It is causing problems when network is in certain, valid, configuration.
The point of it is to workaround a glibc bug that is not present for us.
https://bugs.ruby-lang.org/issues/20208
To reproduce, have IPv6 enabled localhost but only IPv4
interfaces (such as eth0, name and count of them may vary based on system)
to disable public interfaces via sysctl:
`$ sysctl "net.ipv6.conf.<interface>.disable_ipv6=1"`
Where <interface> is NOT lo but can be eth0 or other.
Confirm with command `$ ip addr` command that only `lo`
interface has IPv6 address.
and run the following with Ruby 3.3:
~~~
$ ruby -rnet/http -e 'http = Net::HTTP.new("localhost", 8080); http.local_host = Addrinfo.tcp("localhost", 8080).ip_address; p http.get("/")'
~~~
The error is then related to EAFNOSUPPORT:
```
Errno::EAFNOSUPPORT: Failed to open TCP connection to localhost:37337 (Address family not supported by protocol - bind(2) for "::1" port 45395)
```
running above Ruby command with `strace` suggest that
there is an attempt to bind with AF_INET6
using an AF_INET socket, however, only if the reproducer above
attempts to set the `http.local_host` with Addrinfo.
From upstream issue #20208:
"This is calling getaddrinfo to resolve "localhost" for us to use
it as the local side of the connection. Because Ruby does not know
what you intend to do with this IP address, it does not make the request
with AI_ADDRCONFIG. Thus, you get an IPv6 result returned,
since there is an IPv6 addres for localhost!"
Armv8.3+ capable CPUs might segfault with incorrect compilation options.
See related upstream report: https://bugs.ruby-lang.org/issues/20085
We have hit this on COPR which uses [0] c7g.xlarge AWS flavor for aarch64
architecture builds. This machine flavor seems to fall into the ARMv8.3+
range.
Fedora CFLAGS come with `-mbranch-protection=standard` which means that
both BTI and the PAC protections are used. The option is equivalent
to `-mbranch-protection=pac-ret+bti`.
However, since the upstream configure.ac automatically appends
`-mbranch-protection=pac-ret`, the BTI protection is not used
as the last used option seems to overwrite [1] the Fedora
default for this.
To resolve both of these issues, of BTI being skipped and the potential
segfaults a patch is applied.
To fix segfaults an upstream patch was applied [2].
To fix the issue of overridden option I have patched the
configure.ac file to check for the `=standard` first when searching for a usable
`-mbranch-protection` option.
The overriding of our options was reported upstream:
<https://bugs.ruby-lang.org/issues/20154>
In the same issue I provided the extension of compilation option as an
attachment, to showcase a workaround that fixes the situation in Fedora.
[0] <https://lists.fedoraproject.org/archives/list/devel@lists.fedoraproject.org/message/Y4GPCGQAZT2LJ5CE7MTIEFKGAPP6O2DW/>
[1] <https://lists.fedoraproject.org/archives/list/devel@lists.fedoraproject.org/message/QWLEBS4YQH73HJNNLKCCGIIOU3SSXAYK/>
[2] <02973b78f4>
Tests fail with a LANG=C environment. A new investigation into the
failing tests on x86 revealed exact commits which introduced the issue
and the commits that fixed the situation. Fix is not yet present in Ruby
3.2, therefore the conditional disabling the tests was left in.
A new re-reported ticket replaced the old one. The new upstream ticket
includes better detailed description of the environment causing the
issue.
The tests pass in Rawhide mock (F40), as the LANG there is C.UTF-8
that does not suffer from the issue.
Upstream ticket: https://bugs.ruby-lang.org/issues/20045
libffi since NVR libffi-3.4.4-3.fc39 builds with static trampolines
enabled. This fixes the failures which were the reason for disabling
the tests.
Since it recently landed in Rawhide, we can re-enable the tests.
The rhbz#2043092 is not resolved. Nevertheless, the generation of
package notes should not break builds of rubygem- packages anymore.
Nevertheless, the additional linker options will keep making issues for
`gem install`ed gems. Nothing new here unfortunately.
Fiddle::Closure object is making use of FFI closure
from libffi. When such object is created (instantiated) in Ruby,
and then the process forks on an SELinux-enabled system, the memory
will become corrupted. That is usually not a problem until the
The garbage collector sweeps the object and tries to free it, in which case the
Ruby process will fail with signal SIGABRT.
Tests in test/fiddle/test_closure.rb, test/fiddle/test_func.rb,
and test/fiddle/test_function.rb use the `Fiddle::Closure` class
directly and fiddle/test_import.rb use the class indirectly through
`bind_function` method, therefore they are disabled to prevent
introducing the problematic object into the Ruby GC during test suite
execution instead of relying on that fork and subsequent
garbage collection will not happen.
If an FFI closure object is allocated in Ruby
and the `fork` function is used afterward, the memory
pointing to the closure gets corrupted, and if Ruby GC
tries to collect the object in that state, a SIGABRT
error occurs.
The minimal Ruby reproducer for the issue is the following:
~~~
$ cat fiddle_fork.rb
require 'fiddle/closure'
require 'fiddle/struct'
Fiddle::Closure.new(Fiddle::TYPE_VOID, [])
fork { }
GC.start
~~~
We allocate an unused Closure object,
so it is free for the GC to pick up. Before we call `GC.start`
we fork the process as that corrupts the memory.
Running this with ruby-3.1.2-167.fc37.x86_64 on SELinux enabled system:
~~~
$ ruby fiddle_fork.rb
Aborted (core dumped)
~~~
Such issues may appear at random (depending on the use of forking and GC)
in larger applications that use Fiddle::Closure but can be spotted by the
following functions appearing in the coredump backtrace:
~~~
0x00007f6284d3e5b3 in dlfree (mem=<optimized out>) at ../src/dlmalloc.c:4350
0x00007f6284d6d0b1 in dealloc () from /usr/lib64/ruby/fiddle.so
0x00007f6295e432ec in finalize_list () from /lib64/libruby.so.3.1
0x00007f6295e43420 in finalize_deferred.lto_priv () from /lib64/libruby.so.3.1
0x00007f6295e4ff1c in gc_start_internal.lto_priv () from /lib64/libruby.so.3.1
~~~
Possible solutions to prevent Ruby from crashing:
* Do not use Fiddle::Closure.
* Use the Fiddle::Closure object only in isolated subprocess that
will not fork further.
* Enable static trampolines in libffi as noted in bugzilla comment:
<https://bugzilla.redhat.com/show_bug.cgi?id=2040380#c9>
See related discussion on
<https://bugzilla.redhat.com/show_bug.cgi?id=2040380>
Ruby upstream ticket:
<https://bugs.ruby-lang.org/issues/18914>
Ruby Fiddle ticket:
<https://github.com/ruby/fiddle/issues/102>
RPM 4.18.0-beta1 or later versions remove the build directory
automatically, the build fails on removing temporary directories with
the missing 'w' bit.
RubyGems contain this patch since version 3.3.8.
This was merged to ruby_3_1 branch in upstream ruby, but it is not
included in the v3_1_2 tag, so new Ruby 3.1 release should contain this
change.
Relevant RPM upstream PR: https://github.com/rpm-software-management/rpm/pull/2080
Relevant bug: https://bugzilla.redhat.com/show_bug.cgi?id=2105393
Previous commit changed GC compaction methods to not be implemented
when not supported. However, that commit only does compile time checks,
but there are additional compaction support checks during run time.
This commit changes it so that GC compaction methods aren't defined
also during run time if the platform does not support GC compaction.
The patch is manually backported from following change set:
https://github.com/ruby/ruby/pull/60192c19086323
To apply the patch included with this commit,
first apply `ruby-3.2.0-define-unsupported-gc-compaction-methods-as-rb_f_notimplement.patch`
and then apply the `ruby-3.2.0-detect_compaction_support_during_runtime.patch`.
Related upstream issue: https://bugs.ruby-lang.org/issues/18829
The patch implements the methods as rb_f_notimplement.
To test for compaction users can now use `GC.respond_to?(:compact)`.
The upstream patch relies on the macro `GC_COMPACTION_SUPPORTED`
that is defined when the `__wasi__` is false. The define is defined by
an arch conditional in the specfile, which is not optimal but works in our case.
663833b08fhttps://bugs.ruby-lang.org/issues/18829
This also requires regenerating the gc.rbinc and miniprelude.c
since the patch touches gc.rb. This patch is available in
ruby-3.2.0-define-unsupported-gc-compaction-methods_generated-files.patch.
To regenerate the gc.rbinc and miniprelude.c files patch:
~~~
tar -Jxvf ./ruby-3.1.2.tar.xz
git clone https://github.com/ruby/ruby.git
cd ruby && git checkout v3_1_2
patch -p1 < ../ruby-3.2.0-define-unsupported-gc-compaction-methods-as-rb_f_notimplement.patch
./autogen.sh && ./configure
make gc.rbinc miniprelude.c
cd ..
diff -u {ruby-3.1.2,ruby}/gc.rbinc > ruby-3.2.0-define-unsupported-gc-compaction-methods_generated-files.patch
diff -u {ruby-3.1.2,ruby}/miniprelude.c >> ruby-3.2.0-define-unsupported-gc-compaction-methods_generated-files.patch
~~~
Uptream bug: https://bugs.ruby-lang.org/issues/18779
Upstream PR: https://github.com/ruby/ruby/pull/5934
According to the <https://bugs.ruby-lang.org/issues/18756>, the `Q=` is not
recommended. The `V=1` is recommended to output for the verbose mode.
The `make V=1` suppresses some logs that `make` without `V=1` outputs rather
than just adding verbose logs.
While the `V=1 ECHO0=echo` is more equivalent with `Q=`, the
`make V=1 ECHO0=echo` doesn't really output meaningful logs additionally.
So, the `make V=1` is the best option for us to check the compiler flags.
Because the `%make_build` macro includes the `V=1`. Note there is no actual
difference of build.log between before and after this commit.
```
$ rpm --eval %make_build
/usr/bin/make -O -j8 V=1 VERBOSE=1
```
Because the gem extensions were built in `%{buildroot}` they referred
`BUILDROOT` directory, which would be under normal circumstances
reported by RPM, but there was a bug in RPM due to changes in grep:
https://github.com/rpm-software-management/rpm/issues/1968
