Re: [PATCH RFC tip/core/rcu] Make SRCU be once again optional

[Arnd Bergmann]

On Wed, Jan 17, 2018 at 5:32 PM, Josh Triplett <josh@xxxxxxxxxxxxxxxx> wrote:
> On Wed, Jan 17, 2018 at 11:29:26AM +0100, Arnd Bergmann wrote:
>> On Wed, Jan 17, 2018 at 12:57 AM, Paul E. McKenney
>> <paulmck@xxxxxxxxxxxxxxxxxx> wrote:
>> > On Wed, Jan 17, 2018 at 12:03:18AM +0100, Arnd Bergmann wrote:
>> >> Evidently there is at least one driver that uses SRCU but doesn't 'select SRCU'
>> >> in Kconfig. There are probably others that just haven't been found.
>> >
>> > Does adding "select SRCU" on "config PM_SLEEP" in kernel/power/Kconfig
>> > fix this?
>>
>> I'm sure it does, but the point I was making is that we probably have a number
>> of those, and would never find the other ones through the current build test
>> setup.
>>
>> I've now tried disabling a ridiculous number of options to come up with a
>> setup that never enables SRCU. Interestingly, that also means we don't
>> get the drivers/base/power/wakeup.c problem in 'allmodconfig', though
>> I did get a link-time error:
> [...]
>> Turning off lockdep and kmemleak gives me a working allmodconfig build.
>> I'm doing some more testing on ARM, but it looks like this is a dark corner
>> of the randconfig state space that I'm not sure I want to explore more.
>>
>> Doing an hour of randconfig builds, I already found exactly two missing
>> 'select SRCU':
>>
>> ERROR: "__srcu_read_unlock" [drivers/infiniband/core/ib_uverbs.ko] undefined!
>> drivers/base/power/wakeup.c:68:1: error: type defaults to 'int' in
>> declaration of 'DEFINE_STATIC_SRCU' [-Werror=implicit-int]
>
> I've found that, when trying to make something optional, it doesn't
> suffice to do randconfigs.  You need a configuration with *everything*
> enabled, except for the option you want to test disabling, and anything
> that depends on or selects that option.  And, conversely, when testing
> if a specific option has all the dependencies it needs, you want a
> configuration with that option and its dependencies/selects enabled
> and everything else disabled.
>
> I wonder how easily we could make a kconfig option for an "all except"
> or "none except" config?  Such an option would read a minimal config
> snippet containing only specific options, and then act like
> allyesconfig/allmodconfig/allnoconfig as long as doing so doesn't
> change anything from the minimal config snippet.

Well, that's what I tried with the patch above that marks tons of
options as 'BROKEN', leaving all the options that don't select SRCU.

The funny thing is that the "everything but SRCU" config worked
just fine, but with the same patch, 'randconfig' failed one out of 10
times ;-)

      Arnd