Strange results of ``docker stats`` (because of lots of dentry and ovl_inode slabs)

General
1

We run docker on 4.14.219 (hence cgroups v1).

docker stats reports insanely high memory usage, although memory.stat contains much lower rss.

The problem is high dentry and ovl_inode usage (we found it in slabinfo)
This fact affects both memory.usage_in_bytes and memory.kmem.usage_in_bytes

After echo 2 > /proc/sys/vm/drop_caches everything becomes OK.

Docker doc says:

On Linux, the Docker CLI reports memory usage by subtracting cache usage from the total memory usage.

I assume Docker uses memory.usage_in_bytes as “total memory usage”, am I right?

Kernel documentation says:

If you want to know more exact memory usage, you should use RSS+CACHE(+SWAP)
value in memory.stat(see 5.2).

https://www.kernel.org/doc/Documentation/cgroup-v1/memory.txt

So, my question is: why use obscure usage_in_bytes instead of RSS+SWAP?

In any case, big number of dentry and ovl_inode slabs isn’t something that should affect docker stats, should it?

How to reproduce.

  1. On the container run du --inodes -d1 /
  2. On the host run docker stats: you should see lots of memory used
  3. Check slabs: sed 1,2d memory.kmem.slabinfo | awk '{RES=$2*$4; print RES"\t"RES/1024/1024" MB\t"$1}' | sort -n (in my case alsmost 10 megabytes occupied by ovl_inode
  4. Reclaim memory (via 2 > /proc/sys/vm/drop_caches or 0 > memory.force_empty)
  5. Check docker stats again