apupier opened a new pull request, #30489:
URL: https://github.com/apache/camel/pull/30489

   Fixes [CAMEL-27293](https://issues.apache.org/jira/browse/CAMEL-27293): 
`PQCSABERGenerateEncapsulationAESNoAutowiredTest` is flaky since 7 November.
   
   ## Root cause
   
   Every SABER key of a parameter set points to one static Bouncy Castle 
`SABERParameters` constant, which owns a single `SABEREngine` whose SHA-3 state 
is mutated by each encapsulation and decapsulation. The KEM operations only 
synchronized on the `KeyGenerator` they use, and `PQCProducer` creates a fresh 
`KeyGenerator` per exchange. Concurrent SABER exchanges therefore ran on 
different locks and corrupted each other, with either `IllegalStateException: 
attempt to absorb while squeezing`, or an extracted secret that differs from 
the encapsulated one (the `assertTrue` failure seen in CI).
   
   It surfaced on 7 November because CAMEL-24290 (#29354) turned on parallel 
test classes for camel-pqc, so `PQCSABERGenerateEncapsulationAESTest` and 
`PQCSABERGenerateEncapsulationAESNoAutowiredTest` started to overlap. This is a 
production race, not a test problem: two routes encapsulating with SABER at the 
same time can hit it.
   
   ## Fix
   
   `PQCKemLocks.lockFor(generator, key)` returns one JVM-wide monitor for SABER 
keys and the `KeyGenerator` itself for every other algorithm. `PQCProducer`, 
`PQCDataFormat` and `HybridKEM` use it around their `init()` / `generateKey()` 
sequences.
   
   I checked the other algorithms by running concurrent encapsulation and 
extraction on separate `KeyGenerator` instances (8 threads x 30 round trips 
each): only SABER fails (147 of 240); MLKEM, BIKE, HQC, CMCE, FRODO, NTRU, 
NTRULPRime, SNTRUPrime and KYBER have no failures. The non-stateful signature 
algorithms were checked the same way and are fine. HQC and SmaugT also hold an 
engine in their static parameter constants, but in HQC it only contains final 
sizes and a helper without mutable state.
   
   ## Tests
   
   New deterministic tests, parameterized over every 
`PQCKeyEncapsulationAlgorithms` value, for the producer, the data format and 
the hybrid KEM. They use the existing `InterleavingProbe` / 
`ProbedKeyGenerator`: one exchange is held between `init()` and `generateKey()` 
of its own generator while a second exchange starts on another generator, so 
the result does not depend on thread scheduling. Without the production change 
exactly the SABER case fails in all three; with it everything passes. The full 
camel-pqc suite passes (397 tests).
   
   `PQCKemLocks.sharesEngineAcrossKeyGenerators(algorithm)` is the single 
definition of "this algorithm needs the shared lock", used by the lock and by 
the tests. If a future Bouncy Castle release shares an engine for another 
algorithm, extending it makes the matching test case enforce the lock.
   
   ---
   _AI-generated: Bob (IBM Bob 2.2.1) on behalf of @apupier._
   


-- 
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.

To unsubscribe, e-mail: [email protected]

For queries about this service, please contact Infrastructure at:
[email protected]

Reply via email to