xref: /freebsd/sys/vm/vm_reserv.c (revision 6580f5c38dd5b01aeeaed16b370f1a12423437f0)
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 #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 #include <sys/bitstring.h>
55 #include <sys/counter.h>
56 #include <sys/ktr.h>
57 #include <sys/vmmeter.h>
58 #include <sys/smp.h>
59 
60 #include <vm/vm.h>
61 #include <vm/vm_extern.h>
62 #include <vm/vm_param.h>
63 #include <vm/vm_object.h>
64 #include <vm/vm_page.h>
65 #include <vm/vm_pageout.h>
66 #include <vm/vm_pagequeue.h>
67 #include <vm/vm_phys.h>
68 #include <vm/vm_radix.h>
69 #include <vm/vm_reserv.h>
70 
71 /*
72  * The reservation system supports the speculative allocation of large physical
73  * pages ("superpages").  Speculative allocation enables the fully automatic
74  * utilization of superpages by the virtual memory system.  In other words, no
75  * programmatic directives are required to use superpages.
76  */
77 
78 #if VM_NRESERVLEVEL > 0
79 
80 #ifndef VM_LEVEL_0_ORDER_MAX
81 #define	VM_LEVEL_0_ORDER_MAX	VM_LEVEL_0_ORDER
82 #endif
83 
84 /*
85  * The number of small pages that are contained in a level 0 reservation
86  */
87 #define	VM_LEVEL_0_NPAGES	(1 << VM_LEVEL_0_ORDER)
88 #define	VM_LEVEL_0_NPAGES_MAX	(1 << VM_LEVEL_0_ORDER_MAX)
89 
90 /*
91  * The number of bits by which a physical address is shifted to obtain the
92  * reservation number
93  */
94 #define	VM_LEVEL_0_SHIFT	(VM_LEVEL_0_ORDER + PAGE_SHIFT)
95 
96 /*
97  * The size of a level 0 reservation in bytes
98  */
99 #define	VM_LEVEL_0_SIZE		(1 << VM_LEVEL_0_SHIFT)
100 
101 /*
102  * Computes the index of the small page underlying the given (object, pindex)
103  * within the reservation's array of small pages.
104  */
105 #define	VM_RESERV_INDEX(object, pindex)	\
106     (((object)->pg_color + (pindex)) & (VM_LEVEL_0_NPAGES - 1))
107 
108 /*
109  * Number of elapsed ticks before we update the LRU queue position.  Used
110  * to reduce contention and churn on the list.
111  */
112 #define	PARTPOPSLOP	1
113 
114 /*
115  * The reservation structure
116  *
117  * A reservation structure is constructed whenever a large physical page is
118  * speculatively allocated to an object.  The reservation provides the small
119  * physical pages for the range [pindex, pindex + VM_LEVEL_0_NPAGES) of offsets
120  * within that object.  The reservation's "popcnt" tracks the number of these
121  * small physical pages that are in use at any given time.  When and if the
122  * reservation is not fully utilized, it appears in the queue of partially
123  * populated reservations.  The reservation always appears on the containing
124  * object's list of reservations.
125  *
126  * A partially populated reservation can be broken and reclaimed at any time.
127  *
128  * c - constant after boot
129  * d - vm_reserv_domain_lock
130  * o - vm_reserv_object_lock
131  * r - vm_reserv_lock
132  * s - vm_reserv_domain_scan_lock
133  */
134 struct vm_reserv {
135 	struct mtx	lock;			/* reservation lock. */
136 	TAILQ_ENTRY(vm_reserv) partpopq;	/* (d, r) per-domain queue. */
137 	LIST_ENTRY(vm_reserv) objq;		/* (o, r) object queue */
138 	vm_object_t	object;			/* (o, r) containing object */
139 	vm_pindex_t	pindex;			/* (o, r) offset in object */
140 	vm_page_t	pages;			/* (c) first page  */
141 	uint16_t	popcnt;			/* (r) # of pages in use */
142 	uint8_t		domain;			/* (c) NUMA domain. */
143 	char		inpartpopq;		/* (d, r) */
144 	int		lasttick;		/* (r) last pop update tick. */
145 	bitstr_t	bit_decl(popmap, VM_LEVEL_0_NPAGES_MAX);
146 						/* (r) bit vector, used pages */
147 };
148 
149 TAILQ_HEAD(vm_reserv_queue, vm_reserv);
150 
151 #define	vm_reserv_lockptr(rv)		(&(rv)->lock)
152 #define	vm_reserv_assert_locked(rv)					\
153 	    mtx_assert(vm_reserv_lockptr(rv), MA_OWNED)
154 #define	vm_reserv_lock(rv)		mtx_lock(vm_reserv_lockptr(rv))
155 #define	vm_reserv_trylock(rv)		mtx_trylock(vm_reserv_lockptr(rv))
156 #define	vm_reserv_unlock(rv)		mtx_unlock(vm_reserv_lockptr(rv))
157 
158 /*
159  * The reservation array
160  *
161  * This array is analoguous in function to vm_page_array.  It differs in the
162  * respect that it may contain a greater number of useful reservation
163  * structures than there are (physical) superpages.  These "invalid"
164  * reservation structures exist to trade-off space for time in the
165  * implementation of vm_reserv_from_page().  Invalid reservation structures are
166  * distinguishable from "valid" reservation structures by inspecting the
167  * reservation's "pages" field.  Invalid reservation structures have a NULL
168  * "pages" field.
169  *
170  * vm_reserv_from_page() maps a small (physical) page to an element of this
171  * array by computing a physical reservation number from the page's physical
172  * address.  The physical reservation number is used as the array index.
173  *
174  * An "active" reservation is a valid reservation structure that has a non-NULL
175  * "object" field and a non-zero "popcnt" field.  In other words, every active
176  * reservation belongs to a particular object.  Moreover, every active
177  * reservation has an entry in the containing object's list of reservations.
178  */
179 static vm_reserv_t vm_reserv_array;
180 
181 /*
182  * The per-domain partially populated reservation queues
183  *
184  * These queues enable the fast recovery of an unused free small page from a
185  * partially populated reservation.  The reservation at the head of a queue
186  * is the least recently changed, partially populated reservation.
187  *
188  * Access to this queue is synchronized by the per-domain reservation lock.
189  * Threads reclaiming free pages from the queue must hold the per-domain scan
190  * lock.
191  */
192 struct vm_reserv_domain {
193 	struct mtx 		lock;
194 	struct vm_reserv_queue	partpop;	/* (d) */
195 	struct vm_reserv	marker;		/* (d, s) scan marker/lock */
196 } __aligned(CACHE_LINE_SIZE);
197 
198 static struct vm_reserv_domain vm_rvd[MAXMEMDOM];
199 
200 #define	vm_reserv_domain_lockptr(d)	(&vm_rvd[(d)].lock)
201 #define	vm_reserv_domain_assert_locked(d)	\
202 	mtx_assert(vm_reserv_domain_lockptr(d), MA_OWNED)
203 #define	vm_reserv_domain_lock(d)	mtx_lock(vm_reserv_domain_lockptr(d))
204 #define	vm_reserv_domain_unlock(d)	mtx_unlock(vm_reserv_domain_lockptr(d))
205 
206 #define	vm_reserv_domain_scan_lock(d)	mtx_lock(&vm_rvd[(d)].marker.lock)
207 #define	vm_reserv_domain_scan_unlock(d)	mtx_unlock(&vm_rvd[(d)].marker.lock)
208 
209 static SYSCTL_NODE(_vm, OID_AUTO, reserv, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
210     "Reservation Info");
211 
212 static COUNTER_U64_DEFINE_EARLY(vm_reserv_broken);
213 SYSCTL_COUNTER_U64(_vm_reserv, OID_AUTO, broken, CTLFLAG_RD,
214     &vm_reserv_broken, "Cumulative number of broken reservations");
215 
216 static COUNTER_U64_DEFINE_EARLY(vm_reserv_freed);
217 SYSCTL_COUNTER_U64(_vm_reserv, OID_AUTO, freed, CTLFLAG_RD,
218     &vm_reserv_freed, "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_MPSAFE | CTLFLAG_RD,
223     NULL, 0, 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,
228     CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, 0,
229     sysctl_vm_reserv_partpopq, "A",
230     "Partially populated reservation queues");
231 
232 static COUNTER_U64_DEFINE_EARLY(vm_reserv_reclaimed);
233 SYSCTL_COUNTER_U64(_vm_reserv, OID_AUTO, reclaimed, CTLFLAG_RD,
234     &vm_reserv_reclaimed, "Cumulative number of reclaimed reservations");
235 
236 /*
237  * The object lock pool is used to synchronize the rvq.  We can not use a
238  * pool mutex because it is required before malloc works.
239  *
240  * The "hash" function could be made faster without divide and modulo.
241  */
242 #define	VM_RESERV_OBJ_LOCK_COUNT	MAXCPU
243 
244 struct mtx_padalign vm_reserv_object_mtx[VM_RESERV_OBJ_LOCK_COUNT];
245 
246 #define	vm_reserv_object_lock_idx(object)			\
247 	    (((uintptr_t)object / sizeof(*object)) % VM_RESERV_OBJ_LOCK_COUNT)
248 #define	vm_reserv_object_lock_ptr(object)			\
249 	    &vm_reserv_object_mtx[vm_reserv_object_lock_idx((object))]
250 #define	vm_reserv_object_lock(object)				\
251 	    mtx_lock(vm_reserv_object_lock_ptr((object)))
252 #define	vm_reserv_object_unlock(object)				\
253 	    mtx_unlock(vm_reserv_object_lock_ptr((object)))
254 
255 static void		vm_reserv_break(vm_reserv_t rv);
256 static void		vm_reserv_depopulate(vm_reserv_t rv, int index);
257 static vm_reserv_t	vm_reserv_from_page(vm_page_t m);
258 static boolean_t	vm_reserv_has_pindex(vm_reserv_t rv,
259 			    vm_pindex_t pindex);
260 static void		vm_reserv_populate(vm_reserv_t rv, int index);
261 static void		vm_reserv_reclaim(vm_reserv_t rv);
262 
263 /*
264  * Returns the current number of full reservations.
265  *
266  * Since the number of full reservations is computed without acquiring any
267  * locks, the returned value is inexact.
268  */
269 static int
270 sysctl_vm_reserv_fullpop(SYSCTL_HANDLER_ARGS)
271 {
272 	vm_paddr_t paddr;
273 	struct vm_phys_seg *seg;
274 	vm_reserv_t rv;
275 	int fullpop, segind;
276 
277 	fullpop = 0;
278 	for (segind = 0; segind < vm_phys_nsegs; segind++) {
279 		seg = &vm_phys_segs[segind];
280 		paddr = roundup2(seg->start, VM_LEVEL_0_SIZE);
281 #ifdef VM_PHYSSEG_SPARSE
282 		rv = seg->first_reserv + (paddr >> VM_LEVEL_0_SHIFT) -
283 		    (seg->start >> VM_LEVEL_0_SHIFT);
284 #else
285 		rv = &vm_reserv_array[paddr >> VM_LEVEL_0_SHIFT];
286 #endif
287 		while (paddr + VM_LEVEL_0_SIZE > paddr && paddr +
288 		    VM_LEVEL_0_SIZE <= seg->end) {
289 			fullpop += rv->popcnt == VM_LEVEL_0_NPAGES;
290 			paddr += VM_LEVEL_0_SIZE;
291 			rv++;
292 		}
293 	}
294 	return (sysctl_handle_int(oidp, &fullpop, 0, req));
295 }
296 
297 /*
298  * Describes the current state of the partially populated reservation queue.
299  */
300 static int
301 sysctl_vm_reserv_partpopq(SYSCTL_HANDLER_ARGS)
302 {
303 	struct sbuf sbuf;
304 	vm_reserv_t rv;
305 	int counter, error, domain, level, unused_pages;
306 
307 	error = sysctl_wire_old_buffer(req, 0);
308 	if (error != 0)
309 		return (error);
310 	sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
311 	sbuf_printf(&sbuf, "\nDOMAIN    LEVEL     SIZE  NUMBER\n\n");
312 	for (domain = 0; domain < vm_ndomains; domain++) {
313 		for (level = -1; level <= VM_NRESERVLEVEL - 2; level++) {
314 			counter = 0;
315 			unused_pages = 0;
316 			vm_reserv_domain_lock(domain);
317 			TAILQ_FOREACH(rv, &vm_rvd[domain].partpop, partpopq) {
318 				if (rv == &vm_rvd[domain].marker)
319 					continue;
320 				counter++;
321 				unused_pages += VM_LEVEL_0_NPAGES - rv->popcnt;
322 			}
323 			vm_reserv_domain_unlock(domain);
324 			sbuf_printf(&sbuf, "%6d, %7d, %6dK, %6d\n",
325 			    domain, level,
326 			    unused_pages * ((int)PAGE_SIZE / 1024), counter);
327 		}
328 	}
329 	error = sbuf_finish(&sbuf);
330 	sbuf_delete(&sbuf);
331 	return (error);
332 }
333 
334 /*
335  * Remove a reservation from the object's objq.
336  */
337 static void
338 vm_reserv_remove(vm_reserv_t rv)
339 {
340 	vm_object_t object;
341 
342 	vm_reserv_assert_locked(rv);
343 	CTR5(KTR_VM, "%s: rv %p object %p popcnt %d inpartpop %d",
344 	    __FUNCTION__, rv, rv->object, rv->popcnt, rv->inpartpopq);
345 	KASSERT(rv->object != NULL,
346 	    ("vm_reserv_remove: reserv %p is free", rv));
347 	KASSERT(!rv->inpartpopq,
348 	    ("vm_reserv_remove: reserv %p's inpartpopq is TRUE", rv));
349 	object = rv->object;
350 	vm_reserv_object_lock(object);
351 	LIST_REMOVE(rv, objq);
352 	rv->object = NULL;
353 	vm_reserv_object_unlock(object);
354 }
355 
356 /*
357  * Insert a new reservation into the object's objq.
358  */
359 static void
360 vm_reserv_insert(vm_reserv_t rv, vm_object_t object, vm_pindex_t pindex)
361 {
362 
363 	vm_reserv_assert_locked(rv);
364 	CTR6(KTR_VM,
365 	    "%s: rv %p(%p) object %p new %p popcnt %d",
366 	    __FUNCTION__, rv, rv->pages, rv->object, object,
367 	   rv->popcnt);
368 	KASSERT(rv->object == NULL,
369 	    ("vm_reserv_insert: reserv %p isn't free", rv));
370 	KASSERT(rv->popcnt == 0,
371 	    ("vm_reserv_insert: reserv %p's popcnt is corrupted", rv));
372 	KASSERT(!rv->inpartpopq,
373 	    ("vm_reserv_insert: reserv %p's inpartpopq is TRUE", rv));
374 	KASSERT(bit_ntest(rv->popmap, 0, VM_LEVEL_0_NPAGES - 1, 0),
375 	    ("vm_reserv_insert: reserv %p's popmap is corrupted", rv));
376 	vm_reserv_object_lock(object);
377 	rv->pindex = pindex;
378 	rv->object = object;
379 	rv->lasttick = ticks;
380 	LIST_INSERT_HEAD(&object->rvq, rv, objq);
381 	vm_reserv_object_unlock(object);
382 }
383 
384 /*
385  * Reduces the given reservation's population count.  If the population count
386  * becomes zero, the reservation is destroyed.  Additionally, moves the
387  * reservation to the tail of the partially populated reservation queue if the
388  * population count is non-zero.
389  */
390 static void
391 vm_reserv_depopulate(vm_reserv_t rv, int index)
392 {
393 	struct vm_domain *vmd;
394 
395 	vm_reserv_assert_locked(rv);
396 	CTR5(KTR_VM, "%s: rv %p object %p popcnt %d inpartpop %d",
397 	    __FUNCTION__, rv, rv->object, rv->popcnt, rv->inpartpopq);
398 	KASSERT(rv->object != NULL,
399 	    ("vm_reserv_depopulate: reserv %p is free", rv));
400 	KASSERT(bit_test(rv->popmap, index),
401 	    ("vm_reserv_depopulate: reserv %p's popmap[%d] is clear", rv,
402 	    index));
403 	KASSERT(rv->popcnt > 0,
404 	    ("vm_reserv_depopulate: reserv %p's popcnt is corrupted", rv));
405 	KASSERT(rv->domain < vm_ndomains,
406 	    ("vm_reserv_depopulate: reserv %p's domain is corrupted %d",
407 	    rv, rv->domain));
408 	if (rv->popcnt == VM_LEVEL_0_NPAGES) {
409 		KASSERT(rv->pages->psind == 1,
410 		    ("vm_reserv_depopulate: reserv %p is already demoted",
411 		    rv));
412 		rv->pages->psind = 0;
413 	}
414 	bit_clear(rv->popmap, index);
415 	rv->popcnt--;
416 	if ((unsigned)(ticks - rv->lasttick) >= PARTPOPSLOP ||
417 	    rv->popcnt == 0) {
418 		vm_reserv_domain_lock(rv->domain);
419 		if (rv->inpartpopq) {
420 			TAILQ_REMOVE(&vm_rvd[rv->domain].partpop, rv, partpopq);
421 			rv->inpartpopq = FALSE;
422 		}
423 		if (rv->popcnt != 0) {
424 			rv->inpartpopq = TRUE;
425 			TAILQ_INSERT_TAIL(&vm_rvd[rv->domain].partpop, rv,
426 			    partpopq);
427 		}
428 		vm_reserv_domain_unlock(rv->domain);
429 		rv->lasttick = ticks;
430 	}
431 	vmd = VM_DOMAIN(rv->domain);
432 	if (rv->popcnt == 0) {
433 		vm_reserv_remove(rv);
434 		vm_domain_free_lock(vmd);
435 		vm_phys_free_pages(rv->pages, VM_LEVEL_0_ORDER);
436 		vm_domain_free_unlock(vmd);
437 		counter_u64_add(vm_reserv_freed, 1);
438 	}
439 	vm_domain_freecnt_inc(vmd, 1);
440 }
441 
442 /*
443  * Returns the reservation to which the given page might belong.
444  */
445 static __inline vm_reserv_t
446 vm_reserv_from_page(vm_page_t m)
447 {
448 #ifdef VM_PHYSSEG_SPARSE
449 	struct vm_phys_seg *seg;
450 
451 	seg = &vm_phys_segs[m->segind];
452 	return (seg->first_reserv + (VM_PAGE_TO_PHYS(m) >> VM_LEVEL_0_SHIFT) -
453 	    (seg->start >> VM_LEVEL_0_SHIFT));
454 #else
455 	return (&vm_reserv_array[VM_PAGE_TO_PHYS(m) >> VM_LEVEL_0_SHIFT]);
456 #endif
457 }
458 
459 /*
460  * Returns an existing reservation or NULL and initialized successor pointer.
461  */
462 static vm_reserv_t
463 vm_reserv_from_object(vm_object_t object, vm_pindex_t pindex,
464     vm_page_t mpred, vm_page_t *msuccp)
465 {
466 	vm_reserv_t rv;
467 	vm_page_t msucc;
468 
469 	msucc = NULL;
470 	if (mpred != NULL) {
471 		KASSERT(mpred->object == object,
472 		    ("vm_reserv_from_object: object doesn't contain mpred"));
473 		KASSERT(mpred->pindex < pindex,
474 		    ("vm_reserv_from_object: mpred doesn't precede pindex"));
475 		rv = vm_reserv_from_page(mpred);
476 		if (rv->object == object && vm_reserv_has_pindex(rv, pindex))
477 			goto found;
478 		msucc = TAILQ_NEXT(mpred, listq);
479 	} else
480 		msucc = TAILQ_FIRST(&object->memq);
481 	if (msucc != NULL) {
482 		KASSERT(msucc->pindex > pindex,
483 		    ("vm_reserv_from_object: msucc doesn't succeed pindex"));
484 		rv = vm_reserv_from_page(msucc);
485 		if (rv->object == object && vm_reserv_has_pindex(rv, pindex))
486 			goto found;
487 	}
488 	rv = NULL;
489 
490 found:
491 	*msuccp = msucc;
492 
493 	return (rv);
494 }
495 
496 /*
497  * Returns TRUE if the given reservation contains the given page index and
498  * FALSE otherwise.
499  */
500 static __inline boolean_t
501 vm_reserv_has_pindex(vm_reserv_t rv, vm_pindex_t pindex)
502 {
503 
504 	return (((pindex - rv->pindex) & ~(VM_LEVEL_0_NPAGES - 1)) == 0);
505 }
506 
507 /*
508  * Increases the given reservation's population count.  Moves the reservation
509  * to the tail of the partially populated reservation queue.
510  */
511 static void
512 vm_reserv_populate(vm_reserv_t rv, int index)
513 {
514 
515 	vm_reserv_assert_locked(rv);
516 	CTR5(KTR_VM, "%s: rv %p object %p popcnt %d inpartpop %d",
517 	    __FUNCTION__, rv, rv->object, rv->popcnt, rv->inpartpopq);
518 	KASSERT(rv->object != NULL,
519 	    ("vm_reserv_populate: reserv %p is free", rv));
520 	KASSERT(!bit_test(rv->popmap, index),
521 	    ("vm_reserv_populate: reserv %p's popmap[%d] is set", rv,
522 	    index));
523 	KASSERT(rv->popcnt < VM_LEVEL_0_NPAGES,
524 	    ("vm_reserv_populate: reserv %p is already full", rv));
525 	KASSERT(rv->pages->psind == 0,
526 	    ("vm_reserv_populate: reserv %p is already promoted", rv));
527 	KASSERT(rv->domain < vm_ndomains,
528 	    ("vm_reserv_populate: reserv %p's domain is corrupted %d",
529 	    rv, rv->domain));
530 	bit_set(rv->popmap, index);
531 	rv->popcnt++;
532 	if ((unsigned)(ticks - rv->lasttick) < PARTPOPSLOP &&
533 	    rv->inpartpopq && rv->popcnt != VM_LEVEL_0_NPAGES)
534 		return;
535 	rv->lasttick = ticks;
536 	vm_reserv_domain_lock(rv->domain);
537 	if (rv->inpartpopq) {
538 		TAILQ_REMOVE(&vm_rvd[rv->domain].partpop, rv, partpopq);
539 		rv->inpartpopq = FALSE;
540 	}
541 	if (rv->popcnt < VM_LEVEL_0_NPAGES) {
542 		rv->inpartpopq = TRUE;
543 		TAILQ_INSERT_TAIL(&vm_rvd[rv->domain].partpop, rv, partpopq);
544 	} else {
545 		KASSERT(rv->pages->psind == 0,
546 		    ("vm_reserv_populate: reserv %p is already promoted",
547 		    rv));
548 		rv->pages->psind = 1;
549 	}
550 	vm_reserv_domain_unlock(rv->domain);
551 }
552 
553 /*
554  * Allocates a contiguous set of physical pages of the given size "npages"
555  * from existing or newly created reservations.  All of the physical pages
556  * must be at or above the given physical address "low" and below the given
557  * physical address "high".  The given value "alignment" determines the
558  * alignment of the first physical page in the set.  If the given value
559  * "boundary" is non-zero, then the set of physical pages cannot cross any
560  * physical address boundary that is a multiple of that value.  Both
561  * "alignment" and "boundary" must be a power of two.
562  *
563  * The page "mpred" must immediately precede the offset "pindex" within the
564  * specified object.
565  *
566  * The object must be locked.
567  */
568 vm_page_t
569 vm_reserv_alloc_contig(vm_object_t object, vm_pindex_t pindex, int domain,
570     int req, vm_page_t mpred, u_long npages, vm_paddr_t low, vm_paddr_t high,
571     u_long alignment, vm_paddr_t boundary)
572 {
573 	struct vm_domain *vmd;
574 	vm_paddr_t pa, size;
575 	vm_page_t m, m_ret, msucc;
576 	vm_pindex_t first, leftcap, rightcap;
577 	vm_reserv_t rv;
578 	u_long allocpages, maxpages, minpages;
579 	int i, index, n;
580 
581 	VM_OBJECT_ASSERT_WLOCKED(object);
582 	KASSERT(npages != 0, ("vm_reserv_alloc_contig: npages is 0"));
583 
584 	/*
585 	 * Is a reservation fundamentally impossible?
586 	 */
587 	if (pindex < VM_RESERV_INDEX(object, pindex) ||
588 	    pindex + npages > object->size)
589 		return (NULL);
590 
591 	/*
592 	 * All reservations of a particular size have the same alignment.
593 	 * Assuming that the first page is allocated from a reservation, the
594 	 * least significant bits of its physical address can be determined
595 	 * from its offset from the beginning of the reservation and the size
596 	 * of the reservation.
597 	 *
598 	 * Could the specified index within a reservation of the smallest
599 	 * possible size satisfy the alignment and boundary requirements?
600 	 */
601 	pa = VM_RESERV_INDEX(object, pindex) << PAGE_SHIFT;
602 	size = npages << PAGE_SHIFT;
603 	if (!vm_addr_ok(pa, size, alignment, boundary))
604 		return (NULL);
605 
606 	/*
607 	 * Look for an existing reservation.
608 	 */
609 	rv = vm_reserv_from_object(object, pindex, mpred, &msucc);
610 	if (rv != NULL) {
611 		KASSERT(object != kernel_object || rv->domain == domain,
612 		    ("vm_reserv_alloc_contig: domain mismatch"));
613 		index = VM_RESERV_INDEX(object, pindex);
614 		/* Does the allocation fit within the reservation? */
615 		if (index + npages > VM_LEVEL_0_NPAGES)
616 			return (NULL);
617 		domain = rv->domain;
618 		vmd = VM_DOMAIN(domain);
619 		vm_reserv_lock(rv);
620 		/* Handle reclaim race. */
621 		if (rv->object != object)
622 			goto out;
623 		m = &rv->pages[index];
624 		pa = VM_PAGE_TO_PHYS(m);
625 		if (pa < low || pa + size > high ||
626 		    !vm_addr_ok(pa, size, alignment, boundary))
627 			goto out;
628 		/* Handle vm_page_rename(m, new_object, ...). */
629 		if (!bit_ntest(rv->popmap, index, index + npages - 1, 0))
630 			goto out;
631 		if (!vm_domain_allocate(vmd, req, npages))
632 			goto out;
633 		for (i = 0; i < npages; i++)
634 			vm_reserv_populate(rv, index + i);
635 		vm_reserv_unlock(rv);
636 		return (m);
637 out:
638 		vm_reserv_unlock(rv);
639 		return (NULL);
640 	}
641 
642 	/*
643 	 * Could at least one reservation fit between the first index to the
644 	 * left that can be used ("leftcap") and the first index to the right
645 	 * that cannot be used ("rightcap")?
646 	 *
647 	 * We must synchronize with the reserv object lock to protect the
648 	 * pindex/object of the resulting reservations against rename while
649 	 * we are inspecting.
650 	 */
651 	first = pindex - VM_RESERV_INDEX(object, pindex);
652 	minpages = VM_RESERV_INDEX(object, pindex) + npages;
653 	maxpages = roundup2(minpages, VM_LEVEL_0_NPAGES);
654 	allocpages = maxpages;
655 	vm_reserv_object_lock(object);
656 	if (mpred != NULL) {
657 		if ((rv = vm_reserv_from_page(mpred))->object != object)
658 			leftcap = mpred->pindex + 1;
659 		else
660 			leftcap = rv->pindex + VM_LEVEL_0_NPAGES;
661 		if (leftcap > first) {
662 			vm_reserv_object_unlock(object);
663 			return (NULL);
664 		}
665 	}
666 	if (msucc != NULL) {
667 		if ((rv = vm_reserv_from_page(msucc))->object != object)
668 			rightcap = msucc->pindex;
669 		else
670 			rightcap = rv->pindex;
671 		if (first + maxpages > rightcap) {
672 			if (maxpages == VM_LEVEL_0_NPAGES) {
673 				vm_reserv_object_unlock(object);
674 				return (NULL);
675 			}
676 
677 			/*
678 			 * At least one reservation will fit between "leftcap"
679 			 * and "rightcap".  However, a reservation for the
680 			 * last of the requested pages will not fit.  Reduce
681 			 * the size of the upcoming allocation accordingly.
682 			 */
683 			allocpages = minpages;
684 		}
685 	}
686 	vm_reserv_object_unlock(object);
687 
688 	/*
689 	 * Would the last new reservation extend past the end of the object?
690 	 *
691 	 * If the object is unlikely to grow don't allocate a reservation for
692 	 * the tail.
693 	 */
694 	if ((object->flags & OBJ_ANON) == 0 &&
695 	    first + maxpages > object->size) {
696 		if (maxpages == VM_LEVEL_0_NPAGES)
697 			return (NULL);
698 		allocpages = minpages;
699 	}
700 
701 	/*
702 	 * Allocate the physical pages.  The alignment and boundary specified
703 	 * for this allocation may be different from the alignment and
704 	 * boundary specified for the requested pages.  For instance, the
705 	 * specified index may not be the first page within the first new
706 	 * reservation.
707 	 */
708 	m = NULL;
709 	vmd = VM_DOMAIN(domain);
710 	if (vm_domain_allocate(vmd, req, npages)) {
711 		vm_domain_free_lock(vmd);
712 		m = vm_phys_alloc_contig(domain, allocpages, low, high,
713 		    ulmax(alignment, VM_LEVEL_0_SIZE),
714 		    boundary > VM_LEVEL_0_SIZE ? boundary : 0);
715 		vm_domain_free_unlock(vmd);
716 		if (m == NULL) {
717 			vm_domain_freecnt_inc(vmd, npages);
718 			return (NULL);
719 		}
720 	} else
721 		return (NULL);
722 	KASSERT(vm_page_domain(m) == domain,
723 	    ("vm_reserv_alloc_contig: Page domain does not match requested."));
724 
725 	/*
726 	 * The allocated physical pages always begin at a reservation
727 	 * boundary, but they do not always end at a reservation boundary.
728 	 * Initialize every reservation that is completely covered by the
729 	 * allocated physical pages.
730 	 */
731 	m_ret = NULL;
732 	index = VM_RESERV_INDEX(object, pindex);
733 	do {
734 		rv = vm_reserv_from_page(m);
735 		KASSERT(rv->pages == m,
736 		    ("vm_reserv_alloc_contig: reserv %p's pages is corrupted",
737 		    rv));
738 		vm_reserv_lock(rv);
739 		vm_reserv_insert(rv, object, first);
740 		n = ulmin(VM_LEVEL_0_NPAGES - index, npages);
741 		for (i = 0; i < n; i++)
742 			vm_reserv_populate(rv, index + i);
743 		npages -= n;
744 		if (m_ret == NULL) {
745 			m_ret = &rv->pages[index];
746 			index = 0;
747 		}
748 		vm_reserv_unlock(rv);
749 		m += VM_LEVEL_0_NPAGES;
750 		first += VM_LEVEL_0_NPAGES;
751 		allocpages -= VM_LEVEL_0_NPAGES;
752 	} while (allocpages >= VM_LEVEL_0_NPAGES);
753 	return (m_ret);
754 }
755 
756 /*
757  * Allocate a physical page from an existing or newly created reservation.
758  *
759  * The page "mpred" must immediately precede the offset "pindex" within the
760  * specified object.
761  *
762  * The object must be locked.
763  */
764 vm_page_t
765 vm_reserv_alloc_page(vm_object_t object, vm_pindex_t pindex, int domain,
766     int req, vm_page_t mpred)
767 {
768 	struct vm_domain *vmd;
769 	vm_page_t m, msucc;
770 	vm_pindex_t first, leftcap, rightcap;
771 	vm_reserv_t rv;
772 	int index;
773 
774 	VM_OBJECT_ASSERT_WLOCKED(object);
775 
776 	/*
777 	 * Is a reservation fundamentally impossible?
778 	 */
779 	if (pindex < VM_RESERV_INDEX(object, pindex) ||
780 	    pindex >= object->size)
781 		return (NULL);
782 
783 	/*
784 	 * Look for an existing reservation.
785 	 */
786 	rv = vm_reserv_from_object(object, pindex, mpred, &msucc);
787 	if (rv != NULL) {
788 		KASSERT(object != kernel_object || rv->domain == domain,
789 		    ("vm_reserv_alloc_page: domain mismatch"));
790 		domain = rv->domain;
791 		vmd = VM_DOMAIN(domain);
792 		index = VM_RESERV_INDEX(object, pindex);
793 		m = &rv->pages[index];
794 		vm_reserv_lock(rv);
795 		/* Handle reclaim race. */
796 		if (rv->object != object ||
797 		    /* Handle vm_page_rename(m, new_object, ...). */
798 		    bit_test(rv->popmap, index)) {
799 			m = NULL;
800 			goto out;
801 		}
802 		if (vm_domain_allocate(vmd, req, 1) == 0)
803 			m = NULL;
804 		else
805 			vm_reserv_populate(rv, index);
806 out:
807 		vm_reserv_unlock(rv);
808 		return (m);
809 	}
810 
811 	/*
812 	 * Could a reservation fit between the first index to the left that
813 	 * can be used and the first index to the right that cannot be used?
814 	 *
815 	 * We must synchronize with the reserv object lock to protect the
816 	 * pindex/object of the resulting reservations against rename while
817 	 * we are inspecting.
818 	 */
819 	first = pindex - VM_RESERV_INDEX(object, pindex);
820 	vm_reserv_object_lock(object);
821 	if (mpred != NULL) {
822 		if ((rv = vm_reserv_from_page(mpred))->object != object)
823 			leftcap = mpred->pindex + 1;
824 		else
825 			leftcap = rv->pindex + VM_LEVEL_0_NPAGES;
826 		if (leftcap > first) {
827 			vm_reserv_object_unlock(object);
828 			return (NULL);
829 		}
830 	}
831 	if (msucc != NULL) {
832 		if ((rv = vm_reserv_from_page(msucc))->object != object)
833 			rightcap = msucc->pindex;
834 		else
835 			rightcap = rv->pindex;
836 		if (first + VM_LEVEL_0_NPAGES > rightcap) {
837 			vm_reserv_object_unlock(object);
838 			return (NULL);
839 		}
840 	}
841 	vm_reserv_object_unlock(object);
842 
843 	/*
844 	 * Would the last new reservation extend past the end of the object?
845 	 *
846 	 * If the object is unlikely to grow don't allocate a reservation for
847 	 * the tail.
848 	 */
849 	if ((object->flags & OBJ_ANON) == 0 &&
850 	    first + VM_LEVEL_0_NPAGES > object->size)
851 		return (NULL);
852 
853 	/*
854 	 * Allocate and populate the new reservation.
855 	 */
856 	m = NULL;
857 	vmd = VM_DOMAIN(domain);
858 	if (vm_domain_allocate(vmd, req, 1)) {
859 		vm_domain_free_lock(vmd);
860 		m = vm_phys_alloc_pages(domain, VM_FREEPOOL_DEFAULT,
861 		    VM_LEVEL_0_ORDER);
862 		vm_domain_free_unlock(vmd);
863 		if (m == NULL) {
864 			vm_domain_freecnt_inc(vmd, 1);
865 			return (NULL);
866 		}
867 	} else
868 		return (NULL);
869 	rv = vm_reserv_from_page(m);
870 	vm_reserv_lock(rv);
871 	KASSERT(rv->pages == m,
872 	    ("vm_reserv_alloc_page: reserv %p's pages is corrupted", rv));
873 	vm_reserv_insert(rv, object, first);
874 	index = VM_RESERV_INDEX(object, pindex);
875 	vm_reserv_populate(rv, index);
876 	vm_reserv_unlock(rv);
877 
878 	return (&rv->pages[index]);
879 }
880 
881 /*
882  * Breaks the given reservation.  All free pages in the reservation
883  * are returned to the physical memory allocator.  The reservation's
884  * population count and map are reset to their initial state.
885  *
886  * The given reservation must not be in the partially populated reservation
887  * queue.
888  */
889 static void
890 vm_reserv_break(vm_reserv_t rv)
891 {
892 	int hi, lo, pos;
893 
894 	vm_reserv_assert_locked(rv);
895 	CTR5(KTR_VM, "%s: rv %p object %p popcnt %d inpartpop %d",
896 	    __FUNCTION__, rv, rv->object, rv->popcnt, rv->inpartpopq);
897 	vm_reserv_remove(rv);
898 	rv->pages->psind = 0;
899 	hi = lo = -1;
900 	pos = 0;
901 	for (;;) {
902 		bit_ff_at(rv->popmap, pos, VM_LEVEL_0_NPAGES, lo != hi, &pos);
903 		if (lo == hi) {
904 			if (pos == -1)
905 				break;
906 			lo = pos;
907 			continue;
908 		}
909 		if (pos == -1)
910 			pos = VM_LEVEL_0_NPAGES;
911 		hi = pos;
912 		vm_domain_free_lock(VM_DOMAIN(rv->domain));
913 		vm_phys_enqueue_contig(&rv->pages[lo], hi - lo);
914 		vm_domain_free_unlock(VM_DOMAIN(rv->domain));
915 		lo = hi;
916 	}
917 	bit_nclear(rv->popmap, 0, VM_LEVEL_0_NPAGES - 1);
918 	rv->popcnt = 0;
919 	counter_u64_add(vm_reserv_broken, 1);
920 }
921 
922 /*
923  * Breaks all reservations belonging to the given object.
924  */
925 void
926 vm_reserv_break_all(vm_object_t object)
927 {
928 	vm_reserv_t rv;
929 
930 	/*
931 	 * This access of object->rvq is unsynchronized so that the
932 	 * object rvq lock can nest after the domain_free lock.  We
933 	 * must check for races in the results.  However, the object
934 	 * lock prevents new additions, so we are guaranteed that when
935 	 * it returns NULL the object is properly empty.
936 	 */
937 	while ((rv = LIST_FIRST(&object->rvq)) != NULL) {
938 		vm_reserv_lock(rv);
939 		/* Reclaim race. */
940 		if (rv->object != object) {
941 			vm_reserv_unlock(rv);
942 			continue;
943 		}
944 		vm_reserv_domain_lock(rv->domain);
945 		if (rv->inpartpopq) {
946 			TAILQ_REMOVE(&vm_rvd[rv->domain].partpop, rv, partpopq);
947 			rv->inpartpopq = FALSE;
948 		}
949 		vm_reserv_domain_unlock(rv->domain);
950 		vm_reserv_break(rv);
951 		vm_reserv_unlock(rv);
952 	}
953 }
954 
955 /*
956  * Frees the given page if it belongs to a reservation.  Returns TRUE if the
957  * page is freed and FALSE otherwise.
958  */
959 boolean_t
960 vm_reserv_free_page(vm_page_t m)
961 {
962 	vm_reserv_t rv;
963 	boolean_t ret;
964 
965 	rv = vm_reserv_from_page(m);
966 	if (rv->object == NULL)
967 		return (FALSE);
968 	vm_reserv_lock(rv);
969 	/* Re-validate after lock. */
970 	if (rv->object != NULL) {
971 		vm_reserv_depopulate(rv, m - rv->pages);
972 		ret = TRUE;
973 	} else
974 		ret = FALSE;
975 	vm_reserv_unlock(rv);
976 
977 	return (ret);
978 }
979 
980 /*
981  * Initializes the reservation management system.  Specifically, initializes
982  * the reservation array.
983  *
984  * Requires that vm_page_array and first_page are initialized!
985  */
986 void
987 vm_reserv_init(void)
988 {
989 	vm_paddr_t paddr;
990 	struct vm_phys_seg *seg;
991 	struct vm_reserv *rv;
992 	struct vm_reserv_domain *rvd;
993 #ifdef VM_PHYSSEG_SPARSE
994 	vm_pindex_t used;
995 #endif
996 	int i, segind;
997 
998 	/*
999 	 * Initialize the reservation array.  Specifically, initialize the
1000 	 * "pages" field for every element that has an underlying superpage.
1001 	 */
1002 #ifdef VM_PHYSSEG_SPARSE
1003 	used = 0;
1004 #endif
1005 	for (segind = 0; segind < vm_phys_nsegs; segind++) {
1006 		seg = &vm_phys_segs[segind];
1007 #ifdef VM_PHYSSEG_SPARSE
1008 		seg->first_reserv = &vm_reserv_array[used];
1009 		used += howmany(seg->end, VM_LEVEL_0_SIZE) -
1010 		    seg->start / VM_LEVEL_0_SIZE;
1011 #else
1012 		seg->first_reserv =
1013 		    &vm_reserv_array[seg->start >> VM_LEVEL_0_SHIFT];
1014 #endif
1015 		paddr = roundup2(seg->start, VM_LEVEL_0_SIZE);
1016 		rv = seg->first_reserv + (paddr >> VM_LEVEL_0_SHIFT) -
1017 		    (seg->start >> VM_LEVEL_0_SHIFT);
1018 		while (paddr + VM_LEVEL_0_SIZE > paddr && paddr +
1019 		    VM_LEVEL_0_SIZE <= seg->end) {
1020 			rv->pages = PHYS_TO_VM_PAGE(paddr);
1021 			rv->domain = seg->domain;
1022 			mtx_init(&rv->lock, "vm reserv", NULL, MTX_DEF);
1023 			paddr += VM_LEVEL_0_SIZE;
1024 			rv++;
1025 		}
1026 	}
1027 	for (i = 0; i < MAXMEMDOM; i++) {
1028 		rvd = &vm_rvd[i];
1029 		mtx_init(&rvd->lock, "vm reserv domain", NULL, MTX_DEF);
1030 		TAILQ_INIT(&rvd->partpop);
1031 		mtx_init(&rvd->marker.lock, "vm reserv marker", NULL, MTX_DEF);
1032 
1033 		/*
1034 		 * Fully populated reservations should never be present in the
1035 		 * partially populated reservation queues.
1036 		 */
1037 		rvd->marker.popcnt = VM_LEVEL_0_NPAGES;
1038 		bit_nset(rvd->marker.popmap, 0, VM_LEVEL_0_NPAGES - 1);
1039 	}
1040 
1041 	for (i = 0; i < VM_RESERV_OBJ_LOCK_COUNT; i++)
1042 		mtx_init(&vm_reserv_object_mtx[i], "resv obj lock", NULL,
1043 		    MTX_DEF);
1044 }
1045 
1046 /*
1047  * Returns true if the given page belongs to a reservation and that page is
1048  * free.  Otherwise, returns false.
1049  */
1050 bool
1051 vm_reserv_is_page_free(vm_page_t m)
1052 {
1053 	vm_reserv_t rv;
1054 
1055 	rv = vm_reserv_from_page(m);
1056 	if (rv->object == NULL)
1057 		return (false);
1058 	return (!bit_test(rv->popmap, m - rv->pages));
1059 }
1060 
1061 /*
1062  * Returns true if the given page is part of a block of npages, starting at a
1063  * multiple of npages, that are all allocated.  Otherwise, returns false.
1064  */
1065 bool
1066 vm_reserv_is_populated(vm_page_t m, int npages)
1067 {
1068 	vm_reserv_t rv;
1069 	int index;
1070 
1071 	KASSERT(npages <= VM_LEVEL_0_NPAGES,
1072 	    ("%s: npages %d exceeds VM_LEVEL_0_NPAGES", __func__, npages));
1073 	KASSERT(powerof2(npages),
1074 	    ("%s: npages %d is not a power of 2", __func__, npages));
1075 	rv = vm_reserv_from_page(m);
1076 	if (rv->object == NULL)
1077 		return (false);
1078 	index = rounddown2(m - rv->pages, npages);
1079 	return (bit_ntest(rv->popmap, index, index + npages - 1, 1));
1080 }
1081 
1082 /*
1083  * If the given page belongs to a reservation, returns the level of that
1084  * reservation.  Otherwise, returns -1.
1085  */
1086 int
1087 vm_reserv_level(vm_page_t m)
1088 {
1089 	vm_reserv_t rv;
1090 
1091 	rv = vm_reserv_from_page(m);
1092 	return (rv->object != NULL ? 0 : -1);
1093 }
1094 
1095 /*
1096  * Returns a reservation level if the given page belongs to a fully populated
1097  * reservation and -1 otherwise.
1098  */
1099 int
1100 vm_reserv_level_iffullpop(vm_page_t m)
1101 {
1102 	vm_reserv_t rv;
1103 
1104 	rv = vm_reserv_from_page(m);
1105 	return (rv->popcnt == VM_LEVEL_0_NPAGES ? 0 : -1);
1106 }
1107 
1108 /*
1109  * Remove a partially populated reservation from the queue.
1110  */
1111 static void
1112 vm_reserv_dequeue(vm_reserv_t rv)
1113 {
1114 
1115 	vm_reserv_domain_assert_locked(rv->domain);
1116 	vm_reserv_assert_locked(rv);
1117 	CTR5(KTR_VM, "%s: rv %p object %p popcnt %d inpartpop %d",
1118 	    __FUNCTION__, rv, rv->object, rv->popcnt, rv->inpartpopq);
1119 	KASSERT(rv->inpartpopq,
1120 	    ("vm_reserv_reclaim: reserv %p's inpartpopq is FALSE", rv));
1121 
1122 	TAILQ_REMOVE(&vm_rvd[rv->domain].partpop, rv, partpopq);
1123 	rv->inpartpopq = FALSE;
1124 }
1125 
1126 /*
1127  * Breaks the given partially populated reservation, releasing its free pages
1128  * to the physical memory allocator.
1129  */
1130 static void
1131 vm_reserv_reclaim(vm_reserv_t rv)
1132 {
1133 
1134 	vm_reserv_assert_locked(rv);
1135 	CTR5(KTR_VM, "%s: rv %p object %p popcnt %d inpartpop %d",
1136 	    __FUNCTION__, rv, rv->object, rv->popcnt, rv->inpartpopq);
1137 	if (rv->inpartpopq) {
1138 		vm_reserv_domain_lock(rv->domain);
1139 		vm_reserv_dequeue(rv);
1140 		vm_reserv_domain_unlock(rv->domain);
1141 	}
1142 	vm_reserv_break(rv);
1143 	counter_u64_add(vm_reserv_reclaimed, 1);
1144 }
1145 
1146 /*
1147  * Breaks a reservation near the head of the partially populated reservation
1148  * queue, releasing its free pages to the physical memory allocator.  Returns
1149  * TRUE if a reservation is broken and FALSE otherwise.
1150  */
1151 bool
1152 vm_reserv_reclaim_inactive(int domain)
1153 {
1154 	vm_reserv_t rv;
1155 
1156 	vm_reserv_domain_lock(domain);
1157 	TAILQ_FOREACH(rv, &vm_rvd[domain].partpop, partpopq) {
1158 		/*
1159 		 * A locked reservation is likely being updated or reclaimed,
1160 		 * so just skip ahead.
1161 		 */
1162 		if (rv != &vm_rvd[domain].marker && vm_reserv_trylock(rv)) {
1163 			vm_reserv_dequeue(rv);
1164 			break;
1165 		}
1166 	}
1167 	vm_reserv_domain_unlock(domain);
1168 	if (rv != NULL) {
1169 		vm_reserv_reclaim(rv);
1170 		vm_reserv_unlock(rv);
1171 		return (true);
1172 	}
1173 	return (false);
1174 }
1175 
1176 /*
1177  * Determine whether this reservation has free pages that satisfy the given
1178  * request for contiguous physical memory.  Start searching from the lower
1179  * bound, defined by lo, and stop at the upper bound, hi.  Return the index
1180  * of the first satisfactory free page, or -1 if none is found.
1181  */
1182 static int
1183 vm_reserv_find_contig(vm_reserv_t rv, int npages, int lo,
1184     int hi, int ppn_align, int ppn_bound)
1185 {
1186 
1187 	vm_reserv_assert_locked(rv);
1188 	KASSERT(npages <= VM_LEVEL_0_NPAGES - 1,
1189 	    ("%s: Too many pages", __func__));
1190 	KASSERT(ppn_bound <= VM_LEVEL_0_NPAGES,
1191 	    ("%s: Too big a boundary for reservation size", __func__));
1192 	KASSERT(npages <= ppn_bound,
1193 	    ("%s: Too many pages for given boundary", __func__));
1194 	KASSERT(ppn_align != 0 && powerof2(ppn_align),
1195 	    ("ppn_align is not a positive power of 2"));
1196 	KASSERT(ppn_bound != 0 && powerof2(ppn_bound),
1197 	    ("ppn_bound is not a positive power of 2"));
1198 	while (bit_ffc_area_at(rv->popmap, lo, hi, npages, &lo), lo != -1) {
1199 		if (lo < roundup2(lo, ppn_align)) {
1200 			/* Skip to next aligned page. */
1201 			lo = roundup2(lo, ppn_align);
1202 		} else if (roundup2(lo + 1, ppn_bound) >= lo + npages)
1203 			return (lo);
1204 		if (roundup2(lo + 1, ppn_bound) < lo + npages) {
1205 			/* Skip to next boundary-matching page. */
1206 			lo = roundup2(lo + 1, ppn_bound);
1207 		}
1208 	}
1209 	return (-1);
1210 }
1211 
1212 /*
1213  * Searches the partially populated reservation queue for the least recently
1214  * changed reservation with free pages that satisfy the given request for
1215  * contiguous physical memory.  If a satisfactory reservation is found, it is
1216  * broken.  Returns a page if a reservation is broken and NULL otherwise.
1217  */
1218 vm_page_t
1219 vm_reserv_reclaim_contig(int domain, u_long npages, vm_paddr_t low,
1220     vm_paddr_t high, u_long alignment, vm_paddr_t boundary)
1221 {
1222 	struct vm_reserv_queue *queue;
1223 	vm_paddr_t pa, size;
1224 	vm_page_t m_ret;
1225 	vm_reserv_t marker, rv, rvn;
1226 	int hi, lo, posn, ppn_align, ppn_bound;
1227 
1228 	KASSERT(npages > 0, ("npages is 0"));
1229 	KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
1230 	KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
1231 	if (npages > VM_LEVEL_0_NPAGES - 1)
1232 		return (NULL);
1233 	size = npages << PAGE_SHIFT;
1234 	/*
1235 	 * Ensure that a free range starting at a boundary-multiple
1236 	 * doesn't include a boundary-multiple within it.  Otherwise,
1237 	 * no boundary-constrained allocation is possible.
1238 	 */
1239 	if (!vm_addr_bound_ok(0, size, boundary))
1240 		return (NULL);
1241 	marker = &vm_rvd[domain].marker;
1242 	queue = &vm_rvd[domain].partpop;
1243 	/*
1244 	 * Compute shifted alignment, boundary values for page-based
1245 	 * calculations.  Constrain to range [1, VM_LEVEL_0_NPAGES] to
1246 	 * avoid overflow.
1247 	 */
1248 	ppn_align = (int)(ulmin(ulmax(PAGE_SIZE, alignment),
1249 	    VM_LEVEL_0_SIZE) >> PAGE_SHIFT);
1250 	ppn_bound = boundary == 0 ? VM_LEVEL_0_NPAGES :
1251 	    (int)(MIN(MAX(PAGE_SIZE, boundary),
1252             VM_LEVEL_0_SIZE) >> PAGE_SHIFT);
1253 
1254 	vm_reserv_domain_scan_lock(domain);
1255 	vm_reserv_domain_lock(domain);
1256 	TAILQ_FOREACH_SAFE(rv, queue, partpopq, rvn) {
1257 		pa = VM_PAGE_TO_PHYS(&rv->pages[0]);
1258 		if (pa + VM_LEVEL_0_SIZE - size < low) {
1259 			/* This entire reservation is too low; go to next. */
1260 			continue;
1261 		}
1262 		if (pa + size > high) {
1263 			/* This entire reservation is too high; go to next. */
1264 			continue;
1265 		}
1266 		if (!vm_addr_align_ok(pa, alignment)) {
1267 			/* This entire reservation is unaligned; go to next. */
1268 			continue;
1269 		}
1270 
1271 		if (vm_reserv_trylock(rv) == 0) {
1272 			TAILQ_INSERT_AFTER(queue, rv, marker, partpopq);
1273 			vm_reserv_domain_unlock(domain);
1274 			vm_reserv_lock(rv);
1275 			if (TAILQ_PREV(marker, vm_reserv_queue, partpopq) !=
1276 			    rv) {
1277 				vm_reserv_unlock(rv);
1278 				vm_reserv_domain_lock(domain);
1279 				rvn = TAILQ_NEXT(marker, partpopq);
1280 				TAILQ_REMOVE(queue, marker, partpopq);
1281 				continue;
1282 			}
1283 			vm_reserv_domain_lock(domain);
1284 			TAILQ_REMOVE(queue, marker, partpopq);
1285 		}
1286 		vm_reserv_domain_unlock(domain);
1287 		lo = (pa >= low) ? 0 :
1288 		    (int)((low + PAGE_MASK - pa) >> PAGE_SHIFT);
1289 		hi = (pa + VM_LEVEL_0_SIZE <= high) ? VM_LEVEL_0_NPAGES :
1290 		    (int)((high - pa) >> PAGE_SHIFT);
1291 		posn = vm_reserv_find_contig(rv, (int)npages, lo, hi,
1292 		    ppn_align, ppn_bound);
1293 		if (posn >= 0) {
1294 			vm_reserv_domain_scan_unlock(domain);
1295 			/* Allocate requested space */
1296 			rv->popcnt += npages;
1297 			bit_nset(rv->popmap, posn, posn + npages - 1);
1298 			vm_reserv_reclaim(rv);
1299 			vm_reserv_unlock(rv);
1300 			m_ret = &rv->pages[posn];
1301 			pa = VM_PAGE_TO_PHYS(m_ret);
1302 			KASSERT(vm_addr_ok(pa, size, alignment, boundary),
1303 			    ("%s: adjusted address not aligned/bounded to "
1304 			     "%lx/%jx",
1305 			     __func__, alignment, (uintmax_t)boundary));
1306 			return (m_ret);
1307 		}
1308 		vm_reserv_domain_lock(domain);
1309 		rvn = TAILQ_NEXT(rv, partpopq);
1310 		vm_reserv_unlock(rv);
1311 	}
1312 	vm_reserv_domain_unlock(domain);
1313 	vm_reserv_domain_scan_unlock(domain);
1314 	return (NULL);
1315 }
1316 
1317 /*
1318  * Transfers the reservation underlying the given page to a new object.
1319  *
1320  * The object must be locked.
1321  */
1322 void
1323 vm_reserv_rename(vm_page_t m, vm_object_t new_object, vm_object_t old_object,
1324     vm_pindex_t old_object_offset)
1325 {
1326 	vm_reserv_t rv;
1327 
1328 	VM_OBJECT_ASSERT_WLOCKED(new_object);
1329 	rv = vm_reserv_from_page(m);
1330 	if (rv->object == old_object) {
1331 		vm_reserv_lock(rv);
1332 		CTR6(KTR_VM,
1333 		    "%s: rv %p object %p new %p popcnt %d inpartpop %d",
1334 		    __FUNCTION__, rv, rv->object, new_object, rv->popcnt,
1335 		    rv->inpartpopq);
1336 		if (rv->object == old_object) {
1337 			vm_reserv_object_lock(old_object);
1338 			rv->object = NULL;
1339 			LIST_REMOVE(rv, objq);
1340 			vm_reserv_object_unlock(old_object);
1341 			vm_reserv_object_lock(new_object);
1342 			rv->object = new_object;
1343 			rv->pindex -= old_object_offset;
1344 			LIST_INSERT_HEAD(&new_object->rvq, rv, objq);
1345 			vm_reserv_object_unlock(new_object);
1346 		}
1347 		vm_reserv_unlock(rv);
1348 	}
1349 }
1350 
1351 /*
1352  * Returns the size (in bytes) of a reservation of the specified level.
1353  */
1354 int
1355 vm_reserv_size(int level)
1356 {
1357 
1358 	switch (level) {
1359 	case 0:
1360 		return (VM_LEVEL_0_SIZE);
1361 	case -1:
1362 		return (PAGE_SIZE);
1363 	default:
1364 		return (0);
1365 	}
1366 }
1367 
1368 /*
1369  * Allocates the virtual and physical memory required by the reservation
1370  * management system's data structures, in particular, the reservation array.
1371  */
1372 vm_paddr_t
1373 vm_reserv_startup(vm_offset_t *vaddr, vm_paddr_t end)
1374 {
1375 	vm_paddr_t new_end;
1376 	vm_pindex_t count;
1377 	size_t size;
1378 	int i;
1379 
1380 	count = 0;
1381 	for (i = 0; i < vm_phys_nsegs; i++) {
1382 #ifdef VM_PHYSSEG_SPARSE
1383 		count += howmany(vm_phys_segs[i].end, VM_LEVEL_0_SIZE) -
1384 		    vm_phys_segs[i].start / VM_LEVEL_0_SIZE;
1385 #else
1386 		count = MAX(count,
1387 		    howmany(vm_phys_segs[i].end, VM_LEVEL_0_SIZE));
1388 #endif
1389 	}
1390 
1391 	for (i = 0; phys_avail[i + 1] != 0; i += 2) {
1392 #ifdef VM_PHYSSEG_SPARSE
1393 		count += howmany(phys_avail[i + 1], VM_LEVEL_0_SIZE) -
1394 		    phys_avail[i] / VM_LEVEL_0_SIZE;
1395 #else
1396 		count = MAX(count,
1397 		    howmany(phys_avail[i + 1], VM_LEVEL_0_SIZE));
1398 #endif
1399 	}
1400 
1401 	/*
1402 	 * Calculate the size (in bytes) of the reservation array.  Rounding up
1403 	 * for partial superpages at boundaries, as every small page is mapped
1404 	 * to an element in the reservation array based on its physical address.
1405 	 * Thus, the number of elements in the reservation array can be greater
1406 	 * than the number of superpages.
1407 	 */
1408 	size = count * sizeof(struct vm_reserv);
1409 
1410 	/*
1411 	 * Allocate and map the physical memory for the reservation array.  The
1412 	 * next available virtual address is returned by reference.
1413 	 */
1414 	new_end = end - round_page(size);
1415 	vm_reserv_array = (void *)(uintptr_t)pmap_map(vaddr, new_end, end,
1416 	    VM_PROT_READ | VM_PROT_WRITE);
1417 	bzero(vm_reserv_array, size);
1418 
1419 	/*
1420 	 * Return the next available physical address.
1421 	 */
1422 	return (new_end);
1423 }
1424 
1425 /*
1426  * Returns the superpage containing the given page.
1427  */
1428 vm_page_t
1429 vm_reserv_to_superpage(vm_page_t m)
1430 {
1431 	vm_reserv_t rv;
1432 
1433 	VM_OBJECT_ASSERT_LOCKED(m->object);
1434 	rv = vm_reserv_from_page(m);
1435 	if (rv->object == m->object && rv->popcnt == VM_LEVEL_0_NPAGES)
1436 		m = rv->pages;
1437 	else
1438 		m = NULL;
1439 
1440 	return (m);
1441 }
1442 
1443 #endif	/* VM_NRESERVLEVEL > 0 */
1444