[maleadt]

NEO installs a SIGSEGV handler that forwards faults it doesn't own to Julia's, tracking its position in the handler chain with a counter shared by all threads. Julia's GC makes every thread fault on the safepoint page at the same time, so with more than one thread a GC crashed the process with a bare SIGSEGV. NEO_jll 26.35.39758+1 (JuliaPackaging/Yggdrasil#15164) makes that counter thread-local; the fix is proposed upstream in intel/compute-runtime#1009.

The test runs a 4-thread GC loop in a child process. NEO only installs its handler for shared allocations with device storage, which integrated GPUs don't use, so the test forces that with NEO debug keys; on the TGL runner the allocation then fails, but the handler is installed. With NEO_jll +0 the child dies from SIGSEGV 5/5 times, with +1 it passes 5/5.

The test is skipped when NEO_jll's libze_intel_gpu is redirected outside its artifact, as on Aurora, where the system driver doesn't have the fix yet.

Fixes #607.

[github-actions]

Your PR requires formatting changes to meet the project's style guidelines.
Please consider running Runic (git runic main) to apply these changes.

Click here to view the suggested changes.

diff --git a/test/level-zero.jl b/test/level-zero.jl
index b86ed26..4b0a25d 100644
--- a/test/level-zero.jl
+++ b/test/level-zero.jl
@@ -391,30 +391,30 @@ end
 
 @testset "SIGSEGV handler chaining" begin
 
-# NEO installs a SIGSEGV handler that chains to Julia's. Julia's GC makes every thread fault
-# on its safepoint page at the same time, which crashed NEO before it made chaining
-# thread-safe (oneAPI.jl#607, intel/compute-runtime#1009).
-script = raw"""
-    using oneAPI, oneAPI.oneL0
-    # NEO installs its handler on the first shared allocation that gets device storage.
-    # Integrated GPUs normally don't use it, and fail the allocation when forced to,
-    # but the handler is installed anyway.
-    try
-        free(shared_alloc(context(), device(), 1024))
-    catch err
-        err isa oneL0.OutOfGPUMemoryError || rethrow()
-    end
-    Threads.@threads for i in 1:100
-        x = [rand(100) for _ in 1:1000]
-        GC.gc(false)
-    end
-"""
-cmd = `$(Base.julia_cmd()) --project=$(Base.active_project()) --threads=4 -e $script`
-cmd = addenv(cmd, "NEOReadDebugKeys" => "1", "AllocateSharedAllocationsWithCpuAndGpuStorage" => "1")
-# Only our NEO_jll build carries the fix; skip with a system driver (e.g. on Aurora, where
-# preferences point NEO_jll at /usr/lib64).
-NEO_jll = oneL0.NEO_jll
-system_driver = !startswith(NEO_jll.libze_intel_gpu, NEO_jll.artifact_dir)
-@test success(pipeline(cmd, stdout=stdout, stderr=stderr)) skip=system_driver
+    # NEO installs a SIGSEGV handler that chains to Julia's. Julia's GC makes every thread fault
+    # on its safepoint page at the same time, which crashed NEO before it made chaining
+    # thread-safe (oneAPI.jl#607, intel/compute-runtime#1009).
+    script = raw"""
+        using oneAPI, oneAPI.oneL0
+        # NEO installs its handler on the first shared allocation that gets device storage.
+        # Integrated GPUs normally don't use it, and fail the allocation when forced to,
+        # but the handler is installed anyway.
+        try
+            free(shared_alloc(context(), device(), 1024))
+        catch err
+            err isa oneL0.OutOfGPUMemoryError || rethrow()
+        end
+        Threads.@threads for i in 1:100
+            x = [rand(100) for _ in 1:1000]
+            GC.gc(false)
+        end
+    """
+    cmd = `$(Base.julia_cmd()) --project=$(Base.active_project()) --threads=4 -e $script`
+    cmd = addenv(cmd, "NEOReadDebugKeys" => "1", "AllocateSharedAllocationsWithCpuAndGpuStorage" => "1")
+    # Only our NEO_jll build carries the fix; skip with a system driver (e.g. on Aurora, where
+    # preferences point NEO_jll at /usr/lib64).
+    NEO_jll = oneL0.NEO_jll
+    system_driver = !startswith(NEO_jll.libze_intel_gpu, NEO_jll.artifact_dir)
+    @test success(pipeline(cmd, stdout = stdout, stderr = stderr)) skip = system_driver
 
 end

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 83.44%. Comparing base (8460d3f) to head (13e0764).

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #706      +/-   ##
==========================================
- Coverage   83.70%   83.44%   -0.26%     
==========================================
  Files          52       51       -1     
  Lines        4376     4253     -123     
==========================================
- Hits         3663     3549     -114     
+ Misses        713      704       -9     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.