[PATCH] prandom: make use of smaller types in prandom_u32_max

[PATCH] prandom: make use of smaller types in prandom_u32_max

[Jason A. Donenfeld]

When possible at compile-time, make use of smaller types in
prandom_u32_max(), so that we can use smaller batches from random.c,
which in turn leads to a 2x or 4x performance boost. This makes a
difference, for example, in kfence, which needs a fast stream of small
numbers (booleans).

At the same time, we use the occasion to update the old documentation on
these functions. prandom_u32() and prandom_bytes() have direct
replacements now in random.h, while prandom_u32_max() remains useful as
a prandom.h function, since it's not cryptographically secure by virtue
of not being evenly distributed.

Cc: Marco Elver <elver@google.com>
Cc: Dominik Brodowski <linux@dominikbrodowski.net>
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
---
 include/linux/prandom.h | 17 +++++++++++------
 1 file changed, 11 insertions(+), 6 deletions(-)

diff --git a/include/linux/prandom.h b/include/linux/prandom.h
index deace5fb4e62..78db003bc290 100644
--- a/include/linux/prandom.h
+++ b/include/linux/prandom.h
@@ -12,11 +12,13 @@
 #include <linux/percpu.h>
 #include <linux/random.h>
 
+/* Deprecated: use get_random_u32 instead. */
 static inline u32 prandom_u32(void)
 {
 	return get_random_u32();
 }
 
+/* Deprecated: use get_random_bytes instead. */
 static inline void prandom_bytes(void *buf, size_t nbytes)
 {
 	return get_random_bytes(buf, nbytes);
@@ -37,17 +39,20 @@ void prandom_seed_full_state(struct rnd_state __percpu *pcpu_state);
  * prandom_u32_max - returns a pseudo-random number in interval [0, ep_ro)
  * @ep_ro: right open interval endpoint
  *
- * Returns a pseudo-random number that is in interval [0, ep_ro). Note
- * that the result depends on PRNG being well distributed in [0, ~0U]
- * u32 space. Here we use maximally equidistributed combined Tausworthe
- * generator, that is, prandom_u32(). This is useful when requesting a
- * random index of an array containing ep_ro elements, for example.
+ * Returns a pseudo-random number that is in interval [0, ep_ro). This is
+ * useful when requesting a random index of an array containing ep_ro elements,
+ * for example. The result is somewhat biased when ep_ro is not a power of 2,
+ * so do not use this for cryptographic purposes.
  *
  * Returns: pseudo-random number in interval [0, ep_ro)
  */
 static inline u32 prandom_u32_max(u32 ep_ro)
 {
-	return (u32)(((u64) prandom_u32() * ep_ro) >> 32);
+	if (__builtin_constant_p(ep_ro <= 1U << 8) && ep_ro <= 1U << 8)
+		return (get_random_u8() * ep_ro) >> 8;
+	if (__builtin_constant_p(ep_ro <= 1U << 16) && ep_ro <= 1U << 16)
+		return (get_random_u16() * ep_ro) >> 16;
+	return ((u64)get_random_u32() * ep_ro) >> 32;
 }
 
 /*
-- 
2.37.3

Re: [PATCH] prandom: make use of smaller types in prandom_u32_max

[Marco Elver]

On Thu, 29 Sept 2022 at 12:31, Jason A. Donenfeld <Jason@zx2c4.com> wrote:
>
> When possible at compile-time, make use of smaller types in
> prandom_u32_max(), so that we can use smaller batches from random.c,
> which in turn leads to a 2x or 4x performance boost. This makes a
> difference, for example, in kfence, which needs a fast stream of small
> numbers (booleans).
>
> At the same time, we use the occasion to update the old documentation on
> these functions. prandom_u32() and prandom_bytes() have direct
> replacements now in random.h, while prandom_u32_max() remains useful as
> a prandom.h function, since it's not cryptographically secure by virtue
> of not being evenly distributed.
>
> Cc: Marco Elver <elver@google.com>
> Cc: Dominik Brodowski <linux@dominikbrodowski.net>
> Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>

Acked-by: Marco Elver <elver@google.com>

> ---
>  include/linux/prandom.h | 17 +++++++++++------
>  1 file changed, 11 insertions(+), 6 deletions(-)
>
> diff --git a/include/linux/prandom.h b/include/linux/prandom.h
> index deace5fb4e62..78db003bc290 100644
> --- a/include/linux/prandom.h
> +++ b/include/linux/prandom.h
> @@ -12,11 +12,13 @@
>  #include <linux/percpu.h>
>  #include <linux/random.h>
>
> +/* Deprecated: use get_random_u32 instead. */
>  static inline u32 prandom_u32(void)
>  {
>         return get_random_u32();
>  }
>
> +/* Deprecated: use get_random_bytes instead. */
>  static inline void prandom_bytes(void *buf, size_t nbytes)
>  {
>         return get_random_bytes(buf, nbytes);
> @@ -37,17 +39,20 @@ void prandom_seed_full_state(struct rnd_state __percpu *pcpu_state);
>   * prandom_u32_max - returns a pseudo-random number in interval [0, ep_ro)
>   * @ep_ro: right open interval endpoint
>   *
> - * Returns a pseudo-random number that is in interval [0, ep_ro). Note
> - * that the result depends on PRNG being well distributed in [0, ~0U]
> - * u32 space. Here we use maximally equidistributed combined Tausworthe
> - * generator, that is, prandom_u32(). This is useful when requesting a
> - * random index of an array containing ep_ro elements, for example.
> + * Returns a pseudo-random number that is in interval [0, ep_ro). This is
> + * useful when requesting a random index of an array containing ep_ro elements,
> + * for example. The result is somewhat biased when ep_ro is not a power of 2,
> + * so do not use this for cryptographic purposes.
>   *
>   * Returns: pseudo-random number in interval [0, ep_ro)
>   */
>  static inline u32 prandom_u32_max(u32 ep_ro)
>  {
> -       return (u32)(((u64) prandom_u32() * ep_ro) >> 32);
> +       if (__builtin_constant_p(ep_ro <= 1U << 8) && ep_ro <= 1U << 8)
> +               return (get_random_u8() * ep_ro) >> 8;
> +       if (__builtin_constant_p(ep_ro <= 1U << 16) && ep_ro <= 1U << 16)
> +               return (get_random_u16() * ep_ro) >> 16;
> +       return ((u64)get_random_u32() * ep_ro) >> 32;
>  }
>
>  /*
> --
> 2.37.3
>