Portal:DeveloperDocs/set internals
Jump to navigation
Jump to search
The nftables generalized set infrastructure includes multiple set implementations. The implementation chosen for a given set depends on required set features and operations, and on estimated element lookup time and set memory requirements.
Available nft_set_types
nft_set_type | nft_set_types[] order | nft_set_estimate NFT_SET_CLASS_[order] | # Concatenated fields | Variable # elements | NFT_SET_INTERVAL | NFT_SET_MAP | NFT_SET_TIMEOUT | NFT_SET_OBJECT | NFT_SET_EVAL | Expression support | Notes | |
---|---|---|---|---|---|---|---|---|---|---|---|---|
.lookup | .space | |||||||||||
nft_set_hash_fast_type | 0 | O_1 | O_N | No | No | Yes | No | Yes | No | klen != 4 | ||
nft_set_hash_type | 1 | O_1 | O_N | No | No | Yes | No | Yes | No | klen != 4 | ||
nft_set_rhash_type | 2 | O_1 | O_N | Yes | No | Yes | Yes | Yes | Yes | |||
nft_set_bitmap_type | 3 | O_1 | O_1 | Yes | No | No | No | No | No | No | klen <= 2 | |
nft_set_rbtree_type | 4 | O_LOG_N | O_N | <= 1 | Yes | Yes | Yes | Yes | Yes | No | ||
nft_set_pipapo_avx2_type | 5 | O_LOG_N | O_N | >= 2 | Yes | Yes (Mandatory) | Yes | Yes | Yes | No | ||
nft_set_pipapo_type | 6 | O_LOG_N | O_N | >= 2 | Yes | Yes (Mandatory) | Yes | Yes | Yes | No |
- nft_set_estimate .lookup and .space are in terms of enum nft_set_class, defined in nf_tables.h:
enum nft_set_class {
NFT_SET_CLASS_O_1,
NFT_SET_CLASS_O_LOG_N,
NFT_SET_CLASS_O_N,
};
- nft_select_set_ops() in nf_tables_api.c: chooses which nft_set_type to use. For sets with default performance policy it chooses lower .lookup; for sets with memory policy it chooses lower .space.
- When choosing between two nft_set_types with the same .lookup and .space, nft_select_set_ops() chooses the type that appears first in nft_set_types[].
Hash implementations
Bitmap implementation
nft_set_bitmap.c - contains good documentation
Red-black tree implementation
PIPAPO implementations
- nft_set_pipapo.c - contains excellent documentation
- nft_set_pipapo_avx2.c
PIPAPO is loosely inspired by the Grouper network packet classification algorithm.