Skip to contentSkip to conversation

Makefile: can't build modules outside the kernel tree if it was built using O=

PATCH In review Caveh Jalali 4 participants 14 messages Sep 17, 2009 – Sep 27, 2009 Kbuild

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 Makefile: can't build modules outside the kernel tree if it was built using O=

in 2.6.31, it is no longer possible to build kernel modules for some platforms (the ones that need include2) outside the kernel src tree if the src tree was built using O= to specify an alternate directory for the build output.

i think this broke with the following commit:

commit b8b572e1015f81b4e748417be2629dfe51ab99f9
Author: Stephen Rothwell <[email protected]>
Date:   Fri Aug 1 15:20:30 2008 +1000

   powerpc: Move include files to arch/powerpc/include/asm

when the platform specific include directories were shuffled around without updating the
top level Makefile. however, i think this FAIL is generic, and not limited to powerpc.
here's a patch to the top level Makefile to teach it about the shuffled files:

--- /local/caveh/linux/linux-2.6.31/Makefile    2009-09-09 15:13:59.000000000 -0700
+++ Makefile    2009-09-16 12:23:05.000000000 -0700
@@ -958,9 +958,9 @@ ifneq ($(KBUILD_SRC),)
                echo "  in the '$(srctree)' directory.";\
                /bin/false; \
        fi;
-       $(Q)if [ ! -d include2 ]; then                                  \
-           mkdir -p include2;                                          \
-           ln -fsn $(srctree)/include/asm-$(SRCARCH) include2/asm;     \
+       $(Q)if [ ! -d include2 ]; then                                   \
+           mkdir -p include2;                                           \
+           ln -fsn $(srctree)/arch/$(SRCARCH)/include/asm include2/asm; \
        fi
 endif

Acked-by on v1 1/1
#1 lore raw

On Thu, Sep 17, 2009 at 10:30 AM, Caveh Jalali <[email protected]> wrote:> in 2.6.31, it is no longer possible to build kernel modules for some platforms> (the ones that need include2) outside the kernel src tree if the> src tree was built using O= to specify an alternate directory for the> build output.>> i think this broke with the following commit:>> commit b8b572e1015f81b4e748417be2629dfe51ab99f9> Author: Stephen Rothwell <[email protected]>> Date:   Fri Aug 1 15:20:30 2008 +1000

Only after 2.6.31? This commit was on 2008 Aug 1...

Quoted 1 line
Quoted 1 line

   powerpc: Move include files to arch/powerpc/include/asm>>> when the platform specific include directories were shuffled around> without updating the> top level Makefile.  however, i think this FAIL is generic, and not> limited to powerpc.> here's a patch to the top level Makefile to teach it about the shuffled files:>

You forgot your SOB line.
Anyway, it looks fine.

Quoted 1 line
Quoted 1 line

--- /local/caveh/linux/linux-2.6.31/Makefile    2009-09-09> 15:13:59.000000000 -0700> +++ Makefile    2009-09-16 12:23:05.000000000 -0700> @@ -958,9 +958,9 @@ ifneq ($(KBUILD_SRC),)>                 echo "  in the '$(srctree)' directory.";\>                 /bin/false; \>         fi;> -       $(Q)if [ ! -d include2 ]; then                                  \> -           mkdir -p include2;                                          \> -           ln -fsn $(srctree)/include/asm-$(SRCARCH) include2/asm;     \> +       $(Q)if [ ! -d include2 ]; then                                   \> +           mkdir -p include2;                                           \> +           ln -fsn $(srctree)/arch/$(SRCARCH)/include/asm include2/asm; \>         fi>  endif���{.n�+�������+%�����ݶ��w��{.n�+����{��G�����{ay�ʇڙ�,j��f���h��������z_�(�階�ݢj"���m������G������?����&���~��iO��z��v�^�m���� ������?�I�

#2 lore raw
In reply to Caveh Jalali

it's known good in 2.6.25, borked in 2.6.31. i didn't track down the kernel version number where the offending commit showed up.

anyway, here's my SOB:

Quoted 37 lines · Caveh Jalali

in 2.6.31, it is no longer possible to build kernel modules for some platforms (the ones that need include2) outside the kernel src tree if the src tree was built using O= to specify an alternate directory for the build output.

i think this broke with the following commit:

commit b8b572e1015f81b4e748417be2629dfe51ab99f9
Author: Stephen Rothwell <[email protected]>
Date:   Fri Aug 1 15:20:30 2008 +1000

   powerpc: Move include files to arch/powerpc/include/asm

when the platform specific include directories were shuffled around without updating the
top level Makefile.  however, i think this FAIL is generic, and not limited to powerpc.
here's a patch to the top level Makefile to teach it about the shuffled files:

--- /local/caveh/linux/linux-2.6.31/Makefile    2009-09-09 15:13:59.000000000 -0700
+++ Makefile    2009-09-16 12:23:05.000000000 -0700

(quoted diff) +3 -3
@@ -958,9 +958,9 @@ ifneq ($(KBUILD_SRC),)
                 echo "  in the '$(srctree)' directory.";\
                 /bin/false; \
         fi;
-       $(Q)if [ ! -d include2 ]; then                                  \
-           mkdir -p include2;                                          \
-           ln -fsn $(srctree)/include/asm-$(SRCARCH) include2/asm;     \
+       $(Q)if [ ! -d include2 ]; then                                   \
+           mkdir -p include2;                                           \
+           ln -fsn $(srctree)/arch/$(SRCARCH)/include/asm include2/asm; \

        fi
 endif

Signature · 2 lines
00c
http://www.boincstats.com/signature/user_870433_project5.gif
#3 lore raw
In reply to Caveh Jalali
Quoted 12 lines · Caveh Jalali

when the platform specific include directories were shuffled around without updating the
top level Makefile. however, i think this FAIL is generic, and not limited to powerpc.
here's a patch to the top level Makefile to teach it about the shuffled files:

--- /local/caveh/linux/linux-2.6.31/Makefile 2009-09-09
15:13:59.000000000 -0700
+++ Makefile 2009-09-16 12:23:05.000000000 -0700

(quoted diff) +0 -0
@@ -958,9 +958,9 @@ ifneq ($(KBUILD_SRC),)
                echo "  in the '$(srctree)' directory.";\

Review ifneq ($(KBUILD_SRC),) · line 965

Quoted 10 lines
(quoted diff) +3 -3
                /bin/false; \
        fi;
-       $(Q)if [ ! -d include2 ]; then                                  \
-           mkdir -p include2;                                          \
-           ln -fsn $(srctree)/include/asm-$(SRCARCH) include2/asm;     \
+       $(Q)if [ ! -d include2 ]; then                                   \
+           mkdir -p include2;                                           \
+           ln -fsn $(srctree)/arch/$(SRCARCH)/include/asm include2/asm; \
        fi
 endif

I think we should no longer have the include2 directory at all with new kernels.

LINUXINCLUDE already contains the right path in theory:

LINUXINCLUDE    := -Iinclude \
                   $(if $(KBUILD_SRC),-Iinclude2 -I$(srctree)/include) \
                   -I$(srctree)/arch/$(hdr-arch)/include               \
                   -include include/linux/autoconf.h

I'm not sure what goes wrong here. What is the gcc command line you see?

	Arnd <><
#4 lore raw

sadly, some 3rd party src packages don't quite do the right thing with their kernel include paths and still depend on the include2/asm symlink to find the platform header files.

i see your point about include2 being somewhat redundant/legacy, however, as long as we are creating it, we should have it point to the intended location instead of being a dead link!

Quoted 38 lines · Arnd Bergmann
Quoted 20 lines · Caveh Jalali

when the platform specific include directories were shuffled around without updating the
top level Makefile.  however, i think this FAIL is generic, and not limited to powerpc.
here's a patch to the top level Makefile to teach it about the shuffled files:

--- /local/caveh/linux/linux-2.6.31/Makefile    2009-09-09 15:13:59.000000000 -0700
+++ Makefile    2009-09-16 12:23:05.000000000 -0700

(quoted diff) +3 -3
@@ -958,9 +958,9 @@ ifneq ($(KBUILD_SRC),)
                 echo "  in the '$(srctree)' directory.";\
                 /bin/false; \
         fi;
-       $(Q)if [ ! -d include2 ]; then                                  \
-           mkdir -p include2;                                          \
-           ln -fsn $(srctree)/include/asm-$(SRCARCH) include2/asm;     \
+       $(Q)if [ ! -d include2 ]; then                                   \
+           mkdir -p include2;                                           \
+           ln -fsn $(srctree)/arch/$(SRCARCH)/include/asm include2/asm; \
Quoted 2 lines

        fi
 endif

I think we should no longer have the include2 directory at all with new kernels.

LINUXINCLUDE already contains the right path in theory:

LINUXINCLUDE    := -Iinclude \
                  $(if $(KBUILD_SRC),-Iinclude2 -I$(srctree)/include) \
                  -I$(srctree)/arch/$(hdr-arch)/include               \
                  -include include/linux/autoconf.h

I'm not sure what goes wrong here. What is the gcc command line you see?

       Arnd <><

Signature · 2 lines
00c
http://www.boincstats.com/signature/user_870433_project5.gif
#5 lore raw
Quoted 7 lines · Caveh Jalali

sadly, some 3rd party src packages don't quite do the right thing with their kernel include paths and still depend on the include2/asm symlink to find the platform header files.

i see your point about include2 being somewhat redundant/legacy, however, as long as we are creating it, we should have it point to the intended location instead of being a dead link!

I guess we could just as well put something in there to notify such programs of how to use the kernel Makefiles. External packages should not try to guess the layout of the kernel source tree, they can only get it wrong anyway.

There are apparently also user space packages trying to access /lib/modules/${krel}/build/arch/${arch}/include/, which is even more wrong.

I'm thinking about adding something in <linux/kernel.h> to warn about this when using the kernel headers without either '-D__KERNEL__' or 'make headers_install'. Is there any valid case that this would break? It should probably give a warning that includes a http URL holding a description of what to do.

	Arnd <><
#6 lore raw
In reply to Caveh Jalali
Quoted 7 lines · Caveh Jalali

sadly, some 3rd party src packages don't quite do the right thing with their kernel include paths and still depend on the include2/asm symlink to find the platform header files.

i see your point about include2 being somewhat redundant/legacy, however, as long as we are creating it, we should have it point to the intended location instead of being a dead link!

I guess we could just as well put something in there to notify such programs of how to use the kernel Makefiles. External packages should not try to guess the layout of the kernel source tree, they can only get it wrong anyway.

There are apparently also user space packages trying to access /lib/modules/${krel}/build/arch/${arch}/include/, which is even more wrong.

I'm thinking about adding something in <linux/kernel.h> to warn about this when using the kernel headers without either '-D__KERNEL__' or 'make headers_install'. Is there any valid case that this would break? It should probably give a warning that includes a http URL holding a description of what to do.

	Arnd <><
#7 lore raw
In reply to Arnd Bergmann
Quoted 27 lines · Arnd Bergmann
Quoted 20 lines · Caveh Jalali

when the platform specific include directories were shuffled around without updating the
top level Makefile.  however, i think this FAIL is generic, and not limited to powerpc.
here's a patch to the top level Makefile to teach it about the shuffled files:

--- /local/caveh/linux/linux-2.6.31/Makefile    2009-09-09 15:13:59.000000000 -0700
+++ Makefile    2009-09-16 12:23:05.000000000 -0700

(quoted diff) +3 -3
@@ -958,9 +958,9 @@ ifneq ($(KBUILD_SRC),)
                 echo "  in the '$(srctree)' directory.";\
                 /bin/false; \
         fi;
-       $(Q)if [ ! -d include2 ]; then                                  \
-           mkdir -p include2;                                          \
-           ln -fsn $(srctree)/include/asm-$(SRCARCH) include2/asm;     \
+       $(Q)if [ ! -d include2 ]; then                                   \
+           mkdir -p include2;                                           \
+           ln -fsn $(srctree)/arch/$(SRCARCH)/include/asm include2/asm; \
Quoted 2 lines

        fi
 endif

I think we should no longer have the include2 directory at all with new kernels.

Quoted 6 lines

LINUXINCLUDE already contains the right path in theory:

LINUXINCLUDE    := -Iinclude \
                  $(if $(KBUILD_SRC),-Iinclude2 -I$(srctree)/include) \
                  -I$(srctree)/arch/$(hdr-arch)/include               \
                  -include include/linux/autoconf.h

Hmm, at least in mmotm, we still have include2...

Maybe Sam is queueing the patches to remove include2? ���{.n�+�������+%�����ݶ��w��{.n�+����{��F�W��{ay�ʇڙ�,j��f���h��������z_�(�階�ݢj"���m������G������?����&���~��

#8 lore raw
Quoted 30 lines · Am?rico Wang
Quoted 29 lines · Arnd Bergmann
Quoted 20 lines · Caveh Jalali

when the platform specific include directories were shuffled around without updating the
top level Makefile.  however, i think this FAIL is generic, and not limited to powerpc.
here's a patch to the top level Makefile to teach it about the shuffled files:

--- /local/caveh/linux/linux-2.6.31/Makefile    2009-09-09 15:13:59.000000000 -0700
+++ Makefile    2009-09-16 12:23:05.000000000 -0700

(quoted diff) +3 -3
@@ -958,9 +958,9 @@ ifneq ($(KBUILD_SRC),)
                 echo "  in the '$(srctree)' directory.";\
                 /bin/false; \
         fi;
-       $(Q)if [ ! -d include2 ]; then                                  \
-           mkdir -p include2;                                          \
-           ln -fsn $(srctree)/include/asm-$(SRCARCH) include2/asm;     \
+       $(Q)if [ ! -d include2 ]; then                                   \
+           mkdir -p include2;                                           \
+           ln -fsn $(srctree)/arch/$(SRCARCH)/include/asm include2/asm; \
Quoted 2 lines

        fi
 endif

I think we should no longer have the include2 directory at all with new kernels.

LINUXINCLUDE already contains the right path in theory:

Quoted 8 lines
Quoted 4 lines

LINUXINCLUDE    := -Iinclude \
                  $(if $(KBUILD_SRC),-Iinclude2 -I$(srctree)/include) \
                  -I$(srctree)/arch/$(hdr-arch)/include               \
                  -include include/linux/autoconf.h

Hmm, at least in mmotm, we still have include2...

Maybe Sam is queueing the patches to remove include2?

I have them queued - yes.
But as they never hit -next they will wait another cycle.

Sam

#9 lore raw

PATCH 1/1 warn about use of uninstalled kernel headers

Quoted 9 lines · Sam Ravnborg
Quoted 9 lines
Quoted 8 lines

I think we should no longer have the include2 directory at all with new kernels.

LINUXINCLUDE already contains the right path in theory:

LINUXINCLUDE    := -Iinclude \
                  $(if $(KBUILD_SRC),-Iinclude2 -I$(srctree)/include) \
                  -I$(srctree)/arch/$(hdr-arch)/include               \
                  -include include/linux/autoconf.h
Quoted 5 lines
Quoted 3 lines

Hmm, at least in mmotm, we still have include2...

Maybe Sam is queueing the patches to remove include2?

I have them queued - yes.
But as they never hit -next they will wait another cycle.

Variations of this are starting to turn into a FAQ. How about printing a useful warning when someone tries to use the kernel headers without installing them first?

	Arnd <><

---
User applications frequently hit problems when they try to use the kernel headers directly, rather than the exported headers.

This adds an explicit warning for this case, and points to a URL holding an explanation of why this is wrong and what to do about it.

---
 include/linux/kernel.h     |    6 ++++++
 scripts/headers_install.pl |    2 +-
 2 files changed, 7 insertions(+), 1 deletion(-)
include/linux/kernel.h +6 -0
index d6320a3..f392d72 100644
@@ -656,6 +656,12 @@ extern int do_sysinfo(struct sysinfo *info);
 
 #endif /* __KERNEL__ */
 
+#ifndef __EXPORTED_HEADERS__
+#ifndef __KERNEL__
+#warning Attempt to use kernel headers from user space, see http://kernelnewbies.org/KernelHeaders
+#endif /* __KERNEL__ */
+#endif /* __EXPORTED_HEADERS__ */
+
 #define SI_LOAD_SHIFT	16
 struct sysinfo {
 	long uptime;			/* Seconds since boot */
scripts/headers_install.pl +1 -1
index c6ae405..b89ca2c 100644
@@ -20,7 +20,7 @@ use strict;
 
 my ($readdir, $installdir, $arch, @files) = @ARGV;
 
-my $unifdef = "scripts/unifdef -U__KERNEL__";
+my $unifdef = "scripts/unifdef -U__KERNEL__ -D__EXPORTED_HEADERS__";
 
 foreach my $file (@files) {
 	local *INFILE;
#10 lore raw

Re: [PATCH] warn about use of uninstalled kernel headers

Quoted 6 lines · Arnd Bergmann

User applications frequently hit problems when they try to use the kernel headers directly, rather than the exported headers.

This adds an explicit warning for this case, and points to a URL holding an explanation of why this is wrong and what to do about it.

Like the idea.

Quoted 12 lines
---
 include/linux/kernel.h     |    6 ++++++
 scripts/headers_install.pl |    2 +-
 2 files changed, 7 insertions(+), 1 deletion(-)
include/linux/kernel.h +0 -0
index d6320a3..f392d72 100644

include/linux/kernel.h extern int do_sysinfo(struct sysinfo *info); · line 659

Quoted 4 lines
include/linux/kernel.h +1 -0
@@ -656,6 +656,12 @@ extern int do_sysinfo(struct sysinfo *info);
 
 #endif /* __KERNEL__ */
 
+#ifndef __EXPORTED_HEADERS__

But why do we need this "__EXPORTED_HEADERS__" thing?

?

Sam

#11 lore raw

Re: [PATCH] warn about use of uninstalled kernel headers

Quoted 4 lines · Sam Ravnborg
Quoted 1 line

+#ifndef __EXPORTED_HEADERS__

But why do we need this "__EXPORTED_HEADERS__" thing?

?

The problem is that the warning should only be for headers that are not installed yet, but it needs to be stripped in the installed version. Since we're already postprocessing all files with unifdef, that seemed like the easiest way to strip out the #warning. Obviously, I couldn't use the #ifdef __KERNEL__ logic, because that would either warn give false positives when building kernel code or when building user code from installed headers.

	Arnd <><
#12 lore raw

Re: [PATCH] warn about use of uninstalled kernel headers

In reply to Arnd Bergmann
Quoted 12 lines · Arnd Bergmann
Quoted 11 lines · Sam Ravnborg
Quoted 11 lines
Quoted 8 lines

I think we should no longer have the include2 directory at all with new kernels.

LINUXINCLUDE already contains the right path in theory:

LINUXINCLUDE    := -Iinclude \
                  $(if $(KBUILD_SRC),-Iinclude2 -I$(srctree)/include) \
                  -I$(srctree)/arch/$(hdr-arch)/include               \
                  -include include/linux/autoconf.h

Hmm, at least in mmotm, we still have include2...

Quoted 7 lines
Quoted 3 lines
Quoted 1 line

Maybe Sam is queueing the patches to remove include2?

I have them queued - yes.
But as they never hit -next they will wait another cycle.

Variations of this are starting to turn into a FAQ. How about printing a useful warning when someone tries to use the kernel headers without installing them first?

OK - applied.

Sam

#13 lore raw
In reply to Caveh Jalali
Quoted 11 lines · Caveh Jalali

in 2.6.31, it is no longer possible to build kernel modules for some platforms (the ones that need include2) outside the kernel src tree if the src tree was built using O= to specify an alternate directory for the build output.

i think this broke with the following commit:

commit b8b572e1015f81b4e748417be2629dfe51ab99f9
Author: Stephen Rothwell <[email protected]>
Date:   Fri Aug 1 15:20:30 2008 +1000

Quoted 8 lines

   powerpc: Move include files to arch/powerpc/include/asm

when the platform specific include directories were shuffled around without updating the
top level Makefile. however, i think this FAIL is generic, and not limited to powerpc.
here's a patch to the top level Makefile to teach it about the shuffled files:

This is no good. Nothing outside the kernel should in any way rely on include2. And it will soon be gone anyway.

So we will not try to workaround something that is broken.

Sam