Collecting a context together with its memory segfaults #714

Every Level Zero object is destroyed by its own finalizer, and Julia runs the finalizers of objects that become unreachable in the same collection in no particular order. A ZeContext can thus be destroyed before memory allocated in it, and freeing that memory afterwards crashes in NEO:

using oneAPI, oneAPI.oneL0

function work()
    ctx = ZeContext(oneAPI.driver())
    for _ in 1:3; GC.gc(true); end   # the context is old by the time the array is allocated
    old = context()
    context!(ctx)
    a = oneArray{Float32}(undef, 1_000_000)
    context!(old)
    return
end
work()
GC.gc(true)

[19895] signal 11 (1): Segmentation fault
_ZN2L07Context7freeMemEPKvb at libze_intel_gpu.so.1
zeMemFree at ...

This fails every time on an Iris Xe with NEO 26.35. Without the collections before the allocation the finalizers happen to run in a working order. A context in which a kernel was launched, or small (pooled) arrays were allocated, stays reachable from oneAPI.jl's caches and is never collected, which is probably why this hasn't come up.

Forcing other parent-before-child orders (a queue before its fences, an event pool before its events, a module before its kernels, a command list while it is executing) does not crash on NEO, though the Level Zero spec doesn't allow them either. The context is the one that matters: destroying it frees the allocations made in it.

CUDA.jl avoids this by not destroying contexts from a finalizer at all (CuContext has no finalizer, see also JuliaGPU/CUDA.jl#3404 for releases after an explicit destroy).