On Sun, 20 Nov 2016 19:12:57 +0000 Russell King - ARM Linux <li...@armlinux.org.uk> wrote:
> On Sun, Nov 20, 2016 at 10:32:50AM -0800, Linus Torvalds wrote: > > On Sun, Nov 20, 2016 at 5:21 AM, Russell King - ARM Linux > > <li...@armlinux.org.uk> wrote: > > > On Tue, Oct 25, 2016 at 07:32:00PM +1100, Nicholas Piggin wrote: > > >> > > >> Michal, what's your thoughts? If you merge my patch 2/2 and skip 1/2, it > > >> should not give any new build warnings or errors, so then arch patches > > >> can > > >> go via arch trees. 1/2 could go in after everyone is up to date. > > > > > > So what's the conclusion on this? I've just had a failure due to > > > CONFIG_TRIM_UNUSED_KSYMS reported on ARM, and it looks like (at > > > least some of) patch 1 could resolve it. > > > > Hmm. I've got > > > > cc6acc11cad1 kbuild: be more careful about matching preprocessed asm > > ___EXPORT_SYMBOL > > 4efca4ed05cb kbuild: modversions for EXPORT_SYMBOL() for asm > > > > in my tree. Is that sufficient, or do we still have issues? > > Hmm, those seem to have gone in during the last week, so I haven't > tested it yet (build running, but it'll take a while). However, I > don't think they'll solve _this_ problem. > > Some of the issue here is that we use a mixture of assembly macros > and preprocessor for the ARM bitops - the ARM bitops are created > with an assembly macro which contains some pre-processor expanded > macros (eg, EXPORT_SYMBOL()). > > This means that the actual symbol being exported is not known to > the preprocessor, so doing the "__is_defined(__KSYM_##sym)" inside > "EXPORT_SYMBOL(\name)" becomes "__is_defined(__KSYM_\name)" to the > preprocessor. As "__KSYM_\name" is never defined, it always comes > out as zero, hence we always use __cond_export_sym_0, which omits > the symbol export from the assembly macro definition: > > .macro bitop, name, instr > .globl \name ; .align 0 ; \name: > > ... > > .type \name, %function; .size \name, .-\name > > .endm > > In other words, using preprocessor macros inside an assembly macro > may not work as expected, and now leads to config-specific failures. > Yes, that's a limitation. cpp expansion we can handle, but not gas macros. You will need Arnd's patches for ARM. http://marc.info/?l=linux-kbuild&m=147732160529499&w=2 If that doesn't fix it for you, send me your .config offline and I'll set up a cross compile to work on it. Again, any arch always has the option of going back to doing asm exports in the old style of putting them into a .c file, but hopefully you'll find Arnd's reworked patches to be something you're willing to merge. Thanks, Nick
