xref: /freebsd/sys/vm/vm_phys.c (revision 8df8b2d3e51d1b816201d8a1fe8bc29fe192e562)
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 #include <sys/seq.h>
62 
63 #include <ddb/ddb.h>
64 
65 #include <vm/vm.h>
66 #include <vm/vm_param.h>
67 #include <vm/vm_kern.h>
68 #include <vm/vm_object.h>
69 #include <vm/vm_page.h>
70 #include <vm/vm_phys.h>
71 #include <vm/vm_pagequeue.h>
72 
73 _Static_assert(sizeof(long) * NBBY >= VM_PHYSSEG_MAX,
74     "Too many physsegs.");
75 
76 #ifdef NUMA
77 struct mem_affinity __read_mostly *mem_affinity;
78 int __read_mostly *mem_locality;
79 #endif
80 
81 int __read_mostly vm_ndomains = 1;
82 domainset_t __read_mostly all_domains = DOMAINSET_T_INITIALIZER(0x1);
83 
84 struct vm_phys_seg __read_mostly vm_phys_segs[VM_PHYSSEG_MAX];
85 int __read_mostly vm_phys_nsegs;
86 
87 struct vm_phys_fictitious_seg;
88 static int vm_phys_fictitious_cmp(struct vm_phys_fictitious_seg *,
89     struct vm_phys_fictitious_seg *);
90 
91 RB_HEAD(fict_tree, vm_phys_fictitious_seg) vm_phys_fictitious_tree =
92     RB_INITIALIZER(_vm_phys_fictitious_tree);
93 
94 struct vm_phys_fictitious_seg {
95 	RB_ENTRY(vm_phys_fictitious_seg) node;
96 	/* Memory region data */
97 	vm_paddr_t	start;
98 	vm_paddr_t	end;
99 	vm_page_t	first_page;
100 };
101 
102 RB_GENERATE_STATIC(fict_tree, vm_phys_fictitious_seg, node,
103     vm_phys_fictitious_cmp);
104 
105 static struct rwlock_padalign vm_phys_fictitious_reg_lock;
106 MALLOC_DEFINE(M_FICT_PAGES, "vm_fictitious", "Fictitious VM pages");
107 
108 static struct vm_freelist __aligned(CACHE_LINE_SIZE)
109     vm_phys_free_queues[MAXMEMDOM][VM_NFREELIST][VM_NFREEPOOL]
110     [VM_NFREEORDER_MAX];
111 
112 static int __read_mostly vm_nfreelists;
113 
114 /*
115  * Provides the mapping from VM_FREELIST_* to free list indices (flind).
116  */
117 static int __read_mostly vm_freelist_to_flind[VM_NFREELIST];
118 
119 CTASSERT(VM_FREELIST_DEFAULT == 0);
120 
121 #ifdef VM_FREELIST_DMA32
122 #define	VM_DMA32_BOUNDARY	((vm_paddr_t)1 << 32)
123 #endif
124 
125 /*
126  * Enforce the assumptions made by vm_phys_add_seg() and vm_phys_init() about
127  * the ordering of the free list boundaries.
128  */
129 #if defined(VM_LOWMEM_BOUNDARY) && defined(VM_DMA32_BOUNDARY)
130 CTASSERT(VM_LOWMEM_BOUNDARY < VM_DMA32_BOUNDARY);
131 #endif
132 
133 static int sysctl_vm_phys_free(SYSCTL_HANDLER_ARGS);
134 SYSCTL_OID(_vm, OID_AUTO, phys_free, CTLTYPE_STRING | CTLFLAG_RD,
135     NULL, 0, sysctl_vm_phys_free, "A", "Phys Free Info");
136 
137 static int sysctl_vm_phys_segs(SYSCTL_HANDLER_ARGS);
138 SYSCTL_OID(_vm, OID_AUTO, phys_segs, CTLTYPE_STRING | CTLFLAG_RD,
139     NULL, 0, sysctl_vm_phys_segs, "A", "Phys Seg Info");
140 
141 #ifdef NUMA
142 static int sysctl_vm_phys_locality(SYSCTL_HANDLER_ARGS);
143 SYSCTL_OID(_vm, OID_AUTO, phys_locality, CTLTYPE_STRING | CTLFLAG_RD,
144     NULL, 0, sysctl_vm_phys_locality, "A", "Phys Locality Info");
145 #endif
146 
147 SYSCTL_INT(_vm, OID_AUTO, ndomains, CTLFLAG_RD,
148     &vm_ndomains, 0, "Number of physical memory domains available.");
149 
150 static vm_page_t vm_phys_alloc_seg_contig(struct vm_phys_seg *seg,
151     u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment,
152     vm_paddr_t boundary);
153 static void _vm_phys_create_seg(vm_paddr_t start, vm_paddr_t end, int domain);
154 static void vm_phys_create_seg(vm_paddr_t start, vm_paddr_t end);
155 static void vm_phys_split_pages(vm_page_t m, int oind, struct vm_freelist *fl,
156     int order, int tail);
157 
158 /*
159  * Red-black tree helpers for vm fictitious range management.
160  */
161 static inline int
162 vm_phys_fictitious_in_range(struct vm_phys_fictitious_seg *p,
163     struct vm_phys_fictitious_seg *range)
164 {
165 
166 	KASSERT(range->start != 0 && range->end != 0,
167 	    ("Invalid range passed on search for vm_fictitious page"));
168 	if (p->start >= range->end)
169 		return (1);
170 	if (p->start < range->start)
171 		return (-1);
172 
173 	return (0);
174 }
175 
176 static int
177 vm_phys_fictitious_cmp(struct vm_phys_fictitious_seg *p1,
178     struct vm_phys_fictitious_seg *p2)
179 {
180 
181 	/* Check if this is a search for a page */
182 	if (p1->end == 0)
183 		return (vm_phys_fictitious_in_range(p1, p2));
184 
185 	KASSERT(p2->end != 0,
186     ("Invalid range passed as second parameter to vm fictitious comparison"));
187 
188 	/* Searching to add a new range */
189 	if (p1->end <= p2->start)
190 		return (-1);
191 	if (p1->start >= p2->end)
192 		return (1);
193 
194 	panic("Trying to add overlapping vm fictitious ranges:\n"
195 	    "[%#jx:%#jx] and [%#jx:%#jx]", (uintmax_t)p1->start,
196 	    (uintmax_t)p1->end, (uintmax_t)p2->start, (uintmax_t)p2->end);
197 }
198 
199 int
200 vm_phys_domain_match(int prefer, vm_paddr_t low, vm_paddr_t high)
201 {
202 #ifdef NUMA
203 	domainset_t mask;
204 	int i;
205 
206 	if (vm_ndomains == 1 || mem_affinity == NULL)
207 		return (0);
208 
209 	DOMAINSET_ZERO(&mask);
210 	/*
211 	 * Check for any memory that overlaps low, high.
212 	 */
213 	for (i = 0; mem_affinity[i].end != 0; i++)
214 		if (mem_affinity[i].start <= high &&
215 		    mem_affinity[i].end >= low)
216 			DOMAINSET_SET(mem_affinity[i].domain, &mask);
217 	if (prefer != -1 && DOMAINSET_ISSET(prefer, &mask))
218 		return (prefer);
219 	if (DOMAINSET_EMPTY(&mask))
220 		panic("vm_phys_domain_match:  Impossible constraint");
221 	return (DOMAINSET_FFS(&mask) - 1);
222 #else
223 	return (0);
224 #endif
225 }
226 
227 /*
228  * Outputs the state of the physical memory allocator, specifically,
229  * the amount of physical memory in each free list.
230  */
231 static int
232 sysctl_vm_phys_free(SYSCTL_HANDLER_ARGS)
233 {
234 	struct sbuf sbuf;
235 	struct vm_freelist *fl;
236 	int dom, error, flind, oind, pind;
237 
238 	error = sysctl_wire_old_buffer(req, 0);
239 	if (error != 0)
240 		return (error);
241 	sbuf_new_for_sysctl(&sbuf, NULL, 128 * vm_ndomains, req);
242 	for (dom = 0; dom < vm_ndomains; dom++) {
243 		sbuf_printf(&sbuf,"\nDOMAIN %d:\n", dom);
244 		for (flind = 0; flind < vm_nfreelists; flind++) {
245 			sbuf_printf(&sbuf, "\nFREE LIST %d:\n"
246 			    "\n  ORDER (SIZE)  |  NUMBER"
247 			    "\n              ", flind);
248 			for (pind = 0; pind < VM_NFREEPOOL; pind++)
249 				sbuf_printf(&sbuf, "  |  POOL %d", pind);
250 			sbuf_printf(&sbuf, "\n--            ");
251 			for (pind = 0; pind < VM_NFREEPOOL; pind++)
252 				sbuf_printf(&sbuf, "-- --      ");
253 			sbuf_printf(&sbuf, "--\n");
254 			for (oind = VM_NFREEORDER - 1; oind >= 0; oind--) {
255 				sbuf_printf(&sbuf, "  %2d (%6dK)", oind,
256 				    1 << (PAGE_SHIFT - 10 + oind));
257 				for (pind = 0; pind < VM_NFREEPOOL; pind++) {
258 				fl = vm_phys_free_queues[dom][flind][pind];
259 					sbuf_printf(&sbuf, "  |  %6d",
260 					    fl[oind].lcnt);
261 				}
262 				sbuf_printf(&sbuf, "\n");
263 			}
264 		}
265 	}
266 	error = sbuf_finish(&sbuf);
267 	sbuf_delete(&sbuf);
268 	return (error);
269 }
270 
271 /*
272  * Outputs the set of physical memory segments.
273  */
274 static int
275 sysctl_vm_phys_segs(SYSCTL_HANDLER_ARGS)
276 {
277 	struct sbuf sbuf;
278 	struct vm_phys_seg *seg;
279 	int error, segind;
280 
281 	error = sysctl_wire_old_buffer(req, 0);
282 	if (error != 0)
283 		return (error);
284 	sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
285 	for (segind = 0; segind < vm_phys_nsegs; segind++) {
286 		sbuf_printf(&sbuf, "\nSEGMENT %d:\n\n", segind);
287 		seg = &vm_phys_segs[segind];
288 		sbuf_printf(&sbuf, "start:     %#jx\n",
289 		    (uintmax_t)seg->start);
290 		sbuf_printf(&sbuf, "end:       %#jx\n",
291 		    (uintmax_t)seg->end);
292 		sbuf_printf(&sbuf, "domain:    %d\n", seg->domain);
293 		sbuf_printf(&sbuf, "free list: %p\n", seg->free_queues);
294 	}
295 	error = sbuf_finish(&sbuf);
296 	sbuf_delete(&sbuf);
297 	return (error);
298 }
299 
300 /*
301  * Return affinity, or -1 if there's no affinity information.
302  */
303 int
304 vm_phys_mem_affinity(int f, int t)
305 {
306 
307 #ifdef NUMA
308 	if (mem_locality == NULL)
309 		return (-1);
310 	if (f >= vm_ndomains || t >= vm_ndomains)
311 		return (-1);
312 	return (mem_locality[f * vm_ndomains + t]);
313 #else
314 	return (-1);
315 #endif
316 }
317 
318 #ifdef NUMA
319 /*
320  * Outputs the VM locality table.
321  */
322 static int
323 sysctl_vm_phys_locality(SYSCTL_HANDLER_ARGS)
324 {
325 	struct sbuf sbuf;
326 	int error, i, j;
327 
328 	error = sysctl_wire_old_buffer(req, 0);
329 	if (error != 0)
330 		return (error);
331 	sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
332 
333 	sbuf_printf(&sbuf, "\n");
334 
335 	for (i = 0; i < vm_ndomains; i++) {
336 		sbuf_printf(&sbuf, "%d: ", i);
337 		for (j = 0; j < vm_ndomains; j++) {
338 			sbuf_printf(&sbuf, "%d ", vm_phys_mem_affinity(i, j));
339 		}
340 		sbuf_printf(&sbuf, "\n");
341 	}
342 	error = sbuf_finish(&sbuf);
343 	sbuf_delete(&sbuf);
344 	return (error);
345 }
346 #endif
347 
348 static void
349 vm_freelist_add(struct vm_freelist *fl, vm_page_t m, int order, int tail)
350 {
351 
352 	m->order = order;
353 	if (tail)
354 		TAILQ_INSERT_TAIL(&fl[order].pl, m, listq);
355 	else
356 		TAILQ_INSERT_HEAD(&fl[order].pl, m, listq);
357 	fl[order].lcnt++;
358 }
359 
360 static void
361 vm_freelist_rem(struct vm_freelist *fl, vm_page_t m, int order)
362 {
363 
364 	TAILQ_REMOVE(&fl[order].pl, m, listq);
365 	fl[order].lcnt--;
366 	m->order = VM_NFREEORDER;
367 }
368 
369 /*
370  * Create a physical memory segment.
371  */
372 static void
373 _vm_phys_create_seg(vm_paddr_t start, vm_paddr_t end, int domain)
374 {
375 	struct vm_phys_seg *seg;
376 
377 	KASSERT(vm_phys_nsegs < VM_PHYSSEG_MAX,
378 	    ("vm_phys_create_seg: increase VM_PHYSSEG_MAX"));
379 	KASSERT(domain >= 0 && domain < vm_ndomains,
380 	    ("vm_phys_create_seg: invalid domain provided"));
381 	seg = &vm_phys_segs[vm_phys_nsegs++];
382 	while (seg > vm_phys_segs && (seg - 1)->start >= end) {
383 		*seg = *(seg - 1);
384 		seg--;
385 	}
386 	seg->start = start;
387 	seg->end = end;
388 	seg->domain = domain;
389 }
390 
391 static void
392 vm_phys_create_seg(vm_paddr_t start, vm_paddr_t end)
393 {
394 #ifdef NUMA
395 	int i;
396 
397 	if (mem_affinity == NULL) {
398 		_vm_phys_create_seg(start, end, 0);
399 		return;
400 	}
401 
402 	for (i = 0;; i++) {
403 		if (mem_affinity[i].end == 0)
404 			panic("Reached end of affinity info");
405 		if (mem_affinity[i].end <= start)
406 			continue;
407 		if (mem_affinity[i].start > start)
408 			panic("No affinity info for start %jx",
409 			    (uintmax_t)start);
410 		if (mem_affinity[i].end >= end) {
411 			_vm_phys_create_seg(start, end,
412 			    mem_affinity[i].domain);
413 			break;
414 		}
415 		_vm_phys_create_seg(start, mem_affinity[i].end,
416 		    mem_affinity[i].domain);
417 		start = mem_affinity[i].end;
418 	}
419 #else
420 	_vm_phys_create_seg(start, end, 0);
421 #endif
422 }
423 
424 /*
425  * Add a physical memory segment.
426  */
427 void
428 vm_phys_add_seg(vm_paddr_t start, vm_paddr_t end)
429 {
430 	vm_paddr_t paddr;
431 
432 	KASSERT((start & PAGE_MASK) == 0,
433 	    ("vm_phys_define_seg: start is not page aligned"));
434 	KASSERT((end & PAGE_MASK) == 0,
435 	    ("vm_phys_define_seg: end is not page aligned"));
436 
437 	/*
438 	 * Split the physical memory segment if it spans two or more free
439 	 * list boundaries.
440 	 */
441 	paddr = start;
442 #ifdef	VM_FREELIST_LOWMEM
443 	if (paddr < VM_LOWMEM_BOUNDARY && end > VM_LOWMEM_BOUNDARY) {
444 		vm_phys_create_seg(paddr, VM_LOWMEM_BOUNDARY);
445 		paddr = VM_LOWMEM_BOUNDARY;
446 	}
447 #endif
448 #ifdef	VM_FREELIST_DMA32
449 	if (paddr < VM_DMA32_BOUNDARY && end > VM_DMA32_BOUNDARY) {
450 		vm_phys_create_seg(paddr, VM_DMA32_BOUNDARY);
451 		paddr = VM_DMA32_BOUNDARY;
452 	}
453 #endif
454 	vm_phys_create_seg(paddr, end);
455 }
456 
457 /*
458  * Initialize the physical memory allocator.
459  *
460  * Requires that vm_page_array is initialized!
461  */
462 void
463 vm_phys_init(void)
464 {
465 	struct vm_freelist *fl;
466 	struct vm_phys_seg *end_seg, *prev_seg, *seg, *tmp_seg;
467 	u_long npages;
468 	int dom, flind, freelist, oind, pind, segind;
469 
470 	/*
471 	 * Compute the number of free lists, and generate the mapping from the
472 	 * manifest constants VM_FREELIST_* to the free list indices.
473 	 *
474 	 * Initially, the entries of vm_freelist_to_flind[] are set to either
475 	 * 0 or 1 to indicate which free lists should be created.
476 	 */
477 	npages = 0;
478 	for (segind = vm_phys_nsegs - 1; segind >= 0; segind--) {
479 		seg = &vm_phys_segs[segind];
480 #ifdef	VM_FREELIST_LOWMEM
481 		if (seg->end <= VM_LOWMEM_BOUNDARY)
482 			vm_freelist_to_flind[VM_FREELIST_LOWMEM] = 1;
483 		else
484 #endif
485 #ifdef	VM_FREELIST_DMA32
486 		if (
487 #ifdef	VM_DMA32_NPAGES_THRESHOLD
488 		    /*
489 		     * Create the DMA32 free list only if the amount of
490 		     * physical memory above physical address 4G exceeds the
491 		     * given threshold.
492 		     */
493 		    npages > VM_DMA32_NPAGES_THRESHOLD &&
494 #endif
495 		    seg->end <= VM_DMA32_BOUNDARY)
496 			vm_freelist_to_flind[VM_FREELIST_DMA32] = 1;
497 		else
498 #endif
499 		{
500 			npages += atop(seg->end - seg->start);
501 			vm_freelist_to_flind[VM_FREELIST_DEFAULT] = 1;
502 		}
503 	}
504 	/* Change each entry into a running total of the free lists. */
505 	for (freelist = 1; freelist < VM_NFREELIST; freelist++) {
506 		vm_freelist_to_flind[freelist] +=
507 		    vm_freelist_to_flind[freelist - 1];
508 	}
509 	vm_nfreelists = vm_freelist_to_flind[VM_NFREELIST - 1];
510 	KASSERT(vm_nfreelists > 0, ("vm_phys_init: no free lists"));
511 	/* Change each entry into a free list index. */
512 	for (freelist = 0; freelist < VM_NFREELIST; freelist++)
513 		vm_freelist_to_flind[freelist]--;
514 
515 	/*
516 	 * Initialize the first_page and free_queues fields of each physical
517 	 * memory segment.
518 	 */
519 #ifdef VM_PHYSSEG_SPARSE
520 	npages = 0;
521 #endif
522 	for (segind = 0; segind < vm_phys_nsegs; segind++) {
523 		seg = &vm_phys_segs[segind];
524 #ifdef VM_PHYSSEG_SPARSE
525 		seg->first_page = &vm_page_array[npages];
526 		npages += atop(seg->end - seg->start);
527 #else
528 		seg->first_page = PHYS_TO_VM_PAGE(seg->start);
529 #endif
530 #ifdef	VM_FREELIST_LOWMEM
531 		if (seg->end <= VM_LOWMEM_BOUNDARY) {
532 			flind = vm_freelist_to_flind[VM_FREELIST_LOWMEM];
533 			KASSERT(flind >= 0,
534 			    ("vm_phys_init: LOWMEM flind < 0"));
535 		} else
536 #endif
537 #ifdef	VM_FREELIST_DMA32
538 		if (seg->end <= VM_DMA32_BOUNDARY) {
539 			flind = vm_freelist_to_flind[VM_FREELIST_DMA32];
540 			KASSERT(flind >= 0,
541 			    ("vm_phys_init: DMA32 flind < 0"));
542 		} else
543 #endif
544 		{
545 			flind = vm_freelist_to_flind[VM_FREELIST_DEFAULT];
546 			KASSERT(flind >= 0,
547 			    ("vm_phys_init: DEFAULT flind < 0"));
548 		}
549 		seg->free_queues = &vm_phys_free_queues[seg->domain][flind];
550 	}
551 
552 	/*
553 	 * Coalesce physical memory segments that are contiguous and share the
554 	 * same per-domain free queues.
555 	 */
556 	prev_seg = vm_phys_segs;
557 	seg = &vm_phys_segs[1];
558 	end_seg = &vm_phys_segs[vm_phys_nsegs];
559 	while (seg < end_seg) {
560 		if (prev_seg->end == seg->start &&
561 		    prev_seg->free_queues == seg->free_queues) {
562 			prev_seg->end = seg->end;
563 			KASSERT(prev_seg->domain == seg->domain,
564 			    ("vm_phys_init: free queues cannot span domains"));
565 			vm_phys_nsegs--;
566 			end_seg--;
567 			for (tmp_seg = seg; tmp_seg < end_seg; tmp_seg++)
568 				*tmp_seg = *(tmp_seg + 1);
569 		} else {
570 			prev_seg = seg;
571 			seg++;
572 		}
573 	}
574 
575 	/*
576 	 * Initialize the free queues.
577 	 */
578 	for (dom = 0; dom < vm_ndomains; dom++) {
579 		for (flind = 0; flind < vm_nfreelists; flind++) {
580 			for (pind = 0; pind < VM_NFREEPOOL; pind++) {
581 				fl = vm_phys_free_queues[dom][flind][pind];
582 				for (oind = 0; oind < VM_NFREEORDER; oind++)
583 					TAILQ_INIT(&fl[oind].pl);
584 			}
585 		}
586 	}
587 
588 	rw_init(&vm_phys_fictitious_reg_lock, "vmfctr");
589 }
590 
591 /*
592  * Register info about the NUMA topology of the system.
593  *
594  * Invoked by platform-dependent code prior to vm_phys_init().
595  */
596 void
597 vm_phys_register_domains(int ndomains, struct mem_affinity *affinity,
598     int *locality)
599 {
600 #ifdef NUMA
601 	int d, i;
602 
603 	/*
604 	 * For now the only override value that we support is 1, which
605 	 * effectively disables NUMA-awareness in the allocators.
606 	 */
607 	d = 0;
608 	TUNABLE_INT_FETCH("vm.numa.disabled", &d);
609 	if (d)
610 		ndomains = 1;
611 
612 	if (ndomains > 1) {
613 		vm_ndomains = ndomains;
614 		mem_affinity = affinity;
615 		mem_locality = locality;
616 	}
617 
618 	for (i = 0; i < vm_ndomains; i++)
619 		DOMAINSET_SET(i, &all_domains);
620 #else
621 	(void)ndomains;
622 	(void)affinity;
623 	(void)locality;
624 #endif
625 }
626 
627 /*
628  * Split a contiguous, power of two-sized set of physical pages.
629  *
630  * When this function is called by a page allocation function, the caller
631  * should request insertion at the head unless the order [order, oind) queues
632  * are known to be empty.  The objective being to reduce the likelihood of
633  * long-term fragmentation by promoting contemporaneous allocation and
634  * (hopefully) deallocation.
635  */
636 static __inline void
637 vm_phys_split_pages(vm_page_t m, int oind, struct vm_freelist *fl, int order,
638     int tail)
639 {
640 	vm_page_t m_buddy;
641 
642 	while (oind > order) {
643 		oind--;
644 		m_buddy = &m[1 << oind];
645 		KASSERT(m_buddy->order == VM_NFREEORDER,
646 		    ("vm_phys_split_pages: page %p has unexpected order %d",
647 		    m_buddy, m_buddy->order));
648 		vm_freelist_add(fl, m_buddy, oind, tail);
649         }
650 }
651 
652 /*
653  * Add the physical pages [m, m + npages) at the end of a power-of-two aligned
654  * and sized set to the specified free list.
655  *
656  * When this function is called by a page allocation function, the caller
657  * should request insertion at the head unless the lower-order queues are
658  * known to be empty.  The objective being to reduce the likelihood of long-
659  * term fragmentation by promoting contemporaneous allocation and (hopefully)
660  * deallocation.
661  *
662  * The physical page m's buddy must not be free.
663  */
664 static void
665 vm_phys_enq_range(vm_page_t m, u_int npages, struct vm_freelist *fl, int tail)
666 {
667 	u_int n;
668 	int order;
669 
670 	KASSERT(npages > 0, ("vm_phys_enq_range: npages is 0"));
671 	KASSERT(((VM_PAGE_TO_PHYS(m) + npages * PAGE_SIZE) &
672 	    ((PAGE_SIZE << (fls(npages) - 1)) - 1)) == 0,
673 	    ("vm_phys_enq_range: page %p and npages %u are misaligned",
674 	    m, npages));
675 	do {
676 		KASSERT(m->order == VM_NFREEORDER,
677 		    ("vm_phys_enq_range: page %p has unexpected order %d",
678 		    m, m->order));
679 		order = ffs(npages) - 1;
680 		KASSERT(order < VM_NFREEORDER,
681 		    ("vm_phys_enq_range: order %d is out of range", order));
682 		vm_freelist_add(fl, m, order, tail);
683 		n = 1 << order;
684 		m += n;
685 		npages -= n;
686 	} while (npages > 0);
687 }
688 
689 /*
690  * Tries to allocate the specified number of pages from the specified pool
691  * within the specified domain.  Returns the actual number of allocated pages
692  * and a pointer to each page through the array ma[].
693  *
694  * The returned pages may not be physically contiguous.  However, in contrast
695  * to performing multiple, back-to-back calls to vm_phys_alloc_pages(..., 0),
696  * calling this function once to allocate the desired number of pages will
697  * avoid wasted time in vm_phys_split_pages().
698  *
699  * The free page queues for the specified domain must be locked.
700  */
701 int
702 vm_phys_alloc_npages(int domain, int pool, int npages, vm_page_t ma[])
703 {
704 	struct vm_freelist *alt, *fl;
705 	vm_page_t m;
706 	int avail, end, flind, freelist, i, need, oind, pind;
707 
708 	KASSERT(domain >= 0 && domain < vm_ndomains,
709 	    ("vm_phys_alloc_npages: domain %d is out of range", domain));
710 	KASSERT(pool < VM_NFREEPOOL,
711 	    ("vm_phys_alloc_npages: pool %d is out of range", pool));
712 	KASSERT(npages <= 1 << (VM_NFREEORDER - 1),
713 	    ("vm_phys_alloc_npages: npages %d is out of range", npages));
714 	vm_domain_free_assert_locked(VM_DOMAIN(domain));
715 	i = 0;
716 	for (freelist = 0; freelist < VM_NFREELIST; freelist++) {
717 		flind = vm_freelist_to_flind[freelist];
718 		if (flind < 0)
719 			continue;
720 		fl = vm_phys_free_queues[domain][flind][pool];
721 		for (oind = 0; oind < VM_NFREEORDER; oind++) {
722 			while ((m = TAILQ_FIRST(&fl[oind].pl)) != NULL) {
723 				vm_freelist_rem(fl, m, oind);
724 				avail = 1 << oind;
725 				need = imin(npages - i, avail);
726 				for (end = i + need; i < end;)
727 					ma[i++] = m++;
728 				if (need < avail) {
729 					/*
730 					 * Return excess pages to fl.  Its
731 					 * order [0, oind) queues are empty.
732 					 */
733 					vm_phys_enq_range(m, avail - need, fl,
734 					    1);
735 					return (npages);
736 				} else if (i == npages)
737 					return (npages);
738 			}
739 		}
740 		for (oind = VM_NFREEORDER - 1; oind >= 0; oind--) {
741 			for (pind = 0; pind < VM_NFREEPOOL; pind++) {
742 				alt = vm_phys_free_queues[domain][flind][pind];
743 				while ((m = TAILQ_FIRST(&alt[oind].pl)) !=
744 				    NULL) {
745 					vm_freelist_rem(alt, m, oind);
746 					vm_phys_set_pool(pool, m, oind);
747 					avail = 1 << oind;
748 					need = imin(npages - i, avail);
749 					for (end = i + need; i < end;)
750 						ma[i++] = m++;
751 					if (need < avail) {
752 						/*
753 						 * Return excess pages to fl.
754 						 * Its order [0, oind) queues
755 						 * are empty.
756 						 */
757 						vm_phys_enq_range(m, avail -
758 						    need, fl, 1);
759 						return (npages);
760 					} else if (i == npages)
761 						return (npages);
762 				}
763 			}
764 		}
765 	}
766 	return (i);
767 }
768 
769 /*
770  * Allocate a contiguous, power of two-sized set of physical pages
771  * from the free lists.
772  *
773  * The free page queues must be locked.
774  */
775 vm_page_t
776 vm_phys_alloc_pages(int domain, int pool, int order)
777 {
778 	vm_page_t m;
779 	int freelist;
780 
781 	for (freelist = 0; freelist < VM_NFREELIST; freelist++) {
782 		m = vm_phys_alloc_freelist_pages(domain, freelist, pool, order);
783 		if (m != NULL)
784 			return (m);
785 	}
786 	return (NULL);
787 }
788 
789 /*
790  * Allocate a contiguous, power of two-sized set of physical pages from the
791  * specified free list.  The free list must be specified using one of the
792  * manifest constants VM_FREELIST_*.
793  *
794  * The free page queues must be locked.
795  */
796 vm_page_t
797 vm_phys_alloc_freelist_pages(int domain, int freelist, int pool, int order)
798 {
799 	struct vm_freelist *alt, *fl;
800 	vm_page_t m;
801 	int oind, pind, flind;
802 
803 	KASSERT(domain >= 0 && domain < vm_ndomains,
804 	    ("vm_phys_alloc_freelist_pages: domain %d is out of range",
805 	    domain));
806 	KASSERT(freelist < VM_NFREELIST,
807 	    ("vm_phys_alloc_freelist_pages: freelist %d is out of range",
808 	    freelist));
809 	KASSERT(pool < VM_NFREEPOOL,
810 	    ("vm_phys_alloc_freelist_pages: pool %d is out of range", pool));
811 	KASSERT(order < VM_NFREEORDER,
812 	    ("vm_phys_alloc_freelist_pages: order %d is out of range", order));
813 
814 	flind = vm_freelist_to_flind[freelist];
815 	/* Check if freelist is present */
816 	if (flind < 0)
817 		return (NULL);
818 
819 	vm_domain_free_assert_locked(VM_DOMAIN(domain));
820 	fl = &vm_phys_free_queues[domain][flind][pool][0];
821 	for (oind = order; oind < VM_NFREEORDER; oind++) {
822 		m = TAILQ_FIRST(&fl[oind].pl);
823 		if (m != NULL) {
824 			vm_freelist_rem(fl, m, oind);
825 			/* The order [order, oind) queues are empty. */
826 			vm_phys_split_pages(m, oind, fl, order, 1);
827 			return (m);
828 		}
829 	}
830 
831 	/*
832 	 * The given pool was empty.  Find the largest
833 	 * contiguous, power-of-two-sized set of pages in any
834 	 * pool.  Transfer these pages to the given pool, and
835 	 * use them to satisfy the allocation.
836 	 */
837 	for (oind = VM_NFREEORDER - 1; oind >= order; oind--) {
838 		for (pind = 0; pind < VM_NFREEPOOL; pind++) {
839 			alt = &vm_phys_free_queues[domain][flind][pind][0];
840 			m = TAILQ_FIRST(&alt[oind].pl);
841 			if (m != NULL) {
842 				vm_freelist_rem(alt, m, oind);
843 				vm_phys_set_pool(pool, m, oind);
844 				/* The order [order, oind) queues are empty. */
845 				vm_phys_split_pages(m, oind, fl, order, 1);
846 				return (m);
847 			}
848 		}
849 	}
850 	return (NULL);
851 }
852 
853 /*
854  * Find the vm_page corresponding to the given physical address.
855  */
856 vm_page_t
857 vm_phys_paddr_to_vm_page(vm_paddr_t pa)
858 {
859 	struct vm_phys_seg *seg;
860 	int segind;
861 
862 	for (segind = 0; segind < vm_phys_nsegs; segind++) {
863 		seg = &vm_phys_segs[segind];
864 		if (pa >= seg->start && pa < seg->end)
865 			return (&seg->first_page[atop(pa - seg->start)]);
866 	}
867 	return (NULL);
868 }
869 
870 vm_page_t
871 vm_phys_fictitious_to_vm_page(vm_paddr_t pa)
872 {
873 	struct vm_phys_fictitious_seg tmp, *seg;
874 	vm_page_t m;
875 
876 	m = NULL;
877 	tmp.start = pa;
878 	tmp.end = 0;
879 
880 	rw_rlock(&vm_phys_fictitious_reg_lock);
881 	seg = RB_FIND(fict_tree, &vm_phys_fictitious_tree, &tmp);
882 	rw_runlock(&vm_phys_fictitious_reg_lock);
883 	if (seg == NULL)
884 		return (NULL);
885 
886 	m = &seg->first_page[atop(pa - seg->start)];
887 	KASSERT((m->flags & PG_FICTITIOUS) != 0, ("%p not fictitious", m));
888 
889 	return (m);
890 }
891 
892 static inline void
893 vm_phys_fictitious_init_range(vm_page_t range, vm_paddr_t start,
894     long page_count, vm_memattr_t memattr)
895 {
896 	long i;
897 
898 	bzero(range, page_count * sizeof(*range));
899 	for (i = 0; i < page_count; i++) {
900 		vm_page_initfake(&range[i], start + PAGE_SIZE * i, memattr);
901 		range[i].oflags &= ~VPO_UNMANAGED;
902 		range[i].busy_lock = VPB_UNBUSIED;
903 	}
904 }
905 
906 int
907 vm_phys_fictitious_reg_range(vm_paddr_t start, vm_paddr_t end,
908     vm_memattr_t memattr)
909 {
910 	struct vm_phys_fictitious_seg *seg;
911 	vm_page_t fp;
912 	long page_count;
913 #ifdef VM_PHYSSEG_DENSE
914 	long pi, pe;
915 	long dpage_count;
916 #endif
917 
918 	KASSERT(start < end,
919 	    ("Start of segment isn't less than end (start: %jx end: %jx)",
920 	    (uintmax_t)start, (uintmax_t)end));
921 
922 	page_count = (end - start) / PAGE_SIZE;
923 
924 #ifdef VM_PHYSSEG_DENSE
925 	pi = atop(start);
926 	pe = atop(end);
927 	if (pi >= first_page && (pi - first_page) < vm_page_array_size) {
928 		fp = &vm_page_array[pi - first_page];
929 		if ((pe - first_page) > vm_page_array_size) {
930 			/*
931 			 * We have a segment that starts inside
932 			 * of vm_page_array, but ends outside of it.
933 			 *
934 			 * Use vm_page_array pages for those that are
935 			 * inside of the vm_page_array range, and
936 			 * allocate the remaining ones.
937 			 */
938 			dpage_count = vm_page_array_size - (pi - first_page);
939 			vm_phys_fictitious_init_range(fp, start, dpage_count,
940 			    memattr);
941 			page_count -= dpage_count;
942 			start += ptoa(dpage_count);
943 			goto alloc;
944 		}
945 		/*
946 		 * We can allocate the full range from vm_page_array,
947 		 * so there's no need to register the range in the tree.
948 		 */
949 		vm_phys_fictitious_init_range(fp, start, page_count, memattr);
950 		return (0);
951 	} else if (pe > first_page && (pe - first_page) < vm_page_array_size) {
952 		/*
953 		 * We have a segment that ends inside of vm_page_array,
954 		 * but starts outside of it.
955 		 */
956 		fp = &vm_page_array[0];
957 		dpage_count = pe - first_page;
958 		vm_phys_fictitious_init_range(fp, ptoa(first_page), dpage_count,
959 		    memattr);
960 		end -= ptoa(dpage_count);
961 		page_count -= dpage_count;
962 		goto alloc;
963 	} else if (pi < first_page && pe > (first_page + vm_page_array_size)) {
964 		/*
965 		 * Trying to register a fictitious range that expands before
966 		 * and after vm_page_array.
967 		 */
968 		return (EINVAL);
969 	} else {
970 alloc:
971 #endif
972 		fp = malloc(page_count * sizeof(struct vm_page), M_FICT_PAGES,
973 		    M_WAITOK);
974 #ifdef VM_PHYSSEG_DENSE
975 	}
976 #endif
977 	vm_phys_fictitious_init_range(fp, start, page_count, memattr);
978 
979 	seg = malloc(sizeof(*seg), M_FICT_PAGES, M_WAITOK | M_ZERO);
980 	seg->start = start;
981 	seg->end = end;
982 	seg->first_page = fp;
983 
984 	rw_wlock(&vm_phys_fictitious_reg_lock);
985 	RB_INSERT(fict_tree, &vm_phys_fictitious_tree, seg);
986 	rw_wunlock(&vm_phys_fictitious_reg_lock);
987 
988 	return (0);
989 }
990 
991 void
992 vm_phys_fictitious_unreg_range(vm_paddr_t start, vm_paddr_t end)
993 {
994 	struct vm_phys_fictitious_seg *seg, tmp;
995 #ifdef VM_PHYSSEG_DENSE
996 	long pi, pe;
997 #endif
998 
999 	KASSERT(start < end,
1000 	    ("Start of segment isn't less than end (start: %jx end: %jx)",
1001 	    (uintmax_t)start, (uintmax_t)end));
1002 
1003 #ifdef VM_PHYSSEG_DENSE
1004 	pi = atop(start);
1005 	pe = atop(end);
1006 	if (pi >= first_page && (pi - first_page) < vm_page_array_size) {
1007 		if ((pe - first_page) <= vm_page_array_size) {
1008 			/*
1009 			 * This segment was allocated using vm_page_array
1010 			 * only, there's nothing to do since those pages
1011 			 * were never added to the tree.
1012 			 */
1013 			return;
1014 		}
1015 		/*
1016 		 * We have a segment that starts inside
1017 		 * of vm_page_array, but ends outside of it.
1018 		 *
1019 		 * Calculate how many pages were added to the
1020 		 * tree and free them.
1021 		 */
1022 		start = ptoa(first_page + vm_page_array_size);
1023 	} else if (pe > first_page && (pe - first_page) < vm_page_array_size) {
1024 		/*
1025 		 * We have a segment that ends inside of vm_page_array,
1026 		 * but starts outside of it.
1027 		 */
1028 		end = ptoa(first_page);
1029 	} else if (pi < first_page && pe > (first_page + vm_page_array_size)) {
1030 		/* Since it's not possible to register such a range, panic. */
1031 		panic(
1032 		    "Unregistering not registered fictitious range [%#jx:%#jx]",
1033 		    (uintmax_t)start, (uintmax_t)end);
1034 	}
1035 #endif
1036 	tmp.start = start;
1037 	tmp.end = 0;
1038 
1039 	rw_wlock(&vm_phys_fictitious_reg_lock);
1040 	seg = RB_FIND(fict_tree, &vm_phys_fictitious_tree, &tmp);
1041 	if (seg->start != start || seg->end != end) {
1042 		rw_wunlock(&vm_phys_fictitious_reg_lock);
1043 		panic(
1044 		    "Unregistering not registered fictitious range [%#jx:%#jx]",
1045 		    (uintmax_t)start, (uintmax_t)end);
1046 	}
1047 	RB_REMOVE(fict_tree, &vm_phys_fictitious_tree, seg);
1048 	rw_wunlock(&vm_phys_fictitious_reg_lock);
1049 	free(seg->first_page, M_FICT_PAGES);
1050 	free(seg, M_FICT_PAGES);
1051 }
1052 
1053 /*
1054  * Free a contiguous, power of two-sized set of physical pages.
1055  *
1056  * The free page queues must be locked.
1057  */
1058 void
1059 vm_phys_free_pages(vm_page_t m, int order)
1060 {
1061 	struct vm_freelist *fl;
1062 	struct vm_phys_seg *seg;
1063 	vm_paddr_t pa;
1064 	vm_page_t m_buddy;
1065 
1066 	KASSERT(m->order == VM_NFREEORDER,
1067 	    ("vm_phys_free_pages: page %p has unexpected order %d",
1068 	    m, m->order));
1069 	KASSERT(m->pool < VM_NFREEPOOL,
1070 	    ("vm_phys_free_pages: page %p has unexpected pool %d",
1071 	    m, m->pool));
1072 	KASSERT(order < VM_NFREEORDER,
1073 	    ("vm_phys_free_pages: order %d is out of range", order));
1074 	seg = &vm_phys_segs[m->segind];
1075 	vm_domain_free_assert_locked(VM_DOMAIN(seg->domain));
1076 	if (order < VM_NFREEORDER - 1) {
1077 		pa = VM_PAGE_TO_PHYS(m);
1078 		do {
1079 			pa ^= ((vm_paddr_t)1 << (PAGE_SHIFT + order));
1080 			if (pa < seg->start || pa >= seg->end)
1081 				break;
1082 			m_buddy = &seg->first_page[atop(pa - seg->start)];
1083 			if (m_buddy->order != order)
1084 				break;
1085 			fl = (*seg->free_queues)[m_buddy->pool];
1086 			vm_freelist_rem(fl, m_buddy, order);
1087 			if (m_buddy->pool != m->pool)
1088 				vm_phys_set_pool(m->pool, m_buddy, order);
1089 			order++;
1090 			pa &= ~(((vm_paddr_t)1 << (PAGE_SHIFT + order)) - 1);
1091 			m = &seg->first_page[atop(pa - seg->start)];
1092 		} while (order < VM_NFREEORDER - 1);
1093 	}
1094 	fl = (*seg->free_queues)[m->pool];
1095 	vm_freelist_add(fl, m, order, 1);
1096 }
1097 
1098 /*
1099  * Free a contiguous, arbitrarily sized set of physical pages.
1100  *
1101  * The free page queues must be locked.
1102  */
1103 void
1104 vm_phys_free_contig(vm_page_t m, u_long npages)
1105 {
1106 	u_int n;
1107 	int order;
1108 
1109 	/*
1110 	 * Avoid unnecessary coalescing by freeing the pages in the largest
1111 	 * possible power-of-two-sized subsets.
1112 	 */
1113 	vm_domain_free_assert_locked(vm_pagequeue_domain(m));
1114 	for (;; npages -= n) {
1115 		/*
1116 		 * Unsigned "min" is used here so that "order" is assigned
1117 		 * "VM_NFREEORDER - 1" when "m"'s physical address is zero
1118 		 * or the low-order bits of its physical address are zero
1119 		 * because the size of a physical address exceeds the size of
1120 		 * a long.
1121 		 */
1122 		order = min(ffsl(VM_PAGE_TO_PHYS(m) >> PAGE_SHIFT) - 1,
1123 		    VM_NFREEORDER - 1);
1124 		n = 1 << order;
1125 		if (npages < n)
1126 			break;
1127 		vm_phys_free_pages(m, order);
1128 		m += n;
1129 	}
1130 	/* The residual "npages" is less than "1 << (VM_NFREEORDER - 1)". */
1131 	for (; npages > 0; npages -= n) {
1132 		order = flsl(npages) - 1;
1133 		n = 1 << order;
1134 		vm_phys_free_pages(m, order);
1135 		m += n;
1136 	}
1137 }
1138 
1139 /*
1140  * Scan physical memory between the specified addresses "low" and "high" for a
1141  * run of contiguous physical pages that satisfy the specified conditions, and
1142  * return the lowest page in the run.  The specified "alignment" determines
1143  * the alignment of the lowest physical page in the run.  If the specified
1144  * "boundary" is non-zero, then the run of physical pages cannot span a
1145  * physical address that is a multiple of "boundary".
1146  *
1147  * "npages" must be greater than zero.  Both "alignment" and "boundary" must
1148  * be a power of two.
1149  */
1150 vm_page_t
1151 vm_phys_scan_contig(int domain, u_long npages, vm_paddr_t low, vm_paddr_t high,
1152     u_long alignment, vm_paddr_t boundary, int options)
1153 {
1154 	vm_paddr_t pa_end;
1155 	vm_page_t m_end, m_run, m_start;
1156 	struct vm_phys_seg *seg;
1157 	int segind;
1158 
1159 	KASSERT(npages > 0, ("npages is 0"));
1160 	KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
1161 	KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
1162 	if (low >= high)
1163 		return (NULL);
1164 	for (segind = 0; segind < vm_phys_nsegs; segind++) {
1165 		seg = &vm_phys_segs[segind];
1166 		if (seg->domain != domain)
1167 			continue;
1168 		if (seg->start >= high)
1169 			break;
1170 		if (low >= seg->end)
1171 			continue;
1172 		if (low <= seg->start)
1173 			m_start = seg->first_page;
1174 		else
1175 			m_start = &seg->first_page[atop(low - seg->start)];
1176 		if (high < seg->end)
1177 			pa_end = high;
1178 		else
1179 			pa_end = seg->end;
1180 		if (pa_end - VM_PAGE_TO_PHYS(m_start) < ptoa(npages))
1181 			continue;
1182 		m_end = &seg->first_page[atop(pa_end - seg->start)];
1183 		m_run = vm_page_scan_contig(npages, m_start, m_end,
1184 		    alignment, boundary, options);
1185 		if (m_run != NULL)
1186 			return (m_run);
1187 	}
1188 	return (NULL);
1189 }
1190 
1191 /*
1192  * Set the pool for a contiguous, power of two-sized set of physical pages.
1193  */
1194 void
1195 vm_phys_set_pool(int pool, vm_page_t m, int order)
1196 {
1197 	vm_page_t m_tmp;
1198 
1199 	for (m_tmp = m; m_tmp < &m[1 << order]; m_tmp++)
1200 		m_tmp->pool = pool;
1201 }
1202 
1203 /*
1204  * Search for the given physical page "m" in the free lists.  If the search
1205  * succeeds, remove "m" from the free lists and return TRUE.  Otherwise, return
1206  * FALSE, indicating that "m" is not in the free lists.
1207  *
1208  * The free page queues must be locked.
1209  */
1210 boolean_t
1211 vm_phys_unfree_page(vm_page_t m)
1212 {
1213 	struct vm_freelist *fl;
1214 	struct vm_phys_seg *seg;
1215 	vm_paddr_t pa, pa_half;
1216 	vm_page_t m_set, m_tmp;
1217 	int order;
1218 
1219 	/*
1220 	 * First, find the contiguous, power of two-sized set of free
1221 	 * physical pages containing the given physical page "m" and
1222 	 * assign it to "m_set".
1223 	 */
1224 	seg = &vm_phys_segs[m->segind];
1225 	vm_domain_free_assert_locked(VM_DOMAIN(seg->domain));
1226 	for (m_set = m, order = 0; m_set->order == VM_NFREEORDER &&
1227 	    order < VM_NFREEORDER - 1; ) {
1228 		order++;
1229 		pa = m->phys_addr & (~(vm_paddr_t)0 << (PAGE_SHIFT + order));
1230 		if (pa >= seg->start)
1231 			m_set = &seg->first_page[atop(pa - seg->start)];
1232 		else
1233 			return (FALSE);
1234 	}
1235 	if (m_set->order < order)
1236 		return (FALSE);
1237 	if (m_set->order == VM_NFREEORDER)
1238 		return (FALSE);
1239 	KASSERT(m_set->order < VM_NFREEORDER,
1240 	    ("vm_phys_unfree_page: page %p has unexpected order %d",
1241 	    m_set, m_set->order));
1242 
1243 	/*
1244 	 * Next, remove "m_set" from the free lists.  Finally, extract
1245 	 * "m" from "m_set" using an iterative algorithm: While "m_set"
1246 	 * is larger than a page, shrink "m_set" by returning the half
1247 	 * of "m_set" that does not contain "m" to the free lists.
1248 	 */
1249 	fl = (*seg->free_queues)[m_set->pool];
1250 	order = m_set->order;
1251 	vm_freelist_rem(fl, m_set, order);
1252 	while (order > 0) {
1253 		order--;
1254 		pa_half = m_set->phys_addr ^ (1 << (PAGE_SHIFT + order));
1255 		if (m->phys_addr < pa_half)
1256 			m_tmp = &seg->first_page[atop(pa_half - seg->start)];
1257 		else {
1258 			m_tmp = m_set;
1259 			m_set = &seg->first_page[atop(pa_half - seg->start)];
1260 		}
1261 		vm_freelist_add(fl, m_tmp, order, 0);
1262 	}
1263 	KASSERT(m_set == m, ("vm_phys_unfree_page: fatal inconsistency"));
1264 	return (TRUE);
1265 }
1266 
1267 /*
1268  * Allocate a contiguous set of physical pages of the given size
1269  * "npages" from the free lists.  All of the physical pages must be at
1270  * or above the given physical address "low" and below the given
1271  * physical address "high".  The given value "alignment" determines the
1272  * alignment of the first physical page in the set.  If the given value
1273  * "boundary" is non-zero, then the set of physical pages cannot cross
1274  * any physical address boundary that is a multiple of that value.  Both
1275  * "alignment" and "boundary" must be a power of two.
1276  */
1277 vm_page_t
1278 vm_phys_alloc_contig(int domain, u_long npages, vm_paddr_t low, vm_paddr_t high,
1279     u_long alignment, vm_paddr_t boundary)
1280 {
1281 	vm_paddr_t pa_end, pa_start;
1282 	vm_page_t m_run;
1283 	struct vm_phys_seg *seg;
1284 	int segind;
1285 
1286 	KASSERT(npages > 0, ("npages is 0"));
1287 	KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
1288 	KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
1289 	vm_domain_free_assert_locked(VM_DOMAIN(domain));
1290 	if (low >= high)
1291 		return (NULL);
1292 	m_run = NULL;
1293 	for (segind = vm_phys_nsegs - 1; segind >= 0; segind--) {
1294 		seg = &vm_phys_segs[segind];
1295 		if (seg->start >= high || seg->domain != domain)
1296 			continue;
1297 		if (low >= seg->end)
1298 			break;
1299 		if (low <= seg->start)
1300 			pa_start = seg->start;
1301 		else
1302 			pa_start = low;
1303 		if (high < seg->end)
1304 			pa_end = high;
1305 		else
1306 			pa_end = seg->end;
1307 		if (pa_end - pa_start < ptoa(npages))
1308 			continue;
1309 		m_run = vm_phys_alloc_seg_contig(seg, npages, low, high,
1310 		    alignment, boundary);
1311 		if (m_run != NULL)
1312 			break;
1313 	}
1314 	return (m_run);
1315 }
1316 
1317 /*
1318  * Allocate a run of contiguous physical pages from the free list for the
1319  * specified segment.
1320  */
1321 static vm_page_t
1322 vm_phys_alloc_seg_contig(struct vm_phys_seg *seg, u_long npages,
1323     vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary)
1324 {
1325 	struct vm_freelist *fl;
1326 	vm_paddr_t pa, pa_end, size;
1327 	vm_page_t m, m_ret;
1328 	u_long npages_end;
1329 	int oind, order, pind;
1330 
1331 	KASSERT(npages > 0, ("npages is 0"));
1332 	KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
1333 	KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
1334 	vm_domain_free_assert_locked(VM_DOMAIN(seg->domain));
1335 	/* Compute the queue that is the best fit for npages. */
1336 	order = flsl(npages - 1);
1337 	/* Search for a run satisfying the specified conditions. */
1338 	size = npages << PAGE_SHIFT;
1339 	for (oind = min(order, VM_NFREEORDER - 1); oind < VM_NFREEORDER;
1340 	    oind++) {
1341 		for (pind = 0; pind < VM_NFREEPOOL; pind++) {
1342 			fl = (*seg->free_queues)[pind];
1343 			TAILQ_FOREACH(m_ret, &fl[oind].pl, listq) {
1344 				/*
1345 				 * Is the size of this allocation request
1346 				 * larger than the largest block size?
1347 				 */
1348 				if (order >= VM_NFREEORDER) {
1349 					/*
1350 					 * Determine if a sufficient number of
1351 					 * subsequent blocks to satisfy the
1352 					 * allocation request are free.
1353 					 */
1354 					pa = VM_PAGE_TO_PHYS(m_ret);
1355 					pa_end = pa + size;
1356 					if (pa_end < pa)
1357 						continue;
1358 					for (;;) {
1359 						pa += 1 << (PAGE_SHIFT +
1360 						    VM_NFREEORDER - 1);
1361 						if (pa >= pa_end ||
1362 						    pa < seg->start ||
1363 						    pa >= seg->end)
1364 							break;
1365 						m = &seg->first_page[atop(pa -
1366 						    seg->start)];
1367 						if (m->order != VM_NFREEORDER -
1368 						    1)
1369 							break;
1370 					}
1371 					/* If not, go to the next block. */
1372 					if (pa < pa_end)
1373 						continue;
1374 				}
1375 
1376 				/*
1377 				 * Determine if the blocks are within the
1378 				 * given range, satisfy the given alignment,
1379 				 * and do not cross the given boundary.
1380 				 */
1381 				pa = VM_PAGE_TO_PHYS(m_ret);
1382 				pa_end = pa + size;
1383 				if (pa >= low && pa_end <= high &&
1384 				    (pa & (alignment - 1)) == 0 &&
1385 				    rounddown2(pa ^ (pa_end - 1), boundary) == 0)
1386 					goto done;
1387 			}
1388 		}
1389 	}
1390 	return (NULL);
1391 done:
1392 	for (m = m_ret; m < &m_ret[npages]; m = &m[1 << oind]) {
1393 		fl = (*seg->free_queues)[m->pool];
1394 		vm_freelist_rem(fl, m, oind);
1395 		if (m->pool != VM_FREEPOOL_DEFAULT)
1396 			vm_phys_set_pool(VM_FREEPOOL_DEFAULT, m, oind);
1397 	}
1398 	/* Return excess pages to the free lists. */
1399 	npages_end = roundup2(npages, 1 << oind);
1400 	if (npages < npages_end) {
1401 		fl = (*seg->free_queues)[VM_FREEPOOL_DEFAULT];
1402 		vm_phys_enq_range(&m_ret[npages], npages_end - npages, fl, 0);
1403 	}
1404 	return (m_ret);
1405 }
1406 
1407 #ifdef DDB
1408 /*
1409  * Show the number of physical pages in each of the free lists.
1410  */
1411 DB_SHOW_COMMAND(freepages, db_show_freepages)
1412 {
1413 	struct vm_freelist *fl;
1414 	int flind, oind, pind, dom;
1415 
1416 	for (dom = 0; dom < vm_ndomains; dom++) {
1417 		db_printf("DOMAIN: %d\n", dom);
1418 		for (flind = 0; flind < vm_nfreelists; flind++) {
1419 			db_printf("FREE LIST %d:\n"
1420 			    "\n  ORDER (SIZE)  |  NUMBER"
1421 			    "\n              ", flind);
1422 			for (pind = 0; pind < VM_NFREEPOOL; pind++)
1423 				db_printf("  |  POOL %d", pind);
1424 			db_printf("\n--            ");
1425 			for (pind = 0; pind < VM_NFREEPOOL; pind++)
1426 				db_printf("-- --      ");
1427 			db_printf("--\n");
1428 			for (oind = VM_NFREEORDER - 1; oind >= 0; oind--) {
1429 				db_printf("  %2.2d (%6.6dK)", oind,
1430 				    1 << (PAGE_SHIFT - 10 + oind));
1431 				for (pind = 0; pind < VM_NFREEPOOL; pind++) {
1432 				fl = vm_phys_free_queues[dom][flind][pind];
1433 					db_printf("  |  %6.6d", fl[oind].lcnt);
1434 				}
1435 				db_printf("\n");
1436 			}
1437 			db_printf("\n");
1438 		}
1439 		db_printf("\n");
1440 	}
1441 }
1442 #endif
1443