[riscv64] test_dataloader.py fails to exit normally on k1

[yuzibo]

🐛 Describe the bug

Recently CI workflow on K1, the run_test.py can not exit normal and can be reproduced always:

 root@RwdUeLilRZ130:~# ps -ef | grep run_test.py
jenkins  2615865 2615592  0 10:14 ?        00:00:19 python test/run_test.py --keep-going --shard 25 48
root     2825873 2711147  0 16:17 pts/0    00:00:00 grep --color=auto run_test.py

root@RwdUeLilRZ130:~# pstree -ap 2615865
python,2615865 test/run_test.py --keep-going --shard 25 48
  ├─python,2627633 -c from multiprocessing.resource_tracker import main;main(5)
  ├─python,2637385 -c from multiprocessing.spawn import spawn_main; spawn_main(tracker_fd=6, pipe_handle=14) --multiprocessing-fork
  │   ├─python,2637612 -bb test_dataloader.py -m not serial --shard-id=1 --num-shards=2 -vv -rfEX -p no:xdist --use-pytest -x --reruns=2 --sc=test_dataloader_1_4030aef120e04706 --print-items
  │   │   └─python,2638252 -bb -c from multiprocessing.resource_tracker import main;main(7)
  │   ├─{python},2637397
  │   ├─{python},2637398
  │   ├─{python},2637399
  │   ├─{python},2637400
  │   ├─{python},2637401
  │   ├─{python},2637402
  │   └─{python},2637403
  ├─python,2672092 -c from multiprocessing.spawn import spawn_main; spawn_main(tracker_fd=6, pipe_handle=14) --multiprocessing-fork
  │   ├─{python},2672106
  │   ├─{python},2672107
  │   ├─{python},2672108
  │   ├─{python},2672109
  │   ├─{python},2672110
  │   ├─{python},2672111
  │   └─{python},2672112
  ├─python,2672413 -c from multiprocessing.spawn import spawn_main; spawn_main(tracker_fd=6, pipe_handle=14) --multiprocessing-fork
  │   ├─{python},2672424
  │   ├─{python},2672425
  │   ├─{python},2672426
  │   ├─{python},2672427
  │   ├─{python},2672428
  │   ├─{python},2672429
  │   └─{python},2672430
  ├─{python},2615877
  ├─{python},2615878
  ├─{python},2615879
  ├─{python},2615880
  ├─{python},2615881
  ├─{python},2615882
  ├─{python},2615883
  ├─{python},2627637
  ├─{python},2627638
  └─{python},2627639

This will cause CI to be endless.

The log is here.

Versions

Nope, but will update here later