RE: Random high CPU utilization in blk-mq with the none scheduler
From: Dexuan Cui <decui@microsoft.com>
Date: 2021-12-11 03:44:47
Also in:
lkml
From: Dexuan Cui <decui@microsoft.com>
Date: 2021-12-11 03:44:47
Also in:
lkml
From: Jens Axboe <axboe@kernel.dk> Just out of curiosity, can you do: # perf record -a -g -- sleep 3 when you see the excessive CPU usage, then attach the output of # perf report -g to a reply?
I ran the commands against the 5.15.7 kernel and generated a 771-MB file. After I ran 'bzip2 -9', the file size is still 22 MB, which I guess is too big to be shared via the list. I'll try to put it somewhere and send you a link.
How confident are you in your bisect result? -- Jens Axboe
I'm pretty confident:
1) I can't reproduce the issue with v5.16-rc4 even if I run the test
10 times. Typically the issue repros every time.
2) If I revert the commit
dc5fc361d891 ("block: attempt direct issue of plug list")
and the related patches
ff1552232b36 ("blk-mq: don't issue request directly in case that current is to be blocked")
b22809092c70 ("block: replace always false argument with 'false'")
from v5.16-rc4, I'm able to repro the issue immediately.
Thanks,
Dexuan