Rework pmap to use dynamic P0 and P1 region allocation, instead of allocating

the largest possible page table for every pmap; from NetBSD. This allows the
kernel to use much less memory for page tables.

Significant differences against the NetBSD code are:
- allocation of page table pages is done with a pool instead of allocating
  whole logical pages from uvm and managing the freelist within pmap, never
  releasing allocated pages.
- try to use pt_entry_t * rather than int * whenever possible.
- growth of P0 and P1 regions is allowed to fail, if invoked from
  pmap_enter with the PMAP_CANFAIL flag. This will stall processes until
  memory for the page tables can be obtained, rather than panicing, in
  most cases.
- keep management of mappings for managed pages using pv lists tied to the
  vm_page (using __HAVE_VM_PAGE_MD), rather than a global pv_list head.
- bound check against Sysmap[] in pmap_extract() when asked for a kernel
  address.

As a result of this, bsd.rd can now install a working system on a 12MB machine
without needing to enable swap.

0 files changed, 0 insertions, 0 deletions