[PATCH v2] random: defer fast pool mixing to worker

[Jason A. Donenfeld]

On PREEMPT_RT, it's problematic to take spinlocks from hard irq
handlers. We can fix this by deferring to a work queue the dumping of
the fast pool into the input pool.

We accomplish this with some careful rules on fast_pool->count:

  - When it's incremented to >= 64, we schedule the work.
  - If the top bit is set, we never schedule the work, even if >= 64.
  - The worker is responsible for setting it back to 0 when it's done.

In the worst case, an irq handler is mixing a new irq into the pool at
the same time as the worker is dumping it into the input pool. In this
case, we only ever set the count back to 0 _after_ we're done, so that
subsequent cycles will require a full 64 to dump it in again. In other
words, the result of this race is only ever adding a little bit more
information than normal, but never less, and never crediting any more
for this partial additional information.

Note that this doesn't deal with the spinlocks in crng_fast_load(),
which will have to be dealt with some other way.

Cc: Sebastian Andrzej Siewior <bigeasy@xxxxxxxxxxxxx>
Cc: Thomas Gleixner <tglx@xxxxxxxxxxxxx>
Cc: Peter Zijlstra <peterz@xxxxxxxxxxxxx>
Cc: Theodore Ts'o <tytso@xxxxxxx>
Cc: Sultan Alsawaf <sultan@xxxxxxxxxxxxxxx>
Cc: Jonathan Neuschäfer <j.neuschaefer@xxxxxxx>
Signed-off-by: Jason A. Donenfeld <Jason@xxxxxxxxx>
---
 drivers/char/random.c         | 54 ++++++++++++++++++++---------------
 include/trace/events/random.h |  6 ----
 2 files changed, 31 insertions(+), 29 deletions(-)

diff --git a/drivers/char/random.c b/drivers/char/random.c
index 5d7d6e01bbc4..575616de2e16 100644
--- a/drivers/char/random.c
+++ b/drivers/char/random.c
@@ -383,12 +383,6 @@ static void _mix_pool_bytes(const void *in, int nbytes)
 	blake2s_update(&input_pool.hash, in, nbytes);
 }
 
-static void __mix_pool_bytes(const void *in, int nbytes)
-{
-	trace_mix_pool_bytes_nolock(nbytes, _RET_IP_);
-	_mix_pool_bytes(in, nbytes);
-}
-
 static void mix_pool_bytes(const void *in, int nbytes)
 {
 	unsigned long flags;
@@ -400,11 +394,13 @@ static void mix_pool_bytes(const void *in, int nbytes)
 }
 
 struct fast_pool {
-	u32 pool[4];
+	struct work_struct mix;
 	unsigned long last;
+	u32 pool[4];
+	unsigned int count;
 	u16 reg_idx;
-	u8 count;
 };
+#define FAST_POOL_MIX_INFLIGHT (1U << 31)
 
 /*
  * This is a fast mixing routine used by the interrupt randomness
@@ -434,7 +430,6 @@ static void fast_mix(struct fast_pool *f)
 
 	f->pool[0] = a;  f->pool[1] = b;
 	f->pool[2] = c;  f->pool[3] = d;
-	f->count++;
 }
 
 static void process_random_ready_list(void)
@@ -985,12 +980,30 @@ static u32 get_reg(struct fast_pool *f, struct pt_regs *regs)
 	return *ptr;
 }
 
+static void mix_interrupt_randomness(struct work_struct *work)
+{
+	struct fast_pool *fast_pool = container_of(work, struct fast_pool, mix);
+
+	/*
+	 * Since this is the result of a trip through the scheduler, xor in
+	 * a cycle counter. It can't hurt, and might help.
+	 */
+	fast_pool->pool[3] ^= random_get_entropy();
+
+	mix_pool_bytes(&fast_pool->pool, sizeof(fast_pool->pool));
+	/* We take care to zero out the count only after we're done reading the pool. */
+	WRITE_ONCE(fast_pool->count, 0);
+	fast_pool->last = jiffies;
+	credit_entropy_bits(1);
+}
+
 void add_interrupt_randomness(int irq)
 {
 	struct fast_pool *fast_pool = this_cpu_ptr(&irq_randomness);
 	struct pt_regs *regs = get_irq_regs();
 	unsigned long now = jiffies;
 	cycles_t cycles = random_get_entropy();
+	unsigned int new_count;
 	u32 c_high, j_high;
 	u64 ip;
 
@@ -1007,9 +1020,10 @@ void add_interrupt_randomness(int irq)
 
 	fast_mix(fast_pool);
 	add_interrupt_bench(cycles);
+	new_count = __this_cpu_inc_return(irq_randomness.count);
 
 	if (unlikely(crng_init == 0)) {
-		if ((fast_pool->count >= 64) &&
+		if (new_count >= 64 &&
 		    crng_fast_load((u8 *)fast_pool->pool, sizeof(fast_pool->pool)) > 0) {
 			fast_pool->count = 0;
 			fast_pool->last = now;
@@ -1017,20 +1031,14 @@ void add_interrupt_randomness(int irq)
 		return;
 	}
 
-	if ((fast_pool->count < 64) && !time_after(now, fast_pool->last + HZ))
-		return;
-
-	if (!spin_trylock(&input_pool.lock))
-		return;
-
-	fast_pool->last = now;
-	__mix_pool_bytes(&fast_pool->pool, sizeof(fast_pool->pool));
-	spin_unlock(&input_pool.lock);
-
-	fast_pool->count = 0;
+	if (new_count >= 64 && new_count < FAST_POOL_MIX_INFLIGHT &&
+	    time_after(now, fast_pool->last + HZ)) {
+		if (unlikely(!fast_pool->mix.func))
+			INIT_WORK(&fast_pool->mix, mix_interrupt_randomness);
+		__this_cpu_or(irq_randomness.count, FAST_POOL_MIX_INFLIGHT);
+		queue_work_on(raw_smp_processor_id(), system_highpri_wq, &fast_pool->mix);
 
-	/* award one bit for the contents of the fast pool */
-	credit_entropy_bits(1);
+	}
 }
 EXPORT_SYMBOL_GPL(add_interrupt_randomness);
 
diff --git a/include/trace/events/random.h b/include/trace/events/random.h
index ad149aeaf42c..833f42afc70f 100644
--- a/include/trace/events/random.h
+++ b/include/trace/events/random.h
@@ -52,12 +52,6 @@ DEFINE_EVENT(random__mix_pool_bytes, mix_pool_bytes,
 	TP_ARGS(bytes, IP)
 );
 
-DEFINE_EVENT(random__mix_pool_bytes, mix_pool_bytes_nolock,
-	TP_PROTO(int bytes, unsigned long IP),
-
-	TP_ARGS(bytes, IP)
-);
-
 TRACE_EVENT(credit_entropy_bits,
 	TP_PROTO(int bits, int entropy_count, unsigned long IP),
 
-- 
2.35.0