On Thu, Jul 08, 2021 at 08:52:23PM -0700, Alexei Starovoitov wrote:
> On Thu, Jul 08, 2021 at 06:51:19PM -0700, Martin KaFai Lau wrote:
> > > +
> > > /* Called when map->refcnt goes to zero, either from workqueue or from syscall */
> > > static void array_map_free(struct bpf_map *map)
> > > {
> > > @@ -382,6 +402,7 @@ static void array_map_free(struct bpf_map *map)
> > > if (array->map.map_type == BPF_MAP_TYPE_PERCPU_ARRAY)
> > > bpf_array_free_percpu(array);
> > >
> > > + array_map_free_timers(map);
> > array_map_free() is called when map->refcnt reached 0.
> > By then, map->usercnt should have reached 0 before
> > and array_map_free_timers() should have already been called,
> > so no need to call it here again? The same goes for hashtab.
>
> Not sure it's that simple.
> Currently map->usercnt > 0 check is done for bpf_timer_set_callback only,
> because prog refcnting is what matters to usercnt and map_release_uref scheme.
> bpf_map_init doesn't have this check because there is no circular dependency
> prog->map->timer->prog to worry about.
> So after usercnt reached zero the prog can still do bpf_timer_init.
Ah. right. missed the bpf_timer_init().
> I guess we can add usercnt > 0 to bpf_timer_init as well.
> Need to think whether it's enough and the race between atomic64_read(usercnt)
> and atomic64_dec_and_test(usercnt) is addressed the same way as the race
> in set_callback and cancel_and_free. So far looks like it. Hmm.
hmm... right, checking usercnt > 0 seems ok.
When usercnt is 0, it may be better to also error out instead of allocating
a timer that cannot be used.
I was mostly thinking avoiding changes in map_free could make future map
support a little easier.
>
> >
> > > +static void htab_free_malloced_timers(struct bpf_htab *htab)
> > > +{
> > > + int i;
> > > +
> > > + rcu_read_lock();
> > > + for (i = 0; i < htab->n_buckets; i++) {
> > > + struct hlist_nulls_head *head = select_bucket(htab, i);
> > > + struct hlist_nulls_node *n;
> > > + struct htab_elem *l;
> > > +
> > > + hlist_nulls_for_each_entry(l, n, head, hash_node)
> > May be put rcu_read_lock/unlock() in the loop and do a
> > cond_resched() in case the hashtab is large.
Just recalled cond_resched_rcu() may be cleaner, like:
static void htab_free_malloced_timers(struct bpf_htab *htab)
{
int i;
rcu_read_lock();
for (i = 0; i < htab->n_buckets; i++) {
/* ... */
hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
check_and_free_timer(htab, l);
cond_resched_rcu();
}
rcu_read_unlock();
}
>
> Feels a bit like premature optimization. delete_all_elements()
> loop right above is doing similar work without cond_resched.
> I don't mind cond_resched. I just don't see how to cleanly add it
> without breaking rcu_read_lock and overcomplicating the code.
yep, it can be done later together with delete_all_elements().
