On Wed, 24 Apr 2024 07:14:31 GMT, Andrew Haley <a...@openjdk.org> wrote:
>> Hi, >> >> I would like to propose an implementation to support AES CBC with Ciphertext >> Stealing (CTS) in SunPKCS11, according to what has been specified in >> [JDK-8330843 CSR](https://bugs.openjdk.org/browse/JDK-8330843). >> >> What follows are implementation notes that describe the most relevant >> aspects of the proposed change. >> >> ### Buffering >> >> A buffering mechanism was implemented so PKCS #11 tokens only receive >> multipart input updates (`C_EncryptUpdate` or `C_DecryptUpdate`) in >> multiples of the block size. This mechanism protects against tokens —such as >> NSS— that assume an update with data not multiple of the block size is final >> and do variant ordering between the last 2 blocks, returning an incorrect >> partial output and resetting state. For example, when encrypting, a token >> may receive an update with an input not multiple of the block size and >> prematurely finalize the operation returning the last two blocks of >> ciphertext according to its ordering. By buffering on the Java side, the >> token is not aware of the end of the stream during updates and SunPKCS11 >> retains full control to do the last two blocks at `C_EncryptFinal` or >> `C_DecryptFinal`. A bug in NSS (see >> [1823875](https://bugzilla.mozilla.org/show_bug.cgi?id=1823875#c2)) requires >> buffering three blocks instead of two. If this bug gets fixed, the three >> block b uffering will still work and allow it to support old versions of the NSS library. >> >> ### `implUpdate` implementation >> >> The code added to `P11Cipher::implUpdate` follows the existing three-stage >> strategy: 1) Process data in the `padBuffer`, 2) Process data in the input >> and 3) Buffer to `padBuffer`. Stage #1 for CTS has some additional >> complexity that is worth describing in this section. >> >> The stage begins by calculating the total data available (what is already >> buffered in `padBuffer` + the new input) and assigning this value to the >> variable `totalInLen`. `newPadBufferLen` is the number of bytes that must be >> buffered at the end of the update operation, irrespective of where they come >> from (`padBuffer` or input). This number of bytes is determined out of >> `totalInLen` and based on the fact that each operation must retain bytes >> from the last two —or three, for NSS— blocks in `padBuffer`, or retain >> whatever is available if less than that. `dataForP11Update` is the number of >> bytes to be passed to the token during the operation (`C_EncryptUpdate` or >> `C_DecryptUpdate` calls), irrespective of the stage in which they are passed >> and of the dat... > > src/jdk.crypto.cryptoki/share/classes/sun/security/pkcs11/P11Cipher.java line > 697: > >> 695: k += token.p11.C_EncryptUpdate(session.id(), >> 696: 0, in, inOfs, inLen, >> 697: 0, out, (outOfs + k), (outLen - k)); > > Is this an actual change? @theRealAph: no, it looked to me easier to read if we have input-related parameters separated from output ones. So I did this reordering in a separate commit (0a777e94229723376e1264e87cbf0ba805dc736f), to distinguish it from the core change (in c4419bfecc0b795f6a1bdcbe5c5a6aa0142350ea). We can revert 0a777e94229723376e1264e87cbf0ba805dc736f if it confuses more than it improves. ------------- PR Review Comment: https://git.openjdk.org/jdk/pull/18898#discussion_r1577793343
