Re: [Qemu-devel] [RFC PATCH 2/2] Assign a new irq handler while irqfd enabled

["Michael S. Tsirkin" <mst@redhat.com>]

On Mon, Oct 27, 2014 at 07:04:11PM +0800, Li Liu wrote:
> 
> 
> On 2014/10/26 19:56, Michael S. Tsirkin wrote:
> > On Sat, Oct 25, 2014 at 04:24:54PM +0800, john.liuli wrote:
> >> From: Li Liu <john.liuli@huawei.com>
> >>
> >> This irq handler will get the interrupt reason from a
> >> shared memory. And will be assigned only while irqfd
> >> enabled.
> >>
> >> Signed-off-by: Li Liu <john.liuli@huawei.com>
> >> ---
> >>  drivers/virtio/virtio_mmio.c |   34 ++++++++++++++++++++++++++++++++--
> >>  1 file changed, 32 insertions(+), 2 deletions(-)
> >>
> >> diff --git a/drivers/virtio/virtio_mmio.c b/drivers/virtio/virtio_mmio.c
> >> index 28ddb55..7229605 100644
> >> --- a/drivers/virtio/virtio_mmio.c
> >> +++ b/drivers/virtio/virtio_mmio.c
> >> @@ -259,7 +259,31 @@ static irqreturn_t vm_interrupt(int irq, void *opaque)
> >>  	return ret;
> >>  }
> >>  
> >> +/* Notify all virtqueues on an interrupt. */
> >> +static irqreturn_t vm_interrupt_irqfd(int irq, void *opaque)
> >> +{
> >> +	struct virtio_mmio_device *vm_dev = opaque;
> >> +	struct virtio_mmio_vq_info *info;
> >> +	unsigned long status;
> >> +	unsigned long flags;
> >> +	irqreturn_t ret = IRQ_NONE;
> >>  
> >> +	/* Read the interrupt reason and reset it */
> >> +	status = *vm_dev->isr_mem;
> >> +	*vm_dev->isr_mem = 0x0;
> > 
> > you are reading and modifying shared memory
> > without atomics and any memory barriers.
> > Why is this safe?
> > 
> 
> good catch, a stupid mistake.
> 
> >> +
> >> +	if (unlikely(status & VIRTIO_MMIO_INT_CONFIG)) {
> >> +		virtio_config_changed(&vm_dev->vdev);
> >> +		ret = IRQ_HANDLED;
> >> +	}
> >> +
> >> +	spin_lock_irqsave(&vm_dev->lock, flags);
> >> +	list_for_each_entry(info, &vm_dev->virtqueues, node)
> >> +		ret |= vring_interrupt(irq, info->vq);
> >> +	spin_unlock_irqrestore(&vm_dev->lock, flags);
> >> +
> >> +	return ret;
> >> +}
> >>  
> >>  static void vm_del_vq(struct virtqueue *vq)
> >>  {
> > 
> > So you invoke callbacks for all VQs.
> > This won't scale well as the number of VQs grows, will it?
> > 
> >> @@ -391,6 +415,7 @@ error_available:
> >>  	return ERR_PTR(err);
> >>  }
> >>  
> >> +#define VIRTIO_MMIO_F_IRQFD        (1 << 7)
> >>  static int vm_find_vqs(struct virtio_device *vdev, unsigned nvqs,
> >>  		       struct virtqueue *vqs[],
> >>  		       vq_callback_t *callbacks[],
> >> @@ -400,8 +425,13 @@ static int vm_find_vqs(struct virtio_device *vdev, unsigned nvqs,
> >>  	unsigned int irq = platform_get_irq(vm_dev->pdev, 0);
> >>  	int i, err;
> >>  
> >> -	err = request_irq(irq, vm_interrupt, IRQF_SHARED,
> >> -			dev_name(&vdev->dev), vm_dev);
> >> +	if (*vm_dev->isr_mem & VIRTIO_MMIO_F_IRQFD) {
> >> +		err = request_irq(irq, vm_interrupt_irqfd, IRQF_SHARED,
> >> +				  dev_name(&vdev->dev), vm_dev);
> >> +	} else {
> >> +		err = request_irq(irq, vm_interrupt, IRQF_SHARED,
> >> +				  dev_name(&vdev->dev), vm_dev);
> >> +	}
> >>  	if (err)
> >>  		return err;
> > 
> > 
> > So still a single interrupt for all VQs.
> > Again this doesn't scale: a single CPU has to handle
> > interrupts for all of them.
> > I think you need to find a way to get per-VQ interrupts.
> 
> Yeah, AFAIK it's impossible to distribute works to different CPUs with
> only one irq without MSI-X kind mechanism. Assign multiple gsis to one
> device, obviously it's consumptive and not scalable.

Why not? How many gsis are there on ARM?

> Any ideas? Thx.
> 
> > 
> >> -- 
> >> 1.7.9.5
> >>
> > 
> > .
> >