On Thu, 2019-09-26 at 13:52 +0200, Michal Hocko wrote: > On Thu 26-09-19 07:19:27, Qian Cai wrote: > > > > > > > On Sep 26, 2019, at 3:26 AM, Michal Hocko <mhocko@xxxxxxxxxx> wrote: > > > > > > OK, this is using for_each_online_cpu but why is this a problem? Have > > > you checked what the code actually does? Let's say that online_pages is > > > racing with cpu hotplug. A new CPU appears/disappears from the online > > > mask while we are iterating it, right? Let's start with cpu offlining > > > case. We have two choices, either the cpu is still visible and we update > > > its local node configuration even though it will disappear shortly which > > > is ok because we are not touching any data that disappears (it's all > > > per-cpu). Case when the cpu is no longer there is not really > > > interesting. For the online case we might miss a cpu but that should be > > > tolerateable because that is not any different from triggering the > > > online independently of the memory hotplug. So there has to be a hook > > > from that code path as well. If there is none then this is buggy > > > irrespective of the locking. > > > > > > Makes sense? > > > > This sounds to me requires lots of audits and testing. Also, someone who is more > > familiar with CPU hotplug should review this patch. > > Thomas is on the CC list. > > > Personally, I am no fun of > > operating on an incorrect CPU mask to begin with, things could go wrong really > > quickly... > > Do you have any specific arguments? Just think of cpu and memory > hotplugs being independent operations. There is nothing really > inherently binding them together. If the cpu_online_mask really needs a > special treatment here then I would like to hear about that. Handwaving > doesn't really helps us. That is why I said it needs CPU hotplug experts to confirm that things including if CPU masks are tolerate to this kind of "abuse", or in-depth analysis of each calls sites that access CPU masks in both online_pages() and offline_pages() as well as ideally, more testing data in those areas. However, many kernel commits were merged with the expectations that people are going to deal with the aftermath, so I am not going to insist.
