Re: [PATCH v4] Drivers: hv: vmbus: Expose monitor data only when monitor pages are used

From: Greg KH
Date: Mon Mar 04 2019 - 02:38:49 EST


On Sun, Mar 03, 2019 at 04:11:28PM -0500, Kimberly Brown wrote:
> On Sun, Mar 03, 2019 at 09:05:43AM +0100, Greg KH wrote:
> > On Fri, Mar 01, 2019 at 02:18:24PM -0500, Kimberly Brown wrote:
> > > +/*
> > > + * Channel-level attribute_group callback function. Returns the permission for
> > > + * each attribute, and returns 0 if an attribute is not visible.
> > > + */
> > > +static umode_t vmbus_chan_attr_is_visible(struct kobject *kobj,
> > > + struct attribute *attr, int idx)
> > > +{
> > > + const struct vmbus_channel *channel =
> > > + container_of(kobj, struct vmbus_channel, kobj);
> > > +
> > > + /* Hide the monitor attributes if the monitor mechanism is not used. */
> > > + if (!channel->offermsg.monitor_allocated &&
> > > + (attr == &chan_attr_pending.attr ||
> > > + attr == &chan_attr_latency.attr ||
> > > + attr == &chan_attr_monitor_id.attr))
> > > + return 0;
> > > +
> > > + return attr->mode;
> > > +}
> > > +
> > > +static struct attribute_group vmbus_chan_group = {
> > > + .attrs = vmbus_chan_attrs,
> > > + .is_visible = vmbus_chan_attr_is_visible
> > > +};
> > > +
> > > static struct kobj_type vmbus_chan_ktype = {
> > > .sysfs_ops = &vmbus_chan_sysfs_ops,
> > > .release = vmbus_chan_release,
> > > - .default_attrs = vmbus_chan_attrs,
> >
> > Why did you remove this line?
>
> I removed the default attributes because vmbus_chan_attrs contains
> non-default attributes. You suggested that I use one attribute_group and
> an is_visible() callback for the device-level attributes (see
> https://lore.kernel.org/lkml/20190226081848.GA15659@xxxxxxxxx/). I
> assumed (possibly incorrectly) that I should do the same for these
> channel-level attributes.

That is fine to have a callback, but why did you have to remove the
default attribute pointer? The two should have nothing to do with each
other.

> > > };
> > >
> > > /*
> > > @@ -1571,6 +1624,12 @@ int vmbus_add_channel_kobj(struct hv_device *dev, struct vmbus_channel *channel)
> > > if (ret)
> > > return ret;
> > >
> > > + ret = sysfs_create_group(kobj, &vmbus_chan_group);
> >
> > Why are you adding these "by hand"? What was wrong with using the
> > default attribute group pointer? You also are not removing the
> > attributes :(
>
> Are you referring to default_attrs in kobj_type? It's not an
> attribute_group pointer, it's a pointer to an attribute pointer array.
> The problem with using default_attrs is that all of the attributes are
> visible.

It shouldn't, the is_visable() callback will be called on each attribute
when the group is created by the core. Did that not work properly when
you tested this?

> I'm fairly certain that the monitor attributes are being removed.
> sysfs_create_group() uses the attribute_group's is_visible() callback to
> control the attribute visibility. And, when I look at the sysfs files, I
> can see that the monitor sysyfs files are removed.

I mean you are not removing the group when the device goes away, not
that the individual files are not present. If you leave the pointer to
default_attributes there, the core will properly remove the sysfs
attributes when the device is removed from the system. Otherwise you
just "get lucky" if the attributes are removed or not.

> In v3, I proposed moving the monitor attributes to a special
> attribute_group and adding that group manually when needed. Do you
> prefer that approach for the channel-level monitor attributes?

No need for a special group here, just use the is_visable() callback
like you currently are, all should be fine. I think you are adding more
code than you need to in order to get this to all work properly :)

thanks,

greg k-h