In build_zonelists(), when the fallback list is built for the nodes, the
node load gets reinitialized during each iteration.  This results in
nodes with same distances occupying the same slot in different node
fallback lists rather than appearing in the intended round- robin
manner.  This results in one node getting picked for allocation more
compared to other nodes with the same distance.

As an example, consider a 4 node system with the following distance
matrix.

  Node 0  1  2  3
  ----------------
  0    10 12 32 32
  1    12 10 32 32
  2    32 32 10 12
  3    32 32 12 10

For this case, the node fallback list gets built like this:

  Node  Fallback list
  ---------------------
  0     0 1 2 3
  1     1 0 3 2
  2     2 3 0 1
  3     3 2 0 1 <-- Unexpected fallback order

In the fallback list for nodes 2 and 3, the nodes 0 and 1 appear in the
same order which results in more allocations getting satisfied from node
0 compared to node 1.

The effect of this on remote memory bandwidth as seen by stream
benchmark is shown below:

  Case 1: Bandwidth from cores on nodes 2 & 3 to memory on nodes 0 & 1
	(numactl -m 0,1 ./stream_lowOverhead ... --cores <from 2, 3>)
  Case 2: Bandwidth from cores on nodes 0 & 1 to memory on nodes 2 & 3
	(numactl -m 2,3 ./stream_lowOverhead ... --cores <from 0, 1>)

  ----------------------------------------
		BANDWIDTH (MB/s)
      TEST	Case 1		Case 2
  ----------------------------------------
      COPY	57479.6		110791.8
     SCALE	55372.9		105685.9
       ADD	50460.6		96734.2
    TRIADD	50397.6		97119.1
  ----------------------------------------

The bandwidth drop in Case 1 occurs because most of the allocations get
satisfied by node 0 as it appears first in the fallback order for both
nodes 2 and 3.

This can be fixed by accumulating the node load in build_zonelists()
rather than reinitializing it during each iteration.  With this the
nodes with the same distance rightly get assigned in the round robin
manner.

In fact this was how it was originally until commit f0c0b2b8
("change zonelist order: zonelist order selection logic") dropped the
load accumulation and resorted to initializing the load during each
iteration.

While zonelist ordering was removed by commit c9bff3ee ("mm,
page_alloc: rip out ZONELIST_ORDER_ZONE"), the change to the node load
accumulation in build_zonelists() remained.  So essentially this patch
reverts back to the accumulated node load logic.

After this fix, the fallback order gets built like this:

  Node Fallback list
  ------------------
  0    0 1 2 3
  1    1 0 3 2
  2    2 3 0 1
  3    3 2 1 0 <-- Note the change here

The bandwidth in Case 1 improves and matches Case 2 as shown below.

  ----------------------------------------
		BANDWIDTH (MB/s)
      TEST	Case 1		Case 2
  ----------------------------------------
      COPY	110438.9	110107.2
     SCALE	105930.5	105817.5
       ADD	97005.1		96159.8
    TRIADD	97441.5		96757.1
  ----------------------------------------

The correctness of the fallback list generation has been verified for
the above node configuration where the node 3 starts as memory-less node
and comes up online only during memory hotplug.

[bharata@amd.com: Added changelog, review, test validation]

Link: https://lkml.kernel.org/r/20210830121603.1081-3-bharata@amd.com
Fixes: f0c0b2b8 ("change zonelist order: zonelist order selection logic")
Signed-off-by: Krupa Ramakrishnan <krupa.ramakrishnan@amd.com>
Co-developed-by: Sadagopan Srinivasan <Sadagopan.Srinivasan@amd.com>
Signed-off-by: Sadagopan Srinivasan <Sadagopan.Srinivasan@amd.com>
Signed-off-by: Bharata B Rao <bharata@amd.com>
Acked-by: Mel Gorman <mgorman@suse.de>
Reviewed-by: Anshuman Khandual <anshuman.khandual@arm.com>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Lee Schermerhorn <lee.schermerhorn@hp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>