xref: /freebsd/sys/vm/vm_phys.c (revision eda14cbc264d6969b02f2b1994cef11148e914f1)
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 static struct vm_phys_seg vm_phys_early_segs[8];
86 static int vm_phys_early_nsegs;
87 
88 struct vm_phys_fictitious_seg;
89 static int vm_phys_fictitious_cmp(struct vm_phys_fictitious_seg *,
90     struct vm_phys_fictitious_seg *);
91 
92 RB_HEAD(fict_tree, vm_phys_fictitious_seg) vm_phys_fictitious_tree =
93     RB_INITIALIZER(&vm_phys_fictitious_tree);
94 
95 struct vm_phys_fictitious_seg {
96 	RB_ENTRY(vm_phys_fictitious_seg) node;
97 	/* Memory region data */
98 	vm_paddr_t	start;
99 	vm_paddr_t	end;
100 	vm_page_t	first_page;
101 };
102 
103 RB_GENERATE_STATIC(fict_tree, vm_phys_fictitious_seg, node,
104     vm_phys_fictitious_cmp);
105 
106 static struct rwlock_padalign vm_phys_fictitious_reg_lock;
107 MALLOC_DEFINE(M_FICT_PAGES, "vm_fictitious", "Fictitious VM pages");
108 
109 static struct vm_freelist __aligned(CACHE_LINE_SIZE)
110     vm_phys_free_queues[MAXMEMDOM][VM_NFREELIST][VM_NFREEPOOL]
111     [VM_NFREEORDER_MAX];
112 
113 static int __read_mostly vm_nfreelists;
114 
115 /*
116  * These "avail lists" are globals used to communicate boot-time physical
117  * memory layout to other parts of the kernel.  Each physically contiguous
118  * region of memory is defined by a start address at an even index and an
119  * end address at the following odd index.  Each list is terminated by a
120  * pair of zero entries.
121  *
122  * dump_avail tells the dump code what regions to include in a crash dump, and
123  * phys_avail is all of the remaining physical memory that is available for
124  * the vm system.
125  *
126  * Initially dump_avail and phys_avail are identical.  Boot time memory
127  * allocations remove extents from phys_avail that may still be included
128  * in dumps.
129  */
130 vm_paddr_t phys_avail[PHYS_AVAIL_COUNT];
131 vm_paddr_t dump_avail[PHYS_AVAIL_COUNT];
132 
133 /*
134  * Provides the mapping from VM_FREELIST_* to free list indices (flind).
135  */
136 static int __read_mostly vm_freelist_to_flind[VM_NFREELIST];
137 
138 CTASSERT(VM_FREELIST_DEFAULT == 0);
139 
140 #ifdef VM_FREELIST_DMA32
141 #define	VM_DMA32_BOUNDARY	((vm_paddr_t)1 << 32)
142 #endif
143 
144 /*
145  * Enforce the assumptions made by vm_phys_add_seg() and vm_phys_init() about
146  * the ordering of the free list boundaries.
147  */
148 #if defined(VM_LOWMEM_BOUNDARY) && defined(VM_DMA32_BOUNDARY)
149 CTASSERT(VM_LOWMEM_BOUNDARY < VM_DMA32_BOUNDARY);
150 #endif
151 
152 static int sysctl_vm_phys_free(SYSCTL_HANDLER_ARGS);
153 SYSCTL_OID(_vm, OID_AUTO, phys_free,
154     CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_NEEDGIANT, NULL, 0,
155     sysctl_vm_phys_free, "A",
156     "Phys Free Info");
157 
158 static int sysctl_vm_phys_segs(SYSCTL_HANDLER_ARGS);
159 SYSCTL_OID(_vm, OID_AUTO, phys_segs,
160     CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_NEEDGIANT, NULL, 0,
161     sysctl_vm_phys_segs, "A",
162     "Phys Seg Info");
163 
164 #ifdef NUMA
165 static int sysctl_vm_phys_locality(SYSCTL_HANDLER_ARGS);
166 SYSCTL_OID(_vm, OID_AUTO, phys_locality,
167     CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_NEEDGIANT, NULL, 0,
168     sysctl_vm_phys_locality, "A",
169     "Phys Locality Info");
170 #endif
171 
172 SYSCTL_INT(_vm, OID_AUTO, ndomains, CTLFLAG_RD,
173     &vm_ndomains, 0, "Number of physical memory domains available.");
174 
175 static vm_page_t vm_phys_alloc_seg_contig(struct vm_phys_seg *seg,
176     u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment,
177     vm_paddr_t boundary);
178 static void _vm_phys_create_seg(vm_paddr_t start, vm_paddr_t end, int domain);
179 static void vm_phys_create_seg(vm_paddr_t start, vm_paddr_t end);
180 static void vm_phys_split_pages(vm_page_t m, int oind, struct vm_freelist *fl,
181     int order, int tail);
182 
183 /*
184  * Red-black tree helpers for vm fictitious range management.
185  */
186 static inline int
187 vm_phys_fictitious_in_range(struct vm_phys_fictitious_seg *p,
188     struct vm_phys_fictitious_seg *range)
189 {
190 
191 	KASSERT(range->start != 0 && range->end != 0,
192 	    ("Invalid range passed on search for vm_fictitious page"));
193 	if (p->start >= range->end)
194 		return (1);
195 	if (p->start < range->start)
196 		return (-1);
197 
198 	return (0);
199 }
200 
201 static int
202 vm_phys_fictitious_cmp(struct vm_phys_fictitious_seg *p1,
203     struct vm_phys_fictitious_seg *p2)
204 {
205 
206 	/* Check if this is a search for a page */
207 	if (p1->end == 0)
208 		return (vm_phys_fictitious_in_range(p1, p2));
209 
210 	KASSERT(p2->end != 0,
211     ("Invalid range passed as second parameter to vm fictitious comparison"));
212 
213 	/* Searching to add a new range */
214 	if (p1->end <= p2->start)
215 		return (-1);
216 	if (p1->start >= p2->end)
217 		return (1);
218 
219 	panic("Trying to add overlapping vm fictitious ranges:\n"
220 	    "[%#jx:%#jx] and [%#jx:%#jx]", (uintmax_t)p1->start,
221 	    (uintmax_t)p1->end, (uintmax_t)p2->start, (uintmax_t)p2->end);
222 }
223 
224 int
225 vm_phys_domain_match(int prefer, vm_paddr_t low, vm_paddr_t high)
226 {
227 #ifdef NUMA
228 	domainset_t mask;
229 	int i;
230 
231 	if (vm_ndomains == 1 || mem_affinity == NULL)
232 		return (0);
233 
234 	DOMAINSET_ZERO(&mask);
235 	/*
236 	 * Check for any memory that overlaps low, high.
237 	 */
238 	for (i = 0; mem_affinity[i].end != 0; i++)
239 		if (mem_affinity[i].start <= high &&
240 		    mem_affinity[i].end >= low)
241 			DOMAINSET_SET(mem_affinity[i].domain, &mask);
242 	if (prefer != -1 && DOMAINSET_ISSET(prefer, &mask))
243 		return (prefer);
244 	if (DOMAINSET_EMPTY(&mask))
245 		panic("vm_phys_domain_match:  Impossible constraint");
246 	return (DOMAINSET_FFS(&mask) - 1);
247 #else
248 	return (0);
249 #endif
250 }
251 
252 /*
253  * Outputs the state of the physical memory allocator, specifically,
254  * the amount of physical memory in each free list.
255  */
256 static int
257 sysctl_vm_phys_free(SYSCTL_HANDLER_ARGS)
258 {
259 	struct sbuf sbuf;
260 	struct vm_freelist *fl;
261 	int dom, error, flind, oind, pind;
262 
263 	error = sysctl_wire_old_buffer(req, 0);
264 	if (error != 0)
265 		return (error);
266 	sbuf_new_for_sysctl(&sbuf, NULL, 128 * vm_ndomains, req);
267 	for (dom = 0; dom < vm_ndomains; dom++) {
268 		sbuf_printf(&sbuf,"\nDOMAIN %d:\n", dom);
269 		for (flind = 0; flind < vm_nfreelists; flind++) {
270 			sbuf_printf(&sbuf, "\nFREE LIST %d:\n"
271 			    "\n  ORDER (SIZE)  |  NUMBER"
272 			    "\n              ", flind);
273 			for (pind = 0; pind < VM_NFREEPOOL; pind++)
274 				sbuf_printf(&sbuf, "  |  POOL %d", pind);
275 			sbuf_printf(&sbuf, "\n--            ");
276 			for (pind = 0; pind < VM_NFREEPOOL; pind++)
277 				sbuf_printf(&sbuf, "-- --      ");
278 			sbuf_printf(&sbuf, "--\n");
279 			for (oind = VM_NFREEORDER - 1; oind >= 0; oind--) {
280 				sbuf_printf(&sbuf, "  %2d (%6dK)", oind,
281 				    1 << (PAGE_SHIFT - 10 + oind));
282 				for (pind = 0; pind < VM_NFREEPOOL; pind++) {
283 				fl = vm_phys_free_queues[dom][flind][pind];
284 					sbuf_printf(&sbuf, "  |  %6d",
285 					    fl[oind].lcnt);
286 				}
287 				sbuf_printf(&sbuf, "\n");
288 			}
289 		}
290 	}
291 	error = sbuf_finish(&sbuf);
292 	sbuf_delete(&sbuf);
293 	return (error);
294 }
295 
296 /*
297  * Outputs the set of physical memory segments.
298  */
299 static int
300 sysctl_vm_phys_segs(SYSCTL_HANDLER_ARGS)
301 {
302 	struct sbuf sbuf;
303 	struct vm_phys_seg *seg;
304 	int error, segind;
305 
306 	error = sysctl_wire_old_buffer(req, 0);
307 	if (error != 0)
308 		return (error);
309 	sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
310 	for (segind = 0; segind < vm_phys_nsegs; segind++) {
311 		sbuf_printf(&sbuf, "\nSEGMENT %d:\n\n", segind);
312 		seg = &vm_phys_segs[segind];
313 		sbuf_printf(&sbuf, "start:     %#jx\n",
314 		    (uintmax_t)seg->start);
315 		sbuf_printf(&sbuf, "end:       %#jx\n",
316 		    (uintmax_t)seg->end);
317 		sbuf_printf(&sbuf, "domain:    %d\n", seg->domain);
318 		sbuf_printf(&sbuf, "free list: %p\n", seg->free_queues);
319 	}
320 	error = sbuf_finish(&sbuf);
321 	sbuf_delete(&sbuf);
322 	return (error);
323 }
324 
325 /*
326  * Return affinity, or -1 if there's no affinity information.
327  */
328 int
329 vm_phys_mem_affinity(int f, int t)
330 {
331 
332 #ifdef NUMA
333 	if (mem_locality == NULL)
334 		return (-1);
335 	if (f >= vm_ndomains || t >= vm_ndomains)
336 		return (-1);
337 	return (mem_locality[f * vm_ndomains + t]);
338 #else
339 	return (-1);
340 #endif
341 }
342 
343 #ifdef NUMA
344 /*
345  * Outputs the VM locality table.
346  */
347 static int
348 sysctl_vm_phys_locality(SYSCTL_HANDLER_ARGS)
349 {
350 	struct sbuf sbuf;
351 	int error, i, j;
352 
353 	error = sysctl_wire_old_buffer(req, 0);
354 	if (error != 0)
355 		return (error);
356 	sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
357 
358 	sbuf_printf(&sbuf, "\n");
359 
360 	for (i = 0; i < vm_ndomains; i++) {
361 		sbuf_printf(&sbuf, "%d: ", i);
362 		for (j = 0; j < vm_ndomains; j++) {
363 			sbuf_printf(&sbuf, "%d ", vm_phys_mem_affinity(i, j));
364 		}
365 		sbuf_printf(&sbuf, "\n");
366 	}
367 	error = sbuf_finish(&sbuf);
368 	sbuf_delete(&sbuf);
369 	return (error);
370 }
371 #endif
372 
373 static void
374 vm_freelist_add(struct vm_freelist *fl, vm_page_t m, int order, int tail)
375 {
376 
377 	m->order = order;
378 	if (tail)
379 		TAILQ_INSERT_TAIL(&fl[order].pl, m, listq);
380 	else
381 		TAILQ_INSERT_HEAD(&fl[order].pl, m, listq);
382 	fl[order].lcnt++;
383 }
384 
385 static void
386 vm_freelist_rem(struct vm_freelist *fl, vm_page_t m, int order)
387 {
388 
389 	TAILQ_REMOVE(&fl[order].pl, m, listq);
390 	fl[order].lcnt--;
391 	m->order = VM_NFREEORDER;
392 }
393 
394 /*
395  * Create a physical memory segment.
396  */
397 static void
398 _vm_phys_create_seg(vm_paddr_t start, vm_paddr_t end, int domain)
399 {
400 	struct vm_phys_seg *seg;
401 
402 	KASSERT(vm_phys_nsegs < VM_PHYSSEG_MAX,
403 	    ("vm_phys_create_seg: increase VM_PHYSSEG_MAX"));
404 	KASSERT(domain >= 0 && domain < vm_ndomains,
405 	    ("vm_phys_create_seg: invalid domain provided"));
406 	seg = &vm_phys_segs[vm_phys_nsegs++];
407 	while (seg > vm_phys_segs && (seg - 1)->start >= end) {
408 		*seg = *(seg - 1);
409 		seg--;
410 	}
411 	seg->start = start;
412 	seg->end = end;
413 	seg->domain = domain;
414 }
415 
416 static void
417 vm_phys_create_seg(vm_paddr_t start, vm_paddr_t end)
418 {
419 #ifdef NUMA
420 	int i;
421 
422 	if (mem_affinity == NULL) {
423 		_vm_phys_create_seg(start, end, 0);
424 		return;
425 	}
426 
427 	for (i = 0;; i++) {
428 		if (mem_affinity[i].end == 0)
429 			panic("Reached end of affinity info");
430 		if (mem_affinity[i].end <= start)
431 			continue;
432 		if (mem_affinity[i].start > start)
433 			panic("No affinity info for start %jx",
434 			    (uintmax_t)start);
435 		if (mem_affinity[i].end >= end) {
436 			_vm_phys_create_seg(start, end,
437 			    mem_affinity[i].domain);
438 			break;
439 		}
440 		_vm_phys_create_seg(start, mem_affinity[i].end,
441 		    mem_affinity[i].domain);
442 		start = mem_affinity[i].end;
443 	}
444 #else
445 	_vm_phys_create_seg(start, end, 0);
446 #endif
447 }
448 
449 /*
450  * Add a physical memory segment.
451  */
452 void
453 vm_phys_add_seg(vm_paddr_t start, vm_paddr_t end)
454 {
455 	vm_paddr_t paddr;
456 
457 	KASSERT((start & PAGE_MASK) == 0,
458 	    ("vm_phys_define_seg: start is not page aligned"));
459 	KASSERT((end & PAGE_MASK) == 0,
460 	    ("vm_phys_define_seg: end is not page aligned"));
461 
462 	/*
463 	 * Split the physical memory segment if it spans two or more free
464 	 * list boundaries.
465 	 */
466 	paddr = start;
467 #ifdef	VM_FREELIST_LOWMEM
468 	if (paddr < VM_LOWMEM_BOUNDARY && end > VM_LOWMEM_BOUNDARY) {
469 		vm_phys_create_seg(paddr, VM_LOWMEM_BOUNDARY);
470 		paddr = VM_LOWMEM_BOUNDARY;
471 	}
472 #endif
473 #ifdef	VM_FREELIST_DMA32
474 	if (paddr < VM_DMA32_BOUNDARY && end > VM_DMA32_BOUNDARY) {
475 		vm_phys_create_seg(paddr, VM_DMA32_BOUNDARY);
476 		paddr = VM_DMA32_BOUNDARY;
477 	}
478 #endif
479 	vm_phys_create_seg(paddr, end);
480 }
481 
482 /*
483  * Initialize the physical memory allocator.
484  *
485  * Requires that vm_page_array is initialized!
486  */
487 void
488 vm_phys_init(void)
489 {
490 	struct vm_freelist *fl;
491 	struct vm_phys_seg *end_seg, *prev_seg, *seg, *tmp_seg;
492 	u_long npages;
493 	int dom, flind, freelist, oind, pind, segind;
494 
495 	/*
496 	 * Compute the number of free lists, and generate the mapping from the
497 	 * manifest constants VM_FREELIST_* to the free list indices.
498 	 *
499 	 * Initially, the entries of vm_freelist_to_flind[] are set to either
500 	 * 0 or 1 to indicate which free lists should be created.
501 	 */
502 	npages = 0;
503 	for (segind = vm_phys_nsegs - 1; segind >= 0; segind--) {
504 		seg = &vm_phys_segs[segind];
505 #ifdef	VM_FREELIST_LOWMEM
506 		if (seg->end <= VM_LOWMEM_BOUNDARY)
507 			vm_freelist_to_flind[VM_FREELIST_LOWMEM] = 1;
508 		else
509 #endif
510 #ifdef	VM_FREELIST_DMA32
511 		if (
512 #ifdef	VM_DMA32_NPAGES_THRESHOLD
513 		    /*
514 		     * Create the DMA32 free list only if the amount of
515 		     * physical memory above physical address 4G exceeds the
516 		     * given threshold.
517 		     */
518 		    npages > VM_DMA32_NPAGES_THRESHOLD &&
519 #endif
520 		    seg->end <= VM_DMA32_BOUNDARY)
521 			vm_freelist_to_flind[VM_FREELIST_DMA32] = 1;
522 		else
523 #endif
524 		{
525 			npages += atop(seg->end - seg->start);
526 			vm_freelist_to_flind[VM_FREELIST_DEFAULT] = 1;
527 		}
528 	}
529 	/* Change each entry into a running total of the free lists. */
530 	for (freelist = 1; freelist < VM_NFREELIST; freelist++) {
531 		vm_freelist_to_flind[freelist] +=
532 		    vm_freelist_to_flind[freelist - 1];
533 	}
534 	vm_nfreelists = vm_freelist_to_flind[VM_NFREELIST - 1];
535 	KASSERT(vm_nfreelists > 0, ("vm_phys_init: no free lists"));
536 	/* Change each entry into a free list index. */
537 	for (freelist = 0; freelist < VM_NFREELIST; freelist++)
538 		vm_freelist_to_flind[freelist]--;
539 
540 	/*
541 	 * Initialize the first_page and free_queues fields of each physical
542 	 * memory segment.
543 	 */
544 #ifdef VM_PHYSSEG_SPARSE
545 	npages = 0;
546 #endif
547 	for (segind = 0; segind < vm_phys_nsegs; segind++) {
548 		seg = &vm_phys_segs[segind];
549 #ifdef VM_PHYSSEG_SPARSE
550 		seg->first_page = &vm_page_array[npages];
551 		npages += atop(seg->end - seg->start);
552 #else
553 		seg->first_page = PHYS_TO_VM_PAGE(seg->start);
554 #endif
555 #ifdef	VM_FREELIST_LOWMEM
556 		if (seg->end <= VM_LOWMEM_BOUNDARY) {
557 			flind = vm_freelist_to_flind[VM_FREELIST_LOWMEM];
558 			KASSERT(flind >= 0,
559 			    ("vm_phys_init: LOWMEM flind < 0"));
560 		} else
561 #endif
562 #ifdef	VM_FREELIST_DMA32
563 		if (seg->end <= VM_DMA32_BOUNDARY) {
564 			flind = vm_freelist_to_flind[VM_FREELIST_DMA32];
565 			KASSERT(flind >= 0,
566 			    ("vm_phys_init: DMA32 flind < 0"));
567 		} else
568 #endif
569 		{
570 			flind = vm_freelist_to_flind[VM_FREELIST_DEFAULT];
571 			KASSERT(flind >= 0,
572 			    ("vm_phys_init: DEFAULT flind < 0"));
573 		}
574 		seg->free_queues = &vm_phys_free_queues[seg->domain][flind];
575 	}
576 
577 	/*
578 	 * Coalesce physical memory segments that are contiguous and share the
579 	 * same per-domain free queues.
580 	 */
581 	prev_seg = vm_phys_segs;
582 	seg = &vm_phys_segs[1];
583 	end_seg = &vm_phys_segs[vm_phys_nsegs];
584 	while (seg < end_seg) {
585 		if (prev_seg->end == seg->start &&
586 		    prev_seg->free_queues == seg->free_queues) {
587 			prev_seg->end = seg->end;
588 			KASSERT(prev_seg->domain == seg->domain,
589 			    ("vm_phys_init: free queues cannot span domains"));
590 			vm_phys_nsegs--;
591 			end_seg--;
592 			for (tmp_seg = seg; tmp_seg < end_seg; tmp_seg++)
593 				*tmp_seg = *(tmp_seg + 1);
594 		} else {
595 			prev_seg = seg;
596 			seg++;
597 		}
598 	}
599 
600 	/*
601 	 * Initialize the free queues.
602 	 */
603 	for (dom = 0; dom < vm_ndomains; dom++) {
604 		for (flind = 0; flind < vm_nfreelists; flind++) {
605 			for (pind = 0; pind < VM_NFREEPOOL; pind++) {
606 				fl = vm_phys_free_queues[dom][flind][pind];
607 				for (oind = 0; oind < VM_NFREEORDER; oind++)
608 					TAILQ_INIT(&fl[oind].pl);
609 			}
610 		}
611 	}
612 
613 	rw_init(&vm_phys_fictitious_reg_lock, "vmfctr");
614 }
615 
616 /*
617  * Register info about the NUMA topology of the system.
618  *
619  * Invoked by platform-dependent code prior to vm_phys_init().
620  */
621 void
622 vm_phys_register_domains(int ndomains, struct mem_affinity *affinity,
623     int *locality)
624 {
625 #ifdef NUMA
626 	int d, i;
627 
628 	/*
629 	 * For now the only override value that we support is 1, which
630 	 * effectively disables NUMA-awareness in the allocators.
631 	 */
632 	d = 0;
633 	TUNABLE_INT_FETCH("vm.numa.disabled", &d);
634 	if (d)
635 		ndomains = 1;
636 
637 	if (ndomains > 1) {
638 		vm_ndomains = ndomains;
639 		mem_affinity = affinity;
640 		mem_locality = locality;
641 	}
642 
643 	for (i = 0; i < vm_ndomains; i++)
644 		DOMAINSET_SET(i, &all_domains);
645 #else
646 	(void)ndomains;
647 	(void)affinity;
648 	(void)locality;
649 #endif
650 }
651 
652 int
653 _vm_phys_domain(vm_paddr_t pa)
654 {
655 #ifdef NUMA
656 	int i;
657 
658 	if (vm_ndomains == 1)
659 		return (0);
660 	for (i = 0; mem_affinity[i].end != 0; i++)
661 		if (mem_affinity[i].start <= pa &&
662 		    mem_affinity[i].end >= pa)
663 			return (mem_affinity[i].domain);
664 	return (-1);
665 #else
666 	return (0);
667 #endif
668 }
669 
670 /*
671  * Split a contiguous, power of two-sized set of physical pages.
672  *
673  * When this function is called by a page allocation function, the caller
674  * should request insertion at the head unless the order [order, oind) queues
675  * are known to be empty.  The objective being to reduce the likelihood of
676  * long-term fragmentation by promoting contemporaneous allocation and
677  * (hopefully) deallocation.
678  */
679 static __inline void
680 vm_phys_split_pages(vm_page_t m, int oind, struct vm_freelist *fl, int order,
681     int tail)
682 {
683 	vm_page_t m_buddy;
684 
685 	while (oind > order) {
686 		oind--;
687 		m_buddy = &m[1 << oind];
688 		KASSERT(m_buddy->order == VM_NFREEORDER,
689 		    ("vm_phys_split_pages: page %p has unexpected order %d",
690 		    m_buddy, m_buddy->order));
691 		vm_freelist_add(fl, m_buddy, oind, tail);
692         }
693 }
694 
695 /*
696  * Add the physical pages [m, m + npages) at the end of a power-of-two aligned
697  * and sized set to the specified free list.
698  *
699  * When this function is called by a page allocation function, the caller
700  * should request insertion at the head unless the lower-order queues are
701  * known to be empty.  The objective being to reduce the likelihood of long-
702  * term fragmentation by promoting contemporaneous allocation and (hopefully)
703  * deallocation.
704  *
705  * The physical page m's buddy must not be free.
706  */
707 static void
708 vm_phys_enq_range(vm_page_t m, u_int npages, struct vm_freelist *fl, int tail)
709 {
710 	u_int n;
711 	int order;
712 
713 	KASSERT(npages > 0, ("vm_phys_enq_range: npages is 0"));
714 	KASSERT(((VM_PAGE_TO_PHYS(m) + npages * PAGE_SIZE) &
715 	    ((PAGE_SIZE << (fls(npages) - 1)) - 1)) == 0,
716 	    ("vm_phys_enq_range: page %p and npages %u are misaligned",
717 	    m, npages));
718 	do {
719 		KASSERT(m->order == VM_NFREEORDER,
720 		    ("vm_phys_enq_range: page %p has unexpected order %d",
721 		    m, m->order));
722 		order = ffs(npages) - 1;
723 		KASSERT(order < VM_NFREEORDER,
724 		    ("vm_phys_enq_range: order %d is out of range", order));
725 		vm_freelist_add(fl, m, order, tail);
726 		n = 1 << order;
727 		m += n;
728 		npages -= n;
729 	} while (npages > 0);
730 }
731 
732 /*
733  * Tries to allocate the specified number of pages from the specified pool
734  * within the specified domain.  Returns the actual number of allocated pages
735  * and a pointer to each page through the array ma[].
736  *
737  * The returned pages may not be physically contiguous.  However, in contrast
738  * to performing multiple, back-to-back calls to vm_phys_alloc_pages(..., 0),
739  * calling this function once to allocate the desired number of pages will
740  * avoid wasted time in vm_phys_split_pages().
741  *
742  * The free page queues for the specified domain must be locked.
743  */
744 int
745 vm_phys_alloc_npages(int domain, int pool, int npages, vm_page_t ma[])
746 {
747 	struct vm_freelist *alt, *fl;
748 	vm_page_t m;
749 	int avail, end, flind, freelist, i, need, oind, pind;
750 
751 	KASSERT(domain >= 0 && domain < vm_ndomains,
752 	    ("vm_phys_alloc_npages: domain %d is out of range", domain));
753 	KASSERT(pool < VM_NFREEPOOL,
754 	    ("vm_phys_alloc_npages: pool %d is out of range", pool));
755 	KASSERT(npages <= 1 << (VM_NFREEORDER - 1),
756 	    ("vm_phys_alloc_npages: npages %d is out of range", npages));
757 	vm_domain_free_assert_locked(VM_DOMAIN(domain));
758 	i = 0;
759 	for (freelist = 0; freelist < VM_NFREELIST; freelist++) {
760 		flind = vm_freelist_to_flind[freelist];
761 		if (flind < 0)
762 			continue;
763 		fl = vm_phys_free_queues[domain][flind][pool];
764 		for (oind = 0; oind < VM_NFREEORDER; oind++) {
765 			while ((m = TAILQ_FIRST(&fl[oind].pl)) != NULL) {
766 				vm_freelist_rem(fl, m, oind);
767 				avail = 1 << oind;
768 				need = imin(npages - i, avail);
769 				for (end = i + need; i < end;)
770 					ma[i++] = m++;
771 				if (need < avail) {
772 					/*
773 					 * Return excess pages to fl.  Its
774 					 * order [0, oind) queues are empty.
775 					 */
776 					vm_phys_enq_range(m, avail - need, fl,
777 					    1);
778 					return (npages);
779 				} else if (i == npages)
780 					return (npages);
781 			}
782 		}
783 		for (oind = VM_NFREEORDER - 1; oind >= 0; oind--) {
784 			for (pind = 0; pind < VM_NFREEPOOL; pind++) {
785 				alt = vm_phys_free_queues[domain][flind][pind];
786 				while ((m = TAILQ_FIRST(&alt[oind].pl)) !=
787 				    NULL) {
788 					vm_freelist_rem(alt, m, oind);
789 					vm_phys_set_pool(pool, m, oind);
790 					avail = 1 << oind;
791 					need = imin(npages - i, avail);
792 					for (end = i + need; i < end;)
793 						ma[i++] = m++;
794 					if (need < avail) {
795 						/*
796 						 * Return excess pages to fl.
797 						 * Its order [0, oind) queues
798 						 * are empty.
799 						 */
800 						vm_phys_enq_range(m, avail -
801 						    need, fl, 1);
802 						return (npages);
803 					} else if (i == npages)
804 						return (npages);
805 				}
806 			}
807 		}
808 	}
809 	return (i);
810 }
811 
812 /*
813  * Allocate a contiguous, power of two-sized set of physical pages
814  * from the free lists.
815  *
816  * The free page queues must be locked.
817  */
818 vm_page_t
819 vm_phys_alloc_pages(int domain, int pool, int order)
820 {
821 	vm_page_t m;
822 	int freelist;
823 
824 	for (freelist = 0; freelist < VM_NFREELIST; freelist++) {
825 		m = vm_phys_alloc_freelist_pages(domain, freelist, pool, order);
826 		if (m != NULL)
827 			return (m);
828 	}
829 	return (NULL);
830 }
831 
832 /*
833  * Allocate a contiguous, power of two-sized set of physical pages from the
834  * specified free list.  The free list must be specified using one of the
835  * manifest constants VM_FREELIST_*.
836  *
837  * The free page queues must be locked.
838  */
839 vm_page_t
840 vm_phys_alloc_freelist_pages(int domain, int freelist, int pool, int order)
841 {
842 	struct vm_freelist *alt, *fl;
843 	vm_page_t m;
844 	int oind, pind, flind;
845 
846 	KASSERT(domain >= 0 && domain < vm_ndomains,
847 	    ("vm_phys_alloc_freelist_pages: domain %d is out of range",
848 	    domain));
849 	KASSERT(freelist < VM_NFREELIST,
850 	    ("vm_phys_alloc_freelist_pages: freelist %d is out of range",
851 	    freelist));
852 	KASSERT(pool < VM_NFREEPOOL,
853 	    ("vm_phys_alloc_freelist_pages: pool %d is out of range", pool));
854 	KASSERT(order < VM_NFREEORDER,
855 	    ("vm_phys_alloc_freelist_pages: order %d is out of range", order));
856 
857 	flind = vm_freelist_to_flind[freelist];
858 	/* Check if freelist is present */
859 	if (flind < 0)
860 		return (NULL);
861 
862 	vm_domain_free_assert_locked(VM_DOMAIN(domain));
863 	fl = &vm_phys_free_queues[domain][flind][pool][0];
864 	for (oind = order; oind < VM_NFREEORDER; oind++) {
865 		m = TAILQ_FIRST(&fl[oind].pl);
866 		if (m != NULL) {
867 			vm_freelist_rem(fl, m, oind);
868 			/* The order [order, oind) queues are empty. */
869 			vm_phys_split_pages(m, oind, fl, order, 1);
870 			return (m);
871 		}
872 	}
873 
874 	/*
875 	 * The given pool was empty.  Find the largest
876 	 * contiguous, power-of-two-sized set of pages in any
877 	 * pool.  Transfer these pages to the given pool, and
878 	 * use them to satisfy the allocation.
879 	 */
880 	for (oind = VM_NFREEORDER - 1; oind >= order; oind--) {
881 		for (pind = 0; pind < VM_NFREEPOOL; pind++) {
882 			alt = &vm_phys_free_queues[domain][flind][pind][0];
883 			m = TAILQ_FIRST(&alt[oind].pl);
884 			if (m != NULL) {
885 				vm_freelist_rem(alt, m, oind);
886 				vm_phys_set_pool(pool, m, oind);
887 				/* The order [order, oind) queues are empty. */
888 				vm_phys_split_pages(m, oind, fl, order, 1);
889 				return (m);
890 			}
891 		}
892 	}
893 	return (NULL);
894 }
895 
896 /*
897  * Find the vm_page corresponding to the given physical address.
898  */
899 vm_page_t
900 vm_phys_paddr_to_vm_page(vm_paddr_t pa)
901 {
902 	struct vm_phys_seg *seg;
903 	int segind;
904 
905 	for (segind = 0; segind < vm_phys_nsegs; segind++) {
906 		seg = &vm_phys_segs[segind];
907 		if (pa >= seg->start && pa < seg->end)
908 			return (&seg->first_page[atop(pa - seg->start)]);
909 	}
910 	return (NULL);
911 }
912 
913 vm_page_t
914 vm_phys_fictitious_to_vm_page(vm_paddr_t pa)
915 {
916 	struct vm_phys_fictitious_seg tmp, *seg;
917 	vm_page_t m;
918 
919 	m = NULL;
920 	tmp.start = pa;
921 	tmp.end = 0;
922 
923 	rw_rlock(&vm_phys_fictitious_reg_lock);
924 	seg = RB_FIND(fict_tree, &vm_phys_fictitious_tree, &tmp);
925 	rw_runlock(&vm_phys_fictitious_reg_lock);
926 	if (seg == NULL)
927 		return (NULL);
928 
929 	m = &seg->first_page[atop(pa - seg->start)];
930 	KASSERT((m->flags & PG_FICTITIOUS) != 0, ("%p not fictitious", m));
931 
932 	return (m);
933 }
934 
935 static inline void
936 vm_phys_fictitious_init_range(vm_page_t range, vm_paddr_t start,
937     long page_count, vm_memattr_t memattr)
938 {
939 	long i;
940 
941 	bzero(range, page_count * sizeof(*range));
942 	for (i = 0; i < page_count; i++) {
943 		vm_page_initfake(&range[i], start + PAGE_SIZE * i, memattr);
944 		range[i].oflags &= ~VPO_UNMANAGED;
945 		range[i].busy_lock = VPB_UNBUSIED;
946 	}
947 }
948 
949 int
950 vm_phys_fictitious_reg_range(vm_paddr_t start, vm_paddr_t end,
951     vm_memattr_t memattr)
952 {
953 	struct vm_phys_fictitious_seg *seg;
954 	vm_page_t fp;
955 	long page_count;
956 #ifdef VM_PHYSSEG_DENSE
957 	long pi, pe;
958 	long dpage_count;
959 #endif
960 
961 	KASSERT(start < end,
962 	    ("Start of segment isn't less than end (start: %jx end: %jx)",
963 	    (uintmax_t)start, (uintmax_t)end));
964 
965 	page_count = (end - start) / PAGE_SIZE;
966 
967 #ifdef VM_PHYSSEG_DENSE
968 	pi = atop(start);
969 	pe = atop(end);
970 	if (pi >= first_page && (pi - first_page) < vm_page_array_size) {
971 		fp = &vm_page_array[pi - first_page];
972 		if ((pe - first_page) > vm_page_array_size) {
973 			/*
974 			 * We have a segment that starts inside
975 			 * of vm_page_array, but ends outside of it.
976 			 *
977 			 * Use vm_page_array pages for those that are
978 			 * inside of the vm_page_array range, and
979 			 * allocate the remaining ones.
980 			 */
981 			dpage_count = vm_page_array_size - (pi - first_page);
982 			vm_phys_fictitious_init_range(fp, start, dpage_count,
983 			    memattr);
984 			page_count -= dpage_count;
985 			start += ptoa(dpage_count);
986 			goto alloc;
987 		}
988 		/*
989 		 * We can allocate the full range from vm_page_array,
990 		 * so there's no need to register the range in the tree.
991 		 */
992 		vm_phys_fictitious_init_range(fp, start, page_count, memattr);
993 		return (0);
994 	} else if (pe > first_page && (pe - first_page) < vm_page_array_size) {
995 		/*
996 		 * We have a segment that ends inside of vm_page_array,
997 		 * but starts outside of it.
998 		 */
999 		fp = &vm_page_array[0];
1000 		dpage_count = pe - first_page;
1001 		vm_phys_fictitious_init_range(fp, ptoa(first_page), dpage_count,
1002 		    memattr);
1003 		end -= ptoa(dpage_count);
1004 		page_count -= dpage_count;
1005 		goto alloc;
1006 	} else if (pi < first_page && pe > (first_page + vm_page_array_size)) {
1007 		/*
1008 		 * Trying to register a fictitious range that expands before
1009 		 * and after vm_page_array.
1010 		 */
1011 		return (EINVAL);
1012 	} else {
1013 alloc:
1014 #endif
1015 		fp = malloc(page_count * sizeof(struct vm_page), M_FICT_PAGES,
1016 		    M_WAITOK);
1017 #ifdef VM_PHYSSEG_DENSE
1018 	}
1019 #endif
1020 	vm_phys_fictitious_init_range(fp, start, page_count, memattr);
1021 
1022 	seg = malloc(sizeof(*seg), M_FICT_PAGES, M_WAITOK | M_ZERO);
1023 	seg->start = start;
1024 	seg->end = end;
1025 	seg->first_page = fp;
1026 
1027 	rw_wlock(&vm_phys_fictitious_reg_lock);
1028 	RB_INSERT(fict_tree, &vm_phys_fictitious_tree, seg);
1029 	rw_wunlock(&vm_phys_fictitious_reg_lock);
1030 
1031 	return (0);
1032 }
1033 
1034 void
1035 vm_phys_fictitious_unreg_range(vm_paddr_t start, vm_paddr_t end)
1036 {
1037 	struct vm_phys_fictitious_seg *seg, tmp;
1038 #ifdef VM_PHYSSEG_DENSE
1039 	long pi, pe;
1040 #endif
1041 
1042 	KASSERT(start < end,
1043 	    ("Start of segment isn't less than end (start: %jx end: %jx)",
1044 	    (uintmax_t)start, (uintmax_t)end));
1045 
1046 #ifdef VM_PHYSSEG_DENSE
1047 	pi = atop(start);
1048 	pe = atop(end);
1049 	if (pi >= first_page && (pi - first_page) < vm_page_array_size) {
1050 		if ((pe - first_page) <= vm_page_array_size) {
1051 			/*
1052 			 * This segment was allocated using vm_page_array
1053 			 * only, there's nothing to do since those pages
1054 			 * were never added to the tree.
1055 			 */
1056 			return;
1057 		}
1058 		/*
1059 		 * We have a segment that starts inside
1060 		 * of vm_page_array, but ends outside of it.
1061 		 *
1062 		 * Calculate how many pages were added to the
1063 		 * tree and free them.
1064 		 */
1065 		start = ptoa(first_page + vm_page_array_size);
1066 	} else if (pe > first_page && (pe - first_page) < vm_page_array_size) {
1067 		/*
1068 		 * We have a segment that ends inside of vm_page_array,
1069 		 * but starts outside of it.
1070 		 */
1071 		end = ptoa(first_page);
1072 	} else if (pi < first_page && pe > (first_page + vm_page_array_size)) {
1073 		/* Since it's not possible to register such a range, panic. */
1074 		panic(
1075 		    "Unregistering not registered fictitious range [%#jx:%#jx]",
1076 		    (uintmax_t)start, (uintmax_t)end);
1077 	}
1078 #endif
1079 	tmp.start = start;
1080 	tmp.end = 0;
1081 
1082 	rw_wlock(&vm_phys_fictitious_reg_lock);
1083 	seg = RB_FIND(fict_tree, &vm_phys_fictitious_tree, &tmp);
1084 	if (seg->start != start || seg->end != end) {
1085 		rw_wunlock(&vm_phys_fictitious_reg_lock);
1086 		panic(
1087 		    "Unregistering not registered fictitious range [%#jx:%#jx]",
1088 		    (uintmax_t)start, (uintmax_t)end);
1089 	}
1090 	RB_REMOVE(fict_tree, &vm_phys_fictitious_tree, seg);
1091 	rw_wunlock(&vm_phys_fictitious_reg_lock);
1092 	free(seg->first_page, M_FICT_PAGES);
1093 	free(seg, M_FICT_PAGES);
1094 }
1095 
1096 /*
1097  * Free a contiguous, power of two-sized set of physical pages.
1098  *
1099  * The free page queues must be locked.
1100  */
1101 void
1102 vm_phys_free_pages(vm_page_t m, int order)
1103 {
1104 	struct vm_freelist *fl;
1105 	struct vm_phys_seg *seg;
1106 	vm_paddr_t pa;
1107 	vm_page_t m_buddy;
1108 
1109 	KASSERT(m->order == VM_NFREEORDER,
1110 	    ("vm_phys_free_pages: page %p has unexpected order %d",
1111 	    m, m->order));
1112 	KASSERT(m->pool < VM_NFREEPOOL,
1113 	    ("vm_phys_free_pages: page %p has unexpected pool %d",
1114 	    m, m->pool));
1115 	KASSERT(order < VM_NFREEORDER,
1116 	    ("vm_phys_free_pages: order %d is out of range", order));
1117 	seg = &vm_phys_segs[m->segind];
1118 	vm_domain_free_assert_locked(VM_DOMAIN(seg->domain));
1119 	if (order < VM_NFREEORDER - 1) {
1120 		pa = VM_PAGE_TO_PHYS(m);
1121 		do {
1122 			pa ^= ((vm_paddr_t)1 << (PAGE_SHIFT + order));
1123 			if (pa < seg->start || pa >= seg->end)
1124 				break;
1125 			m_buddy = &seg->first_page[atop(pa - seg->start)];
1126 			if (m_buddy->order != order)
1127 				break;
1128 			fl = (*seg->free_queues)[m_buddy->pool];
1129 			vm_freelist_rem(fl, m_buddy, order);
1130 			if (m_buddy->pool != m->pool)
1131 				vm_phys_set_pool(m->pool, m_buddy, order);
1132 			order++;
1133 			pa &= ~(((vm_paddr_t)1 << (PAGE_SHIFT + order)) - 1);
1134 			m = &seg->first_page[atop(pa - seg->start)];
1135 		} while (order < VM_NFREEORDER - 1);
1136 	}
1137 	fl = (*seg->free_queues)[m->pool];
1138 	vm_freelist_add(fl, m, order, 1);
1139 }
1140 
1141 /*
1142  * Return the largest possible order of a set of pages starting at m.
1143  */
1144 static int
1145 max_order(vm_page_t m)
1146 {
1147 
1148 	/*
1149 	 * Unsigned "min" is used here so that "order" is assigned
1150 	 * "VM_NFREEORDER - 1" when "m"'s physical address is zero
1151 	 * or the low-order bits of its physical address are zero
1152 	 * because the size of a physical address exceeds the size of
1153 	 * a long.
1154 	 */
1155 	return (min(ffsl(VM_PAGE_TO_PHYS(m) >> PAGE_SHIFT) - 1,
1156 	    VM_NFREEORDER - 1));
1157 }
1158 
1159 /*
1160  * Free a contiguous, arbitrarily sized set of physical pages, without
1161  * merging across set boundaries.
1162  *
1163  * The free page queues must be locked.
1164  */
1165 void
1166 vm_phys_enqueue_contig(vm_page_t m, u_long npages)
1167 {
1168 	struct vm_freelist *fl;
1169 	struct vm_phys_seg *seg;
1170 	vm_page_t m_end;
1171 	int order;
1172 
1173 	/*
1174 	 * Avoid unnecessary coalescing by freeing the pages in the largest
1175 	 * possible power-of-two-sized subsets.
1176 	 */
1177 	vm_domain_free_assert_locked(vm_pagequeue_domain(m));
1178 	seg = &vm_phys_segs[m->segind];
1179 	fl = (*seg->free_queues)[m->pool];
1180 	m_end = m + npages;
1181 	/* Free blocks of increasing size. */
1182 	while ((order = max_order(m)) < VM_NFREEORDER - 1 &&
1183 	    m + (1 << order) <= m_end) {
1184 		KASSERT(seg == &vm_phys_segs[m->segind],
1185 		    ("%s: page range [%p,%p) spans multiple segments",
1186 		    __func__, m_end - npages, m));
1187 		vm_freelist_add(fl, m, order, 1);
1188 		m += 1 << order;
1189 	}
1190 	/* Free blocks of maximum size. */
1191 	while (m + (1 << order) <= m_end) {
1192 		KASSERT(seg == &vm_phys_segs[m->segind],
1193 		    ("%s: page range [%p,%p) spans multiple segments",
1194 		    __func__, m_end - npages, m));
1195 		vm_freelist_add(fl, m, order, 1);
1196 		m += 1 << order;
1197 	}
1198 	/* Free blocks of diminishing size. */
1199 	while (m < m_end) {
1200 		KASSERT(seg == &vm_phys_segs[m->segind],
1201 		    ("%s: page range [%p,%p) spans multiple segments",
1202 		    __func__, m_end - npages, m));
1203 		order = flsl(m_end - m) - 1;
1204 		vm_freelist_add(fl, m, order, 1);
1205 		m += 1 << order;
1206 	}
1207 }
1208 
1209 /*
1210  * Free a contiguous, arbitrarily sized set of physical pages.
1211  *
1212  * The free page queues must be locked.
1213  */
1214 void
1215 vm_phys_free_contig(vm_page_t m, u_long npages)
1216 {
1217 	int order_start, order_end;
1218 	vm_page_t m_start, m_end;
1219 
1220 	vm_domain_free_assert_locked(vm_pagequeue_domain(m));
1221 
1222 	m_start = m;
1223 	order_start = max_order(m_start);
1224 	if (order_start < VM_NFREEORDER - 1)
1225 		m_start += 1 << order_start;
1226 	m_end = m + npages;
1227 	order_end = max_order(m_end);
1228 	if (order_end < VM_NFREEORDER - 1)
1229 		m_end -= 1 << order_end;
1230 	/*
1231 	 * Avoid unnecessary coalescing by freeing the pages at the start and
1232 	 * end of the range last.
1233 	 */
1234 	if (m_start < m_end)
1235 		vm_phys_enqueue_contig(m_start, m_end - m_start);
1236 	if (order_start < VM_NFREEORDER - 1)
1237 		vm_phys_free_pages(m, order_start);
1238 	if (order_end < VM_NFREEORDER - 1)
1239 		vm_phys_free_pages(m_end, order_end);
1240 }
1241 
1242 /*
1243  * Scan physical memory between the specified addresses "low" and "high" for a
1244  * run of contiguous physical pages that satisfy the specified conditions, and
1245  * return the lowest page in the run.  The specified "alignment" determines
1246  * the alignment of the lowest physical page in the run.  If the specified
1247  * "boundary" is non-zero, then the run of physical pages cannot span a
1248  * physical address that is a multiple of "boundary".
1249  *
1250  * "npages" must be greater than zero.  Both "alignment" and "boundary" must
1251  * be a power of two.
1252  */
1253 vm_page_t
1254 vm_phys_scan_contig(int domain, u_long npages, vm_paddr_t low, vm_paddr_t high,
1255     u_long alignment, vm_paddr_t boundary, int options)
1256 {
1257 	vm_paddr_t pa_end;
1258 	vm_page_t m_end, m_run, m_start;
1259 	struct vm_phys_seg *seg;
1260 	int segind;
1261 
1262 	KASSERT(npages > 0, ("npages is 0"));
1263 	KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
1264 	KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
1265 	if (low >= high)
1266 		return (NULL);
1267 	for (segind = 0; segind < vm_phys_nsegs; segind++) {
1268 		seg = &vm_phys_segs[segind];
1269 		if (seg->domain != domain)
1270 			continue;
1271 		if (seg->start >= high)
1272 			break;
1273 		if (low >= seg->end)
1274 			continue;
1275 		if (low <= seg->start)
1276 			m_start = seg->first_page;
1277 		else
1278 			m_start = &seg->first_page[atop(low - seg->start)];
1279 		if (high < seg->end)
1280 			pa_end = high;
1281 		else
1282 			pa_end = seg->end;
1283 		if (pa_end - VM_PAGE_TO_PHYS(m_start) < ptoa(npages))
1284 			continue;
1285 		m_end = &seg->first_page[atop(pa_end - seg->start)];
1286 		m_run = vm_page_scan_contig(npages, m_start, m_end,
1287 		    alignment, boundary, options);
1288 		if (m_run != NULL)
1289 			return (m_run);
1290 	}
1291 	return (NULL);
1292 }
1293 
1294 /*
1295  * Set the pool for a contiguous, power of two-sized set of physical pages.
1296  */
1297 void
1298 vm_phys_set_pool(int pool, vm_page_t m, int order)
1299 {
1300 	vm_page_t m_tmp;
1301 
1302 	for (m_tmp = m; m_tmp < &m[1 << order]; m_tmp++)
1303 		m_tmp->pool = pool;
1304 }
1305 
1306 /*
1307  * Search for the given physical page "m" in the free lists.  If the search
1308  * succeeds, remove "m" from the free lists and return TRUE.  Otherwise, return
1309  * FALSE, indicating that "m" is not in the free lists.
1310  *
1311  * The free page queues must be locked.
1312  */
1313 boolean_t
1314 vm_phys_unfree_page(vm_page_t m)
1315 {
1316 	struct vm_freelist *fl;
1317 	struct vm_phys_seg *seg;
1318 	vm_paddr_t pa, pa_half;
1319 	vm_page_t m_set, m_tmp;
1320 	int order;
1321 
1322 	/*
1323 	 * First, find the contiguous, power of two-sized set of free
1324 	 * physical pages containing the given physical page "m" and
1325 	 * assign it to "m_set".
1326 	 */
1327 	seg = &vm_phys_segs[m->segind];
1328 	vm_domain_free_assert_locked(VM_DOMAIN(seg->domain));
1329 	for (m_set = m, order = 0; m_set->order == VM_NFREEORDER &&
1330 	    order < VM_NFREEORDER - 1; ) {
1331 		order++;
1332 		pa = m->phys_addr & (~(vm_paddr_t)0 << (PAGE_SHIFT + order));
1333 		if (pa >= seg->start)
1334 			m_set = &seg->first_page[atop(pa - seg->start)];
1335 		else
1336 			return (FALSE);
1337 	}
1338 	if (m_set->order < order)
1339 		return (FALSE);
1340 	if (m_set->order == VM_NFREEORDER)
1341 		return (FALSE);
1342 	KASSERT(m_set->order < VM_NFREEORDER,
1343 	    ("vm_phys_unfree_page: page %p has unexpected order %d",
1344 	    m_set, m_set->order));
1345 
1346 	/*
1347 	 * Next, remove "m_set" from the free lists.  Finally, extract
1348 	 * "m" from "m_set" using an iterative algorithm: While "m_set"
1349 	 * is larger than a page, shrink "m_set" by returning the half
1350 	 * of "m_set" that does not contain "m" to the free lists.
1351 	 */
1352 	fl = (*seg->free_queues)[m_set->pool];
1353 	order = m_set->order;
1354 	vm_freelist_rem(fl, m_set, order);
1355 	while (order > 0) {
1356 		order--;
1357 		pa_half = m_set->phys_addr ^ (1 << (PAGE_SHIFT + order));
1358 		if (m->phys_addr < pa_half)
1359 			m_tmp = &seg->first_page[atop(pa_half - seg->start)];
1360 		else {
1361 			m_tmp = m_set;
1362 			m_set = &seg->first_page[atop(pa_half - seg->start)];
1363 		}
1364 		vm_freelist_add(fl, m_tmp, order, 0);
1365 	}
1366 	KASSERT(m_set == m, ("vm_phys_unfree_page: fatal inconsistency"));
1367 	return (TRUE);
1368 }
1369 
1370 /*
1371  * Allocate a contiguous set of physical pages of the given size
1372  * "npages" from the free lists.  All of the physical pages must be at
1373  * or above the given physical address "low" and below the given
1374  * physical address "high".  The given value "alignment" determines the
1375  * alignment of the first physical page in the set.  If the given value
1376  * "boundary" is non-zero, then the set of physical pages cannot cross
1377  * any physical address boundary that is a multiple of that value.  Both
1378  * "alignment" and "boundary" must be a power of two.
1379  */
1380 vm_page_t
1381 vm_phys_alloc_contig(int domain, u_long npages, vm_paddr_t low, vm_paddr_t high,
1382     u_long alignment, vm_paddr_t boundary)
1383 {
1384 	vm_paddr_t pa_end, pa_start;
1385 	vm_page_t m_run;
1386 	struct vm_phys_seg *seg;
1387 	int segind;
1388 
1389 	KASSERT(npages > 0, ("npages is 0"));
1390 	KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
1391 	KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
1392 	vm_domain_free_assert_locked(VM_DOMAIN(domain));
1393 	if (low >= high)
1394 		return (NULL);
1395 	m_run = NULL;
1396 	for (segind = vm_phys_nsegs - 1; segind >= 0; segind--) {
1397 		seg = &vm_phys_segs[segind];
1398 		if (seg->start >= high || seg->domain != domain)
1399 			continue;
1400 		if (low >= seg->end)
1401 			break;
1402 		if (low <= seg->start)
1403 			pa_start = seg->start;
1404 		else
1405 			pa_start = low;
1406 		if (high < seg->end)
1407 			pa_end = high;
1408 		else
1409 			pa_end = seg->end;
1410 		if (pa_end - pa_start < ptoa(npages))
1411 			continue;
1412 		m_run = vm_phys_alloc_seg_contig(seg, npages, low, high,
1413 		    alignment, boundary);
1414 		if (m_run != NULL)
1415 			break;
1416 	}
1417 	return (m_run);
1418 }
1419 
1420 /*
1421  * Allocate a run of contiguous physical pages from the free list for the
1422  * specified segment.
1423  */
1424 static vm_page_t
1425 vm_phys_alloc_seg_contig(struct vm_phys_seg *seg, u_long npages,
1426     vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary)
1427 {
1428 	struct vm_freelist *fl;
1429 	vm_paddr_t pa, pa_end, size;
1430 	vm_page_t m, m_ret;
1431 	u_long npages_end;
1432 	int oind, order, pind;
1433 
1434 	KASSERT(npages > 0, ("npages is 0"));
1435 	KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
1436 	KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
1437 	vm_domain_free_assert_locked(VM_DOMAIN(seg->domain));
1438 	/* Compute the queue that is the best fit for npages. */
1439 	order = flsl(npages - 1);
1440 	/* Search for a run satisfying the specified conditions. */
1441 	size = npages << PAGE_SHIFT;
1442 	for (oind = min(order, VM_NFREEORDER - 1); oind < VM_NFREEORDER;
1443 	    oind++) {
1444 		for (pind = 0; pind < VM_NFREEPOOL; pind++) {
1445 			fl = (*seg->free_queues)[pind];
1446 			TAILQ_FOREACH(m_ret, &fl[oind].pl, listq) {
1447 				/*
1448 				 * Is the size of this allocation request
1449 				 * larger than the largest block size?
1450 				 */
1451 				if (order >= VM_NFREEORDER) {
1452 					/*
1453 					 * Determine if a sufficient number of
1454 					 * subsequent blocks to satisfy the
1455 					 * allocation request are free.
1456 					 */
1457 					pa = VM_PAGE_TO_PHYS(m_ret);
1458 					pa_end = pa + size;
1459 					if (pa_end < pa)
1460 						continue;
1461 					for (;;) {
1462 						pa += 1 << (PAGE_SHIFT +
1463 						    VM_NFREEORDER - 1);
1464 						if (pa >= pa_end ||
1465 						    pa < seg->start ||
1466 						    pa >= seg->end)
1467 							break;
1468 						m = &seg->first_page[atop(pa -
1469 						    seg->start)];
1470 						if (m->order != VM_NFREEORDER -
1471 						    1)
1472 							break;
1473 					}
1474 					/* If not, go to the next block. */
1475 					if (pa < pa_end)
1476 						continue;
1477 				}
1478 
1479 				/*
1480 				 * Determine if the blocks are within the
1481 				 * given range, satisfy the given alignment,
1482 				 * and do not cross the given boundary.
1483 				 */
1484 				pa = VM_PAGE_TO_PHYS(m_ret);
1485 				pa_end = pa + size;
1486 				if (pa >= low && pa_end <= high &&
1487 				    (pa & (alignment - 1)) == 0 &&
1488 				    rounddown2(pa ^ (pa_end - 1), boundary) == 0)
1489 					goto done;
1490 			}
1491 		}
1492 	}
1493 	return (NULL);
1494 done:
1495 	for (m = m_ret; m < &m_ret[npages]; m = &m[1 << oind]) {
1496 		fl = (*seg->free_queues)[m->pool];
1497 		vm_freelist_rem(fl, m, oind);
1498 		if (m->pool != VM_FREEPOOL_DEFAULT)
1499 			vm_phys_set_pool(VM_FREEPOOL_DEFAULT, m, oind);
1500 	}
1501 	/* Return excess pages to the free lists. */
1502 	npages_end = roundup2(npages, 1 << oind);
1503 	if (npages < npages_end) {
1504 		fl = (*seg->free_queues)[VM_FREEPOOL_DEFAULT];
1505 		vm_phys_enq_range(&m_ret[npages], npages_end - npages, fl, 0);
1506 	}
1507 	return (m_ret);
1508 }
1509 
1510 /*
1511  * Return the index of the first unused slot which may be the terminating
1512  * entry.
1513  */
1514 static int
1515 vm_phys_avail_count(void)
1516 {
1517 	int i;
1518 
1519 	for (i = 0; phys_avail[i + 1]; i += 2)
1520 		continue;
1521 	if (i > PHYS_AVAIL_ENTRIES)
1522 		panic("Improperly terminated phys_avail %d entries", i);
1523 
1524 	return (i);
1525 }
1526 
1527 /*
1528  * Assert that a phys_avail entry is valid.
1529  */
1530 static void
1531 vm_phys_avail_check(int i)
1532 {
1533 	if (phys_avail[i] & PAGE_MASK)
1534 		panic("Unaligned phys_avail[%d]: %#jx", i,
1535 		    (intmax_t)phys_avail[i]);
1536 	if (phys_avail[i+1] & PAGE_MASK)
1537 		panic("Unaligned phys_avail[%d + 1]: %#jx", i,
1538 		    (intmax_t)phys_avail[i]);
1539 	if (phys_avail[i + 1] < phys_avail[i])
1540 		panic("phys_avail[%d] start %#jx < end %#jx", i,
1541 		    (intmax_t)phys_avail[i], (intmax_t)phys_avail[i+1]);
1542 }
1543 
1544 /*
1545  * Return the index of an overlapping phys_avail entry or -1.
1546  */
1547 #ifdef NUMA
1548 static int
1549 vm_phys_avail_find(vm_paddr_t pa)
1550 {
1551 	int i;
1552 
1553 	for (i = 0; phys_avail[i + 1]; i += 2)
1554 		if (phys_avail[i] <= pa && phys_avail[i + 1] > pa)
1555 			return (i);
1556 	return (-1);
1557 }
1558 #endif
1559 
1560 /*
1561  * Return the index of the largest entry.
1562  */
1563 int
1564 vm_phys_avail_largest(void)
1565 {
1566 	vm_paddr_t sz, largesz;
1567 	int largest;
1568 	int i;
1569 
1570 	largest = 0;
1571 	largesz = 0;
1572 	for (i = 0; phys_avail[i + 1]; i += 2) {
1573 		sz = vm_phys_avail_size(i);
1574 		if (sz > largesz) {
1575 			largesz = sz;
1576 			largest = i;
1577 		}
1578 	}
1579 
1580 	return (largest);
1581 }
1582 
1583 vm_paddr_t
1584 vm_phys_avail_size(int i)
1585 {
1586 
1587 	return (phys_avail[i + 1] - phys_avail[i]);
1588 }
1589 
1590 /*
1591  * Split an entry at the address 'pa'.  Return zero on success or errno.
1592  */
1593 static int
1594 vm_phys_avail_split(vm_paddr_t pa, int i)
1595 {
1596 	int cnt;
1597 
1598 	vm_phys_avail_check(i);
1599 	if (pa <= phys_avail[i] || pa >= phys_avail[i + 1])
1600 		panic("vm_phys_avail_split: invalid address");
1601 	cnt = vm_phys_avail_count();
1602 	if (cnt >= PHYS_AVAIL_ENTRIES)
1603 		return (ENOSPC);
1604 	memmove(&phys_avail[i + 2], &phys_avail[i],
1605 	    (cnt - i) * sizeof(phys_avail[0]));
1606 	phys_avail[i + 1] = pa;
1607 	phys_avail[i + 2] = pa;
1608 	vm_phys_avail_check(i);
1609 	vm_phys_avail_check(i+2);
1610 
1611 	return (0);
1612 }
1613 
1614 void
1615 vm_phys_early_add_seg(vm_paddr_t start, vm_paddr_t end)
1616 {
1617 	struct vm_phys_seg *seg;
1618 
1619 	if (vm_phys_early_nsegs == -1)
1620 		panic("%s: called after initialization", __func__);
1621 	if (vm_phys_early_nsegs == nitems(vm_phys_early_segs))
1622 		panic("%s: ran out of early segments", __func__);
1623 
1624 	seg = &vm_phys_early_segs[vm_phys_early_nsegs++];
1625 	seg->start = start;
1626 	seg->end = end;
1627 }
1628 
1629 /*
1630  * This routine allocates NUMA node specific memory before the page
1631  * allocator is bootstrapped.
1632  */
1633 vm_paddr_t
1634 vm_phys_early_alloc(int domain, size_t alloc_size)
1635 {
1636 	int i, mem_index, biggestone;
1637 	vm_paddr_t pa, mem_start, mem_end, size, biggestsize, align;
1638 
1639 	KASSERT(domain == -1 || (domain >= 0 && domain < vm_ndomains),
1640 	    ("%s: invalid domain index %d", __func__, domain));
1641 
1642 	/*
1643 	 * Search the mem_affinity array for the biggest address
1644 	 * range in the desired domain.  This is used to constrain
1645 	 * the phys_avail selection below.
1646 	 */
1647 	biggestsize = 0;
1648 	mem_index = 0;
1649 	mem_start = 0;
1650 	mem_end = -1;
1651 #ifdef NUMA
1652 	if (mem_affinity != NULL) {
1653 		for (i = 0;; i++) {
1654 			size = mem_affinity[i].end - mem_affinity[i].start;
1655 			if (size == 0)
1656 				break;
1657 			if (domain != -1 && mem_affinity[i].domain != domain)
1658 				continue;
1659 			if (size > biggestsize) {
1660 				mem_index = i;
1661 				biggestsize = size;
1662 			}
1663 		}
1664 		mem_start = mem_affinity[mem_index].start;
1665 		mem_end = mem_affinity[mem_index].end;
1666 	}
1667 #endif
1668 
1669 	/*
1670 	 * Now find biggest physical segment in within the desired
1671 	 * numa domain.
1672 	 */
1673 	biggestsize = 0;
1674 	biggestone = 0;
1675 	for (i = 0; phys_avail[i + 1] != 0; i += 2) {
1676 		/* skip regions that are out of range */
1677 		if (phys_avail[i+1] - alloc_size < mem_start ||
1678 		    phys_avail[i+1] > mem_end)
1679 			continue;
1680 		size = vm_phys_avail_size(i);
1681 		if (size > biggestsize) {
1682 			biggestone = i;
1683 			biggestsize = size;
1684 		}
1685 	}
1686 	alloc_size = round_page(alloc_size);
1687 
1688 	/*
1689 	 * Grab single pages from the front to reduce fragmentation.
1690 	 */
1691 	if (alloc_size == PAGE_SIZE) {
1692 		pa = phys_avail[biggestone];
1693 		phys_avail[biggestone] += PAGE_SIZE;
1694 		vm_phys_avail_check(biggestone);
1695 		return (pa);
1696 	}
1697 
1698 	/*
1699 	 * Naturally align large allocations.
1700 	 */
1701 	align = phys_avail[biggestone + 1] & (alloc_size - 1);
1702 	if (alloc_size + align > biggestsize)
1703 		panic("cannot find a large enough size\n");
1704 	if (align != 0 &&
1705 	    vm_phys_avail_split(phys_avail[biggestone + 1] - align,
1706 	    biggestone) != 0)
1707 		/* Wasting memory. */
1708 		phys_avail[biggestone + 1] -= align;
1709 
1710 	phys_avail[biggestone + 1] -= alloc_size;
1711 	vm_phys_avail_check(biggestone);
1712 	pa = phys_avail[biggestone + 1];
1713 	return (pa);
1714 }
1715 
1716 void
1717 vm_phys_early_startup(void)
1718 {
1719 	struct vm_phys_seg *seg;
1720 	int i;
1721 
1722 	for (i = 0; phys_avail[i + 1] != 0; i += 2) {
1723 		phys_avail[i] = round_page(phys_avail[i]);
1724 		phys_avail[i + 1] = trunc_page(phys_avail[i + 1]);
1725 	}
1726 
1727 	for (i = 0; i < vm_phys_early_nsegs; i++) {
1728 		seg = &vm_phys_early_segs[i];
1729 		vm_phys_add_seg(seg->start, seg->end);
1730 	}
1731 	vm_phys_early_nsegs = -1;
1732 
1733 #ifdef NUMA
1734 	/* Force phys_avail to be split by domain. */
1735 	if (mem_affinity != NULL) {
1736 		int idx;
1737 
1738 		for (i = 0; mem_affinity[i].end != 0; i++) {
1739 			idx = vm_phys_avail_find(mem_affinity[i].start);
1740 			if (idx != -1 &&
1741 			    phys_avail[idx] != mem_affinity[i].start)
1742 				vm_phys_avail_split(mem_affinity[i].start, idx);
1743 			idx = vm_phys_avail_find(mem_affinity[i].end);
1744 			if (idx != -1 &&
1745 			    phys_avail[idx] != mem_affinity[i].end)
1746 				vm_phys_avail_split(mem_affinity[i].end, idx);
1747 		}
1748 	}
1749 #endif
1750 }
1751 
1752 #ifdef DDB
1753 /*
1754  * Show the number of physical pages in each of the free lists.
1755  */
1756 DB_SHOW_COMMAND(freepages, db_show_freepages)
1757 {
1758 	struct vm_freelist *fl;
1759 	int flind, oind, pind, dom;
1760 
1761 	for (dom = 0; dom < vm_ndomains; dom++) {
1762 		db_printf("DOMAIN: %d\n", dom);
1763 		for (flind = 0; flind < vm_nfreelists; flind++) {
1764 			db_printf("FREE LIST %d:\n"
1765 			    "\n  ORDER (SIZE)  |  NUMBER"
1766 			    "\n              ", flind);
1767 			for (pind = 0; pind < VM_NFREEPOOL; pind++)
1768 				db_printf("  |  POOL %d", pind);
1769 			db_printf("\n--            ");
1770 			for (pind = 0; pind < VM_NFREEPOOL; pind++)
1771 				db_printf("-- --      ");
1772 			db_printf("--\n");
1773 			for (oind = VM_NFREEORDER - 1; oind >= 0; oind--) {
1774 				db_printf("  %2.2d (%6.6dK)", oind,
1775 				    1 << (PAGE_SHIFT - 10 + oind));
1776 				for (pind = 0; pind < VM_NFREEPOOL; pind++) {
1777 				fl = vm_phys_free_queues[dom][flind][pind];
1778 					db_printf("  |  %6.6d", fl[oind].lcnt);
1779 				}
1780 				db_printf("\n");
1781 			}
1782 			db_printf("\n");
1783 		}
1784 		db_printf("\n");
1785 	}
1786 }
1787 #endif
1788