Hi folks,
As promised, here's minutes from the call we had this afternoon. Spoiler: the result we've agreed is
/lib/ld-linux-armhf.so.3
And here's a transcription of the minutes from
https://wiki.linaro.org/OfficeofCTO/HardFloat/LinkerPathCallApr2012
========================================
Meeting: 13th April 2012, 15:00 UTC
Agenda ------
* Debian/Ubuntu have so far built using /lib/arm-linux-gnueabihf/ld-linux.so.3 * Some other distros (Fedora, OpenSUSE) are still using /lib/ld-linux.so.3 option which matches the older soft-float ABI * Some people are proposing /libhf/ld-linux.so.3 or /libhfp/ld-linux.so.3 (multilib) * Some people proposed /lib/ld-arm-linux-gnueabihf.so.3 (similar to x86_64, libs still in /lib, from Michael Hope) * What should we do as a community?
Present -------
Name Affiliations
Steve McIntyre ARM, Debian, Linaro Wookey ARM, Debian, Linaro Richard Earnshaw ARM, gcc Jeff Law Fedora, Red Hat, gcc, glibc Jon Masters Fedora, Red Hat Andrew Haley Fedora, Red Hat, gcc Andreas Jaeger SUSE, openSUSE, glibc Carlos O'Donnell Mentor, gcc Steve Langasek Canonical, Ubuntu, Debian Dann Frazier Canonical, Ubuntu, Debian Adam Conrad Canonical, Ubuntu, Debian Matthias Klose Canonical, Ubuntu, Debian Mike Frysinger Gentoo Dennis Gilmore Fedora, Red Hat
Discussion ----------
We started with a couple of questions up front to establish the grounds for discussion:
* We believed that decision makers were present for all the important parties, i.e. all the arm hard-float distros, plus toolchain developers. This meant that a decision taken at the meeting could be implemented without needing further arguments/negotiations.
* All the people present understood the importance of cross-distro binary compatibility, and they all wanted it. This led to agreement that we needed to agree on a standard path for the runtime linker for ARM hard-float Linux binaries.
Debian and Ubuntu had so far been using the "multi-arch" path of /lib/arm-linux-gnueabihf/ld-linux.so.3. Fedora and OpenSUSE were thus far using /lib/ld-linux.so.3, the same as the soft-float ABI. Others had proposed alternative paths such as /libhf/ld-linux.so.3 or /libhfp/ld-linux.so.3 (multilib) or /lib/ld-arm-linux-gnueabihf.so.3. Discussion showed that none of these were found to be universally acceptable.
Two parties were likely to be soon affected by an agreement here:
1. Ubuntu 12.04 (releasing with armhf in ~2 weeks)
Adam/Steve L agreed that all efforts would be put in to switch the compilers in Ubuntu to a new path before release. Default things like gcc would be correct, but less common tools might still be targetting the old path /lib/arm-linux-gnueabihf/ld-linux.so. at release. They could be fixed in the longer term and would not stop progress here.
2. Mentor (Codebench due for release in ~1 week)
Carlos mentioned this - Codebench has been using /lib/ld-linux.so.3 for hard-float binaries for some time and it was too late to change that for this upcoming release. Next release due in October. Suggested and accepted that this should be mentioned in release notes: if people want to use Codebench on some systems (Debian/Ubuntu and derivatives), they'll need to tweak their system setup. He may be able to do the linker change and rebuild in a point release in a few weeks.
It was briefly suggested that the soft-float linker should be renamed away from /lib/ld-linux.so.3 as well at this time, but that idea was quickly shot down.
Proposal #1: /lib/ld-armhf.so.3 (not generally liked)
Proposal #2: /lib/ld-linux-armhf.so.3 (not favourite, but considered an acceptable compromise by all)
No need to go any further.
Conclusion ----------
All the people in the meeting agreed: the new runtime linker path for ARM hard-float Linux binaries was to be
/lib/ld-linux-armhf.so.3
ACTION: People at the meeting to present this decision to their companies / communities and get the appropriate changes made.
Further discussion ------------------
General unhappiness with the mess that led to this meeting. Future planning needs to be better between the various groups for the next time we have a new CPU/ABI/whatever.
ACTION: Jon Masters to talk to the Linux Foundation about setting up a forum for such discussions.
In the meantime, strong consensus to use the cross-distro@lists.linaro.org mailing list for any more conversations now we have people in contact.
ACTION: Steve McIntyre to write up the minutes and circulate. Include an updated linker path patch for gcc to match the decision made here.
More discussion about triplets and naming, but nothing came of it in the end. Distro folks have already decided what they're using and have patched various software to build appropriately. Richard wants to move gcc's config.guess to use arm-linux-gnueabihf; no strong objections to that.
Linker path patch for gcc -------------------------
Adapted from earlier work by Dann Frazier dann.frazier@canonical.com and Michael Hope michael.hope@linaro.org
2012-04-13 Steve McIntyre steve.mcintyre@linaro.org
* config/arm/linux-eabi.h (GLIBC_DYNAMIC_LINKER_HARD_FLOAT): Define. (GLIBC_DYNAMIC_LINKER): Redefine to use the hard float path.
diff --git a/gcc/config/arm/linux-eabi.h b/gcc/config/arm/linux-eabi.h index 80bd825..8c9d2e7 100644 --- a/gcc/config/arm/linux-eabi.h +++ b/gcc/config/arm/linux-eabi.h @@ -62,7 +62,11 @@ /* Use ld-linux.so.3 so that it will be possible to run "classic" GNU/Linux binaries on an EABI system. */ #undef GLIBC_DYNAMIC_LINKER -#define GLIBC_DYNAMIC_LINKER "/lib/ld-linux.so.3" +#define GLIBC_DYNAMIC_LINKER_SOFT_FLOAT "/lib/ld-linux.so.3" +#define GLIBC_DYNAMIC_LINKER_HARD_FLOAT "/lib/ld-linux-armhf.so.3" +#define GLIBC_DYNAMIC_LINKER \ + "%{mfloat-abi=hard:" GLIBC_DYNAMIC_LINKER_HARD_FLOAT "} \ + %{!mfloat-abi=hard:" GLIBC_DYNAMIC_LINKER_SOFT_FLOAT "}"
/* At this point, bpabi.h will have clobbered LINK_SPEC. We want to use the GNU/Linux version, not the generic BPABI version. */
Post-meeting on IRC -------------------
Suggested that Richard should push the change into gcc trunk ASAP. Steve McIntyre agreed to work on that with Richard.
Also suggested that we want to get a patch into glibc too to change the installation path for ARM hard-float. Andrew agreed to push glibc upstream for that.
Cheers,
On Fri, Apr 13, 2012 at 06:37:11PM +0100, Steve McIntyre wrote:
Hi folks,
As promised, here's minutes from the call we had this afternoon. Spoiler: the result we've agreed is
/lib/ld-linux-armhf.so.3
And here's a transcription of the minutes from
https://wiki.linaro.org/OfficeofCTO/HardFloat/LinkerPathCallApr2012
And there's a slight wrinkle we've found (of course). The change of the filename equates to a change of soname for the linker too. Adam found this a couple of days back when doing rebuilds of the glibc and toolchain packages in Ubuntu, and between us we've come up with a (hacky, ugly, *never* to be upstreamed) solution. It seems to work fine for us, to at least allow for supporting old binaries until packages are all rebuilt to use the new linker path/soname. Just talking to Andreas, the openSUSE folks are running with this too.
/me waits for the barrage of rotten fruit...
============================
For backward compatibility with armhf binaries built with the old linker SONAME, we need to fake out the linker to believe the new is the old, until such a point as everything is rebuilt
--- eglibc-2.15.orig/elf/dl-load.c 2012-04-14 12:11:37.000000000 +0000 +++ eglibc-2.15/elf/dl-load.c 2012-04-15 00:39:36.558765502 +0000 @@ -2303,10 +2303,13 @@ soname = ((const char *) D_PTR (l, l_info[DT_STRTAB]) + l->l_info[DT_SONAME]->d_un.d_val); if (strcmp (name, soname) != 0) +#ifdef __arm__ + if (strcmp(name, "ld-linux.so.3") || strcmp(soname, "ld-linux-armhf.so.3")) +#endif continue;
/* We have a match on a new name -- cache it. */ - add_name_to_object (l, soname); + add_name_to_object (l, name); l->l_soname_added = 1; }
Cheers,
-
Steve McIntyre