Enable LTO support on Linux by default again.
Add patch to fix dependentlibs.list generation under LTO. This is
necessary for fixing firefox-wayland crashing when built with LTO.
Add makeFlags which set ar, ranlib, and nm to be llvm-ar, llvm-ranlib
and llvm-nm when building with llvm-based LTO. (bmo#1480005)
This was required to solve the XPCOMGlueLoad error when building with
LTO. However, it turns out libxul.so is supposed to have some libraries
that are reported as not found by ldd. Setting the RPATH worked around
the error as it forced dependency resolution but failed to fix the real
issue of broken generation of dependentlibs.list.
The libraries that are reported as not found by ldd are supposed to be
dlopened through the logic found in nsXPCOMGlue.cpp. However since the
generation of dependentlibs.list is broken under LTO this did not
happen. Instead of pulling libwayland-client.so from the GTK libraries
it found the stub library first (libmozwayland.so). The stub library
causes (as it should) wl_display_connect to always return NULL which is
the cause of the segmentation fault and LTO breaking wayland support.
Remove the hardcoded path used for the XPCOMGlueLoad error workaround
in NIX_LDFLAGS. libunwind is still unfortunately needed. Once the issue
of the generation of dependentlibs.list being borked is fixed it should
remedy the wayland crash issue on LTO.
Firefox has a number of optional dependencies that get dlopened.
Instead of using patchelf to set the RPATH use LD_LIBRARY_PATH.
The motivation for this is we already set LD_LIBRARY_PATH in the
wrapper on Linux.
It only affected FF80 so place an upper bound restriction. See
bmo#1661096 for details.
This fixes substituteStream() warnings about missing patterns which
appeared in the logs.
It was added for nspr and nss back in the 55.0.3 to 56.0 upgrade. It
also served as a workaround for an undeclared gio-unix-2.0 dependency.
Sometime afterwards nspr was removed, leaving just the two. Since then,
upstream has added a declaration for gio-unix-2.0 (in FF62). As for the
nss include it seemingly has no purpose since current firefox builds
with it removed.
After the fedora patches for screen sharing using pipewire got updated
for Firefox 83 (pipewire was inlined there), the nixpkgs buildInput
pipewire got stripped from the resulting firefox binary and so firefox
was unable to actually get the shared stream from the running pipewire
service.
Adding pipewire to the firefox binary with `patchelf --add-needed`
makes it atually get the stream from the service.
Fixes: #106812
libcubeb has dlopened libraries for awhile now. In nixpkgs there was
support for the PulseAudio backend doing this, however the ALSA backend
support was missed and caused issue #79310 (no sound with ALSA). This
gives ALSA users the ability to hear sound once again.
As discussed in #101429 firefox 82 started crashing when used with
wayland. A brief investigation showed that this appears to be rooted
within the LTO support that was recently added to the package. For the
time being, until someone figures out where the crashes are coming from,
we can just disable LTO.
This ensures that we aren't applying any of the experiemental pipewire
patches when the dependencies aren't enabled. As of now pipewire only
works with wayland and webrtc. If either of them are not activated we
can't build with pipewireSupport and we should not.
Regression introduced by bce5268a21.
The bit size of the initialisation vector for AES GCM has been
introduced in NSS version 3.52 in the CK_GCM_PARMS struct via the
ulIvBits field.
Unfortunately, Firefox 68.8.0 and 76.0 do not set this field and thus it
gets initialised to zero, which in turn causes IV generation to fail.
I found out about this because WebRTC stopped working after updating to
NSS 3.52 and so I started bisecting.
Since there wasn't an obvious error in Firefox hinting towards NSS but
instead just the video stream ended up as a "null" stream, I didn't
suspect the NSS update to be the culprit at first. So I verified a few
times and then also started bisecting the actual commit in NSS that
caused the issue.
This turned out to be the problematic change:
https://phabricator.services.mozilla.com/D63241
> One notable change was caused by an inconsistancy between the spec and
> the released headers in PKCS#11 v2.40. CK_GCM_PARAMS had an extra
> field in the header that was not in the spec. OASIS considers the
> header file to be normative, so PKCS#11 v3.0 resolved the issue in
> favor of the header file definition.
Since the test I've used[1] was a bit flaky, I still didn't believe the
result of the bisect to be accurate, but after running the test several
times leading same results I dug through the above change line by line
to get more clues.
It fortunately didn't take that long to stumble upon the ulIvBits change
(which is actually documented in the NSS 3.52 release notes[4], but I
managed to blatantly ignore it for some reason) and started checking the
Firefox source tree for changes regarding that field.
Initialisation of that new field has been introduced[2] in preparation
for the 76 release, but subsequently got reverted[3] prior to the
release, because Firefox 76 is expected to be shipped with NSS 3.51,
which didn't have the ulIvBits field.
The patch I'm adding here is just a reintroduction of that change,
because we're using NSS 3.52. Not initialising that field will break
WebRTC and WebCrypto, which I think the former seems to gain in
popularity these days ;-)
Tested the change against the mentioned VM test[1] and also by testing
manually using Jitsi Meet and Nextcloud Talk.
[1]: https://github.com/aszlig/avonc/tree/884315838b6f0ebb32b/tests/talk
[2]: https://hg.mozilla.org/mozilla-central/rev/3ed30e6b6de1
[3]: https://hg.mozilla.org/mozilla-central/rev/665137da70ee
[4]: https://developer.mozilla.org/en-US/docs/Mozilla/Projects/NSS/NSS_3.52_release_notes
Signed-off-by: aszlig <aszlig@nix.build>
