The hash, hash_fast, rhash and bitwise sets may indicate no result even though a matching element exists during a short time window while other cpu is finalizing the transaction. This happens when the hash lookup/bitwise lookup function has picked up the old genbit, right before it was toggled by nf_tables_commit(), but then the same cpu managed to unlink the matching old element from the hash table: cpu0 cpu1 has added new elements to clone has marked elements as being inactive in new generation perform lookup in the set enters commit phase: A) observes old genbit increments base_seq I) increments the genbit II) removes old element from the set B) finds matching element C) returns no match: found element is not valid in old generation Next lookup observes new genbit and finds matching e2. Consider a packet matching element e1, e2. cpu0 processes following transaction: 1. remove e1 2. adds e2, which has same key as e1. P matches both e1 and e2. Therefore, cpu1 should always find a match for P. Due to above race, this is not the case: cpu1 observed the old genbit. e2 will not be considered once it is found. The element e1 is not found anymore if cpu0 managed to unlink it from the hlist before cpu1 found it during list traversal. The situation only occurs for a brief time period, lookups happening after I) observe new genbit and return e2. This problem exists in all set types except nft_set_pipapo, so fix it once in nft_lookup rather than each set ops individually. Sample the base sequence counter, which gets incremented right before the genbit is changed. Then, if no match is found, retry the lookup if the base sequence was altered in between. If the base sequence hasn't changed: - No update took place: no-match result is expected. This is the common case. or: - nf_tables_commit() hasn't progressed to genbit update yet. Old elements were still visible and nomatch result is expected, or: - nf_tables_commit updated the genbit: We picked up the new base_seq, so the lookup function also picked up the new genbit, no-match result is expected. If the old genbit was observed, then nft_lookup also picked up the old base_seq: nft_lookup_should_retry() returns true and relookup is performed in the new generation. This problem was added when the unconditional synchronize_rcu() call that followed the current/next generation bit toggle was removed. Thanks to Pablo Neira Ayuso for reviewing an earlier version of this patchset, for suggesting re-use of existing base_seq and placement of the restart loop in nft_set_do_lookup(). Fixes: 0cbc06b3faba ("netfilter: nf_tables: remove synchronize_rcu in commit phase") Signed-off-by: Florian Westphal --- net/netfilter/nf_tables_api.c | 3 ++- net/netfilter/nft_lookup.c | 31 ++++++++++++++++++++++++++++++- 2 files changed, 32 insertions(+), 2 deletions(-) diff --git a/net/netfilter/nf_tables_api.c b/net/netfilter/nf_tables_api.c index 9518b50695ba..c3c73411c40c 100644 --- a/net/netfilter/nf_tables_api.c +++ b/net/netfilter/nf_tables_api.c @@ -10973,7 +10973,8 @@ static int nf_tables_commit(struct net *net, struct sk_buff *skb) while (++base_seq == 0) ; - WRITE_ONCE(net->nft.base_seq, base_seq); + /* pairs with smp_load_acquire in nft_lookup_eval */ + smp_store_release(&net->nft.base_seq, base_seq); gc_seq = nft_gc_seq_begin(nft_net); diff --git a/net/netfilter/nft_lookup.c b/net/netfilter/nft_lookup.c index 2c6909bf1b40..58c5b14889c4 100644 --- a/net/netfilter/nft_lookup.c +++ b/net/netfilter/nft_lookup.c @@ -55,11 +55,40 @@ __nft_set_do_lookup(const struct net *net, const struct nft_set *set, return set->ops->lookup(net, set, key); } +static unsigned int nft_base_seq(const struct net *net) +{ + /* pairs with smp_store_release() in nf_tables_commit() */ + return smp_load_acquire(&net->nft.base_seq); +} + +static bool nft_lookup_should_retry(const struct net *net, unsigned int seq) +{ + return unlikely(seq != nft_base_seq(net)); +} + const struct nft_set_ext * nft_set_do_lookup(const struct net *net, const struct nft_set *set, const u32 *key) { - return __nft_set_do_lookup(net, set, key); + const struct nft_set_ext *ext; + unsigned int base_seq; + + do { + base_seq = nft_base_seq(net); + + ext = __nft_set_do_lookup(net, set, key); + if (ext) + break; + /* No match? There is a small chance that lookup was + * performed in the old generation, but nf_tables_commit() + * already unlinked a (matching) element. + * + * We need to repeat the lookup to make sure that we didn't + * miss a matching element in the new generation. + */ + } while (nft_lookup_should_retry(net, base_seq)); + + return ext; } EXPORT_SYMBOL_GPL(nft_set_do_lookup); -- 2.49.1