Micro-optimize kmem_unback().

[markj]

Summary

As with vm_page_grab_pages(), we can avoid a radix tree lookup on each
page in the range by using the object memq.

[kib]

sys/vm/vm_kern.c

400 ↗

(On Diff #31842)

for (; pindex++ < end; m = next) {

[alc]

To be clear, I'm happy with the concept, and in general eliminating vm_page_lookup() calls inside of loops.

sys/vm/vm_kern.c

390 ↗

(On Diff #31842)

Switching from vm_offset_t to vm_pindex_t is a pessimization for 32-bit architectures because it forces them to perform 64-bit arithmetic where before they used 32-bit arithmetic.

401 ↗

(On Diff #31842)

Just to be clear, the reason to use vm_page_next() rather than TAILQ_NEXT() here is that it will preserve the current behavior of panic'ing on a NULL pointer dereference if a page is missing.

[markj]

• Let kmem_unback() use vm_offset_t instead of vm_pindex_t
• Use a for-loop

[markj]

sys/vm/vm_kern.c

401 ↗

(On Diff #31842)

Right. I had originally written

next = TAILQ_NEXT(m, listq);
MPASS(next->pindex == pindex);

and then I remembered that vm_page_next() exists.