_ON could
> happen because you change the global first.
The re-computing of batch number comes after this sync, so at this
point the batch is still the bigger one, and won't trigger the warning.
> Probably you would need another global that says "i'm currently changing
> the mode" and then skip the WARN_ON in that window. Maybe a sequence lock.
>
> Seems all overkill to me. Better to kill the warning.
Yes, the cost is high, schedule_on_each_cpu is labeled as "very slow"
in the code comments.
Thanks,
Feng
>
> -Andi
