Re: [PATCH RFC V11 15/18] kvm : Paravirtual ticketlocks support for linux guests running on KVM hypervisor

[Raghavendra K T <raghavendra.kt@linux.vnet.ibm.com>]

On 07/24/2013 06:06 PM, Raghavendra K T wrote:
> On 07/24/2013 05:36 PM, Gleb Natapov wrote:
>> On Wed, Jul 24, 2013 at 05:30:20PM +0530, Raghavendra K T wrote:
>>> On 07/24/2013 04:09 PM, Gleb Natapov wrote:
>>>> On Wed, Jul 24, 2013 at 03:15:50PM +0530, Raghavendra K T wrote:
>>>>> On 07/23/2013 08:37 PM, Gleb Natapov wrote:
>>>>>> On Mon, Jul 22, 2013 at 11:50:16AM +0530, Raghavendra K T wrote:
>>>>>>> +static void kvm_lock_spinning(struct arch_spinlock *lock,
>>>>>>> __ticket_t want)
>>>>> [...]
>>>>>>> +
>>>>>>> +    /*
>>>>>>> +     * halt until it's our turn and kicked. Note that we do safe
>>>>>>> halt
>>>>>>> +     * for irq enabled case to avoid hang when lock info is
>>>>>>> overwritten
>>>>>>> +     * in irq spinlock slowpath and no spurious interrupt occur
>>>>>>> to save us.
>>>>>>> +     */
>>>>>>> +    if (arch_irqs_disabled_flags(flags))
>>>>>>> +        halt();
>>>>>>> +    else
>>>>>>> +        safe_halt();
>>>>>>> +
>>>>>>> +out:
>>>>>> So here now interrupts can be either disabled or enabled. Previous
>>>>>> version disabled interrupts here, so are we sure it is safe to
>>>>>> have them
>>>>>> enabled at this point? I do not see any problem yet, will keep
>>>>>> thinking.
>>>>>
>>>>> If we enable interrupt here, then
>>>>>
>>>>>
>>>>>>> +    cpumask_clear_cpu(cpu, &waiting_cpus);
>>>>>
>>>>> and if we start serving lock for an interrupt that came here,
>>>>> cpumask clear and w->lock=null may not happen atomically.
>>>>> if irq spinlock does not take slow path we would have non null value
>>>>> for lock, but with no information in waitingcpu.
>>>>>
>>>>> I am still thinking what would be problem with that.
>>>>>
>>>> Exactly, for kicker waiting_cpus and w->lock updates are
>>>> non atomic anyway.
>>>>
>>>>>>> +    w->lock = NULL;
>>>>>>> +    local_irq_restore(flags);
>>>>>>> +    spin_time_accum_blocked(start);
>>>>>>> +}
>>>>>>> +PV_CALLEE_SAVE_REGS_THUNK(kvm_lock_spinning);
>>>>>>> +
>>>>>>> +/* Kick vcpu waiting on @lock->head to reach value @ticket */
>>>>>>> +static void kvm_unlock_kick(struct arch_spinlock *lock,
>>>>>>> __ticket_t ticket)
>>>>>>> +{
>>>>>>> +    int cpu;
>>>>>>> +
>>>>>>> +    add_stats(RELEASED_SLOW, 1);
>>>>>>> +    for_each_cpu(cpu, &waiting_cpus) {
>>>>>>> +        const struct kvm_lock_waiting *w =
>>>>>>> &per_cpu(lock_waiting, cpu);
>>>>>>> +        if (ACCESS_ONCE(w->lock) == lock &&
>>>>>>> +            ACCESS_ONCE(w->want) == ticket) {
>>>>>>> +            add_stats(RELEASED_SLOW_KICKED, 1);
>>>>>>> +            kvm_kick_cpu(cpu);
>>>>>> What about using NMI to wake sleepers? I think it was discussed, but
>>>>>> forgot why it was dismissed.
>>>>>
>>>>> I think I have missed that discussion. 'll go back and check. so
>>>>> what is the idea here? we can easily wake up the halted vcpus that
>>>>> have interrupt disabled?
>>>> We can of course. IIRC the objection was that NMI handling path is very
>>>> fragile and handling NMI on each wakeup will be more expensive then
>>>> waking up a guest without injecting an event, but it is still
>>>> interesting
>>>> to see the numbers.
>>>>
>>>
>>> Haam, now I remember, We had tried request based mechanism. (new
>>> request like REQ_UNHALT) and process that. It had worked, but had some
>>> complex hacks in vcpu_enter_guest to avoid guest hang in case of
>>> request cleared.  So had left it there..
>>>
>>> https://lkml.org/lkml/2012/4/30/67
>>>
>>> But I do not remember performance impact though.
>> No, this is something different. Wakeup with NMI does not need KVM
>> changes at
>> all. Instead of kvm_kick_cpu(cpu) in kvm_unlock_kick you send NMI IPI.
>>
>
> True. It was not NMI.
> just to confirm, are you talking about something like this to be tried ?
>
> apic->send_IPI_mask(cpumask_of(cpu), APIC_DM_NMI);

When I started benchmark, I started seeing
"Dazed and confused, but trying to continue" from unknown nmi error
handling.
Did I miss anything (because we did not register any NMI handler)? or
is it that spurious NMIs are trouble because we could get spurious NMIs
if next waiter already acquired the lock.

(note: I tried sending APIC_DM_REMRD IPI directly, which worked fine
but hypercall way of handling still performed well from the results I
saw).