On 2022-08-29 17:48:05 [+0200], Maurizio Lombardi wrote: > diff --git a/mm/slub.c b/mm/slub.c > index 862dbd9af4f5..d46ee90651d2 100644 > --- a/mm/slub.c > +++ b/mm/slub.c > @@ -2681,30 +2681,34 @@ struct slub_flush_work { > bool skip; > }; > > +static void flush_cpu_slab(void *d) > +{ > + struct kmem_cache *s = d; > + struct kmem_cache_cpu *c = this_cpu_ptr(s->cpu_slab); > + > + if (c->slab) > + flush_slab(s, c); > + > + unfreeze_partials(s); > +} … > @@ -2721,13 +2725,18 @@ static void flush_all_cpus_locked(struct kmem_cache *s) > lockdep_assert_cpus_held(); > mutex_lock(&flush_lock); > > + if (in_task()) { > + on_each_cpu_cond(has_cpu_slab, flush_cpu_slab, s, 1); This blocks with disabled preemption until it completes flush_cpu_slab() on all CPUs. That function acquires a local_lock_t which can not be acquired from in-IRQ which is where this function will be invoked due to on_each_cpu_cond(). Couldn't we instead use a workqueue with that WQ_MEM_RECLAIM bit? It may reclaim memory after all ;) Sebastian