grpc-flamingo 1.11.0 → 1.15.0

data/third_party/boringssl/crypto/fipsmodule/bn/generic.c

@@ -64,7 +64,8 @@
 // This file has two other implementations: x86 assembly language in
 // asm/bn-586.pl and x86_64 inline assembly in asm/x86_64-gcc.c.
 #if defined(OPENSSL_NO_ASM) || \
-    !(defined(OPENSSL_X86) || (defined(OPENSSL_X86_64) && defined(__GNUC__)))
+    !(defined(OPENSSL_X86) ||  \
+      (defined(OPENSSL_X86_64) && (defined(__GNUC__) || defined(__clang__))))
 
 #ifdef BN_ULLONG
 #define mul_add(r, a, w, c)               \

data/third_party/boringssl/crypto/fipsmodule/bn/internal.h

@@ -140,9 +140,12 @@ extern "C" {
 
 #if defined(OPENSSL_64_BIT)
 
-#if !defined(_MSC_VER)
+#if defined(BORINGSSL_HAS_UINT128)
 // MSVC doesn't support two-word integers on 64-bit.
 #define BN_ULLONG uint128_t
+#if defined(BORINGSSL_CAN_DIVIDE_UINT128)
+#define BN_CAN_DIVIDE_ULLONG
+#endif
 #endif
 
 #define BN_BITS2 64
@@ -160,6 +163,7 @@ extern "C" {
 #elif defined(OPENSSL_32_BIT)
 
 #define BN_ULLONG uint64_t
+#define BN_CAN_DIVIDE_ULLONG
 #define BN_BITS2 32
 #define BN_BYTES 4
 #define BN_BITS4 16
@@ -193,9 +197,13 @@ extern "C" {
 #define Hw(t) ((BN_ULONG)((t) >> BN_BITS2))
 #endif
 
-// bn_correct_top decrements |bn->top| until |bn->d[top-1]| is non-zero or
-// until |top| is zero. If |bn| is zero, |bn->neg| is set to zero.
-void bn_correct_top(BIGNUM *bn);
+// bn_minimal_width returns the minimal value of |bn->top| which fits the
+// value of |bn|.
+int bn_minimal_width(const BIGNUM *bn);
+
+// bn_set_minimal_width sets |bn->width| to |bn_minimal_width(bn)|. If |bn| is
+// zero, |bn->neg| is set to zero.
+void bn_set_minimal_width(BIGNUM *bn);
 
 // bn_wexpand ensures that |bn| has at least |words| works of space without
 // altering its value. It returns one on success or zero on allocation
@@ -206,10 +214,27 @@ int bn_wexpand(BIGNUM *bn, size_t words);
 // than a number of words.
 int bn_expand(BIGNUM *bn, size_t bits);
 
+// bn_resize_words adjusts |bn->top| to be |words|. It returns one on success
+// and zero on allocation error or if |bn|'s value is too large.
+OPENSSL_EXPORT int bn_resize_words(BIGNUM *bn, size_t words);
+
+// bn_select_words sets |r| to |a| if |mask| is all ones or |b| if |mask| is
+// all zeros.
+void bn_select_words(BN_ULONG *r, BN_ULONG mask, const BN_ULONG *a,
+                     const BN_ULONG *b, size_t num);
+
 // bn_set_words sets |bn| to the value encoded in the |num| words in |words|,
 // least significant word first.
 int bn_set_words(BIGNUM *bn, const BN_ULONG *words, size_t num);
 
+// bn_fits_in_words returns one if |bn| may be represented in |num| words, plus
+// a sign bit, and zero otherwise.
+int bn_fits_in_words(const BIGNUM *bn, size_t num);
+
+// bn_copy_words copies the value of |bn| to |out| and returns one if the value
+// is representable in |num| words. Otherwise, it returns zero.
+int bn_copy_words(BN_ULONG *out, size_t num, const BIGNUM *bn);
+
 // bn_mul_add_words multiples |ap| by |w|, adds the result to |rp|, and places
 // the result in |rp|. |ap| and |rp| must both be |num| words long. It returns
 // the carry word of the operation. |ap| and |rp| may be equal but otherwise may
@@ -255,25 +280,13 @@ void bn_sqr_comba8(BN_ULONG r[16], const BN_ULONG a[4]);
 // bn_sqr_comba4 sets |r| to |a|^2.
 void bn_sqr_comba4(BN_ULONG r[8], const BN_ULONG a[4]);
 
-// bn_cmp_words returns a value less than, equal to or greater than zero if
-// the, length |n|, array |a| is less than, equal to or greater than |b|.
-int bn_cmp_words(const BN_ULONG *a, const BN_ULONG *b, int n);
-
-// bn_cmp_words returns a value less than, equal to or greater than zero if the
-// array |a| is less than, equal to or greater than |b|. The arrays can be of
-// different lengths: |cl| gives the minimum of the two lengths and |dl| gives
-// the length of |a| minus the length of |b|.
-int bn_cmp_part_words(const BN_ULONG *a, const BN_ULONG *b, int cl, int dl);
-
 // bn_less_than_words returns one if |a| < |b| and zero otherwise, where |a|
 // and |b| both are |len| words long. It runs in constant time.
 int bn_less_than_words(const BN_ULONG *a, const BN_ULONG *b, size_t len);
 
 // bn_in_range_words returns one if |min_inclusive| <= |a| < |max_exclusive|,
-// where |a| and |max_exclusive| both are |len| words long. This function leaks
-// which of [0, min_inclusive), [min_inclusive, max_exclusive), and
-// [max_exclusive, 2^(BN_BITS2*len)) contains |a|, but otherwise the value of
-// |a| is secret.
+// where |a| and |max_exclusive| both are |len| words long. |a| and
+// |max_exclusive| are treated as secret.
 int bn_in_range_words(const BN_ULONG *a, BN_ULONG min_inclusive,
                       const BN_ULONG *max_exclusive, size_t len);
 
@@ -288,11 +301,38 @@ int bn_rand_range_words(BN_ULONG *out, BN_ULONG min_inclusive,
                         const BN_ULONG *max_exclusive, size_t len,
                         const uint8_t additional_data[32]);
 
+// bn_range_secret_range behaves like |BN_rand_range_ex|, but treats
+// |max_exclusive| as secret. Because of this constraint, the distribution of
+// values returned is more complex.
+//
+// Rather than repeatedly generating values until one is in range, which would
+// leak information, it generates one value. If the value is in range, it sets
+// |*out_is_uniform| to one. Otherwise, it sets |*out_is_uniform| to zero,
+// fixing up the value to force it in range.
+//
+// The subset of calls to |bn_rand_secret_range| which set |*out_is_uniform| to
+// one are uniformly distributed in the target range. Calls overall are not.
+// This function is intended for use in situations where the extra values are
+// still usable and where the number of iterations needed to reach the target
+// number of uniform outputs may be blinded for negligible probabilities of
+// timing leaks.
+//
+// Although this function treats |max_exclusive| as secret, it treats the number
+// of bits in |max_exclusive| as public.
+int bn_rand_secret_range(BIGNUM *r, int *out_is_uniform, BN_ULONG min_inclusive,
+                         const BIGNUM *max_exclusive);
+
 int bn_mul_mont(BN_ULONG *rp, const BN_ULONG *ap, const BN_ULONG *bp,
                 const BN_ULONG *np, const BN_ULONG *n0, int num);
 
 uint64_t bn_mont_n0(const BIGNUM *n);
-int bn_mod_exp_base_2_vartime(BIGNUM *r, unsigned p, const BIGNUM *n);
+
+// bn_mod_exp_base_2_consttime calculates r = 2**p (mod n). |p| must be larger
+// than log_2(n); i.e. 2**p must be larger than |n|. |n| must be positive and
+// odd. |p| and the bit width of |n| are assumed public, but |n| is otherwise
+// treated as secret.
+int bn_mod_exp_base_2_consttime(BIGNUM *r, unsigned p, const BIGNUM *n,
+                                BN_CTX *ctx);
 
 #if defined(OPENSSL_X86_64) && defined(_MSC_VER)
 #define BN_UMULT_LOHI(low, high, a, b) ((low) = _umul128((a), (b), &(high)))
@@ -302,6 +342,127 @@ int bn_mod_exp_base_2_vartime(BIGNUM *r, unsigned p, const BIGNUM *n);
 #error "Either BN_ULLONG or BN_UMULT_LOHI must be defined on every platform."
 #endif
 
+// bn_jacobi returns the Jacobi symbol of |a| and |b| (which is -1, 0 or 1), or
+// -2 on error.
+int bn_jacobi(const BIGNUM *a, const BIGNUM *b, BN_CTX *ctx);
+
+// bn_is_bit_set_words returns one if bit |bit| is set in |a| and zero
+// otherwise.
+int bn_is_bit_set_words(const BN_ULONG *a, size_t num, unsigned bit);
+
+// bn_one_to_montgomery sets |r| to one in Montgomery form. It returns one on
+// success and zero on error. This function treats the bit width of the modulus
+// as public.
+int bn_one_to_montgomery(BIGNUM *r, const BN_MONT_CTX *mont, BN_CTX *ctx);
+
+// bn_less_than_montgomery_R returns one if |bn| is less than the Montgomery R
+// value for |mont| and zero otherwise.
+int bn_less_than_montgomery_R(const BIGNUM *bn, const BN_MONT_CTX *mont);
+
+// bn_mod_u16_consttime returns |bn| mod |d|, ignoring |bn|'s sign bit. It runs
+// in time independent of the value of |bn|, but it treats |d| as public.
+OPENSSL_EXPORT uint16_t bn_mod_u16_consttime(const BIGNUM *bn, uint16_t d);
+
+// bn_odd_number_is_obviously_composite returns one if |bn| is divisible by one
+// of the first several odd primes and zero otherwise.
+int bn_odd_number_is_obviously_composite(const BIGNUM *bn);
+
+// bn_rshift1_words sets |r| to |a| >> 1, where both arrays are |num| bits wide.
+void bn_rshift1_words(BN_ULONG *r, const BN_ULONG *a, size_t num);
+
+// bn_rshift_secret_shift behaves like |BN_rshift| but runs in time independent
+// of both |a| and |n|.
+OPENSSL_EXPORT int bn_rshift_secret_shift(BIGNUM *r, const BIGNUM *a,
+                                          unsigned n, BN_CTX *ctx);
+
+
+// Constant-time non-modular arithmetic.
+//
+// The following functions implement non-modular arithmetic in constant-time
+// and pessimally set |r->width| to the largest possible word size.
+//
+// Note this means that, e.g., repeatedly multiplying by one will cause widths
+// to increase without bound. The corresponding public API functions minimize
+// their outputs to avoid regressing calculator consumers.
+
+// bn_uadd_consttime behaves like |BN_uadd|, but it pessimally sets
+// |r->width| = |a->width| + |b->width| + 1.
+int bn_uadd_consttime(BIGNUM *r, const BIGNUM *a, const BIGNUM *b);
+
+// bn_usub_consttime behaves like |BN_usub|, but it pessimally sets
+// |r->width| = |a->width|.
+int bn_usub_consttime(BIGNUM *r, const BIGNUM *a, const BIGNUM *b);
+
+// bn_abs_sub_consttime sets |r| to the absolute value of |a| - |b|, treating
+// both inputs as secret. It returns one on success and zero on error.
+OPENSSL_EXPORT int bn_abs_sub_consttime(BIGNUM *r, const BIGNUM *a,
+                                        const BIGNUM *b, BN_CTX *ctx);
+
+// bn_mul_consttime behaves like |BN_mul|, but it rejects negative inputs and
+// pessimally sets |r->width| to |a->width| + |b->width|, to avoid leaking
+// information about |a| and |b|.
+int bn_mul_consttime(BIGNUM *r, const BIGNUM *a, const BIGNUM *b, BN_CTX *ctx);
+
+// bn_sqrt_consttime behaves like |BN_sqrt|, but it pessimally sets |r->width|
+// to 2*|a->width|, to avoid leaking information about |a| and |b|.
+int bn_sqr_consttime(BIGNUM *r, const BIGNUM *a, BN_CTX *ctx);
+
+// bn_div_consttime behaves like |BN_div|, but it rejects negative inputs and
+// treats both inputs, including their magnitudes, as secret. It is, as a
+// result, much slower than |BN_div| and should only be used for rare operations
+// where Montgomery reduction is not available.
+//
+// Note that |quotient->width| will be set pessimally to |numerator->width|.
+OPENSSL_EXPORT int bn_div_consttime(BIGNUM *quotient, BIGNUM *remainder,
+                                    const BIGNUM *numerator,
+                                    const BIGNUM *divisor, BN_CTX *ctx);
+
+// bn_is_relatively_prime checks whether GCD(|x|, |y|) is one. On success, it
+// returns one and sets |*out_relatively_prime| to one if the GCD was one and
+// zero otherwise. On error, it returns zero.
+OPENSSL_EXPORT int bn_is_relatively_prime(int *out_relatively_prime,
+                                          const BIGNUM *x, const BIGNUM *y,
+                                          BN_CTX *ctx);
+
+// bn_lcm_consttime sets |r| to LCM(|a|, |b|). It returns one and success and
+// zero on error. |a| and |b| are both treated as secret.
+OPENSSL_EXPORT int bn_lcm_consttime(BIGNUM *r, const BIGNUM *a, const BIGNUM *b,
+                                    BN_CTX *ctx);
+
+
+// Constant-time modular arithmetic.
+//
+// The following functions implement basic constant-time modular arithmetic.
+
+// bn_mod_add_consttime acts like |BN_mod_add_quick| but takes a |BN_CTX|.
+int bn_mod_add_consttime(BIGNUM *r, const BIGNUM *a, const BIGNUM *b,
+                         const BIGNUM *m, BN_CTX *ctx);
+
+// bn_mod_sub_consttime acts like |BN_mod_sub_quick| but takes a |BN_CTX|.
+int bn_mod_sub_consttime(BIGNUM *r, const BIGNUM *a, const BIGNUM *b,
+                         const BIGNUM *m, BN_CTX *ctx);
+
+// bn_mod_lshift1_consttime acts like |BN_mod_lshift1_quick| but takes a
+// |BN_CTX|.
+int bn_mod_lshift1_consttime(BIGNUM *r, const BIGNUM *a, const BIGNUM *m,
+                             BN_CTX *ctx);
+
+// bn_mod_lshift_consttime acts like |BN_mod_lshift_quick| but takes a |BN_CTX|.
+int bn_mod_lshift_consttime(BIGNUM *r, const BIGNUM *a, int n, const BIGNUM *m,
+                            BN_CTX *ctx);
+
+// bn_mod_inverse_consttime sets |r| to |a|^-1, mod |n|. |a| must be non-
+// negative and less than |n|. It returns one on success and zero on error. On
+// failure, if the failure was caused by |a| having no inverse mod |n| then
+// |*out_no_inverse| will be set to one; otherwise it will be set to zero.
+//
+// This function treats both |a| and |n| as secret, provided they are both non-
+// zero and the inverse exists. It should only be used for even moduli where
+// none of the less general implementations are applicable.
+OPENSSL_EXPORT int bn_mod_inverse_consttime(BIGNUM *r, int *out_no_inverse,
+                                            const BIGNUM *a, const BIGNUM *n,
+                                            BN_CTX *ctx);
+
 // bn_mod_inverse_prime sets |out| to the modular inverse of |a| modulo |p|,
 // computed with Fermat's Little Theorem. It returns one on success and zero on
 // error. If |mont_p| is NULL, one will be computed temporarily.
@@ -314,14 +475,6 @@ int bn_mod_inverse_prime(BIGNUM *out, const BIGNUM *a, const BIGNUM *p,
 int bn_mod_inverse_secret_prime(BIGNUM *out, const BIGNUM *a, const BIGNUM *p,
                                 BN_CTX *ctx, const BN_MONT_CTX *mont_p);
 
-// bn_jacobi returns the Jacobi symbol of |a| and |b| (which is -1, 0 or 1), or
-// -2 on error.
-int bn_jacobi(const BIGNUM *a, const BIGNUM *b, BN_CTX *ctx);
-
-// bn_is_bit_set_words returns one if bit |bit| is set in |a| and zero
-// otherwise.
-int bn_is_bit_set_words(const BN_ULONG *a, size_t num, unsigned bit);
-
 
 // Low-level operations for small numbers.
 //
@@ -368,6 +521,13 @@ int bn_to_montgomery_small(BN_ULONG *r, size_t num_r, const BN_ULONG *a,
 int bn_from_montgomery_small(BN_ULONG *r, size_t num_r, const BN_ULONG *a,
                              size_t num_a, const BN_MONT_CTX *mont);
 
+// bn_one_to_montgomery_small sets |r| to one in Montgomery form. It returns one
+// on success and zero on error. |num_r| must be the length of the modulus,
+// which is |mont->N.top|. This function treats the bit width of the modulus as
+// public.
+int bn_one_to_montgomery_small(BN_ULONG *r, size_t num_r,
+                               const BN_MONT_CTX *mont);
+
 // bn_mod_mul_montgomery_small sets |r| to |a| * |b| mod |mont->N|. Both inputs
 // and outputs are in the Montgomery domain. |num_r| must be the length of the
 // modulus, which is |mont->N.top|. This function returns one on success and

data/third_party/boringssl/crypto/fipsmodule/bn/jacobi.c

@@ -58,7 +58,7 @@
 
 
 // least significant word
-#define BN_lsw(n) (((n)->top == 0) ? (BN_ULONG) 0 : (n)->d[0])
+#define BN_lsw(n) (((n)->width == 0) ? (BN_ULONG) 0 : (n)->d[0])
 
 int bn_jacobi(const BIGNUM *a, const BIGNUM *b, BN_CTX *ctx) {
   // In 'tab', only odd-indexed entries are relevant:

data/third_party/boringssl/crypto/fipsmodule/bn/montgomery.c

@@ -126,10 +126,6 @@
 #define OPENSSL_BN_ASM_MONT
 #endif
 
-static int bn_mod_mul_montgomery_fallback(BIGNUM *r, const BIGNUM *a,
-                                          const BIGNUM *b,
-                                          const BN_MONT_CTX *mont, BN_CTX *ctx);
-
 
 BN_MONT_CTX *BN_MONT_CTX_new(void) {
   BN_MONT_CTX *ret = OPENSSL_malloc(sizeof(BN_MONT_CTX));
@@ -193,6 +189,10 @@ int BN_MONT_CTX_set(BN_MONT_CTX *mont, const BIGNUM *mod, BN_CTX *ctx) {
     OPENSSL_PUT_ERROR(BN, ERR_R_INTERNAL_ERROR);
     return 0;
   }
+  // |mont->N| is always stored minimally. Computing RR efficiently leaks the
+  // size of the modulus. While the modulus may be private in RSA (one of the
+  // primes), their sizes are public, so this is fine.
+  bn_set_minimal_width(&mont->N);
 
   // Find n0 such that n0 * N == -1 (mod r).
   //
@@ -200,7 +200,7 @@ int BN_MONT_CTX_set(BN_MONT_CTX *mont, const BIGNUM *mod, BN_CTX *ctx) {
   // others, we could use a shorter R value and use faster |BN_ULONG|-based
   // math instead of |uint64_t|-based math, which would be double-precision.
   // However, currently only the assembler files know which is which.
-  uint64_t n0 = bn_mont_n0(mod);
+  uint64_t n0 = bn_mont_n0(&mont->N);
   mont->n0[0] = (BN_ULONG)n0;
 #if BN_MONT_CTX_N0_LIMBS == 2
   mont->n0[1] = (BN_ULONG)(n0 >> BN_BITS2);
@@ -208,19 +208,34 @@ int BN_MONT_CTX_set(BN_MONT_CTX *mont, const BIGNUM *mod, BN_CTX *ctx) {
   mont->n0[1] = 0;
 #endif
 
+  BN_CTX *new_ctx = NULL;
+  if (ctx == NULL) {
+    new_ctx = BN_CTX_new();
+    if (new_ctx == NULL) {
+      return 0;
+    }
+    ctx = new_ctx;
+  }
+
   // Save RR = R**2 (mod N). R is the smallest power of 2**BN_BITS2 such that R
   // > mod. Even though the assembly on some 32-bit platforms works with 64-bit
   // values, using |BN_BITS2| here, rather than |BN_MONT_CTX_N0_LIMBS *
   // BN_BITS2|, is correct because R**2 will still be a multiple of the latter
   // as |BN_MONT_CTX_N0_LIMBS| is either one or two.
-  //
-  // XXX: This is not constant time with respect to |mont->N|, but it should be.
-  unsigned lgBigR = (BN_num_bits(mod) + (BN_BITS2 - 1)) / BN_BITS2 * BN_BITS2;
-  if (!bn_mod_exp_base_2_vartime(&mont->RR, lgBigR * 2, &mont->N)) {
-    return 0;
-  }
+  unsigned lgBigR = mont->N.width * BN_BITS2;
+  int ok = bn_mod_exp_base_2_consttime(&mont->RR, lgBigR * 2, &mont->N, ctx);
+  BN_CTX_free(new_ctx);
+  return ok;
+}
 
-  return 1;
+BN_MONT_CTX *BN_MONT_CTX_new_for_modulus(const BIGNUM *mod, BN_CTX *ctx) {
+  BN_MONT_CTX *mont = BN_MONT_CTX_new();
+  if (mont == NULL ||
+      !BN_MONT_CTX_set(mont, mod, ctx)) {
+    BN_MONT_CTX_free(mont);
+    return NULL;
+  }
+  return mont;
 }
 
 int BN_MONT_CTX_set_locked(BN_MONT_CTX **pmont, CRYPTO_MUTEX *lock,
@@ -234,25 +249,12 @@ int BN_MONT_CTX_set_locked(BN_MONT_CTX **pmont, CRYPTO_MUTEX *lock,
   }
 
   CRYPTO_MUTEX_lock_write(lock);
-  ctx = *pmont;
-  if (ctx) {
-    goto out;
+  if (*pmont == NULL) {
+    *pmont = BN_MONT_CTX_new_for_modulus(mod, bn_ctx);
   }
-
-  ctx = BN_MONT_CTX_new();
-  if (ctx == NULL) {
-    goto out;
-  }
-  if (!BN_MONT_CTX_set(ctx, mod, bn_ctx)) {
-    BN_MONT_CTX_free(ctx);
-    ctx = NULL;
-    goto out;
-  }
-  *pmont = ctx;
-
-out:
+  const int ok = *pmont != NULL;
   CRYPTO_MUTEX_unlock_write(lock);
-  return ctx != NULL;
+  return ok;
 }
 
 int BN_to_montgomery(BIGNUM *ret, const BIGNUM *a, const BN_MONT_CTX *mont,
@@ -263,7 +265,7 @@ int BN_to_montgomery(BIGNUM *ret, const BIGNUM *a, const BN_MONT_CTX *mont,
 static int bn_from_montgomery_in_place(BN_ULONG *r, size_t num_r, BN_ULONG *a,
                                        size_t num_a, const BN_MONT_CTX *mont) {
   const BN_ULONG *n = mont->N.d;
-  size_t num_n = mont->N.top;
+  size_t num_n = mont->N.width;
   if (num_r != num_n || num_a != 2 * num_n) {
     OPENSSL_PUT_ERROR(BN, ERR_R_SHOULD_NOT_HAVE_BEEN_CALLED);
     return 0;
@@ -304,32 +306,26 @@ static int bn_from_montgomery_in_place(BN_ULONG *r, size_t num_r, BN_ULONG *a,
 
 static int BN_from_montgomery_word(BIGNUM *ret, BIGNUM *r,
                                    const BN_MONT_CTX *mont) {
-  const BIGNUM *n = &mont->N;
-  if (n->top == 0) {
-    ret->top = 0;
-    return 1;
-  }
-
-  int max = (2 * n->top);  // carry is stored separately
-  if (!bn_wexpand(r, max) ||
-      !bn_wexpand(ret, n->top)) {
+  if (r->neg) {
+    OPENSSL_PUT_ERROR(BN, BN_R_NEGATIVE_NUMBER);
     return 0;
   }
-  // Clear the top words of |r|.
-  if (max > r->top) {
-    OPENSSL_memset(r->d + r->top, 0, (max - r->top) * sizeof(BN_ULONG));
+
+  const BIGNUM *n = &mont->N;
+  if (n->width == 0) {
+    ret->width = 0;
+    return 1;
   }
-  r->top = max;
-  ret->top = n->top;
 
-  if (!bn_from_montgomery_in_place(ret->d, ret->top, r->d, r->top, mont)) {
+  int max = 2 * n->width;  // carry is stored separately
+  if (!bn_resize_words(r, max) ||
+      !bn_wexpand(ret, n->width)) {
     return 0;
   }
-  ret->neg = r->neg;
 
-  bn_correct_top(r);
-  bn_correct_top(ret);
-  return 1;
+  ret->width = n->width;
+  ret->neg = 0;
+  return bn_from_montgomery_in_place(ret->d, ret->width, r->d, r->width, mont);
 }
 
 int BN_from_montgomery(BIGNUM *r, const BIGNUM *a, const BN_MONT_CTX *mont,
@@ -352,35 +348,24 @@ err:
   return ret;
 }
 
-int BN_mod_mul_montgomery(BIGNUM *r, const BIGNUM *a, const BIGNUM *b,
-                          const BN_MONT_CTX *mont, BN_CTX *ctx) {
-#if !defined(OPENSSL_BN_ASM_MONT)
-  return bn_mod_mul_montgomery_fallback(r, a, b, mont, ctx);
-#else
-  int num = mont->N.top;
-
-  // |bn_mul_mont| requires at least 128 bits of limbs, at least for x86.
-  if (num < (128 / BN_BITS2) ||
-      a->top != num ||
-      b->top != num) {
-    return bn_mod_mul_montgomery_fallback(r, a, b, mont, ctx);
-  }
-
-  if (!bn_wexpand(r, num)) {
-    return 0;
-  }
-  if (!bn_mul_mont(r->d, a->d, b->d, mont->N.d, mont->n0, num)) {
-    // The check above ensures this won't happen.
-    assert(0);
-    OPENSSL_PUT_ERROR(BN, ERR_R_INTERNAL_ERROR);
-    return 0;
+int bn_one_to_montgomery(BIGNUM *r, const BN_MONT_CTX *mont, BN_CTX *ctx) {
+  // If the high bit of |n| is set, R = 2^(width*BN_BITS2) < 2 * |n|, so we
+  // compute R - |n| rather than perform Montgomery reduction.
+  const BIGNUM *n = &mont->N;
+  if (n->width > 0 && (n->d[n->width - 1] >> (BN_BITS2 - 1)) != 0) {
+    if (!bn_wexpand(r, n->width)) {
+      return 0;
+    }
+    r->d[0] = 0 - n->d[0];
+    for (int i = 1; i < n->width; i++) {
+      r->d[i] = ~n->d[i];
+    }
+    r->width = n->width;
+    r->neg = 0;
+    return 1;
   }
-  r->neg = a->neg ^ b->neg;
-  r->top = num;
-  bn_correct_top(r);
 
-  return 1;
-#endif
+  return BN_from_montgomery(r, &mont->RR, mont, ctx);
 }
 
 static int bn_mod_mul_montgomery_fallback(BIGNUM *r, const BIGNUM *a,
@@ -396,11 +381,11 @@ static int bn_mod_mul_montgomery_fallback(BIGNUM *r, const BIGNUM *a,
   }
 
   if (a == b) {
-    if (!BN_sqr(tmp, a, ctx)) {
+    if (!bn_sqr_consttime(tmp, a, ctx)) {
       goto err;
     }
   } else {
-    if (!BN_mul(tmp, a, b, ctx)) {
+    if (!bn_mul_consttime(tmp, a, b, ctx)) {
       goto err;
     }
   }
@@ -417,15 +402,51 @@ err:
   return ret;
 }
 
+int BN_mod_mul_montgomery(BIGNUM *r, const BIGNUM *a, const BIGNUM *b,
+                          const BN_MONT_CTX *mont, BN_CTX *ctx) {
+  if (a->neg || b->neg) {
+    OPENSSL_PUT_ERROR(BN, BN_R_NEGATIVE_NUMBER);
+    return 0;
+  }
+
+#if defined(OPENSSL_BN_ASM_MONT)
+  // |bn_mul_mont| requires at least 128 bits of limbs, at least for x86.
+  int num = mont->N.width;
+  if (num >= (128 / BN_BITS2) &&
+      a->width == num &&
+      b->width == num) {
+    if (!bn_wexpand(r, num)) {
+      return 0;
+    }
+    if (!bn_mul_mont(r->d, a->d, b->d, mont->N.d, mont->n0, num)) {
+      // The check above ensures this won't happen.
+      assert(0);
+      OPENSSL_PUT_ERROR(BN, ERR_R_INTERNAL_ERROR);
+      return 0;
+    }
+    r->neg = 0;
+    r->width = num;
+    return 1;
+  }
+#endif
+
+  return bn_mod_mul_montgomery_fallback(r, a, b, mont, ctx);
+}
+
+int bn_less_than_montgomery_R(const BIGNUM *bn, const BN_MONT_CTX *mont) {
+  return !BN_is_negative(bn) &&
+         bn_fits_in_words(bn, mont->N.width);
+}
+
 int bn_to_montgomery_small(BN_ULONG *r, size_t num_r, const BN_ULONG *a,
                            size_t num_a, const BN_MONT_CTX *mont) {
   return bn_mod_mul_montgomery_small(r, num_r, a, num_a, mont->RR.d,
-                                     mont->RR.top, mont);
+                                     mont->RR.width, mont);
 }
 
 int bn_from_montgomery_small(BN_ULONG *r, size_t num_r, const BN_ULONG *a,
                              size_t num_a, const BN_MONT_CTX *mont) {
-  size_t num_n = mont->N.top;
+  size_t num_n = mont->N.width;
   if (num_a > 2 * num_n || num_r != num_n || num_n > BN_SMALL_MAX_WORDS) {
     OPENSSL_PUT_ERROR(BN, ERR_R_SHOULD_NOT_HAVE_BEEN_CALLED);
     return 0;
@@ -439,10 +460,32 @@ int bn_from_montgomery_small(BN_ULONG *r, size_t num_r, const BN_ULONG *a,
   return ret;
 }
 
+int bn_one_to_montgomery_small(BN_ULONG *r, size_t num_r,
+                               const BN_MONT_CTX *mont) {
+  const BN_ULONG *n = mont->N.d;
+  size_t num_n = mont->N.width;
+  if (num_n == 0 || num_r != num_n) {
+    OPENSSL_PUT_ERROR(BN, ERR_R_SHOULD_NOT_HAVE_BEEN_CALLED);
+    return 0;
+  }
+
+  // If the high bit of |n| is set, R = 2^(num_n*BN_BITS2) < 2 * |n|, so we
+  // compute R - |n| rather than perform Montgomery reduction.
+  if (num_n > 0 && (n[num_n - 1] >> (BN_BITS2 - 1)) != 0) {
+    r[0] = 0 - n[0];
+    for (size_t i = 1; i < num_n; i++) {
+      r[i] = ~n[i];
+    }
+    return 1;
+  }
+
+  return bn_from_montgomery_small(r, num_r, mont->RR.d, mont->RR.width, mont);
+}
+
 int bn_mod_mul_montgomery_small(BN_ULONG *r, size_t num_r, const BN_ULONG *a,
                                 size_t num_a, const BN_ULONG *b, size_t num_b,
                                 const BN_MONT_CTX *mont) {
-  size_t num_n = mont->N.top;
+  size_t num_n = mont->N.width;
   if (num_r != num_n || num_a + num_b > 2 * num_n ||
       num_n > BN_SMALL_MAX_WORDS) {
     OPENSSL_PUT_ERROR(BN, ERR_R_SHOULD_NOT_HAVE_BEEN_CALLED);