GENERATE LINUX KERNEL COMPILATION DATABASE

Prerequisites

A modern Linux host

Though Ubuntu 16.04 used here, any modern Linux Distributions should work. Tested on Arch Linux.

compiledb

python2 or python3 compiledb command

$ pip install compiledb

Linux 2.0.27

1. decompress the source tar ball

   $ mkdir linux-2.0.27 && tar -xvf linux-2.0.27.tar.xz -C linux-2.0.27 --strip-components 1
       

2. make menuconfig

   $ make menuconfig
       

   Notice: the path of the ncurses library on your host system may be not at the position as old linux expected.

   make -C scripts/lxdialog all
   -e
   >> Unable to find the Ncurses libraries.
   >>
   >> You must have Ncurses installed in order
   >> to use 'make menuconfig'
   Makefile:34: recipe for target 'ncurses' failed
       

   To fix it, install ncurses lib via your package manager and then modify scripts/lxdialog/Makefile:29 as follows:

   all: ncurses lxdialog -> all: lxdialog
       

   Let gcc find libncurses automatically. Then rerun make menuconfig.

   

3. config the kernel as you want, or use the linux-x.x.x-config in this repo as a basis
4. tune two Makefiles to support dry-run without error

   a. fs/Makefile:
      11 # L_OBJS    = $(join $(SUB_DIRS),$(SUB_DIRS:%=/%.o))
      12 L_OBJS     =
   
   b. net/Makefile:
      52 # L_OBJS             := socket.o protocols.o sysctl_net.o $(join $(SUB_DIRS),$(SUB_DIRS:%=/%.o))
      53 L_OBJS       := socket.o protocols.o sysctl_net.o
       

5. generate build log

   $ LANGUAGE=en make V=1 -j1 bzImage modules --dry-run |& tee build-log.txt
       

   This does not compile the kernel actually, just dryrun.

6. generate compile_commands.json

   $ compiledb < build-log.txt
       

7. [OPTIONAL] tune compile_commands.json

   Add more general CFLAGS, some index backend engines might need them.

   $ sed -i.bak '/m486/a \ \ \ "-march=i386",\n\ \ \ "-m32",' compile_commands.json
       

Linux 2.0.33

1. decompress the source tar ball

   $ mkdir linux-2.0.33 && tar -xvf linux-2.0.33.tar.xz -C linux-2.0.33 --strip-components 1
       

2. make menuconfig

   $ make menuconfig
       

   Notice: the path of the ncurses library on your host system may be not at the position as old linux expected.

   make -C scripts/lxdialog all
   -e
   >> Unable to find the Ncurses libraries.
   >>
   >> You must have Ncurses installed in order
   >> to use 'make menuconfig'
   Makefile:34: recipe for target 'ncurses' failed
       

   To fix it, install ncurses lib via your system package manager and then modify scripts/lxdialog/Makefile:29 as follows:

   all: ncurses lxdialog -> all: lxdialog
       

   Let gcc find libncurses automatically. Fix a bug in scripts/lxdialog/lxdialog.c:60 before going on.

   const char *title = NULL; -> char *title = NULL;
       

   Then rerun make menuconfig.

3. config the kernel as you want, or use the linux-x.x.x-config in this repo as a basis
4. tune two Makefiles to support dry-run without error

   a. fs/Makefile:
      11 # L_OBJS    = $(join $(SUB_DIRS),$(SUB_DIRS:%=/%.o))
      12 L_OBJS     =
   
   b. net/Makefile:
      52 # L_OBJS             := socket.o protocols.o sysctl_net.o $(join $(SUB_DIRS),$(SUB_DIRS:%=/%.o))
      53 L_OBJS       := socket.o protocols.o sysctl_net.o
       

5. generate build log

   $ LANGUAGE=en make V=1 -j1 bzImage modules --dry-run |& tee build-log.txt
       

   This does not compile the kernel actually, just dryrun.

6. generate compile_commands.json

   $ compiledb < build-log.txt
       

7. [OPTIONAL] tune compile_commands.json

   Add more general CFLAGS, some index backend engines might need them.

   $ sed -i.bak '/m486/a \ \ \ "-march=i386",\n\ \ \ "-m32",' compile_commands.json
       

Linux 2.2.14

1. decompress the source tar ball

   $ mkdir linux-2.2.14 && tar -xvf linux-2.2.14.tar.xz -C linux-2.2.14 --strip-components 1
       

2. make menuconfig

   $ make ARCH=i386 menuconfig
       

   Notice: the path of the ncurses library on your host system may be not at the position as old linux expected.

   make -C scripts/lxdialog all
   -e
   >> Unable to find the Ncurses libraries.
   >>
   >> You must have Ncurses installed in order
   >> to use 'make menuconfig'
   Makefile:34: recipe for target 'ncurses' failed
       

   To fix it, install ncurses lib via your system package manager and then modify scripts/lxdialog/Makefile:28 as follows:

   all: ncurses lxdialog -> all: lxdialog
       

   Let gcc find libncurses automatically. Then rerun make ARCH=i386 menuconfig.

3. config the kernel as you want, or use the linux-x.x.x-config in this repo as a basis
4. generate build log

   $ LANGUAGE=en make V=1 ARCH=i386 -j1 bzImage modules --dry-run |& tee build-log.txt
       

   This does not compile the kernel actually, just dryrun.

5. generate compile_commands.json

   $ compiledb < build-log.txt
       

6. [OPTIONAL] tune compile_commands.json

   Add more general CFLAGS, some index backend engines might need them.

   $ sed -i.bak '/m486/a \ \ \ "-march=i386",\n\ \ \ "-m32",' compile_commands.json
       

Linux 2.4.0

1. decompress the source tar ball

   $ mkdir linux-2.4.0 && tar -xvf linux-2.4.0.tar.xz -C linux-2.4.0 --strip-components 1
       

2. make menuconfig

   $ make ARCH=i386 menuconfig
       

3. config the kernel as you want, or use the linux-x.x.x-config in this repo as a basis
4. generate build log

   $ LANGUAGE=en make V=1 ARCH=i386 -j1 bzImage modules --dry-run |& tee build-log.txt
       

   This does not compile the kernel actually, just dryrun.

5. generate compile_commands.json

   $ compiledb < build-log.txt
       

6. tune compile_commands.json

   Some index backend engines may not work well using this compile_commands.json, since some CFLAGS needed by modern x86_64 compilers are missing in old kernel config. Add -m32 to the compile_commands.json:

   $ sed -i.bak '/march=i686/a \ \ \ "-m32",' compile_commands.json
       

Linux 2.4.18

1. decompress the source tar ball

   $ mkdir linux-2.4.18 && tar -xvf linux-2.4.18.tar.xz -C linux-2.4.18 --strip-components 1
       

2. make menuconfig

   $ make ARCH=i386 menuconfig
       

3. config the kernel as you want, or use the linux-x.x.x-config in this repo as a basis
4. generate build log

   $ LANGUAGE=en make V=1 ARCH=i386 -j1 bzImage modules --dry-run |& tee build-log.txt
       

   This does not compile the kernel actually, just dryrun.

5. generate compile_commands.json

   $ compiledb < build-log.txt
       

6. tune compile_commands.json

   Some index backend engines may not work well using this compile_commands.json, since some CFLAGS needed by modern x86_64 compilers are missing in old kernel config. Add -m32 to the compile_commands.json:

   $ sed -i.bak '/march=i686/a \ \ \ "-m32",' compile_commands.json
       

Linux 2.6.11

1. decompress the source tar ball

   $ mkdir linux-2.6.11 && tar -xvf linux-2.6.11.tar.xz -C linux-2.6.11 --strip-components 1
       

2. make menuconfig

   $ make ARCH=i386 menuconfig
       

3. config the kernel as you want, or use the linux-x.x.x-config in this repo as a basis
4. fix a bug in Makefile

   drivers/media/dvb/b2c2/Makefile:4: *** missing separator.  Stop.
   scripts/Makefile.build:311: recipe for target 'drivers/media/dvb/b2c2' failed
   
     4 # obj-$(CONFIG_DVB_B2C2_USB) + = b2c2-usb.o
     5 obj-$(CONFIG_DVB_B2C2_USB) += b2c2-usb.o
       

5. generate build log

   $ LANGUAGE=en make V=1 ARCH=i386 -j1 --dry-run |& tee build-log.txt
       

   This does not compile the kernel actually, just dryrun. The fail of the final linkage of vmlinux does not matter, since the total compilation has finished.

6. generate compile_commands.json

   $ compiledb < build-log.txt
       

Linux 2.6.24

1. decompress the source tar ball

   $ mkdir linux-2.6.24 && tar -xvf linux-2.6.24.tar.xz -C linux-2.6.24 --strip-components 1
       

2. make menuconfig

   choose ARCH, i386 or x86_64

   $ make ARCH=i386 menuconfig
       

   or just use a common default config, and skip step 3.

   $ make ARCH=i386 defconfig
       

   Note: The Makefile in src root dir has syntax error using modern make, fix that first.

   434 config %config: scripts_basic outputmakefile FORCE
   435         $(Q)mkdir -p include/linux include/config
   436         $(Q)$(MAKE) $(build)=scripts/kconfig $@
   
   -->
   
   config: scripts_basic outputmakefile FORCE
           $(Q)mkdir -p include/linux include/config
           $(Q)$(MAKE) $(build)=scripts/kconfig $@
   %config: scripts_basic outputmakefile FORCE
           $(Q)mkdir -p include/linux include/config
           $(Q)$(MAKE) $(build)=scripts/kconfig $@
   
   1506 / %/: prepare scripts FORCE
   1507         $(cmd_crmodverdir)
   1508         $(Q)$(MAKE) KBUILD_MODULES=$(if $(CONFIG_MODULES),1) \
   1509         $(build)=$(build-dir)
   
   -->
   
   /: prepare scripts FORCE
           $(cmd_crmodverdir)
           $(Q)$(MAKE) KBUILD_MODULES=$(if $(CONFIG_MODULES),1) \
           $(build)=$(build-dir)
   %/: prepare scripts FORCE
           $(cmd_crmodverdir)
           $(Q)$(MAKE) KBUILD_MODULES=$(if $(CONFIG_MODULES),1) \
           $(build)=$(build-dir)
       

3. config the kernel as you want, or use the linux-x.x.x-config in this repo as a basis
4. generate build log

   $ LANGUAGE=en make V=1 ARCH=i386 -j1 --dry-run |& tee build-log.txt
       

   This does not compile the kernel actually, just dryrun. The fail of the final linkage of vmlinux does not matter, since the total compilation has finished.

5. generate compile_commands.json

   $ compiledb < build-log.txt
       

Linux 2.6.34

1. decompress the source tar ball

   $ mkdir linux-2.6.34 && tar -xvf linux-2.6.34.tar.xz -C linux-2.6.34 --strip-components 1
       

2. make menuconfig

   choose ARCH, i386 or x86_64

   $ make ARCH=i386 menuconfig
       

   This config will base on your host’s /boot/config of the host kernel.

   or just use a common default config

   $ make ARCH=i386 defconfig
       

3. config the kernel as you want, or use the linux-x.x.x-config in this repo as a basis

   You may need to make menuconfig again after make defconfig to disable the “Device Drivers -> Graphics support -> Bootup logo”, which causes the dryrun fail prematurely. If you want a real compilation of the kernel source, just skip this step after make defconfig.

4. [OPTIONAL] prepare a real compilation of the kernel if you want

   a. install gcc-4.x multilib to support the compiling, here I use 4.6, 4.9 should be ok, not tested

   b. modify a Makefile to support gcc 4.x to compile

   arch/x86/vdso/Makefile
    28 # VDSO_LDFLAGS_vdso.lds = -m elf_x86_64 -Wl,-soname=linux-vdso.so.1 \
    29 VDSO_LDFLAGS_vdso.lds = -m64 -Wl,-soname=linux-vdso.so.1 \
   
    72 # VDSO_LDFLAGS_vdso32.lds = -m elf_i386 -Wl,-soname=linux-gate.so.1
    73 VDSO_LDFLAGS_vdso32.lds = -m32 -Wl,-soname=linux-gate.so.1
       

   c. perl scritps may need to be modified to support more recent perl interpreter, e.g.

   kernel/timeconst.pl
    373         # if (!defined(@val)) {
    374         if (!@val) {
       

5. generate build log

   note: [a] and [b] are exclusive, use one of them according to real compile or dryrun.

   [a]. do a real compilation of the kernel and get the build log, this requires step[4]

   $ LANGUAGE=en make V=1 CC=gcc-4.6 ARCH=i386 -j4 |& tee build-log.txt
       

   ARCH x86_64 should be the same as i386, the gcc-4.x multilib version should be used if both i386 and x86_64 need to be supported.

   [b]. get the build log using make dryrun, if not real compilation

   Before we can dryrun, “Device Drivers -> Graphics support -> Bootup logo” should be disabled.

6. generate compile_commands.json

   $ compiledb < build-log.txt
       

Linux 3.x - latest

1. The 3.x and 4.x versions should be the same as 2.6.34. If dryrun fails, fix the problems or JUST DO A REAL [CROSS] COMPILATION on your host. The difference may be that the gcc version used is varied.
2. Since kernel v5, scripts/gen_compile_commands.py can be used to generate the compile_commands.json natively. Just compile the kernel, and run the script. e.g.

   $ make ARCH=x86_64 defconfig
   $ make -j8
   $ scripts/gen_compile_commands.py
       

Bonus: Linux 0.12

1. decompress the source tar ball

   $ tar -xvf linux-0.12.tar.gz
       

2. generate build log

   note: [a] and [b] are exclusive, use one of them.

   [a]. use linux-0.12-gen_build_log.sh to generate the build log

   $ cp /PATH/TO/linux-0.12-gen_build_log.sh linux-0.12/
   $ cd linux-0.12 && bash ./linux-0.12-gen_build_log.sh
       

   [b]. or use linux-0.12-gen_build_log.mk to generate the build log

   $ cp /PATH/TO/linux-0.12-gen_build_log.mk linux-0.12/
   $ cd linux-0.12 && make -f linux-0.12-gen_build_log.mk |& tee build-log.txt
       

3. generate compile_commands.json

   $ compiledb < build-log.txt
       

4. [OPTIONAL] add -m32

   $ sed -i.bak '/nostdinc/a \ \ \ "-m32",' compile_commands.json
       

Dry-Run vs Real Compilation

The compile database of a real compilation get all the files involved in the compilation. The compile database of dry-run might miss some seperate targets despite all the kernel vmlinux compilation commands that successfully generated. The missing targets are mostly in arch/$ARCH/boot/, and some helping tools and scripts. Files in arch/$ARCH/boot of ancient kernel source are mostly ASM files, which are not able to be indexed by clang based C/C++ indexers. Routines or symbols in the .S asm files can be easily found via grep tools like ripgrep . The arch/$ARCH/boot of relatively new kernels can be indexed via a real compilation.

So we consider the compile_commands.json from dry-run a good enough compilation database when indexing ancient kernel source.