This is a list of the changes to the vm system as of 14 Jan 94:

This is a list of the changes to the vm system as of 14 Jan 94:

1)	Upages are now in process address space.  They can be paged out when
	the process is swapped out.  The majority of changes are in
	vm_glue.c and pmap.c.  Upages are always recorded to be modified.

2)	Page table pages are not high level vm wired except to insert pages
	pte's into the pmap.  The pmap subsystem has a wiring mechanism of
	its own now, and releases page table pages to the vm system when the
	number of pages pointed to by the page table page drops to zero.
	This mechanism allows the physical memory for process page tables
	to be reclaimed for use by other processes during memory shortages.
	The majority of changes are in trap.c, pmap.c, kern__physio.c,
	procfs_subr.c.  Page table pages are never recorded to be modified.

3)	Page table pages are faulted in by referring to the page table page
	from kernel address space.  Temporary wiring of the page table page
	is performed by removing the page from the inactive and active
	queues.  This is much more efficient than many other techniques
	(explicit vm_fault's.)

4)	Pmap.c has major sections rewritten for efficiency.  String operations
	are used when possible.  Also, pmap_remove (called often) has
	been rewritten to be *much* more efficient.

5)	Pmap_attributes has been removed.  The side effect of storing the
	modified and used (referenced) bits is difficult to control.  Mods
	have been made to the rest of the vm system to not assume carryover
	of the modified bits.  It is believed that this fixes a problem
	where pages were often written to the backing store unnecessarily.

6)	Pageout queue management is *much* more efficient:

6a)	Pages that are deactivated are removed from pmaps.  This allows the
	system to fault immediately when a page is used again.  Thereby
	providing LRU heuristics used in [6b].

6b)	The active page queue is logically segmented into slices, and a
	page has to be scanned as unused for multiple passes to be deactivated.
	Pages that are used are placed immediately at the end of the active
	queue, thereby deferring deactivation for a "long time".

6c)	A limited form of biasing page inactivation towards processes exceeding
	RSS limits has been included.  Hooks have been added to support hard
	RSS limits that may be added in the future.

7)	Vm_fault now supports pagers that can do multiple block input.  The
	logic makes sure that the readahead/behind falls in the same object
	and pager (used in [8] and [9].)  Readaheads/behinds are inhibited
	under very low memory conditions.  It is left to the pager to decide
	whether to actually perform the readahead/behind.

8)	The swap_pager supports contiguous block allocations and multi-block
	input can be peformed.

9)	The vnode_pager bypasses the buffer-caching mechanism, thereby
	conserving buffer cache and eliminating page memory copies under
	most circumstances.  Also, the vnode_pager supports contiguous
	readaheads/behinds.

10)	The vm_object_collapse code has be improved significantly, and the
	swap_pager allocations have been modified to support the improved
	operation.  (This improves the swap space allocation problem.)
	Under rare circumstances it is still possible to run into the
	VM swap space deallocation problems related to the object
	paging_in_progress flag, but a fix is in the works ( and is
	relatively simple, but tedious.)

11)	The swap space full condition is handled in a rather severe way.
	If a process tries to allocate another page in its working
	set when swap space is zero and no backing store exists, it is killed.
	(I know that this is not good, but it is worse to hang the system.)

12)	Because of the swap space being allocated in blocks, it is possible
	for a page that is allocated on swap space to never be used.  An
	optimization has been made to the swap pager to reclaim the space,
	thereby almost eliminating internal fragmentation of the swap
	blocks.

13)	The algorithms associated with vm_pageout have been improved to
	decrease significantly the amount of CPU time used.  (Wakeups
	of the pageout daemon have been minimized.)
13a)	vm_page_alloc woke-up the page daemon MUCH too often.
13b)	the swap pager output completion woke-up the page daemon too often.
13c)	the probability of a page at the beginning of a queue needing
	to be (deactivated or freed) is higher

14)	The system has been modified to support discontiguous allocation of
	physical memory.  This allows use of memory below 640K when the kernel
	is loaded at 1MB.  Now, the system is configured to load only at 1MB,
	because the cost is minimal (1/4 page table page, instead of 640K) and
	every kernel loads at the same address.

15)	There now is an internal option to tune the location of the kernel in
	the virtual address space.  Additionally, the default base kernel
	virtual address is now 0xF0000000, moved from 0xFE000000, allowing
	up to a 256MB kernel virtual address space.  Currently we support
	a usable kernel address space of (60-1)MB (as determined by the actual
	number of allocated kernel ptes (15).)  (Actually, now the kernel uses
	virtual addresses starting at 0xF0100000 because it is loaded at
	1MB.)

	John Dyson
	David Greenman