[maleadt]

At exit, Julia runs the remaining finalizers newest first. The AsyncCondition that delivers queued library messages can therefore be finalized before objects whose finalizers still call into a library and make it log, e.g. cuDNN descriptors created before cuDNN.enable_logging(). enqueue_log then passed the condition's null handle to uv_async_send and Julia segfaulted. Skip the wake-up instead, since nothing can deliver the message at that point.

On main this no longer triggers with the reporter's example, because #3318 defers descriptor destruction from finalizers, but any library callback during exit finalization still crashes. The regression test reproduces it with a plain finalizer in a subprocess.

Fixes #3379

[github-actions]

CUDA.jl Benchmarks

Details

Benchmark suite	Current: b327b1f	Previous: 212af1e	Ratio
array/accumulate/Float32/1d	64609 ns	65317 ns	0.99
array/accumulate/Float32/dims=1	208848 ns	208785 ns	1.00
array/accumulate/Float32/dims=1L	231528 ns	231362 ns	1.00
array/accumulate/Float32/dims=2	92902 ns	92936 ns	1.00
array/accumulate/Float32/dims=2L	4414669 ns	4414162 ns	1.00
array/accumulate/Int64/1d	106567 ns	106759 ns	1.00
array/accumulate/Int64/dims=1	223934 ns	224599 ns	1.00
array/accumulate/Int64/dims=1L	404020 ns	404787 ns	1.00
array/accumulate/Int64/dims=2	112184 ns	112409 ns	1.00
array/accumulate/Int64/dims=2L	5303946 ns	5304186 ns	1.00
array/broadcast	10940 ns	11284 ns	0.97
array/broadcast launch	7079.75 ns	7388.5 ns	0.96
array/construct	943.1739130434783 ns	966.34375 ns	0.98
array/copy	16888 ns	17285 ns	0.98
array/copyto!/cpu_to_gpu	208837 ns	206598 ns	1.01
array/copyto!/gpu_to_cpu	241314 ns	240623 ns	1.00
array/copyto!/gpu_to_gpu	8884 ns	8960 ns	0.99
array/iteration/findall/bool	49071 ns	49634 ns	0.99
array/iteration/findall/int	71391 ns	70475 ns	1.01
array/iteration/findfirst/bool	115883 ns	116138 ns	1.00
array/iteration/findfirst/int	113280 ns	113800 ns	1.00
array/iteration/findmin/1d	67294 ns	66791 ns	1.01
array/iteration/findmin/2d	69284 ns	67619 ns	1.02
array/iteration/logical	53091 ns	52939 ns	1.00
array/iteration/scalar	60314 ns	61169 ns	0.99
array/permutedims/2d	47958 ns	48338 ns	0.99
array/permutedims/3d	45229 ns	45659 ns	0.99
array/permutedims/4d	49014 ns	48717 ns	1.01
array/random/rand/Float32	10985 ns	11309 ns	0.97
array/random/rand/Int64	18397 ns	18964 ns	0.97
array/random/rand!/Float32	8230.666666666666 ns	8438.333333333334 ns	0.98
array/random/rand!/Int64	16192 ns	16530 ns	0.98
array/random/randn/Float32	34136 ns	34478 ns	0.99
array/random/randn!/Float32	24998 ns	25061 ns	1.00
array/reductions/mapreduce/Float32/1d	50008 ns	50429 ns	0.99
array/reductions/mapreduce/Float32/dims=1	44093 ns	43107 ns	1.02
array/reductions/mapreduce/Float32/dims=1L	74118 ns	72540 ns	1.02
array/reductions/mapreduce/Float32/dims=2	46005 ns	44886 ns	1.02
array/reductions/mapreduce/Float32/dims=2L	74633 ns	73143 ns	1.02
array/reductions/mapreduce/Int64/1d	51444 ns	51687 ns	1.00
array/reductions/mapreduce/Int64/dims=1	46127 ns	45426 ns	1.02
array/reductions/mapreduce/Int64/dims=1L	113606 ns	112852 ns	1.01
array/reductions/mapreduce/Int64/dims=2	46287 ns	46742 ns	0.99
array/reductions/mapreduce/Int64/dims=2L	101842 ns	100935 ns	1.01
array/reductions/reduce/Float32/1d	50776 ns	50222 ns	1.01
array/reductions/reduce/Float32/dims=1	44427 ns	43415 ns	1.02
array/reductions/reduce/Float32/dims=1L	74170 ns	73141 ns	1.01
array/reductions/reduce/Float32/dims=2	46199 ns	44957 ns	1.03
array/reductions/reduce/Float32/dims=2L	75139 ns	73558 ns	1.02
array/reductions/reduce/Int64/1d	51973 ns	52140 ns	1.00
array/reductions/reduce/Int64/dims=1	46913 ns	45836 ns	1.02
array/reductions/reduce/Int64/dims=1L	113548 ns	113326 ns	1.00
array/reductions/reduce/Int64/dims=2	46644 ns	46418 ns	1.00
array/reductions/reduce/Int64/dims=2L	101461 ns	100281 ns	1.01
array/reverse/1d	23408 ns	23893 ns	0.98
array/reverse/1dL	129738 ns	129747 ns	1.00
array/reverse/1dL_inplace	71114 ns	71493 ns	0.99
array/reverse/1d_inplace	13045 ns	13317 ns	0.98
array/reverse/2d	23325 ns	23517 ns	0.99
array/reverse/2dL	81828 ns	82043 ns	1.00
array/reverse/2dL_inplace	71695 ns	71810 ns	1.00
array/reverse/2d_inplace	15628 ns	15843 ns	0.99
array/sorting/1d	753167 ns	751635 ns	1.00
array/sorting/2d	195755 ns	195077 ns	1.00
array/sorting/by	932460 ns	928868 ns	1.00
cuda/graph/capture	42257 ns	41925 ns	1.01
cuda/graph/eager	39871 ns	42248 ns	0.94
cuda/graph/launch	13923 ns	14019 ns	0.99
cuda/synchronization/context/auto	7932 ns	7997.75 ns	0.99
cuda/synchronization/context/blocking	908.8235294117648 ns	895.8 ns	1.01
cuda/synchronization/context/nonblocking	7893 ns	7974.75 ns	0.99
cuda/synchronization/stream/auto	832.8 ns	886.9019607843137 ns	0.94
cuda/synchronization/stream/blocking	766.2477876106195 ns	780.3818181818182 ns	0.98
cuda/synchronization/stream/nonblocking	7836.333333333333 ns	7939.75 ns	0.99
integration/byval/reference	147631 ns	147881 ns	1.00
integration/byval/slices=1	148637 ns	148971 ns	1.00
integration/byval/slices=2	290364 ns	290729 ns	1.00
integration/byval/slices=3	432045 ns	432610 ns	1.00
integration/cudadevrt	105159 ns	105333 ns	1.00
integration/volumerhs	9609694 ns	9609784 ns	1.00
kernel/indexing	11564 ns	11657 ns	0.99
kernel/indexing_checked	12215 ns	12194 ns	1.00
kernel/launch	2256.222222222222 ns	2508 ns	0.90
kernel/occupancy	904.953488372093 ns	878.25 ns	1.03
kernel/rand	15171 ns	15391 ns	0.99
latency/import	4769772213 ns	4759695773 ns	1.00
latency/precompile	5396089777 ns	5390889215 ns	1.00
latency/ttfp	5249709853 ns	5247400815 ns	1.00

This comment was automatically generated by workflow using github-action-benchmark.