vm_page: Retire its listq fieldHEADmain

Over the life cycle of a vm_page, its listq field has been used for two
distinct purposes.  First, linking together all of the pages allocated
to a vm_object.  Recently, c8d56817b80f ("vm_object: drop memq field")
completed the elimination of this use case, using pctrie iterators in
place of iteration over the listq.  Second, linking together power-of-
two-sized chunks of free pages within vm_phys.  This change eliminates
that use case.  In essence, this change reverts vm_phys back to using
the plinks.q field, like it did before 5cd29d0f3cda ("Improve VM page
queue scalability."), but with a twist to maintain scalability.  Just
before vm_phys uses the plinks.q field, it ensures that any lazy dequeue
from a paging queue, e.g., PQ_ACTIVE, has completed.  Typically, the
dequeue has completed, so vm_page_dequeue() is infrequently called by
vm_freelist_add().  The reason being that vm_phys only needs to use the
plinks.q field within the first page of any power-of-two-sized chunk,
so the rest of the pages can still have pending dequeues until the
chunk is split.

This change saves a non-trivial amount of memory, since we have an
instance of struct vm_page for every dynamically allocatable physical
page.

Bump __FreeBSD_version, since third-party modules that use the inline
accessors in vm_page.h may need to be recompiled.

Reviewed by:	dougm, kib, markj
Differential Revision:	https://reviews.freebsd.org/D50515

0 files changed, 0 insertions, 0 deletions