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: 0cbc06b3fa ("netfilter: nf_tables: remove synchronize_rcu in commit phase")
Signed-off-by: Florian Westphal <fw@strlen.de>
b2f742c846