This partially fixes a bug where, on x86_64, BN_mod_exp_mont_consttime
would sometimes return m, the modulus, when it should have returned
zero. Thanks to Guido Vranken for reporting it. It is only a partial fix
because the same bug also exists in the "rsaz" codepath. That will be
fixed in the subsequent CL. (See the commented out test.)

The bug only affects zero outputs (with non-zero inputs), so we believe
it has no security impact on our cryptographic functions. BoringSSL
calls BN_mod_exp_mont_consttime in the following cases:

- RSA private key operations
- Primality testing, raising the witness to the odd part of p-1
- DSA keygen and key import, pub = g^priv (mod p)
- DSA signing, r = g^k (mod p)
- DH keygen, pub = g^priv (mod p)
- Diffie-Hellman, secret = peer^priv (mod p)

It is not possible in the RSA private key operation, provided p and q
are primes. If using CRT, we are working modulo a prime, so zero output
with non-zero input is impossible. If not using CRT, we work mod n.
While there are nilpotent values mod n, none of them hit zero by
exponentiating. (Both p and q would need to divide the input, which
means n divides the input.)

In primality testing, this can only be hit when the input was composite.
But as the rest of the loop cannot then hit 1, we'll correctly report it
as composite anyway.

DSA and DH work modulo a prime, where this case cannot happen.

Analysis:

This bug is the result of sloppiness with the looser bounds from "almost
Montgomery multiplication", described in
https://eprint.iacr.org/2011/239. Prior to upstream's
ec9cc70f72454b8d4a84247c86159613cee83b81, I believe x86_64-mont5.pl
implemented standard Montgomery reduction (the left half of figure 3 in
the paper).

Though it did not document this, ec9cc70f7245 changed it to implement
the "almost" variant (the right half of the figure.) The difference is
that, rather than subtracting if T >= m, it subtracts if T >= R. In
code, it is the difference between something like our bn_reduce_once,
vs. subtracting based only on T's carry bit. (Interestingly, the
.Lmul_enter branch of bn_mul_mont_gather5 seems to still implement
normal reduction, but the .Lmul4x_enter branch is an almost reduction.)

That means none of the intermediate values here are bounded by m. They
are only bounded by R. Accordingly, Figure 2 in the paper ends with
step 10: REDUCE h modulo m. BN_mod_exp_mont_consttime is missing this
step. The bn_from_montgomery call only implements step 9, AMM(h, 1).
(x86_64-mont5.pl's bn_from_montgomery only implements an almost
reduction.)

The impact depends on how unreduced AMM(h, 1) can be. Remark 1 of the
paper discusses this, but is ambiguous about the scope of its 2^(n-1) <
m < 2^n precondition. The m+1 bound appears to be unconditional:

Montgomery reduction ultimately adds some 0 <= Y < m*R to T, to get a
multiple of R, and then divides by R. The output, pre-subtraction, is
thus less than m + T/R. MM works because T < mR => T' < m + mR/R = 2m.
A single subtraction of m if T' >= m gives T'' < m. AMM works because
T < R^2 => T' < m + R^2/R = m + R. A single subtraction of m if T' >= R
gives T'' < R. See also Lemma 1, Section 3 and Section 4 of the paper,
though their formulation is more complicated to capture the word-by-word
algorithm. It's ultimately the same adjustment to T.

But in AMM(h, 1), T = h*1 = h < R, so AMM(h, 1) < m + R/R = m + 1. That
is, AMM(h, 1) <= m. So the only case when AMM(h, 1) isn't fully reduced
is if it outputs m. Thus, our limited impact. Indeed, Remark 1 mentions
step 10 isn't necessary because m is a prime and the inputs are
non-zero. But that doesn't apply here because BN_mod_exp_mont_consttime
may be called elsewhere.

Fix:

To fix this, we could add the missing step 10, but a full division would
not be constant-time. The analysis above says it could be a single
subtraction, bn_reduce_once, but then we could integrate it into
the subtraction already in plain Montgomery reduction, implemented by
uppercase BN_from_montgomery. h*1 = h < R <= m*R, so we are within
bounds.

Thus, we delete lowercase bn_from_montgomery altogether, and have the
mont5 path use the same BN_from_montgomery ending as the non-mont5 path.
This only impacts the final step of the whole exponentiation and has no
measurable perf impact.

In doing so, add comments describing these looser bounds.  This includes
one subtlety that BN_mod_exp_mont_consttime actually mixes bn_mul_mont
(MM) with bn_mul_mont_gather5/bn_power5 (AMM). But this is fine because
MM is AMM-compatible; when passed AMM's looser inputs, it will still
produce a correct looser output.

Ideally we'd drop the "almost" reduction and stick to the more
straightforward bounds. As this only impacts the final subtraction in
each reduction, I would be surprised if it actually had a real
performance impact. But this would involve deeper change to
x86_64-mont5.pl, so I haven't tried this yet.

I believe this is basically the same bug as
golang/go#13907 from Go.

Change-Id: I06f879777bb2ef181e9da7632ec858582e2afa38
Reviewed-on: https://boringssl-review.googlesource.com/c/boringssl/+/52825
Commit-Queue: David Benjamin <davidben@google.com>
Reviewed-by: Adam Langley <agl@google.com>

…normal Montgomery reduction.

The impact analysis is the same as BoringSSL's. There is further
protection provided by the fact that in the RSA computation we always
double-check the result against faults ("Verify the result" in
`private_exponentiate`). In the other use, the result is immediately
followed by a multiplication that would fully reduce the result.

All the test vectors from BoringSSL were imported except the ones with
moduli with bit sizes that are not a multiple of 512, since that is an
enforced preresquisite in *ring*. Since *ring* doesn't use RSAZ the
test vector that BoringSSL had temporarily commented out is also
enabled.

```
git diff \
      13c9d5c:crypto/fipsmodule/bn/bn_tests.txt \
      src/arithmetic/bigint_elem_exp_consttime_tests.txt
```