vm_phys: pick the largest phys_avail segment of the requested domain

[wanpengqian_gmail.com]

Summary

vm_phys_early_alloc(domain, alloc_size) restricts its search of
phys_avail to the single largest SRAT range of the domain (or, for
domain -1, the largest range of all). When phys_avail has little or
nothing inside that range the search fails and the kernel panics early
in boot with "cannot find a large enough size", although other ranges
of the same domain may have plenty of free memory.

How we found it: keelOS's hypervisor keeps most of the RAM for its
guest pool and gives its FreeBSD root partition a few GB of it. On a
two-node test bed (12 GB, 6 GB per node, 2 GB given to the root, all of
it below 4 GB) every boot panicked in pmap_page_array_startup() ->
vm_phys_early_alloc(): node 0's largest SRAT range is 4G-8G and
phys_avail has nothing there. Stock FreeBSD reaches the same state on
a NUMA machine when hw.physmem limits the usable memory to the low part
of node 0 (the test plan reproduces it in QEMU that way), and in general
whenever the firmware or a hypervisor presents a node whose largest
range is mostly unusable.

Drop the SRAT range pre-selection and pick the largest phys_avail
segment that belongs to the requested domain. vm_phys_early_startup()
has already split phys_avail at the domain boundaries, so every segment
lies in one domain; it is looked up by its last byte because
vm_phys_domain() treats a range's end as inclusive and a segment may
start exactly where the previous domain ends.

Test Plan

Stock QEMU (KVM), FreeBSD main GENERIC, 8 vCPUs, 12 GB. Unpatched = main of 2026-10-02 (vm_phys_early_alloc() unchanged since); patched = main 6d705f2bf + this change.

1. Two nodes of 6 GB, hw.physmem="4100M". Node 0's SRAT ranges are 0-2G and 4G-8G, so its largest range is 4G-8G, and hw.physmem leaves only a few MB of phys_avail in it (node 1 starts at 8G, above hw.physmem, and srat.c ignores it).
   • Unpatched: every boot panics while the vm_page array is laid out: panic: cannot find a large enough size vm_phys_early_alloc() pmap_page_array_startup() vm_page_startup() vm_mem_init() mi_startup()
   • Patched: boots; vm.ndomains 2, hw.physmem 2105053184, the last vm.phys_segs is 0x100000000 (domain 0).

2. Two nodes of 6 GB, no hw.physmem: patched boots, vm.phys_segs above 4G is identical to the unpatched kernel's (0x100001000 domain 0, 0x200000000 and 0x37fc00000 domain 1), free pages domain 0 1478102 / domain 1 1511183.

3. Four nodes of 3 GB, hw.physmem="4100M" (domains 1-3 have no memory): patched boots, vm.ndomains 4.

[kib]

Then why search the largest address range at all? Would it be more reasonable to change the function to search for largest phys_avail in the requested domain?

[wanpengqian_gmail.com]

Pick the largest phys_avail segment of the requested domain instead of the largest SRAT range, as suggested by kib; the fallback is gone.