The selftest I provided can reproduce a panic: './test_progs -a cgroup_storage_update'
When we attach a program to cgroup and if prog->aux->cgroup_storage exists, which means the cgroup_storage map is used in the program, we will then allocate storage by bpf_cgroup_storages_alloc() and assign it to pl->storage.
At the end, pl->storage will be assigned to cgrp->bpf.effective[atype]->cgroup_storage by xxx_effective_progs().
But when we attach a program without the cgroup_storage map being used (prog->aux->cgroup_storage is empty), the cgroup_storage in struct bpf_prog_array_item is empty.
Then, if we use BPF_LINK_UPDATE to replace the old program with a new one that uses the cgroup_storage map, we miss the cgroup_storage being initialized.
This causes a panic when accessing storage in bpf_get_local_storage.
Jiayuan Chen (2): bpf: Create cgroup storage if needed when updating link selftests/bpf: Add link update test for cgroup_storage
kernel/bpf/cgroup.c | 24 +++++++--- .../selftests/bpf/prog_tests/cgroup_storage.c | 45 +++++++++++++++++++ .../selftests/bpf/progs/cgroup_storage.c | 6 +++ 3 files changed, 70 insertions(+), 5 deletions(-)