random: use a different mixing algorithm for add_device_randomness()

add_device_randomness() use of crng_fast_load() was highly
problematic.  Some callers of add_device_randomness() can pass in a
large amount of static information.  This would immediately promote
the crng_init state from 0 to 1, without really doing much to
initialize the primary_crng's internal state with something even
vaguely unpredictable.

Since we don't have the speed constraints of add_interrupt_randomness(),
we can do a better job mixing in the what unpredictability a device
driver or architecture maintainer might see fit to give us, and do it
in a way which does not bump the crng_init_cnt variable.

Also, since add_device_randomness() doesn't bump any entropy
accounting in crng_init state 0, mix the device randomness into the
input_pool entropy pool as well.  This is related to CVE-2018-1108.
Reported-by: Jann Horn <jannh@google.com>
Fixes: ee7998c5 ("random: do not ignore early device randomness")
Cc: stable@kernel.org # 4.13+
Signed-off-by: Theodore Ts'o <tytso@mit.edu>

drivers/char/random.c

@@ -787,6 +787,10 @@ static void crng_initialize(struct crng_state *crng)
 	crng->init_time = jiffies - CRNG_RESEED_INTERVAL - 1;
 }
 
+/*
+ * crng_fast_load() can be called by code in the interrupt service
+ * path.  So we can't afford to dilly-dally.
+ */
 static int crng_fast_load(const char *cp, size_t len)
 {
 	unsigned long flags;
@@ -813,6 +817,51 @@ static int crng_fast_load(const char *cp, size_t len)
 	return 1;
 }
 
+/*
+ * crng_slow_load() is called by add_device_randomness, which has two
+ * attributes.  (1) We can't trust the buffer passed to it is
+ * guaranteed to be unpredictable (so it might not have any entropy at
+ * all), and (2) it doesn't have the performance constraints of
+ * crng_fast_load().
+ *
+ * So we do something more comprehensive which is guaranteed to touch
+ * all of the primary_crng's state, and which uses a LFSR with a
+ * period of 255 as part of the mixing algorithm.  Finally, we do
+ * *not* advance crng_init_cnt since buffer we may get may be something
+ * like a fixed DMI table (for example), which might very well be
+ * unique to the machine, but is otherwise unvarying.
+ */
+static int crng_slow_load(const char *cp, size_t len)
+{
+	unsigned long		flags;
+	static unsigned char	lfsr = 1;
+	unsigned char		tmp;
+	unsigned		i, max = CHACHA20_KEY_SIZE;
+	const char *		src_buf = cp;
+	char *			dest_buf = (char *) &primary_crng.state[4];
+
+	if (!spin_trylock_irqsave(&primary_crng.lock, flags))
+		return 0;
+	if (crng_init != 0) {
+		spin_unlock_irqrestore(&primary_crng.lock, flags);
+		return 0;
+	}
+	if (len > max)
+		max = len;
+
+	for (i = 0; i < max ; i++) {
+		tmp = lfsr;
+		lfsr >>= 1;
+		if (tmp & 1)
+			lfsr ^= 0xE1;
+		tmp = dest_buf[i % CHACHA20_KEY_SIZE];
+		dest_buf[i % CHACHA20_KEY_SIZE] ^= src_buf[i % len] ^ lfsr;
+		lfsr += (tmp << 3) | (tmp >> 5);
+	}
+	spin_unlock_irqrestore(&primary_crng.lock, flags);
+	return 1;
+}
+
 static void crng_reseed(struct crng_state *crng, struct entropy_store *r)
 {
 	unsigned long	flags;
@@ -981,10 +1030,8 @@ void add_device_randomness(const void *buf, unsigned int size)
 	unsigned long time = random_get_entropy() ^ jiffies;
 	unsigned long flags;
 
-	if (!crng_ready()) {
-		crng_fast_load(buf, size);
-		return;
-	}
+	if (!crng_ready() && size)
+		crng_slow_load(buf, size);
 
 	trace_add_device_randomness(size, _RET_IP_);
 	spin_lock_irqsave(&input_pool.lock, flags);