Re: Handling s390 libc ABI change in Debian
On Tue, Jul 15, 2014 at 1:18 AM, Aurelien Jarno <aurel32@debian.org> wrote:
> On Mon, Jul 14, 2014 at 11:14:42PM -0400, Carlos O'Donell wrote:
>> On Mon, Jul 14, 2014 at 4:36 PM, Aurelien Jarno <aurel32@debian.org> wrote:
>> > glibc 2.19 has changed the libc ABI on s390, more specifically the
>> > setjmp/longjmp functions [1] [2]. Symbol versioning is used to handle
>> > some cases, but it doesn't work when a jmp_buf variable is embedded
>> > into a structure, as it changes the size of the structure. The result
>> > is that mixing programs or libraries built with 2.18 with ones built
>> > with 2.19 do not work anymore, usually they end up with a segmentation
>> > fault. Some persons from this list have experienced that with perl.
>>
>> That is not true. This is an over generalization of the problem. You
>> can use libraries built with 2.18 and 2.19 and they work just fine.
>
> I agree I probably a bit over exaggerated here, but the problem is real,
> breakages do happen, and some persons on this mailing list have already
> experienced them.
>
>> The extent of the problem in correct language is listed here:
>> https://sourceware.org/glibc/wiki/Release/2.19#Packaging_Changes
>
> This seems to minimize the problem, listing only perl. In practice we
> have seen much more breakages, part of them being due to the change of
> the __pthread_unwind_buf_t struct.
That is a change that nobody reported. You're the first to mention it
and that does make it more serious. We have discussed this upstream
and I agree that we need more versioning of the interfaces there to
support the change fully.
>> > We first thought it was limited to a few packages (even if all perl is
>> > already more than that), but as time goes more and more issues are
>> > found. libpng and gauche are also affected, the issue with mono is
>> > also likely due to this ABI change.
>>
>> That is new information, and it is important for distributions to
>> relay this information back upstream where the decision for a SO bump
>> can be made.
>
> I can follow up with a list affected packages, but we are slowly
> discovering them one by one, so it might takes time. So far we have:
>
> * Mixing modules/libraries built with pre-2.19 and 2.19 libc
> - perl
> - libpng
You can never support a mixed-ABI environment with versioning.
You must update all of those packages at once.
The best we could do is warn the user of the incompatibility at
runtime and refuse to load the module via dlopen, or refuse to start
the application at startup.
> * Using libc 2.19 without rebuilding anything:
> - gauche
> - mono
This we believe to be pthread issues.
>> > According to upstream [3], the problem is that Debian doesn't do a mass
>> > rebuild, which is the strategy chosen by Red Hat to handle^Wworkaround
>> > this issue. This means some programs might segfault during the upgrade,
>> > or on partially upgraded systems.
>>
>> I apologize if you took what I wrote to mean that. I did not mean it
>> was Debian's problem, but rather that Debian suffered the most because
>> they don't do rebuilds. The two are orthogonal. You face a situation
>> that is unique to the framework used to build the distribution.
>>
>> Please engage upstream to champion a SO name bump for libc for
>
> I think that would be the correct solution. That said as it is not
> something trivial and thus not done often, it's an opportunity to push
> for more ABI changes if some others are envisaged in the future.
The problems are worse.
I just tried to simulate this on x86-64 and there are serious problems.
In most libraries you can load multiple different copies and it won't conflict.
Here libc.so.6 and libc.so.7 or libc.so.6.1 all conflict in the same
namespace and worse control aspects of the implementation like TLS. It
doesn't work to bump the SONAME.
We would have to implement a coordination framework amongst all the
SONAME bumped libc's for all of the basic functionality that had to
keep working. That would force future libcs to stay compatible
internally with other libcs and that would be very difficult to
maintain.
I am starting to think that a tooling option to fail to load mixed-ABI
objects is the only option, with user rebuilds happening after that.
>> > Now we have to chose a strategy for Debian. I see multiple options:
>> >
>> > 1) Ignore the issue and just rebuild (binNMU) the packages that seems
>> > affected when we discover them. This means partial upgrades will likely
>> > be broken, and that we might discover some broken packages only after
>> > the jessie release.
>> >
>> > 2) Rebuild (binNMU) all packages. This means partial upgrades will
>> > likely be broken.
>> >
>> > 3) Bump the soname of affected packages and rebuild their reverse
>> > dependencies. It is the solution that is currently being implemented for
>> > perl. It clearly won't scale if more broken packages (and even for
>> > libpng) are discovered as it requires a source upload and a transition
>> > handled by the release team. It also means breaking the ABI compatibility
>> > with other distributions.
>> >
>> > 4) Bump the libc soname to libc.so.6.1 and do a libc transition. This is
>> > probably what upstream should have done instead of breaking the ABI.
>> > This is a huge work though, and this also means breaking the ABI
>> > compatibility with other distributions.
>> >
>> > 5) Revert the ABI change. This is likely just postponing the problem as
>> > the change is required to support future hardware. This also means
>> > breaking the ABI compatibility with other distributions.
>> >
>> > 6) simply drop the s390x port and tell users to either use an other
>> > distribution or use Debian on other hardware.
>> >
>> > Any opinion? Any other ideas how to handle that?
>>
>> Option (6) is the nuclear option, and clearly a little excessive given
>> the situation. If user's install from an installer they get a
>> perfectly working system. Punishing those users because partial
>> upgrades don't work seems excessive.
>
> If we are not able to solve this problem by ensuring all the packages
> in the next release are using a consistent ABI, I think it is an option
> to consider. It's probably better being honest with the users telling
> them "we fail" than shipping a half broken release. Consider that we
> have limited human resources to maintain this port.
I disagree, but I'm not assisting Debian in any way and you have your
own choices to make.
>> Option (5) postpones the problem until newer s390 hardware arrives.
>>
>> Option (4) is likely what upstream should have done. Why do you think
>> it's a huge amount of work? If all s390-supporting distributions agree
>> then we just change the SO name?
>
> It's a huge work for Debian, maybe not for other distribution, as it
> basically means we have to rebootstrap everything. This includes manual
> bootstrapping of self-dependent languages (haskell, gnat, ...) and
> manual handling of some dependencies loop. In addition it's something
> which hasn't been done since the libc5 transition, so we might discover
> some unexpected issues.
Why do you have to do that? Is it just like for rpm where the
packaging system encodes the SONAME as a dependency? We would also
need a manual bootstrap in Fedora because of this issue.
>> Option (3), (2) or (1) all look like reasonable "fire and forget"
>> options that require users of s390 to simply move forward with their
>> installs to full new versions of Debian. It's an option for RHEL
>> because we don't allow partial-upgrades and we are working with users
>> to notify them of the breakage during a major RHEL upgrade e.g. RHEL7
>> to RHEL8.
>
> Does it mean on RHEL, a full reinstall would be needed? A system being
> upgraded from one version to another will transition though partial
> ugprades, and we have to ensure that at least the upgrade won't break.
Yes. For RHEL you always reinstall a new version.
For Fedora you use FedUp (a special initial root disk) to carry out
the upgrade safely (after FedUp downloads it the data it needs
locally).
AFAIK there is no partial upgrade case for Fedora and RHEL.
We still have the problem that user built modules will need to be recompiled.
>> In my opinion it's not too late to do option (4).
>
> Agreed.
Unfortunately the more I test this the more impossible this option
seems. The SONAME bump allows two libcs to be loaded at once and it's
not working.
I think the best we can do is diagnose the ABI mixing and refuse to
load or abort a starting program.
Cheers,
Carlos.
Reply to: