Re: [PATCH v2 2/5] locking/pvqspinlock: Fix missed PV wakeup problem

From: Pan Xinhui
Date: Fri Jul 15 2016 - 05:39:36 EST




å 16/7/15 16:47, Peter Zijlstra åé:

So the reason I never get around to this is because the patch stinks.

It simply doesn't make sense... Remember, the harder you make a reviewer
work the less likely the review will be done.

Present things in clear concise language and draw a picture.

On Tue, May 31, 2016 at 12:53:48PM -0400, Waiman Long wrote:
Currently, calling pv_hash() and setting _Q_SLOW_VAL is only
done once for any pv_node. It is either in pv_kick_node() or in
pv_wait_head_or_lock().

So far so good....

Because of lock stealing, a pv_kick'ed node is
not guaranteed to get the lock before the spinning threshold expires
and has to call pv_wait() again. As a result, the new lock holder
won't see _Q_SLOW_VAL and so won't wake up the sleeping vCPU.

waiman, it might be "as a result, the head node will not really enter wait state because ->locked is not_Q_SLOW_VAL, the pv_wait will return directly."

*brain melts* what!? pv_kick'ed node reads like pv_kick_node() and that
doesn't make any kind of sense.

I'm thinking you're trying to say this:


CPU0 CPU1 CPU2

__pv_queued_spin_unlock_slowpath()
...
smp_store_release(&l->locked, 0);
__pv_queued_spin_lock_slowpath()
...
pv_queued_spin_steal_lock()
cmpxchg(&l->locked, 0, _Q_LOCKED_VAL) == 0


pv_wait_head_or_lock()

pv_kick(node->cpu); ----------------------> pv_wait(&l->locked, _Q_SLOW_VAL);

__pv_queued_spin_unlock()
cmpxchg(&l->locked, _Q_LOCKED_VAL, 0) == _Q_LOCKED_VAL

for () {
trylock_clear_pending();
cpu_relax();
}

pv_wait(&l->locked, _Q_SLOW_VAL);


Which is indeed 'bad', but not fatal, note that the later pv_wait() will
not in fact go wait, since l->locked will _not_ be _Q_SLOW_VAL.

hi, Peter

the problem is that "this later pv_wait will do nothing as l->locked is not _Q_SLOW_VAL",
So it is not paravirt friendly then. we will go into the trylock loop again and again until the lock is unlocked.

So if we are kicked by the unlock_slowpath, and the lock is stealed by someone else, we need hash its node again and set l->locked to _Q_SLOW_VAL, then enter pv_wait.

but I am worried about lock stealing. could the node in the queue starve for a long time? I notice the latency of pv_wait on an over-commited guest can be bigger than 300us. I have not seen such starving case, but I think it is possible to happen.

thanks
xinhui

Is this indeed the 3 CPU scenario you tried to describe in a scant 4
lines of text, or is there more to it?