ac4afe746f
Looks like ubuntu 17.04 is broken now and it fails to link on 17.10 and 18.04. It seems to work on stretch though! |
||
---|---|---|
.. | ||
docker | ||
ios | ||
android-install-ndk.sh | ||
android-install-sdk.sh | ||
android-sysimage.sh | ||
dox.sh | ||
emscripten-entry.sh | ||
emscripten.sh | ||
landing-page-footer.html | ||
landing-page-head.html | ||
linux-s390x.sh | ||
linux-sparc64.sh | ||
README.md | ||
run-docker.sh | ||
run-qemu.sh | ||
run.sh | ||
runtest-android.rs | ||
style.rs | ||
test-runner-linux |
The goal of the libc crate is to have CI running everywhere to have the strongest guarantees about the definitions that this library contains, and as a result the CI is pretty complicated and also pretty large! Hopefully this can serve as a guide through the sea of scripts in this directory and elsewhere in this project.
Files
First up, let's talk about the files in this directory:
-
run-travis.sh
- a shell script run by all Travis builders, this is responsible for setting up the rest of the environment such as installing new packages, downloading Rust target libraries, etc. -
run.sh
- the actual script which runs tests for a particular architecture. Called from therun-travis.sh
script this will run all tests for the target specified. -
cargo-config
- Cargo configuration of linkers to use copied into place by therun-travis.sh
script before builds are run. -
dox.sh
- script called fromrun-travis.sh
on only the linux 64-bit nightly Travis bots to build documentation for this crate. -
landing-page-*.html
- used bydox.sh
to generate a landing page for all architectures' documentation. -
run-qemu.sh
- see discussion about QEMU below -
mips
,rumprun
- instructions to build the docker image for each respective CI target
CI Systems
Currently this repository leverages a combination of Travis CI and AppVeyor for running tests. The triples tested are:
- AppVeyor
{i686,x86_64}-pc-windows-{msvc,gnu}
- Travis
{i686,x86_64,mips,aarch64}-unknown-linux-gnu
{x86_64,aarch64}-unknown-linux-musl
arm-unknown-linux-gnueabihf
arm-linux-androideabi
{i686,x86_64}-apple-{darwin,ios}
x86_64-rumprun-netbsd
x86_64-unknown-freebsd
x86_64-unknown-openbsd
The Windows triples are all pretty standard, they just set up their environment then run tests, no need for downloading any extra target libs (we just download the right installer). The Intel Linux/OSX builds are similar in that we just download the right target libs and run tests. Note that the Intel Linux/OSX builds are run on stable/beta/nightly, but are the only ones that do so.
The remaining architectures look like:
- Android runs in a docker image with an emulator, the NDK, and the SDK already set up. The entire build happens within the docker image.
- The MIPS, ARM, and AArch64 builds all use the QEMU userspace emulator to run the generated binary to actually verify the tests pass.
- The MUSL build just has to download a MUSL compiler and target libraries and then otherwise runs tests normally.
- iOS builds need an extra linker flag currently, but beyond that they're built as standard as everything else.
- The rumprun target builds an entire kernel from the test suite and then runs it inside QEMU using the serial console to test whether it succeeded or failed.
- The BSD builds, currently OpenBSD and FreeBSD, use QEMU to boot up a system and compile/run tests. More information on that below.
QEMU
Lots of the architectures tested here use QEMU in the tests, so it's worth going over all the crazy capabilities QEMU has and the various flavors in which we use it!
First up, QEMU has userspace emulation where it doesn't boot a full kernel, it
just runs a binary from another architecture (using the qemu-<arch>
wrappers).
We provide it the runtime path for the dynamically loaded system libraries,
however. This strategy is used for all Linux architectures that aren't intel.
Note that one downside of this QEMU system is that threads are barely
implemented, so we're careful to not spawn many threads.
For the rumprun target the only output is a kernel image, so we just use that
plus the rumpbake
command to create a full kernel image which is then run from
within QEMU.
Finally, the fun part, the BSDs. Quite a few hoops are jumped through to get CI working for these platforms, but the gist of it looks like:
- Cross compiling from Linux to any of the BSDs seems to be quite non-standard. We may be able to get it working but it might be difficult at that point to ensure that the libc definitions align with what you'd get on the BSD itself. As a result, we try to do compiles within the BSD distro.
- On Travis we can't run a VM-in-a-VM, so we resort to userspace emulation (QEMU).
- Unfortunately on Travis we also can't use KVM, so the emulation is super slow.
With all that in mind, the way BSD is tested looks like:
- Download a pre-prepared image for the OS being tested.
- Generate the tests for the OS being tested. This involves running the
ctest
library over libc to generate a Rust file and a C file which will then be compiled into the final test. - Generate a disk image which will later be mounted by the OS being tested. This image is mostly just the libc directory, but some modifications are made to compile the generated files from step 2.
- The kernel is booted in QEMU, and it is configured to detect the libc-test image being available, run the test script, and then shut down afterwards.
- Look for whether the tests passed in the serial console output of the kernel.
There's some pretty specific instructions for setting up each image (detailed
below), but the main gist of this is that we must avoid a vanilla cargo run
inside of the libc-test
directory (which is what it's intended for) because
that would compile syntex_syntax
, a large library, with userspace emulation.
This invariably times out on Travis, so we can't do that.
Once all those hoops are jumped through, however, we can be happy that we're testing almost everything!
Below are some details of how to set up the initial OS images which are
downloaded. Each image must be enabled have input/output over the serial
console, log in automatically at the serial console, detect if a second drive in
QEMU is available, and if so mount it, run a script (it'll specifically be
run-qemu.sh
in this folder which is copied into the generated image talked
about above), and then shut down.
QEMU setup - FreeBSD
- Download CD installer (most minimal is fine)
qemu-img create -f qcow2 foo.qcow2 2G
qemu -cdrom foo.iso -drive if=virtio,file=foo.qcow2 -net nic,model=virtio -net user
- run installer
echo 'console="comconsole"' >> /boot/loader.conf
echo 'autoboot_delay="0"' >> /boot/loader.conf
- look at /etc/ttys, see what getty argument is for ttyu0
- edit /etc/gettytab, look for ttyu0 argument, prepend
:al=root
to line beneath
(note that the current image has a freebsd
user, but this isn't really
necessary)
Once that's done, arrange for this script to run at login:
#!/bin/sh
sudo kldload ext2fs
[ -e /dev/vtbd1 ] || exit 0
sudo mount -t ext2fs /dev/vtbd1 /mnt
sh /mnt/run.sh /mnt
sudo poweroff
Helpful links
- https://en.wikibooks.org/wiki/QEMU/Images
- https://blog.nekoconeko.nl/blog/2015/06/04/creating-an-openstack-freebsd-image.html
- https://www.freebsd.org/doc/handbook/serialconsole-setup.html
QEMU setup - OpenBSD
- Download CD installer
qemu-img create -f qcow2 foo.qcow2 2G
qemu -cdrom foo.iso -drive if=virtio,file=foo.qcow2 -net nic,model=virtio -net user
- run installer
echo 'set tty com0' >> /etc/boot.conf
echo 'boot' >> /etc/boot.conf
- Modify /etc/ttys, change the
tty00
at the end from 'unknown off' to 'vt220 on secure' - Modify same line in /etc/ttys to have
"/root/foo.sh"
as the shell - Add this script to
/root/foo.sh
#!/bin/sh
exec 1>/dev/tty00
exec 2>&1
if mount -t ext2fs /dev/sd1c /mnt; then
sh /mnt/run.sh /mnt
shutdown -ph now
fi
# limited shell...
exec /bin/sh < /dev/tty00
chmod +x /root/foo.sh
Helpful links:
Questions?
Hopefully that's at least somewhat of an introduction to everything going on here, and feel free to ping @alexcrichton with questions!