vm_reserv.c revision 174982
1174982Salc/*-
2174982Salc * Copyright (c) 2002-2006 Rice University
3174982Salc * Copyright (c) 2007 Alan L. Cox <alc@cs.rice.edu>
4174982Salc * All rights reserved.
5174982Salc *
6174982Salc * This software was developed for the FreeBSD Project by Alan L. Cox,
7174982Salc * Olivier Crameri, Peter Druschel, Sitaram Iyer, and Juan Navarro.
8174982Salc *
9174982Salc * Redistribution and use in source and binary forms, with or without
10174982Salc * modification, are permitted provided that the following conditions
11174982Salc * are met:
12174982Salc * 1. Redistributions of source code must retain the above copyright
13174982Salc *    notice, this list of conditions and the following disclaimer.
14174982Salc * 2. Redistributions in binary form must reproduce the above copyright
15174982Salc *    notice, this list of conditions and the following disclaimer in the
16174982Salc *    documentation and/or other materials provided with the distribution.
17174982Salc *
18174982Salc * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
19174982Salc * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
20174982Salc * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
21174982Salc * A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE COPYRIGHT
22174982Salc * HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
23174982Salc * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
24174982Salc * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS
25174982Salc * OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
26174982Salc * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
27174982Salc * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY
28174982Salc * WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
29174982Salc * POSSIBILITY OF SUCH DAMAGE.
30174982Salc */
31174982Salc
32174982Salc/*
33174982Salc *	Superpage reservation management module
34174982Salc */
35174982Salc
36174982Salc#include <sys/cdefs.h>
37174982Salc__FBSDID("$FreeBSD: head/sys/vm/vm_reserv.c 174982 2007-12-29 19:53:04Z alc $");
38174982Salc
39174982Salc#include "opt_vm.h"
40174982Salc
41174982Salc#include <sys/param.h>
42174982Salc#include <sys/kernel.h>
43174982Salc#include <sys/lock.h>
44174982Salc#include <sys/malloc.h>
45174982Salc#include <sys/mutex.h>
46174982Salc#include <sys/queue.h>
47174982Salc#include <sys/sbuf.h>
48174982Salc#include <sys/sysctl.h>
49174982Salc#include <sys/systm.h>
50174982Salc
51174982Salc#include <vm/vm.h>
52174982Salc#include <vm/vm_param.h>
53174982Salc#include <vm/vm_object.h>
54174982Salc#include <vm/vm_page.h>
55174982Salc#include <vm/vm_phys.h>
56174982Salc#include <vm/vm_reserv.h>
57174982Salc
58174982Salc/*
59174982Salc * The reservation system supports the speculative allocation of large physical
60174982Salc * pages ("superpages").  Speculative allocation enables the fully-automatic
61174982Salc * utilization of superpages by the virtual memory system.  In other words, no
62174982Salc * programmatic directives are required to use superpages.
63174982Salc */
64174982Salc
65174982Salc#if VM_NRESERVLEVEL > 0
66174982Salc
67174982Salc/*
68174982Salc * The number of small pages that are contained in a level 0 reservation
69174982Salc */
70174982Salc#define	VM_LEVEL_0_NPAGES	(1 << VM_LEVEL_0_ORDER)
71174982Salc
72174982Salc/*
73174982Salc * The number of bits by which a physical address is shifted to obtain the
74174982Salc * reservation number
75174982Salc */
76174982Salc#define	VM_LEVEL_0_SHIFT	(VM_LEVEL_0_ORDER + PAGE_SHIFT)
77174982Salc
78174982Salc/*
79174982Salc * The size of a level 0 reservation in bytes
80174982Salc */
81174982Salc#define	VM_LEVEL_0_SIZE		(1 << VM_LEVEL_0_SHIFT)
82174982Salc
83174982Salc/*
84174982Salc * Computes the index of the small page underlying the given (object, pindex)
85174982Salc * within the reservation's array of small pages.
86174982Salc */
87174982Salc#define	VM_RESERV_INDEX(object, pindex)	\
88174982Salc    (((object)->pg_color + (pindex)) & (VM_LEVEL_0_NPAGES - 1))
89174982Salc
90174982Salc/*
91174982Salc * The reservation structure
92174982Salc *
93174982Salc * A reservation structure is constructed whenever a large physical page is
94174982Salc * speculatively allocated to an object.  The reservation provides the small
95174982Salc * physical pages for the range [pindex, pindex + VM_LEVEL_0_NPAGES) of offsets
96174982Salc * within that object.  The reservation's "popcnt" tracks the number of these
97174982Salc * small physical pages that are in use at any given time.  When and if the
98174982Salc * reservation is not fully utilized, it appears in the queue of partially-
99174982Salc * populated reservations.  The reservation always appears on the containing
100174982Salc * object's list of reservations.
101174982Salc *
102174982Salc * A partially-populated reservation can be broken and reclaimed at any time.
103174982Salc */
104174982Salcstruct vm_reserv {
105174982Salc	TAILQ_ENTRY(vm_reserv) partpopq;
106174982Salc	LIST_ENTRY(vm_reserv) objq;
107174982Salc	vm_object_t	object;			/* containing object */
108174982Salc	vm_pindex_t	pindex;			/* offset within object */
109174982Salc	vm_page_t	pages;			/* first page of a superpage */
110174982Salc	int		popcnt;			/* # of pages in use */
111174982Salc	char		inpartpopq;
112174982Salc};
113174982Salc
114174982Salc/*
115174982Salc * The reservation array
116174982Salc *
117174982Salc * This array is analoguous in function to vm_page_array.  It differs in the
118174982Salc * respect that it may contain a greater number of useful reservation
119174982Salc * structures than there are (physical) superpages.  These "invalid"
120174982Salc * reservation structures exist to trade-off space for time in the
121174982Salc * implementation of vm_reserv_from_page().  Invalid reservation structures are
122174982Salc * distinguishable from "valid" reservation structures by inspecting the
123174982Salc * reservation's "pages" field.  Invalid reservation structures have a NULL
124174982Salc * "pages" field.
125174982Salc *
126174982Salc * vm_reserv_from_page() maps a small (physical) page to an element of this
127174982Salc * array by computing a physical reservation number from the page's physical
128174982Salc * address.  The physical reservation number is used as the array index.
129174982Salc *
130174982Salc * An "active" reservation is a valid reservation structure that has a non-NULL
131174982Salc * "object" field and a non-zero "popcnt" field.  In other words, every active
132174982Salc * reservation belongs to a particular object.  Moreover, every active
133174982Salc * reservation has an entry in the containing object's list of reservations.
134174982Salc */
135174982Salcstatic vm_reserv_t vm_reserv_array;
136174982Salc
137174982Salc/*
138174982Salc * The partially-populated reservation queue
139174982Salc *
140174982Salc * This queue enables the fast recovery of an unused cached or free small page
141174982Salc * from a partially-populated reservation.  The head of this queue is either
142174982Salc * the least-recently-populated or most-recently-depopulated reservation.
143174982Salc *
144174982Salc * Access to this queue is synchronized by the free page queue lock.
145174982Salc */
146174982Salcstatic TAILQ_HEAD(, vm_reserv) vm_rvq_partpop =
147174982Salc			    TAILQ_HEAD_INITIALIZER(vm_rvq_partpop);
148174982Salc
149174982Salcstatic SYSCTL_NODE(_vm, OID_AUTO, reserv, CTLFLAG_RD, 0, "Reservation Info");
150174982Salc
151174982Salcstatic long vm_reserv_broken;
152174982SalcSYSCTL_LONG(_vm_reserv, OID_AUTO, broken, CTLFLAG_RD,
153174982Salc    &vm_reserv_broken, 0, "Cumulative number of broken reservations");
154174982Salc
155174982Salcstatic long vm_reserv_freed;
156174982SalcSYSCTL_LONG(_vm_reserv, OID_AUTO, freed, CTLFLAG_RD,
157174982Salc    &vm_reserv_freed, 0, "Cumulative number of freed reservations");
158174982Salc
159174982Salcstatic int sysctl_vm_reserv_partpopq(SYSCTL_HANDLER_ARGS);
160174982Salc
161174982SalcSYSCTL_OID(_vm_reserv, OID_AUTO, partpopq, CTLTYPE_STRING | CTLFLAG_RD, NULL, 0,
162174982Salc    sysctl_vm_reserv_partpopq, "A", "Partially-populated reservation queues");
163174982Salc
164174982Salcstatic long vm_reserv_reclaimed;
165174982SalcSYSCTL_LONG(_vm_reserv, OID_AUTO, reclaimed, CTLFLAG_RD,
166174982Salc    &vm_reserv_reclaimed, 0, "Cumulative number of reclaimed reservations");
167174982Salc
168174982Salcstatic void		vm_reserv_depopulate(vm_reserv_t rv);
169174982Salcstatic vm_reserv_t	vm_reserv_from_page(vm_page_t m);
170174982Salcstatic boolean_t	vm_reserv_has_pindex(vm_reserv_t rv,
171174982Salc			    vm_pindex_t pindex);
172174982Salcstatic void		vm_reserv_populate(vm_reserv_t rv);
173174982Salc
174174982Salc/*
175174982Salc * Describes the current state of the partially-populated reservation queue.
176174982Salc */
177174982Salcstatic int
178174982Salcsysctl_vm_reserv_partpopq(SYSCTL_HANDLER_ARGS)
179174982Salc{
180174982Salc	struct sbuf sbuf;
181174982Salc	vm_reserv_t rv;
182174982Salc	char *cbuf;
183174982Salc	const int cbufsize = (VM_NRESERVLEVEL + 1) * 81;
184174982Salc	int counter, error, level, unused_pages;
185174982Salc
186174982Salc	cbuf = malloc(cbufsize, M_TEMP, M_WAITOK | M_ZERO);
187174982Salc	sbuf_new(&sbuf, cbuf, cbufsize, SBUF_FIXEDLEN);
188174982Salc	sbuf_printf(&sbuf, "\nLEVEL     SIZE  NUMBER\n\n");
189174982Salc	for (level = -1; level <= VM_NRESERVLEVEL - 2; level++) {
190174982Salc		counter = 0;
191174982Salc		unused_pages = 0;
192174982Salc		mtx_lock(&vm_page_queue_free_mtx);
193174982Salc		TAILQ_FOREACH(rv, &vm_rvq_partpop/*[level]*/, partpopq) {
194174982Salc			counter++;
195174982Salc			unused_pages += VM_LEVEL_0_NPAGES - rv->popcnt;
196174982Salc		}
197174982Salc		mtx_unlock(&vm_page_queue_free_mtx);
198174982Salc		sbuf_printf(&sbuf, "%5.5d: %6.6dK, %6.6d\n", level,
199174982Salc		    unused_pages * (PAGE_SIZE / 1024), counter);
200174982Salc	}
201174982Salc	sbuf_finish(&sbuf);
202174982Salc	error = SYSCTL_OUT(req, sbuf_data(&sbuf), sbuf_len(&sbuf));
203174982Salc	sbuf_delete(&sbuf);
204174982Salc	free(cbuf, M_TEMP);
205174982Salc	return (error);
206174982Salc}
207174982Salc
208174982Salc/*
209174982Salc * Reduces the given reservation's population count.  If the population count
210174982Salc * becomes zero, the reservation is destroyed.  Additionally, moves the
211174982Salc * reservation to the head of the partially-populated reservations queue if the
212174982Salc * population count is non-zero.
213174982Salc *
214174982Salc * The free page queue lock must be held.
215174982Salc */
216174982Salcstatic void
217174982Salcvm_reserv_depopulate(vm_reserv_t rv)
218174982Salc{
219174982Salc
220174982Salc	mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
221174982Salc	KASSERT(rv->object != NULL,
222174982Salc	    ("vm_reserv_depopulate: reserv %p is free", rv));
223174982Salc	KASSERT(rv->popcnt > 0,
224174982Salc	    ("vm_reserv_depopulate: reserv %p's popcnt is corrupted", rv));
225174982Salc	if (rv->inpartpopq) {
226174982Salc		TAILQ_REMOVE(&vm_rvq_partpop, rv, partpopq);
227174982Salc		rv->inpartpopq = FALSE;
228174982Salc	}
229174982Salc	rv->popcnt--;
230174982Salc	if (rv->popcnt == 0) {
231174982Salc		LIST_REMOVE(rv, objq);
232174982Salc		rv->object = NULL;
233174982Salc		vm_phys_free_pages(rv->pages, VM_LEVEL_0_ORDER);
234174982Salc		vm_reserv_freed++;
235174982Salc	} else {
236174982Salc		rv->inpartpopq = TRUE;
237174982Salc		TAILQ_INSERT_HEAD(&vm_rvq_partpop, rv, partpopq);
238174982Salc	}
239174982Salc}
240174982Salc
241174982Salc/*
242174982Salc * Returns the reservation to which the given page might belong.
243174982Salc */
244174982Salcstatic __inline vm_reserv_t
245174982Salcvm_reserv_from_page(vm_page_t m)
246174982Salc{
247174982Salc
248174982Salc	return (&vm_reserv_array[VM_PAGE_TO_PHYS(m) >> VM_LEVEL_0_SHIFT]);
249174982Salc}
250174982Salc
251174982Salc/*
252174982Salc * Returns TRUE if the given reservation contains the given page index and
253174982Salc * FALSE otherwise.
254174982Salc */
255174982Salcstatic __inline boolean_t
256174982Salcvm_reserv_has_pindex(vm_reserv_t rv, vm_pindex_t pindex)
257174982Salc{
258174982Salc
259174982Salc	return (((pindex - rv->pindex) & ~(VM_LEVEL_0_NPAGES - 1)) == 0);
260174982Salc}
261174982Salc
262174982Salc/*
263174982Salc * Increases the given reservation's population count.  Moves the reservation
264174982Salc * to the tail of the partially-populated reservation queue.
265174982Salc *
266174982Salc * The free page queue must be locked.
267174982Salc */
268174982Salcstatic void
269174982Salcvm_reserv_populate(vm_reserv_t rv)
270174982Salc{
271174982Salc
272174982Salc	mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
273174982Salc	KASSERT(rv->object != NULL,
274174982Salc	    ("vm_reserv_populate: reserv %p is free", rv));
275174982Salc	KASSERT(rv->popcnt < VM_LEVEL_0_NPAGES,
276174982Salc	    ("vm_reserv_populate: reserv %p is already full", rv));
277174982Salc	if (rv->inpartpopq) {
278174982Salc		TAILQ_REMOVE(&vm_rvq_partpop, rv, partpopq);
279174982Salc		rv->inpartpopq = FALSE;
280174982Salc	}
281174982Salc	rv->popcnt++;
282174982Salc	if (rv->popcnt < VM_LEVEL_0_NPAGES) {
283174982Salc		rv->inpartpopq = TRUE;
284174982Salc		TAILQ_INSERT_TAIL(&vm_rvq_partpop, rv, partpopq);
285174982Salc	}
286174982Salc}
287174982Salc
288174982Salc/*
289174982Salc * Allocates a page from an existing or newly-created reservation.
290174982Salc *
291174982Salc * The object and free page queue must be locked.
292174982Salc */
293174982Salcvm_page_t
294174982Salcvm_reserv_alloc_page(vm_object_t object, vm_pindex_t pindex)
295174982Salc{
296174982Salc	vm_page_t m, mpred, msucc;
297174982Salc	vm_pindex_t first, leftcap, rightcap;
298174982Salc	vm_reserv_t rv;
299174982Salc
300174982Salc	mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
301174982Salc
302174982Salc	/*
303174982Salc	 * Is a reservation fundamentally not possible?
304174982Salc	 */
305174982Salc	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
306174982Salc	if (pindex < VM_RESERV_INDEX(object, pindex) ||
307174982Salc	    pindex >= object->size)
308174982Salc		return (NULL);
309174982Salc
310174982Salc	/*
311174982Salc	 * Look for an existing reservation.
312174982Salc	 */
313174982Salc	msucc = NULL;
314174982Salc	mpred = object->root;
315174982Salc	while (mpred != NULL) {
316174982Salc		KASSERT(mpred->pindex != pindex,
317174982Salc		    ("vm_reserv_alloc_page: pindex already allocated"));
318174982Salc		rv = vm_reserv_from_page(mpred);
319174982Salc		if (rv->object == object && vm_reserv_has_pindex(rv, pindex)) {
320174982Salc			m = &rv->pages[VM_RESERV_INDEX(object, pindex)];
321174982Salc			/* Handle vm_page_rename(m, new_object, ...). */
322174982Salc			if ((m->flags & (PG_CACHED | PG_FREE)) == 0)
323174982Salc				return (NULL);
324174982Salc			vm_reserv_populate(rv);
325174982Salc			return (m);
326174982Salc		} else if (mpred->pindex < pindex) {
327174982Salc			if (msucc != NULL ||
328174982Salc			    (msucc = TAILQ_NEXT(mpred, listq)) == NULL)
329174982Salc				break;
330174982Salc			KASSERT(msucc->pindex != pindex,
331174982Salc			    ("vm_reserv_alloc_page: pindex already allocated"));
332174982Salc			rv = vm_reserv_from_page(msucc);
333174982Salc			if (rv->object == object &&
334174982Salc			    vm_reserv_has_pindex(rv, pindex)) {
335174982Salc				m = &rv->pages[VM_RESERV_INDEX(object, pindex)];
336174982Salc				/* Handle vm_page_rename(m, new_object, ...). */
337174982Salc				if ((m->flags & (PG_CACHED | PG_FREE)) == 0)
338174982Salc					return (NULL);
339174982Salc				vm_reserv_populate(rv);
340174982Salc				return (m);
341174982Salc			} else if (pindex < msucc->pindex)
342174982Salc				break;
343174982Salc		} else if (msucc == NULL) {
344174982Salc			msucc = mpred;
345174982Salc			mpred = TAILQ_PREV(msucc, pglist, listq);
346174982Salc			continue;
347174982Salc		}
348174982Salc		msucc = NULL;
349174982Salc		mpred = object->root = vm_page_splay(pindex, object->root);
350174982Salc	}
351174982Salc
352174982Salc	/*
353174982Salc	 * Determine the first index to the left that can be used.
354174982Salc	 */
355174982Salc	if (mpred == NULL)
356174982Salc		leftcap = 0;
357174982Salc	else if ((rv = vm_reserv_from_page(mpred))->object != object)
358174982Salc		leftcap = mpred->pindex + 1;
359174982Salc	else
360174982Salc		leftcap = rv->pindex + VM_LEVEL_0_NPAGES;
361174982Salc
362174982Salc	/*
363174982Salc	 * Determine the first index to the right that cannot be used.
364174982Salc	 */
365174982Salc	if (msucc == NULL)
366174982Salc		rightcap = pindex + VM_LEVEL_0_NPAGES;
367174982Salc	else if ((rv = vm_reserv_from_page(msucc))->object != object)
368174982Salc		rightcap = msucc->pindex;
369174982Salc	else
370174982Salc		rightcap = rv->pindex;
371174982Salc
372174982Salc	/*
373174982Salc	 * Determine if a reservation fits between the first index to
374174982Salc	 * the left that can be used and the first index to the right
375174982Salc	 * that cannot be used.
376174982Salc	 */
377174982Salc	first = pindex - VM_RESERV_INDEX(object, pindex);
378174982Salc	if (first < leftcap || first + VM_LEVEL_0_NPAGES > rightcap)
379174982Salc		return (NULL);
380174982Salc
381174982Salc	/*
382174982Salc	 * Would a new reservation extend past the end of the given object?
383174982Salc	 */
384174982Salc	if (object->size < first + VM_LEVEL_0_NPAGES) {
385174982Salc		/*
386174982Salc		 * Don't allocate a new reservation if the object is a vnode or
387174982Salc		 * backed by another object that is a vnode.
388174982Salc		 */
389174982Salc		if (object->type == OBJT_VNODE ||
390174982Salc		    (object->backing_object != NULL &&
391174982Salc		    object->backing_object->type == OBJT_VNODE))
392174982Salc			return (NULL);
393174982Salc		/* Speculate that the object may grow. */
394174982Salc	}
395174982Salc
396174982Salc	/*
397174982Salc	 * Allocate a new reservation.
398174982Salc	 */
399174982Salc	m = vm_phys_alloc_pages(VM_FREEPOOL_DEFAULT, VM_LEVEL_0_ORDER);
400174982Salc	if (m != NULL) {
401174982Salc		rv = vm_reserv_from_page(m);
402174982Salc		KASSERT(rv->pages == m,
403174982Salc		    ("vm_reserv_alloc_page: reserv %p's pages is corrupted",
404174982Salc		    rv));
405174982Salc		KASSERT(rv->object == NULL,
406174982Salc		    ("vm_reserv_alloc_page: reserv %p isn't free", rv));
407174982Salc		LIST_INSERT_HEAD(&object->rvq, rv, objq);
408174982Salc		rv->object = object;
409174982Salc		rv->pindex = first;
410174982Salc		KASSERT(rv->popcnt == 0,
411174982Salc		    ("vm_reserv_alloc_page: reserv %p's popcnt is corrupted",
412174982Salc		    rv));
413174982Salc		KASSERT(!rv->inpartpopq,
414174982Salc		    ("vm_reserv_alloc_page: reserv %p's inpartpopq is TRUE",
415174982Salc		    rv));
416174982Salc		vm_reserv_populate(rv);
417174982Salc		m = &rv->pages[VM_RESERV_INDEX(object, pindex)];
418174982Salc	}
419174982Salc	return (m);
420174982Salc}
421174982Salc
422174982Salc/*
423174982Salc * Breaks all reservations belonging to the given object.
424174982Salc */
425174982Salcvoid
426174982Salcvm_reserv_break_all(vm_object_t object)
427174982Salc{
428174982Salc	vm_reserv_t rv;
429174982Salc	int i;
430174982Salc
431174982Salc	mtx_lock(&vm_page_queue_free_mtx);
432174982Salc	while ((rv = LIST_FIRST(&object->rvq)) != NULL) {
433174982Salc		KASSERT(rv->object == object,
434174982Salc		    ("vm_reserv_break_all: reserv %p is corrupted", rv));
435174982Salc		if (rv->inpartpopq) {
436174982Salc			TAILQ_REMOVE(&vm_rvq_partpop, rv, partpopq);
437174982Salc			rv->inpartpopq = FALSE;
438174982Salc		}
439174982Salc		LIST_REMOVE(rv, objq);
440174982Salc		rv->object = NULL;
441174982Salc		for (i = 0; i < VM_LEVEL_0_NPAGES; i++) {
442174982Salc			if ((rv->pages[i].flags & (PG_CACHED | PG_FREE)) != 0)
443174982Salc				vm_phys_free_pages(&rv->pages[i], 0);
444174982Salc			else
445174982Salc				rv->popcnt--;
446174982Salc		}
447174982Salc		KASSERT(rv->popcnt == 0,
448174982Salc		    ("vm_reserv_break_all: reserv %p's popcnt is corrupted",
449174982Salc		    rv));
450174982Salc		vm_reserv_broken++;
451174982Salc	}
452174982Salc	mtx_unlock(&vm_page_queue_free_mtx);
453174982Salc}
454174982Salc
455174982Salc/*
456174982Salc * Frees the given page if it belongs to a reservation.  Returns TRUE if the
457174982Salc * page is freed and FALSE otherwise.
458174982Salc *
459174982Salc * The free page queue lock must be held.
460174982Salc */
461174982Salcboolean_t
462174982Salcvm_reserv_free_page(vm_page_t m)
463174982Salc{
464174982Salc	vm_reserv_t rv;
465174982Salc
466174982Salc	mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
467174982Salc	rv = vm_reserv_from_page(m);
468174982Salc	if (rv->object != NULL) {
469174982Salc		vm_reserv_depopulate(rv);
470174982Salc		return (TRUE);
471174982Salc	}
472174982Salc	return (FALSE);
473174982Salc}
474174982Salc
475174982Salc/*
476174982Salc * Initializes the reservation management system.  Specifically, initializes
477174982Salc * the reservation array.
478174982Salc *
479174982Salc * Requires that vm_page_array and first_page are initialized!
480174982Salc */
481174982Salcvoid
482174982Salcvm_reserv_init(void)
483174982Salc{
484174982Salc	vm_paddr_t paddr;
485174982Salc	int i;
486174982Salc
487174982Salc	/*
488174982Salc	 * Initialize the reservation array.  Specifically, initialize the
489174982Salc	 * "pages" field for every element that has an underlying superpage.
490174982Salc	 */
491174982Salc	for (i = 0; phys_avail[i + 1] != 0; i += 2) {
492174982Salc		paddr = roundup2(phys_avail[i], VM_LEVEL_0_SIZE);
493174982Salc		while (paddr + VM_LEVEL_0_SIZE <= phys_avail[i + 1]) {
494174982Salc			vm_reserv_array[paddr >> VM_LEVEL_0_SHIFT].pages =
495174982Salc			    PHYS_TO_VM_PAGE(paddr);
496174982Salc			paddr += VM_LEVEL_0_SIZE;
497174982Salc		}
498174982Salc	}
499174982Salc}
500174982Salc
501174982Salc/*
502174982Salc * Returns a reservation level if the given page belongs to a fully-populated
503174982Salc * reservation and -1 otherwise.
504174982Salc */
505174982Salcint
506174982Salcvm_reserv_level_iffullpop(vm_page_t m)
507174982Salc{
508174982Salc	vm_reserv_t rv;
509174982Salc
510174982Salc	rv = vm_reserv_from_page(m);
511174982Salc	return (rv->popcnt == VM_LEVEL_0_NPAGES ? 0 : -1);
512174982Salc}
513174982Salc
514174982Salc/*
515174982Salc * Prepare for the reactivation of a cached page.
516174982Salc *
517174982Salc * First, suppose that the given page "m" was allocated individually, i.e., not
518174982Salc * as part of a reservation, and cached.  Then, suppose a reservation
519174982Salc * containing "m" is allocated by the same object.  Although "m" and the
520174982Salc * reservation belong to the same object, "m"'s pindex may not match the
521174982Salc * reservation's.
522174982Salc *
523174982Salc * The free page queue must be locked.
524174982Salc */
525174982Salcboolean_t
526174982Salcvm_reserv_reactivate_page(vm_page_t m)
527174982Salc{
528174982Salc	vm_reserv_t rv;
529174982Salc	int i, m_index;
530174982Salc
531174982Salc	mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
532174982Salc	rv = vm_reserv_from_page(m);
533174982Salc	if (rv->object == NULL)
534174982Salc		return (FALSE);
535174982Salc	KASSERT((m->flags & PG_CACHED) != 0,
536174982Salc	    ("vm_reserv_uncache_page: page %p is not cached", m));
537174982Salc	if (m->object == rv->object &&
538174982Salc	    m->pindex - rv->pindex == VM_RESERV_INDEX(m->object, m->pindex))
539174982Salc		vm_reserv_populate(rv);
540174982Salc	else {
541174982Salc		KASSERT(rv->inpartpopq,
542174982Salc		    ("vm_reserv_uncache_page: reserv %p's inpartpopq is FALSE",
543174982Salc		    rv));
544174982Salc		TAILQ_REMOVE(&vm_rvq_partpop, rv, partpopq);
545174982Salc		rv->inpartpopq = FALSE;
546174982Salc		LIST_REMOVE(rv, objq);
547174982Salc		rv->object = NULL;
548174982Salc		/* Don't vm_phys_free_pages(m, 0). */
549174982Salc		m_index = m - rv->pages;
550174982Salc		for (i = 0; i < m_index; i++) {
551174982Salc			if ((rv->pages[i].flags & (PG_CACHED | PG_FREE)) != 0)
552174982Salc				vm_phys_free_pages(&rv->pages[i], 0);
553174982Salc			else
554174982Salc				rv->popcnt--;
555174982Salc		}
556174982Salc		for (i++; i < VM_LEVEL_0_NPAGES; i++) {
557174982Salc			if ((rv->pages[i].flags & (PG_CACHED | PG_FREE)) != 0)
558174982Salc				vm_phys_free_pages(&rv->pages[i], 0);
559174982Salc			else
560174982Salc				rv->popcnt--;
561174982Salc		}
562174982Salc		KASSERT(rv->popcnt == 0,
563174982Salc		    ("vm_reserv_uncache_page: reserv %p's popcnt is corrupted",
564174982Salc		    rv));
565174982Salc		vm_reserv_broken++;
566174982Salc	}
567174982Salc	return (TRUE);
568174982Salc}
569174982Salc
570174982Salc/*
571174982Salc * Breaks the reservation at the head of the partially-populated reservation
572174982Salc * queue, releasing its cached and free pages to the physical memory
573174982Salc * allocator.  Returns TRUE if a reservation is broken and FALSE otherwise.
574174982Salc *
575174982Salc * The free page queue lock must be held.
576174982Salc */
577174982Salcboolean_t
578174982Salcvm_reserv_reclaim(void)
579174982Salc{
580174982Salc	vm_reserv_t rv;
581174982Salc	int i;
582174982Salc
583174982Salc	mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
584174982Salc	if ((rv = TAILQ_FIRST(&vm_rvq_partpop)) != NULL) {
585174982Salc		KASSERT(rv->inpartpopq,
586174982Salc		    ("vm_reserv_reclaim: reserv %p's inpartpopq is corrupted",
587174982Salc		    rv));
588174982Salc		TAILQ_REMOVE(&vm_rvq_partpop, rv, partpopq);
589174982Salc		rv->inpartpopq = FALSE;
590174982Salc		KASSERT(rv->object != NULL,
591174982Salc		    ("vm_reserv_reclaim: reserv %p is free", rv));
592174982Salc		LIST_REMOVE(rv, objq);
593174982Salc		rv->object = NULL;
594174982Salc		for (i = 0; i < VM_LEVEL_0_NPAGES; i++) {
595174982Salc			if ((rv->pages[i].flags & (PG_CACHED | PG_FREE)) != 0)
596174982Salc				vm_phys_free_pages(&rv->pages[i], 0);
597174982Salc			else
598174982Salc				rv->popcnt--;
599174982Salc		}
600174982Salc		KASSERT(rv->popcnt == 0,
601174982Salc		    ("vm_reserv_reclaim: reserv %p's popcnt is corrupted",
602174982Salc		    rv));
603174982Salc		vm_reserv_reclaimed++;
604174982Salc		return (TRUE);
605174982Salc	}
606174982Salc	return (FALSE);
607174982Salc}
608174982Salc
609174982Salc/*
610174982Salc * Transfers the reservation underlying the given page to a new object.
611174982Salc *
612174982Salc * The object must be locked.
613174982Salc */
614174982Salcvoid
615174982Salcvm_reserv_rename(vm_page_t m, vm_object_t new_object, vm_object_t old_object,
616174982Salc    vm_pindex_t old_object_offset)
617174982Salc{
618174982Salc	vm_reserv_t rv;
619174982Salc
620174982Salc	VM_OBJECT_LOCK_ASSERT(new_object, MA_OWNED);
621174982Salc	rv = vm_reserv_from_page(m);
622174982Salc	if (rv->object == old_object) {
623174982Salc		mtx_lock(&vm_page_queue_free_mtx);
624174982Salc		if (rv->object == old_object) {
625174982Salc			LIST_REMOVE(rv, objq);
626174982Salc			LIST_INSERT_HEAD(&new_object->rvq, rv, objq);
627174982Salc			rv->object = new_object;
628174982Salc			rv->pindex -= old_object_offset;
629174982Salc		}
630174982Salc		mtx_unlock(&vm_page_queue_free_mtx);
631174982Salc	}
632174982Salc}
633174982Salc
634174982Salc/*
635174982Salc * Allocates the virtual and physical memory required by the reservation
636174982Salc * management system's data structures, in particular, the reservation array.
637174982Salc */
638174982Salcvm_paddr_t
639174982Salcvm_reserv_startup(vm_offset_t *vaddr, vm_paddr_t end, vm_paddr_t high_water)
640174982Salc{
641174982Salc	vm_paddr_t new_end;
642174982Salc	size_t size;
643174982Salc
644174982Salc	/*
645174982Salc	 * Calculate the size (in bytes) of the reservation array.  Round up
646174982Salc	 * from "high_water" because every small page is mapped to an element
647174982Salc	 * in the reservation array based on its physical address.  Thus, the
648174982Salc	 * number of elements in the reservation array can be greater than the
649174982Salc	 * number of superpages.
650174982Salc	 */
651174982Salc	size = howmany(high_water, VM_LEVEL_0_SIZE) * sizeof(struct vm_reserv);
652174982Salc
653174982Salc	/*
654174982Salc	 * Allocate and map the physical memory for the reservation array.  The
655174982Salc	 * next available virtual address is returned by reference.
656174982Salc	 */
657174982Salc	new_end = end - round_page(size);
658174982Salc	vm_reserv_array = (void *)(uintptr_t)pmap_map(vaddr, new_end, end,
659174982Salc	    VM_PROT_READ | VM_PROT_WRITE);
660174982Salc	bzero(vm_reserv_array, size);
661174982Salc
662174982Salc	/*
663174982Salc	 * Return the next available physical address.
664174982Salc	 */
665174982Salc	return (new_end);
666174982Salc}
667174982Salc
668174982Salc#endif	/* VM_NRESERVLEVEL > 0 */
669