CipherSpi includes 3 overloads of engineDoFinal. OpenSSLCipherRSA overloads 2 of them, but not the 3rd one engineDoFinal(ByteBuffer, ByteBuffer). This method is called by Cipher when invoking doFinal.

Unfortunately, the CipherSpi implementation of engineDoFinal calls Cipher.bufferCrypt, which is slow. It makes array copies of the inputs if they are not backed by arrays. When using direct ByteBuffers, multiple unnecessary allocations and copies are made.

CC: @nmittler