Re: [PATCH RFC v2 2/2] driver core: Don't allow passing a -ENOMEM to dev_err_probe()

[Uwe Kleine-König]

Hello Geert,

On Fri, Jun 14, 2024 at 09:26:52AM +0200, Geert Uytterhoeven wrote:
> On Thu, Jun 13, 2024 at 11:24 PM Uwe Kleine-König <ukleinek@xxxxxxxxxx> wrote:
> > From: Uwe Kleine-König <u.kleine-koenig@xxxxxxxxxxxx>
> >
> > If a function returns the error code -ENOMEM, there should be no error
> > output, because a failing allocation is already quite talkative and
> > adding another indication only makes it harder to determine the actual
> > problem.
> >
> > So the construct:
> >
> >         ret = some_function(...);
> >         if (ret)
> >                 return dev_err_probe(dev, ret, ...);
> >
> > is questionable if some_function() can only succeed or return -ENODEV.
> >
> > Catch some of these failures during compile time.
> >
> > Suggested-by: Andy Shevchenko <andriy.shevchenko@xxxxxxxxxxxxxxx>
> > Signed-off-by: Uwe Kleine-König <u.kleine-koenig@xxxxxxxxxxxx>
> 
> Thanks for your patch!
> 
> > I have some concerns about this patch, I only implemented it because in
> > reply to the first submission of patch #1 Andy wrote that he thinks this
> > should be done, too. So the idea of this patch is only to keep the
> > discussion about handling a constant -ENOMEM to dev_err_probe() away
> > from patch 1, in the hope to make application of patch 1 more likely :-)
> >
> > So, I think this patch 2/2 is a bad idea, because:
> >
> >  - Let's assume there are functions, that return either success or
> >    -ENOMEM. (I'm not aware of such a function, but I didn't search for
> >    one and probably something like that exists.) Probably the compiler
> >    won't be able to know that, and so doesn't catch that "problem".
> 
> You can find several in public header files:
> 
>     git grep -W "return\s*-ENOMEM\>" -- include/
> 
> I expect there are more in static code all over the place.
> 
> >  - Using dev_err_probe() to handle the return code of some_function() is
> >    convenient. First to make error handling in the calling function
> >    uniform, and second, to not create a patch opportunity for all
> >    callers when some_function() might return another error code in the
> >    future. So dev_err_probe() can just be used without caring for the
> >    details of the handled error.
> 
> IMHO this is the only drawback.
> And things may change: a static (inline) function that can only return
> zero or -ENOMEM now, can return other error codes tomorrow.
> Also, some dummies (e.g. dma_mapping_error()) return -ENOMEM, so it
> depends on kernel configuration too.

Huh, I didn't spot the dependency on kernel configuration. That makes it
quite bad.

> >  - In the presence of patch #1, there is no real problem with calling
> >    dev_err_probe(dev, -ENOMEM, ...), because this is an error path and
> >    so not performance critical, and no error message is emitted.
> 
> There's still the issue of increased kernel size, mainly due to the
> presence of the error message string.
> 
> > Given these, the more complicated implementation for dev_err_probe()
> > isn't really justified IMHO.
> 
> My initial reaction was quite positive, until I discovered the dummies...
> 
> > --- a/include/linux/dev_printk.h
> > +++ b/include/linux/dev_printk.h
> > @@ -275,6 +275,12 @@ do {                                                                       \
> >         WARN_ONCE(condition, "%s %s: " format, \
> >                         dev_driver_string(dev), dev_name(dev), ## arg)
> >
> > -__printf(3, 4) int dev_err_probe(const struct device *dev, int err, const char *fmt, ...);
> > +__printf(3, 4) int __dev_err_probe(const struct device *dev, int err, const char *fmt, ...);
> > +#define dev_err_probe(dev, err, ...)                                           \
> > +       ({                                                                      \
> > +               int __err = (err);                                              \
> > +               BUILD_BUG_ON(__builtin_constant_p(__err) && __err == -ENOMEM);  \
> > +               __dev_err_probe((dev), __err, __VA_ARGS__);                     \
> > +        })
> >
> >  #endif /* _DEVICE_PRINTK_H_ */
> 
> Looks like dev_err_probe() does not have a dummy for the !CONFIG_PRINTK
> case yet, while it could definitely use one.

Would you want to drop

	device_set_deferred_probe_reason(dev, &vaf);

from dev_err_probe() for !CONFIG_PRINTK, too? If not, you can throw away
the string only if __builtin_constant_p(__err != -EPROBE_DEFER) && __err
!= -EPROBE_DEFER. I agree such an improvement would be nice, but that's
orthogonal to this series.

Best regards
Uwe

Attachment: signature.asc
Description: PGP signature