Message-ID: <20170418081001.GA2411@ozzy.nask.waw.pl>
Date:   Tue, 18 Apr 2017 10:10:01 +0200
From:   Michał Kępień <kernel@...pniu.pl>
To:     Darren Hart <dvhart@...radead.org>,
        Jonathan Woithe <jwoithe@...t42.net>
Cc:     Andy Shevchenko <andy@...radead.org>,
        platform-driver-x86@...r.kernel.org, linux-kernel@...r.kernel.org
Subject: Re: [PATCH 5/6] platform/x86: fujitsu-laptop: do not log LED
 registration failures

Jonathan, I hope this response to Darren's message also addresses your
concerns.  Feel free to let me know if it does not.

> On Fri, Apr 07, 2017 at 03:07:12PM +0200, Michał Kępień wrote:
> > If acpi_fujitsu_laptop_leds_register() returns an error, the latter will
> > become the return value of acpi_fujitsu_laptop_add(), which in turn will
> > be reported by driver core.  Simplify code by replacing pr_err() calls
> > with return statements.  Return 0 instead of result when no errors occur
> > in order to make the code easier to read.
> 
> Hi Michał,
> 
> Jonathan's comment regarding the information loss of removing the pr_err
> statements seems valid to me. Based on the outer if block, I take it each
> registration only fails in true error scenarios and not because some laptop
> might have one but not another LED in the list.

Correct.

> If so, then the pr_err messages
> would only appear when there was a legitimate problem. I think they're worth

I am not hell-bent on removing these pr_err() calls, but allow me to
briefly walk you through my thought process.

devm_led_classdev_register() is basically a managed wrapper for
led_classdev_register(), so let's see under what circumstances the
latter may fail.  While it does quite a bit, its return value can only
be different than zero for one of two reasons:

  - there is already a LED with the same name present in the system, so
    the kernel automatically renames the one we are registering and the
    length of the generated name exceeds LED_MAX_NAME_SIZE,

  - device_create_with_groups() fails, either because we are out of
    memory or the device hierarchy is screwed up.

The first case will never happen, given that the longest LED name that
fujitsu-laptop tries to register is 18 bytes long, the counter used for
auto-incrementation is an unsigned int and LED_MAX_NAME_SIZE is 64.

In the second case, we are likely to be notified by driver core about
the exact nature of the failure, but more importantly, logging which LED
"caused" the failure makes us none the wiser.  Actions taken by the
kernel in response to each of the devm_led_classdev_register() calls are
virtually identical and if any of these fails, we are more than likely
to have problems way more severe than non-functioning LEDs.

Have I missed anything or perhaps assumed something I should have not?

> This seems to introduce a behavior change as well. Previously only the last
> LED registered would determine the result - which is wrong of course and I
> believe you noted a related bug in an early patch. Previously, however, if
> LOGOLAMP_POWERON failed, for example, the KEYBOARD_LAMPS would still be attempted.
> 
> So the question really comes down to this: Is there a legitimate situation in
> which one LEDs registration fails and another succeeds? If so, then this would
> constitute a regression for such systems.

The behavior change you mentioned is intentional.  As pointed out above,
if any devm_led_classdev_register() call fails, it means we have reached
some inconsistent state which is really unlikely to be improved by
further attempts to register even more devices.

What do you guys think?

-- 
Best regards,
Michał Kępień

Powered by blists - more mailing lists