xref: /freebsd/sys/vm/vm_reserv.c (revision c3755aa30cbddc30cbdc26707aac2606e9cd6ec5)
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/rwlock.h>
51 #include <sys/sbuf.h>
52 #include <sys/sysctl.h>
53 #include <sys/systm.h>
54 
55 #include <vm/vm.h>
56 #include <vm/vm_param.h>
57 #include <vm/vm_object.h>
58 #include <vm/vm_page.h>
59 #include <vm/vm_phys.h>
60 #include <vm/vm_radix.h>
61 #include <vm/vm_reserv.h>
62 
63 /*
64  * The reservation system supports the speculative allocation of large physical
65  * pages ("superpages").  Speculative allocation enables the fully-automatic
66  * utilization of superpages by the virtual memory system.  In other words, no
67  * programmatic directives are required to use superpages.
68  */
69 
70 #if VM_NRESERVLEVEL > 0
71 
72 /*
73  * The number of small pages that are contained in a level 0 reservation
74  */
75 #define	VM_LEVEL_0_NPAGES	(1 << VM_LEVEL_0_ORDER)
76 
77 /*
78  * The number of bits by which a physical address is shifted to obtain the
79  * reservation number
80  */
81 #define	VM_LEVEL_0_SHIFT	(VM_LEVEL_0_ORDER + PAGE_SHIFT)
82 
83 /*
84  * The size of a level 0 reservation in bytes
85  */
86 #define	VM_LEVEL_0_SIZE		(1 << VM_LEVEL_0_SHIFT)
87 
88 /*
89  * Computes the index of the small page underlying the given (object, pindex)
90  * within the reservation's array of small pages.
91  */
92 #define	VM_RESERV_INDEX(object, pindex)	\
93     (((object)->pg_color + (pindex)) & (VM_LEVEL_0_NPAGES - 1))
94 
95 /*
96  * The reservation structure
97  *
98  * A reservation structure is constructed whenever a large physical page is
99  * speculatively allocated to an object.  The reservation provides the small
100  * physical pages for the range [pindex, pindex + VM_LEVEL_0_NPAGES) of offsets
101  * within that object.  The reservation's "popcnt" tracks the number of these
102  * small physical pages that are in use at any given time.  When and if the
103  * reservation is not fully utilized, it appears in the queue of partially-
104  * populated reservations.  The reservation always appears on the containing
105  * object's list of reservations.
106  *
107  * A partially-populated reservation can be broken and reclaimed at any time.
108  */
109 struct vm_reserv {
110 	TAILQ_ENTRY(vm_reserv) partpopq;
111 	LIST_ENTRY(vm_reserv) objq;
112 	vm_object_t	object;			/* containing object */
113 	vm_pindex_t	pindex;			/* offset within object */
114 	vm_page_t	pages;			/* first page of a superpage */
115 	int		popcnt;			/* # of pages in use */
116 	char		inpartpopq;
117 };
118 
119 /*
120  * The reservation array
121  *
122  * This array is analoguous in function to vm_page_array.  It differs in the
123  * respect that it may contain a greater number of useful reservation
124  * structures than there are (physical) superpages.  These "invalid"
125  * reservation structures exist to trade-off space for time in the
126  * implementation of vm_reserv_from_page().  Invalid reservation structures are
127  * distinguishable from "valid" reservation structures by inspecting the
128  * reservation's "pages" field.  Invalid reservation structures have a NULL
129  * "pages" field.
130  *
131  * vm_reserv_from_page() maps a small (physical) page to an element of this
132  * array by computing a physical reservation number from the page's physical
133  * address.  The physical reservation number is used as the array index.
134  *
135  * An "active" reservation is a valid reservation structure that has a non-NULL
136  * "object" field and a non-zero "popcnt" field.  In other words, every active
137  * reservation belongs to a particular object.  Moreover, every active
138  * reservation has an entry in the containing object's list of reservations.
139  */
140 static vm_reserv_t vm_reserv_array;
141 
142 /*
143  * The partially-populated reservation queue
144  *
145  * This queue enables the fast recovery of an unused cached or free small page
146  * from a partially-populated reservation.  The reservation at the head of
147  * this queue is the least-recently-changed, partially-populated reservation.
148  *
149  * Access to this queue is synchronized by the free page queue lock.
150  */
151 static TAILQ_HEAD(, vm_reserv) vm_rvq_partpop =
152 			    TAILQ_HEAD_INITIALIZER(vm_rvq_partpop);
153 
154 static SYSCTL_NODE(_vm, OID_AUTO, reserv, CTLFLAG_RD, 0, "Reservation Info");
155 
156 static long vm_reserv_broken;
157 SYSCTL_LONG(_vm_reserv, OID_AUTO, broken, CTLFLAG_RD,
158     &vm_reserv_broken, 0, "Cumulative number of broken reservations");
159 
160 static long vm_reserv_freed;
161 SYSCTL_LONG(_vm_reserv, OID_AUTO, freed, CTLFLAG_RD,
162     &vm_reserv_freed, 0, "Cumulative number of freed reservations");
163 
164 static int sysctl_vm_reserv_partpopq(SYSCTL_HANDLER_ARGS);
165 
166 SYSCTL_OID(_vm_reserv, OID_AUTO, partpopq, CTLTYPE_STRING | CTLFLAG_RD, NULL, 0,
167     sysctl_vm_reserv_partpopq, "A", "Partially-populated reservation queues");
168 
169 static long vm_reserv_reclaimed;
170 SYSCTL_LONG(_vm_reserv, OID_AUTO, reclaimed, CTLFLAG_RD,
171     &vm_reserv_reclaimed, 0, "Cumulative number of reclaimed reservations");
172 
173 static void		vm_reserv_depopulate(vm_reserv_t rv);
174 static vm_reserv_t	vm_reserv_from_page(vm_page_t m);
175 static boolean_t	vm_reserv_has_pindex(vm_reserv_t rv,
176 			    vm_pindex_t pindex);
177 static void		vm_reserv_populate(vm_reserv_t rv);
178 static void		vm_reserv_reclaim(vm_reserv_t rv);
179 
180 /*
181  * Describes the current state of the partially-populated reservation queue.
182  */
183 static int
184 sysctl_vm_reserv_partpopq(SYSCTL_HANDLER_ARGS)
185 {
186 	struct sbuf sbuf;
187 	vm_reserv_t rv;
188 	int counter, error, level, unused_pages;
189 
190 	error = sysctl_wire_old_buffer(req, 0);
191 	if (error != 0)
192 		return (error);
193 	sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
194 	sbuf_printf(&sbuf, "\nLEVEL     SIZE  NUMBER\n\n");
195 	for (level = -1; level <= VM_NRESERVLEVEL - 2; level++) {
196 		counter = 0;
197 		unused_pages = 0;
198 		mtx_lock(&vm_page_queue_free_mtx);
199 		TAILQ_FOREACH(rv, &vm_rvq_partpop/*[level]*/, partpopq) {
200 			counter++;
201 			unused_pages += VM_LEVEL_0_NPAGES - rv->popcnt;
202 		}
203 		mtx_unlock(&vm_page_queue_free_mtx);
204 		sbuf_printf(&sbuf, "%5d: %6dK, %6d\n", level,
205 		    unused_pages * ((int)PAGE_SIZE / 1024), counter);
206 	}
207 	error = sbuf_finish(&sbuf);
208 	sbuf_delete(&sbuf);
209 	return (error);
210 }
211 
212 /*
213  * Reduces the given reservation's population count.  If the population count
214  * becomes zero, the reservation is destroyed.  Additionally, moves the
215  * reservation to the tail of the partially-populated reservations queue if the
216  * population count is non-zero.
217  *
218  * The free page queue lock must be held.
219  */
220 static void
221 vm_reserv_depopulate(vm_reserv_t rv)
222 {
223 
224 	mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
225 	KASSERT(rv->object != NULL,
226 	    ("vm_reserv_depopulate: reserv %p is free", rv));
227 	KASSERT(rv->popcnt > 0,
228 	    ("vm_reserv_depopulate: reserv %p's popcnt is corrupted", rv));
229 	if (rv->inpartpopq) {
230 		TAILQ_REMOVE(&vm_rvq_partpop, rv, partpopq);
231 		rv->inpartpopq = FALSE;
232 	}
233 	rv->popcnt--;
234 	if (rv->popcnt == 0) {
235 		LIST_REMOVE(rv, objq);
236 		rv->object = NULL;
237 		vm_phys_free_pages(rv->pages, VM_LEVEL_0_ORDER);
238 		vm_reserv_freed++;
239 	} else {
240 		rv->inpartpopq = TRUE;
241 		TAILQ_INSERT_TAIL(&vm_rvq_partpop, rv, partpopq);
242 	}
243 }
244 
245 /*
246  * Returns the reservation to which the given page might belong.
247  */
248 static __inline vm_reserv_t
249 vm_reserv_from_page(vm_page_t m)
250 {
251 
252 	return (&vm_reserv_array[VM_PAGE_TO_PHYS(m) >> VM_LEVEL_0_SHIFT]);
253 }
254 
255 /*
256  * Returns TRUE if the given reservation contains the given page index and
257  * FALSE otherwise.
258  */
259 static __inline boolean_t
260 vm_reserv_has_pindex(vm_reserv_t rv, vm_pindex_t pindex)
261 {
262 
263 	return (((pindex - rv->pindex) & ~(VM_LEVEL_0_NPAGES - 1)) == 0);
264 }
265 
266 /*
267  * Increases the given reservation's population count.  Moves the reservation
268  * to the tail of the partially-populated reservation queue.
269  *
270  * The free page queue must be locked.
271  */
272 static void
273 vm_reserv_populate(vm_reserv_t rv)
274 {
275 
276 	mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
277 	KASSERT(rv->object != NULL,
278 	    ("vm_reserv_populate: reserv %p is free", rv));
279 	KASSERT(rv->popcnt < VM_LEVEL_0_NPAGES,
280 	    ("vm_reserv_populate: reserv %p is already full", rv));
281 	if (rv->inpartpopq) {
282 		TAILQ_REMOVE(&vm_rvq_partpop, rv, partpopq);
283 		rv->inpartpopq = FALSE;
284 	}
285 	rv->popcnt++;
286 	if (rv->popcnt < VM_LEVEL_0_NPAGES) {
287 		rv->inpartpopq = TRUE;
288 		TAILQ_INSERT_TAIL(&vm_rvq_partpop, rv, partpopq);
289 	}
290 }
291 
292 /*
293  * Allocates a contiguous set of physical pages of the given size "npages"
294  * from an existing or newly-created reservation.  All of the physical pages
295  * must be at or above the given physical address "low" and below the given
296  * physical address "high".  The given value "alignment" determines the
297  * alignment of the first physical page in the set.  If the given value
298  * "boundary" is non-zero, then the set of physical pages cannot cross any
299  * physical address boundary that is a multiple of that value.  Both
300  * "alignment" and "boundary" must be a power of two.
301  *
302  * The object and free page queue must be locked.
303  */
304 vm_page_t
305 vm_reserv_alloc_contig(vm_object_t object, vm_pindex_t pindex, u_long npages,
306     vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary)
307 {
308 	vm_paddr_t pa, size;
309 	vm_page_t m, m_ret, mpred, msucc;
310 	vm_pindex_t first, leftcap, rightcap;
311 	vm_reserv_t rv;
312 	u_long allocpages, maxpages, minpages;
313 	int i, index, n;
314 
315 	mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
316 	VM_OBJECT_ASSERT_WLOCKED(object);
317 	KASSERT(npages != 0, ("vm_reserv_alloc_contig: npages is 0"));
318 
319 	/*
320 	 * Is a reservation fundamentally impossible?
321 	 */
322 	if (pindex < VM_RESERV_INDEX(object, pindex) ||
323 	    pindex + npages > object->size)
324 		return (NULL);
325 
326 	/*
327 	 * All reservations of a particular size have the same alignment.
328 	 * Assuming that the first page is allocated from a reservation, the
329 	 * least significant bits of its physical address can be determined
330 	 * from its offset from the beginning of the reservation and the size
331 	 * of the reservation.
332 	 *
333 	 * Could the specified index within a reservation of the smallest
334 	 * possible size satisfy the alignment and boundary requirements?
335 	 */
336 	pa = VM_RESERV_INDEX(object, pindex) << PAGE_SHIFT;
337 	if ((pa & (alignment - 1)) != 0)
338 		return (NULL);
339 	size = npages << PAGE_SHIFT;
340 	if (((pa ^ (pa + size - 1)) & ~(boundary - 1)) != 0)
341 		return (NULL);
342 
343 	/*
344 	 * Look for an existing reservation.
345 	 */
346 	mpred = vm_radix_lookup_le(&object->rtree, pindex);
347 	if (mpred != NULL) {
348 		KASSERT(mpred->pindex < pindex,
349 		    ("vm_reserv_alloc_contig: pindex already allocated"));
350 		rv = vm_reserv_from_page(mpred);
351 		if (rv->object == object && vm_reserv_has_pindex(rv, pindex))
352 			goto found;
353 		msucc = TAILQ_NEXT(mpred, listq);
354 	} else
355 		msucc = TAILQ_FIRST(&object->memq);
356 	if (msucc != NULL) {
357 		KASSERT(msucc->pindex > pindex,
358 		    ("vm_reserv_alloc_page: pindex already allocated"));
359 		rv = vm_reserv_from_page(msucc);
360 		if (rv->object == object && vm_reserv_has_pindex(rv, pindex))
361 			goto found;
362 	}
363 
364 	/*
365 	 * Could at least one reservation fit between the first index to the
366 	 * left that can be used and the first index to the right that cannot
367 	 * be used?
368 	 */
369 	first = pindex - VM_RESERV_INDEX(object, pindex);
370 	if (mpred != NULL) {
371 		if ((rv = vm_reserv_from_page(mpred))->object != object)
372 			leftcap = mpred->pindex + 1;
373 		else
374 			leftcap = rv->pindex + VM_LEVEL_0_NPAGES;
375 		if (leftcap > first)
376 			return (NULL);
377 	}
378 	minpages = VM_RESERV_INDEX(object, pindex) + npages;
379 	maxpages = roundup2(minpages, VM_LEVEL_0_NPAGES);
380 	allocpages = maxpages;
381 	if (msucc != NULL) {
382 		if ((rv = vm_reserv_from_page(msucc))->object != object)
383 			rightcap = msucc->pindex;
384 		else
385 			rightcap = rv->pindex;
386 		if (first + maxpages > rightcap) {
387 			if (maxpages == VM_LEVEL_0_NPAGES)
388 				return (NULL);
389 			allocpages = minpages;
390 		}
391 	}
392 
393 	/*
394 	 * Would the last new reservation extend past the end of the object?
395 	 */
396 	if (first + maxpages > object->size) {
397 		/*
398 		 * Don't allocate the last new reservation if the object is a
399 		 * vnode or backed by another object that is a vnode.
400 		 */
401 		if (object->type == OBJT_VNODE ||
402 		    (object->backing_object != NULL &&
403 		    object->backing_object->type == OBJT_VNODE)) {
404 			if (maxpages == VM_LEVEL_0_NPAGES)
405 				return (NULL);
406 			allocpages = minpages;
407 		}
408 		/* Speculate that the object may grow. */
409 	}
410 
411 	/*
412 	 * Allocate and populate the new reservations.  The alignment and
413 	 * boundary specified for this allocation may be different from the
414 	 * alignment and boundary specified for the requested pages.  For
415 	 * instance, the specified index may not be the first page within the
416 	 * first new reservation.
417 	 */
418 	m = vm_phys_alloc_contig(allocpages, low, high, ulmax(alignment,
419 	    VM_LEVEL_0_SIZE), boundary > VM_LEVEL_0_SIZE ? boundary : 0);
420 	if (m == NULL)
421 		return (NULL);
422 	m_ret = NULL;
423 	index = VM_RESERV_INDEX(object, pindex);
424 	do {
425 		rv = vm_reserv_from_page(m);
426 		KASSERT(rv->pages == m,
427 		    ("vm_reserv_alloc_contig: reserv %p's pages is corrupted",
428 		    rv));
429 		KASSERT(rv->object == NULL,
430 		    ("vm_reserv_alloc_contig: reserv %p isn't free", rv));
431 		LIST_INSERT_HEAD(&object->rvq, rv, objq);
432 		rv->object = object;
433 		rv->pindex = first;
434 		KASSERT(rv->popcnt == 0,
435 		    ("vm_reserv_alloc_contig: reserv %p's popcnt is corrupted",
436 		    rv));
437 		KASSERT(!rv->inpartpopq,
438 		    ("vm_reserv_alloc_contig: reserv %p's inpartpopq is TRUE",
439 		    rv));
440 		n = ulmin(VM_LEVEL_0_NPAGES - index, npages);
441 		for (i = 0; i < n; i++)
442 			vm_reserv_populate(rv);
443 		npages -= n;
444 		if (m_ret == NULL) {
445 			m_ret = &rv->pages[index];
446 			index = 0;
447 		}
448 		m += VM_LEVEL_0_NPAGES;
449 		first += VM_LEVEL_0_NPAGES;
450 		allocpages -= VM_LEVEL_0_NPAGES;
451 	} while (allocpages > 0);
452 	return (m_ret);
453 
454 	/*
455 	 * Found a matching reservation.
456 	 */
457 found:
458 	index = VM_RESERV_INDEX(object, pindex);
459 	/* Does the allocation fit within the reservation? */
460 	if (index + npages > VM_LEVEL_0_NPAGES)
461 		return (NULL);
462 	m = &rv->pages[index];
463 	pa = VM_PAGE_TO_PHYS(m);
464 	if (pa < low || pa + size > high || (pa & (alignment - 1)) != 0 ||
465 	    ((pa ^ (pa + size - 1)) & ~(boundary - 1)) != 0)
466 		return (NULL);
467 	/* Handle vm_page_rename(m, new_object, ...). */
468 	for (i = 0; i < npages; i++)
469 		if ((rv->pages[index + i].flags & (PG_CACHED | PG_FREE)) == 0)
470 			return (NULL);
471 	for (i = 0; i < npages; i++)
472 		vm_reserv_populate(rv);
473 	return (m);
474 }
475 
476 /*
477  * Allocates a page from an existing or newly-created reservation.
478  *
479  * The page "mpred" must immediately precede the offset "pindex" within the
480  * specified object.
481  *
482  * The object and free page queue must be locked.
483  */
484 vm_page_t
485 vm_reserv_alloc_page(vm_object_t object, vm_pindex_t pindex, vm_page_t mpred)
486 {
487 	vm_page_t m, msucc;
488 	vm_pindex_t first, leftcap, rightcap;
489 	vm_reserv_t rv;
490 
491 	mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
492 	VM_OBJECT_ASSERT_WLOCKED(object);
493 
494 	/*
495 	 * Is a reservation fundamentally impossible?
496 	 */
497 	if (pindex < VM_RESERV_INDEX(object, pindex) ||
498 	    pindex >= object->size)
499 		return (NULL);
500 
501 	/*
502 	 * Look for an existing reservation.
503 	 */
504 	if (mpred != NULL) {
505 		KASSERT(mpred->object == object ||
506 		    (mpred->flags & PG_SLAB) != 0,
507 		    ("vm_reserv_alloc_page: object doesn't contain mpred"));
508 		KASSERT(mpred->pindex < pindex,
509 		    ("vm_reserv_alloc_page: mpred doesn't precede pindex"));
510 		rv = vm_reserv_from_page(mpred);
511 		if (rv->object == object && vm_reserv_has_pindex(rv, pindex))
512 			goto found;
513 		msucc = TAILQ_NEXT(mpred, listq);
514 	} else
515 		msucc = TAILQ_FIRST(&object->memq);
516 	if (msucc != NULL) {
517 		KASSERT(msucc->pindex > pindex,
518 		    ("vm_reserv_alloc_page: msucc doesn't succeed pindex"));
519 		rv = vm_reserv_from_page(msucc);
520 		if (rv->object == object && vm_reserv_has_pindex(rv, pindex))
521 			goto found;
522 	}
523 
524 	/*
525 	 * Could a reservation fit between the first index to the left that
526 	 * can be used and the first index to the right that cannot be used?
527 	 */
528 	first = pindex - VM_RESERV_INDEX(object, pindex);
529 	if (mpred != NULL) {
530 		if ((rv = vm_reserv_from_page(mpred))->object != object)
531 			leftcap = mpred->pindex + 1;
532 		else
533 			leftcap = rv->pindex + VM_LEVEL_0_NPAGES;
534 		if (leftcap > first)
535 			return (NULL);
536 	}
537 	if (msucc != NULL) {
538 		if ((rv = vm_reserv_from_page(msucc))->object != object)
539 			rightcap = msucc->pindex;
540 		else
541 			rightcap = rv->pindex;
542 		if (first + VM_LEVEL_0_NPAGES > rightcap)
543 			return (NULL);
544 	}
545 
546 	/*
547 	 * Would a new reservation extend past the end of the object?
548 	 */
549 	if (first + VM_LEVEL_0_NPAGES > object->size) {
550 		/*
551 		 * Don't allocate a new reservation if the object is a vnode or
552 		 * backed by another object that is a vnode.
553 		 */
554 		if (object->type == OBJT_VNODE ||
555 		    (object->backing_object != NULL &&
556 		    object->backing_object->type == OBJT_VNODE))
557 			return (NULL);
558 		/* Speculate that the object may grow. */
559 	}
560 
561 	/*
562 	 * Allocate and populate the new reservation.
563 	 */
564 	m = vm_phys_alloc_pages(VM_FREEPOOL_DEFAULT, VM_LEVEL_0_ORDER);
565 	if (m == NULL)
566 		return (NULL);
567 	rv = vm_reserv_from_page(m);
568 	KASSERT(rv->pages == m,
569 	    ("vm_reserv_alloc_page: reserv %p's pages is corrupted", rv));
570 	KASSERT(rv->object == NULL,
571 	    ("vm_reserv_alloc_page: reserv %p isn't free", rv));
572 	LIST_INSERT_HEAD(&object->rvq, rv, objq);
573 	rv->object = object;
574 	rv->pindex = first;
575 	KASSERT(rv->popcnt == 0,
576 	    ("vm_reserv_alloc_page: reserv %p's popcnt is corrupted", rv));
577 	KASSERT(!rv->inpartpopq,
578 	    ("vm_reserv_alloc_page: reserv %p's inpartpopq is TRUE", rv));
579 	vm_reserv_populate(rv);
580 	return (&rv->pages[VM_RESERV_INDEX(object, pindex)]);
581 
582 	/*
583 	 * Found a matching reservation.
584 	 */
585 found:
586 	m = &rv->pages[VM_RESERV_INDEX(object, pindex)];
587 	/* Handle vm_page_rename(m, new_object, ...). */
588 	if ((m->flags & (PG_CACHED | PG_FREE)) == 0)
589 		return (NULL);
590 	vm_reserv_populate(rv);
591 	return (m);
592 }
593 
594 /*
595  * Breaks all reservations belonging to the given object.
596  */
597 void
598 vm_reserv_break_all(vm_object_t object)
599 {
600 	vm_reserv_t rv;
601 	int i;
602 
603 	mtx_lock(&vm_page_queue_free_mtx);
604 	while ((rv = LIST_FIRST(&object->rvq)) != NULL) {
605 		KASSERT(rv->object == object,
606 		    ("vm_reserv_break_all: reserv %p is corrupted", rv));
607 		if (rv->inpartpopq) {
608 			TAILQ_REMOVE(&vm_rvq_partpop, rv, partpopq);
609 			rv->inpartpopq = FALSE;
610 		}
611 		LIST_REMOVE(rv, objq);
612 		rv->object = NULL;
613 		for (i = 0; i < VM_LEVEL_0_NPAGES; i++) {
614 			if ((rv->pages[i].flags & (PG_CACHED | PG_FREE)) != 0)
615 				vm_phys_free_pages(&rv->pages[i], 0);
616 			else
617 				rv->popcnt--;
618 		}
619 		KASSERT(rv->popcnt == 0,
620 		    ("vm_reserv_break_all: reserv %p's popcnt is corrupted",
621 		    rv));
622 		vm_reserv_broken++;
623 	}
624 	mtx_unlock(&vm_page_queue_free_mtx);
625 }
626 
627 /*
628  * Frees the given page if it belongs to a reservation.  Returns TRUE if the
629  * page is freed and FALSE otherwise.
630  *
631  * The free page queue lock must be held.
632  */
633 boolean_t
634 vm_reserv_free_page(vm_page_t m)
635 {
636 	vm_reserv_t rv;
637 
638 	mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
639 	rv = vm_reserv_from_page(m);
640 	if (rv->object == NULL)
641 		return (FALSE);
642 	if ((m->flags & PG_CACHED) != 0 && m->pool != VM_FREEPOOL_CACHE)
643 		vm_phys_set_pool(VM_FREEPOOL_CACHE, rv->pages,
644 		    VM_LEVEL_0_ORDER);
645 	vm_reserv_depopulate(rv);
646 	return (TRUE);
647 }
648 
649 /*
650  * Initializes the reservation management system.  Specifically, initializes
651  * the reservation array.
652  *
653  * Requires that vm_page_array and first_page are initialized!
654  */
655 void
656 vm_reserv_init(void)
657 {
658 	vm_paddr_t paddr;
659 	int i;
660 
661 	/*
662 	 * Initialize the reservation array.  Specifically, initialize the
663 	 * "pages" field for every element that has an underlying superpage.
664 	 */
665 	for (i = 0; phys_avail[i + 1] != 0; i += 2) {
666 		paddr = roundup2(phys_avail[i], VM_LEVEL_0_SIZE);
667 		while (paddr + VM_LEVEL_0_SIZE <= phys_avail[i + 1]) {
668 			vm_reserv_array[paddr >> VM_LEVEL_0_SHIFT].pages =
669 			    PHYS_TO_VM_PAGE(paddr);
670 			paddr += VM_LEVEL_0_SIZE;
671 		}
672 	}
673 }
674 
675 /*
676  * Returns a reservation level if the given page belongs to a fully-populated
677  * reservation and -1 otherwise.
678  */
679 int
680 vm_reserv_level_iffullpop(vm_page_t m)
681 {
682 	vm_reserv_t rv;
683 
684 	rv = vm_reserv_from_page(m);
685 	return (rv->popcnt == VM_LEVEL_0_NPAGES ? 0 : -1);
686 }
687 
688 /*
689  * Prepare for the reactivation of a cached page.
690  *
691  * First, suppose that the given page "m" was allocated individually, i.e., not
692  * as part of a reservation, and cached.  Then, suppose a reservation
693  * containing "m" is allocated by the same object.  Although "m" and the
694  * reservation belong to the same object, "m"'s pindex may not match the
695  * reservation's.
696  *
697  * The free page queue must be locked.
698  */
699 boolean_t
700 vm_reserv_reactivate_page(vm_page_t m)
701 {
702 	vm_reserv_t rv;
703 	int i, m_index;
704 
705 	mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
706 	rv = vm_reserv_from_page(m);
707 	if (rv->object == NULL)
708 		return (FALSE);
709 	KASSERT((m->flags & PG_CACHED) != 0,
710 	    ("vm_reserv_uncache_page: page %p is not cached", m));
711 	if (m->object == rv->object &&
712 	    m->pindex - rv->pindex == VM_RESERV_INDEX(m->object, m->pindex))
713 		vm_reserv_populate(rv);
714 	else {
715 		KASSERT(rv->inpartpopq,
716 		    ("vm_reserv_uncache_page: reserv %p's inpartpopq is FALSE",
717 		    rv));
718 		TAILQ_REMOVE(&vm_rvq_partpop, rv, partpopq);
719 		rv->inpartpopq = FALSE;
720 		LIST_REMOVE(rv, objq);
721 		rv->object = NULL;
722 		/* Don't vm_phys_free_pages(m, 0). */
723 		m_index = m - rv->pages;
724 		for (i = 0; i < m_index; i++) {
725 			if ((rv->pages[i].flags & (PG_CACHED | PG_FREE)) != 0)
726 				vm_phys_free_pages(&rv->pages[i], 0);
727 			else
728 				rv->popcnt--;
729 		}
730 		for (i++; i < VM_LEVEL_0_NPAGES; i++) {
731 			if ((rv->pages[i].flags & (PG_CACHED | PG_FREE)) != 0)
732 				vm_phys_free_pages(&rv->pages[i], 0);
733 			else
734 				rv->popcnt--;
735 		}
736 		KASSERT(rv->popcnt == 0,
737 		    ("vm_reserv_uncache_page: reserv %p's popcnt is corrupted",
738 		    rv));
739 		vm_reserv_broken++;
740 	}
741 	return (TRUE);
742 }
743 
744 /*
745  * Breaks the given partially-populated reservation, releasing its cached and
746  * free pages to the physical memory allocator.
747  *
748  * The free page queue lock must be held.
749  */
750 static void
751 vm_reserv_reclaim(vm_reserv_t rv)
752 {
753 	int i;
754 
755 	mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
756 	KASSERT(rv->inpartpopq,
757 	    ("vm_reserv_reclaim: reserv %p's inpartpopq is corrupted", rv));
758 	TAILQ_REMOVE(&vm_rvq_partpop, rv, partpopq);
759 	rv->inpartpopq = FALSE;
760 	KASSERT(rv->object != NULL,
761 	    ("vm_reserv_reclaim: reserv %p is free", rv));
762 	LIST_REMOVE(rv, objq);
763 	rv->object = NULL;
764 	for (i = 0; i < VM_LEVEL_0_NPAGES; i++) {
765 		if ((rv->pages[i].flags & (PG_CACHED | PG_FREE)) != 0)
766 			vm_phys_free_pages(&rv->pages[i], 0);
767 		else
768 			rv->popcnt--;
769 	}
770 	KASSERT(rv->popcnt == 0,
771 	    ("vm_reserv_reclaim: reserv %p's popcnt is corrupted", rv));
772 	vm_reserv_reclaimed++;
773 }
774 
775 /*
776  * Breaks the reservation at the head of the partially-populated reservation
777  * queue, releasing its cached and free pages to the physical memory
778  * allocator.  Returns TRUE if a reservation is broken and FALSE otherwise.
779  *
780  * The free page queue lock must be held.
781  */
782 boolean_t
783 vm_reserv_reclaim_inactive(void)
784 {
785 	vm_reserv_t rv;
786 
787 	mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
788 	if ((rv = TAILQ_FIRST(&vm_rvq_partpop)) != NULL) {
789 		vm_reserv_reclaim(rv);
790 		return (TRUE);
791 	}
792 	return (FALSE);
793 }
794 
795 /*
796  * Searches the partially-populated reservation queue for the least recently
797  * active reservation with unused pages, i.e., cached or free, that satisfy the
798  * given request for contiguous physical memory.  If a satisfactory reservation
799  * is found, it is broken.  Returns TRUE if a reservation is broken and FALSE
800  * otherwise.
801  *
802  * The free page queue lock must be held.
803  */
804 boolean_t
805 vm_reserv_reclaim_contig(u_long npages, vm_paddr_t low, vm_paddr_t high,
806     u_long alignment, vm_paddr_t boundary)
807 {
808 	vm_paddr_t pa, pa_length, size;
809 	vm_reserv_t rv;
810 	int i;
811 
812 	mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
813 	if (npages > VM_LEVEL_0_NPAGES - 1)
814 		return (FALSE);
815 	size = npages << PAGE_SHIFT;
816 	TAILQ_FOREACH(rv, &vm_rvq_partpop, partpopq) {
817 		pa = VM_PAGE_TO_PHYS(&rv->pages[VM_LEVEL_0_NPAGES - 1]);
818 		if (pa + PAGE_SIZE - size < low) {
819 			/* this entire reservation is too low; go to next */
820 			continue;
821 		}
822 		pa_length = 0;
823 		for (i = 0; i < VM_LEVEL_0_NPAGES; i++)
824 			if ((rv->pages[i].flags & (PG_CACHED | PG_FREE)) != 0) {
825 				pa_length += PAGE_SIZE;
826 				if (pa_length == PAGE_SIZE) {
827 					pa = VM_PAGE_TO_PHYS(&rv->pages[i]);
828 					if (pa + size > high) {
829 						/* skip to next reservation */
830 						break;
831 					} else if (pa < low ||
832 					    (pa & (alignment - 1)) != 0 ||
833 					    ((pa ^ (pa + size - 1)) &
834 					    ~(boundary - 1)) != 0)
835 						pa_length = 0;
836 				}
837 				if (pa_length >= size) {
838 					vm_reserv_reclaim(rv);
839 					return (TRUE);
840 				}
841 			} else
842 				pa_length = 0;
843 	}
844 	return (FALSE);
845 }
846 
847 /*
848  * Transfers the reservation underlying the given page to a new object.
849  *
850  * The object must be locked.
851  */
852 void
853 vm_reserv_rename(vm_page_t m, vm_object_t new_object, vm_object_t old_object,
854     vm_pindex_t old_object_offset)
855 {
856 	vm_reserv_t rv;
857 
858 	VM_OBJECT_ASSERT_WLOCKED(new_object);
859 	rv = vm_reserv_from_page(m);
860 	if (rv->object == old_object) {
861 		mtx_lock(&vm_page_queue_free_mtx);
862 		if (rv->object == old_object) {
863 			LIST_REMOVE(rv, objq);
864 			LIST_INSERT_HEAD(&new_object->rvq, rv, objq);
865 			rv->object = new_object;
866 			rv->pindex -= old_object_offset;
867 		}
868 		mtx_unlock(&vm_page_queue_free_mtx);
869 	}
870 }
871 
872 /*
873  * Allocates the virtual and physical memory required by the reservation
874  * management system's data structures, in particular, the reservation array.
875  */
876 vm_paddr_t
877 vm_reserv_startup(vm_offset_t *vaddr, vm_paddr_t end, vm_paddr_t high_water)
878 {
879 	vm_paddr_t new_end;
880 	size_t size;
881 
882 	/*
883 	 * Calculate the size (in bytes) of the reservation array.  Round up
884 	 * from "high_water" because every small page is mapped to an element
885 	 * in the reservation array based on its physical address.  Thus, the
886 	 * number of elements in the reservation array can be greater than the
887 	 * number of superpages.
888 	 */
889 	size = howmany(high_water, VM_LEVEL_0_SIZE) * sizeof(struct vm_reserv);
890 
891 	/*
892 	 * Allocate and map the physical memory for the reservation array.  The
893 	 * next available virtual address is returned by reference.
894 	 */
895 	new_end = end - round_page(size);
896 	vm_reserv_array = (void *)(uintptr_t)pmap_map(vaddr, new_end, end,
897 	    VM_PROT_READ | VM_PROT_WRITE);
898 	bzero(vm_reserv_array, size);
899 
900 	/*
901 	 * Return the next available physical address.
902 	 */
903 	return (new_end);
904 }
905 
906 #endif	/* VM_NRESERVLEVEL > 0 */
907