On Fri, Apr 16, 2021 at 04:00:48PM +0800, Zhang Yi wrote:

Now, we use already use "if (bdev->bd_super)" to prevent call into
->bdev_try_to_free_page unless the super is alive, and the problem is
bd_super becomes NULL concurrently after this check. So, IIUC, I think it's
the same to switch to check the superblock is active or not. The acvive
flag also could becomes inactive (raced by umount) after we call into
bdev_try_to_free_page().

Indeed.

In order to close this race, One solution is introduce a lock to synchronize
the active state between kill_block_super() and blkdev_releasepage(), but
the releasing page process have to try to acquire this lock in
blkdev_releasepage() for each page, and the umount process still need to wait
until the page release if some one invoke into ->bdev_try_to_free_page().
I think this solution may affect performace and is not a good way.
Think about it in depth, use percpu refcount seems have the smallest
performance effect on blkdev_releasepage().

If you don't like the refcount, maybe we could add synchronize_rcu_expedited()
in ext4_put_super(), it also could prevent this race. Any suggestions?

I really don't like to put a lot of overhead into the core VFS and block
device code.  ext4/jbd does not own the block device inode and really
has no business controlling releasepage for it.  I suspect the right
answer might be to simply revert the commit that added this hook.