Re: [RFC] My research agenda for 2.7

[Martin J. Bligh]

>> > I also wonder if moving kernel pages is really worth the hassle.
>> 
>> That's the question of course.  The benefit is getting rid of high order
>> allocation failures, and gaining some confidence that larger filesystem
>> blocksizes will work reliably, however the workload evolves.

Oh, BTW ... I suspect you've realised this already, but ....

The buddy allocator is not a good system for getting rid of fragmentation. 
If I group pages together in aligned pairs, and F is free and A is 
allocated, it'll not do anything useful with this:

F A   A F   F A   A F   F A   A F   F A   A F   F A   F A 

because the adjacent "F"s aren't "buddies". It seems that the purpose of
the buddy allocator was to be quick at allocating pages. Now that we stuck
a front end cache on it, in the form of hot & cold pages, that goal no
longer seems paramount - altering it to reduce fragmentation at the source,
rather than actively defrag afterwards would seem like a good goal to me.

M.