On Tue, 12 Sep 2023 21:25:24 +0800
Yu Kuai [off-list ref] wrote:

Hi,

在 2023/09/12 21:01, Mariusz Tkaczyk 写道:

quoted

During working on changes proposed by Kuai [1], I determined that
mddev->active is continusly decremented for array marked by MD_CLOSING.
It brought me to md_seq_next() changed by [2]. I determined the regression
here, if mddev_get() fails we updated mddev pointer and as a result we
_put failed device.  

This mddev is decremented while there is another mddev increased, that's
why AceLan said that single array can't reporduce the problem.

And because mddev->active is leaked, then del_gendisk() will never be
called for the mddev while closing the array, that's why user will
always see this array, cause infiniate loop open -> stop array -> close
for systemd-shutdown.

Ohh, I see the scenario now...
First array can be successfully stopped. We marked MD_DELETED and proceed
with scheduling wq but in the middle of that md_seq_next() increased active for
other array and decrement active for the one with MD_DELETED.
For this next array we are unable to reach active == 0 anymore.

Song, let me know if you need description like that in commit message.

Thanks!
Mariusz