On Thu, Oct 29, 2020 at 12:09:23PM +0100, Hans de Goede wrote:
Hi,
On 10/29/20 11:59 AM, Coiby Xu wrote:
Hi Hans,
Thank you for reviewing this patch!
On Thu, Oct 29, 2020 at 11:04:36AM +0100, Hans de Goede wrote:
Hi,
On 10/29/20 8:41 AM, Coiby Xu wrote:
SIMPLE_DEV_PM_OPS has already took good care of CONFIG_PM_CONFIG.
No it does not, when CONFIG_PM_SLEEP is not set then the
SET_SYSTEM_SLEEP_PM_OPS macro which SIMPLE_DEV_PM_OPS uses
is a no-op, so nothing will reference xo15_sci_resume leading to
a compiler warning when CONFIG_PM_SLEEP is not set.
You could drop the ifdef and add __maybe_unused to the definition
of xo15_sci_resume, but that feels like needless churn, best to
just keep this as is IMHO.
Actually, this is a tree-wide change by some semi-automation scripts.
Thank you for pointing out the issue to prevent me from releasing
another ~150 emails to flood other mailing lists.
Currently there are 929 drivers has device PM callbacks,
$ grep -rI "\.pm = &" --include=*.c ./|wc -l
929
I put all files having device PM callbacks into four categories
based on weather a file has CONFIG_PM_SLEEP or PM macro like
SET_SYSTEM_SLEEP_PM_OPS, here are the statistics,
1. have both CONFIG_PM_SLEEP and PM_OPS macro: 213
2. have CONFIG_PM_SLEEP but no PM_OPS macro: 19
3. have PM macro but not CONFIG_PM_SLEEP: 347
4. no PM macro or CONFIG_PM_SLEEP: 302
Some drivers which have PM macro but not CONFIG_PM_SLEEP like
sound/x86/intel_hdmi_audio.c indeed use __maybe_unused to eliminate
the compiling warning. In 2011, there's a patch proposing to remove
ONFIG_PM altogether but an objection was turning CONFIG_PM on would
increase the kernel size [1]. So __maybe_unused also have this issue.
I would expect the compiler to remove the unused function, it knows
it is unused, that is why __maybe_unused is necessary to suppress
the warning and compilers are pretty smart and agressive wrt remove
unnecessary code these days.
Then __maybe_unused is a good solution and there's also convincing
reason to prefer __maybe_unused over CONFIG_PM_SLEEP according to
Arnd Bergmann [2],
> By and large, drivers handle this by using a CONFIG_PM_SLEEP ifdef.
>
> Unless you can make an extremely convincing argument why not to do
> so here, I'd like you to handle it that way instead.
[adding linux-pm to Cc]
The main reason is that everyone gets the #ifdef wrong, I run into
half a dozen new build regressions with linux-next every week on
average, the typical problems being:
- testing CONFIG_PM_SLEEP instead of CONFIG_PM, leading to an unused
function warning
- testing CONFIG_PM instead of CONFIG_PM_SLEEP, leading to a build
failure
- calling a function outside of the #ifdef only from inside an
otherwise correct #ifdef, again leading to an unused function
warning
- causing a warning inside of the #ifdef but only testing if that
is disabled, leading to a problem if the macro is set (this is
rare these days for CONFIG_PM as that is normally enabled)
Using __maybe_unused avoids all of the above.
Regards,
Hans
[2] https://lore.kernel.org/patchwork/comment/919944/
--
Best regards,
Coiby
