device_node lifetime (was: Re: [PATCH 1/7] phy: brcmstb-sata: add missing of_node_put)

device_node lifetime (was: Re: [PATCH 1/7] phy: brcmstb-sata: add missing of_node_put)

[Brian Norris]

(changing subject, add devicetree-u79uwXL29TY76Z2rM5mHXA@public.gmane.org)

On Tue, Nov 17, 2015 at 11:33:25PM +0100, Julia Lawall wrote:
> On Tue, 17 Nov 2015, Brian Norris wrote:
> > On Tue, Nov 17, 2015 at 06:48:39PM +0100, Julia Lawall wrote:
> > > Is this something that should be checked for elsewhere?
> > 
> > I expect the same sort of problem shows up plenty of other places. I
> > don't think many people use CONFIG_OF_DYNAMIC, so the effects of these
> > failures probably aren't felt by many.
> 
> I tried the following semantic patch:
> 
> @@
> struct device_node *e;
> expression e1;
> identifier fld;
> @@
> 
>  ... when != of_node_get(...)
> *(<+...e1->fld...+>) = e
>  ... when != of_node_get(...)
>  return e1;
> 
> basically, this says that a structure field is initilized to a device node 
> value, the structure is returned by the containing function, and the 
> containing function contains no of_node_get at all.  Certainly this is 
> quite constrained, but it does produce a number of examples.
> 
> I looked at a few of them:
> 
> drivers/clk/ingenic/cgu.c, ingenic_cgu_new
> clk/pistachio/clk.c, pistachio_clk_alloc_provider

It looks like the clock core (drivers/clk/clk.c) initially grabs the clk
provider node in of_clk_init(), then drops it after it's initialized,
but most of these providers use of_clk_add_provider(), which seems to
manage the device_node lifetime for the user. So I think these are OK.

> drivers/mfd/syscon.c, of_syscon_register

This one looks potentially suspect. Syscon nodes aren't usually directly
managed by a single driver, and the device_node pointer is used for
lookups later...so I think it should keep a kref, and it doesn't.

> drivers/of/pdt.c, function of_pdt_create_node

Not real sure about this one.

> Any idea whether these need of_node_get?  In all cases the device node 
> value comes in as a parameter.

I'm really not an expert on this stuff. I just saw a potential problem
that I happen to be looking at in other subsystems, and I wanted to know
what others thought. I think this discussion should include the DT folks
and the subsystems in question. For one, I'm as interested as anyone in
getting this todo clarified:

Documentation/devicetree/todo.txt
- Document node lifecycle for CONFIG_OF_DYNAMIC

Regards,
Brian
--
To unsubscribe from this list: send the line "unsubscribe devicetree" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Re: device_node lifetime (was: Re: [PATCH 1/7] phy: brcmstb-sata: add missing of_node_put)

[Julia Lawall]

On Wed, 18 Nov 2015, Brian Norris wrote:

> (changing subject, add devicetree-u79uwXL29TY76Z2rM5mHXA@public.gmane.org)
> 
> On Tue, Nov 17, 2015 at 11:33:25PM +0100, Julia Lawall wrote:
> > On Tue, 17 Nov 2015, Brian Norris wrote:
> > > On Tue, Nov 17, 2015 at 06:48:39PM +0100, Julia Lawall wrote:
> > > > Is this something that should be checked for elsewhere?
> > > 
> > > I expect the same sort of problem shows up plenty of other places. I
> > > don't think many people use CONFIG_OF_DYNAMIC, so the effects of these
> > > failures probably aren't felt by many.
> > 
> > I tried the following semantic patch:
> > 
> > @@
> > struct device_node *e;
> > expression e1;
> > identifier fld;
> > @@
> > 
> >  ... when != of_node_get(...)
> > *(<+...e1->fld...+>) = e
> >  ... when != of_node_get(...)
> >  return e1;
> > 
> > basically, this says that a structure field is initilized to a device node 
> > value, the structure is returned by the containing function, and the 
> > containing function contains no of_node_get at all.  Certainly this is 
> > quite constrained, but it does produce a number of examples.
> > 
> > I looked at a few of them:
> > 
> > drivers/clk/ingenic/cgu.c, ingenic_cgu_new
> > clk/pistachio/clk.c, pistachio_clk_alloc_provider
> 
> It looks like the clock core (drivers/clk/clk.c) initially grabs the clk
> provider node in of_clk_init(), then drops it after it's initialized,
> but most of these providers use of_clk_add_provider(), which seems to
> manage the device_node lifetime for the user. So I think these are OK.
> 
> > drivers/mfd/syscon.c, of_syscon_register
> 
> This one looks potentially suspect. Syscon nodes aren't usually directly
> managed by a single driver, and the device_node pointer is used for
> lookups later...so I think it should keep a kref, and it doesn't.
> 
> > drivers/of/pdt.c, function of_pdt_create_node
> 
> Not real sure about this one.
> 
> > Any idea whether these need of_node_get?  In all cases the device node 
> > value comes in as a parameter.
> 
> I'm really not an expert on this stuff. I just saw a potential problem
> that I happen to be looking at in other subsystems, and I wanted to know
> what others thought.

Thanks for the analysis.  I will look into them a bit more.  Hopefully at 
least the maintainer of each file will know what should be done.

julia

> I think this discussion should include the DT folks
> and the subsystems in question. For one, I'm as interested as anyone in
> getting this todo clarified:
> 
> Documentation/devicetree/todo.txt
> - Document node lifecycle for CONFIG_OF_DYNAMIC
> 
> Regards,
> Brian
> 
--
To unsubscribe from this list: send the line "unsubscribe devicetree" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Re: device_node lifetime (was: Re: [PATCH 1/7] phy: brcmstb-sata: add missing of_node_put)

[Rob Herring]

On Wed, Nov 18, 2015 at 1:05 PM, Brian Norris
<computersforpeace@gmail.com> wrote:
> (changing subject, add devicetree@vger.kernel.org)
>
> On Tue, Nov 17, 2015 at 11:33:25PM +0100, Julia Lawall wrote:
>> On Tue, 17 Nov 2015, Brian Norris wrote:
>> > On Tue, Nov 17, 2015 at 06:48:39PM +0100, Julia Lawall wrote:
>> > > Is this something that should be checked for elsewhere?
>> >
>> > I expect the same sort of problem shows up plenty of other places. I
>> > don't think many people use CONFIG_OF_DYNAMIC, so the effects of these
>> > failures probably aren't felt by many.

The "problem" is non-existent because either CONFIG_OF_DYNAMIC is off
or where it is used is limited (memory and cpus on PSeries) and now
overlays. Overlays have the potential to be problematic, but we should
manage ref counting for overlays in a completely different way. What
that looks like, I don't know. I'll leave that to the person that
cares about removing overlays.

>> basically, this says that a structure field is initilized to a device node
>> value, the structure is returned by the containing function, and the
>> containing function contains no of_node_get at all.  Certainly this is
>> quite constrained, but it does produce a number of examples.

I've got no idea if this is right or not.

>> drivers/of/pdt.c, function of_pdt_create_node
>
> Not real sure about this one.

SPARC. Stay away.

>
>> Any idea whether these need of_node_get?  In all cases the device node
>> value comes in as a parameter.
>
> I'm really not an expert on this stuff. I just saw a potential problem
> that I happen to be looking at in other subsystems, and I wanted to know
> what others thought. I think this discussion should include the DT folks
> and the subsystems in question. For one, I'm as interested as anyone in
> getting this todo clarified:
>
> Documentation/devicetree/todo.txt
> - Document node lifecycle for CONFIG_OF_DYNAMIC

Step 2 after figuring out it can't be documented is "define a new way
to handle dynamic DT refcounting aka how to get rid of
of_node_get/put."

Rob

Re: device_node lifetime (was: Re: [PATCH 1/7] phy: brcmstb-sata: add missing of_node_put)

[Russell King - ARM Linux]

On Thu, Nov 19, 2015 at 12:44:11PM -0600, Rob Herring wrote:
> On Wed, Nov 18, 2015 at 1:05 PM, Brian Norris
> <computersforpeace@gmail.com> wrote:
> > (changing subject, add devicetree@vger.kernel.org)
> >
> > On Tue, Nov 17, 2015 at 11:33:25PM +0100, Julia Lawall wrote:
> >> On Tue, 17 Nov 2015, Brian Norris wrote:
> >> > On Tue, Nov 17, 2015 at 06:48:39PM +0100, Julia Lawall wrote:
> >> > > Is this something that should be checked for elsewhere?
> >> >
> >> > I expect the same sort of problem shows up plenty of other places. I
> >> > don't think many people use CONFIG_OF_DYNAMIC, so the effects of these
> >> > failures probably aren't felt by many.
> 
> The "problem" is non-existent because either CONFIG_OF_DYNAMIC is off
> or where it is used is limited (memory and cpus on PSeries) and now
> overlays. Overlays have the potential to be problematic, but we should
> manage ref counting for overlays in a completely different way. What
> that looks like, I don't know. I'll leave that to the person that
> cares about removing overlays.

So are you saying we should just forget about of_node_put and delete all
of_node_put/of_node_get references in code outside drivers/of ?  That
seems pretty obtuse given that we do have the overlay code merged, and
sounds to me like a very bad idea.

Expecting those who want to use overlays to run around checking that
the refcounting is correct in drivers is a really silly idea IMHO -
the existing API is refcounted, so either people really ought to be
using it correctly as it's already been designed (in other words, with
correct refcounting, and we shouldn't be shovelling this problem onto
other people) or the refcounting should be completely killed.

The existing half-way house of "we have refcounting, but we don't care
about it" is really insane.

Either we have refcounting, and it's used properly, or we don't have
refcounting.  No middle ground IMHO.

-- 
FTTC broadband for 0.8mile line: currently at 9.6Mbps down 400kbps up
according to speedtest.net.

Re: device_node lifetime (was: Re: [PATCH 1/7] phy: brcmstb-sata: add missing of_node_put)

[Rob Herring]

On Thu, Nov 19, 2015 at 1:14 PM, Russell King - ARM Linux <linux@arm.linux.org.uk> wrote:
> On Thu, Nov 19, 2015 at 12:44:11PM -0600, Rob Herring wrote:
>> On Wed, Nov 18, 2015 at 1:05 PM, Brian Norris
>> <computersforpeace@gmail.com> wrote:
>> > (changing subject, add devicetree@vger.kernel.org)
>> >
>> > On Tue, Nov 17, 2015 at 11:33:25PM +0100, Julia Lawall wrote:
>> >> On Tue, 17 Nov 2015, Brian Norris wrote:
>> >> > On Tue, Nov 17, 2015 at 06:48:39PM +0100, Julia Lawall wrote:
>> >> > > Is this something that should be checked for elsewhere?
>> >> >
>> >> > I expect the same sort of problem shows up plenty of other places. I
>> >> > don't think many people use CONFIG_OF_DYNAMIC, so the effects of these
>> >> > failures probably aren't felt by many.
>>
>> The "problem" is non-existent because either CONFIG_OF_DYNAMIC is off
>> or where it is used is limited (memory and cpus on PSeries) and now
>> overlays. Overlays have the potential to be problematic, but we should
>> manage ref counting for overlays in a completely different way. What
>> that looks like, I don't know. I'll leave that to the person that
>> cares about removing overlays.
>
> So are you saying we should just forget about of_node_put and delete all
> of_node_put/of_node_get references in code outside drivers/of ?  That
> seems pretty obtuse given that we do have the overlay code merged, and
> sounds to me like a very bad idea.

I wasn't suggesting going as far as that, but I do think we can't rely on drivers to get ref counting right.

> Expecting those who want to use overlays to run around checking that
> the refcounting is correct in drivers is a really silly idea IMHO -
> the existing API is refcounted, so either people really ought to be
> using it correctly as it's already been designed (in other words, with
> correct refcounting, and we shouldn't be shovelling this problem onto
> other people) or the refcounting should be completely killed.

I'm only saying stop shoveling until ref

This is only the tip of the iceberg. What happens when you start hot plugging GPIO controllers for example.

> The existing half-way house of "we have refcounting, but we don't care
> about it" is really insane.
>
> Either we have refcounting, and it's used properly, or we don't have
> refcounting.  No middle ground IMHO.


>
> --
> FTTC broadband for 0.8mile line: currently at 9.6Mbps down 400kbps up
> according to speedtest.net.