Re: [RFC] Generalize prio_tree (1/3)

From: Werner Almesberger
Date: Mon Nov 15 2004 - 18:09:51 EST


Rajesh Venkatasubramanian wrote:
> No, that's not possible. Whenever we change vm_pgoff, vm_start, or
> vm_end, we reshuffle the prio_tree. Check mm/mmap.c:vma_adjust().

Excellent, that makes it a lot easier.

> If you are thinking vm_set.head or vm_set.parent is free to use
> for h_index, then it is incorrect. No field is free. If a vma
> is a tree node (and in this discussion we care only about tree
> nodes), then every field in vma->shared is used. We cannot reuse
> them for storing h_index.

Oh, I see. I thought it was that either vm_set or prio_tree_node
was used. (Which I found somewhat confusing, but attributed my
confusion to simply not understanding all the strange ways of MM.)
Sorry about the confusion.

So yes, one more word then :-(

> If we impose that there can be only 2 types of prio_tree, then
> we can simply add an if-else condition in the GET_INDEX macro.
> IMHO, that will not lead to any noticeable performance drop.

Yes, that sounds better. It would also allow for a later transition
of VMA_PRIO_TREE to GENERIC_PRIO_TREE.

Now, if we want to prepare things already now for a later migration,
it would be nice to call the generic one "struct prio_tree_node",
since it would eventually become the only node type anyway.

Perhaps something along these lines:

struct prio_tree_node {
struct vma_prio_tree_node prio_tree_node;
unsigned long r_index, h_index;
};

Or would you consider this as premature optimization ?

> But, I agree with you that changing the layout of vm_area_struct
> is a better (but intrusive) approach.

I also wonder how expensive the calculations in HEAP_INDEX are.
Probably not very.

To make the intrusive change a bit more palatable, vm_pgoff could
become #define vm_pgoff(vma) ((vma)->shared.prio_tree_node.r_index)

Okay, so now we only need someone who has meaningful MM tests to
tell us how badly this would hurt performance :-)

Thanks,
- Werner

--
_________________________________________________________________________
/ Werner Almesberger, Buenos Aires, Argentina werner@xxxxxxxxxxxxxxx /
/_http://www.almesberger.net/____________________________________________/
-
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/