Skip to contentSkip to conversation

Proper use for linking foo.o_shipped after 69ea912fda74 ("kbuild: remove unneeded link_multi_deps")?

DISCUSSION Florian Fainelli 2 participants 5 messages May 6, 2020 Kbuild

lore
Share & tools
lore
#0 lore raw

Hi Masahiro, Michal,

While updating our systems from 4.9 to 5.4, we noticed that one of the kernel modules that we build, which is done by linking an object that we pre-compile out of Kbuild stopped working.

I bisected it down to:

commit 69ea912fda74a673d330d23595385e5b73e3a2b9 (refs/bisect/bad)

Author: Masahiro Yamada <[email protected]>
Date:   Thu Oct 4 13:25:19 2018 +0900

kbuild: remove unneeded link_multi_deps

Since commit c8589d1e9e01 ("kbuild: handle multi-objs dependency appropriately"), $^ really represents all the prerequisite of the composite object being built.

Hence, $(filter %.o,$^) contains all the objects to link together, which is much simpler than link_multi_deps calculation.

Please note $(filter-out FORCE,$^) does not work here. When a single object module is turned into a multi object module, $^ will contain header files that were previously included for building the single object, and recorded in the .*.cmd file. To filter out such headers, $(filter %.o,$^) should be used here.

and the linker now fails with the following:

mkdir -p /home/florian/dev/lkm/.tmp_versions ; rm -f /home/florian/dev/lkm/.tmp_versions/*

WARNING: Symbol version dump ./Module.symvers
is missing; modules will have no dependencies and modversions.

make -f ./scripts/Makefile.build obj=/home/florian/dev/lkm
(cat /dev/null; echo kernel//home/florian/dev/lkm/hello.ko;) >
/home/florian/dev/lkm/modules.order
ld -m elf_x86_64 -z max-page-size=0x200000 -r -o
/home/florian/dev/lkm/hello.o
ld: no input files
make[1]: *** [scripts/Makefile.build:492: /home/florian/dev/lkm/hello.o] Error 1
make: *** [Makefile:1530: _module_/home/florian/dev/lkm] Error 2

and here are some steps to reproduce this:

Kbuild:
obj-m := hello.o
hello-y := test.o_shipped

test.c can be a simple hello world, and you can compile it using a standard Kbuild file first, and then move test.o as test.o_shipped.

I am afraid I do not speak Kbuild fluently enough to recommend a fix for that.

Thanks!

Signature · 1 line
Florian
#1 lore raw
Quoted 48 lines · Florian Fainelli

Hi Masahiro, Michal,

While updating our systems from 4.9 to 5.4, we noticed that one of the kernel modules that we build, which is done by linking an object that we pre-compile out of Kbuild stopped working.

I bisected it down to:

commit 69ea912fda74a673d330d23595385e5b73e3a2b9 (refs/bisect/bad)

Author: Masahiro Yamada <[email protected]>
Date:   Thu Oct 4 13:25:19 2018 +0900

kbuild: remove unneeded link_multi_deps

Since commit c8589d1e9e01 ("kbuild: handle multi-objs dependency appropriately"), $^ really represents all the prerequisite of the composite object being built.

Hence, $(filter %.o,$^) contains all the objects to link together, which is much simpler than link_multi_deps calculation.

Please note $(filter-out FORCE,$^) does not work here. When a single object module is turned into a multi object module, $^ will contain header files that were previously included for building the single object, and recorded in the .*.cmd file. To filter out such headers, $(filter %.o,$^) should be used here.

and the linker now fails with the following:

mkdir -p /home/florian/dev/lkm/.tmp_versions ; rm -f /home/florian/dev/lkm/.tmp_versions/*

WARNING: Symbol version dump ./Module.symvers
is missing; modules will have no dependencies and modversions.

make -f ./scripts/Makefile.build obj=/home/florian/dev/lkm
(cat /dev/null; echo kernel//home/florian/dev/lkm/hello.ko;) >
/home/florian/dev/lkm/modules.order
ld -m elf_x86_64 -z max-page-size=0x200000 -r -o
/home/florian/dev/lkm/hello.o
ld: no input files
make[1]: *** [scripts/Makefile.build:492: /home/florian/dev/lkm/hello.o] Error 1
make: *** [Makefile:1530: _module_/home/florian/dev/lkm] Error 2

Quoted 8 lines

and here are some steps to reproduce this:

Kbuild:
obj-m := hello.o
hello-y := test.o_shipped

test.c can be a simple hello world, and you can compile it using a standard Kbuild file first, and then move test.o as test.o_shipped.

Why don't you do like this?

obj-m := hello.o
hello-y := test.o

Quoted 6 lines

I am afraid I do not speak Kbuild fluently enough to recommend a fix for that.

Thanks!
--
Florian

Signature · 2 lines
Best Regards
Masahiro Yamada
#2 lore raw
Quoted 60 lines · Masahiro Yamada
Quoted 48 lines · Florian Fainelli

Hi Masahiro, Michal,

While updating our systems from 4.9 to 5.4, we noticed that one of the kernel modules that we build, which is done by linking an object that we pre-compile out of Kbuild stopped working.

I bisected it down to:

commit 69ea912fda74a673d330d23595385e5b73e3a2b9 (refs/bisect/bad)

Author: Masahiro Yamada <[email protected]>
Date:   Thu Oct 4 13:25:19 2018 +0900

kbuild: remove unneeded link_multi_deps

Since commit c8589d1e9e01 ("kbuild: handle multi-objs dependency appropriately"), $^ really represents all the prerequisite of the composite object being built.

Hence, $(filter %.o,$^) contains all the objects to link together, which is much simpler than link_multi_deps calculation.

Please note $(filter-out FORCE,$^) does not work here. When a single object module is turned into a multi object module, $^ will contain header files that were previously included for building the single object, and recorded in the .*.cmd file. To filter out such headers, $(filter %.o,$^) should be used here.

and the linker now fails with the following:

mkdir -p /home/florian/dev/lkm/.tmp_versions ; rm -f /home/florian/dev/lkm/.tmp_versions/*

WARNING: Symbol version dump ./Module.symvers
is missing; modules will have no dependencies and modversions.

make -f ./scripts/Makefile.build obj=/home/florian/dev/lkm
(cat /dev/null; echo kernel//home/florian/dev/lkm/hello.ko;) >
/home/florian/dev/lkm/modules.order
ld -m elf_x86_64 -z max-page-size=0x200000 -r -o
/home/florian/dev/lkm/hello.o
ld: no input files
make[1]: *** [scripts/Makefile.build:492: /home/florian/dev/lkm/hello.o] Error 1
make: *** [Makefile:1530: _module_/home/florian/dev/lkm] Error 2

Quoted 8 lines

and here are some steps to reproduce this:

Kbuild:
obj-m := hello.o
hello-y := test.o_shipped

test.c can be a simple hello world, and you can compile it using a standard Kbuild file first, and then move test.o as test.o_shipped.

Quoted 4 lines

Why don't you do like this?

obj-m := hello.o
hello-y := test.o

I tried it in the original environment where it failed, not my contrived test case, and this did not work, as we really need test.o and test.o_shipped to be separate objects, doing what you suggest results in a circular dependency.

To me this is a regression, as it used to work and now it does not, thus we should be fixing it, any idea about how we go about it without doing a plain revert?

Thank you!

Signature · 1 line
Florian
#3 lore raw
Quoted 66 lines · Florian Fainelli
Quoted 62 lines · Masahiro Yamada
Quoted 48 lines · Florian Fainelli

Hi Masahiro, Michal,

While updating our systems from 4.9 to 5.4, we noticed that one of the kernel modules that we build, which is done by linking an object that we pre-compile out of Kbuild stopped working.

I bisected it down to:

commit 69ea912fda74a673d330d23595385e5b73e3a2b9 (refs/bisect/bad)

Author: Masahiro Yamada <[email protected]>
Date:   Thu Oct 4 13:25:19 2018 +0900

kbuild: remove unneeded link_multi_deps

Since commit c8589d1e9e01 ("kbuild: handle multi-objs dependency appropriately"), $^ really represents all the prerequisite of the composite object being built.

Hence, $(filter %.o,$^) contains all the objects to link together, which is much simpler than link_multi_deps calculation.

Please note $(filter-out FORCE,$^) does not work here. When a single object module is turned into a multi object module, $^ will contain header files that were previously included for building the single object, and recorded in the .*.cmd file. To filter out such headers, $(filter %.o,$^) should be used here.

and the linker now fails with the following:

mkdir -p /home/florian/dev/lkm/.tmp_versions ; rm -f /home/florian/dev/lkm/.tmp_versions/*

WARNING: Symbol version dump ./Module.symvers
is missing; modules will have no dependencies and modversions.

make -f ./scripts/Makefile.build obj=/home/florian/dev/lkm
(cat /dev/null; echo kernel//home/florian/dev/lkm/hello.ko;) >
/home/florian/dev/lkm/modules.order
ld -m elf_x86_64 -z max-page-size=0x200000 -r -o
/home/florian/dev/lkm/hello.o
ld: no input files
make[1]: *** [scripts/Makefile.build:492: /home/florian/dev/lkm/hello.o] Error 1
make: *** [Makefile:1530: _module_/home/florian/dev/lkm] Error 2

Quoted 8 lines

and here are some steps to reproduce this:

Kbuild:
obj-m := hello.o
hello-y := test.o_shipped

test.c can be a simple hello world, and you can compile it using a standard Kbuild file first, and then move test.o as test.o_shipped.

Why don't you do like this?

Quoted 7 lines
Quoted 2 lines

obj-m := hello.o
hello-y := test.o

I tried it in the original environment where it failed, not my contrived test case, and this did not work, as we really need test.o and test.o_shipped to be separate objects, doing what you suggest results in a circular dependency.

I do not understand why you are trying to link *.o_shipped directly to the module.

Did you notice the following code in scripts/Makefile.lib ?

quiet_cmd_shipped = SHIPPED $@
cmd_shipped = cat $< > $@

$(obj)/%: $(src)/%_shipped

        $(call cmd,shipped)

*_shipped should be copied to *
before the use.

Quoted 3 lines

To me this is a regression, as it used to work and now it does not, thus we should be fixing it, any idea about how we go about it without doing a plain revert?

In fact, a patch exists.

https://patchwork.kernel.org/patch/11318691/

It allows you to link whatever file name to a module, but please read the comments from Christoph and Greg.

In this case, your Makefile is doing strange.

I'd recommend to fix your Makefile.

Quoted 4 lines

Thank you!
--
Florian

Signature · 2 lines
Best Regards
Masahiro Yamada
#4 lore raw
Quoted 9 lines · Masahiro Yamada
Quoted 3 lines

To me this is a regression, as it used to work and now it does not, thus we should be fixing it, any idea about how we go about it without doing a plain revert?

In fact, a patch exists.

https://patchwork.kernel.org/patch/11318691/

Quoted 7 lines

It allows you to link whatever file name to a module, but please read the comments from Christoph and Greg.

In this case, your Makefile is doing strange.

I'd recommend to fix your Makefile.

Yes, I ended up finding a way to fix it, thank you for your time.

Signature · 1 line
Florian