On 6/1/2017 5:18 PM, Bryan Drewery wrote: > On 5/23/2017 12:00 PM, O. Hartmann wrote: >> Am Tue, 23 May 2017 12:52:35 -0500 >> Pedro Giffuni <p...@freebsd.org> schrieb: >> >>> On 23/05/2017 12:12, O. Hartmann wrote: >>>> Am Tue, 23 May 2017 16:38:10 +0000 (UTC) >>>> "Pedro F. Giffuni" <p...@freebsd.org> schrieb: >>>> >>>>> Author: pfg >>>>> Date: Tue May 23 16:38:10 2017 >>>>> New Revision: 318750 >>>>> URL: https://svnweb.freebsd.org/changeset/base/318750 >>>>> >>>>> Log: >>>>> Bring some rough support for FreeBSD S/390 to the GNU toolchain. >>>>> >>>>> This is no-op and only for reference: the S/390 port seems to be >>>>> elusive >>>>> in the BSDs so it is convenient to keep some trace from past efforts. >>>>> It is likely newer attempts will focus on a newer toolchain using clang >>>>> instead. >>>>> >>>>> Obtained from: Perforce depot/projects/s390 >>>>> >>>>> Added: >>>>> head/contrib/binutils/ld/emulparams/elf64_s390_fbsd.sh (contents, >>>>> props changed) >>>>> head/contrib/binutils/ld/emulparams/elf_s390_fbsd.sh (contents, >>>>> props changed) >>>>> head/contrib/gcc/config/s390/freebsd.h >>>>> - copied, changed from r318546, head/contrib/gcc/config/s390/linux.h >>>>> Modified: >>>>> head/contrib/binutils/bfd/config.bfd >>>>> head/contrib/binutils/ld/configure.tgt >>>>> head/contrib/gcc/config.gcc >>> ... >>>>> Buildworld fails on r318751 with a "Segmentation fault" error as shown >>>>> below: >>>>> >>>>> Building /usr/obj/usr/src/lib/libkiconv/_libinstall >>>>> --- lib/libmd__L --- >>>>> --- skein_block_asm.o --- >>>>> Segmentation fault >>>>> *** [skein_block_asm.o] Error code 139 >>>>> >>>>> make[4]: stopped in /usr/src/lib/libmd >>>>> .ERROR_TARGET='skein_block_asm.o' >>>>> .ERROR_META_FILE='/usr/obj/usr/src/lib/libmd/skein_block_asm.o.meta' >>>>> .MAKE.LEVEL='4' >>>>> MAKEFILE='' >>>>> >>>>> >>>>> Host is running recent CURRENT: FreeBSD 12.0-CURRENT #124 r318748: Tue >>>>> May 23 >>>>> 18:52:59 CEST 2017 amd64 >>> >>> It shouldn't be related to this change: >>> >>> 1) This only affects s390 configuration which is never activated >>> 2) I did run a tinderbox build to make sure nothing was affected (the >>> only thing failing was an unrelated powerpc warning that was fixed). >>> >>> Pedro. >> >> Hello, >> >> the problem could be resolved by deleting the /usr/obj folder and start a >> clean build >> again. >> > > I think this is fallout from ino64 combined with META_MODE. META_MODE > assumes that host tools will be ABI-compatible and generally does not > [force] rebuild them very often. So the act of upgrading to ino64 host > and then doing another build, your ld and various other host tools will > still be running pre-ino64 binaries via COMPAT_FREEBSD11. I think the > bug in this system is that it is possible for some of these tools to get > mixed-ABI objects linked together resulting in differing ideas of what > 'struct stat' is for example. This could be possible since META_MODE > won't force rebuild all objects for a directory/tool but it may rebuild > 1 object for some reason or just relink the resulting binary. > > Here's an example of what I'm talking about with the host ld object files: > >> -rwxr-xr-x 1 root wheel 1857312 Jun 1 12:59 ld.bfd* >> -rwxr-xr-x 1 root wheel 931304 Jun 1 12:59 ld.bfd.debug* >> -rw-r--r-- 1 root wheel 978 Jun 1 12:59 ld.bfd.debug.meta >> -rwxr-xr-x 1 root wheel 2600311 Jun 1 12:59 ld.bfd.full* >> -rw-r--r-- 1 root wheel 3988 Jun 1 12:59 ld.bfd.full.meta >> -rw-r--r-- 1 root wheel 999 Jun 1 12:59 ld.bfd.meta >> -rw-r--r-- 1 root wheel 154400 Apr 18 16:12 ldcref.o >> -rw-r--r-- 1 root wheel 4553 Apr 18 16:12 ldcref.o.meta >> -rw-r--r-- 1 root wheel 137088 May 11 14:14 ldctor.o >> -rw-r--r-- 1 root wheel 4348 May 11 14:14 ldctor.o.meta >> -rw-r--r-- 1 root wheel 205 Oct 11 2016 ldemul-list.h >> -rw-r--r-- 1 root wheel 814 Oct 11 2016 ldemul-list.h.meta >> -rw-r--r-- 1 root wheel 144088 Oct 13 2016 ldemul.o >> -rw-r--r-- 1 root wheel 4374 Oct 13 2016 ldemul.o.meta > > The object files predate ino64 but the linked binaries do not. I did > not dig into these object files more but I suspect somewhere there are > mixed-ABI object files hitting this bug or that just linking pre-ino64 > objects may cause a problem. I don't think linking would be a problem > though.
This commit did cause a mixed-ABI object issue that I predicted. Built before, updated, built again and targets.o is rebuilt. > Building /usr/obj/root/svn/base/gnu/usr.bin/binutils/libbfd/targmatch.h > Skipping meta for .depend: no commands > Skipping meta for afterdepend: .PHONY > Skipping meta for depend: .PHONY > Skipping meta for objwarn: .PHONY > Skipping meta for beforebuild: .PHONY > Skipping meta for .WAIT_1: .PHONY > /usr/obj/root/svn/base/gnu/usr.bin/binutils/libbfd/targets.o.meta: 81: file > '/usr/obj/root/svn/base/gnu/usr.bin/binutils/libbfd/./targmatch.h' is newer > than the target... > Building /usr/obj/root/svn/base/gnu/usr.bin/binutils/libbfd/targets.o targets.o is the only relevant object file I could find in the broken objtree given to me that was post-ino64. It doesn't directly use any of the changes structs that I can see, but the mixed-ABI proof is enough for me to put the understanding of this to rest. In case I wasn't clear, this commit is perfectly fine. It just triggered a META_MODE bug. > > Anyway the fix for this would be to either 'make cleanworld' after > upgrading to ino64, use -DNO_META_IGNORE_HOST for the first build after, > or wait for my fix. I will commit a fix to force rebuild host tools > through known major ABI changes to avoid this problem. > > For discussion of why META_MODE tries to not rebuild host tools see > r301467. The gist is that a simple > 'buildworld->installworld->buildworld' causes everything to rebuild due > to changed host file timestamps. Really it would be better if > filemon/META_MODE used file content hashing like ccache did. Then > timestamps wouldn't cause such a problem here. > -- Regards, Bryan Drewery
signature.asc
Description: OpenPGP digital signature
