Masahiro Yamada masahiroy@kernel.org writes:
On Thu, Aug 26, 2021 at 12:42 AM Boris Kolpackov boris@codesynthesis.com wrote:
Masahiro Yamada masahiroy@kernel.org writes:
Kconfig (syncconfig) generates include/generated/autoconf.h to make CONFIG options available to the pre-processor.
The macros are suffixed with '_MODULE' for symbols with the value 'm'.
Here is a conflict; CONFIG_FOO=m results in '#define CONFIG_FOO_MODULE 1', but CONFIG_FOO_MODULE=y also results in the same define.
fixdep always assumes CONFIG_FOO_MODULE comes from CONFIG_FOO=m, so the dependency is not properly tracked for symbols that end with '_MODULE'.
It seem to me the problem is in autoconf.h/fixdep, not in the Kconfig language.
So, what is your suggestion for doing this correctly? (of course without breaking the compatibility because this is how the kernel is configured/built for more than 20 years)
Yes, I appreciate that fixing this properly may not be an option due to backwards-compatibility. How about then moving the check from the language closer to the place where it will actually be an issue. Specifically, can the error be triggered when we are about to write #define to autoconf.h and see that the name ends with _MODULE?
I know you don't care, but I will voice my objection, for the record: Kconfig is used by projects other than the Linux kernel and some of them do not use the autoconf.h functionality. For such projects this restriction seems arbitrary and potentially backwards-incompatible.
I am not sure what your worry is, but this check resides in "if (modules_sym)" conditional, so projects using Kconfig but not module functionality (e.g. buildroot) will not be affected.
The Kconfig module semantics is actually general enough that a project other than the Linux kernel could reuse it. (I've written more on this possibility here[1]).
[1] https://build2.org/libbuild2-kconfig/doc/build2-kconfig-manual.xhtml#lang-mo...