On Mon, 01 Feb 2016, Peter Zijlstra wrote:
Subject: locking/mutex: Avoid spinner vs waiter starvation
From: Peter Zijlstra <peterz@xxxxxxxxxxxxx>
Date: Fri, 22 Jan 2016 12:06:53 +0100
Ding Tianhong reported that under his load the optimistic spinners
would totally starve a task that ended up on the wait list.
Fix this by ensuring the top waiter also partakes in the optimistic
spin queue.
There are a few subtle differences between the assumed state of
regular optimistic spinners and those already on the wait list, which
result in the @acquired complication of the acquire path.
Most notable are:
- waiters are on the wait list and need to be taken off
- mutex_optimistic_spin() sets the lock->count to 0 on acquire
even though there might be more tasks on the wait list.
Right, the main impact I see with these complications are that the
window of when a waiter takes the lock via spinning and then acquires
the wait_lock to remove itself from the list, will allow an unlock
thread to set the lock as available in the fastpath which could in
turn allow a third thread the steal the lock. With high contention,
this window will be come obviously larger as we contend for the
wait_lock.
CPU-0 CPU-1 CPU-3
__mutex_lock_common mutex_optimistic_spin
(->count now 0)
__mutex_fastpath_unlock
(->count now 1) __mutex_fastpath_lock
(stolen)
spin_lock_mutex(&lock->wait_lock, flags);
But we've always been bad when it comes to counter and waiters.
Thanks,
Davidlohr