Re: [PATCH] ext4: fix potential deadlock in ext4_nonda_switch()

[Theodore Ts'o <tytso@mit.edu>]

Ah, you're right.  I *can* easily see the problem; somehow I just
didn't notice it, but when I look back at my test logs, it's
definitely there.  I thought sb_start_write() took s_umount, but it
doesn't, because it was trying to solve the exact same lock ordering
problem.  Drat....

Unfortunately, there doesn't seem to be a good solution here.  The two
options I see is either (a) create another version of
writeback_inodes_sb() which doesn't have the WARN_ON check in
fs/fs-writeback.c (the warning is a false positive, since there are
other mechanisms which protect the superblock from being unmounted
while the write system call is in progress), or (b) call
writeback_inodes_sb() out of a workqueue.

		      	       - Ted