Re: [PATCH v3] IO: Intelligent device lookup on bus

[Sasha Levin <levinsasha928@gmail.com>]

On Wed, 2011-07-27 at 14:35 +0300, Avi Kivity wrote:
> On 07/24/2011 09:15 AM, Sasha Levin wrote:
> > Currently the method of dealing with an IO operation on a bus (PIO/MMIO)
> > is to call the read or write callback for each device registered
> > on the bus until we find a device which handles it.
> >
> > Since the number of devices on a bus can be significant due to ioeventfds
> > and coalesced MMIO zones, this leads to a lot of overhead on each IO
> > operation.
> >
> > Instead of registering devices, we now register ranges which points to
> > a device. Lookup is done using an efficient bsearch instead of a linear
> > search.
> >
> > Performance test was conducted by comparing exit count per second with
> > 200 ioeventfds created on one byte and the guest is trying to access a
> > different byte continuously (triggering usermode exits).
> > Before the patch the guest has achieved 259k exits per second, after the
> > patch the guest does 274k exits per second.
> >
> > diff --git a/arch/x86/kvm/i8254.c b/arch/x86/kvm/i8254.c
> > index efad723..094e057 100644
> > --- a/arch/x86/kvm/i8254.c
> > +++ b/arch/x86/kvm/i8254.c
> > @@ -713,14 +713,15 @@ struct kvm_pit *kvm_create_pit(struct kvm *kvm, u32 flags)
> >   	kvm_register_irq_mask_notifier(kvm, 0,&pit->mask_notifier);
> >
> >   	kvm_iodevice_init(&pit->dev,&pit_dev_ops);
> > -	ret = kvm_io_bus_register_dev(kvm, KVM_PIO_BUS,&pit->dev);
> > +	ret = kvm_io_bus_register_dev(kvm, KVM_PIO_BUS, KVM_PIT_BASE_ADDRESS, KVM_PIT_MEM_LENGTH,&pit->dev);
> 
> Long line.
> 

Will fix.

> >
> > -static inline struct kvm_pic *to_pic(struct kvm_io_device *dev)
> > +static inline struct kvm_pic *to_pic(struct kvm_io_device *dev, gpa_t addr)
> >   {
> > -	return container_of(dev, struct kvm_pic, dev);
> > +	switch (addr) {
> > +	case 0x20:
> > +	case 0x21:
> > +		return container_of(dev, struct kvm_pic, dev_master);
> > +	case 0xa0:
> > +	case 0xa1:
> > +		return container_of(dev, struct kvm_pic, dev_slave);
> > +	case 0x4d0:
> > +	case 0x4d1:
> > +		return container_of(dev, struct kvm_pic, dev_eclr);
> > +	}
> > +
> > +	return NULL;
> >   }
> 
> Somewhat ugly.  I think
> 
>      int picdev_write_master(...)
>     {
>          return pcidev_write(container_of(...), ...);
>      }
> 
> is nicer, no?

It would mean we need a total of 6 wrappers for master, slave and eclr
instead of this switch, if that sounds ok I'll change it.

> 
> > @@ -560,16 +572,36 @@ struct kvm_pic *kvm_create_pic(struct kvm *kvm)
> >   	/*
> >   	 * Initialize PIO device
> >   	 */
> > -	kvm_iodevice_init(&s->dev,&picdev_ops);
> > +	kvm_iodevice_init(&s->dev_master,&picdev_ops);
> > +	kvm_iodevice_init(&s->dev_slave,&picdev_ops);
> > +	kvm_iodevice_init(&s->dev_eclr,&picdev_ops);
> >   	mutex_lock(&kvm->slots_lock);
> > -	ret = kvm_io_bus_register_dev(kvm, KVM_PIO_BUS,&s->dev);
> > +	ret = kvm_io_bus_register_dev(kvm, KVM_PIO_BUS, 0x20, 2,&s->dev_master);
> > +	if (ret<  0)
> > +		goto fail_unlock;
> > +
> > +	ret = kvm_io_bus_register_dev(kvm, KVM_PIO_BUS, 0xa0, 2,&s->dev_slave);
> > +	if (ret<  0)
> > +		goto fail_unlock;
> > +
> > +	ret = kvm_io_bus_register_dev(kvm, KVM_PIO_BUS, 0x4d0, 2,&s->dev_eclr);
> > +	if (ret<  0)
> > +		goto fail_unlock;
> > +
> >   	mutex_unlock(&kvm->slots_lock);
> > -	if (ret<  0) {
> > -		kfree(s);
> > -		return NULL;
> > -	}
> >
> >   	return s;
> > +
> > +fail_unlock:
> > +
> > +	mutex_unlock(&kvm->slots_lock);
> > +	kvm_io_bus_unregister_dev(kvm, KVM_PIO_BUS,&s->dev_master);
> > +	kvm_io_bus_unregister_dev(kvm, KVM_PIO_BUS,&s->dev_slave);
> > +	kvm_io_bus_unregister_dev(kvm, KVM_PIO_BUS,&s->dev_eclr);
> > +
> > +	kfree(s);
> > +
> > +	return NULL;
> >   }
> 
> You're unregistering devices that were never registered.  It may work 
> now, but it's fragile.

I'll fix that.

> 
> 
> >   	if (ret<  0)
> >   		goto out_free_dev;
> >   	list_add_tail(&dev->list,&kvm->coalesced_zones);
> > diff --git a/virt/kvm/eventfd.c b/virt/kvm/eventfd.c
> > index 73358d2..f59c1e8 100644
> > --- a/virt/kvm/eventfd.c
> > +++ b/virt/kvm/eventfd.c
> > @@ -586,7 +586,8 @@ kvm_assign_ioeventfd(struct kvm *kvm, struct kvm_ioeventfd *args)
> >
> >   	kvm_iodevice_init(&p->dev,&ioeventfd_ops);
> >
> > -	ret = kvm_io_bus_register_dev(kvm, bus_idx,&p->dev);
> > +	ret = kvm_io_bus_register_dev(kvm, bus_idx, p->addr, p->length,
> > +				&p->dev);
> 
> Should this be p->length or 1?

We register p->length since when we process a write, the operation
should be fully contained within the IO space of the device.

We verify that the write happens on the first byte within ioeventfd
write handler.

> 
> >   #include<asm/processor.h>
> >   #include<asm/io.h>
> > @@ -2391,24 +2393,94 @@ static void kvm_io_bus_destroy(struct kvm_io_bus *bus)
> >   	int i;
> >
> >   	for (i = 0; i<  bus->dev_count; i++) {
> > -		struct kvm_io_device *pos = bus->devs[i];
> > +		struct kvm_io_device *pos = bus->range[i].dev;
> >
> 
> This will call the destructor three times for the PIC.  Is this safe?

PIC doesn't have a destructor for devices, the code above just does
nothing for PIC devices.

-- 

Sasha.