I think there's a mistake either in the implementation of 
uvm_page_physload() or in every caller of it in the tree.  After creating 
the vm_page object for a page, uvm_page_physload() does this:

                        if (atop(paddr) >= avail_start &&
                            atop(paddr) <= avail_end)
                                uvm_pagefree(&pgs[lcv]);


Similarly, uvm_page_init() does this:

                        if (atop(paddr) >= vm_physmem[lcv].avail_start &&
                            atop(paddr) <= vm_physmem[lcv].avail_end) {
                                uvmexp.npages++;
                                /* add page to free pool */
                                uvm_pagefree(&vm_physmem[lcv].pgs[i]);
                        }

Thus in both cases, "avail_end" is the page number of the last available 
page rather than the first unavailable page (which would be more 
conventional, and consistent with the "end" argument, which is the number 
of the first page above the segment being loaded).  Every caller of 
uvm_page_physload, though, passes the same value for "end" and 
"avail_end", so they clearly think they have the same semantics.  Would 
anyone object if I made uvm_page_physload behave accordingly (by changing 
two <=s to <)?

While I'm there, I'm rather suspicious of only counting initially-free 
pages in uvmexp.npages, but I probably ought to leave that for another 
time.

-- 
Ben Harris                                                   <bjh21@NetBSD.org>
Portmaster, NetBSD/acorn26           <URL:http://www.NetBSD.org/Ports/acorn26/>