xref: /freebsd/sys/vm/vm_reserv.c (revision 895f86f15fbf6540071feb9328c3c50ed1f027b8)
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_fullpop(SYSCTL_HANDLER_ARGS);
221 
222 SYSCTL_PROC(_vm_reserv, OID_AUTO, fullpop, CTLTYPE_INT | CTLFLAG_RD, NULL, 0,
223     sysctl_vm_reserv_fullpop, "I", "Current number of full reservations");
224 
225 static int sysctl_vm_reserv_partpopq(SYSCTL_HANDLER_ARGS);
226 
227 SYSCTL_OID(_vm_reserv, OID_AUTO, partpopq, CTLTYPE_STRING | CTLFLAG_RD, NULL, 0,
228     sysctl_vm_reserv_partpopq, "A", "Partially-populated reservation queues");
229 
230 static long vm_reserv_reclaimed;
231 SYSCTL_LONG(_vm_reserv, OID_AUTO, reclaimed, CTLFLAG_RD,
232     &vm_reserv_reclaimed, 0, "Cumulative number of reclaimed reservations");
233 
234 static void		vm_reserv_break(vm_reserv_t rv, vm_page_t m);
235 static void		vm_reserv_depopulate(vm_reserv_t rv, int index);
236 static vm_reserv_t	vm_reserv_from_page(vm_page_t m);
237 static boolean_t	vm_reserv_has_pindex(vm_reserv_t rv,
238 			    vm_pindex_t pindex);
239 static void		vm_reserv_populate(vm_reserv_t rv, int index);
240 static void		vm_reserv_reclaim(vm_reserv_t rv);
241 
242 /*
243  * Returns the current number of full reservations.
244  *
245  * Since the number of full reservations is computed without acquiring the
246  * free page queue lock, the returned value may be inexact.
247  */
248 static int
249 sysctl_vm_reserv_fullpop(SYSCTL_HANDLER_ARGS)
250 {
251 	vm_paddr_t paddr;
252 	struct vm_phys_seg *seg;
253 	vm_reserv_t rv;
254 	int fullpop, segind;
255 
256 	fullpop = 0;
257 	for (segind = 0; segind < vm_phys_nsegs; segind++) {
258 		seg = &vm_phys_segs[segind];
259 		paddr = roundup2(seg->start, VM_LEVEL_0_SIZE);
260 		while (paddr + VM_LEVEL_0_SIZE <= seg->end) {
261 			rv = &vm_reserv_array[paddr >> VM_LEVEL_0_SHIFT];
262 			fullpop += rv->popcnt == VM_LEVEL_0_NPAGES;
263 			paddr += VM_LEVEL_0_SIZE;
264 		}
265 	}
266 	return (sysctl_handle_int(oidp, &fullpop, 0, req));
267 }
268 
269 /*
270  * Describes the current state of the partially-populated reservation queue.
271  */
272 static int
273 sysctl_vm_reserv_partpopq(SYSCTL_HANDLER_ARGS)
274 {
275 	struct sbuf sbuf;
276 	vm_reserv_t rv;
277 	int counter, error, level, unused_pages;
278 
279 	error = sysctl_wire_old_buffer(req, 0);
280 	if (error != 0)
281 		return (error);
282 	sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
283 	sbuf_printf(&sbuf, "\nLEVEL     SIZE  NUMBER\n\n");
284 	for (level = -1; level <= VM_NRESERVLEVEL - 2; level++) {
285 		counter = 0;
286 		unused_pages = 0;
287 		mtx_lock(&vm_page_queue_free_mtx);
288 		TAILQ_FOREACH(rv, &vm_rvq_partpop/*[level]*/, partpopq) {
289 			counter++;
290 			unused_pages += VM_LEVEL_0_NPAGES - rv->popcnt;
291 		}
292 		mtx_unlock(&vm_page_queue_free_mtx);
293 		sbuf_printf(&sbuf, "%5d: %6dK, %6d\n", level,
294 		    unused_pages * ((int)PAGE_SIZE / 1024), counter);
295 	}
296 	error = sbuf_finish(&sbuf);
297 	sbuf_delete(&sbuf);
298 	return (error);
299 }
300 
301 /*
302  * Reduces the given reservation's population count.  If the population count
303  * becomes zero, the reservation is destroyed.  Additionally, moves the
304  * reservation to the tail of the partially-populated reservation queue if the
305  * population count is non-zero.
306  *
307  * The free page queue lock must be held.
308  */
309 static void
310 vm_reserv_depopulate(vm_reserv_t rv, int index)
311 {
312 
313 	mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
314 	KASSERT(rv->object != NULL,
315 	    ("vm_reserv_depopulate: reserv %p is free", rv));
316 	KASSERT(popmap_is_set(rv->popmap, index),
317 	    ("vm_reserv_depopulate: reserv %p's popmap[%d] is clear", rv,
318 	    index));
319 	KASSERT(rv->popcnt > 0,
320 	    ("vm_reserv_depopulate: reserv %p's popcnt is corrupted", rv));
321 	if (rv->inpartpopq) {
322 		TAILQ_REMOVE(&vm_rvq_partpop, rv, partpopq);
323 		rv->inpartpopq = FALSE;
324 	} else {
325 		KASSERT(rv->pages->psind == 1,
326 		    ("vm_reserv_depopulate: reserv %p is already demoted",
327 		    rv));
328 		rv->pages->psind = 0;
329 	}
330 	popmap_clear(rv->popmap, index);
331 	rv->popcnt--;
332 	if (rv->popcnt == 0) {
333 		LIST_REMOVE(rv, objq);
334 		rv->object = NULL;
335 		vm_phys_free_pages(rv->pages, VM_LEVEL_0_ORDER);
336 		vm_reserv_freed++;
337 	} else {
338 		rv->inpartpopq = TRUE;
339 		TAILQ_INSERT_TAIL(&vm_rvq_partpop, rv, partpopq);
340 	}
341 }
342 
343 /*
344  * Returns the reservation to which the given page might belong.
345  */
346 static __inline vm_reserv_t
347 vm_reserv_from_page(vm_page_t m)
348 {
349 
350 	return (&vm_reserv_array[VM_PAGE_TO_PHYS(m) >> VM_LEVEL_0_SHIFT]);
351 }
352 
353 /*
354  * Returns TRUE if the given reservation contains the given page index and
355  * FALSE otherwise.
356  */
357 static __inline boolean_t
358 vm_reserv_has_pindex(vm_reserv_t rv, vm_pindex_t pindex)
359 {
360 
361 	return (((pindex - rv->pindex) & ~(VM_LEVEL_0_NPAGES - 1)) == 0);
362 }
363 
364 /*
365  * Increases the given reservation's population count.  Moves the reservation
366  * to the tail of the partially-populated reservation queue.
367  *
368  * The free page queue must be locked.
369  */
370 static void
371 vm_reserv_populate(vm_reserv_t rv, int index)
372 {
373 
374 	mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
375 	KASSERT(rv->object != NULL,
376 	    ("vm_reserv_populate: reserv %p is free", rv));
377 	KASSERT(popmap_is_clear(rv->popmap, index),
378 	    ("vm_reserv_populate: reserv %p's popmap[%d] is set", rv,
379 	    index));
380 	KASSERT(rv->popcnt < VM_LEVEL_0_NPAGES,
381 	    ("vm_reserv_populate: reserv %p is already full", rv));
382 	KASSERT(rv->pages->psind == 0,
383 	    ("vm_reserv_populate: reserv %p is already promoted", rv));
384 	if (rv->inpartpopq) {
385 		TAILQ_REMOVE(&vm_rvq_partpop, rv, partpopq);
386 		rv->inpartpopq = FALSE;
387 	}
388 	popmap_set(rv->popmap, index);
389 	rv->popcnt++;
390 	if (rv->popcnt < VM_LEVEL_0_NPAGES) {
391 		rv->inpartpopq = TRUE;
392 		TAILQ_INSERT_TAIL(&vm_rvq_partpop, rv, partpopq);
393 	} else
394 		rv->pages->psind = 1;
395 }
396 
397 /*
398  * Allocates a contiguous set of physical pages of the given size "npages"
399  * from existing or newly created reservations.  All of the physical pages
400  * must be at or above the given physical address "low" and below the given
401  * physical address "high".  The given value "alignment" determines the
402  * alignment of the first physical page in the set.  If the given value
403  * "boundary" is non-zero, then the set of physical pages cannot cross any
404  * physical address boundary that is a multiple of that value.  Both
405  * "alignment" and "boundary" must be a power of two.
406  *
407  * The object and free page queue must be locked.
408  */
409 vm_page_t
410 vm_reserv_alloc_contig(vm_object_t object, vm_pindex_t pindex, u_long npages,
411     vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary)
412 {
413 	vm_paddr_t pa, size;
414 	vm_page_t m, m_ret, mpred, msucc;
415 	vm_pindex_t first, leftcap, rightcap;
416 	vm_reserv_t rv;
417 	u_long allocpages, maxpages, minpages;
418 	int i, index, n;
419 
420 	mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
421 	VM_OBJECT_ASSERT_WLOCKED(object);
422 	KASSERT(npages != 0, ("vm_reserv_alloc_contig: npages is 0"));
423 
424 	/*
425 	 * Is a reservation fundamentally impossible?
426 	 */
427 	if (pindex < VM_RESERV_INDEX(object, pindex) ||
428 	    pindex + npages > object->size)
429 		return (NULL);
430 
431 	/*
432 	 * All reservations of a particular size have the same alignment.
433 	 * Assuming that the first page is allocated from a reservation, the
434 	 * least significant bits of its physical address can be determined
435 	 * from its offset from the beginning of the reservation and the size
436 	 * of the reservation.
437 	 *
438 	 * Could the specified index within a reservation of the smallest
439 	 * possible size satisfy the alignment and boundary requirements?
440 	 */
441 	pa = VM_RESERV_INDEX(object, pindex) << PAGE_SHIFT;
442 	if ((pa & (alignment - 1)) != 0)
443 		return (NULL);
444 	size = npages << PAGE_SHIFT;
445 	if (((pa ^ (pa + size - 1)) & ~(boundary - 1)) != 0)
446 		return (NULL);
447 
448 	/*
449 	 * Look for an existing reservation.
450 	 */
451 	mpred = vm_radix_lookup_le(&object->rtree, pindex);
452 	if (mpred != NULL) {
453 		KASSERT(mpred->pindex < pindex,
454 		    ("vm_reserv_alloc_contig: pindex already allocated"));
455 		rv = vm_reserv_from_page(mpred);
456 		if (rv->object == object && vm_reserv_has_pindex(rv, pindex))
457 			goto found;
458 		msucc = TAILQ_NEXT(mpred, listq);
459 	} else
460 		msucc = TAILQ_FIRST(&object->memq);
461 	if (msucc != NULL) {
462 		KASSERT(msucc->pindex > pindex,
463 		    ("vm_reserv_alloc_contig: pindex already allocated"));
464 		rv = vm_reserv_from_page(msucc);
465 		if (rv->object == object && vm_reserv_has_pindex(rv, pindex))
466 			goto found;
467 	}
468 
469 	/*
470 	 * Could at least one reservation fit between the first index to the
471 	 * left that can be used ("leftcap") and the first index to the right
472 	 * that cannot be used ("rightcap")?
473 	 */
474 	first = pindex - VM_RESERV_INDEX(object, pindex);
475 	if (mpred != NULL) {
476 		if ((rv = vm_reserv_from_page(mpred))->object != object)
477 			leftcap = mpred->pindex + 1;
478 		else
479 			leftcap = rv->pindex + VM_LEVEL_0_NPAGES;
480 		if (leftcap > first)
481 			return (NULL);
482 	}
483 	minpages = VM_RESERV_INDEX(object, pindex) + npages;
484 	maxpages = roundup2(minpages, VM_LEVEL_0_NPAGES);
485 	allocpages = maxpages;
486 	if (msucc != NULL) {
487 		if ((rv = vm_reserv_from_page(msucc))->object != object)
488 			rightcap = msucc->pindex;
489 		else
490 			rightcap = rv->pindex;
491 		if (first + maxpages > rightcap) {
492 			if (maxpages == VM_LEVEL_0_NPAGES)
493 				return (NULL);
494 
495 			/*
496 			 * At least one reservation will fit between "leftcap"
497 			 * and "rightcap".  However, a reservation for the
498 			 * last of the requested pages will not fit.  Reduce
499 			 * the size of the upcoming allocation accordingly.
500 			 */
501 			allocpages = minpages;
502 		}
503 	}
504 
505 	/*
506 	 * Would the last new reservation extend past the end of the object?
507 	 */
508 	if (first + maxpages > object->size) {
509 		/*
510 		 * Don't allocate the last new reservation if the object is a
511 		 * vnode or backed by another object that is a vnode.
512 		 */
513 		if (object->type == OBJT_VNODE ||
514 		    (object->backing_object != NULL &&
515 		    object->backing_object->type == OBJT_VNODE)) {
516 			if (maxpages == VM_LEVEL_0_NPAGES)
517 				return (NULL);
518 			allocpages = minpages;
519 		}
520 		/* Speculate that the object may grow. */
521 	}
522 
523 	/*
524 	 * Allocate the physical pages.  The alignment and boundary specified
525 	 * for this allocation may be different from the alignment and
526 	 * boundary specified for the requested pages.  For instance, the
527 	 * specified index may not be the first page within the first new
528 	 * reservation.
529 	 */
530 	m = vm_phys_alloc_contig(allocpages, low, high, ulmax(alignment,
531 	    VM_LEVEL_0_SIZE), boundary > VM_LEVEL_0_SIZE ? boundary : 0);
532 	if (m == NULL)
533 		return (NULL);
534 
535 	/*
536 	 * The allocated physical pages always begin at a reservation
537 	 * boundary, but they do not always end at a reservation boundary.
538 	 * Initialize every reservation that is completely covered by the
539 	 * allocated physical pages.
540 	 */
541 	m_ret = NULL;
542 	index = VM_RESERV_INDEX(object, pindex);
543 	do {
544 		rv = vm_reserv_from_page(m);
545 		KASSERT(rv->pages == m,
546 		    ("vm_reserv_alloc_contig: reserv %p's pages is corrupted",
547 		    rv));
548 		KASSERT(rv->object == NULL,
549 		    ("vm_reserv_alloc_contig: reserv %p isn't free", rv));
550 		LIST_INSERT_HEAD(&object->rvq, rv, objq);
551 		rv->object = object;
552 		rv->pindex = first;
553 		KASSERT(rv->popcnt == 0,
554 		    ("vm_reserv_alloc_contig: reserv %p's popcnt is corrupted",
555 		    rv));
556 		KASSERT(!rv->inpartpopq,
557 		    ("vm_reserv_alloc_contig: reserv %p's inpartpopq is TRUE",
558 		    rv));
559 		for (i = 0; i < NPOPMAP; i++)
560 			KASSERT(rv->popmap[i] == 0,
561 		    ("vm_reserv_alloc_contig: reserv %p's popmap is corrupted",
562 			    rv));
563 		n = ulmin(VM_LEVEL_0_NPAGES - index, npages);
564 		for (i = 0; i < n; i++)
565 			vm_reserv_populate(rv, index + i);
566 		npages -= n;
567 		if (m_ret == NULL) {
568 			m_ret = &rv->pages[index];
569 			index = 0;
570 		}
571 		m += VM_LEVEL_0_NPAGES;
572 		first += VM_LEVEL_0_NPAGES;
573 		allocpages -= VM_LEVEL_0_NPAGES;
574 	} while (allocpages >= VM_LEVEL_0_NPAGES);
575 	return (m_ret);
576 
577 	/*
578 	 * Found a matching reservation.
579 	 */
580 found:
581 	index = VM_RESERV_INDEX(object, pindex);
582 	/* Does the allocation fit within the reservation? */
583 	if (index + npages > VM_LEVEL_0_NPAGES)
584 		return (NULL);
585 	m = &rv->pages[index];
586 	pa = VM_PAGE_TO_PHYS(m);
587 	if (pa < low || pa + size > high || (pa & (alignment - 1)) != 0 ||
588 	    ((pa ^ (pa + size - 1)) & ~(boundary - 1)) != 0)
589 		return (NULL);
590 	/* Handle vm_page_rename(m, new_object, ...). */
591 	for (i = 0; i < npages; i++)
592 		if (popmap_is_set(rv->popmap, index + i))
593 			return (NULL);
594 	for (i = 0; i < npages; i++)
595 		vm_reserv_populate(rv, index + i);
596 	return (m);
597 }
598 
599 /*
600  * Allocates a page from an existing or newly-created reservation.
601  *
602  * The page "mpred" must immediately precede the offset "pindex" within the
603  * specified object.
604  *
605  * The object and free page queue must be locked.
606  */
607 vm_page_t
608 vm_reserv_alloc_page(vm_object_t object, vm_pindex_t pindex, vm_page_t mpred)
609 {
610 	vm_page_t m, msucc;
611 	vm_pindex_t first, leftcap, rightcap;
612 	vm_reserv_t rv;
613 	int i, index;
614 
615 	mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
616 	VM_OBJECT_ASSERT_WLOCKED(object);
617 
618 	/*
619 	 * Is a reservation fundamentally impossible?
620 	 */
621 	if (pindex < VM_RESERV_INDEX(object, pindex) ||
622 	    pindex >= object->size)
623 		return (NULL);
624 
625 	/*
626 	 * Look for an existing reservation.
627 	 */
628 	if (mpred != NULL) {
629 		KASSERT(mpred->object == object,
630 		    ("vm_reserv_alloc_page: object doesn't contain mpred"));
631 		KASSERT(mpred->pindex < pindex,
632 		    ("vm_reserv_alloc_page: mpred doesn't precede pindex"));
633 		rv = vm_reserv_from_page(mpred);
634 		if (rv->object == object && vm_reserv_has_pindex(rv, pindex))
635 			goto found;
636 		msucc = TAILQ_NEXT(mpred, listq);
637 	} else
638 		msucc = TAILQ_FIRST(&object->memq);
639 	if (msucc != NULL) {
640 		KASSERT(msucc->pindex > pindex,
641 		    ("vm_reserv_alloc_page: msucc doesn't succeed pindex"));
642 		rv = vm_reserv_from_page(msucc);
643 		if (rv->object == object && vm_reserv_has_pindex(rv, pindex))
644 			goto found;
645 	}
646 
647 	/*
648 	 * Could a reservation fit between the first index to the left that
649 	 * can be used and the first index to the right that cannot be used?
650 	 */
651 	first = pindex - VM_RESERV_INDEX(object, pindex);
652 	if (mpred != NULL) {
653 		if ((rv = vm_reserv_from_page(mpred))->object != object)
654 			leftcap = mpred->pindex + 1;
655 		else
656 			leftcap = rv->pindex + VM_LEVEL_0_NPAGES;
657 		if (leftcap > first)
658 			return (NULL);
659 	}
660 	if (msucc != NULL) {
661 		if ((rv = vm_reserv_from_page(msucc))->object != object)
662 			rightcap = msucc->pindex;
663 		else
664 			rightcap = rv->pindex;
665 		if (first + VM_LEVEL_0_NPAGES > rightcap)
666 			return (NULL);
667 	}
668 
669 	/*
670 	 * Would a new reservation extend past the end of the object?
671 	 */
672 	if (first + VM_LEVEL_0_NPAGES > object->size) {
673 		/*
674 		 * Don't allocate a new reservation if the object is a vnode or
675 		 * backed by another object that is a vnode.
676 		 */
677 		if (object->type == OBJT_VNODE ||
678 		    (object->backing_object != NULL &&
679 		    object->backing_object->type == OBJT_VNODE))
680 			return (NULL);
681 		/* Speculate that the object may grow. */
682 	}
683 
684 	/*
685 	 * Allocate and populate the new reservation.
686 	 */
687 	m = vm_phys_alloc_pages(VM_FREEPOOL_DEFAULT, VM_LEVEL_0_ORDER);
688 	if (m == NULL)
689 		return (NULL);
690 	rv = vm_reserv_from_page(m);
691 	KASSERT(rv->pages == m,
692 	    ("vm_reserv_alloc_page: reserv %p's pages is corrupted", rv));
693 	KASSERT(rv->object == NULL,
694 	    ("vm_reserv_alloc_page: reserv %p isn't free", rv));
695 	LIST_INSERT_HEAD(&object->rvq, rv, objq);
696 	rv->object = object;
697 	rv->pindex = first;
698 	KASSERT(rv->popcnt == 0,
699 	    ("vm_reserv_alloc_page: reserv %p's popcnt is corrupted", rv));
700 	KASSERT(!rv->inpartpopq,
701 	    ("vm_reserv_alloc_page: reserv %p's inpartpopq is TRUE", rv));
702 	for (i = 0; i < NPOPMAP; i++)
703 		KASSERT(rv->popmap[i] == 0,
704 		    ("vm_reserv_alloc_page: reserv %p's popmap is corrupted",
705 		    rv));
706 	index = VM_RESERV_INDEX(object, pindex);
707 	vm_reserv_populate(rv, index);
708 	return (&rv->pages[index]);
709 
710 	/*
711 	 * Found a matching reservation.
712 	 */
713 found:
714 	index = VM_RESERV_INDEX(object, pindex);
715 	m = &rv->pages[index];
716 	/* Handle vm_page_rename(m, new_object, ...). */
717 	if (popmap_is_set(rv->popmap, index))
718 		return (NULL);
719 	vm_reserv_populate(rv, index);
720 	return (m);
721 }
722 
723 /*
724  * Breaks the given reservation.  Except for the specified cached or free
725  * page, all cached and free pages in the reservation are returned to the
726  * physical memory allocator.  The reservation's population count and map are
727  * reset to their initial state.
728  *
729  * The given reservation must not be in the partially-populated reservation
730  * queue.  The free page queue lock must be held.
731  */
732 static void
733 vm_reserv_break(vm_reserv_t rv, vm_page_t m)
734 {
735 	int begin_zeroes, hi, i, lo;
736 
737 	mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
738 	KASSERT(rv->object != NULL,
739 	    ("vm_reserv_break: reserv %p is free", rv));
740 	KASSERT(!rv->inpartpopq,
741 	    ("vm_reserv_break: reserv %p's inpartpopq is TRUE", rv));
742 	LIST_REMOVE(rv, objq);
743 	rv->object = NULL;
744 	if (m != NULL) {
745 		/*
746 		 * Since the reservation is being broken, there is no harm in
747 		 * abusing the population map to stop "m" from being returned
748 		 * to the physical memory allocator.
749 		 */
750 		i = m - rv->pages;
751 		KASSERT(popmap_is_clear(rv->popmap, i),
752 		    ("vm_reserv_break: reserv %p's popmap is corrupted", rv));
753 		popmap_set(rv->popmap, i);
754 		rv->popcnt++;
755 	}
756 	i = hi = 0;
757 	do {
758 		/* Find the next 0 bit.  Any previous 0 bits are < "hi". */
759 		lo = ffsl(~(((1UL << hi) - 1) | rv->popmap[i]));
760 		if (lo == 0) {
761 			/* Redundantly clears bits < "hi". */
762 			rv->popmap[i] = 0;
763 			rv->popcnt -= NBPOPMAP - hi;
764 			while (++i < NPOPMAP) {
765 				lo = ffsl(~rv->popmap[i]);
766 				if (lo == 0) {
767 					rv->popmap[i] = 0;
768 					rv->popcnt -= NBPOPMAP;
769 				} else
770 					break;
771 			}
772 			if (i == NPOPMAP)
773 				break;
774 			hi = 0;
775 		}
776 		KASSERT(lo > 0, ("vm_reserv_break: lo is %d", lo));
777 		/* Convert from ffsl() to ordinary bit numbering. */
778 		lo--;
779 		if (lo > 0) {
780 			/* Redundantly clears bits < "hi". */
781 			rv->popmap[i] &= ~((1UL << lo) - 1);
782 			rv->popcnt -= lo - hi;
783 		}
784 		begin_zeroes = NBPOPMAP * i + lo;
785 		/* Find the next 1 bit. */
786 		do
787 			hi = ffsl(rv->popmap[i]);
788 		while (hi == 0 && ++i < NPOPMAP);
789 		if (i != NPOPMAP)
790 			/* Convert from ffsl() to ordinary bit numbering. */
791 			hi--;
792 		vm_phys_free_contig(&rv->pages[begin_zeroes], NBPOPMAP * i +
793 		    hi - begin_zeroes);
794 	} while (i < NPOPMAP);
795 	KASSERT(rv->popcnt == 0,
796 	    ("vm_reserv_break: reserv %p's popcnt is corrupted", rv));
797 	vm_reserv_broken++;
798 }
799 
800 /*
801  * Breaks all reservations belonging to the given object.
802  */
803 void
804 vm_reserv_break_all(vm_object_t object)
805 {
806 	vm_reserv_t rv;
807 
808 	mtx_lock(&vm_page_queue_free_mtx);
809 	while ((rv = LIST_FIRST(&object->rvq)) != NULL) {
810 		KASSERT(rv->object == object,
811 		    ("vm_reserv_break_all: reserv %p is corrupted", rv));
812 		if (rv->inpartpopq) {
813 			TAILQ_REMOVE(&vm_rvq_partpop, rv, partpopq);
814 			rv->inpartpopq = FALSE;
815 		}
816 		vm_reserv_break(rv, NULL);
817 	}
818 	mtx_unlock(&vm_page_queue_free_mtx);
819 }
820 
821 /*
822  * Frees the given page if it belongs to a reservation.  Returns TRUE if the
823  * page is freed and FALSE otherwise.
824  *
825  * The free page queue lock must be held.
826  */
827 boolean_t
828 vm_reserv_free_page(vm_page_t m)
829 {
830 	vm_reserv_t rv;
831 
832 	mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
833 	rv = vm_reserv_from_page(m);
834 	if (rv->object == NULL)
835 		return (FALSE);
836 	vm_reserv_depopulate(rv, m - rv->pages);
837 	return (TRUE);
838 }
839 
840 /*
841  * Initializes the reservation management system.  Specifically, initializes
842  * the reservation array.
843  *
844  * Requires that vm_page_array and first_page are initialized!
845  */
846 void
847 vm_reserv_init(void)
848 {
849 	vm_paddr_t paddr;
850 	struct vm_phys_seg *seg;
851 	int segind;
852 
853 	/*
854 	 * Initialize the reservation array.  Specifically, initialize the
855 	 * "pages" field for every element that has an underlying superpage.
856 	 */
857 	for (segind = 0; segind < vm_phys_nsegs; segind++) {
858 		seg = &vm_phys_segs[segind];
859 		paddr = roundup2(seg->start, VM_LEVEL_0_SIZE);
860 		while (paddr + VM_LEVEL_0_SIZE <= seg->end) {
861 			vm_reserv_array[paddr >> VM_LEVEL_0_SHIFT].pages =
862 			    PHYS_TO_VM_PAGE(paddr);
863 			paddr += VM_LEVEL_0_SIZE;
864 		}
865 	}
866 }
867 
868 /*
869  * Returns a reservation level if the given page belongs to a fully-populated
870  * reservation and -1 otherwise.
871  */
872 int
873 vm_reserv_level_iffullpop(vm_page_t m)
874 {
875 	vm_reserv_t rv;
876 
877 	rv = vm_reserv_from_page(m);
878 	return (rv->popcnt == VM_LEVEL_0_NPAGES ? 0 : -1);
879 }
880 
881 /*
882  * Prepare for the reactivation of a cached page.
883  *
884  * First, suppose that the given page "m" was allocated individually, i.e., not
885  * as part of a reservation, and cached.  Then, suppose a reservation
886  * containing "m" is allocated by the same object.  Although "m" and the
887  * reservation belong to the same object, "m"'s pindex may not match the
888  * reservation's.
889  *
890  * The free page queue must be locked.
891  */
892 boolean_t
893 vm_reserv_reactivate_page(vm_page_t m)
894 {
895 	vm_reserv_t rv;
896 	int index;
897 
898 	mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
899 	rv = vm_reserv_from_page(m);
900 	if (rv->object == NULL)
901 		return (FALSE);
902 	KASSERT((m->flags & PG_CACHED) != 0,
903 	    ("vm_reserv_reactivate_page: page %p is not cached", m));
904 	if (m->object == rv->object &&
905 	    m->pindex - rv->pindex == (index = VM_RESERV_INDEX(m->object,
906 	    m->pindex)))
907 		vm_reserv_populate(rv, index);
908 	else {
909 		KASSERT(rv->inpartpopq,
910 	    ("vm_reserv_reactivate_page: reserv %p's inpartpopq is FALSE",
911 		    rv));
912 		TAILQ_REMOVE(&vm_rvq_partpop, rv, partpopq);
913 		rv->inpartpopq = FALSE;
914 		/* Don't release "m" to the physical memory allocator. */
915 		vm_reserv_break(rv, m);
916 	}
917 	return (TRUE);
918 }
919 
920 /*
921  * Breaks the given partially-populated reservation, releasing its cached and
922  * free pages to the physical memory allocator.
923  *
924  * The free page queue lock must be held.
925  */
926 static void
927 vm_reserv_reclaim(vm_reserv_t rv)
928 {
929 
930 	mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
931 	KASSERT(rv->inpartpopq,
932 	    ("vm_reserv_reclaim: reserv %p's inpartpopq is FALSE", rv));
933 	TAILQ_REMOVE(&vm_rvq_partpop, rv, partpopq);
934 	rv->inpartpopq = FALSE;
935 	vm_reserv_break(rv, NULL);
936 	vm_reserv_reclaimed++;
937 }
938 
939 /*
940  * Breaks the reservation at the head of the partially-populated reservation
941  * queue, releasing its cached and free pages to the physical memory
942  * allocator.  Returns TRUE if a reservation is broken and FALSE otherwise.
943  *
944  * The free page queue lock must be held.
945  */
946 boolean_t
947 vm_reserv_reclaim_inactive(void)
948 {
949 	vm_reserv_t rv;
950 
951 	mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
952 	if ((rv = TAILQ_FIRST(&vm_rvq_partpop)) != NULL) {
953 		vm_reserv_reclaim(rv);
954 		return (TRUE);
955 	}
956 	return (FALSE);
957 }
958 
959 /*
960  * Searches the partially-populated reservation queue for the least recently
961  * active reservation with unused pages, i.e., cached or free, that satisfy the
962  * given request for contiguous physical memory.  If a satisfactory reservation
963  * is found, it is broken.  Returns TRUE if a reservation is broken and FALSE
964  * otherwise.
965  *
966  * The free page queue lock must be held.
967  */
968 boolean_t
969 vm_reserv_reclaim_contig(u_long npages, vm_paddr_t low, vm_paddr_t high,
970     u_long alignment, vm_paddr_t boundary)
971 {
972 	vm_paddr_t pa, size;
973 	vm_reserv_t rv;
974 	int hi, i, lo, low_index, next_free;
975 
976 	mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
977 	if (npages > VM_LEVEL_0_NPAGES - 1)
978 		return (FALSE);
979 	size = npages << PAGE_SHIFT;
980 	TAILQ_FOREACH(rv, &vm_rvq_partpop, partpopq) {
981 		pa = VM_PAGE_TO_PHYS(&rv->pages[VM_LEVEL_0_NPAGES - 1]);
982 		if (pa + PAGE_SIZE - size < low) {
983 			/* This entire reservation is too low; go to next. */
984 			continue;
985 		}
986 		pa = VM_PAGE_TO_PHYS(&rv->pages[0]);
987 		if (pa + size > high) {
988 			/* This entire reservation is too high; go to next. */
989 			continue;
990 		}
991 		if (pa < low) {
992 			/* Start the search for free pages at "low". */
993 			low_index = (low + PAGE_MASK - pa) >> PAGE_SHIFT;
994 			i = low_index / NBPOPMAP;
995 			hi = low_index % NBPOPMAP;
996 		} else
997 			i = hi = 0;
998 		do {
999 			/* Find the next free page. */
1000 			lo = ffsl(~(((1UL << hi) - 1) | rv->popmap[i]));
1001 			while (lo == 0 && ++i < NPOPMAP)
1002 				lo = ffsl(~rv->popmap[i]);
1003 			if (i == NPOPMAP)
1004 				break;
1005 			/* Convert from ffsl() to ordinary bit numbering. */
1006 			lo--;
1007 			next_free = NBPOPMAP * i + lo;
1008 			pa = VM_PAGE_TO_PHYS(&rv->pages[next_free]);
1009 			KASSERT(pa >= low,
1010 			    ("vm_reserv_reclaim_contig: pa is too low"));
1011 			if (pa + size > high) {
1012 				/* The rest of this reservation is too high. */
1013 				break;
1014 			} else if ((pa & (alignment - 1)) != 0 ||
1015 			    ((pa ^ (pa + size - 1)) & ~(boundary - 1)) != 0) {
1016 				/*
1017 				 * The current page doesn't meet the alignment
1018 				 * and/or boundary requirements.  Continue
1019 				 * searching this reservation until the rest
1020 				 * of its free pages are either excluded or
1021 				 * exhausted.
1022 				 */
1023 				hi = lo + 1;
1024 				if (hi >= NBPOPMAP) {
1025 					hi = 0;
1026 					i++;
1027 				}
1028 				continue;
1029 			}
1030 			/* Find the next used page. */
1031 			hi = ffsl(rv->popmap[i] & ~((1UL << lo) - 1));
1032 			while (hi == 0 && ++i < NPOPMAP) {
1033 				if ((NBPOPMAP * i - next_free) * PAGE_SIZE >=
1034 				    size) {
1035 					vm_reserv_reclaim(rv);
1036 					return (TRUE);
1037 				}
1038 				hi = ffsl(rv->popmap[i]);
1039 			}
1040 			/* Convert from ffsl() to ordinary bit numbering. */
1041 			if (i != NPOPMAP)
1042 				hi--;
1043 			if ((NBPOPMAP * i + hi - next_free) * PAGE_SIZE >=
1044 			    size) {
1045 				vm_reserv_reclaim(rv);
1046 				return (TRUE);
1047 			}
1048 		} while (i < NPOPMAP);
1049 	}
1050 	return (FALSE);
1051 }
1052 
1053 /*
1054  * Transfers the reservation underlying the given page to a new object.
1055  *
1056  * The object must be locked.
1057  */
1058 void
1059 vm_reserv_rename(vm_page_t m, vm_object_t new_object, vm_object_t old_object,
1060     vm_pindex_t old_object_offset)
1061 {
1062 	vm_reserv_t rv;
1063 
1064 	VM_OBJECT_ASSERT_WLOCKED(new_object);
1065 	rv = vm_reserv_from_page(m);
1066 	if (rv->object == old_object) {
1067 		mtx_lock(&vm_page_queue_free_mtx);
1068 		if (rv->object == old_object) {
1069 			LIST_REMOVE(rv, objq);
1070 			LIST_INSERT_HEAD(&new_object->rvq, rv, objq);
1071 			rv->object = new_object;
1072 			rv->pindex -= old_object_offset;
1073 		}
1074 		mtx_unlock(&vm_page_queue_free_mtx);
1075 	}
1076 }
1077 
1078 /*
1079  * Allocates the virtual and physical memory required by the reservation
1080  * management system's data structures, in particular, the reservation array.
1081  */
1082 vm_paddr_t
1083 vm_reserv_startup(vm_offset_t *vaddr, vm_paddr_t end, vm_paddr_t high_water)
1084 {
1085 	vm_paddr_t new_end;
1086 	size_t size;
1087 
1088 	/*
1089 	 * Calculate the size (in bytes) of the reservation array.  Round up
1090 	 * from "high_water" because every small page is mapped to an element
1091 	 * in the reservation array based on its physical address.  Thus, the
1092 	 * number of elements in the reservation array can be greater than the
1093 	 * number of superpages.
1094 	 */
1095 	size = howmany(high_water, VM_LEVEL_0_SIZE) * sizeof(struct vm_reserv);
1096 
1097 	/*
1098 	 * Allocate and map the physical memory for the reservation array.  The
1099 	 * next available virtual address is returned by reference.
1100 	 */
1101 	new_end = end - round_page(size);
1102 	vm_reserv_array = (void *)(uintptr_t)pmap_map(vaddr, new_end, end,
1103 	    VM_PROT_READ | VM_PROT_WRITE);
1104 	bzero(vm_reserv_array, size);
1105 
1106 	/*
1107 	 * Return the next available physical address.
1108 	 */
1109 	return (new_end);
1110 }
1111 
1112 #endif	/* VM_NRESERVLEVEL > 0 */
1113