Re: [PATCH v3 2/5] thermal/drivers/armada: convert to use devm_request_threaded_irq_emsg()

[Miquel Raynal]

Hi Uwe,

u.kleine-koenig@xxxxxxxxxxxxxx wrote on Tue, 4 Jul 2023 16:22:27 +0200:

> Hello Miquel,
> 
> On Tue, Jul 04, 2023 at 10:46:08AM +0200, Miquel Raynal wrote:
> > Hi Yangtao,
> > 
> > frank.li@xxxxxxxx wrote on Mon,  3 Jul 2023 17:04:51 +0800:
> >   
> > > There are more than 700 calls to the devm_request_threaded_irq method.
> > > Most drivers only request one interrupt resource, and these error
> > > messages are basically the same. If error messages are printed
> > > everywhere, more than 1000 lines of code can be saved by removing the
> > > msg in the driver.
> > > 
> > > And tglx point out that:
> > > 
> > >   If we actually look at the call sites of
> > >   devm_request_threaded_irq() then the vast majority of them print more or
> > >   less lousy error messages. A quick grep/sed/awk/sort/uniq revealed
> > > 
> > >      519 messages total (there are probably more)
> > > 
> > >      352 unique messages
> > > 
> > >      323 unique messages after lower casing
> > > 
> > >          Those 323 are mostly just variants of the same patterns with
> > >          slight modifications in formatting and information provided.
> > > 
> > >      186 of these messages do not deliver any useful information,
> > >          e.g. "no irq", "
> > > 
> > >      The most useful one of all is: "could request wakeup irq: %d"
> > > 
> > >   So there is certainly an argument to be made that this particular
> > >   function should print a well formatted and informative error message.
> > > 
> > >   It's not a general allocator like kmalloc(). It's specialized and in the
> > >   vast majority of cases failing to request the interrupt causes the
> > >   device probe to fail. So having proper and consistent information why
> > >   the device cannot be used _is_ useful.
> > > 
> > > Let's use devm_request_threaded_irq_emsg(), which ensure that all error
> > > handling branches print error information. In this way, when this function
> > > fails, the upper-layer functions can directly return an error code without
> > > missing debugging information. Otherwise, the error message will be
> > > printed redundantly or missing.
> > > 
> > > Cc: Thomas Gleixner <tglx@xxxxxxxxxxxxx>
> > > Signed-off-by: Yangtao Li <frank.li@xxxxxxxx>
> > > ---
> > >  drivers/thermal/armada_thermal.c | 13 +++++--------
> > >  1 file changed, 5 insertions(+), 8 deletions(-)
> > > 
> > > diff --git a/drivers/thermal/armada_thermal.c b/drivers/thermal/armada_thermal.c
> > > index 9f6dc4fc9112..a5e140643f00 100644
> > > --- a/drivers/thermal/armada_thermal.c
> > > +++ b/drivers/thermal/armada_thermal.c
> > > @@ -913,15 +913,12 @@ static int armada_thermal_probe(struct platform_device *pdev)
> > >  
> > >  	/* The overheat interrupt feature is not mandatory */
> > >  	if (irq > 0) {
> > > -		ret = devm_request_threaded_irq(&pdev->dev, irq,
> > > -						armada_overheat_isr,
> > > -						armada_overheat_isr_thread,
> > > -						0, NULL, priv);
> > > -		if (ret) {
> > > -			dev_err(&pdev->dev, "Cannot request threaded IRQ %d\n",
> > > -				irq);
> > > +		ret = devm_request_threaded_irq_emsg(&pdev->dev, irq,
> > > +						     armada_overheat_isr,
> > > +						     armada_overheat_isr_thread,
> > > +						     0, NULL, priv, NULL);
> > > +		if (ret)  
> > 
> > I don't see a patch renaming this helper with s/emsg//, do you plan to
> > keep it like that? I bet nobody outside of this series will notice the
> > new helper and will continue to add error messages because it kind
> > of feels "right" to do so.
> > 
> > I would rather prefer returning to the original function name which
> > everybody knows/uses.  
> 
> See https://lore.kernel.org/lkml/87h6qpyzkd.ffs@tglx for why there is a
> new function name.

Yes of course, I fully understand Thomas' concerns, but I am
questioning the usefulness of creating such helper if it's not the
default. People will continue to use [devm_]request_threaded_irq()
without noticing the new helper. If we want to make this change useful,
I believe we should:
- target all users
- at the end of the series: atomically include the error message in
  request_threaded_irq() and rename all callers of the _verbose variant
  which will eventually vanish.

Otherwise this is a lot of noise for very little progress, IMHO :)

Thanks,
Miquèl