[PATCH 0/2] x86/numa: Fix NUMA node overlap & init failure

[alison.schofield@intel.com]

From: Alison Schofield <alison.schofield@intel.com>

A previously posted single patch [1] is obsoleted by this set. The
feedback from that review is applied and noted in Patch 1.

While trying to attribute a CXL user report to the bad selection of
overlapping memblks, as fixed in Patch 1, I found that two issues,
in sequence, lead to NUMA Node overlap and NUMA init failure.

An overlapping NUMA node occurs when a non-overlapping memblk is
selected to fill (Patch 1), and then a bad sort (Patch 2) puts the
memblk with the greater address ahead of the lesser address memblk
in the fill list.

It looked like this:

Existing memblks:
	node 6 [mem 0xb90000000-0xc90000000]
	node 7 [mem 0xc90000000-0xd90000000]

Call to numa_fill_memblks(b90000000,c90000000)

Error (Patch 1): collects 2 blks  
	blk[0] node 6 [0xb90000000-0xc90000000]
	blk[1] node 7 [0xc90000000-0xd90000000]

Error (Patch 2): bad sort of the 2 blks
	blk[0] node 7 [0xc90000000-0xd90000000]
	blk[1] node 6 [0xb90000000-0xc90000000]

Seals the deal with a bad fill:
	blk[0] node 7 [0xb90000000-0xd90000000]

Boom: numa_clean_meminfo() discovers the overlap in Nodes 6 & 7
and NUMA init fails.

Since the scenario above is not solely attributed to either patch,
the story is explicity shared here.

[1] https://lore.kernel.org/linux-cxl/20240102213206.1493733-1-alison.schofield@intel.com/

Alison Schofield (2):
  x86/numa: Fix the address overlap check in numa_fill_memblks()
  x86/numa: Fix the sort compare func used in numa_fill_memblks()

 arch/x86/mm/numa.c       | 21 ++++++++-------------
 include/linux/memblock.h |  2 ++
 mm/memblock.c            |  5 +++--
 3 files changed, 13 insertions(+), 15 deletions(-)


base-commit: bd009225e8cbb6e18ad3389328fa640e4887dd9e
-- 
2.37.3