I've run the stress2 swap tests without incident.
With this test patch: {F97661278}, I've tested the time it take to do 1 million swp_page_meta_build calls (and a smaller number of swp_pager_getswapspace calls) with and without this change. The results, analyzed by ministat, are here:
```
x old
+ new
+------------------------------------------------------------------------------+
| + |
| ++ |
|+++ + x xx |
|+++ ++ xx xxxx x x x|
||MA_| |_____M__A________| |
+------------------------------------------------------------------------------+
N Min Max Median Avg Stddev
x 12 45506632 50173397 46252604 46751133 1558878.7
+ 12 36901703 37918690 37119202 37277605 376800.52
Difference at 95.0% confidence
-9.47353e+06 +/- 960197
-20.2637% +/- 1.66335%
(Student's t, pooled s = 1.13404e+06)
```