Ah, ok so my problem had been typographical rather than philosophical.

It's cute to observe that Atalla having no RSA-specific form of mod_exp
causes a DSA server to achieve about 6 times as many signatures per
second than an RSA server. :-)

diff --git a/crypto/engine/hw_atalla.c b/crypto/engine/hw_atalla.c
index e0806cb146c2af1772393531f2551185ee5af155..d83f4acafd452836f21c095cd3397c1e5fbb9be3 100644 (file)
--- a/crypto/engine/hw_atalla.c
+++ b/crypto/engine/hw_atalla.c
@@ -384,15 +384,16 @@ err:
        }
 
 /* This code was liberated and adapted from the commented-out code in
- * dsa_ossl.c but doesn't work (for me at least). So I need to work out
- * some mod_exp-based variant of BN_mod_exp2_mont ... for now we resort
- * to software which should only hurt DSA clients, or DSA servers doing
- * client-authentication. */
+ * dsa_ossl.c. Because of the unoptimised form of the Atalla acceleration
+ * (it doesn't have a CRT form for RSA), this function means that an
+ * Atalla system running with a DSA server certificate can handshake
+ * around 5 or 6 times faster/more than an equivalent system running with
+ * RSA. Just check out the "signs" statistics from the RSA and DSA parts
+ * of "openssl speed -engine atalla dsa1024 rsa1024". */
 static int atalla_dsa_mod_exp(DSA *dsa, BIGNUM *rr, BIGNUM *a1,
                BIGNUM *p1, BIGNUM *a2, BIGNUM *p2, BIGNUM *m,
                BN_CTX *ctx, BN_MONT_CTX *in_mont)
        {
-#if 0
        BIGNUM t;
        int to_return = 0;
  
@@ -401,15 +402,12 @@ static int atalla_dsa_mod_exp(DSA *dsa, BIGNUM *rr, BIGNUM *a1,
        if (!atalla_mod_exp(rr,a1,p1,m,ctx)) goto end;
        /* let t = a2 ^ p2 mod m */
        if (!atalla_mod_exp(&t,a2,p2,m,ctx)) goto end;
-       /* let p1 = rr * t mod m */
-       if (!BN_mod_mul(p1,rr,&t,m,ctx)) goto end;
+       /* let rr = rr * t mod m */
+       if (!BN_mod_mul(rr,rr,&t,m,ctx)) goto end;
        to_return = 1;
 end:
        BN_free(&t);
        return to_return;
-#else
-       return BN_mod_exp2_mont(rr, a1, p1, a2, p2, m, ctx, in_mont);
-#endif
        }