netfilter: nf_tables: store new sets in dedicated list
Subsystems (2)
Florian Westphal
maintainerPATCH 1/1 netfilter: nf_tables: store new sets in dedicated list
nft_set_lookup_byid() is very slow when transaction becomes large, due to walk of the transaction list.
Add a dedicated list that contains only the new sets.
Before: nft -f ruleset 0.07s user 0.00s system 0% cpu 1:04.84 total
After: nft -f ruleset 0.07s user 0.00s system 0% cpu 30.115 total
.. where ruleset contains ~10 sets with ~100k elements. The above number is for a combined flush+reload of the ruleset.
With previous flush, even the first NEWELEM has to walk through a few hundred thousands of DELSET(ELEM) transactions before the first NEWSET object. To cope with random-order-newset-newsetelem we'd need to replace commit_set_list with a hashtable.
Expectation is that a NEWELEM operation refers to the most recently added set, so last entry of the dedicated list should be the set we want.
NB: This is not a bug fix per se (functionality is fine), but with larger transaction batches list search takes forever, so it would be nice to speed this up for -stable too, hence adding a "fixes" tag.
--- include/net/netfilter/nf_tables.h | 2 ++ net/netfilter/nf_tables_api.c | 29 ++++++++++++++++++++--------- 2 files changed, 22 insertions(+), 9 deletions(-)
include/net/netfilter/nf_tables.h +2 -0
@@ -1674,6 +1674,7 @@ struct nft_trans_rule {struct nft_trans_set {struct nft_trans_binding nft_trans_binding; +struct list_head list_trans_newset;struct nft_set *set;u32 set_id;u32 gc_int;
@@ -1875,6 +1876,7 @@ static inline int nft_request_module(struct net *net, const char *fmt, ...) { restruct nftables_pernet {struct list_head tables;struct list_head commit_list; +struct list_head commit_set_list;struct list_head binding_list;struct list_head module_list;struct list_head notify_list;
net/netfilter/nf_tables_api.c +20 -9
@@ -393,6 +393,7 @@ static void nft_trans_commit_list_add_tail(struct net *net, struct nft_trans *tr {struct nftables_pernet *nft_net =nft_pernet (net);struct nft_trans_binding *binding; +struct nft_trans_set *trans_set;list_add_tail (&trans->list, &nft_net->commit_list);
@@ -402,9 +403,13 @@ static void nft_trans_commit_list_add_tail(struct net *net, struct nft_trans *trswitch (trans->msg_type) {case NFT_MSG_NEWSET : + trans_set =nft_trans_container_set (trans); +if (!nft_trans_set_update (trans) &&nft_set_is_anonymous (nft_trans_set (trans)))list_add_tail (&binding->binding_list, &nft_net->binding_list); + +list_add_tail (&trans_set->list_trans_newset, &nft_net->commit_set_list);break ;case NFT_MSG_NEWCHAIN : if (!nft_trans_chain_update(trans) &&
@@ -611,6 +616,7 @@ static int __nft_trans_set_add(const struct nft_ctx *ctx, int msg_type, trans_set =nft_trans_container_set (trans);INIT_LIST_HEAD (&trans_set->nft_trans_binding.binding_list); +INIT_LIST_HEAD (&trans_set->list_trans_newset);if (msg_type ==NFT_MSG_NEWSET && ctx->nla[NFTA_SET_ID ] && !desc) { nft_trans_set_id(trans) =
@@ -4473,17 +4479,16 @@ static struct nft_set *nft_set_lookup_byid(const struct net *net, {struct nftables_pernet *nft_net =nft_pernet (net);u32 id =ntohl (nla_get_be32 (nla)); -struct nft_trans *trans; +struct nft_trans_set *trans; -list_for_each_entry (trans, &nft_net->commit_list, list) { -if (trans->msg_type ==NFT_MSG_NEWSET ) { -struct nft_set *set =nft_trans_set (trans); +/* its likely the id we need is at the tail, not at start */ +list_for_each_entry_reverse (trans, &nft_net->commit_set_list, list_trans_newset) { +struct nft_set *set = trans->set; -if (id ==nft_trans_set_id (trans) && - set->table == table && -nft_active_genmask (set, genmask)) -return set; - } +if (id == trans->set_id && + set->table == table && +nft_active_genmask (set, genmask)) +return set; }return ERR_PTR (-ENOENT ); }
@@ -10378,6 +10383,7 @@ static int nf_tables_commit(struct net *net, struct sk_buff *skb)nft_flow_rule_destroy (nft_trans_flow_rule (trans));break ;case NFT_MSG_NEWSET : +list_del (&nft_trans_container_set (trans)->list_trans_newset);if (nft_trans_set_update (trans)) {struct nft_set *set =nft_trans_set (trans);
@@ -10686,6 +10692,7 @@ static int __nf_tables_abort(struct net *net, enum nfnl_abort_action action)nft_trans_destroy (trans);break ;case NFT_MSG_NEWSET : +list_del (&nft_trans_container_set (trans)->list_trans_newset);if (nft_trans_set_update (trans)) {nft_trans_destroy (trans);break ;
@@ -10781,6 +10788,8 @@ static int __nf_tables_abort(struct net *net, enum nfnl_abort_action action) } } +WARN_ON_ONCE (!list_empty (&nft_net->commit_set_list)); +nft_set_abort_update (&set_update_list);synchronize_rcu ();
@@ -11595,6 +11604,7 @@ static int __net_init nf_tables_init_net(struct net *net)INIT_LIST_HEAD (&nft_net->tables);INIT_LIST_HEAD (&nft_net->commit_list); +INIT_LIST_HEAD (&nft_net->commit_set_list);INIT_LIST_HEAD (&nft_net->binding_list);INIT_LIST_HEAD (&nft_net->module_list);INIT_LIST_HEAD (&nft_net->notify_list);
@@ -11625,6 +11635,7 @@ static void __net_exit nf_tables_exit_net(struct net *net) gc_seq =nft_gc_seq_begin (nft_net);WARN_ON_ONCE (!list_empty (&nft_net->commit_list)); +WARN_ON_ONCE (!list_empty (&nft_net->commit_set_list));if (!list_empty (&nft_net->module_list))nf_tables_module_autoload_cleanup (net);
Signature · 1 line
2.44.2
Quoted 16 lines · Florian Westphal
nft_set_lookup_byid()is very slow when transaction becomes large, due to walk of the transaction list.Add a dedicated list that contains only the new sets.
Before: nft -f ruleset 0.07s user 0.00s system 0% cpu 1:04.84 total
After: nft -f ruleset 0.07s user 0.00s system 0% cpu 30.115 total.. where ruleset contains ~10 sets with ~100k elements. The above number is for a combined flush+reload of the ruleset.
With previous flush, even the first NEWELEM has to walk through a few hundred thousands of DELSET(ELEM) transactions before the first NEWSET object. To cope with random-order-newset-newsetelem we'd need to replace
commit_set_listwith a hashtable.
Quoted 6 lines
Expectation is that a NEWELEM operation refers to the most recently added set, so last entry of the dedicated list should be the set we want.
NB: This is not a bug fix per se (functionality is fine), but with larger transaction batches list search takes forever, so it would be nice to speed this up for -stable too, hence adding a "fixes" tag.
applied to nf-next, thanks