[LTP] [RESEND PATCH 1/4] controllers/memcg: account per-node kernel memory

[Richard Palethorpe <rpalethorpe@suse.de>]

Hello Krzysztof,

Krzysztof Kozlowski <krzysztof.kozlowski@canonical.com> writes:

> Recent Linux kernels () charge groups also with kernel memory.  This is
> not limited only to process-allocated memory but also cgroup-handling
> code memory as well.
>
> For example since kernel v5.9 with commit 3e38e0aaca9e ("mm: memcg:
> charge memcg percpu memory to the parent cgroup") creating a subgroup
> causes several kernel allocations towards this group.
>
> These additional kernel memory allocations are proportional to number of
> CPUs and number of nodes.
>
> On c4.8xlarge AWS instance with 36 cores in two nodes with v5.11 Linux
> kernel the memcg_subgroup_charge and memcg_use_hierarchy_test tests were
> failing:
>
>     memcg_use_hierarchy_test 1 TINFO: timeout per run is 0h 5m 0s
>     memcg_use_hierarchy_test 1 TINFO: set /dev/memcg/memory.use_hierarchy to 0 failed
>     memcg_use_hierarchy_test 1 TINFO: test if one of the ancestors goes over its limit, the proces will be killed
>     mkdir: cannot create directory ?subgroup?: Cannot allocate memory
>     /home/ubuntu/ltp-install/testcases/bin/memcg_use_hierarchy_test.sh: 26: cd: can't cd to subgroup
>     memcg_use_hierarchy_test 1 TINFO: Running memcg_process --mmap-lock1 -s 8192
>     memcg_use_hierarchy_test 1 TFAIL: process  is not killed
>     rmdir: failed to remove 'subgroup': No such file or directory
>
> The kernel was unable to create the subgroup (mkdir returned -ENOMEM)
> due to this additional per-node kernel memory allocations.
>
> Signed-off-by: Krzysztof Kozlowski <krzysztof.kozlowski@canonical.com>
> ---
>  .../controllers/memcg/functional/memcg_lib.sh | 44 +++++++++++++++++++
>  .../memcg/functional/memcg_subgroup_charge.sh |  8 +---
>  .../functional/memcg_use_hierarchy_test.sh    |  8 +++-
>  3 files changed, 52 insertions(+), 8 deletions(-)
>
> diff --git a/testcases/kernel/controllers/memcg/functional/memcg_lib.sh b/testcases/kernel/controllers/memcg/functional/memcg_lib.sh
> index dad66c798e19..700e9e367bff 100755
> --- a/testcases/kernel/controllers/memcg/functional/memcg_lib.sh
> +++ b/testcases/kernel/controllers/memcg/functional/memcg_lib.sh
> @@ -63,6 +63,50 @@ memcg_require_hierarchy_disabled()
>  	fi
>  }
>  
> +# Kernel memory allocated for the process is also charged.  It might depend on
> +# the number of CPUs and number of nodes. For example on kernel v5.11
> +# additionally total_cpus (plus 1 or 2) pages are charged to the group via
> +# kernel memory.  For a two-node machine, additional 108 pages kernel memory
> +# are charged to the group.
> +#
> +# Adjust the limit to account such per-CPU and per-node kernel memory.
> +# $1 - variable name with limit to adjust
> +memcg_adjust_limit_for_kmem()
> +{
> +	[ $# -ne 1 ] && tst_brk TBROK "memcg_adjust_limit_for_kmem expects 1 parameter"
> +	eval "local _limit=\$$1"

Could we do this a simpler way?

It would be much easier to read if we just returned the value which
needed to be added.

> +
> +	# Total number of CPUs
> +	local total_cpus=`tst_ncpus`
> +
> +	# Get the number of NODES

Is it acceptable or necessary to use /sys/devices/system/node/possible
(or online) instead?

> +	if [ -f "/sys/devices/system/node/has_high_memory" ]; then
> +		local mem_string="`cat /sys/devices/system/node/has_high_memory`"
> +	else
> +		local mem_string="`cat /sys/devices/system/node/has_normal_memory`"
> +	fi
> +
> +	local total_nodes="`echo $mem_string | tr ',' ' '`"
> +	local count=0
> +	for item in $total_nodes; do
> +		local delta=1
> +		if [ "${item#*-*}" != "$item" ]; then
> +			delta=$((${item#*-*} - ${item%*-*} + 1))
> +		fi
> +		count=$((count + $delta))
> +	done

Or perhaps we could count the number of 'node[0-9]+' directories? I
think that would be easier to understand.

> +	total_nodes=$count
> +	# Additional nodes impose charging the kmem, not having regular one node
> +	local node_mem=0
> +	if [ $total_nodes -gt 1 ]; then
> +		node_mem=$((total_nodes - 1))
> +		node_mem=$((node_mem * PAGESIZE * 128))
> +	fi
> +
> +	eval "$1='$((_limit + 4 * PAGESIZE + total_cpus * PAGESIZE + node_mem))'"
> +	return 0
> +}

Otherwise looks good.

-- 
Thank you,
Richard.