More 2.2.17pre9 VM issues

[Marcelo Tosatti]

> There's no stability issue.
> 
> But the current fix is suboptimal. Suppose we have somebody that is
> freeing memory in try_to_free_pages (so free_before_allocate is 1). Then
> one task with a RSS of 1giga exits and release 1giga of memory. If
> somebody will try to allocate memory at this point he will have to block
> too, while there's 1 giga of memory free. Avoiding the block if
> nr_free_pages > freepages.high as in 2.2.16 disables the logic so that's
> not a viable solution.

I see. 
Do you have any possible solution in mind for that problem? 

Another thing which I would like to receive comments from you and other
people is my patch to fix the IO completion wait problem at: 

http://bazar.conectiva.com.br/~marcelo/patches/v2.2/2.2.17pre9/wait-dirty.patch

Thanks