kbuild: remove duplicate dependencies from .mod files
PATCH 1/1 kbuild: remove duplicate dependencies from .mod files
With CONFIG_TRIM_UNUSED_SYMS, if a module has enough dependencies to exceed the default xargs command line size limit, the output is split into multiple lines, which can result in used symbols getting trimmed.
This change removes duplicate dependencies, which will reduce the probability of this happening and makes .mod files smaller and easier to read.
--- scripts/Makefile.build | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)
scripts/Makefile.build +1 -1
@@ -257,7 +257,7 @@ endef # List module undefined symbols (or empty line if not enabled) ifdef CONFIG_TRIM_UNUSED_KSYMS -cmd_undef_syms = $(NM) $< | sed -n 's/^ *U //p' | xargs echo +cmd_undef_syms = $(NM) $< | sed -n 's/^ *U //p' | sort -u | xargs echo else cmd_undef_syms = echo endif
base-commit: bb6d3fb354c5ee8d6bde2d576eb7220ea09862b9
Signature · 1 line
2.25.0.341.g760bfbb309-goog
Hi Sami,
Quoted 19 lines · Sami Tolvanen
With
CONFIG_TRIM_UNUSED_SYMS, if a module has enough dependencies to exceed the default xargs command line size limit, the output is split into multiple lines, which can result in used symbols getting trimmed.This change removes duplicate dependencies, which will reduce the probability of this happening and makes .mod files smaller and easier to read.
Signed-off-by Sami Tolvanen <[email protected]>--- scripts/Makefile.build | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)scripts/Makefile.build +0
-0
scripts/Makefile.build endef · line 260
Quoted 5 lines
scripts/Makefile.build +1
-1@@ -257,7 +257,7 @@ endef # List module undefined symbols (or empty line if not enabled) ifdef CONFIG_TRIM_UNUSED_KSYMS -cmd_undef_syms = $(NM) $< | sed -n 's/^ *U //p' | xargs echo +cmd_undef_syms = $(NM) $< | sed -n 's/^ *U //p' | sort -u | xargs echo
In which case are undefined symbols duplicated?
Do you have a .config to reproduce it?
Quoted 8 lines
scripts/Makefile.build +0
-0else cmd_undef_syms = echo endifbase-commit:
bb6d3fb354c5ee8d6bde2d576eb7220ea09862b9--
2.25.0.341.g760bfbb309-goog
Signature · 2 lines
Best Regards Masahiro Yamada
Quoted 1 line · Masahiro Yamada
In which case are undefined symbols duplicated?
When a module consists of multiple compilation units, which depend on the same external symbols. In Android, we ran into this when adding hardening features that all depend on an external error handler function with a rather long name. When CONFIG_TRIM_UNUSED_SYMS was later enabled, we ran into this:
$ llvm-nm drivers/gpu/drm/nouveau/nouveau.o | sed -n 's/^ *U //p' |
xargs echo | wc
2 9136 168660xargs defaults to 128kiB limit for command line size, so the output was split into two lines, which means some of the dependencies were dropped and we ran into modpost errors. One method of fixing this is to increase the limit:
$ llvm-nm drivers/gpu/drm/nouveau/nouveau.o | sed -n 's/^ *U //p' |
xargs -s 262144 echo | wc
1 9136 168660But it seems removing duplicates is a better solution as the length of the dependency list is reduced significantly:
$ llvm-nm drivers/gpu/drm/nouveau/nouveau.o | sed -n 's/^ *U //p' |
sort -u | xargs echo | wc
1 2716 50461Quoted 1 line
Do you have a .config to reproduce it?
I can currently reproduce this on an Android kernel that has Control-Flow Integrity (CFI) enabled. While this feature is not upstreamed yet, there's nothing that would prevent us from hitting the command line limit with sufficiently large modules otherwise as well.
Sami
Hi Sami,
Quoted 23 lines · Sami Tolvanen
Quoted 1 line · Masahiro Yamada
In which case are undefined symbols duplicated?
When a module consists of multiple compilation units, which depend on the same external symbols. In Android, we ran into this when adding hardening features that all depend on an external error handler function with a rather long name. When
CONFIG_TRIM_UNUSED_SYMSwas later enabled, we ran into this:$ llvm-nm drivers/gpu/drm/nouveau/nouveau.o | sed -n 's/^ *U //p' | xargs echo | wc 2 9136 168660xargs defaults to 128kiB limit for command line size, so the output was split into two lines, which means some of the dependencies were dropped and we ran into modpost errors. One method of fixing this is to increase the limit:
$ llvm-nm drivers/gpu/drm/nouveau/nouveau.o | sed -n 's/^ *U //p' | xargs -s 262144 echo | wc 1 9136 168660
Quoted 6 lines
But it seems removing duplicates is a better solution as the length of the dependency list is reduced significantly:
$ llvm-nm drivers/gpu/drm/nouveau/nouveau.o | sed -n 's/^ *U //p' | sort -u | xargs echo | wc 1 2716 50461
At least, I am unable to reproduce this in upstream.
This is my result for x86 allmodconfig builds.
masahiro@grover:~/workspace/linux-kbuild$ nm
drivers/gpu/drm/nouveau/nouveau.o | sed -n 's/^ *U //p' | wc
572 572 11478
masahiro@grover:~/workspace/linux-kbuild$ nm
drivers/gpu/drm/nouveau/nouveau.o | sed -n 's/^ *U //p' | sort -u |
wc
572 572 11478I see no difference with/without 'sort -u'.
I also tried llvm-nm instead of GNU nm,
but the result is the same.
masahiro@grover:~/workspace/linux-kbuild$ llvm-nm
drivers/gpu/drm/nouveau/nouveau.o | sed -n 's/^ *U //p' | wc
572 572 11478
masahiro@grover:~/workspace/linux-kbuild$ llvm-nm
drivers/gpu/drm/nouveau/nouveau.o | sed -n 's/^ *U //p' | sort -u |
wc
572 572 11478Quoted 6 lines
Quoted 1 line
Do you have a .config to reproduce it?
I can currently reproduce this on an Android kernel that has Control-Flow Integrity (CFI) enabled. While this feature is not upstreamed yet, there's nothing that would prevent us from hitting the command line limit with sufficiently large modules otherwise as well.
Does ACK do this differently?
I think it would be strange
if $(NM) duplicated undefined symbols.
Signature · 2 lines
Best Regards Masahiro Yamada
Hi Masahiro,
Quoted 4 lines · Masahiro Yamada
masahiro@grover:~/workspace/linux-kbuild$ llvm-nm
drivers/gpu/drm/nouveau/nouveau.o | sed -n 's/^ *U //p' | sort -u |
wc
572 572 11478Thank you for testing this!
Quoted 1 line
Does ACK do this differently?
Yes, the difference is that we use LTO and after looking into this a bit more, it turns out that when the individual files in the thin archive are LLVM bitcode, llvm-nm behaves differently and prints out the symbols for each file separately.
$ llvm-nm drivers/gpu/drm/nouveau/nouveau.o
...
nvif/client.o:
---------------- W __cfi_check
---------------- W __cfi_check_fail
U __cfi_slowpath_diag
U __ubsan_handle_cfi_check_fail
---------------- T nvif_client_fini
---------------- T nvif_client_init
---------------- T nvif_client_ioctl
---------------- T nvif_client_resume
---------------- T nvif_client_suspend
U nvif_object_fini
U nvif_object_init
U strncpynvif/device.o:
---------------- W __cfi_check
---------------- W __cfi_check_fail
U __ubsan_handle_cfi_check_fail
U kfree
---------------- T nvif_device_fini
---------------- T nvif_device_init
---------------- T nvif_device_time
U nvif_object_fini
U nvif_object_init
U nvif_object_mthd
U nvif_user_fini
...
While this output format still works for us, it does generate a lot of duplicates. Anyway, I think we can come back to this when LTO is closer to getting upstreamed.
Sami