-Cc: non-lists
On Tue, Mar 04, 2025 at 03:20:58PM +0100, Michal Koutný mkoutny@suse.com wrote:
Actually, I noticed test_memcontrol failure yesterday (with ~mainline kernel) but I remember they used to work also rather recently. I haven't got time to look into that but at least that one may be a regression (in code or test).
So I'm seeing (with v6.15-rc1):
| not ok 1 test_kmem_basic | ok 2 test_kmem_memcg_deletion | ok 3 test_kmem_proc_kpagecgroup | ok 4 test_kmem_kernel_stacks | not ok 5 test_kmem_dead_cgroups | memory.current 8130560 [ <- 1 vCPU ] 13168640 | percpu 5040000 [ 4 vCPUs ->] 10080000 | not ok 6 test_percpu_basic
not ok 1 By a quick look I suspect that negative dentries that are used to boost memory consumption aren't enough (since some kernel changes, test assumes at least 10B/dentry) -- presumably inappropriate test in new dentry environment, not memcg bug proper.
not ok 5 A dying memcg pinned by something indefinitely, didn't look deeper into that. Little suspicious.
not ok 6 That looks like the test doesn't take into account non-percpu allocations of memcg (e.g. struct memcg alone is a ~2KiB + struct mem_cgroup_per_node). The test needs better boundaries, not a memcg bug.
HTH, Michal