Re: [RFC/PATCH 1/3] pm: make PM macros more smart

[David Cohen <david.a.cohen@linux.intel.com>]

On Tue, Jan 14, 2014 at 02:42:11PM -0800, David Cohen wrote:
> On Fri, Dec 20, 2013 at 12:23:36PM -0800, David Cohen wrote:
> > On Fri, Dec 20, 2013 at 08:55:27PM +0100, Pavel Machek wrote:
> > > On Sun 2013-12-15 11:25:08, David Cohen wrote:
> > > > On Sun, Dec 15, 2013 at 06:51:12PM +0100, Pavel Machek wrote:
> > > > > On Thu 2013-12-12 21:18:23, David Cohen wrote:
> > > > > > This patch makes SET_SYSTEM_SLEEP_PM_OPS() and SET_RUNTIME_PM_OPS() more
> > > > > > smart.
> > > > > > 
> > > > > > Despite those macros check for '#ifdef CONFIG_PM_SLEEP/RUNTIME' to avoid
> > > > > > setting the callbacks when such #ifdef's aren't defined, they don't
> > > > > > handle compiler to avoid messages like that:
> > > > > > 
> > > > > > drivers/usb/host/xhci-plat.c:200:12: warning: ???xhci_plat_suspend??? defined but not used [-Wunused-function]
> > > > > > drivers/usb/host/xhci-plat.c:208:12: warning: ???xhci_plat_resume??? defined but not used [-Wunused-function]
> > > > > > 
> > > > > > As result, those macros get rid of #ifdef's when setting callbacks but
> > > > > > not when implementing them.
> > > > > > 
> > > > > > With this patch, drivers using SET_*_PM_OPS() macros don't need to #ifdef
> > > > > > the callbacks implementation as well.
> > > > > 
> > > > > Well... Interesting trickery, but it means that resulting kernel
> > > > > will be bigge due to the dead functions no?
> > > > 
> > > > Actually, it doesn't get bigger. Before sending the patch I did this
> > > > dummy test app:
> > > > 
> > > > --------------------------------------------------------------------------------
> > > > #include <stdio.h>
> > > > 
> > > > #define USE_IT_OR_LOOSE_IT(fn)  ((void *)((unsigned long)(fn) - (unsigned long)(fn)))
> > > > 
> > > > #ifdef MAKE_ME_NULL
> > > > static int func1(int a)
> > > > {
> > > >         printf("Hey!!\n");
> > > >         return 0;
> > > > }
> > > > #endif
> > > 
> > > I thought that point of this patch series was getting rid of the
> > > #ifdefs around the function...? Now I'm confused.
> > 
> > Maybe you're misinterpreting the test :)
> > 
> > This #ifdef is used to make this same test code to replicate both
> > scenarios according to -DMAKE_ME_NULL (just pay attention to actual
> > resulting code after #ifdef's are tested. the #ifdef here is nor related
> > to actual #ifdef on kernel). Here are both scenarios:
> > 
> > (1) Not using my trickery (which needs the function to not be present).
> > (2) Using my trickery (which needs to function to stay).
> > 
> > With -DMAKE_ME_NULL we replicate (2), then the function *is* there but
> > gcc gets rid of it on resulting binary without warnings if used with -O2.
> > 
> > Without -DMAKE_ME_NULL we replicate (1). The #ifdef will fail and then
> > remove the function which is an obvious scenario the function won't be
> > part of resulting binary.
> > 
> > If we use -S option to have human readable resulting assembly code
> > (which is kind of 1:1 for resulting binary), we can compare the result
> > of (1) and (2) and check they are pretty similar.
> > This proves gcc behaves as expected with my patch: do not need #ifdef
> > and do not generate dead codes to resulting binary.
> > 
> > > 
> > > > struct global_data {
> > > >         int (*func)(int);
> > > > };
> > > > 
> > > > static struct global_data gd = {
> > > > #ifdef MAKE_ME_NULL
> > > >         .func = USE_IT_OR_LOOSE_IT(func1),
> > > 
> > > If you have ifdef around the function, why do you need magic here? Why
> > > not
> > 
> > This #ifdef is necessary to prevent the function to be used when it
> > doesn't exist due to above #ifdef. But once again: don't misinterpret
> > the #ifdefs in this test app with the ones in kernel. They are not
> > related at all. If it's still confusing you just make 2 test apps
> > without #ifdeds out of this one where one keeps the code inside #ifdefs
> > and the other doesn't.
> > 
> > > 
> > > 	.func = func1
> > > 
> > > ?
> > > 
> > > Basically the warning tells you that you want the ifdef around the
> > > function, too... (Otherwise you waste space). That seems like good
> > > warning.
> > 
> > Just check my first explanation.
> 
> Ping :)
> 
> Comments here?

I found few problems to be fixed prior to be possible this optimization.

Many drivers are calling SET_*_PM_OPS() functions and passing
non-defined symbols as argument. It's not triggering compilation issue
currently because the drivers synchronize when the symbols are not
defined with SET_*_PM_OPS() not using them. But IMHO it's a violation of
scopes: all drivers calling SET_*_PM_OPS() should give valid symbols,
since it creates an unwanted cross-dependence between those PM functions
and their users (why SET_*_PM_OPS() can't use symbols given to them as
argument?).

I can work on fixing SET_*_PM_OPS() users beforehand in case my proposal
here is accepted.

Br, David Cohen

> 
> Br, David Cohen