xref: /freebsd/sys/vm/vm_phys.c (revision 123af6ec70016f5556da5972d4d63c7d175c06d3)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3  *
4  * Copyright (c) 2002-2006 Rice University
5  * Copyright (c) 2007 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  *	Physical memory system implementation
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_ddb.h"
45 #include "opt_vm.h"
46 
47 #include <sys/param.h>
48 #include <sys/systm.h>
49 #include <sys/domainset.h>
50 #include <sys/lock.h>
51 #include <sys/kernel.h>
52 #include <sys/malloc.h>
53 #include <sys/mutex.h>
54 #include <sys/proc.h>
55 #include <sys/queue.h>
56 #include <sys/rwlock.h>
57 #include <sys/sbuf.h>
58 #include <sys/sysctl.h>
59 #include <sys/tree.h>
60 #include <sys/vmmeter.h>
61 
62 #include <ddb/ddb.h>
63 
64 #include <vm/vm.h>
65 #include <vm/vm_param.h>
66 #include <vm/vm_kern.h>
67 #include <vm/vm_object.h>
68 #include <vm/vm_page.h>
69 #include <vm/vm_phys.h>
70 #include <vm/vm_pagequeue.h>
71 
72 _Static_assert(sizeof(long) * NBBY >= VM_PHYSSEG_MAX,
73     "Too many physsegs.");
74 
75 #ifdef NUMA
76 struct mem_affinity __read_mostly *mem_affinity;
77 int __read_mostly *mem_locality;
78 #endif
79 
80 int __read_mostly vm_ndomains = 1;
81 domainset_t __read_mostly all_domains = DOMAINSET_T_INITIALIZER(0x1);
82 
83 struct vm_phys_seg __read_mostly vm_phys_segs[VM_PHYSSEG_MAX];
84 int __read_mostly vm_phys_nsegs;
85 
86 struct vm_phys_fictitious_seg;
87 static int vm_phys_fictitious_cmp(struct vm_phys_fictitious_seg *,
88     struct vm_phys_fictitious_seg *);
89 
90 RB_HEAD(fict_tree, vm_phys_fictitious_seg) vm_phys_fictitious_tree =
91     RB_INITIALIZER(_vm_phys_fictitious_tree);
92 
93 struct vm_phys_fictitious_seg {
94 	RB_ENTRY(vm_phys_fictitious_seg) node;
95 	/* Memory region data */
96 	vm_paddr_t	start;
97 	vm_paddr_t	end;
98 	vm_page_t	first_page;
99 };
100 
101 RB_GENERATE_STATIC(fict_tree, vm_phys_fictitious_seg, node,
102     vm_phys_fictitious_cmp);
103 
104 static struct rwlock_padalign vm_phys_fictitious_reg_lock;
105 MALLOC_DEFINE(M_FICT_PAGES, "vm_fictitious", "Fictitious VM pages");
106 
107 static struct vm_freelist __aligned(CACHE_LINE_SIZE)
108     vm_phys_free_queues[MAXMEMDOM][VM_NFREELIST][VM_NFREEPOOL]
109     [VM_NFREEORDER_MAX];
110 
111 static int __read_mostly vm_nfreelists;
112 
113 /*
114  * Provides the mapping from VM_FREELIST_* to free list indices (flind).
115  */
116 static int __read_mostly vm_freelist_to_flind[VM_NFREELIST];
117 
118 CTASSERT(VM_FREELIST_DEFAULT == 0);
119 
120 #ifdef VM_FREELIST_DMA32
121 #define	VM_DMA32_BOUNDARY	((vm_paddr_t)1 << 32)
122 #endif
123 
124 /*
125  * Enforce the assumptions made by vm_phys_add_seg() and vm_phys_init() about
126  * the ordering of the free list boundaries.
127  */
128 #if defined(VM_LOWMEM_BOUNDARY) && defined(VM_DMA32_BOUNDARY)
129 CTASSERT(VM_LOWMEM_BOUNDARY < VM_DMA32_BOUNDARY);
130 #endif
131 
132 static int sysctl_vm_phys_free(SYSCTL_HANDLER_ARGS);
133 SYSCTL_OID(_vm, OID_AUTO, phys_free, CTLTYPE_STRING | CTLFLAG_RD,
134     NULL, 0, sysctl_vm_phys_free, "A", "Phys Free Info");
135 
136 static int sysctl_vm_phys_segs(SYSCTL_HANDLER_ARGS);
137 SYSCTL_OID(_vm, OID_AUTO, phys_segs, CTLTYPE_STRING | CTLFLAG_RD,
138     NULL, 0, sysctl_vm_phys_segs, "A", "Phys Seg Info");
139 
140 #ifdef NUMA
141 static int sysctl_vm_phys_locality(SYSCTL_HANDLER_ARGS);
142 SYSCTL_OID(_vm, OID_AUTO, phys_locality, CTLTYPE_STRING | CTLFLAG_RD,
143     NULL, 0, sysctl_vm_phys_locality, "A", "Phys Locality Info");
144 #endif
145 
146 SYSCTL_INT(_vm, OID_AUTO, ndomains, CTLFLAG_RD,
147     &vm_ndomains, 0, "Number of physical memory domains available.");
148 
149 static vm_page_t vm_phys_alloc_seg_contig(struct vm_phys_seg *seg,
150     u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment,
151     vm_paddr_t boundary);
152 static void _vm_phys_create_seg(vm_paddr_t start, vm_paddr_t end, int domain);
153 static void vm_phys_create_seg(vm_paddr_t start, vm_paddr_t end);
154 static void vm_phys_split_pages(vm_page_t m, int oind, struct vm_freelist *fl,
155     int order, int tail);
156 
157 /*
158  * Red-black tree helpers for vm fictitious range management.
159  */
160 static inline int
161 vm_phys_fictitious_in_range(struct vm_phys_fictitious_seg *p,
162     struct vm_phys_fictitious_seg *range)
163 {
164 
165 	KASSERT(range->start != 0 && range->end != 0,
166 	    ("Invalid range passed on search for vm_fictitious page"));
167 	if (p->start >= range->end)
168 		return (1);
169 	if (p->start < range->start)
170 		return (-1);
171 
172 	return (0);
173 }
174 
175 static int
176 vm_phys_fictitious_cmp(struct vm_phys_fictitious_seg *p1,
177     struct vm_phys_fictitious_seg *p2)
178 {
179 
180 	/* Check if this is a search for a page */
181 	if (p1->end == 0)
182 		return (vm_phys_fictitious_in_range(p1, p2));
183 
184 	KASSERT(p2->end != 0,
185     ("Invalid range passed as second parameter to vm fictitious comparison"));
186 
187 	/* Searching to add a new range */
188 	if (p1->end <= p2->start)
189 		return (-1);
190 	if (p1->start >= p2->end)
191 		return (1);
192 
193 	panic("Trying to add overlapping vm fictitious ranges:\n"
194 	    "[%#jx:%#jx] and [%#jx:%#jx]", (uintmax_t)p1->start,
195 	    (uintmax_t)p1->end, (uintmax_t)p2->start, (uintmax_t)p2->end);
196 }
197 
198 int
199 vm_phys_domain_match(int prefer, vm_paddr_t low, vm_paddr_t high)
200 {
201 #ifdef NUMA
202 	domainset_t mask;
203 	int i;
204 
205 	if (vm_ndomains == 1 || mem_affinity == NULL)
206 		return (0);
207 
208 	DOMAINSET_ZERO(&mask);
209 	/*
210 	 * Check for any memory that overlaps low, high.
211 	 */
212 	for (i = 0; mem_affinity[i].end != 0; i++)
213 		if (mem_affinity[i].start <= high &&
214 		    mem_affinity[i].end >= low)
215 			DOMAINSET_SET(mem_affinity[i].domain, &mask);
216 	if (prefer != -1 && DOMAINSET_ISSET(prefer, &mask))
217 		return (prefer);
218 	if (DOMAINSET_EMPTY(&mask))
219 		panic("vm_phys_domain_match:  Impossible constraint");
220 	return (DOMAINSET_FFS(&mask) - 1);
221 #else
222 	return (0);
223 #endif
224 }
225 
226 /*
227  * Outputs the state of the physical memory allocator, specifically,
228  * the amount of physical memory in each free list.
229  */
230 static int
231 sysctl_vm_phys_free(SYSCTL_HANDLER_ARGS)
232 {
233 	struct sbuf sbuf;
234 	struct vm_freelist *fl;
235 	int dom, error, flind, oind, pind;
236 
237 	error = sysctl_wire_old_buffer(req, 0);
238 	if (error != 0)
239 		return (error);
240 	sbuf_new_for_sysctl(&sbuf, NULL, 128 * vm_ndomains, req);
241 	for (dom = 0; dom < vm_ndomains; dom++) {
242 		sbuf_printf(&sbuf,"\nDOMAIN %d:\n", dom);
243 		for (flind = 0; flind < vm_nfreelists; flind++) {
244 			sbuf_printf(&sbuf, "\nFREE LIST %d:\n"
245 			    "\n  ORDER (SIZE)  |  NUMBER"
246 			    "\n              ", flind);
247 			for (pind = 0; pind < VM_NFREEPOOL; pind++)
248 				sbuf_printf(&sbuf, "  |  POOL %d", pind);
249 			sbuf_printf(&sbuf, "\n--            ");
250 			for (pind = 0; pind < VM_NFREEPOOL; pind++)
251 				sbuf_printf(&sbuf, "-- --      ");
252 			sbuf_printf(&sbuf, "--\n");
253 			for (oind = VM_NFREEORDER - 1; oind >= 0; oind--) {
254 				sbuf_printf(&sbuf, "  %2d (%6dK)", oind,
255 				    1 << (PAGE_SHIFT - 10 + oind));
256 				for (pind = 0; pind < VM_NFREEPOOL; pind++) {
257 				fl = vm_phys_free_queues[dom][flind][pind];
258 					sbuf_printf(&sbuf, "  |  %6d",
259 					    fl[oind].lcnt);
260 				}
261 				sbuf_printf(&sbuf, "\n");
262 			}
263 		}
264 	}
265 	error = sbuf_finish(&sbuf);
266 	sbuf_delete(&sbuf);
267 	return (error);
268 }
269 
270 /*
271  * Outputs the set of physical memory segments.
272  */
273 static int
274 sysctl_vm_phys_segs(SYSCTL_HANDLER_ARGS)
275 {
276 	struct sbuf sbuf;
277 	struct vm_phys_seg *seg;
278 	int error, segind;
279 
280 	error = sysctl_wire_old_buffer(req, 0);
281 	if (error != 0)
282 		return (error);
283 	sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
284 	for (segind = 0; segind < vm_phys_nsegs; segind++) {
285 		sbuf_printf(&sbuf, "\nSEGMENT %d:\n\n", segind);
286 		seg = &vm_phys_segs[segind];
287 		sbuf_printf(&sbuf, "start:     %#jx\n",
288 		    (uintmax_t)seg->start);
289 		sbuf_printf(&sbuf, "end:       %#jx\n",
290 		    (uintmax_t)seg->end);
291 		sbuf_printf(&sbuf, "domain:    %d\n", seg->domain);
292 		sbuf_printf(&sbuf, "free list: %p\n", seg->free_queues);
293 	}
294 	error = sbuf_finish(&sbuf);
295 	sbuf_delete(&sbuf);
296 	return (error);
297 }
298 
299 /*
300  * Return affinity, or -1 if there's no affinity information.
301  */
302 int
303 vm_phys_mem_affinity(int f, int t)
304 {
305 
306 #ifdef NUMA
307 	if (mem_locality == NULL)
308 		return (-1);
309 	if (f >= vm_ndomains || t >= vm_ndomains)
310 		return (-1);
311 	return (mem_locality[f * vm_ndomains + t]);
312 #else
313 	return (-1);
314 #endif
315 }
316 
317 #ifdef NUMA
318 /*
319  * Outputs the VM locality table.
320  */
321 static int
322 sysctl_vm_phys_locality(SYSCTL_HANDLER_ARGS)
323 {
324 	struct sbuf sbuf;
325 	int error, i, j;
326 
327 	error = sysctl_wire_old_buffer(req, 0);
328 	if (error != 0)
329 		return (error);
330 	sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
331 
332 	sbuf_printf(&sbuf, "\n");
333 
334 	for (i = 0; i < vm_ndomains; i++) {
335 		sbuf_printf(&sbuf, "%d: ", i);
336 		for (j = 0; j < vm_ndomains; j++) {
337 			sbuf_printf(&sbuf, "%d ", vm_phys_mem_affinity(i, j));
338 		}
339 		sbuf_printf(&sbuf, "\n");
340 	}
341 	error = sbuf_finish(&sbuf);
342 	sbuf_delete(&sbuf);
343 	return (error);
344 }
345 #endif
346 
347 static void
348 vm_freelist_add(struct vm_freelist *fl, vm_page_t m, int order, int tail)
349 {
350 
351 	m->order = order;
352 	if (tail)
353 		TAILQ_INSERT_TAIL(&fl[order].pl, m, listq);
354 	else
355 		TAILQ_INSERT_HEAD(&fl[order].pl, m, listq);
356 	fl[order].lcnt++;
357 }
358 
359 static void
360 vm_freelist_rem(struct vm_freelist *fl, vm_page_t m, int order)
361 {
362 
363 	TAILQ_REMOVE(&fl[order].pl, m, listq);
364 	fl[order].lcnt--;
365 	m->order = VM_NFREEORDER;
366 }
367 
368 /*
369  * Create a physical memory segment.
370  */
371 static void
372 _vm_phys_create_seg(vm_paddr_t start, vm_paddr_t end, int domain)
373 {
374 	struct vm_phys_seg *seg;
375 
376 	KASSERT(vm_phys_nsegs < VM_PHYSSEG_MAX,
377 	    ("vm_phys_create_seg: increase VM_PHYSSEG_MAX"));
378 	KASSERT(domain >= 0 && domain < vm_ndomains,
379 	    ("vm_phys_create_seg: invalid domain provided"));
380 	seg = &vm_phys_segs[vm_phys_nsegs++];
381 	while (seg > vm_phys_segs && (seg - 1)->start >= end) {
382 		*seg = *(seg - 1);
383 		seg--;
384 	}
385 	seg->start = start;
386 	seg->end = end;
387 	seg->domain = domain;
388 }
389 
390 static void
391 vm_phys_create_seg(vm_paddr_t start, vm_paddr_t end)
392 {
393 #ifdef NUMA
394 	int i;
395 
396 	if (mem_affinity == NULL) {
397 		_vm_phys_create_seg(start, end, 0);
398 		return;
399 	}
400 
401 	for (i = 0;; i++) {
402 		if (mem_affinity[i].end == 0)
403 			panic("Reached end of affinity info");
404 		if (mem_affinity[i].end <= start)
405 			continue;
406 		if (mem_affinity[i].start > start)
407 			panic("No affinity info for start %jx",
408 			    (uintmax_t)start);
409 		if (mem_affinity[i].end >= end) {
410 			_vm_phys_create_seg(start, end,
411 			    mem_affinity[i].domain);
412 			break;
413 		}
414 		_vm_phys_create_seg(start, mem_affinity[i].end,
415 		    mem_affinity[i].domain);
416 		start = mem_affinity[i].end;
417 	}
418 #else
419 	_vm_phys_create_seg(start, end, 0);
420 #endif
421 }
422 
423 /*
424  * Add a physical memory segment.
425  */
426 void
427 vm_phys_add_seg(vm_paddr_t start, vm_paddr_t end)
428 {
429 	vm_paddr_t paddr;
430 
431 	KASSERT((start & PAGE_MASK) == 0,
432 	    ("vm_phys_define_seg: start is not page aligned"));
433 	KASSERT((end & PAGE_MASK) == 0,
434 	    ("vm_phys_define_seg: end is not page aligned"));
435 
436 	/*
437 	 * Split the physical memory segment if it spans two or more free
438 	 * list boundaries.
439 	 */
440 	paddr = start;
441 #ifdef	VM_FREELIST_LOWMEM
442 	if (paddr < VM_LOWMEM_BOUNDARY && end > VM_LOWMEM_BOUNDARY) {
443 		vm_phys_create_seg(paddr, VM_LOWMEM_BOUNDARY);
444 		paddr = VM_LOWMEM_BOUNDARY;
445 	}
446 #endif
447 #ifdef	VM_FREELIST_DMA32
448 	if (paddr < VM_DMA32_BOUNDARY && end > VM_DMA32_BOUNDARY) {
449 		vm_phys_create_seg(paddr, VM_DMA32_BOUNDARY);
450 		paddr = VM_DMA32_BOUNDARY;
451 	}
452 #endif
453 	vm_phys_create_seg(paddr, end);
454 }
455 
456 /*
457  * Initialize the physical memory allocator.
458  *
459  * Requires that vm_page_array is initialized!
460  */
461 void
462 vm_phys_init(void)
463 {
464 	struct vm_freelist *fl;
465 	struct vm_phys_seg *end_seg, *prev_seg, *seg, *tmp_seg;
466 	u_long npages;
467 	int dom, flind, freelist, oind, pind, segind;
468 
469 	/*
470 	 * Compute the number of free lists, and generate the mapping from the
471 	 * manifest constants VM_FREELIST_* to the free list indices.
472 	 *
473 	 * Initially, the entries of vm_freelist_to_flind[] are set to either
474 	 * 0 or 1 to indicate which free lists should be created.
475 	 */
476 	npages = 0;
477 	for (segind = vm_phys_nsegs - 1; segind >= 0; segind--) {
478 		seg = &vm_phys_segs[segind];
479 #ifdef	VM_FREELIST_LOWMEM
480 		if (seg->end <= VM_LOWMEM_BOUNDARY)
481 			vm_freelist_to_flind[VM_FREELIST_LOWMEM] = 1;
482 		else
483 #endif
484 #ifdef	VM_FREELIST_DMA32
485 		if (
486 #ifdef	VM_DMA32_NPAGES_THRESHOLD
487 		    /*
488 		     * Create the DMA32 free list only if the amount of
489 		     * physical memory above physical address 4G exceeds the
490 		     * given threshold.
491 		     */
492 		    npages > VM_DMA32_NPAGES_THRESHOLD &&
493 #endif
494 		    seg->end <= VM_DMA32_BOUNDARY)
495 			vm_freelist_to_flind[VM_FREELIST_DMA32] = 1;
496 		else
497 #endif
498 		{
499 			npages += atop(seg->end - seg->start);
500 			vm_freelist_to_flind[VM_FREELIST_DEFAULT] = 1;
501 		}
502 	}
503 	/* Change each entry into a running total of the free lists. */
504 	for (freelist = 1; freelist < VM_NFREELIST; freelist++) {
505 		vm_freelist_to_flind[freelist] +=
506 		    vm_freelist_to_flind[freelist - 1];
507 	}
508 	vm_nfreelists = vm_freelist_to_flind[VM_NFREELIST - 1];
509 	KASSERT(vm_nfreelists > 0, ("vm_phys_init: no free lists"));
510 	/* Change each entry into a free list index. */
511 	for (freelist = 0; freelist < VM_NFREELIST; freelist++)
512 		vm_freelist_to_flind[freelist]--;
513 
514 	/*
515 	 * Initialize the first_page and free_queues fields of each physical
516 	 * memory segment.
517 	 */
518 #ifdef VM_PHYSSEG_SPARSE
519 	npages = 0;
520 #endif
521 	for (segind = 0; segind < vm_phys_nsegs; segind++) {
522 		seg = &vm_phys_segs[segind];
523 #ifdef VM_PHYSSEG_SPARSE
524 		seg->first_page = &vm_page_array[npages];
525 		npages += atop(seg->end - seg->start);
526 #else
527 		seg->first_page = PHYS_TO_VM_PAGE(seg->start);
528 #endif
529 #ifdef	VM_FREELIST_LOWMEM
530 		if (seg->end <= VM_LOWMEM_BOUNDARY) {
531 			flind = vm_freelist_to_flind[VM_FREELIST_LOWMEM];
532 			KASSERT(flind >= 0,
533 			    ("vm_phys_init: LOWMEM flind < 0"));
534 		} else
535 #endif
536 #ifdef	VM_FREELIST_DMA32
537 		if (seg->end <= VM_DMA32_BOUNDARY) {
538 			flind = vm_freelist_to_flind[VM_FREELIST_DMA32];
539 			KASSERT(flind >= 0,
540 			    ("vm_phys_init: DMA32 flind < 0"));
541 		} else
542 #endif
543 		{
544 			flind = vm_freelist_to_flind[VM_FREELIST_DEFAULT];
545 			KASSERT(flind >= 0,
546 			    ("vm_phys_init: DEFAULT flind < 0"));
547 		}
548 		seg->free_queues = &vm_phys_free_queues[seg->domain][flind];
549 	}
550 
551 	/*
552 	 * Coalesce physical memory segments that are contiguous and share the
553 	 * same per-domain free queues.
554 	 */
555 	prev_seg = vm_phys_segs;
556 	seg = &vm_phys_segs[1];
557 	end_seg = &vm_phys_segs[vm_phys_nsegs];
558 	while (seg < end_seg) {
559 		if (prev_seg->end == seg->start &&
560 		    prev_seg->free_queues == seg->free_queues) {
561 			prev_seg->end = seg->end;
562 			KASSERT(prev_seg->domain == seg->domain,
563 			    ("vm_phys_init: free queues cannot span domains"));
564 			vm_phys_nsegs--;
565 			end_seg--;
566 			for (tmp_seg = seg; tmp_seg < end_seg; tmp_seg++)
567 				*tmp_seg = *(tmp_seg + 1);
568 		} else {
569 			prev_seg = seg;
570 			seg++;
571 		}
572 	}
573 
574 	/*
575 	 * Initialize the free queues.
576 	 */
577 	for (dom = 0; dom < vm_ndomains; dom++) {
578 		for (flind = 0; flind < vm_nfreelists; flind++) {
579 			for (pind = 0; pind < VM_NFREEPOOL; pind++) {
580 				fl = vm_phys_free_queues[dom][flind][pind];
581 				for (oind = 0; oind < VM_NFREEORDER; oind++)
582 					TAILQ_INIT(&fl[oind].pl);
583 			}
584 		}
585 	}
586 
587 	rw_init(&vm_phys_fictitious_reg_lock, "vmfctr");
588 }
589 
590 /*
591  * Register info about the NUMA topology of the system.
592  *
593  * Invoked by platform-dependent code prior to vm_phys_init().
594  */
595 void
596 vm_phys_register_domains(int ndomains, struct mem_affinity *affinity,
597     int *locality)
598 {
599 #ifdef NUMA
600 	int d, i;
601 
602 	/*
603 	 * For now the only override value that we support is 1, which
604 	 * effectively disables NUMA-awareness in the allocators.
605 	 */
606 	d = 0;
607 	TUNABLE_INT_FETCH("vm.numa.disabled", &d);
608 	if (d)
609 		ndomains = 1;
610 
611 	if (ndomains > 1) {
612 		vm_ndomains = ndomains;
613 		mem_affinity = affinity;
614 		mem_locality = locality;
615 	}
616 
617 	for (i = 0; i < vm_ndomains; i++)
618 		DOMAINSET_SET(i, &all_domains);
619 #else
620 	(void)ndomains;
621 	(void)affinity;
622 	(void)locality;
623 #endif
624 }
625 
626 /*
627  * Split a contiguous, power of two-sized set of physical pages.
628  *
629  * When this function is called by a page allocation function, the caller
630  * should request insertion at the head unless the order [order, oind) queues
631  * are known to be empty.  The objective being to reduce the likelihood of
632  * long-term fragmentation by promoting contemporaneous allocation and
633  * (hopefully) deallocation.
634  */
635 static __inline void
636 vm_phys_split_pages(vm_page_t m, int oind, struct vm_freelist *fl, int order,
637     int tail)
638 {
639 	vm_page_t m_buddy;
640 
641 	while (oind > order) {
642 		oind--;
643 		m_buddy = &m[1 << oind];
644 		KASSERT(m_buddy->order == VM_NFREEORDER,
645 		    ("vm_phys_split_pages: page %p has unexpected order %d",
646 		    m_buddy, m_buddy->order));
647 		vm_freelist_add(fl, m_buddy, oind, tail);
648         }
649 }
650 
651 /*
652  * Add the physical pages [m, m + npages) at the end of a power-of-two aligned
653  * and sized set to the specified free list.
654  *
655  * When this function is called by a page allocation function, the caller
656  * should request insertion at the head unless the lower-order queues are
657  * known to be empty.  The objective being to reduce the likelihood of long-
658  * term fragmentation by promoting contemporaneous allocation and (hopefully)
659  * deallocation.
660  *
661  * The physical page m's buddy must not be free.
662  */
663 static void
664 vm_phys_enq_range(vm_page_t m, u_int npages, struct vm_freelist *fl, int tail)
665 {
666 	u_int n;
667 	int order;
668 
669 	KASSERT(npages > 0, ("vm_phys_enq_range: npages is 0"));
670 	KASSERT(((VM_PAGE_TO_PHYS(m) + npages * PAGE_SIZE) &
671 	    ((PAGE_SIZE << (fls(npages) - 1)) - 1)) == 0,
672 	    ("vm_phys_enq_range: page %p and npages %u are misaligned",
673 	    m, npages));
674 	do {
675 		KASSERT(m->order == VM_NFREEORDER,
676 		    ("vm_phys_enq_range: page %p has unexpected order %d",
677 		    m, m->order));
678 		order = ffs(npages) - 1;
679 		KASSERT(order < VM_NFREEORDER,
680 		    ("vm_phys_enq_range: order %d is out of range", order));
681 		vm_freelist_add(fl, m, order, tail);
682 		n = 1 << order;
683 		m += n;
684 		npages -= n;
685 	} while (npages > 0);
686 }
687 
688 /*
689  * Tries to allocate the specified number of pages from the specified pool
690  * within the specified domain.  Returns the actual number of allocated pages
691  * and a pointer to each page through the array ma[].
692  *
693  * The returned pages may not be physically contiguous.  However, in contrast
694  * to performing multiple, back-to-back calls to vm_phys_alloc_pages(..., 0),
695  * calling this function once to allocate the desired number of pages will
696  * avoid wasted time in vm_phys_split_pages().
697  *
698  * The free page queues for the specified domain must be locked.
699  */
700 int
701 vm_phys_alloc_npages(int domain, int pool, int npages, vm_page_t ma[])
702 {
703 	struct vm_freelist *alt, *fl;
704 	vm_page_t m;
705 	int avail, end, flind, freelist, i, need, oind, pind;
706 
707 	KASSERT(domain >= 0 && domain < vm_ndomains,
708 	    ("vm_phys_alloc_npages: domain %d is out of range", domain));
709 	KASSERT(pool < VM_NFREEPOOL,
710 	    ("vm_phys_alloc_npages: pool %d is out of range", pool));
711 	KASSERT(npages <= 1 << (VM_NFREEORDER - 1),
712 	    ("vm_phys_alloc_npages: npages %d is out of range", npages));
713 	vm_domain_free_assert_locked(VM_DOMAIN(domain));
714 	i = 0;
715 	for (freelist = 0; freelist < VM_NFREELIST; freelist++) {
716 		flind = vm_freelist_to_flind[freelist];
717 		if (flind < 0)
718 			continue;
719 		fl = vm_phys_free_queues[domain][flind][pool];
720 		for (oind = 0; oind < VM_NFREEORDER; oind++) {
721 			while ((m = TAILQ_FIRST(&fl[oind].pl)) != NULL) {
722 				vm_freelist_rem(fl, m, oind);
723 				avail = 1 << oind;
724 				need = imin(npages - i, avail);
725 				for (end = i + need; i < end;)
726 					ma[i++] = m++;
727 				if (need < avail) {
728 					/*
729 					 * Return excess pages to fl.  Its
730 					 * order [0, oind) queues are empty.
731 					 */
732 					vm_phys_enq_range(m, avail - need, fl,
733 					    1);
734 					return (npages);
735 				} else if (i == npages)
736 					return (npages);
737 			}
738 		}
739 		for (oind = VM_NFREEORDER - 1; oind >= 0; oind--) {
740 			for (pind = 0; pind < VM_NFREEPOOL; pind++) {
741 				alt = vm_phys_free_queues[domain][flind][pind];
742 				while ((m = TAILQ_FIRST(&alt[oind].pl)) !=
743 				    NULL) {
744 					vm_freelist_rem(alt, m, oind);
745 					vm_phys_set_pool(pool, m, oind);
746 					avail = 1 << oind;
747 					need = imin(npages - i, avail);
748 					for (end = i + need; i < end;)
749 						ma[i++] = m++;
750 					if (need < avail) {
751 						/*
752 						 * Return excess pages to fl.
753 						 * Its order [0, oind) queues
754 						 * are empty.
755 						 */
756 						vm_phys_enq_range(m, avail -
757 						    need, fl, 1);
758 						return (npages);
759 					} else if (i == npages)
760 						return (npages);
761 				}
762 			}
763 		}
764 	}
765 	return (i);
766 }
767 
768 /*
769  * Allocate a contiguous, power of two-sized set of physical pages
770  * from the free lists.
771  *
772  * The free page queues must be locked.
773  */
774 vm_page_t
775 vm_phys_alloc_pages(int domain, int pool, int order)
776 {
777 	vm_page_t m;
778 	int freelist;
779 
780 	for (freelist = 0; freelist < VM_NFREELIST; freelist++) {
781 		m = vm_phys_alloc_freelist_pages(domain, freelist, pool, order);
782 		if (m != NULL)
783 			return (m);
784 	}
785 	return (NULL);
786 }
787 
788 /*
789  * Allocate a contiguous, power of two-sized set of physical pages from the
790  * specified free list.  The free list must be specified using one of the
791  * manifest constants VM_FREELIST_*.
792  *
793  * The free page queues must be locked.
794  */
795 vm_page_t
796 vm_phys_alloc_freelist_pages(int domain, int freelist, int pool, int order)
797 {
798 	struct vm_freelist *alt, *fl;
799 	vm_page_t m;
800 	int oind, pind, flind;
801 
802 	KASSERT(domain >= 0 && domain < vm_ndomains,
803 	    ("vm_phys_alloc_freelist_pages: domain %d is out of range",
804 	    domain));
805 	KASSERT(freelist < VM_NFREELIST,
806 	    ("vm_phys_alloc_freelist_pages: freelist %d is out of range",
807 	    freelist));
808 	KASSERT(pool < VM_NFREEPOOL,
809 	    ("vm_phys_alloc_freelist_pages: pool %d is out of range", pool));
810 	KASSERT(order < VM_NFREEORDER,
811 	    ("vm_phys_alloc_freelist_pages: order %d is out of range", order));
812 
813 	flind = vm_freelist_to_flind[freelist];
814 	/* Check if freelist is present */
815 	if (flind < 0)
816 		return (NULL);
817 
818 	vm_domain_free_assert_locked(VM_DOMAIN(domain));
819 	fl = &vm_phys_free_queues[domain][flind][pool][0];
820 	for (oind = order; oind < VM_NFREEORDER; oind++) {
821 		m = TAILQ_FIRST(&fl[oind].pl);
822 		if (m != NULL) {
823 			vm_freelist_rem(fl, m, oind);
824 			/* The order [order, oind) queues are empty. */
825 			vm_phys_split_pages(m, oind, fl, order, 1);
826 			return (m);
827 		}
828 	}
829 
830 	/*
831 	 * The given pool was empty.  Find the largest
832 	 * contiguous, power-of-two-sized set of pages in any
833 	 * pool.  Transfer these pages to the given pool, and
834 	 * use them to satisfy the allocation.
835 	 */
836 	for (oind = VM_NFREEORDER - 1; oind >= order; oind--) {
837 		for (pind = 0; pind < VM_NFREEPOOL; pind++) {
838 			alt = &vm_phys_free_queues[domain][flind][pind][0];
839 			m = TAILQ_FIRST(&alt[oind].pl);
840 			if (m != NULL) {
841 				vm_freelist_rem(alt, m, oind);
842 				vm_phys_set_pool(pool, m, oind);
843 				/* The order [order, oind) queues are empty. */
844 				vm_phys_split_pages(m, oind, fl, order, 1);
845 				return (m);
846 			}
847 		}
848 	}
849 	return (NULL);
850 }
851 
852 /*
853  * Find the vm_page corresponding to the given physical address.
854  */
855 vm_page_t
856 vm_phys_paddr_to_vm_page(vm_paddr_t pa)
857 {
858 	struct vm_phys_seg *seg;
859 	int segind;
860 
861 	for (segind = 0; segind < vm_phys_nsegs; segind++) {
862 		seg = &vm_phys_segs[segind];
863 		if (pa >= seg->start && pa < seg->end)
864 			return (&seg->first_page[atop(pa - seg->start)]);
865 	}
866 	return (NULL);
867 }
868 
869 vm_page_t
870 vm_phys_fictitious_to_vm_page(vm_paddr_t pa)
871 {
872 	struct vm_phys_fictitious_seg tmp, *seg;
873 	vm_page_t m;
874 
875 	m = NULL;
876 	tmp.start = pa;
877 	tmp.end = 0;
878 
879 	rw_rlock(&vm_phys_fictitious_reg_lock);
880 	seg = RB_FIND(fict_tree, &vm_phys_fictitious_tree, &tmp);
881 	rw_runlock(&vm_phys_fictitious_reg_lock);
882 	if (seg == NULL)
883 		return (NULL);
884 
885 	m = &seg->first_page[atop(pa - seg->start)];
886 	KASSERT((m->flags & PG_FICTITIOUS) != 0, ("%p not fictitious", m));
887 
888 	return (m);
889 }
890 
891 static inline void
892 vm_phys_fictitious_init_range(vm_page_t range, vm_paddr_t start,
893     long page_count, vm_memattr_t memattr)
894 {
895 	long i;
896 
897 	bzero(range, page_count * sizeof(*range));
898 	for (i = 0; i < page_count; i++) {
899 		vm_page_initfake(&range[i], start + PAGE_SIZE * i, memattr);
900 		range[i].oflags &= ~VPO_UNMANAGED;
901 		range[i].busy_lock = VPB_UNBUSIED;
902 	}
903 }
904 
905 int
906 vm_phys_fictitious_reg_range(vm_paddr_t start, vm_paddr_t end,
907     vm_memattr_t memattr)
908 {
909 	struct vm_phys_fictitious_seg *seg;
910 	vm_page_t fp;
911 	long page_count;
912 #ifdef VM_PHYSSEG_DENSE
913 	long pi, pe;
914 	long dpage_count;
915 #endif
916 
917 	KASSERT(start < end,
918 	    ("Start of segment isn't less than end (start: %jx end: %jx)",
919 	    (uintmax_t)start, (uintmax_t)end));
920 
921 	page_count = (end - start) / PAGE_SIZE;
922 
923 #ifdef VM_PHYSSEG_DENSE
924 	pi = atop(start);
925 	pe = atop(end);
926 	if (pi >= first_page && (pi - first_page) < vm_page_array_size) {
927 		fp = &vm_page_array[pi - first_page];
928 		if ((pe - first_page) > vm_page_array_size) {
929 			/*
930 			 * We have a segment that starts inside
931 			 * of vm_page_array, but ends outside of it.
932 			 *
933 			 * Use vm_page_array pages for those that are
934 			 * inside of the vm_page_array range, and
935 			 * allocate the remaining ones.
936 			 */
937 			dpage_count = vm_page_array_size - (pi - first_page);
938 			vm_phys_fictitious_init_range(fp, start, dpage_count,
939 			    memattr);
940 			page_count -= dpage_count;
941 			start += ptoa(dpage_count);
942 			goto alloc;
943 		}
944 		/*
945 		 * We can allocate the full range from vm_page_array,
946 		 * so there's no need to register the range in the tree.
947 		 */
948 		vm_phys_fictitious_init_range(fp, start, page_count, memattr);
949 		return (0);
950 	} else if (pe > first_page && (pe - first_page) < vm_page_array_size) {
951 		/*
952 		 * We have a segment that ends inside of vm_page_array,
953 		 * but starts outside of it.
954 		 */
955 		fp = &vm_page_array[0];
956 		dpage_count = pe - first_page;
957 		vm_phys_fictitious_init_range(fp, ptoa(first_page), dpage_count,
958 		    memattr);
959 		end -= ptoa(dpage_count);
960 		page_count -= dpage_count;
961 		goto alloc;
962 	} else if (pi < first_page && pe > (first_page + vm_page_array_size)) {
963 		/*
964 		 * Trying to register a fictitious range that expands before
965 		 * and after vm_page_array.
966 		 */
967 		return (EINVAL);
968 	} else {
969 alloc:
970 #endif
971 		fp = malloc(page_count * sizeof(struct vm_page), M_FICT_PAGES,
972 		    M_WAITOK);
973 #ifdef VM_PHYSSEG_DENSE
974 	}
975 #endif
976 	vm_phys_fictitious_init_range(fp, start, page_count, memattr);
977 
978 	seg = malloc(sizeof(*seg), M_FICT_PAGES, M_WAITOK | M_ZERO);
979 	seg->start = start;
980 	seg->end = end;
981 	seg->first_page = fp;
982 
983 	rw_wlock(&vm_phys_fictitious_reg_lock);
984 	RB_INSERT(fict_tree, &vm_phys_fictitious_tree, seg);
985 	rw_wunlock(&vm_phys_fictitious_reg_lock);
986 
987 	return (0);
988 }
989 
990 void
991 vm_phys_fictitious_unreg_range(vm_paddr_t start, vm_paddr_t end)
992 {
993 	struct vm_phys_fictitious_seg *seg, tmp;
994 #ifdef VM_PHYSSEG_DENSE
995 	long pi, pe;
996 #endif
997 
998 	KASSERT(start < end,
999 	    ("Start of segment isn't less than end (start: %jx end: %jx)",
1000 	    (uintmax_t)start, (uintmax_t)end));
1001 
1002 #ifdef VM_PHYSSEG_DENSE
1003 	pi = atop(start);
1004 	pe = atop(end);
1005 	if (pi >= first_page && (pi - first_page) < vm_page_array_size) {
1006 		if ((pe - first_page) <= vm_page_array_size) {
1007 			/*
1008 			 * This segment was allocated using vm_page_array
1009 			 * only, there's nothing to do since those pages
1010 			 * were never added to the tree.
1011 			 */
1012 			return;
1013 		}
1014 		/*
1015 		 * We have a segment that starts inside
1016 		 * of vm_page_array, but ends outside of it.
1017 		 *
1018 		 * Calculate how many pages were added to the
1019 		 * tree and free them.
1020 		 */
1021 		start = ptoa(first_page + vm_page_array_size);
1022 	} else if (pe > first_page && (pe - first_page) < vm_page_array_size) {
1023 		/*
1024 		 * We have a segment that ends inside of vm_page_array,
1025 		 * but starts outside of it.
1026 		 */
1027 		end = ptoa(first_page);
1028 	} else if (pi < first_page && pe > (first_page + vm_page_array_size)) {
1029 		/* Since it's not possible to register such a range, panic. */
1030 		panic(
1031 		    "Unregistering not registered fictitious range [%#jx:%#jx]",
1032 		    (uintmax_t)start, (uintmax_t)end);
1033 	}
1034 #endif
1035 	tmp.start = start;
1036 	tmp.end = 0;
1037 
1038 	rw_wlock(&vm_phys_fictitious_reg_lock);
1039 	seg = RB_FIND(fict_tree, &vm_phys_fictitious_tree, &tmp);
1040 	if (seg->start != start || seg->end != end) {
1041 		rw_wunlock(&vm_phys_fictitious_reg_lock);
1042 		panic(
1043 		    "Unregistering not registered fictitious range [%#jx:%#jx]",
1044 		    (uintmax_t)start, (uintmax_t)end);
1045 	}
1046 	RB_REMOVE(fict_tree, &vm_phys_fictitious_tree, seg);
1047 	rw_wunlock(&vm_phys_fictitious_reg_lock);
1048 	free(seg->first_page, M_FICT_PAGES);
1049 	free(seg, M_FICT_PAGES);
1050 }
1051 
1052 /*
1053  * Free a contiguous, power of two-sized set of physical pages.
1054  *
1055  * The free page queues must be locked.
1056  */
1057 void
1058 vm_phys_free_pages(vm_page_t m, int order)
1059 {
1060 	struct vm_freelist *fl;
1061 	struct vm_phys_seg *seg;
1062 	vm_paddr_t pa;
1063 	vm_page_t m_buddy;
1064 
1065 	KASSERT(m->order == VM_NFREEORDER,
1066 	    ("vm_phys_free_pages: page %p has unexpected order %d",
1067 	    m, m->order));
1068 	KASSERT(m->pool < VM_NFREEPOOL,
1069 	    ("vm_phys_free_pages: page %p has unexpected pool %d",
1070 	    m, m->pool));
1071 	KASSERT(order < VM_NFREEORDER,
1072 	    ("vm_phys_free_pages: order %d is out of range", order));
1073 	seg = &vm_phys_segs[m->segind];
1074 	vm_domain_free_assert_locked(VM_DOMAIN(seg->domain));
1075 	if (order < VM_NFREEORDER - 1) {
1076 		pa = VM_PAGE_TO_PHYS(m);
1077 		do {
1078 			pa ^= ((vm_paddr_t)1 << (PAGE_SHIFT + order));
1079 			if (pa < seg->start || pa >= seg->end)
1080 				break;
1081 			m_buddy = &seg->first_page[atop(pa - seg->start)];
1082 			if (m_buddy->order != order)
1083 				break;
1084 			fl = (*seg->free_queues)[m_buddy->pool];
1085 			vm_freelist_rem(fl, m_buddy, order);
1086 			if (m_buddy->pool != m->pool)
1087 				vm_phys_set_pool(m->pool, m_buddy, order);
1088 			order++;
1089 			pa &= ~(((vm_paddr_t)1 << (PAGE_SHIFT + order)) - 1);
1090 			m = &seg->first_page[atop(pa - seg->start)];
1091 		} while (order < VM_NFREEORDER - 1);
1092 	}
1093 	fl = (*seg->free_queues)[m->pool];
1094 	vm_freelist_add(fl, m, order, 1);
1095 }
1096 
1097 /*
1098  * Free a contiguous, arbitrarily sized set of physical pages.
1099  *
1100  * The free page queues must be locked.
1101  */
1102 void
1103 vm_phys_free_contig(vm_page_t m, u_long npages)
1104 {
1105 	u_int n;
1106 	int order;
1107 
1108 	/*
1109 	 * Avoid unnecessary coalescing by freeing the pages in the largest
1110 	 * possible power-of-two-sized subsets.
1111 	 */
1112 	vm_domain_free_assert_locked(vm_pagequeue_domain(m));
1113 	for (;; npages -= n) {
1114 		/*
1115 		 * Unsigned "min" is used here so that "order" is assigned
1116 		 * "VM_NFREEORDER - 1" when "m"'s physical address is zero
1117 		 * or the low-order bits of its physical address are zero
1118 		 * because the size of a physical address exceeds the size of
1119 		 * a long.
1120 		 */
1121 		order = min(ffsl(VM_PAGE_TO_PHYS(m) >> PAGE_SHIFT) - 1,
1122 		    VM_NFREEORDER - 1);
1123 		n = 1 << order;
1124 		if (npages < n)
1125 			break;
1126 		vm_phys_free_pages(m, order);
1127 		m += n;
1128 	}
1129 	/* The residual "npages" is less than "1 << (VM_NFREEORDER - 1)". */
1130 	for (; npages > 0; npages -= n) {
1131 		order = flsl(npages) - 1;
1132 		n = 1 << order;
1133 		vm_phys_free_pages(m, order);
1134 		m += n;
1135 	}
1136 }
1137 
1138 /*
1139  * Scan physical memory between the specified addresses "low" and "high" for a
1140  * run of contiguous physical pages that satisfy the specified conditions, and
1141  * return the lowest page in the run.  The specified "alignment" determines
1142  * the alignment of the lowest physical page in the run.  If the specified
1143  * "boundary" is non-zero, then the run of physical pages cannot span a
1144  * physical address that is a multiple of "boundary".
1145  *
1146  * "npages" must be greater than zero.  Both "alignment" and "boundary" must
1147  * be a power of two.
1148  */
1149 vm_page_t
1150 vm_phys_scan_contig(int domain, u_long npages, vm_paddr_t low, vm_paddr_t high,
1151     u_long alignment, vm_paddr_t boundary, int options)
1152 {
1153 	vm_paddr_t pa_end;
1154 	vm_page_t m_end, m_run, m_start;
1155 	struct vm_phys_seg *seg;
1156 	int segind;
1157 
1158 	KASSERT(npages > 0, ("npages is 0"));
1159 	KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
1160 	KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
1161 	if (low >= high)
1162 		return (NULL);
1163 	for (segind = 0; segind < vm_phys_nsegs; segind++) {
1164 		seg = &vm_phys_segs[segind];
1165 		if (seg->domain != domain)
1166 			continue;
1167 		if (seg->start >= high)
1168 			break;
1169 		if (low >= seg->end)
1170 			continue;
1171 		if (low <= seg->start)
1172 			m_start = seg->first_page;
1173 		else
1174 			m_start = &seg->first_page[atop(low - seg->start)];
1175 		if (high < seg->end)
1176 			pa_end = high;
1177 		else
1178 			pa_end = seg->end;
1179 		if (pa_end - VM_PAGE_TO_PHYS(m_start) < ptoa(npages))
1180 			continue;
1181 		m_end = &seg->first_page[atop(pa_end - seg->start)];
1182 		m_run = vm_page_scan_contig(npages, m_start, m_end,
1183 		    alignment, boundary, options);
1184 		if (m_run != NULL)
1185 			return (m_run);
1186 	}
1187 	return (NULL);
1188 }
1189 
1190 /*
1191  * Set the pool for a contiguous, power of two-sized set of physical pages.
1192  */
1193 void
1194 vm_phys_set_pool(int pool, vm_page_t m, int order)
1195 {
1196 	vm_page_t m_tmp;
1197 
1198 	for (m_tmp = m; m_tmp < &m[1 << order]; m_tmp++)
1199 		m_tmp->pool = pool;
1200 }
1201 
1202 /*
1203  * Search for the given physical page "m" in the free lists.  If the search
1204  * succeeds, remove "m" from the free lists and return TRUE.  Otherwise, return
1205  * FALSE, indicating that "m" is not in the free lists.
1206  *
1207  * The free page queues must be locked.
1208  */
1209 boolean_t
1210 vm_phys_unfree_page(vm_page_t m)
1211 {
1212 	struct vm_freelist *fl;
1213 	struct vm_phys_seg *seg;
1214 	vm_paddr_t pa, pa_half;
1215 	vm_page_t m_set, m_tmp;
1216 	int order;
1217 
1218 	/*
1219 	 * First, find the contiguous, power of two-sized set of free
1220 	 * physical pages containing the given physical page "m" and
1221 	 * assign it to "m_set".
1222 	 */
1223 	seg = &vm_phys_segs[m->segind];
1224 	vm_domain_free_assert_locked(VM_DOMAIN(seg->domain));
1225 	for (m_set = m, order = 0; m_set->order == VM_NFREEORDER &&
1226 	    order < VM_NFREEORDER - 1; ) {
1227 		order++;
1228 		pa = m->phys_addr & (~(vm_paddr_t)0 << (PAGE_SHIFT + order));
1229 		if (pa >= seg->start)
1230 			m_set = &seg->first_page[atop(pa - seg->start)];
1231 		else
1232 			return (FALSE);
1233 	}
1234 	if (m_set->order < order)
1235 		return (FALSE);
1236 	if (m_set->order == VM_NFREEORDER)
1237 		return (FALSE);
1238 	KASSERT(m_set->order < VM_NFREEORDER,
1239 	    ("vm_phys_unfree_page: page %p has unexpected order %d",
1240 	    m_set, m_set->order));
1241 
1242 	/*
1243 	 * Next, remove "m_set" from the free lists.  Finally, extract
1244 	 * "m" from "m_set" using an iterative algorithm: While "m_set"
1245 	 * is larger than a page, shrink "m_set" by returning the half
1246 	 * of "m_set" that does not contain "m" to the free lists.
1247 	 */
1248 	fl = (*seg->free_queues)[m_set->pool];
1249 	order = m_set->order;
1250 	vm_freelist_rem(fl, m_set, order);
1251 	while (order > 0) {
1252 		order--;
1253 		pa_half = m_set->phys_addr ^ (1 << (PAGE_SHIFT + order));
1254 		if (m->phys_addr < pa_half)
1255 			m_tmp = &seg->first_page[atop(pa_half - seg->start)];
1256 		else {
1257 			m_tmp = m_set;
1258 			m_set = &seg->first_page[atop(pa_half - seg->start)];
1259 		}
1260 		vm_freelist_add(fl, m_tmp, order, 0);
1261 	}
1262 	KASSERT(m_set == m, ("vm_phys_unfree_page: fatal inconsistency"));
1263 	return (TRUE);
1264 }
1265 
1266 /*
1267  * Allocate a contiguous set of physical pages of the given size
1268  * "npages" from the free lists.  All of the physical pages must be at
1269  * or above the given physical address "low" and below the given
1270  * physical address "high".  The given value "alignment" determines the
1271  * alignment of the first physical page in the set.  If the given value
1272  * "boundary" is non-zero, then the set of physical pages cannot cross
1273  * any physical address boundary that is a multiple of that value.  Both
1274  * "alignment" and "boundary" must be a power of two.
1275  */
1276 vm_page_t
1277 vm_phys_alloc_contig(int domain, u_long npages, vm_paddr_t low, vm_paddr_t high,
1278     u_long alignment, vm_paddr_t boundary)
1279 {
1280 	vm_paddr_t pa_end, pa_start;
1281 	vm_page_t m_run;
1282 	struct vm_phys_seg *seg;
1283 	int segind;
1284 
1285 	KASSERT(npages > 0, ("npages is 0"));
1286 	KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
1287 	KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
1288 	vm_domain_free_assert_locked(VM_DOMAIN(domain));
1289 	if (low >= high)
1290 		return (NULL);
1291 	m_run = NULL;
1292 	for (segind = vm_phys_nsegs - 1; segind >= 0; segind--) {
1293 		seg = &vm_phys_segs[segind];
1294 		if (seg->start >= high || seg->domain != domain)
1295 			continue;
1296 		if (low >= seg->end)
1297 			break;
1298 		if (low <= seg->start)
1299 			pa_start = seg->start;
1300 		else
1301 			pa_start = low;
1302 		if (high < seg->end)
1303 			pa_end = high;
1304 		else
1305 			pa_end = seg->end;
1306 		if (pa_end - pa_start < ptoa(npages))
1307 			continue;
1308 		m_run = vm_phys_alloc_seg_contig(seg, npages, low, high,
1309 		    alignment, boundary);
1310 		if (m_run != NULL)
1311 			break;
1312 	}
1313 	return (m_run);
1314 }
1315 
1316 /*
1317  * Allocate a run of contiguous physical pages from the free list for the
1318  * specified segment.
1319  */
1320 static vm_page_t
1321 vm_phys_alloc_seg_contig(struct vm_phys_seg *seg, u_long npages,
1322     vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary)
1323 {
1324 	struct vm_freelist *fl;
1325 	vm_paddr_t pa, pa_end, size;
1326 	vm_page_t m, m_ret;
1327 	u_long npages_end;
1328 	int oind, order, pind;
1329 
1330 	KASSERT(npages > 0, ("npages is 0"));
1331 	KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
1332 	KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
1333 	vm_domain_free_assert_locked(VM_DOMAIN(seg->domain));
1334 	/* Compute the queue that is the best fit for npages. */
1335 	order = flsl(npages - 1);
1336 	/* Search for a run satisfying the specified conditions. */
1337 	size = npages << PAGE_SHIFT;
1338 	for (oind = min(order, VM_NFREEORDER - 1); oind < VM_NFREEORDER;
1339 	    oind++) {
1340 		for (pind = 0; pind < VM_NFREEPOOL; pind++) {
1341 			fl = (*seg->free_queues)[pind];
1342 			TAILQ_FOREACH(m_ret, &fl[oind].pl, listq) {
1343 				/*
1344 				 * Is the size of this allocation request
1345 				 * larger than the largest block size?
1346 				 */
1347 				if (order >= VM_NFREEORDER) {
1348 					/*
1349 					 * Determine if a sufficient number of
1350 					 * subsequent blocks to satisfy the
1351 					 * allocation request are free.
1352 					 */
1353 					pa = VM_PAGE_TO_PHYS(m_ret);
1354 					pa_end = pa + size;
1355 					if (pa_end < pa)
1356 						continue;
1357 					for (;;) {
1358 						pa += 1 << (PAGE_SHIFT +
1359 						    VM_NFREEORDER - 1);
1360 						if (pa >= pa_end ||
1361 						    pa < seg->start ||
1362 						    pa >= seg->end)
1363 							break;
1364 						m = &seg->first_page[atop(pa -
1365 						    seg->start)];
1366 						if (m->order != VM_NFREEORDER -
1367 						    1)
1368 							break;
1369 					}
1370 					/* If not, go to the next block. */
1371 					if (pa < pa_end)
1372 						continue;
1373 				}
1374 
1375 				/*
1376 				 * Determine if the blocks are within the
1377 				 * given range, satisfy the given alignment,
1378 				 * and do not cross the given boundary.
1379 				 */
1380 				pa = VM_PAGE_TO_PHYS(m_ret);
1381 				pa_end = pa + size;
1382 				if (pa >= low && pa_end <= high &&
1383 				    (pa & (alignment - 1)) == 0 &&
1384 				    rounddown2(pa ^ (pa_end - 1), boundary) == 0)
1385 					goto done;
1386 			}
1387 		}
1388 	}
1389 	return (NULL);
1390 done:
1391 	for (m = m_ret; m < &m_ret[npages]; m = &m[1 << oind]) {
1392 		fl = (*seg->free_queues)[m->pool];
1393 		vm_freelist_rem(fl, m, oind);
1394 		if (m->pool != VM_FREEPOOL_DEFAULT)
1395 			vm_phys_set_pool(VM_FREEPOOL_DEFAULT, m, oind);
1396 	}
1397 	/* Return excess pages to the free lists. */
1398 	npages_end = roundup2(npages, 1 << oind);
1399 	if (npages < npages_end) {
1400 		fl = (*seg->free_queues)[VM_FREEPOOL_DEFAULT];
1401 		vm_phys_enq_range(&m_ret[npages], npages_end - npages, fl, 0);
1402 	}
1403 	return (m_ret);
1404 }
1405 
1406 #ifdef DDB
1407 /*
1408  * Show the number of physical pages in each of the free lists.
1409  */
1410 DB_SHOW_COMMAND(freepages, db_show_freepages)
1411 {
1412 	struct vm_freelist *fl;
1413 	int flind, oind, pind, dom;
1414 
1415 	for (dom = 0; dom < vm_ndomains; dom++) {
1416 		db_printf("DOMAIN: %d\n", dom);
1417 		for (flind = 0; flind < vm_nfreelists; flind++) {
1418 			db_printf("FREE LIST %d:\n"
1419 			    "\n  ORDER (SIZE)  |  NUMBER"
1420 			    "\n              ", flind);
1421 			for (pind = 0; pind < VM_NFREEPOOL; pind++)
1422 				db_printf("  |  POOL %d", pind);
1423 			db_printf("\n--            ");
1424 			for (pind = 0; pind < VM_NFREEPOOL; pind++)
1425 				db_printf("-- --      ");
1426 			db_printf("--\n");
1427 			for (oind = VM_NFREEORDER - 1; oind >= 0; oind--) {
1428 				db_printf("  %2.2d (%6.6dK)", oind,
1429 				    1 << (PAGE_SHIFT - 10 + oind));
1430 				for (pind = 0; pind < VM_NFREEPOOL; pind++) {
1431 				fl = vm_phys_free_queues[dom][flind][pind];
1432 					db_printf("  |  %6.6d", fl[oind].lcnt);
1433 				}
1434 				db_printf("\n");
1435 			}
1436 			db_printf("\n");
1437 		}
1438 		db_printf("\n");
1439 	}
1440 }
1441 #endif
1442