Speed up iteration in vm_object_madvise().

[markj]

Summary

Some tracing indicates that vm_object_madvise(MADV_FREE) is frequently
invoked by jemalloc on anonymous objects with no backing object. With
r311346 this occurs during each execve as well. In this case we can
iterate over the range more quickly by using the object's memq rather
than performing a radix tree lookup on each pindex in the range.
Moreover, we can skip over non-resident subranges in a single iteration.
This case occurs frequently with both execve and jemalloc - firefox, for
instance, appears to frequently purge arenas using mallctlbymib(3), and
the corresponding VA ranges are not always backed by resident pages.

While here, rename "advise" to "advice" for consistency with related
functions and because it makes more sense, and move MADV_WILLNEED
handling into vm_page_advise() (vm_object_madvise() is its only caller).

[kib]

sys/vm/vm_object.c

1148 ↗

(On Diff #23752)

If you write these lines as

} else if ((m = vm_page_lookup(tobject, tpindex)) == NULL) {

you can keep existing indent (reduce indent in the patch).

[markj]

Reduce indentation.

[alc]

sys/vm/vm_object.c

1145 ↗

(On Diff #23813)

Isn't the assignment to "tpindex" redundant?

[markj]

sys/vm/vm_object.c

1145 ↗

(On Diff #23813)

Yes, you're right. It made sense in an earlier revision where I didn't set tpindex until this point - we must have tpindex = m->pindex in order for the swap_pager_freespace() call at the end of the loop to be correct.

I'll change it to "pindex = tpindex".

[markj]

Remove a redundant assignment.

[alc]

This change looks okay. Go ahead and commit it. After that, please take a look at how vm_object_unwire() is written. The same approach might be applicable here.

[alc]

To be clear, my point is that I think that you can always avoid the vm_page_lookup() on the top-level object, and not just in the case that there is no backing object.

[markj]

In D9098#189866, @alc wrote:

To be clear, my point is that I think that you can always avoid the vm_page_lookup() on the top-level object, and not just in the case that there is no backing object.

Ok, I see. I'll give this a try.