xref: /freebsd/sys/vm/vm_phys.c (revision 662e7fa8d9fb296379f594ebb6a7f231c1ea8a2f)
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][VM_NFREEORDER];
110 
111 static int __read_mostly vm_nfreelists;
112 
113 /*
114  * Provides the mapping from VM_FREELIST_* to free list indices (flind).
115  */
116 static int __read_mostly vm_freelist_to_flind[VM_NFREELIST];
117 
118 CTASSERT(VM_FREELIST_DEFAULT == 0);
119 
120 #ifdef VM_FREELIST_DMA32
121 #define	VM_DMA32_BOUNDARY	((vm_paddr_t)1 << 32)
122 #endif
123 
124 /*
125  * Enforce the assumptions made by vm_phys_add_seg() and vm_phys_init() about
126  * the ordering of the free list boundaries.
127  */
128 #if defined(VM_LOWMEM_BOUNDARY) && defined(VM_DMA32_BOUNDARY)
129 CTASSERT(VM_LOWMEM_BOUNDARY < VM_DMA32_BOUNDARY);
130 #endif
131 
132 static int sysctl_vm_phys_free(SYSCTL_HANDLER_ARGS);
133 SYSCTL_OID(_vm, OID_AUTO, phys_free, CTLTYPE_STRING | CTLFLAG_RD,
134     NULL, 0, sysctl_vm_phys_free, "A", "Phys Free Info");
135 
136 static int sysctl_vm_phys_segs(SYSCTL_HANDLER_ARGS);
137 SYSCTL_OID(_vm, OID_AUTO, phys_segs, CTLTYPE_STRING | CTLFLAG_RD,
138     NULL, 0, sysctl_vm_phys_segs, "A", "Phys Seg Info");
139 
140 #ifdef NUMA
141 static int sysctl_vm_phys_locality(SYSCTL_HANDLER_ARGS);
142 SYSCTL_OID(_vm, OID_AUTO, phys_locality, CTLTYPE_STRING | CTLFLAG_RD,
143     NULL, 0, sysctl_vm_phys_locality, "A", "Phys Locality Info");
144 #endif
145 
146 SYSCTL_INT(_vm, OID_AUTO, ndomains, CTLFLAG_RD,
147     &vm_ndomains, 0, "Number of physical memory domains available.");
148 
149 static vm_page_t vm_phys_alloc_seg_contig(struct vm_phys_seg *seg,
150     u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment,
151     vm_paddr_t boundary);
152 static void _vm_phys_create_seg(vm_paddr_t start, vm_paddr_t end, int domain);
153 static void vm_phys_create_seg(vm_paddr_t start, vm_paddr_t end);
154 static void vm_phys_split_pages(vm_page_t m, int oind, struct vm_freelist *fl,
155     int order, int tail);
156 
157 /*
158  * Red-black tree helpers for vm fictitious range management.
159  */
160 static inline int
161 vm_phys_fictitious_in_range(struct vm_phys_fictitious_seg *p,
162     struct vm_phys_fictitious_seg *range)
163 {
164 
165 	KASSERT(range->start != 0 && range->end != 0,
166 	    ("Invalid range passed on search for vm_fictitious page"));
167 	if (p->start >= range->end)
168 		return (1);
169 	if (p->start < range->start)
170 		return (-1);
171 
172 	return (0);
173 }
174 
175 static int
176 vm_phys_fictitious_cmp(struct vm_phys_fictitious_seg *p1,
177     struct vm_phys_fictitious_seg *p2)
178 {
179 
180 	/* Check if this is a search for a page */
181 	if (p1->end == 0)
182 		return (vm_phys_fictitious_in_range(p1, p2));
183 
184 	KASSERT(p2->end != 0,
185     ("Invalid range passed as second parameter to vm fictitious comparison"));
186 
187 	/* Searching to add a new range */
188 	if (p1->end <= p2->start)
189 		return (-1);
190 	if (p1->start >= p2->end)
191 		return (1);
192 
193 	panic("Trying to add overlapping vm fictitious ranges:\n"
194 	    "[%#jx:%#jx] and [%#jx:%#jx]", (uintmax_t)p1->start,
195 	    (uintmax_t)p1->end, (uintmax_t)p2->start, (uintmax_t)p2->end);
196 }
197 
198 int
199 vm_phys_domain_match(int prefer, vm_paddr_t low, vm_paddr_t high)
200 {
201 #ifdef NUMA
202 	domainset_t mask;
203 	int i;
204 
205 	if (vm_ndomains == 1 || mem_affinity == NULL)
206 		return (0);
207 
208 	DOMAINSET_ZERO(&mask);
209 	/*
210 	 * Check for any memory that overlaps low, high.
211 	 */
212 	for (i = 0; mem_affinity[i].end != 0; i++)
213 		if (mem_affinity[i].start <= high &&
214 		    mem_affinity[i].end >= low)
215 			DOMAINSET_SET(mem_affinity[i].domain, &mask);
216 	if (prefer != -1 && DOMAINSET_ISSET(prefer, &mask))
217 		return (prefer);
218 	if (DOMAINSET_EMPTY(&mask))
219 		panic("vm_phys_domain_match:  Impossible constraint");
220 	return (DOMAINSET_FFS(&mask) - 1);
221 #else
222 	return (0);
223 #endif
224 }
225 
226 /*
227  * Outputs the state of the physical memory allocator, specifically,
228  * the amount of physical memory in each free list.
229  */
230 static int
231 sysctl_vm_phys_free(SYSCTL_HANDLER_ARGS)
232 {
233 	struct sbuf sbuf;
234 	struct vm_freelist *fl;
235 	int dom, error, flind, oind, pind;
236 
237 	error = sysctl_wire_old_buffer(req, 0);
238 	if (error != 0)
239 		return (error);
240 	sbuf_new_for_sysctl(&sbuf, NULL, 128 * vm_ndomains, req);
241 	for (dom = 0; dom < vm_ndomains; dom++) {
242 		sbuf_printf(&sbuf,"\nDOMAIN %d:\n", dom);
243 		for (flind = 0; flind < vm_nfreelists; flind++) {
244 			sbuf_printf(&sbuf, "\nFREE LIST %d:\n"
245 			    "\n  ORDER (SIZE)  |  NUMBER"
246 			    "\n              ", flind);
247 			for (pind = 0; pind < VM_NFREEPOOL; pind++)
248 				sbuf_printf(&sbuf, "  |  POOL %d", pind);
249 			sbuf_printf(&sbuf, "\n--            ");
250 			for (pind = 0; pind < VM_NFREEPOOL; pind++)
251 				sbuf_printf(&sbuf, "-- --      ");
252 			sbuf_printf(&sbuf, "--\n");
253 			for (oind = VM_NFREEORDER - 1; oind >= 0; oind--) {
254 				sbuf_printf(&sbuf, "  %2d (%6dK)", oind,
255 				    1 << (PAGE_SHIFT - 10 + oind));
256 				for (pind = 0; pind < VM_NFREEPOOL; pind++) {
257 				fl = vm_phys_free_queues[dom][flind][pind];
258 					sbuf_printf(&sbuf, "  |  %6d",
259 					    fl[oind].lcnt);
260 				}
261 				sbuf_printf(&sbuf, "\n");
262 			}
263 		}
264 	}
265 	error = sbuf_finish(&sbuf);
266 	sbuf_delete(&sbuf);
267 	return (error);
268 }
269 
270 /*
271  * Outputs the set of physical memory segments.
272  */
273 static int
274 sysctl_vm_phys_segs(SYSCTL_HANDLER_ARGS)
275 {
276 	struct sbuf sbuf;
277 	struct vm_phys_seg *seg;
278 	int error, segind;
279 
280 	error = sysctl_wire_old_buffer(req, 0);
281 	if (error != 0)
282 		return (error);
283 	sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
284 	for (segind = 0; segind < vm_phys_nsegs; segind++) {
285 		sbuf_printf(&sbuf, "\nSEGMENT %d:\n\n", segind);
286 		seg = &vm_phys_segs[segind];
287 		sbuf_printf(&sbuf, "start:     %#jx\n",
288 		    (uintmax_t)seg->start);
289 		sbuf_printf(&sbuf, "end:       %#jx\n",
290 		    (uintmax_t)seg->end);
291 		sbuf_printf(&sbuf, "domain:    %d\n", seg->domain);
292 		sbuf_printf(&sbuf, "free list: %p\n", seg->free_queues);
293 	}
294 	error = sbuf_finish(&sbuf);
295 	sbuf_delete(&sbuf);
296 	return (error);
297 }
298 
299 /*
300  * Return affinity, or -1 if there's no affinity information.
301  */
302 int
303 vm_phys_mem_affinity(int f, int t)
304 {
305 
306 #ifdef NUMA
307 	if (mem_locality == NULL)
308 		return (-1);
309 	if (f >= vm_ndomains || t >= vm_ndomains)
310 		return (-1);
311 	return (mem_locality[f * vm_ndomains + t]);
312 #else
313 	return (-1);
314 #endif
315 }
316 
317 #ifdef NUMA
318 /*
319  * Outputs the VM locality table.
320  */
321 static int
322 sysctl_vm_phys_locality(SYSCTL_HANDLER_ARGS)
323 {
324 	struct sbuf sbuf;
325 	int error, i, j;
326 
327 	error = sysctl_wire_old_buffer(req, 0);
328 	if (error != 0)
329 		return (error);
330 	sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
331 
332 	sbuf_printf(&sbuf, "\n");
333 
334 	for (i = 0; i < vm_ndomains; i++) {
335 		sbuf_printf(&sbuf, "%d: ", i);
336 		for (j = 0; j < vm_ndomains; j++) {
337 			sbuf_printf(&sbuf, "%d ", vm_phys_mem_affinity(i, j));
338 		}
339 		sbuf_printf(&sbuf, "\n");
340 	}
341 	error = sbuf_finish(&sbuf);
342 	sbuf_delete(&sbuf);
343 	return (error);
344 }
345 #endif
346 
347 static void
348 vm_freelist_add(struct vm_freelist *fl, vm_page_t m, int order, int tail)
349 {
350 
351 	m->order = order;
352 	if (tail)
353 		TAILQ_INSERT_TAIL(&fl[order].pl, m, listq);
354 	else
355 		TAILQ_INSERT_HEAD(&fl[order].pl, m, listq);
356 	fl[order].lcnt++;
357 }
358 
359 static void
360 vm_freelist_rem(struct vm_freelist *fl, vm_page_t m, int order)
361 {
362 
363 	TAILQ_REMOVE(&fl[order].pl, m, listq);
364 	fl[order].lcnt--;
365 	m->order = VM_NFREEORDER;
366 }
367 
368 /*
369  * Create a physical memory segment.
370  */
371 static void
372 _vm_phys_create_seg(vm_paddr_t start, vm_paddr_t end, int domain)
373 {
374 	struct vm_phys_seg *seg;
375 
376 	KASSERT(vm_phys_nsegs < VM_PHYSSEG_MAX,
377 	    ("vm_phys_create_seg: increase VM_PHYSSEG_MAX"));
378 	KASSERT(domain >= 0 && domain < vm_ndomains,
379 	    ("vm_phys_create_seg: invalid domain provided"));
380 	seg = &vm_phys_segs[vm_phys_nsegs++];
381 	while (seg > vm_phys_segs && (seg - 1)->start >= end) {
382 		*seg = *(seg - 1);
383 		seg--;
384 	}
385 	seg->start = start;
386 	seg->end = end;
387 	seg->domain = domain;
388 }
389 
390 static void
391 vm_phys_create_seg(vm_paddr_t start, vm_paddr_t end)
392 {
393 #ifdef NUMA
394 	int i;
395 
396 	if (mem_affinity == NULL) {
397 		_vm_phys_create_seg(start, end, 0);
398 		return;
399 	}
400 
401 	for (i = 0;; i++) {
402 		if (mem_affinity[i].end == 0)
403 			panic("Reached end of affinity info");
404 		if (mem_affinity[i].end <= start)
405 			continue;
406 		if (mem_affinity[i].start > start)
407 			panic("No affinity info for start %jx",
408 			    (uintmax_t)start);
409 		if (mem_affinity[i].end >= end) {
410 			_vm_phys_create_seg(start, end,
411 			    mem_affinity[i].domain);
412 			break;
413 		}
414 		_vm_phys_create_seg(start, mem_affinity[i].end,
415 		    mem_affinity[i].domain);
416 		start = mem_affinity[i].end;
417 	}
418 #else
419 	_vm_phys_create_seg(start, end, 0);
420 #endif
421 }
422 
423 /*
424  * Add a physical memory segment.
425  */
426 void
427 vm_phys_add_seg(vm_paddr_t start, vm_paddr_t end)
428 {
429 	vm_paddr_t paddr;
430 
431 	KASSERT((start & PAGE_MASK) == 0,
432 	    ("vm_phys_define_seg: start is not page aligned"));
433 	KASSERT((end & PAGE_MASK) == 0,
434 	    ("vm_phys_define_seg: end is not page aligned"));
435 
436 	/*
437 	 * Split the physical memory segment if it spans two or more free
438 	 * list boundaries.
439 	 */
440 	paddr = start;
441 #ifdef	VM_FREELIST_LOWMEM
442 	if (paddr < VM_LOWMEM_BOUNDARY && end > VM_LOWMEM_BOUNDARY) {
443 		vm_phys_create_seg(paddr, VM_LOWMEM_BOUNDARY);
444 		paddr = VM_LOWMEM_BOUNDARY;
445 	}
446 #endif
447 #ifdef	VM_FREELIST_DMA32
448 	if (paddr < VM_DMA32_BOUNDARY && end > VM_DMA32_BOUNDARY) {
449 		vm_phys_create_seg(paddr, VM_DMA32_BOUNDARY);
450 		paddr = VM_DMA32_BOUNDARY;
451 	}
452 #endif
453 	vm_phys_create_seg(paddr, end);
454 }
455 
456 /*
457  * Initialize the physical memory allocator.
458  *
459  * Requires that vm_page_array is initialized!
460  */
461 void
462 vm_phys_init(void)
463 {
464 	struct vm_freelist *fl;
465 	struct vm_phys_seg *end_seg, *prev_seg, *seg, *tmp_seg;
466 	u_long npages;
467 	int dom, flind, freelist, oind, pind, segind;
468 
469 	/*
470 	 * Compute the number of free lists, and generate the mapping from the
471 	 * manifest constants VM_FREELIST_* to the free list indices.
472 	 *
473 	 * Initially, the entries of vm_freelist_to_flind[] are set to either
474 	 * 0 or 1 to indicate which free lists should be created.
475 	 */
476 	npages = 0;
477 	for (segind = vm_phys_nsegs - 1; segind >= 0; segind--) {
478 		seg = &vm_phys_segs[segind];
479 #ifdef	VM_FREELIST_LOWMEM
480 		if (seg->end <= VM_LOWMEM_BOUNDARY)
481 			vm_freelist_to_flind[VM_FREELIST_LOWMEM] = 1;
482 		else
483 #endif
484 #ifdef	VM_FREELIST_DMA32
485 		if (
486 #ifdef	VM_DMA32_NPAGES_THRESHOLD
487 		    /*
488 		     * Create the DMA32 free list only if the amount of
489 		     * physical memory above physical address 4G exceeds the
490 		     * given threshold.
491 		     */
492 		    npages > VM_DMA32_NPAGES_THRESHOLD &&
493 #endif
494 		    seg->end <= VM_DMA32_BOUNDARY)
495 			vm_freelist_to_flind[VM_FREELIST_DMA32] = 1;
496 		else
497 #endif
498 		{
499 			npages += atop(seg->end - seg->start);
500 			vm_freelist_to_flind[VM_FREELIST_DEFAULT] = 1;
501 		}
502 	}
503 	/* Change each entry into a running total of the free lists. */
504 	for (freelist = 1; freelist < VM_NFREELIST; freelist++) {
505 		vm_freelist_to_flind[freelist] +=
506 		    vm_freelist_to_flind[freelist - 1];
507 	}
508 	vm_nfreelists = vm_freelist_to_flind[VM_NFREELIST - 1];
509 	KASSERT(vm_nfreelists > 0, ("vm_phys_init: no free lists"));
510 	/* Change each entry into a free list index. */
511 	for (freelist = 0; freelist < VM_NFREELIST; freelist++)
512 		vm_freelist_to_flind[freelist]--;
513 
514 	/*
515 	 * Initialize the first_page and free_queues fields of each physical
516 	 * memory segment.
517 	 */
518 #ifdef VM_PHYSSEG_SPARSE
519 	npages = 0;
520 #endif
521 	for (segind = 0; segind < vm_phys_nsegs; segind++) {
522 		seg = &vm_phys_segs[segind];
523 #ifdef VM_PHYSSEG_SPARSE
524 		seg->first_page = &vm_page_array[npages];
525 		npages += atop(seg->end - seg->start);
526 #else
527 		seg->first_page = PHYS_TO_VM_PAGE(seg->start);
528 #endif
529 #ifdef	VM_FREELIST_LOWMEM
530 		if (seg->end <= VM_LOWMEM_BOUNDARY) {
531 			flind = vm_freelist_to_flind[VM_FREELIST_LOWMEM];
532 			KASSERT(flind >= 0,
533 			    ("vm_phys_init: LOWMEM flind < 0"));
534 		} else
535 #endif
536 #ifdef	VM_FREELIST_DMA32
537 		if (seg->end <= VM_DMA32_BOUNDARY) {
538 			flind = vm_freelist_to_flind[VM_FREELIST_DMA32];
539 			KASSERT(flind >= 0,
540 			    ("vm_phys_init: DMA32 flind < 0"));
541 		} else
542 #endif
543 		{
544 			flind = vm_freelist_to_flind[VM_FREELIST_DEFAULT];
545 			KASSERT(flind >= 0,
546 			    ("vm_phys_init: DEFAULT flind < 0"));
547 		}
548 		seg->free_queues = &vm_phys_free_queues[seg->domain][flind];
549 	}
550 
551 	/*
552 	 * Coalesce physical memory segments that are contiguous and share the
553 	 * same per-domain free queues.
554 	 */
555 	prev_seg = vm_phys_segs;
556 	seg = &vm_phys_segs[1];
557 	end_seg = &vm_phys_segs[vm_phys_nsegs];
558 	while (seg < end_seg) {
559 		if (prev_seg->end == seg->start &&
560 		    prev_seg->free_queues == seg->free_queues) {
561 			prev_seg->end = seg->end;
562 			KASSERT(prev_seg->domain == seg->domain,
563 			    ("vm_phys_init: free queues cannot span domains"));
564 			vm_phys_nsegs--;
565 			end_seg--;
566 			for (tmp_seg = seg; tmp_seg < end_seg; tmp_seg++)
567 				*tmp_seg = *(tmp_seg + 1);
568 		} else {
569 			prev_seg = seg;
570 			seg++;
571 		}
572 	}
573 
574 	/*
575 	 * Initialize the free queues.
576 	 */
577 	for (dom = 0; dom < vm_ndomains; dom++) {
578 		for (flind = 0; flind < vm_nfreelists; flind++) {
579 			for (pind = 0; pind < VM_NFREEPOOL; pind++) {
580 				fl = vm_phys_free_queues[dom][flind][pind];
581 				for (oind = 0; oind < VM_NFREEORDER; oind++)
582 					TAILQ_INIT(&fl[oind].pl);
583 			}
584 		}
585 	}
586 
587 	rw_init(&vm_phys_fictitious_reg_lock, "vmfctr");
588 }
589 
590 /*
591  * Register info about the NUMA topology of the system.
592  *
593  * Invoked by platform-dependent code prior to vm_phys_init().
594  */
595 void
596 vm_phys_register_domains(int ndomains, struct mem_affinity *affinity,
597     int *locality)
598 {
599 #ifdef NUMA
600 	int i;
601 
602 	vm_ndomains = ndomains;
603 	mem_affinity = affinity;
604 	mem_locality = locality;
605 
606 	for (i = 0; i < vm_ndomains; i++)
607 		DOMAINSET_SET(i, &all_domains);
608 
609 	domainset_init();
610 #else
611 	(void)ndomains;
612 	(void)affinity;
613 	(void)locality;
614 #endif
615 }
616 
617 /*
618  * Split a contiguous, power of two-sized set of physical pages.
619  *
620  * When this function is called by a page allocation function, the caller
621  * should request insertion at the head unless the order [order, oind) queues
622  * are known to be empty.  The objective being to reduce the likelihood of
623  * long-term fragmentation by promoting contemporaneous allocation and
624  * (hopefully) deallocation.
625  */
626 static __inline void
627 vm_phys_split_pages(vm_page_t m, int oind, struct vm_freelist *fl, int order,
628     int tail)
629 {
630 	vm_page_t m_buddy;
631 
632 	while (oind > order) {
633 		oind--;
634 		m_buddy = &m[1 << oind];
635 		KASSERT(m_buddy->order == VM_NFREEORDER,
636 		    ("vm_phys_split_pages: page %p has unexpected order %d",
637 		    m_buddy, m_buddy->order));
638 		vm_freelist_add(fl, m_buddy, oind, tail);
639         }
640 }
641 
642 /*
643  * Add the physical pages [m, m + npages) at the end of a power-of-two aligned
644  * and sized set to the specified free list.
645  *
646  * When this function is called by a page allocation function, the caller
647  * should request insertion at the head unless the lower-order queues are
648  * known to be empty.  The objective being to reduce the likelihood of long-
649  * term fragmentation by promoting contemporaneous allocation and (hopefully)
650  * deallocation.
651  *
652  * The physical page m's buddy must not be free.
653  */
654 static void
655 vm_phys_enq_range(vm_page_t m, u_int npages, struct vm_freelist *fl, int tail)
656 {
657 	u_int n;
658 	int order;
659 
660 	KASSERT(npages > 0, ("vm_phys_enq_range: npages is 0"));
661 	KASSERT(((VM_PAGE_TO_PHYS(m) + npages * PAGE_SIZE) &
662 	    ((PAGE_SIZE << (fls(npages) - 1)) - 1)) == 0,
663 	    ("vm_phys_enq_range: page %p and npages %u are misaligned",
664 	    m, npages));
665 	do {
666 		KASSERT(m->order == VM_NFREEORDER,
667 		    ("vm_phys_enq_range: page %p has unexpected order %d",
668 		    m, m->order));
669 		order = ffs(npages) - 1;
670 		KASSERT(order < VM_NFREEORDER,
671 		    ("vm_phys_enq_range: order %d is out of range", order));
672 		vm_freelist_add(fl, m, order, tail);
673 		n = 1 << order;
674 		m += n;
675 		npages -= n;
676 	} while (npages > 0);
677 }
678 
679 /*
680  * Tries to allocate the specified number of pages from the specified pool
681  * within the specified domain.  Returns the actual number of allocated pages
682  * and a pointer to each page through the array ma[].
683  *
684  * The returned pages may not be physically contiguous.  However, in contrast
685  * to performing multiple, back-to-back calls to vm_phys_alloc_pages(..., 0),
686  * calling this function once to allocate the desired number of pages will
687  * avoid wasted time in vm_phys_split_pages().
688  *
689  * The free page queues for the specified domain must be locked.
690  */
691 int
692 vm_phys_alloc_npages(int domain, int pool, int npages, vm_page_t ma[])
693 {
694 	struct vm_freelist *alt, *fl;
695 	vm_page_t m;
696 	int avail, end, flind, freelist, i, need, oind, pind;
697 
698 	KASSERT(domain >= 0 && domain < vm_ndomains,
699 	    ("vm_phys_alloc_npages: domain %d is out of range", domain));
700 	KASSERT(pool < VM_NFREEPOOL,
701 	    ("vm_phys_alloc_npages: pool %d is out of range", pool));
702 	KASSERT(npages <= 1 << (VM_NFREEORDER - 1),
703 	    ("vm_phys_alloc_npages: npages %d is out of range", npages));
704 	vm_domain_free_assert_locked(VM_DOMAIN(domain));
705 	i = 0;
706 	for (freelist = 0; freelist < VM_NFREELIST; freelist++) {
707 		flind = vm_freelist_to_flind[freelist];
708 		if (flind < 0)
709 			continue;
710 		fl = vm_phys_free_queues[domain][flind][pool];
711 		for (oind = 0; oind < VM_NFREEORDER; oind++) {
712 			while ((m = TAILQ_FIRST(&fl[oind].pl)) != NULL) {
713 				vm_freelist_rem(fl, m, oind);
714 				avail = 1 << oind;
715 				need = imin(npages - i, avail);
716 				for (end = i + need; i < end;)
717 					ma[i++] = m++;
718 				if (need < avail) {
719 					/*
720 					 * Return excess pages to fl.  Its
721 					 * order [0, oind) queues are empty.
722 					 */
723 					vm_phys_enq_range(m, avail - need, fl,
724 					    1);
725 					return (npages);
726 				} else if (i == npages)
727 					return (npages);
728 			}
729 		}
730 		for (oind = VM_NFREEORDER - 1; oind >= 0; oind--) {
731 			for (pind = 0; pind < VM_NFREEPOOL; pind++) {
732 				alt = vm_phys_free_queues[domain][flind][pind];
733 				while ((m = TAILQ_FIRST(&alt[oind].pl)) !=
734 				    NULL) {
735 					vm_freelist_rem(alt, m, oind);
736 					vm_phys_set_pool(pool, m, oind);
737 					avail = 1 << oind;
738 					need = imin(npages - i, avail);
739 					for (end = i + need; i < end;)
740 						ma[i++] = m++;
741 					if (need < avail) {
742 						/*
743 						 * Return excess pages to fl.
744 						 * Its order [0, oind) queues
745 						 * are empty.
746 						 */
747 						vm_phys_enq_range(m, avail -
748 						    need, fl, 1);
749 						return (npages);
750 					} else if (i == npages)
751 						return (npages);
752 				}
753 			}
754 		}
755 	}
756 	return (i);
757 }
758 
759 /*
760  * Allocate a contiguous, power of two-sized set of physical pages
761  * from the free lists.
762  *
763  * The free page queues must be locked.
764  */
765 vm_page_t
766 vm_phys_alloc_pages(int domain, int pool, int order)
767 {
768 	vm_page_t m;
769 	int freelist;
770 
771 	for (freelist = 0; freelist < VM_NFREELIST; freelist++) {
772 		m = vm_phys_alloc_freelist_pages(domain, freelist, pool, order);
773 		if (m != NULL)
774 			return (m);
775 	}
776 	return (NULL);
777 }
778 
779 /*
780  * Allocate a contiguous, power of two-sized set of physical pages from the
781  * specified free list.  The free list must be specified using one of the
782  * manifest constants VM_FREELIST_*.
783  *
784  * The free page queues must be locked.
785  */
786 vm_page_t
787 vm_phys_alloc_freelist_pages(int domain, int freelist, int pool, int order)
788 {
789 	struct vm_freelist *alt, *fl;
790 	vm_page_t m;
791 	int oind, pind, flind;
792 
793 	KASSERT(domain >= 0 && domain < vm_ndomains,
794 	    ("vm_phys_alloc_freelist_pages: domain %d is out of range",
795 	    domain));
796 	KASSERT(freelist < VM_NFREELIST,
797 	    ("vm_phys_alloc_freelist_pages: freelist %d is out of range",
798 	    freelist));
799 	KASSERT(pool < VM_NFREEPOOL,
800 	    ("vm_phys_alloc_freelist_pages: pool %d is out of range", pool));
801 	KASSERT(order < VM_NFREEORDER,
802 	    ("vm_phys_alloc_freelist_pages: order %d is out of range", order));
803 
804 	flind = vm_freelist_to_flind[freelist];
805 	/* Check if freelist is present */
806 	if (flind < 0)
807 		return (NULL);
808 
809 	vm_domain_free_assert_locked(VM_DOMAIN(domain));
810 	fl = &vm_phys_free_queues[domain][flind][pool][0];
811 	for (oind = order; oind < VM_NFREEORDER; oind++) {
812 		m = TAILQ_FIRST(&fl[oind].pl);
813 		if (m != NULL) {
814 			vm_freelist_rem(fl, m, oind);
815 			/* The order [order, oind) queues are empty. */
816 			vm_phys_split_pages(m, oind, fl, order, 1);
817 			return (m);
818 		}
819 	}
820 
821 	/*
822 	 * The given pool was empty.  Find the largest
823 	 * contiguous, power-of-two-sized set of pages in any
824 	 * pool.  Transfer these pages to the given pool, and
825 	 * use them to satisfy the allocation.
826 	 */
827 	for (oind = VM_NFREEORDER - 1; oind >= order; oind--) {
828 		for (pind = 0; pind < VM_NFREEPOOL; pind++) {
829 			alt = &vm_phys_free_queues[domain][flind][pind][0];
830 			m = TAILQ_FIRST(&alt[oind].pl);
831 			if (m != NULL) {
832 				vm_freelist_rem(alt, m, oind);
833 				vm_phys_set_pool(pool, m, oind);
834 				/* The order [order, oind) queues are empty. */
835 				vm_phys_split_pages(m, oind, fl, order, 1);
836 				return (m);
837 			}
838 		}
839 	}
840 	return (NULL);
841 }
842 
843 /*
844  * Find the vm_page corresponding to the given physical address.
845  */
846 vm_page_t
847 vm_phys_paddr_to_vm_page(vm_paddr_t pa)
848 {
849 	struct vm_phys_seg *seg;
850 	int segind;
851 
852 	for (segind = 0; segind < vm_phys_nsegs; segind++) {
853 		seg = &vm_phys_segs[segind];
854 		if (pa >= seg->start && pa < seg->end)
855 			return (&seg->first_page[atop(pa - seg->start)]);
856 	}
857 	return (NULL);
858 }
859 
860 vm_page_t
861 vm_phys_fictitious_to_vm_page(vm_paddr_t pa)
862 {
863 	struct vm_phys_fictitious_seg tmp, *seg;
864 	vm_page_t m;
865 
866 	m = NULL;
867 	tmp.start = pa;
868 	tmp.end = 0;
869 
870 	rw_rlock(&vm_phys_fictitious_reg_lock);
871 	seg = RB_FIND(fict_tree, &vm_phys_fictitious_tree, &tmp);
872 	rw_runlock(&vm_phys_fictitious_reg_lock);
873 	if (seg == NULL)
874 		return (NULL);
875 
876 	m = &seg->first_page[atop(pa - seg->start)];
877 	KASSERT((m->flags & PG_FICTITIOUS) != 0, ("%p not fictitious", m));
878 
879 	return (m);
880 }
881 
882 static inline void
883 vm_phys_fictitious_init_range(vm_page_t range, vm_paddr_t start,
884     long page_count, vm_memattr_t memattr)
885 {
886 	long i;
887 
888 	bzero(range, page_count * sizeof(*range));
889 	for (i = 0; i < page_count; i++) {
890 		vm_page_initfake(&range[i], start + PAGE_SIZE * i, memattr);
891 		range[i].oflags &= ~VPO_UNMANAGED;
892 		range[i].busy_lock = VPB_UNBUSIED;
893 	}
894 }
895 
896 int
897 vm_phys_fictitious_reg_range(vm_paddr_t start, vm_paddr_t end,
898     vm_memattr_t memattr)
899 {
900 	struct vm_phys_fictitious_seg *seg;
901 	vm_page_t fp;
902 	long page_count;
903 #ifdef VM_PHYSSEG_DENSE
904 	long pi, pe;
905 	long dpage_count;
906 #endif
907 
908 	KASSERT(start < end,
909 	    ("Start of segment isn't less than end (start: %jx end: %jx)",
910 	    (uintmax_t)start, (uintmax_t)end));
911 
912 	page_count = (end - start) / PAGE_SIZE;
913 
914 #ifdef VM_PHYSSEG_DENSE
915 	pi = atop(start);
916 	pe = atop(end);
917 	if (pi >= first_page && (pi - first_page) < vm_page_array_size) {
918 		fp = &vm_page_array[pi - first_page];
919 		if ((pe - first_page) > vm_page_array_size) {
920 			/*
921 			 * We have a segment that starts inside
922 			 * of vm_page_array, but ends outside of it.
923 			 *
924 			 * Use vm_page_array pages for those that are
925 			 * inside of the vm_page_array range, and
926 			 * allocate the remaining ones.
927 			 */
928 			dpage_count = vm_page_array_size - (pi - first_page);
929 			vm_phys_fictitious_init_range(fp, start, dpage_count,
930 			    memattr);
931 			page_count -= dpage_count;
932 			start += ptoa(dpage_count);
933 			goto alloc;
934 		}
935 		/*
936 		 * We can allocate the full range from vm_page_array,
937 		 * so there's no need to register the range in the tree.
938 		 */
939 		vm_phys_fictitious_init_range(fp, start, page_count, memattr);
940 		return (0);
941 	} else if (pe > first_page && (pe - first_page) < vm_page_array_size) {
942 		/*
943 		 * We have a segment that ends inside of vm_page_array,
944 		 * but starts outside of it.
945 		 */
946 		fp = &vm_page_array[0];
947 		dpage_count = pe - first_page;
948 		vm_phys_fictitious_init_range(fp, ptoa(first_page), dpage_count,
949 		    memattr);
950 		end -= ptoa(dpage_count);
951 		page_count -= dpage_count;
952 		goto alloc;
953 	} else if (pi < first_page && pe > (first_page + vm_page_array_size)) {
954 		/*
955 		 * Trying to register a fictitious range that expands before
956 		 * and after vm_page_array.
957 		 */
958 		return (EINVAL);
959 	} else {
960 alloc:
961 #endif
962 		fp = malloc(page_count * sizeof(struct vm_page), M_FICT_PAGES,
963 		    M_WAITOK);
964 #ifdef VM_PHYSSEG_DENSE
965 	}
966 #endif
967 	vm_phys_fictitious_init_range(fp, start, page_count, memattr);
968 
969 	seg = malloc(sizeof(*seg), M_FICT_PAGES, M_WAITOK | M_ZERO);
970 	seg->start = start;
971 	seg->end = end;
972 	seg->first_page = fp;
973 
974 	rw_wlock(&vm_phys_fictitious_reg_lock);
975 	RB_INSERT(fict_tree, &vm_phys_fictitious_tree, seg);
976 	rw_wunlock(&vm_phys_fictitious_reg_lock);
977 
978 	return (0);
979 }
980 
981 void
982 vm_phys_fictitious_unreg_range(vm_paddr_t start, vm_paddr_t end)
983 {
984 	struct vm_phys_fictitious_seg *seg, tmp;
985 #ifdef VM_PHYSSEG_DENSE
986 	long pi, pe;
987 #endif
988 
989 	KASSERT(start < end,
990 	    ("Start of segment isn't less than end (start: %jx end: %jx)",
991 	    (uintmax_t)start, (uintmax_t)end));
992 
993 #ifdef VM_PHYSSEG_DENSE
994 	pi = atop(start);
995 	pe = atop(end);
996 	if (pi >= first_page && (pi - first_page) < vm_page_array_size) {
997 		if ((pe - first_page) <= vm_page_array_size) {
998 			/*
999 			 * This segment was allocated using vm_page_array
1000 			 * only, there's nothing to do since those pages
1001 			 * were never added to the tree.
1002 			 */
1003 			return;
1004 		}
1005 		/*
1006 		 * We have a segment that starts inside
1007 		 * of vm_page_array, but ends outside of it.
1008 		 *
1009 		 * Calculate how many pages were added to the
1010 		 * tree and free them.
1011 		 */
1012 		start = ptoa(first_page + vm_page_array_size);
1013 	} else if (pe > first_page && (pe - first_page) < vm_page_array_size) {
1014 		/*
1015 		 * We have a segment that ends inside of vm_page_array,
1016 		 * but starts outside of it.
1017 		 */
1018 		end = ptoa(first_page);
1019 	} else if (pi < first_page && pe > (first_page + vm_page_array_size)) {
1020 		/* Since it's not possible to register such a range, panic. */
1021 		panic(
1022 		    "Unregistering not registered fictitious range [%#jx:%#jx]",
1023 		    (uintmax_t)start, (uintmax_t)end);
1024 	}
1025 #endif
1026 	tmp.start = start;
1027 	tmp.end = 0;
1028 
1029 	rw_wlock(&vm_phys_fictitious_reg_lock);
1030 	seg = RB_FIND(fict_tree, &vm_phys_fictitious_tree, &tmp);
1031 	if (seg->start != start || seg->end != end) {
1032 		rw_wunlock(&vm_phys_fictitious_reg_lock);
1033 		panic(
1034 		    "Unregistering not registered fictitious range [%#jx:%#jx]",
1035 		    (uintmax_t)start, (uintmax_t)end);
1036 	}
1037 	RB_REMOVE(fict_tree, &vm_phys_fictitious_tree, seg);
1038 	rw_wunlock(&vm_phys_fictitious_reg_lock);
1039 	free(seg->first_page, M_FICT_PAGES);
1040 	free(seg, M_FICT_PAGES);
1041 }
1042 
1043 /*
1044  * Free a contiguous, power of two-sized set of physical pages.
1045  *
1046  * The free page queues must be locked.
1047  */
1048 void
1049 vm_phys_free_pages(vm_page_t m, int order)
1050 {
1051 	struct vm_freelist *fl;
1052 	struct vm_phys_seg *seg;
1053 	vm_paddr_t pa;
1054 	vm_page_t m_buddy;
1055 
1056 	KASSERT(m->order == VM_NFREEORDER,
1057 	    ("vm_phys_free_pages: page %p has unexpected order %d",
1058 	    m, m->order));
1059 	KASSERT(m->pool < VM_NFREEPOOL,
1060 	    ("vm_phys_free_pages: page %p has unexpected pool %d",
1061 	    m, m->pool));
1062 	KASSERT(order < VM_NFREEORDER,
1063 	    ("vm_phys_free_pages: order %d is out of range", order));
1064 	seg = &vm_phys_segs[m->segind];
1065 	vm_domain_free_assert_locked(VM_DOMAIN(seg->domain));
1066 	if (order < VM_NFREEORDER - 1) {
1067 		pa = VM_PAGE_TO_PHYS(m);
1068 		do {
1069 			pa ^= ((vm_paddr_t)1 << (PAGE_SHIFT + order));
1070 			if (pa < seg->start || pa >= seg->end)
1071 				break;
1072 			m_buddy = &seg->first_page[atop(pa - seg->start)];
1073 			if (m_buddy->order != order)
1074 				break;
1075 			fl = (*seg->free_queues)[m_buddy->pool];
1076 			vm_freelist_rem(fl, m_buddy, order);
1077 			if (m_buddy->pool != m->pool)
1078 				vm_phys_set_pool(m->pool, m_buddy, order);
1079 			order++;
1080 			pa &= ~(((vm_paddr_t)1 << (PAGE_SHIFT + order)) - 1);
1081 			m = &seg->first_page[atop(pa - seg->start)];
1082 		} while (order < VM_NFREEORDER - 1);
1083 	}
1084 	fl = (*seg->free_queues)[m->pool];
1085 	vm_freelist_add(fl, m, order, 1);
1086 }
1087 
1088 /*
1089  * Free a contiguous, arbitrarily sized set of physical pages.
1090  *
1091  * The free page queues must be locked.
1092  */
1093 void
1094 vm_phys_free_contig(vm_page_t m, u_long npages)
1095 {
1096 	u_int n;
1097 	int order;
1098 
1099 	/*
1100 	 * Avoid unnecessary coalescing by freeing the pages in the largest
1101 	 * possible power-of-two-sized subsets.
1102 	 */
1103 	vm_domain_free_assert_locked(vm_pagequeue_domain(m));
1104 	for (;; npages -= n) {
1105 		/*
1106 		 * Unsigned "min" is used here so that "order" is assigned
1107 		 * "VM_NFREEORDER - 1" when "m"'s physical address is zero
1108 		 * or the low-order bits of its physical address are zero
1109 		 * because the size of a physical address exceeds the size of
1110 		 * a long.
1111 		 */
1112 		order = min(ffsl(VM_PAGE_TO_PHYS(m) >> PAGE_SHIFT) - 1,
1113 		    VM_NFREEORDER - 1);
1114 		n = 1 << order;
1115 		if (npages < n)
1116 			break;
1117 		vm_phys_free_pages(m, order);
1118 		m += n;
1119 	}
1120 	/* The residual "npages" is less than "1 << (VM_NFREEORDER - 1)". */
1121 	for (; npages > 0; npages -= n) {
1122 		order = flsl(npages) - 1;
1123 		n = 1 << order;
1124 		vm_phys_free_pages(m, order);
1125 		m += n;
1126 	}
1127 }
1128 
1129 /*
1130  * Scan physical memory between the specified addresses "low" and "high" for a
1131  * run of contiguous physical pages that satisfy the specified conditions, and
1132  * return the lowest page in the run.  The specified "alignment" determines
1133  * the alignment of the lowest physical page in the run.  If the specified
1134  * "boundary" is non-zero, then the run of physical pages cannot span a
1135  * physical address that is a multiple of "boundary".
1136  *
1137  * "npages" must be greater than zero.  Both "alignment" and "boundary" must
1138  * be a power of two.
1139  */
1140 vm_page_t
1141 vm_phys_scan_contig(int domain, u_long npages, vm_paddr_t low, vm_paddr_t high,
1142     u_long alignment, vm_paddr_t boundary, int options)
1143 {
1144 	vm_paddr_t pa_end;
1145 	vm_page_t m_end, m_run, m_start;
1146 	struct vm_phys_seg *seg;
1147 	int segind;
1148 
1149 	KASSERT(npages > 0, ("npages is 0"));
1150 	KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
1151 	KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
1152 	if (low >= high)
1153 		return (NULL);
1154 	for (segind = 0; segind < vm_phys_nsegs; segind++) {
1155 		seg = &vm_phys_segs[segind];
1156 		if (seg->domain != domain)
1157 			continue;
1158 		if (seg->start >= high)
1159 			break;
1160 		if (low >= seg->end)
1161 			continue;
1162 		if (low <= seg->start)
1163 			m_start = seg->first_page;
1164 		else
1165 			m_start = &seg->first_page[atop(low - seg->start)];
1166 		if (high < seg->end)
1167 			pa_end = high;
1168 		else
1169 			pa_end = seg->end;
1170 		if (pa_end - VM_PAGE_TO_PHYS(m_start) < ptoa(npages))
1171 			continue;
1172 		m_end = &seg->first_page[atop(pa_end - seg->start)];
1173 		m_run = vm_page_scan_contig(npages, m_start, m_end,
1174 		    alignment, boundary, options);
1175 		if (m_run != NULL)
1176 			return (m_run);
1177 	}
1178 	return (NULL);
1179 }
1180 
1181 /*
1182  * Set the pool for a contiguous, power of two-sized set of physical pages.
1183  */
1184 void
1185 vm_phys_set_pool(int pool, vm_page_t m, int order)
1186 {
1187 	vm_page_t m_tmp;
1188 
1189 	for (m_tmp = m; m_tmp < &m[1 << order]; m_tmp++)
1190 		m_tmp->pool = pool;
1191 }
1192 
1193 /*
1194  * Search for the given physical page "m" in the free lists.  If the search
1195  * succeeds, remove "m" from the free lists and return TRUE.  Otherwise, return
1196  * FALSE, indicating that "m" is not in the free lists.
1197  *
1198  * The free page queues must be locked.
1199  */
1200 boolean_t
1201 vm_phys_unfree_page(vm_page_t m)
1202 {
1203 	struct vm_freelist *fl;
1204 	struct vm_phys_seg *seg;
1205 	vm_paddr_t pa, pa_half;
1206 	vm_page_t m_set, m_tmp;
1207 	int order;
1208 
1209 	/*
1210 	 * First, find the contiguous, power of two-sized set of free
1211 	 * physical pages containing the given physical page "m" and
1212 	 * assign it to "m_set".
1213 	 */
1214 	seg = &vm_phys_segs[m->segind];
1215 	vm_domain_free_assert_locked(VM_DOMAIN(seg->domain));
1216 	for (m_set = m, order = 0; m_set->order == VM_NFREEORDER &&
1217 	    order < VM_NFREEORDER - 1; ) {
1218 		order++;
1219 		pa = m->phys_addr & (~(vm_paddr_t)0 << (PAGE_SHIFT + order));
1220 		if (pa >= seg->start)
1221 			m_set = &seg->first_page[atop(pa - seg->start)];
1222 		else
1223 			return (FALSE);
1224 	}
1225 	if (m_set->order < order)
1226 		return (FALSE);
1227 	if (m_set->order == VM_NFREEORDER)
1228 		return (FALSE);
1229 	KASSERT(m_set->order < VM_NFREEORDER,
1230 	    ("vm_phys_unfree_page: page %p has unexpected order %d",
1231 	    m_set, m_set->order));
1232 
1233 	/*
1234 	 * Next, remove "m_set" from the free lists.  Finally, extract
1235 	 * "m" from "m_set" using an iterative algorithm: While "m_set"
1236 	 * is larger than a page, shrink "m_set" by returning the half
1237 	 * of "m_set" that does not contain "m" to the free lists.
1238 	 */
1239 	fl = (*seg->free_queues)[m_set->pool];
1240 	order = m_set->order;
1241 	vm_freelist_rem(fl, m_set, order);
1242 	while (order > 0) {
1243 		order--;
1244 		pa_half = m_set->phys_addr ^ (1 << (PAGE_SHIFT + order));
1245 		if (m->phys_addr < pa_half)
1246 			m_tmp = &seg->first_page[atop(pa_half - seg->start)];
1247 		else {
1248 			m_tmp = m_set;
1249 			m_set = &seg->first_page[atop(pa_half - seg->start)];
1250 		}
1251 		vm_freelist_add(fl, m_tmp, order, 0);
1252 	}
1253 	KASSERT(m_set == m, ("vm_phys_unfree_page: fatal inconsistency"));
1254 	return (TRUE);
1255 }
1256 
1257 /*
1258  * Allocate a contiguous set of physical pages of the given size
1259  * "npages" from the free lists.  All of the physical pages must be at
1260  * or above the given physical address "low" and below the given
1261  * physical address "high".  The given value "alignment" determines the
1262  * alignment of the first physical page in the set.  If the given value
1263  * "boundary" is non-zero, then the set of physical pages cannot cross
1264  * any physical address boundary that is a multiple of that value.  Both
1265  * "alignment" and "boundary" must be a power of two.
1266  */
1267 vm_page_t
1268 vm_phys_alloc_contig(int domain, u_long npages, vm_paddr_t low, vm_paddr_t high,
1269     u_long alignment, vm_paddr_t boundary)
1270 {
1271 	vm_paddr_t pa_end, pa_start;
1272 	vm_page_t m_run;
1273 	struct vm_phys_seg *seg;
1274 	int segind;
1275 
1276 	KASSERT(npages > 0, ("npages is 0"));
1277 	KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
1278 	KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
1279 	vm_domain_free_assert_locked(VM_DOMAIN(domain));
1280 	if (low >= high)
1281 		return (NULL);
1282 	m_run = NULL;
1283 	for (segind = vm_phys_nsegs - 1; segind >= 0; segind--) {
1284 		seg = &vm_phys_segs[segind];
1285 		if (seg->start >= high || seg->domain != domain)
1286 			continue;
1287 		if (low >= seg->end)
1288 			break;
1289 		if (low <= seg->start)
1290 			pa_start = seg->start;
1291 		else
1292 			pa_start = low;
1293 		if (high < seg->end)
1294 			pa_end = high;
1295 		else
1296 			pa_end = seg->end;
1297 		if (pa_end - pa_start < ptoa(npages))
1298 			continue;
1299 		m_run = vm_phys_alloc_seg_contig(seg, npages, low, high,
1300 		    alignment, boundary);
1301 		if (m_run != NULL)
1302 			break;
1303 	}
1304 	return (m_run);
1305 }
1306 
1307 /*
1308  * Allocate a run of contiguous physical pages from the free list for the
1309  * specified segment.
1310  */
1311 static vm_page_t
1312 vm_phys_alloc_seg_contig(struct vm_phys_seg *seg, u_long npages,
1313     vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary)
1314 {
1315 	struct vm_freelist *fl;
1316 	vm_paddr_t pa, pa_end, size;
1317 	vm_page_t m, m_ret;
1318 	u_long npages_end;
1319 	int oind, order, pind;
1320 
1321 	KASSERT(npages > 0, ("npages is 0"));
1322 	KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
1323 	KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
1324 	vm_domain_free_assert_locked(VM_DOMAIN(seg->domain));
1325 	/* Compute the queue that is the best fit for npages. */
1326 	order = flsl(npages - 1);
1327 	/* Search for a run satisfying the specified conditions. */
1328 	size = npages << PAGE_SHIFT;
1329 	for (oind = min(order, VM_NFREEORDER - 1); oind < VM_NFREEORDER;
1330 	    oind++) {
1331 		for (pind = 0; pind < VM_NFREEPOOL; pind++) {
1332 			fl = (*seg->free_queues)[pind];
1333 			TAILQ_FOREACH(m_ret, &fl[oind].pl, listq) {
1334 				/*
1335 				 * Is the size of this allocation request
1336 				 * larger than the largest block size?
1337 				 */
1338 				if (order >= VM_NFREEORDER) {
1339 					/*
1340 					 * Determine if a sufficient number of
1341 					 * subsequent blocks to satisfy the
1342 					 * allocation request are free.
1343 					 */
1344 					pa = VM_PAGE_TO_PHYS(m_ret);
1345 					pa_end = pa + size;
1346 					if (pa_end < pa)
1347 						continue;
1348 					for (;;) {
1349 						pa += 1 << (PAGE_SHIFT +
1350 						    VM_NFREEORDER - 1);
1351 						if (pa >= pa_end ||
1352 						    pa < seg->start ||
1353 						    pa >= seg->end)
1354 							break;
1355 						m = &seg->first_page[atop(pa -
1356 						    seg->start)];
1357 						if (m->order != VM_NFREEORDER -
1358 						    1)
1359 							break;
1360 					}
1361 					/* If not, go to the next block. */
1362 					if (pa < pa_end)
1363 						continue;
1364 				}
1365 
1366 				/*
1367 				 * Determine if the blocks are within the
1368 				 * given range, satisfy the given alignment,
1369 				 * and do not cross the given boundary.
1370 				 */
1371 				pa = VM_PAGE_TO_PHYS(m_ret);
1372 				pa_end = pa + size;
1373 				if (pa >= low && pa_end <= high &&
1374 				    (pa & (alignment - 1)) == 0 &&
1375 				    rounddown2(pa ^ (pa_end - 1), boundary) == 0)
1376 					goto done;
1377 			}
1378 		}
1379 	}
1380 	return (NULL);
1381 done:
1382 	for (m = m_ret; m < &m_ret[npages]; m = &m[1 << oind]) {
1383 		fl = (*seg->free_queues)[m->pool];
1384 		vm_freelist_rem(fl, m, oind);
1385 		if (m->pool != VM_FREEPOOL_DEFAULT)
1386 			vm_phys_set_pool(VM_FREEPOOL_DEFAULT, m, oind);
1387 	}
1388 	/* Return excess pages to the free lists. */
1389 	npages_end = roundup2(npages, 1 << oind);
1390 	if (npages < npages_end) {
1391 		fl = (*seg->free_queues)[VM_FREEPOOL_DEFAULT];
1392 		vm_phys_enq_range(&m_ret[npages], npages_end - npages, fl, 0);
1393 	}
1394 	return (m_ret);
1395 }
1396 
1397 #ifdef DDB
1398 /*
1399  * Show the number of physical pages in each of the free lists.
1400  */
1401 DB_SHOW_COMMAND(freepages, db_show_freepages)
1402 {
1403 	struct vm_freelist *fl;
1404 	int flind, oind, pind, dom;
1405 
1406 	for (dom = 0; dom < vm_ndomains; dom++) {
1407 		db_printf("DOMAIN: %d\n", dom);
1408 		for (flind = 0; flind < vm_nfreelists; flind++) {
1409 			db_printf("FREE LIST %d:\n"
1410 			    "\n  ORDER (SIZE)  |  NUMBER"
1411 			    "\n              ", flind);
1412 			for (pind = 0; pind < VM_NFREEPOOL; pind++)
1413 				db_printf("  |  POOL %d", pind);
1414 			db_printf("\n--            ");
1415 			for (pind = 0; pind < VM_NFREEPOOL; pind++)
1416 				db_printf("-- --      ");
1417 			db_printf("--\n");
1418 			for (oind = VM_NFREEORDER - 1; oind >= 0; oind--) {
1419 				db_printf("  %2.2d (%6.6dK)", oind,
1420 				    1 << (PAGE_SHIFT - 10 + oind));
1421 				for (pind = 0; pind < VM_NFREEPOOL; pind++) {
1422 				fl = vm_phys_free_queues[dom][flind][pind];
1423 					db_printf("  |  %6.6d", fl[oind].lcnt);
1424 				}
1425 				db_printf("\n");
1426 			}
1427 			db_printf("\n");
1428 		}
1429 		db_printf("\n");
1430 	}
1431 }
1432 #endif
1433