Travis CI Android emulator-based tests for Android are broken

[briansmith]

It seems like the Android SDK changed in some non-backward-compatible way. If somebody has a fix, I'd take a PR for it. Otherwise I think I might try rewriting all the non-Windows CI stuff using Docker and then using the Docker-ified stuff, ETA in Feburary.

[briansmith]

See rust-lang/rust#45580 for some inspiration.

[andrewtj]

I think a change in Rust might have inadvertently broken the test suite. This glob in travis.sh matches two files so adb push attempts to place them both in the directory /data/ring-test instead of just copying the test executable to that path. Patching that gets the run a little further to an Illegal instruction failure.

[andrewtj]

And here's a tombstone.

[briansmith]

@andrewtj That's great news! It seems like there were several problems.

The problem I was experiencing previously was "adb: command not found" and "emulator: command not found" errors on Travis CI. Now it seems those errors aren't occurring! Maybe the problem was just that the Travis CI cache was in a corrupted state or something.

Because of those "command not found" errors, I wasn't able to notice the problem that your patch fixes, which apparently is due to a change in rustc. Thanks for providing an example fix.

I went back to tag 0.12.1 and applied that fix to the 0.12.1 tag and kicked off a Travis CI build. The illegal instruction error still occurs. However, back when I released 0.12.1, the build passed just fine; see https://travis-ci.org/briansmith/ring/builds/271520581. That indicates to me that there is probably a change in the emulator, or a change in rustc or some other part of the toolchain that makes it incompatible with the emulator.

[andrewtj]

Okay I've run a few more builds and it looks like 1.21.0 works but 1.22.1 does not. So I guess something must have changed in Rust's jemalloc fork in 1.22?

[briansmith]

The Rust 1.22.0/1.22.1 release notes are at https://github.com/rust-lang/rust/blob/master/RELEASES.md#version-1220-2017-11-22

The most clearly relevant changes listed in the release notes are:

• enable strict alignment (+strict-align) on ARMv6 rust-lang/rust#45094: enable strict alignment (+strict-align) on ARMv6
• [beta] ci: Upgrade Android SDK/NDK and refactor to use sdkmanager/avdmanager. rust-lang/rust#45656: The minimum Android version that rustc can build for has been bumped to 4.0 from 2.3 / Upgrade Android SDK/NDK and refactor to use sdkmanager/avdmanager.

In Rust 1.21.0, jemalloc was upgraded to version 4.5.0, according to those release notes. However, you said 1.21.0 works fine.

See also:

• armv5te hello world fails to run on qemu and pi3 rust-lang/rust#46822
• arm-unknown-linux-gnueabihf generates armv7 instructions rust-lang/rust#45284
• 1.17.0 armv7-gnueabihf Debian: SIGILL: illegal instruction rust-lang/rust#42477
• [beta] armv7-unknown-linux-gnueabihf bootstrap SIGILL rust-lang/rust#41291

[andrewtj]

I've narrowed it down to something between nightly-2017-10-29 and nightly-2017-10-30:

#!/bin/sh
set -x
for REL in nightly-2017-10-29 nightly-2017-10-30; do
	rustup override set ${REL} 
	cargo build --target=armv7-linux-androideabi
	adb push target/armv7-linux-androideabi/debug/hello_world /data
	adb shell '/data/hello_world; echo $?'
done

$ ./test_run.sh 
+ rustup override set nightly-2017-10-29
info: using existing install for 'nightly-2017-10-29-x86_64-unknown-linux-gnu'
info: override toolchain for '/home/atj/hello_world' set to 'nightly-2017-10-29-x86_64-unknown-linux-gnu'

  nightly-2017-10-29-x86_64-unknown-linux-gnu unchanged - rustc 1.23.0-nightly (269cf5026 2017-10-28)

+ cargo build --target=armv7-linux-androideabi
    Finished dev [unoptimized + debuginfo] target(s) in 0.0 secs
+ adb push target/armv7-linux-androideabi/debug/hello_world /data
target/armv7-linux-androideabi/debug/hello_world: 1 file pushed. 10.6 MB/s (3949796 bytes in 0.356s)
+ adb shell /data/hello_world; echo $?
Hello, world!
0
+ rustup override set nightly-2017-10-30
info: using existing install for 'nightly-2017-10-30-x86_64-unknown-linux-gnu'
info: override toolchain for '/home/atj/hello_world' set to 'nightly-2017-10-30-x86_64-unknown-linux-gnu'

  nightly-2017-10-30-x86_64-unknown-linux-gnu unchanged - rustc 1.23.0-nightly (90ef3372e 2017-10-29)

+ cargo build --target=armv7-linux-androideabi
    Finished dev [unoptimized + debuginfo] target(s) in 0.0 secs
+ adb push target/armv7-linux-androideabi/debug/hello_world /data
target/armv7-linux-androideabi/debug/hello_world: 1 file pushed. 9.4 MB/s (3271284 bytes in 0.333s)
+ adb shell /data/hello_world; echo $?
Illegal instruction 
132

It looks like rust-lang/rust#45656 landed in that window. I'm not familiar with the Android toolchain so I'm going to leave it here for the minute.

[briansmith]

I'm guessing that the version of the Android SDK that were'e using on Travis CI, in particular its emulator, is simply incompatible with the one that rustc uses. The Android SDK does warn us that it is out of date during installation. I think we need to switch to using a sdkmanager-based installation. I'll take a look at doing that.

[andrewtj]

That sounds reasonable. It's noted in rust-lang/rust#45580 that the minimum supported version is 14 and the current script sets up a version 18 emulator, but emulator and device are probably different ball games.

[briansmith]

Commit 9e2bcb6 disables the Android emulator tests on Travis CI pending a solution to this.