[PATCH v4 00/21] net/tcp: Add TCP-AO support

From: Dmitry Safonov
Date: Wed Feb 15 2023 - 13:34:06 EST


In TODO (expect in next versions):
- Documentation/ page about TCP-AO kernel design, UAPI
- setsockopt(TCP_REPAIR) with TCP-AO
- getsockopt() to return TCP-AO counters (per-MKT & per-ao_info)
- tcp_{v4,v6}_send_reset() to support TCP-AO RST signing on request sockets
- TCP-AO and TCP-MD5 interraction in non/default VRFs: more selftests
- check getsockopt() with VRFs (currently untested)

This is also available as a git branch for pulling:
https://github.com/0x7f454c46/linux/tree/tcp-ao-v4
And another branch with selftests, that will be sent later separately:
https://github.com/0x7f454c46/linux/tree/tcp-ao-v4-with-selftests

Changes from v3:
- TCP_MD5 dynamic static key enable/disable patches merged separately [4]
- crypto_pool patches were nacked [5], so instead this patch set extends
TCP-MD5-sigpool to be used for TCP-AO as well as for TCP-MD5
- Added missing `static' for tcp_v6_ao_calc_key()
(kernel test robot <lkp@xxxxxxxxx>)
- Removed CONFIG_TCP_AO default=y and added "If unsure, say N."
- Don't leak ao_info and don't create an unsigned TCP socket if there was
a TCP-AO key during handshake, but it was removed from listening socket
while the connection was being established
- Migrate to use static_key_fast_inc_not_disabled() and check return
code of static_branch_inc()
- Change some return codes to EAFNOSUPPORT for error-pathes where
family is neither AF_INET nor AF_INET6
- setsockopt()s on a closed/listen socket might have created stray ao_info,
remove it if connect() is called with a correct TCP-MD5 key, the same
for the reverse situation: remove md5sig_info straight away from the
socket if it's going to be TCP-AO connection
- IPv4-mapped-IPv6 addresses + selftest in fcnal-test.sh (by Salam)
- fix using uninitialized sisn/disn from stack - it would only make
non-SYN packets fail verification on a listen socket, which are not
expected anyway (kernel test robot <lkp@xxxxxxxxx>)
- implicit padding in UAPI TCP-AO structures converted to explicit
(spotted-by David Laight)
- Some selftests missed zero-initializers for uapi structs on stack
- Removed tcp_ao_do_lookup_rcvid() and tcp_ao_do_lookup_sndid() in
favor of unified tcp_ao_matched_key()
- Disallowed setting current/rnext keys on listen sockets - that wasn't
supported and didn't affect anything, cleanup for the UAPI
- VRFs support for TCP-AO

Changes from v2:
- Added more missing `static' declarations for local functions
(kernel test robot <lkp@xxxxxxxxx>)
- Building now with CONFIG_TCP_AO=n and CONFIG_TCP_MD5SIG=n
(kernel test robot <lkp@xxxxxxxxx>)
- Now setsockopt(TCP_AO) is allowed when it's TCP_LISTEN or TCP_CLOSE
state OR the key added is not the first key on a socket (by Salam)
- CONFIG_TCP_AO does not depend on CONFIG_TCP_MD5SIG anymore
- Don't leak tcp_md5_needed static branch counter when TCP-MD5 key
is modified/changed
- TCP-AO lookups are dynamically enabled/disabled with static key when
there is ao_info in the system (and when it is destroyed)
- Wired SYN cookies up to TCP-AO (by Salam)
- Fix verification for possible re-transmitted SYN packets (by Salam)
- use sockopt_lock_sock() instead of lock_sock()
(from v6.1 rebase, commit d51bbff2aba7)
- use sockptr_t in getsockopt(TCP_AO_GET)
(from v6.1 rebase, commit 34704ef024ae)
- Fixed reallocating crypto_pool's scratch area by IPI while
crypto_pool_get() was get by another CPU
- selftests on older kernels (or with CONFIG_TCP_AO=n) should exit with
SKIP, not FAIL (Shuah Khan <shuah@xxxxxxxxxx>)
- selftests that check interaction between TCP-AO and TCP-MD5 now
SKIP when CONFIG_TCP_MD5SIG=n
- Measured the performance of different hashing algorithms for TCP-AO
and compare with TCP-MD5 performance. This is done with hacky patches
to iperf (see [3]). At this moment I've done it in qemu/KVM with CPU
affinities set on Intel(R) Core(TM) i7-7600U CPU @ 2.80GHz.
No performance degradation was noticed before/after patches, but given
the measures were done in a VM, without measuring it on a physical dut
it only gives a hint of relative speed for different hash algorithms
with TCP-AO. Here are results, averaging on 30 measures each:
TCP: 3.51Gbits/sec
TCP-MD5: 1.12Gbits/sec
TCP-AO(HMAC(SHA1)): 1.53Gbits/sec
TCP-AO(CMAC(AES128)): 621Mbits/sec
TCP-AO(HMAC(SHA512)): 1.21Gbits/sec
TCP-AO(HMAC(SHA384)): 1.20Gbits/sec
TCP-AO(HMAC(SHA224)): 961Mbits/sec
TCP-AO(HMAC(SHA3-512)): 157Mbits/sec
TCP-AO(HMAC(RMD160)): 659Mbits/sec
TCP-AO(HMAC(MD5): 1.12Gbits/sec
(the last one is just for fun, but may make sense as it provides
the same security as TCP-MD5, but allows multiple keys and a mechanism
to change them from RFC5925)

Changes from v1:
- Building now with CONFIG_IPV6=n (kernel test robot <lkp@xxxxxxxxx>)
- Added missing static declarations for local functions
(kernel test robot <lkp@xxxxxxxxx>)
- Addressed static analyzer and review comments by Dan Carpenter
(thanks, they were very useful!)
- Fix elif without defined() for !CONFIG_TCP_AO
- Recursively build selftests/net/tcp_ao (Shuah Khan), patches in:
https://lore.kernel.org/all/20220919201958.279545-1-dima@xxxxxxxxxx/T/#u
- Don't leak crypto_pool reference when TCP-MD5 key is modified/changed
- Add TCP-AO support for nettest.c and fcnal-test.sh
(will be used for VRF testing in later versions)

Comparison between Leonard proposal and this (overview):
https://lore.kernel.org/all/3cf03d51-74db-675c-b392-e4647fa5b5a6@xxxxxxxxxx/T/#u

Version 2: https://lore.kernel.org/all/20220923201319.493208-1-dima@xxxxxxxxxx/T/#u
Version 1: https://lore.kernel.org/all/20220818170005.747015-1-dima@xxxxxxxxxx/T/#u

This patchset implements the TCP-AO option as described in RFC5925. There
is a request from industry to move away from TCP-MD5SIG and it seems the time
is right to have a TCP-AO upstreamed. This TCP option is meant to replace
the TCP MD5 option and address its shortcomings. Specifically, it provides
more secure hashing, key rotation and support for long-lived connections
(see the summary of TCP-AO advantages over TCP-MD5 in (1.3) of RFC5925).
The patch series starts with six patches that are not specific to TCP-AO
but implement a general crypto facility that we thought is useful
to eliminate code duplication between TCP-MD5SIG and TCP-AO as well as other
crypto users. These six patches are being submitted separately in
a different patchset [1]. Including them here will show better the gain
in code sharing. Next are 18 patches that implement the actual TCP-AO option,
followed by patches implementing selftests.

The patch set was written as a collaboration of three authors (in alphabetical
order): Dmitry Safonov, Francesco Ruggeri and Salam Noureddine. Additional
credits should be given to Prasad Koya, who was involved in early prototyping
a few years back. There is also a separate submission done by Leonard Crestez
whom we thank for his efforts getting an implementation of RFC5925 submitted
for review upstream [2]. This is an independent implementation that makes
different design decisions.

For example, we chose a similar design to the TCP-MD5SIG implementation and
used setsockopts to program per-socket keys, avoiding the extra complexity
of managing a centralized key database in the kernel. A centralized database
in the kernel has dubious benefits since it doesn’t eliminate per-socket
setsockopts needed to specify which sockets need TCP-AO and what are the
currently preferred keys. It also complicates traffic key caching and
preventing deletion of in-use keys.

In this implementation, a centralized database of keys can be thought of
as living in user space and user applications would have to program those
keys on matching sockets. On the server side, the user application programs
keys (MKTS in TCP-AO nomenclature) on the listening socket for all peers that
are expected to connect. Prefix matching on the peer address is supported.
When a peer issues a successful connect, all the MKTs matching the IP address
of the peer are copied to the newly created socket. On the active side,
when a connect() is issued all MKTs that do not match the peer are deleted
from the socket since they will never match the peer. This implementation
uses three setsockopt()s for adding, deleting and modifying keys on a socket.
All three setsockopt()s have extensive sanity checks that prevent
inconsistencies in the keys on a given socket. A getsockopt() is provided
to get key information from any given socket.

Few things to note about this implementation:
- Traffic keys are cached for established connections avoiding the cost of
such calculation for each packet received or sent.
- Great care has been taken to avoid deleting in-use MKTs
as required by the RFC.
- Any crypto algorithm supported by the Linux kernel can be used
to calculate packet hashes.
- Fastopen works with TCP-AO but hasn’t been tested extensively.
- Tested for interop with other major networking vendors (on linux-4.19),
including testing for key rotation and long lived connections.

[1]: https://lore.kernel.org/all/20220726201600.1715505-1-dima@xxxxxxxxxx/
[2]: https://lore.kernel.org/all/cover.1658815925.git.cdleonard@xxxxxxxxx/
[3]: https://github.com/0x7f454c46/iperf/tree/tcp-md5-ao
[4]: https://lore.kernel.org/all/166995421700.16716.17446147162780881407.git-patchwork-notify@xxxxxxxxxx/T/#u
[5]: https://lore.kernel.org/all/Y8kSkW4X4vQdFyOl@xxxxxxxxxxxxxxxxxxx/T/#u

Cc: Andy Lutomirski <luto@xxxxxxxxxxxxxx>
Cc: Ard Biesheuvel <ardb@xxxxxxxxxx>
Cc: Bob Gilligan <gilligan@xxxxxxxxxx>
Cc: Dan Carpenter <dan.carpenter@xxxxxxxxxx>
Cc: David Ahern <dsahern@xxxxxxxxxx>
Cc: David Laight <David.Laight@xxxxxxxxxx>
Cc: "David S. Miller" <davem@xxxxxxxxxxxxx>
Cc: Dmitry Safonov <0x7f454c46@xxxxxxxxx>
Cc: Eric Biggers <ebiggers@xxxxxxxxxx>
Cc: Eric Dumazet <edumazet@xxxxxxxxxx>
Cc: "Eric W. Biederman" <ebiederm@xxxxxxxxxxxx>
Cc: Francesco Ruggeri <fruggeri05@xxxxxxxxx>
Cc: Herbert Xu <herbert@xxxxxxxxxxxxxxxxxxx>
Cc: Hideaki YOSHIFUJI <yoshfuji@xxxxxxxxxxxxxx>
Cc: Ivan Delalande <colona@xxxxxxxxxx>
Cc: Jakub Kicinski <kuba@xxxxxxxxxx>
Cc: Leonard Crestez <cdleonard@xxxxxxxxx>
Cc: Paolo Abeni <pabeni@xxxxxxxxxx>
Cc: Salam Noureddine <noureddine@xxxxxxxxxx>
Cc: netdev@xxxxxxxxxxxxxxx
Cc: linux-kernel@xxxxxxxxxxxxxxx

Dmitry Safonov (21):
net/tcp: Prepare tcp_md5sig_pool for TCP-AO
tcp: Add TCP-AO config and structures
net/tcp: Introduce TCP_AO setsockopt()s
net/tcp: Prevent TCP-MD5 with TCP-AO being set
net/tcp: Calculate TCP-AO traffic keys
net/tcp: Add TCP-AO sign to outgoing packets
net/tcp: Add tcp_parse_auth_options()
net/tcp: Add AO sign to RST packets
net/tcp: Add TCP-AO sign to twsk
net/tcp: Wire TCP-AO to request sockets
net/tcp: Sign SYN-ACK segments with TCP-AO
net/tcp: Verify inbound TCP-AO signed segments
net/tcp: Add TCP-AO segments counters
net/tcp: Add TCP-AO SNE support
net/tcp: Add tcp_hash_fail() ratelimited logs
net/tcp: Ignore specific ICMPs for TCP-AO connections
net/tcp: Add option for TCP-AO to (not) hash header
net/tcp: Add getsockopt(TCP_AO_GET)
net/tcp: Allow asynchronous delete for TCP-AO keys (MKTs)
net/tcp-ao: Add static_key for TCP-AO
net/tcp-ao: Wire up l3index to TCP-AO

include/linux/sockptr.h | 23 +
include/linux/tcp.h | 30 +-
include/net/dropreason.h | 25 +
include/net/tcp.h | 207 +++-
include/net/tcp_ao.h | 306 ++++++
include/uapi/linux/snmp.h | 5 +
include/uapi/linux/tcp.h | 73 ++
net/ipv4/Kconfig | 17 +
net/ipv4/Makefile | 2 +
net/ipv4/proc.c | 5 +
net/ipv4/syncookies.c | 4 +
net/ipv4/tcp.c | 192 ++--
net/ipv4/tcp_ao.c | 2130 +++++++++++++++++++++++++++++++++++++
net/ipv4/tcp_input.c | 105 +-
net/ipv4/tcp_ipv4.c | 354 ++++--
net/ipv4/tcp_minisocks.c | 35 +-
net/ipv4/tcp_output.c | 211 +++-
net/ipv4/tcp_sigpool.c | 333 ++++++
net/ipv6/Makefile | 1 +
net/ipv6/syncookies.c | 5 +
net/ipv6/tcp_ao.c | 149 +++
net/ipv6/tcp_ipv6.c | 372 +++++--
22 files changed, 4251 insertions(+), 333 deletions(-)
create mode 100644 include/net/tcp_ao.h
create mode 100644 net/ipv4/tcp_ao.c
create mode 100644 net/ipv4/tcp_sigpool.c
create mode 100644 net/ipv6/tcp_ao.c


base-commit: e1c04510f521e853019afeca2a5991a5ef8d6a5b
--
2.39.1