From: Eric Biggers ebiggers@google.com
Currently, registering an algorithm with the crypto API always causes a notification to be posted to the "cryptomgr", which then creates a kthread to self-test the algorithm. However, if self-tests are disabled in the kconfig (as is the default option), then this kthread just notifies waiters that the algorithm has been tested, then exits.
This causes a significant amount of overhead, especially in the kthread creation and destruction, which is not necessary at all. For example, in a quick test I found that booting a "minimum" x86_64 kernel with all the crypto options enabled (except for the self-tests) takes about 400ms until PID 1 can start. Of that, a full 13ms is spent just doing this pointless dance, involving a kthread being created, run, and destroyed over 200 times. That's over 3% of the entire kernel start time.
Fix this by just skipping the creation of the test larval and the posting of the registration notification entirely, when self-tests are disabled. Also compile out the unnecessary code in algboss.c.
While this patch is an optimization and not a "fix" per se, I've marked it as for stable, due to the large improvement it can make to boot time.
Cc: stable@vger.kernel.org Signed-off-by: Eric Biggers ebiggers@google.com --- crypto/algapi.c | 3 ++- crypto/algboss.c | 11 +++++++---- 2 files changed, 9 insertions(+), 5 deletions(-)
diff --git a/crypto/algapi.c b/crypto/algapi.c index 5c69ff8e8fa5c..018935cb8417d 100644 --- a/crypto/algapi.c +++ b/crypto/algapi.c @@ -226,7 +226,8 @@ static struct crypto_larval *crypto_alloc_test_larval(struct crypto_alg *alg) { struct crypto_larval *larval;
- if (!IS_ENABLED(CONFIG_CRYPTO_MANAGER)) + if (!IS_ENABLED(CONFIG_CRYPTO_MANAGER) || + IS_ENABLED(CONFIG_CRYPTO_MANAGER_DISABLE_TESTS)) return NULL;
larval = crypto_larval_alloc(alg->cra_name, diff --git a/crypto/algboss.c b/crypto/algboss.c index eb5fe84efb83e..e6f0443d08048 100644 --- a/crypto/algboss.c +++ b/crypto/algboss.c @@ -171,16 +171,18 @@ static int cryptomgr_schedule_probe(struct crypto_larval *larval) return NOTIFY_OK; }
+#ifdef CONFIG_CRYPTO_MANAGER_DISABLE_TESTS +static int cryptomgr_schedule_test(struct crypto_alg *alg) +{ + return 0; +} +#else static int cryptomgr_test(void *data) { struct crypto_test_param *param = data; u32 type = param->type; int err = 0;
-#ifdef CONFIG_CRYPTO_MANAGER_DISABLE_TESTS - goto skiptest; -#endif - if (type & CRYPTO_ALG_TESTED) goto skiptest;
@@ -229,6 +231,7 @@ static int cryptomgr_schedule_test(struct crypto_alg *alg) err: return NOTIFY_OK; } +#endif /* !CONFIG_CRYPTO_MANAGER_DISABLE_TESTS */
static int cryptomgr_notify(struct notifier_block *this, unsigned long msg, void *data)
base-commit: f67dd6ce0723ad013395f20a3f79d8a437d3f455
-----Original Message----- From: Eric Biggers ebiggers@kernel.org Sent: Wednesday, November 9, 2022 8:38 PM Subject: [PATCH] crypto: avoid unnecessary work when self-tests are disabled
Currently, registering an algorithm with the crypto API always causes a notification to be posted to the "cryptomgr", which then creates a kthread to self-test the algorithm. However, if self-tests are disabled in the kconfig (as is the default option), then this kthread just notifies waiters that the algorithm has been tested, then exits.
This causes a significant amount of overhead, especially in the kthread creation and destruction, which is not necessary at all. For example, in a quick test I found that booting a "minimum" x86_64 kernel with all the crypto options enabled (except for the self-tests) takes about 400ms until PID 1 can start. Of that, a full 13ms is spent just doing this pointless dance, involving a kthread being created, run, and destroyed over 200 times. That's over 3% of the entire kernel start time.
Fix this by just skipping the creation of the test larval and the posting of the registration notification entirely, when self-tests are disabled. Also compile out the unnecessary code in algboss.c.
...
+#ifdef CONFIG_CRYPTO_MANAGER_DISABLE_TESTS +static int cryptomgr_schedule_test(struct crypto_alg *alg) +{
- return 0;
+} +#else
The crypto/kdf_sp800108.c init function currently ignores both CONFIG_CRYPTO_MANAGER_DISABLE_TESTS and the cryptomgr module's notests module parameter and always runs its self-test, as described in https://lore.kernel.org/lkml/MW5PR84MB1842811C4EECC0F4B35B5FB3AB709@MW5PR84M...
Paul reported that taking 262 ms on his system; I measured 1.4 s on my system.
It'd be nice if a patch series improving how DISABLE_TESTS is honored would tackle that module too.
On Thu, Nov 10, 2022 at 04:36:17AM +0000, Elliott, Robert (Servers) wrote:
-----Original Message----- From: Eric Biggers ebiggers@kernel.org Sent: Wednesday, November 9, 2022 8:38 PM Subject: [PATCH] crypto: avoid unnecessary work when self-tests are disabled
Currently, registering an algorithm with the crypto API always causes a notification to be posted to the "cryptomgr", which then creates a kthread to self-test the algorithm. However, if self-tests are disabled in the kconfig (as is the default option), then this kthread just notifies waiters that the algorithm has been tested, then exits.
This causes a significant amount of overhead, especially in the kthread creation and destruction, which is not necessary at all. For example, in a quick test I found that booting a "minimum" x86_64 kernel with all the crypto options enabled (except for the self-tests) takes about 400ms until PID 1 can start. Of that, a full 13ms is spent just doing this pointless dance, involving a kthread being created, run, and destroyed over 200 times. That's over 3% of the entire kernel start time.
Fix this by just skipping the creation of the test larval and the posting of the registration notification entirely, when self-tests are disabled. Also compile out the unnecessary code in algboss.c.
...
+#ifdef CONFIG_CRYPTO_MANAGER_DISABLE_TESTS +static int cryptomgr_schedule_test(struct crypto_alg *alg) +{
- return 0;
+} +#else
The crypto/kdf_sp800108.c init function currently ignores both CONFIG_CRYPTO_MANAGER_DISABLE_TESTS and the cryptomgr module's notests module parameter and always runs its self-test, as described in https://lore.kernel.org/lkml/MW5PR84MB1842811C4EECC0F4B35B5FB3AB709@MW5PR84M...
Paul reported that taking 262 ms on his system; I measured 1.4 s on my system.
It'd be nice if a patch series improving how DISABLE_TESTS is honored would tackle that module too.
That should be a separate patch, but yes, it should only run the test if !IS_ENABLED(CONFIG_CRYPTO_MANAGER_DISABLE_TESTS), like what everywhere else does.
- Eric
On Wed, Nov 09, 2022 at 06:37:38PM -0800, Eric Biggers wrote:
From: Eric Biggers ebiggers@google.com
Currently, registering an algorithm with the crypto API always causes a notification to be posted to the "cryptomgr", which then creates a kthread to self-test the algorithm. However, if self-tests are disabled in the kconfig (as is the default option), then this kthread just notifies waiters that the algorithm has been tested, then exits.
This causes a significant amount of overhead, especially in the kthread creation and destruction, which is not necessary at all. For example, in a quick test I found that booting a "minimum" x86_64 kernel with all the crypto options enabled (except for the self-tests) takes about 400ms until PID 1 can start. Of that, a full 13ms is spent just doing this pointless dance, involving a kthread being created, run, and destroyed over 200 times. That's over 3% of the entire kernel start time.
Fix this by just skipping the creation of the test larval and the posting of the registration notification entirely, when self-tests are disabled. Also compile out the unnecessary code in algboss.c.
While this patch is an optimization and not a "fix" per se, I've marked it as for stable, due to the large improvement it can make to boot time.
Cc: stable@vger.kernel.org Signed-off-by: Eric Biggers ebiggers@google.com
Unfortunately, this patch won't work because it breaks templates.
There should still be a solution that at least avoids having to spawn kthreads, though...
- Eric
linux-stable-mirror@lists.linaro.org
-
Elliott, Robert (Servers) -
Eric Biggers