I've got caught in the not-so-sunny world of cross-compilation.
I'm trying to compile a simple hello world application for my BeagleBone Black (which runs a TI Cortex-A8 processor).
First of all, I compiled and ran successfully the hello world application on x86 with gcc
Then I changed my compilation settings to the following:
arm-linux-gnueabi-gcc -c -O0 -g3 -Wall main.c -o bin/obj/main.o
arm-linux-gnueabi-gcc bin/obj/main.o -o bin/hello_world
I transferred the file via SCP to the BeagleBone, and set executable permissions with chmod +x hello_world
Upon running it (./hello_world
), my only response is:
-bash: ./hello_world: No such file or directory
The output of file
matches that of /sbin/init
as I would expect:
$ file hello_world
hello_world: ELF 32-bit LSB executable, ARM, version 1 (SYSV), dynamically linked (uses shared libs), for GNU/Linux 2.6.32, BuildID[sha1]=0x24b659b7a41fe043a6f4649d4ebfb5e692ebf0c7, not stripped
$ file /sbin/init
/sbin/init: ELF 32-bit LSB executable, ARM, version 1 (SYSV), dynamically linked (uses shared libs), for GNU/Linux 2.6.26, BuildID[sha1]=0xd21f6957ec031a27d567b3d5e6aa14b9e0c30c37, stripped
The result of ldd
is:
$ ldd hello_world
not a dynamic executable
I tried adding a suitable platform and CPU type, changing my compilation to:
arm-linux-gnueabi-gcc -c -O0 -g3 -Wall -march=armv7-a -mtune=cortex-a8 main.c -o bin/obj/main.o
arm-linux-gnueabi-gcc bin/obj/main.o -o bin/hello_world
This initially started giving me a new error: Text file busy
, but I have since been unable to get that error back again as it now returns No such file or directory
. I'm guessing that particular attempt was just a bad transfer or something.
Since nobody from the comments posted the answer, I guess I get the pleasure ;)
No such file or directory
comes from when the kernel tries to invoke the dynamic linker specified by the ELF executable's .interp
field, but no such file exists.
The .interp
field can be found with the following command:
objdump -j .interp -s ./hello_world
In the case of this example, the executable's .interp
field was /lib/ld-linux.so.3
, but the name of the dynamic linker on the BeagleBone Black is /lib/ld-linux-armhf.so.3
.
This happened because the program was compiled with a slightly different toolchain to the one required for the platform. It should be arm-linux-gnueabihf-*
rather than arm-linux-gnueabi-*
.
The difference between the two is that the Cortex-A8 uses specific floating point registers with the hard-float version (armhf
) of the EABI, but the original EABI (armel
) uses integer registers for passing around floating point numbers. As a result, armel
programs will run on armhf
(provided the dynamic linker is set to the correct path!), but not vice versa.
Simply adding a symbolic link ln -s /lib/ld-linux-armhf.so.3 /lib/ld-linux.so.3
is enough to resolve this issue, but the correct fix is to use the right toolchain when compiling the program in the first place.