On Sat, Aug 11, 2018 at 12:25:39AM +1000, Rashmica Gupta wrote:
> On Fri, Aug 10, 2018 at 11:00 PM, Michal Hocko <mhocko@xxxxxxxxxx> wrote:
> > On Fri 10-08-18 16:55:40, Rashmica Gupta wrote:
> > [...]
> >> Most memory hotplug/hotremove seems to be block or section based, and
> >> always adds and removes memory at the same place.
> >
> > Yes and that is hard wired to the memory hotplug code. It is not easy to
> > make it work outside of section units restriction. So whatever your
> > memtrace is doing and if it relies on subsection hotplug it cannot
> > possibly work with the current code.
> >
> > I didn't get to review your patch but if it is only needed for an
> > unmerged code I would rather incline to not merge it unless it is a
> > clear win to the resource subsystem. A report from Oscar shows that this
> > is not the case though.
> >
>
> Yup, makes sense. I'll work on it and see if I can not break things.
In __case__ we really need this patch, I think that one way to fix this is
to only call merge_node_resources() in case the node is already online.
Something like this (completely untested):
+struct resource *request_resource_and_merge(struct resource *parent,
+ struct resource *new, int nid)
+{
+ struct resource *conflict;
+
+ conflict = request_resource_conflict(parent, new);
+
+ if (conflict)
+ return conflict;
+
+#ifdef CONFIG_MEMORY_HOTREMOVE
+ /* We do not need to merge any resources on a node that is being
+ * hot-added together with its memory.
+ * The node will be allocated later.
+ */
+ if (node_online(nid))
+ merge_node_resources(nid, parent);
+#endif /* CONFIG_MEMORY_HOTREMOVE */
Although as Michal said, all memory-hotplug code is section-oriented, so
whatever it is that interacts with it should expect that.
Otherwise it can fail soon or later.
--
Oscar Salvador
SUSE L3
