On Tue, Jun 15, 2021 at 10:50:31AM -0400, Qian Cai wrote:

On 6/15/2021 9:19 AM, Mark Rutland wrote:

quoted

Looking some more, it looks like that's correct in isolation, but it
clashes with commit:

  5831eedad2ac6f38 ("mm: replace CONFIG_NEED_MULTIPLE_NODES with CONFIG_NUMA")

Just a data point. Reverting the commit alone fixed the same crash for me.

Yeah, that commit didn't take into the account the change in
pgdat_to_phys().

The patch below should fix it. In the long run I think we should get rid of
contig_page_data and allocate NODE_DATA(0) for !NUMA case as well.

Andrew, can you please add this as a fixup to "mm: replace
CONFIG_NEED_MULTIPLE_NODES with CONFIG_NUMA"?

diff --git a/mm/sparse.c b/mm/sparse.c
index a0e9cdb5bc38..6326cdf36c4f 100644
--- a/mm/sparse.c
+++ b/mm/sparse.c

@@ -347,7 +347,7 @@ size_t mem_section_usage_size(void)
 
 static inline phys_addr_t pgdat_to_phys(struct pglist_data *pgdat)
 {
-#ifndef CONFIG_NEED_MULTIPLE_NODES
+#ifndef CONFIG_NUMA
 	return __pa_symbol(pgdat);
 #else
 	return __pa(pgdat);

-- 
Sincerely yours,
Mike.