Comment 0 for bug 1836816

BugLink: https://bugs.launchpad.net/bugs/

[Impact]

There are a number of races in nf_conntrack when multiple packets are sent from
the same origin to the same destination at the same time, where there is no
pre-existing confirmed conntrack entry, in a NAT environment.

This is most frequently seen in users with kubernetes workloads, where dns
requests are sent from the same socket at the same time, from different threads.
One request is for a A record, the other AAAA. A conntrack race happens, which
leads to one request being dropped, which leads to 5 second dns timeouts.

This problem is not specific to kubernetes, as any multi-threaded process
which sends UDP packets from the same socket at the same time from different
threads is affected.

Note that since UDP is connectionless, no packet is sent when connect() is
called, so no conntrack entries are created until a packet is first sent.

In the scenario where two UDP packets are sent at the same time from the same
socket on different threads, there are the following possible races:

1.) Neither of the packets find a confirmed conntrack entry. This leads to two
    conntrack entries being created with the same tuples.
2.) Same as 1), but a conntrack entry is confirmed for one of the packets before
    the other calls get_unique_tuple(). The other packet gets a different reply
    tuple with the source port changed.

The outcomes of the races are the same, one of the packets is dropped when it
comes time to confirm conntrack entries with __nf_conntrack_confirm().

For more information, read the technical blog post:
https://www.weave.works/blog/racy-conntrack-and-dns-lookup-timeouts

[Fix]

The fix to race 1) was included in 4.19 upstream by this commit:

netfilter: nf_conntrack: resolve clash for matching conntracks
commit ed07d9a021df6da53456663a76999189badc432a
Author: Martynas Pumputis <email address hidden>
Date: Mon Jul 2 16:52:14 2018 +0200

This commit adds an extra check to see if the two entries are the same and to
merge the entries if they are.

The fix to race 2) was included in 5.0 upstream by this commit:

netfilter: nf_nat: skip nat clash resolution for same-origin entries
commit 4e35c1cb9460240e983a01745b5f29fe3a4d8e39
Author: Martynas Pumputis <email address hidden>
Date: Tue Jan 29 15:51:42 2019 +0100

This commit ensures that a new source port is not allocated to a duplicated
entry, and forwards the duplicated entries to be resolved later on.

Note that 4e35c1cb9460240e983a01745b5f29fe3a4d8e39 was also backported to stable
releases 4.9.163, 4.14.106, 4.19.29, 4.20.16.

Both commits cherry-pick to 4.15 bionic cleanly. Please cherry-pick both commits
to all bionic kernels.

[Testcase]

The reproducer has been provided by the author of the fixes, and is best viewed
here:

https://github.com/brb/conntrack-race

The dmesg logs and packet traces in the above github repo are best viewed while
reading upstream discussion:

https://patchwork.ozlabs.org/patch/937963/

I have built a test kernel for xenial HWE, which can be found here:
https://launchpad.net/~mruffell/+archive/ubuntu/sf00227747-test-kernel

The test kernel has been tested with a kubernetes workload and aided with
resolution of kubernetes problems.

[Regression Potential]

As noted previously, 4e35c1cb9460240e983a01745b5f29fe3a4d8e39 has been
backported to stable releases 4.9.163, 4.14.106, 4.19.29, 4.20.16 and has no
changes to the primary commit. This commit is well tested and considered stable
by the community.

Both commits have also been present in the -azure kernels since 4.15.0-1030.31
and 4.18.0-1004.4, as per LP #1795493, http://bugs.launchpad.net/bugs/1795493
They have received wide testing, and no problems have been reported.

The changes are specifically limited to resolving clashes between duplicate
conntrack entries, and any regressions will be limited to that scenario. The
overall risk of regression is very low.

[Notes]
The commits in the -azure kernels will need to be reverted before importing the
generic bionic updates, since they are the pre-upstream commits, and have had
several revisions since they were imported. Replacing them with the
actual upstream commits will make things easier in the future to maintain.

http://bugs.launchpad.net/bugs/1795493

For race 1):
Azure: https://patchwork.ozlabs.org/patch/937963/
Upstream: ed07d9a021df6da53456663a76999189badc432a
Commits are the same.

For race 2):
Azure: https://patchwork.ozlabs.org/patch/952939/
Upstream: 4e35c1cb9460240e983a01745b5f29fe3a4d8e39
Note the difference, old commit must be reverted first.