kbuild: add ability to generate BTF type info for vmlinux

This patch adds new config option to trigger generation of BTF type information from DWARF debuginfo for vmlinux and kernel modules through pahole, which in turn relies on libbpf for btf_dedup() algorithm.

The intent is to record compact type information of all types used inside kernel, including all the structs/unions/typedefs/etc. This enables BPF's compile-once-run-everywhere ([0]) approach, in which tracing programs that are inspecting kernel's internal data (e.g., struct task_struct) can be compiled on a system running some kernel version, but would be possible to run on other kernel versions (and configurations) without recompilation, even if the layout of structs changed and/or some of the fields were added, removed, or renamed.

This is only possible if BPF loader can get kernel type info to adjust all the offsets correctly. This patch is a first time in this direction, making sure that BTF type info is part of Linux kernel image in non-loadable ELF section.

BTF deduplication ([1]) algorithm typically provides 100x savings compared to DWARF data, so resulting .BTF section is not big as is typically about 2MB in size.

[0] http://vger.kernel.org/lpc-bpf2018.html#session-2
[1] https://facebookmicrosites.github.io/bpf/blog/2018/11/14/btf-enhancement.html

Cc Masahiro Yamada <[email protected]>

Cc Arnaldo Carvalho de Melo <[email protected]>

Cc Daniel Borkmann <[email protected]>

Cc Alexei Starovoitov <[email protected]>

Cc Yonghong Song <[email protected]>

Cc Martin KaFai Lau <[email protected]>

Signed-off-by Andrii Nakryiko <[email protected]>

---
  Makefile                |  3 ++-
  lib/Kconfig.debug       |  8 ++++++++
  scripts/link-vmlinux.sh | 20 +++++++++++++++++++-
  3 files changed, 29 insertions(+), 2 deletions(-)

Makefile +2 -1

index d5713e7b1e50..a55308147a09 100644
@@ -387,6 +387,7 @@ NM		= $(CROSS_COMPILE)nm
  STRIP		= $(CROSS_COMPILE)strip
  OBJCOPY		= $(CROSS_COMPILE)objcopy
  OBJDUMP		= $(CROSS_COMPILE)objdump
+PAHOLE		= pahole
  LEX		= flex
  YACC		= bison
  AWK		= awk

@@ -442,7 +443,7 @@ KBUILD_LDFLAGS :=
  GCC_PLUGINS_CFLAGS :=
 
  export ARCH SRCARCH CONFIG_SHELL HOSTCC KBUILD_HOSTCFLAGS CROSS_COMPILE AS LD CC
-export CPP AR NM STRIP OBJCOPY OBJDUMP KBUILD_HOSTLDFLAGS KBUILD_HOSTLDLIBS
+export CPP AR NM STRIP OBJCOPY OBJDUMP PAHOLE KBUILD_HOSTLDFLAGS KBUILD_HOSTLDLIBS
  export MAKE LEX YACC AWK GENKSYMS INSTALLKERNEL PERL PYTHON PYTHON2 PYTHON3 UTS_MACHINE
  export HOSTCXX KBUILD_HOSTCXXFLAGS LDFLAGS_MODULE CHECK CHECKFLAGS
 

lib/Kconfig.debug +8 -0

index d4df5b24d75e..cce78dcd19a2 100644
@@ -200,6 +200,14 @@ config DEBUG_INFO_DWARF4
  	  But it significantly improves the success of resolving
  	  variables in gdb on optimized code.
 
+config DEBUG_INFO_BTF
+	bool "Generate BTF typeinfo"
+	depends on DEBUG_INFO
+	help
+	  Generate deduplicated BTF type information from DWARF debug info.
+	  Turning this on expects presence of pahole tool, which will convert
+	  DWARF type info into equivalent deduplicated BTF type info.
+
  config GDB_SCRIPTS
  	bool "Provide GDB scripts for kernel debugging"
  	depends on DEBUG_INFO

scripts/link-vmlinux.sh +19 -1

index c8cf45362bd6..73bb7dfdc2c9 100755
@@ -40,7 +40,7 @@ set -e
  info()
  {
  	if [ "${quiet}" != "silent_" ]; then
-		printf "  %-7s %s\n" ${1} ${2}
+		printf "  %-7s %s\n" "${1}" "${2}"
  	fi
  }
 

@@ -114,6 +114,20 @@ vmlinux_link()
  	fi
  }
 
+# generate .BTF typeinfo from DWARF debuginfo
+gen_btf()
+{
+	local pahole_ver;
+
+	pahole_ver=$(${PAHOLE} --version | sed -E 's/v([0-9]+)\.([0-9]+)/\1\2/')
+	if [ "${pahole_ver}" -lt "113" ]; then
+		info "BTF" "${1}: pahole version $(${PAHOLE} --version) is too old, need at least v1.13"
+		exit 0
+	fi
+
+	info "BTF" ${1}
+	LLVM_OBJCOPY=${OBJCOPY} ${PAHOLE} -J ${1}
+}
 
  # Create ${2} .o file with all symbols from the ${1} object file
  kallsyms()

@@ -281,6 +295,10 @@ fi
  info LD vmlinux
  vmlinux_link "${kallsymso}" vmlinux
 
+if [ -n "${CONFIG_DEBUG_INFO_BTF}" ]; then
+	gen_btf vmlinux
+fi
+
  if [ -n "${CONFIG_BUILDTIME_EXTABLE_SORT}" ]; then
  	info SORTEX vmlinux
  	sortextable vmlinux

Quoted 25 lines · Andrii Nakryiko

This patch adds new config option to trigger generation of BTF type information from DWARF debuginfo for vmlinux and kernel modules through pahole, which in turn relies on libbpf for btf_dedup() algorithm.

The intent is to record compact type information of all types used inside kernel, including all the structs/unions/typedefs/etc. This enables BPF's compile-once-run-everywhere ([0]) approach, in which tracing programs that are inspecting kernel's internal data (e.g., struct task_struct) can be compiled on a system running some kernel version, but would be possible to run on other kernel versions (and configurations) without recompilation, even if the layout of structs changed and/or some of the fields were added, removed, or renamed.

This is only possible if BPF loader can get kernel type info to adjust all the offsets correctly. This patch is a first time in this direction, making sure that BTF type info is part of Linux kernel image in non-loadable ELF section.

BTF deduplication ([1]) algorithm typically provides 100x savings compared to DWARF data, so resulting .BTF section is not big as is typically about 2MB in size.

[0] http://vger.kernel.org/lpc-bpf2018.html#session-2
[1] https://facebookmicrosites.github.io/bpf/blog/2018/11/14/btf-enhancement.html

Quoted 1 line

Cc Masahiro Yamada <[email protected]>

Masahiro,

could you please review this patch?
It's been pending for 10 days!