xref: /freebsd/sys/vm/vm_reserv.c (revision 23090366f729c56cab62de74c7a51792357e98a9)
1 /*-
2  * Copyright (c) 2002-2006 Rice University
3  * Copyright (c) 2007-2008 Alan L. Cox <alc@cs.rice.edu>
4  * All rights reserved.
5  *
6  * This software was developed for the FreeBSD Project by Alan L. Cox,
7  * Olivier Crameri, Peter Druschel, Sitaram Iyer, and Juan Navarro.
8  *
9  * Redistribution and use in source and binary forms, with or without
10  * modification, are permitted provided that the following conditions
11  * are met:
12  * 1. Redistributions of source code must retain the above copyright
13  *    notice, this list of conditions and the following disclaimer.
14  * 2. Redistributions in binary form must reproduce the above copyright
15  *    notice, this list of conditions and the following disclaimer in the
16  *    documentation and/or other materials provided with the distribution.
17  *
18  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
19  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
20  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
21  * A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE COPYRIGHT
22  * HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
23  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
24  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS
25  * OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
26  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
27  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY
28  * WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
29  * POSSIBILITY OF SUCH DAMAGE.
30  */
31 
32 /*
33  *	Superpage reservation management module
34  *
35  * Any external functions defined by this module are only to be used by the
36  * virtual memory system.
37  */
38 
39 #include <sys/cdefs.h>
40 __FBSDID("$FreeBSD$");
41 
42 #include "opt_vm.h"
43 
44 #include <sys/param.h>
45 #include <sys/kernel.h>
46 #include <sys/lock.h>
47 #include <sys/malloc.h>
48 #include <sys/mutex.h>
49 #include <sys/queue.h>
50 #include <sys/sbuf.h>
51 #include <sys/sysctl.h>
52 #include <sys/systm.h>
53 
54 #include <vm/vm.h>
55 #include <vm/vm_param.h>
56 #include <vm/vm_object.h>
57 #include <vm/vm_page.h>
58 #include <vm/vm_phys.h>
59 #include <vm/vm_reserv.h>
60 
61 /*
62  * The reservation system supports the speculative allocation of large physical
63  * pages ("superpages").  Speculative allocation enables the fully-automatic
64  * utilization of superpages by the virtual memory system.  In other words, no
65  * programmatic directives are required to use superpages.
66  */
67 
68 #if VM_NRESERVLEVEL > 0
69 
70 /*
71  * The number of small pages that are contained in a level 0 reservation
72  */
73 #define	VM_LEVEL_0_NPAGES	(1 << VM_LEVEL_0_ORDER)
74 
75 /*
76  * The number of bits by which a physical address is shifted to obtain the
77  * reservation number
78  */
79 #define	VM_LEVEL_0_SHIFT	(VM_LEVEL_0_ORDER + PAGE_SHIFT)
80 
81 /*
82  * The size of a level 0 reservation in bytes
83  */
84 #define	VM_LEVEL_0_SIZE		(1 << VM_LEVEL_0_SHIFT)
85 
86 /*
87  * Computes the index of the small page underlying the given (object, pindex)
88  * within the reservation's array of small pages.
89  */
90 #define	VM_RESERV_INDEX(object, pindex)	\
91     (((object)->pg_color + (pindex)) & (VM_LEVEL_0_NPAGES - 1))
92 
93 /*
94  * The reservation structure
95  *
96  * A reservation structure is constructed whenever a large physical page is
97  * speculatively allocated to an object.  The reservation provides the small
98  * physical pages for the range [pindex, pindex + VM_LEVEL_0_NPAGES) of offsets
99  * within that object.  The reservation's "popcnt" tracks the number of these
100  * small physical pages that are in use at any given time.  When and if the
101  * reservation is not fully utilized, it appears in the queue of partially-
102  * populated reservations.  The reservation always appears on the containing
103  * object's list of reservations.
104  *
105  * A partially-populated reservation can be broken and reclaimed at any time.
106  */
107 struct vm_reserv {
108 	TAILQ_ENTRY(vm_reserv) partpopq;
109 	LIST_ENTRY(vm_reserv) objq;
110 	vm_object_t	object;			/* containing object */
111 	vm_pindex_t	pindex;			/* offset within object */
112 	vm_page_t	pages;			/* first page of a superpage */
113 	int		popcnt;			/* # of pages in use */
114 	char		inpartpopq;
115 };
116 
117 /*
118  * The reservation array
119  *
120  * This array is analoguous in function to vm_page_array.  It differs in the
121  * respect that it may contain a greater number of useful reservation
122  * structures than there are (physical) superpages.  These "invalid"
123  * reservation structures exist to trade-off space for time in the
124  * implementation of vm_reserv_from_page().  Invalid reservation structures are
125  * distinguishable from "valid" reservation structures by inspecting the
126  * reservation's "pages" field.  Invalid reservation structures have a NULL
127  * "pages" field.
128  *
129  * vm_reserv_from_page() maps a small (physical) page to an element of this
130  * array by computing a physical reservation number from the page's physical
131  * address.  The physical reservation number is used as the array index.
132  *
133  * An "active" reservation is a valid reservation structure that has a non-NULL
134  * "object" field and a non-zero "popcnt" field.  In other words, every active
135  * reservation belongs to a particular object.  Moreover, every active
136  * reservation has an entry in the containing object's list of reservations.
137  */
138 static vm_reserv_t vm_reserv_array;
139 
140 /*
141  * The partially-populated reservation queue
142  *
143  * This queue enables the fast recovery of an unused cached or free small page
144  * from a partially-populated reservation.  The reservation at the head of
145  * this queue is the least-recently-changed, partially-populated reservation.
146  *
147  * Access to this queue is synchronized by the free page queue lock.
148  */
149 static TAILQ_HEAD(, vm_reserv) vm_rvq_partpop =
150 			    TAILQ_HEAD_INITIALIZER(vm_rvq_partpop);
151 
152 static SYSCTL_NODE(_vm, OID_AUTO, reserv, CTLFLAG_RD, 0, "Reservation Info");
153 
154 static long vm_reserv_broken;
155 SYSCTL_LONG(_vm_reserv, OID_AUTO, broken, CTLFLAG_RD,
156     &vm_reserv_broken, 0, "Cumulative number of broken reservations");
157 
158 static long vm_reserv_freed;
159 SYSCTL_LONG(_vm_reserv, OID_AUTO, freed, CTLFLAG_RD,
160     &vm_reserv_freed, 0, "Cumulative number of freed reservations");
161 
162 static int sysctl_vm_reserv_partpopq(SYSCTL_HANDLER_ARGS);
163 
164 SYSCTL_OID(_vm_reserv, OID_AUTO, partpopq, CTLTYPE_STRING | CTLFLAG_RD, NULL, 0,
165     sysctl_vm_reserv_partpopq, "A", "Partially-populated reservation queues");
166 
167 static long vm_reserv_reclaimed;
168 SYSCTL_LONG(_vm_reserv, OID_AUTO, reclaimed, CTLFLAG_RD,
169     &vm_reserv_reclaimed, 0, "Cumulative number of reclaimed reservations");
170 
171 static void		vm_reserv_depopulate(vm_reserv_t rv);
172 static vm_reserv_t	vm_reserv_from_page(vm_page_t m);
173 static boolean_t	vm_reserv_has_pindex(vm_reserv_t rv,
174 			    vm_pindex_t pindex);
175 static void		vm_reserv_populate(vm_reserv_t rv);
176 static void		vm_reserv_reclaim(vm_reserv_t rv);
177 
178 /*
179  * Describes the current state of the partially-populated reservation queue.
180  */
181 static int
182 sysctl_vm_reserv_partpopq(SYSCTL_HANDLER_ARGS)
183 {
184 	struct sbuf sbuf;
185 	vm_reserv_t rv;
186 	int counter, error, level, unused_pages;
187 
188 	error = sysctl_wire_old_buffer(req, 0);
189 	if (error != 0)
190 		return (error);
191 	sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
192 	sbuf_printf(&sbuf, "\nLEVEL     SIZE  NUMBER\n\n");
193 	for (level = -1; level <= VM_NRESERVLEVEL - 2; level++) {
194 		counter = 0;
195 		unused_pages = 0;
196 		mtx_lock(&vm_page_queue_free_mtx);
197 		TAILQ_FOREACH(rv, &vm_rvq_partpop/*[level]*/, partpopq) {
198 			counter++;
199 			unused_pages += VM_LEVEL_0_NPAGES - rv->popcnt;
200 		}
201 		mtx_unlock(&vm_page_queue_free_mtx);
202 		sbuf_printf(&sbuf, "%5d: %6dK, %6d\n", level,
203 		    unused_pages * ((int)PAGE_SIZE / 1024), counter);
204 	}
205 	error = sbuf_finish(&sbuf);
206 	sbuf_delete(&sbuf);
207 	return (error);
208 }
209 
210 /*
211  * Reduces the given reservation's population count.  If the population count
212  * becomes zero, the reservation is destroyed.  Additionally, moves the
213  * reservation to the tail of the partially-populated reservations queue if the
214  * population count is non-zero.
215  *
216  * The free page queue lock must be held.
217  */
218 static void
219 vm_reserv_depopulate(vm_reserv_t rv)
220 {
221 
222 	mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
223 	KASSERT(rv->object != NULL,
224 	    ("vm_reserv_depopulate: reserv %p is free", rv));
225 	KASSERT(rv->popcnt > 0,
226 	    ("vm_reserv_depopulate: reserv %p's popcnt is corrupted", rv));
227 	if (rv->inpartpopq) {
228 		TAILQ_REMOVE(&vm_rvq_partpop, rv, partpopq);
229 		rv->inpartpopq = FALSE;
230 	}
231 	rv->popcnt--;
232 	if (rv->popcnt == 0) {
233 		LIST_REMOVE(rv, objq);
234 		rv->object = NULL;
235 		vm_phys_free_pages(rv->pages, VM_LEVEL_0_ORDER);
236 		vm_reserv_freed++;
237 	} else {
238 		rv->inpartpopq = TRUE;
239 		TAILQ_INSERT_TAIL(&vm_rvq_partpop, rv, partpopq);
240 	}
241 }
242 
243 /*
244  * Returns the reservation to which the given page might belong.
245  */
246 static __inline vm_reserv_t
247 vm_reserv_from_page(vm_page_t m)
248 {
249 
250 	return (&vm_reserv_array[VM_PAGE_TO_PHYS(m) >> VM_LEVEL_0_SHIFT]);
251 }
252 
253 /*
254  * Returns TRUE if the given reservation contains the given page index and
255  * FALSE otherwise.
256  */
257 static __inline boolean_t
258 vm_reserv_has_pindex(vm_reserv_t rv, vm_pindex_t pindex)
259 {
260 
261 	return (((pindex - rv->pindex) & ~(VM_LEVEL_0_NPAGES - 1)) == 0);
262 }
263 
264 /*
265  * Increases the given reservation's population count.  Moves the reservation
266  * to the tail of the partially-populated reservation queue.
267  *
268  * The free page queue must be locked.
269  */
270 static void
271 vm_reserv_populate(vm_reserv_t rv)
272 {
273 
274 	mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
275 	KASSERT(rv->object != NULL,
276 	    ("vm_reserv_populate: reserv %p is free", rv));
277 	KASSERT(rv->popcnt < VM_LEVEL_0_NPAGES,
278 	    ("vm_reserv_populate: reserv %p is already full", rv));
279 	if (rv->inpartpopq) {
280 		TAILQ_REMOVE(&vm_rvq_partpop, rv, partpopq);
281 		rv->inpartpopq = FALSE;
282 	}
283 	rv->popcnt++;
284 	if (rv->popcnt < VM_LEVEL_0_NPAGES) {
285 		rv->inpartpopq = TRUE;
286 		TAILQ_INSERT_TAIL(&vm_rvq_partpop, rv, partpopq);
287 	}
288 }
289 
290 /*
291  * Allocates a contiguous set of physical pages of the given size "npages"
292  * from an existing or newly-created reservation.  All of the physical pages
293  * must be at or above the given physical address "low" and below the given
294  * physical address "high".  The given value "alignment" determines the
295  * alignment of the first physical page in the set.  If the given value
296  * "boundary" is non-zero, then the set of physical pages cannot cross any
297  * physical address boundary that is a multiple of that value.  Both
298  * "alignment" and "boundary" must be a power of two.
299  *
300  * The object and free page queue must be locked.
301  */
302 vm_page_t
303 vm_reserv_alloc_contig(vm_object_t object, vm_pindex_t pindex, u_long npages,
304     vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary)
305 {
306 	vm_paddr_t pa, size;
307 	vm_page_t m, m_ret, mpred, msucc;
308 	vm_pindex_t first, leftcap, rightcap;
309 	vm_reserv_t rv;
310 	u_long allocpages, maxpages, minpages;
311 	int i, index, n;
312 
313 	mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
314 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
315 	KASSERT(npages != 0, ("vm_reserv_alloc_contig: npages is 0"));
316 
317 	/*
318 	 * Is a reservation fundamentally impossible?
319 	 */
320 	if (pindex < VM_RESERV_INDEX(object, pindex) ||
321 	    pindex + npages > object->size)
322 		return (NULL);
323 
324 	/*
325 	 * All reservations of a particular size have the same alignment.
326 	 * Assuming that the first page is allocated from a reservation, the
327 	 * least significant bits of its physical address can be determined
328 	 * from its offset from the beginning of the reservation and the size
329 	 * of the reservation.
330 	 *
331 	 * Could the specified index within a reservation of the smallest
332 	 * possible size satisfy the alignment and boundary requirements?
333 	 */
334 	pa = VM_RESERV_INDEX(object, pindex) << PAGE_SHIFT;
335 	if ((pa & (alignment - 1)) != 0)
336 		return (NULL);
337 	size = npages << PAGE_SHIFT;
338 	if (((pa ^ (pa + size - 1)) & ~(boundary - 1)) != 0)
339 		return (NULL);
340 
341 	/*
342 	 * Look for an existing reservation.
343 	 */
344 	msucc = NULL;
345 	mpred = object->root;
346 	while (mpred != NULL) {
347 		KASSERT(mpred->pindex != pindex,
348 		    ("vm_reserv_alloc_contig: pindex already allocated"));
349 		rv = vm_reserv_from_page(mpred);
350 		if (rv->object == object && vm_reserv_has_pindex(rv, pindex))
351 			goto found;
352 		else if (mpred->pindex < pindex) {
353 			if (msucc != NULL ||
354 			    (msucc = TAILQ_NEXT(mpred, listq)) == NULL)
355 				break;
356 			KASSERT(msucc->pindex != pindex,
357 		    ("vm_reserv_alloc_contig: pindex already allocated"));
358 			rv = vm_reserv_from_page(msucc);
359 			if (rv->object == object &&
360 			    vm_reserv_has_pindex(rv, pindex))
361 				goto found;
362 			else if (pindex < msucc->pindex)
363 				break;
364 		} else if (msucc == NULL) {
365 			msucc = mpred;
366 			mpred = TAILQ_PREV(msucc, pglist, listq);
367 			continue;
368 		}
369 		msucc = NULL;
370 		mpred = object->root = vm_page_splay(pindex, object->root);
371 	}
372 
373 	/*
374 	 * Could at least one reservation fit between the first index to the
375 	 * left that can be used and the first index to the right that cannot
376 	 * be used?
377 	 */
378 	first = pindex - VM_RESERV_INDEX(object, pindex);
379 	if (mpred != NULL) {
380 		if ((rv = vm_reserv_from_page(mpred))->object != object)
381 			leftcap = mpred->pindex + 1;
382 		else
383 			leftcap = rv->pindex + VM_LEVEL_0_NPAGES;
384 		if (leftcap > first)
385 			return (NULL);
386 	}
387 	minpages = VM_RESERV_INDEX(object, pindex) + npages;
388 	maxpages = roundup2(minpages, VM_LEVEL_0_NPAGES);
389 	allocpages = maxpages;
390 	if (msucc != NULL) {
391 		if ((rv = vm_reserv_from_page(msucc))->object != object)
392 			rightcap = msucc->pindex;
393 		else
394 			rightcap = rv->pindex;
395 		if (first + maxpages > rightcap) {
396 			if (maxpages == VM_LEVEL_0_NPAGES)
397 				return (NULL);
398 			allocpages = minpages;
399 		}
400 	}
401 
402 	/*
403 	 * Would the last new reservation extend past the end of the object?
404 	 */
405 	if (first + maxpages > object->size) {
406 		/*
407 		 * Don't allocate the last new reservation if the object is a
408 		 * vnode or backed by another object that is a vnode.
409 		 */
410 		if (object->type == OBJT_VNODE ||
411 		    (object->backing_object != NULL &&
412 		    object->backing_object->type == OBJT_VNODE)) {
413 			if (maxpages == VM_LEVEL_0_NPAGES)
414 				return (NULL);
415 			allocpages = minpages;
416 		}
417 		/* Speculate that the object may grow. */
418 	}
419 
420 	/*
421 	 * Allocate and populate the new reservations.  The alignment and
422 	 * boundary specified for this allocation may be different from the
423 	 * alignment and boundary specified for the requested pages.  For
424 	 * instance, the specified index may not be the first page within the
425 	 * first new reservation.
426 	 */
427 	m = vm_phys_alloc_contig(allocpages, low, high, ulmax(alignment,
428 	    VM_LEVEL_0_SIZE), boundary > VM_LEVEL_0_SIZE ? boundary : 0);
429 	if (m == NULL)
430 		return (NULL);
431 	m_ret = NULL;
432 	index = VM_RESERV_INDEX(object, pindex);
433 	do {
434 		rv = vm_reserv_from_page(m);
435 		KASSERT(rv->pages == m,
436 		    ("vm_reserv_alloc_contig: reserv %p's pages is corrupted",
437 		    rv));
438 		KASSERT(rv->object == NULL,
439 		    ("vm_reserv_alloc_contig: reserv %p isn't free", rv));
440 		LIST_INSERT_HEAD(&object->rvq, rv, objq);
441 		rv->object = object;
442 		rv->pindex = first;
443 		KASSERT(rv->popcnt == 0,
444 		    ("vm_reserv_alloc_contig: reserv %p's popcnt is corrupted",
445 		    rv));
446 		KASSERT(!rv->inpartpopq,
447 		    ("vm_reserv_alloc_contig: reserv %p's inpartpopq is TRUE",
448 		    rv));
449 		n = ulmin(VM_LEVEL_0_NPAGES - index, npages);
450 		for (i = 0; i < n; i++)
451 			vm_reserv_populate(rv);
452 		npages -= n;
453 		if (m_ret == NULL) {
454 			m_ret = &rv->pages[index];
455 			index = 0;
456 		}
457 		m += VM_LEVEL_0_NPAGES;
458 		first += VM_LEVEL_0_NPAGES;
459 		allocpages -= VM_LEVEL_0_NPAGES;
460 	} while (allocpages > 0);
461 	return (m_ret);
462 
463 	/*
464 	 * Found a matching reservation.
465 	 */
466 found:
467 	index = VM_RESERV_INDEX(object, pindex);
468 	/* Does the allocation fit within the reservation? */
469 	if (index + npages > VM_LEVEL_0_NPAGES)
470 		return (NULL);
471 	m = &rv->pages[index];
472 	pa = VM_PAGE_TO_PHYS(m);
473 	if (pa < low || pa + size > high || (pa & (alignment - 1)) != 0 ||
474 	    ((pa ^ (pa + size - 1)) & ~(boundary - 1)) != 0)
475 		return (NULL);
476 	/* Handle vm_page_rename(m, new_object, ...). */
477 	for (i = 0; i < npages; i++)
478 		if ((rv->pages[index + i].flags & (PG_CACHED | PG_FREE)) == 0)
479 			return (NULL);
480 	for (i = 0; i < npages; i++)
481 		vm_reserv_populate(rv);
482 	return (m);
483 }
484 
485 /*
486  * Allocates a page from an existing or newly-created reservation.
487  *
488  * The object and free page queue must be locked.
489  */
490 vm_page_t
491 vm_reserv_alloc_page(vm_object_t object, vm_pindex_t pindex)
492 {
493 	vm_page_t m, mpred, msucc;
494 	vm_pindex_t first, leftcap, rightcap;
495 	vm_reserv_t rv;
496 
497 	mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
498 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
499 
500 	/*
501 	 * Is a reservation fundamentally impossible?
502 	 */
503 	if (pindex < VM_RESERV_INDEX(object, pindex) ||
504 	    pindex >= object->size)
505 		return (NULL);
506 
507 	/*
508 	 * Look for an existing reservation.
509 	 */
510 	msucc = NULL;
511 	mpred = object->root;
512 	while (mpred != NULL) {
513 		KASSERT(mpred->pindex != pindex,
514 		    ("vm_reserv_alloc_page: pindex already allocated"));
515 		rv = vm_reserv_from_page(mpred);
516 		if (rv->object == object && vm_reserv_has_pindex(rv, pindex))
517 			goto found;
518 		else if (mpred->pindex < pindex) {
519 			if (msucc != NULL ||
520 			    (msucc = TAILQ_NEXT(mpred, listq)) == NULL)
521 				break;
522 			KASSERT(msucc->pindex != pindex,
523 			    ("vm_reserv_alloc_page: pindex already allocated"));
524 			rv = vm_reserv_from_page(msucc);
525 			if (rv->object == object &&
526 			    vm_reserv_has_pindex(rv, pindex))
527 				goto found;
528 			else if (pindex < msucc->pindex)
529 				break;
530 		} else if (msucc == NULL) {
531 			msucc = mpred;
532 			mpred = TAILQ_PREV(msucc, pglist, listq);
533 			continue;
534 		}
535 		msucc = NULL;
536 		mpred = object->root = vm_page_splay(pindex, object->root);
537 	}
538 
539 	/*
540 	 * Could a reservation fit between the first index to the left that
541 	 * can be used and the first index to the right that cannot be used?
542 	 */
543 	first = pindex - VM_RESERV_INDEX(object, pindex);
544 	if (mpred != NULL) {
545 		if ((rv = vm_reserv_from_page(mpred))->object != object)
546 			leftcap = mpred->pindex + 1;
547 		else
548 			leftcap = rv->pindex + VM_LEVEL_0_NPAGES;
549 		if (leftcap > first)
550 			return (NULL);
551 	}
552 	if (msucc != NULL) {
553 		if ((rv = vm_reserv_from_page(msucc))->object != object)
554 			rightcap = msucc->pindex;
555 		else
556 			rightcap = rv->pindex;
557 		if (first + VM_LEVEL_0_NPAGES > rightcap)
558 			return (NULL);
559 	}
560 
561 	/*
562 	 * Would a new reservation extend past the end of the object?
563 	 */
564 	if (first + VM_LEVEL_0_NPAGES > object->size) {
565 		/*
566 		 * Don't allocate a new reservation if the object is a vnode or
567 		 * backed by another object that is a vnode.
568 		 */
569 		if (object->type == OBJT_VNODE ||
570 		    (object->backing_object != NULL &&
571 		    object->backing_object->type == OBJT_VNODE))
572 			return (NULL);
573 		/* Speculate that the object may grow. */
574 	}
575 
576 	/*
577 	 * Allocate and populate the new reservation.
578 	 */
579 	m = vm_phys_alloc_pages(VM_FREEPOOL_DEFAULT, VM_LEVEL_0_ORDER);
580 	if (m == NULL)
581 		return (NULL);
582 	rv = vm_reserv_from_page(m);
583 	KASSERT(rv->pages == m,
584 	    ("vm_reserv_alloc_page: reserv %p's pages is corrupted", rv));
585 	KASSERT(rv->object == NULL,
586 	    ("vm_reserv_alloc_page: reserv %p isn't free", rv));
587 	LIST_INSERT_HEAD(&object->rvq, rv, objq);
588 	rv->object = object;
589 	rv->pindex = first;
590 	KASSERT(rv->popcnt == 0,
591 	    ("vm_reserv_alloc_page: reserv %p's popcnt is corrupted", rv));
592 	KASSERT(!rv->inpartpopq,
593 	    ("vm_reserv_alloc_page: reserv %p's inpartpopq is TRUE", rv));
594 	vm_reserv_populate(rv);
595 	return (&rv->pages[VM_RESERV_INDEX(object, pindex)]);
596 
597 	/*
598 	 * Found a matching reservation.
599 	 */
600 found:
601 	m = &rv->pages[VM_RESERV_INDEX(object, pindex)];
602 	/* Handle vm_page_rename(m, new_object, ...). */
603 	if ((m->flags & (PG_CACHED | PG_FREE)) == 0)
604 		return (NULL);
605 	vm_reserv_populate(rv);
606 	return (m);
607 }
608 
609 /*
610  * Breaks all reservations belonging to the given object.
611  */
612 void
613 vm_reserv_break_all(vm_object_t object)
614 {
615 	vm_reserv_t rv;
616 	int i;
617 
618 	mtx_lock(&vm_page_queue_free_mtx);
619 	while ((rv = LIST_FIRST(&object->rvq)) != NULL) {
620 		KASSERT(rv->object == object,
621 		    ("vm_reserv_break_all: reserv %p is corrupted", rv));
622 		if (rv->inpartpopq) {
623 			TAILQ_REMOVE(&vm_rvq_partpop, rv, partpopq);
624 			rv->inpartpopq = FALSE;
625 		}
626 		LIST_REMOVE(rv, objq);
627 		rv->object = NULL;
628 		for (i = 0; i < VM_LEVEL_0_NPAGES; i++) {
629 			if ((rv->pages[i].flags & (PG_CACHED | PG_FREE)) != 0)
630 				vm_phys_free_pages(&rv->pages[i], 0);
631 			else
632 				rv->popcnt--;
633 		}
634 		KASSERT(rv->popcnt == 0,
635 		    ("vm_reserv_break_all: reserv %p's popcnt is corrupted",
636 		    rv));
637 		vm_reserv_broken++;
638 	}
639 	mtx_unlock(&vm_page_queue_free_mtx);
640 }
641 
642 /*
643  * Frees the given page if it belongs to a reservation.  Returns TRUE if the
644  * page is freed and FALSE otherwise.
645  *
646  * The free page queue lock must be held.
647  */
648 boolean_t
649 vm_reserv_free_page(vm_page_t m)
650 {
651 	vm_reserv_t rv;
652 
653 	mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
654 	rv = vm_reserv_from_page(m);
655 	if (rv->object == NULL)
656 		return (FALSE);
657 	if ((m->flags & PG_CACHED) != 0 && m->pool != VM_FREEPOOL_CACHE)
658 		vm_phys_set_pool(VM_FREEPOOL_CACHE, rv->pages,
659 		    VM_LEVEL_0_ORDER);
660 	vm_reserv_depopulate(rv);
661 	return (TRUE);
662 }
663 
664 /*
665  * Initializes the reservation management system.  Specifically, initializes
666  * the reservation array.
667  *
668  * Requires that vm_page_array and first_page are initialized!
669  */
670 void
671 vm_reserv_init(void)
672 {
673 	vm_paddr_t paddr;
674 	int i;
675 
676 	/*
677 	 * Initialize the reservation array.  Specifically, initialize the
678 	 * "pages" field for every element that has an underlying superpage.
679 	 */
680 	for (i = 0; phys_avail[i + 1] != 0; i += 2) {
681 		paddr = roundup2(phys_avail[i], VM_LEVEL_0_SIZE);
682 		while (paddr + VM_LEVEL_0_SIZE <= phys_avail[i + 1]) {
683 			vm_reserv_array[paddr >> VM_LEVEL_0_SHIFT].pages =
684 			    PHYS_TO_VM_PAGE(paddr);
685 			paddr += VM_LEVEL_0_SIZE;
686 		}
687 	}
688 }
689 
690 /*
691  * Returns a reservation level if the given page belongs to a fully-populated
692  * reservation and -1 otherwise.
693  */
694 int
695 vm_reserv_level_iffullpop(vm_page_t m)
696 {
697 	vm_reserv_t rv;
698 
699 	rv = vm_reserv_from_page(m);
700 	return (rv->popcnt == VM_LEVEL_0_NPAGES ? 0 : -1);
701 }
702 
703 /*
704  * Prepare for the reactivation of a cached page.
705  *
706  * First, suppose that the given page "m" was allocated individually, i.e., not
707  * as part of a reservation, and cached.  Then, suppose a reservation
708  * containing "m" is allocated by the same object.  Although "m" and the
709  * reservation belong to the same object, "m"'s pindex may not match the
710  * reservation's.
711  *
712  * The free page queue must be locked.
713  */
714 boolean_t
715 vm_reserv_reactivate_page(vm_page_t m)
716 {
717 	vm_reserv_t rv;
718 	int i, m_index;
719 
720 	mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
721 	rv = vm_reserv_from_page(m);
722 	if (rv->object == NULL)
723 		return (FALSE);
724 	KASSERT((m->flags & PG_CACHED) != 0,
725 	    ("vm_reserv_uncache_page: page %p is not cached", m));
726 	if (m->object == rv->object &&
727 	    m->pindex - rv->pindex == VM_RESERV_INDEX(m->object, m->pindex))
728 		vm_reserv_populate(rv);
729 	else {
730 		KASSERT(rv->inpartpopq,
731 		    ("vm_reserv_uncache_page: reserv %p's inpartpopq is FALSE",
732 		    rv));
733 		TAILQ_REMOVE(&vm_rvq_partpop, rv, partpopq);
734 		rv->inpartpopq = FALSE;
735 		LIST_REMOVE(rv, objq);
736 		rv->object = NULL;
737 		/* Don't vm_phys_free_pages(m, 0). */
738 		m_index = m - rv->pages;
739 		for (i = 0; i < m_index; i++) {
740 			if ((rv->pages[i].flags & (PG_CACHED | PG_FREE)) != 0)
741 				vm_phys_free_pages(&rv->pages[i], 0);
742 			else
743 				rv->popcnt--;
744 		}
745 		for (i++; i < VM_LEVEL_0_NPAGES; i++) {
746 			if ((rv->pages[i].flags & (PG_CACHED | PG_FREE)) != 0)
747 				vm_phys_free_pages(&rv->pages[i], 0);
748 			else
749 				rv->popcnt--;
750 		}
751 		KASSERT(rv->popcnt == 0,
752 		    ("vm_reserv_uncache_page: reserv %p's popcnt is corrupted",
753 		    rv));
754 		vm_reserv_broken++;
755 	}
756 	return (TRUE);
757 }
758 
759 /*
760  * Breaks the given partially-populated reservation, releasing its cached and
761  * free pages to the physical memory allocator.
762  *
763  * The free page queue lock must be held.
764  */
765 static void
766 vm_reserv_reclaim(vm_reserv_t rv)
767 {
768 	int i;
769 
770 	mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
771 	KASSERT(rv->inpartpopq,
772 	    ("vm_reserv_reclaim: reserv %p's inpartpopq is corrupted", rv));
773 	TAILQ_REMOVE(&vm_rvq_partpop, rv, partpopq);
774 	rv->inpartpopq = FALSE;
775 	KASSERT(rv->object != NULL,
776 	    ("vm_reserv_reclaim: reserv %p is free", rv));
777 	LIST_REMOVE(rv, objq);
778 	rv->object = NULL;
779 	for (i = 0; i < VM_LEVEL_0_NPAGES; i++) {
780 		if ((rv->pages[i].flags & (PG_CACHED | PG_FREE)) != 0)
781 			vm_phys_free_pages(&rv->pages[i], 0);
782 		else
783 			rv->popcnt--;
784 	}
785 	KASSERT(rv->popcnt == 0,
786 	    ("vm_reserv_reclaim: reserv %p's popcnt is corrupted", rv));
787 	vm_reserv_reclaimed++;
788 }
789 
790 /*
791  * Breaks the reservation at the head of the partially-populated reservation
792  * queue, releasing its cached and free pages to the physical memory
793  * allocator.  Returns TRUE if a reservation is broken and FALSE otherwise.
794  *
795  * The free page queue lock must be held.
796  */
797 boolean_t
798 vm_reserv_reclaim_inactive(void)
799 {
800 	vm_reserv_t rv;
801 
802 	mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
803 	if ((rv = TAILQ_FIRST(&vm_rvq_partpop)) != NULL) {
804 		vm_reserv_reclaim(rv);
805 		return (TRUE);
806 	}
807 	return (FALSE);
808 }
809 
810 /*
811  * Searches the partially-populated reservation queue for the least recently
812  * active reservation with unused pages, i.e., cached or free, that satisfy the
813  * given request for contiguous physical memory.  If a satisfactory reservation
814  * is found, it is broken.  Returns TRUE if a reservation is broken and FALSE
815  * otherwise.
816  *
817  * The free page queue lock must be held.
818  */
819 boolean_t
820 vm_reserv_reclaim_contig(u_long npages, vm_paddr_t low, vm_paddr_t high,
821     u_long alignment, vm_paddr_t boundary)
822 {
823 	vm_paddr_t pa, pa_length, size;
824 	vm_reserv_t rv;
825 	int i;
826 
827 	mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
828 	if (npages > VM_LEVEL_0_NPAGES - 1)
829 		return (FALSE);
830 	size = npages << PAGE_SHIFT;
831 	TAILQ_FOREACH(rv, &vm_rvq_partpop, partpopq) {
832 		pa = VM_PAGE_TO_PHYS(&rv->pages[VM_LEVEL_0_NPAGES - 1]);
833 		if (pa + PAGE_SIZE - size < low) {
834 			/* this entire reservation is too low; go to next */
835 			continue;
836 		}
837 		pa_length = 0;
838 		for (i = 0; i < VM_LEVEL_0_NPAGES; i++)
839 			if ((rv->pages[i].flags & (PG_CACHED | PG_FREE)) != 0) {
840 				pa_length += PAGE_SIZE;
841 				if (pa_length == PAGE_SIZE) {
842 					pa = VM_PAGE_TO_PHYS(&rv->pages[i]);
843 					if (pa + size > high) {
844 						/* skip to next reservation */
845 						break;
846 					} else if (pa < low ||
847 					    (pa & (alignment - 1)) != 0 ||
848 					    ((pa ^ (pa + size - 1)) &
849 					    ~(boundary - 1)) != 0)
850 						pa_length = 0;
851 				}
852 				if (pa_length >= size) {
853 					vm_reserv_reclaim(rv);
854 					return (TRUE);
855 				}
856 			} else
857 				pa_length = 0;
858 	}
859 	return (FALSE);
860 }
861 
862 /*
863  * Transfers the reservation underlying the given page to a new object.
864  *
865  * The object must be locked.
866  */
867 void
868 vm_reserv_rename(vm_page_t m, vm_object_t new_object, vm_object_t old_object,
869     vm_pindex_t old_object_offset)
870 {
871 	vm_reserv_t rv;
872 
873 	VM_OBJECT_LOCK_ASSERT(new_object, MA_OWNED);
874 	rv = vm_reserv_from_page(m);
875 	if (rv->object == old_object) {
876 		mtx_lock(&vm_page_queue_free_mtx);
877 		if (rv->object == old_object) {
878 			LIST_REMOVE(rv, objq);
879 			LIST_INSERT_HEAD(&new_object->rvq, rv, objq);
880 			rv->object = new_object;
881 			rv->pindex -= old_object_offset;
882 		}
883 		mtx_unlock(&vm_page_queue_free_mtx);
884 	}
885 }
886 
887 /*
888  * Allocates the virtual and physical memory required by the reservation
889  * management system's data structures, in particular, the reservation array.
890  */
891 vm_paddr_t
892 vm_reserv_startup(vm_offset_t *vaddr, vm_paddr_t end, vm_paddr_t high_water)
893 {
894 	vm_paddr_t new_end;
895 	size_t size;
896 
897 	/*
898 	 * Calculate the size (in bytes) of the reservation array.  Round up
899 	 * from "high_water" because every small page is mapped to an element
900 	 * in the reservation array based on its physical address.  Thus, the
901 	 * number of elements in the reservation array can be greater than the
902 	 * number of superpages.
903 	 */
904 	size = howmany(high_water, VM_LEVEL_0_SIZE) * sizeof(struct vm_reserv);
905 
906 	/*
907 	 * Allocate and map the physical memory for the reservation array.  The
908 	 * next available virtual address is returned by reference.
909 	 */
910 	new_end = end - round_page(size);
911 	vm_reserv_array = (void *)(uintptr_t)pmap_map(vaddr, new_end, end,
912 	    VM_PROT_READ | VM_PROT_WRITE);
913 	bzero(vm_reserv_array, size);
914 
915 	/*
916 	 * Return the next available physical address.
917 	 */
918 	return (new_end);
919 }
920 
921 #endif	/* VM_NRESERVLEVEL > 0 */
922