xref: /freebsd/sys/vm/vm_phys.c (revision 01e115ab83a4b82216e02198dbcc7c794a752711)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3  *
4  * Copyright (c) 2002-2006 Rice University
5  * Copyright (c) 2007 Alan L. Cox <alc@cs.rice.edu>
6  * All rights reserved.
7  *
8  * This software was developed for the FreeBSD Project by Alan L. Cox,
9  * Olivier Crameri, Peter Druschel, Sitaram Iyer, and Juan Navarro.
10  *
11  * Redistribution and use in source and binary forms, with or without
12  * modification, are permitted provided that the following conditions
13  * are met:
14  * 1. Redistributions of source code must retain the above copyright
15  *    notice, this list of conditions and the following disclaimer.
16  * 2. Redistributions in binary form must reproduce the above copyright
17  *    notice, this list of conditions and the following disclaimer in the
18  *    documentation and/or other materials provided with the distribution.
19  *
20  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
21  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
22  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
23  * A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE COPYRIGHT
24  * HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
25  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
26  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS
27  * OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
28  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
29  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY
30  * WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
31  * POSSIBILITY OF SUCH DAMAGE.
32  */
33 
34 /*
35  *	Physical memory system implementation
36  *
37  * Any external functions defined by this module are only to be used by the
38  * virtual memory system.
39  */
40 
41 #include <sys/cdefs.h>
42 __FBSDID("$FreeBSD$");
43 
44 #include "opt_ddb.h"
45 #include "opt_vm.h"
46 
47 #include <sys/param.h>
48 #include <sys/systm.h>
49 #include <sys/domainset.h>
50 #include <sys/lock.h>
51 #include <sys/kernel.h>
52 #include <sys/malloc.h>
53 #include <sys/mutex.h>
54 #include <sys/proc.h>
55 #include <sys/queue.h>
56 #include <sys/rwlock.h>
57 #include <sys/sbuf.h>
58 #include <sys/sysctl.h>
59 #include <sys/tree.h>
60 #include <sys/vmmeter.h>
61 
62 #include <ddb/ddb.h>
63 
64 #include <vm/vm.h>
65 #include <vm/vm_extern.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 static struct vm_phys_seg vm_phys_early_segs[8];
87 static int vm_phys_early_nsegs;
88 
89 struct vm_phys_fictitious_seg;
90 static int vm_phys_fictitious_cmp(struct vm_phys_fictitious_seg *,
91     struct vm_phys_fictitious_seg *);
92 
93 RB_HEAD(fict_tree, vm_phys_fictitious_seg) vm_phys_fictitious_tree =
94     RB_INITIALIZER(&vm_phys_fictitious_tree);
95 
96 struct vm_phys_fictitious_seg {
97 	RB_ENTRY(vm_phys_fictitious_seg) node;
98 	/* Memory region data */
99 	vm_paddr_t	start;
100 	vm_paddr_t	end;
101 	vm_page_t	first_page;
102 };
103 
104 RB_GENERATE_STATIC(fict_tree, vm_phys_fictitious_seg, node,
105     vm_phys_fictitious_cmp);
106 
107 static struct rwlock_padalign vm_phys_fictitious_reg_lock;
108 MALLOC_DEFINE(M_FICT_PAGES, "vm_fictitious", "Fictitious VM pages");
109 
110 static struct vm_freelist __aligned(CACHE_LINE_SIZE)
111     vm_phys_free_queues[MAXMEMDOM][VM_NFREELIST][VM_NFREEPOOL]
112     [VM_NFREEORDER_MAX];
113 
114 static int __read_mostly vm_nfreelists;
115 
116 /*
117  * These "avail lists" are globals used to communicate boot-time physical
118  * memory layout to other parts of the kernel.  Each physically contiguous
119  * region of memory is defined by a start address at an even index and an
120  * end address at the following odd index.  Each list is terminated by a
121  * pair of zero entries.
122  *
123  * dump_avail tells the dump code what regions to include in a crash dump, and
124  * phys_avail is all of the remaining physical memory that is available for
125  * the vm system.
126  *
127  * Initially dump_avail and phys_avail are identical.  Boot time memory
128  * allocations remove extents from phys_avail that may still be included
129  * in dumps.
130  */
131 vm_paddr_t phys_avail[PHYS_AVAIL_COUNT];
132 vm_paddr_t dump_avail[PHYS_AVAIL_COUNT];
133 
134 /*
135  * Provides the mapping from VM_FREELIST_* to free list indices (flind).
136  */
137 static int __read_mostly vm_freelist_to_flind[VM_NFREELIST];
138 
139 CTASSERT(VM_FREELIST_DEFAULT == 0);
140 
141 #ifdef VM_FREELIST_DMA32
142 #define	VM_DMA32_BOUNDARY	((vm_paddr_t)1 << 32)
143 #endif
144 
145 /*
146  * Enforce the assumptions made by vm_phys_add_seg() and vm_phys_init() about
147  * the ordering of the free list boundaries.
148  */
149 #if defined(VM_LOWMEM_BOUNDARY) && defined(VM_DMA32_BOUNDARY)
150 CTASSERT(VM_LOWMEM_BOUNDARY < VM_DMA32_BOUNDARY);
151 #endif
152 
153 static int sysctl_vm_phys_free(SYSCTL_HANDLER_ARGS);
154 SYSCTL_OID(_vm, OID_AUTO, phys_free,
155     CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, 0,
156     sysctl_vm_phys_free, "A",
157     "Phys Free Info");
158 
159 static int sysctl_vm_phys_segs(SYSCTL_HANDLER_ARGS);
160 SYSCTL_OID(_vm, OID_AUTO, phys_segs,
161     CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, 0,
162     sysctl_vm_phys_segs, "A",
163     "Phys Seg Info");
164 
165 #ifdef NUMA
166 static int sysctl_vm_phys_locality(SYSCTL_HANDLER_ARGS);
167 SYSCTL_OID(_vm, OID_AUTO, phys_locality,
168     CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, 0,
169     sysctl_vm_phys_locality, "A",
170     "Phys Locality Info");
171 #endif
172 
173 SYSCTL_INT(_vm, OID_AUTO, ndomains, CTLFLAG_RD,
174     &vm_ndomains, 0, "Number of physical memory domains available.");
175 
176 static vm_page_t vm_phys_alloc_seg_contig(struct vm_phys_seg *seg,
177     u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment,
178     vm_paddr_t boundary);
179 static void _vm_phys_create_seg(vm_paddr_t start, vm_paddr_t end, int domain);
180 static void vm_phys_create_seg(vm_paddr_t start, vm_paddr_t end);
181 static void vm_phys_split_pages(vm_page_t m, int oind, struct vm_freelist *fl,
182     int order, int tail);
183 
184 /*
185  * Red-black tree helpers for vm fictitious range management.
186  */
187 static inline int
188 vm_phys_fictitious_in_range(struct vm_phys_fictitious_seg *p,
189     struct vm_phys_fictitious_seg *range)
190 {
191 
192 	KASSERT(range->start != 0 && range->end != 0,
193 	    ("Invalid range passed on search for vm_fictitious page"));
194 	if (p->start >= range->end)
195 		return (1);
196 	if (p->start < range->start)
197 		return (-1);
198 
199 	return (0);
200 }
201 
202 static int
203 vm_phys_fictitious_cmp(struct vm_phys_fictitious_seg *p1,
204     struct vm_phys_fictitious_seg *p2)
205 {
206 
207 	/* Check if this is a search for a page */
208 	if (p1->end == 0)
209 		return (vm_phys_fictitious_in_range(p1, p2));
210 
211 	KASSERT(p2->end != 0,
212     ("Invalid range passed as second parameter to vm fictitious comparison"));
213 
214 	/* Searching to add a new range */
215 	if (p1->end <= p2->start)
216 		return (-1);
217 	if (p1->start >= p2->end)
218 		return (1);
219 
220 	panic("Trying to add overlapping vm fictitious ranges:\n"
221 	    "[%#jx:%#jx] and [%#jx:%#jx]", (uintmax_t)p1->start,
222 	    (uintmax_t)p1->end, (uintmax_t)p2->start, (uintmax_t)p2->end);
223 }
224 
225 int
226 vm_phys_domain_match(int prefer, vm_paddr_t low, vm_paddr_t high)
227 {
228 #ifdef NUMA
229 	domainset_t mask;
230 	int i;
231 
232 	if (vm_ndomains == 1 || mem_affinity == NULL)
233 		return (0);
234 
235 	DOMAINSET_ZERO(&mask);
236 	/*
237 	 * Check for any memory that overlaps low, high.
238 	 */
239 	for (i = 0; mem_affinity[i].end != 0; i++)
240 		if (mem_affinity[i].start <= high &&
241 		    mem_affinity[i].end >= low)
242 			DOMAINSET_SET(mem_affinity[i].domain, &mask);
243 	if (prefer != -1 && DOMAINSET_ISSET(prefer, &mask))
244 		return (prefer);
245 	if (DOMAINSET_EMPTY(&mask))
246 		panic("vm_phys_domain_match:  Impossible constraint");
247 	return (DOMAINSET_FFS(&mask) - 1);
248 #else
249 	return (0);
250 #endif
251 }
252 
253 /*
254  * Outputs the state of the physical memory allocator, specifically,
255  * the amount of physical memory in each free list.
256  */
257 static int
258 sysctl_vm_phys_free(SYSCTL_HANDLER_ARGS)
259 {
260 	struct sbuf sbuf;
261 	struct vm_freelist *fl;
262 	int dom, error, flind, oind, pind;
263 
264 	error = sysctl_wire_old_buffer(req, 0);
265 	if (error != 0)
266 		return (error);
267 	sbuf_new_for_sysctl(&sbuf, NULL, 128 * vm_ndomains, req);
268 	for (dom = 0; dom < vm_ndomains; dom++) {
269 		sbuf_printf(&sbuf,"\nDOMAIN %d:\n", dom);
270 		for (flind = 0; flind < vm_nfreelists; flind++) {
271 			sbuf_printf(&sbuf, "\nFREE LIST %d:\n"
272 			    "\n  ORDER (SIZE)  |  NUMBER"
273 			    "\n              ", flind);
274 			for (pind = 0; pind < VM_NFREEPOOL; pind++)
275 				sbuf_printf(&sbuf, "  |  POOL %d", pind);
276 			sbuf_printf(&sbuf, "\n--            ");
277 			for (pind = 0; pind < VM_NFREEPOOL; pind++)
278 				sbuf_printf(&sbuf, "-- --      ");
279 			sbuf_printf(&sbuf, "--\n");
280 			for (oind = VM_NFREEORDER - 1; oind >= 0; oind--) {
281 				sbuf_printf(&sbuf, "  %2d (%6dK)", oind,
282 				    1 << (PAGE_SHIFT - 10 + oind));
283 				for (pind = 0; pind < VM_NFREEPOOL; pind++) {
284 				fl = vm_phys_free_queues[dom][flind][pind];
285 					sbuf_printf(&sbuf, "  |  %6d",
286 					    fl[oind].lcnt);
287 				}
288 				sbuf_printf(&sbuf, "\n");
289 			}
290 		}
291 	}
292 	error = sbuf_finish(&sbuf);
293 	sbuf_delete(&sbuf);
294 	return (error);
295 }
296 
297 /*
298  * Outputs the set of physical memory segments.
299  */
300 static int
301 sysctl_vm_phys_segs(SYSCTL_HANDLER_ARGS)
302 {
303 	struct sbuf sbuf;
304 	struct vm_phys_seg *seg;
305 	int error, segind;
306 
307 	error = sysctl_wire_old_buffer(req, 0);
308 	if (error != 0)
309 		return (error);
310 	sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
311 	for (segind = 0; segind < vm_phys_nsegs; segind++) {
312 		sbuf_printf(&sbuf, "\nSEGMENT %d:\n\n", segind);
313 		seg = &vm_phys_segs[segind];
314 		sbuf_printf(&sbuf, "start:     %#jx\n",
315 		    (uintmax_t)seg->start);
316 		sbuf_printf(&sbuf, "end:       %#jx\n",
317 		    (uintmax_t)seg->end);
318 		sbuf_printf(&sbuf, "domain:    %d\n", seg->domain);
319 		sbuf_printf(&sbuf, "free list: %p\n", seg->free_queues);
320 	}
321 	error = sbuf_finish(&sbuf);
322 	sbuf_delete(&sbuf);
323 	return (error);
324 }
325 
326 /*
327  * Return affinity, or -1 if there's no affinity information.
328  */
329 int
330 vm_phys_mem_affinity(int f, int t)
331 {
332 
333 #ifdef NUMA
334 	if (mem_locality == NULL)
335 		return (-1);
336 	if (f >= vm_ndomains || t >= vm_ndomains)
337 		return (-1);
338 	return (mem_locality[f * vm_ndomains + t]);
339 #else
340 	return (-1);
341 #endif
342 }
343 
344 #ifdef NUMA
345 /*
346  * Outputs the VM locality table.
347  */
348 static int
349 sysctl_vm_phys_locality(SYSCTL_HANDLER_ARGS)
350 {
351 	struct sbuf sbuf;
352 	int error, i, j;
353 
354 	error = sysctl_wire_old_buffer(req, 0);
355 	if (error != 0)
356 		return (error);
357 	sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
358 
359 	sbuf_printf(&sbuf, "\n");
360 
361 	for (i = 0; i < vm_ndomains; i++) {
362 		sbuf_printf(&sbuf, "%d: ", i);
363 		for (j = 0; j < vm_ndomains; j++) {
364 			sbuf_printf(&sbuf, "%d ", vm_phys_mem_affinity(i, j));
365 		}
366 		sbuf_printf(&sbuf, "\n");
367 	}
368 	error = sbuf_finish(&sbuf);
369 	sbuf_delete(&sbuf);
370 	return (error);
371 }
372 #endif
373 
374 static void
375 vm_freelist_add(struct vm_freelist *fl, vm_page_t m, int order, int tail)
376 {
377 
378 	m->order = order;
379 	if (tail)
380 		TAILQ_INSERT_TAIL(&fl[order].pl, m, listq);
381 	else
382 		TAILQ_INSERT_HEAD(&fl[order].pl, m, listq);
383 	fl[order].lcnt++;
384 }
385 
386 static void
387 vm_freelist_rem(struct vm_freelist *fl, vm_page_t m, int order)
388 {
389 
390 	TAILQ_REMOVE(&fl[order].pl, m, listq);
391 	fl[order].lcnt--;
392 	m->order = VM_NFREEORDER;
393 }
394 
395 /*
396  * Create a physical memory segment.
397  */
398 static void
399 _vm_phys_create_seg(vm_paddr_t start, vm_paddr_t end, int domain)
400 {
401 	struct vm_phys_seg *seg;
402 
403 	KASSERT(vm_phys_nsegs < VM_PHYSSEG_MAX,
404 	    ("vm_phys_create_seg: increase VM_PHYSSEG_MAX"));
405 	KASSERT(domain >= 0 && domain < vm_ndomains,
406 	    ("vm_phys_create_seg: invalid domain provided"));
407 	seg = &vm_phys_segs[vm_phys_nsegs++];
408 	while (seg > vm_phys_segs && (seg - 1)->start >= end) {
409 		*seg = *(seg - 1);
410 		seg--;
411 	}
412 	seg->start = start;
413 	seg->end = end;
414 	seg->domain = domain;
415 }
416 
417 static void
418 vm_phys_create_seg(vm_paddr_t start, vm_paddr_t end)
419 {
420 #ifdef NUMA
421 	int i;
422 
423 	if (mem_affinity == NULL) {
424 		_vm_phys_create_seg(start, end, 0);
425 		return;
426 	}
427 
428 	for (i = 0;; i++) {
429 		if (mem_affinity[i].end == 0)
430 			panic("Reached end of affinity info");
431 		if (mem_affinity[i].end <= start)
432 			continue;
433 		if (mem_affinity[i].start > start)
434 			panic("No affinity info for start %jx",
435 			    (uintmax_t)start);
436 		if (mem_affinity[i].end >= end) {
437 			_vm_phys_create_seg(start, end,
438 			    mem_affinity[i].domain);
439 			break;
440 		}
441 		_vm_phys_create_seg(start, mem_affinity[i].end,
442 		    mem_affinity[i].domain);
443 		start = mem_affinity[i].end;
444 	}
445 #else
446 	_vm_phys_create_seg(start, end, 0);
447 #endif
448 }
449 
450 /*
451  * Add a physical memory segment.
452  */
453 void
454 vm_phys_add_seg(vm_paddr_t start, vm_paddr_t end)
455 {
456 	vm_paddr_t paddr;
457 
458 	KASSERT((start & PAGE_MASK) == 0,
459 	    ("vm_phys_define_seg: start is not page aligned"));
460 	KASSERT((end & PAGE_MASK) == 0,
461 	    ("vm_phys_define_seg: end is not page aligned"));
462 
463 	/*
464 	 * Split the physical memory segment if it spans two or more free
465 	 * list boundaries.
466 	 */
467 	paddr = start;
468 #ifdef	VM_FREELIST_LOWMEM
469 	if (paddr < VM_LOWMEM_BOUNDARY && end > VM_LOWMEM_BOUNDARY) {
470 		vm_phys_create_seg(paddr, VM_LOWMEM_BOUNDARY);
471 		paddr = VM_LOWMEM_BOUNDARY;
472 	}
473 #endif
474 #ifdef	VM_FREELIST_DMA32
475 	if (paddr < VM_DMA32_BOUNDARY && end > VM_DMA32_BOUNDARY) {
476 		vm_phys_create_seg(paddr, VM_DMA32_BOUNDARY);
477 		paddr = VM_DMA32_BOUNDARY;
478 	}
479 #endif
480 	vm_phys_create_seg(paddr, end);
481 }
482 
483 /*
484  * Initialize the physical memory allocator.
485  *
486  * Requires that vm_page_array is initialized!
487  */
488 void
489 vm_phys_init(void)
490 {
491 	struct vm_freelist *fl;
492 	struct vm_phys_seg *end_seg, *prev_seg, *seg, *tmp_seg;
493 	u_long npages;
494 	int dom, flind, freelist, oind, pind, segind;
495 
496 	/*
497 	 * Compute the number of free lists, and generate the mapping from the
498 	 * manifest constants VM_FREELIST_* to the free list indices.
499 	 *
500 	 * Initially, the entries of vm_freelist_to_flind[] are set to either
501 	 * 0 or 1 to indicate which free lists should be created.
502 	 */
503 	npages = 0;
504 	for (segind = vm_phys_nsegs - 1; segind >= 0; segind--) {
505 		seg = &vm_phys_segs[segind];
506 #ifdef	VM_FREELIST_LOWMEM
507 		if (seg->end <= VM_LOWMEM_BOUNDARY)
508 			vm_freelist_to_flind[VM_FREELIST_LOWMEM] = 1;
509 		else
510 #endif
511 #ifdef	VM_FREELIST_DMA32
512 		if (
513 #ifdef	VM_DMA32_NPAGES_THRESHOLD
514 		    /*
515 		     * Create the DMA32 free list only if the amount of
516 		     * physical memory above physical address 4G exceeds the
517 		     * given threshold.
518 		     */
519 		    npages > VM_DMA32_NPAGES_THRESHOLD &&
520 #endif
521 		    seg->end <= VM_DMA32_BOUNDARY)
522 			vm_freelist_to_flind[VM_FREELIST_DMA32] = 1;
523 		else
524 #endif
525 		{
526 			npages += atop(seg->end - seg->start);
527 			vm_freelist_to_flind[VM_FREELIST_DEFAULT] = 1;
528 		}
529 	}
530 	/* Change each entry into a running total of the free lists. */
531 	for (freelist = 1; freelist < VM_NFREELIST; freelist++) {
532 		vm_freelist_to_flind[freelist] +=
533 		    vm_freelist_to_flind[freelist - 1];
534 	}
535 	vm_nfreelists = vm_freelist_to_flind[VM_NFREELIST - 1];
536 	KASSERT(vm_nfreelists > 0, ("vm_phys_init: no free lists"));
537 	/* Change each entry into a free list index. */
538 	for (freelist = 0; freelist < VM_NFREELIST; freelist++)
539 		vm_freelist_to_flind[freelist]--;
540 
541 	/*
542 	 * Initialize the first_page and free_queues fields of each physical
543 	 * memory segment.
544 	 */
545 #ifdef VM_PHYSSEG_SPARSE
546 	npages = 0;
547 #endif
548 	for (segind = 0; segind < vm_phys_nsegs; segind++) {
549 		seg = &vm_phys_segs[segind];
550 #ifdef VM_PHYSSEG_SPARSE
551 		seg->first_page = &vm_page_array[npages];
552 		npages += atop(seg->end - seg->start);
553 #else
554 		seg->first_page = PHYS_TO_VM_PAGE(seg->start);
555 #endif
556 #ifdef	VM_FREELIST_LOWMEM
557 		if (seg->end <= VM_LOWMEM_BOUNDARY) {
558 			flind = vm_freelist_to_flind[VM_FREELIST_LOWMEM];
559 			KASSERT(flind >= 0,
560 			    ("vm_phys_init: LOWMEM flind < 0"));
561 		} else
562 #endif
563 #ifdef	VM_FREELIST_DMA32
564 		if (seg->end <= VM_DMA32_BOUNDARY) {
565 			flind = vm_freelist_to_flind[VM_FREELIST_DMA32];
566 			KASSERT(flind >= 0,
567 			    ("vm_phys_init: DMA32 flind < 0"));
568 		} else
569 #endif
570 		{
571 			flind = vm_freelist_to_flind[VM_FREELIST_DEFAULT];
572 			KASSERT(flind >= 0,
573 			    ("vm_phys_init: DEFAULT flind < 0"));
574 		}
575 		seg->free_queues = &vm_phys_free_queues[seg->domain][flind];
576 	}
577 
578 	/*
579 	 * Coalesce physical memory segments that are contiguous and share the
580 	 * same per-domain free queues.
581 	 */
582 	prev_seg = vm_phys_segs;
583 	seg = &vm_phys_segs[1];
584 	end_seg = &vm_phys_segs[vm_phys_nsegs];
585 	while (seg < end_seg) {
586 		if (prev_seg->end == seg->start &&
587 		    prev_seg->free_queues == seg->free_queues) {
588 			prev_seg->end = seg->end;
589 			KASSERT(prev_seg->domain == seg->domain,
590 			    ("vm_phys_init: free queues cannot span domains"));
591 			vm_phys_nsegs--;
592 			end_seg--;
593 			for (tmp_seg = seg; tmp_seg < end_seg; tmp_seg++)
594 				*tmp_seg = *(tmp_seg + 1);
595 		} else {
596 			prev_seg = seg;
597 			seg++;
598 		}
599 	}
600 
601 	/*
602 	 * Initialize the free queues.
603 	 */
604 	for (dom = 0; dom < vm_ndomains; dom++) {
605 		for (flind = 0; flind < vm_nfreelists; flind++) {
606 			for (pind = 0; pind < VM_NFREEPOOL; pind++) {
607 				fl = vm_phys_free_queues[dom][flind][pind];
608 				for (oind = 0; oind < VM_NFREEORDER; oind++)
609 					TAILQ_INIT(&fl[oind].pl);
610 			}
611 		}
612 	}
613 
614 	rw_init(&vm_phys_fictitious_reg_lock, "vmfctr");
615 }
616 
617 /*
618  * Register info about the NUMA topology of the system.
619  *
620  * Invoked by platform-dependent code prior to vm_phys_init().
621  */
622 void
623 vm_phys_register_domains(int ndomains, struct mem_affinity *affinity,
624     int *locality)
625 {
626 #ifdef NUMA
627 	int d, i;
628 
629 	/*
630 	 * For now the only override value that we support is 1, which
631 	 * effectively disables NUMA-awareness in the allocators.
632 	 */
633 	d = 0;
634 	TUNABLE_INT_FETCH("vm.numa.disabled", &d);
635 	if (d)
636 		ndomains = 1;
637 
638 	if (ndomains > 1) {
639 		vm_ndomains = ndomains;
640 		mem_affinity = affinity;
641 		mem_locality = locality;
642 	}
643 
644 	for (i = 0; i < vm_ndomains; i++)
645 		DOMAINSET_SET(i, &all_domains);
646 #else
647 	(void)ndomains;
648 	(void)affinity;
649 	(void)locality;
650 #endif
651 }
652 
653 /*
654  * Split a contiguous, power of two-sized set of physical pages.
655  *
656  * When this function is called by a page allocation function, the caller
657  * should request insertion at the head unless the order [order, oind) queues
658  * are known to be empty.  The objective being to reduce the likelihood of
659  * long-term fragmentation by promoting contemporaneous allocation and
660  * (hopefully) deallocation.
661  */
662 static __inline void
663 vm_phys_split_pages(vm_page_t m, int oind, struct vm_freelist *fl, int order,
664     int tail)
665 {
666 	vm_page_t m_buddy;
667 
668 	while (oind > order) {
669 		oind--;
670 		m_buddy = &m[1 << oind];
671 		KASSERT(m_buddy->order == VM_NFREEORDER,
672 		    ("vm_phys_split_pages: page %p has unexpected order %d",
673 		    m_buddy, m_buddy->order));
674 		vm_freelist_add(fl, m_buddy, oind, tail);
675         }
676 }
677 
678 /*
679  * Add the physical pages [m, m + npages) at the end of a power-of-two aligned
680  * and sized set to the specified free list.
681  *
682  * When this function is called by a page allocation function, the caller
683  * should request insertion at the head unless the lower-order queues are
684  * known to be empty.  The objective being to reduce the likelihood of long-
685  * term fragmentation by promoting contemporaneous allocation and (hopefully)
686  * deallocation.
687  *
688  * The physical page m's buddy must not be free.
689  */
690 static void
691 vm_phys_enq_range(vm_page_t m, u_int npages, struct vm_freelist *fl, int tail)
692 {
693 	u_int n;
694 	int order;
695 
696 	KASSERT(npages > 0, ("vm_phys_enq_range: npages is 0"));
697 	KASSERT(((VM_PAGE_TO_PHYS(m) + npages * PAGE_SIZE) &
698 	    ((PAGE_SIZE << (fls(npages) - 1)) - 1)) == 0,
699 	    ("vm_phys_enq_range: page %p and npages %u are misaligned",
700 	    m, npages));
701 	do {
702 		KASSERT(m->order == VM_NFREEORDER,
703 		    ("vm_phys_enq_range: page %p has unexpected order %d",
704 		    m, m->order));
705 		order = ffs(npages) - 1;
706 		KASSERT(order < VM_NFREEORDER,
707 		    ("vm_phys_enq_range: order %d is out of range", order));
708 		vm_freelist_add(fl, m, order, tail);
709 		n = 1 << order;
710 		m += n;
711 		npages -= n;
712 	} while (npages > 0);
713 }
714 
715 /*
716  * Set the pool for a contiguous, power of two-sized set of physical pages.
717  */
718 static void
719 vm_phys_set_pool(int pool, vm_page_t m, int order)
720 {
721 	vm_page_t m_tmp;
722 
723 	for (m_tmp = m; m_tmp < &m[1 << order]; m_tmp++)
724 		m_tmp->pool = pool;
725 }
726 
727 /*
728  * Tries to allocate the specified number of pages from the specified pool
729  * within the specified domain.  Returns the actual number of allocated pages
730  * and a pointer to each page through the array ma[].
731  *
732  * The returned pages may not be physically contiguous.  However, in contrast
733  * to performing multiple, back-to-back calls to vm_phys_alloc_pages(..., 0),
734  * calling this function once to allocate the desired number of pages will
735  * avoid wasted time in vm_phys_split_pages().
736  *
737  * The free page queues for the specified domain must be locked.
738  */
739 int
740 vm_phys_alloc_npages(int domain, int pool, int npages, vm_page_t ma[])
741 {
742 	struct vm_freelist *alt, *fl;
743 	vm_page_t m;
744 	int avail, end, flind, freelist, i, need, oind, pind;
745 
746 	KASSERT(domain >= 0 && domain < vm_ndomains,
747 	    ("vm_phys_alloc_npages: domain %d is out of range", domain));
748 	KASSERT(pool < VM_NFREEPOOL,
749 	    ("vm_phys_alloc_npages: pool %d is out of range", pool));
750 	KASSERT(npages <= 1 << (VM_NFREEORDER - 1),
751 	    ("vm_phys_alloc_npages: npages %d is out of range", npages));
752 	vm_domain_free_assert_locked(VM_DOMAIN(domain));
753 	i = 0;
754 	for (freelist = 0; freelist < VM_NFREELIST; freelist++) {
755 		flind = vm_freelist_to_flind[freelist];
756 		if (flind < 0)
757 			continue;
758 		fl = vm_phys_free_queues[domain][flind][pool];
759 		for (oind = 0; oind < VM_NFREEORDER; oind++) {
760 			while ((m = TAILQ_FIRST(&fl[oind].pl)) != NULL) {
761 				vm_freelist_rem(fl, m, oind);
762 				avail = 1 << oind;
763 				need = imin(npages - i, avail);
764 				for (end = i + need; i < end;)
765 					ma[i++] = m++;
766 				if (need < avail) {
767 					/*
768 					 * Return excess pages to fl.  Its
769 					 * order [0, oind) queues are empty.
770 					 */
771 					vm_phys_enq_range(m, avail - need, fl,
772 					    1);
773 					return (npages);
774 				} else if (i == npages)
775 					return (npages);
776 			}
777 		}
778 		for (oind = VM_NFREEORDER - 1; oind >= 0; oind--) {
779 			for (pind = 0; pind < VM_NFREEPOOL; pind++) {
780 				alt = vm_phys_free_queues[domain][flind][pind];
781 				while ((m = TAILQ_FIRST(&alt[oind].pl)) !=
782 				    NULL) {
783 					vm_freelist_rem(alt, m, oind);
784 					vm_phys_set_pool(pool, m, oind);
785 					avail = 1 << oind;
786 					need = imin(npages - i, avail);
787 					for (end = i + need; i < end;)
788 						ma[i++] = m++;
789 					if (need < avail) {
790 						/*
791 						 * Return excess pages to fl.
792 						 * Its order [0, oind) queues
793 						 * are empty.
794 						 */
795 						vm_phys_enq_range(m, avail -
796 						    need, fl, 1);
797 						return (npages);
798 					} else if (i == npages)
799 						return (npages);
800 				}
801 			}
802 		}
803 	}
804 	return (i);
805 }
806 
807 /*
808  * Allocate a contiguous, power of two-sized set of physical pages
809  * from the free lists.
810  *
811  * The free page queues must be locked.
812  */
813 vm_page_t
814 vm_phys_alloc_pages(int domain, int pool, int order)
815 {
816 	vm_page_t m;
817 	int freelist;
818 
819 	for (freelist = 0; freelist < VM_NFREELIST; freelist++) {
820 		m = vm_phys_alloc_freelist_pages(domain, freelist, pool, order);
821 		if (m != NULL)
822 			return (m);
823 	}
824 	return (NULL);
825 }
826 
827 /*
828  * Allocate a contiguous, power of two-sized set of physical pages from the
829  * specified free list.  The free list must be specified using one of the
830  * manifest constants VM_FREELIST_*.
831  *
832  * The free page queues must be locked.
833  */
834 vm_page_t
835 vm_phys_alloc_freelist_pages(int domain, int freelist, int pool, int order)
836 {
837 	struct vm_freelist *alt, *fl;
838 	vm_page_t m;
839 	int oind, pind, flind;
840 
841 	KASSERT(domain >= 0 && domain < vm_ndomains,
842 	    ("vm_phys_alloc_freelist_pages: domain %d is out of range",
843 	    domain));
844 	KASSERT(freelist < VM_NFREELIST,
845 	    ("vm_phys_alloc_freelist_pages: freelist %d is out of range",
846 	    freelist));
847 	KASSERT(pool < VM_NFREEPOOL,
848 	    ("vm_phys_alloc_freelist_pages: pool %d is out of range", pool));
849 	KASSERT(order < VM_NFREEORDER,
850 	    ("vm_phys_alloc_freelist_pages: order %d is out of range", order));
851 
852 	flind = vm_freelist_to_flind[freelist];
853 	/* Check if freelist is present */
854 	if (flind < 0)
855 		return (NULL);
856 
857 	vm_domain_free_assert_locked(VM_DOMAIN(domain));
858 	fl = &vm_phys_free_queues[domain][flind][pool][0];
859 	for (oind = order; oind < VM_NFREEORDER; oind++) {
860 		m = TAILQ_FIRST(&fl[oind].pl);
861 		if (m != NULL) {
862 			vm_freelist_rem(fl, m, oind);
863 			/* The order [order, oind) queues are empty. */
864 			vm_phys_split_pages(m, oind, fl, order, 1);
865 			return (m);
866 		}
867 	}
868 
869 	/*
870 	 * The given pool was empty.  Find the largest
871 	 * contiguous, power-of-two-sized set of pages in any
872 	 * pool.  Transfer these pages to the given pool, and
873 	 * use them to satisfy the allocation.
874 	 */
875 	for (oind = VM_NFREEORDER - 1; oind >= order; oind--) {
876 		for (pind = 0; pind < VM_NFREEPOOL; pind++) {
877 			alt = &vm_phys_free_queues[domain][flind][pind][0];
878 			m = TAILQ_FIRST(&alt[oind].pl);
879 			if (m != NULL) {
880 				vm_freelist_rem(alt, m, oind);
881 				vm_phys_set_pool(pool, m, oind);
882 				/* The order [order, oind) queues are empty. */
883 				vm_phys_split_pages(m, oind, fl, order, 1);
884 				return (m);
885 			}
886 		}
887 	}
888 	return (NULL);
889 }
890 
891 /*
892  * Find the vm_page corresponding to the given physical address.
893  */
894 vm_page_t
895 vm_phys_paddr_to_vm_page(vm_paddr_t pa)
896 {
897 	struct vm_phys_seg *seg;
898 	int segind;
899 
900 	for (segind = 0; segind < vm_phys_nsegs; segind++) {
901 		seg = &vm_phys_segs[segind];
902 		if (pa >= seg->start && pa < seg->end)
903 			return (&seg->first_page[atop(pa - seg->start)]);
904 	}
905 	return (NULL);
906 }
907 
908 vm_page_t
909 vm_phys_fictitious_to_vm_page(vm_paddr_t pa)
910 {
911 	struct vm_phys_fictitious_seg tmp, *seg;
912 	vm_page_t m;
913 
914 	m = NULL;
915 	tmp.start = pa;
916 	tmp.end = 0;
917 
918 	rw_rlock(&vm_phys_fictitious_reg_lock);
919 	seg = RB_FIND(fict_tree, &vm_phys_fictitious_tree, &tmp);
920 	rw_runlock(&vm_phys_fictitious_reg_lock);
921 	if (seg == NULL)
922 		return (NULL);
923 
924 	m = &seg->first_page[atop(pa - seg->start)];
925 	KASSERT((m->flags & PG_FICTITIOUS) != 0, ("%p not fictitious", m));
926 
927 	return (m);
928 }
929 
930 static inline void
931 vm_phys_fictitious_init_range(vm_page_t range, vm_paddr_t start,
932     long page_count, vm_memattr_t memattr)
933 {
934 	long i;
935 
936 	bzero(range, page_count * sizeof(*range));
937 	for (i = 0; i < page_count; i++) {
938 		vm_page_initfake(&range[i], start + PAGE_SIZE * i, memattr);
939 		range[i].oflags &= ~VPO_UNMANAGED;
940 		range[i].busy_lock = VPB_UNBUSIED;
941 	}
942 }
943 
944 int
945 vm_phys_fictitious_reg_range(vm_paddr_t start, vm_paddr_t end,
946     vm_memattr_t memattr)
947 {
948 	struct vm_phys_fictitious_seg *seg;
949 	vm_page_t fp;
950 	long page_count;
951 #ifdef VM_PHYSSEG_DENSE
952 	long pi, pe;
953 	long dpage_count;
954 #endif
955 
956 	KASSERT(start < end,
957 	    ("Start of segment isn't less than end (start: %jx end: %jx)",
958 	    (uintmax_t)start, (uintmax_t)end));
959 
960 	page_count = (end - start) / PAGE_SIZE;
961 
962 #ifdef VM_PHYSSEG_DENSE
963 	pi = atop(start);
964 	pe = atop(end);
965 	if (pi >= first_page && (pi - first_page) < vm_page_array_size) {
966 		fp = &vm_page_array[pi - first_page];
967 		if ((pe - first_page) > vm_page_array_size) {
968 			/*
969 			 * We have a segment that starts inside
970 			 * of vm_page_array, but ends outside of it.
971 			 *
972 			 * Use vm_page_array pages for those that are
973 			 * inside of the vm_page_array range, and
974 			 * allocate the remaining ones.
975 			 */
976 			dpage_count = vm_page_array_size - (pi - first_page);
977 			vm_phys_fictitious_init_range(fp, start, dpage_count,
978 			    memattr);
979 			page_count -= dpage_count;
980 			start += ptoa(dpage_count);
981 			goto alloc;
982 		}
983 		/*
984 		 * We can allocate the full range from vm_page_array,
985 		 * so there's no need to register the range in the tree.
986 		 */
987 		vm_phys_fictitious_init_range(fp, start, page_count, memattr);
988 		return (0);
989 	} else if (pe > first_page && (pe - first_page) < vm_page_array_size) {
990 		/*
991 		 * We have a segment that ends inside of vm_page_array,
992 		 * but starts outside of it.
993 		 */
994 		fp = &vm_page_array[0];
995 		dpage_count = pe - first_page;
996 		vm_phys_fictitious_init_range(fp, ptoa(first_page), dpage_count,
997 		    memattr);
998 		end -= ptoa(dpage_count);
999 		page_count -= dpage_count;
1000 		goto alloc;
1001 	} else if (pi < first_page && pe > (first_page + vm_page_array_size)) {
1002 		/*
1003 		 * Trying to register a fictitious range that expands before
1004 		 * and after vm_page_array.
1005 		 */
1006 		return (EINVAL);
1007 	} else {
1008 alloc:
1009 #endif
1010 		fp = malloc(page_count * sizeof(struct vm_page), M_FICT_PAGES,
1011 		    M_WAITOK);
1012 #ifdef VM_PHYSSEG_DENSE
1013 	}
1014 #endif
1015 	vm_phys_fictitious_init_range(fp, start, page_count, memattr);
1016 
1017 	seg = malloc(sizeof(*seg), M_FICT_PAGES, M_WAITOK | M_ZERO);
1018 	seg->start = start;
1019 	seg->end = end;
1020 	seg->first_page = fp;
1021 
1022 	rw_wlock(&vm_phys_fictitious_reg_lock);
1023 	RB_INSERT(fict_tree, &vm_phys_fictitious_tree, seg);
1024 	rw_wunlock(&vm_phys_fictitious_reg_lock);
1025 
1026 	return (0);
1027 }
1028 
1029 void
1030 vm_phys_fictitious_unreg_range(vm_paddr_t start, vm_paddr_t end)
1031 {
1032 	struct vm_phys_fictitious_seg *seg, tmp;
1033 #ifdef VM_PHYSSEG_DENSE
1034 	long pi, pe;
1035 #endif
1036 
1037 	KASSERT(start < end,
1038 	    ("Start of segment isn't less than end (start: %jx end: %jx)",
1039 	    (uintmax_t)start, (uintmax_t)end));
1040 
1041 #ifdef VM_PHYSSEG_DENSE
1042 	pi = atop(start);
1043 	pe = atop(end);
1044 	if (pi >= first_page && (pi - first_page) < vm_page_array_size) {
1045 		if ((pe - first_page) <= vm_page_array_size) {
1046 			/*
1047 			 * This segment was allocated using vm_page_array
1048 			 * only, there's nothing to do since those pages
1049 			 * were never added to the tree.
1050 			 */
1051 			return;
1052 		}
1053 		/*
1054 		 * We have a segment that starts inside
1055 		 * of vm_page_array, but ends outside of it.
1056 		 *
1057 		 * Calculate how many pages were added to the
1058 		 * tree and free them.
1059 		 */
1060 		start = ptoa(first_page + vm_page_array_size);
1061 	} else if (pe > first_page && (pe - first_page) < vm_page_array_size) {
1062 		/*
1063 		 * We have a segment that ends inside of vm_page_array,
1064 		 * but starts outside of it.
1065 		 */
1066 		end = ptoa(first_page);
1067 	} else if (pi < first_page && pe > (first_page + vm_page_array_size)) {
1068 		/* Since it's not possible to register such a range, panic. */
1069 		panic(
1070 		    "Unregistering not registered fictitious range [%#jx:%#jx]",
1071 		    (uintmax_t)start, (uintmax_t)end);
1072 	}
1073 #endif
1074 	tmp.start = start;
1075 	tmp.end = 0;
1076 
1077 	rw_wlock(&vm_phys_fictitious_reg_lock);
1078 	seg = RB_FIND(fict_tree, &vm_phys_fictitious_tree, &tmp);
1079 	if (seg->start != start || seg->end != end) {
1080 		rw_wunlock(&vm_phys_fictitious_reg_lock);
1081 		panic(
1082 		    "Unregistering not registered fictitious range [%#jx:%#jx]",
1083 		    (uintmax_t)start, (uintmax_t)end);
1084 	}
1085 	RB_REMOVE(fict_tree, &vm_phys_fictitious_tree, seg);
1086 	rw_wunlock(&vm_phys_fictitious_reg_lock);
1087 	free(seg->first_page, M_FICT_PAGES);
1088 	free(seg, M_FICT_PAGES);
1089 }
1090 
1091 /*
1092  * Free a contiguous, power of two-sized set of physical pages.
1093  *
1094  * The free page queues must be locked.
1095  */
1096 void
1097 vm_phys_free_pages(vm_page_t m, int order)
1098 {
1099 	struct vm_freelist *fl;
1100 	struct vm_phys_seg *seg;
1101 	vm_paddr_t pa;
1102 	vm_page_t m_buddy;
1103 
1104 	KASSERT(m->order == VM_NFREEORDER,
1105 	    ("vm_phys_free_pages: page %p has unexpected order %d",
1106 	    m, m->order));
1107 	KASSERT(m->pool < VM_NFREEPOOL,
1108 	    ("vm_phys_free_pages: page %p has unexpected pool %d",
1109 	    m, m->pool));
1110 	KASSERT(order < VM_NFREEORDER,
1111 	    ("vm_phys_free_pages: order %d is out of range", order));
1112 	seg = &vm_phys_segs[m->segind];
1113 	vm_domain_free_assert_locked(VM_DOMAIN(seg->domain));
1114 	if (order < VM_NFREEORDER - 1) {
1115 		pa = VM_PAGE_TO_PHYS(m);
1116 		do {
1117 			pa ^= ((vm_paddr_t)1 << (PAGE_SHIFT + order));
1118 			if (pa < seg->start || pa >= seg->end)
1119 				break;
1120 			m_buddy = &seg->first_page[atop(pa - seg->start)];
1121 			if (m_buddy->order != order)
1122 				break;
1123 			fl = (*seg->free_queues)[m_buddy->pool];
1124 			vm_freelist_rem(fl, m_buddy, order);
1125 			if (m_buddy->pool != m->pool)
1126 				vm_phys_set_pool(m->pool, m_buddy, order);
1127 			order++;
1128 			pa &= ~(((vm_paddr_t)1 << (PAGE_SHIFT + order)) - 1);
1129 			m = &seg->first_page[atop(pa - seg->start)];
1130 		} while (order < VM_NFREEORDER - 1);
1131 	}
1132 	fl = (*seg->free_queues)[m->pool];
1133 	vm_freelist_add(fl, m, order, 1);
1134 }
1135 
1136 /*
1137  * Return the largest possible order of a set of pages starting at m.
1138  */
1139 static int
1140 max_order(vm_page_t m)
1141 {
1142 
1143 	/*
1144 	 * Unsigned "min" is used here so that "order" is assigned
1145 	 * "VM_NFREEORDER - 1" when "m"'s physical address is zero
1146 	 * or the low-order bits of its physical address are zero
1147 	 * because the size of a physical address exceeds the size of
1148 	 * a long.
1149 	 */
1150 	return (min(ffsl(VM_PAGE_TO_PHYS(m) >> PAGE_SHIFT) - 1,
1151 	    VM_NFREEORDER - 1));
1152 }
1153 
1154 /*
1155  * Free a contiguous, arbitrarily sized set of physical pages, without
1156  * merging across set boundaries.
1157  *
1158  * The free page queues must be locked.
1159  */
1160 void
1161 vm_phys_enqueue_contig(vm_page_t m, u_long npages)
1162 {
1163 	struct vm_freelist *fl;
1164 	struct vm_phys_seg *seg;
1165 	vm_page_t m_end;
1166 	int order;
1167 
1168 	/*
1169 	 * Avoid unnecessary coalescing by freeing the pages in the largest
1170 	 * possible power-of-two-sized subsets.
1171 	 */
1172 	vm_domain_free_assert_locked(vm_pagequeue_domain(m));
1173 	seg = &vm_phys_segs[m->segind];
1174 	fl = (*seg->free_queues)[m->pool];
1175 	m_end = m + npages;
1176 	/* Free blocks of increasing size. */
1177 	while ((order = max_order(m)) < VM_NFREEORDER - 1 &&
1178 	    m + (1 << order) <= m_end) {
1179 		KASSERT(seg == &vm_phys_segs[m->segind],
1180 		    ("%s: page range [%p,%p) spans multiple segments",
1181 		    __func__, m_end - npages, m));
1182 		vm_freelist_add(fl, m, order, 1);
1183 		m += 1 << order;
1184 	}
1185 	/* Free blocks of maximum size. */
1186 	while (m + (1 << order) <= m_end) {
1187 		KASSERT(seg == &vm_phys_segs[m->segind],
1188 		    ("%s: page range [%p,%p) spans multiple segments",
1189 		    __func__, m_end - npages, m));
1190 		vm_freelist_add(fl, m, order, 1);
1191 		m += 1 << order;
1192 	}
1193 	/* Free blocks of diminishing size. */
1194 	while (m < m_end) {
1195 		KASSERT(seg == &vm_phys_segs[m->segind],
1196 		    ("%s: page range [%p,%p) spans multiple segments",
1197 		    __func__, m_end - npages, m));
1198 		order = flsl(m_end - m) - 1;
1199 		vm_freelist_add(fl, m, order, 1);
1200 		m += 1 << order;
1201 	}
1202 }
1203 
1204 /*
1205  * Free a contiguous, arbitrarily sized set of physical pages.
1206  *
1207  * The free page queues must be locked.
1208  */
1209 void
1210 vm_phys_free_contig(vm_page_t m, u_long npages)
1211 {
1212 	int order_start, order_end;
1213 	vm_page_t m_start, m_end;
1214 
1215 	vm_domain_free_assert_locked(vm_pagequeue_domain(m));
1216 
1217 	m_start = m;
1218 	order_start = max_order(m_start);
1219 	if (order_start < VM_NFREEORDER - 1)
1220 		m_start += 1 << order_start;
1221 	m_end = m + npages;
1222 	order_end = max_order(m_end);
1223 	if (order_end < VM_NFREEORDER - 1)
1224 		m_end -= 1 << order_end;
1225 	/*
1226 	 * Avoid unnecessary coalescing by freeing the pages at the start and
1227 	 * end of the range last.
1228 	 */
1229 	if (m_start < m_end)
1230 		vm_phys_enqueue_contig(m_start, m_end - m_start);
1231 	if (order_start < VM_NFREEORDER - 1)
1232 		vm_phys_free_pages(m, order_start);
1233 	if (order_end < VM_NFREEORDER - 1)
1234 		vm_phys_free_pages(m_end, order_end);
1235 }
1236 
1237 /*
1238  * Scan physical memory between the specified addresses "low" and "high" for a
1239  * run of contiguous physical pages that satisfy the specified conditions, and
1240  * return the lowest page in the run.  The specified "alignment" determines
1241  * the alignment of the lowest physical page in the run.  If the specified
1242  * "boundary" is non-zero, then the run of physical pages cannot span a
1243  * physical address that is a multiple of "boundary".
1244  *
1245  * "npages" must be greater than zero.  Both "alignment" and "boundary" must
1246  * be a power of two.
1247  */
1248 vm_page_t
1249 vm_phys_scan_contig(int domain, u_long npages, vm_paddr_t low, vm_paddr_t high,
1250     u_long alignment, vm_paddr_t boundary, int options)
1251 {
1252 	vm_paddr_t pa_end;
1253 	vm_page_t m_end, m_run, m_start;
1254 	struct vm_phys_seg *seg;
1255 	int segind;
1256 
1257 	KASSERT(npages > 0, ("npages is 0"));
1258 	KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
1259 	KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
1260 	if (low >= high)
1261 		return (NULL);
1262 	for (segind = 0; segind < vm_phys_nsegs; segind++) {
1263 		seg = &vm_phys_segs[segind];
1264 		if (seg->domain != domain)
1265 			continue;
1266 		if (seg->start >= high)
1267 			break;
1268 		if (low >= seg->end)
1269 			continue;
1270 		if (low <= seg->start)
1271 			m_start = seg->first_page;
1272 		else
1273 			m_start = &seg->first_page[atop(low - seg->start)];
1274 		if (high < seg->end)
1275 			pa_end = high;
1276 		else
1277 			pa_end = seg->end;
1278 		if (pa_end - VM_PAGE_TO_PHYS(m_start) < ptoa(npages))
1279 			continue;
1280 		m_end = &seg->first_page[atop(pa_end - seg->start)];
1281 		m_run = vm_page_scan_contig(npages, m_start, m_end,
1282 		    alignment, boundary, options);
1283 		if (m_run != NULL)
1284 			return (m_run);
1285 	}
1286 	return (NULL);
1287 }
1288 
1289 /*
1290  * Search for the given physical page "m" in the free lists.  If the search
1291  * succeeds, remove "m" from the free lists and return TRUE.  Otherwise, return
1292  * FALSE, indicating that "m" is not in the free lists.
1293  *
1294  * The free page queues must be locked.
1295  */
1296 boolean_t
1297 vm_phys_unfree_page(vm_page_t m)
1298 {
1299 	struct vm_freelist *fl;
1300 	struct vm_phys_seg *seg;
1301 	vm_paddr_t pa, pa_half;
1302 	vm_page_t m_set, m_tmp;
1303 	int order;
1304 
1305 	/*
1306 	 * First, find the contiguous, power of two-sized set of free
1307 	 * physical pages containing the given physical page "m" and
1308 	 * assign it to "m_set".
1309 	 */
1310 	seg = &vm_phys_segs[m->segind];
1311 	vm_domain_free_assert_locked(VM_DOMAIN(seg->domain));
1312 	for (m_set = m, order = 0; m_set->order == VM_NFREEORDER &&
1313 	    order < VM_NFREEORDER - 1; ) {
1314 		order++;
1315 		pa = m->phys_addr & (~(vm_paddr_t)0 << (PAGE_SHIFT + order));
1316 		if (pa >= seg->start)
1317 			m_set = &seg->first_page[atop(pa - seg->start)];
1318 		else
1319 			return (FALSE);
1320 	}
1321 	if (m_set->order < order)
1322 		return (FALSE);
1323 	if (m_set->order == VM_NFREEORDER)
1324 		return (FALSE);
1325 	KASSERT(m_set->order < VM_NFREEORDER,
1326 	    ("vm_phys_unfree_page: page %p has unexpected order %d",
1327 	    m_set, m_set->order));
1328 
1329 	/*
1330 	 * Next, remove "m_set" from the free lists.  Finally, extract
1331 	 * "m" from "m_set" using an iterative algorithm: While "m_set"
1332 	 * is larger than a page, shrink "m_set" by returning the half
1333 	 * of "m_set" that does not contain "m" to the free lists.
1334 	 */
1335 	fl = (*seg->free_queues)[m_set->pool];
1336 	order = m_set->order;
1337 	vm_freelist_rem(fl, m_set, order);
1338 	while (order > 0) {
1339 		order--;
1340 		pa_half = m_set->phys_addr ^ (1 << (PAGE_SHIFT + order));
1341 		if (m->phys_addr < pa_half)
1342 			m_tmp = &seg->first_page[atop(pa_half - seg->start)];
1343 		else {
1344 			m_tmp = m_set;
1345 			m_set = &seg->first_page[atop(pa_half - seg->start)];
1346 		}
1347 		vm_freelist_add(fl, m_tmp, order, 0);
1348 	}
1349 	KASSERT(m_set == m, ("vm_phys_unfree_page: fatal inconsistency"));
1350 	return (TRUE);
1351 }
1352 
1353 /*
1354  * Allocate a contiguous set of physical pages of the given size
1355  * "npages" from the free lists.  All of the physical pages must be at
1356  * or above the given physical address "low" and below the given
1357  * physical address "high".  The given value "alignment" determines the
1358  * alignment of the first physical page in the set.  If the given value
1359  * "boundary" is non-zero, then the set of physical pages cannot cross
1360  * any physical address boundary that is a multiple of that value.  Both
1361  * "alignment" and "boundary" must be a power of two.
1362  */
1363 vm_page_t
1364 vm_phys_alloc_contig(int domain, u_long npages, vm_paddr_t low, vm_paddr_t high,
1365     u_long alignment, vm_paddr_t boundary)
1366 {
1367 	vm_paddr_t pa_end, pa_start;
1368 	vm_page_t m_run;
1369 	struct vm_phys_seg *seg;
1370 	int segind;
1371 
1372 	KASSERT(npages > 0, ("npages is 0"));
1373 	KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
1374 	KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
1375 	vm_domain_free_assert_locked(VM_DOMAIN(domain));
1376 	if (low >= high)
1377 		return (NULL);
1378 	m_run = NULL;
1379 	for (segind = vm_phys_nsegs - 1; segind >= 0; segind--) {
1380 		seg = &vm_phys_segs[segind];
1381 		if (seg->start >= high || seg->domain != domain)
1382 			continue;
1383 		if (low >= seg->end)
1384 			break;
1385 		if (low <= seg->start)
1386 			pa_start = seg->start;
1387 		else
1388 			pa_start = low;
1389 		if (high < seg->end)
1390 			pa_end = high;
1391 		else
1392 			pa_end = seg->end;
1393 		if (pa_end - pa_start < ptoa(npages))
1394 			continue;
1395 		m_run = vm_phys_alloc_seg_contig(seg, npages, low, high,
1396 		    alignment, boundary);
1397 		if (m_run != NULL)
1398 			break;
1399 	}
1400 	return (m_run);
1401 }
1402 
1403 /*
1404  * Allocate a run of contiguous physical pages from the free list for the
1405  * specified segment.
1406  */
1407 static vm_page_t
1408 vm_phys_alloc_seg_contig(struct vm_phys_seg *seg, u_long npages,
1409     vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary)
1410 {
1411 	struct vm_freelist *fl;
1412 	vm_paddr_t pa, pa_end, size;
1413 	vm_page_t m, m_ret;
1414 	u_long npages_end;
1415 	int oind, order, pind;
1416 
1417 	KASSERT(npages > 0, ("npages is 0"));
1418 	KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
1419 	KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
1420 	vm_domain_free_assert_locked(VM_DOMAIN(seg->domain));
1421 	/* Compute the queue that is the best fit for npages. */
1422 	order = flsl(npages - 1);
1423 	/* Search for a run satisfying the specified conditions. */
1424 	size = npages << PAGE_SHIFT;
1425 	for (oind = min(order, VM_NFREEORDER - 1); oind < VM_NFREEORDER;
1426 	    oind++) {
1427 		for (pind = 0; pind < VM_NFREEPOOL; pind++) {
1428 			fl = (*seg->free_queues)[pind];
1429 			TAILQ_FOREACH(m_ret, &fl[oind].pl, listq) {
1430 				/*
1431 				 * Is the size of this allocation request
1432 				 * larger than the largest block size?
1433 				 */
1434 				if (order >= VM_NFREEORDER) {
1435 					/*
1436 					 * Determine if a sufficient number of
1437 					 * subsequent blocks to satisfy the
1438 					 * allocation request are free.
1439 					 */
1440 					pa = VM_PAGE_TO_PHYS(m_ret);
1441 					pa_end = pa + size;
1442 					if (pa_end < pa)
1443 						continue;
1444 					for (;;) {
1445 						pa += 1 << (PAGE_SHIFT +
1446 						    VM_NFREEORDER - 1);
1447 						if (pa >= pa_end ||
1448 						    pa < seg->start ||
1449 						    pa >= seg->end)
1450 							break;
1451 						m = &seg->first_page[atop(pa -
1452 						    seg->start)];
1453 						if (m->order != VM_NFREEORDER -
1454 						    1)
1455 							break;
1456 					}
1457 					/* If not, go to the next block. */
1458 					if (pa < pa_end)
1459 						continue;
1460 				}
1461 
1462 				/*
1463 				 * Determine if the blocks are within the
1464 				 * given range, satisfy the given alignment,
1465 				 * and do not cross the given boundary.
1466 				 */
1467 				pa = VM_PAGE_TO_PHYS(m_ret);
1468 				pa_end = pa + size;
1469 				if (pa >= low && pa_end <= high &&
1470 				    vm_addr_ok(pa, size, alignment, boundary))
1471 					goto done;
1472 			}
1473 		}
1474 	}
1475 	return (NULL);
1476 done:
1477 	for (m = m_ret; m < &m_ret[npages]; m = &m[1 << oind]) {
1478 		fl = (*seg->free_queues)[m->pool];
1479 		vm_freelist_rem(fl, m, oind);
1480 		if (m->pool != VM_FREEPOOL_DEFAULT)
1481 			vm_phys_set_pool(VM_FREEPOOL_DEFAULT, m, oind);
1482 	}
1483 	/* Return excess pages to the free lists. */
1484 	npages_end = roundup2(npages, 1 << oind);
1485 	if (npages < npages_end) {
1486 		fl = (*seg->free_queues)[VM_FREEPOOL_DEFAULT];
1487 		vm_phys_enq_range(&m_ret[npages], npages_end - npages, fl, 0);
1488 	}
1489 	return (m_ret);
1490 }
1491 
1492 /*
1493  * Return the index of the first unused slot which may be the terminating
1494  * entry.
1495  */
1496 static int
1497 vm_phys_avail_count(void)
1498 {
1499 	int i;
1500 
1501 	for (i = 0; phys_avail[i + 1]; i += 2)
1502 		continue;
1503 	if (i > PHYS_AVAIL_ENTRIES)
1504 		panic("Improperly terminated phys_avail %d entries", i);
1505 
1506 	return (i);
1507 }
1508 
1509 /*
1510  * Assert that a phys_avail entry is valid.
1511  */
1512 static void
1513 vm_phys_avail_check(int i)
1514 {
1515 	if (phys_avail[i] & PAGE_MASK)
1516 		panic("Unaligned phys_avail[%d]: %#jx", i,
1517 		    (intmax_t)phys_avail[i]);
1518 	if (phys_avail[i+1] & PAGE_MASK)
1519 		panic("Unaligned phys_avail[%d + 1]: %#jx", i,
1520 		    (intmax_t)phys_avail[i]);
1521 	if (phys_avail[i + 1] < phys_avail[i])
1522 		panic("phys_avail[%d] start %#jx < end %#jx", i,
1523 		    (intmax_t)phys_avail[i], (intmax_t)phys_avail[i+1]);
1524 }
1525 
1526 /*
1527  * Return the index of an overlapping phys_avail entry or -1.
1528  */
1529 #ifdef NUMA
1530 static int
1531 vm_phys_avail_find(vm_paddr_t pa)
1532 {
1533 	int i;
1534 
1535 	for (i = 0; phys_avail[i + 1]; i += 2)
1536 		if (phys_avail[i] <= pa && phys_avail[i + 1] > pa)
1537 			return (i);
1538 	return (-1);
1539 }
1540 #endif
1541 
1542 /*
1543  * Return the index of the largest entry.
1544  */
1545 int
1546 vm_phys_avail_largest(void)
1547 {
1548 	vm_paddr_t sz, largesz;
1549 	int largest;
1550 	int i;
1551 
1552 	largest = 0;
1553 	largesz = 0;
1554 	for (i = 0; phys_avail[i + 1]; i += 2) {
1555 		sz = vm_phys_avail_size(i);
1556 		if (sz > largesz) {
1557 			largesz = sz;
1558 			largest = i;
1559 		}
1560 	}
1561 
1562 	return (largest);
1563 }
1564 
1565 vm_paddr_t
1566 vm_phys_avail_size(int i)
1567 {
1568 
1569 	return (phys_avail[i + 1] - phys_avail[i]);
1570 }
1571 
1572 /*
1573  * Split an entry at the address 'pa'.  Return zero on success or errno.
1574  */
1575 static int
1576 vm_phys_avail_split(vm_paddr_t pa, int i)
1577 {
1578 	int cnt;
1579 
1580 	vm_phys_avail_check(i);
1581 	if (pa <= phys_avail[i] || pa >= phys_avail[i + 1])
1582 		panic("vm_phys_avail_split: invalid address");
1583 	cnt = vm_phys_avail_count();
1584 	if (cnt >= PHYS_AVAIL_ENTRIES)
1585 		return (ENOSPC);
1586 	memmove(&phys_avail[i + 2], &phys_avail[i],
1587 	    (cnt - i) * sizeof(phys_avail[0]));
1588 	phys_avail[i + 1] = pa;
1589 	phys_avail[i + 2] = pa;
1590 	vm_phys_avail_check(i);
1591 	vm_phys_avail_check(i+2);
1592 
1593 	return (0);
1594 }
1595 
1596 /*
1597  * Check if a given physical address can be included as part of a crash dump.
1598  */
1599 bool
1600 vm_phys_is_dumpable(vm_paddr_t pa)
1601 {
1602 	vm_page_t m;
1603 	int i;
1604 
1605 	if ((m = vm_phys_paddr_to_vm_page(pa)) != NULL)
1606 		return ((m->flags & PG_NODUMP) == 0);
1607 
1608 	for (i = 0; dump_avail[i] != 0 || dump_avail[i + 1] != 0; i += 2) {
1609 		if (pa >= dump_avail[i] && pa < dump_avail[i + 1])
1610 			return (true);
1611 	}
1612 	return (false);
1613 }
1614 
1615 void
1616 vm_phys_early_add_seg(vm_paddr_t start, vm_paddr_t end)
1617 {
1618 	struct vm_phys_seg *seg;
1619 
1620 	if (vm_phys_early_nsegs == -1)
1621 		panic("%s: called after initialization", __func__);
1622 	if (vm_phys_early_nsegs == nitems(vm_phys_early_segs))
1623 		panic("%s: ran out of early segments", __func__);
1624 
1625 	seg = &vm_phys_early_segs[vm_phys_early_nsegs++];
1626 	seg->start = start;
1627 	seg->end = end;
1628 }
1629 
1630 /*
1631  * This routine allocates NUMA node specific memory before the page
1632  * allocator is bootstrapped.
1633  */
1634 vm_paddr_t
1635 vm_phys_early_alloc(int domain, size_t alloc_size)
1636 {
1637 	int i, mem_index, biggestone;
1638 	vm_paddr_t pa, mem_start, mem_end, size, biggestsize, align;
1639 
1640 	KASSERT(domain == -1 || (domain >= 0 && domain < vm_ndomains),
1641 	    ("%s: invalid domain index %d", __func__, domain));
1642 
1643 	/*
1644 	 * Search the mem_affinity array for the biggest address
1645 	 * range in the desired domain.  This is used to constrain
1646 	 * the phys_avail selection below.
1647 	 */
1648 	biggestsize = 0;
1649 	mem_start = 0;
1650 	mem_end = -1;
1651 #ifdef NUMA
1652 	mem_index = 0;
1653 	if (mem_affinity != NULL) {
1654 		for (i = 0;; i++) {
1655 			size = mem_affinity[i].end - mem_affinity[i].start;
1656 			if (size == 0)
1657 				break;
1658 			if (domain != -1 && mem_affinity[i].domain != domain)
1659 				continue;
1660 			if (size > biggestsize) {
1661 				mem_index = i;
1662 				biggestsize = size;
1663 			}
1664 		}
1665 		mem_start = mem_affinity[mem_index].start;
1666 		mem_end = mem_affinity[mem_index].end;
1667 	}
1668 #endif
1669 
1670 	/*
1671 	 * Now find biggest physical segment in within the desired
1672 	 * numa domain.
1673 	 */
1674 	biggestsize = 0;
1675 	biggestone = 0;
1676 	for (i = 0; phys_avail[i + 1] != 0; i += 2) {
1677 		/* skip regions that are out of range */
1678 		if (phys_avail[i+1] - alloc_size < mem_start ||
1679 		    phys_avail[i+1] > mem_end)
1680 			continue;
1681 		size = vm_phys_avail_size(i);
1682 		if (size > biggestsize) {
1683 			biggestone = i;
1684 			biggestsize = size;
1685 		}
1686 	}
1687 	alloc_size = round_page(alloc_size);
1688 
1689 	/*
1690 	 * Grab single pages from the front to reduce fragmentation.
1691 	 */
1692 	if (alloc_size == PAGE_SIZE) {
1693 		pa = phys_avail[biggestone];
1694 		phys_avail[biggestone] += PAGE_SIZE;
1695 		vm_phys_avail_check(biggestone);
1696 		return (pa);
1697 	}
1698 
1699 	/*
1700 	 * Naturally align large allocations.
1701 	 */
1702 	align = phys_avail[biggestone + 1] & (alloc_size - 1);
1703 	if (alloc_size + align > biggestsize)
1704 		panic("cannot find a large enough size\n");
1705 	if (align != 0 &&
1706 	    vm_phys_avail_split(phys_avail[biggestone + 1] - align,
1707 	    biggestone) != 0)
1708 		/* Wasting memory. */
1709 		phys_avail[biggestone + 1] -= align;
1710 
1711 	phys_avail[biggestone + 1] -= alloc_size;
1712 	vm_phys_avail_check(biggestone);
1713 	pa = phys_avail[biggestone + 1];
1714 	return (pa);
1715 }
1716 
1717 void
1718 vm_phys_early_startup(void)
1719 {
1720 	struct vm_phys_seg *seg;
1721 	int i;
1722 
1723 	for (i = 0; phys_avail[i + 1] != 0; i += 2) {
1724 		phys_avail[i] = round_page(phys_avail[i]);
1725 		phys_avail[i + 1] = trunc_page(phys_avail[i + 1]);
1726 	}
1727 
1728 	for (i = 0; i < vm_phys_early_nsegs; i++) {
1729 		seg = &vm_phys_early_segs[i];
1730 		vm_phys_add_seg(seg->start, seg->end);
1731 	}
1732 	vm_phys_early_nsegs = -1;
1733 
1734 #ifdef NUMA
1735 	/* Force phys_avail to be split by domain. */
1736 	if (mem_affinity != NULL) {
1737 		int idx;
1738 
1739 		for (i = 0; mem_affinity[i].end != 0; i++) {
1740 			idx = vm_phys_avail_find(mem_affinity[i].start);
1741 			if (idx != -1 &&
1742 			    phys_avail[idx] != mem_affinity[i].start)
1743 				vm_phys_avail_split(mem_affinity[i].start, idx);
1744 			idx = vm_phys_avail_find(mem_affinity[i].end);
1745 			if (idx != -1 &&
1746 			    phys_avail[idx] != mem_affinity[i].end)
1747 				vm_phys_avail_split(mem_affinity[i].end, idx);
1748 		}
1749 	}
1750 #endif
1751 }
1752 
1753 #ifdef DDB
1754 /*
1755  * Show the number of physical pages in each of the free lists.
1756  */
1757 DB_SHOW_COMMAND(freepages, db_show_freepages)
1758 {
1759 	struct vm_freelist *fl;
1760 	int flind, oind, pind, dom;
1761 
1762 	for (dom = 0; dom < vm_ndomains; dom++) {
1763 		db_printf("DOMAIN: %d\n", dom);
1764 		for (flind = 0; flind < vm_nfreelists; flind++) {
1765 			db_printf("FREE LIST %d:\n"
1766 			    "\n  ORDER (SIZE)  |  NUMBER"
1767 			    "\n              ", flind);
1768 			for (pind = 0; pind < VM_NFREEPOOL; pind++)
1769 				db_printf("  |  POOL %d", pind);
1770 			db_printf("\n--            ");
1771 			for (pind = 0; pind < VM_NFREEPOOL; pind++)
1772 				db_printf("-- --      ");
1773 			db_printf("--\n");
1774 			for (oind = VM_NFREEORDER - 1; oind >= 0; oind--) {
1775 				db_printf("  %2.2d (%6.6dK)", oind,
1776 				    1 << (PAGE_SHIFT - 10 + oind));
1777 				for (pind = 0; pind < VM_NFREEPOOL; pind++) {
1778 				fl = vm_phys_free_queues[dom][flind][pind];
1779 					db_printf("  |  %6.6d", fl[oind].lcnt);
1780 				}
1781 				db_printf("\n");
1782 			}
1783 			db_printf("\n");
1784 		}
1785 		db_printf("\n");
1786 	}
1787 }
1788 #endif
1789