On Wed, 1 Aug 2018 15:04:16 -0500 Jeremy Linton <jeremy.lin...@arm.com> wrote:
> The thread "avoid alloc memory on offline node" > > https://lkml.org/lkml/2018/6/7/251 > > Asked at one point why the kzalloc_node was crashing rather than > returning memory from a valid node. The thread ended up fixing > the immediate causes of the crash but left open the case of bad > proximity values being in DSDT tables without corrisponding > SRAT/SLIT entries as is happening on another machine. > > Its also easy to fix that, but we should also harden the allocator > sufficiently that it doesn't crash when passed an invalid node id. > There are a couple possible ways to do this, and i've attached two > separate patches which individually fix that problem. > > The first detects the offline node before calling > the new_slab code path when it becomes apparent that the allocation isn't > going to succeed. The second actually hardens node_zonelist() and > prepare_alloc_pages() in the face of NODE_DATA(nid) returning a NULL > zonelist. This latter case happens if the node has never been initialized > or is possibly out of range. There are other places (NODE_DATA & > online_node) which should be checking if the node id's are > MAX_NUMNODES. > What is it that leads to a caller requesting memory from an invalid node? A race against offlining? If so then that's a lack of appropriate locking, isn't it? I don't see a problem with emitting a warning and then selecting a different node so we can keep running. But we do want that warning, so we can understand the root cause and fix it?
