On Fri 23-01-15 09:17:44, Christoph Lameter wrote: > On Fri, 23 Jan 2015, Johannes Weiner wrote: > > > Is the assumption of this patch wrong? Does the specified node have > > to be online for the fallback to work? Admittedly, I was checking only SLAB allocator when reviewing and assuming SLUB would behave in the same way :/ But maybe I have misinterpreted the slab code as well and get_node(struct kmem_cache *, int node) returns non-NULL for !online nodes. > Nodes that are offline have no control structures allocated and thus > allocations will likely segfault when the address of the controls > structure for the node is accessed. > > If we wanted to prevent that then every allocation would have to add a > check to see if the nodes are online which would impact performance. I have briefly checked the code and it seems that many users are aware of this and use the same construct Johannes used in the end or they use cpu_to_node. But then there are other users doing: net/openvswitch/flow_table.c: /* Initialize the default stat node. */ stats = kmem_cache_alloc_node(flow_stats_cache, GFP_KERNEL | __GFP_ZERO, 0); and this can blow up if Node0 is not online. I haven't checked other callers but are we sure they all are aware of !online nodes? E.g. dev_to_node() will return a node which is assigned to a device. I do not see where exactly this is set to anything else than -1 (I got quickly lost in set_dev_node callers). E.g. PCI bus sets its affinity from bus->sysdata which seems to be initialized in pci_acpi_scan_root and that is checking for an online node. Is it possible that some devices will get the node from BIOS or by other means? That being said I have no problem with checking node_online in the memcg code which was reported to blow up here. I am just thinking whether it is safe to simply blow up like that. -- Michal Hocko SUSE Labs -- To unsubscribe from this list: send the line "unsubscribe linux-next" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html