btrfs: backref: Fix soft lockup in __merge_refs function

When over 1000 file extents refers to one extent, find_parent_nodes()
will be obviously slow, due to the O(n^2)~O(n^3) loops inside
__merge_refs().

The following ftrace shows the cubic growth of execution time:

256 refs
 5) + 91.768 us   |  __add_keyed_refs.isra.12 [btrfs]();
 5)   1.447 us    |  __add_missing_keys.isra.13 [btrfs]();
 5) ! 114.544 us  |  __merge_refs [btrfs]();
 5) ! 136.399 us  |  __merge_refs [btrfs]();

512 refs
 6) ! 279.859 us  |  __add_keyed_refs.isra.12 [btrfs]();
 6)   3.164 us    |  __add_missing_keys.isra.13 [btrfs]();
 6) ! 442.498 us  |  __merge_refs [btrfs]();
 6) # 2091.073 us |  __merge_refs [btrfs]();

and 1024 refs
 7) ! 368.683 us  |  __add_keyed_refs.isra.12 [btrfs]();
 7)   4.810 us    |  __add_missing_keys.isra.13 [btrfs]();
 7) # 2043.428 us |  __merge_refs [btrfs]();
 7) * 18964.23 us |  __merge_refs [btrfs]();

And sort them into the following char:
(Unit: us)
------------------------------------------------------------------------
 Trace function        | 256 ref        | 512 refs      | 1024 refs    |
------------------------------------------------------------------------
 __add_keyed_refs      | 91             | 249           | 368          |
 __add_missing_keys    | 1              | 3             | 4            |
 __merge_refs 1st call | 114            | 442           | 2043         |
 __merge_refs 2nd call | 136            | 2091          | 18964        |
------------------------------------------------------------------------

We can see the that __add_keyed_refs() grows almost in linear behavior.
And __add_missing_keys() in this case doesn't change much or takes much
time.

While for the 1st __merge_refs() it's square growth
for the 2nd __merge_refs() call it's cubic growth.

It's no doubt that merge_refs() will take a long long time to execute if
the number of refs continues its grows.

So add a cond_resced() into the loop of __merge_refs().

Although this will solve the problem of soft lockup, we need to use the
new rb_tree based structure introduced by Lu Fengqi to really solve the
long execution time.

Signed-off-by: Qu Wenruo <quwenruo@cn.fujitsu.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
Signed-off-by: Chris Mason <clm@fb.com>