Re: [patch 5/18] mark swapout pages PageWriteback()

[Linus Torvalds]

On Sun, 26 May 2002, Andrew Morton wrote:
>
> But I recall you saying that there was advantage in keeping swapout pages
> locked so that aggressive memory users would throttle against their
> own swapout.  What's the story there?

The advantage is not the lock itself, as much as having people who page in
swap pages be delayed on them - which ends up slowing down processes that
swap a lot.

BUT: that could equally well be done by doing a "wait_on_writeback()" or
similar, and it could also be a tunable thing (ie wait on writeback only
when we actually need to slow them down). In particular, _not_ slowing
them down does improve throughput, it just makes it really really nasty
from an interactive standpoint under some loads.

I don't know. I have this feeling that it would be good to try to share
all the semantics between swap pages and shared file mappings, but at the
same time I also have to admit to believing that swap _is_ special in some
ways, so if we don't ever really unify them I won't be shedding any huge
tears.

		Linus