Skip to content

L
linux
Switch branch/tag
  1. 17 Jun, 2019 2 commits
    • Florian Westphal's avatar
      netfilter: conntrack: small conntrack lookup optimization · 87e389b4
      Florian Westphal authored 
      ____nf_conntrack_find() performs checks on the conntrack objects in
this order:

1. if (nf_ct_is_expired(ct))

This fetches ct->timeout, in third cache line.

The hnnode that is used to store the list pointers resides in the first
(origin) or second (reply tuple) cache lines.

This test rarely passes, but its necessary to reap obsolete entries.

2. if (nf_ct_is_dying(ct))

This fetches ct->status, also in third cache line.

The test is useless, and can be removed:
  Consider:
     cpu0                                           cpu1
    ct = ____nf_conntrack_find()
    atomic_inc_not_zero(ct) -> ok
    nf_ct_key_equal -> ok
    is_dying -> DYING bit not set, ok
                                                    set_bit(ct, DYING);
						    ... unhash ... etc.
    return ct
    -> returning a ct with dying bit set, despite
    having a test for it.

This (unlikely) case is fine - refcount prevents ct from getting free'd.

3. if (nf_ct_key_equal(h, tuple, zone, net))

nf_ct_key_equal checks in following order:

1. Tuple equal (first or second cacheline)
2. Zone equal (third cacheline)
3. confirmed bit set (->status, third cacheline)
4. net namespace match (third cacheline).

Swapping "timeout" and "cpu" places timeout in the first cacheline.
This has two advantages:

1. For a conntrack that won't even match the original tuple,
   we will now only fetch the first and maybe the second cacheline
   instead of always accessing the 3rd one as well.

2.  in case of TCP ct->timeout changes frequently because we
    reduce/increase it when there are packets outstanding in the network.

The first cacheline contains both the reference count and the ct spinlock,
i.e. moving timeout there avoids writes to 3rd cacheline.

The restart sequence in __nf_conntrack_find() is removed, if we found a
candidate, but then fail to increment the refcount or discover the tuple
has changed (object recycling), just pretend we did not find an entry.

A second lookup won't find anything until another CPU adds a new conntrack
with identical tuple into the hash table, which is very unlikely.

We have the confirmation-time checks (when we hold hash lock) that deal
with identical entries and even perform clash resolution in some cases.
Signed-off-by: default avatarFlorian Westphal <fw@strlen.de>
Signed-off-by: default avatarPablo Neira Ayuso <pablo@netfilter.org>
      87e389b4
    • Stéphane Veyret's avatar
      netfilter: nft_ct: add ct expectations support · 857b4602
      Stéphane Veyret authored 
      This patch allows to add, list and delete expectations via nft objref
infrastructure and assigning these expectations via nft rule.

This allows manual port triggering when no helper is defined to manage a
specific protocol. For example, if I have an online game which protocol
is based on initial connection to TCP port 9753 of the server, and where
the server opens a connection to port 9876, I can set rules as follow:

table ip filter {
    ct expectation mygame {
        protocol udp;
        dport 9876;
        timeout 2m;
        size 1;
    }

    chain input {
        type filter hook input priority 0; policy drop;
        tcp dport 9753 ct expectation set "mygame";
    }

    chain output {
        type filter hook output priority 0; policy drop;
        udp dport 9876 ct status expected accept;
    }
}
Signed-off-by: default avatarStéphane Veyret <sveyret@gmail.com>
Signed-off-by: default avatarPablo Neira Ayuso <pablo@netfilter.org>
      857b4602
  3. 06 Jun, 2019 9 commits
  5. 05 Jun, 2019 29 commits
GitLab Nexedi Edition | About GitLab | About Nexedi | 沪ICP备2021021310号-2 | 沪ICP备2021021310号-7