Re: [PATCH v2 08/12] drivers: export device_is_bound()

[Bartosz Golaszewski]

On Fri, Mar 5, 2021 at 12:27 PM Greg KH <gregkh@xxxxxxxxxxxxxxxxxxx> wrote:
>
> On Fri, Mar 05, 2021 at 11:58:18AM +0100, Bartosz Golaszewski wrote:
> > On Fri, Mar 5, 2021 at 11:24 AM Greg KH <gregkh@xxxxxxxxxxxxxxxxxxx> wrote:
> > >
> > > On Fri, Mar 05, 2021 at 10:16:10AM +0100, Bartosz Golaszewski wrote:
> > > > On Fri, Mar 5, 2021 at 9:55 AM Greg KH <gregkh@xxxxxxxxxxxxxxxxxxx> wrote:
> > > > >
> > > > > On Fri, Mar 05, 2021 at 09:45:41AM +0100, Bartosz Golaszewski wrote:
> > > > > > On Fri, Mar 5, 2021 at 9:34 AM Greg KH <gregkh@xxxxxxxxxxxxxxxxxxx> wrote:
> > > > > > >
> > > > > > > On Fri, Mar 05, 2021 at 09:18:30AM +0100, Geert Uytterhoeven wrote:
> > > > > > > > CC Greg
> > > > > > > >
> > > > > > > > On Thu, Mar 4, 2021 at 11:30 AM Bartosz Golaszewski <brgl@xxxxxxxx> wrote:
> > > > > > > > >
> > > > > > > > > From: Bartosz Golaszewski <bgolaszewski@xxxxxxxxxxxx>
> > > > > > > > >
> > > > > > > > > Export the symbol for device_is_bound() so that we can use it in gpio-sim
> > > > > > > > > to check if the simulated GPIO chip is bound before fetching its driver
> > > > > > > > > data from configfs callbacks in order to retrieve the name of the GPIO
> > > > > > > > > chip device.
> > > > > > > > >
> > > > > > > > > Signed-off-by: Bartosz Golaszewski <bgolaszewski@xxxxxxxxxxxx>
> > > > > > > > > ---
> > > > > > > > >  drivers/base/dd.c | 1 +
> > > > > > > > >  1 file changed, 1 insertion(+)
> > > > > > > > >
> > > > > > > > > diff --git a/drivers/base/dd.c b/drivers/base/dd.c
> > > > > > > > > index 9179825ff646..c62c02e3490a 100644
> > > > > > > > > --- a/drivers/base/dd.c
> > > > > > > > > +++ b/drivers/base/dd.c
> > > > > > > > > @@ -353,6 +353,7 @@ bool device_is_bound(struct device *dev)
> > > > > > > > >  {
> > > > > > > > >         return dev->p && klist_node_attached(&dev->p->knode_driver);
> > > > > > > > >  }
> > > > > > > > > +EXPORT_SYMBOL_GPL(device_is_bound);
> > > > > > >
> > > > > > > No.  Please no.  Why is this needed?  Feels like someone is doing
> > > > > > > something really wrong...
> > > > > > >
> > > > > > > NACK.
> > > > > > >
> > > > > >
> > > > > > I should have Cc'ed you the entire series, my bad.
> > > > > >
> > > > > > This is the patch that uses this change - it's a new, improved testing
> > > > > > module for GPIO using configfs & sysfs as you (I think) suggested a
> > > > > > while ago:
> > > > > >
> > > > > > https://lkml.org/lkml/2021/3/4/355
> > > > > >
> > > > > > The story goes like this: committing the configfs item registers a
> > > > > > platform device.
> > > > >
> > > > > Ick, no, stop there, that's not a "real" device, please do not abuse
> > > > > platform devices like that, you all know I hate this :(
> > > > >
> > > > > Use the virtbus code instead perhaps?
> > > > >
> > > >
> > > > I have no idea what virtbus is and grepping for it only returns three
> > > > hits in: ./drivers/pci/iov.c and it's a function argument.
> > > >
> > > > If it stands for virtual bus then for sure it sounds like the right
> > > > thing but I need to find more info on this.
> > >
> > > Sorry, wrong name, see Documentation/driver-api/auxiliary_bus.rst for
> > > the details.  "virtbus" was what I think about it as that was my
> > > original name for it, but it eventually got merged with a different
> > > name.
> > >

Unless I'm not seeing something - it completely doesn't look like the
right solution. This auxiliary bus sounds like MFD with extra steps.
Its aim seems to be to provide virtual devices for sub-modules of real
devices.

What I have here really is a dummy device for which no HW exists.
Also: while the preferred way is to use configfs to instantiate these
simulated devices, then can still be registered from device-tree (this
is a feature that was requested and eventually implemented in
gpio-mockup which we want to phase out so we can't just drop it).
AFAIK only platform devices can be populated from DT.

I guess we could create something like a "virtual bus" that would be
there for devices that don't exist on any physical bus but this would
end up in big part being the same thing as platform devices.

> > > > > > As far as I understand - there's no guarantee that
> > > > > > the device will be bound to a driver before the commit callback (or
> > > > > > more specifically platform_device_register_full() in this case)
> > > > > > returns so the user may try to retrieve the name of the device
> > > > > > immediately (normally user-space should wait for the associated uevent
> > > > > > but nobody can force that) by doing:
> > > > > >
> > > > > > mv /sys/kernel/config/gpio-sim/pending/foo /sys/kernel/config/gpio-sim/live/
> > > > > > cat /sys/kernel/config/gpio-sim/live/foo/dev_name
> > > > > >
> > > > > > If the device is not bound at this point, we'll have a crash in the
> > > > > > kernel as opposed to just returning -ENODEV.
> > > > >
> > > > > How will the kernel crash?  What has created the dev_name sysfs file
> > > > > before it is possible to be read from?  That feels like the root
> > > > > problem.
> > > > >
> > > >
> > > > It's not sysfs - it's in configfs. Each chip has a read-only configfs
> > > > attribute that returns the name of the device - I don't really have a
> > > > better idea to map the configfs items to devices that committing
> > > > creates.
> > >
> > > Same question, why are you exporting a configfs attribute that can not
> > > be read from?  Only export it when your driver is bound to the device.
> > >
> >
> > The device doesn't know anything about configfs. Why would it? The
> > configuration of a GPIO chip can't be changed after it's instantiated,
> > this is why we have committable items.
> >
> > We export a directory in configfs: gpio-sim -> user creates a new
> > directory (item) in gpio-sim/pending/foo and it's not tied to any
> > device yet but exports attributes which we use to configure the device
> > (label, number of lines, line names etc.), then we mv
> > gpio-sim/pending/foo gpio-sim/live and this is when the device gets
> > created and registered with the subsystem. We take all the configured
> > attributes and put them into device properties for both the driver and
> > gpiolib core (for standard properties) to read - just like we would
> > with a regular GPIO driver because this is the goal: test the core
> > code.
>
> Ok, but they why are you trying to have dev_name be an exported thing?
> I don't understand an attribute here that is visable but can not be read
> from.
>

Because once the associated configfs item is committed and the device
created, it will become readable. The list of attributes is fixed in
configfs. I'm not sure what the better approach would be - return
"none" if the device handle is NULL?

> And why not just use the default device name function: dev_name(), which
> will always return a string that will work no matter if the device is
> bound to a driver or not.
>

I can do this but then it's possible that user-space gets the name of
the device which doesn't exist in sysfs. I guess we can mention that
in the documentation.

Bartosz