Latx32 1.6.4 hangs on i386 apt-get --version

[luzeng87]

Summary

Latx32 1.6.4 does not return when running the i386 apt-get --version command. The hang is reproducible both through the host i386 binfmt_misc registration and by invoking Latx32 explicitly. No network access is involved.

Environment

• Host architecture: loongarch64
• Kernel: 6.18.27, 16K page size
• Latx32: lat-i386 1.6.4
• Latx32 SHA-256: d18eb64f804359369fdc4406a7ea865561536e23db41843fe9cbd7b25e7e7e83
• Latx32 Build ID: 62acc734612c72f911fad4dab9ed12816fd40a0d
• Guest: dynamically linked i386 user space; OCI image ID sha256:586aff9495ac496e8fe0e09b2dcfc19b65be4b832cdcc0163617f75458e8c778

Reproduction

The complete cleanup-safe reproducer is:

https://github.com/luzeng87/loongrun/blob/main/tools/reproduce-latx32-apt-failure.sh

Run it with an i386 image containing Latx32, its native dependency closure, apt-get, and dpkg:

DOCKER_COMMAND="sudo docker" \
  ./tools/reproduce-latx32-apt-failure.sh <i386-image>

Equivalent failing commands are:

# Through binfmt_misc
docker run --rm --platform linux/386 --network none \
  <i386-image> /usr/bin/apt-get --version

# Explicit Latx32 entry
docker run --rm --platform linux/386 --network none \
  --entrypoint /opt/loongrun/native/lib64/ld-linux-loongarch-lp64d.so.1 \
  <i386-image> \
  --library-path /opt/loongrun/native/lib \
  /opt/loongrun/native/bin/latx-i386-1.6 \
  /usr/bin/apt-get --version

Both remain running until forcibly killed. The reproducer stops each container after 8 seconds and records exit code 124.

Controls

In the same image and invocation environment:

• /bin/true: exits 0
• /usr/bin/dpkg --version: exits 0
• /usr/bin/gpgv --version: exits 0
• dynamically linked i386 Python, X11 clients, PulseAudio clients, and dbus-send: run successfully

Observed summary:

reproduction_summary control=pass dpkg=pass binfmt_apt=timeout explicit_apt=timeout

An earlier apt-get update attempt also terminated with signal 11, but the network-free apt-get --version hang is the deterministic minimal case.

Expected result

apt-get --version should print version information and exit 0, as it does on a native i386 system.

[LaurenIsACoder]

I found a concrete bug in the same apt path on current master. The reproduced symptom was a SIGSEGV rather than a hang: lea edi,[0] returned a zero IR operand without actually clearing EDI, so apt later used a stale EDI value in the following TLS address calculation.

I sent a fix in #387. With that change, i386 apt-get --version exits normally with both LATX_AOT=0 LATX_TU=0 and the default translation settings. I also checked the exact LEA case by temporarily reversing the fix: it fails before the change and passes after it.

All runs invoked latx-i386 explicitly. The binfmt registration was not used or modified, and this apt command did not fork or exec another guest program during the test.

I do not want to call the original report fixed yet because the reproduced symptom was not the same timeout. Please try #387 with the original reproducer so we can confirm whether both failures have the same root cause.