Re: [RFC][PATCH 0/3] beat kswapd with the proverbial clue-bat

[Martin J. Bligh]

>> Hmm, and the crowning argument for not stopping at order 3 is that if we
>> never use higher order allocations, nothing will care about their watermarks
>> anyway. I think I had myself confused when that question in the first place.
>> 
>> So yeah, stopping at a fixed number isn't required, and as you say it keeps
>> things general and special cases minimal.
> 
> Hey, please refute my "you need 20% free" to get even to order-3 for most
> cases first.
> 
> It's probably acceptable to have a _very_ backgrounded job that does
> freeing if order-3 isn't available, but it had better be pretty
> slow-moving, I suspect. On the order of "It's probably ok to try to aim
> for up to 25% free 'overnight' if the machine is idle" but it's almost
> certainly not ok to aggressively push things out to that degree..

IIRC, the way we free pages is basically random shootdown. We don't even 
attempt to free pages in a contiguous block, so it's unsuprising it doesn't
work. Now we have rmap, we should be able to walk through a block, see if
everything in it looks swappable, and then try to shoot it down. If not,
move onto the next one. I think that'd be a damned sight more likely
to free things up for higher order allocs if they do happen.

M.

-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/