Re: [PATCH v3] mm: fix panic in __alloc_pages

[Michal Hocko <mhocko@suse.com>]

On Tue 07-12-21 17:09:50, David Hildenbrand wrote:
> On 07.12.21 16:56, Michal Hocko wrote:
> > On Tue 07-12-21 16:34:30, David Hildenbrand wrote:
> >> On 07.12.21 16:29, Michal Hocko wrote:
> >>> On Tue 07-12-21 16:09:39, David Hildenbrand wrote:
> >>>> On 07.12.21 14:23, Michal Hocko wrote:
> >>>>> On Tue 07-12-21 13:28:31, David Hildenbrand wrote:
> >>>>> [...]
> >>>>>> But maybe I am missing something important regarding online vs. offline
> >>>>>> nodes that your patch changes?
> >>>>>
> >>>>> I am relying on alloc_node_data setting the node online. But if we are
> >>>>> to change the call to arch_alloc_node_data then the patch needs to be
> >>>>> more involved. Here is what I have right now. If this happens to be the
> >>>>> right way then there is some additional work to sync up with the hotplug
> >>>>> code.
> >>>>>
> >>>>> diff --git a/mm/page_alloc.c b/mm/page_alloc.c
> >>>>> index c5952749ad40..a296e934ad2f 100644
> >>>>> --- a/mm/page_alloc.c
> >>>>> +++ b/mm/page_alloc.c
> >>>>> @@ -8032,8 +8032,23 @@ void __init free_area_init(unsigned long *max_zone_pfn)
> >>>>>  	/* Initialise every node */
> >>>>>  	mminit_verify_pageflags_layout();
> >>>>>  	setup_nr_node_ids();
> >>>>> -	for_each_online_node(nid) {
> >>>>> -		pg_data_t *pgdat = NODE_DATA(nid);
> >>>>> +	for_each_node(nid) {
> >>>>> +		pg_data_t *pgdat;
> >>>>> +
> >>>>> +		if (!node_online(nid)) {
> >>>>> +			pr_warn("Node %d uninitialized by the platform. Please report with memory map.\n", nid);
> >>>>> +			pgdat = arch_alloc_nodedata(nid);
> >>>>> +			pgdat->per_cpu_nodestats = alloc_percpu(struct per_cpu_nodestat);
> >>>>> +			arch_refresh_nodedata(nid, pgdat);
> >>>>> +			node_set_online(nid);
> >>>>
> >>>> Setting all possible nodes online might result in quite some QE noice,
> >>>> because all these nodes will then be visible in the sysfs and
> >>>> try_offline_nodes() is essentially for the trash.
> >>>
> >>> I am not sure I follow. I believe sysfs will not get populate because I
> >>> do not call register_one_node.
> >>
> >> arch/x86/kernel/topology.c:topology_init()
> >>
> >> for_each_online_node(i)
> >> 	register_one_node(i);
> > 
> > Right you are.
> >  
> >>> You are right that try_offline_nodes will be reduce which is good imho.
> >>> More changes will be possible (hopefully to drop some ugly code) on top
> >>> of this change (or any other that achieves that there are no NULL pgdat
> >>> for possible nodes).
> >>>
> >>
> >> No to exposing actually offline nodes to user space via sysfs.
> > 
> > Why is that a problem with the sysfs for non-populated nodes?
> > 
> 
> https://lore.kernel.org/linuxppc-dev/20200428093836.27190-1-srikar@linux.vnet.ibm.com/t/

Thanks. It is good to be reminded that we are in cicling around this
problem for quite some time without really forward much.

> Contains some points -- certainly nothing unfixable but it clearly shows
> that users expect only nodes with actual memory and cpus to be online --
> that's why we export the possible+online state to user space. My point
> is to be careful with such drastic changes and do one step at a time.
>
> I think preallocation of the pgdat is a reasonable thing to have without
> changing user-space visible semantics or even in-kernel semantics.

So your proposal is to drop set_node_online from the patch and add it as
a separate one which handles 
	- sysfs part (i.e. do not register a node which doesn't span a
	  physical address space)
	- hotplug side of (drop the pgd allocation, register node lazily
	  when a first memblocks are registered)

Makes sense?
-- 
Michal Hocko
SUSE Labs