Keep the vm_page array mapped into KVA on amd64.

[markj]

Summary

r351198 allows the kernel to use domain-local memory to back the vm_page
array (up to 2MB boundaries) and reserves a separate PML4 entry for that
purpose. One consequence of that change is that the vm_page array is no
longer present in minidumps, since the minidump code only adds pages
mapped above VM_MIN_KERNEL_ADDRESS.

To avoid the friction caused by having kernel data structures mapped
below VM_MIN_KERNEL_ADDRESS, map the vm_page array starting at
VM_MIN_KERNEL_ADDRESS on amd64. The main change is that we have to
remove the mapping from kernel_map in kmem_init().

Test Plan

I booted in a 1GB VM, a 2-socket system with 64GB of RAM, and pig1 (4 sockets
and 1TB of RAM), and verified that kgdb can access the page array.

[markj]

sys/amd64/amd64/pmap.c

3855 ↗

(On Diff #61561)

I think we can do better and use 1GB pages when possible, which is when we have domains with >= ~40GB.

sys/vm/vm_kern.c

777 ↗

(On Diff #61561)

The roundup is because pmap_page_array_startup() always uses 2MB pages, but of course the vm_page array might not be a multiple of 2MB in size. KVA_QUANTUM isn't really the right constant, but I'm not sure what else to use. VM_LEVEL_0_ORDER is not defined on some platforms which might use large pages to map the array, like powerpc.

[markj]

Fix the !PMAP_HAS_PAGE_ARRAY build.

[kib]

sys/amd64/include/vmparam.h

167 ↗

(On Diff #61564)

I think it worth noting where the vm_page_array is located.

sys/vm/vm_kern.c

771 ↗

(On Diff #61564)

Does ppc need this space reservation ? I suspect it is more correct to use ifdef amd64 there.

775 ↗

(On Diff #61564)

Extra space between ) and 0. That said, do you need to cast 0 ?

[markj]

sys/vm/vm_kern.c

771 ↗

(On Diff #61564)

I think you're right, powerpc will adjust virtual_avail.

775 ↗

(On Diff #61564)

Blindly copied and pasted from above, will fix.

[markj]

Updates from kib's feedback.

[markj]

Use round_2mpage() instead of rounding up to KVA_QUANTUM, now that
the kmem_init() change is amd64-only.

[alc]

I believe that PA_MIN_ADDRESS can be removed, as it was introduced by the original version of this code.

head/sys/amd64/amd64/pmap.c

3853

ctob(pfn) -> ptoa(pfn)

ctob() is archaic, and never used in Mach-derived VM code. For example, it appears nowhere else in this file or under vm/ except in some stack growth code.

3858

-> pagezero((void *)PHYS_TO_DMAP(pa));