RFC: Page size on PPC/PPC64 builders

[John Reiser]

Lubomir Kundrak wrote:
> If we going the way of not doing things that do expose bugs, we would
> not ever randomize library load addresses, or use canaries on stack, or
> check lengths of str*() arguments, etc... I would not be happy with
> this :)

I agree that doing those things has been fruitful in some ways.
However, the changes have sacrificed legal and useful properties.
For instance, randomizing library load addresses has decreased the usable
contiguous address space on x86, causing pain for applications which
want to use all of what formerly was available.  Similarly, placing the vdso
in the "middle" of nowhere fragments the address space without recourse.

> [snip]
> I never knew this [64KB pages] is possible in x86_64 :)
> Is there a strong reason against 64k pages there? :)

There are only 32768 pages of size 64KB in 2GB of RAM.  Fragmentation
can cost a *lot*.  Just when 2GB (or 4GB, etc.) of physical RAM becomes
inexpensive, the advantage of using it *all* gets denied by thoughts
that 64KB is "nothing."  Systematic use of "nothing" often costs
much more than expected.

--