Remove FFDH in TLS 1.2

[mpg]

There's a mismatch between what TLS 1.2 expects and what PSA Crypto provides regarding FFDH. See the documentation on PSA limitations for details. (Note: this is only a problem for (D)TLS 1.2, not (D)TLS 1.3.)

This task is:

1. Come to a decision. Done: we're fully removing FFDH in TLS 1.2.
2. Create a series of estimated tasks implementing that decision.

[mpg]

Note that the relevance of FFDH-with-TLS-1.2 is currently quite low and will presumably only decrease over time as TLS 1.3 becomes dominant. Here are some stats from a crawler on top web sites:

Total: 561143
TLS 1.3: 380793 (67.86%)
ECDH: 177593 (31.65% total, 98.47% pre-1.3)
FFDH: 1338 (0.24% total, 0.74% pre-1.3)
other: 1419 (0.25% total, 0.79% pre-1.3)

Even restricting ourselves to versions of TLS lower than 1.3, FFDH does not even represent 1% of the connections. Presumably this would be even less in the constrained space, as people have a stronger incentive to switch to ECDH which uses less RAM and computation time / energy.

For reference, here's the script that was used to generate those stats:

#!/usr/bin/python3

# Usage:
# wget -q -O- https://crawler.ninja/files/ciphers.txt | python3 cs.py

import re
import sys

(total, tls13, ecdh, ffdh, other) = (0, 0, 0, 0, 0)

for l in sys.stdin.readlines():
    if l.strip() == "Cipher Suites:":
        continue

    (cs, n) = l.strip().split()

    n = int(re.sub(",", "", n))
    total += n

    if cs.startswith("TLS_"):
        tls13 += n
    elif cs.startswith("ECDHE-"):
        ecdh += n
    elif cs.startswith("DHE-"):
        ffdh += n
    else:
        other += n

pre13 = total - tls13

print("Total:", total)
print("TLS 1.3: {} ({:.2f}%)".format(tls13, tls13 / total * 100))
print("ECDH: {} ({:.2f}% total, {:.2f}% pre-1.3)".format(
    ecdh, ecdh / total * 100, ecdh / pre13 * 100))
print("FFDH: {} ({:.2f}% total, {:.2f}% pre-1.3)".format(
    ffdh, ffdh / total * 100, ffdh / pre13 * 100))
print("other: {} ({:.2f}% total, {:.2f}% pre-1.3)".format(
    other, other / total * 100, other / pre13 * 100))

[mpg]

Based on this data, it is tempting to:

• do nothing until 4.0, that is TLS-DHE-xxx ciphersuites keep using the legacy API even when USE_PSA_CRYPTO is enabled;
• in 4.0, simply remove support for these ciphersuites in TLS 1.2 (FFDH can still be used in TLS 1.3 for people who want algorithm variety in case ECC collapses)

[mpg]

Reminder to self: advertise this plan on the mailing-list and ask for feedback.

[mpg]

Note: FFDH support will be added to TLS 1.3 (based on PSA) by #5979 - currently planned for the next quarter.

[mpg]

Sent an email to the list asking people to speak up if the above plan would cause trouble for them.

[mpg]

Labeling "api-break" based on the current plan, so that we don't forget about it when preparing 4.0.

[thomas-fossati]

Just noting that this is in line with the upcoming TLS BCP -- see in particular the bit starting with "However, [...]" at https://www.ietf.org/archive/id/draft-ietf-uta-rfc7525bis-08.html#section-4.1-2.7.1

[mpg]

Just noting that this is in line with the upcoming TLS BCP -- see in particular the bit starting with "However, [...]" at https://www.ietf.org/archive/id/draft-ietf-uta-rfc7525bis-08.html#section-4.1-2.7.1

Note: the draft mentioned is now an RFC (since November 2022) with official Best Current Practice status.

Now there's also a WG draft to formally deprecate FFDH in TLS 1.2 (as well a RSA-encryption-based key exchanges).

[daverodgman]

Marking as SHOULD because this removes the final public-facing dependency on bignum.h: #6792 (comment)

[gilles-peskine-arm]

Based on the (lack of) feedback on the 2022 mailing list thread and on the reasons cited in this issue, we have decided to remove FFDH support in TLS 1.2 in Mbed TLS 4.0.

Next steps: identify what code to remove:

• Remove the options MBEDTLS_KEY_EXCHANGE_DHE_PSK_ENABLED and MBEDTLS_KEY_EXCHANGE_DHE_RSA_ENABLED, and all the code guarded by these options alone (in library, unit tests, test programs and ssl-opt.sh).
• Remove all code guarded by MBEDTLS_DHM_C in TLS (including the dhm_ctx field in struct mbedtls_ssl_handshake_params). (TLS 1.3, which keeps supporting FFDH, uses PSA, not mbedtls_dhm_ functions.)
• Remove DHM cipher suite support from compat.sh.
• More? E.g. are there buffer sizes that should become shorter now that only ECDH is supported for TLS 1.2 key exchange?

[mpg]

More? E.g. are there buffer sizes that should become shorter now that only ECDH is supported for TLS 1.2 key exchange?

I don't think extra work should be needed for that. All buffer sizes that depend on whether DHM is enabled should already be guarded by MBEDTLS_KEY_EXCHANGE_DHE_RSA_ENABLED or MBEDTLS_DHM_C. If that isn't the case, we're already using suboptimal buffer sizes in configs without FFDH, which I'd consider a pre-existing bug that removal of FFDH-in-1.2 isn't making worse.

Though of course we should keep an eye out for this, because if we have such bugs, then this would be an excellent opportunity to catch them.

[mpg]

Note: there are some tests cases in ssl-opt.sh that use DHE-RSA for testing other things (like checking of the keyUsage extension in certificates) that will need to be replace - presumably just moving them use use ECDHE-RSA instead.

[gilles-peskine-arm]

This has been broken down into:

• Remove DHE-PSK key exchange #9684
• Migrate DHE test cases to ECDHE #9688
• Remove the DHE-RSA key exchange #9685