Testing/LTP

Testing Linux usermode emulation with the Linux Test Project

You can use the Linux Test Project's syscall tests to test QEMU's linux-user mode support. These instructions are for how to do that using a Debian chroot running under an x86 Ubuntu or Debian host (I tested with Ubuntu Precise).

These instructions set up an armhf chroot; they should in theory work for any architecture supported by both Debian and QEMU.

Setting up the testsuite environment

First we need to set up the chroot. This takes a while but once you've done it you can reuse the chroot for multiple test runs. We build the LTP testsuite inside the chroot, which is slow but simpler than setting up a cross-build environment. The finished chroot including the compiled testsuite needs about 620MB of disk space. [This could almost certainly be trimmed down but disk is cheap and time is expensive :-)]

Download the LTP source tarball; I used the "January 2013 Stable" release: https://sourceforge.net/projects/ltp/files/LTP%20Source/ltp-20130109/ltp-full-20130109.bz2/download

As root in the host system:

  mkdir /srv/chroot/ltp-armhf
  qemu-debootstrap --arch armhf wheezy /srv/chroot/ltp-armhf
  echo "deb http://ftp.us.debian.org/debian wheezy main" > /srv/chroot/ltp-armhf/etc/apt/sources.list
  cd /srv/chroot/ltp-armhf
  tar xvjf ltp-full-20130109.bz2
  chroot /srv/chroot/ltp-armhf

Then in the chroot (the compile stage will take an hour or two):

  apt-get update
  apt-get install build-essential
  cd ltp-full-20130109
  ./configure && make && make install

This will install the tests in /opt/ltp/ inside the chroot.

Create an /opt/ltp/qemu.skiplist file inside the chroot with the following contents:

# skiplist for QEMU testing
# This is a list of tests which hang completely under QEMU
# or are otherwise badly behaved (as opposed to merely failing).
# We should probably investigate them more closely at some point.
# 
# Skip all the clone tests, QEMU threading support is known to be broken
# and one of the clone tests seems to cause the LTP test harness
# to bail out entirely.
clone01
clone02
clone03
clone04
clone05
clone06
clone07
# Seems to hang
fork13
# These tests get in a total mess with signals
kill10
kill11
# This runs OK but thrashes the machine with lots of processes
msgctl11
# These three seem to hang
msgrcv03
nanosleep04
splice02
# these tests try to restart syslogd!?!
syslog01
syslog02
syslog03
syslog04
syslog05
syslog06
syslog07
syslog08
syslog09
syslog10
syslog11
syslog12
# hangs
waitpid02

Running tests

OK, now we're ready to actually do a test run! You'll need to build the QEMU to test as a static executable, for example:

   ./configure --target-list=arm-linux-user --static && make -j2

Then copy the arm-linux-user/qemu-arm binary into /srv/chroot/ltp-armhf/usr/bin/qemu-arm-static (you'll need to make sure you don't still have a shell open in the chroot or the copy will fail).

To run tests in the chroot:

   [ -e /proc/cpuinfo ] || mount proc /proc -t proc
   cd /opt/ltp
   ./runltp -p -l "qemu-$(date +%FT%T).log" -o "qemu-$(date +%FT%T).out" -f /opt/ltp/runtest/syscalls -S /opt/ltp/qemu.skiplist 

This will take an hour or so, and writes a human readable results summary to a file in /opt/ltp/results/, and the complete test output dump to a file in /opt/ltp/output/. You can track its progress by tailing the output file that is created in /opt/ltp/output/.

(The LTP test runner appends results to existing log files rather than overwriting them, which is why we make sure to include a date/timestamp in the filenames.)

Note that the test suite often considers "syscall unimplemented" as a PASS condition. To find out whether QEMU is just missing syscalls completely you'll need to look for "qemu: Unsupported syscall:" lines in the output file.

You can investigate a failure by running a single test like this:

   ./runltp -f /opt/ltp/runtest/syscalls -s accept4

(the -s option takes a regex specifying tests to run).

Current status

As of v1.4.0-rc1 the ARM QEMU ran 959 tests with 86 failures. If anybody would like to set up an automated system to run these tests nightly and produce pretty web pages of regressions/progressions that would be cool :-)