The inode eviction can be very slow, because during eviction we
tell the VFS to truncate all of the inode's pages. This results
in calls to btrfs_invalidatepage() which in turn does calls to
lock_extent_bits() and clear_extent_bit(). These calls result in
too many merges and splits of extent_state structures, which
consume a lot of time and cpu when the inode has many pages. In
some scenarios I have experienced umount times higher than 15
minutes, even when there's no pending IO (after a btrfs fs sync).

A quick way to reproduce this issue:

$ mkfs.btrfs -f /dev/sdb3
$ mount /dev/sdb3 /mnt/btrfs
$ cd /mnt/btrfs
$ sysbench --test=fileio --file-num=128 --file-total-size=16G \
    --file-test-mode=seqwr --num-threads=128 \
    --file-block-size=16384 --max-time=60 --max-requests=0 run
$ time btrfs fi sync .
FSSync '.'

real	0m25.457s
user	0m0.000s
sys	0m0.092s
$ cd ..
$ time umount /mnt/btrfs

real	1m38.234s
user	0m0.000s
sys	1m25.760s

The same test on ext4 runs much faster:

$ mkfs.ext4 /dev/sdb3
$ mount /dev/sdb3 /mnt/ext4
$ cd /mnt/ext4
$ sysbench --test=fileio --file-num=128 --file-total-size=16G \
    --file-test-mode=seqwr --num-threads=128 \
    --file-block-size=16384 --max-time=60 --max-requests=0 run
$ sync
$ cd ..
$ time umount /mnt/ext4

real	0m3.626s
user	0m0.004s
sys	0m3.012s

After this patch, the unmount (inode evictions) is much faster:

$ mkfs.btrfs -f /dev/sdb3
$ mount /dev/sdb3 /mnt/btrfs
$ cd /mnt/btrfs
$ sysbench --test=fileio --file-num=128 --file-total-size=16G \
    --file-test-mode=seqwr --num-threads=128 \
    --file-block-size=16384 --max-time=60 --max-requests=0 run
$ time btrfs fi sync .
FSSync '.'

real	0m26.774s
user	0m0.000s
sys	0m0.084s
$ cd ..
$ time umount /mnt/btrfs

real	0m1.811s
user	0m0.000s
sys	0m1.564s
Signed-off-by: Filipe David Borba Manana <fdmanana@gmail.com>
Signed-off-by: Josef Bacik <jbacik@fb.com>
Signed-off-by: Chris Mason <clm@fb.com>