Adapt to JuliaLang/julia#38487.

[maleadt]

No description provided.

[maleadt]

Annoying how the threading test doesn't fail after removing the protections (but without the PR in Julia being merged)... Maybe now we device_reset! at the beginning of every test there isn't enough memory pressure.

[maleadt]

Benchmarks caught it!

WARNING: Error while freeing CuPtr{Nothing}(0x00007fff63d8b600):
AssertionError(msg="!(islocked(nrl.rl)) || nrl.rl.locked_by !== current_task()")

[vchuravy]

We still need the @spinlock? Maybe worth adding to Julia base?

[maleadt]

We still need the @spinlock?

I think so. We can't wait from within a finalizer, because that would cause a task switch. So if another thread is in free (which can happen from non-finalizer code because of unsafe_free) we just need to wait until that's done. So we spin.

This seems pretty specific, so I'm not sure it's useful for Base (where there already is a spinlock for when all uses should spin).

[vchuravy]

Yeah @jpsamaroo is running into this with trying to make Distributed.jl thread-safe, but maybe we can use Spinlocks there.

[codecov]

Codecov Report

Merging #568 (8e3d94e) into master (8eb218d) will decrease coverage by 0.01%.
The diff coverage is 58.33%.

@@            Coverage Diff             @@
##           master     #568      +/-   ##
==========================================
- Coverage   79.94%   79.93%   -0.02%     
==========================================
  Files         116      116              
  Lines        6870     6861       -9     
==========================================
- Hits         5492     5484       -8     
+ Misses       1378     1377       -1     

Impacted Files	Coverage Δ
src/pool.jl	73.02% <50.00%> (-0.58%)	⬇️
src/pool/simple.jl	87.30% <50.00%> (ø)
src/pool/binned.jl	89.56% <66.66%> (ø)
src/pool/split.jl	93.29% <66.66%> (ø)

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update aa591bb...8e3d94e. Read the comment docs.