On Mon, Jun 28, 2021 at 11:40:39AM +0200, "Enrico Weigelt, metux IT consult" [off-list ref] wrote:

Could you please tell a bit more how this really works ?
(maybe some pointers to the code)

When cgroup's consumption is about to cross the configured limit,
reclaim is started
https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/mm/memcontrol.c?id=62fb9874f5da54fdb243003b386128037319b219#n2579

that may evict old entries
https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/mm/vmscan.c?id=62fb9874f5da54fdb243003b386128037319b219#n2852

and if there's still no success freeing some space the dentry allocation
can fail
https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/mm/slab.h?id=62fb9874f5da54fdb243003b386128037319b219#n277

(This describes just one code path, the subject isn't always a dentry.)

I'm curios what happens if those cache objects are used by different
cgroups - are they accounted to multiple times (once per cgroup) ?
What happens when one cgroup using some cache object reaching its limit,
wile another one does not ?

That's explained here
https://www.kernel.org/doc/html/v5.13/admin-guide/cgroup-v2.html#memory-ownership

Michal