xref: /freebsd/sys/vm/vm_phys.c (revision bdafb02fcb88389fd1ab684cfe734cb429d35618)
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/lock.h>
50 #include <sys/kernel.h>
51 #include <sys/malloc.h>
52 #include <sys/mutex.h>
53 #include <sys/proc.h>
54 #include <sys/queue.h>
55 #include <sys/rwlock.h>
56 #include <sys/sbuf.h>
57 #include <sys/sysctl.h>
58 #include <sys/tree.h>
59 #include <sys/vmmeter.h>
60 #include <sys/seq.h>
61 
62 #include <ddb/ddb.h>
63 
64 #include <vm/vm.h>
65 #include <vm/vm_param.h>
66 #include <vm/vm_kern.h>
67 #include <vm/vm_object.h>
68 #include <vm/vm_page.h>
69 #include <vm/vm_phys.h>
70 #include <vm/vm_pagequeue.h>
71 
72 _Static_assert(sizeof(long) * NBBY >= VM_PHYSSEG_MAX,
73     "Too many physsegs.");
74 
75 #ifdef NUMA
76 struct mem_affinity __read_mostly *mem_affinity;
77 int __read_mostly *mem_locality;
78 #endif
79 
80 int __read_mostly vm_ndomains = 1;
81 domainset_t __read_mostly all_domains = DOMAINSET_T_INITIALIZER(0x1);
82 
83 struct vm_phys_seg __read_mostly vm_phys_segs[VM_PHYSSEG_MAX];
84 int __read_mostly vm_phys_nsegs;
85 
86 struct vm_phys_fictitious_seg;
87 static int vm_phys_fictitious_cmp(struct vm_phys_fictitious_seg *,
88     struct vm_phys_fictitious_seg *);
89 
90 RB_HEAD(fict_tree, vm_phys_fictitious_seg) vm_phys_fictitious_tree =
91     RB_INITIALIZER(_vm_phys_fictitious_tree);
92 
93 struct vm_phys_fictitious_seg {
94 	RB_ENTRY(vm_phys_fictitious_seg) node;
95 	/* Memory region data */
96 	vm_paddr_t	start;
97 	vm_paddr_t	end;
98 	vm_page_t	first_page;
99 };
100 
101 RB_GENERATE_STATIC(fict_tree, vm_phys_fictitious_seg, node,
102     vm_phys_fictitious_cmp);
103 
104 static struct rwlock_padalign vm_phys_fictitious_reg_lock;
105 MALLOC_DEFINE(M_FICT_PAGES, "vm_fictitious", "Fictitious VM pages");
106 
107 static struct vm_freelist __aligned(CACHE_LINE_SIZE)
108     vm_phys_free_queues[MAXMEMDOM][VM_NFREELIST][VM_NFREEPOOL][VM_NFREEORDER];
109 
110 static int __read_mostly vm_nfreelists;
111 
112 /*
113  * Provides the mapping from VM_FREELIST_* to free list indices (flind).
114  */
115 static int __read_mostly vm_freelist_to_flind[VM_NFREELIST];
116 
117 CTASSERT(VM_FREELIST_DEFAULT == 0);
118 
119 #ifdef VM_FREELIST_DMA32
120 #define	VM_DMA32_BOUNDARY	((vm_paddr_t)1 << 32)
121 #endif
122 
123 /*
124  * Enforce the assumptions made by vm_phys_add_seg() and vm_phys_init() about
125  * the ordering of the free list boundaries.
126  */
127 #if defined(VM_LOWMEM_BOUNDARY) && defined(VM_DMA32_BOUNDARY)
128 CTASSERT(VM_LOWMEM_BOUNDARY < VM_DMA32_BOUNDARY);
129 #endif
130 
131 static int sysctl_vm_phys_free(SYSCTL_HANDLER_ARGS);
132 SYSCTL_OID(_vm, OID_AUTO, phys_free, CTLTYPE_STRING | CTLFLAG_RD,
133     NULL, 0, sysctl_vm_phys_free, "A", "Phys Free Info");
134 
135 static int sysctl_vm_phys_segs(SYSCTL_HANDLER_ARGS);
136 SYSCTL_OID(_vm, OID_AUTO, phys_segs, CTLTYPE_STRING | CTLFLAG_RD,
137     NULL, 0, sysctl_vm_phys_segs, "A", "Phys Seg Info");
138 
139 #ifdef NUMA
140 static int sysctl_vm_phys_locality(SYSCTL_HANDLER_ARGS);
141 SYSCTL_OID(_vm, OID_AUTO, phys_locality, CTLTYPE_STRING | CTLFLAG_RD,
142     NULL, 0, sysctl_vm_phys_locality, "A", "Phys Locality Info");
143 #endif
144 
145 SYSCTL_INT(_vm, OID_AUTO, ndomains, CTLFLAG_RD,
146     &vm_ndomains, 0, "Number of physical memory domains available.");
147 
148 static vm_page_t vm_phys_alloc_seg_contig(struct vm_phys_seg *seg,
149     u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment,
150     vm_paddr_t boundary);
151 static void _vm_phys_create_seg(vm_paddr_t start, vm_paddr_t end, int domain);
152 static void vm_phys_create_seg(vm_paddr_t start, vm_paddr_t end);
153 static void vm_phys_split_pages(vm_page_t m, int oind, struct vm_freelist *fl,
154     int order, int tail);
155 
156 /*
157  * Red-black tree helpers for vm fictitious range management.
158  */
159 static inline int
160 vm_phys_fictitious_in_range(struct vm_phys_fictitious_seg *p,
161     struct vm_phys_fictitious_seg *range)
162 {
163 
164 	KASSERT(range->start != 0 && range->end != 0,
165 	    ("Invalid range passed on search for vm_fictitious page"));
166 	if (p->start >= range->end)
167 		return (1);
168 	if (p->start < range->start)
169 		return (-1);
170 
171 	return (0);
172 }
173 
174 static int
175 vm_phys_fictitious_cmp(struct vm_phys_fictitious_seg *p1,
176     struct vm_phys_fictitious_seg *p2)
177 {
178 
179 	/* Check if this is a search for a page */
180 	if (p1->end == 0)
181 		return (vm_phys_fictitious_in_range(p1, p2));
182 
183 	KASSERT(p2->end != 0,
184     ("Invalid range passed as second parameter to vm fictitious comparison"));
185 
186 	/* Searching to add a new range */
187 	if (p1->end <= p2->start)
188 		return (-1);
189 	if (p1->start >= p2->end)
190 		return (1);
191 
192 	panic("Trying to add overlapping vm fictitious ranges:\n"
193 	    "[%#jx:%#jx] and [%#jx:%#jx]", (uintmax_t)p1->start,
194 	    (uintmax_t)p1->end, (uintmax_t)p2->start, (uintmax_t)p2->end);
195 }
196 
197 int
198 vm_phys_domain_match(int prefer, vm_paddr_t low, vm_paddr_t high)
199 {
200 #ifdef NUMA
201 	domainset_t mask;
202 	int i;
203 
204 	if (vm_ndomains == 1 || mem_affinity == NULL)
205 		return (0);
206 
207 	DOMAINSET_ZERO(&mask);
208 	/*
209 	 * Check for any memory that overlaps low, high.
210 	 */
211 	for (i = 0; mem_affinity[i].end != 0; i++)
212 		if (mem_affinity[i].start <= high &&
213 		    mem_affinity[i].end >= low)
214 			DOMAINSET_SET(mem_affinity[i].domain, &mask);
215 	if (prefer != -1 && DOMAINSET_ISSET(prefer, &mask))
216 		return (prefer);
217 	if (DOMAINSET_EMPTY(&mask))
218 		panic("vm_phys_domain_match:  Impossible constraint");
219 	return (DOMAINSET_FFS(&mask) - 1);
220 #else
221 	return (0);
222 #endif
223 }
224 
225 /*
226  * Outputs the state of the physical memory allocator, specifically,
227  * the amount of physical memory in each free list.
228  */
229 static int
230 sysctl_vm_phys_free(SYSCTL_HANDLER_ARGS)
231 {
232 	struct sbuf sbuf;
233 	struct vm_freelist *fl;
234 	int dom, error, flind, oind, pind;
235 
236 	error = sysctl_wire_old_buffer(req, 0);
237 	if (error != 0)
238 		return (error);
239 	sbuf_new_for_sysctl(&sbuf, NULL, 128 * vm_ndomains, req);
240 	for (dom = 0; dom < vm_ndomains; dom++) {
241 		sbuf_printf(&sbuf,"\nDOMAIN %d:\n", dom);
242 		for (flind = 0; flind < vm_nfreelists; flind++) {
243 			sbuf_printf(&sbuf, "\nFREE LIST %d:\n"
244 			    "\n  ORDER (SIZE)  |  NUMBER"
245 			    "\n              ", flind);
246 			for (pind = 0; pind < VM_NFREEPOOL; pind++)
247 				sbuf_printf(&sbuf, "  |  POOL %d", pind);
248 			sbuf_printf(&sbuf, "\n--            ");
249 			for (pind = 0; pind < VM_NFREEPOOL; pind++)
250 				sbuf_printf(&sbuf, "-- --      ");
251 			sbuf_printf(&sbuf, "--\n");
252 			for (oind = VM_NFREEORDER - 1; oind >= 0; oind--) {
253 				sbuf_printf(&sbuf, "  %2d (%6dK)", oind,
254 				    1 << (PAGE_SHIFT - 10 + oind));
255 				for (pind = 0; pind < VM_NFREEPOOL; pind++) {
256 				fl = vm_phys_free_queues[dom][flind][pind];
257 					sbuf_printf(&sbuf, "  |  %6d",
258 					    fl[oind].lcnt);
259 				}
260 				sbuf_printf(&sbuf, "\n");
261 			}
262 		}
263 	}
264 	error = sbuf_finish(&sbuf);
265 	sbuf_delete(&sbuf);
266 	return (error);
267 }
268 
269 /*
270  * Outputs the set of physical memory segments.
271  */
272 static int
273 sysctl_vm_phys_segs(SYSCTL_HANDLER_ARGS)
274 {
275 	struct sbuf sbuf;
276 	struct vm_phys_seg *seg;
277 	int error, segind;
278 
279 	error = sysctl_wire_old_buffer(req, 0);
280 	if (error != 0)
281 		return (error);
282 	sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
283 	for (segind = 0; segind < vm_phys_nsegs; segind++) {
284 		sbuf_printf(&sbuf, "\nSEGMENT %d:\n\n", segind);
285 		seg = &vm_phys_segs[segind];
286 		sbuf_printf(&sbuf, "start:     %#jx\n",
287 		    (uintmax_t)seg->start);
288 		sbuf_printf(&sbuf, "end:       %#jx\n",
289 		    (uintmax_t)seg->end);
290 		sbuf_printf(&sbuf, "domain:    %d\n", seg->domain);
291 		sbuf_printf(&sbuf, "free list: %p\n", seg->free_queues);
292 	}
293 	error = sbuf_finish(&sbuf);
294 	sbuf_delete(&sbuf);
295 	return (error);
296 }
297 
298 /*
299  * Return affinity, or -1 if there's no affinity information.
300  */
301 int
302 vm_phys_mem_affinity(int f, int t)
303 {
304 
305 #ifdef NUMA
306 	if (mem_locality == NULL)
307 		return (-1);
308 	if (f >= vm_ndomains || t >= vm_ndomains)
309 		return (-1);
310 	return (mem_locality[f * vm_ndomains + t]);
311 #else
312 	return (-1);
313 #endif
314 }
315 
316 #ifdef NUMA
317 /*
318  * Outputs the VM locality table.
319  */
320 static int
321 sysctl_vm_phys_locality(SYSCTL_HANDLER_ARGS)
322 {
323 	struct sbuf sbuf;
324 	int error, i, j;
325 
326 	error = sysctl_wire_old_buffer(req, 0);
327 	if (error != 0)
328 		return (error);
329 	sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
330 
331 	sbuf_printf(&sbuf, "\n");
332 
333 	for (i = 0; i < vm_ndomains; i++) {
334 		sbuf_printf(&sbuf, "%d: ", i);
335 		for (j = 0; j < vm_ndomains; j++) {
336 			sbuf_printf(&sbuf, "%d ", vm_phys_mem_affinity(i, j));
337 		}
338 		sbuf_printf(&sbuf, "\n");
339 	}
340 	error = sbuf_finish(&sbuf);
341 	sbuf_delete(&sbuf);
342 	return (error);
343 }
344 #endif
345 
346 static void
347 vm_freelist_add(struct vm_freelist *fl, vm_page_t m, int order, int tail)
348 {
349 
350 	m->order = order;
351 	if (tail)
352 		TAILQ_INSERT_TAIL(&fl[order].pl, m, listq);
353 	else
354 		TAILQ_INSERT_HEAD(&fl[order].pl, m, listq);
355 	fl[order].lcnt++;
356 }
357 
358 static void
359 vm_freelist_rem(struct vm_freelist *fl, vm_page_t m, int order)
360 {
361 
362 	TAILQ_REMOVE(&fl[order].pl, m, listq);
363 	fl[order].lcnt--;
364 	m->order = VM_NFREEORDER;
365 }
366 
367 /*
368  * Create a physical memory segment.
369  */
370 static void
371 _vm_phys_create_seg(vm_paddr_t start, vm_paddr_t end, int domain)
372 {
373 	struct vm_phys_seg *seg;
374 
375 	KASSERT(vm_phys_nsegs < VM_PHYSSEG_MAX,
376 	    ("vm_phys_create_seg: increase VM_PHYSSEG_MAX"));
377 	KASSERT(domain >= 0 && domain < vm_ndomains,
378 	    ("vm_phys_create_seg: invalid domain provided"));
379 	seg = &vm_phys_segs[vm_phys_nsegs++];
380 	while (seg > vm_phys_segs && (seg - 1)->start >= end) {
381 		*seg = *(seg - 1);
382 		seg--;
383 	}
384 	seg->start = start;
385 	seg->end = end;
386 	seg->domain = domain;
387 }
388 
389 static void
390 vm_phys_create_seg(vm_paddr_t start, vm_paddr_t end)
391 {
392 #ifdef NUMA
393 	int i;
394 
395 	if (mem_affinity == NULL) {
396 		_vm_phys_create_seg(start, end, 0);
397 		return;
398 	}
399 
400 	for (i = 0;; i++) {
401 		if (mem_affinity[i].end == 0)
402 			panic("Reached end of affinity info");
403 		if (mem_affinity[i].end <= start)
404 			continue;
405 		if (mem_affinity[i].start > start)
406 			panic("No affinity info for start %jx",
407 			    (uintmax_t)start);
408 		if (mem_affinity[i].end >= end) {
409 			_vm_phys_create_seg(start, end,
410 			    mem_affinity[i].domain);
411 			break;
412 		}
413 		_vm_phys_create_seg(start, mem_affinity[i].end,
414 		    mem_affinity[i].domain);
415 		start = mem_affinity[i].end;
416 	}
417 #else
418 	_vm_phys_create_seg(start, end, 0);
419 #endif
420 }
421 
422 /*
423  * Add a physical memory segment.
424  */
425 void
426 vm_phys_add_seg(vm_paddr_t start, vm_paddr_t end)
427 {
428 	vm_paddr_t paddr;
429 
430 	KASSERT((start & PAGE_MASK) == 0,
431 	    ("vm_phys_define_seg: start is not page aligned"));
432 	KASSERT((end & PAGE_MASK) == 0,
433 	    ("vm_phys_define_seg: end is not page aligned"));
434 
435 	/*
436 	 * Split the physical memory segment if it spans two or more free
437 	 * list boundaries.
438 	 */
439 	paddr = start;
440 #ifdef	VM_FREELIST_LOWMEM
441 	if (paddr < VM_LOWMEM_BOUNDARY && end > VM_LOWMEM_BOUNDARY) {
442 		vm_phys_create_seg(paddr, VM_LOWMEM_BOUNDARY);
443 		paddr = VM_LOWMEM_BOUNDARY;
444 	}
445 #endif
446 #ifdef	VM_FREELIST_DMA32
447 	if (paddr < VM_DMA32_BOUNDARY && end > VM_DMA32_BOUNDARY) {
448 		vm_phys_create_seg(paddr, VM_DMA32_BOUNDARY);
449 		paddr = VM_DMA32_BOUNDARY;
450 	}
451 #endif
452 	vm_phys_create_seg(paddr, end);
453 }
454 
455 /*
456  * Initialize the physical memory allocator.
457  *
458  * Requires that vm_page_array is initialized!
459  */
460 void
461 vm_phys_init(void)
462 {
463 	struct vm_freelist *fl;
464 	struct vm_phys_seg *end_seg, *prev_seg, *seg, *tmp_seg;
465 	u_long npages;
466 	int dom, flind, freelist, oind, pind, segind;
467 
468 	/*
469 	 * Compute the number of free lists, and generate the mapping from the
470 	 * manifest constants VM_FREELIST_* to the free list indices.
471 	 *
472 	 * Initially, the entries of vm_freelist_to_flind[] are set to either
473 	 * 0 or 1 to indicate which free lists should be created.
474 	 */
475 	npages = 0;
476 	for (segind = vm_phys_nsegs - 1; segind >= 0; segind--) {
477 		seg = &vm_phys_segs[segind];
478 #ifdef	VM_FREELIST_LOWMEM
479 		if (seg->end <= VM_LOWMEM_BOUNDARY)
480 			vm_freelist_to_flind[VM_FREELIST_LOWMEM] = 1;
481 		else
482 #endif
483 #ifdef	VM_FREELIST_DMA32
484 		if (
485 #ifdef	VM_DMA32_NPAGES_THRESHOLD
486 		    /*
487 		     * Create the DMA32 free list only if the amount of
488 		     * physical memory above physical address 4G exceeds the
489 		     * given threshold.
490 		     */
491 		    npages > VM_DMA32_NPAGES_THRESHOLD &&
492 #endif
493 		    seg->end <= VM_DMA32_BOUNDARY)
494 			vm_freelist_to_flind[VM_FREELIST_DMA32] = 1;
495 		else
496 #endif
497 		{
498 			npages += atop(seg->end - seg->start);
499 			vm_freelist_to_flind[VM_FREELIST_DEFAULT] = 1;
500 		}
501 	}
502 	/* Change each entry into a running total of the free lists. */
503 	for (freelist = 1; freelist < VM_NFREELIST; freelist++) {
504 		vm_freelist_to_flind[freelist] +=
505 		    vm_freelist_to_flind[freelist - 1];
506 	}
507 	vm_nfreelists = vm_freelist_to_flind[VM_NFREELIST - 1];
508 	KASSERT(vm_nfreelists > 0, ("vm_phys_init: no free lists"));
509 	/* Change each entry into a free list index. */
510 	for (freelist = 0; freelist < VM_NFREELIST; freelist++)
511 		vm_freelist_to_flind[freelist]--;
512 
513 	/*
514 	 * Initialize the first_page and free_queues fields of each physical
515 	 * memory segment.
516 	 */
517 #ifdef VM_PHYSSEG_SPARSE
518 	npages = 0;
519 #endif
520 	for (segind = 0; segind < vm_phys_nsegs; segind++) {
521 		seg = &vm_phys_segs[segind];
522 #ifdef VM_PHYSSEG_SPARSE
523 		seg->first_page = &vm_page_array[npages];
524 		npages += atop(seg->end - seg->start);
525 #else
526 		seg->first_page = PHYS_TO_VM_PAGE(seg->start);
527 #endif
528 #ifdef	VM_FREELIST_LOWMEM
529 		if (seg->end <= VM_LOWMEM_BOUNDARY) {
530 			flind = vm_freelist_to_flind[VM_FREELIST_LOWMEM];
531 			KASSERT(flind >= 0,
532 			    ("vm_phys_init: LOWMEM flind < 0"));
533 		} else
534 #endif
535 #ifdef	VM_FREELIST_DMA32
536 		if (seg->end <= VM_DMA32_BOUNDARY) {
537 			flind = vm_freelist_to_flind[VM_FREELIST_DMA32];
538 			KASSERT(flind >= 0,
539 			    ("vm_phys_init: DMA32 flind < 0"));
540 		} else
541 #endif
542 		{
543 			flind = vm_freelist_to_flind[VM_FREELIST_DEFAULT];
544 			KASSERT(flind >= 0,
545 			    ("vm_phys_init: DEFAULT flind < 0"));
546 		}
547 		seg->free_queues = &vm_phys_free_queues[seg->domain][flind];
548 	}
549 
550 	/*
551 	 * Coalesce physical memory segments that are contiguous and share the
552 	 * same per-domain free queues.
553 	 */
554 	prev_seg = vm_phys_segs;
555 	seg = &vm_phys_segs[1];
556 	end_seg = &vm_phys_segs[vm_phys_nsegs];
557 	while (seg < end_seg) {
558 		if (prev_seg->end == seg->start &&
559 		    prev_seg->free_queues == seg->free_queues) {
560 			prev_seg->end = seg->end;
561 			KASSERT(prev_seg->domain == seg->domain,
562 			    ("vm_phys_init: free queues cannot span domains"));
563 			vm_phys_nsegs--;
564 			end_seg--;
565 			for (tmp_seg = seg; tmp_seg < end_seg; tmp_seg++)
566 				*tmp_seg = *(tmp_seg + 1);
567 		} else {
568 			prev_seg = seg;
569 			seg++;
570 		}
571 	}
572 
573 	/*
574 	 * Initialize the free queues.
575 	 */
576 	for (dom = 0; dom < vm_ndomains; dom++) {
577 		for (flind = 0; flind < vm_nfreelists; flind++) {
578 			for (pind = 0; pind < VM_NFREEPOOL; pind++) {
579 				fl = vm_phys_free_queues[dom][flind][pind];
580 				for (oind = 0; oind < VM_NFREEORDER; oind++)
581 					TAILQ_INIT(&fl[oind].pl);
582 			}
583 		}
584 	}
585 
586 	rw_init(&vm_phys_fictitious_reg_lock, "vmfctr");
587 }
588 
589 /*
590  * Split a contiguous, power of two-sized set of physical pages.
591  *
592  * When this function is called by a page allocation function, the caller
593  * should request insertion at the head unless the order [order, oind) queues
594  * are known to be empty.  The objective being to reduce the likelihood of
595  * long-term fragmentation by promoting contemporaneous allocation and
596  * (hopefully) deallocation.
597  */
598 static __inline void
599 vm_phys_split_pages(vm_page_t m, int oind, struct vm_freelist *fl, int order,
600     int tail)
601 {
602 	vm_page_t m_buddy;
603 
604 	while (oind > order) {
605 		oind--;
606 		m_buddy = &m[1 << oind];
607 		KASSERT(m_buddy->order == VM_NFREEORDER,
608 		    ("vm_phys_split_pages: page %p has unexpected order %d",
609 		    m_buddy, m_buddy->order));
610 		vm_freelist_add(fl, m_buddy, oind, tail);
611         }
612 }
613 
614 /*
615  * Add the physical pages [m, m + npages) at the end of a power-of-two aligned
616  * and sized set to the specified free list.
617  *
618  * When this function is called by a page allocation function, the caller
619  * should request insertion at the head unless the lower-order queues are
620  * known to be empty.  The objective being to reduce the likelihood of long-
621  * term fragmentation by promoting contemporaneous allocation and (hopefully)
622  * deallocation.
623  *
624  * The physical page m's buddy must not be free.
625  */
626 static void
627 vm_phys_enq_range(vm_page_t m, u_int npages, struct vm_freelist *fl, int tail)
628 {
629 	u_int n;
630 	int order;
631 
632 	KASSERT(npages > 0, ("vm_phys_enq_range: npages is 0"));
633 	KASSERT(((VM_PAGE_TO_PHYS(m) + npages * PAGE_SIZE) &
634 	    ((PAGE_SIZE << (fls(npages) - 1)) - 1)) == 0,
635 	    ("vm_phys_enq_range: page %p and npages %u are misaligned",
636 	    m, npages));
637 	do {
638 		KASSERT(m->order == VM_NFREEORDER,
639 		    ("vm_phys_enq_range: page %p has unexpected order %d",
640 		    m, m->order));
641 		order = ffs(npages) - 1;
642 		KASSERT(order < VM_NFREEORDER,
643 		    ("vm_phys_enq_range: order %d is out of range", order));
644 		vm_freelist_add(fl, m, order, tail);
645 		n = 1 << order;
646 		m += n;
647 		npages -= n;
648 	} while (npages > 0);
649 }
650 
651 /*
652  * Tries to allocate the specified number of pages from the specified pool
653  * within the specified domain.  Returns the actual number of allocated pages
654  * and a pointer to each page through the array ma[].
655  *
656  * The returned pages may not be physically contiguous.  However, in contrast
657  * to performing multiple, back-to-back calls to vm_phys_alloc_pages(..., 0),
658  * calling this function once to allocate the desired number of pages will
659  * avoid wasted time in vm_phys_split_pages().
660  *
661  * The free page queues for the specified domain must be locked.
662  */
663 int
664 vm_phys_alloc_npages(int domain, int pool, int npages, vm_page_t ma[])
665 {
666 	struct vm_freelist *alt, *fl;
667 	vm_page_t m;
668 	int avail, end, flind, freelist, i, need, oind, pind;
669 
670 	KASSERT(domain >= 0 && domain < vm_ndomains,
671 	    ("vm_phys_alloc_npages: domain %d is out of range", domain));
672 	KASSERT(pool < VM_NFREEPOOL,
673 	    ("vm_phys_alloc_npages: pool %d is out of range", pool));
674 	KASSERT(npages <= 1 << (VM_NFREEORDER - 1),
675 	    ("vm_phys_alloc_npages: npages %d is out of range", npages));
676 	vm_domain_free_assert_locked(VM_DOMAIN(domain));
677 	i = 0;
678 	for (freelist = 0; freelist < VM_NFREELIST; freelist++) {
679 		flind = vm_freelist_to_flind[freelist];
680 		if (flind < 0)
681 			continue;
682 		fl = vm_phys_free_queues[domain][flind][pool];
683 		for (oind = 0; oind < VM_NFREEORDER; oind++) {
684 			while ((m = TAILQ_FIRST(&fl[oind].pl)) != NULL) {
685 				vm_freelist_rem(fl, m, oind);
686 				avail = 1 << oind;
687 				need = imin(npages - i, avail);
688 				for (end = i + need; i < end;)
689 					ma[i++] = m++;
690 				if (need < avail) {
691 					/*
692 					 * Return excess pages to fl.  Its
693 					 * order [0, oind) queues are empty.
694 					 */
695 					vm_phys_enq_range(m, avail - need, fl,
696 					    1);
697 					return (npages);
698 				} else if (i == npages)
699 					return (npages);
700 			}
701 		}
702 		for (oind = VM_NFREEORDER - 1; oind >= 0; oind--) {
703 			for (pind = 0; pind < VM_NFREEPOOL; pind++) {
704 				alt = vm_phys_free_queues[domain][flind][pind];
705 				while ((m = TAILQ_FIRST(&alt[oind].pl)) !=
706 				    NULL) {
707 					vm_freelist_rem(alt, m, oind);
708 					vm_phys_set_pool(pool, m, oind);
709 					avail = 1 << oind;
710 					need = imin(npages - i, avail);
711 					for (end = i + need; i < end;)
712 						ma[i++] = m++;
713 					if (need < avail) {
714 						/*
715 						 * Return excess pages to fl.
716 						 * Its order [0, oind) queues
717 						 * are empty.
718 						 */
719 						vm_phys_enq_range(m, avail -
720 						    need, fl, 1);
721 						return (npages);
722 					} else if (i == npages)
723 						return (npages);
724 				}
725 			}
726 		}
727 	}
728 	return (i);
729 }
730 
731 /*
732  * Allocate a contiguous, power of two-sized set of physical pages
733  * from the free lists.
734  *
735  * The free page queues must be locked.
736  */
737 vm_page_t
738 vm_phys_alloc_pages(int domain, int pool, int order)
739 {
740 	vm_page_t m;
741 	int freelist;
742 
743 	for (freelist = 0; freelist < VM_NFREELIST; freelist++) {
744 		m = vm_phys_alloc_freelist_pages(domain, freelist, pool, order);
745 		if (m != NULL)
746 			return (m);
747 	}
748 	return (NULL);
749 }
750 
751 /*
752  * Allocate a contiguous, power of two-sized set of physical pages from the
753  * specified free list.  The free list must be specified using one of the
754  * manifest constants VM_FREELIST_*.
755  *
756  * The free page queues must be locked.
757  */
758 vm_page_t
759 vm_phys_alloc_freelist_pages(int domain, int freelist, int pool, int order)
760 {
761 	struct vm_freelist *alt, *fl;
762 	vm_page_t m;
763 	int oind, pind, flind;
764 
765 	KASSERT(domain >= 0 && domain < vm_ndomains,
766 	    ("vm_phys_alloc_freelist_pages: domain %d is out of range",
767 	    domain));
768 	KASSERT(freelist < VM_NFREELIST,
769 	    ("vm_phys_alloc_freelist_pages: freelist %d is out of range",
770 	    freelist));
771 	KASSERT(pool < VM_NFREEPOOL,
772 	    ("vm_phys_alloc_freelist_pages: pool %d is out of range", pool));
773 	KASSERT(order < VM_NFREEORDER,
774 	    ("vm_phys_alloc_freelist_pages: order %d is out of range", order));
775 
776 	flind = vm_freelist_to_flind[freelist];
777 	/* Check if freelist is present */
778 	if (flind < 0)
779 		return (NULL);
780 
781 	vm_domain_free_assert_locked(VM_DOMAIN(domain));
782 	fl = &vm_phys_free_queues[domain][flind][pool][0];
783 	for (oind = order; oind < VM_NFREEORDER; oind++) {
784 		m = TAILQ_FIRST(&fl[oind].pl);
785 		if (m != NULL) {
786 			vm_freelist_rem(fl, m, oind);
787 			/* The order [order, oind) queues are empty. */
788 			vm_phys_split_pages(m, oind, fl, order, 1);
789 			return (m);
790 		}
791 	}
792 
793 	/*
794 	 * The given pool was empty.  Find the largest
795 	 * contiguous, power-of-two-sized set of pages in any
796 	 * pool.  Transfer these pages to the given pool, and
797 	 * use them to satisfy the allocation.
798 	 */
799 	for (oind = VM_NFREEORDER - 1; oind >= order; oind--) {
800 		for (pind = 0; pind < VM_NFREEPOOL; pind++) {
801 			alt = &vm_phys_free_queues[domain][flind][pind][0];
802 			m = TAILQ_FIRST(&alt[oind].pl);
803 			if (m != NULL) {
804 				vm_freelist_rem(alt, m, oind);
805 				vm_phys_set_pool(pool, m, oind);
806 				/* The order [order, oind) queues are empty. */
807 				vm_phys_split_pages(m, oind, fl, order, 1);
808 				return (m);
809 			}
810 		}
811 	}
812 	return (NULL);
813 }
814 
815 /*
816  * Find the vm_page corresponding to the given physical address.
817  */
818 vm_page_t
819 vm_phys_paddr_to_vm_page(vm_paddr_t pa)
820 {
821 	struct vm_phys_seg *seg;
822 	int segind;
823 
824 	for (segind = 0; segind < vm_phys_nsegs; segind++) {
825 		seg = &vm_phys_segs[segind];
826 		if (pa >= seg->start && pa < seg->end)
827 			return (&seg->first_page[atop(pa - seg->start)]);
828 	}
829 	return (NULL);
830 }
831 
832 vm_page_t
833 vm_phys_fictitious_to_vm_page(vm_paddr_t pa)
834 {
835 	struct vm_phys_fictitious_seg tmp, *seg;
836 	vm_page_t m;
837 
838 	m = NULL;
839 	tmp.start = pa;
840 	tmp.end = 0;
841 
842 	rw_rlock(&vm_phys_fictitious_reg_lock);
843 	seg = RB_FIND(fict_tree, &vm_phys_fictitious_tree, &tmp);
844 	rw_runlock(&vm_phys_fictitious_reg_lock);
845 	if (seg == NULL)
846 		return (NULL);
847 
848 	m = &seg->first_page[atop(pa - seg->start)];
849 	KASSERT((m->flags & PG_FICTITIOUS) != 0, ("%p not fictitious", m));
850 
851 	return (m);
852 }
853 
854 static inline void
855 vm_phys_fictitious_init_range(vm_page_t range, vm_paddr_t start,
856     long page_count, vm_memattr_t memattr)
857 {
858 	long i;
859 
860 	bzero(range, page_count * sizeof(*range));
861 	for (i = 0; i < page_count; i++) {
862 		vm_page_initfake(&range[i], start + PAGE_SIZE * i, memattr);
863 		range[i].oflags &= ~VPO_UNMANAGED;
864 		range[i].busy_lock = VPB_UNBUSIED;
865 	}
866 }
867 
868 int
869 vm_phys_fictitious_reg_range(vm_paddr_t start, vm_paddr_t end,
870     vm_memattr_t memattr)
871 {
872 	struct vm_phys_fictitious_seg *seg;
873 	vm_page_t fp;
874 	long page_count;
875 #ifdef VM_PHYSSEG_DENSE
876 	long pi, pe;
877 	long dpage_count;
878 #endif
879 
880 	KASSERT(start < end,
881 	    ("Start of segment isn't less than end (start: %jx end: %jx)",
882 	    (uintmax_t)start, (uintmax_t)end));
883 
884 	page_count = (end - start) / PAGE_SIZE;
885 
886 #ifdef VM_PHYSSEG_DENSE
887 	pi = atop(start);
888 	pe = atop(end);
889 	if (pi >= first_page && (pi - first_page) < vm_page_array_size) {
890 		fp = &vm_page_array[pi - first_page];
891 		if ((pe - first_page) > vm_page_array_size) {
892 			/*
893 			 * We have a segment that starts inside
894 			 * of vm_page_array, but ends outside of it.
895 			 *
896 			 * Use vm_page_array pages for those that are
897 			 * inside of the vm_page_array range, and
898 			 * allocate the remaining ones.
899 			 */
900 			dpage_count = vm_page_array_size - (pi - first_page);
901 			vm_phys_fictitious_init_range(fp, start, dpage_count,
902 			    memattr);
903 			page_count -= dpage_count;
904 			start += ptoa(dpage_count);
905 			goto alloc;
906 		}
907 		/*
908 		 * We can allocate the full range from vm_page_array,
909 		 * so there's no need to register the range in the tree.
910 		 */
911 		vm_phys_fictitious_init_range(fp, start, page_count, memattr);
912 		return (0);
913 	} else if (pe > first_page && (pe - first_page) < vm_page_array_size) {
914 		/*
915 		 * We have a segment that ends inside of vm_page_array,
916 		 * but starts outside of it.
917 		 */
918 		fp = &vm_page_array[0];
919 		dpage_count = pe - first_page;
920 		vm_phys_fictitious_init_range(fp, ptoa(first_page), dpage_count,
921 		    memattr);
922 		end -= ptoa(dpage_count);
923 		page_count -= dpage_count;
924 		goto alloc;
925 	} else if (pi < first_page && pe > (first_page + vm_page_array_size)) {
926 		/*
927 		 * Trying to register a fictitious range that expands before
928 		 * and after vm_page_array.
929 		 */
930 		return (EINVAL);
931 	} else {
932 alloc:
933 #endif
934 		fp = malloc(page_count * sizeof(struct vm_page), M_FICT_PAGES,
935 		    M_WAITOK);
936 #ifdef VM_PHYSSEG_DENSE
937 	}
938 #endif
939 	vm_phys_fictitious_init_range(fp, start, page_count, memattr);
940 
941 	seg = malloc(sizeof(*seg), M_FICT_PAGES, M_WAITOK | M_ZERO);
942 	seg->start = start;
943 	seg->end = end;
944 	seg->first_page = fp;
945 
946 	rw_wlock(&vm_phys_fictitious_reg_lock);
947 	RB_INSERT(fict_tree, &vm_phys_fictitious_tree, seg);
948 	rw_wunlock(&vm_phys_fictitious_reg_lock);
949 
950 	return (0);
951 }
952 
953 void
954 vm_phys_fictitious_unreg_range(vm_paddr_t start, vm_paddr_t end)
955 {
956 	struct vm_phys_fictitious_seg *seg, tmp;
957 #ifdef VM_PHYSSEG_DENSE
958 	long pi, pe;
959 #endif
960 
961 	KASSERT(start < end,
962 	    ("Start of segment isn't less than end (start: %jx end: %jx)",
963 	    (uintmax_t)start, (uintmax_t)end));
964 
965 #ifdef VM_PHYSSEG_DENSE
966 	pi = atop(start);
967 	pe = atop(end);
968 	if (pi >= first_page && (pi - first_page) < vm_page_array_size) {
969 		if ((pe - first_page) <= vm_page_array_size) {
970 			/*
971 			 * This segment was allocated using vm_page_array
972 			 * only, there's nothing to do since those pages
973 			 * were never added to the tree.
974 			 */
975 			return;
976 		}
977 		/*
978 		 * We have a segment that starts inside
979 		 * of vm_page_array, but ends outside of it.
980 		 *
981 		 * Calculate how many pages were added to the
982 		 * tree and free them.
983 		 */
984 		start = ptoa(first_page + vm_page_array_size);
985 	} else if (pe > first_page && (pe - first_page) < vm_page_array_size) {
986 		/*
987 		 * We have a segment that ends inside of vm_page_array,
988 		 * but starts outside of it.
989 		 */
990 		end = ptoa(first_page);
991 	} else if (pi < first_page && pe > (first_page + vm_page_array_size)) {
992 		/* Since it's not possible to register such a range, panic. */
993 		panic(
994 		    "Unregistering not registered fictitious range [%#jx:%#jx]",
995 		    (uintmax_t)start, (uintmax_t)end);
996 	}
997 #endif
998 	tmp.start = start;
999 	tmp.end = 0;
1000 
1001 	rw_wlock(&vm_phys_fictitious_reg_lock);
1002 	seg = RB_FIND(fict_tree, &vm_phys_fictitious_tree, &tmp);
1003 	if (seg->start != start || seg->end != end) {
1004 		rw_wunlock(&vm_phys_fictitious_reg_lock);
1005 		panic(
1006 		    "Unregistering not registered fictitious range [%#jx:%#jx]",
1007 		    (uintmax_t)start, (uintmax_t)end);
1008 	}
1009 	RB_REMOVE(fict_tree, &vm_phys_fictitious_tree, seg);
1010 	rw_wunlock(&vm_phys_fictitious_reg_lock);
1011 	free(seg->first_page, M_FICT_PAGES);
1012 	free(seg, M_FICT_PAGES);
1013 }
1014 
1015 /*
1016  * Free a contiguous, power of two-sized set of physical pages.
1017  *
1018  * The free page queues must be locked.
1019  */
1020 void
1021 vm_phys_free_pages(vm_page_t m, int order)
1022 {
1023 	struct vm_freelist *fl;
1024 	struct vm_phys_seg *seg;
1025 	vm_paddr_t pa;
1026 	vm_page_t m_buddy;
1027 
1028 	KASSERT(m->order == VM_NFREEORDER,
1029 	    ("vm_phys_free_pages: page %p has unexpected order %d",
1030 	    m, m->order));
1031 	KASSERT(m->pool < VM_NFREEPOOL,
1032 	    ("vm_phys_free_pages: page %p has unexpected pool %d",
1033 	    m, m->pool));
1034 	KASSERT(order < VM_NFREEORDER,
1035 	    ("vm_phys_free_pages: order %d is out of range", order));
1036 	seg = &vm_phys_segs[m->segind];
1037 	vm_domain_free_assert_locked(VM_DOMAIN(seg->domain));
1038 	if (order < VM_NFREEORDER - 1) {
1039 		pa = VM_PAGE_TO_PHYS(m);
1040 		do {
1041 			pa ^= ((vm_paddr_t)1 << (PAGE_SHIFT + order));
1042 			if (pa < seg->start || pa >= seg->end)
1043 				break;
1044 			m_buddy = &seg->first_page[atop(pa - seg->start)];
1045 			if (m_buddy->order != order)
1046 				break;
1047 			fl = (*seg->free_queues)[m_buddy->pool];
1048 			vm_freelist_rem(fl, m_buddy, order);
1049 			if (m_buddy->pool != m->pool)
1050 				vm_phys_set_pool(m->pool, m_buddy, order);
1051 			order++;
1052 			pa &= ~(((vm_paddr_t)1 << (PAGE_SHIFT + order)) - 1);
1053 			m = &seg->first_page[atop(pa - seg->start)];
1054 		} while (order < VM_NFREEORDER - 1);
1055 	}
1056 	fl = (*seg->free_queues)[m->pool];
1057 	vm_freelist_add(fl, m, order, 1);
1058 }
1059 
1060 /*
1061  * Free a contiguous, arbitrarily sized set of physical pages.
1062  *
1063  * The free page queues must be locked.
1064  */
1065 void
1066 vm_phys_free_contig(vm_page_t m, u_long npages)
1067 {
1068 	u_int n;
1069 	int order;
1070 
1071 	/*
1072 	 * Avoid unnecessary coalescing by freeing the pages in the largest
1073 	 * possible power-of-two-sized subsets.
1074 	 */
1075 	vm_domain_free_assert_locked(vm_pagequeue_domain(m));
1076 	for (;; npages -= n) {
1077 		/*
1078 		 * Unsigned "min" is used here so that "order" is assigned
1079 		 * "VM_NFREEORDER - 1" when "m"'s physical address is zero
1080 		 * or the low-order bits of its physical address are zero
1081 		 * because the size of a physical address exceeds the size of
1082 		 * a long.
1083 		 */
1084 		order = min(ffsl(VM_PAGE_TO_PHYS(m) >> PAGE_SHIFT) - 1,
1085 		    VM_NFREEORDER - 1);
1086 		n = 1 << order;
1087 		if (npages < n)
1088 			break;
1089 		vm_phys_free_pages(m, order);
1090 		m += n;
1091 	}
1092 	/* The residual "npages" is less than "1 << (VM_NFREEORDER - 1)". */
1093 	for (; npages > 0; npages -= n) {
1094 		order = flsl(npages) - 1;
1095 		n = 1 << order;
1096 		vm_phys_free_pages(m, order);
1097 		m += n;
1098 	}
1099 }
1100 
1101 /*
1102  * Scan physical memory between the specified addresses "low" and "high" for a
1103  * run of contiguous physical pages that satisfy the specified conditions, and
1104  * return the lowest page in the run.  The specified "alignment" determines
1105  * the alignment of the lowest physical page in the run.  If the specified
1106  * "boundary" is non-zero, then the run of physical pages cannot span a
1107  * physical address that is a multiple of "boundary".
1108  *
1109  * "npages" must be greater than zero.  Both "alignment" and "boundary" must
1110  * be a power of two.
1111  */
1112 vm_page_t
1113 vm_phys_scan_contig(int domain, u_long npages, vm_paddr_t low, vm_paddr_t high,
1114     u_long alignment, vm_paddr_t boundary, int options)
1115 {
1116 	vm_paddr_t pa_end;
1117 	vm_page_t m_end, m_run, m_start;
1118 	struct vm_phys_seg *seg;
1119 	int segind;
1120 
1121 	KASSERT(npages > 0, ("npages is 0"));
1122 	KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
1123 	KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
1124 	if (low >= high)
1125 		return (NULL);
1126 	for (segind = 0; segind < vm_phys_nsegs; segind++) {
1127 		seg = &vm_phys_segs[segind];
1128 		if (seg->domain != domain)
1129 			continue;
1130 		if (seg->start >= high)
1131 			break;
1132 		if (low >= seg->end)
1133 			continue;
1134 		if (low <= seg->start)
1135 			m_start = seg->first_page;
1136 		else
1137 			m_start = &seg->first_page[atop(low - seg->start)];
1138 		if (high < seg->end)
1139 			pa_end = high;
1140 		else
1141 			pa_end = seg->end;
1142 		if (pa_end - VM_PAGE_TO_PHYS(m_start) < ptoa(npages))
1143 			continue;
1144 		m_end = &seg->first_page[atop(pa_end - seg->start)];
1145 		m_run = vm_page_scan_contig(npages, m_start, m_end,
1146 		    alignment, boundary, options);
1147 		if (m_run != NULL)
1148 			return (m_run);
1149 	}
1150 	return (NULL);
1151 }
1152 
1153 /*
1154  * Set the pool for a contiguous, power of two-sized set of physical pages.
1155  */
1156 void
1157 vm_phys_set_pool(int pool, vm_page_t m, int order)
1158 {
1159 	vm_page_t m_tmp;
1160 
1161 	for (m_tmp = m; m_tmp < &m[1 << order]; m_tmp++)
1162 		m_tmp->pool = pool;
1163 }
1164 
1165 /*
1166  * Search for the given physical page "m" in the free lists.  If the search
1167  * succeeds, remove "m" from the free lists and return TRUE.  Otherwise, return
1168  * FALSE, indicating that "m" is not in the free lists.
1169  *
1170  * The free page queues must be locked.
1171  */
1172 boolean_t
1173 vm_phys_unfree_page(vm_page_t m)
1174 {
1175 	struct vm_freelist *fl;
1176 	struct vm_phys_seg *seg;
1177 	vm_paddr_t pa, pa_half;
1178 	vm_page_t m_set, m_tmp;
1179 	int order;
1180 
1181 	/*
1182 	 * First, find the contiguous, power of two-sized set of free
1183 	 * physical pages containing the given physical page "m" and
1184 	 * assign it to "m_set".
1185 	 */
1186 	seg = &vm_phys_segs[m->segind];
1187 	vm_domain_free_assert_locked(VM_DOMAIN(seg->domain));
1188 	for (m_set = m, order = 0; m_set->order == VM_NFREEORDER &&
1189 	    order < VM_NFREEORDER - 1; ) {
1190 		order++;
1191 		pa = m->phys_addr & (~(vm_paddr_t)0 << (PAGE_SHIFT + order));
1192 		if (pa >= seg->start)
1193 			m_set = &seg->first_page[atop(pa - seg->start)];
1194 		else
1195 			return (FALSE);
1196 	}
1197 	if (m_set->order < order)
1198 		return (FALSE);
1199 	if (m_set->order == VM_NFREEORDER)
1200 		return (FALSE);
1201 	KASSERT(m_set->order < VM_NFREEORDER,
1202 	    ("vm_phys_unfree_page: page %p has unexpected order %d",
1203 	    m_set, m_set->order));
1204 
1205 	/*
1206 	 * Next, remove "m_set" from the free lists.  Finally, extract
1207 	 * "m" from "m_set" using an iterative algorithm: While "m_set"
1208 	 * is larger than a page, shrink "m_set" by returning the half
1209 	 * of "m_set" that does not contain "m" to the free lists.
1210 	 */
1211 	fl = (*seg->free_queues)[m_set->pool];
1212 	order = m_set->order;
1213 	vm_freelist_rem(fl, m_set, order);
1214 	while (order > 0) {
1215 		order--;
1216 		pa_half = m_set->phys_addr ^ (1 << (PAGE_SHIFT + order));
1217 		if (m->phys_addr < pa_half)
1218 			m_tmp = &seg->first_page[atop(pa_half - seg->start)];
1219 		else {
1220 			m_tmp = m_set;
1221 			m_set = &seg->first_page[atop(pa_half - seg->start)];
1222 		}
1223 		vm_freelist_add(fl, m_tmp, order, 0);
1224 	}
1225 	KASSERT(m_set == m, ("vm_phys_unfree_page: fatal inconsistency"));
1226 	return (TRUE);
1227 }
1228 
1229 /*
1230  * Allocate a contiguous set of physical pages of the given size
1231  * "npages" from the free lists.  All of the physical pages must be at
1232  * or above the given physical address "low" and below the given
1233  * physical address "high".  The given value "alignment" determines the
1234  * alignment of the first physical page in the set.  If the given value
1235  * "boundary" is non-zero, then the set of physical pages cannot cross
1236  * any physical address boundary that is a multiple of that value.  Both
1237  * "alignment" and "boundary" must be a power of two.
1238  */
1239 vm_page_t
1240 vm_phys_alloc_contig(int domain, u_long npages, vm_paddr_t low, vm_paddr_t high,
1241     u_long alignment, vm_paddr_t boundary)
1242 {
1243 	vm_paddr_t pa_end, pa_start;
1244 	vm_page_t m_run;
1245 	struct vm_phys_seg *seg;
1246 	int segind;
1247 
1248 	KASSERT(npages > 0, ("npages is 0"));
1249 	KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
1250 	KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
1251 	vm_domain_free_assert_locked(VM_DOMAIN(domain));
1252 	if (low >= high)
1253 		return (NULL);
1254 	m_run = NULL;
1255 	for (segind = vm_phys_nsegs - 1; segind >= 0; segind--) {
1256 		seg = &vm_phys_segs[segind];
1257 		if (seg->start >= high || seg->domain != domain)
1258 			continue;
1259 		if (low >= seg->end)
1260 			break;
1261 		if (low <= seg->start)
1262 			pa_start = seg->start;
1263 		else
1264 			pa_start = low;
1265 		if (high < seg->end)
1266 			pa_end = high;
1267 		else
1268 			pa_end = seg->end;
1269 		if (pa_end - pa_start < ptoa(npages))
1270 			continue;
1271 		m_run = vm_phys_alloc_seg_contig(seg, npages, low, high,
1272 		    alignment, boundary);
1273 		if (m_run != NULL)
1274 			break;
1275 	}
1276 	return (m_run);
1277 }
1278 
1279 /*
1280  * Allocate a run of contiguous physical pages from the free list for the
1281  * specified segment.
1282  */
1283 static vm_page_t
1284 vm_phys_alloc_seg_contig(struct vm_phys_seg *seg, u_long npages,
1285     vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary)
1286 {
1287 	struct vm_freelist *fl;
1288 	vm_paddr_t pa, pa_end, size;
1289 	vm_page_t m, m_ret;
1290 	u_long npages_end;
1291 	int oind, order, pind;
1292 
1293 	KASSERT(npages > 0, ("npages is 0"));
1294 	KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
1295 	KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
1296 	vm_domain_free_assert_locked(VM_DOMAIN(seg->domain));
1297 	/* Compute the queue that is the best fit for npages. */
1298 	order = flsl(npages - 1);
1299 	/* Search for a run satisfying the specified conditions. */
1300 	size = npages << PAGE_SHIFT;
1301 	for (oind = min(order, VM_NFREEORDER - 1); oind < VM_NFREEORDER;
1302 	    oind++) {
1303 		for (pind = 0; pind < VM_NFREEPOOL; pind++) {
1304 			fl = (*seg->free_queues)[pind];
1305 			TAILQ_FOREACH(m_ret, &fl[oind].pl, listq) {
1306 				/*
1307 				 * Is the size of this allocation request
1308 				 * larger than the largest block size?
1309 				 */
1310 				if (order >= VM_NFREEORDER) {
1311 					/*
1312 					 * Determine if a sufficient number of
1313 					 * subsequent blocks to satisfy the
1314 					 * allocation request are free.
1315 					 */
1316 					pa = VM_PAGE_TO_PHYS(m_ret);
1317 					pa_end = pa + size;
1318 					if (pa_end < pa)
1319 						continue;
1320 					for (;;) {
1321 						pa += 1 << (PAGE_SHIFT +
1322 						    VM_NFREEORDER - 1);
1323 						if (pa >= pa_end ||
1324 						    pa < seg->start ||
1325 						    pa >= seg->end)
1326 							break;
1327 						m = &seg->first_page[atop(pa -
1328 						    seg->start)];
1329 						if (m->order != VM_NFREEORDER -
1330 						    1)
1331 							break;
1332 					}
1333 					/* If not, go to the next block. */
1334 					if (pa < pa_end)
1335 						continue;
1336 				}
1337 
1338 				/*
1339 				 * Determine if the blocks are within the
1340 				 * given range, satisfy the given alignment,
1341 				 * and do not cross the given boundary.
1342 				 */
1343 				pa = VM_PAGE_TO_PHYS(m_ret);
1344 				pa_end = pa + size;
1345 				if (pa >= low && pa_end <= high &&
1346 				    (pa & (alignment - 1)) == 0 &&
1347 				    rounddown2(pa ^ (pa_end - 1), boundary) == 0)
1348 					goto done;
1349 			}
1350 		}
1351 	}
1352 	return (NULL);
1353 done:
1354 	for (m = m_ret; m < &m_ret[npages]; m = &m[1 << oind]) {
1355 		fl = (*seg->free_queues)[m->pool];
1356 		vm_freelist_rem(fl, m, oind);
1357 		if (m->pool != VM_FREEPOOL_DEFAULT)
1358 			vm_phys_set_pool(VM_FREEPOOL_DEFAULT, m, oind);
1359 	}
1360 	/* Return excess pages to the free lists. */
1361 	npages_end = roundup2(npages, 1 << oind);
1362 	if (npages < npages_end) {
1363 		fl = (*seg->free_queues)[VM_FREEPOOL_DEFAULT];
1364 		vm_phys_enq_range(&m_ret[npages], npages_end - npages, fl, 0);
1365 	}
1366 	return (m_ret);
1367 }
1368 
1369 #ifdef DDB
1370 /*
1371  * Show the number of physical pages in each of the free lists.
1372  */
1373 DB_SHOW_COMMAND(freepages, db_show_freepages)
1374 {
1375 	struct vm_freelist *fl;
1376 	int flind, oind, pind, dom;
1377 
1378 	for (dom = 0; dom < vm_ndomains; dom++) {
1379 		db_printf("DOMAIN: %d\n", dom);
1380 		for (flind = 0; flind < vm_nfreelists; flind++) {
1381 			db_printf("FREE LIST %d:\n"
1382 			    "\n  ORDER (SIZE)  |  NUMBER"
1383 			    "\n              ", flind);
1384 			for (pind = 0; pind < VM_NFREEPOOL; pind++)
1385 				db_printf("  |  POOL %d", pind);
1386 			db_printf("\n--            ");
1387 			for (pind = 0; pind < VM_NFREEPOOL; pind++)
1388 				db_printf("-- --      ");
1389 			db_printf("--\n");
1390 			for (oind = VM_NFREEORDER - 1; oind >= 0; oind--) {
1391 				db_printf("  %2.2d (%6.6dK)", oind,
1392 				    1 << (PAGE_SHIFT - 10 + oind));
1393 				for (pind = 0; pind < VM_NFREEPOOL; pind++) {
1394 				fl = vm_phys_free_queues[dom][flind][pind];
1395 					db_printf("  |  %6.6d", fl[oind].lcnt);
1396 				}
1397 				db_printf("\n");
1398 			}
1399 			db_printf("\n");
1400 		}
1401 		db_printf("\n");
1402 	}
1403 }
1404 #endif
1405