xref: /freebsd/sys/vm/vm_reserv.c (revision 04803ab2177f1c57467b12f497a7ab8eabc3e8d2)
1 /*-
2  * Copyright (c) 2002-2006 Rice University
3  * Copyright (c) 2007-2011 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 size of a population map entry
97  */
98 typedef	u_long		popmap_t;
99 
100 /*
101  * The number of bits in a population map entry
102  */
103 #define	NBPOPMAP	(NBBY * sizeof(popmap_t))
104 
105 /*
106  * The number of population map entries in a reservation
107  */
108 #define	NPOPMAP		howmany(VM_LEVEL_0_NPAGES, NBPOPMAP)
109 
110 /*
111  * Clear a bit in the population map.
112  */
113 static __inline void
114 popmap_clear(popmap_t popmap[], int i)
115 {
116 
117 	popmap[i / NBPOPMAP] &= ~(1UL << (i % NBPOPMAP));
118 }
119 
120 /*
121  * Set a bit in the population map.
122  */
123 static __inline void
124 popmap_set(popmap_t popmap[], int i)
125 {
126 
127 	popmap[i / NBPOPMAP] |= 1UL << (i % NBPOPMAP);
128 }
129 
130 /*
131  * Is a bit in the population map clear?
132  */
133 static __inline boolean_t
134 popmap_is_clear(popmap_t popmap[], int i)
135 {
136 
137 	return ((popmap[i / NBPOPMAP] & (1UL << (i % NBPOPMAP))) == 0);
138 }
139 
140 /*
141  * Is a bit in the population map set?
142  */
143 static __inline boolean_t
144 popmap_is_set(popmap_t popmap[], int i)
145 {
146 
147 	return ((popmap[i / NBPOPMAP] & (1UL << (i % NBPOPMAP))) != 0);
148 }
149 
150 /*
151  * The reservation structure
152  *
153  * A reservation structure is constructed whenever a large physical page is
154  * speculatively allocated to an object.  The reservation provides the small
155  * physical pages for the range [pindex, pindex + VM_LEVEL_0_NPAGES) of offsets
156  * within that object.  The reservation's "popcnt" tracks the number of these
157  * small physical pages that are in use at any given time.  When and if the
158  * reservation is not fully utilized, it appears in the queue of partially-
159  * populated reservations.  The reservation always appears on the containing
160  * object's list of reservations.
161  *
162  * A partially-populated reservation can be broken and reclaimed at any time.
163  */
164 struct vm_reserv {
165 	TAILQ_ENTRY(vm_reserv) partpopq;
166 	LIST_ENTRY(vm_reserv) objq;
167 	vm_object_t	object;			/* containing object */
168 	vm_pindex_t	pindex;			/* offset within object */
169 	vm_page_t	pages;			/* first page of a superpage */
170 	int		popcnt;			/* # of pages in use */
171 	char		inpartpopq;
172 	popmap_t	popmap[NPOPMAP];	/* bit vector of used pages */
173 };
174 
175 /*
176  * The reservation array
177  *
178  * This array is analoguous in function to vm_page_array.  It differs in the
179  * respect that it may contain a greater number of useful reservation
180  * structures than there are (physical) superpages.  These "invalid"
181  * reservation structures exist to trade-off space for time in the
182  * implementation of vm_reserv_from_page().  Invalid reservation structures are
183  * distinguishable from "valid" reservation structures by inspecting the
184  * reservation's "pages" field.  Invalid reservation structures have a NULL
185  * "pages" field.
186  *
187  * vm_reserv_from_page() maps a small (physical) page to an element of this
188  * array by computing a physical reservation number from the page's physical
189  * address.  The physical reservation number is used as the array index.
190  *
191  * An "active" reservation is a valid reservation structure that has a non-NULL
192  * "object" field and a non-zero "popcnt" field.  In other words, every active
193  * reservation belongs to a particular object.  Moreover, every active
194  * reservation has an entry in the containing object's list of reservations.
195  */
196 static vm_reserv_t vm_reserv_array;
197 
198 /*
199  * The partially-populated reservation queue
200  *
201  * This queue enables the fast recovery of an unused cached or free small page
202  * from a partially-populated reservation.  The reservation at the head of
203  * this queue is the least-recently-changed, partially-populated reservation.
204  *
205  * Access to this queue is synchronized by the free page queue lock.
206  */
207 static TAILQ_HEAD(, vm_reserv) vm_rvq_partpop =
208 			    TAILQ_HEAD_INITIALIZER(vm_rvq_partpop);
209 
210 static SYSCTL_NODE(_vm, OID_AUTO, reserv, CTLFLAG_RD, 0, "Reservation Info");
211 
212 static long vm_reserv_broken;
213 SYSCTL_LONG(_vm_reserv, OID_AUTO, broken, CTLFLAG_RD,
214     &vm_reserv_broken, 0, "Cumulative number of broken reservations");
215 
216 static long vm_reserv_freed;
217 SYSCTL_LONG(_vm_reserv, OID_AUTO, freed, CTLFLAG_RD,
218     &vm_reserv_freed, 0, "Cumulative number of freed reservations");
219 
220 static int sysctl_vm_reserv_partpopq(SYSCTL_HANDLER_ARGS);
221 
222 SYSCTL_OID(_vm_reserv, OID_AUTO, partpopq, CTLTYPE_STRING | CTLFLAG_RD, NULL, 0,
223     sysctl_vm_reserv_partpopq, "A", "Partially-populated reservation queues");
224 
225 static long vm_reserv_reclaimed;
226 SYSCTL_LONG(_vm_reserv, OID_AUTO, reclaimed, CTLFLAG_RD,
227     &vm_reserv_reclaimed, 0, "Cumulative number of reclaimed reservations");
228 
229 static void		vm_reserv_break(vm_reserv_t rv, vm_page_t m);
230 static void		vm_reserv_depopulate(vm_reserv_t rv, int index);
231 static vm_reserv_t	vm_reserv_from_page(vm_page_t m);
232 static boolean_t	vm_reserv_has_pindex(vm_reserv_t rv,
233 			    vm_pindex_t pindex);
234 static void		vm_reserv_populate(vm_reserv_t rv, int index);
235 static void		vm_reserv_reclaim(vm_reserv_t rv);
236 
237 /*
238  * Describes the current state of the partially-populated reservation queue.
239  */
240 static int
241 sysctl_vm_reserv_partpopq(SYSCTL_HANDLER_ARGS)
242 {
243 	struct sbuf sbuf;
244 	vm_reserv_t rv;
245 	int counter, error, level, unused_pages;
246 
247 	error = sysctl_wire_old_buffer(req, 0);
248 	if (error != 0)
249 		return (error);
250 	sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
251 	sbuf_printf(&sbuf, "\nLEVEL     SIZE  NUMBER\n\n");
252 	for (level = -1; level <= VM_NRESERVLEVEL - 2; level++) {
253 		counter = 0;
254 		unused_pages = 0;
255 		mtx_lock(&vm_page_queue_free_mtx);
256 		TAILQ_FOREACH(rv, &vm_rvq_partpop/*[level]*/, partpopq) {
257 			counter++;
258 			unused_pages += VM_LEVEL_0_NPAGES - rv->popcnt;
259 		}
260 		mtx_unlock(&vm_page_queue_free_mtx);
261 		sbuf_printf(&sbuf, "%5d: %6dK, %6d\n", level,
262 		    unused_pages * ((int)PAGE_SIZE / 1024), counter);
263 	}
264 	error = sbuf_finish(&sbuf);
265 	sbuf_delete(&sbuf);
266 	return (error);
267 }
268 
269 /*
270  * Reduces the given reservation's population count.  If the population count
271  * becomes zero, the reservation is destroyed.  Additionally, moves the
272  * reservation to the tail of the partially-populated reservation queue if the
273  * population count is non-zero.
274  *
275  * The free page queue lock must be held.
276  */
277 static void
278 vm_reserv_depopulate(vm_reserv_t rv, int index)
279 {
280 
281 	mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
282 	KASSERT(rv->object != NULL,
283 	    ("vm_reserv_depopulate: reserv %p is free", rv));
284 	KASSERT(popmap_is_set(rv->popmap, index),
285 	    ("vm_reserv_depopulate: reserv %p's popmap[%d] is clear", rv,
286 	    index));
287 	KASSERT(rv->popcnt > 0,
288 	    ("vm_reserv_depopulate: reserv %p's popcnt is corrupted", rv));
289 	if (rv->inpartpopq) {
290 		TAILQ_REMOVE(&vm_rvq_partpop, rv, partpopq);
291 		rv->inpartpopq = FALSE;
292 	} else {
293 		KASSERT(rv->pages->psind == 1,
294 		    ("vm_reserv_depopulate: reserv %p is already demoted",
295 		    rv));
296 		rv->pages->psind = 0;
297 	}
298 	popmap_clear(rv->popmap, index);
299 	rv->popcnt--;
300 	if (rv->popcnt == 0) {
301 		LIST_REMOVE(rv, objq);
302 		rv->object = NULL;
303 		vm_phys_free_pages(rv->pages, VM_LEVEL_0_ORDER);
304 		vm_reserv_freed++;
305 	} else {
306 		rv->inpartpopq = TRUE;
307 		TAILQ_INSERT_TAIL(&vm_rvq_partpop, rv, partpopq);
308 	}
309 }
310 
311 /*
312  * Returns the reservation to which the given page might belong.
313  */
314 static __inline vm_reserv_t
315 vm_reserv_from_page(vm_page_t m)
316 {
317 
318 	return (&vm_reserv_array[VM_PAGE_TO_PHYS(m) >> VM_LEVEL_0_SHIFT]);
319 }
320 
321 /*
322  * Returns TRUE if the given reservation contains the given page index and
323  * FALSE otherwise.
324  */
325 static __inline boolean_t
326 vm_reserv_has_pindex(vm_reserv_t rv, vm_pindex_t pindex)
327 {
328 
329 	return (((pindex - rv->pindex) & ~(VM_LEVEL_0_NPAGES - 1)) == 0);
330 }
331 
332 /*
333  * Increases the given reservation's population count.  Moves the reservation
334  * to the tail of the partially-populated reservation queue.
335  *
336  * The free page queue must be locked.
337  */
338 static void
339 vm_reserv_populate(vm_reserv_t rv, int index)
340 {
341 
342 	mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
343 	KASSERT(rv->object != NULL,
344 	    ("vm_reserv_populate: reserv %p is free", rv));
345 	KASSERT(popmap_is_clear(rv->popmap, index),
346 	    ("vm_reserv_populate: reserv %p's popmap[%d] is set", rv,
347 	    index));
348 	KASSERT(rv->popcnt < VM_LEVEL_0_NPAGES,
349 	    ("vm_reserv_populate: reserv %p is already full", rv));
350 	KASSERT(rv->pages->psind == 0,
351 	    ("vm_reserv_populate: reserv %p is already promoted", rv));
352 	if (rv->inpartpopq) {
353 		TAILQ_REMOVE(&vm_rvq_partpop, rv, partpopq);
354 		rv->inpartpopq = FALSE;
355 	}
356 	popmap_set(rv->popmap, index);
357 	rv->popcnt++;
358 	if (rv->popcnt < VM_LEVEL_0_NPAGES) {
359 		rv->inpartpopq = TRUE;
360 		TAILQ_INSERT_TAIL(&vm_rvq_partpop, rv, partpopq);
361 	} else
362 		rv->pages->psind = 1;
363 }
364 
365 /*
366  * Allocates a contiguous set of physical pages of the given size "npages"
367  * from existing or newly created reservations.  All of the physical pages
368  * must be at or above the given physical address "low" and below the given
369  * physical address "high".  The given value "alignment" determines the
370  * alignment of the first physical page in the set.  If the given value
371  * "boundary" is non-zero, then the set of physical pages cannot cross any
372  * physical address boundary that is a multiple of that value.  Both
373  * "alignment" and "boundary" must be a power of two.
374  *
375  * The object and free page queue must be locked.
376  */
377 vm_page_t
378 vm_reserv_alloc_contig(vm_object_t object, vm_pindex_t pindex, u_long npages,
379     vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary)
380 {
381 	vm_paddr_t pa, size;
382 	vm_page_t m, m_ret, mpred, msucc;
383 	vm_pindex_t first, leftcap, rightcap;
384 	vm_reserv_t rv;
385 	u_long allocpages, maxpages, minpages;
386 	int i, index, n;
387 
388 	mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
389 	VM_OBJECT_ASSERT_WLOCKED(object);
390 	KASSERT(npages != 0, ("vm_reserv_alloc_contig: npages is 0"));
391 
392 	/*
393 	 * Is a reservation fundamentally impossible?
394 	 */
395 	if (pindex < VM_RESERV_INDEX(object, pindex) ||
396 	    pindex + npages > object->size)
397 		return (NULL);
398 
399 	/*
400 	 * All reservations of a particular size have the same alignment.
401 	 * Assuming that the first page is allocated from a reservation, the
402 	 * least significant bits of its physical address can be determined
403 	 * from its offset from the beginning of the reservation and the size
404 	 * of the reservation.
405 	 *
406 	 * Could the specified index within a reservation of the smallest
407 	 * possible size satisfy the alignment and boundary requirements?
408 	 */
409 	pa = VM_RESERV_INDEX(object, pindex) << PAGE_SHIFT;
410 	if ((pa & (alignment - 1)) != 0)
411 		return (NULL);
412 	size = npages << PAGE_SHIFT;
413 	if (((pa ^ (pa + size - 1)) & ~(boundary - 1)) != 0)
414 		return (NULL);
415 
416 	/*
417 	 * Look for an existing reservation.
418 	 */
419 	mpred = vm_radix_lookup_le(&object->rtree, pindex);
420 	if (mpred != NULL) {
421 		KASSERT(mpred->pindex < pindex,
422 		    ("vm_reserv_alloc_contig: pindex already allocated"));
423 		rv = vm_reserv_from_page(mpred);
424 		if (rv->object == object && vm_reserv_has_pindex(rv, pindex))
425 			goto found;
426 		msucc = TAILQ_NEXT(mpred, listq);
427 	} else
428 		msucc = TAILQ_FIRST(&object->memq);
429 	if (msucc != NULL) {
430 		KASSERT(msucc->pindex > pindex,
431 		    ("vm_reserv_alloc_page: pindex already allocated"));
432 		rv = vm_reserv_from_page(msucc);
433 		if (rv->object == object && vm_reserv_has_pindex(rv, pindex))
434 			goto found;
435 	}
436 
437 	/*
438 	 * Could at least one reservation fit between the first index to the
439 	 * left that can be used ("leftcap") and the first index to the right
440 	 * that cannot be used ("rightcap")?
441 	 */
442 	first = pindex - VM_RESERV_INDEX(object, pindex);
443 	if (mpred != NULL) {
444 		if ((rv = vm_reserv_from_page(mpred))->object != object)
445 			leftcap = mpred->pindex + 1;
446 		else
447 			leftcap = rv->pindex + VM_LEVEL_0_NPAGES;
448 		if (leftcap > first)
449 			return (NULL);
450 	}
451 	minpages = VM_RESERV_INDEX(object, pindex) + npages;
452 	maxpages = roundup2(minpages, VM_LEVEL_0_NPAGES);
453 	allocpages = maxpages;
454 	if (msucc != NULL) {
455 		if ((rv = vm_reserv_from_page(msucc))->object != object)
456 			rightcap = msucc->pindex;
457 		else
458 			rightcap = rv->pindex;
459 		if (first + maxpages > rightcap) {
460 			if (maxpages == VM_LEVEL_0_NPAGES)
461 				return (NULL);
462 
463 			/*
464 			 * At least one reservation will fit between "leftcap"
465 			 * and "rightcap".  However, a reservation for the
466 			 * last of the requested pages will not fit.  Reduce
467 			 * the size of the upcoming allocation accordingly.
468 			 */
469 			allocpages = minpages;
470 		}
471 	}
472 
473 	/*
474 	 * Would the last new reservation extend past the end of the object?
475 	 */
476 	if (first + maxpages > object->size) {
477 		/*
478 		 * Don't allocate the last new reservation if the object is a
479 		 * vnode or backed by another object that is a vnode.
480 		 */
481 		if (object->type == OBJT_VNODE ||
482 		    (object->backing_object != NULL &&
483 		    object->backing_object->type == OBJT_VNODE)) {
484 			if (maxpages == VM_LEVEL_0_NPAGES)
485 				return (NULL);
486 			allocpages = minpages;
487 		}
488 		/* Speculate that the object may grow. */
489 	}
490 
491 	/*
492 	 * Allocate the physical pages.  The alignment and boundary specified
493 	 * for this allocation may be different from the alignment and
494 	 * boundary specified for the requested pages.  For instance, the
495 	 * specified index may not be the first page within the first new
496 	 * reservation.
497 	 */
498 	m = vm_phys_alloc_contig(allocpages, low, high, ulmax(alignment,
499 	    VM_LEVEL_0_SIZE), boundary > VM_LEVEL_0_SIZE ? boundary : 0);
500 	if (m == NULL)
501 		return (NULL);
502 
503 	/*
504 	 * The allocated physical pages always begin at a reservation
505 	 * boundary, but they do not always end at a reservation boundary.
506 	 * Initialize every reservation that is completely covered by the
507 	 * allocated physical pages.
508 	 */
509 	m_ret = NULL;
510 	index = VM_RESERV_INDEX(object, pindex);
511 	do {
512 		rv = vm_reserv_from_page(m);
513 		KASSERT(rv->pages == m,
514 		    ("vm_reserv_alloc_contig: reserv %p's pages is corrupted",
515 		    rv));
516 		KASSERT(rv->object == NULL,
517 		    ("vm_reserv_alloc_contig: reserv %p isn't free", rv));
518 		LIST_INSERT_HEAD(&object->rvq, rv, objq);
519 		rv->object = object;
520 		rv->pindex = first;
521 		KASSERT(rv->popcnt == 0,
522 		    ("vm_reserv_alloc_contig: reserv %p's popcnt is corrupted",
523 		    rv));
524 		KASSERT(!rv->inpartpopq,
525 		    ("vm_reserv_alloc_contig: reserv %p's inpartpopq is TRUE",
526 		    rv));
527 		for (i = 0; i < NPOPMAP; i++)
528 			KASSERT(rv->popmap[i] == 0,
529 		    ("vm_reserv_alloc_contig: reserv %p's popmap is corrupted",
530 			    rv));
531 		n = ulmin(VM_LEVEL_0_NPAGES - index, npages);
532 		for (i = 0; i < n; i++)
533 			vm_reserv_populate(rv, index + i);
534 		npages -= n;
535 		if (m_ret == NULL) {
536 			m_ret = &rv->pages[index];
537 			index = 0;
538 		}
539 		m += VM_LEVEL_0_NPAGES;
540 		first += VM_LEVEL_0_NPAGES;
541 		allocpages -= VM_LEVEL_0_NPAGES;
542 	} while (allocpages >= VM_LEVEL_0_NPAGES);
543 	return (m_ret);
544 
545 	/*
546 	 * Found a matching reservation.
547 	 */
548 found:
549 	index = VM_RESERV_INDEX(object, pindex);
550 	/* Does the allocation fit within the reservation? */
551 	if (index + npages > VM_LEVEL_0_NPAGES)
552 		return (NULL);
553 	m = &rv->pages[index];
554 	pa = VM_PAGE_TO_PHYS(m);
555 	if (pa < low || pa + size > high || (pa & (alignment - 1)) != 0 ||
556 	    ((pa ^ (pa + size - 1)) & ~(boundary - 1)) != 0)
557 		return (NULL);
558 	/* Handle vm_page_rename(m, new_object, ...). */
559 	for (i = 0; i < npages; i++)
560 		if (popmap_is_set(rv->popmap, index + i))
561 			return (NULL);
562 	for (i = 0; i < npages; i++)
563 		vm_reserv_populate(rv, index + i);
564 	return (m);
565 }
566 
567 /*
568  * Allocates a page from an existing or newly-created reservation.
569  *
570  * The page "mpred" must immediately precede the offset "pindex" within the
571  * specified object.
572  *
573  * The object and free page queue must be locked.
574  */
575 vm_page_t
576 vm_reserv_alloc_page(vm_object_t object, vm_pindex_t pindex, vm_page_t mpred)
577 {
578 	vm_page_t m, msucc;
579 	vm_pindex_t first, leftcap, rightcap;
580 	vm_reserv_t rv;
581 	int i, index;
582 
583 	mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
584 	VM_OBJECT_ASSERT_WLOCKED(object);
585 
586 	/*
587 	 * Is a reservation fundamentally impossible?
588 	 */
589 	if (pindex < VM_RESERV_INDEX(object, pindex) ||
590 	    pindex >= object->size)
591 		return (NULL);
592 
593 	/*
594 	 * Look for an existing reservation.
595 	 */
596 	if (mpred != NULL) {
597 		KASSERT(mpred->object == object,
598 		    ("vm_reserv_alloc_page: object doesn't contain mpred"));
599 		KASSERT(mpred->pindex < pindex,
600 		    ("vm_reserv_alloc_page: mpred doesn't precede pindex"));
601 		rv = vm_reserv_from_page(mpred);
602 		if (rv->object == object && vm_reserv_has_pindex(rv, pindex))
603 			goto found;
604 		msucc = TAILQ_NEXT(mpred, listq);
605 	} else
606 		msucc = TAILQ_FIRST(&object->memq);
607 	if (msucc != NULL) {
608 		KASSERT(msucc->pindex > pindex,
609 		    ("vm_reserv_alloc_page: msucc doesn't succeed pindex"));
610 		rv = vm_reserv_from_page(msucc);
611 		if (rv->object == object && vm_reserv_has_pindex(rv, pindex))
612 			goto found;
613 	}
614 
615 	/*
616 	 * Could a reservation fit between the first index to the left that
617 	 * can be used and the first index to the right that cannot be used?
618 	 */
619 	first = pindex - VM_RESERV_INDEX(object, pindex);
620 	if (mpred != NULL) {
621 		if ((rv = vm_reserv_from_page(mpred))->object != object)
622 			leftcap = mpred->pindex + 1;
623 		else
624 			leftcap = rv->pindex + VM_LEVEL_0_NPAGES;
625 		if (leftcap > first)
626 			return (NULL);
627 	}
628 	if (msucc != NULL) {
629 		if ((rv = vm_reserv_from_page(msucc))->object != object)
630 			rightcap = msucc->pindex;
631 		else
632 			rightcap = rv->pindex;
633 		if (first + VM_LEVEL_0_NPAGES > rightcap)
634 			return (NULL);
635 	}
636 
637 	/*
638 	 * Would a new reservation extend past the end of the object?
639 	 */
640 	if (first + VM_LEVEL_0_NPAGES > object->size) {
641 		/*
642 		 * Don't allocate a new reservation if the object is a vnode or
643 		 * backed by another object that is a vnode.
644 		 */
645 		if (object->type == OBJT_VNODE ||
646 		    (object->backing_object != NULL &&
647 		    object->backing_object->type == OBJT_VNODE))
648 			return (NULL);
649 		/* Speculate that the object may grow. */
650 	}
651 
652 	/*
653 	 * Allocate and populate the new reservation.
654 	 */
655 	m = vm_phys_alloc_pages(VM_FREEPOOL_DEFAULT, VM_LEVEL_0_ORDER);
656 	if (m == NULL)
657 		return (NULL);
658 	rv = vm_reserv_from_page(m);
659 	KASSERT(rv->pages == m,
660 	    ("vm_reserv_alloc_page: reserv %p's pages is corrupted", rv));
661 	KASSERT(rv->object == NULL,
662 	    ("vm_reserv_alloc_page: reserv %p isn't free", rv));
663 	LIST_INSERT_HEAD(&object->rvq, rv, objq);
664 	rv->object = object;
665 	rv->pindex = first;
666 	KASSERT(rv->popcnt == 0,
667 	    ("vm_reserv_alloc_page: reserv %p's popcnt is corrupted", rv));
668 	KASSERT(!rv->inpartpopq,
669 	    ("vm_reserv_alloc_page: reserv %p's inpartpopq is TRUE", rv));
670 	for (i = 0; i < NPOPMAP; i++)
671 		KASSERT(rv->popmap[i] == 0,
672 		    ("vm_reserv_alloc_page: reserv %p's popmap is corrupted",
673 		    rv));
674 	index = VM_RESERV_INDEX(object, pindex);
675 	vm_reserv_populate(rv, index);
676 	return (&rv->pages[index]);
677 
678 	/*
679 	 * Found a matching reservation.
680 	 */
681 found:
682 	index = VM_RESERV_INDEX(object, pindex);
683 	m = &rv->pages[index];
684 	/* Handle vm_page_rename(m, new_object, ...). */
685 	if (popmap_is_set(rv->popmap, index))
686 		return (NULL);
687 	vm_reserv_populate(rv, index);
688 	return (m);
689 }
690 
691 /*
692  * Breaks the given reservation.  Except for the specified cached or free
693  * page, all cached and free pages in the reservation are returned to the
694  * physical memory allocator.  The reservation's population count and map are
695  * reset to their initial state.
696  *
697  * The given reservation must not be in the partially-populated reservation
698  * queue.  The free page queue lock must be held.
699  */
700 static void
701 vm_reserv_break(vm_reserv_t rv, vm_page_t m)
702 {
703 	int begin_zeroes, hi, i, lo;
704 
705 	mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
706 	KASSERT(rv->object != NULL,
707 	    ("vm_reserv_break: reserv %p is free", rv));
708 	KASSERT(!rv->inpartpopq,
709 	    ("vm_reserv_break: reserv %p's inpartpopq is TRUE", rv));
710 	LIST_REMOVE(rv, objq);
711 	rv->object = NULL;
712 	if (m != NULL) {
713 		/*
714 		 * Since the reservation is being broken, there is no harm in
715 		 * abusing the population map to stop "m" from being returned
716 		 * to the physical memory allocator.
717 		 */
718 		i = m - rv->pages;
719 		KASSERT(popmap_is_clear(rv->popmap, i),
720 		    ("vm_reserv_break: reserv %p's popmap is corrupted", rv));
721 		popmap_set(rv->popmap, i);
722 		rv->popcnt++;
723 	}
724 	i = hi = 0;
725 	do {
726 		/* Find the next 0 bit.  Any previous 0 bits are < "hi". */
727 		lo = ffsl(~(((1UL << hi) - 1) | rv->popmap[i]));
728 		if (lo == 0) {
729 			/* Redundantly clears bits < "hi". */
730 			rv->popmap[i] = 0;
731 			rv->popcnt -= NBPOPMAP - hi;
732 			while (++i < NPOPMAP) {
733 				lo = ffsl(~rv->popmap[i]);
734 				if (lo == 0) {
735 					rv->popmap[i] = 0;
736 					rv->popcnt -= NBPOPMAP;
737 				} else
738 					break;
739 			}
740 			if (i == NPOPMAP)
741 				break;
742 			hi = 0;
743 		}
744 		KASSERT(lo > 0, ("vm_reserv_break: lo is %d", lo));
745 		/* Convert from ffsl() to ordinary bit numbering. */
746 		lo--;
747 		if (lo > 0) {
748 			/* Redundantly clears bits < "hi". */
749 			rv->popmap[i] &= ~((1UL << lo) - 1);
750 			rv->popcnt -= lo - hi;
751 		}
752 		begin_zeroes = NBPOPMAP * i + lo;
753 		/* Find the next 1 bit. */
754 		do
755 			hi = ffsl(rv->popmap[i]);
756 		while (hi == 0 && ++i < NPOPMAP);
757 		if (i != NPOPMAP)
758 			/* Convert from ffsl() to ordinary bit numbering. */
759 			hi--;
760 		vm_phys_free_contig(&rv->pages[begin_zeroes], NBPOPMAP * i +
761 		    hi - begin_zeroes);
762 	} while (i < NPOPMAP);
763 	KASSERT(rv->popcnt == 0,
764 	    ("vm_reserv_break: reserv %p's popcnt is corrupted", rv));
765 	vm_reserv_broken++;
766 }
767 
768 /*
769  * Breaks all reservations belonging to the given object.
770  */
771 void
772 vm_reserv_break_all(vm_object_t object)
773 {
774 	vm_reserv_t rv;
775 
776 	mtx_lock(&vm_page_queue_free_mtx);
777 	while ((rv = LIST_FIRST(&object->rvq)) != NULL) {
778 		KASSERT(rv->object == object,
779 		    ("vm_reserv_break_all: reserv %p is corrupted", rv));
780 		if (rv->inpartpopq) {
781 			TAILQ_REMOVE(&vm_rvq_partpop, rv, partpopq);
782 			rv->inpartpopq = FALSE;
783 		}
784 		vm_reserv_break(rv, NULL);
785 	}
786 	mtx_unlock(&vm_page_queue_free_mtx);
787 }
788 
789 /*
790  * Frees the given page if it belongs to a reservation.  Returns TRUE if the
791  * page is freed and FALSE otherwise.
792  *
793  * The free page queue lock must be held.
794  */
795 boolean_t
796 vm_reserv_free_page(vm_page_t m)
797 {
798 	vm_reserv_t rv;
799 
800 	mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
801 	rv = vm_reserv_from_page(m);
802 	if (rv->object == NULL)
803 		return (FALSE);
804 	if ((m->flags & PG_CACHED) != 0 && m->pool != VM_FREEPOOL_CACHE)
805 		vm_phys_set_pool(VM_FREEPOOL_CACHE, rv->pages,
806 		    VM_LEVEL_0_ORDER);
807 	vm_reserv_depopulate(rv, m - rv->pages);
808 	return (TRUE);
809 }
810 
811 /*
812  * Initializes the reservation management system.  Specifically, initializes
813  * the reservation array.
814  *
815  * Requires that vm_page_array and first_page are initialized!
816  */
817 void
818 vm_reserv_init(void)
819 {
820 	vm_paddr_t paddr;
821 	struct vm_phys_seg *seg;
822 	int segind;
823 
824 	/*
825 	 * Initialize the reservation array.  Specifically, initialize the
826 	 * "pages" field for every element that has an underlying superpage.
827 	 */
828 	for (segind = 0; segind < vm_phys_nsegs; segind++) {
829 		seg = &vm_phys_segs[segind];
830 		paddr = roundup2(seg->start, VM_LEVEL_0_SIZE);
831 		while (paddr + VM_LEVEL_0_SIZE <= seg->end) {
832 			vm_reserv_array[paddr >> VM_LEVEL_0_SHIFT].pages =
833 			    PHYS_TO_VM_PAGE(paddr);
834 			paddr += VM_LEVEL_0_SIZE;
835 		}
836 	}
837 }
838 
839 /*
840  * Returns a reservation level if the given page belongs to a fully-populated
841  * reservation and -1 otherwise.
842  */
843 int
844 vm_reserv_level_iffullpop(vm_page_t m)
845 {
846 	vm_reserv_t rv;
847 
848 	rv = vm_reserv_from_page(m);
849 	return (rv->popcnt == VM_LEVEL_0_NPAGES ? 0 : -1);
850 }
851 
852 /*
853  * Prepare for the reactivation of a cached page.
854  *
855  * First, suppose that the given page "m" was allocated individually, i.e., not
856  * as part of a reservation, and cached.  Then, suppose a reservation
857  * containing "m" is allocated by the same object.  Although "m" and the
858  * reservation belong to the same object, "m"'s pindex may not match the
859  * reservation's.
860  *
861  * The free page queue must be locked.
862  */
863 boolean_t
864 vm_reserv_reactivate_page(vm_page_t m)
865 {
866 	vm_reserv_t rv;
867 	int index;
868 
869 	mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
870 	rv = vm_reserv_from_page(m);
871 	if (rv->object == NULL)
872 		return (FALSE);
873 	KASSERT((m->flags & PG_CACHED) != 0,
874 	    ("vm_reserv_reactivate_page: page %p is not cached", m));
875 	if (m->object == rv->object &&
876 	    m->pindex - rv->pindex == (index = VM_RESERV_INDEX(m->object,
877 	    m->pindex)))
878 		vm_reserv_populate(rv, index);
879 	else {
880 		KASSERT(rv->inpartpopq,
881 	    ("vm_reserv_reactivate_page: reserv %p's inpartpopq is FALSE",
882 		    rv));
883 		TAILQ_REMOVE(&vm_rvq_partpop, rv, partpopq);
884 		rv->inpartpopq = FALSE;
885 		/* Don't release "m" to the physical memory allocator. */
886 		vm_reserv_break(rv, m);
887 	}
888 	return (TRUE);
889 }
890 
891 /*
892  * Breaks the given partially-populated reservation, releasing its cached and
893  * free pages to the physical memory allocator.
894  *
895  * The free page queue lock must be held.
896  */
897 static void
898 vm_reserv_reclaim(vm_reserv_t rv)
899 {
900 
901 	mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
902 	KASSERT(rv->inpartpopq,
903 	    ("vm_reserv_reclaim: reserv %p's inpartpopq is FALSE", rv));
904 	TAILQ_REMOVE(&vm_rvq_partpop, rv, partpopq);
905 	rv->inpartpopq = FALSE;
906 	vm_reserv_break(rv, NULL);
907 	vm_reserv_reclaimed++;
908 }
909 
910 /*
911  * Breaks the reservation at the head of the partially-populated reservation
912  * queue, releasing its cached and free pages to the physical memory
913  * allocator.  Returns TRUE if a reservation is broken and FALSE otherwise.
914  *
915  * The free page queue lock must be held.
916  */
917 boolean_t
918 vm_reserv_reclaim_inactive(void)
919 {
920 	vm_reserv_t rv;
921 
922 	mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
923 	if ((rv = TAILQ_FIRST(&vm_rvq_partpop)) != NULL) {
924 		vm_reserv_reclaim(rv);
925 		return (TRUE);
926 	}
927 	return (FALSE);
928 }
929 
930 /*
931  * Searches the partially-populated reservation queue for the least recently
932  * active reservation with unused pages, i.e., cached or free, that satisfy the
933  * given request for contiguous physical memory.  If a satisfactory reservation
934  * is found, it is broken.  Returns TRUE if a reservation is broken and FALSE
935  * otherwise.
936  *
937  * The free page queue lock must be held.
938  */
939 boolean_t
940 vm_reserv_reclaim_contig(u_long npages, vm_paddr_t low, vm_paddr_t high,
941     u_long alignment, vm_paddr_t boundary)
942 {
943 	vm_paddr_t pa, size;
944 	vm_reserv_t rv;
945 	int hi, i, lo, next_free;
946 
947 	mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
948 	if (npages > VM_LEVEL_0_NPAGES - 1)
949 		return (FALSE);
950 	size = npages << PAGE_SHIFT;
951 	TAILQ_FOREACH(rv, &vm_rvq_partpop, partpopq) {
952 		pa = VM_PAGE_TO_PHYS(&rv->pages[VM_LEVEL_0_NPAGES - 1]);
953 		if (pa + PAGE_SIZE - size < low) {
954 			/* This entire reservation is too low; go to next. */
955 			continue;
956 		}
957 		pa = VM_PAGE_TO_PHYS(&rv->pages[0]);
958 		if (pa + size > high) {
959 			/* This entire reservation is too high; go to next. */
960 			continue;
961 		}
962 		if (pa < low) {
963 			/* Start the search for free pages at "low". */
964 			i = (low - pa) / NBPOPMAP;
965 			hi = (low - pa) % NBPOPMAP;
966 		} else
967 			i = hi = 0;
968 		do {
969 			/* Find the next free page. */
970 			lo = ffsl(~(((1UL << hi) - 1) | rv->popmap[i]));
971 			while (lo == 0 && ++i < NPOPMAP)
972 				lo = ffsl(~rv->popmap[i]);
973 			if (i == NPOPMAP)
974 				break;
975 			/* Convert from ffsl() to ordinary bit numbering. */
976 			lo--;
977 			next_free = NBPOPMAP * i + lo;
978 			pa = VM_PAGE_TO_PHYS(&rv->pages[next_free]);
979 			KASSERT(pa >= low,
980 			    ("vm_reserv_reclaim_contig: pa is too low"));
981 			if (pa + size > high) {
982 				/* The rest of this reservation is too high. */
983 				break;
984 			} else if ((pa & (alignment - 1)) != 0 ||
985 			    ((pa ^ (pa + size - 1)) & ~(boundary - 1)) != 0) {
986 				/* Continue with this reservation. */
987 				hi = lo;
988 				continue;
989 			}
990 			/* Find the next used page. */
991 			hi = ffsl(rv->popmap[i] & ~((1UL << lo) - 1));
992 			while (hi == 0 && ++i < NPOPMAP) {
993 				if ((NBPOPMAP * i - next_free) * PAGE_SIZE >=
994 				    size) {
995 					vm_reserv_reclaim(rv);
996 					return (TRUE);
997 				}
998 				hi = ffsl(rv->popmap[i]);
999 			}
1000 			/* Convert from ffsl() to ordinary bit numbering. */
1001 			if (i != NPOPMAP)
1002 				hi--;
1003 			if ((NBPOPMAP * i + hi - next_free) * PAGE_SIZE >=
1004 			    size) {
1005 				vm_reserv_reclaim(rv);
1006 				return (TRUE);
1007 			}
1008 		} while (i < NPOPMAP);
1009 	}
1010 	return (FALSE);
1011 }
1012 
1013 /*
1014  * Transfers the reservation underlying the given page to a new object.
1015  *
1016  * The object must be locked.
1017  */
1018 void
1019 vm_reserv_rename(vm_page_t m, vm_object_t new_object, vm_object_t old_object,
1020     vm_pindex_t old_object_offset)
1021 {
1022 	vm_reserv_t rv;
1023 
1024 	VM_OBJECT_ASSERT_WLOCKED(new_object);
1025 	rv = vm_reserv_from_page(m);
1026 	if (rv->object == old_object) {
1027 		mtx_lock(&vm_page_queue_free_mtx);
1028 		if (rv->object == old_object) {
1029 			LIST_REMOVE(rv, objq);
1030 			LIST_INSERT_HEAD(&new_object->rvq, rv, objq);
1031 			rv->object = new_object;
1032 			rv->pindex -= old_object_offset;
1033 		}
1034 		mtx_unlock(&vm_page_queue_free_mtx);
1035 	}
1036 }
1037 
1038 /*
1039  * Allocates the virtual and physical memory required by the reservation
1040  * management system's data structures, in particular, the reservation array.
1041  */
1042 vm_paddr_t
1043 vm_reserv_startup(vm_offset_t *vaddr, vm_paddr_t end, vm_paddr_t high_water)
1044 {
1045 	vm_paddr_t new_end;
1046 	size_t size;
1047 
1048 	/*
1049 	 * Calculate the size (in bytes) of the reservation array.  Round up
1050 	 * from "high_water" because every small page is mapped to an element
1051 	 * in the reservation array based on its physical address.  Thus, the
1052 	 * number of elements in the reservation array can be greater than the
1053 	 * number of superpages.
1054 	 */
1055 	size = howmany(high_water, VM_LEVEL_0_SIZE) * sizeof(struct vm_reserv);
1056 
1057 	/*
1058 	 * Allocate and map the physical memory for the reservation array.  The
1059 	 * next available virtual address is returned by reference.
1060 	 */
1061 	new_end = end - round_page(size);
1062 	vm_reserv_array = (void *)(uintptr_t)pmap_map(vaddr, new_end, end,
1063 	    VM_PROT_READ | VM_PROT_WRITE);
1064 	bzero(vm_reserv_array, size);
1065 
1066 	/*
1067 	 * Return the next available physical address.
1068 	 */
1069 	return (new_end);
1070 }
1071 
1072 #endif	/* VM_NRESERVLEVEL > 0 */
1073