xref: /freebsd/sys/vm/vm_reserv.c (revision ba3c1f5972d7b90feb6e6da47905ff2757e0fe57)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause
3  *
4  * Copyright (c) 2002-2006 Rice University
5  * Copyright (c) 2007-2011 Alan L. Cox <alc@cs.rice.edu>
6  * All rights reserved.
7  *
8  * This software was developed for the FreeBSD Project by Alan L. Cox,
9  * Olivier Crameri, Peter Druschel, Sitaram Iyer, and Juan Navarro.
10  *
11  * Redistribution and use in source and binary forms, with or without
12  * modification, are permitted provided that the following conditions
13  * are met:
14  * 1. Redistributions of source code must retain the above copyright
15  *    notice, this list of conditions and the following disclaimer.
16  * 2. Redistributions in binary form must reproduce the above copyright
17  *    notice, this list of conditions and the following disclaimer in the
18  *    documentation and/or other materials provided with the distribution.
19  *
20  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
21  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
22  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
23  * A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE COPYRIGHT
24  * HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
25  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
26  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS
27  * OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
28  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
29  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY
30  * WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
31  * POSSIBILITY OF SUCH DAMAGE.
32  */
33 
34 /*
35  *	Superpage reservation management module
36  *
37  * Any external functions defined by this module are only to be used by the
38  * virtual memory system.
39  */
40 
41 #include <sys/cdefs.h>
42 __FBSDID("$FreeBSD$");
43 
44 #include "opt_vm.h"
45 
46 #include <sys/param.h>
47 #include <sys/kernel.h>
48 #include <sys/lock.h>
49 #include <sys/malloc.h>
50 #include <sys/mutex.h>
51 #include <sys/queue.h>
52 #include <sys/rwlock.h>
53 #include <sys/sbuf.h>
54 #include <sys/sysctl.h>
55 #include <sys/systm.h>
56 #include <sys/bitstring.h>
57 #include <sys/counter.h>
58 #include <sys/ktr.h>
59 #include <sys/vmmeter.h>
60 #include <sys/smp.h>
61 
62 #include <vm/vm.h>
63 #include <vm/vm_extern.h>
64 #include <vm/vm_param.h>
65 #include <vm/vm_object.h>
66 #include <vm/vm_page.h>
67 #include <vm/vm_pageout.h>
68 #include <vm/vm_pagequeue.h>
69 #include <vm/vm_phys.h>
70 #include <vm/vm_radix.h>
71 #include <vm/vm_reserv.h>
72 
73 /*
74  * The reservation system supports the speculative allocation of large physical
75  * pages ("superpages").  Speculative allocation enables the fully automatic
76  * utilization of superpages by the virtual memory system.  In other words, no
77  * programmatic directives are required to use superpages.
78  */
79 
80 #if VM_NRESERVLEVEL > 0
81 
82 #ifndef VM_LEVEL_0_ORDER_MAX
83 #define	VM_LEVEL_0_ORDER_MAX	VM_LEVEL_0_ORDER
84 #endif
85 
86 /*
87  * The number of small pages that are contained in a level 0 reservation
88  */
89 #define	VM_LEVEL_0_NPAGES	(1 << VM_LEVEL_0_ORDER)
90 #define	VM_LEVEL_0_NPAGES_MAX	(1 << VM_LEVEL_0_ORDER_MAX)
91 
92 /*
93  * The number of bits by which a physical address is shifted to obtain the
94  * reservation number
95  */
96 #define	VM_LEVEL_0_SHIFT	(VM_LEVEL_0_ORDER + PAGE_SHIFT)
97 
98 /*
99  * The size of a level 0 reservation in bytes
100  */
101 #define	VM_LEVEL_0_SIZE		(1 << VM_LEVEL_0_SHIFT)
102 
103 /*
104  * Computes the index of the small page underlying the given (object, pindex)
105  * within the reservation's array of small pages.
106  */
107 #define	VM_RESERV_INDEX(object, pindex)	\
108     (((object)->pg_color + (pindex)) & (VM_LEVEL_0_NPAGES - 1))
109 
110 /*
111  * Number of elapsed ticks before we update the LRU queue position.  Used
112  * to reduce contention and churn on the list.
113  */
114 #define	PARTPOPSLOP	1
115 
116 /*
117  * The reservation structure
118  *
119  * A reservation structure is constructed whenever a large physical page is
120  * speculatively allocated to an object.  The reservation provides the small
121  * physical pages for the range [pindex, pindex + VM_LEVEL_0_NPAGES) of offsets
122  * within that object.  The reservation's "popcnt" tracks the number of these
123  * small physical pages that are in use at any given time.  When and if the
124  * reservation is not fully utilized, it appears in the queue of partially
125  * populated reservations.  The reservation always appears on the containing
126  * object's list of reservations.
127  *
128  * A partially populated reservation can be broken and reclaimed at any time.
129  *
130  * c - constant after boot
131  * d - vm_reserv_domain_lock
132  * o - vm_reserv_object_lock
133  * r - vm_reserv_lock
134  * s - vm_reserv_domain_scan_lock
135  */
136 struct vm_reserv {
137 	struct mtx	lock;			/* reservation lock. */
138 	TAILQ_ENTRY(vm_reserv) partpopq;	/* (d, r) per-domain queue. */
139 	LIST_ENTRY(vm_reserv) objq;		/* (o, r) object queue */
140 	vm_object_t	object;			/* (o, r) containing object */
141 	vm_pindex_t	pindex;			/* (o, r) offset in object */
142 	vm_page_t	pages;			/* (c) first page  */
143 	uint16_t	popcnt;			/* (r) # of pages in use */
144 	uint8_t		domain;			/* (c) NUMA domain. */
145 	char		inpartpopq;		/* (d, r) */
146 	int		lasttick;		/* (r) last pop update tick. */
147 	bitstr_t	bit_decl(popmap, VM_LEVEL_0_NPAGES_MAX);
148 						/* (r) bit vector, used pages */
149 };
150 
151 TAILQ_HEAD(vm_reserv_queue, vm_reserv);
152 
153 #define	vm_reserv_lockptr(rv)		(&(rv)->lock)
154 #define	vm_reserv_assert_locked(rv)					\
155 	    mtx_assert(vm_reserv_lockptr(rv), MA_OWNED)
156 #define	vm_reserv_lock(rv)		mtx_lock(vm_reserv_lockptr(rv))
157 #define	vm_reserv_trylock(rv)		mtx_trylock(vm_reserv_lockptr(rv))
158 #define	vm_reserv_unlock(rv)		mtx_unlock(vm_reserv_lockptr(rv))
159 
160 /*
161  * The reservation array
162  *
163  * This array is analoguous in function to vm_page_array.  It differs in the
164  * respect that it may contain a greater number of useful reservation
165  * structures than there are (physical) superpages.  These "invalid"
166  * reservation structures exist to trade-off space for time in the
167  * implementation of vm_reserv_from_page().  Invalid reservation structures are
168  * distinguishable from "valid" reservation structures by inspecting the
169  * reservation's "pages" field.  Invalid reservation structures have a NULL
170  * "pages" field.
171  *
172  * vm_reserv_from_page() maps a small (physical) page to an element of this
173  * array by computing a physical reservation number from the page's physical
174  * address.  The physical reservation number is used as the array index.
175  *
176  * An "active" reservation is a valid reservation structure that has a non-NULL
177  * "object" field and a non-zero "popcnt" field.  In other words, every active
178  * reservation belongs to a particular object.  Moreover, every active
179  * reservation has an entry in the containing object's list of reservations.
180  */
181 static vm_reserv_t vm_reserv_array;
182 
183 /*
184  * The per-domain partially populated reservation queues
185  *
186  * These queues enable the fast recovery of an unused free small page from a
187  * partially populated reservation.  The reservation at the head of a queue
188  * is the least recently changed, partially populated reservation.
189  *
190  * Access to this queue is synchronized by the per-domain reservation lock.
191  * Threads reclaiming free pages from the queue must hold the per-domain scan
192  * lock.
193  */
194 struct vm_reserv_domain {
195 	struct mtx 		lock;
196 	struct vm_reserv_queue	partpop;	/* (d) */
197 	struct vm_reserv	marker;		/* (d, s) scan marker/lock */
198 } __aligned(CACHE_LINE_SIZE);
199 
200 static struct vm_reserv_domain vm_rvd[MAXMEMDOM];
201 
202 #define	vm_reserv_domain_lockptr(d)	(&vm_rvd[(d)].lock)
203 #define	vm_reserv_domain_assert_locked(d)	\
204 	mtx_assert(vm_reserv_domain_lockptr(d), MA_OWNED)
205 #define	vm_reserv_domain_lock(d)	mtx_lock(vm_reserv_domain_lockptr(d))
206 #define	vm_reserv_domain_unlock(d)	mtx_unlock(vm_reserv_domain_lockptr(d))
207 
208 #define	vm_reserv_domain_scan_lock(d)	mtx_lock(&vm_rvd[(d)].marker.lock)
209 #define	vm_reserv_domain_scan_unlock(d)	mtx_unlock(&vm_rvd[(d)].marker.lock)
210 
211 static SYSCTL_NODE(_vm, OID_AUTO, reserv, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
212     "Reservation Info");
213 
214 static COUNTER_U64_DEFINE_EARLY(vm_reserv_broken);
215 SYSCTL_COUNTER_U64(_vm_reserv, OID_AUTO, broken, CTLFLAG_RD,
216     &vm_reserv_broken, "Cumulative number of broken reservations");
217 
218 static COUNTER_U64_DEFINE_EARLY(vm_reserv_freed);
219 SYSCTL_COUNTER_U64(_vm_reserv, OID_AUTO, freed, CTLFLAG_RD,
220     &vm_reserv_freed, "Cumulative number of freed reservations");
221 
222 static int sysctl_vm_reserv_fullpop(SYSCTL_HANDLER_ARGS);
223 
224 SYSCTL_PROC(_vm_reserv, OID_AUTO, fullpop, CTLTYPE_INT | CTLFLAG_MPSAFE | CTLFLAG_RD,
225     NULL, 0, sysctl_vm_reserv_fullpop, "I", "Current number of full reservations");
226 
227 static int sysctl_vm_reserv_partpopq(SYSCTL_HANDLER_ARGS);
228 
229 SYSCTL_OID(_vm_reserv, OID_AUTO, partpopq,
230     CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, 0,
231     sysctl_vm_reserv_partpopq, "A",
232     "Partially populated reservation queues");
233 
234 static COUNTER_U64_DEFINE_EARLY(vm_reserv_reclaimed);
235 SYSCTL_COUNTER_U64(_vm_reserv, OID_AUTO, reclaimed, CTLFLAG_RD,
236     &vm_reserv_reclaimed, "Cumulative number of reclaimed reservations");
237 
238 /*
239  * The object lock pool is used to synchronize the rvq.  We can not use a
240  * pool mutex because it is required before malloc works.
241  *
242  * The "hash" function could be made faster without divide and modulo.
243  */
244 #define	VM_RESERV_OBJ_LOCK_COUNT	MAXCPU
245 
246 struct mtx_padalign vm_reserv_object_mtx[VM_RESERV_OBJ_LOCK_COUNT];
247 
248 #define	vm_reserv_object_lock_idx(object)			\
249 	    (((uintptr_t)object / sizeof(*object)) % VM_RESERV_OBJ_LOCK_COUNT)
250 #define	vm_reserv_object_lock_ptr(object)			\
251 	    &vm_reserv_object_mtx[vm_reserv_object_lock_idx((object))]
252 #define	vm_reserv_object_lock(object)				\
253 	    mtx_lock(vm_reserv_object_lock_ptr((object)))
254 #define	vm_reserv_object_unlock(object)				\
255 	    mtx_unlock(vm_reserv_object_lock_ptr((object)))
256 
257 static void		vm_reserv_break(vm_reserv_t rv);
258 static void		vm_reserv_depopulate(vm_reserv_t rv, int index);
259 static vm_reserv_t	vm_reserv_from_page(vm_page_t m);
260 static boolean_t	vm_reserv_has_pindex(vm_reserv_t rv,
261 			    vm_pindex_t pindex);
262 static void		vm_reserv_populate(vm_reserv_t rv, int index);
263 static void		vm_reserv_reclaim(vm_reserv_t rv);
264 
265 /*
266  * Returns the current number of full reservations.
267  *
268  * Since the number of full reservations is computed without acquiring any
269  * locks, the returned value is inexact.
270  */
271 static int
272 sysctl_vm_reserv_fullpop(SYSCTL_HANDLER_ARGS)
273 {
274 	vm_paddr_t paddr;
275 	struct vm_phys_seg *seg;
276 	vm_reserv_t rv;
277 	int fullpop, segind;
278 
279 	fullpop = 0;
280 	for (segind = 0; segind < vm_phys_nsegs; segind++) {
281 		seg = &vm_phys_segs[segind];
282 		paddr = roundup2(seg->start, VM_LEVEL_0_SIZE);
283 #ifdef VM_PHYSSEG_SPARSE
284 		rv = seg->first_reserv + (paddr >> VM_LEVEL_0_SHIFT) -
285 		    (seg->start >> VM_LEVEL_0_SHIFT);
286 #else
287 		rv = &vm_reserv_array[paddr >> VM_LEVEL_0_SHIFT];
288 #endif
289 		while (paddr + VM_LEVEL_0_SIZE > paddr && paddr +
290 		    VM_LEVEL_0_SIZE <= seg->end) {
291 			fullpop += rv->popcnt == VM_LEVEL_0_NPAGES;
292 			paddr += VM_LEVEL_0_SIZE;
293 			rv++;
294 		}
295 	}
296 	return (sysctl_handle_int(oidp, &fullpop, 0, req));
297 }
298 
299 /*
300  * Describes the current state of the partially populated reservation queue.
301  */
302 static int
303 sysctl_vm_reserv_partpopq(SYSCTL_HANDLER_ARGS)
304 {
305 	struct sbuf sbuf;
306 	vm_reserv_t rv;
307 	int counter, error, domain, level, unused_pages;
308 
309 	error = sysctl_wire_old_buffer(req, 0);
310 	if (error != 0)
311 		return (error);
312 	sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
313 	sbuf_printf(&sbuf, "\nDOMAIN    LEVEL     SIZE  NUMBER\n\n");
314 	for (domain = 0; domain < vm_ndomains; domain++) {
315 		for (level = -1; level <= VM_NRESERVLEVEL - 2; level++) {
316 			counter = 0;
317 			unused_pages = 0;
318 			vm_reserv_domain_lock(domain);
319 			TAILQ_FOREACH(rv, &vm_rvd[domain].partpop, partpopq) {
320 				if (rv == &vm_rvd[domain].marker)
321 					continue;
322 				counter++;
323 				unused_pages += VM_LEVEL_0_NPAGES - rv->popcnt;
324 			}
325 			vm_reserv_domain_unlock(domain);
326 			sbuf_printf(&sbuf, "%6d, %7d, %6dK, %6d\n",
327 			    domain, level,
328 			    unused_pages * ((int)PAGE_SIZE / 1024), counter);
329 		}
330 	}
331 	error = sbuf_finish(&sbuf);
332 	sbuf_delete(&sbuf);
333 	return (error);
334 }
335 
336 /*
337  * Remove a reservation from the object's objq.
338  */
339 static void
340 vm_reserv_remove(vm_reserv_t rv)
341 {
342 	vm_object_t object;
343 
344 	vm_reserv_assert_locked(rv);
345 	CTR5(KTR_VM, "%s: rv %p object %p popcnt %d inpartpop %d",
346 	    __FUNCTION__, rv, rv->object, rv->popcnt, rv->inpartpopq);
347 	KASSERT(rv->object != NULL,
348 	    ("vm_reserv_remove: reserv %p is free", rv));
349 	KASSERT(!rv->inpartpopq,
350 	    ("vm_reserv_remove: reserv %p's inpartpopq is TRUE", rv));
351 	object = rv->object;
352 	vm_reserv_object_lock(object);
353 	LIST_REMOVE(rv, objq);
354 	rv->object = NULL;
355 	vm_reserv_object_unlock(object);
356 }
357 
358 /*
359  * Insert a new reservation into the object's objq.
360  */
361 static void
362 vm_reserv_insert(vm_reserv_t rv, vm_object_t object, vm_pindex_t pindex)
363 {
364 
365 	vm_reserv_assert_locked(rv);
366 	CTR6(KTR_VM,
367 	    "%s: rv %p(%p) object %p new %p popcnt %d",
368 	    __FUNCTION__, rv, rv->pages, rv->object, object,
369 	   rv->popcnt);
370 	KASSERT(rv->object == NULL,
371 	    ("vm_reserv_insert: reserv %p isn't free", rv));
372 	KASSERT(rv->popcnt == 0,
373 	    ("vm_reserv_insert: reserv %p's popcnt is corrupted", rv));
374 	KASSERT(!rv->inpartpopq,
375 	    ("vm_reserv_insert: reserv %p's inpartpopq is TRUE", rv));
376 	KASSERT(bit_ntest(rv->popmap, 0, VM_LEVEL_0_NPAGES - 1, 0),
377 	    ("vm_reserv_insert: reserv %p's popmap is corrupted", rv));
378 	vm_reserv_object_lock(object);
379 	rv->pindex = pindex;
380 	rv->object = object;
381 	rv->lasttick = ticks;
382 	LIST_INSERT_HEAD(&object->rvq, rv, objq);
383 	vm_reserv_object_unlock(object);
384 }
385 
386 /*
387  * Reduces the given reservation's population count.  If the population count
388  * becomes zero, the reservation is destroyed.  Additionally, moves the
389  * reservation to the tail of the partially populated reservation queue if the
390  * population count is non-zero.
391  */
392 static void
393 vm_reserv_depopulate(vm_reserv_t rv, int index)
394 {
395 	struct vm_domain *vmd;
396 
397 	vm_reserv_assert_locked(rv);
398 	CTR5(KTR_VM, "%s: rv %p object %p popcnt %d inpartpop %d",
399 	    __FUNCTION__, rv, rv->object, rv->popcnt, rv->inpartpopq);
400 	KASSERT(rv->object != NULL,
401 	    ("vm_reserv_depopulate: reserv %p is free", rv));
402 	KASSERT(bit_test(rv->popmap, index),
403 	    ("vm_reserv_depopulate: reserv %p's popmap[%d] is clear", rv,
404 	    index));
405 	KASSERT(rv->popcnt > 0,
406 	    ("vm_reserv_depopulate: reserv %p's popcnt is corrupted", rv));
407 	KASSERT(rv->domain < vm_ndomains,
408 	    ("vm_reserv_depopulate: reserv %p's domain is corrupted %d",
409 	    rv, rv->domain));
410 	if (rv->popcnt == VM_LEVEL_0_NPAGES) {
411 		KASSERT(rv->pages->psind == 1,
412 		    ("vm_reserv_depopulate: reserv %p is already demoted",
413 		    rv));
414 		rv->pages->psind = 0;
415 	}
416 	bit_clear(rv->popmap, index);
417 	rv->popcnt--;
418 	if ((unsigned)(ticks - rv->lasttick) >= PARTPOPSLOP ||
419 	    rv->popcnt == 0) {
420 		vm_reserv_domain_lock(rv->domain);
421 		if (rv->inpartpopq) {
422 			TAILQ_REMOVE(&vm_rvd[rv->domain].partpop, rv, partpopq);
423 			rv->inpartpopq = FALSE;
424 		}
425 		if (rv->popcnt != 0) {
426 			rv->inpartpopq = TRUE;
427 			TAILQ_INSERT_TAIL(&vm_rvd[rv->domain].partpop, rv,
428 			    partpopq);
429 		}
430 		vm_reserv_domain_unlock(rv->domain);
431 		rv->lasttick = ticks;
432 	}
433 	vmd = VM_DOMAIN(rv->domain);
434 	if (rv->popcnt == 0) {
435 		vm_reserv_remove(rv);
436 		vm_domain_free_lock(vmd);
437 		vm_phys_free_pages(rv->pages, VM_LEVEL_0_ORDER);
438 		vm_domain_free_unlock(vmd);
439 		counter_u64_add(vm_reserv_freed, 1);
440 	}
441 	vm_domain_freecnt_inc(vmd, 1);
442 }
443 
444 /*
445  * Returns the reservation to which the given page might belong.
446  */
447 static __inline vm_reserv_t
448 vm_reserv_from_page(vm_page_t m)
449 {
450 #ifdef VM_PHYSSEG_SPARSE
451 	struct vm_phys_seg *seg;
452 
453 	seg = &vm_phys_segs[m->segind];
454 	return (seg->first_reserv + (VM_PAGE_TO_PHYS(m) >> VM_LEVEL_0_SHIFT) -
455 	    (seg->start >> VM_LEVEL_0_SHIFT));
456 #else
457 	return (&vm_reserv_array[VM_PAGE_TO_PHYS(m) >> VM_LEVEL_0_SHIFT]);
458 #endif
459 }
460 
461 /*
462  * Returns an existing reservation or NULL and initialized successor pointer.
463  */
464 static vm_reserv_t
465 vm_reserv_from_object(vm_object_t object, vm_pindex_t pindex,
466     vm_page_t mpred, vm_page_t *msuccp)
467 {
468 	vm_reserv_t rv;
469 	vm_page_t msucc;
470 
471 	msucc = NULL;
472 	if (mpred != NULL) {
473 		KASSERT(mpred->object == object,
474 		    ("vm_reserv_from_object: object doesn't contain mpred"));
475 		KASSERT(mpred->pindex < pindex,
476 		    ("vm_reserv_from_object: mpred doesn't precede pindex"));
477 		rv = vm_reserv_from_page(mpred);
478 		if (rv->object == object && vm_reserv_has_pindex(rv, pindex))
479 			goto found;
480 		msucc = TAILQ_NEXT(mpred, listq);
481 	} else
482 		msucc = TAILQ_FIRST(&object->memq);
483 	if (msucc != NULL) {
484 		KASSERT(msucc->pindex > pindex,
485 		    ("vm_reserv_from_object: msucc doesn't succeed pindex"));
486 		rv = vm_reserv_from_page(msucc);
487 		if (rv->object == object && vm_reserv_has_pindex(rv, pindex))
488 			goto found;
489 	}
490 	rv = NULL;
491 
492 found:
493 	*msuccp = msucc;
494 
495 	return (rv);
496 }
497 
498 /*
499  * Returns TRUE if the given reservation contains the given page index and
500  * FALSE otherwise.
501  */
502 static __inline boolean_t
503 vm_reserv_has_pindex(vm_reserv_t rv, vm_pindex_t pindex)
504 {
505 
506 	return (((pindex - rv->pindex) & ~(VM_LEVEL_0_NPAGES - 1)) == 0);
507 }
508 
509 /*
510  * Increases the given reservation's population count.  Moves the reservation
511  * to the tail of the partially populated reservation queue.
512  */
513 static void
514 vm_reserv_populate(vm_reserv_t rv, int index)
515 {
516 
517 	vm_reserv_assert_locked(rv);
518 	CTR5(KTR_VM, "%s: rv %p object %p popcnt %d inpartpop %d",
519 	    __FUNCTION__, rv, rv->object, rv->popcnt, rv->inpartpopq);
520 	KASSERT(rv->object != NULL,
521 	    ("vm_reserv_populate: reserv %p is free", rv));
522 	KASSERT(!bit_test(rv->popmap, index),
523 	    ("vm_reserv_populate: reserv %p's popmap[%d] is set", rv,
524 	    index));
525 	KASSERT(rv->popcnt < VM_LEVEL_0_NPAGES,
526 	    ("vm_reserv_populate: reserv %p is already full", rv));
527 	KASSERT(rv->pages->psind == 0,
528 	    ("vm_reserv_populate: reserv %p is already promoted", rv));
529 	KASSERT(rv->domain < vm_ndomains,
530 	    ("vm_reserv_populate: reserv %p's domain is corrupted %d",
531 	    rv, rv->domain));
532 	bit_set(rv->popmap, index);
533 	rv->popcnt++;
534 	if ((unsigned)(ticks - rv->lasttick) < PARTPOPSLOP &&
535 	    rv->inpartpopq && rv->popcnt != VM_LEVEL_0_NPAGES)
536 		return;
537 	rv->lasttick = ticks;
538 	vm_reserv_domain_lock(rv->domain);
539 	if (rv->inpartpopq) {
540 		TAILQ_REMOVE(&vm_rvd[rv->domain].partpop, rv, partpopq);
541 		rv->inpartpopq = FALSE;
542 	}
543 	if (rv->popcnt < VM_LEVEL_0_NPAGES) {
544 		rv->inpartpopq = TRUE;
545 		TAILQ_INSERT_TAIL(&vm_rvd[rv->domain].partpop, rv, partpopq);
546 	} else {
547 		KASSERT(rv->pages->psind == 0,
548 		    ("vm_reserv_populate: reserv %p is already promoted",
549 		    rv));
550 		rv->pages->psind = 1;
551 	}
552 	vm_reserv_domain_unlock(rv->domain);
553 }
554 
555 /*
556  * Allocates a contiguous set of physical pages of the given size "npages"
557  * from existing or newly created reservations.  All of the physical pages
558  * must be at or above the given physical address "low" and below the given
559  * physical address "high".  The given value "alignment" determines the
560  * alignment of the first physical page in the set.  If the given value
561  * "boundary" is non-zero, then the set of physical pages cannot cross any
562  * physical address boundary that is a multiple of that value.  Both
563  * "alignment" and "boundary" must be a power of two.
564  *
565  * The page "mpred" must immediately precede the offset "pindex" within the
566  * specified object.
567  *
568  * The object must be locked.
569  */
570 vm_page_t
571 vm_reserv_alloc_contig(vm_object_t object, vm_pindex_t pindex, int domain,
572     int req, vm_page_t mpred, u_long npages, vm_paddr_t low, vm_paddr_t high,
573     u_long alignment, vm_paddr_t boundary)
574 {
575 	struct vm_domain *vmd;
576 	vm_paddr_t pa, size;
577 	vm_page_t m, m_ret, msucc;
578 	vm_pindex_t first, leftcap, rightcap;
579 	vm_reserv_t rv;
580 	u_long allocpages, maxpages, minpages;
581 	int i, index, n;
582 
583 	VM_OBJECT_ASSERT_WLOCKED(object);
584 	KASSERT(npages != 0, ("vm_reserv_alloc_contig: npages is 0"));
585 
586 	/*
587 	 * Is a reservation fundamentally impossible?
588 	 */
589 	if (pindex < VM_RESERV_INDEX(object, pindex) ||
590 	    pindex + npages > object->size)
591 		return (NULL);
592 
593 	/*
594 	 * All reservations of a particular size have the same alignment.
595 	 * Assuming that the first page is allocated from a reservation, the
596 	 * least significant bits of its physical address can be determined
597 	 * from its offset from the beginning of the reservation and the size
598 	 * of the reservation.
599 	 *
600 	 * Could the specified index within a reservation of the smallest
601 	 * possible size satisfy the alignment and boundary requirements?
602 	 */
603 	pa = VM_RESERV_INDEX(object, pindex) << PAGE_SHIFT;
604 	size = npages << PAGE_SHIFT;
605 	if (!vm_addr_ok(pa, size, alignment, boundary))
606 		return (NULL);
607 
608 	/*
609 	 * Look for an existing reservation.
610 	 */
611 	rv = vm_reserv_from_object(object, pindex, mpred, &msucc);
612 	if (rv != NULL) {
613 		KASSERT(object != kernel_object || rv->domain == domain,
614 		    ("vm_reserv_alloc_contig: domain mismatch"));
615 		index = VM_RESERV_INDEX(object, pindex);
616 		/* Does the allocation fit within the reservation? */
617 		if (index + npages > VM_LEVEL_0_NPAGES)
618 			return (NULL);
619 		domain = rv->domain;
620 		vmd = VM_DOMAIN(domain);
621 		vm_reserv_lock(rv);
622 		/* Handle reclaim race. */
623 		if (rv->object != object)
624 			goto out;
625 		m = &rv->pages[index];
626 		pa = VM_PAGE_TO_PHYS(m);
627 		if (pa < low || pa + size > high ||
628 		    !vm_addr_ok(pa, size, alignment, boundary))
629 			goto out;
630 		/* Handle vm_page_rename(m, new_object, ...). */
631 		if (!bit_ntest(rv->popmap, index, index + npages - 1, 0))
632 			goto out;
633 		if (!vm_domain_allocate(vmd, req, npages))
634 			goto out;
635 		for (i = 0; i < npages; i++)
636 			vm_reserv_populate(rv, index + i);
637 		vm_reserv_unlock(rv);
638 		return (m);
639 out:
640 		vm_reserv_unlock(rv);
641 		return (NULL);
642 	}
643 
644 	/*
645 	 * Could at least one reservation fit between the first index to the
646 	 * left that can be used ("leftcap") and the first index to the right
647 	 * that cannot be used ("rightcap")?
648 	 *
649 	 * We must synchronize with the reserv object lock to protect the
650 	 * pindex/object of the resulting reservations against rename while
651 	 * we are inspecting.
652 	 */
653 	first = pindex - VM_RESERV_INDEX(object, pindex);
654 	minpages = VM_RESERV_INDEX(object, pindex) + npages;
655 	maxpages = roundup2(minpages, VM_LEVEL_0_NPAGES);
656 	allocpages = maxpages;
657 	vm_reserv_object_lock(object);
658 	if (mpred != NULL) {
659 		if ((rv = vm_reserv_from_page(mpred))->object != object)
660 			leftcap = mpred->pindex + 1;
661 		else
662 			leftcap = rv->pindex + VM_LEVEL_0_NPAGES;
663 		if (leftcap > first) {
664 			vm_reserv_object_unlock(object);
665 			return (NULL);
666 		}
667 	}
668 	if (msucc != NULL) {
669 		if ((rv = vm_reserv_from_page(msucc))->object != object)
670 			rightcap = msucc->pindex;
671 		else
672 			rightcap = rv->pindex;
673 		if (first + maxpages > rightcap) {
674 			if (maxpages == VM_LEVEL_0_NPAGES) {
675 				vm_reserv_object_unlock(object);
676 				return (NULL);
677 			}
678 
679 			/*
680 			 * At least one reservation will fit between "leftcap"
681 			 * and "rightcap".  However, a reservation for the
682 			 * last of the requested pages will not fit.  Reduce
683 			 * the size of the upcoming allocation accordingly.
684 			 */
685 			allocpages = minpages;
686 		}
687 	}
688 	vm_reserv_object_unlock(object);
689 
690 	/*
691 	 * Would the last new reservation extend past the end of the object?
692 	 *
693 	 * If the object is unlikely to grow don't allocate a reservation for
694 	 * the tail.
695 	 */
696 	if ((object->flags & OBJ_ANON) == 0 &&
697 	    first + maxpages > object->size) {
698 		if (maxpages == VM_LEVEL_0_NPAGES)
699 			return (NULL);
700 		allocpages = minpages;
701 	}
702 
703 	/*
704 	 * Allocate the physical pages.  The alignment and boundary specified
705 	 * for this allocation may be different from the alignment and
706 	 * boundary specified for the requested pages.  For instance, the
707 	 * specified index may not be the first page within the first new
708 	 * reservation.
709 	 */
710 	m = NULL;
711 	vmd = VM_DOMAIN(domain);
712 	if (vm_domain_allocate(vmd, req, npages)) {
713 		vm_domain_free_lock(vmd);
714 		m = vm_phys_alloc_contig(domain, allocpages, low, high,
715 		    ulmax(alignment, VM_LEVEL_0_SIZE),
716 		    boundary > VM_LEVEL_0_SIZE ? boundary : 0);
717 		vm_domain_free_unlock(vmd);
718 		if (m == NULL) {
719 			vm_domain_freecnt_inc(vmd, npages);
720 			return (NULL);
721 		}
722 	} else
723 		return (NULL);
724 	KASSERT(vm_page_domain(m) == domain,
725 	    ("vm_reserv_alloc_contig: Page domain does not match requested."));
726 
727 	/*
728 	 * The allocated physical pages always begin at a reservation
729 	 * boundary, but they do not always end at a reservation boundary.
730 	 * Initialize every reservation that is completely covered by the
731 	 * allocated physical pages.
732 	 */
733 	m_ret = NULL;
734 	index = VM_RESERV_INDEX(object, pindex);
735 	do {
736 		rv = vm_reserv_from_page(m);
737 		KASSERT(rv->pages == m,
738 		    ("vm_reserv_alloc_contig: reserv %p's pages is corrupted",
739 		    rv));
740 		vm_reserv_lock(rv);
741 		vm_reserv_insert(rv, object, first);
742 		n = ulmin(VM_LEVEL_0_NPAGES - index, npages);
743 		for (i = 0; i < n; i++)
744 			vm_reserv_populate(rv, index + i);
745 		npages -= n;
746 		if (m_ret == NULL) {
747 			m_ret = &rv->pages[index];
748 			index = 0;
749 		}
750 		vm_reserv_unlock(rv);
751 		m += VM_LEVEL_0_NPAGES;
752 		first += VM_LEVEL_0_NPAGES;
753 		allocpages -= VM_LEVEL_0_NPAGES;
754 	} while (allocpages >= VM_LEVEL_0_NPAGES);
755 	return (m_ret);
756 }
757 
758 /*
759  * Allocate a physical page from an existing or newly created reservation.
760  *
761  * The page "mpred" must immediately precede the offset "pindex" within the
762  * specified object.
763  *
764  * The object must be locked.
765  */
766 vm_page_t
767 vm_reserv_alloc_page(vm_object_t object, vm_pindex_t pindex, int domain,
768     int req, vm_page_t mpred)
769 {
770 	struct vm_domain *vmd;
771 	vm_page_t m, msucc;
772 	vm_pindex_t first, leftcap, rightcap;
773 	vm_reserv_t rv;
774 	int index;
775 
776 	VM_OBJECT_ASSERT_WLOCKED(object);
777 
778 	/*
779 	 * Is a reservation fundamentally impossible?
780 	 */
781 	if (pindex < VM_RESERV_INDEX(object, pindex) ||
782 	    pindex >= object->size)
783 		return (NULL);
784 
785 	/*
786 	 * Look for an existing reservation.
787 	 */
788 	rv = vm_reserv_from_object(object, pindex, mpred, &msucc);
789 	if (rv != NULL) {
790 		KASSERT(object != kernel_object || rv->domain == domain,
791 		    ("vm_reserv_alloc_page: domain mismatch"));
792 		domain = rv->domain;
793 		vmd = VM_DOMAIN(domain);
794 		index = VM_RESERV_INDEX(object, pindex);
795 		m = &rv->pages[index];
796 		vm_reserv_lock(rv);
797 		/* Handle reclaim race. */
798 		if (rv->object != object ||
799 		    /* Handle vm_page_rename(m, new_object, ...). */
800 		    bit_test(rv->popmap, index)) {
801 			m = NULL;
802 			goto out;
803 		}
804 		if (vm_domain_allocate(vmd, req, 1) == 0)
805 			m = NULL;
806 		else
807 			vm_reserv_populate(rv, index);
808 out:
809 		vm_reserv_unlock(rv);
810 		return (m);
811 	}
812 
813 	/*
814 	 * Could a reservation fit between the first index to the left that
815 	 * can be used and the first index to the right that cannot be used?
816 	 *
817 	 * We must synchronize with the reserv object lock to protect the
818 	 * pindex/object of the resulting reservations against rename while
819 	 * we are inspecting.
820 	 */
821 	first = pindex - VM_RESERV_INDEX(object, pindex);
822 	vm_reserv_object_lock(object);
823 	if (mpred != NULL) {
824 		if ((rv = vm_reserv_from_page(mpred))->object != object)
825 			leftcap = mpred->pindex + 1;
826 		else
827 			leftcap = rv->pindex + VM_LEVEL_0_NPAGES;
828 		if (leftcap > first) {
829 			vm_reserv_object_unlock(object);
830 			return (NULL);
831 		}
832 	}
833 	if (msucc != NULL) {
834 		if ((rv = vm_reserv_from_page(msucc))->object != object)
835 			rightcap = msucc->pindex;
836 		else
837 			rightcap = rv->pindex;
838 		if (first + VM_LEVEL_0_NPAGES > rightcap) {
839 			vm_reserv_object_unlock(object);
840 			return (NULL);
841 		}
842 	}
843 	vm_reserv_object_unlock(object);
844 
845 	/*
846 	 * Would the last new reservation extend past the end of the object?
847 	 *
848 	 * If the object is unlikely to grow don't allocate a reservation for
849 	 * the tail.
850 	 */
851 	if ((object->flags & OBJ_ANON) == 0 &&
852 	    first + VM_LEVEL_0_NPAGES > object->size)
853 		return (NULL);
854 
855 	/*
856 	 * Allocate and populate the new reservation.
857 	 */
858 	m = NULL;
859 	vmd = VM_DOMAIN(domain);
860 	if (vm_domain_allocate(vmd, req, 1)) {
861 		vm_domain_free_lock(vmd);
862 		m = vm_phys_alloc_pages(domain, VM_FREEPOOL_DEFAULT,
863 		    VM_LEVEL_0_ORDER);
864 		vm_domain_free_unlock(vmd);
865 		if (m == NULL) {
866 			vm_domain_freecnt_inc(vmd, 1);
867 			return (NULL);
868 		}
869 	} else
870 		return (NULL);
871 	rv = vm_reserv_from_page(m);
872 	vm_reserv_lock(rv);
873 	KASSERT(rv->pages == m,
874 	    ("vm_reserv_alloc_page: reserv %p's pages is corrupted", rv));
875 	vm_reserv_insert(rv, object, first);
876 	index = VM_RESERV_INDEX(object, pindex);
877 	vm_reserv_populate(rv, index);
878 	vm_reserv_unlock(rv);
879 
880 	return (&rv->pages[index]);
881 }
882 
883 /*
884  * Breaks the given reservation.  All free pages in the reservation
885  * are returned to the physical memory allocator.  The reservation's
886  * population count and map are reset to their initial state.
887  *
888  * The given reservation must not be in the partially populated reservation
889  * queue.
890  */
891 static void
892 vm_reserv_break(vm_reserv_t rv)
893 {
894 	int hi, lo, pos;
895 
896 	vm_reserv_assert_locked(rv);
897 	CTR5(KTR_VM, "%s: rv %p object %p popcnt %d inpartpop %d",
898 	    __FUNCTION__, rv, rv->object, rv->popcnt, rv->inpartpopq);
899 	vm_reserv_remove(rv);
900 	rv->pages->psind = 0;
901 	hi = lo = -1;
902 	pos = 0;
903 	for (;;) {
904 		bit_ff_at(rv->popmap, pos, VM_LEVEL_0_NPAGES, lo != hi, &pos);
905 		if (lo == hi) {
906 			if (pos == -1)
907 				break;
908 			lo = pos;
909 			continue;
910 		}
911 		if (pos == -1)
912 			pos = VM_LEVEL_0_NPAGES;
913 		hi = pos;
914 		vm_domain_free_lock(VM_DOMAIN(rv->domain));
915 		vm_phys_enqueue_contig(&rv->pages[lo], hi - lo);
916 		vm_domain_free_unlock(VM_DOMAIN(rv->domain));
917 		lo = hi;
918 	}
919 	bit_nclear(rv->popmap, 0, VM_LEVEL_0_NPAGES - 1);
920 	rv->popcnt = 0;
921 	counter_u64_add(vm_reserv_broken, 1);
922 }
923 
924 /*
925  * Breaks all reservations belonging to the given object.
926  */
927 void
928 vm_reserv_break_all(vm_object_t object)
929 {
930 	vm_reserv_t rv;
931 
932 	/*
933 	 * This access of object->rvq is unsynchronized so that the
934 	 * object rvq lock can nest after the domain_free lock.  We
935 	 * must check for races in the results.  However, the object
936 	 * lock prevents new additions, so we are guaranteed that when
937 	 * it returns NULL the object is properly empty.
938 	 */
939 	while ((rv = LIST_FIRST(&object->rvq)) != NULL) {
940 		vm_reserv_lock(rv);
941 		/* Reclaim race. */
942 		if (rv->object != object) {
943 			vm_reserv_unlock(rv);
944 			continue;
945 		}
946 		vm_reserv_domain_lock(rv->domain);
947 		if (rv->inpartpopq) {
948 			TAILQ_REMOVE(&vm_rvd[rv->domain].partpop, rv, partpopq);
949 			rv->inpartpopq = FALSE;
950 		}
951 		vm_reserv_domain_unlock(rv->domain);
952 		vm_reserv_break(rv);
953 		vm_reserv_unlock(rv);
954 	}
955 }
956 
957 /*
958  * Frees the given page if it belongs to a reservation.  Returns TRUE if the
959  * page is freed and FALSE otherwise.
960  */
961 boolean_t
962 vm_reserv_free_page(vm_page_t m)
963 {
964 	vm_reserv_t rv;
965 	boolean_t ret;
966 
967 	rv = vm_reserv_from_page(m);
968 	if (rv->object == NULL)
969 		return (FALSE);
970 	vm_reserv_lock(rv);
971 	/* Re-validate after lock. */
972 	if (rv->object != NULL) {
973 		vm_reserv_depopulate(rv, m - rv->pages);
974 		ret = TRUE;
975 	} else
976 		ret = FALSE;
977 	vm_reserv_unlock(rv);
978 
979 	return (ret);
980 }
981 
982 /*
983  * Initializes the reservation management system.  Specifically, initializes
984  * the reservation array.
985  *
986  * Requires that vm_page_array and first_page are initialized!
987  */
988 void
989 vm_reserv_init(void)
990 {
991 	vm_paddr_t paddr;
992 	struct vm_phys_seg *seg;
993 	struct vm_reserv *rv;
994 	struct vm_reserv_domain *rvd;
995 #ifdef VM_PHYSSEG_SPARSE
996 	vm_pindex_t used;
997 #endif
998 	int i, segind;
999 
1000 	/*
1001 	 * Initialize the reservation array.  Specifically, initialize the
1002 	 * "pages" field for every element that has an underlying superpage.
1003 	 */
1004 #ifdef VM_PHYSSEG_SPARSE
1005 	used = 0;
1006 #endif
1007 	for (segind = 0; segind < vm_phys_nsegs; segind++) {
1008 		seg = &vm_phys_segs[segind];
1009 #ifdef VM_PHYSSEG_SPARSE
1010 		seg->first_reserv = &vm_reserv_array[used];
1011 		used += howmany(seg->end, VM_LEVEL_0_SIZE) -
1012 		    seg->start / VM_LEVEL_0_SIZE;
1013 #else
1014 		seg->first_reserv =
1015 		    &vm_reserv_array[seg->start >> VM_LEVEL_0_SHIFT];
1016 #endif
1017 		paddr = roundup2(seg->start, VM_LEVEL_0_SIZE);
1018 		rv = seg->first_reserv + (paddr >> VM_LEVEL_0_SHIFT) -
1019 		    (seg->start >> VM_LEVEL_0_SHIFT);
1020 		while (paddr + VM_LEVEL_0_SIZE > paddr && paddr +
1021 		    VM_LEVEL_0_SIZE <= seg->end) {
1022 			rv->pages = PHYS_TO_VM_PAGE(paddr);
1023 			rv->domain = seg->domain;
1024 			mtx_init(&rv->lock, "vm reserv", NULL, MTX_DEF);
1025 			paddr += VM_LEVEL_0_SIZE;
1026 			rv++;
1027 		}
1028 	}
1029 	for (i = 0; i < MAXMEMDOM; i++) {
1030 		rvd = &vm_rvd[i];
1031 		mtx_init(&rvd->lock, "vm reserv domain", NULL, MTX_DEF);
1032 		TAILQ_INIT(&rvd->partpop);
1033 		mtx_init(&rvd->marker.lock, "vm reserv marker", NULL, MTX_DEF);
1034 
1035 		/*
1036 		 * Fully populated reservations should never be present in the
1037 		 * partially populated reservation queues.
1038 		 */
1039 		rvd->marker.popcnt = VM_LEVEL_0_NPAGES;
1040 		bit_nset(rvd->marker.popmap, 0, VM_LEVEL_0_NPAGES - 1);
1041 	}
1042 
1043 	for (i = 0; i < VM_RESERV_OBJ_LOCK_COUNT; i++)
1044 		mtx_init(&vm_reserv_object_mtx[i], "resv obj lock", NULL,
1045 		    MTX_DEF);
1046 }
1047 
1048 /*
1049  * Returns true if the given page belongs to a reservation and that page is
1050  * free.  Otherwise, returns false.
1051  */
1052 bool
1053 vm_reserv_is_page_free(vm_page_t m)
1054 {
1055 	vm_reserv_t rv;
1056 
1057 	rv = vm_reserv_from_page(m);
1058 	if (rv->object == NULL)
1059 		return (false);
1060 	return (!bit_test(rv->popmap, m - rv->pages));
1061 }
1062 
1063 /*
1064  * If the given page belongs to a reservation, returns the level of that
1065  * reservation.  Otherwise, returns -1.
1066  */
1067 int
1068 vm_reserv_level(vm_page_t m)
1069 {
1070 	vm_reserv_t rv;
1071 
1072 	rv = vm_reserv_from_page(m);
1073 	return (rv->object != NULL ? 0 : -1);
1074 }
1075 
1076 /*
1077  * Returns a reservation level if the given page belongs to a fully populated
1078  * reservation and -1 otherwise.
1079  */
1080 int
1081 vm_reserv_level_iffullpop(vm_page_t m)
1082 {
1083 	vm_reserv_t rv;
1084 
1085 	rv = vm_reserv_from_page(m);
1086 	return (rv->popcnt == VM_LEVEL_0_NPAGES ? 0 : -1);
1087 }
1088 
1089 /*
1090  * Remove a partially populated reservation from the queue.
1091  */
1092 static void
1093 vm_reserv_dequeue(vm_reserv_t rv)
1094 {
1095 
1096 	vm_reserv_domain_assert_locked(rv->domain);
1097 	vm_reserv_assert_locked(rv);
1098 	CTR5(KTR_VM, "%s: rv %p object %p popcnt %d inpartpop %d",
1099 	    __FUNCTION__, rv, rv->object, rv->popcnt, rv->inpartpopq);
1100 	KASSERT(rv->inpartpopq,
1101 	    ("vm_reserv_reclaim: reserv %p's inpartpopq is FALSE", rv));
1102 
1103 	TAILQ_REMOVE(&vm_rvd[rv->domain].partpop, rv, partpopq);
1104 	rv->inpartpopq = FALSE;
1105 }
1106 
1107 /*
1108  * Breaks the given partially populated reservation, releasing its free pages
1109  * to the physical memory allocator.
1110  */
1111 static void
1112 vm_reserv_reclaim(vm_reserv_t rv)
1113 {
1114 
1115 	vm_reserv_assert_locked(rv);
1116 	CTR5(KTR_VM, "%s: rv %p object %p popcnt %d inpartpop %d",
1117 	    __FUNCTION__, rv, rv->object, rv->popcnt, rv->inpartpopq);
1118 	if (rv->inpartpopq) {
1119 		vm_reserv_domain_lock(rv->domain);
1120 		vm_reserv_dequeue(rv);
1121 		vm_reserv_domain_unlock(rv->domain);
1122 	}
1123 	vm_reserv_break(rv);
1124 	counter_u64_add(vm_reserv_reclaimed, 1);
1125 }
1126 
1127 /*
1128  * Breaks a reservation near the head of the partially populated reservation
1129  * queue, releasing its free pages to the physical memory allocator.  Returns
1130  * TRUE if a reservation is broken and FALSE otherwise.
1131  */
1132 bool
1133 vm_reserv_reclaim_inactive(int domain)
1134 {
1135 	vm_reserv_t rv;
1136 
1137 	vm_reserv_domain_lock(domain);
1138 	TAILQ_FOREACH(rv, &vm_rvd[domain].partpop, partpopq) {
1139 		/*
1140 		 * A locked reservation is likely being updated or reclaimed,
1141 		 * so just skip ahead.
1142 		 */
1143 		if (rv != &vm_rvd[domain].marker && vm_reserv_trylock(rv)) {
1144 			vm_reserv_dequeue(rv);
1145 			break;
1146 		}
1147 	}
1148 	vm_reserv_domain_unlock(domain);
1149 	if (rv != NULL) {
1150 		vm_reserv_reclaim(rv);
1151 		vm_reserv_unlock(rv);
1152 		return (true);
1153 	}
1154 	return (false);
1155 }
1156 
1157 /*
1158  * Determine whether this reservation has free pages that satisfy the given
1159  * request for contiguous physical memory.  Start searching from the lower
1160  * bound, defined by lo, and stop at the upper bound, hi.  Return the index
1161  * of the first satisfactory free page, or -1 if none is found.
1162  */
1163 static int
1164 vm_reserv_find_contig(vm_reserv_t rv, int npages, int lo,
1165     int hi, int ppn_align, int ppn_bound)
1166 {
1167 
1168 	vm_reserv_assert_locked(rv);
1169 	KASSERT(npages <= VM_LEVEL_0_NPAGES - 1,
1170 	    ("%s: Too many pages", __func__));
1171 	KASSERT(ppn_bound <= VM_LEVEL_0_NPAGES,
1172 	    ("%s: Too big a boundary for reservation size", __func__));
1173 	KASSERT(npages <= ppn_bound,
1174 	    ("%s: Too many pages for given boundary", __func__));
1175 	KASSERT(ppn_align != 0 && powerof2(ppn_align),
1176 	    ("ppn_align is not a positive power of 2"));
1177 	KASSERT(ppn_bound != 0 && powerof2(ppn_bound),
1178 	    ("ppn_bound is not a positive power of 2"));
1179 	while (bit_ffc_area_at(rv->popmap, lo, hi, npages, &lo), lo != -1) {
1180 		if (lo < roundup2(lo, ppn_align)) {
1181 			/* Skip to next aligned page. */
1182 			lo = roundup2(lo, ppn_align);
1183 		} else if (roundup2(lo + 1, ppn_bound) >= lo + npages)
1184 			return (lo);
1185 		if (roundup2(lo + 1, ppn_bound) < lo + npages) {
1186 			/* Skip to next boundary-matching page. */
1187 			lo = roundup2(lo + 1, ppn_bound);
1188 		}
1189 	}
1190 	return (-1);
1191 }
1192 
1193 /*
1194  * Searches the partially populated reservation queue for the least recently
1195  * changed reservation with free pages that satisfy the given request for
1196  * contiguous physical memory.  If a satisfactory reservation is found, it is
1197  * broken.  Returns true if a reservation is broken and false otherwise.
1198  */
1199 vm_page_t
1200 vm_reserv_reclaim_contig(int domain, u_long npages, vm_paddr_t low,
1201     vm_paddr_t high, u_long alignment, vm_paddr_t boundary)
1202 {
1203 	struct vm_reserv_queue *queue;
1204 	vm_paddr_t pa, size;
1205 	vm_page_t m_ret;
1206 	vm_reserv_t marker, rv, rvn;
1207 	int hi, lo, posn, ppn_align, ppn_bound;
1208 
1209 	KASSERT(npages > 0, ("npages is 0"));
1210 	KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
1211 	KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
1212 	if (npages > VM_LEVEL_0_NPAGES - 1)
1213 		return (false);
1214 	size = npages << PAGE_SHIFT;
1215 	/*
1216 	 * Ensure that a free range starting at a boundary-multiple
1217 	 * doesn't include a boundary-multiple within it.  Otherwise,
1218 	 * no boundary-constrained allocation is possible.
1219 	 */
1220 	if (!vm_addr_bound_ok(0, size, boundary))
1221 		return (NULL);
1222 	marker = &vm_rvd[domain].marker;
1223 	queue = &vm_rvd[domain].partpop;
1224 	/*
1225 	 * Compute shifted alignment, boundary values for page-based
1226 	 * calculations.  Constrain to range [1, VM_LEVEL_0_NPAGES] to
1227 	 * avoid overflow.
1228 	 */
1229 	ppn_align = (int)(ulmin(ulmax(PAGE_SIZE, alignment),
1230 	    VM_LEVEL_0_SIZE) >> PAGE_SHIFT);
1231 	ppn_bound = boundary == 0 ? VM_LEVEL_0_NPAGES :
1232 	    (int)(MIN(MAX(PAGE_SIZE, boundary),
1233             VM_LEVEL_0_SIZE) >> PAGE_SHIFT);
1234 
1235 	vm_reserv_domain_scan_lock(domain);
1236 	vm_reserv_domain_lock(domain);
1237 	TAILQ_FOREACH_SAFE(rv, queue, partpopq, rvn) {
1238 		pa = VM_PAGE_TO_PHYS(&rv->pages[0]);
1239 		if (pa + VM_LEVEL_0_SIZE - size < low) {
1240 			/* This entire reservation is too low; go to next. */
1241 			continue;
1242 		}
1243 		if (pa + size > high) {
1244 			/* This entire reservation is too high; go to next. */
1245 			continue;
1246 		}
1247 		if (!vm_addr_align_ok(pa, alignment)) {
1248 			/* This entire reservation is unaligned; go to next. */
1249 			continue;
1250 		}
1251 
1252 		if (vm_reserv_trylock(rv) == 0) {
1253 			TAILQ_INSERT_AFTER(queue, rv, marker, partpopq);
1254 			vm_reserv_domain_unlock(domain);
1255 			vm_reserv_lock(rv);
1256 			if (TAILQ_PREV(marker, vm_reserv_queue, partpopq) !=
1257 			    rv) {
1258 				vm_reserv_unlock(rv);
1259 				vm_reserv_domain_lock(domain);
1260 				rvn = TAILQ_NEXT(marker, partpopq);
1261 				TAILQ_REMOVE(queue, marker, partpopq);
1262 				continue;
1263 			}
1264 			vm_reserv_domain_lock(domain);
1265 			TAILQ_REMOVE(queue, marker, partpopq);
1266 		}
1267 		vm_reserv_domain_unlock(domain);
1268 		lo = (pa >= low) ? 0 :
1269 		    (int)((low + PAGE_MASK - pa) >> PAGE_SHIFT);
1270 		hi = (pa + VM_LEVEL_0_SIZE <= high) ? VM_LEVEL_0_NPAGES :
1271 		    (int)((high - pa) >> PAGE_SHIFT);
1272 		posn = vm_reserv_find_contig(rv, (int)npages, lo, hi,
1273 		    ppn_align, ppn_bound);
1274 		if (posn >= 0) {
1275 			vm_reserv_domain_scan_unlock(domain);
1276 			/* Allocate requested space */
1277 			rv->popcnt += npages;
1278 			bit_nset(rv->popmap, posn, posn + npages - 1);
1279 			vm_reserv_reclaim(rv);
1280 			vm_reserv_unlock(rv);
1281 			m_ret = &rv->pages[posn];
1282 			pa = VM_PAGE_TO_PHYS(m_ret);
1283 			KASSERT(vm_addr_ok(pa, size, alignment, boundary),
1284 			    ("%s: adjusted address not aligned/bounded to "
1285 			     "%lx/%jx",
1286 			     __func__, alignment, (uintmax_t)boundary));
1287 			return (m_ret);
1288 		}
1289 		vm_reserv_domain_lock(domain);
1290 		rvn = TAILQ_NEXT(rv, partpopq);
1291 		vm_reserv_unlock(rv);
1292 	}
1293 	vm_reserv_domain_unlock(domain);
1294 	vm_reserv_domain_scan_unlock(domain);
1295 	return (NULL);
1296 }
1297 
1298 /*
1299  * Transfers the reservation underlying the given page to a new object.
1300  *
1301  * The object must be locked.
1302  */
1303 void
1304 vm_reserv_rename(vm_page_t m, vm_object_t new_object, vm_object_t old_object,
1305     vm_pindex_t old_object_offset)
1306 {
1307 	vm_reserv_t rv;
1308 
1309 	VM_OBJECT_ASSERT_WLOCKED(new_object);
1310 	rv = vm_reserv_from_page(m);
1311 	if (rv->object == old_object) {
1312 		vm_reserv_lock(rv);
1313 		CTR6(KTR_VM,
1314 		    "%s: rv %p object %p new %p popcnt %d inpartpop %d",
1315 		    __FUNCTION__, rv, rv->object, new_object, rv->popcnt,
1316 		    rv->inpartpopq);
1317 		if (rv->object == old_object) {
1318 			vm_reserv_object_lock(old_object);
1319 			rv->object = NULL;
1320 			LIST_REMOVE(rv, objq);
1321 			vm_reserv_object_unlock(old_object);
1322 			vm_reserv_object_lock(new_object);
1323 			rv->object = new_object;
1324 			rv->pindex -= old_object_offset;
1325 			LIST_INSERT_HEAD(&new_object->rvq, rv, objq);
1326 			vm_reserv_object_unlock(new_object);
1327 		}
1328 		vm_reserv_unlock(rv);
1329 	}
1330 }
1331 
1332 /*
1333  * Returns the size (in bytes) of a reservation of the specified level.
1334  */
1335 int
1336 vm_reserv_size(int level)
1337 {
1338 
1339 	switch (level) {
1340 	case 0:
1341 		return (VM_LEVEL_0_SIZE);
1342 	case -1:
1343 		return (PAGE_SIZE);
1344 	default:
1345 		return (0);
1346 	}
1347 }
1348 
1349 /*
1350  * Allocates the virtual and physical memory required by the reservation
1351  * management system's data structures, in particular, the reservation array.
1352  */
1353 vm_paddr_t
1354 vm_reserv_startup(vm_offset_t *vaddr, vm_paddr_t end)
1355 {
1356 	vm_paddr_t new_end;
1357 	vm_pindex_t count;
1358 	size_t size;
1359 	int i;
1360 
1361 	count = 0;
1362 	for (i = 0; i < vm_phys_nsegs; i++) {
1363 #ifdef VM_PHYSSEG_SPARSE
1364 		count += howmany(vm_phys_segs[i].end, VM_LEVEL_0_SIZE) -
1365 		    vm_phys_segs[i].start / VM_LEVEL_0_SIZE;
1366 #else
1367 		count = MAX(count,
1368 		    howmany(vm_phys_segs[i].end, VM_LEVEL_0_SIZE));
1369 #endif
1370 	}
1371 
1372 	for (i = 0; phys_avail[i + 1] != 0; i += 2) {
1373 #ifdef VM_PHYSSEG_SPARSE
1374 		count += howmany(phys_avail[i + 1], VM_LEVEL_0_SIZE) -
1375 		    phys_avail[i] / VM_LEVEL_0_SIZE;
1376 #else
1377 		count = MAX(count,
1378 		    howmany(phys_avail[i + 1], VM_LEVEL_0_SIZE));
1379 #endif
1380 	}
1381 
1382 	/*
1383 	 * Calculate the size (in bytes) of the reservation array.  Rounding up
1384 	 * for partial superpages at boundaries, as every small page is mapped
1385 	 * to an element in the reservation array based on its physical address.
1386 	 * Thus, the number of elements in the reservation array can be greater
1387 	 * than the number of superpages.
1388 	 */
1389 	size = count * sizeof(struct vm_reserv);
1390 
1391 	/*
1392 	 * Allocate and map the physical memory for the reservation array.  The
1393 	 * next available virtual address is returned by reference.
1394 	 */
1395 	new_end = end - round_page(size);
1396 	vm_reserv_array = (void *)(uintptr_t)pmap_map(vaddr, new_end, end,
1397 	    VM_PROT_READ | VM_PROT_WRITE);
1398 	bzero(vm_reserv_array, size);
1399 
1400 	/*
1401 	 * Return the next available physical address.
1402 	 */
1403 	return (new_end);
1404 }
1405 
1406 /*
1407  * Returns the superpage containing the given page.
1408  */
1409 vm_page_t
1410 vm_reserv_to_superpage(vm_page_t m)
1411 {
1412 	vm_reserv_t rv;
1413 
1414 	VM_OBJECT_ASSERT_LOCKED(m->object);
1415 	rv = vm_reserv_from_page(m);
1416 	if (rv->object == m->object && rv->popcnt == VM_LEVEL_0_NPAGES)
1417 		m = rv->pages;
1418 	else
1419 		m = NULL;
1420 
1421 	return (m);
1422 }
1423 
1424 #endif	/* VM_NRESERVLEVEL > 0 */
1425