On Monday, July 18, 2016 12:13:38 PM CEST Thierry Reding wrote:

On Mon, Jul 18, 2016 at 12:09:39PM +0200, Arnd Bergmann wrote:

quoted

On Monday, July 18, 2016 11:33:28 AM CEST Thierry Reding wrote:

quoted

quoted

+
+     return 0;
+}
+
+static SIMPLE_DEV_PM_OPS(tango_thermal_pm, NULL, tango_thermal_resume);
+
+#define DEV_PM_OPS   &tango_thermal_pm
+#else
+#define DEV_PM_OPS   NULL
+#endif

In my experience it's often not useful to #ifdef the struct pm_ops.
These days you almost certainly want PM enabled, and the conditional
doesn't save you all that much in the first place, because it's not
unlikely for this to fit into some of the space that would be padded
out anyway.

This will also generate a warning when CONFIG_PM_SLEEP is not set.
Better write this as

#define DEV_PM_OPS (IS_ENABLED(CONFIG_PM_SLEEP) ? &tango_thermal_pm : NULL)

so the compiler can drop the variable definition when it's not
needed.

My suggestion was to define tango_thermal_pm unconditionally to avoid
any of these tricks. For any real use-case in which the 92 bytes for the
struct dev_pm_ops would matter you most likely want PM_SLEEP anyway, so
I don't really see why we would even want to make it optional.

Sure, leaving it unconditional works too. 

quoted

quoted

As a side-note, I've noticed that this driver has the following
dependencies:

        depends on ARCH_TANGO || COMPILE_TEST

which, last I checked, is probably going to fail on some architectures
because you need at least another one on HAS_IOMEM (for readl() and
writel()). That's a pre-existing problem, of course, so should be fixed
in a separate patch.

No need, we just merged a patch to no longer allow COMPILE_TEST on
arch/um/, so we can safely rely on MMIO to be available for COMPILE_TEST.

I thought at least S390 didn't have readl() and writel() either, at
least when PCI wasn't enabled, or some such.

Yes, but they've never complained about COMPILE_TEST breakage because
of that. Tile is in the same boat too in some configurations.

	Arnd