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

[David Hildenbrand <david@redhat.com>]

On 07.12.21 11:54, Michal Hocko wrote:
> Hi,
> I didn't have much time to dive into this deeper and I have hit some
> problems handling this in an arch specific code so I have tried to play
> with this instead:
> 
> diff --git a/mm/page_alloc.c b/mm/page_alloc.c
> index c5952749ad40..4d71759d0d9b 100644
> --- a/mm/page_alloc.c
> +++ b/mm/page_alloc.c
> @@ -8032,8 +8032,16 @@ 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) {
> +	for_each_node(nid) {
>  		pg_data_t *pgdat = NODE_DATA(nid);
> +
> +		if (!node_online(nid)) {
> +			pr_warn("Node %d uninitialized by the platform. Please report with memory map.\n");
> +			alloc_node_data(nid);

That's x86 specific and not exposed to generic code -- at least in my
code base. I think we'd want an arch_alloc_nodedata() variant that
allocates via memblock -- and initializes all fields to 0. So
essentially a generic alloc_node_data().

> +			free_area_init_memoryless_node(nid);

That's really just free_area_init_node() below, I do wonder what value
free_area_init_memoryless_node() has as of today.

> +			continue;
> +		}
> +
>  		free_area_init_node(nid);
>  
>  		/* Any memory on that node */
> 
> Could you give it a try? I do not have any machine which would exhibit
> the problem so I cannot really test this out. I hope build_zone_info
> will not choke on this. I assume the node distance table is
> uninitialized for these nodes and IIUC this should lead to an assumption
> that all other nodes are close. But who knows that can blow up there.
> 
> Btw. does this make any sense at all to others?
> 

__build_all_zonelists() has to update the zonelists of all nodes I think.

-- 
Thanks,

David / dhildenb