build: gate amdgpu plugin on amdgpu dev support

[cheese-cakee]

Problem

CONFIG_AMDGPU was enabled whenever libdrm was detected, even on systems where libdrm is present without amdgpu development support (libdrm/amdgpu.h / -ldrm_amdgpu).

This causes build failures in plugins/amdgpu/amdgpu_plugin.c.

Change

• Add FEATURE_TEST_LIBDRM_AMDGPU in scripts/feature-tests.mak to verify both headers and linkability for amdgpu (-ldrm -ldrm_amdgpu).
• Gate CONFIG_AMDGPU in Makefile.config on that feature test instead of generic pkg-config-check,libdrm.
• Keep behavior minimal: amdgpu plugin is built only when amdgpu dev support is available.

Notes

This is a surgical build-detection fix with no runtime logic changes.

Fixes: #1900

[Copilot]

Pull request overview

This PR fixes build failures when libdrm is installed without amdgpu development support (issue #1900). Previously, CONFIG_AMDGPU was gated on a generic pkg-config-check,libdrm, which succeeded even without amdgpu headers/libraries.

Changes:

• Added a FEATURE_TEST_LIBDRM_AMDGPU compile test that verifies amdgpu headers and linkability
• Replaced the pkg-config-check,libdrm gate with a try-cc feature test using -ldrm -ldrm_amdgpu
• Updated the help message to mention "amdgpu support"

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated no comments.

File	Description
scripts/feature-tests.mak	New FEATURE_TEST_LIBDRM_AMDGPU compile test checking amdgpu headers and API
Makefile.config	Gate CONFIG_AMDGPU on the new feature test instead of generic libdrm pkg-config

---

You can also share your feedback on Copilot code review. Take the survey.

[rst0git]

@cheese-cakee To build the AMD GPU plugin, we currently require libdrm-devel / libdrm-dev as a build dependency, and without this package the plugin will be skipped.

CONFIG_AMDGPU was enabled whenever libdrm was detected, even on systems where libdrm is present without amdgpu development support (libdrm/amdgpu.h / -ldrm_amdgpu). This causes build failures in plugins/amdgpu/amdgpu_plugin.c.

Would you be able to provide more information on how to reproduce this error?

[cheese-cakee]

Thanks for the question @rst0git . Repro is the one documented in #1900 : libdrm installed but AMDGPU development bits absent (libdrm/amdgpu.h and -ldrm_amdgpu), while generic libdrm detection still enabled CONFIG_AMDGPU; build then fails in plugins/amdgpu/amdgpu_plugin.c. This patch makes detection explicit (header + link), so the plugin is built only when AMDGPU dev support is present. If this case is no longer possible in current CI/package constraints, I'm happy to close as obsolete.

[rst0git]

Repro is the one documented in #1900

@cheese-cakee This issue was opened in 2022, and the code has changed significantly since then. Were you able to reproduce the error yourself?

[cheese-cakee]

I reproduced this on current code in a clean Ubuntu 24.04 container.

What I tested:

• Installed build-essential pkg-config libdrm-dev
• Simulated the problematic package state (libdrm present, AMDGPU dev bits absent) by removing:
  • /usr/include/libdrm/amdgpu.h
  • /usr/lib/x86_64-linux-gnu/libdrm_amdgpu.so

Observed results:

• On current criu-dev logic:
  • make -pn | grep CONFIG_AMDGPU -> CONFIG_AMDGPU = y
  • plugin compile fails:
    • plugins/amdgpu/amdgpu_plugin.c:22:10: fatal error: libdrm/amdgpu.h: No such file or directory
• On this PR branch:
  • make -pn | grep CONFIG_AMDGPU -> no CONFIG_AMDGPU (disabled as expected)

So with libdrm detected but missing AMDGPU dev support, current logic still enables AMDGPU, while this patch gates it correctly.

[cheese-cakee]

for clarity:

docker run --rm -it \
  -v "$PWD:/src:ro" ubuntu:24.04 bash -lc '
set -eux
apt-get update -qq
DEBIAN_FRONTEND=noninteractive apt-get install -y -qq build-essential pkg-config libdrm-dev
cp -a /src /work

# Simulate: libdrm present, AMDGPU dev bits missing
mv /usr/include/libdrm/amdgpu.h /usr/include/libdrm/amdgpu.h.bak || true
mv /usr/lib/x86_64-linux-gnu/libdrm_amdgpu.so /usr/lib/x86_64-linux-gnu/libdrm_amdgpu.so.bak || true

# Current criu-dev behavior: generic libdrm check still enables AMDGPU
(cd /work && make -pn 2>/dev/null | grep -m1 CONFIG_AMDGPU || true)

# This then fails in amdgpu plugin compilation
(gcc $(pkg-config --cflags libdrm) -c /work/plugins/amdgpu/amdgpu_plugin.c -o /tmp/amdgpu.o) || true
'

[rst0git]

# Simulate: libdrm present, AMDGPU dev bits missing
mv /usr/include/libdrm/amdgpu.h /usr/include/libdrm/amdgpu.h.bak || true
mv /usr/lib/x86_64-linux-gnu/libdrm_amdgpu.so /usr/lib/x86_64-linux-gnu/libdrm_amdgpu.so.bak || true

These lines look like intentionally breaking parts of the AMD GPU userspace stack :)

[cheese-cakee]

@rst0git I reran this with a package-native setup (no manual deletion of distro files):

• Built and installed libdrm from source with AMDGPU disabled (-Damdgpu=disabled) into an isolated prefix.
• Verified in that prefix:
  • pkg-config --exists libdrm -> yes
  • include/libdrm/amdgpu.h -> missing
  • libdrm_amdgpu.so -> missing

Results:

• Old logic equivalent (pkg-config-check,libdrm) would enable CONFIG_AMDGPU.
• New feature-test logic fails as expected:
  • fatal error: libdrm/amdgpu.h: No such file or directory
  • therefore CONFIG_AMDGPU stays disabled.

So this reproduces the exact mismatch using a real package build configuration, without mutating installed distro package contents.

[cheese-cakee]

Adding exact code context that this PR changes:

Old gate (current criu-dev):

ifeq ($(call pkg-config-check,libdrm),y)
export CONFIG_AMDGPU := y
endif

This only checks generic libdrm presence.

New gate (Makefile.config):

LIBDRM_CFLAGS := $(shell $(PKG_CONFIG) --cflags libdrm 2>/dev/null)
ifeq ($(call try-cc,$(FEATURE_TEST_LIBDRM_AMDGPU),$(LIBDRM_CFLAGS) -ldrm -ldrm_amdgpu),true)
export CONFIG_AMDGPU := y
endif

Feature test (scripts/feature-tests.mak):

#include <xf86drm.h>
#include <libdrm/amdgpu.h>
...
return amdgpu_device_initialize(-1, &major, &minor, &dev);

So we now require both headers and linkability to AMDGPU userspace before enabling the plugin.

[cheese-cakee]

Hi @fdavid-amd, following up on the review request from @avagin. The change is a surgical build-detection fix: adds a compile-time check for libdrm/amdgpu.h and ldrm_amdgpu before enabling CONFIG_AMDGPU.

Happy to answer any questions.