On Fri, Jan 11, 2019 at 09:12:57AM +0100, Daniel Borkmann wrote:
> Hi Kris,
>
> On 01/11/2019 06:08 AM, Kris Van Hees wrote:
> > Maybe I am missing something trivial here, but it looks to me that there is
> > a leak of htab elements in htab_map_update_elem when you are updating an
> > existing element. After the new element is linked into the bucket list, the
> > following code snippet is found:
> >
> > if (l_old) {
> > hlist_nulls_del_rcu(&l_old->hash_node);
> > if (!htab_is_prealloc(htab))
> > free_htab_elem(htab, l_old);
> > }
> >
> > Nothing is done with l_old in the remainder of the function, and to me this
> > looks like that element is be leaked if the htab is preallocated because we
> > never add it to the free list. In fact, free_htab_elem() contains the very
> > conditional that handles the two cases (preallocated vs non-preallocated.
>
> In this case in alloc_htab_elem() we are swapping out the per-cpu extra
> element
> with the existing one to avoid freelist_pop/push combination in order to have
> a fast replace, meaning we cannot call free_htab_elem() on it like in plain
> htab_map_delete_elem() case.
Aha, and since it is a per-cpu extra element, you never run into a problem
with two updates using this trick because only one update can ever be ongoing
per cpu. Nifty!
Thanks for explaining!
Kris
