Message-ID: <CAHmME9r9y+A89+=ErYDYOfRkGD2=GZaik57PMaV_MHwKzJwDzw@mail.gmail.com>
Date:   Mon, 27 Feb 2023 20:00:51 +0100
From:   "Jason A. Donenfeld" <Jason@...c4.com>
To:     Eric Biggers <ebiggers@...nel.org>
Cc:     linux-crypto@...r.kernel.org,
        Herbert Xu <herbert@...dor.apana.org.au>,
        linux-kernel@...r.kernel.org, Yann Droneaud <ydroneaud@...eya.com>,
        stable@...r.kernel.org
Subject: Re: [PATCH v2] crypto: testmgr - fix RNG performance in fuzz tests

On Mon, Feb 27, 2023 at 7:30 PM Eric Biggers <ebiggers@...nel.org> wrote:
>
> From: Eric Biggers <ebiggers@...gle.com>
>
> The performance of the crypto fuzz tests has greatly regressed since
> v5.18.  When booting a kernel on an arm64 dev board with all software
> crypto algorithms and CONFIG_CRYPTO_MANAGER_EXTRA_TESTS enabled, the
> fuzz tests now take about 200 seconds to run, or about 325 seconds with
> lockdep enabled, compared to about 5 seconds before.
>
> The root cause is that the random number generation has become much
> slower due to commit d4150779e60f ("random32: use real rng for
> non-deterministic randomness").  On my same arm64 dev board, at the time
> the fuzz tests are run, get_random_u8() is about 345x slower than
> prandom_u32_state(), or about 469x if lockdep is enabled.
>
> Lockdep makes a big difference, but much of the rest comes from the
> get_random_*() functions taking a *very* slow path when the CRNG is not
> yet initialized.  Since the crypto self-tests run early during boot,
> even having a hardware RNG driver enabled (CONFIG_CRYPTO_DEV_QCOM_RNG in
> my case) doesn't prevent this.  x86 systems don't have this issue, but
> they still see a significant regression if lockdep is enabled.
>
> Converting the "Fully random bytes" case in generate_random_bytes() to
> use get_random_bytes() helps significantly, improving the test time to
> about 27 seconds.  But that's still over 5x slower than before.
>
> This is all a bit silly, though, since the fuzz tests don't actually
> need cryptographically secure random numbers.  So let's just make them
> use a non-cryptographically-secure RNG as they did before.  The original
> prandom_u32() is gone now, so let's use prandom_u32_state() instead,
> with an explicitly managed state, like various other self-tests in the
> kernel source tree (rbtree_test.c, test_scanf.c, etc.) already do.  This
> also has the benefit that no locking is required anymore, so performance
> should be even better than the original version that used prandom_u32().
>
> Fixes: d4150779e60f ("random32: use real rng for non-deterministic randomness")
> Cc: stable@...r.kernel.org
> Signed-off-by: Eric Biggers <ebiggers@...gle.com>
> ---
>
> v2: made init_rnd_state() use get_random_u64()
>
>  crypto/testmgr.c | 266 ++++++++++++++++++++++++++++++-----------------
>  1 file changed, 169 insertions(+), 97 deletions(-)
>
> diff --git a/crypto/testmgr.c b/crypto/testmgr.c
> index c91e93ece20b..b160eeb12c8e 100644
> --- a/crypto/testmgr.c
> +++ b/crypto/testmgr.c
> @@ -860,12 +860,50 @@ static int prepare_keybuf(const u8 *key, unsigned int ksize,
>
>  #ifdef CONFIG_CRYPTO_MANAGER_EXTRA_TESTS
>
> +/*
> + * The fuzz tests use prandom instead of the normal Linux RNG since they don't
> + * need cryptographically secure random numbers.  This greatly improves the
> + * performance of these tests, especially if they are run before the Linux RNG
> + * has been initialized or if they are run on a lockdep-enabled kernel.
> + */
> +
> +static inline void init_rnd_state(struct rnd_state *rng)
> +{
> +       prandom_seed_state(rng, get_random_u64());

i915 does something similar and prints it out with `kunit_info(suite,
"Testing DRM buddy manager, with random_seed=0x%x\n", random_seed);`,
so that you can repeat the test if necessary. Not saying you have to
do this now, but it may be a cool feature to keep in mind for the
future.

Jason

Powered by blists - more mailing lists