The following command line will leave the aio-stress process unkillable
on an ext4 file system (in my case, mounted on /mnt/test):

aio-stress -t 20 -s 10 -O -S -o 2 -I 1000 /mnt/test/aiostress.3561.4 /mnt/test/aiostress.3561.4.20 /mnt/test/aiostress.3561.4.19 /mnt/test/aiostress.3561.4.18 /mnt/test/aiostress.3561.4.17 /mnt/test/aiostress.3561.4.16 /mnt/test/aiostress.3561.4.15 /mnt/test/aiostress.3561.4.14 /mnt/test/aiostress.3561.4.13 /mnt/test/aiostress.3561.4.12 /mnt/test/aiostress.3561.4.11 /mnt/test/aiostress.3561.4.10 /mnt/test/aiostress.3561.4.9 /mnt/test/aiostress.3561.4.8 /mnt/test/aiostress.3561.4.7 /mnt/test/aiostress.3561.4.6 /mnt/test/aiostress.3561.4.5 /mnt/test/aiostress.3561.4.4 /mnt/test/aiostress.3561.4.3 /mnt/test/aiostress.3561.4.2

This is using the aio-stress program from the xfstests test suite.
That particular command line tells aio-stress to do random writes to
20 files from 20 threads (one thread per file).  The files are NOT
preallocated, so you will get writes to random offsets within the
file, thus creating holes and extending i_size.  It also opens the
file with O_DIRECT and O_SYNC.

On to the problem.  When an I/O requires unwritten extent conversion,
it is queued onto the completed_io_list for the ext4 inode.  Two code
paths will pull work items from this list.  The first is the
ext4_end_io_work routine, and the second is ext4_flush_completed_IO,
which is called via the fsync path (and O_SYNC handling, as well).
There are two issues I've found in these code paths.  First, if the
fsync path beats the work routine to a particular I/O, the work
routine will free the io_end structure!  It does not take into account
the fact that the io_end may still be in use by the fsync path.  I've
fixed this issue by adding yet another IO_END flag, indicating that
the io_end is being processed by the fsync path.

The second problem is that the work routine will make an assignment to
io->flag outside of the lock.  I have witnessed this result in a hang
at umount.  Moving the flag setting inside the lock resolved that
problem.

The problem was introduced by commit b82e384c ("ext4: optimize
locking for end_io extent conversion"), which first appeared in 3.2.
As such, the fix should be backported to that release (probably along
with the unwritten extent conversion race fix).
Signed-off-by: Jeff Moyer <jmoyer@redhat.com>
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
CC: stable@kernel.org