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

[Yongji Xie]

On Tue, Dec 20, 2022 at 2:29 PM Jason Wang <jasowang@xxxxxxxxxx> wrote:
>
> On Mon, Dec 19, 2022 at 1:16 PM Yongji Xie <xieyongji@xxxxxxxxxxxxx> wrote:
> >
> > On Fri, Dec 16, 2022 at 1:35 PM Jason Wang <jasowang@xxxxxxxxxx> wrote:
> > >
> > > On Mon, Dec 5, 2022 at 5:03 PM Xie Yongji <xieyongji@xxxxxxxxxxxxx> 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@xxxxxxxxxxxxx>
> > > > ---
> > > >  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)
>

Yes, I think we need to make the sysfs interface for irq affinity
writable. The effective irq affinity can be removed now since we
choose using round-robin to spread IRQs between CPUs.

Thanks,
Yongji