[dangowrt]

Some validation tests check things the host kernel cannot provide, so they fail against every runtime on such hosts. This makes them adapt to the host instead:

• validation: skip the hugetlb tests without hugetlb support: linux_cgroups_hugetlb and linux_cgroups_relative_hugetlb now skip unless /proc/cgroups lists the hugetlb controller as enabled and /sys/kernel/mm/hugepages offers at least one page size. Without either, no runtime can apply linux.resources.hugepageLimits; the tests used to fail reading /sys/kernel/mm/hugepages. /proc/cgroups is created for both cgroup versions, so the check needs neither a cgroup v1 mount nor a v2 reader.
• validation/linux_cgroups_hugetlb: report failures to start a container: when the runtime refused the container, testHugetlbCgroups() returned early, its deferred AutoPlan() printed 1..0 and main() exited through util.Fatal(), so the failure never showed up as a failed test and the invalid page size check did not run. Both checks now share one TAP stream and a failure to run the container is a failed test.
• validation/process_rlimits: skip limits the host cannot hold: a 32-bit kernel stores rlimits in an unsigned long and turns any value of 4 GiB or more into RLIM_INFINITY, so the hard limits for RLIMIT_CORE, RLIMIT_DATA, RLIMIT_FSIZE and RLIMIT_STACK could never match there. The test now probes the host by setting a 4 GiB soft RLIMIT_FSIZE on itself and reading it back, and leaves out the limits of 4 GiB and more when the kernel cannot hold them. A 32-bit userland on a 64-bit kernel still tests all of them. The limits are left out rather than scaled down, because on a 32-bit kernel RLIMIT_STACK also sets the gap below the stack, and a finite multi-GiB value would leave the container little address space.

Tested with ujail on OpenWrt in QEMU: on x86_64 and i386 kernels without the hugetlb controller both hugetlb tests skip; on an x86_64 kernel with it, both run and pass (2MB limit set, the invalid 3MB page size refused); on i386 process_rlimits passes with the large limits left out, and on a 64-bit kernel it still sets all of them. go test ./validation/util/ covers the /proc/cgroups parser, and that the rlimit probe restores the original limit.