Linux Virtualization - Mar 2014 - [PATCH v6 05/11] pvqspinlock, x86: Allow unfair spinlock in a PV guest

On Thu, Mar 13, 2014 at 04:05:19PM -0400, Waiman Long
wrote:
> On 03/13/2014 11:15 AM, Peter Zijlstra wrote:
> >On Wed, Mar 12, 2014 at 02:54:52PM -0400, Waiman Long wrote:
> >>+static inline void arch_spin_lock(struct qspinlock *lock)
> >>+{
> >>+	if (static_key_false(&paravirt_unfairlocks_enabled))
> >>+		queue_spin_lock_unfair(lock);
> >>+	else
> >>+		queue_spin_lock(lock);
> >>+}
> >So I would have expected something like:
> >
> >	if (static_key_false(&paravirt_spinlock)) {
> >		while (!queue_spin_trylock(lock))
> >			cpu_relax();
> >		return;
> >	}
> >
> >At the top of queue_spin_lock_slowpath().
> 
> I don't like the idea of constantly spinning on the lock. That can
cause all
> sort of performance issues.
Its bloody virt; _that_ is a performance issue to begin with.

Anybody half sane stops using virt (esp. if they care about
performance).
> My version of the unfair lock tries to grab the
> lock ignoring if there are others waiting in the queue or not. So instead
of
> the doing a cmpxchg of the whole 32-bit word, I just do a cmpxchg of the
> lock byte in the unfair version. A CPU has only one chance to steal the
> lock. If it can't, it will be lined up in the queue just like the fair
> version. It is not as unfair as the other unfair locking schemes that spins
> on the lock repetitively. So lock starvation should be less a problem.
> 
> On the other hand, it may not perform as well as the other unfair locking
> schemes. It is a compromise to provide some lock unfairness without
> sacrificing the good cacheline behavior of the queue spinlock.
But but but,.. any kind of queueing gets you into a world of hurt with
virt.

The simple test-and-set lock (as per the above) still sucks due to lock
holder preemption, but at least the suckage doesn't queue. Because with
queueing you not only have to worry about the lock holder getting
preemption, but also the waiter(s).

Take the situation of 3 (v)CPUs where cpu0 holds the lock but is
preempted. cpu1 queues, cpu2 queues. Then cpu1 gets preempted, after
which cpu0 gets back online.

The simple test-and-set lock will now let cpu2 acquire. Your queue
however will just sit there spinning, waiting for cpu1 to come back from
holiday.

I think you're way over engineering this. Just do the simple
test-and-set lock for virt && !paravirt (as I think Paolo Bonzini
suggested RHEL6 already does).

Il 14/03/2014 09:30, Peter Zijlstra ha scritto:
> Take the situation of 3 (v)CPUs where cpu0 holds the lock but is
> preempted. cpu1 queues, cpu2 queues. Then cpu1 gets preempted, after
> which cpu0 gets back online.
>
> The simple test-and-set lock will now let cpu2 acquire. Your queue
> however will just sit there spinning, waiting for cpu1 to come back from
> holiday.
>
> I think you're way over engineering this. Just do the simple
> test-and-set lock for virt && !paravirt (as I think Paolo Bonzini
> suggested RHEL6 already does).
Exactly.

Paolo

On 03/14/2014 04:30 AM, Peter Zijlstra wrote:
> On Thu, Mar 13, 2014 at 04:05:19PM -0400, Waiman Long wrote:
>> On 03/13/2014 11:15 AM, Peter Zijlstra wrote:
>>> On Wed, Mar 12, 2014 at 02:54:52PM -0400, Waiman Long wrote:
>>>> +static inline void arch_spin_lock(struct qspinlock *lock)
>>>> +{
>>>> +	if (static_key_false(&paravirt_unfairlocks_enabled))
>>>> +		queue_spin_lock_unfair(lock);
>>>> +	else
>>>> +		queue_spin_lock(lock);
>>>> +}
>>> So I would have expected something like:
>>>
>>> 	if (static_key_false(&paravirt_spinlock)) {
>>> 		while (!queue_spin_trylock(lock))
>>> 			cpu_relax();
>>> 		return;
>>> 	}
>>>
>>> At the top of queue_spin_lock_slowpath().
>> I don't like the idea of constantly spinning on the lock. That can
cause all
>> sort of performance issues.
> Its bloody virt; _that_ is a performance issue to begin with.
>
> Anybody half sane stops using virt (esp. if they care about
> performance).
>
>> My version of the unfair lock tries to grab the
>> lock ignoring if there are others waiting in the queue or not. So
instead of
>> the doing a cmpxchg of the whole 32-bit word, I just do a cmpxchg of
the
>> lock byte in the unfair version. A CPU has only one chance to steal the
>> lock. If it can't, it will be lined up in the queue just like the
fair
>> version. It is not as unfair as the other unfair locking schemes that
spins
>> on the lock repetitively. So lock starvation should be less a problem.
>>
>> On the other hand, it may not perform as well as the other unfair
locking
>> schemes. It is a compromise to provide some lock unfairness without
>> sacrificing the good cacheline behavior of the queue spinlock.
> But but but,.. any kind of queueing gets you into a world of hurt with
> virt.
>
> The simple test-and-set lock (as per the above) still sucks due to lock
> holder preemption, but at least the suckage doesn't queue. Because with
> queueing you not only have to worry about the lock holder getting
> preemption, but also the waiter(s).
>
> Take the situation of 3 (v)CPUs where cpu0 holds the lock but is
> preempted. cpu1 queues, cpu2 queues. Then cpu1 gets preempted, after
> which cpu0 gets back online.
>
> The simple test-and-set lock will now let cpu2 acquire. Your queue
> however will just sit there spinning, waiting for cpu1 to come back from
> holiday.
>
> I think you're way over engineering this. Just do the simple
> test-and-set lock for virt&&  !paravirt (as I think Paolo Bonzini
> suggested RHEL6 already does).
The PV ticketlock code was designed to handle lock holder preemption by 
redirecting CPU resources in a preempted guest to another guest that can 
better use it and then return the preempted CPU back sooner.

Using a simple test-and-set lock will not allow us to enable this PV 
spinlock functionality as there is no structure to decide who does what. 
I can extend the current unfair lock code to allow those waiting in the 
queue to also attempt to steal the lock, though at a lesser frequency so 
that the queue head has a higher chance of getting the lock. This will 
solve the lock waiter preemption problem that you worry about. This does 
make the code a bit more complex, but it allow us to enable both the 
unfair lock and the PV spinlock code together to solve the lock waiter 
and lock holder preemption problems.

-Longman

On Mon, Mar 17, 2014 at 01:44:34PM -0400, Waiman Long
wrote:
> The PV ticketlock code was designed to handle lock holder preemption by
> redirecting CPU resources in a preempted guest to another guest that can
> better use it and then return the preempted CPU back sooner.
But that's the PV code, not the unfair bit. And your fuller PV thing
doesn't need the unfair option.

On Mon, Mar 17, 2014 at 01:44:34PM -0400, Waiman Long
wrote:
> On 03/14/2014 04:30 AM, Peter Zijlstra wrote:
> >On Thu, Mar 13, 2014 at 04:05:19PM -0400, Waiman Long wrote:
> >>On 03/13/2014 11:15 AM, Peter Zijlstra wrote:
> >>>On Wed, Mar 12, 2014 at 02:54:52PM -0400, Waiman Long wrote:
> >>>>+static inline void arch_spin_lock(struct qspinlock *lock)
> >>>>+{
> >>>>+	if (static_key_false(&paravirt_unfairlocks_enabled))
> >>>>+		queue_spin_lock_unfair(lock);
> >>>>+	else
> >>>>+		queue_spin_lock(lock);
> >>>>+}
> >>>So I would have expected something like:
> >>>
> >>>	if (static_key_false(&paravirt_spinlock)) {
> >>>		while (!queue_spin_trylock(lock))
> >>>			cpu_relax();
> >>>		return;
> >>>	}
> >>>
> >>>At the top of queue_spin_lock_slowpath().
> >>I don't like the idea of constantly spinning on the lock. That
can cause all
> >>sort of performance issues.
> >Its bloody virt; _that_ is a performance issue to begin with.
> >
> >Anybody half sane stops using virt (esp. if they care about
> >performance).
> >
> >>My version of the unfair lock tries to grab the
> >>lock ignoring if there are others waiting in the queue or not. So
instead of
> >>the doing a cmpxchg of the whole 32-bit word, I just do a cmpxchg
of the
> >>lock byte in the unfair version. A CPU has only one chance to steal
the
> >>lock. If it can't, it will be lined up in the queue just like
the fair
> >>version. It is not as unfair as the other unfair locking schemes
that spins
> >>on the lock repetitively. So lock starvation should be less a
problem.
> >>
> >>On the other hand, it may not perform as well as the other unfair
locking
> >>schemes. It is a compromise to provide some lock unfairness without
> >>sacrificing the good cacheline behavior of the queue spinlock.
> >But but but,.. any kind of queueing gets you into a world of hurt with
> >virt.
> >
> >The simple test-and-set lock (as per the above) still sucks due to lock
> >holder preemption, but at least the suckage doesn't queue. Because
with
> >queueing you not only have to worry about the lock holder getting
> >preemption, but also the waiter(s).
> >
> >Take the situation of 3 (v)CPUs where cpu0 holds the lock but is
> >preempted. cpu1 queues, cpu2 queues. Then cpu1 gets preempted, after
> >which cpu0 gets back online.
> >
> >The simple test-and-set lock will now let cpu2 acquire. Your queue
> >however will just sit there spinning, waiting for cpu1 to come back
from
> >holiday.
> >
> >I think you're way over engineering this. Just do the simple
> >test-and-set lock for virt&&  !paravirt (as I think Paolo
Bonzini
> >suggested RHEL6 already does).
> 
> The PV ticketlock code was designed to handle lock holder preemption
> by redirecting CPU resources in a preempted guest to another guest
> that can better use it and then return the preempted CPU back
> sooner.
> 
> Using a simple test-and-set lock will not allow us to enable this PV
> spinlock functionality as there is no structure to decide who does
> what. I can extend the current unfair lock code to allow those
And what would be needed to do 'decide who does what'?
> waiting in the queue to also attempt to steal the lock, though at a
> lesser frequency so that the queue head has a higher chance of
> getting the lock. This will solve the lock waiter preemption problem
> that you worry about. This does make the code a bit more complex,
> but it allow us to enable both the unfair lock and the PV spinlock
> code together to solve the lock waiter and lock holder preemption
> problems.
> 
> -Longman
>