K 10
svn:author
V 3
alc
K 8
svn:date
V 27
2012-06-14T16:53:08.879729Z
K 7
svn:log
V 962
Speed up the destruction of superpage mappings on amd64 by eliminating the
frequent need for pmap_remove_pages() and pmap_remove_pde() to access each
struct vm_page that makes up a superpage.  This is accomplished through a
combination of two changes:

1) Replace PGA_WRITEABLE with an exact count of write mappings at both the
   4KB and 2MB granularity so that pmap_remove_pages() and pmap_remove_pde()
   must no longer access each struct vm_page in order to clear
   PGA_WRITEABLE.  (These counts are stored in space that was previously lost
   to alignment in struct md_page, so struct vm_page is no larger.)

2) Call vm_page_dirty() during promotion of write mappings rather than
   pmap_remove_pages() and pmap_remove_pde().  Each struct vm_page is
   already accessed on promotion, so adding a call to vm_page_dirty()
   shouldn't result in additional cache misses.

For "buildworld", this reduces the average running time of
pmap_remove_pages() by 23%.

END