On Mon, Jun 26, 2023 at 10:55 AM David Vernet <void@xxxxxxxxxxxxx> wrote:
>
> > > > +
> > > > + migrate_disable();
> > > > + bpf_mem_cache_free_rcu(&bpf_cpumask_ma, cpumask);
> > > > + migrate_enable();
> > >
> > > The fact that callers have to disable migration like this in order to
> > > safely free the memory feels a bit leaky. Is there any reason we can't
> > > move this into bpf_mem_{cache_}free_rcu()?
> >
> > migrate_disable/enable() are actually not necessary here.
> > We can call bpf_mem_cache_free_rcu() directly from any kfunc.
>
> Could you please clarify why? Can't we migrate if the kfunc is called
> from a sleepable struct_ops callback?
migration is disabled for all bpf progs including sleepable.
> If migration is always disabled
> for any kfunc then I agree these migrate_{en,dis}able() calls can be
> removed. Otherwise from my reading of the code we'd race between calling
> this_cpu_ptr() and the local_irq_save() in unit_free().
>
> Thanks,
> David
>
> > Explicit migrate_disable() is only necessary from syscall.
> >
> > I believe rcu callbacks also cannot migrate, so the existing
> > code probably doesn't need them either.
