put_device(&adap->dev) might free the memory pointed to by `adap`,
so we shouldn't read adap->owner after that.

Fix by saving module pointer before calling put_device().

Signed-off-by: Michał Mirosław <mirq-linux@rere.qmqm.pl>
---
 drivers/i2c/i2c-core-base.c | 5 ++++-
 1 file changed, 4 insertions(+), 1 deletion(-)

diff --git a/drivers/i2c/i2c-core-base.c b/drivers/i2c/i2c-core-base.c
index 2c59dd748a49..5d694f8ce9ef 100644
--- a/drivers/i2c/i2c-core-base.c
+++ b/drivers/i2c/i2c-core-base.c
@@ -2464,11 +2464,14 @@ EXPORT_SYMBOL(i2c_get_adapter);
 
 void i2c_put_adapter(struct i2c_adapter *adap)
 {
+	struct module *owner;
+
 	if (!adap)
 		return;
 
+	owner = adap->owner;
 	put_device(&adap->dev);
-	module_put(adap->owner);
+	module_put(owner);
 }
 EXPORT_SYMBOL(i2c_put_adapter);
 
-- 
2.30.2

Hi Michał,

I finally had a look at your patch...

> put_device(&adap->dev) might free the memory pointed to by `adap`,
> so we shouldn't read adap->owner after that.
> 
> Fix by saving module pointer before calling put_device().

... and found a different approach for this problem from 2019:

http://patchwork.ozlabs.org/project/linux-i2c/patch/1577439272-10362-1-git-send-email-vulab@iscas.ac.cn/

I think this is also proper. I found other subsystems in the kernel
first putting the module, then the device. Do you see problems with the
above patch?

Thanks for looking into the issue!

   Wolfram

> 
> Signed-off-by: Michał Mirosław <mirq-linux@rere.qmqm.pl>
> ---
>  drivers/i2c/i2c-core-base.c | 5 ++++-
>  1 file changed, 4 insertions(+), 1 deletion(-)
> 
> diff --git a/drivers/i2c/i2c-core-base.c b/drivers/i2c/i2c-core-base.c
> index 2c59dd748a49..5d694f8ce9ef 100644
> --- a/drivers/i2c/i2c-core-base.c
> +++ b/drivers/i2c/i2c-core-base.c
> @@ -2464,11 +2464,14 @@ EXPORT_SYMBOL(i2c_get_adapter);
>  
>  void i2c_put_adapter(struct i2c_adapter *adap)
>  {
> +	struct module *owner;
> +
>  	if (!adap)
>  		return;
>  
> +	owner = adap->owner;
>  	put_device(&adap->dev);
> -	module_put(adap->owner);
> +	module_put(owner);
>  }
>  EXPORT_SYMBOL(i2c_put_adapter);
>  
> -- 
> 2.30.2
> 

On Tue, Jun 14, 2022 at 09:45:42PM +0200, Wolfram Sang wrote:
> Hi Michał,
> 
> I finally had a look at your patch...
> 
> > put_device(&adap->dev) might free the memory pointed to by `adap`,
> > so we shouldn't read adap->owner after that.
> > 
> > Fix by saving module pointer before calling put_device().
> 
> ... and found a different approach for this problem from 2019:
> 
> http://patchwork.ozlabs.org/project/linux-i2c/patch/1577439272-10362-1-git-send-email-vulab@iscas.ac.cn/
> 
> I think this is also proper. I found other subsystems in the kernel
> first putting the module, then the device. Do you see problems with the
> above patch?
> 
> Thanks for looking into the issue!

Hi!

I looked briefly at the kobject machinery and it seems to ignore module
dependencies. So while both approaches might work, I'd usually reverse
the order the init code is using: in this case module_get+device_get,
so on release: device_put+module_put. I don't know what keeps the kernel
from unloading the module after module_put() and before the function
returns, but I assume that would blow up for both patches.

Best Regards
Michał Mirosław

> > 
> > Signed-off-by: Michał Mirosław <mirq-linux@rere.qmqm.pl>
> > ---
> >  drivers/i2c/i2c-core-base.c | 5 ++++-
> >  1 file changed, 4 insertions(+), 1 deletion(-)
> > 
> > diff --git a/drivers/i2c/i2c-core-base.c b/drivers/i2c/i2c-core-base.c
> > index 2c59dd748a49..5d694f8ce9ef 100644
> > --- a/drivers/i2c/i2c-core-base.c
> > +++ b/drivers/i2c/i2c-core-base.c
> > @@ -2464,11 +2464,14 @@ EXPORT_SYMBOL(i2c_get_adapter);
> >  
> >  void i2c_put_adapter(struct i2c_adapter *adap)
> >  {
> > +	struct module *owner;
> > +
> >  	if (!adap)
> >  		return;
> >  
> > +	owner = adap->owner;
> >  	put_device(&adap->dev);
> > -	module_put(adap->owner);
> > +	module_put(owner);
> >  }
> >  EXPORT_SYMBOL(i2c_put_adapter);
> >  
> > -- 
> > 2.30.2
> > 



-- 
Michał Mirosław

Hi Michał,

> I looked briefly at the kobject machinery and it seems to ignore module
> dependencies. So while both approaches might work, I'd usually reverse

Thanks for checking!

> the order the init code is using: in this case module_get+device_get,
> so on release: device_put+module_put. I don't know what keeps the kernel

I agree this is good style. I'll add a comment why we reverse the order.
This will be also good to avoid regressions.

> from unloading the module after module_put() and before the function
> returns, but I assume that would blow up for both patches.

Yes. There are other users in the kernel doing it like this (RTC and
regmap IIRC), so I think problems would have become visible by then.

Thank you for your help!

   Wolfram