Re: [PATCH 0/2] Add device_create_files() and device_remove_files() helpers

[Greg Kroah-Hartman <gregkh@linuxfoundation.org>]

On Fri, Jan 30, 2015 at 05:31:51PM +0100, Takashi Iwai wrote:
> At Thu, 29 Jan 2015 20:26:26 -0800,
> Greg Kroah-Hartman wrote:
> > 
> > On Thu, Jan 29, 2015 at 12:11:21AM +0100, Takashi Iwai wrote:
> > > At Wed, 28 Jan 2015 14:28:51 -0800,
> > > Greg Kroah-Hartman wrote:
> > > > 
> > > > On Wed, Jan 28, 2015 at 11:18:57PM +0100, Takashi Iwai wrote:
> > > > > At Wed, 28 Jan 2015 13:34:21 -0800,
> > > > > Greg Kroah-Hartman wrote:
> > > > > > 
> > > > > > On Wed, Jan 28, 2015 at 10:26:28PM +0100, Takashi Iwai wrote:
> > > > > > > At Wed, 28 Jan 2015 13:05:47 -0800,
> > > > > > > Greg Kroah-Hartman wrote:
> > > > > > > > 
> > > > > > > > On Wed, Jan 28, 2015 at 09:46:12PM +0100, Takashi Iwai wrote:
> > > > > > > > > Hi,
> > > > > > > > > 
> > > > > > > > > this is a simple patch to add device_create_files() and
> > > > > > > > > device_remove_files() to replace multiple device_create_file() or
> > > > > > > > > _remove() calls with a single shot with the device_attr list.
> > > > > > > > > 
> > > > > > > > > It's basically just a clean up, but also helps to simplify the error
> > > > > > > > > handling a lot in many existing codes since the function itself does
> > > > > > > > > rollback at error.
> > > > > > > > > 
> > > > > > > > > The series contains a patch to apply these to drivers/base/node.c.
> > > > > > > > > I have lots of patches (up to 30) to use these in the whole tree, but
> > > > > > > > > maybe it'd be easier too apply once after this stuff is merged at
> > > > > > > > > first.  It's just a cleanup so no urgent task, after all.
> > > > > > > > 
> > > > > > > > I'd like to some day be able to drop device_create_file entirely, as it
> > > > > > > > is almost always used in a racy way (but not always, so we can't get rid
> > > > > > > > of it today.)
> > > > > > > > 
> > > > > > > > A driver should be using an attribute group and be created/registered
> > > > > > > > with it if they want any files associated with it, so giving people the
> > > > > > > > ability to add large numbers of files all at once seems like the wrong
> > > > > > > > thing to do :)
> > > > > > > 
> > > > > > > Well, through the glance over many codes using device_create_file(),
> > > > > > > I think the problem of the attribute group is that there is little
> > > > > > > help for generating the entries dynamically.  For example, if you have
> > > > > > > two groups you want to enable conditionally, what would be the best
> > > > > > > way to implement?
> > > > > > 
> > > > > > Use the is_visable() function callback, that's what it is there for.
> > > > > 
> > > > > But if the entries are determined dynamically?  Selecting the enabled
> > > > > elements from the static list is one way, it'd work in many cases, but
> > > > > it's not always the most straightforward way.  It often would be
> > > > > easier to build up the list dynamically.
> > > > 
> > > > Do you have an example of this?  Wouldn't it be the same thing to list
> > > > them all in an attribute group, but only say "this is valid" in the
> > > > is_visable() callback for those that would have been built up
> > > > dynamically?
> > > 
> > > One common scene is the case where a device has already the static
> > > group defined in the core helper module while a driver wants to put
> > > additional sysfs entries on it.
> > > 
> > > A complex case is something in drivers/leds/trigger/ledtrig-*.c.
> > > 
> > > Another interesting example is drivers/regulator/core.c.  It creates a
> > > bunch of various sysfs files depending on the client's ops presence.
> > > It might be implemented via is_visible, but then it'd become more
> > > lengthy (too many small callback functions).
> > 
> > Yeah, I'm not saying it's easy, or simple, it's just the only way I know
> > how to do this in a race-free way.  We have to create the files before
> > the uevent happens, not after, like these drivers are doing.
> > 
> > If you can think of a way that we can do this in a simpler way, that
> > would be great.
> 
> The latter one (regulator/core.c) is actually a case where is_visible
> callback would work better, I noticed after studying mode code.
> Thanks for hints.  I'm going to submit a patch later.
> 
> OTOH, the leds class looks not intuitive.  Need more investigation.
> 
> > > Also, multiple drivers seem calling device_create_file() from the
> > > array of attributes in a loop.  One reason might be that it's easier
> > > to write for a bunch of entries, without defining many piece of
> > > structs.  An example is found in drivers/gpu/drm/drm_sysfs.c.
> > 
> > That one should just be adding the whole attribute group, using
> > device_add_groups, which we have in the driver core, but I didn't export
> > publicly.  That is if those are being added in a race-free way, I
> > couldn't unwind the drm mess to see if the uevent is happening after the
> > files are added or before.
> 
> If we export device_add_groups() and device_remove_groups(), is it
> safe to call it before device_add()?  If yes, some drivers/subsystems
> can have a code flow like:
> 
> some_subsystem_init(struct device *dev)
> {
> 	device_initialize(dev);
> 	devs->groups = subsystem_groups;
> 	....
> }
> 
> driver_init(struct device *dev)
> {
> 	some_subsystem_init(dev);
> 	device_add_groups(dev, additional_groups);
> 	....
> 	device_add(dev);
> 	....
> }
> 
> The network device has a own multi dev_groups array so that the driver
> can put an own group while the net core fills common groups
> dynamically just before the device registration call.  I though of
> implementing similar for others (including the sound stuff), but if
> the scheme above works, the rewrite will become smaller.
> 
> Of corse, the drawback of the explicit device_add_groups() call would
> be that you'll have to call device_remove_groups() at removal or error
> paths.

Right now, no, you can't call device_add_groups() until after
device_add() happens, as it device_initialize() doesn't do enough sysfs
work in order to be able to create the files.

> > > > > What if having a link to the chained group for appending entries
> > > > > dynamically?  Just a wild idea, but it might make things easier.
> > > > 
> > > > We have the ability to pass a group list pointer to device_create
> > > > already, and the attribute pointer is a list of groups as well, how can
> > > > we change this to be "easier"?
> > > 
> > > I guess the order is the problem.  In many cases, you know the
> > > additional entries only after the device creation.  The device
> > > creation is often done by a helper code.  So the driver has no control
> > > to it, just gets the resultant device.
> > 
> > Yeah, that's the problem.  And another problem is drivers adding
> > attributes to devices after they are bound to a device, which is kind of
> > pointless, as the uevent is long past at that point in time.  I've
> > cleaned up a bunch of those, but odds are there are still more to fix.
> 
> Right, there are a bunch of drivers doing it.  I guess partly because
> they don't need uevents for creation, but also partly because there is
> no way to give attribute groups properly in some cases.  For example,
> misc_register() or register_framebuffer() calls device_create() so the
> caller can't pass groups.
> 
> It'd be trivial to extend struct miscdevice to carry an optional group
> field and change the call to device_create_with_groups().  But,
> fb_info has also common sysfs entries, so it'd need also the solution
> above with device_add_groups() in addition.

Your patch to do that looks good, I'll queue them all up after 3.20-rc1
is out as it's too close to 3.19 at the moment.

thanks,

greg k-h