Hi again,

On Sat, Oct 21, 2023 at 01:15:10PM +0200, Julien Olivain wrote:
> Hi Marcus,
> 
> Thanks for the updated patch!
> 
> I confirm the normal test-pkg is passing. Then, I ran a "test-pkg -a"
> (which test more toolchains/arches) and observed few new minor
> failures. I'll comment a bit each failures and provide hints for
> improvements, when possible.
> 
>     cat >criu.config <<EOF
>     BR2_PACKAGE_HOST_PYTHON3=y
>     BR2_PACKAGE_CRIU=y
>     EOF
>     utils/test-pkg -a -c criu.config -p criu
>                      arm-aarch64 [ 1/45]: OK
>            bootlin-aarch64-glibc [ 2/45]: FAILED
> 
> criu uses restartable sequences, see:
> https://github.com/checkpoint-restore/criu/blob/v3.18/criu/include/linux/rseq.h#L7
> 
> Criu seems to have a header detection on "sys/rseq.h" (i.e. on the
> libc side). But the sys/rseq.h then includes linux/rseq.h (i.e. on the
> kernel side), which is not checked. Some configurations like
> bootlin-aarch64-glibc includes the glibc header, but also have older
> kernel. Restartable sequences were introduced in kernel >= 4.18. So
> this can be handled by adding a dependency on kernel.
> 
>        bootlin-arcle-hs38-uclibc [ 3/45]: SKIPPED
>             bootlin-armv5-uclibc [ 4/45]: SKIPPED
>              bootlin-armv7-glibc [ 5/45]: OK
>            bootlin-armv7m-uclibc [ 6/45]: SKIPPED
>               bootlin-armv7-musl [ 7/45]: OK
>         bootlin-m68k-5208-uclibc [ 8/45]: SKIPPED
>        bootlin-m68k-68040-uclibc [ 9/45]: SKIPPED
>      bootlin-microblazeel-uclibc [10/45]: SKIPPED
>         bootlin-mipsel32r6-glibc [11/45]: SKIPPED
>            bootlin-mipsel-uclibc [12/45]: SKIPPED
>              bootlin-nios2-glibc [13/45]: SKIPPED
>          bootlin-openrisc-uclibc [14/45]: SKIPPED
> bootlin-powerpc64le-power8-glibc [15/45]: SKIPPED
>    bootlin-powerpc-e500mc-uclibc [16/45]: SKIPPED
>            bootlin-riscv32-glibc [17/45]: SKIPPED
>            bootlin-riscv64-glibc [18/45]: SKIPPED
>             bootlin-riscv64-musl [19/45]: SKIPPED
>          bootlin-s390x-z13-glibc [20/45]: FAILED
> 
> Build fails with output:
> 
>     Makefile:23: *** "The architecture s390x isn't supported".  Stop.

Another problem that occured...
After I use $(BR2_NORMALIZED_ARCH) instead of $(BR2_ARCH), the arch
is at least supported. But it fails with

Error (compel/src/lib/handle-elf-host.c:161): Unsupported header detected
make[3]: *** [criu/pie/Makefile:58: criu/pie/parasite-blob.h] Error 255

After some more investigation, it ended up here:

static bool is_header_supported(Elf_Ehdr *hdr)
{
	if (!arch_is_machine_supported(hdr->e_machine))
		return false;
	if ((hdr->e_type != ET_REL
#ifdef NO_RELOCS
	     && hdr->e_type != ET_EXEC
#endif
	     ) ||
	    hdr->e_version != EV_CURRENT)
		return false;
	return true;
}

I've added debug trace to see why it fails and got this:
Error (compel/src/lib/handle-elf-host.c:155): 	type 0x100 machine 0x1600 version 0x1000000
Error (compel/src/lib/handle-elf-host.c:158): 	ET_REL 0x1 EM_s390 0x16, EV_CURRENT 0x1

It is a tool that runs on host (little endian), but read data of an elf
compiled for s390 (big endian), so bytes is read swapped.

Right now I thinking of not list s390 as a compatible architecture for
this package.

Do you think it is reasonable?


/Marcus