On Fri, 2020-04-17 at 12:24 +0200, Arnd Bergmann wrote: > On Fri, Apr 17, 2020 at 3:12 AM Saeed Mahameed <saeedm@xxxxxxxxxxxx> > wrote: > > Due to the changes to the semantics of imply keyword [1], which now > > doesn't force any config options to the implied configs any more. > > > > A module (FOO) that has a weak dependency on some other modules > > (BAR) > > is now broken if it was using imply to force dependency > > restrictions. > > e.g.: FOO needs BAR to be reachable, especially when FOO=y and > > BAR=m. > > Which might now introduce build/link errors. > > > > There are two options to solve this: > > 1. use IS_REACHABLE(BAR), everywhere BAR is referenced inside FOO. > > 2. in FOO's Kconfig add: depends on (BAR || !BAR) > > > > The first option is not desirable, and will leave the user confused > > when > > setting FOO=y and BAR=m, FOO will never reach BAR even though both > > are > > compiled. > > > > The 2nd one is the preferred approach, and will guarantee BAR is > > always > > reachable by FOO if both are compiled. But, (BAR || !BAR) is really > > confusing for those who don't really get how kconfig tristate > > arithmetics > > work. > > > > To solve this and hide this weird expression and to avoid > > repetition > > across the tree, we introduce new keyword "uses" to the Kconfig > > options > > family. > > > > uses BAR: > > Equivalent to: depends on symbol || !symbol > > Semantically it means, if FOO is enabled (y/m) and has the option: > > uses BAR, make sure it can reach/use BAR when possible. > > > > For example: if FOO=y and BAR=m, FOO will be forced to m. > > > > [1] > > https://lore.kernel.org/linux-doc/20200302062340.21453-1-masahiroy@xxxxxxxxxx/ > > Thanks a lot for getting this done. I've tried it out on my > randconfig > build tree > and can confirm that this works together with your second patch to > address the > specific MLX5 problem. > > I also tried out replacing all other instances of 'depends on FOO || > !FOO', using > this oneline script: > > git ls-files | grep Kconfig | xargs sed -i > 's:depends.on.\([A-Z0-9_a-z]\+\) || \(\1 \?= \?n\|!\1\):uses \1:' > > Unfortunately, this immediately crashes with: > > $ make -skj30 > how to free type 0? > double free or corruption (fasttop) > make[6]: *** [/git/arm-soc/scripts/kconfig/Makefile:71: olddefconfig] > Aborted (core dumped) > make[5]: *** [/git/arm-soc/Makefile:587: olddefconfig] Error 2 > make[4]: *** [/git/arm-soc/scripts/kconfig/Makefile:95: > allrandom.config] Error 2 > make[3]: *** [/git/arm-soc/Makefile:587: allrandom.config] Error 2 > make[2]: *** [Makefile:180: sub-make] Error 2 > make[2]: Target 'allrandom.config' not remade because of errors. > make[1]: *** [makefile:127: allrandom.config] Error 2 > > It's probably easy to fix, but I did not look any deeper into the > bug. > Ahh, I know what it is, i am allocating only one expression for the two symbols (FOO || !FOO) .. in the rule action in parser.y, i must allocate two individual instances per each of the FOO appearances .. something like: struct expr *symexpr1 = expr_alloc_symbol($2); struct expr *symexpr2 = expr_alloc_symbol($2); menu_add_dep(expr_alloc_two(E_OR, symexpr1, expr_alloc_one(E_NOT, symexpr2))); Thanks Arnd for testing this ! I will test this and send V2 later.
