Comment # 18 on bug 1021307 from Jan Kara

As Mel says gathering 'iostat -x 5' and seeing the numbers around stall time is
a good idea. We'll get idea how much the disk is busy around that time. You can
check which IO scheduler you are using in /sys/block/<device>/queue/scheduler
however on 42.2 for non-rotating disk it should already be set to 'deadline'.
But it doesn't hurt to verify it on your machine.

Finally, I'm also adding David to CC as a btrfs guy. Maybe he has an idea why
fsync(2) can take 5 minutes for the root btrfs filesystem.