Re: page_launder() bug

[David S. Miller]

Linus Torvalds writes:
 > Maybe it's academic. Do we know that any of this actually makes any
 > performance difference at all?

We know that dirty swap pages can accumulate to the point where the
swapper starves before it gets to enough of the "second pass" cases of
the page_launder loop to run in order to get rid of them.

That was the behavior I saw that let me to hacking up my broken patch.

Even simple "memory hog" test programs can trigger this behavior.
Just write a program which "grows almost completely into swap then
subsides" over and over again.

Later,
David S. Miller
[e-mail]