On Tue, Nov 8, 2022 at 5:28 AM Hou Tao <houtao@xxxxxxxxxxxxxxx> wrote: > On 11/8/2022 3:03 PM, Yonghong Song wrote: > > On 11/7/22 8:08 PM, Hou Tao wrote: > >> On 11/8/2022 10:11 AM, Hao Luo wrote: <...> > >>> There is an alternative: does it make sense to have the iterator hold > >>> a ref of the link? When the link is closed, my assumption is that the > >>> program is already detached from the cgroup. After that, it makes no > >>> sense to still allow iterating the cgroup. IIUC, holding a ref to the > >>> link in the iterator also fixes for other types of objects. > >> > >> Also considered the alternative solution when fixing the similar problem in bpf > >> map element iterator [0]. The problem is not all of bpf iterators need the > >> pinning (e.g., bpf map iterator). Because bpf prog is also pinned by iterator fd > >> in iter_open(), so closing the fd of iterator link doesn't release the bpf > >> program. > >> > >> [0]: https://lore.kernel.org/bpf/20220810080538.1845898-2-houtao@xxxxxxxxxxxxxxx/ > > > > Okay, let us do the solution to hold a reference to the link for the iterator. > > For cgroup_iter, that means, both prog and cgroup will be present so we should > > be okay then. > > > The reason I did not use the solution is that it will create unnecessary > dependency between iterator fd and iterator link and many bpf iterators also > don't need that. If we use the solution, should I revert the fixes to bpf map > iterator done before or keep it as-is ? > > Hou Tao, on the contrary, I do think the dependency is necessary. My understanding is, the lifetime of an iterator should not go beyond the lifetime of the link who generates the iterator. You mention that many bpf iterators don't need that. I suspect there are bugs due to lack of such dependencies. Hypothetically, if the link is released, it may cause the program to be released as well. Then, how could we still iterate the objects and call the program on the objects? Please correct me if there is anything I missed.
