random: optimize add_interrupt_randomness

add_interrupt_randomess always wakes up
code blocking on /dev/random. This wake up is done
unconditionally. Unfortunately this means all interrupts
take the wait queue spinlock, which can be rather expensive
on large systems processing lots of interrupts.

We saw 1% cpu time spinning on this on a large macro workload
running on a large system.

I believe it's a recent regression (?)

Always check if there is a waiter on the wait queue
before waking up. This check can be done without
taking a spinlock.

1.06%         10460  [kernel.vmlinux] [k] native_queued_spin_lock_slowpath
         |
         ---native_queued_spin_lock_slowpath
            |
             --0.57%--_raw_spin_lock_irqsave
                       |
                        --0.56%--__wake_up_common_lock
                                  credit_entropy_bits
                                  add_interrupt_randomness
                                  handle_irq_event_percpu
                                  handle_irq_event
                                  handle_edge_irq
                                  handle_irq
                                  do_IRQ
                                  common_interrupt

Signed-off-by: Andi Kleen <ak@linux.intel.com>
Signed-off-by: Theodore Ts'o <tytso@mit.edu>
This commit is contained in:
Andi Kleen 2018-02-28 13:43:28 -08:00 committed by Theodore Ts'o
parent 25e3fca492
commit e8e8a2e47d

View File

@ -709,7 +709,8 @@ static void credit_entropy_bits(struct entropy_store *r, int nbits)
}
/* should we wake readers? */
if (entropy_bits >= random_read_wakeup_bits) {
if (entropy_bits >= random_read_wakeup_bits &&
wq_has_sleeper(&random_read_wait)) {
wake_up_interruptible(&random_read_wait);
kill_fasync(&fasync, SIGIO, POLL_IN);
}