visaljs 0.0.2 → 0.0.3

package/runtime/third_party/boringssl/crypto/fipsmodule/rsa/rsa_impl.cc.inc

@@ -0,0 +1,1000 @@
+// Copyright 1995-2016 The OpenSSL Project Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     https://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include <openssl/rsa.h>
+
+#include <assert.h>
+#include <limits.h>
+#include <string.h>
+
+#include <utility>
+
+#include <openssl/bn.h>
+#include <openssl/err.h>
+#include <openssl/mem.h>
+
+#include "../../bcm_support.h"
+#include "../../internal.h"
+#include "../../mem_internal.h"
+#include "../bn/internal.h"
+#include "../delocate.h"
+#include "../service_indicator/internal.h"
+#include "internal.h"
+
+
+using namespace bssl;
+
+static_assert(OPENSSL_RSA_MAX_MODULUS_BITS <=
+                  BN_MONTGOMERY_MAX_WORDS * BN_BITS2,
+              "Max RSA size too big for Montgomery arithmetic");
+
+int bssl::rsa_check_public_key(const RSA *rsa) {
+  auto *impl = FromOpaque(rsa);
+
+  if (impl->n == nullptr) {
+    OPENSSL_PUT_ERROR(RSA, RSA_R_VALUE_MISSING);
+    return 0;
+  }
+
+  unsigned n_bits = BN_num_bits(impl->n.get());
+  if (n_bits > OPENSSL_RSA_MAX_MODULUS_BITS) {
+    OPENSSL_PUT_ERROR(RSA, RSA_R_MODULUS_TOO_LARGE);
+    return 0;
+  }
+
+  if (n_bits < OPENSSL_RSA_MIN_MODULUS_BITS) {
+    OPENSSL_PUT_ERROR(RSA, RSA_R_KEY_SIZE_TOO_SMALL);
+    return 0;
+  }
+
+  // RSA moduli must be positive and odd. In addition to being necessary for RSA
+  // in general, we cannot setup Montgomery reduction with even moduli.
+  if (!BN_is_odd(impl->n.get()) || BN_is_negative(impl->n.get())) {
+    OPENSSL_PUT_ERROR(RSA, RSA_R_BAD_RSA_PARAMETERS);
+    return 0;
+  }
+
+  static const unsigned kMaxExponentBits = 33;
+  if (impl->e != nullptr) {
+    // Reject e = 1, negative e, and even e. e must be odd to be relatively
+    // prime with phi(n).
+    unsigned e_bits = BN_num_bits(impl->e.get());
+    if (e_bits < 2 || BN_is_negative(impl->e.get()) ||
+        !BN_is_odd(impl->e.get())) {
+      OPENSSL_PUT_ERROR(RSA, RSA_R_BAD_E_VALUE);
+      return 0;
+    }
+    if (impl->flags & RSA_FLAG_LARGE_PUBLIC_EXPONENT) {
+      // The caller has requested disabling DoS protections. Still, e must be
+      // less than n.
+      if (BN_ucmp(impl->n.get(), impl->e.get()) <= 0) {
+        OPENSSL_PUT_ERROR(RSA, RSA_R_BAD_E_VALUE);
+        return 0;
+      }
+    } else {
+      // Mitigate DoS attacks by limiting the exponent size. 33 bits was chosen
+      // as the limit based on the recommendations in [1] and [2]. Windows
+      // CryptoAPI doesn't support values larger than 32 bits [3], so it is
+      // unlikely that exponents larger than 32 bits are being used for anything
+      // Windows commonly does.
+      //
+      // [1] https://www.imperialviolet.org/2012/03/16/rsae.html
+      // [2] https://www.imperialviolet.org/2012/03/17/rsados.html
+      // [3] https://msdn.microsoft.com/en-us/library/aa387685(VS.85).aspx
+      if (e_bits > kMaxExponentBits) {
+        OPENSSL_PUT_ERROR(RSA, RSA_R_BAD_E_VALUE);
+        return 0;
+      }
+
+      // The upper bound on `e_bits` and lower bound on `n_bits` imply e is
+      // bounded by n.
+      assert(BN_ucmp(impl->n.get(), impl->e.get()) > 0);
+    }
+  } else if (!(impl->flags & RSA_FLAG_NO_PUBLIC_EXPONENT)) {
+    OPENSSL_PUT_ERROR(RSA, RSA_R_VALUE_MISSING);
+    return 0;
+  }
+
+  return 1;
+}
+
+static int ensure_fixed_copy(UniquePtr<BIGNUM> *out, const BIGNUM *in,
+                             int width) {
+  if (*out != nullptr) {
+    return 1;
+  }
+  UniquePtr<BIGNUM> copy(BN_dup(in));
+  if (copy == nullptr || !bn_resize_words(copy.get(), width)) {
+    return 0;
+  }
+  bn_secret(copy.get());
+  *out = std::move(copy);
+  return 1;
+}
+
+// freeze_private_key finishes initializing `rsa`'s private key components.
+// After this function has returned, `rsa` may not be changed. This is needed
+// because `RSA` is a public struct and, additionally, OpenSSL 1.1.0 opaquified
+// it wrong (see https://github.com/openssl/openssl/issues/5158).
+static int freeze_private_key(RSAImpl *rsa, BN_CTX *ctx) {
+  rsa->lock.LockRead();
+  int frozen = rsa->private_key_frozen;
+  rsa->lock.UnlockRead();
+  if (frozen) {
+    return 1;
+  }
+
+  const BIGNUM *n_fixed;
+  MutexWriteLock lock(&rsa->lock);
+  if (rsa->private_key_frozen) {
+    return 1;
+  }
+
+  // Check the public components are within DoS bounds.
+  if (!rsa_check_public_key(rsa)) {
+    return 0;
+  }
+
+  // Pre-compute various intermediate values, as well as copies of private
+  // exponents with correct widths. Note that other threads may concurrently
+  // read from `rsa->n`, `rsa->e`, etc., so any fixes must be in separate
+  // copies. We use `mont_n->N`, `mont_p->N`, and `mont_q->N` as copies of `n`,
+  // `p`, and `q` with the correct minimal widths.
+
+  if (rsa->mont_n == nullptr) {
+    rsa->mont_n.reset(BN_MONT_CTX_new_for_modulus(rsa->n.get(), ctx));
+    if (rsa->mont_n == nullptr) {
+      return 0;
+    }
+  }
+  n_fixed = &rsa->mont_n->N;
+
+  // The only public upper-bound of `rsa->d` is the bit length of `rsa->n`. The
+  // ASN.1 serialization of RSA private keys unfortunately leaks the byte length
+  // of `rsa->d`, but normalize it so we only leak it once, rather than per
+  // operation.
+  if (rsa->d != nullptr &&
+      !ensure_fixed_copy(&rsa->d_fixed, rsa->d.get(), n_fixed->width)) {
+    return 0;
+  }
+
+  if (rsa->e != nullptr && rsa->p != nullptr && rsa->q != nullptr) {
+    // TODO: p and q are also CONSTTIME_SECRET but not yet marked as such
+    // because the Montgomery code does things like test whether or not values
+    // are zero. So the secret marking probably needs to happen inside that
+    // code.
+
+    if (rsa->mont_p == nullptr) {
+      rsa->mont_p.reset(BN_MONT_CTX_new_consttime(rsa->p.get(), ctx));
+      if (rsa->mont_p == nullptr) {
+        return 0;
+      }
+    }
+
+    if (rsa->mont_q == nullptr) {
+      rsa->mont_q.reset(BN_MONT_CTX_new_consttime(rsa->q.get(), ctx));
+      if (rsa->mont_q == nullptr) {
+        return 0;
+      }
+    }
+
+    if (rsa->dmp1 != nullptr && rsa->dmq1 != nullptr && rsa->iqmp != nullptr) {
+      // CRT components are only publicly bounded by their corresponding
+      // moduli's bit lengths.
+      const BIGNUM *p_fixed = &rsa->mont_p->N;
+      const BIGNUM *q_fixed = &rsa->mont_q->N;
+      if (!ensure_fixed_copy(&rsa->dmp1_fixed, rsa->dmp1.get(),
+                             p_fixed->width) ||
+          !ensure_fixed_copy(&rsa->dmq1_fixed, rsa->dmq1.get(),
+                             q_fixed->width)) {
+        return 0;
+      }
+
+      // Compute `iqmp_mont`, which is `iqmp` in Montgomery form and with the
+      // correct bit width.
+      if (rsa->iqmp_mont == nullptr) {
+        UniquePtr<BIGNUM> iqmp_mont(BN_new());
+        if (iqmp_mont == nullptr ||
+            !BN_to_montgomery(iqmp_mont.get(), rsa->iqmp.get(),
+                              rsa->mont_p.get(), ctx)) {
+          return 0;
+        }
+        rsa->iqmp_mont = std::move(iqmp_mont);
+        bn_secret(rsa->iqmp_mont.get());
+      }
+    }
+  }
+
+  rsa->private_key_frozen = 1;
+  return 1;
+}
+
+void bssl::rsa_invalidate_key(RSA *rsa) {
+  auto *impl = FromOpaque(rsa);
+  impl->private_key_frozen = 0;
+  impl->mont_n = nullptr;
+  impl->mont_p = nullptr;
+  impl->mont_q = nullptr;
+  impl->d_fixed = nullptr;
+  impl->dmp1_fixed = nullptr;
+  impl->dmq1_fixed = nullptr;
+  impl->iqmp_mont = nullptr;
+}
+
+int bssl::rsa_default_sign_raw(RSA *rsa, size_t *out_len, uint8_t *out,
+                               size_t max_out, const uint8_t *in, size_t in_len,
+                               int padding) {
+  const unsigned rsa_size = RSA_size(rsa);
+  uint8_t *buf = nullptr;
+  int i, ret = 0;
+
+  if (max_out < rsa_size) {
+    OPENSSL_PUT_ERROR(RSA, RSA_R_OUTPUT_BUFFER_TOO_SMALL);
+    return 0;
+  }
+
+  buf = reinterpret_cast<uint8_t *>(OPENSSL_malloc(rsa_size));
+  if (buf == nullptr) {
+    goto err;
+  }
+
+  switch (padding) {
+    case RSA_PKCS1_PADDING:
+      i = RSA_padding_add_PKCS1_type_1(buf, rsa_size, in, in_len);
+      break;
+    case RSA_NO_PADDING:
+      i = RSA_padding_add_none(buf, rsa_size, in, in_len);
+      break;
+    default:
+      OPENSSL_PUT_ERROR(RSA, RSA_R_UNKNOWN_PADDING_TYPE);
+      goto err;
+  }
+
+  if (i <= 0) {
+    goto err;
+  }
+
+  if (!rsa_private_transform_no_self_test(rsa, out, buf, rsa_size)) {
+    goto err;
+  }
+
+  CONSTTIME_DECLASSIFY(out, rsa_size);
+  *out_len = rsa_size;
+  ret = 1;
+
+err:
+  OPENSSL_free(buf);
+
+  return ret;
+}
+
+
+static int rsa_mod_exp_crt(BIGNUM *r0, const BIGNUM *I, RSAImpl *rsa,
+                           BN_CTX *ctx);
+
+int bssl::rsa_verify_raw_no_self_test(RSA *rsa, size_t *out_len, uint8_t *out,
+                                      size_t max_out, const uint8_t *in,
+                                      size_t in_len, int padding) {
+  auto *impl = FromOpaque(rsa);
+
+  if (impl->n == nullptr || impl->e == nullptr) {
+    OPENSSL_PUT_ERROR(RSA, RSA_R_VALUE_MISSING);
+    return 0;
+  }
+
+  if (!rsa_check_public_key(rsa)) {
+    return 0;
+  }
+
+  const unsigned rsa_size = RSA_size(rsa);
+  if (max_out < rsa_size) {
+    OPENSSL_PUT_ERROR(RSA, RSA_R_OUTPUT_BUFFER_TOO_SMALL);
+    return 0;
+  }
+
+  if (in_len != rsa_size) {
+    OPENSSL_PUT_ERROR(RSA, RSA_R_DATA_LEN_NOT_EQUAL_TO_MOD_LEN);
+    return 0;
+  }
+
+  UniquePtr<BN_CTX> ctx(BN_CTX_new());
+  if (ctx == nullptr) {
+    return 0;
+  }
+
+  int ret = 0;
+  uint8_t *buf = nullptr;
+  BN_CTXScope scope(ctx.get());
+  BIGNUM *f = BN_CTX_get(ctx.get());
+  BIGNUM *result = BN_CTX_get(ctx.get());
+  if (f == nullptr || result == nullptr) {
+    goto err;
+  }
+
+  if (padding == RSA_NO_PADDING) {
+    buf = out;
+  } else {
+    // Allocate a temporary buffer to hold the padded plaintext.
+    buf = reinterpret_cast<uint8_t *>(OPENSSL_malloc(rsa_size));
+    if (buf == nullptr) {
+      goto err;
+    }
+  }
+
+  if (BN_bin2bn(in, in_len, f) == nullptr) {
+    goto err;
+  }
+
+  if (BN_ucmp(f, impl->n.get()) >= 0) {
+    OPENSSL_PUT_ERROR(RSA, RSA_R_DATA_TOO_LARGE_FOR_MODULUS);
+    goto err;
+  }
+
+  if (!BN_MONT_CTX_set_locked(&impl->mont_n, &impl->lock, impl->n.get(),
+                              ctx.get()) ||
+      !BN_mod_exp_mont(result, f, impl->e.get(), &impl->mont_n->N, ctx.get(),
+                       impl->mont_n.get())) {
+    goto err;
+  }
+
+  if (!BN_bn2bin_padded(buf, rsa_size, result)) {
+    OPENSSL_PUT_ERROR(RSA, ERR_R_INTERNAL_ERROR);
+    goto err;
+  }
+
+  switch (padding) {
+    case RSA_PKCS1_PADDING:
+      ret =
+          RSA_padding_check_PKCS1_type_1(out, out_len, rsa_size, buf, rsa_size);
+      break;
+    case RSA_NO_PADDING:
+      ret = 1;
+      *out_len = rsa_size;
+      break;
+    default:
+      OPENSSL_PUT_ERROR(RSA, RSA_R_UNKNOWN_PADDING_TYPE);
+      goto err;
+  }
+
+  if (!ret) {
+    OPENSSL_PUT_ERROR(RSA, RSA_R_PADDING_CHECK_FAILED);
+    goto err;
+  }
+
+err:
+  if (buf != out) {
+    OPENSSL_free(buf);
+  }
+  return ret;
+}
+
+int RSA_verify_raw(RSA *rsa, size_t *out_len, uint8_t *out, size_t max_out,
+                   const uint8_t *in, size_t in_len, int padding) {
+  boringssl_ensure_rsa_verify_self_test();
+  return rsa_verify_raw_no_self_test(rsa, out_len, out, max_out, in, in_len,
+                                     padding);
+}
+
+int bssl::rsa_default_private_transform(RSA *rsa, uint8_t *out,
+                                        const uint8_t *in, size_t len) {
+  auto *impl = FromOpaque(rsa);
+
+  if (impl->n == nullptr || impl->d == nullptr) {
+    OPENSSL_PUT_ERROR(RSA, RSA_R_VALUE_MISSING);
+    return 0;
+  }
+
+  UniquePtr<BN_CTX> ctx(BN_CTX_new());
+  if (ctx == nullptr) {
+    return 0;
+  }
+  BN_CTXScope scope(ctx.get());
+  BIGNUM *f = BN_CTX_get(ctx.get());
+  BIGNUM *result = BN_CTX_get(ctx.get());
+  if (f == nullptr || result == nullptr) {
+    return 0;
+  }
+
+  // The caller should have ensured this.
+  assert(len == BN_num_bytes(impl->n.get()));
+  if (BN_bin2bn(in, len, f) == nullptr) {
+    return 0;
+  }
+
+  // The input to the RSA private transform may be secret, but padding is
+  // expected to construct a value within range, so we can leak this comparison.
+  if (constant_time_declassify_int(BN_ucmp(f, impl->n.get()) >= 0)) {
+    // Usually the padding functions would catch this.
+    OPENSSL_PUT_ERROR(RSA, RSA_R_DATA_TOO_LARGE_FOR_MODULUS);
+    return 0;
+  }
+
+  if (!freeze_private_key(impl, ctx.get())) {
+    OPENSSL_PUT_ERROR(RSA, ERR_R_INTERNAL_ERROR);
+    return 0;
+  }
+
+  if (impl->e == nullptr && (impl->flags & RSA_FLAG_NO_PUBLIC_EXPONENT) == 0) {
+    // Unless the private key was specifically created with an API like
+    // `RSA_new_private_key_no_e`, don't allow RSA keys to be missing the public
+    // exponent, which disables some fault attack mitigations. (It should not be
+    // possible to construct such an `RSA` object in the public API.)
+    OPENSSL_PUT_ERROR(RSA, RSA_R_NO_PUBLIC_EXPONENT);
+    return 0;
+  }
+
+  if (impl->p != nullptr && impl->q != nullptr && impl->e != nullptr &&
+      impl->dmp1 != nullptr && impl->dmq1 != nullptr && impl->iqmp != nullptr &&
+      // Require that we can reduce `f` by `impl->p` and `impl->q` in constant
+      // time, which requires primes be the same size, rounded to the Montgomery
+      // coefficient. (See `mod_montgomery`.) This is not required by RFC 8017,
+      // but it is true for keys generated by us and all common implementations.
+      bn_less_than_montgomery_R(impl->q.get(), impl->mont_p.get()) &&
+      bn_less_than_montgomery_R(impl->p.get(), impl->mont_q.get())) {
+    if (!rsa_mod_exp_crt(result, f, impl, ctx.get())) {
+      return 0;
+    }
+  } else if (!BN_mod_exp_mont_consttime(result, f, impl->d_fixed.get(),
+                                        impl->n.get(), ctx.get(),
+                                        impl->mont_n.get())) {
+    return 0;
+  }
+
+  // Verify the result to protect against fault attacks as described in the
+  // 1997 paper "On the Importance of Checking Cryptographic Protocols for
+  // Faults" by Dan Boneh, Richard A. DeMillo, and Richard J. Lipton. Some
+  // implementations do this only when the CRT is used, but we do it in all
+  // cases. Section 6 of the aforementioned paper describes an attack that
+  // works when the CRT isn't used. That attack is much less likely to succeed
+  // than the CRT attack, but there have likely been improvements since 1997.
+  //
+  // This check is cheap assuming `e` is small, which we require in
+  // `rsa_check_public_key`.
+  if (impl->e != nullptr) {
+    BIGNUM *vrfy = BN_CTX_get(ctx.get());
+    if (vrfy == nullptr ||
+        !BN_mod_exp_mont(vrfy, result, impl->e.get(), impl->n.get(), ctx.get(),
+                         impl->mont_n.get()) ||
+        !constant_time_declassify_int(BN_equal_consttime(vrfy, f))) {
+      OPENSSL_PUT_ERROR(RSA, ERR_R_INTERNAL_ERROR);
+      return 0;
+    }
+  }
+
+  // The computation should have left `result` as a maximally-wide number, so
+  // that it and serializing does not leak information about the magnitude of
+  // the result.
+  //
+  // See Falko Strenzke, "Manger's Attack revisited", ICICS 2010.
+  assert(result->width == impl->mont_n->N.width);
+  bn_assert_fits_in_bytes(result, len);
+  if (!BN_bn2bin_padded(out, len, result)) {
+    OPENSSL_PUT_ERROR(RSA, ERR_R_INTERNAL_ERROR);
+    return 0;
+  }
+
+  return 1;
+}
+
+// mod_montgomery sets `r` to `I` mod `p`. `I` must already be fully reduced
+// modulo `p` times `q`. It returns one on success and zero on error.
+static int mod_montgomery(BIGNUM *r, const BIGNUM *I, const BIGNUM *p,
+                          const BN_MONT_CTX *mont_p, const BIGNUM *q,
+                          BN_CTX *ctx) {
+  // Reducing in constant-time with Montgomery reduction requires I <= p * R. We
+  // have I < p * q, so this follows if q < R. The caller should have checked
+  // this already.
+  if (!bn_less_than_montgomery_R(q, mont_p)) {
+    OPENSSL_PUT_ERROR(RSA, ERR_R_INTERNAL_ERROR);
+    return 0;
+  }
+
+  if (  // Reduce mod p with Montgomery reduction. This computes I * R^-1 mod p.
+      !BN_from_montgomery(r, I, mont_p, ctx) ||
+      // Multiply by R² and do another Montgomery reduction to compute
+      // I * R^-1 * R² * R^-1 = I mod p.
+      !BN_to_montgomery(r, r, mont_p, ctx)) {
+    return 0;
+  }
+
+  // By precomputing R³ mod p (normally `BN_MONT_CTX` only uses R² mod p) and
+  // adjusting the API for `BN_mod_exp_mont_consttime`, we could instead compute
+  // I * R mod p here and save a reduction per prime. But this would require
+  // changing the RSAZ code and may not be worth it. Note that the RSAZ code
+  // uses a different radix, so it uses R' = 2^1044. There we'd actually want
+  // R² * R', and would further benefit from a precomputed R'². It currently
+  // converts `mont_p->RR` to R'².
+  return 1;
+}
+
+static int rsa_mod_exp_crt(BIGNUM *r0, const BIGNUM *I, RSAImpl *rsa,
+                           BN_CTX *ctx) {
+  assert(ctx != nullptr);
+
+  assert(rsa->n != nullptr);
+  assert(rsa->e != nullptr);
+  assert(rsa->d != nullptr);
+  assert(rsa->p != nullptr);
+  assert(rsa->q != nullptr);
+  assert(rsa->dmp1 != nullptr);
+  assert(rsa->dmq1 != nullptr);
+  assert(rsa->iqmp != nullptr);
+
+  BN_CTXScope scope(ctx);
+  BIGNUM *r1 = BN_CTX_get(ctx);
+  BIGNUM *m1 = BN_CTX_get(ctx);
+  if (r1 == nullptr || m1 == nullptr) {
+    return 0;
+  }
+
+  // Use the minimal-width versions of `n`, `p`, and `q`. Either works, but if
+  // someone gives us non-minimal values, these will be slightly more efficient
+  // on the non-Montgomery operations.
+  BIGNUM *n = &rsa->mont_n->N;
+  BIGNUM *p = &rsa->mont_p->N;
+  BIGNUM *q = &rsa->mont_q->N;
+
+  // This is a pre-condition for `mod_montgomery`. It was already checked by the
+  // caller.
+  declassify_assert(BN_ucmp(I, n) < 0);
+
+  if (  // `m1` is the result modulo `q`.
+      !mod_montgomery(r1, I, q, rsa->mont_q.get(), p, ctx) ||
+      !BN_mod_exp_mont_consttime(m1, r1, rsa->dmq1_fixed.get(), q, ctx,
+                                 rsa->mont_q.get()) ||
+      // `r0` is the result modulo `p`.
+      !mod_montgomery(r1, I, p, rsa->mont_p.get(), q, ctx) ||
+      !BN_mod_exp_mont_consttime(r0, r1, rsa->dmp1_fixed.get(), p, ctx,
+                                 rsa->mont_p.get()) ||
+      // Compute r0 = r0 - m1 mod p. `m1` is reduced mod `q`, not `p`, so we
+      // just run `mod_montgomery` again for srsaicity. This could be more
+      // efficient with more cases: if `p > q`, `m1` is already reduced. If
+      // `p < q` but they have the same bit width, `bn_reduce_once` suffices.
+      // However, compared to over 2048 Montgomery multiplications above, this
+      // difference is not measurable.
+      !mod_montgomery(r1, m1, p, rsa->mont_p.get(), q, ctx) ||
+      !bn_mod_sub_consttime(r0, r0, r1, p, ctx) ||
+      // r0 = r0 * iqmp mod p. We use Montgomery multiplication to compute this
+      // in constant time. `iqmp_mont` is in Montgomery form and r0 is not, so
+      // the result is taken out of Montgomery form.
+      !BN_mod_mul_montgomery(r0, r0, rsa->iqmp_mont.get(), rsa->mont_p.get(),
+                             ctx) ||
+      // r0 = r0 * q + m1 gives the final result. Reducing modulo q gives m1, so
+      // it is correct mod p. Reducing modulo p gives (r0-m1)*iqmp*q + m1 = r0,
+      // so it is correct mod q. Finally, the result is bounded by [m1, n + m1),
+      // and the result is at least `m1`, so this must be the unique answer in
+      // [0, n).
+      !bn_mul_consttime(r0, r0, q, ctx) ||  //
+      !bn_uadd_consttime(r0, r0, m1)) {
+    return 0;
+  }
+
+  // The result should be bounded by `n`, but fixed-width operations may
+  // bound the width slightly higher, so fix it. This trips constant-time checks
+  // because a naive data flow analysis does not realize the excess words are
+  // publicly zero.
+  declassify_assert(BN_cmp(r0, n) < 0);
+  bn_assert_fits_in_bytes(r0, BN_num_bytes(n));
+  if (!bn_resize_words(r0, n->width)) {
+    return 0;
+  }
+
+  return 1;
+}
+
+static int ensure_bignum(UniquePtr<BIGNUM> *out) {
+  if (*out == nullptr) {
+    out->reset(BN_new());
+  }
+  return *out != nullptr;
+}
+
+// generate_prime sets `out` to a prime with length `bits` such that `out`-1 is
+// relatively prime to `e`. If `p` is non-NULL, `out` will also not be close to
+// `p`. `pow2_bits_100` must be 2^(bits-100).
+//
+// This function fails with probability around 2^-21.
+static int generate_prime(BIGNUM *out, int bits, const BIGNUM *e,
+                          const BIGNUM *p, const BIGNUM *pow2_bits_100,
+                          BN_CTX *ctx, BN_GENCB *cb) {
+  if (bits < 128 || (bits % BN_BITS2) != 0) {
+    OPENSSL_PUT_ERROR(RSA, ERR_R_INTERNAL_ERROR);
+    return 0;
+  }
+  assert(BN_is_pow2(pow2_bits_100));
+  assert(BN_is_bit_set(pow2_bits_100, bits - 100));
+
+  // See FIPS 186-5 appendix A.1.3, steps 4 and 5. Note `bits` here is nlen/2.
+
+  // Use the limit from steps 4.7 and 5.8 for most values of `e`. When `e` is 3,
+  // the 186-5 limit is too low, so we use a higher one. Note this case is not
+  // reachable from `RSA_generate_key_fips`.
+  //
+  // `limit` determines the failure probability. We must find a prime that is
+  // not 1 mod `e`. By the prime number theorem, we'll find one with probability
+  // p = (e-1)/e * 2/(ln(2)*bits). Note the second term is doubled because we
+  // discard even numbers.
+  //
+  // The failure probability is thus (1-p)^limit. To convert that to a power of
+  // two, we take logs. -log_2((1-p)^limit) = -limit * ln(1-p) / ln(2).
+  //
+  // >>> def f(bits, e, limit):
+  // ...   p = (e-1.0)/e * 2.0/(math.log(2)*bits)
+  // ...   return -limit * math.log(1 - p) / math.log(2)
+  // ...
+  // >>> f(1024, 65537, 5*1024)
+  // 20.842750558272634
+  // >>> f(1536, 65537, 5*1536)
+  // 20.83294549602474
+  // >>> f(2048, 65537, 5*2048)
+  // 20.828047576234948
+  // >>> f(1024, 3, 8*1024)
+  // 22.222147925962307
+  // >>> f(1536, 3, 8*1536)
+  // 22.21518251065506
+  // >>> f(2048, 3, 8*2048)
+  // 22.211701985875937
+  if (bits >= INT_MAX / 32) {
+    OPENSSL_PUT_ERROR(RSA, RSA_R_MODULUS_TOO_LARGE);
+    return 0;
+  }
+  int limit = BN_is_word(e, 3) ? bits * 8 : bits * 5;
+
+  int tries = 0, rand_tries = 0;
+  BN_CTXScope scope(ctx);
+  BIGNUM *tmp = BN_CTX_get(ctx);
+  if (tmp == nullptr) {
+    return 0;
+  }
+
+  for (;;) {
+    // Generate a random number of length `bits` where the bottom bit is set and
+    // top two bits are set (steps 4.2–4.4 and 5.2–5.4):
+    //
+    // - Setting the top two bits is permitted by steps 4.2.1 and 5.2.1. Doing
+    //   so implements steps 4.4 and 5.4 by making this case impossible because
+    //   √2 < 1.5.
+    //
+    // - Setting the bottom bit implements steps 4.3 and 5.3.
+    if (!BN_rand(out, bits, BN_RAND_TOP_TWO, BN_RAND_BOTTOM_ODD) ||
+        !BN_GENCB_call(cb, BN_GENCB_GENERATED, rand_tries++)) {
+      return 0;
+    }
+
+    if (p != nullptr) {
+      // If `p` and `out` are too close, try again (step 5.5).
+      if (!bn_abs_sub_consttime(tmp, out, p, ctx)) {
+        return 0;
+      }
+      if (BN_cmp(tmp, pow2_bits_100) <= 0) {
+        continue;
+      }
+    }
+
+    // RSA key generation's bottleneck is discarding composites. If it fails
+    // trial division, do not bother computing a GCD or performing Miller-Rabin.
+    if (!bn_odd_number_is_obviously_composite(out)) {
+      // Check gcd(out-1, e) is one (steps 4.5 and 5.6). Leaking the final
+      // result of this comparison is safe because, if not relatively prime, the
+      // value will be discarded.
+      int relatively_prime;
+      if (!bn_usub_consttime(tmp, out, BN_value_one()) ||
+          !bn_is_relatively_prime(&relatively_prime, tmp, e, ctx)) {
+        return 0;
+      }
+      if (constant_time_declassify_int(relatively_prime)) {
+        // Test `out` for primality (steps 4.5.1 and 5.6.1).
+        int is_probable_prime;
+        if (!BN_primality_test(&is_probable_prime, out,
+                               BN_prime_checks_for_generation, ctx, 0, cb)) {
+          return 0;
+        }
+        if (is_probable_prime) {
+          return 1;
+        }
+      }
+    }
+
+    // If we've tried too many times to find a prime, abort (steps 4.7 and 5.8).
+    tries++;
+    if (tries >= limit) {
+      OPENSSL_PUT_ERROR(RSA, RSA_R_TOO_MANY_ITERATIONS);
+      return 0;
+    }
+    if (!BN_GENCB_call(cb, 2, tries)) {
+      return 0;
+    }
+  }
+}
+
+// rsa_generate_key_impl generates an RSA key using a generalized version of
+// FIPS 186-5 appendix A.1.3. `RSA_generate_key_fips` performs additional checks
+// for FIPS-compliant key generation.
+//
+// This function returns one on success and zero on failure. It has a failure
+// probability of about 2^-20.
+static int rsa_generate_key_impl(RSAImpl *rsa, int bits, const BIGNUM *e_value,
+                                 BN_GENCB *cb) {
+  if (bits > OPENSSL_RSA_MAX_MODULUS_BITS) {
+    OPENSSL_PUT_ERROR(RSA, RSA_R_MODULUS_TOO_LARGE);
+    return 0;
+  }
+
+  // Always generate RSA keys which are a multiple of 128 bits. Round `bits`
+  // down as needed.
+  bits &= ~127;
+
+  // Reject excessively small keys.
+  if (bits < OPENSSL_RSA_MIN_MODULUS_BITS) {
+    OPENSSL_PUT_ERROR(RSA, RSA_R_KEY_SIZE_TOO_SMALL);
+    return 0;
+  }
+
+  // Reject excessively large public exponents. Windows CryptoAPI and Go don't
+  // support values larger than 32 bits, so match their limits for generating
+  // keys. (`rsa_check_public_key` uses a slightly more conservative value, but
+  // we don't need to support generating such keys.)
+  // https://github.com/golang/go/issues/3161
+  // https://msdn.microsoft.com/en-us/library/aa387685(VS.85).aspx
+  if (BN_num_bits(e_value) > 32) {
+    OPENSSL_PUT_ERROR(RSA, RSA_R_BAD_E_VALUE);
+    return 0;
+  }
+
+  // The smallest reasonable RSA exponent is 3, and it definitely must be odd.
+  // Catching these here prevents endless loops or slow computation when trying
+  // to generate keys later, and results in a better error code.
+  if (
+      // Would fail in `bn_lcm_consttime` as it only allows positive integers.
+      BN_is_negative(e_value) ||
+      // Would fail in `generate_prime` as only one `rsa->p`-1 is coprime with
+      // an even `e_value` and that one is a little bit short. (The R in RSA
+      // doesn't stand for Rabin.)
+      !BN_is_odd(e_value) ||
+      // Would loop endlessly because it'll always compute an `rsa->d` exponent
+      // of 1, which is too small.
+      BN_is_one(e_value)) {
+    OPENSSL_PUT_ERROR(RSA, RSA_R_BAD_E_VALUE);
+    return 0;
+  }
+
+  UniquePtr<BN_CTX> ctx(BN_CTX_new());
+  if (ctx == nullptr) {
+    OPENSSL_PUT_ERROR(RSA, ERR_LIB_BN);
+    return 0;
+  }
+
+  int prime_bits = bits / 2;
+  BN_CTXScope scope(ctx.get());
+  BIGNUM *totient = BN_CTX_get(ctx.get());
+  BIGNUM *pm1 = BN_CTX_get(ctx.get());
+  BIGNUM *qm1 = BN_CTX_get(ctx.get());
+  BIGNUM *pow2_prime_bits_100 = BN_CTX_get(ctx.get());
+  BIGNUM *pow2_prime_bits = BN_CTX_get(ctx.get());
+  if (totient == nullptr || pm1 == nullptr || qm1 == nullptr ||
+      pow2_prime_bits_100 == nullptr || pow2_prime_bits == nullptr ||
+      !BN_set_bit(pow2_prime_bits_100, prime_bits - 100) ||
+      !BN_set_bit(pow2_prime_bits, prime_bits)) {
+    OPENSSL_PUT_ERROR(RSA, ERR_LIB_BN);
+    return 0;
+  }
+
+  // We need the RSA components non-null.
+  if (!ensure_bignum(&rsa->n) ||     //
+      !ensure_bignum(&rsa->d) ||     //
+      !ensure_bignum(&rsa->e) ||     //
+      !ensure_bignum(&rsa->p) ||     //
+      !ensure_bignum(&rsa->q) ||     //
+      !ensure_bignum(&rsa->dmp1) ||  //
+      !ensure_bignum(&rsa->dmq1) ||  //
+      !ensure_bignum(&rsa->iqmp)) {
+    OPENSSL_PUT_ERROR(RSA, ERR_LIB_BN);
+    return 0;
+  }
+
+  if (!BN_copy(rsa->e.get(), e_value)) {
+    OPENSSL_PUT_ERROR(RSA, ERR_LIB_BN);
+    return 0;
+  }
+
+  do {
+    // Generate p and q, each of size `prime_bits`, using the steps outlined in
+    // appendix FIPS 186-5 appendix C.3.3.
+    //
+    // Each call to `generate_prime` fails with probability p = 2^-21. The
+    // probability that either call fails is 1 - (1-p)², which is around 2^-20.
+    if (!generate_prime(rsa->p.get(), prime_bits, rsa->e.get(), nullptr,
+                        pow2_prime_bits_100, ctx.get(), cb) ||
+        !BN_GENCB_call(cb, 3, 0) ||
+        !generate_prime(rsa->q.get(), prime_bits, rsa->e.get(), rsa->p.get(),
+                        pow2_prime_bits_100, ctx.get(), cb) ||
+        !BN_GENCB_call(cb, 3, 1)) {
+      OPENSSL_PUT_ERROR(RSA, ERR_LIB_BN);
+      return 0;
+    }
+
+    if (BN_cmp(rsa->p.get(), rsa->q.get()) < 0) {
+      std::swap(rsa->p, rsa->q);
+    }
+
+    // Calculate d = e^(-1) (mod lcm(p-1, q-1)), per FIPS 186-5. This differs
+    // from typical RSA rsaementations which use (p-1)*(q-1).
+    //
+    // Note this means the size of d might reveal information about p-1 and
+    // q-1. However, we do operations with Chinese Remainder Theorem, so we only
+    // use d (mod p-1) and d (mod q-1) as exponents. Using a minimal totient
+    // does not affect those two values.
+    int no_inverse;
+    if (!bn_usub_consttime(pm1, rsa->p.get(), BN_value_one()) ||
+        !bn_usub_consttime(qm1, rsa->q.get(), BN_value_one()) ||
+        !bn_lcm_consttime(totient, pm1, qm1, ctx.get()) ||
+        !bn_mod_inverse_consttime(rsa->d.get(), &no_inverse, rsa->e.get(),
+                                  totient, ctx.get())) {
+      OPENSSL_PUT_ERROR(RSA, ERR_LIB_BN);
+      return 0;
+    }
+
+    // Retry if `rsa->d` <= 2^`prime_bits`. See appendix A.3.1's guidance on
+    // values for d. When we retry, p and q are discarded, so it is safe to leak
+    // this comparison.
+  } while (
+      constant_time_declassify_int(BN_cmp(rsa->d.get(), pow2_prime_bits) <= 0));
+
+  assert(BN_num_bits(pm1) == (unsigned)prime_bits);
+  assert(BN_num_bits(qm1) == (unsigned)prime_bits);
+  if (  // Calculate n.
+      !bn_mul_consttime(rsa->n.get(), rsa->p.get(), rsa->q.get(), ctx.get()) ||
+      // Calculate d mod (p-1).
+      !bn_div_consttime(nullptr, rsa->dmp1.get(), rsa->d.get(), pm1, prime_bits,
+                        ctx.get()) ||
+      // Calculate d mod (q-1)
+      !bn_div_consttime(nullptr, rsa->dmq1.get(), rsa->d.get(), qm1, prime_bits,
+                        ctx.get())) {
+    OPENSSL_PUT_ERROR(RSA, ERR_LIB_BN);
+    return 0;
+  }
+  bn_set_minimal_width(rsa->n.get());
+
+  // `rsa->n` is computed from the private key, but is public.
+  bn_declassify(rsa->n.get());
+
+  // Calculate q^-1 mod p.
+  rsa->mont_p.reset(BN_MONT_CTX_new_consttime(rsa->p.get(), ctx.get()));
+  if (rsa->mont_p == nullptr ||  //
+      !bn_mod_inverse_secret_prime(rsa->iqmp.get(), rsa->q.get(), rsa->p.get(),
+                                   ctx.get(), rsa->mont_p.get())) {
+    OPENSSL_PUT_ERROR(RSA, ERR_LIB_BN);
+    return 0;
+  }
+
+  // Sanity-check that `rsa->n` has the specified size. This is rsaied by
+  // `generate_prime`'s bounds.
+  if (BN_num_bits(rsa->n.get()) != (unsigned)bits) {
+    OPENSSL_PUT_ERROR(RSA, ERR_R_INTERNAL_ERROR);
+    return 0;
+  }
+
+  // The key generation process is complex and thus error-prone. It could be
+  // disastrous to generate and then use a bad key so double-check that the key
+  // makes sense. Also, while `rsa` is mutable, fill in the cached components.
+  if (!RSA_check_key(rsa) || !freeze_private_key(rsa, ctx.get())) {
+    OPENSSL_PUT_ERROR(RSA, RSA_R_INTERNAL_ERROR);
+    return 0;
+  }
+
+  return 1;
+}
+
+static int RSA_generate_key_ex_maybe_fips(RSAImpl *rsa, int bits,
+                                          const BIGNUM *e_value, BN_GENCB *cb,
+                                          int check_fips) {
+  boringssl_ensure_rsa_sign_self_test();
+  boringssl_ensure_rsa_verify_self_test();
+
+  if (rsa == nullptr) {
+    OPENSSL_PUT_ERROR(EC, ERR_R_PASSED_NULL_PARAMETER);
+    return 0;
+  }
+
+  UniquePtr<RSAImpl> tmp;
+
+  // `rsa_generate_key_impl`'s 2^-20 failure probability is too high at scale,
+  // so we run the FIPS algorithm four times, bringing it down to 2^-80. We
+  // should just adjust the retry limit, but FIPS 186-5 prescribes that value
+  // and thus results in unnecessary complexity.
+  int failures = 0;
+  do {
+    ERR_clear_error();
+    // Generate into scratch space, to avoid leaving partial work on failure.
+    tmp.reset(FromOpaque(RSA_new()));
+    if (tmp == nullptr) {
+      return 0;
+    }
+
+    if (rsa_generate_key_impl(tmp.get(), bits, e_value, cb)) {
+      break;
+    }
+
+    tmp = nullptr;
+    failures++;
+
+    // Only retry on `RSA_R_TOO_MANY_ITERATIONS`. This is so a caller-induced
+    // failure in `BN_GENCB_call` is still fatal.
+  } while (failures < 4 && ERR_equals(ERR_peek_error(), ERR_LIB_RSA,
+                                      RSA_R_TOO_MANY_ITERATIONS));
+
+  if (tmp == nullptr || (check_fips && !RSA_check_fips(tmp.get()))) {
+    return 0;
+  }
+
+  rsa_invalidate_key(rsa);
+  rsa->n = std::move(tmp->n);
+  rsa->e = std::move(tmp->e);
+  rsa->d = std::move(tmp->d);
+  rsa->p = std::move(tmp->p);
+  rsa->q = std::move(tmp->q);
+  rsa->dmp1 = std::move(tmp->dmp1);
+  rsa->dmq1 = std::move(tmp->dmq1);
+  rsa->iqmp = std::move(tmp->iqmp);
+  rsa->mont_n = std::move(tmp->mont_n);
+  rsa->mont_p = std::move(tmp->mont_p);
+  rsa->mont_q = std::move(tmp->mont_q);
+  rsa->d_fixed = std::move(tmp->d_fixed);
+  rsa->dmp1_fixed = std::move(tmp->dmp1_fixed);
+  rsa->dmq1_fixed = std::move(tmp->dmq1_fixed);
+  rsa->iqmp_mont = std::move(tmp->iqmp_mont);
+  rsa->private_key_frozen = tmp->private_key_frozen;
+  return 1;
+}
+
+int RSA_generate_key_ex(RSA *rsa, int bits, const BIGNUM *e_value,
+                        BN_GENCB *cb) {
+  return RSA_generate_key_ex_maybe_fips(FromOpaque(rsa), bits, e_value, cb,
+                                        /*check_fips=*/0);
+}
+
+int RSA_generate_key_fips(RSA *rsa, int bits, BN_GENCB *cb) {
+  // FIPS 186-4 allowed 2048-bit and 3072-bit RSA keys (1024-bit and 1536-bit
+  // primes, respectively) with the prime generation method we use.
+  // Subsequently, IG A.14 stated that larger modulus sizes can be used and ACVP
+  // testing supports 4096 bits, and FIPS 186-5 allowed all key sizes at least
+  // 2048.
+  if (bits != 2048 && bits != 3072 && bits != 4096) {
+    OPENSSL_PUT_ERROR(RSA, RSA_R_BAD_RSA_PARAMETERS);
+    return 0;
+  }
+
+  BIGNUM *e = BN_new();
+  int ret = e != nullptr && BN_set_word(e, RSA_F4) &&
+            RSA_generate_key_ex_maybe_fips(FromOpaque(rsa), bits, e, cb,
+                                           /*check_fips=*/1);
+  BN_free(e);
+
+  if (ret) {
+    FIPS_service_indicator_update_state();
+  }
+  return ret;
+}
+
+BSSL_NAMESPACE_BEGIN
+
+DEFINE_METHOD_FUNCTION(RSA_METHOD, RSA_default_method) {
+  // All of the methods are NULL to make it easier for the compiler/linker to
+  // drop unused functions. The wrapper functions will select the appropriate
+  // `rsa_default_*` implementation.
+  OPENSSL_memset(out, 0, sizeof(RSA_METHOD));
+  out->common.is_static = 1;
+}
+
+BSSL_NAMESPACE_END