Re: [PATCH] base: soc: Export soc_device_to_device API

[Greg KH]

On Mon, Sep 23, 2019 at 02:35:33PM -0700, mnalajal@xxxxxxxxxxxxxx wrote:
> On 2019-09-19 23:10, Greg KH wrote:
> > On Thu, Sep 19, 2019 at 08:36:51PM -0700, Bjorn Andersson wrote:
> > > On Thu 19 Sep 15:45 PDT 2019, Greg KH wrote:
> > > 
> > > > On Thu, Sep 19, 2019 at 03:40:17PM -0700, Bjorn Andersson wrote:
> > > > > On Thu 19 Sep 15:25 PDT 2019, Greg KH wrote:
> > > > >
> > > > > > On Thu, Sep 19, 2019 at 03:14:56PM -0700, Bjorn Andersson wrote:
> > > > > > > On Thu 19 Sep 14:58 PDT 2019, Greg KH wrote:
> > > > > > >
> > > > > > > > On Thu, Sep 19, 2019 at 02:53:00PM -0700, Bjorn Andersson wrote:
> > > > > > > > > On Thu 19 Sep 14:32 PDT 2019, Greg KH wrote:
> > > > > > > > >
> > > > > > > > > > On Thu, Sep 19, 2019 at 02:13:44PM -0700, Murali Nalajala wrote:
> > > > > > > > > > > If the soc drivers want to add custom sysfs entries it needs to
> > > > > > > > > > > access "dev" field in "struct soc_device". This can be achieved
> > > > > > > > > > > by "soc_device_to_device" API. Soc drivers which are built as a
> > > > > > > > > > > module they need above API to be exported. Otherwise one can
> > > > > > > > > > > observe compilation issues.
> > > > > > > > > > >
> > > > > > > > > > > Signed-off-by: Murali Nalajala <mnalajal@xxxxxxxxxxxxxx>
> > > > > > > > > > > ---
> > > > > > > > > > >  drivers/base/soc.c | 1 +
> > > > > > > > > > >  1 file changed, 1 insertion(+)
> > > > > > > > > > >
> > > > > > > > > > > diff --git a/drivers/base/soc.c b/drivers/base/soc.c
> > > > > > > > > > > index 7c0c5ca..4ad52f6 100644
> > > > > > > > > > > --- a/drivers/base/soc.c
> > > > > > > > > > > +++ b/drivers/base/soc.c
> > > > > > > > > > > @@ -41,6 +41,7 @@ struct device *soc_device_to_device(struct soc_device *soc_dev)
> > > > > > > > > > >  {
> > > > > > > > > > >  	return &soc_dev->dev;
> > > > > > > > > > >  }
> > > > > > > > > > > +EXPORT_SYMBOL_GPL(soc_device_to_device);
> > > > > > > > > > >
> > > > > > > > > > >  static umode_t soc_attribute_mode(struct kobject *kobj,
> > > > > > > > > > >  				struct attribute *attr,
> > > > > > > > > >
> > > > > > > > > > What in-kernel driver needs this?
> > > > > > > > > >
> > > > > > > > >
> > > > > > > > > Half of the drivers interacting with the soc driver calls this API,
> > > > > > > > > several of these I see no reason for being builtin (e.g.
> > > > > > > > > ux500 andversatile). So I think this patch makes sense to allow us to
> > > > > > > > > build these as modules.
> > > > > > > > >
> > > > > > > > > > Is linux-next breaking without this?
> > > > > > > > > >
> > > > > > > > >
> > > > > > > > > No, we postponed the addition of any sysfs attributes in the Qualcomm
> > > > > > > > > socinfo driver.
> > > > > > > > >
> > > > > > > > > > We don't export things unless we have a user of the export.
> > > > > > > > > >
> > > > > > > > > > Also, adding "custom" sysfs attributes is almost always not the correct
> > > > > > > > > > thing to do at all.  The driver should be doing it, by setting up the
> > > > > > > > > > attribute group properly so that the driver core can do it automatically
> > > > > > > > > > for it.
> > > > > > > > > >
> > > > > > > > > > No driver should be doing individual add/remove of sysfs files.  If it
> > > > > > > > > > does so, it is almost guaranteed to be doing it incorrectly and racing
> > > > > > > > > > userspace.
> > > > > > > > > >
> > > > > > > > >
> > > > > > > > > The problem here is that the attributes are expected to be attached to
> > > > > > > > > the soc driver, which is separate from the platform-specific drivers. So
> > > > > > > > > there's no way to do platform specific attributes the right way.
> > > > > > > > >
> > > > > > > > > > And yes, there's loads of in-kernel examples of doing this wrong, I've
> > > > > > > > > > been working on fixing that up, look at the patches now in Linus's tree
> > > > > > > > > > for platform and USB drivers that do this as examples of how to do it
> > > > > > > > > > right.
> > > > > > > > > >
> > > > > > > > >
> > > > > > > > > Agreed, this patch should not be used as an approval for any crazy
> > > > > > > > > attributes; but it's necessary in order to extend the soc device's
> > > > > > > > > attributes, per the current design.
> > > > > > > >
> > > > > > > > Wait, no, let's not let the "current design" remain if it is broken!
> > > > > > > >
> > > > > > > > Why can't the soc driver handle the attributes properly so that the
> > > > > > > > individual driver doesn't have to do the create/remove?
> > > > > > > >
> > > > > > >
> > > > > > > The custom attributes that these drivers want to add to the common ones
> > > > > > > are known in advance, so I presume we could have them passed into
> > > > > > > soc_device_register() and registered together with the common
> > > > > > > attributes...
> > > > > > >
> > > > > > > It sounds like it's worth a prototype.
> > > > > >
> > > > > > Do you have an in-kernel example I can look at to get an idea of what is
> > > > > > needed here?
> > > > > >
> > > > >
> > > > > realview_soc_probe(), in drivers/soc/versatile/soc-realview.c,
> > > > > implements the current mechanism of acquiring the soc's struct device
> > > > > and then issuing a few device_create_file calls on that.
> > > >
> > > > That looks to be a trivial driver to fix up.  Look at 6d03c140db2e
> > > > ("USB: phy: fsl-usb: convert platform driver to use dev_groups") as an
> > > > example of how to do this.
> > > >
> > > 
> > > The difference between the two cases is that in the fsl-usb case it's
> > > attributes of the device itself, while in the soc case the
> > > realview-soc
> > > driver (or the others doing this) calls soc_device_register() to
> > > register a new (dangling) soc device, which it then adds its
> > > attributes
> > > onto.
> > 
> > That sounds really really odd.  Why can't the soc device do the creation
> > "automatically" when the device is registered?  The soc core should
> > handle this for the soc "drivers", that's what it is there for.
> > 
> Clients are registering to soc framework using "soce_device_register()"
> with "soc_device_attribute". This attribute structure does not have all
> the sysfs fields what client are interested. Hence clients are handling
> their required sysfs fields in their drivers.
> https://elixir.bootlin.com/linux/v5.3/source/drivers/base/soc.c#L114

Then you should fix that :)

> > > We can't use dev_groups, because the soc_device (soc.c) isn't
> > > actually a
> > > driver and the list of attributes is a combination of things from
> > > soc.c
> > > and e.g. soc-realview.c.
> > > 
> > > But if we pass a struct attribute_group into soc_device_register() and
> > > then have that register both groups using dev.groups, this should be
> > > much cleaner at least.
> > 
> > Don't you have a structure you can store these in as well?
> At present client is populating entries one-by-one
> https://android.googlesource.com/kernel/msm/+/android-7.1.0_r0.2/drivers/soc/qcom/socinfo.c#1254

And that is known to be broken and racy and will cause problems with
userspace.  This should be fixed...

thanks,

greg k-h