xref: /freebsd/sys/vm/vm_reserv.c (revision 0bd5d367989b3d2de0e8d8ceaa2e31d3f0d96536)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
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/counter.h>
57 #include <sys/ktr.h>
58 #include <sys/vmmeter.h>
59 #include <sys/smp.h>
60 
61 #include <vm/vm.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_phys.h>
67 #include <vm/vm_pagequeue.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  * The size of a population map entry
110  */
111 typedef	u_long		popmap_t;
112 
113 /*
114  * The number of bits in a population map entry
115  */
116 #define	NBPOPMAP	(NBBY * sizeof(popmap_t))
117 
118 /*
119  * The number of population map entries in a reservation
120  */
121 #define	NPOPMAP		howmany(VM_LEVEL_0_NPAGES, NBPOPMAP)
122 #define	NPOPMAP_MAX	howmany(VM_LEVEL_0_NPAGES_MAX, NBPOPMAP)
123 
124 /*
125  * Number of elapsed ticks before we update the LRU queue position.  Used
126  * to reduce contention and churn on the list.
127  */
128 #define	PARTPOPSLOP	1
129 
130 /*
131  * Clear a bit in the population map.
132  */
133 static __inline void
134 popmap_clear(popmap_t popmap[], int i)
135 {
136 
137 	popmap[i / NBPOPMAP] &= ~(1UL << (i % NBPOPMAP));
138 }
139 
140 /*
141  * Set a bit in the population map.
142  */
143 static __inline void
144 popmap_set(popmap_t popmap[], int i)
145 {
146 
147 	popmap[i / NBPOPMAP] |= 1UL << (i % NBPOPMAP);
148 }
149 
150 /*
151  * Is a bit in the population map clear?
152  */
153 static __inline boolean_t
154 popmap_is_clear(popmap_t popmap[], int i)
155 {
156 
157 	return ((popmap[i / NBPOPMAP] & (1UL << (i % NBPOPMAP))) == 0);
158 }
159 
160 /*
161  * Is a bit in the population map set?
162  */
163 static __inline boolean_t
164 popmap_is_set(popmap_t popmap[], int i)
165 {
166 
167 	return ((popmap[i / NBPOPMAP] & (1UL << (i % NBPOPMAP))) != 0);
168 }
169 
170 /*
171  * The reservation structure
172  *
173  * A reservation structure is constructed whenever a large physical page is
174  * speculatively allocated to an object.  The reservation provides the small
175  * physical pages for the range [pindex, pindex + VM_LEVEL_0_NPAGES) of offsets
176  * within that object.  The reservation's "popcnt" tracks the number of these
177  * small physical pages that are in use at any given time.  When and if the
178  * reservation is not fully utilized, it appears in the queue of partially
179  * populated reservations.  The reservation always appears on the containing
180  * object's list of reservations.
181  *
182  * A partially populated reservation can be broken and reclaimed at any time.
183  *
184  * r - vm_reserv_lock
185  * d - vm_reserv_domain_lock
186  * o - vm_reserv_object_lock
187  * c - constant after boot
188  */
189 struct vm_reserv {
190 	struct mtx	lock;			/* reservation lock. */
191 	TAILQ_ENTRY(vm_reserv) partpopq;	/* (d) per-domain queue. */
192 	LIST_ENTRY(vm_reserv) objq;		/* (o, r) object queue */
193 	vm_object_t	object;			/* (o, r) containing object */
194 	vm_pindex_t	pindex;			/* (o, r) offset in object */
195 	vm_page_t	pages;			/* (c) first page  */
196 	uint16_t	domain;			/* (c) NUMA domain. */
197 	uint16_t	popcnt;			/* (r) # of pages in use */
198 	int		lasttick;		/* (r) last pop update tick. */
199 	char		inpartpopq;		/* (d) */
200 	popmap_t	popmap[NPOPMAP_MAX];	/* (r) bit vector, used pages */
201 };
202 
203 #define	vm_reserv_lockptr(rv)		(&(rv)->lock)
204 #define	vm_reserv_assert_locked(rv)					\
205 	    mtx_assert(vm_reserv_lockptr(rv), MA_OWNED)
206 #define	vm_reserv_lock(rv)		mtx_lock(vm_reserv_lockptr(rv))
207 #define	vm_reserv_trylock(rv)		mtx_trylock(vm_reserv_lockptr(rv))
208 #define	vm_reserv_unlock(rv)		mtx_unlock(vm_reserv_lockptr(rv))
209 
210 static struct mtx_padalign vm_reserv_domain_locks[MAXMEMDOM];
211 
212 #define	vm_reserv_domain_lockptr(d)	&vm_reserv_domain_locks[(d)]
213 #define	vm_reserv_domain_lock(d)	mtx_lock(vm_reserv_domain_lockptr(d))
214 #define	vm_reserv_domain_unlock(d)	mtx_unlock(vm_reserv_domain_lockptr(d))
215 
216 /*
217  * The reservation array
218  *
219  * This array is analoguous in function to vm_page_array.  It differs in the
220  * respect that it may contain a greater number of useful reservation
221  * structures than there are (physical) superpages.  These "invalid"
222  * reservation structures exist to trade-off space for time in the
223  * implementation of vm_reserv_from_page().  Invalid reservation structures are
224  * distinguishable from "valid" reservation structures by inspecting the
225  * reservation's "pages" field.  Invalid reservation structures have a NULL
226  * "pages" field.
227  *
228  * vm_reserv_from_page() maps a small (physical) page to an element of this
229  * array by computing a physical reservation number from the page's physical
230  * address.  The physical reservation number is used as the array index.
231  *
232  * An "active" reservation is a valid reservation structure that has a non-NULL
233  * "object" field and a non-zero "popcnt" field.  In other words, every active
234  * reservation belongs to a particular object.  Moreover, every active
235  * reservation has an entry in the containing object's list of reservations.
236  */
237 static vm_reserv_t vm_reserv_array;
238 
239 /*
240  * The partially populated reservation queue
241  *
242  * This queue enables the fast recovery of an unused free small page from a
243  * partially populated reservation.  The reservation at the head of this queue
244  * is the least recently changed, partially populated reservation.
245  *
246  * Access to this queue is synchronized by the free page queue lock.
247  */
248 static TAILQ_HEAD(, vm_reserv) vm_rvq_partpop[MAXMEMDOM];
249 
250 static SYSCTL_NODE(_vm, OID_AUTO, reserv, CTLFLAG_RD, 0, "Reservation Info");
251 
252 static counter_u64_t vm_reserv_broken = EARLY_COUNTER;
253 SYSCTL_COUNTER_U64(_vm_reserv, OID_AUTO, broken, CTLFLAG_RD,
254     &vm_reserv_broken, "Cumulative number of broken reservations");
255 
256 static counter_u64_t vm_reserv_freed = EARLY_COUNTER;
257 SYSCTL_COUNTER_U64(_vm_reserv, OID_AUTO, freed, CTLFLAG_RD,
258     &vm_reserv_freed, "Cumulative number of freed reservations");
259 
260 static int sysctl_vm_reserv_fullpop(SYSCTL_HANDLER_ARGS);
261 
262 SYSCTL_PROC(_vm_reserv, OID_AUTO, fullpop, CTLTYPE_INT | CTLFLAG_RD, NULL, 0,
263     sysctl_vm_reserv_fullpop, "I", "Current number of full reservations");
264 
265 static int sysctl_vm_reserv_partpopq(SYSCTL_HANDLER_ARGS);
266 
267 SYSCTL_OID(_vm_reserv, OID_AUTO, partpopq, CTLTYPE_STRING | CTLFLAG_RD, NULL, 0,
268     sysctl_vm_reserv_partpopq, "A", "Partially populated reservation queues");
269 
270 static counter_u64_t vm_reserv_reclaimed = EARLY_COUNTER;
271 SYSCTL_COUNTER_U64(_vm_reserv, OID_AUTO, reclaimed, CTLFLAG_RD,
272     &vm_reserv_reclaimed, "Cumulative number of reclaimed reservations");
273 
274 /*
275  * The object lock pool is used to synchronize the rvq.  We can not use a
276  * pool mutex because it is required before malloc works.
277  *
278  * The "hash" function could be made faster without divide and modulo.
279  */
280 #define	VM_RESERV_OBJ_LOCK_COUNT	MAXCPU
281 
282 struct mtx_padalign vm_reserv_object_mtx[VM_RESERV_OBJ_LOCK_COUNT];
283 
284 #define	vm_reserv_object_lock_idx(object)			\
285 	    (((uintptr_t)object / sizeof(*object)) % VM_RESERV_OBJ_LOCK_COUNT)
286 #define	vm_reserv_object_lock_ptr(object)			\
287 	    &vm_reserv_object_mtx[vm_reserv_object_lock_idx((object))]
288 #define	vm_reserv_object_lock(object)				\
289 	    mtx_lock(vm_reserv_object_lock_ptr((object)))
290 #define	vm_reserv_object_unlock(object)				\
291 	    mtx_unlock(vm_reserv_object_lock_ptr((object)))
292 
293 static void		vm_reserv_break(vm_reserv_t rv);
294 static void		vm_reserv_depopulate(vm_reserv_t rv, int index);
295 static vm_reserv_t	vm_reserv_from_page(vm_page_t m);
296 static boolean_t	vm_reserv_has_pindex(vm_reserv_t rv,
297 			    vm_pindex_t pindex);
298 static void		vm_reserv_populate(vm_reserv_t rv, int index);
299 static void		vm_reserv_reclaim(vm_reserv_t rv);
300 
301 /*
302  * Returns the current number of full reservations.
303  *
304  * Since the number of full reservations is computed without acquiring the
305  * free page queue lock, the returned value may be inexact.
306  */
307 static int
308 sysctl_vm_reserv_fullpop(SYSCTL_HANDLER_ARGS)
309 {
310 	vm_paddr_t paddr;
311 	struct vm_phys_seg *seg;
312 	vm_reserv_t rv;
313 	int fullpop, segind;
314 
315 	fullpop = 0;
316 	for (segind = 0; segind < vm_phys_nsegs; segind++) {
317 		seg = &vm_phys_segs[segind];
318 		paddr = roundup2(seg->start, VM_LEVEL_0_SIZE);
319 		while (paddr + VM_LEVEL_0_SIZE > paddr && paddr +
320 		    VM_LEVEL_0_SIZE <= seg->end) {
321 			rv = &vm_reserv_array[paddr >> VM_LEVEL_0_SHIFT];
322 			fullpop += rv->popcnt == VM_LEVEL_0_NPAGES;
323 			paddr += VM_LEVEL_0_SIZE;
324 		}
325 	}
326 	return (sysctl_handle_int(oidp, &fullpop, 0, req));
327 }
328 
329 /*
330  * Describes the current state of the partially populated reservation queue.
331  */
332 static int
333 sysctl_vm_reserv_partpopq(SYSCTL_HANDLER_ARGS)
334 {
335 	struct sbuf sbuf;
336 	vm_reserv_t rv;
337 	int counter, error, domain, level, unused_pages;
338 
339 	error = sysctl_wire_old_buffer(req, 0);
340 	if (error != 0)
341 		return (error);
342 	sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
343 	sbuf_printf(&sbuf, "\nDOMAIN    LEVEL     SIZE  NUMBER\n\n");
344 	for (domain = 0; domain < vm_ndomains; domain++) {
345 		for (level = -1; level <= VM_NRESERVLEVEL - 2; level++) {
346 			counter = 0;
347 			unused_pages = 0;
348 			vm_reserv_domain_lock(domain);
349 			TAILQ_FOREACH(rv, &vm_rvq_partpop[domain], partpopq) {
350 				counter++;
351 				unused_pages += VM_LEVEL_0_NPAGES - rv->popcnt;
352 			}
353 			vm_reserv_domain_unlock(domain);
354 			sbuf_printf(&sbuf, "%6d, %7d, %6dK, %6d\n",
355 			    domain, level,
356 			    unused_pages * ((int)PAGE_SIZE / 1024), counter);
357 		}
358 	}
359 	error = sbuf_finish(&sbuf);
360 	sbuf_delete(&sbuf);
361 	return (error);
362 }
363 
364 /*
365  * Remove a reservation from the object's objq.
366  */
367 static void
368 vm_reserv_remove(vm_reserv_t rv)
369 {
370 	vm_object_t object;
371 
372 	vm_reserv_assert_locked(rv);
373 	CTR5(KTR_VM, "%s: rv %p object %p popcnt %d inpartpop %d",
374 	    __FUNCTION__, rv, rv->object, rv->popcnt, rv->inpartpopq);
375 	KASSERT(rv->object != NULL,
376 	    ("vm_reserv_remove: reserv %p is free", rv));
377 	KASSERT(!rv->inpartpopq,
378 	    ("vm_reserv_remove: reserv %p's inpartpopq is TRUE", rv));
379 	object = rv->object;
380 	vm_reserv_object_lock(object);
381 	LIST_REMOVE(rv, objq);
382 	rv->object = NULL;
383 	vm_reserv_object_unlock(object);
384 }
385 
386 /*
387  * Insert a new reservation into the object's objq.
388  */
389 static void
390 vm_reserv_insert(vm_reserv_t rv, vm_object_t object, vm_pindex_t pindex)
391 {
392 	int i;
393 
394 	vm_reserv_assert_locked(rv);
395 	CTR6(KTR_VM,
396 	    "%s: rv %p(%p) object %p new %p popcnt %d",
397 	    __FUNCTION__, rv, rv->pages, rv->object, object,
398 	   rv->popcnt);
399 	KASSERT(rv->object == NULL,
400 	    ("vm_reserv_insert: reserv %p isn't free", rv));
401 	KASSERT(rv->popcnt == 0,
402 	    ("vm_reserv_insert: reserv %p's popcnt is corrupted", rv));
403 	KASSERT(!rv->inpartpopq,
404 	    ("vm_reserv_insert: reserv %p's inpartpopq is TRUE", rv));
405 	for (i = 0; i < NPOPMAP; i++)
406 		KASSERT(rv->popmap[i] == 0,
407 		    ("vm_reserv_insert: reserv %p's popmap is corrupted", rv));
408 	vm_reserv_object_lock(object);
409 	rv->pindex = pindex;
410 	rv->object = object;
411 	rv->lasttick = ticks;
412 	LIST_INSERT_HEAD(&object->rvq, rv, objq);
413 	vm_reserv_object_unlock(object);
414 }
415 
416 /*
417  * Reduces the given reservation's population count.  If the population count
418  * becomes zero, the reservation is destroyed.  Additionally, moves the
419  * reservation to the tail of the partially populated reservation queue if the
420  * population count is non-zero.
421  */
422 static void
423 vm_reserv_depopulate(vm_reserv_t rv, int index)
424 {
425 	struct vm_domain *vmd;
426 
427 	vm_reserv_assert_locked(rv);
428 	CTR5(KTR_VM, "%s: rv %p object %p popcnt %d inpartpop %d",
429 	    __FUNCTION__, rv, rv->object, rv->popcnt, rv->inpartpopq);
430 	KASSERT(rv->object != NULL,
431 	    ("vm_reserv_depopulate: reserv %p is free", rv));
432 	KASSERT(popmap_is_set(rv->popmap, index),
433 	    ("vm_reserv_depopulate: reserv %p's popmap[%d] is clear", rv,
434 	    index));
435 	KASSERT(rv->popcnt > 0,
436 	    ("vm_reserv_depopulate: reserv %p's popcnt is corrupted", rv));
437 	KASSERT(rv->domain < vm_ndomains,
438 	    ("vm_reserv_depopulate: reserv %p's domain is corrupted %d",
439 	    rv, rv->domain));
440 	if (rv->popcnt == VM_LEVEL_0_NPAGES) {
441 		KASSERT(rv->pages->psind == 1,
442 		    ("vm_reserv_depopulate: reserv %p is already demoted",
443 		    rv));
444 		rv->pages->psind = 0;
445 	}
446 	popmap_clear(rv->popmap, index);
447 	rv->popcnt--;
448 	if ((unsigned)(ticks - rv->lasttick) >= PARTPOPSLOP ||
449 	    rv->popcnt == 0) {
450 		vm_reserv_domain_lock(rv->domain);
451 		if (rv->inpartpopq) {
452 			TAILQ_REMOVE(&vm_rvq_partpop[rv->domain], rv, partpopq);
453 			rv->inpartpopq = FALSE;
454 		}
455 		if (rv->popcnt != 0) {
456 			rv->inpartpopq = TRUE;
457 			TAILQ_INSERT_TAIL(&vm_rvq_partpop[rv->domain], rv, partpopq);
458 		}
459 		vm_reserv_domain_unlock(rv->domain);
460 		rv->lasttick = ticks;
461 	}
462 	vmd = VM_DOMAIN(rv->domain);
463 	if (rv->popcnt == 0) {
464 		vm_reserv_remove(rv);
465 		vm_domain_free_lock(vmd);
466 		vm_phys_free_pages(rv->pages, VM_LEVEL_0_ORDER);
467 		vm_domain_free_unlock(vmd);
468 		counter_u64_add(vm_reserv_freed, 1);
469 	}
470 	vm_domain_freecnt_inc(vmd, 1);
471 }
472 
473 /*
474  * Returns the reservation to which the given page might belong.
475  */
476 static __inline vm_reserv_t
477 vm_reserv_from_page(vm_page_t m)
478 {
479 
480 	return (&vm_reserv_array[VM_PAGE_TO_PHYS(m) >> VM_LEVEL_0_SHIFT]);
481 }
482 
483 /*
484  * Returns an existing reservation or NULL and initialized successor pointer.
485  */
486 static vm_reserv_t
487 vm_reserv_from_object(vm_object_t object, vm_pindex_t pindex,
488     vm_page_t mpred, vm_page_t *msuccp)
489 {
490 	vm_reserv_t rv;
491 	vm_page_t msucc;
492 
493 	msucc = NULL;
494 	if (mpred != NULL) {
495 		KASSERT(mpred->object == object,
496 		    ("vm_reserv_from_object: object doesn't contain mpred"));
497 		KASSERT(mpred->pindex < pindex,
498 		    ("vm_reserv_from_object: mpred doesn't precede pindex"));
499 		rv = vm_reserv_from_page(mpred);
500 		if (rv->object == object && vm_reserv_has_pindex(rv, pindex))
501 			goto found;
502 		msucc = TAILQ_NEXT(mpred, listq);
503 	} else
504 		msucc = TAILQ_FIRST(&object->memq);
505 	if (msucc != NULL) {
506 		KASSERT(msucc->pindex > pindex,
507 		    ("vm_reserv_from_object: msucc doesn't succeed pindex"));
508 		rv = vm_reserv_from_page(msucc);
509 		if (rv->object == object && vm_reserv_has_pindex(rv, pindex))
510 			goto found;
511 	}
512 	rv = NULL;
513 
514 found:
515 	*msuccp = msucc;
516 
517 	return (rv);
518 }
519 
520 /*
521  * Returns TRUE if the given reservation contains the given page index and
522  * FALSE otherwise.
523  */
524 static __inline boolean_t
525 vm_reserv_has_pindex(vm_reserv_t rv, vm_pindex_t pindex)
526 {
527 
528 	return (((pindex - rv->pindex) & ~(VM_LEVEL_0_NPAGES - 1)) == 0);
529 }
530 
531 /*
532  * Increases the given reservation's population count.  Moves the reservation
533  * to the tail of the partially populated reservation queue.
534  *
535  * The free page queue must be locked.
536  */
537 static void
538 vm_reserv_populate(vm_reserv_t rv, int index)
539 {
540 
541 	vm_reserv_assert_locked(rv);
542 	CTR5(KTR_VM, "%s: rv %p object %p popcnt %d inpartpop %d",
543 	    __FUNCTION__, rv, rv->object, rv->popcnt, rv->inpartpopq);
544 	KASSERT(rv->object != NULL,
545 	    ("vm_reserv_populate: reserv %p is free", rv));
546 	KASSERT(popmap_is_clear(rv->popmap, index),
547 	    ("vm_reserv_populate: reserv %p's popmap[%d] is set", rv,
548 	    index));
549 	KASSERT(rv->popcnt < VM_LEVEL_0_NPAGES,
550 	    ("vm_reserv_populate: reserv %p is already full", rv));
551 	KASSERT(rv->pages->psind == 0,
552 	    ("vm_reserv_populate: reserv %p is already promoted", rv));
553 	KASSERT(rv->domain < vm_ndomains,
554 	    ("vm_reserv_populate: reserv %p's domain is corrupted %d",
555 	    rv, rv->domain));
556 	popmap_set(rv->popmap, index);
557 	rv->popcnt++;
558 	if ((unsigned)(ticks - rv->lasttick) < PARTPOPSLOP &&
559 	    rv->inpartpopq && rv->popcnt != VM_LEVEL_0_NPAGES)
560 		return;
561 	rv->lasttick = ticks;
562 	vm_reserv_domain_lock(rv->domain);
563 	if (rv->inpartpopq) {
564 		TAILQ_REMOVE(&vm_rvq_partpop[rv->domain], rv, partpopq);
565 		rv->inpartpopq = FALSE;
566 	}
567 	if (rv->popcnt < VM_LEVEL_0_NPAGES) {
568 		rv->inpartpopq = TRUE;
569 		TAILQ_INSERT_TAIL(&vm_rvq_partpop[rv->domain], rv, partpopq);
570 	} else {
571 		KASSERT(rv->pages->psind == 0,
572 		    ("vm_reserv_populate: reserv %p is already promoted",
573 		    rv));
574 		rv->pages->psind = 1;
575 	}
576 	vm_reserv_domain_unlock(rv->domain);
577 }
578 
579 /*
580  * Allocates a contiguous set of physical pages of the given size "npages"
581  * from existing or newly created reservations.  All of the physical pages
582  * must be at or above the given physical address "low" and below the given
583  * physical address "high".  The given value "alignment" determines the
584  * alignment of the first physical page in the set.  If the given value
585  * "boundary" is non-zero, then the set of physical pages cannot cross any
586  * physical address boundary that is a multiple of that value.  Both
587  * "alignment" and "boundary" must be a power of two.
588  *
589  * The page "mpred" must immediately precede the offset "pindex" within the
590  * specified object.
591  *
592  * The object must be locked.
593  */
594 vm_page_t
595 vm_reserv_alloc_contig(vm_object_t object, vm_pindex_t pindex, int domain,
596     int req, vm_page_t mpred, u_long npages, vm_paddr_t low, vm_paddr_t high,
597     u_long alignment, vm_paddr_t boundary)
598 {
599 	struct vm_domain *vmd;
600 	vm_paddr_t pa, size;
601 	vm_page_t m, m_ret, msucc;
602 	vm_pindex_t first, leftcap, rightcap;
603 	vm_reserv_t rv;
604 	u_long allocpages, maxpages, minpages;
605 	int i, index, n;
606 
607 	VM_OBJECT_ASSERT_WLOCKED(object);
608 	KASSERT(npages != 0, ("vm_reserv_alloc_contig: npages is 0"));
609 
610 	/*
611 	 * Is a reservation fundamentally impossible?
612 	 */
613 	if (pindex < VM_RESERV_INDEX(object, pindex) ||
614 	    pindex + npages > object->size)
615 		return (NULL);
616 
617 	/*
618 	 * All reservations of a particular size have the same alignment.
619 	 * Assuming that the first page is allocated from a reservation, the
620 	 * least significant bits of its physical address can be determined
621 	 * from its offset from the beginning of the reservation and the size
622 	 * of the reservation.
623 	 *
624 	 * Could the specified index within a reservation of the smallest
625 	 * possible size satisfy the alignment and boundary requirements?
626 	 */
627 	pa = VM_RESERV_INDEX(object, pindex) << PAGE_SHIFT;
628 	if ((pa & (alignment - 1)) != 0)
629 		return (NULL);
630 	size = npages << PAGE_SHIFT;
631 	if (((pa ^ (pa + size - 1)) & ~(boundary - 1)) != 0)
632 		return (NULL);
633 
634 	/*
635 	 * Look for an existing reservation.
636 	 */
637 	rv = vm_reserv_from_object(object, pindex, mpred, &msucc);
638 	if (rv != NULL) {
639 		KASSERT(object != kernel_object || rv->domain == domain,
640 		    ("vm_reserv_alloc_contig: domain mismatch"));
641 		index = VM_RESERV_INDEX(object, pindex);
642 		/* Does the allocation fit within the reservation? */
643 		if (index + npages > VM_LEVEL_0_NPAGES)
644 			return (NULL);
645 		domain = rv->domain;
646 		vmd = VM_DOMAIN(domain);
647 		vm_reserv_lock(rv);
648 		/* Handle reclaim race. */
649 		if (rv->object != object)
650 			goto out;
651 		m = &rv->pages[index];
652 		pa = VM_PAGE_TO_PHYS(m);
653 		if (pa < low || pa + size > high ||
654 		    (pa & (alignment - 1)) != 0 ||
655 		    ((pa ^ (pa + size - 1)) & ~(boundary - 1)) != 0)
656 			goto out;
657 		/* Handle vm_page_rename(m, new_object, ...). */
658 		for (i = 0; i < npages; i++)
659 			if (popmap_is_set(rv->popmap, index + i))
660 				goto out;
661 		if (!vm_domain_allocate(vmd, req, npages))
662 			goto out;
663 		for (i = 0; i < npages; i++)
664 			vm_reserv_populate(rv, index + i);
665 		vm_reserv_unlock(rv);
666 		return (m);
667 out:
668 		vm_reserv_unlock(rv);
669 		return (NULL);
670 	}
671 
672 	/*
673 	 * Could at least one reservation fit between the first index to the
674 	 * left that can be used ("leftcap") and the first index to the right
675 	 * that cannot be used ("rightcap")?
676 	 *
677 	 * We must synchronize with the reserv object lock to protect the
678 	 * pindex/object of the resulting reservations against rename while
679 	 * we are inspecting.
680 	 */
681 	first = pindex - VM_RESERV_INDEX(object, pindex);
682 	minpages = VM_RESERV_INDEX(object, pindex) + npages;
683 	maxpages = roundup2(minpages, VM_LEVEL_0_NPAGES);
684 	allocpages = maxpages;
685 	vm_reserv_object_lock(object);
686 	if (mpred != NULL) {
687 		if ((rv = vm_reserv_from_page(mpred))->object != object)
688 			leftcap = mpred->pindex + 1;
689 		else
690 			leftcap = rv->pindex + VM_LEVEL_0_NPAGES;
691 		if (leftcap > first) {
692 			vm_reserv_object_unlock(object);
693 			return (NULL);
694 		}
695 	}
696 	if (msucc != NULL) {
697 		if ((rv = vm_reserv_from_page(msucc))->object != object)
698 			rightcap = msucc->pindex;
699 		else
700 			rightcap = rv->pindex;
701 		if (first + maxpages > rightcap) {
702 			if (maxpages == VM_LEVEL_0_NPAGES) {
703 				vm_reserv_object_unlock(object);
704 				return (NULL);
705 			}
706 
707 			/*
708 			 * At least one reservation will fit between "leftcap"
709 			 * and "rightcap".  However, a reservation for the
710 			 * last of the requested pages will not fit.  Reduce
711 			 * the size of the upcoming allocation accordingly.
712 			 */
713 			allocpages = minpages;
714 		}
715 	}
716 	vm_reserv_object_unlock(object);
717 
718 	/*
719 	 * Would the last new reservation extend past the end of the object?
720 	 */
721 	if (first + maxpages > object->size) {
722 		/*
723 		 * Don't allocate the last new reservation if the object is a
724 		 * vnode or backed by another object that is a vnode.
725 		 */
726 		if (object->type == OBJT_VNODE ||
727 		    (object->backing_object != NULL &&
728 		    object->backing_object->type == OBJT_VNODE)) {
729 			if (maxpages == VM_LEVEL_0_NPAGES)
730 				return (NULL);
731 			allocpages = minpages;
732 		}
733 		/* Speculate that the object may grow. */
734 	}
735 
736 	/*
737 	 * Allocate the physical pages.  The alignment and boundary specified
738 	 * for this allocation may be different from the alignment and
739 	 * boundary specified for the requested pages.  For instance, the
740 	 * specified index may not be the first page within the first new
741 	 * reservation.
742 	 */
743 	m = NULL;
744 	vmd = VM_DOMAIN(domain);
745 	if (vm_domain_allocate(vmd, req, npages)) {
746 		vm_domain_free_lock(vmd);
747 		m = vm_phys_alloc_contig(domain, allocpages, low, high,
748 		    ulmax(alignment, VM_LEVEL_0_SIZE),
749 		    boundary > VM_LEVEL_0_SIZE ? boundary : 0);
750 		vm_domain_free_unlock(vmd);
751 		if (m == NULL) {
752 			vm_domain_freecnt_inc(vmd, npages);
753 			return (NULL);
754 		}
755 	} else
756 		return (NULL);
757 	KASSERT(vm_phys_domain(m) == domain,
758 	    ("vm_reserv_alloc_contig: Page domain does not match requested."));
759 
760 	/*
761 	 * The allocated physical pages always begin at a reservation
762 	 * boundary, but they do not always end at a reservation boundary.
763 	 * Initialize every reservation that is completely covered by the
764 	 * allocated physical pages.
765 	 */
766 	m_ret = NULL;
767 	index = VM_RESERV_INDEX(object, pindex);
768 	do {
769 		rv = vm_reserv_from_page(m);
770 		KASSERT(rv->pages == m,
771 		    ("vm_reserv_alloc_contig: reserv %p's pages is corrupted",
772 		    rv));
773 		vm_reserv_lock(rv);
774 		vm_reserv_insert(rv, object, first);
775 		n = ulmin(VM_LEVEL_0_NPAGES - index, npages);
776 		for (i = 0; i < n; i++)
777 			vm_reserv_populate(rv, index + i);
778 		npages -= n;
779 		if (m_ret == NULL) {
780 			m_ret = &rv->pages[index];
781 			index = 0;
782 		}
783 		vm_reserv_unlock(rv);
784 		m += VM_LEVEL_0_NPAGES;
785 		first += VM_LEVEL_0_NPAGES;
786 		allocpages -= VM_LEVEL_0_NPAGES;
787 	} while (allocpages >= VM_LEVEL_0_NPAGES);
788 	return (m_ret);
789 }
790 
791 /*
792  * Allocate a physical page from an existing or newly created reservation.
793  *
794  * The page "mpred" must immediately precede the offset "pindex" within the
795  * specified object.
796  *
797  * The object must be locked.
798  */
799 vm_page_t
800 vm_reserv_alloc_page(vm_object_t object, vm_pindex_t pindex, int domain,
801     int req, vm_page_t mpred)
802 {
803 	struct vm_domain *vmd;
804 	vm_page_t m, msucc;
805 	vm_pindex_t first, leftcap, rightcap;
806 	vm_reserv_t rv;
807 	int index;
808 
809 	VM_OBJECT_ASSERT_WLOCKED(object);
810 
811 	/*
812 	 * Is a reservation fundamentally impossible?
813 	 */
814 	if (pindex < VM_RESERV_INDEX(object, pindex) ||
815 	    pindex >= object->size)
816 		return (NULL);
817 
818 	/*
819 	 * Look for an existing reservation.
820 	 */
821 	rv = vm_reserv_from_object(object, pindex, mpred, &msucc);
822 	if (rv != NULL) {
823 		KASSERT(object != kernel_object || rv->domain == domain,
824 		    ("vm_reserv_alloc_page: domain mismatch"));
825 		domain = rv->domain;
826 		vmd = VM_DOMAIN(domain);
827 		index = VM_RESERV_INDEX(object, pindex);
828 		m = &rv->pages[index];
829 		vm_reserv_lock(rv);
830 		/* Handle reclaim race. */
831 		if (rv->object != object ||
832 		    /* Handle vm_page_rename(m, new_object, ...). */
833 		    popmap_is_set(rv->popmap, index)) {
834 			m = NULL;
835 			goto out;
836 		}
837 		if (vm_domain_allocate(vmd, req, 1) == 0)
838 			m = NULL;
839 		else
840 			vm_reserv_populate(rv, index);
841 out:
842 		vm_reserv_unlock(rv);
843 		return (m);
844 	}
845 
846 	/*
847 	 * Could a reservation fit between the first index to the left that
848 	 * can be used and the first index to the right that cannot be used?
849 	 *
850 	 * We must synchronize with the reserv object lock to protect the
851 	 * pindex/object of the resulting reservations against rename while
852 	 * we are inspecting.
853 	 */
854 	first = pindex - VM_RESERV_INDEX(object, pindex);
855 	vm_reserv_object_lock(object);
856 	if (mpred != NULL) {
857 		if ((rv = vm_reserv_from_page(mpred))->object != object)
858 			leftcap = mpred->pindex + 1;
859 		else
860 			leftcap = rv->pindex + VM_LEVEL_0_NPAGES;
861 		if (leftcap > first) {
862 			vm_reserv_object_unlock(object);
863 			return (NULL);
864 		}
865 	}
866 	if (msucc != NULL) {
867 		if ((rv = vm_reserv_from_page(msucc))->object != object)
868 			rightcap = msucc->pindex;
869 		else
870 			rightcap = rv->pindex;
871 		if (first + VM_LEVEL_0_NPAGES > rightcap) {
872 			vm_reserv_object_unlock(object);
873 			return (NULL);
874 		}
875 	}
876 	vm_reserv_object_unlock(object);
877 
878 	/*
879 	 * Would a new reservation extend past the end of the object?
880 	 */
881 	if (first + VM_LEVEL_0_NPAGES > object->size) {
882 		/*
883 		 * Don't allocate a new reservation if the object is a vnode or
884 		 * backed by another object that is a vnode.
885 		 */
886 		if (object->type == OBJT_VNODE ||
887 		    (object->backing_object != NULL &&
888 		    object->backing_object->type == OBJT_VNODE))
889 			return (NULL);
890 		/* Speculate that the object may grow. */
891 	}
892 
893 	/*
894 	 * Allocate and populate the new reservation.
895 	 */
896 	m = NULL;
897 	vmd = VM_DOMAIN(domain);
898 	if (vm_domain_allocate(vmd, req, 1)) {
899 		vm_domain_free_lock(vmd);
900 		m = vm_phys_alloc_pages(domain, VM_FREEPOOL_DEFAULT,
901 		    VM_LEVEL_0_ORDER);
902 		vm_domain_free_unlock(vmd);
903 		if (m == NULL) {
904 			vm_domain_freecnt_inc(vmd, 1);
905 			return (NULL);
906 		}
907 	} else
908 		return (NULL);
909 	rv = vm_reserv_from_page(m);
910 	vm_reserv_lock(rv);
911 	KASSERT(rv->pages == m,
912 	    ("vm_reserv_alloc_page: reserv %p's pages is corrupted", rv));
913 	vm_reserv_insert(rv, object, first);
914 	index = VM_RESERV_INDEX(object, pindex);
915 	vm_reserv_populate(rv, index);
916 	vm_reserv_unlock(rv);
917 
918 	return (&rv->pages[index]);
919 }
920 
921 /*
922  * Breaks the given reservation.  All free pages in the reservation
923  * are returned to the physical memory allocator.  The reservation's
924  * population count and map are reset to their initial state.
925  *
926  * The given reservation must not be in the partially populated reservation
927  * queue.  The free page queue lock must be held.
928  */
929 static void
930 vm_reserv_break(vm_reserv_t rv)
931 {
932 	u_long changes;
933 	int bitpos, hi, i, lo;
934 
935 	vm_reserv_assert_locked(rv);
936 	CTR5(KTR_VM, "%s: rv %p object %p popcnt %d inpartpop %d",
937 	    __FUNCTION__, rv, rv->object, rv->popcnt, rv->inpartpopq);
938 	vm_reserv_remove(rv);
939 	rv->pages->psind = 0;
940 	hi = lo = -1;
941 	for (i = 0; i <= NPOPMAP; i++) {
942 		/*
943 		 * "changes" is a bitmask that marks where a new sequence of
944 		 * 0s or 1s begins in popmap[i], with last bit in popmap[i-1]
945 		 * considered to be 1 if and only if lo == hi.  The bits of
946 		 * popmap[-1] and popmap[NPOPMAP] are considered all 1s.
947 		 */
948 		if (i == NPOPMAP)
949 			changes = lo != hi;
950 		else {
951 			changes = rv->popmap[i];
952 			changes ^= (changes << 1) | (lo == hi);
953 			rv->popmap[i] = 0;
954 		}
955 		while (changes != 0) {
956 			/*
957 			 * If the next change marked begins a run of 0s, set
958 			 * lo to mark that position.  Otherwise set hi and
959 			 * free pages from lo up to hi.
960 			 */
961 			bitpos = ffsl(changes) - 1;
962 			changes ^= 1UL << bitpos;
963 			if (lo == hi)
964 				lo = NBPOPMAP * i + bitpos;
965 			else {
966 				hi = NBPOPMAP * i + bitpos;
967 				vm_domain_free_lock(VM_DOMAIN(rv->domain));
968 				vm_phys_enqueue_contig(&rv->pages[lo], hi - lo);
969 				vm_domain_free_unlock(VM_DOMAIN(rv->domain));
970 				lo = hi;
971 			}
972 		}
973 	}
974 	rv->popcnt = 0;
975 	counter_u64_add(vm_reserv_broken, 1);
976 }
977 
978 /*
979  * Breaks all reservations belonging to the given object.
980  */
981 void
982 vm_reserv_break_all(vm_object_t object)
983 {
984 	vm_reserv_t rv;
985 
986 	/*
987 	 * This access of object->rvq is unsynchronized so that the
988 	 * object rvq lock can nest after the domain_free lock.  We
989 	 * must check for races in the results.  However, the object
990 	 * lock prevents new additions, so we are guaranteed that when
991 	 * it returns NULL the object is properly empty.
992 	 */
993 	while ((rv = LIST_FIRST(&object->rvq)) != NULL) {
994 		vm_reserv_lock(rv);
995 		/* Reclaim race. */
996 		if (rv->object != object) {
997 			vm_reserv_unlock(rv);
998 			continue;
999 		}
1000 		vm_reserv_domain_lock(rv->domain);
1001 		if (rv->inpartpopq) {
1002 			TAILQ_REMOVE(&vm_rvq_partpop[rv->domain], rv, partpopq);
1003 			rv->inpartpopq = FALSE;
1004 		}
1005 		vm_reserv_domain_unlock(rv->domain);
1006 		vm_reserv_break(rv);
1007 		vm_reserv_unlock(rv);
1008 	}
1009 }
1010 
1011 /*
1012  * Frees the given page if it belongs to a reservation.  Returns TRUE if the
1013  * page is freed and FALSE otherwise.
1014  *
1015  * The free page queue lock must be held.
1016  */
1017 boolean_t
1018 vm_reserv_free_page(vm_page_t m)
1019 {
1020 	vm_reserv_t rv;
1021 	boolean_t ret;
1022 
1023 	rv = vm_reserv_from_page(m);
1024 	if (rv->object == NULL)
1025 		return (FALSE);
1026 	vm_reserv_lock(rv);
1027 	/* Re-validate after lock. */
1028 	if (rv->object != NULL) {
1029 		vm_reserv_depopulate(rv, m - rv->pages);
1030 		ret = TRUE;
1031 	} else
1032 		ret = FALSE;
1033 	vm_reserv_unlock(rv);
1034 
1035 	return (ret);
1036 }
1037 
1038 /*
1039  * Initializes the reservation management system.  Specifically, initializes
1040  * the reservation array.
1041  *
1042  * Requires that vm_page_array and first_page are initialized!
1043  */
1044 void
1045 vm_reserv_init(void)
1046 {
1047 	vm_paddr_t paddr;
1048 	struct vm_phys_seg *seg;
1049 	struct vm_reserv *rv;
1050 	int i, segind;
1051 
1052 	/*
1053 	 * Initialize the reservation array.  Specifically, initialize the
1054 	 * "pages" field for every element that has an underlying superpage.
1055 	 */
1056 	for (segind = 0; segind < vm_phys_nsegs; segind++) {
1057 		seg = &vm_phys_segs[segind];
1058 		paddr = roundup2(seg->start, VM_LEVEL_0_SIZE);
1059 		while (paddr + VM_LEVEL_0_SIZE > paddr && paddr +
1060 		    VM_LEVEL_0_SIZE <= seg->end) {
1061 			rv = &vm_reserv_array[paddr >> VM_LEVEL_0_SHIFT];
1062 			rv->pages = PHYS_TO_VM_PAGE(paddr);
1063 			rv->domain = seg->domain;
1064 			mtx_init(&rv->lock, "vm reserv", NULL, MTX_DEF);
1065 			paddr += VM_LEVEL_0_SIZE;
1066 		}
1067 	}
1068 	for (i = 0; i < MAXMEMDOM; i++) {
1069 		mtx_init(&vm_reserv_domain_locks[i], "VM reserv domain", NULL,
1070 		    MTX_DEF);
1071 		TAILQ_INIT(&vm_rvq_partpop[i]);
1072 	}
1073 
1074 	for (i = 0; i < VM_RESERV_OBJ_LOCK_COUNT; i++)
1075 		mtx_init(&vm_reserv_object_mtx[i], "resv obj lock", NULL,
1076 		    MTX_DEF);
1077 }
1078 
1079 /*
1080  * Returns true if the given page belongs to a reservation and that page is
1081  * free.  Otherwise, returns false.
1082  */
1083 bool
1084 vm_reserv_is_page_free(vm_page_t m)
1085 {
1086 	vm_reserv_t rv;
1087 
1088 	rv = vm_reserv_from_page(m);
1089 	if (rv->object == NULL)
1090 		return (false);
1091 	return (popmap_is_clear(rv->popmap, m - rv->pages));
1092 }
1093 
1094 /*
1095  * If the given page belongs to a reservation, returns the level of that
1096  * reservation.  Otherwise, returns -1.
1097  */
1098 int
1099 vm_reserv_level(vm_page_t m)
1100 {
1101 	vm_reserv_t rv;
1102 
1103 	rv = vm_reserv_from_page(m);
1104 	return (rv->object != NULL ? 0 : -1);
1105 }
1106 
1107 /*
1108  * Returns a reservation level if the given page belongs to a fully populated
1109  * reservation and -1 otherwise.
1110  */
1111 int
1112 vm_reserv_level_iffullpop(vm_page_t m)
1113 {
1114 	vm_reserv_t rv;
1115 
1116 	rv = vm_reserv_from_page(m);
1117 	return (rv->popcnt == VM_LEVEL_0_NPAGES ? 0 : -1);
1118 }
1119 
1120 /*
1121  * Breaks the given partially populated reservation, releasing its free pages
1122  * to the physical memory allocator.
1123  *
1124  * The free page queue lock must be held.
1125  */
1126 static void
1127 vm_reserv_reclaim(vm_reserv_t rv)
1128 {
1129 
1130 	vm_reserv_assert_locked(rv);
1131 	CTR5(KTR_VM, "%s: rv %p object %p popcnt %d inpartpop %d",
1132 	    __FUNCTION__, rv, rv->object, rv->popcnt, rv->inpartpopq);
1133 	vm_reserv_domain_lock(rv->domain);
1134 	KASSERT(rv->inpartpopq,
1135 	    ("vm_reserv_reclaim: reserv %p's inpartpopq is FALSE", rv));
1136 	KASSERT(rv->domain < vm_ndomains,
1137 	    ("vm_reserv_reclaim: reserv %p's domain is corrupted %d",
1138 	    rv, rv->domain));
1139 	TAILQ_REMOVE(&vm_rvq_partpop[rv->domain], rv, partpopq);
1140 	rv->inpartpopq = FALSE;
1141 	vm_reserv_domain_unlock(rv->domain);
1142 	vm_reserv_break(rv);
1143 	counter_u64_add(vm_reserv_reclaimed, 1);
1144 }
1145 
1146 /*
1147  * Breaks the reservation at 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  * The free page queue lock must be held.
1152  */
1153 boolean_t
1154 vm_reserv_reclaim_inactive(int domain)
1155 {
1156 	vm_reserv_t rv;
1157 
1158 	while ((rv = TAILQ_FIRST(&vm_rvq_partpop[domain])) != NULL) {
1159 		vm_reserv_lock(rv);
1160 		if (rv != TAILQ_FIRST(&vm_rvq_partpop[domain])) {
1161 			vm_reserv_unlock(rv);
1162 			continue;
1163 		}
1164 		vm_reserv_reclaim(rv);
1165 		vm_reserv_unlock(rv);
1166 		return (TRUE);
1167 	}
1168 	return (FALSE);
1169 }
1170 
1171 /*
1172  * Determine whether this reservation has free pages that satisfy the given
1173  * request for contiguous physical memory.  Start searching from the lower
1174  * bound, defined by low_index.
1175  *
1176  * The free page queue lock must be held.
1177  */
1178 static bool
1179 vm_reserv_test_contig(vm_reserv_t rv, u_long npages, vm_paddr_t low,
1180     vm_paddr_t high, u_long alignment, vm_paddr_t boundary)
1181 {
1182 	vm_paddr_t pa, size;
1183 	u_long changes;
1184 	int bitpos, bits_left, i, hi, lo, n;
1185 
1186 	vm_reserv_assert_locked(rv);
1187 	size = npages << PAGE_SHIFT;
1188 	pa = VM_PAGE_TO_PHYS(&rv->pages[0]);
1189 	lo = (pa < low) ?
1190 	    ((low + PAGE_MASK - pa) >> PAGE_SHIFT) : 0;
1191 	i = lo / NBPOPMAP;
1192 	changes = rv->popmap[i] | ((1UL << (lo % NBPOPMAP)) - 1);
1193 	hi = (pa + VM_LEVEL_0_SIZE > high) ?
1194 	    ((high + PAGE_MASK - pa) >> PAGE_SHIFT) : VM_LEVEL_0_NPAGES;
1195 	n = hi / NBPOPMAP;
1196 	bits_left = hi % NBPOPMAP;
1197 	hi = lo = -1;
1198 	for (;;) {
1199 		/*
1200 		 * "changes" is a bitmask that marks where a new sequence of
1201 		 * 0s or 1s begins in popmap[i], with last bit in popmap[i-1]
1202 		 * considered to be 1 if and only if lo == hi.  The bits of
1203 		 * popmap[-1] and popmap[NPOPMAP] are considered all 1s.
1204 		 */
1205 		changes ^= (changes << 1) | (lo == hi);
1206 		while (changes != 0) {
1207 			/*
1208 			 * If the next change marked begins a run of 0s, set
1209 			 * lo to mark that position.  Otherwise set hi and
1210 			 * look for a satisfactory first page from lo up to hi.
1211 			 */
1212 			bitpos = ffsl(changes) - 1;
1213 			changes ^= 1UL << bitpos;
1214 			if (lo == hi) {
1215 				lo = NBPOPMAP * i + bitpos;
1216 				continue;
1217 			}
1218 			hi = NBPOPMAP * i + bitpos;
1219 			pa = VM_PAGE_TO_PHYS(&rv->pages[lo]);
1220 			if ((pa & (alignment - 1)) != 0) {
1221 				/* Skip to next aligned page. */
1222 				lo += (((pa - 1) | (alignment - 1)) + 1) >>
1223 				    PAGE_SHIFT;
1224 				if (lo >= VM_LEVEL_0_NPAGES)
1225 					return (false);
1226 				pa = VM_PAGE_TO_PHYS(&rv->pages[lo]);
1227 			}
1228 			if (((pa ^ (pa + size - 1)) & ~(boundary - 1)) != 0) {
1229 				/* Skip to next boundary-matching page. */
1230 				lo += (((pa - 1) | (boundary - 1)) + 1) >>
1231 				    PAGE_SHIFT;
1232 				if (lo >= VM_LEVEL_0_NPAGES)
1233 					return (false);
1234 				pa = VM_PAGE_TO_PHYS(&rv->pages[lo]);
1235 			}
1236 			if (lo * PAGE_SIZE + size <= hi * PAGE_SIZE)
1237 				return (true);
1238 			lo = hi;
1239 		}
1240 		if (++i < n)
1241 			changes = rv->popmap[i];
1242 		else if (i == n)
1243 			changes = bits_left == 0 ? -1UL :
1244 			    (rv->popmap[n] | (-1UL << bits_left));
1245 		else
1246 			return (false);
1247 	}
1248 }
1249 
1250 /*
1251  * Searches the partially populated reservation queue for the least recently
1252  * changed reservation with free pages that satisfy the given request for
1253  * contiguous physical memory.  If a satisfactory reservation is found, it is
1254  * broken.  Returns true if a reservation is broken and false otherwise.
1255  *
1256  * The free page queue lock must be held.
1257  */
1258 boolean_t
1259 vm_reserv_reclaim_contig(int domain, u_long npages, vm_paddr_t low,
1260     vm_paddr_t high, u_long alignment, vm_paddr_t boundary)
1261 {
1262 	vm_paddr_t pa, size;
1263 	vm_reserv_t rv, rvn;
1264 
1265 	if (npages > VM_LEVEL_0_NPAGES - 1)
1266 		return (false);
1267 	size = npages << PAGE_SHIFT;
1268 	vm_reserv_domain_lock(domain);
1269 again:
1270 	for (rv = TAILQ_FIRST(&vm_rvq_partpop[domain]); rv != NULL; rv = rvn) {
1271 		rvn = TAILQ_NEXT(rv, partpopq);
1272 		pa = VM_PAGE_TO_PHYS(&rv->pages[0]);
1273 		if (pa + VM_LEVEL_0_SIZE - size < low) {
1274 			/* This entire reservation is too low; go to next. */
1275 			continue;
1276 		}
1277 		if (pa + size > high) {
1278 			/* This entire reservation is too high; go to next. */
1279 			continue;
1280 		}
1281 		if (vm_reserv_trylock(rv) == 0) {
1282 			vm_reserv_domain_unlock(domain);
1283 			vm_reserv_lock(rv);
1284 			if (!rv->inpartpopq) {
1285 				vm_reserv_domain_lock(domain);
1286 				if (!rvn->inpartpopq)
1287 					goto again;
1288 				continue;
1289 			}
1290 		} else
1291 			vm_reserv_domain_unlock(domain);
1292 		if (vm_reserv_test_contig(rv, npages, low, high,
1293 		    alignment, boundary)) {
1294 			vm_reserv_reclaim(rv);
1295 			vm_reserv_unlock(rv);
1296 			return (true);
1297 		}
1298 		vm_reserv_unlock(rv);
1299 		vm_reserv_domain_lock(domain);
1300 		if (rvn != NULL && !rvn->inpartpopq)
1301 			goto again;
1302 	}
1303 	vm_reserv_domain_unlock(domain);
1304 	return (false);
1305 }
1306 
1307 /*
1308  * Transfers the reservation underlying the given page to a new object.
1309  *
1310  * The object must be locked.
1311  */
1312 void
1313 vm_reserv_rename(vm_page_t m, vm_object_t new_object, vm_object_t old_object,
1314     vm_pindex_t old_object_offset)
1315 {
1316 	vm_reserv_t rv;
1317 
1318 	VM_OBJECT_ASSERT_WLOCKED(new_object);
1319 	rv = vm_reserv_from_page(m);
1320 	if (rv->object == old_object) {
1321 		vm_reserv_lock(rv);
1322 		CTR6(KTR_VM,
1323 		    "%s: rv %p object %p new %p popcnt %d inpartpop %d",
1324 		    __FUNCTION__, rv, rv->object, new_object, rv->popcnt,
1325 		    rv->inpartpopq);
1326 		if (rv->object == old_object) {
1327 			vm_reserv_object_lock(old_object);
1328 			rv->object = NULL;
1329 			LIST_REMOVE(rv, objq);
1330 			vm_reserv_object_unlock(old_object);
1331 			vm_reserv_object_lock(new_object);
1332 			rv->object = new_object;
1333 			rv->pindex -= old_object_offset;
1334 			LIST_INSERT_HEAD(&new_object->rvq, rv, objq);
1335 			vm_reserv_object_unlock(new_object);
1336 		}
1337 		vm_reserv_unlock(rv);
1338 	}
1339 }
1340 
1341 /*
1342  * Returns the size (in bytes) of a reservation of the specified level.
1343  */
1344 int
1345 vm_reserv_size(int level)
1346 {
1347 
1348 	switch (level) {
1349 	case 0:
1350 		return (VM_LEVEL_0_SIZE);
1351 	case -1:
1352 		return (PAGE_SIZE);
1353 	default:
1354 		return (0);
1355 	}
1356 }
1357 
1358 /*
1359  * Allocates the virtual and physical memory required by the reservation
1360  * management system's data structures, in particular, the reservation array.
1361  */
1362 vm_paddr_t
1363 vm_reserv_startup(vm_offset_t *vaddr, vm_paddr_t end)
1364 {
1365 	vm_paddr_t new_end, high_water;
1366 	size_t size;
1367 	int i;
1368 
1369 	high_water = phys_avail[1];
1370 	for (i = 0; i < vm_phys_nsegs; i++) {
1371 		if (vm_phys_segs[i].end > high_water)
1372 			high_water = vm_phys_segs[i].end;
1373 	}
1374 
1375 	/* Skip the first chunk.  It is already accounted for. */
1376 	for (i = 2; phys_avail[i + 1] != 0; i += 2) {
1377 		if (phys_avail[i + 1] > high_water)
1378 			high_water = phys_avail[i + 1];
1379 	}
1380 
1381 	/*
1382 	 * Calculate the size (in bytes) of the reservation array.  Round up
1383 	 * from "high_water" because every small page is mapped to an element
1384 	 * in the reservation array based on its physical address.  Thus, the
1385 	 * number of elements in the reservation array can be greater than the
1386 	 * number of superpages.
1387 	 */
1388 	size = howmany(high_water, VM_LEVEL_0_SIZE) * sizeof(struct vm_reserv);
1389 
1390 	/*
1391 	 * Allocate and map the physical memory for the reservation array.  The
1392 	 * next available virtual address is returned by reference.
1393 	 */
1394 	new_end = end - round_page(size);
1395 	vm_reserv_array = (void *)(uintptr_t)pmap_map(vaddr, new_end, end,
1396 	    VM_PROT_READ | VM_PROT_WRITE);
1397 	bzero(vm_reserv_array, size);
1398 
1399 	/*
1400 	 * Return the next available physical address.
1401 	 */
1402 	return (new_end);
1403 }
1404 
1405 /*
1406  * Initializes the reservation management system.  Specifically, initializes
1407  * the reservation counters.
1408  */
1409 static void
1410 vm_reserv_counter_init(void *unused)
1411 {
1412 
1413 	vm_reserv_freed = counter_u64_alloc(M_WAITOK);
1414 	vm_reserv_broken = counter_u64_alloc(M_WAITOK);
1415 	vm_reserv_reclaimed = counter_u64_alloc(M_WAITOK);
1416 }
1417 SYSINIT(vm_reserv_counter_init, SI_SUB_CPU, SI_ORDER_ANY,
1418     vm_reserv_counter_init, NULL);
1419 
1420 /*
1421  * Returns the superpage containing the given page.
1422  */
1423 vm_page_t
1424 vm_reserv_to_superpage(vm_page_t m)
1425 {
1426 	vm_reserv_t rv;
1427 
1428 	VM_OBJECT_ASSERT_LOCKED(m->object);
1429 	rv = vm_reserv_from_page(m);
1430 	if (rv->object == m->object && rv->popcnt == VM_LEVEL_0_NPAGES)
1431 		m = rv->pages;
1432 	else
1433 		m = NULL;
1434 
1435 	return (m);
1436 }
1437 
1438 #endif	/* VM_NRESERVLEVEL > 0 */
1439