autopkgtest fails in hirsute on armhf

Indeed this failure started happening on armhf recently (I'd say about a week ago or so) for all builds of firefox, affecting not just hirsute but also xenial, bionic, focal and groovy.

The google search URL that's printed suggests a race condition: selenium doesn't have a built-in method to wait until the current URL changes to a given value, instead it allows watching the URL until it changes from the current value. This combined with a slow test environment might explain why the google redirection (to their consent popup) happens before the real search URL is captured by the test.

I cannot reproduce the test failure locally, nor when running the test in an armhf LXC container on an arm64 cloud instance.

When running locally, I can see the google consent popup overlaid on top of the search results page, but the URL isn't rewritten, and the network tab of devtools doesn't show any query to consent.google.com either. There's definitely something different in the test environment that causes this bug.

I am seeing the redirection when using firefox on android. It might be that Google uses the UA string to serve a mobile version of the consent page, and for some reason it thinks the UA string of the desktop version on armhf is that of a mobile device. Not sure why I'm not able to reproduce this in an armhf container though.

This is the UA string of the desktop version on armhf:

    Mozilla/5.0 (X11; Ubuntu; Linux armv8l; rv:87.0) Gecko/20100101 Firefox/87.0

But I'm still unable to reproduce the test failure on armhf.

In the autopkgtest infrastructure, on armhf the UA string is:

    Mozilla/5.0 (X11; Ubuntu; Linux armv7l; rv:88.0) Gecko/20100101 Firefox/88.0

i.e. the difference with my tests in an armhf container is the architecture ("armv7l" vs "armv8l")

And I can confirm that if I spoof the user agent using the "armv7l" token, the google search page is redirected to the consent page, which is consistent with what I'm seeing with firefox on android.

That token in the UA string corresponds to the machine field of the result of the uname syscall (== `uname -m`).

Tentative fix: https://bazaar.launchpad.net/~mozillateam/firefox/firefox.hirsute/revision/1484.

I built firefox with the fix above in a PPA and ran the autopkgtests in the official infrastructure, and I can confirm that they now pass on armhf.

The fix will be released with the next firefox stable update, version 88.0, due out next week.

This bug was fixed in the package firefox - 88.0+build2-0ubuntu0.20.10.1

---------------
firefox (88.0+build2-0ubuntu0.20.10.1) groovy; urgency=medium

  * New upstream release (88.0+build2)

  [ Olivier Tilloy ]
  * On armhf, override the UA string used by an autopkgtest to prevent Google
    from serving mobile content, which would break the test's expectations
    (LP: #1923090)
    - debian/tests/search-engines

  [ Rico Tzschichholz ]
  * Update cbindgen to 0.19.0
    - debian/build/create-tarball.py
  * Use clang 12 if available
    - debian/control{,.in}
    - debian/build/rules.mk
  * Update patches
    - debian/patches/rust-drop-dll-checksums.patch

 -- Olivier Tilloy <email address hidden> Fri, 16 Apr 2021 16:07:30 +0200

This bug was fixed in the package firefox - 88.0+build2-0ubuntu0.20.04.1

---------------
firefox (88.0+build2-0ubuntu0.20.04.1) focal; urgency=medium

  * New upstream release (88.0+build2)

  [ Olivier Tilloy ]
  * On armhf, override the UA string used by an autopkgtest to prevent Google
    from serving mobile content, which would break the test's expectations
    (LP: #1923090)
    - debian/tests/search-engines

  [ Rico Tzschichholz ]
  * Update cbindgen to 0.19.0
    - debian/build/create-tarball.py
  * Use clang 12 if available
    - debian/control{,.in}
    - debian/build/rules.mk
  * Update patches
    - debian/patches/rust-drop-dll-checksums.patch

 -- Olivier Tilloy <email address hidden> Fri, 16 Apr 2021 16:00:27 +0200