crypto: api - Fix algorithm test race that broke aead initialisation

When we complete a test we'll notify everyone waiting on it, drop
the mutex, and then remove the test larval (after reacquiring the
mutex).  If one of the notified parties tries to register another
algorithm with the same driver name prior to the removal of the
test larval, they will fail with EEXIST as only one algorithm of
a given name can be tested at any time.

This broke the initialisation of aead and givcipher algorithms as
they will register two algorithms with the same driver name, in
sequence.

This patch fixes the problem by marking the larval as dead before
we drop the mutex, and also ignoring all dead or dying algorithms
on the registration path.

Tested-by: Andreas Steffen <andreas.steffen@strongswan.org>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>

crypto/algapi.c

@@ -149,6 +149,9 @@ static struct crypto_larval *__crypto_register_alg(struct crypto_alg *alg)
 		if (q == alg)
 			goto err;
 
+		if (crypto_is_moribund(q))
+			continue;
+
 		if (crypto_is_larval(q)) {
 			if (!strcmp(alg->cra_driver_name, q->cra_driver_name))
 				goto err;
@@ -197,7 +200,7 @@ void crypto_alg_tested(const char *name, int err)
 
 	down_write(&crypto_alg_sem);
 	list_for_each_entry(q, &crypto_alg_list, cra_list) {
-		if (!crypto_is_larval(q))
+		if (crypto_is_moribund(q) || !crypto_is_larval(q))
 			continue;
 
 		test = (struct crypto_larval *)q;
@@ -210,6 +213,7 @@ void crypto_alg_tested(const char *name, int err)
 	goto unlock;
 
 found:
+	q->cra_flags |= CRYPTO_ALG_DEAD;
 	alg = test->adult;
 	if (err || list_empty(&alg->cra_list))
 		goto complete;