Kconfig STOP_MACHINE goofiness, 2.6.39+

[wcg]

I am getting this warning message from "make menuconfig"
(or "make oldconfig") when configuring kernels from 2.6.39
to 3.1.6 (newest one that I have tried):

[Hu]

You can enable CPU hotplug in software even if it is not supported by your hardware. On systems that support suspend-to-disk, CPU hotplug is forcibly enabled, because the hibernation code uses it to disable non-boot CPUs during creation of the hibernation image.

[wcg]

Ok, thanks.

It is not so much that it is doing unreasonable things with
text_poke_smp() or stop_machine et al. They are used either
in early boot code or debugging code like kprobes to
patch jump targets, etc, in the running kernel.
(See arch/x86/kernel/alternative.c, arch/x86/include/asm/fixmap.h,
and kernel/stop_machine.c.) The #ifdef CONFIG_HOTPLUG_CPU
sections are just there for necessary locking if cpus can come
and go while these functions run.

It is just that the code seems to mash the _init and module
load/unload cases together so that it won't need separate
functions or a bunch of #ifdef/#endif blocks if loadable modules
are not enabled.

I was hoping it was only a Kconfig silliness, listing a
dependency that it was not really dependent on in the
actual code.

"Everyone using smp uses loadable modules."

Wrong assumption.
_________________
TIA

[Hu]

Indeed. I have systems that are SMP and yet have MODULES=n.

[wcg]

So, setting CONFIG_HOTPLUG_CPU=y in the kernel 3.1.6 .config
worked. I did not need to enable loadable modules as well. I got
no warnings from menuconfig, the kernel compiled ok, and
it is now running. (There is evidence of an ACPI bug in BIOS
related to plug-n-play, but having the kernel probe and then
disable that at boot is only a noisy log entry in dmesg.)

I still find the "loadable modules *or* cpu hotplug" dependency
of STOP_MACHINE in Kconfig suspicious. By what stretch of the
imagination are those two alternative implementations of the
same functionality? Likely they share a side-effect that the functions
enabled by STOP_MACHINE rely on that should be factored out
into a common implementation of something that any subsystem
can use (imho).
_________________
TIA

[wcg]

Kconfig could have an option called CONFIG_STRICT_TXT_LOCK
which is what STOP_MACHINE wants to keep threads running on other
cpus from stepping on its toes when the caller of its functions
needs to change something in the kernel code. SMP and loadable
modules would both automatically select it and depend on it.
Hotplug cpu would depend on it. STOP_MACHINE would depend on it.
Kprobes et al could select it and depend on it. HAVE_TXT_POKE_SMP
would depend on it. The jump labels stuff would depend on it.

Then there is only the one Kconfig dependency there for these
different kernel config options that need that same funtionality.
The only way to find it not selected in contexts where it is
needed is to alter the .config file with a text editor.

(More explicit and less "kernel gestalt voodoo" Kconfig depencies
makes it easier for users configuring kernels to understand the warnings,
Kconfig files, and so on.)
_________________
TIA

[libertytrek]

Hello,

Ok, I just noticed this warning when configuring a new kernel, and am unsure of what I should do - if anything...

My system is just like OPs - dual AMD64, motherboard doesn't have hotplug support, and I have loadable modules disable...

So... what should I do about this warning? Can it be safely ignored? Or should I enable CPU Hotplug?

Thanks