Manuel Pégourié-Gonnard | 06adca4 | 2024-05-23 09:12:24 +0200 | [diff] [blame] | 1 | This document explains feature guards macros to be used during the transition |
| 2 | from legacy to PSA in order to determine whether a given cryptographic |
| 3 | mechanism is available in the current build. |
| 4 | |
| 5 | We currently (as of Mbed TLS 3.6) have three sets of feature macros: |
| 6 | - `PSA_WANT` macros; |
| 7 | - legacy `MBEDTLS_xxx` macros; |
| 8 | - transitional `MBEDTLS_xxx` macros that stem from the desire to be able to |
| 9 | use crypto mechanisms that are only provided by a driver (G5 in |
| 10 | `strategy.md`). |
| 11 | |
| 12 | This document's goal is to shed some light on when to use which. It is mostly |
| 13 | intended for maintainers. |
| 14 | |
| 15 | Since most transition macros come from driver-only work, it can be useful to |
| 16 | check `docs/driver-only-builds.md` as well for background. (Note: as |
| 17 | maintainers, for the best precision about what's supported of not with |
| 18 | drivers, check the relevant `component_test_psa_crypto_config_accel_xxx`'s |
| 19 | configuration, as well as the corresponding exclude list in |
| 20 | `analyze_outcomes.py`.) |
| 21 | |
| 22 | General considerations |
| 23 | ====================== |
| 24 | |
| 25 | This document only applies to Mbed TLS 3.6 TLS. By contrast: |
| 26 | - in 2.28 we have no driver-only support, so the legacy guards `MBEDTLS_XXX` |
| 27 | should be used everywhere; |
| 28 | - in 4.0 configuration will be purely based on PSA, so `PSA_WANT` macros |
| 29 | should be used everywhere. |
| 30 | |
| 31 | It is useful to consider the following domains: |
| 32 | - The PSA domain: things declared in `include/psa/*.h`, implemented in |
| 33 | `library/psa_*.c` and tested in `tests/suites/test_suite_psa*`. |
| 34 | - The pure TLS 1.3 domain: the parts of TLS 1.3 that are not in the `USE_PSA` |
| 35 | domain (see below). Those use PSA APIs unconditionally. |
Manuel Pégourié-Gonnard | 4975232 | 2024-07-23 12:58:35 +0200 | [diff] [blame^] | 36 | - The `USE_PSA` domain (that is, code that calls PSA crypto APIs when |
| 37 | `USE_PSA` is enabled, and legacy crypto APIs otherwise): that's PK, X.509, |
| 38 | most of TLS 1.2 and the parts of TLS 1.3 that are common with TLS 1.2 or are |
| 39 | about public/private keys (see `docs/use-psa-crypto.md` for details). |
Manuel Pégourié-Gonnard | 06adca4 | 2024-05-23 09:12:24 +0200 | [diff] [blame] | 40 | - The legacy crypto domain: a number of modules there will use crypto from |
| 41 | other modules, for example RSA and entropy will use hashes, PEM will use |
| 42 | hashes and ciphers (from encrypted PEM), etc. |
| 43 | |
| 44 | The first two categories (PSA domain, pure TLS 1.3 domain) are simple: as a |
| 45 | general rule, use `PSA_WANT` macros. (With very few exceptions, see |
Manuel Pégourié-Gonnard | 4975232 | 2024-07-23 12:58:35 +0200 | [diff] [blame^] | 46 | `component_check_test_dependencies` in `all.sh`.) In the rare instances where it is necessary to |
Manuel Pégourié-Gonnard | 06adca4 | 2024-05-23 09:12:24 +0200 | [diff] [blame] | 47 | check whether a mechanism is built-in or provided by a driver, |
| 48 | `MBEDTLS_PSA_BUILTIN_xxx` and `MBEDTLS_PSA_ACCEL_xxx` macros should be used |
| 49 | (but not legacy `MBEDTLS_xxx` macros). |
| 50 | |
Manuel Pégourié-Gonnard | 4975232 | 2024-07-23 12:58:35 +0200 | [diff] [blame^] | 51 | For the `USE_PSA` domain, it should always be correct to use expressions like |
| 52 | `(!USE_PSA && MBEDTLS_xxx) || (USE_PSA && PSA_WANT_xxx)`. Sometimes, macros |
| 53 | are defined in order to avoid using long expressions everywhere; they will be |
| 54 | mentioned in the following sections. |
Manuel Pégourié-Gonnard | 06adca4 | 2024-05-23 09:12:24 +0200 | [diff] [blame] | 55 | |
Manuel Pégourié-Gonnard | 4975232 | 2024-07-23 12:58:35 +0200 | [diff] [blame^] | 56 | The remaining category, the legacy domain, tends to be more complex. There are |
| 57 | different rules for different families of mechanisms, as detailed in the |
| 58 | following sections. |
Manuel Pégourié-Gonnard | 06adca4 | 2024-05-23 09:12:24 +0200 | [diff] [blame] | 59 | |
| 60 | Symmetric crypto |
| 61 | ================ |
| 62 | |
| 63 | Hashes |
| 64 | ------ |
| 65 | |
| 66 | **Hash vs HMAC:** Historically (since 2.0) we've had the generic hash |
| 67 | interface, and the implementation of HMAC, in the same file controlled by a |
Manuel Pégourié-Gonnard | 3342e80 | 2024-07-23 09:57:57 +0200 | [diff] [blame] | 68 | single feature macro: `MBEDTLS_MD_C`. This has now been split in two: |
Manuel Pégourié-Gonnard | 06adca4 | 2024-05-23 09:12:24 +0200 | [diff] [blame] | 69 | - `MBEDTLS_MD_LIGHT` is about the generic hash interface; we could think of it |
| 70 | as `MBEDTLS_HASH_C`. |
Manuel Pégourié-Gonnard | 3342e80 | 2024-07-23 09:57:57 +0200 | [diff] [blame] | 71 | - `MBEDTLS_MD_C` is about the HMAC implementation; we could think of it as |
Manuel Pégourié-Gonnard | 06adca4 | 2024-05-23 09:12:24 +0200 | [diff] [blame] | 72 | `MBEDTLS_HMAC_C` (auto-enabling `MBEDTLS_HASH_C`). |
Manuel Pégourié-Gonnard | 3342e80 | 2024-07-23 09:57:57 +0200 | [diff] [blame] | 73 | |
Manuel Pégourié-Gonnard | 06adca4 | 2024-05-23 09:12:24 +0200 | [diff] [blame] | 74 | (In fact, this is not the whole story: `MD_LIGHT` is the _core_ of the generic |
| 75 | hash interface, excluding functions such as `mbedtls_md_list()` and |
| 76 | `mbedtls_md_info_from_string()`, `mbedtls_md_file()`, etc. But I think the |
| 77 | above should still provide a good intuition as first approximation.) |
| 78 | |
Manuel Pégourié-Gonnard | 4975232 | 2024-07-23 12:58:35 +0200 | [diff] [blame^] | 79 | Note that all users of hashes in the library use either the PSA Crypto API or the `md.h` API. |
| 80 | That is, no user in the library, even in the legacy domain, uses the low-level hash APIs |
| 81 | (`mbedtls_sha256` etc). (That's not true of all example programs, though.) |
Manuel Pégourié-Gonnard | 06adca4 | 2024-05-23 09:12:24 +0200 | [diff] [blame] | 82 | |
| 83 | **Helper macros:** in `config_adjust_legacy_crypto.h` we define a family of |
| 84 | macro `MBEDTLS_MD_CAN_xxx`. These macros are defined (for available hashes) as |
| 85 | soon as `MBEDTLS_MD_LIGHT` is enabled. This subset of `MD` is automatically |
| 86 | enabled as soon as something from the legacy domain, or from the `USE_PSA` |
Manuel Pégourié-Gonnard | 3342e80 | 2024-07-23 09:57:57 +0200 | [diff] [blame] | 87 | domain, needs a hash. (Note that this includes `ENTROPY_C`, so in practice |
Manuel Pégourié-Gonnard | 06adca4 | 2024-05-23 09:12:24 +0200 | [diff] [blame] | 88 | `MD_LIGHT` is enabled in most builds.) |
| 89 | |
| 90 | Note that there is a rule, enforced by `config_adjust_psa_superset_legacy.h`, |
| 91 | that all hashes that are enabled on the legacy side are also enabled on the |
| 92 | PSA side. So, in practice, when `MD_LIGHT` is enabled, `PSA_WANT_ALG_xxx` and |
| 93 | `MBEDTLS_MD_CAN_xxx` are equivalent. |
| 94 | |
| 95 | **Legacy and `USE_PSA` domains:** for hashes, `MBEDTLS_MD_CAN_xxx` (where |
| 96 | `xxx` is the legacy name of the hash) can be used everywhere (except in the |
| 97 | PSA domain which should use `PSA_WANT` as usual). No special include is |
| 98 | required, `build_info.h` or `common.h` is enough. |
| 99 | |
| 100 | **Pure TLS 1.3 domain:** it is not easy to know which uses of hashes fall in |
| 101 | this domain as opposed to the `USE_PSA` domain which looking at the code. |
| 102 | Fortunately, `MD_CAN` and `PSA_WANT` macros can be used interchangeably, as |
| 103 | per the note above. |
| 104 | |
| 105 | |
| 106 | |
| 107 | HMAC |
| 108 | ---- |
| 109 | |
| 110 | **Legacy domain:** the code is using the `md.h` API. For this domain, |
| 111 | availability of HMAC-xxx is determined by `MBEDTLS_MD_C && MBEDTLS_MD_CAN_xxx` |
| 112 | (see previous subsection about `MD_CAN`). Modules in this domain that may use |
| 113 | HMAC are PKCS5, PKCS7, HKDF, HMAC-DRBG and ECDSA deterministic. |
| 114 | |
Manuel Pégourié-Gonnard | 4975232 | 2024-07-23 12:58:35 +0200 | [diff] [blame^] | 115 | **`USE_PSA` domain:** code will use the `md.h` API when `USE_PSA` is disabled, |
| 116 | and the `psa_mac` API when `USE_PSA` is enabled. It should check for the |
| 117 | availability of HMAC-xxx with either: |
Manuel Pégourié-Gonnard | 06adca4 | 2024-05-23 09:12:24 +0200 | [diff] [blame] | 118 | ``` |
| 119 | ((!MBEDTLS_USE_PSA_CRYPTO && MBEDTLS_MD_C) || |
| 120 | (MBEDTLS_USE_PSA_CRYPTO && PSA_WANT_ALG_HMAC)) && |
| 121 | MBEDTLS_MD_CAN_xxx |
| 122 | ``` |
| 123 | or |
| 124 | ``` |
| 125 | (!MBEDTLS_USE_PSA_CRYPTO && MBEDTLS_MD_C && MBEDTLS_xxx_C) || |
| 126 | (MBEDTLS_USE_PSA_CRYPTO && PSA_WANT_ALG_HMAC && PSA_WANT_ALG_xxx) |
| 127 | ``` |
| 128 | or any equivalent condition (see note at the end of the previous section). |
| 129 | The only module in this case is TLS, which currently depends on |
| 130 | `USE_PSA_CRYPTO || MD_C`. |
| 131 | |
| 132 | Note: while writing this, it occurs to me that TLS 1.2 does not seem to be |
| 133 | checking for `PSA_WANT_ALG_HMAC` before enabling CBC ciphersuites when |
| 134 | `USE_PSA` is enabled, which I think it should. Builds with `USE_PSA` enabled, |
| 135 | `PSA_WANT_ALG_HMAC` disabled and other requirements for CBC ciphersuites |
| 136 | enabled, are probably broken (perhaps only at runtime when a CBC ciphersuite |
| 137 | is negotiated). |
| 138 | |
| 139 | **Pure TLS 1.3 domain:** HMAC is used for the Finished message via PSA Crypto |
| 140 | APIs. So, TLS 1.3 should depend on `PSA_WANT_ALG_HMAC` - doesn't seem to be |
Manuel Pégourié-Gonnard | 4975232 | 2024-07-23 12:58:35 +0200 | [diff] [blame^] | 141 | enforced by `check_config.h`, or documented in `mbedtls_config.h`, at the |
| 142 | moment. |
Manuel Pégourié-Gonnard | 06adca4 | 2024-05-23 09:12:24 +0200 | [diff] [blame] | 143 | |
| 144 | Ciphers (AEAD and unauthenticated) |
| 145 | ---------------------------------- |
| 146 | |
| 147 | **Overview of existing (internal) APIs:** we currently have 4 (families of) |
| 148 | APIs for ciphers in the library: |
| 149 | - Low-level API: `mbedtls_aes_xxx` etc. - used by `cipher.c` and some other |
| 150 | modules in the legacy domain. |
| 151 | - Internal abstraction layer `block_cipher` for AES, ARIA and Camellia |
| 152 | primitives - used only by `gcm.c` and `ccm.c`, only when `CIPHER_C` is not |
| 153 | enabled (for compatibility reasons). |
| 154 | - Cipher: used by some modules in the legacy domain, and by the built-in PSA |
| 155 | implementation. |
| 156 | - PSA: used by the `USE_PSA` domain when `MBEDTLS_USE_PSA_CRYPTO` is enabled. |
| 157 | |
| 158 | **Legacy domain:** most code here is using either `cipher.h` or low-level APIs |
| 159 | like `aes.h`, and should use legacy macros like `MBEDTLS_AES_C` and |
Manuel Pégourié-Gonnard | 3342e80 | 2024-07-23 09:57:57 +0200 | [diff] [blame] | 160 | `MBEDTLS_CIPHER_MODE_CBC`. This includes NIST-KW, CMAC, PKCS5/PKCS12 en/decryption |
Manuel Pégourié-Gonnard | 06adca4 | 2024-05-23 09:12:24 +0200 | [diff] [blame] | 161 | functions, PEM decryption, PK parsing of encrypted keys. The only exceptions |
| 162 | are `GCM` and `CCM` which use the internal abstraction layer `block_cipher` |
| 163 | and check for availability of block ciphers using `MBEDTLS_CCM_GCM_CAN_xxx` |
| 164 | macros defined in `config_adjut_legacy_crypto.h`. |
| 165 | |
| 166 | **`USE_PSA` domain:** here we should use conditions like the following in |
| 167 | order to test for availability of ciphers and associated modes. |
| 168 | ``` |
| 169 | // is AES available? |
| 170 | (!defined(MBEDTLS_USE_PSA_CRYPTO) && defined(MBEDTLS_AES_C)) || \ |
| 171 | (defined(MBEDTLS_USE_PSA_CRYPTO) && defined(PSA_WANT_KEY_TYPE_AES)) |
| 172 | // is CBC available? |
| 173 | (!defined(MBEDTLS_USE_PSA_CRYPTO) && defined(MBEDTLS_CIPHER_MODE_CBC)) || \ |
| 174 | (defined(MBEDTLS_USE_PSA_CRYPTO) && defined(PSA_WANT_ALG_CBC_NO_PADDING)) |
| 175 | // is GCM available? |
| 176 | (!defined(MBEDTLS_USE_PSA_CRYPTO) && defined(MBEDTLS_GCM_C)) || \ |
| 177 | (defined(MBEDTLS_USE_PSA_CRYPTO) && defined(PSA_WANT_ALG_GCM)) |
| 178 | ``` |
| 179 | Note: TLS is the only user of ciphers in the `USE_PSA` domain, and it defines |
| 180 | `MBEDTLS_SSL_HAVE_xxx` macros in `config_adjust_legacy_crypto.h` for the |
| 181 | ciphers and modes it needs to know about. |
| 182 | |
| 183 | **Pure TLS 1.3 domain:** none. All from TLS 1.3 are in the `USE_PSA` domain |
| 184 | (common to TLS 1.2). |
| 185 | |
| 186 | Key derivation |
| 187 | -------------- |
| 188 | |
| 189 | **Legacy and `USE_PSA` domains:** no users here. |
| 190 | |
| 191 | **Pure TLS 1.3 domain:** TLS 1.3 is using HKDF via PSA Crypto APIs. We already |
| 192 | enforce in `check_config.h` that TLS 1.3 depends on the appropriate `PSA_WANT` |
| 193 | macros. |
| 194 | |
| 195 | Asymmetric crypto |
| 196 | ================= |
| 197 | |
| 198 | RSA |
| 199 | --- |
| 200 | |
| 201 | **Legacy domain and `USE_PSA` domain:** use `RSA_C` everywhere. (Note: there's |
| 202 | no user of RSA in the legacy domain, and the only direct user in the `USE_PSA` |
| 203 | domain is PK - both X.509 and TLS will only RSA via PK.) |
| 204 | |
| 205 | **Pure TLS 1.3 domain:** no use of RSA in this domain. All TLS 1.3 uses of RSA |
| 206 | go through PK, hence are in the `USE_PSA` domain. |
| 207 | |
| 208 | FFDH |
| 209 | ---- |
| 210 | |
| 211 | **Legacy domain and `USE_PSA` domain:** use `DHM_C`. The only user is TLS 1.2 |
| 212 | which is actually in the legacy domain - this is an exception where `USE_PSA` |
| 213 | has no effect, because PSA doesn't cover the needs of TLS 1.2 here. |
| 214 | |
| 215 | **Pure TLS 1.3 domain:** use `PSA_WANT`. The TLS 1.3 code for Diffie-Hellman |
| 216 | is common to ECDH and FFDH thanks to PSA Crypto APIs being generic enough. The |
| 217 | parts about FFDH are guarded with `PSA_WANT_ALG_FFDH` (with the reasoning that |
| 218 | this implies support for the corresponding key type). |
| 219 | |
| 220 | ECC |
| 221 | --- |
| 222 | |
| 223 | **Curves:** in `config_adjut_psa_superset_legacy.h` we ensure that all |
| 224 | curves that are supported on the legacy side (`MBEDTLS_ECP_DP_xxx_ENABLED`) |
| 225 | are also supported on the PSA side (`PSA_WANT_ECC_xxx`). |
| 226 | |
| 227 | In `config_adjust_legacy_crypto.h` we define macros `MBEDTLS_ECP_HAVE_xxx`. |
| 228 | These macros are useful for data and functions that have users in several |
| 229 | domains, such as `mbedtls_ecc_group_to_psa()`, or that have users only in the |
| 230 | `USE_PSA` domain but want a simpler (if sub-optimal) condition, such as |
| 231 | `mbedtls_oid_get_ec_grp()`. |
| 232 | |
| 233 | Strictly speaking, code in the `USE_PSA` domain should not use the above |
| 234 | `MBEDTLS_ECP_HAVE_xxx` macros but conditions like |
| 235 | ``` |
| 236 | (!MBEDTLS_USE_PSA_CRYPTO && MBEDTLS_ECP_DP_xxx_ENABLED) || |
| 237 | (MBEDTLS_USE_PSA_CRYPTO && PSA_WANT_ECC_xxx) |
| 238 | ``` |
| 239 | Note while writing: a lot of tests for things in the `USE_PSA` domain appear |
| 240 | to be using `MBEDTLS_ECP_HAVE_xxx`. IMO this is incorrect, but not caught by |
| 241 | the CI because I guess we don't run tests in configurations that have both |
| 242 | `USE_PSA_CRYPTO` disabled, and some curves enabled only on the PSA side. My |
| 243 | initial feeling is we don't care about such configurations as this point, and |
| 244 | can leave the dependencies as they are until they're replaced with `PSA_WANT` |
| 245 | macros in 4.0 anyway. |
| 246 | |
| 247 | **Legacy domain:** use the legacy macros `ECP_C`, `ECDH_C`, `ECDSA_C`, |
| 248 | `ECJPAKE_C`, `MBEDTLS_ECP_DP_xxx_ENABLED`. (This is mostly just ECDH, ECDSA |
| 249 | and EC J-PAKE using ECP.) |
| 250 | |
| 251 | **Key management, `USE_PSA` domain:** `MBEDTLS_PK_HAVE_ECC_KEYS` means that PK |
| 252 | supports ECC key parsing and writing (and storage). It does not imply support |
| 253 | for doing crypto operation with such keys - see `MBEDTLS_PK_CAN_ECDSA_xxx` |
| 254 | above for that. |
| 255 | |
| 256 | **ECDH, `USE_PSA` domain:** this is just TLS 1.2. It's using the helper macro |
| 257 | `MBEDTLS_CAN_ECDH` defined in `config_adjust_legacy_crypto.h` (which should |
| 258 | probably be called `MBEDTLS_SSL_TLS1_2_CAN_ECDH` as it's only for TLS 1.2). |
| 259 | (Note: the macro is not used directly in the code, it's only used as a |
| 260 | dependency for relevant TLS 1.2 key exchanges. Then the code uses the guards |
| 261 | for the key exchanges.) |
| 262 | |
| 263 | **ECDH, pure TLS 1.3 domain:** using `PSA_WANT_ALG_ECDH`. |
| 264 | |
| 265 | **ECDSA, `USE_PSA` domain:** should use the macros |
| 266 | `MBEDTLS_PK_CAN_ECDSA_{SIGN,VERIFY,SOME}` that indicate support for signature |
| 267 | generation, verification, or at least one of those, respectively. To check for |
| 268 | support for signatures with a specific hash, combine |
| 269 | `MBEDTLS_PK_CAN_ECDSA_xxx` with `MBEDTLS_MD_CAN_xxx`. |
| 270 | |
| 271 | **ECDSA, pure TLS 1.3 domain:** none - everything goes through PK. |
| 272 | |
| 273 | **EC J-PAKE, `USE_PSA` domain:** only used by TLS 1.2. The code is guarded by |
| 274 | the corresponding `KEY_EXCHANGE` macro, which in `check_config.h` depends on |
| 275 | the appropriate macros depending on whether `USE_PSA` is on or off. |
| 276 | |
| 277 | **EC J-PAKE, pure TLS 1.3 domain:** none - EC J-PAKE is TLS 1.2 (so far). |
| 278 | |
| 279 | **Related internal macros:** |
| 280 | - `MBEDTLS_PK_USE_PSA_EC_DATA` is an internal switch of the PK module. When |
| 281 | it's not defined, PK stores ECC keys as a `struct mbedtls_ecxxx_keypair`; |
| 282 | when it's defined, PK stores in a PSA -friendly format instead (PSA key slot |
| 283 | for private keys, metadata + array of bytes with the PSA import/export format |
| 284 | for the public part). This macro is only defined when `ECP_C` is not and |
| 285 | `USE_PSA` is, see comments above its definition in `pk.h` for details. |
| 286 | - `MBEDTLS_ECP_LIGHT` enables only a subset of `ecp.c`. This subset is pretty |
| 287 | much ad hoc: it's basically everything that doesn't depend on scalar |
| 288 | multiplication (_the_ complex expensive operation in ECC arithmetic). |
| 289 | Basically, this subset gives access to curve data (constants), key storage, |
| 290 | basic parsing and writing. It is auto-enabled in some driver-only |
| 291 | configurations where the user has disabled `ECP_C` because they have drivers |
| 292 | for the crypto operations they use, but they've also asked for some things |
| 293 | that are not supported by drivers yet, such as deterministic key derivation, |
| 294 | or parsing of compressed keys - on those cases, `ECP_LIGHT` will support this |
| 295 | needs without bringing back the full `ECP_C`. |