Re: [PATCH v2 01/11] genirq/affinity:: Export irq_create_affinity_masks()

Re: [PATCH v2 01/11] genirq/affinity:: Export irq_create_affinity_masks()

[Christoph Hellwig]

On Mon, Dec 05, 2022 at 04:41:17PM +0800, Xie Yongji wrote:
> Export irq_create_affinity_masks() so that some modules
> can make use of it to implement interrupt affinity
> spreading mechanism.

I don't think driver should be building low-level affinity masks.
_______________________________________________
Virtualization mailing list
Virtualization@lists.linux-foundation.org
https://lists.linuxfoundation.org/mailman/listinfo/virtualization

Re: [PATCH v2 01/11] genirq/affinity:: Export irq_create_affinity_masks()

[Christoph Hellwig]

On Tue, Dec 06, 2022 at 04:40:37PM +0800, Yongji Xie wrote:
> With the vDPA framework, some drivers (vduse, vdpa-sim) can create
> software-defined virtio devices and attach them to the virtio bus.
> This kind of virtio device is not a pci device or a platform device.
> So it would be needed to export this function if we want to implement
> the automatic affinity management for the virtio device driver which
> is binded to this device.

Why are these devices even using interrupts?  The whjole vdpa thing
is a mess, I also still need to fix up the horrible abuse of the DMA
API for something that isn't even DMA, and this just seems to spread
that same mistake even further.
_______________________________________________
Virtualization mailing list
Virtualization@lists.linux-foundation.org
https://lists.linuxfoundation.org/mailman/listinfo/virtualization

Re: [PATCH v2 01/11] genirq/affinity:: Export irq_create_affinity_masks()

[Jason Wang]

On Tue, Dec 6, 2022 at 5:28 PM Yongji Xie <xieyongji@bytedance.com> wrote:
>
> On Tue, Dec 6, 2022 at 4:47 PM Christoph Hellwig <hch@lst.de> wrote:
> >
> > On Tue, Dec 06, 2022 at 04:40:37PM +0800, Yongji Xie wrote:
> > > With the vDPA framework, some drivers (vduse, vdpa-sim) can create
> > > software-defined virtio devices and attach them to the virtio bus.
> > > This kind of virtio device is not a pci device or a platform device.
> > > So it would be needed to export this function if we want to implement
> > > the automatic affinity management for the virtio device driver which
> > > is binded to this device.
> >
> > Why are these devices even using interrupts?
>
> They don't use interrupt. But they use a bound workqueue to run the
> interrupt callback. So the driver needs an algorithm to choose which
> cpu to run the interrupt callback. Then we found the existing
> interrupt affinity spreading mechanism is very suitable for this
> scenario, so we try to export this function to reuse it.
>
> > The whjole vdpa thing
> > is a mess, I also still need to fix up the horrible abuse of the DMA
> > API for something that isn't even DMA, and this just seems to spread
> > that same mistake even further.

I think it's mostly an issue of some vDPA parents, not the vDPA
itself. I had patches to get rid of the DMA API for vDPA simulators.
Will post.

>
> We just want to reuse this algorithm. And it is completely independent
> of the IRQ subsystem. I guess it would not mess things up.

I think so, it's about which CPU do we want to run the callback and
the callback is not necessarily triggered by an IRQ.

Thanks

>
> Thanks,
> Yongji
>

_______________________________________________
Virtualization mailing list
Virtualization@lists.linux-foundation.org
https://lists.linuxfoundation.org/mailman/listinfo/virtualization

Re: [PATCH v2 01/11] genirq/affinity:: Export irq_create_affinity_masks()

[Michael S. Tsirkin]

On Mon, Dec 05, 2022 at 04:41:17PM +0800, Xie Yongji wrote:
> Export irq_create_affinity_masks() so that some modules
> can make use of it to implement interrupt affinity
> spreading mechanism.
> 
> Signed-off-by: Xie Yongji <xieyongji@bytedance.com>

So this got nacked, what's the plan now?

> ---
>  kernel/irq/affinity.c | 1 +
>  1 file changed, 1 insertion(+)
> 
> diff --git a/kernel/irq/affinity.c b/kernel/irq/affinity.c
> index d9a5c1d65a79..f074a7707c6d 100644
> --- a/kernel/irq/affinity.c
> +++ b/kernel/irq/affinity.c
> @@ -487,6 +487,7 @@ irq_create_affinity_masks(unsigned int nvecs, struct irq_affinity *affd)
>  
>  	return masks;
>  }
> +EXPORT_SYMBOL_GPL(irq_create_affinity_masks);
>  
>  /**
>   * irq_calc_affinity_vectors - Calculate the optimal number of vectors
> -- 
> 2.20.1

_______________________________________________
Virtualization mailing list
Virtualization@lists.linux-foundation.org
https://lists.linuxfoundation.org/mailman/listinfo/virtualization

Re: [PATCH v2 01/11] genirq/affinity:: Export irq_create_affinity_masks()

[Michael S. Tsirkin]

On Mon, Dec 19, 2022 at 05:36:02PM +0800, Yongji Xie wrote:
> On Mon, Dec 19, 2022 at 3:33 PM Michael S. Tsirkin <mst@redhat.com> wrote:
> >
> > On Mon, Dec 05, 2022 at 04:41:17PM +0800, Xie Yongji wrote:
> > > Export irq_create_affinity_masks() so that some modules
> > > can make use of it to implement interrupt affinity
> > > spreading mechanism.
> > >
> > > Signed-off-by: Xie Yongji <xieyongji@bytedance.com>
> >
> > So this got nacked, what's the plan now?
> >
> 
> I‘d like to check with Christoph again first.
> 
> Hi Christoph,
> 
> Jason will post some patches to get rid of the DMA API for vDPA
> simulators. And the irq affinity algorithm is independent of the IRQ
> subsystem IIUC. So could you allow this patch so that we can reuse the
> algorithm to select the best CPU (per-cpu affinity if possible, or at
> least per-node) to run the virtqueue's irq callback.
> 
> Thanks,
> Yongji

I think you need to explain why you are building low level
affinity masks.  what's the plan now?

> > > ---
> > >  kernel/irq/affinity.c | 1 +
> > >  1 file changed, 1 insertion(+)
> > >
> > > diff --git a/kernel/irq/affinity.c b/kernel/irq/affinity.c
> > > index d9a5c1d65a79..f074a7707c6d 100644
> > > --- a/kernel/irq/affinity.c
> > > +++ b/kernel/irq/affinity.c
> > > @@ -487,6 +487,7 @@ irq_create_affinity_masks(unsigned int nvecs, struct irq_affinity *affd)
> > >
> > >       return masks;
> > >  }
> > > +EXPORT_SYMBOL_GPL(irq_create_affinity_masks);
> > >
> > >  /**
> > >   * irq_calc_affinity_vectors - Calculate the optimal number of vectors
> > > --
> > > 2.20.1
> >

_______________________________________________
Virtualization mailing list
Virtualization@lists.linux-foundation.org
https://lists.linuxfoundation.org/mailman/listinfo/virtualization

Re: [PATCH v2 01/11] genirq/affinity:: Export irq_create_affinity_masks()

[Michael S. Tsirkin]

On Mon, Jan 30, 2023 at 07:53:55PM +0800, Yongji Xie wrote:
> On Fri, Jan 27, 2023 at 4:22 PM Michael S. Tsirkin <mst@redhat.com> wrote:
> >
> > On Mon, Dec 19, 2022 at 05:36:02PM +0800, Yongji Xie wrote:
> > > On Mon, Dec 19, 2022 at 3:33 PM Michael S. Tsirkin <mst@redhat.com> wrote:
> > > >
> > > > On Mon, Dec 05, 2022 at 04:41:17PM +0800, Xie Yongji wrote:
> > > > > Export irq_create_affinity_masks() so that some modules
> > > > > can make use of it to implement interrupt affinity
> > > > > spreading mechanism.
> > > > >
> > > > > Signed-off-by: Xie Yongji <xieyongji@bytedance.com>
> > > >
> > > > So this got nacked, what's the plan now?
> > > >
> > >
> > > I‘d like to check with Christoph again first.
> > >
> > > Hi Christoph,
> > >
> > > Jason will post some patches to get rid of the DMA API for vDPA
> > > simulators. And the irq affinity algorithm is independent of the IRQ
> > > subsystem IIUC. So could you allow this patch so that we can reuse the
> > > algorithm to select the best CPU (per-cpu affinity if possible, or at
> > > least per-node) to run the virtqueue's irq callback.
> > >
> > > Thanks,
> > > Yongji
> >
> > I think you need to explain why you are building low level
> > affinity masks.
> 
> In VDUSE case, we use workqueue to run the virtqueue's irq callback.
> Now I want to queue the irq callback kwork to one specific CPU to get
> per-cpu affinity if possible, or at least per-node. So I need to use
> this function to build the low level affinity masks for each
> virtqueue.
> 
> > what's the plan now?
> >
> 
> If there is no objection, I'll post a new version.
> 
> Thanks,
> Yongji

I doubt you made a convicing case here - I think Christoph was saying if
it is not an irq it should not use an irq affinity API.
So a new API possibly sharing implementation with irq affinity
is called for then? Maybe.


-- 
MST

_______________________________________________
Virtualization mailing list
Virtualization@lists.linux-foundation.org
https://lists.linuxfoundation.org/mailman/listinfo/virtualization

Re: [PATCH v2 01/11] genirq/affinity:: Export irq_create_affinity_masks()

[Thomas Gleixner]

On Mon, Feb 13 2023 at 22:50, Yongji Xie wrote:
> On Mon, Feb 13, 2023 at 8:00 PM Michael S. Tsirkin <mst@redhat.com> wrote:
> I can try to split irq_create_affinity_masks() into a common part and
> an irq specific part, and move the common part to a common dir such as
> /lib and export it. Then we can use the common part to build a new API
> for usage.

  https://lore.kernel.org/all/20221227022905.352674-1-ming.lei@redhat.com/

Thanks,

        tglx
_______________________________________________
Virtualization mailing list
Virtualization@lists.linux-foundation.org
https://lists.linuxfoundation.org/mailman/listinfo/virtualization

Re: [PATCH v2 02/11] vdpa: Add set/get_vq_affinity callbacks in vdpa_config_ops

[Jason Wang]

On Mon, Dec 5, 2022 at 4:43 PM Xie Yongji <xieyongji@bytedance.com> wrote:
>
> This introduces set/get_vq_affinity callbacks in
> vdpa_config_ops to support interrupt affinity
> management for vdpa device drivers.
>
> Signed-off-by: Xie Yongji <xieyongji@bytedance.com>

Acked-by: Jason Wang <jasowang@redhat.com>

Thanks

> ---
>  drivers/virtio/virtio_vdpa.c | 28 ++++++++++++++++++++++++++++
>  include/linux/vdpa.h         | 13 +++++++++++++
>  2 files changed, 41 insertions(+)
>
> diff --git a/drivers/virtio/virtio_vdpa.c b/drivers/virtio/virtio_vdpa.c
> index 9670cc79371d..08084b49e5a1 100644
> --- a/drivers/virtio/virtio_vdpa.c
> +++ b/drivers/virtio/virtio_vdpa.c
> @@ -330,6 +330,32 @@ static const char *virtio_vdpa_bus_name(struct virtio_device *vdev)
>         return dev_name(&vdpa->dev);
>  }
>
> +static int virtio_vdpa_set_vq_affinity(struct virtqueue *vq,
> +                                      const struct cpumask *cpu_mask)
> +{
> +       struct virtio_vdpa_device *vd_dev = to_virtio_vdpa_device(vq->vdev);
> +       struct vdpa_device *vdpa = vd_dev->vdpa;
> +       const struct vdpa_config_ops *ops = vdpa->config;
> +       unsigned int index = vq->index;
> +
> +       if (ops->set_vq_affinity)
> +               return ops->set_vq_affinity(vdpa, index, cpu_mask);
> +
> +       return 0;
> +}
> +
> +static const struct cpumask *
> +virtio_vdpa_get_vq_affinity(struct virtio_device *vdev, int index)
> +{
> +       struct vdpa_device *vdpa = vd_get_vdpa(vdev);
> +       const struct vdpa_config_ops *ops = vdpa->config;
> +
> +       if (ops->get_vq_affinity)
> +               return ops->get_vq_affinity(vdpa, index);
> +
> +       return NULL;
> +}
> +
>  static const struct virtio_config_ops virtio_vdpa_config_ops = {
>         .get            = virtio_vdpa_get,
>         .set            = virtio_vdpa_set,
> @@ -342,6 +368,8 @@ static const struct virtio_config_ops virtio_vdpa_config_ops = {
>         .get_features   = virtio_vdpa_get_features,
>         .finalize_features = virtio_vdpa_finalize_features,
>         .bus_name       = virtio_vdpa_bus_name,
> +       .set_vq_affinity = virtio_vdpa_set_vq_affinity,
> +       .get_vq_affinity = virtio_vdpa_get_vq_affinity,
>  };
>
>  static void virtio_vdpa_release_dev(struct device *_d)
> diff --git a/include/linux/vdpa.h b/include/linux/vdpa.h
> index 6d0f5e4e82c2..0ff6c9363356 100644
> --- a/include/linux/vdpa.h
> +++ b/include/linux/vdpa.h
> @@ -247,6 +247,15 @@ struct vdpa_map_file {
>   *                             @vdev: vdpa device
>   *                             Returns the iova range supported by
>   *                             the device.
> + * @set_vq_affinity:           Set the irq affinity of virtqueue (optional)
> + *                             @vdev: vdpa device
> + *                             @idx: virtqueue index
> + *                             @cpu_mask: irq affinity mask
> + *                             Returns integer: success (0) or error (< 0)
> + * @get_vq_affinity:           Get the irq affinity of virtqueue (optional)
> + *                             @vdev: vdpa device
> + *                             @idx: virtqueue index
> + *                             Returns the irq affinity mask
>   * @set_group_asid:            Set address space identifier for a
>   *                             virtqueue group (optional)
>   *                             @vdev: vdpa device
> @@ -331,6 +340,10 @@ struct vdpa_config_ops {
>                            const void *buf, unsigned int len);
>         u32 (*get_generation)(struct vdpa_device *vdev);
>         struct vdpa_iova_range (*get_iova_range)(struct vdpa_device *vdev);
> +       int (*set_vq_affinity)(struct vdpa_device *vdev, u16 idx,
> +                              const struct cpumask *cpu_mask);
> +       const struct cpumask *(*get_vq_affinity)(struct vdpa_device *vdev,
> +                                                u16 idx);
>
>         /* DMA ops */
>         int (*set_map)(struct vdpa_device *vdev, unsigned int asid,
> --
> 2.20.1
>

_______________________________________________
Virtualization mailing list
Virtualization@lists.linux-foundation.org
https://lists.linuxfoundation.org/mailman/listinfo/virtualization

Re: [PATCH v2 03/11] vdpa: Add set_irq_affinity callback in vdpa_config_ops

[Jason Wang]

On Mon, Dec 5, 2022 at 4:43 PM Xie Yongji <xieyongji@bytedance.com> wrote:
>
> This introduces set_irq_affinity callback in
> vdpa_config_ops so that vdpa device driver can
> get the interrupt affinity hint from the virtio
> device driver. The interrupt affinity hint would
> be needed by the interrupt affinity spreading
> mechanism.
>
> Signed-off-by: Xie Yongji <xieyongji@bytedance.com>
> ---
>  drivers/virtio/virtio_vdpa.c | 4 ++++
>  include/linux/vdpa.h         | 8 ++++++++
>  2 files changed, 12 insertions(+)
>
> diff --git a/drivers/virtio/virtio_vdpa.c b/drivers/virtio/virtio_vdpa.c
> index 08084b49e5a1..4731e4616ee0 100644
> --- a/drivers/virtio/virtio_vdpa.c
> +++ b/drivers/virtio/virtio_vdpa.c
> @@ -275,9 +275,13 @@ static int virtio_vdpa_find_vqs(struct virtio_device *vdev, unsigned int nvqs,
>         struct virtio_vdpa_device *vd_dev = to_virtio_vdpa_device(vdev);
>         struct vdpa_device *vdpa = vd_get_vdpa(vdev);
>         const struct vdpa_config_ops *ops = vdpa->config;
> +       struct irq_affinity default_affd = { 0 };
>         struct vdpa_callback cb;
>         int i, err, queue_idx = 0;
>
> +       if (ops->set_irq_affinity)
> +               ops->set_irq_affinity(vdpa, desc ? desc : &default_affd);

I wonder if we need to do this in vhost-vDPA. Or it's better to have a
default affinity by the vDPA parent itself.

(Looking at virtio-pci, it doesn't do something like this).

Thanks

> +
>         for (i = 0; i < nvqs; ++i) {
>                 if (!names[i]) {
>                         vqs[i] = NULL;
> diff --git a/include/linux/vdpa.h b/include/linux/vdpa.h
> index 0ff6c9363356..482ff7d0206f 100644
> --- a/include/linux/vdpa.h
> +++ b/include/linux/vdpa.h
> @@ -256,6 +256,12 @@ struct vdpa_map_file {
>   *                             @vdev: vdpa device
>   *                             @idx: virtqueue index
>   *                             Returns the irq affinity mask
> + * @set_irq_affinity:          Pass the irq affinity hint from the virtio
> + *                             device driver to vdpa driver (optional).
> + *                             Needed by the interrupt affinity spreading
> + *                             mechanism.
> + *                             @vdev: vdpa device
> + *                             @desc: irq affinity hint
>   * @set_group_asid:            Set address space identifier for a
>   *                             virtqueue group (optional)
>   *                             @vdev: vdpa device
> @@ -344,6 +350,8 @@ struct vdpa_config_ops {
>                                const struct cpumask *cpu_mask);
>         const struct cpumask *(*get_vq_affinity)(struct vdpa_device *vdev,
>                                                  u16 idx);
> +       void (*set_irq_affinity)(struct vdpa_device *vdev,
> +                                struct irq_affinity *desc);
>
>         /* DMA ops */
>         int (*set_map)(struct vdpa_device *vdev, unsigned int asid,
> --
> 2.20.1
>

_______________________________________________
Virtualization mailing list
Virtualization@lists.linux-foundation.org
https://lists.linuxfoundation.org/mailman/listinfo/virtualization

Re: [PATCH v2 03/11] vdpa: Add set_irq_affinity callback in vdpa_config_ops

[Jason Wang]

On Mon, Dec 19, 2022 at 12:39 PM Yongji Xie <xieyongji@bytedance.com> wrote:
>
> On Fri, Dec 16, 2022 at 11:58 AM Jason Wang <jasowang@redhat.com> wrote:
> >
> > On Mon, Dec 5, 2022 at 4:43 PM Xie Yongji <xieyongji@bytedance.com> wrote:
> > >
> > > This introduces set_irq_affinity callback in
> > > vdpa_config_ops so that vdpa device driver can
> > > get the interrupt affinity hint from the virtio
> > > device driver. The interrupt affinity hint would
> > > be needed by the interrupt affinity spreading
> > > mechanism.
> > >
> > > Signed-off-by: Xie Yongji <xieyongji@bytedance.com>
> > > ---
> > >  drivers/virtio/virtio_vdpa.c | 4 ++++
> > >  include/linux/vdpa.h         | 8 ++++++++
> > >  2 files changed, 12 insertions(+)
> > >
> > > diff --git a/drivers/virtio/virtio_vdpa.c b/drivers/virtio/virtio_vdpa.c
> > > index 08084b49e5a1..4731e4616ee0 100644
> > > --- a/drivers/virtio/virtio_vdpa.c
> > > +++ b/drivers/virtio/virtio_vdpa.c
> > > @@ -275,9 +275,13 @@ static int virtio_vdpa_find_vqs(struct virtio_device *vdev, unsigned int nvqs,
> > >         struct virtio_vdpa_device *vd_dev = to_virtio_vdpa_device(vdev);
> > >         struct vdpa_device *vdpa = vd_get_vdpa(vdev);
> > >         const struct vdpa_config_ops *ops = vdpa->config;
> > > +       struct irq_affinity default_affd = { 0 };
> > >         struct vdpa_callback cb;
> > >         int i, err, queue_idx = 0;
> > >
> > > +       if (ops->set_irq_affinity)
> > > +               ops->set_irq_affinity(vdpa, desc ? desc : &default_affd);
> >
> > I wonder if we need to do this in vhost-vDPA.
>
> I don't get why we need to do this in vhost-vDPA? Should this be done in VM?

If I was not wrong, this tries to set affinity on the host instead of
the guest. More below.

>
> > Or it's better to have a
> > default affinity by the vDPA parent
> >
>
> I think both are OK. But the default value should always be zero, so I
> put it in a common place.

I think we should either:

1) document the zero default value in vdpa.c
2) set the zero in both vhost-vdpa and virtio-vdpa, or in the vdpa core

>
> > (Looking at virtio-pci, it doesn't do something like this).
> >
>
> Yes, but we did something like this in the pci layer:
> pci_alloc_irq_vectors_affinity().

Ok.

Thanks

>
> Thanks,
> Yongji
>

_______________________________________________
Virtualization mailing list
Virtualization@lists.linux-foundation.org
https://lists.linuxfoundation.org/mailman/listinfo/virtualization

Re: [PATCH v2 03/11] vdpa: Add set_irq_affinity callback in vdpa_config_ops

[Jason Wang]

On Mon, Dec 19, 2022 at 3:12 PM Yongji Xie <xieyongji@bytedance.com> wrote:
>
> On Mon, Dec 19, 2022 at 2:06 PM Jason Wang <jasowang@redhat.com> wrote:
> >
> > On Mon, Dec 19, 2022 at 12:39 PM Yongji Xie <xieyongji@bytedance.com> wrote:
> > >
> > > On Fri, Dec 16, 2022 at 11:58 AM Jason Wang <jasowang@redhat.com> wrote:
> > > >
> > > > On Mon, Dec 5, 2022 at 4:43 PM Xie Yongji <xieyongji@bytedance.com> wrote:
> > > > >
> > > > > This introduces set_irq_affinity callback in
> > > > > vdpa_config_ops so that vdpa device driver can
> > > > > get the interrupt affinity hint from the virtio
> > > > > device driver. The interrupt affinity hint would
> > > > > be needed by the interrupt affinity spreading
> > > > > mechanism.
> > > > >
> > > > > Signed-off-by: Xie Yongji <xieyongji@bytedance.com>
> > > > > ---
> > > > >  drivers/virtio/virtio_vdpa.c | 4 ++++
> > > > >  include/linux/vdpa.h         | 8 ++++++++
> > > > >  2 files changed, 12 insertions(+)
> > > > >
> > > > > diff --git a/drivers/virtio/virtio_vdpa.c b/drivers/virtio/virtio_vdpa.c
> > > > > index 08084b49e5a1..4731e4616ee0 100644
> > > > > --- a/drivers/virtio/virtio_vdpa.c
> > > > > +++ b/drivers/virtio/virtio_vdpa.c
> > > > > @@ -275,9 +275,13 @@ static int virtio_vdpa_find_vqs(struct virtio_device *vdev, unsigned int nvqs,
> > > > >         struct virtio_vdpa_device *vd_dev = to_virtio_vdpa_device(vdev);
> > > > >         struct vdpa_device *vdpa = vd_get_vdpa(vdev);
> > > > >         const struct vdpa_config_ops *ops = vdpa->config;
> > > > > +       struct irq_affinity default_affd = { 0 };
> > > > >         struct vdpa_callback cb;
> > > > >         int i, err, queue_idx = 0;
> > > > >
> > > > > +       if (ops->set_irq_affinity)
> > > > > +               ops->set_irq_affinity(vdpa, desc ? desc : &default_affd);
> > > >
> > > > I wonder if we need to do this in vhost-vDPA.
> > >
> > > I don't get why we need to do this in vhost-vDPA? Should this be done in VM?
> >
> > If I was not wrong, this tries to set affinity on the host instead of
> > the guest. More below.
> >
>
> Yes, it's host stuff. This is used by the virtio device driver to pass
> the irq affinity hint (tell which irq vectors don't need affinity
> management) to the irq affinity manager. In the VM case, it should
> only be related to the guest's virtio device driver and pci irq
> affinity manager. So I don't get why we need to do this in vhost-vDPA.

It's not necessarily the VM, do we have the same requirement for
userspace (like DPDK) drivers?

Thanks

>
> > >
> > > > Or it's better to have a
> > > > default affinity by the vDPA parent
> > > >
> > >
> > > I think both are OK. But the default value should always be zero, so I
> > > put it in a common place.
> >
> > I think we should either:
> >
> > 1) document the zero default value in vdpa.c
> > 2) set the zero in both vhost-vdpa and virtio-vdpa, or in the vdpa core
> >
>
> Can we only call it in the virtio-vdpa case? Thus the vdpa device
> driver can know whether it needs to do the automatic irq affinity
> management or not. In the vhost-vdpa case, we actually don't need the
> irq affinity management.
>
> Thanks,
> Yongji
>

_______________________________________________
Virtualization mailing list
Virtualization@lists.linux-foundation.org
https://lists.linuxfoundation.org/mailman/listinfo/virtualization

Re: [PATCH v2 03/11] vdpa: Add set_irq_affinity callback in vdpa_config_ops

[Jason Wang]

On Tue, Dec 20, 2022 at 6:14 PM Yongji Xie <xieyongji@bytedance.com> wrote:
>
> On Tue, Dec 20, 2022 at 2:31 PM Jason Wang <jasowang@redhat.com> wrote:
> >
> > On Mon, Dec 19, 2022 at 3:12 PM Yongji Xie <xieyongji@bytedance.com> wrote:
> > >
> > > On Mon, Dec 19, 2022 at 2:06 PM Jason Wang <jasowang@redhat.com> wrote:
> > > >
> > > > On Mon, Dec 19, 2022 at 12:39 PM Yongji Xie <xieyongji@bytedance.com> wrote:
> > > > >
> > > > > On Fri, Dec 16, 2022 at 11:58 AM Jason Wang <jasowang@redhat.com> wrote:
> > > > > >
> > > > > > On Mon, Dec 5, 2022 at 4:43 PM Xie Yongji <xieyongji@bytedance.com> wrote:
> > > > > > >
> > > > > > > This introduces set_irq_affinity callback in
> > > > > > > vdpa_config_ops so that vdpa device driver can
> > > > > > > get the interrupt affinity hint from the virtio
> > > > > > > device driver. The interrupt affinity hint would
> > > > > > > be needed by the interrupt affinity spreading
> > > > > > > mechanism.
> > > > > > >
> > > > > > > Signed-off-by: Xie Yongji <xieyongji@bytedance.com>
> > > > > > > ---
> > > > > > >  drivers/virtio/virtio_vdpa.c | 4 ++++
> > > > > > >  include/linux/vdpa.h         | 8 ++++++++
> > > > > > >  2 files changed, 12 insertions(+)
> > > > > > >
> > > > > > > diff --git a/drivers/virtio/virtio_vdpa.c b/drivers/virtio/virtio_vdpa.c
> > > > > > > index 08084b49e5a1..4731e4616ee0 100644
> > > > > > > --- a/drivers/virtio/virtio_vdpa.c
> > > > > > > +++ b/drivers/virtio/virtio_vdpa.c
> > > > > > > @@ -275,9 +275,13 @@ static int virtio_vdpa_find_vqs(struct virtio_device *vdev, unsigned int nvqs,
> > > > > > >         struct virtio_vdpa_device *vd_dev = to_virtio_vdpa_device(vdev);
> > > > > > >         struct vdpa_device *vdpa = vd_get_vdpa(vdev);
> > > > > > >         const struct vdpa_config_ops *ops = vdpa->config;
> > > > > > > +       struct irq_affinity default_affd = { 0 };
> > > > > > >         struct vdpa_callback cb;
> > > > > > >         int i, err, queue_idx = 0;
> > > > > > >
> > > > > > > +       if (ops->set_irq_affinity)
> > > > > > > +               ops->set_irq_affinity(vdpa, desc ? desc : &default_affd);
> > > > > >
> > > > > > I wonder if we need to do this in vhost-vDPA.
> > > > >
> > > > > I don't get why we need to do this in vhost-vDPA? Should this be done in VM?
> > > >
> > > > If I was not wrong, this tries to set affinity on the host instead of
> > > > the guest. More below.
> > > >
> > >
> > > Yes, it's host stuff. This is used by the virtio device driver to pass
> > > the irq affinity hint (tell which irq vectors don't need affinity
> > > management) to the irq affinity manager. In the VM case, it should
> > > only be related to the guest's virtio device driver and pci irq
> > > affinity manager. So I don't get why we need to do this in vhost-vDPA.
> >
> > It's not necessarily the VM, do we have the same requirement for
> > userspace (like DPDK) drivers?
> >
>
> IIUC the vhost-vdpa's irq callback just signals the eventfd. I didn't
> see how to use the irq affinity hint in vdpa device driver. The real
> irq callback should be called in DPDK internally.

I agree.

Thanks

>
> Thanks,
> Yongji
>

_______________________________________________
Virtualization mailing list
Virtualization@lists.linux-foundation.org
https://lists.linuxfoundation.org/mailman/listinfo/virtualization

Re: [PATCH v2 04/11] vduse: Refactor allocation for vduse virtqueues

[Jason Wang]

On Mon, Dec 5, 2022 at 4:44 PM Xie Yongji <xieyongji@bytedance.com> wrote:
>
> Allocate memory for vduse virtqueues one by one instead of
> doing one allocation for all of them.
>
> This is a preparation for adding sysfs interface for virtqueues.
>
> Signed-off-by: Xie Yongji <xieyongji@bytedance.com>

Acked-by: Jason Wang <jasowang@redhat.com>

Thanks

> ---
>  drivers/vdpa/vdpa_user/vduse_dev.c | 98 ++++++++++++++++++++----------
>  1 file changed, 66 insertions(+), 32 deletions(-)
>
> diff --git a/drivers/vdpa/vdpa_user/vduse_dev.c b/drivers/vdpa/vdpa_user/vduse_dev.c
> index 35dceee3ed56..37809bfcb7ef 100644
> --- a/drivers/vdpa/vdpa_user/vduse_dev.c
> +++ b/drivers/vdpa/vdpa_user/vduse_dev.c
> @@ -76,7 +76,7 @@ struct vduse_umem {
>  struct vduse_dev {
>         struct vduse_vdpa *vdev;
>         struct device *dev;
> -       struct vduse_virtqueue *vqs;
> +       struct vduse_virtqueue **vqs;
>         struct vduse_iova_domain *domain;
>         char *name;
>         struct mutex lock;
> @@ -434,7 +434,7 @@ static void vduse_dev_reset(struct vduse_dev *dev)
>         flush_work(&dev->inject);
>
>         for (i = 0; i < dev->vq_num; i++) {
> -               struct vduse_virtqueue *vq = &dev->vqs[i];
> +               struct vduse_virtqueue *vq = dev->vqs[i];
>
>                 vq->ready = false;
>                 vq->desc_addr = 0;
> @@ -466,7 +466,7 @@ static int vduse_vdpa_set_vq_address(struct vdpa_device *vdpa, u16 idx,
>                                 u64 device_area)
>  {
>         struct vduse_dev *dev = vdpa_to_vduse(vdpa);
> -       struct vduse_virtqueue *vq = &dev->vqs[idx];
> +       struct vduse_virtqueue *vq = dev->vqs[idx];
>
>         vq->desc_addr = desc_area;
>         vq->driver_addr = driver_area;
> @@ -500,7 +500,7 @@ static void vduse_vq_kick_work(struct work_struct *work)
>  static void vduse_vdpa_kick_vq(struct vdpa_device *vdpa, u16 idx)
>  {
>         struct vduse_dev *dev = vdpa_to_vduse(vdpa);
> -       struct vduse_virtqueue *vq = &dev->vqs[idx];
> +       struct vduse_virtqueue *vq = dev->vqs[idx];
>
>         if (!eventfd_signal_allowed()) {
>                 schedule_work(&vq->kick);
> @@ -513,7 +513,7 @@ static void vduse_vdpa_set_vq_cb(struct vdpa_device *vdpa, u16 idx,
>                               struct vdpa_callback *cb)
>  {
>         struct vduse_dev *dev = vdpa_to_vduse(vdpa);
> -       struct vduse_virtqueue *vq = &dev->vqs[idx];
> +       struct vduse_virtqueue *vq = dev->vqs[idx];
>
>         spin_lock(&vq->irq_lock);
>         vq->cb.callback = cb->callback;
> @@ -524,7 +524,7 @@ static void vduse_vdpa_set_vq_cb(struct vdpa_device *vdpa, u16 idx,
>  static void vduse_vdpa_set_vq_num(struct vdpa_device *vdpa, u16 idx, u32 num)
>  {
>         struct vduse_dev *dev = vdpa_to_vduse(vdpa);
> -       struct vduse_virtqueue *vq = &dev->vqs[idx];
> +       struct vduse_virtqueue *vq = dev->vqs[idx];
>
>         vq->num = num;
>  }
> @@ -533,7 +533,7 @@ static void vduse_vdpa_set_vq_ready(struct vdpa_device *vdpa,
>                                         u16 idx, bool ready)
>  {
>         struct vduse_dev *dev = vdpa_to_vduse(vdpa);
> -       struct vduse_virtqueue *vq = &dev->vqs[idx];
> +       struct vduse_virtqueue *vq = dev->vqs[idx];
>
>         vq->ready = ready;
>  }
> @@ -541,7 +541,7 @@ static void vduse_vdpa_set_vq_ready(struct vdpa_device *vdpa,
>  static bool vduse_vdpa_get_vq_ready(struct vdpa_device *vdpa, u16 idx)
>  {
>         struct vduse_dev *dev = vdpa_to_vduse(vdpa);
> -       struct vduse_virtqueue *vq = &dev->vqs[idx];
> +       struct vduse_virtqueue *vq = dev->vqs[idx];
>
>         return vq->ready;
>  }
> @@ -550,7 +550,7 @@ static int vduse_vdpa_set_vq_state(struct vdpa_device *vdpa, u16 idx,
>                                 const struct vdpa_vq_state *state)
>  {
>         struct vduse_dev *dev = vdpa_to_vduse(vdpa);
> -       struct vduse_virtqueue *vq = &dev->vqs[idx];
> +       struct vduse_virtqueue *vq = dev->vqs[idx];
>
>         if (dev->driver_features & BIT_ULL(VIRTIO_F_RING_PACKED)) {
>                 vq->state.packed.last_avail_counter =
> @@ -569,7 +569,7 @@ static int vduse_vdpa_get_vq_state(struct vdpa_device *vdpa, u16 idx,
>                                 struct vdpa_vq_state *state)
>  {
>         struct vduse_dev *dev = vdpa_to_vduse(vdpa);
> -       struct vduse_virtqueue *vq = &dev->vqs[idx];
> +       struct vduse_virtqueue *vq = dev->vqs[idx];
>
>         if (dev->driver_features & BIT_ULL(VIRTIO_F_RING_PACKED))
>                 return vduse_dev_get_vq_state_packed(dev, vq, &state->packed);
> @@ -624,8 +624,8 @@ static u16 vduse_vdpa_get_vq_num_max(struct vdpa_device *vdpa)
>         int i;
>
>         for (i = 0; i < dev->vq_num; i++)
> -               if (num_max < dev->vqs[i].num_max)
> -                       num_max = dev->vqs[i].num_max;
> +               if (num_max < dev->vqs[i]->num_max)
> +                       num_max = dev->vqs[i]->num_max;
>
>         return num_max;
>  }
> @@ -863,7 +863,7 @@ static int vduse_kickfd_setup(struct vduse_dev *dev,
>                 return -EINVAL;
>
>         index = array_index_nospec(eventfd->index, dev->vq_num);
> -       vq = &dev->vqs[index];
> +       vq = dev->vqs[index];
>         if (eventfd->fd >= 0) {
>                 ctx = eventfd_ctx_fdget(eventfd->fd);
>                 if (IS_ERR(ctx))
> @@ -889,7 +889,7 @@ static bool vduse_dev_is_ready(struct vduse_dev *dev)
>         int i;
>
>         for (i = 0; i < dev->vq_num; i++)
> -               if (!dev->vqs[i].num_max)
> +               if (!dev->vqs[i]->num_max)
>                         return false;
>
>         return true;
> @@ -1130,7 +1130,7 @@ static long vduse_dev_ioctl(struct file *file, unsigned int cmd,
>                         break;
>
>                 index = array_index_nospec(config.index, dev->vq_num);
> -               dev->vqs[index].num_max = config.max_size;
> +               dev->vqs[index]->num_max = config.max_size;
>                 ret = 0;
>                 break;
>         }
> @@ -1148,7 +1148,7 @@ static long vduse_dev_ioctl(struct file *file, unsigned int cmd,
>                         break;
>
>                 index = array_index_nospec(vq_info.index, dev->vq_num);
> -               vq = &dev->vqs[index];
> +               vq = dev->vqs[index];
>                 vq_info.desc_addr = vq->desc_addr;
>                 vq_info.driver_addr = vq->driver_addr;
>                 vq_info.device_addr = vq->device_addr;
> @@ -1198,7 +1198,7 @@ static long vduse_dev_ioctl(struct file *file, unsigned int cmd,
>                         break;
>
>                 index = array_index_nospec(index, dev->vq_num);
> -               ret = vduse_dev_queue_irq_work(dev, &dev->vqs[index].inject);
> +               ret = vduse_dev_queue_irq_work(dev, &dev->vqs[index]->inject);
>                 break;
>         }
>         case VDUSE_IOTLB_REG_UMEM: {
> @@ -1339,6 +1339,49 @@ static const struct file_operations vduse_dev_fops = {
>         .llseek         = noop_llseek,
>  };
>
> +static void vduse_dev_deinit_vqs(struct vduse_dev *dev)
> +{
> +       int i;
> +
> +       if (!dev->vqs)
> +               return;
> +
> +       for (i = 0; i < dev->vq_num; i++)
> +               kfree(dev->vqs[i]);
> +       kfree(dev->vqs);
> +}
> +
> +static int vduse_dev_init_vqs(struct vduse_dev *dev, u32 vq_align, u32 vq_num)
> +{
> +       int i;
> +
> +       dev->vq_align = vq_align;
> +       dev->vq_num = vq_num;
> +       dev->vqs = kcalloc(dev->vq_num, sizeof(*dev->vqs), GFP_KERNEL);
> +       if (!dev->vqs)
> +               return -ENOMEM;
> +
> +       for (i = 0; i < vq_num; i++) {
> +               dev->vqs[i] = kzalloc(sizeof(*dev->vqs[i]), GFP_KERNEL);
> +               if (!dev->vqs[i])
> +                       goto err;
> +
> +               dev->vqs[i]->index = i;
> +               INIT_WORK(&dev->vqs[i]->inject, vduse_vq_irq_inject);
> +               INIT_WORK(&dev->vqs[i]->kick, vduse_vq_kick_work);
> +               spin_lock_init(&dev->vqs[i]->kick_lock);
> +               spin_lock_init(&dev->vqs[i]->irq_lock);
> +       }
> +
> +       return 0;
> +err:
> +       while (i--)
> +               kfree(dev->vqs[i]);
> +       kfree(dev->vqs);
> +       dev->vqs = NULL;
> +       return -ENOMEM;
> +}
> +
>  static struct vduse_dev *vduse_dev_create(void)
>  {
>         struct vduse_dev *dev = kzalloc(sizeof(*dev), GFP_KERNEL);
> @@ -1396,7 +1439,7 @@ static int vduse_destroy_dev(char *name)
>         device_destroy(vduse_class, MKDEV(MAJOR(vduse_major), dev->minor));
>         idr_remove(&vduse_idr, dev->minor);
>         kvfree(dev->config);
> -       kfree(dev->vqs);
> +       vduse_dev_deinit_vqs(dev);
>         vduse_domain_destroy(dev->domain);
>         kfree(dev->name);
>         vduse_dev_destroy(dev);
> @@ -1483,7 +1526,7 @@ ATTRIBUTE_GROUPS(vduse_dev);
>  static int vduse_create_dev(struct vduse_dev_config *config,
>                             void *config_buf, u64 api_version)
>  {
> -       int i, ret;
> +       int ret;
>         struct vduse_dev *dev;
>
>         ret = -EEXIST;
> @@ -1510,19 +1553,10 @@ static int vduse_create_dev(struct vduse_dev_config *config,
>
>         dev->config = config_buf;
>         dev->config_size = config->config_size;
> -       dev->vq_align = config->vq_align;
> -       dev->vq_num = config->vq_num;
> -       dev->vqs = kcalloc(dev->vq_num, sizeof(*dev->vqs), GFP_KERNEL);
> -       if (!dev->vqs)
> -               goto err_vqs;
>
> -       for (i = 0; i < dev->vq_num; i++) {
> -               dev->vqs[i].index = i;
> -               INIT_WORK(&dev->vqs[i].inject, vduse_vq_irq_inject);
> -               INIT_WORK(&dev->vqs[i].kick, vduse_vq_kick_work);
> -               spin_lock_init(&dev->vqs[i].kick_lock);
> -               spin_lock_init(&dev->vqs[i].irq_lock);
> -       }
> +       ret = vduse_dev_init_vqs(dev, config->vq_align, config->vq_num);
> +       if (ret)
> +               goto err_vqs;
>
>         ret = idr_alloc(&vduse_idr, dev, 1, VDUSE_DEV_MAX, GFP_KERNEL);
>         if (ret < 0)
> @@ -1543,7 +1577,7 @@ static int vduse_create_dev(struct vduse_dev_config *config,
>  err_dev:
>         idr_remove(&vduse_idr, dev->minor);
>  err_idr:
> -       kfree(dev->vqs);
> +       vduse_dev_deinit_vqs(dev);
>  err_vqs:
>         vduse_domain_destroy(dev->domain);
>  err_domain:
> --
> 2.20.1
>

_______________________________________________
Virtualization mailing list
Virtualization@lists.linux-foundation.org
https://lists.linuxfoundation.org/mailman/listinfo/virtualization

Re: [PATCH v2 05/11] vduse: Introduce bound workqueue for irq injection

[Jason Wang]

On Mon, Dec 5, 2022 at 4:44 PM Xie Yongji <xieyongji@bytedance.com> wrote:
>
> This introduces a bound workqueue to support running
> irq callback in a specified cpu.
>
> Signed-off-by: Xie Yongji <xieyongji@bytedance.com>
> ---
>  drivers/vdpa/vdpa_user/vduse_dev.c | 29 ++++++++++++++++++++++-------
>  1 file changed, 22 insertions(+), 7 deletions(-)
>
> diff --git a/drivers/vdpa/vdpa_user/vduse_dev.c b/drivers/vdpa/vdpa_user/vduse_dev.c
> index 37809bfcb7ef..d126f3e32a20 100644
> --- a/drivers/vdpa/vdpa_user/vduse_dev.c
> +++ b/drivers/vdpa/vdpa_user/vduse_dev.c
> @@ -57,6 +57,7 @@ struct vduse_virtqueue {
>         struct vdpa_callback cb;
>         struct work_struct inject;
>         struct work_struct kick;
> +       int irq_effective_cpu;

I wonder why it's a cpu number instead of a cpumask. The latter seems
more flexible, e.g when using NUMA.

>  };
>
>  struct vduse_dev;
> @@ -128,6 +129,7 @@ static struct class *vduse_class;
>  static struct cdev vduse_ctrl_cdev;
>  static struct cdev vduse_cdev;
>  static struct workqueue_struct *vduse_irq_wq;
> +static struct workqueue_struct *vduse_irq_bound_wq;
>
>  static u32 allowed_device_id[] = {
>         VIRTIO_ID_BLOCK,
> @@ -917,7 +919,8 @@ static void vduse_vq_irq_inject(struct work_struct *work)
>  }
>
>  static int vduse_dev_queue_irq_work(struct vduse_dev *dev,
> -                                   struct work_struct *irq_work)
> +                                   struct work_struct *irq_work,
> +                                   int irq_effective_cpu)
>  {
>         int ret = -EINVAL;
>
> @@ -926,7 +929,11 @@ static int vduse_dev_queue_irq_work(struct vduse_dev *dev,
>                 goto unlock;
>
>         ret = 0;
> -       queue_work(vduse_irq_wq, irq_work);
> +       if (irq_effective_cpu == -1)

Is it better to have a macro for this magic number?

Thanks

_______________________________________________
Virtualization mailing list
Virtualization@lists.linux-foundation.org
https://lists.linuxfoundation.org/mailman/listinfo/virtualization

Re: [PATCH v2 05/11] vduse: Introduce bound workqueue for irq injection

[Jason Wang]

On Mon, Dec 19, 2022 at 1:04 PM Yongji Xie <xieyongji@bytedance.com> wrote:
>
> On Fri, Dec 16, 2022 at 12:02 PM Jason Wang <jasowang@redhat.com> wrote:
> >
> > On Mon, Dec 5, 2022 at 4:44 PM Xie Yongji <xieyongji@bytedance.com> wrote:
> > >
> > > This introduces a bound workqueue to support running
> > > irq callback in a specified cpu.
> > >
> > > Signed-off-by: Xie Yongji <xieyongji@bytedance.com>
> > > ---
> > >  drivers/vdpa/vdpa_user/vduse_dev.c | 29 ++++++++++++++++++++++-------
> > >  1 file changed, 22 insertions(+), 7 deletions(-)
> > >
> > > diff --git a/drivers/vdpa/vdpa_user/vduse_dev.c b/drivers/vdpa/vdpa_user/vduse_dev.c
> > > index 37809bfcb7ef..d126f3e32a20 100644
> > > --- a/drivers/vdpa/vdpa_user/vduse_dev.c
> > > +++ b/drivers/vdpa/vdpa_user/vduse_dev.c
> > > @@ -57,6 +57,7 @@ struct vduse_virtqueue {
> > >         struct vdpa_callback cb;
> > >         struct work_struct inject;
> > >         struct work_struct kick;
> > > +       int irq_effective_cpu;
> >
> > I wonder why it's a cpu number instead of a cpumask. The latter seems
> > more flexible, e.g when using NUMA.
> >
>
> This variable represents the CPU that runs the interrupt callback
> rather than CPU affinity.

Ok, but for some reason it only gets updated when a new affinity is set?

(Btw, I don't see how the code deals with cpu hotplug, do we need
cpuhot notifier?)

Thanks

>
> > >  };
> > >
> > >  struct vduse_dev;
> > > @@ -128,6 +129,7 @@ static struct class *vduse_class;
> > >  static struct cdev vduse_ctrl_cdev;
> > >  static struct cdev vduse_cdev;
> > >  static struct workqueue_struct *vduse_irq_wq;
> > > +static struct workqueue_struct *vduse_irq_bound_wq;
> > >
> > >  static u32 allowed_device_id[] = {
> > >         VIRTIO_ID_BLOCK,
> > > @@ -917,7 +919,8 @@ static void vduse_vq_irq_inject(struct work_struct *work)
> > >  }
> > >
> > >  static int vduse_dev_queue_irq_work(struct vduse_dev *dev,
> > > -                                   struct work_struct *irq_work)
> > > +                                   struct work_struct *irq_work,
> > > +                                   int irq_effective_cpu)
> > >  {
> > >         int ret = -EINVAL;
> > >
> > > @@ -926,7 +929,11 @@ static int vduse_dev_queue_irq_work(struct vduse_dev *dev,
> > >                 goto unlock;
> > >
> > >         ret = 0;
> > > -       queue_work(vduse_irq_wq, irq_work);
> > > +       if (irq_effective_cpu == -1)
> >
> > Is it better to have a macro for this magic number?
> >
>
> It makes sense to me.
>
> Thanks,
> Yongji
>

_______________________________________________
Virtualization mailing list
Virtualization@lists.linux-foundation.org
https://lists.linuxfoundation.org/mailman/listinfo/virtualization

Re: [PATCH v2 05/11] vduse: Introduce bound workqueue for irq injection

[Jason Wang]

On Tue, Dec 20, 2022 at 6:02 PM Yongji Xie <xieyongji@bytedance.com> wrote:
>
> On Tue, Dec 20, 2022 at 2:28 PM Jason Wang <jasowang@redhat.com> wrote:
> >
> > On Mon, Dec 19, 2022 at 1:04 PM Yongji Xie <xieyongji@bytedance.com> wrote:
> > >
> > > On Fri, Dec 16, 2022 at 12:02 PM Jason Wang <jasowang@redhat.com> wrote:
> > > >
> > > > On Mon, Dec 5, 2022 at 4:44 PM Xie Yongji <xieyongji@bytedance.com> wrote:
> > > > >
> > > > > This introduces a bound workqueue to support running
> > > > > irq callback in a specified cpu.
> > > > >
> > > > > Signed-off-by: Xie Yongji <xieyongji@bytedance.com>
> > > > > ---
> > > > >  drivers/vdpa/vdpa_user/vduse_dev.c | 29 ++++++++++++++++++++++-------
> > > > >  1 file changed, 22 insertions(+), 7 deletions(-)
> > > > >
> > > > > diff --git a/drivers/vdpa/vdpa_user/vduse_dev.c b/drivers/vdpa/vdpa_user/vduse_dev.c
> > > > > index 37809bfcb7ef..d126f3e32a20 100644
> > > > > --- a/drivers/vdpa/vdpa_user/vduse_dev.c
> > > > > +++ b/drivers/vdpa/vdpa_user/vduse_dev.c
> > > > > @@ -57,6 +57,7 @@ struct vduse_virtqueue {
> > > > >         struct vdpa_callback cb;
> > > > >         struct work_struct inject;
> > > > >         struct work_struct kick;
> > > > > +       int irq_effective_cpu;
> > > >
> > > > I wonder why it's a cpu number instead of a cpumask. The latter seems
> > > > more flexible, e.g when using NUMA.
> > > >
> > >
> > > This variable represents the CPU that runs the interrupt callback
> > > rather than CPU affinity.
> >
> > Ok, but for some reason it only gets updated when a new affinity is set?
> >
>
> Yes, since we don't use round-robin now. And if affinity is not set,
> we rollback to the default behavior (use un-bounded workqueue to run
> irq callback).
>
> > (Btw, I don't see how the code deals with cpu hotplug, do we need
> > cpuhot notifier?)
> >
>
> Currently the queue_work_on() can handle the cpu hotplug case, so I
> think we can simply check whether the CPU is online each time queuing
> the kwork, then update the affinity if needed.

Right.

Thanks

>
> Thanks,
> Yongji
>

_______________________________________________
Virtualization mailing list
Virtualization@lists.linux-foundation.org
https://lists.linuxfoundation.org/mailman/listinfo/virtualization

Re: [PATCH v2 06/11] vduse: Support automatic irq callback affinity

[Jason Wang]

On Mon, Dec 5, 2022 at 4:59 PM Xie Yongji <xieyongji@bytedance.com> wrote:
>
> This brings current interrupt affinity spreading mechanism
> to vduse device. We will make use of irq_create_affinity_masks()
> to create an irq callback affinity mask for each virtqueue of
> vduse device. Then we will choose the CPU which has the lowest
> number of interrupt allocated in the affinity mask to run the
> irq callback.

This seems a balance mechanism but it might not be the semantic of the
affinity or any reason we need to do this? I guess we should use at
least round-robin in this case.

>
> Signed-off-by: Xie Yongji <xieyongji@bytedance.com>
> ---
>  drivers/vdpa/vdpa_user/vduse_dev.c | 50 ++++++++++++++++++++++++++++++
>  1 file changed, 50 insertions(+)
>
> diff --git a/drivers/vdpa/vdpa_user/vduse_dev.c b/drivers/vdpa/vdpa_user/vduse_dev.c
> index d126f3e32a20..90c2896039d9 100644
> --- a/drivers/vdpa/vdpa_user/vduse_dev.c
> +++ b/drivers/vdpa/vdpa_user/vduse_dev.c
> @@ -23,6 +23,7 @@
>  #include <linux/nospec.h>
>  #include <linux/vmalloc.h>
>  #include <linux/sched/mm.h>
> +#include <linux/interrupt.h>
>  #include <uapi/linux/vduse.h>
>  #include <uapi/linux/vdpa.h>
>  #include <uapi/linux/virtio_config.h>
> @@ -58,6 +59,8 @@ struct vduse_virtqueue {
>         struct work_struct inject;
>         struct work_struct kick;
>         int irq_effective_cpu;
> +       struct cpumask irq_affinity;
> +       spinlock_t irq_affinity_lock;

Ok, I'd suggest to squash this into patch 5 to make it more easier to
be reviewed.

>  };
>
>  struct vduse_dev;
> @@ -123,6 +126,7 @@ struct vduse_control {
>
>  static DEFINE_MUTEX(vduse_lock);
>  static DEFINE_IDR(vduse_idr);
> +static DEFINE_PER_CPU(unsigned long, vduse_allocated_irq);
>
>  static dev_t vduse_major;
>  static struct class *vduse_class;
> @@ -710,6 +714,49 @@ static u32 vduse_vdpa_get_generation(struct vdpa_device *vdpa)
>         return dev->generation;
>  }
>
> +static void vduse_vq_update_effective_cpu(struct vduse_virtqueue *vq)
> +{
> +       unsigned int cpu, best_cpu;
> +       unsigned long allocated, allocated_min = UINT_MAX;
> +
> +       spin_lock(&vq->irq_affinity_lock);
> +
> +       best_cpu = vq->irq_effective_cpu;
> +       if (best_cpu != -1)
> +               per_cpu(vduse_allocated_irq, best_cpu) -= 1;
> +
> +       for_each_cpu(cpu, &vq->irq_affinity) {
> +               allocated = per_cpu(vduse_allocated_irq, cpu);
> +               if (!cpu_online(cpu) || allocated >= allocated_min)
> +                       continue;
> +
> +               best_cpu = cpu;
> +               allocated_min = allocated;
> +       }
> +       vq->irq_effective_cpu = best_cpu;
> +       per_cpu(vduse_allocated_irq, best_cpu) += 1;
> +
> +       spin_unlock(&vq->irq_affinity_lock);
> +}
> +
> +static void vduse_vdpa_set_irq_affinity(struct vdpa_device *vdpa,
> +                                       struct irq_affinity *desc)
> +{
> +       struct vduse_dev *dev = vdpa_to_vduse(vdpa);
> +       struct irq_affinity_desc *affd = NULL;
> +       int i;
> +
> +       affd = irq_create_affinity_masks(dev->vq_num, desc);
> +       if (!affd)

Let's add a comment on the vdpa config ops to say set_irq_affinity()
is best effort.

Thanks

> +               return;
> +
> +       for (i = 0; i < dev->vq_num; i++) {
> +               cpumask_copy(&dev->vqs[i]->irq_affinity, &affd[i].mask);
> +               vduse_vq_update_effective_cpu(dev->vqs[i]);
> +       }
> +       kfree(affd);
> +}
> +
>  static int vduse_vdpa_set_map(struct vdpa_device *vdpa,
>                                 unsigned int asid,
>                                 struct vhost_iotlb *iotlb)
> @@ -760,6 +807,7 @@ static const struct vdpa_config_ops vduse_vdpa_config_ops = {
>         .get_config             = vduse_vdpa_get_config,
>         .set_config             = vduse_vdpa_set_config,
>         .get_generation         = vduse_vdpa_get_generation,
> +       .set_irq_affinity       = vduse_vdpa_set_irq_affinity,
>         .reset                  = vduse_vdpa_reset,
>         .set_map                = vduse_vdpa_set_map,
>         .free                   = vduse_vdpa_free,
> @@ -1380,6 +1428,8 @@ static int vduse_dev_init_vqs(struct vduse_dev *dev, u32 vq_align, u32 vq_num)
>                 INIT_WORK(&dev->vqs[i]->kick, vduse_vq_kick_work);
>                 spin_lock_init(&dev->vqs[i]->kick_lock);
>                 spin_lock_init(&dev->vqs[i]->irq_lock);
> +               spin_lock_init(&dev->vqs[i]->irq_affinity_lock);
> +               cpumask_setall(&dev->vqs[i]->irq_affinity);
>         }
>
>         return 0;
> --
> 2.20.1
>

_______________________________________________
Virtualization mailing list
Virtualization@lists.linux-foundation.org
https://lists.linuxfoundation.org/mailman/listinfo/virtualization

Re: [PATCH v2 06/11] vduse: Support automatic irq callback affinity

[Jason Wang]

On Mon, Dec 19, 2022 at 12:56 PM Yongji Xie <xieyongji@bytedance.com> wrote:
>
> On Fri, Dec 16, 2022 at 1:30 PM Jason Wang <jasowang@redhat.com> wrote:
> >
> > On Mon, Dec 5, 2022 at 4:59 PM Xie Yongji <xieyongji@bytedance.com> wrote:
> > >
> > > This brings current interrupt affinity spreading mechanism
> > > to vduse device. We will make use of irq_create_affinity_masks()
> > > to create an irq callback affinity mask for each virtqueue of
> > > vduse device. Then we will choose the CPU which has the lowest
> > > number of interrupt allocated in the affinity mask to run the
> > > irq callback.
> >
> > This seems a balance mechanism but it might not be the semantic of the
> > affinity or any reason we need to do this? I guess we should use at
> > least round-robin in this case.
> >
>
> Here we try to follow the pci interrupt management mechanism. In VM
> cases, the interrupt should always be triggered to one specific CPU
> rather than to each CPU in turn.

If I was not wrong, when using MSI, most arch allows not only the
cpuid as the destination but policy like rr and low priority first.

Thanks

>
> > >
> > > Signed-off-by: Xie Yongji <xieyongji@bytedance.com>
> > > ---
> > >  drivers/vdpa/vdpa_user/vduse_dev.c | 50 ++++++++++++++++++++++++++++++
> > >  1 file changed, 50 insertions(+)
> > >
> > > diff --git a/drivers/vdpa/vdpa_user/vduse_dev.c b/drivers/vdpa/vdpa_user/vduse_dev.c
> > > index d126f3e32a20..90c2896039d9 100644
> > > --- a/drivers/vdpa/vdpa_user/vduse_dev.c
> > > +++ b/drivers/vdpa/vdpa_user/vduse_dev.c
> > > @@ -23,6 +23,7 @@
> > >  #include <linux/nospec.h>
> > >  #include <linux/vmalloc.h>
> > >  #include <linux/sched/mm.h>
> > > +#include <linux/interrupt.h>
> > >  #include <uapi/linux/vduse.h>
> > >  #include <uapi/linux/vdpa.h>
> > >  #include <uapi/linux/virtio_config.h>
> > > @@ -58,6 +59,8 @@ struct vduse_virtqueue {
> > >         struct work_struct inject;
> > >         struct work_struct kick;
> > >         int irq_effective_cpu;
> > > +       struct cpumask irq_affinity;
> > > +       spinlock_t irq_affinity_lock;
> >
> > Ok, I'd suggest to squash this into patch 5 to make it more easier to
> > be reviewed.
> >
>
> OK.
>
> > >  };
> > >
> > >  struct vduse_dev;
> > > @@ -123,6 +126,7 @@ struct vduse_control {
> > >
> > >  static DEFINE_MUTEX(vduse_lock);
> > >  static DEFINE_IDR(vduse_idr);
> > > +static DEFINE_PER_CPU(unsigned long, vduse_allocated_irq);
> > >
> > >  static dev_t vduse_major;
> > >  static struct class *vduse_class;
> > > @@ -710,6 +714,49 @@ static u32 vduse_vdpa_get_generation(struct vdpa_device *vdpa)
> > >         return dev->generation;
> > >  }
> > >
> > > +static void vduse_vq_update_effective_cpu(struct vduse_virtqueue *vq)
> > > +{
> > > +       unsigned int cpu, best_cpu;
> > > +       unsigned long allocated, allocated_min = UINT_MAX;
> > > +
> > > +       spin_lock(&vq->irq_affinity_lock);
> > > +
> > > +       best_cpu = vq->irq_effective_cpu;
> > > +       if (best_cpu != -1)
> > > +               per_cpu(vduse_allocated_irq, best_cpu) -= 1;
> > > +
> > > +       for_each_cpu(cpu, &vq->irq_affinity) {
> > > +               allocated = per_cpu(vduse_allocated_irq, cpu);
> > > +               if (!cpu_online(cpu) || allocated >= allocated_min)
> > > +                       continue;
> > > +
> > > +               best_cpu = cpu;
> > > +               allocated_min = allocated;
> > > +       }
> > > +       vq->irq_effective_cpu = best_cpu;
> > > +       per_cpu(vduse_allocated_irq, best_cpu) += 1;
> > > +
> > > +       spin_unlock(&vq->irq_affinity_lock);
> > > +}
> > > +
> > > +static void vduse_vdpa_set_irq_affinity(struct vdpa_device *vdpa,
> > > +                                       struct irq_affinity *desc)
> > > +{
> > > +       struct vduse_dev *dev = vdpa_to_vduse(vdpa);
> > > +       struct irq_affinity_desc *affd = NULL;
> > > +       int i;
> > > +
> > > +       affd = irq_create_affinity_masks(dev->vq_num, desc);
> > > +       if (!affd)
> >
> > Let's add a comment on the vdpa config ops to say set_irq_affinity()
> > is best effort.
> >
>
> OK.
>
> Thanks,
> Yongji
>

_______________________________________________
Virtualization mailing list
Virtualization@lists.linux-foundation.org
https://lists.linuxfoundation.org/mailman/listinfo/virtualization

Re: [PATCH v2 08/11] vduse: Add sysfs interface for irq callback affinity

[Jason Wang]

On Mon, Dec 5, 2022 at 5:03 PM Xie Yongji <xieyongji@bytedance.com> wrote:
>
> Add sysfs interface for each vduse virtqueue to
> show the affinity and effective affinity for irq
> callback.
>
> And we can also use this interface to change the
> effective affinity which must be a subset of the
> irq callback affinity mask for the virtqueue. This
> might be useful for performance tuning when the irq
> callback affinity mask contains more than one CPU.
>
> Signed-off-by: Xie Yongji <xieyongji@bytedance.com>
> ---
>  drivers/vdpa/vdpa_user/vduse_dev.c | 148 ++++++++++++++++++++++++++---
>  1 file changed, 137 insertions(+), 11 deletions(-)
>
> diff --git a/drivers/vdpa/vdpa_user/vduse_dev.c b/drivers/vdpa/vdpa_user/vduse_dev.c
> index 6507a78abc9d..c65f84100e30 100644
> --- a/drivers/vdpa/vdpa_user/vduse_dev.c
> +++ b/drivers/vdpa/vdpa_user/vduse_dev.c
> @@ -61,6 +61,7 @@ struct vduse_virtqueue {
>         int irq_effective_cpu;
>         struct cpumask irq_affinity;
>         spinlock_t irq_affinity_lock;
> +       struct kobject kobj;
>  };
>
>  struct vduse_dev;
> @@ -1419,6 +1420,120 @@ static const struct file_operations vduse_dev_fops = {
>         .llseek         = noop_llseek,
>  };
>
> +static ssize_t irq_cb_affinity_show(struct vduse_virtqueue *vq, char *buf)
> +{
> +       return sprintf(buf, "%*pb\n", cpumask_pr_args(&vq->irq_affinity));
> +}
> +
> +static ssize_t irq_cb_effective_affinity_show(struct vduse_virtqueue *vq,
> +                                             char *buf)
> +{
> +       struct cpumask all_mask = CPU_MASK_ALL;
> +       const struct cpumask *mask = &all_mask;
> +
> +       if (vq->irq_effective_cpu != -1)
> +               mask = get_cpu_mask(vq->irq_effective_cpu);

Shouldn't this be vq->irq_affinity?

> +
> +       return sprintf(buf, "%*pb\n", cpumask_pr_args(mask));
> +}
> +
> +static ssize_t irq_cb_effective_affinity_store(struct vduse_virtqueue *vq,
> +                                              const char *buf, size_t count)
> +{
> +       cpumask_var_t new_value;
> +       int ret;
> +
> +       if (!zalloc_cpumask_var(&new_value, GFP_KERNEL))
> +               return -ENOMEM;
> +
> +       ret = cpumask_parse(buf, new_value);
> +       if (ret)
> +               goto free_mask;
> +
> +       ret = -EINVAL;
> +       if (!cpumask_intersects(new_value, &vq->irq_affinity))
> +               goto free_mask;
> +
> +       spin_lock(&vq->irq_affinity_lock);
> +
> +       if (vq->irq_effective_cpu != -1)
> +               per_cpu(vduse_allocated_irq, vq->irq_effective_cpu) -= 1;
> +
> +       vq->irq_effective_cpu = cpumask_first(new_value);

Does this mean except for the first cpu, the rest of the mask is unused?

Thanks

> +       per_cpu(vduse_allocated_irq, vq->irq_effective_cpu) += 1;
> +
> +       spin_unlock(&vq->irq_affinity_lock);
> +       ret = count;
> +
> +free_mask:
> +       free_cpumask_var(new_value);
> +       return ret;
> +}
> +
> +struct vq_sysfs_entry {
> +       struct attribute attr;
> +       ssize_t (*show)(struct vduse_virtqueue *vq, char *buf);
> +       ssize_t (*store)(struct vduse_virtqueue *vq, const char *buf,
> +                        size_t count);
> +};
> +
> +static struct vq_sysfs_entry irq_cb_affinity_attr = __ATTR_RO(irq_cb_affinity);
> +static struct vq_sysfs_entry irq_cb_effective_affinity_attr =
> +                                       __ATTR_RW(irq_cb_effective_affinity);
> +
> +static struct attribute *vq_attrs[] = {
> +       &irq_cb_affinity_attr.attr,
> +       &irq_cb_effective_affinity_attr.attr,
> +       NULL,
> +};
> +ATTRIBUTE_GROUPS(vq);
> +
> +static ssize_t vq_attr_show(struct kobject *kobj, struct attribute *attr,
> +                           char *buf)
> +{
> +       struct vduse_virtqueue *vq = container_of(kobj,
> +                                       struct vduse_virtqueue, kobj);
> +       struct vq_sysfs_entry *entry = container_of(attr,
> +                                       struct vq_sysfs_entry, attr);
> +
> +       if (!entry->show)
> +               return -EIO;
> +
> +       return entry->show(vq, buf);
> +}
> +
> +static ssize_t vq_attr_store(struct kobject *kobj, struct attribute *attr,
> +                            const char *buf, size_t count)
> +{
> +       struct vduse_virtqueue *vq = container_of(kobj,
> +                                       struct vduse_virtqueue, kobj);
> +       struct vq_sysfs_entry *entry = container_of(attr,
> +                                       struct vq_sysfs_entry, attr);
> +
> +       if (!entry->store)
> +               return -EIO;
> +
> +       return entry->store(vq, buf, count);
> +}
> +
> +static const struct sysfs_ops vq_sysfs_ops = {
> +       .show = vq_attr_show,
> +       .store = vq_attr_store,
> +};
> +
> +static void vq_release(struct kobject *kobj)
> +{
> +       struct vduse_virtqueue *vq = container_of(kobj,
> +                                       struct vduse_virtqueue, kobj);
> +       kfree(vq);
> +}
> +
> +static struct kobj_type vq_type = {
> +       .release        = vq_release,
> +       .sysfs_ops      = &vq_sysfs_ops,
> +       .default_groups = vq_groups,
> +};
> +
>  static void vduse_dev_deinit_vqs(struct vduse_dev *dev)
>  {
>         int i;
> @@ -1427,13 +1542,13 @@ static void vduse_dev_deinit_vqs(struct vduse_dev *dev)
>                 return;
>
>         for (i = 0; i < dev->vq_num; i++)
> -               kfree(dev->vqs[i]);
> +               kobject_put(&dev->vqs[i]->kobj);
>         kfree(dev->vqs);
>  }
>
>  static int vduse_dev_init_vqs(struct vduse_dev *dev, u32 vq_align, u32 vq_num)
>  {
> -       int i;
> +       int ret, i;
>
>         dev->vq_align = vq_align;
>         dev->vq_num = vq_num;
> @@ -1443,8 +1558,10 @@ static int vduse_dev_init_vqs(struct vduse_dev *dev, u32 vq_align, u32 vq_num)
>
>         for (i = 0; i < vq_num; i++) {
>                 dev->vqs[i] = kzalloc(sizeof(*dev->vqs[i]), GFP_KERNEL);
> -               if (!dev->vqs[i])
> +               if (!dev->vqs[i]) {
> +                       ret = -ENOMEM;
>                         goto err;
> +               }
>
>                 dev->vqs[i]->index = i;
>                 dev->vqs[i]->irq_effective_cpu = -1;
> @@ -1454,15 +1571,23 @@ static int vduse_dev_init_vqs(struct vduse_dev *dev, u32 vq_align, u32 vq_num)
>                 spin_lock_init(&dev->vqs[i]->irq_lock);
>                 spin_lock_init(&dev->vqs[i]->irq_affinity_lock);
>                 cpumask_setall(&dev->vqs[i]->irq_affinity);
> +
> +               kobject_init(&dev->vqs[i]->kobj, &vq_type);
> +               ret = kobject_add(&dev->vqs[i]->kobj,
> +                                 &dev->dev->kobj, "vq%d", i);
> +               if (ret) {
> +                       kfree(dev->vqs[i]);
> +                       goto err;
> +               }
>         }
>
>         return 0;
>  err:
>         while (i--)
> -               kfree(dev->vqs[i]);
> +               kobject_put(&dev->vqs[i]->kobj);
>         kfree(dev->vqs);
>         dev->vqs = NULL;
> -       return -ENOMEM;
> +       return ret;
>  }
>
>  static struct vduse_dev *vduse_dev_create(void)
> @@ -1637,10 +1762,6 @@ static int vduse_create_dev(struct vduse_dev_config *config,
>         dev->config = config_buf;
>         dev->config_size = config->config_size;
>
> -       ret = vduse_dev_init_vqs(dev, config->vq_align, config->vq_num);
> -       if (ret)
> -               goto err_vqs;
> -
>         ret = idr_alloc(&vduse_idr, dev, 1, VDUSE_DEV_MAX, GFP_KERNEL);
>         if (ret < 0)
>                 goto err_idr;
> @@ -1654,14 +1775,19 @@ static int vduse_create_dev(struct vduse_dev_config *config,
>                 ret = PTR_ERR(dev->dev);
>                 goto err_dev;
>         }
> +
> +       ret = vduse_dev_init_vqs(dev, config->vq_align, config->vq_num);
> +       if (ret)
> +               goto err_vqs;
> +
>         __module_get(THIS_MODULE);
>
>         return 0;
> +err_vqs:
> +       device_destroy(vduse_class, MKDEV(MAJOR(vduse_major), dev->minor));
>  err_dev:
>         idr_remove(&vduse_idr, dev->minor);
>  err_idr:
> -       vduse_dev_deinit_vqs(dev);
> -err_vqs:
>         vduse_domain_destroy(dev->domain);
>  err_domain:
>         kfree(dev->name);
> --
> 2.20.1
>

_______________________________________________
Virtualization mailing list
Virtualization@lists.linux-foundation.org
https://lists.linuxfoundation.org/mailman/listinfo/virtualization

Re: [PATCH v2 08/11] vduse: Add sysfs interface for irq callback affinity

[Jason Wang]

On Mon, Dec 19, 2022 at 1:16 PM Yongji Xie <xieyongji@bytedance.com> wrote:
>
> On Fri, Dec 16, 2022 at 1:35 PM Jason Wang <jasowang@redhat.com> wrote:
> >
> > On Mon, Dec 5, 2022 at 5:03 PM Xie Yongji <xieyongji@bytedance.com> wrote:
> > >
> > > Add sysfs interface for each vduse virtqueue to
> > > show the affinity and effective affinity for irq
> > > callback.
> > >
> > > And we can also use this interface to change the
> > > effective affinity which must be a subset of the
> > > irq callback affinity mask for the virtqueue. This
> > > might be useful for performance tuning when the irq
> > > callback affinity mask contains more than one CPU.
> > >
> > > Signed-off-by: Xie Yongji <xieyongji@bytedance.com>
> > > ---
> > >  drivers/vdpa/vdpa_user/vduse_dev.c | 148 ++++++++++++++++++++++++++---
> > >  1 file changed, 137 insertions(+), 11 deletions(-)
> > >
> > > diff --git a/drivers/vdpa/vdpa_user/vduse_dev.c b/drivers/vdpa/vdpa_user/vduse_dev.c
> > > index 6507a78abc9d..c65f84100e30 100644
> > > --- a/drivers/vdpa/vdpa_user/vduse_dev.c
> > > +++ b/drivers/vdpa/vdpa_user/vduse_dev.c
> > > @@ -61,6 +61,7 @@ struct vduse_virtqueue {
> > >         int irq_effective_cpu;
> > >         struct cpumask irq_affinity;
> > >         spinlock_t irq_affinity_lock;
> > > +       struct kobject kobj;
> > >  };
> > >
> > >  struct vduse_dev;
> > > @@ -1419,6 +1420,120 @@ static const struct file_operations vduse_dev_fops = {
> > >         .llseek         = noop_llseek,
> > >  };
> > >
> > > +static ssize_t irq_cb_affinity_show(struct vduse_virtqueue *vq, char *buf)
> > > +{
> > > +       return sprintf(buf, "%*pb\n", cpumask_pr_args(&vq->irq_affinity));
> > > +}
> > > +
> > > +static ssize_t irq_cb_effective_affinity_show(struct vduse_virtqueue *vq,
> > > +                                             char *buf)
> > > +{
> > > +       struct cpumask all_mask = CPU_MASK_ALL;
> > > +       const struct cpumask *mask = &all_mask;
> > > +
> > > +       if (vq->irq_effective_cpu != -1)
> > > +               mask = get_cpu_mask(vq->irq_effective_cpu);
> >
> > Shouldn't this be vq->irq_affinity?
> >
>
> This sysfs interface is provided for effective irq affinity rather
> than irq affinity. We created another read-only sysfs interface for
> irq affinity.
>
> > > +
> > > +       return sprintf(buf, "%*pb\n", cpumask_pr_args(mask));
> > > +}
> > > +
> > > +static ssize_t irq_cb_effective_affinity_store(struct vduse_virtqueue *vq,
> > > +                                              const char *buf, size_t count)
> > > +{
> > > +       cpumask_var_t new_value;
> > > +       int ret;
> > > +
> > > +       if (!zalloc_cpumask_var(&new_value, GFP_KERNEL))
> > > +               return -ENOMEM;
> > > +
> > > +       ret = cpumask_parse(buf, new_value);
> > > +       if (ret)
> > > +               goto free_mask;
> > > +
> > > +       ret = -EINVAL;
> > > +       if (!cpumask_intersects(new_value, &vq->irq_affinity))
> > > +               goto free_mask;
> > > +
> > > +       spin_lock(&vq->irq_affinity_lock);
> > > +
> > > +       if (vq->irq_effective_cpu != -1)
> > > +               per_cpu(vduse_allocated_irq, vq->irq_effective_cpu) -= 1;
> > > +
> > > +       vq->irq_effective_cpu = cpumask_first(new_value);
> >
> > Does this mean except for the first cpu, the rest of the mask is unused?
> >
>
> Yes, but the user should always specify a mask that only contains one
> CPU to make it work as expected.

This doesn't seem to be the way that the IRQ affinity{hint} exported
via /sys work. Any reason for doing this?

(E.g we may have the require to limit the IRQ/callback to a NUMA node
instead of a specific cpu)

Thanks

> This sysfs interface is used to
> specify the effective irq affinity rather than irq affinity.
>
> Thanks,
> Yongji
>

_______________________________________________
Virtualization mailing list
Virtualization@lists.linux-foundation.org
https://lists.linuxfoundation.org/mailman/listinfo/virtualization

Re: [PATCH v2 09/11] vduse: Add enable_irq_wq sysfs interface for virtqueues

[Jason Wang]

On Mon, Dec 5, 2022 at 5:03 PM Xie Yongji <xieyongji@bytedance.com> wrote:
>
> Add enable_irq_wq sysfs interface to control whether
> use workqueue to inject irq or not. The vhost-vdpa case
> can benefit from it.

Do we have a benchmark result for this?

Or I wonder if we can extend set_vq_cb() by associating an eventfd
then VDUSE can signal that eventfd directly?

Thanks

>
> Signed-off-by: Xie Yongji <xieyongji@bytedance.com>
> ---
>  drivers/vdpa/vdpa_user/vduse_dev.c | 50 +++++++++++++++++++++++++++++-
>  1 file changed, 49 insertions(+), 1 deletion(-)
>
> diff --git a/drivers/vdpa/vdpa_user/vduse_dev.c b/drivers/vdpa/vdpa_user/vduse_dev.c
> index c65f84100e30..ed06c7afd484 100644
> --- a/drivers/vdpa/vdpa_user/vduse_dev.c
> +++ b/drivers/vdpa/vdpa_user/vduse_dev.c
> @@ -62,6 +62,7 @@ struct vduse_virtqueue {
>         struct cpumask irq_affinity;
>         spinlock_t irq_affinity_lock;
>         struct kobject kobj;
> +       bool enable_irq_wq;
>  };
>
>  struct vduse_dev;
> @@ -1013,6 +1014,26 @@ static int vduse_dev_queue_irq_work(struct vduse_dev *dev,
>         return ret;
>  }
>
> +static int vduse_dev_inject_vq_irq(struct vduse_dev *dev,
> +                                  struct vduse_virtqueue *vq)
> +{
> +       int ret = -EINVAL;
> +
> +       down_read(&dev->rwsem);
> +       if (!(dev->status & VIRTIO_CONFIG_S_DRIVER_OK))
> +               goto unlock;
> +
> +       ret = 0;
> +       spin_lock_irq(&vq->irq_lock);
> +       if (vq->ready && vq->cb.callback)
> +               vq->cb.callback(vq->cb.private);
> +       spin_unlock_irq(&vq->irq_lock);
> +unlock:
> +       up_read(&dev->rwsem);
> +
> +       return ret;
> +}
> +
>  static int vduse_dev_dereg_umem(struct vduse_dev *dev,
>                                 u64 iova, u64 size)
>  {
> @@ -1278,8 +1299,12 @@ static long vduse_dev_ioctl(struct file *file, unsigned int cmd,
>                         break;
>
>                 index = array_index_nospec(index, dev->vq_num);
> -               ret = vduse_dev_queue_irq_work(dev, &dev->vqs[index]->inject,
> +               if (dev->vqs[index]->enable_irq_wq)
> +                       ret = vduse_dev_queue_irq_work(dev,
> +                                       &dev->vqs[index]->inject,
>                                         dev->vqs[index]->irq_effective_cpu);
> +               else
> +                       ret = vduse_dev_inject_vq_irq(dev, dev->vqs[index]);
>                 break;
>         }
>         case VDUSE_IOTLB_REG_UMEM: {
> @@ -1420,6 +1445,26 @@ static const struct file_operations vduse_dev_fops = {
>         .llseek         = noop_llseek,
>  };
>
> +static ssize_t enable_irq_wq_show(struct vduse_virtqueue *vq, char *buf)
> +{
> +       return sprintf(buf, "%d\n", vq->enable_irq_wq);
> +}
> +
> +static ssize_t enable_irq_wq_store(struct vduse_virtqueue *vq,
> +                                  const char *buf, size_t count)
> +{
> +       bool enabled;
> +       int ret;
> +
> +       ret = kstrtobool(buf, &enabled);
> +       if (ret)
> +               return ret;
> +
> +       vq->enable_irq_wq = enabled;
> +
> +       return count;
> +}
> +
>  static ssize_t irq_cb_affinity_show(struct vduse_virtqueue *vq, char *buf)
>  {
>         return sprintf(buf, "%*pb\n", cpumask_pr_args(&vq->irq_affinity));
> @@ -1480,10 +1525,12 @@ struct vq_sysfs_entry {
>  static struct vq_sysfs_entry irq_cb_affinity_attr = __ATTR_RO(irq_cb_affinity);
>  static struct vq_sysfs_entry irq_cb_effective_affinity_attr =
>                                         __ATTR_RW(irq_cb_effective_affinity);
> +static struct vq_sysfs_entry enable_irq_wq_attr = __ATTR_RW(enable_irq_wq);
>
>  static struct attribute *vq_attrs[] = {
>         &irq_cb_affinity_attr.attr,
>         &irq_cb_effective_affinity_attr.attr,
> +       &enable_irq_wq_attr.attr,
>         NULL,
>  };
>  ATTRIBUTE_GROUPS(vq);
> @@ -1565,6 +1612,7 @@ static int vduse_dev_init_vqs(struct vduse_dev *dev, u32 vq_align, u32 vq_num)
>
>                 dev->vqs[i]->index = i;
>                 dev->vqs[i]->irq_effective_cpu = -1;
> +               dev->vqs[i]->enable_irq_wq = true;
>                 INIT_WORK(&dev->vqs[i]->inject, vduse_vq_irq_inject);
>                 INIT_WORK(&dev->vqs[i]->kick, vduse_vq_kick_work);
>                 spin_lock_init(&dev->vqs[i]->kick_lock);
> --
> 2.20.1
>

_______________________________________________
Virtualization mailing list
Virtualization@lists.linux-foundation.org
https://lists.linuxfoundation.org/mailman/listinfo/virtualization

Re: [PATCH v2 10/11] vduse: Delay iova domain creation

[Jason Wang]

On Mon, Dec 5, 2022 at 5:03 PM Xie Yongji <xieyongji@bytedance.com> wrote:
>
> Delay creating iova domain until the vduse device is binded
> to vdpa bus.

s/binded/registered/

Other than this.

Acked-by: Jason Wang <jasowang@redhat.com>

Thanks

>
> This is a preparation for adding sysfs interface to support
> specifying bounce buffer size for the iova domain.
>
> Signed-off-by: Xie Yongji <xieyongji@bytedance.com>
> ---
>  drivers/vdpa/vdpa_user/vduse_dev.c | 75 +++++++++++++++++++++---------
>  1 file changed, 53 insertions(+), 22 deletions(-)
>
> diff --git a/drivers/vdpa/vdpa_user/vduse_dev.c b/drivers/vdpa/vdpa_user/vduse_dev.c
> index ed06c7afd484..bd1ba6c33e09 100644
> --- a/drivers/vdpa/vdpa_user/vduse_dev.c
> +++ b/drivers/vdpa/vdpa_user/vduse_dev.c
> @@ -112,6 +112,8 @@ struct vduse_dev {
>         u32 vq_align;
>         struct vduse_umem *umem;
>         struct mutex mem_lock;
> +       unsigned int bounce_size;
> +       struct mutex domain_lock;
>  };
>
>  struct vduse_dev_msg {
> @@ -427,7 +429,7 @@ static void vduse_dev_reset(struct vduse_dev *dev)
>         struct vduse_iova_domain *domain = dev->domain;
>
>         /* The coherent mappings are handled in vduse_dev_free_coherent() */
> -       if (domain->bounce_map)
> +       if (domain && domain->bounce_map)
>                 vduse_domain_reset_bounce_map(domain);
>
>         down_write(&dev->rwsem);
> @@ -1045,6 +1047,9 @@ static int vduse_dev_dereg_umem(struct vduse_dev *dev,
>                 goto unlock;
>
>         ret = -EINVAL;
> +       if (!dev->domain)
> +               goto unlock;
> +
>         if (dev->umem->iova != iova || size != dev->domain->bounce_size)
>                 goto unlock;
>
> @@ -1071,7 +1076,7 @@ static int vduse_dev_reg_umem(struct vduse_dev *dev,
>         unsigned long npages, lock_limit;
>         int ret;
>
> -       if (!dev->domain->bounce_map ||
> +       if (!dev->domain || !dev->domain->bounce_map ||
>             size != dev->domain->bounce_size ||
>             iova != 0 || uaddr & ~PAGE_MASK)
>                 return -EINVAL;
> @@ -1145,7 +1150,6 @@ static long vduse_dev_ioctl(struct file *file, unsigned int cmd,
>                 struct vduse_iotlb_entry entry;
>                 struct vhost_iotlb_map *map;
>                 struct vdpa_map_file *map_file;
> -               struct vduse_iova_domain *domain = dev->domain;
>                 struct file *f = NULL;
>
>                 ret = -EFAULT;
> @@ -1156,8 +1160,13 @@ static long vduse_dev_ioctl(struct file *file, unsigned int cmd,
>                 if (entry.start > entry.last)
>                         break;
>
> -               spin_lock(&domain->iotlb_lock);
> -               map = vhost_iotlb_itree_first(domain->iotlb,
> +               mutex_lock(&dev->domain_lock);
> +               if (!dev->domain) {
> +                       mutex_unlock(&dev->domain_lock);
> +                       break;
> +               }
> +               spin_lock(&dev->domain->iotlb_lock);
> +               map = vhost_iotlb_itree_first(dev->domain->iotlb,
>                                               entry.start, entry.last);
>                 if (map) {
>                         map_file = (struct vdpa_map_file *)map->opaque;
> @@ -1167,7 +1176,8 @@ static long vduse_dev_ioctl(struct file *file, unsigned int cmd,
>                         entry.last = map->last;
>                         entry.perm = map->perm;
>                 }
> -               spin_unlock(&domain->iotlb_lock);
> +               spin_unlock(&dev->domain->iotlb_lock);
> +               mutex_unlock(&dev->domain_lock);
>                 ret = -EINVAL;
>                 if (!f)
>                         break;
> @@ -1319,8 +1329,10 @@ static long vduse_dev_ioctl(struct file *file, unsigned int cmd,
>                                  sizeof(umem.reserved)))
>                         break;
>
> +               mutex_lock(&dev->domain_lock);
>                 ret = vduse_dev_reg_umem(dev, umem.iova,
>                                          umem.uaddr, umem.size);
> +               mutex_unlock(&dev->domain_lock);
>                 break;
>         }
>         case VDUSE_IOTLB_DEREG_UMEM: {
> @@ -1334,15 +1346,15 @@ static long vduse_dev_ioctl(struct file *file, unsigned int cmd,
>                 if (!is_mem_zero((const char *)umem.reserved,
>                                  sizeof(umem.reserved)))
>                         break;
> -
> +               mutex_lock(&dev->domain_lock);
>                 ret = vduse_dev_dereg_umem(dev, umem.iova,
>                                            umem.size);
> +               mutex_unlock(&dev->domain_lock);
>                 break;
>         }
>         case VDUSE_IOTLB_GET_INFO: {
>                 struct vduse_iova_info info;
>                 struct vhost_iotlb_map *map;
> -               struct vduse_iova_domain *domain = dev->domain;
>
>                 ret = -EFAULT;
>                 if (copy_from_user(&info, argp, sizeof(info)))
> @@ -1356,18 +1368,24 @@ static long vduse_dev_ioctl(struct file *file, unsigned int cmd,
>                                  sizeof(info.reserved)))
>                         break;
>
> -               spin_lock(&domain->iotlb_lock);
> -               map = vhost_iotlb_itree_first(domain->iotlb,
> +               mutex_lock(&dev->domain_lock);
> +               if (!dev->domain) {
> +                       mutex_unlock(&dev->domain_lock);
> +                       break;
> +               }
> +               spin_lock(&dev->domain->iotlb_lock);
> +               map = vhost_iotlb_itree_first(dev->domain->iotlb,
>                                               info.start, info.last);
>                 if (map) {
>                         info.start = map->start;
>                         info.last = map->last;
>                         info.capability = 0;
> -                       if (domain->bounce_map && map->start == 0 &&
> -                           map->last == domain->bounce_size - 1)
> +                       if (dev->domain->bounce_map && map->start == 0 &&
> +                           map->last == dev->domain->bounce_size - 1)
>                                 info.capability |= VDUSE_IOVA_CAP_UMEM;
>                 }
> -               spin_unlock(&domain->iotlb_lock);
> +               spin_unlock(&dev->domain->iotlb_lock);
> +               mutex_unlock(&dev->domain_lock);
>                 if (!map)
>                         break;
>
> @@ -1390,7 +1408,10 @@ static int vduse_dev_release(struct inode *inode, struct file *file)
>  {
>         struct vduse_dev *dev = file->private_data;
>
> -       vduse_dev_dereg_umem(dev, 0, dev->domain->bounce_size);
> +       mutex_lock(&dev->domain_lock);
> +       if (dev->domain)
> +               vduse_dev_dereg_umem(dev, 0, dev->domain->bounce_size);
> +       mutex_unlock(&dev->domain_lock);
>         spin_lock(&dev->msg_lock);
>         /* Make sure the inflight messages can processed after reconncection */
>         list_splice_init(&dev->recv_list, &dev->send_list);
> @@ -1647,6 +1668,7 @@ static struct vduse_dev *vduse_dev_create(void)
>
>         mutex_init(&dev->lock);
>         mutex_init(&dev->mem_lock);
> +       mutex_init(&dev->domain_lock);
>         spin_lock_init(&dev->msg_lock);
>         INIT_LIST_HEAD(&dev->send_list);
>         INIT_LIST_HEAD(&dev->recv_list);
> @@ -1696,7 +1718,8 @@ static int vduse_destroy_dev(char *name)
>         idr_remove(&vduse_idr, dev->minor);
>         kvfree(dev->config);
>         vduse_dev_deinit_vqs(dev);
> -       vduse_domain_destroy(dev->domain);
> +       if (dev->domain)
> +               vduse_domain_destroy(dev->domain);
>         kfree(dev->name);
>         vduse_dev_destroy(dev);
>         module_put(THIS_MODULE);
> @@ -1802,11 +1825,7 @@ static int vduse_create_dev(struct vduse_dev_config *config,
>         if (!dev->name)
>                 goto err_str;
>
> -       dev->domain = vduse_domain_create(VDUSE_IOVA_SIZE - 1,
> -                                         VDUSE_BOUNCE_SIZE);
> -       if (!dev->domain)
> -               goto err_domain;
> -
> +       dev->bounce_size = VDUSE_BOUNCE_SIZE;
>         dev->config = config_buf;
>         dev->config_size = config->config_size;
>
> @@ -1836,8 +1855,6 @@ static int vduse_create_dev(struct vduse_dev_config *config,
>  err_dev:
>         idr_remove(&vduse_idr, dev->minor);
>  err_idr:
> -       vduse_domain_destroy(dev->domain);
> -err_domain:
>         kfree(dev->name);
>  err_str:
>         vduse_dev_destroy(dev);
> @@ -2004,9 +2021,23 @@ static int vdpa_dev_add(struct vdpa_mgmt_dev *mdev, const char *name,
>         if (ret)
>                 return ret;
>
> +       mutex_lock(&dev->domain_lock);
> +       if (!dev->domain)
> +               dev->domain = vduse_domain_create(VDUSE_IOVA_SIZE - 1,
> +                                                 dev->bounce_size);
> +       mutex_unlock(&dev->domain_lock);
> +       if (!dev->domain) {
> +               put_device(&dev->vdev->vdpa.dev);
> +               return -ENOMEM;
> +       }
> +
>         ret = _vdpa_register_device(&dev->vdev->vdpa, dev->vq_num);
>         if (ret) {
>                 put_device(&dev->vdev->vdpa.dev);
> +               mutex_lock(&dev->domain_lock);
> +               vduse_domain_destroy(dev->domain);
> +               dev->domain = NULL;
> +               mutex_unlock(&dev->domain_lock);
>                 return ret;
>         }
>
> --
> 2.20.1
>

_______________________________________________
Virtualization mailing list
Virtualization@lists.linux-foundation.org
https://lists.linuxfoundation.org/mailman/listinfo/virtualization

Re: [PATCH v2 11/11] vduse: Support specifying bounce buffer size via sysfs

[Jason Wang]

On Mon, Dec 5, 2022 at 5:03 PM Xie Yongji <xieyongji@bytedance.com> wrote:
>
> As discussed in [1], this adds sysfs interface to support
> specifying bounce buffer size in virtio-vdpa case. It would
> be a performance tuning parameter for high throughput workloads.
>
> [1] https://www.spinics.net/lists/netdev/msg754288.html
>
> Signed-off-by: Xie Yongji <xieyongji@bytedance.com>

Acked-by: Jason Wang <jasowang@redhat.com>

Thanks

> ---
>  drivers/vdpa/vdpa_user/vduse_dev.c | 45 +++++++++++++++++++++++++++++-
>  1 file changed, 44 insertions(+), 1 deletion(-)
>
> diff --git a/drivers/vdpa/vdpa_user/vduse_dev.c b/drivers/vdpa/vdpa_user/vduse_dev.c
> index bd1ba6c33e09..0458d61cee1f 100644
> --- a/drivers/vdpa/vdpa_user/vduse_dev.c
> +++ b/drivers/vdpa/vdpa_user/vduse_dev.c
> @@ -38,8 +38,11 @@
>  #define DRV_LICENSE  "GPL v2"
>
>  #define VDUSE_DEV_MAX (1U << MINORBITS)
> +#define VDUSE_MAX_BOUNCE_SIZE (1024 * 1024 * 1024)
> +#define VDUSE_MIN_BOUNCE_SIZE (1024 * 1024)
>  #define VDUSE_BOUNCE_SIZE (64 * 1024 * 1024)
> -#define VDUSE_IOVA_SIZE (128 * 1024 * 1024)
> +/* 128 MB reserved for virtqueue creation */
> +#define VDUSE_IOVA_SIZE (VDUSE_MAX_BOUNCE_SIZE + 128 * 1024 * 1024)
>  #define VDUSE_MSG_DEFAULT_TIMEOUT 30
>
>  struct vduse_virtqueue {
> @@ -1795,8 +1798,48 @@ static ssize_t msg_timeout_store(struct device *device,
>
>  static DEVICE_ATTR_RW(msg_timeout);
>
> +static ssize_t bounce_size_show(struct device *device,
> +                               struct device_attribute *attr, char *buf)
> +{
> +       struct vduse_dev *dev = dev_get_drvdata(device);
> +
> +       return sysfs_emit(buf, "%u\n", dev->bounce_size);
> +}
> +
> +static ssize_t bounce_size_store(struct device *device,
> +                                struct device_attribute *attr,
> +                                const char *buf, size_t count)
> +{
> +       struct vduse_dev *dev = dev_get_drvdata(device);
> +       unsigned int bounce_size;
> +       int ret;
> +
> +       ret = -EPERM;
> +       mutex_lock(&dev->domain_lock);
> +       if (dev->domain)
> +               goto unlock;
> +
> +       ret = kstrtouint(buf, 10, &bounce_size);
> +       if (ret < 0)
> +               goto unlock;
> +
> +       ret = -EINVAL;
> +       if (bounce_size > VDUSE_MAX_BOUNCE_SIZE ||
> +           bounce_size < VDUSE_MIN_BOUNCE_SIZE)
> +               goto unlock;
> +
> +       dev->bounce_size = bounce_size & PAGE_MASK;
> +       ret = count;
> +unlock:
> +       mutex_unlock(&dev->domain_lock);
> +       return ret;
> +}
> +
> +static DEVICE_ATTR_RW(bounce_size);
> +
>  static struct attribute *vduse_dev_attrs[] = {
>         &dev_attr_msg_timeout.attr,
> +       &dev_attr_bounce_size.attr,
>         NULL
>  };
>
> --
> 2.20.1
>

_______________________________________________
Virtualization mailing list
Virtualization@lists.linux-foundation.org
https://lists.linuxfoundation.org/mailman/listinfo/virtualization