Re: [PATCH] cfq: clear queue pointers from cfqg after unpinning them in cfq_pd_offline

On 17.08.2018 19:30, Paolo Valente wrote:

quoted

Il giorno 17 ago 2018, alle ore 19:28, Maciej S. Szmigiero [off-list ref] ha scritto:

The current linux-block, 4.18 and 4.17 can reliably be crashed within few
minutes by running the following bash snippet:

mkfs.ext4 -v /dev/sda3 && mount /dev/sda3 /mnt/test/ -t ext4;
while true; do
	mkdir /sys/fs/cgroup/unified/test/;
	echo $$ >/sys/fs/cgroup/unified/test/cgroup.procs;
	dd if=/dev/zero of=/mnt/test/test-$(( RANDOM * 10 / 32768 )) bs=1M count=1024 &
	echo $$ >/sys/fs/cgroup/unified/cgroup.procs;
	sleep 1;
	kill -KILL $!; wait $!;
	rmdir /sys/fs/cgroup/unified/test;
done

# cat /sys/block/sda/queue/scheduler
noop [cfq]
# cat /sys/block/sda/queue/rotational
1
# cat /sys/fs/cgroup/unified/cgroup.subtree_control
cpu io memory pids

The backtraces vary but often they are NULL pointer dereferences due to
various cfqq fields being NULL.
Or BUG_ON(cfqq->ref <= 0) in cfq_put_queue() triggers due to cfqq reference
count being zero.

Bisection points at
commit 4c6994806f70 ("blk-throttle: fix race between blkcg_bio_issue_check() and cgroup_rmdir()").
The prime suspect looked like .pd_offline_fn() method being called multiple
times, but from analyzing the mentioned commit this didn't seem possible
and runtime trials have confirmed that.

However, CFQ's cfq_pd_offline() implementation of the above method were
leaving queue pointers intact in cfqg after unpinning them.
After making sure that they are cleared to NULL in this function I can no
longer reproduce the crash.

By chance, did you check whether is BFQ is ok in this respect?

I wasn't able to crash BFQ with the above test and in fact had run my
machines on BFQ until I was able to find a fix for this in CFQ.

Also, BFQ has a bit similar code in bfq_put_async_queues() that is
called from bfq_pd_offline() that is already NULL-ing the passed
pointer.

Thanks,
Paolo

Regards,
Maciej