lib: optimize cpumask_local_spread()

lib: optimize cpumask_local_spread()

In multi-processor and NUMA system, I/O driver will find cpu cores that
which shall be bound IRQ.  When cpu cores in the local numa have been used
up, it is better to find the node closest to the local numa node for
performance, instead of choosing any online cpu immediately.

On arm64 or x86 platform that has 2-sockets and 4-NUMA nodes, if the
network card is located in node2 of socket1, while the number queues of
network card is greater than the number of cores of node2, when all cores
of node2 has been bound to the queues, the remaining queues will be bound
to the cores of node0 which is further than NUMA node3.  It is not
friendly for performance or Intel's DDIO (Data Direct I/O Technology) when
if the user enables SNC (sub-NUMA-clustering).  Let's improve it and find
the nearest unused node through NUMA distance for the non-local NUMA
nodes.

On Huawei Kunpeng 920 server, there are 4 NUMA node(0 - 3) in the 2-cpu
system(0 - 1). The topology of this server is followed:
available: 4 nodes (0-3)
node 0 cpus: 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23
node 0 size: 63379 MB
node 0 free: 61899 MB
node 1 cpus: 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47
node 1 size: 64509 MB
node 1 free: 63942 MB
node 2 cpus: 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71
node 2 size: 64509 MB
node 2 free: 63056 MB
node 3 cpus: 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95
node 3 size: 63997 MB
node 3 free: 63420 MB
node distances:
node   0   1   2   3
  0:  10  16  32  33
  1:  16  10  25  32
  2:  32  25  10  16
  3:  33  32  16  10

We perform PS (parameter server) business test, the behavior of the
service is that the client initiates a request through the network card,
the server responds to the request after calculation.  When two PS
processes run on node2 and node3 separately and the network card is
located on 'node2' which is in cpu1, the performance of node2 (26W QPS)
and node3 (22W QPS) is different.

It is better that the NIC queues are bound to the cpu1 cores in turn, then
XPS will also be properly initialized, while cpumask_local_spread only
considers the local node.  When the number of NIC queues exceeds the
number of cores in the local node, it returns to the online core directly.
So when PS runs on node3 sending a calculated request, the performance is
not as good as the node2.

The IRQ from 369-392 will be bound from NUMA node0 to NUMA node3 with this
patch, before the patch:

Euler:/sys/bus/pci # cat /proc/irq/369/smp_affinity_list
0
Euler:/sys/bus/pci # cat /proc/irq/370/smp_affinity_list
1
...
Euler:/sys/bus/pci # cat /proc/irq/391/smp_affinity_list
22
Euler:/sys/bus/pci # cat /proc/irq/392/smp_affinity_list
23
After the patch:
Euler:/sys/bus/pci # cat /proc/irq/369/smp_affinity_list
72
Euler:/sys/bus/pci # cat /proc/irq/370/smp_affinity_list
73
...
Euler:/sys/bus/pci # cat /proc/irq/391/smp_affinity_list
94
Euler:/sys/bus/pci # cat /proc/irq/392/smp_affinity_list
95

So the performance of the node3 is the same as node2 that is 26W QPS when
the network card is still in 'node2' with the patch.

Link: https://lkml.kernel.org/r/1605668072-44780-1-git-send-email-zhangshaokun@hisilicon.com
Signed-off-by: Yuqi Jin <jinyuqi@huawei.com>
Signed-off-by: Shaokun Zhang <zhangshaokun@hisilicon.com>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: Rusty Russell <rusty@rustcorp.com.au>
Cc: Juergen Gross <jgross@suse.com>
Cc: Paul Burton <paul.burton@mips.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Michael Ellerman <mpe@ellerman.id.au>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Anshuman Khandual <anshuman.khandual@arm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>