once: add DO_ONCE_SLOW() for sleepable contexts

Christophe Leroy reported a ~80ms latency spike
happening at first TCP connect() time.

This is because __inet_hash_connect() uses get_random_once()
to populate a perturbation table which became quite big
after commit 4c2c8f0 ("tcp: increase source port perturb table to 2^16")

get_random_once() uses DO_ONCE(), which block hard irqs for the duration
of the operation.

This patch adds DO_ONCE_SLOW() which uses a mutex instead of a spinlock
for operations where we prefer to stay in process context.

Then __inet_hash_connect() can use get_random_slow_once()
to populate its perturbation table.

Fixes: 4c2c8f0 ("tcp: increase source port perturb table to 2^16")
Fixes: 190cc82 ("tcp: change source port randomizarion at connect() time")
Reported-by: Christophe Leroy <christophe.leroy@csgroup.eu>
Link: https://lore.kernel.org/netdev/CANn89iLAEYBaoYajy0Y9UmGFff5GPxDUoG-ErVB2jDdRNQ5Tug@mail.gmail.com/T/#t
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Willy Tarreau <w@1wt.eu>
Tested-by: Christophe Leroy <christophe.leroy@csgroup.eu>
Signed-off-by: David S. Miller <davem@davemloft.net>

include/linux/once.h

@@ -5,10 +5,18 @@
 #include <linux/types.h>
 #include <linux/jump_label.h>
 
+/* Helpers used from arbitrary contexts.
+ * Hard irqs are blocked, be cautious.
+ */
 bool __do_once_start(bool *done, unsigned long *flags);
 void __do_once_done(bool *done, struct static_key_true *once_key,
 		    unsigned long *flags, struct module *mod);
 
+/* Variant for process contexts only. */
+bool __do_once_slow_start(bool *done);
+void __do_once_slow_done(bool *done, struct static_key_true *once_key,
+			 struct module *mod);
+
 /* Call a function exactly once. The idea of DO_ONCE() is to perform
  * a function call such as initialization of random seeds, etc, only
  * once, where DO_ONCE() can live in the fast-path. After @func has
@@ -52,7 +60,27 @@ void __do_once_done(bool *done, struct static_key_true *once_key,
 		___ret;							     \
 	})
 
+/* Variant of DO_ONCE() for process/sleepable contexts. */
+#define DO_ONCE_SLOW(func, ...)						     \
+	({								     \
+		bool ___ret = false;					     \
+		static bool __section(".data.once") ___done = false;	     \
+		static DEFINE_STATIC_KEY_TRUE(___once_key);		     \
+		if (static_branch_unlikely(&___once_key)) {		     \
+			___ret = __do_once_slow_start(&___done);	     \
+			if (unlikely(___ret)) {				     \
+				func(__VA_ARGS__);			     \
+				__do_once_slow_done(&___done, &___once_key,  \
+						    THIS_MODULE);	     \
+			}						     \
+		}							     \
+		___ret;							     \
+	})
+
 #define get_random_once(buf, nbytes)					     \
 	DO_ONCE(get_random_bytes, (buf), (nbytes))
 
+#define get_random_slow_once(buf, nbytes)				     \
+	DO_ONCE_SLOW(get_random_bytes, (buf), (nbytes))
+
 #endif /* _LINUX_ONCE_H */

lib/once.c

@@ -66,3 +66,33 @@ void __do_once_done(bool *done, struct static_key_true *once_key,
 	once_disable_jump(once_key, mod);
 }
 EXPORT_SYMBOL(__do_once_done);
+
+static DEFINE_MUTEX(once_mutex);
+
+bool __do_once_slow_start(bool *done)
+	__acquires(once_mutex)
+{
+	mutex_lock(&once_mutex);
+	if (*done) {
+		mutex_unlock(&once_mutex);
+		/* Keep sparse happy by restoring an even lock count on
+		 * this mutex. In case we return here, we don't call into
+		 * __do_once_done but return early in the DO_ONCE_SLOW() macro.
+		 */
+		__acquire(once_mutex);
+		return false;
+	}
+
+	return true;
+}
+EXPORT_SYMBOL(__do_once_slow_start);
+
+void __do_once_slow_done(bool *done, struct static_key_true *once_key,
+			 struct module *mod)
+	__releases(once_mutex)
+{
+	*done = true;
+	mutex_unlock(&once_mutex);
+	once_disable_jump(once_key, mod);
+}
+EXPORT_SYMBOL(__do_once_slow_done);

net/ipv4/inet_hashtables.c

@@ -958,8 +958,8 @@ int __inet_hash_connect(struct inet_timewait_death_row *death_row,
 	if (likely(remaining > 1))
 		remaining &= ~1U;
 
-	net_get_random_once(table_perturb,
-			    INET_TABLE_PERTURB_SIZE * sizeof(*table_perturb));
+	get_random_slow_once(table_perturb,
+			     INET_TABLE_PERTURB_SIZE * sizeof(*table_perturb));
 	index = port_offset & (INET_TABLE_PERTURB_SIZE - 1);
 
 	offset = READ_ONCE(table_perturb[index]) + (port_offset >> 32);