oom when enabling memory cgroups without swap on IBM power 8

The system has enough memory for testing even without swap.
# free -m

              total        used        free      shared  buff/cache   available
Mem:          31653         492       28329           4        2830       29414
Swap:             0           0           0

# runc -v runc version 1.0.0-rc2 commit: 6d6086f29040d5d98dd8c67f2073010f10514ee3 spec: 1.0.0-rc3

Use the rootfs from docker.io/caiqian/rhel7

# runc spec
Make changes as below to enable memory cgroup,

...
       "platform": {
                "os": "linux",
                "arch": "ppc64le"
...
       "process": {
                "terminal": false,
                "args": [
                        "sleep", "6000"
                ],
...
        "linux": {
                "resources": {
                        "devices": [
                                {
                                        "allow": false,
                                        "access": "rwm"
                                }
                        ],
                        "memory": {
                                "limit": 16742640,
                                "reservation": 19996672
                        }
                },

# runc run -d root
# runc exec root ls ls: error while loading shared libraries: libpcre.so.1: failed to map segment from shared object: Cannot allocate memory

# dmesg
...

[175567.920712]   cache: kmalloc-512(0:root), object size: 512, buffer size: 512, default order: 0, min order: 0
[175567.920717]   node 2: slabs: 6, objs: 768, free: 0
[175567.941567] SLUB: Unable to allocate memory on node -1 (gfp=0x80d0)
[175567.941572]   cache: vm_area_struct(0:root), object size: 216, buffer size: 216, default order: 0, min order: 0
[175567.941576]   node 2: slabs: 5, objs: 1515, free: 0

BTW, the page size is 64kb on those systems.

Reproduce the configuration using runc spec, runc run -d root, and runc exec root ls with the supplied RHEL7 rootfs, memory limits, ppc64le platform, and no swap. Start by comparing the cgroup and loader failures with the dmesg allocation errors on a 64 KB page-size system; done means identifying the cause and defining a verified runc fix or limitation.