kbuild: pass objects instead of archives to linker
Subsystems (2)
RFC PATCH 1/1 kbuild: pass objects instead of archives to linker
This is an experimental patch, driven by the feedback from Jiri Slaby and Michael Matz. [1]
Michael Matz says:
"I know of no linker (outside LTO-like modes) that processes
archives in a different order than first-to-last-member (under whole-archive), but that's not guaranteed anywhere. So relying on member-order within archives is always brittle."
It is pretty easy to pass the list of objects instead of a thin archive because the linker supports the '@file' syntax, where command line arguments are read from 'file'.
Without this patch, the linker receives
--whole-archive vmlinux.a --no-whole-archive
With this patch, the linker will receive
@vmlinux.order
Here, vmlinux.order is a text file that lists built-in objects in the correct link order.
I am not a toolchain expert. I just want to know if this makes any difference from the linker perspective and from (non-upstreamed) GCC-LTO perspective.
(I know this patch does not work for Clang LTO because I did not touch scripts/generate_initcall_order.pl)
This patch may be unneeded because more correct patches were submitted [2] but I am still curious about "thin archive vs direct object list".
[1]: https://lore.kernel.org/linux-kbuild/[email protected]/
[2]: https://lore.kernel.org/all/[email protected]/
---
Makefile | 21 ++++++++++----------- scripts/Makefile.modpost | 5 +++-- scripts/Makefile.vmlinux_o | 6 +++--- scripts/clang-tools/gen_compile_commands.py | 21 ++++++++++++++++++++- scripts/link-vmlinux.sh | 8 ++++---- 5 files changed, 40 insertions(+), 21 deletions(-)
Makefile +10 -11
@@ -1213,19 +1213,18 @@ quiet_cmd_autoksyms_h = GEN $@ $(autoksyms_h): $(call cmd,autoksyms_h) -# '$(AR) mPi' needs 'T' to workaround the bug of llvm-ar <= 14 -quiet_cmd_ar_vmlinux.a = AR $@ - cmd_ar_vmlinux.a = \ - rm -f $@; \ - $(AR) cDPrST $@ $(KBUILD_VMLINUX_OBJS); \ - $(AR) mPiT $$($(AR) t $@ | sed -n 1p) $@ $$($(AR) t $@ | grep -F -f $(srctree)/scripts/head-object-list.txt) +quiet_cmd_vmlinux_order = GEN $@ + cmd_vmlinux_order = \ + { $(foreach m, $(KBUILD_VMLINUX_OBJS), $(AR) t $m;) :; } > $(tmp-target) ; \ + grep -F -f $(srctree)/scripts/head-object-list.txt $(tmp-target) > $@; \ + grep -F -f $(srctree)/scripts/head-object-list.txt $(tmp-target) -v >> $@ -targets += vmlinux.a -vmlinux.a: $(KBUILD_VMLINUX_OBJS) scripts/head-object-list.txt autoksyms_recursive FORCE - $(call if_changed,ar_vmlinux.a) +targets += vmlinux.order +vmlinux.order: $(KBUILD_VMLINUX_OBJS) scripts/head-object-list.txt autoksyms_recursive FORCE + $(call if_changed,vmlinux_order) PHONY += vmlinux_o -vmlinux_o: vmlinux.a $(KBUILD_VMLINUX_LIBS) +vmlinux_o: vmlinux.order $(KBUILD_VMLINUX_LIBS) $(Q)$(MAKE) -f $(srctree)/scripts/Makefile.vmlinux_o vmlinux.o modules.builtin.modinfo modules.builtin: vmlinux_o
@@ -2037,7 +2036,7 @@ quiet_cmd_gen_compile_commands = GEN $@ cmd_gen_compile_commands = $(PYTHON3) $< -a $(AR) -o $@ $(filter-out $<, $(real-prereqs)) $(extmod_prefix)compile_commands.json: scripts/clang-tools/gen_compile_commands.py \ - $(if $(KBUILD_EXTMOD),, vmlinux.a $(KBUILD_VMLINUX_LIBS)) \ + $(if $(KBUILD_EXTMOD),, vmlinux.order $(KBUILD_VMLINUX_LIBS)) \ $(if $(CONFIG_MODULES), $(MODORDER)) FORCE $(call if_changed,gen_compile_commands)
scripts/Makefile.modpost +3 -2
@@ -70,12 +70,13 @@ quiet_cmd_vmlinux_objs = GEN $@ for f in $(real-prereqs); do \ case $${f} in \ *libgcc.a) ;; \ - *) $(AR) t $${f} ;; \ + *.a) $(AR) t $${f} ;; \ + *) cat $${f} ;; \ esac \ done > $@ targets += .vmlinux.objs -.vmlinux.objs: vmlinux.a $(KBUILD_VMLINUX_LIBS) FORCE +.vmlinux.objs: vmlinux.order $(KBUILD_VMLINUX_LIBS) FORCE $(call if_changed,vmlinux_objs) vmlinux.o-if-present := $(wildcard vmlinux.o)
scripts/Makefile.vmlinux_o +3 -3
@@ -18,7 +18,7 @@ quiet_cmd_gen_initcalls_lds = GEN $@ $(PERL) $(real-prereqs) > $@ .tmp_initcalls.lds: $(srctree)/scripts/generate_initcall_order.pl \ - vmlinux.a $(KBUILD_VMLINUX_LIBS) FORCE + vmlinux.order $(KBUILD_VMLINUX_LIBS) FORCE $(call if_changed,gen_initcalls_lds) targets := .tmp_initcalls.lds
@@ -48,7 +48,7 @@ quiet_cmd_ld_vmlinux.o = LD $@ cmd_ld_vmlinux.o = \ $(LD) ${KBUILD_LDFLAGS} -r -o $@ \ $(addprefix -T , $(initcalls-lds)) \ - --whole-archive vmlinux.a --no-whole-archive \ + @vmlinux.order \ --start-group $(KBUILD_VMLINUX_LIBS) --end-group \ $(cmd_objtool)
@@ -57,7 +57,7 @@ define rule_ld_vmlinux.o $(call cmd,gen_objtooldep) endef -vmlinux.o: $(initcalls-lds) vmlinux.a $(KBUILD_VMLINUX_LIBS) FORCE +vmlinux.o: $(initcalls-lds) vmlinux.order $(KBUILD_VMLINUX_LIBS) FORCE $(call if_changed_rule,ld_vmlinux.o) targets += vmlinux.o
scripts/clang-tools/gen_compile_commands.py +20 -1
@@ -57,7 +57,7 @@ def parse_arguments(): parser.add_argument('-a', '--ar', type=str, default='llvm-ar', help=ar_help) paths_help = ('directories to search or files to parse ' - '(files should be *.o, *.a, or modules.order). ' + '(files should be *.o, *.a, or *.order). ' 'If nothing is specified, the current directory is searched') parser.add_argument('paths', type=str, nargs='*', help=paths_help)
@@ -124,6 +124,23 @@ def cmdfiles_for_a(archive, ar): yield to_cmdfile(obj) +def cmdfiles_for_vmlinux_order(vmlinux_order): + """Generate the iterator of .cmd files associated with the vmlinux.order. + + Parse the given vmlinux.order, and yield every .cmd file used to build the + contained modules. + + Args: + vmlinux_order: The vmlinux.order file to parse + + Yields: + The path to every .cmd file found + """ + with open(vmlinux_order) as f: + for line in f: + yield to_cmdfile(line.rstrip()) + + def cmdfiles_for_modorder(modorder): """Generate the iterator of .cmd files associated with the modules.order.
@@ -203,6 +220,8 @@ def main(): cmdfiles = cmdfiles_in_dir(path) elif path.endswith('.a'): cmdfiles = cmdfiles_for_a(path, ar) + elif path.endswith('vmlinux.order'): + cmdfiles = cmdfiles_for_vmlinux_order(path) elif path.endswith('modules.order'): cmdfiles = cmdfiles_for_modorder(path) else:
scripts/link-vmlinux.sh +4 -4
@@ -3,15 +3,15 @@ # # link vmlinux # -# vmlinux is linked from the objects in vmlinux.a and $(KBUILD_VMLINUX_LIBS). -# vmlinux.a contains objects that are linked unconditionally. +# vmlinux is linked from the objects in vmlinux.order and $(KBUILD_VMLINUX_LIBS). +# vmlinux.order is a text file that contains objects that are linked unconditionally. # $(KBUILD_VMLINUX_LIBS) are archives which are linked conditionally # (not within --whole-archive), and do not require symbol indexes added. # # vmlinux # ^ # | -# +--< vmlinux.a +# +--< vmlinux.order # | # +--< $(KBUILD_VMLINUX_LIBS) # | +--< lib/lib.a + more
@@ -65,7 +65,7 @@ vmlinux_link() objs=vmlinux.o libs= else - objs=vmlinux.a + [email protected] libs="${KBUILD_VMLINUX_LIBS}" fi
Signature · 1 line
2.34.1
Hello Masahiro,
Quoted 10 lines · Masahiro Yamada
This is an experimental patch, driven by the feedback from Jiri Slaby and Michael Matz. [1]
Michael Matz says:
"I know of no linker (outside LTO-like modes) that processes
archives in a different order than first-to-last-member (under whole-archive), but that's not guaranteed anywhere. So relying on member-order within archives is always brittle."
Quoted 4 lines
It is pretty easy to pass the list of objects instead of a thin archive because the linker supports the '@file' syntax, where command line arguments are read from 'file'.
Can you explain which problem is solved by doing this?
If we can only produce a working kernel if each object is linked in the order it appears in the archive, I think we have bigger problems that need solving regardless. And for the .head.text objects that need to appear at the start of the binary image, I think the reported issue with __head annotated C functions on x86 needs to be addressed by getting rid of __head entirely (which seems to have been introduced without proper justification)
Quoted 208 lines
Without this patch, the linker receives
--whole-archive vmlinux.a --no-whole-archiveWith this patch, the linker will receive
@vmlinux.order
Here, vmlinux.order is a text file that lists built-in objects in the correct link order.
I am not a toolchain expert. I just want to know if this makes any difference from the linker perspective and from (non-upstreamed) GCC-LTO perspective.
(I know this patch does not work for Clang LTO because I did not touch
scripts/generate_initcall_order.pl)This patch may be unneeded because more correct patches were submitted [2] but I am still curious about "thin archive vs direct object list".
[1]: https://lore.kernel.org/linux-kbuild/[email protected]/
[2]: https://lore.kernel.org/all/[email protected]/Signed-off-by Masahiro Yamada <[email protected]>---
Makefile | 21 ++++++++++----------- scripts/Makefile.modpost | 5 +++-- scripts/Makefile.vmlinux_o | 6 +++--- scripts/clang-tools/gen_compile_commands.py | 21 ++++++++++++++++++++- scripts/link-vmlinux.sh | 8 ++++---- 5 files changed, 40 insertions(+), 21 deletions(-)Makefile +10
-11@@ -1213,19 +1213,18 @@ quiet_cmd_autoksyms_h = GEN $@ $(autoksyms_h): $(call cmd,autoksyms_h) -# '$(AR) mPi' needs 'T' to workaround the bug of llvm-ar <= 14 -quiet_cmd_ar_vmlinux.a = AR $@ - cmd_ar_vmlinux.a = \ - rm -f $@; \ - $(AR) cDPrST $@ $(KBUILD_VMLINUX_OBJS); \ - $(AR) mPiT $$($(AR) t $@ | sed -n 1p) $@ $$($(AR) t $@ | grep -F -f $(srctree)/scripts/head-object-list.txt) +quiet_cmd_vmlinux_order = GEN $@ + cmd_vmlinux_order = \ + { $(foreach m, $(KBUILD_VMLINUX_OBJS), $(AR) t $m;) :; } > $(tmp-target) ; \ + grep -F -f $(srctree)/scripts/head-object-list.txt $(tmp-target) > $@; \ + grep -F -f $(srctree)/scripts/head-object-list.txt $(tmp-target) -v >> $@ -targets += vmlinux.a -vmlinux.a: $(KBUILD_VMLINUX_OBJS) scripts/head-object-list.txt autoksyms_recursive FORCE - $(call if_changed,ar_vmlinux.a) +targets += vmlinux.order +vmlinux.order: $(KBUILD_VMLINUX_OBJS) scripts/head-object-list.txt autoksyms_recursive FORCE + $(call if_changed,vmlinux_order) PHONY += vmlinux_o -vmlinux_o: vmlinux.a $(KBUILD_VMLINUX_LIBS) +vmlinux_o: vmlinux.order $(KBUILD_VMLINUX_LIBS) $(Q)$(MAKE) -f $(srctree)/scripts/Makefile.vmlinux_o vmlinux.o modules.builtin.modinfo modules.builtin: vmlinux_o@@ -2037,7 +2036,7 @@ quiet_cmd_gen_compile_commands = GEN $@ cmd_gen_compile_commands = $(PYTHON3) $< -a $(AR) -o $@ $(filter-out $<, $(real-prereqs)) $(extmod_prefix)compile_commands.json: scripts/clang-tools/gen_compile_commands.py \ - $(if $(KBUILD_EXTMOD),, vmlinux.a $(KBUILD_VMLINUX_LIBS)) \ + $(if $(KBUILD_EXTMOD),, vmlinux.order $(KBUILD_VMLINUX_LIBS)) \ $(if $(CONFIG_MODULES), $(MODORDER)) FORCE $(call if_changed,gen_compile_commands)scripts/Makefile.modpost +3
-2@@ -70,12 +70,13 @@ quiet_cmd_vmlinux_objs = GEN $@ for f in $(real-prereqs); do \ case $${f} in \ *libgcc.a) ;; \ - *) $(AR) t $${f} ;; \ + *.a) $(AR) t $${f} ;; \ + *) cat $${f} ;; \ esac \ done > $@ targets += .vmlinux.objs -.vmlinux.objs: vmlinux.a $(KBUILD_VMLINUX_LIBS) FORCE +.vmlinux.objs: vmlinux.order $(KBUILD_VMLINUX_LIBS) FORCE $(call if_changed,vmlinux_objs) vmlinux.o-if-present := $(wildcard vmlinux.o)scripts/Makefile.vmlinux_o +3
-3@@ -18,7 +18,7 @@ quiet_cmd_gen_initcalls_lds = GEN $@ $(PERL) $(real-prereqs) > $@ .tmp_initcalls.lds: $(srctree)/scripts/generate_initcall_order.pl \ - vmlinux.a $(KBUILD_VMLINUX_LIBS) FORCE + vmlinux.order $(KBUILD_VMLINUX_LIBS) FORCE $(call if_changed,gen_initcalls_lds) targets := .tmp_initcalls.lds@@ -48,7 +48,7 @@ quiet_cmd_ld_vmlinux.o = LD $@ cmd_ld_vmlinux.o = \ $(LD) ${KBUILD_LDFLAGS} -r -o $@ \ $(addprefix -T , $(initcalls-lds)) \ - --whole-archive vmlinux.a --no-whole-archive \ + @vmlinux.order \ --start-group $(KBUILD_VMLINUX_LIBS) --end-group \ $(cmd_objtool)@@ -57,7 +57,7 @@ define rule_ld_vmlinux.o $(call cmd,gen_objtooldep) endef -vmlinux.o: $(initcalls-lds) vmlinux.a $(KBUILD_VMLINUX_LIBS) FORCE +vmlinux.o: $(initcalls-lds) vmlinux.order $(KBUILD_VMLINUX_LIBS) FORCE $(call if_changed_rule,ld_vmlinux.o) targets += vmlinux.oscripts/clang-tools/gen_compile_commands.py +20
-1@@ -57,7 +57,7 @@ def parse_arguments(): parser.add_argument('-a', '--ar', type=str, default='llvm-ar', help=ar_help) paths_help = ('directories to search or files to parse ' - '(files should be *.o, *.a, or modules.order). ' + '(files should be *.o, *.a, or *.order). ' 'If nothing is specified, the current directory is searched') parser.add_argument('paths', type=str, nargs='*', help=paths_help)@@ -124,6 +124,23 @@ def cmdfiles_for_a(archive, ar): yield to_cmdfile(obj) +def cmdfiles_for_vmlinux_order(vmlinux_order): + """Generate the iterator of .cmd files associated with the vmlinux.order. + + Parse the given vmlinux.order, and yield every .cmd file used to build the + contained modules. + + Args: + vmlinux_order: The vmlinux.order file to parse + + Yields: + The path to every .cmd file found + """ + with open(vmlinux_order) as f: + for line in f: + yield to_cmdfile(line.rstrip()) + + def cmdfiles_for_modorder(modorder): """Generate the iterator of .cmd files associated with the modules.order.@@ -203,6 +220,8 @@ def main(): cmdfiles = cmdfiles_in_dir(path) elif path.endswith('.a'): cmdfiles = cmdfiles_for_a(path, ar) + elif path.endswith('vmlinux.order'): + cmdfiles = cmdfiles_for_vmlinux_order(path) elif path.endswith('modules.order'): cmdfiles = cmdfiles_for_modorder(path) else:scripts/link-vmlinux.sh +4
-5@@ -3,15 +3,15 @@ # # link vmlinux # -# vmlinux is linked from the objects in vmlinux.a and $(KBUILD_VMLINUX_LIBS). -# vmlinux.a contains objects that are linked unconditionally. +# vmlinux is linked from the objects in vmlinux.order and $(KBUILD_VMLINUX_LIBS). +# vmlinux.order is a text file that contains objects that are linked unconditionally. # $(KBUILD_VMLINUX_LIBS) are archives which are linked conditionally # (not within --whole-archive), and do not require symbol indexes added. # # vmlinux # ^ # | -# +--< vmlinux.a +# +--< vmlinux.order # | # +--< $(KBUILD_VMLINUX_LIBS) # | +--< lib/lib.a + more@@ -65,7 +65,7 @@ vmlinux_link() objs=vmlinux.o libs= else - objs=vmlinux.a + [email protected] libs="${KBUILD_VMLINUX_LIBS}" fi --2.34.1
Quoted 14 lines · Ard Biesheuvel
Hello Masahiro,
Quoted 10 lines · Masahiro Yamada
This is an experimental patch, driven by the feedback from Jiri Slaby and Michael Matz. [1]
Michael Matz says:
"I know of no linker (outside LTO-like modes) that processes
archives in a different order than first-to-last-member (under whole-archive), but that's not guaranteed anywhere. So relying on member-order within archives is always brittle."
Quoted 6 lines
Quoted 4 lines
It is pretty easy to pass the list of objects instead of a thin archive because the linker supports the '@file' syntax, where command line arguments are read from 'file'.
Can you explain which problem is solved by doing this?
Jiri Slaby reported that the (not-upstreamed) GCC-LTO tree got broken due to 321648455061 ("kbuild: use obj-y instead extra-y for objects placed at the head")
https://lore.kernel.org/linux-kbuild/[email protected]/
I am not pretty sure because I did not check the downstream code.
If I understood his report correctly, the reason for the breakage is because I put all objects into the thin archive, expecting the linker would preserve the object order in the archive.
By specifying the object order directly in the command line, GCC-LTO should get back working again.
Quoted 8 lines
If we can only produce a working kernel if each object is linked in the order it appears in the archive, I think we have bigger problems that need solving regardless. And for the .head.text objects that need to appear at the start of the binary image, I think the reported issue with
__headannotated C functions on x86 needs to be addressed by getting rid of__headentirely (which seems to have been introduced without proper justification)
I agree that it is the correct approach.
I think my patch is unneeded (hence RFC), but I just wanted to know if linkers (gnu ld and lld) see any difference.
--
Best Regards
Masahiro Yamada
Quoted 16 lines · Masahiro Yamada
Jiri Slaby reported that the (not-upstreamed) GCC-LTO tree got broken due to 321648455061 ("kbuild: use obj-y instead extra-y for objects placed at the head")
https://lore.kernel.org/linux-kbuild/[email protected]/
I am not pretty sure because I did not check the downstream code.
If I understood his report correctly, the reason for the breakage is because I put all objects into the thin archive, expecting the linker would preserve the object order in the archive.
By specifying the object order directly in the command line, GCC-LTO should get back working again.
...
Quoted 2 lines
I think my patch is unneeded (hence RFC), but I just wanted to know if linkers (gnu ld and lld) see any difference.
Yes, including your patch (instead of vmlinux.S and /compressed/* changes) also fixes the issue. So the linker (gcc-ld) indeed respects command line, but not order of objects in .a.
thanks,
Signature · 2 lines
js suse labs
Quoted 19 lines · Jiri Slaby
Quoted 10 lines · Masahiro Yamada
Jiri Slaby reported that the (not-upstreamed) GCC-LTO tree got broken due to 321648455061 ("kbuild: use obj-y instead extra-y for objects placed at the head")
https://lore.kernel.org/linux-kbuild/[email protected]/
I am not pretty sure because I did not check the downstream code.
Quoted 6 lines
If I understood his report correctly, the reason for the breakage is because I put all objects into the thin archive, expecting the linker would preserve the object order in the archive.
By specifying the object order directly in the command line, GCC-LTO should get back working again.
Quoted 9 lines
...
Quoted 2 lines
I think my patch is unneeded (hence RFC), but I just wanted to know if linkers (gnu ld and lld) see any difference.
Yes, including your patch (instead of
vmlinux.Sand /compressed/* changes) also fixes the issue. So the linker (gcc-ld) indeed respects command line, but not order of objects in .a.
Can I see your LTO implementation somewhere in public?
I know Andy Kleen's GCC-LTO, the latest version is 5.13-ish.
https://github.com/andikleen/linux-misc.git
Quoted 5 lines
thanks,
--
js
suse labs
Signature · 2 lines
Best Regards Masahiro Yamada
Quoted 1 line · Masahiro Yamada
Can I see your LTO implementation somewhere in public?
Sure:
https://git.kernel.org/pub/scm/linux/kernel/git/jirislaby/linux.git/commit/?h=lto
FWIW, I plan to send a v1 for comments soon as I finished the patch order, CCs and commit logs finally.
thanks,
Signature · 2 lines
js suse labs