| Message ID | 20220831184408.2778264-5-ndesaulniers@google.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | fix debug info for asm and DEBUG_INFO_SPLIT | expand |
On Wed, Aug 31, 2022 at 11:44:07AM -0700, Nick Desaulniers wrote: > Alexey reported that the fraction of unknown filename instances in > kallsyms grew from ~0.3% to ~10% recently; Bill and Greg tracked it down > to assembler defined symbols, which regressed as a result of: > > commit b8a9092330da ("Kbuild: do not emit debug info for assembly with LLVM_IAS=1") > > In that commit, I allude to restoring debug info for assembler defined > symbols in a follow up patch, but it seems I forgot to do so in > > commit a66049e2cf0e ("Kbuild: make DWARF version a choice") > > This patch does a few things: > 1. Add -g to KBUILD_AFLAGS. This will instruct the compiler to instruct > the assembler to emit debug info. But this can cause an issue for > folks using a newer compiler but older assembler, because the > implicit default DWARF version changed from v4 to v5 in gcc-11 and > clang-14. > 2. If the user is using CONFIG_DEBUG_INFO_DWARF_TOOLCHAIN_DEFAULT, use a > version check to explicitly set -Wa,-gdwarf-<version> for the > assembler. There's another problem with this; GAS only gained support > for explicit DWARF versions 3-5 in the 2.36 GNU binutils release. > 3. Wrap -Wa,-gdwarf-<version> in as-option call to test whether the > assembler supports that explicit DWARF version. > > Link: https://sourceware.org/git/gitweb.cgi?p=binutils-gdb.git;h=31bf18645d98b4d3d7357353be840e320649a67d > Fixes: b8a9092330da ("Kbuild: do not emit debug info for assembly with LLVM_IAS=1") > Reported-by: Alexey Alexandrov <aalexand@google.com> > Reported-by: Bill Wendling <morbo@google.com> > Reported-by: Greg Thelen <gthelen@google.com> > Signed-off-by: Nick Desaulniers <ndesaulniers@google.com> Reviewed-by: Nathan Chancellor <nathan@kernel.org> > --- > Changes v1 -> v2: > * Use newly added compiler-specific macros, as per Bill. > > scripts/Makefile.debug | 22 ++++++++++++++++++---- > 1 file changed, 18 insertions(+), 4 deletions(-) > > diff --git a/scripts/Makefile.debug b/scripts/Makefile.debug > index 9f39b0130551..46e88f0ca998 100644 > --- a/scripts/Makefile.debug > +++ b/scripts/Makefile.debug > @@ -4,18 +4,32 @@ ifdef CONFIG_DEBUG_INFO_SPLIT > DEBUG_CFLAGS += -gsplit-dwarf > else > DEBUG_CFLAGS += -g > +KBUILD_AFLAGS += -g > endif > > -ifndef CONFIG_AS_IS_LLVM > -KBUILD_AFLAGS += -Wa,-gdwarf-2 > +ifdef CONFIG_DEBUG_INFO_DWARF_TOOLCHAIN_DEFAULT > +# gcc-11+, clang-14+ > +ifeq ($(call cc-min-version, 110000, 140000),y) > +dwarf-version-y := 5 > +else > +dwarf-version-y := 4 > endif > - > -ifndef CONFIG_DEBUG_INFO_DWARF_TOOLCHAIN_DEFAULT > +else # !CONFIG_DEBUG_INFO_DWARF_TOOLCHAIN_DEFAULT > dwarf-version-$(CONFIG_DEBUG_INFO_DWARF4) := 4 > dwarf-version-$(CONFIG_DEBUG_INFO_DWARF5) := 5 > DEBUG_CFLAGS += -gdwarf-$(dwarf-version-y) > endif > > +# Binutils 2.35+ (or clang) required for -gdwarf-{4|5}. > +# https://sourceware.org/git/gitweb.cgi?p=binutils-gdb.git;h=31bf18645d98b4d3d7357353be840e320649a67d > +ifneq ($(call as-option,-Wa$(comma)-gdwarf-$(dwarf-version-y)),) > +KBUILD_AFLAGS += -Wa,-gdwarf-$(dwarf-version-y) > +else > +ifndef CONFIG_AS_IS_LLVM > +KBUILD_AFLAGS += -Wa,-gdwarf-2 > +endif > +endif > + > ifdef CONFIG_DEBUG_INFO_REDUCED > DEBUG_CFLAGS += -fno-var-tracking > ifdef CONFIG_CC_IS_GCC > -- > 2.37.2.672.g94769d06f0-goog >
On Thu, Sep 1, 2022 at 3:44 AM Nick Desaulniers <ndesaulniers@google.com> wrote: > > Alexey reported that the fraction of unknown filename instances in > kallsyms grew from ~0.3% to ~10% recently; Bill and Greg tracked it down > to assembler defined symbols, which regressed as a result of: > > commit b8a9092330da ("Kbuild: do not emit debug info for assembly with LLVM_IAS=1") > > In that commit, I allude to restoring debug info for assembler defined > symbols in a follow up patch, but it seems I forgot to do so in > > commit a66049e2cf0e ("Kbuild: make DWARF version a choice") > > This patch does a few things: > 1. Add -g to KBUILD_AFLAGS. This will instruct the compiler to instruct > the assembler to emit debug info. But this can cause an issue for > folks using a newer compiler but older assembler, because the > implicit default DWARF version changed from v4 to v5 in gcc-11 and > clang-14. What kind of bad things happen for "KBUILD_AFLAGS += -g"? I think 'gcc -g -c -o foo.o foo.S' will invoke 'as --gdwarf-2' as the backend if gcc is configured to work with old binutils. > diff --git a/scripts/Makefile.debug b/scripts/Makefile.debug > index 9f39b0130551..46e88f0ca998 100644 > --- a/scripts/Makefile.debug > +++ b/scripts/Makefile.debug > @@ -4,18 +4,32 @@ ifdef CONFIG_DEBUG_INFO_SPLIT > DEBUG_CFLAGS += -gsplit-dwarf > else > DEBUG_CFLAGS += -g > +KBUILD_AFLAGS += -g > endif > > -ifndef CONFIG_AS_IS_LLVM > -KBUILD_AFLAGS += -Wa,-gdwarf-2 > +ifdef CONFIG_DEBUG_INFO_DWARF_TOOLCHAIN_DEFAULT > +# gcc-11+, clang-14+ > +ifeq ($(call cc-min-version, 110000, 140000),y) > +dwarf-version-y := 5 > +else > +dwarf-version-y := 4 If you explicitly specify the DWARF version for CONFIG_DEBUG_INFO_DWARF_TOOLCHAIN_DEFAULT, what is the point of CONFIG_DEBUG_INFO_DWARF_TOOLCHAIN_DEFAULT? When CONFIG_DEBUG_INFO_DWARF_TOOLCHAIN_DEFAULT=y, I believe the right thing to do is to pass only -g, and let the tool do whatever it thinks is appropriate. > endif > - > -ifndef CONFIG_DEBUG_INFO_DWARF_TOOLCHAIN_DEFAULT > +else # !CONFIG_DEBUG_INFO_DWARF_TOOLCHAIN_DEFAULT > dwarf-version-$(CONFIG_DEBUG_INFO_DWARF4) := 4 > dwarf-version-$(CONFIG_DEBUG_INFO_DWARF5) := 5 > DEBUG_CFLAGS += -gdwarf-$(dwarf-version-y) > endif > > +# Binutils 2.35+ (or clang) required for -gdwarf-{4|5}. > +# https://sourceware.org/git/gitweb.cgi?p=binutils-gdb.git;h=31bf18645d98b4d3d7357353be840e320649a67d > +ifneq ($(call as-option,-Wa$(comma)-gdwarf-$(dwarf-version-y)),) When is this as-option supposed to fail? Binutils <= 2.34 always accepts whatever -gdwarf-* option. Surprisingly or not, it accepts -gdwarf-6, -gdwarf-7, ... No matter what DWARF version you specify, GAS silently downgrades it to DWARF-2. masahiro@zoe:~/tools/binutils-2.34/bin$ ./as --version | head -n 1 GNU assembler (GNU Binutils) 2.34 masahiro@zoe:~/tools/binutils-2.34/bin$ cat /dev/null | ./as -gdwarf-5 -o /dev/null - masahiro@zoe:~/tools/binutils-2.34/bin$ echo $? 0 masahiro@zoe:~/tools/binutils-2.34/bin$ cat /dev/null | ./as -gdwarf-100 -o /dev/null - masahiro@zoe:~/tools/binutils-2.34/bin$ echo $? 0 Overall, I am not convinced with this patch. Please see the attached patch. Is there any problem with writing this more simply? > +KBUILD_AFLAGS += -Wa,-gdwarf-$(dwarf-version-y) > +else > +ifndef CONFIG_AS_IS_LLVM > +KBUILD_AFLAGS += -Wa,-gdwarf-2 > +endif > +endif > + > ifdef CONFIG_DEBUG_INFO_REDUCED > DEBUG_CFLAGS += -fno-var-tracking > ifdef CONFIG_CC_IS_GCC > -- > 2.37.2.672.g94769d06f0-goog > -- Best Regards Masahiro Yamada
On Mon, Sep 5, 2022 at 12:50 AM Masahiro Yamada <masahiroy@kernel.org> wrote: > > On Thu, Sep 1, 2022 at 3:44 AM Nick Desaulniers <ndesaulniers@google.com> wrote: > > > > Alexey reported that the fraction of unknown filename instances in > > kallsyms grew from ~0.3% to ~10% recently; Bill and Greg tracked it down > > to assembler defined symbols, which regressed as a result of: > > > > commit b8a9092330da ("Kbuild: do not emit debug info for assembly with LLVM_IAS=1") > > > > In that commit, I allude to restoring debug info for assembler defined > > symbols in a follow up patch, but it seems I forgot to do so in > > > > commit a66049e2cf0e ("Kbuild: make DWARF version a choice") > > > > This patch does a few things: > > 1. Add -g to KBUILD_AFLAGS. This will instruct the compiler to instruct > > the assembler to emit debug info. But this can cause an issue for > > folks using a newer compiler but older assembler, because the > > implicit default DWARF version changed from v4 to v5 in gcc-11 and > > clang-14. > > > > What kind of bad things happen for "KBUILD_AFLAGS += -g"? > > > I think 'gcc -g -c -o foo.o foo.S' will invoke 'as --gdwarf-2' as the backend > if gcc is configured to work with old binutils. That's fine for CONFIG_DEBUG_INFO_DWARF_TOOLCHAIN_DEFAULT I think? What other problems were you envisioning? > > > > > > diff --git a/scripts/Makefile.debug b/scripts/Makefile.debug > > index 9f39b0130551..46e88f0ca998 100644 > > --- a/scripts/Makefile.debug > > +++ b/scripts/Makefile.debug > > @@ -4,18 +4,32 @@ ifdef CONFIG_DEBUG_INFO_SPLIT > > DEBUG_CFLAGS += -gsplit-dwarf > > else > > DEBUG_CFLAGS += -g > > +KBUILD_AFLAGS += -g > > endif > > > > -ifndef CONFIG_AS_IS_LLVM > > -KBUILD_AFLAGS += -Wa,-gdwarf-2 > > +ifdef CONFIG_DEBUG_INFO_DWARF_TOOLCHAIN_DEFAULT > > +# gcc-11+, clang-14+ > > +ifeq ($(call cc-min-version, 110000, 140000),y) > > +dwarf-version-y := 5 > > +else > > +dwarf-version-y := 4 > > > > If you explicitly specify the DWARF version > for CONFIG_DEBUG_INFO_DWARF_TOOLCHAIN_DEFAULT, > what is the point of CONFIG_DEBUG_INFO_DWARF_TOOLCHAIN_DEFAULT? > > > When CONFIG_DEBUG_INFO_DWARF_TOOLCHAIN_DEFAULT=y, > I believe the right thing to do is to pass only -g, > and let the tool do whatever it thinks is appropriate. Ok, sure, I will revise. > > > > > > > > endif > > - > > -ifndef CONFIG_DEBUG_INFO_DWARF_TOOLCHAIN_DEFAULT > > +else # !CONFIG_DEBUG_INFO_DWARF_TOOLCHAIN_DEFAULT > > dwarf-version-$(CONFIG_DEBUG_INFO_DWARF4) := 4 > > dwarf-version-$(CONFIG_DEBUG_INFO_DWARF5) := 5 > > DEBUG_CFLAGS += -gdwarf-$(dwarf-version-y) > > endif > > > > +# Binutils 2.35+ (or clang) required for -gdwarf-{4|5}. > > +# https://sourceware.org/git/gitweb.cgi?p=binutils-gdb.git;h=31bf18645d98b4d3d7357353be840e320649a67d > > +ifneq ($(call as-option,-Wa$(comma)-gdwarf-$(dwarf-version-y)),) > > > > When is this as-option supposed to fail? > > > Binutils <= 2.34 always accepts whatever -gdwarf-* option. > Surprisingly or not, it accepts -gdwarf-6, -gdwarf-7, ... > > No matter what DWARF version you specify, GAS silently downgrades > it to DWARF-2. > > > masahiro@zoe:~/tools/binutils-2.34/bin$ ./as --version | head -n 1 > GNU assembler (GNU Binutils) 2.34 > masahiro@zoe:~/tools/binutils-2.34/bin$ cat /dev/null | ./as -gdwarf-5 > -o /dev/null - > masahiro@zoe:~/tools/binutils-2.34/bin$ echo $? > 0 > masahiro@zoe:~/tools/binutils-2.34/bin$ cat /dev/null | ./as > -gdwarf-100 -o /dev/null - > masahiro@zoe:~/tools/binutils-2.34/bin$ echo $? > 0 ah, right. Maybe an explicit version check is necessary then. > > > > > Overall, I am not convinced with this patch. > > > > Please see the attached patch. > Is there any problem with writing this more simply? Thanks for the inspiration, I will use that as an inspiration/base for a new patch.
diff --git a/scripts/Makefile.debug b/scripts/Makefile.debug index 9f39b0130551..46e88f0ca998 100644 --- a/scripts/Makefile.debug +++ b/scripts/Makefile.debug @@ -4,18 +4,32 @@ ifdef CONFIG_DEBUG_INFO_SPLIT DEBUG_CFLAGS += -gsplit-dwarf else DEBUG_CFLAGS += -g +KBUILD_AFLAGS += -g endif -ifndef CONFIG_AS_IS_LLVM -KBUILD_AFLAGS += -Wa,-gdwarf-2 +ifdef CONFIG_DEBUG_INFO_DWARF_TOOLCHAIN_DEFAULT +# gcc-11+, clang-14+ +ifeq ($(call cc-min-version, 110000, 140000),y) +dwarf-version-y := 5 +else +dwarf-version-y := 4 endif - -ifndef CONFIG_DEBUG_INFO_DWARF_TOOLCHAIN_DEFAULT +else # !CONFIG_DEBUG_INFO_DWARF_TOOLCHAIN_DEFAULT dwarf-version-$(CONFIG_DEBUG_INFO_DWARF4) := 4 dwarf-version-$(CONFIG_DEBUG_INFO_DWARF5) := 5 DEBUG_CFLAGS += -gdwarf-$(dwarf-version-y) endif +# Binutils 2.35+ (or clang) required for -gdwarf-{4|5}. +# https://sourceware.org/git/gitweb.cgi?p=binutils-gdb.git;h=31bf18645d98b4d3d7357353be840e320649a67d +ifneq ($(call as-option,-Wa$(comma)-gdwarf-$(dwarf-version-y)),) +KBUILD_AFLAGS += -Wa,-gdwarf-$(dwarf-version-y) +else +ifndef CONFIG_AS_IS_LLVM +KBUILD_AFLAGS += -Wa,-gdwarf-2 +endif +endif + ifdef CONFIG_DEBUG_INFO_REDUCED DEBUG_CFLAGS += -fno-var-tracking ifdef CONFIG_CC_IS_GCC
Alexey reported that the fraction of unknown filename instances in kallsyms grew from ~0.3% to ~10% recently; Bill and Greg tracked it down to assembler defined symbols, which regressed as a result of: commit b8a9092330da ("Kbuild: do not emit debug info for assembly with LLVM_IAS=1") In that commit, I allude to restoring debug info for assembler defined symbols in a follow up patch, but it seems I forgot to do so in commit a66049e2cf0e ("Kbuild: make DWARF version a choice") This patch does a few things: 1. Add -g to KBUILD_AFLAGS. This will instruct the compiler to instruct the assembler to emit debug info. But this can cause an issue for folks using a newer compiler but older assembler, because the implicit default DWARF version changed from v4 to v5 in gcc-11 and clang-14. 2. If the user is using CONFIG_DEBUG_INFO_DWARF_TOOLCHAIN_DEFAULT, use a version check to explicitly set -Wa,-gdwarf-<version> for the assembler. There's another problem with this; GAS only gained support for explicit DWARF versions 3-5 in the 2.36 GNU binutils release. 3. Wrap -Wa,-gdwarf-<version> in as-option call to test whether the assembler supports that explicit DWARF version. Link: https://sourceware.org/git/gitweb.cgi?p=binutils-gdb.git;h=31bf18645d98b4d3d7357353be840e320649a67d Fixes: b8a9092330da ("Kbuild: do not emit debug info for assembly with LLVM_IAS=1") Reported-by: Alexey Alexandrov <aalexand@google.com> Reported-by: Bill Wendling <morbo@google.com> Reported-by: Greg Thelen <gthelen@google.com> Signed-off-by: Nick Desaulniers <ndesaulniers@google.com> --- Changes v1 -> v2: * Use newly added compiler-specific macros, as per Bill. scripts/Makefile.debug | 22 ++++++++++++++++++---- 1 file changed, 18 insertions(+), 4 deletions(-)