Hugh Dickins wrote:

> It's clearly superior to what David and I had, in branching
> less (other than in your BUG_ONs), and I do believe your
> "if (end - ceiling - 1 < P*_SIZE - 1)" is correct and efficient.
> 

Well the BUG_ONs were more to just satisfy me that my assumptions
were correct. Not to mention only contained in the top level, so
they shouldn't hurt performance. But they could go.

> But I still find it harder to understand than ours; and don't
> understand at all your comment "end can't have approached ceiling
> from above...." - but I think you're bravely trying to explain the case
> I sidestepped with a lordly unexplained "end can't go down to 0 there".
> 

Yes, say ceiling is 0 - something less than P*_SIZE, you might
get the feeling that end may be able to come within our limit
of it if it were a very small number.

This can't happen because 0 is actually the top of address space,
and end can't be *greater* than ceiling before any rounding. If it
is not 0, then it must be at least 1, in which case it will always
be rounded up to the next P*_SIZE boundary. So no problem.

This may have been obvious to you from the start, in which case my
extra rambling may have confused you... actually on re-reading it,
it would have confused you no matter what. See if the next version
is better.

> Let others decide.
> 
> One thing I believe is outright wrong, at least with out-of-tree
> patches: your change from "if (addr > end - 1)" to "if (addr >= end)",
> after you've just rounded up end (perhaps to 0).
> 

Oh yes, good catch. I don't know why I did that :(

> (And let me astonish you by asking for the blank lines back before
> pmd_offset and pud_offset!)
> 

Hugh? What have you done with Hugh?