| Message ID | 20170318215823.GS4152@decadent.org.uk (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
+To Arnd +To Michal 2017-03-19 6:58 GMT+09:00 Ben Hutchings <ben@decadent.org.uk>: > systemtap currently fails to build modules when the kernel source and > object trees are separate. > > systemtap adds something like -I"/usr/share/systemtap/runtime" to > EXTRA_CFLAGS, and addtree should not adjust this as it's specifying an > absolute directory. But since make has no understanding of shell > quoting, it does anyway. > > For a long time this didn't matter, because addtree would still emit > the original -I option after the adjusted one. However, commit > db547ef19064 ("Kbuild: don't add obj tree in additional includes") > changed it to remove the original -I option. > > Remove quotes (both double and single) before matching against the > excluded patterns. > > References: https://bugs.debian.org/856474 > Reported-and-tested-by: Jack Henschel <jackdev@mailbox.org> > Reported-and-tested-by: Ritesh Raj Sarraf <rrs@debian.org> > Fixes: db547ef19064 ("Kbuild: don't add obj tree in additional includes") > Cc: stable@vger.kernel.org # 4.8+ > Signed-off-by: Ben Hutchings <ben@decadent.org.uk> > --- > --- a/scripts/Kbuild.include > +++ b/scripts/Kbuild.include > @@ -207,7 +207,7 @@ hdr-inst := -f $(srctree)/scripts/Makefi > # Prefix -I with $(srctree) if it is not an absolute path. > # skip if -I has no parameter > addtree = $(if $(patsubst -I%,%,$(1)), \ > -$(if $(filter-out -I/% -I./% -I../%,$(1)),$(patsubst -I%,-I$(srctree)/%,$(1)),$(1))) > +$(if $(filter-out -I/% -I./% -I../%,$(subst $(quote),,$(subst $(squote),,$(1)))),$(patsubst -I%,-I$(srctree)/%,$(1)),$(1))) > > # Find all -I options and call addtree > flags = $(foreach o,$($(1)),$(if $(filter -I%,$(o)),$(call addtree,$(o)),$(o))) I have been thinking about this patch. We touched addtree every time a new problem popped up. It would be easy to pick up this patch, but it looks like a crazy idea to play pattern-matching in the addtree even more. [1] -Ifoo/bar/baz [2] -I"foo/bar/baz" [3] -I foo/bar/baz (whitespace(s) after -I) All of these are valid notation, and should be handled in the same way. However, addtree expects only [1], (then this patch is going to add [2] in the soup). Each Makefile knows it wants to see additional headers in the source tree, or objtree. I am guessing the right approach in a long run is, we require -I to specify $(srctree) or $(objtree) explicitly. ccflags-y := -I$(srctree)/foo/bar/baz or ccflags-y := -I$(objtree)/foo/bar/baz (For the latter, we can omit $(objtree)/ as it is ./) Then, delete $(call flags,_c_flags) after the conversion. Any comments?
On 2017-04-03 09:42, Masahiro Yamada wrote: > Each Makefile knows it wants to see > additional headers in the source tree, or objtree. > > I am guessing the right approach in a long run is, > we require -I to specify $(srctree) or $(objtree) explicitly. > > ccflags-y := -I$(srctree)/foo/bar/baz > > or > > ccflags-y := -I$(objtree)/foo/bar/baz > > > (For the latter, we can omit $(objtree)/ as it is ./) > > > Then, delete $(call flags,_c_flags) after the conversion. Agreed. The addtree function is more of a hack to make things just work with O=, but AFAIK there is no clean way to implement VPATH for -I arguments. So it's sensible to get rid of the hack. It looks like it's going to be lot of work though: $ git grep -e '-I' -- '*Makefile*' | wc -l 732 $ git grep -e '-I *\$(\(src\|obj\)tree)' -- '*Makefile*' | wc -l 166 Michal -- To unsubscribe from this list: send the line "unsubscribe linux-kbuild" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On Mon, Apr 03, 2017 at 03:25:10PM +0200, Michal Marek wrote: > On 2017-04-03 09:42, Masahiro Yamada wrote: > > Each Makefile knows it wants to see > > additional headers in the source tree, or objtree. > > > > I am guessing the right approach in a long run is, > > we require -I to specify $(srctree) or $(objtree) explicitly. > > > > ccflags-y := -I$(srctree)/foo/bar/baz > > > > or > > > > ccflags-y := -I$(objtree)/foo/bar/baz > > > > > > (For the latter, we can omit $(objtree)/ as it is ./) > > > > > > Then, delete $(call flags,_c_flags) after the conversion. > > Agreed. The addtree function is more of a hack to make things just work > with O=, but AFAIK there is no clean way to implement VPATH for -I > arguments. So it's sensible to get rid of the hack. It looks like it's > going to be lot of work though: > > $ git grep -e '-I' -- '*Makefile*' | wc -l > 732 > $ git grep -e '-I *\$(\(src\|obj\)tree)' -- '*Makefile*' | wc -l > 166 There was a goal long time ago that moving the kernel source should not trigger a rebuild. Any hardcoded path would violate this (like $(srctree), $(objtree)) I dunno if this is really something to aim for today. I have personally from time to time renamed the directory where I have kernel soruce (which is seen like moving the kernel source), and would not be happy if this always triggered a full rebuild. But this is frankly a corner case. Sam -- To unsubscribe from this list: send the line "unsubscribe linux-kbuild" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
Hi Sam, 2017-04-04 5:20 GMT+09:00 Sam Ravnborg <sam@ravnborg.org>: > On Mon, Apr 03, 2017 at 03:25:10PM +0200, Michal Marek wrote: >> On 2017-04-03 09:42, Masahiro Yamada wrote: >> > Each Makefile knows it wants to see >> > additional headers in the source tree, or objtree. >> > >> > I am guessing the right approach in a long run is, >> > we require -I to specify $(srctree) or $(objtree) explicitly. >> > >> > ccflags-y := -I$(srctree)/foo/bar/baz >> > >> > or >> > >> > ccflags-y := -I$(objtree)/foo/bar/baz >> > >> > >> > (For the latter, we can omit $(objtree)/ as it is ./) >> > >> > >> > Then, delete $(call flags,_c_flags) after the conversion. >> >> Agreed. The addtree function is more of a hack to make things just work >> with O=, but AFAIK there is no clean way to implement VPATH for -I >> arguments. So it's sensible to get rid of the hack. It looks like it's >> going to be lot of work though: >> >> $ git grep -e '-I' -- '*Makefile*' | wc -l >> 732 >> $ git grep -e '-I *\$(\(src\|obj\)tree)' -- '*Makefile*' | wc -l >> 166 > > There was a goal long time ago that moving the kernel source should > not trigger a rebuild. > Any hardcoded path would violate this (like $(srctree), $(objtree)) Even if we avoid hard-coding $(srctree) in each Makefile, "addtree" will modify the -I paths and the modified paths will be recorded in .*.cmd files. If we really want to achieve the goal, my old patch seemed to have a point: https://patchwork.kernel.org/patch/4511991/ This was rejected due to side effects, anyway.
Dne 3.4.2017 v 22:20 Sam Ravnborg napsal(a): > On Mon, Apr 03, 2017 at 03:25:10PM +0200, Michal Marek wrote: >> On 2017-04-03 09:42, Masahiro Yamada wrote: >>> Each Makefile knows it wants to see >>> additional headers in the source tree, or objtree. >>> >>> I am guessing the right approach in a long run is, >>> we require -I to specify $(srctree) or $(objtree) explicitly. >>> >>> ccflags-y := -I$(srctree)/foo/bar/baz >>> >>> or >>> >>> ccflags-y := -I$(objtree)/foo/bar/baz >>> >>> >>> (For the latter, we can omit $(objtree)/ as it is ./) >>> >>> >>> Then, delete $(call flags,_c_flags) after the conversion. >> >> Agreed. The addtree function is more of a hack to make things just work >> with O=, but AFAIK there is no clean way to implement VPATH for -I >> arguments. So it's sensible to get rid of the hack. It looks like it's >> going to be lot of work though: >> >> $ git grep -e '-I' -- '*Makefile*' | wc -l >> 732 >> $ git grep -e '-I *\$(\(src\|obj\)tree)' -- '*Makefile*' | wc -l >> 166 > > There was a goal long time ago that moving the kernel source should > not trigger a rebuild. > Any hardcoded path would violate this (like $(srctree), $(objtree)) > > I dunno if this is really something to aim for today. > > I have personally from time to time renamed the directory where > I have kernel soruce (which is seen like moving the kernel source), > and would not be happy if this always triggered a full rebuild. > But this is frankly a corner case. This should be working nowadays if you build without O= or if the buildtree points to a subdirectory of the source tree. The srctree variable is set to . or .. in such case. Although when I tried it now, it still does not look perfect: $ cd /dev/shm/mmarek/linux-2.6 $ make O=build -s defconfig $ make O=build -s -j64 Setup is 15500 bytes (padded to 15872 bytes). System is 6530 kB CRC 4bc8f6e2 Kernel: arch/x86/boot/bzImage is ready (#1) $ grep -r $PWD build/ Binary file build/arch/x86/boot/setup.elf matches Binary file build/arch/x86/boot/header.o matches Binary file build/arch/x86/boot/video.o matches Binary file build/arch/x86/boot/cpu.o matches Binary file build/arch/x86/boot/printf.o matches Binary file build/arch/x86/boot/string.o matches Binary file build/arch/x86/boot/video-mode.o matches Binary file build/arch/x86/boot/video-vga.o matches Binary file build/arch/x86/boot/early_serial_console.o matches Binary file build/arch/x86/boot/video-vesa.o matches Binary file build/arch/x86/boot/cpucheck.o matches Binary file build/arch/x86/boot/video-bios.o matches Binary file build/arch/x86/boot/tty.o matches Binary file build/arch/x86/boot/memory.o matches Binary file build/arch/x86/boot/pm.o matches Binary file build/arch/x86/boot/cmdline.o matches Binary file build/arch/x86/boot/main.o matches Binary file build/arch/x86/boot/a20.o matches Binary file build/arch/x86/boot/regs.o matches Binary file build/arch/x86/boot/version.o matches Binary file build/arch/x86/boot/edd.o matches Binary file build/arch/x86/boot/cpuflags.o matches Binary file build/arch/x86/boot/pmjump.o matches Binary file build/arch/x86/boot/copy.o matches Binary file build/arch/x86/boot/bioscall.o matches Binary file build/arch/x86/realmode/rm/video-bios.o matches Binary file build/arch/x86/realmode/rm/video-mode.o matches Binary file build/arch/x86/realmode/rm/video-vga.o matches Binary file build/arch/x86/realmode/rm/video-vesa.o matches Binary file build/arch/x86/realmode/rm/wakemain.o matches Binary file build/arch/x86/realmode/rm/regs.o matches Binary file build/arch/x86/realmode/rm/trampoline_64.o matches Binary file build/arch/x86/realmode/rm/wakeup_asm.o matches Binary file build/arch/x86/realmode/rm/copy.o matches Binary file build/arch/x86/realmode/rm/bioscall.o matches Binary file build/arch/x86/realmode/rm/reboot.o matches All the 32bit object files contain the full path, but none of the .cmd files do?! And we do trigger rebuild of a few files after moving the tree, but we do so even in the original location, so it looks like somebody got a custom rule wrong again. Michal -- To unsubscribe from this list: send the line "unsubscribe linux-kbuild" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
--- a/scripts/Kbuild.include +++ b/scripts/Kbuild.include @@ -207,7 +207,7 @@ hdr-inst := -f $(srctree)/scripts/Makefi # Prefix -I with $(srctree) if it is not an absolute path. # skip if -I has no parameter addtree = $(if $(patsubst -I%,%,$(1)), \ -$(if $(filter-out -I/% -I./% -I../%,$(1)),$(patsubst -I%,-I$(srctree)/%,$(1)),$(1))) +$(if $(filter-out -I/% -I./% -I../%,$(subst $(quote),,$(subst $(squote),,$(1)))),$(patsubst -I%,-I$(srctree)/%,$(1)),$(1))) # Find all -I options and call addtree flags = $(foreach o,$($(1)),$(if $(filter -I%,$(o)),$(call addtree,$(o)),$(o)))
systemtap currently fails to build modules when the kernel source and object trees are separate. systemtap adds something like -I"/usr/share/systemtap/runtime" to EXTRA_CFLAGS, and addtree should not adjust this as it's specifying an absolute directory. But since make has no understanding of shell quoting, it does anyway. For a long time this didn't matter, because addtree would still emit the original -I option after the adjusted one. However, commit db547ef19064 ("Kbuild: don't add obj tree in additional includes") changed it to remove the original -I option. Remove quotes (both double and single) before matching against the excluded patterns. References: https://bugs.debian.org/856474 Reported-and-tested-by: Jack Henschel <jackdev@mailbox.org> Reported-and-tested-by: Ritesh Raj Sarraf <rrs@debian.org> Fixes: db547ef19064 ("Kbuild: don't add obj tree in additional includes") Cc: stable@vger.kernel.org # 4.8+ Signed-off-by: Ben Hutchings <ben@decadent.org.uk> ---