Incorrect expected behavior of BPF_MAP_TYPE_LRU_HASH
Opened this issue · 2 comments
According to the documentation for eBPF hash maps (https://docs.kernel.org/bpf/map_hash.html), an LRU hash map is expected to evict the least recently used item when attempting to insert an element when the map is full.
Despite this description, the implementation for the LRU hash map behaves in a way that causes maps with a small number of max entries to behave unexpectedly. See https://stackoverflow.com/questions/75882443/elements-incorrectly-evicted-from-ebpf-lru-hash-map for an example and discussion.
Tracee uses a few small capacity LRU hash maps (io_file_path_cache_map, elf_files_map, recent_deleted_module_map). They should be reviewed to make sure that they don't rely on incorrect assumptions about the behavior of the LRU map.
__bpf_lru_list_shrink_inactive() will evict a node batch. It will depend on the type of the map, if local (LOCAL_FREE_TARGET), 128, if percpu (PERCPU_FREE_TARGET), 4.
Since it's a special LRU and we can't count with the removal of a single element per eviction triggered, one possibility is to convert the local map to a percpu type trying at least to reduce the removed batch.