xref: /freebsd/sys/vm/vm_phys.c (revision 9c999a259f00b35f0467acd351fea9157ed7e1e4)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3  *
4  * Copyright (c) 2002-2006 Rice University
5  * Copyright (c) 2007 Alan L. Cox <alc@cs.rice.edu>
6  * All rights reserved.
7  *
8  * This software was developed for the FreeBSD Project by Alan L. Cox,
9  * Olivier Crameri, Peter Druschel, Sitaram Iyer, and Juan Navarro.
10  *
11  * Redistribution and use in source and binary forms, with or without
12  * modification, are permitted provided that the following conditions
13  * are met:
14  * 1. Redistributions of source code must retain the above copyright
15  *    notice, this list of conditions and the following disclaimer.
16  * 2. Redistributions in binary form must reproduce the above copyright
17  *    notice, this list of conditions and the following disclaimer in the
18  *    documentation and/or other materials provided with the distribution.
19  *
20  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
21  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
22  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
23  * A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE COPYRIGHT
24  * HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
25  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
26  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS
27  * OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
28  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
29  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY
30  * WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
31  * POSSIBILITY OF SUCH DAMAGE.
32  */
33 
34 /*
35  *	Physical memory system implementation
36  *
37  * Any external functions defined by this module are only to be used by the
38  * virtual memory system.
39  */
40 
41 #include <sys/cdefs.h>
42 __FBSDID("$FreeBSD$");
43 
44 #include "opt_ddb.h"
45 #include "opt_vm.h"
46 
47 #include <sys/param.h>
48 #include <sys/systm.h>
49 #include <sys/domainset.h>
50 #include <sys/lock.h>
51 #include <sys/kernel.h>
52 #include <sys/malloc.h>
53 #include <sys/mutex.h>
54 #include <sys/proc.h>
55 #include <sys/queue.h>
56 #include <sys/rwlock.h>
57 #include <sys/sbuf.h>
58 #include <sys/sysctl.h>
59 #include <sys/tree.h>
60 #include <sys/vmmeter.h>
61 
62 #include <ddb/ddb.h>
63 
64 #include <vm/vm.h>
65 #include <vm/vm_param.h>
66 #include <vm/vm_kern.h>
67 #include <vm/vm_object.h>
68 #include <vm/vm_page.h>
69 #include <vm/vm_phys.h>
70 #include <vm/vm_pagequeue.h>
71 
72 _Static_assert(sizeof(long) * NBBY >= VM_PHYSSEG_MAX,
73     "Too many physsegs.");
74 
75 #ifdef NUMA
76 struct mem_affinity __read_mostly *mem_affinity;
77 int __read_mostly *mem_locality;
78 #endif
79 
80 int __read_mostly vm_ndomains = 1;
81 domainset_t __read_mostly all_domains = DOMAINSET_T_INITIALIZER(0x1);
82 
83 struct vm_phys_seg __read_mostly vm_phys_segs[VM_PHYSSEG_MAX];
84 int __read_mostly vm_phys_nsegs;
85 static struct vm_phys_seg vm_phys_early_segs[8];
86 static int vm_phys_early_nsegs;
87 
88 struct vm_phys_fictitious_seg;
89 static int vm_phys_fictitious_cmp(struct vm_phys_fictitious_seg *,
90     struct vm_phys_fictitious_seg *);
91 
92 RB_HEAD(fict_tree, vm_phys_fictitious_seg) vm_phys_fictitious_tree =
93     RB_INITIALIZER(&vm_phys_fictitious_tree);
94 
95 struct vm_phys_fictitious_seg {
96 	RB_ENTRY(vm_phys_fictitious_seg) node;
97 	/* Memory region data */
98 	vm_paddr_t	start;
99 	vm_paddr_t	end;
100 	vm_page_t	first_page;
101 };
102 
103 RB_GENERATE_STATIC(fict_tree, vm_phys_fictitious_seg, node,
104     vm_phys_fictitious_cmp);
105 
106 static struct rwlock_padalign vm_phys_fictitious_reg_lock;
107 MALLOC_DEFINE(M_FICT_PAGES, "vm_fictitious", "Fictitious VM pages");
108 
109 static struct vm_freelist __aligned(CACHE_LINE_SIZE)
110     vm_phys_free_queues[MAXMEMDOM][VM_NFREELIST][VM_NFREEPOOL]
111     [VM_NFREEORDER_MAX];
112 
113 static int __read_mostly vm_nfreelists;
114 
115 /*
116  * These "avail lists" are globals used to communicate boot-time physical
117  * memory layout to other parts of the kernel.  Each physically contiguous
118  * region of memory is defined by a start address at an even index and an
119  * end address at the following odd index.  Each list is terminated by a
120  * pair of zero entries.
121  *
122  * dump_avail tells the dump code what regions to include in a crash dump, and
123  * phys_avail is all of the remaining physical memory that is available for
124  * the vm system.
125  *
126  * Initially dump_avail and phys_avail are identical.  Boot time memory
127  * allocations remove extents from phys_avail that may still be included
128  * in dumps.
129  */
130 vm_paddr_t phys_avail[PHYS_AVAIL_COUNT];
131 vm_paddr_t dump_avail[PHYS_AVAIL_COUNT];
132 
133 /*
134  * Provides the mapping from VM_FREELIST_* to free list indices (flind).
135  */
136 static int __read_mostly vm_freelist_to_flind[VM_NFREELIST];
137 
138 CTASSERT(VM_FREELIST_DEFAULT == 0);
139 
140 #ifdef VM_FREELIST_DMA32
141 #define	VM_DMA32_BOUNDARY	((vm_paddr_t)1 << 32)
142 #endif
143 
144 /*
145  * Enforce the assumptions made by vm_phys_add_seg() and vm_phys_init() about
146  * the ordering of the free list boundaries.
147  */
148 #if defined(VM_LOWMEM_BOUNDARY) && defined(VM_DMA32_BOUNDARY)
149 CTASSERT(VM_LOWMEM_BOUNDARY < VM_DMA32_BOUNDARY);
150 #endif
151 
152 static int sysctl_vm_phys_free(SYSCTL_HANDLER_ARGS);
153 SYSCTL_OID(_vm, OID_AUTO, phys_free,
154     CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, 0,
155     sysctl_vm_phys_free, "A",
156     "Phys Free Info");
157 
158 static int sysctl_vm_phys_segs(SYSCTL_HANDLER_ARGS);
159 SYSCTL_OID(_vm, OID_AUTO, phys_segs,
160     CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, 0,
161     sysctl_vm_phys_segs, "A",
162     "Phys Seg Info");
163 
164 #ifdef NUMA
165 static int sysctl_vm_phys_locality(SYSCTL_HANDLER_ARGS);
166 SYSCTL_OID(_vm, OID_AUTO, phys_locality,
167     CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, 0,
168     sysctl_vm_phys_locality, "A",
169     "Phys Locality Info");
170 #endif
171 
172 SYSCTL_INT(_vm, OID_AUTO, ndomains, CTLFLAG_RD,
173     &vm_ndomains, 0, "Number of physical memory domains available.");
174 
175 static vm_page_t vm_phys_alloc_seg_contig(struct vm_phys_seg *seg,
176     u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment,
177     vm_paddr_t boundary);
178 static void _vm_phys_create_seg(vm_paddr_t start, vm_paddr_t end, int domain);
179 static void vm_phys_create_seg(vm_paddr_t start, vm_paddr_t end);
180 static void vm_phys_split_pages(vm_page_t m, int oind, struct vm_freelist *fl,
181     int order, int tail);
182 
183 /*
184  * Red-black tree helpers for vm fictitious range management.
185  */
186 static inline int
187 vm_phys_fictitious_in_range(struct vm_phys_fictitious_seg *p,
188     struct vm_phys_fictitious_seg *range)
189 {
190 
191 	KASSERT(range->start != 0 && range->end != 0,
192 	    ("Invalid range passed on search for vm_fictitious page"));
193 	if (p->start >= range->end)
194 		return (1);
195 	if (p->start < range->start)
196 		return (-1);
197 
198 	return (0);
199 }
200 
201 static int
202 vm_phys_fictitious_cmp(struct vm_phys_fictitious_seg *p1,
203     struct vm_phys_fictitious_seg *p2)
204 {
205 
206 	/* Check if this is a search for a page */
207 	if (p1->end == 0)
208 		return (vm_phys_fictitious_in_range(p1, p2));
209 
210 	KASSERT(p2->end != 0,
211     ("Invalid range passed as second parameter to vm fictitious comparison"));
212 
213 	/* Searching to add a new range */
214 	if (p1->end <= p2->start)
215 		return (-1);
216 	if (p1->start >= p2->end)
217 		return (1);
218 
219 	panic("Trying to add overlapping vm fictitious ranges:\n"
220 	    "[%#jx:%#jx] and [%#jx:%#jx]", (uintmax_t)p1->start,
221 	    (uintmax_t)p1->end, (uintmax_t)p2->start, (uintmax_t)p2->end);
222 }
223 
224 int
225 vm_phys_domain_match(int prefer, vm_paddr_t low, vm_paddr_t high)
226 {
227 #ifdef NUMA
228 	domainset_t mask;
229 	int i;
230 
231 	if (vm_ndomains == 1 || mem_affinity == NULL)
232 		return (0);
233 
234 	DOMAINSET_ZERO(&mask);
235 	/*
236 	 * Check for any memory that overlaps low, high.
237 	 */
238 	for (i = 0; mem_affinity[i].end != 0; i++)
239 		if (mem_affinity[i].start <= high &&
240 		    mem_affinity[i].end >= low)
241 			DOMAINSET_SET(mem_affinity[i].domain, &mask);
242 	if (prefer != -1 && DOMAINSET_ISSET(prefer, &mask))
243 		return (prefer);
244 	if (DOMAINSET_EMPTY(&mask))
245 		panic("vm_phys_domain_match:  Impossible constraint");
246 	return (DOMAINSET_FFS(&mask) - 1);
247 #else
248 	return (0);
249 #endif
250 }
251 
252 /*
253  * Outputs the state of the physical memory allocator, specifically,
254  * the amount of physical memory in each free list.
255  */
256 static int
257 sysctl_vm_phys_free(SYSCTL_HANDLER_ARGS)
258 {
259 	struct sbuf sbuf;
260 	struct vm_freelist *fl;
261 	int dom, error, flind, oind, pind;
262 
263 	error = sysctl_wire_old_buffer(req, 0);
264 	if (error != 0)
265 		return (error);
266 	sbuf_new_for_sysctl(&sbuf, NULL, 128 * vm_ndomains, req);
267 	for (dom = 0; dom < vm_ndomains; dom++) {
268 		sbuf_printf(&sbuf,"\nDOMAIN %d:\n", dom);
269 		for (flind = 0; flind < vm_nfreelists; flind++) {
270 			sbuf_printf(&sbuf, "\nFREE LIST %d:\n"
271 			    "\n  ORDER (SIZE)  |  NUMBER"
272 			    "\n              ", flind);
273 			for (pind = 0; pind < VM_NFREEPOOL; pind++)
274 				sbuf_printf(&sbuf, "  |  POOL %d", pind);
275 			sbuf_printf(&sbuf, "\n--            ");
276 			for (pind = 0; pind < VM_NFREEPOOL; pind++)
277 				sbuf_printf(&sbuf, "-- --      ");
278 			sbuf_printf(&sbuf, "--\n");
279 			for (oind = VM_NFREEORDER - 1; oind >= 0; oind--) {
280 				sbuf_printf(&sbuf, "  %2d (%6dK)", oind,
281 				    1 << (PAGE_SHIFT - 10 + oind));
282 				for (pind = 0; pind < VM_NFREEPOOL; pind++) {
283 				fl = vm_phys_free_queues[dom][flind][pind];
284 					sbuf_printf(&sbuf, "  |  %6d",
285 					    fl[oind].lcnt);
286 				}
287 				sbuf_printf(&sbuf, "\n");
288 			}
289 		}
290 	}
291 	error = sbuf_finish(&sbuf);
292 	sbuf_delete(&sbuf);
293 	return (error);
294 }
295 
296 /*
297  * Outputs the set of physical memory segments.
298  */
299 static int
300 sysctl_vm_phys_segs(SYSCTL_HANDLER_ARGS)
301 {
302 	struct sbuf sbuf;
303 	struct vm_phys_seg *seg;
304 	int error, segind;
305 
306 	error = sysctl_wire_old_buffer(req, 0);
307 	if (error != 0)
308 		return (error);
309 	sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
310 	for (segind = 0; segind < vm_phys_nsegs; segind++) {
311 		sbuf_printf(&sbuf, "\nSEGMENT %d:\n\n", segind);
312 		seg = &vm_phys_segs[segind];
313 		sbuf_printf(&sbuf, "start:     %#jx\n",
314 		    (uintmax_t)seg->start);
315 		sbuf_printf(&sbuf, "end:       %#jx\n",
316 		    (uintmax_t)seg->end);
317 		sbuf_printf(&sbuf, "domain:    %d\n", seg->domain);
318 		sbuf_printf(&sbuf, "free list: %p\n", seg->free_queues);
319 	}
320 	error = sbuf_finish(&sbuf);
321 	sbuf_delete(&sbuf);
322 	return (error);
323 }
324 
325 /*
326  * Return affinity, or -1 if there's no affinity information.
327  */
328 int
329 vm_phys_mem_affinity(int f, int t)
330 {
331 
332 #ifdef NUMA
333 	if (mem_locality == NULL)
334 		return (-1);
335 	if (f >= vm_ndomains || t >= vm_ndomains)
336 		return (-1);
337 	return (mem_locality[f * vm_ndomains + t]);
338 #else
339 	return (-1);
340 #endif
341 }
342 
343 #ifdef NUMA
344 /*
345  * Outputs the VM locality table.
346  */
347 static int
348 sysctl_vm_phys_locality(SYSCTL_HANDLER_ARGS)
349 {
350 	struct sbuf sbuf;
351 	int error, i, j;
352 
353 	error = sysctl_wire_old_buffer(req, 0);
354 	if (error != 0)
355 		return (error);
356 	sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
357 
358 	sbuf_printf(&sbuf, "\n");
359 
360 	for (i = 0; i < vm_ndomains; i++) {
361 		sbuf_printf(&sbuf, "%d: ", i);
362 		for (j = 0; j < vm_ndomains; j++) {
363 			sbuf_printf(&sbuf, "%d ", vm_phys_mem_affinity(i, j));
364 		}
365 		sbuf_printf(&sbuf, "\n");
366 	}
367 	error = sbuf_finish(&sbuf);
368 	sbuf_delete(&sbuf);
369 	return (error);
370 }
371 #endif
372 
373 static void
374 vm_freelist_add(struct vm_freelist *fl, vm_page_t m, int order, int tail)
375 {
376 
377 	m->order = order;
378 	if (tail)
379 		TAILQ_INSERT_TAIL(&fl[order].pl, m, listq);
380 	else
381 		TAILQ_INSERT_HEAD(&fl[order].pl, m, listq);
382 	fl[order].lcnt++;
383 }
384 
385 static void
386 vm_freelist_rem(struct vm_freelist *fl, vm_page_t m, int order)
387 {
388 
389 	TAILQ_REMOVE(&fl[order].pl, m, listq);
390 	fl[order].lcnt--;
391 	m->order = VM_NFREEORDER;
392 }
393 
394 /*
395  * Create a physical memory segment.
396  */
397 static void
398 _vm_phys_create_seg(vm_paddr_t start, vm_paddr_t end, int domain)
399 {
400 	struct vm_phys_seg *seg;
401 
402 	KASSERT(vm_phys_nsegs < VM_PHYSSEG_MAX,
403 	    ("vm_phys_create_seg: increase VM_PHYSSEG_MAX"));
404 	KASSERT(domain >= 0 && domain < vm_ndomains,
405 	    ("vm_phys_create_seg: invalid domain provided"));
406 	seg = &vm_phys_segs[vm_phys_nsegs++];
407 	while (seg > vm_phys_segs && (seg - 1)->start >= end) {
408 		*seg = *(seg - 1);
409 		seg--;
410 	}
411 	seg->start = start;
412 	seg->end = end;
413 	seg->domain = domain;
414 }
415 
416 static void
417 vm_phys_create_seg(vm_paddr_t start, vm_paddr_t end)
418 {
419 #ifdef NUMA
420 	int i;
421 
422 	if (mem_affinity == NULL) {
423 		_vm_phys_create_seg(start, end, 0);
424 		return;
425 	}
426 
427 	for (i = 0;; i++) {
428 		if (mem_affinity[i].end == 0)
429 			panic("Reached end of affinity info");
430 		if (mem_affinity[i].end <= start)
431 			continue;
432 		if (mem_affinity[i].start > start)
433 			panic("No affinity info for start %jx",
434 			    (uintmax_t)start);
435 		if (mem_affinity[i].end >= end) {
436 			_vm_phys_create_seg(start, end,
437 			    mem_affinity[i].domain);
438 			break;
439 		}
440 		_vm_phys_create_seg(start, mem_affinity[i].end,
441 		    mem_affinity[i].domain);
442 		start = mem_affinity[i].end;
443 	}
444 #else
445 	_vm_phys_create_seg(start, end, 0);
446 #endif
447 }
448 
449 /*
450  * Add a physical memory segment.
451  */
452 void
453 vm_phys_add_seg(vm_paddr_t start, vm_paddr_t end)
454 {
455 	vm_paddr_t paddr;
456 
457 	KASSERT((start & PAGE_MASK) == 0,
458 	    ("vm_phys_define_seg: start is not page aligned"));
459 	KASSERT((end & PAGE_MASK) == 0,
460 	    ("vm_phys_define_seg: end is not page aligned"));
461 
462 	/*
463 	 * Split the physical memory segment if it spans two or more free
464 	 * list boundaries.
465 	 */
466 	paddr = start;
467 #ifdef	VM_FREELIST_LOWMEM
468 	if (paddr < VM_LOWMEM_BOUNDARY && end > VM_LOWMEM_BOUNDARY) {
469 		vm_phys_create_seg(paddr, VM_LOWMEM_BOUNDARY);
470 		paddr = VM_LOWMEM_BOUNDARY;
471 	}
472 #endif
473 #ifdef	VM_FREELIST_DMA32
474 	if (paddr < VM_DMA32_BOUNDARY && end > VM_DMA32_BOUNDARY) {
475 		vm_phys_create_seg(paddr, VM_DMA32_BOUNDARY);
476 		paddr = VM_DMA32_BOUNDARY;
477 	}
478 #endif
479 	vm_phys_create_seg(paddr, end);
480 }
481 
482 /*
483  * Initialize the physical memory allocator.
484  *
485  * Requires that vm_page_array is initialized!
486  */
487 void
488 vm_phys_init(void)
489 {
490 	struct vm_freelist *fl;
491 	struct vm_phys_seg *end_seg, *prev_seg, *seg, *tmp_seg;
492 	u_long npages;
493 	int dom, flind, freelist, oind, pind, segind;
494 
495 	/*
496 	 * Compute the number of free lists, and generate the mapping from the
497 	 * manifest constants VM_FREELIST_* to the free list indices.
498 	 *
499 	 * Initially, the entries of vm_freelist_to_flind[] are set to either
500 	 * 0 or 1 to indicate which free lists should be created.
501 	 */
502 	npages = 0;
503 	for (segind = vm_phys_nsegs - 1; segind >= 0; segind--) {
504 		seg = &vm_phys_segs[segind];
505 #ifdef	VM_FREELIST_LOWMEM
506 		if (seg->end <= VM_LOWMEM_BOUNDARY)
507 			vm_freelist_to_flind[VM_FREELIST_LOWMEM] = 1;
508 		else
509 #endif
510 #ifdef	VM_FREELIST_DMA32
511 		if (
512 #ifdef	VM_DMA32_NPAGES_THRESHOLD
513 		    /*
514 		     * Create the DMA32 free list only if the amount of
515 		     * physical memory above physical address 4G exceeds the
516 		     * given threshold.
517 		     */
518 		    npages > VM_DMA32_NPAGES_THRESHOLD &&
519 #endif
520 		    seg->end <= VM_DMA32_BOUNDARY)
521 			vm_freelist_to_flind[VM_FREELIST_DMA32] = 1;
522 		else
523 #endif
524 		{
525 			npages += atop(seg->end - seg->start);
526 			vm_freelist_to_flind[VM_FREELIST_DEFAULT] = 1;
527 		}
528 	}
529 	/* Change each entry into a running total of the free lists. */
530 	for (freelist = 1; freelist < VM_NFREELIST; freelist++) {
531 		vm_freelist_to_flind[freelist] +=
532 		    vm_freelist_to_flind[freelist - 1];
533 	}
534 	vm_nfreelists = vm_freelist_to_flind[VM_NFREELIST - 1];
535 	KASSERT(vm_nfreelists > 0, ("vm_phys_init: no free lists"));
536 	/* Change each entry into a free list index. */
537 	for (freelist = 0; freelist < VM_NFREELIST; freelist++)
538 		vm_freelist_to_flind[freelist]--;
539 
540 	/*
541 	 * Initialize the first_page and free_queues fields of each physical
542 	 * memory segment.
543 	 */
544 #ifdef VM_PHYSSEG_SPARSE
545 	npages = 0;
546 #endif
547 	for (segind = 0; segind < vm_phys_nsegs; segind++) {
548 		seg = &vm_phys_segs[segind];
549 #ifdef VM_PHYSSEG_SPARSE
550 		seg->first_page = &vm_page_array[npages];
551 		npages += atop(seg->end - seg->start);
552 #else
553 		seg->first_page = PHYS_TO_VM_PAGE(seg->start);
554 #endif
555 #ifdef	VM_FREELIST_LOWMEM
556 		if (seg->end <= VM_LOWMEM_BOUNDARY) {
557 			flind = vm_freelist_to_flind[VM_FREELIST_LOWMEM];
558 			KASSERT(flind >= 0,
559 			    ("vm_phys_init: LOWMEM flind < 0"));
560 		} else
561 #endif
562 #ifdef	VM_FREELIST_DMA32
563 		if (seg->end <= VM_DMA32_BOUNDARY) {
564 			flind = vm_freelist_to_flind[VM_FREELIST_DMA32];
565 			KASSERT(flind >= 0,
566 			    ("vm_phys_init: DMA32 flind < 0"));
567 		} else
568 #endif
569 		{
570 			flind = vm_freelist_to_flind[VM_FREELIST_DEFAULT];
571 			KASSERT(flind >= 0,
572 			    ("vm_phys_init: DEFAULT flind < 0"));
573 		}
574 		seg->free_queues = &vm_phys_free_queues[seg->domain][flind];
575 	}
576 
577 	/*
578 	 * Coalesce physical memory segments that are contiguous and share the
579 	 * same per-domain free queues.
580 	 */
581 	prev_seg = vm_phys_segs;
582 	seg = &vm_phys_segs[1];
583 	end_seg = &vm_phys_segs[vm_phys_nsegs];
584 	while (seg < end_seg) {
585 		if (prev_seg->end == seg->start &&
586 		    prev_seg->free_queues == seg->free_queues) {
587 			prev_seg->end = seg->end;
588 			KASSERT(prev_seg->domain == seg->domain,
589 			    ("vm_phys_init: free queues cannot span domains"));
590 			vm_phys_nsegs--;
591 			end_seg--;
592 			for (tmp_seg = seg; tmp_seg < end_seg; tmp_seg++)
593 				*tmp_seg = *(tmp_seg + 1);
594 		} else {
595 			prev_seg = seg;
596 			seg++;
597 		}
598 	}
599 
600 	/*
601 	 * Initialize the free queues.
602 	 */
603 	for (dom = 0; dom < vm_ndomains; dom++) {
604 		for (flind = 0; flind < vm_nfreelists; flind++) {
605 			for (pind = 0; pind < VM_NFREEPOOL; pind++) {
606 				fl = vm_phys_free_queues[dom][flind][pind];
607 				for (oind = 0; oind < VM_NFREEORDER; oind++)
608 					TAILQ_INIT(&fl[oind].pl);
609 			}
610 		}
611 	}
612 
613 	rw_init(&vm_phys_fictitious_reg_lock, "vmfctr");
614 }
615 
616 /*
617  * Register info about the NUMA topology of the system.
618  *
619  * Invoked by platform-dependent code prior to vm_phys_init().
620  */
621 void
622 vm_phys_register_domains(int ndomains, struct mem_affinity *affinity,
623     int *locality)
624 {
625 #ifdef NUMA
626 	int d, i;
627 
628 	/*
629 	 * For now the only override value that we support is 1, which
630 	 * effectively disables NUMA-awareness in the allocators.
631 	 */
632 	d = 0;
633 	TUNABLE_INT_FETCH("vm.numa.disabled", &d);
634 	if (d)
635 		ndomains = 1;
636 
637 	if (ndomains > 1) {
638 		vm_ndomains = ndomains;
639 		mem_affinity = affinity;
640 		mem_locality = locality;
641 	}
642 
643 	for (i = 0; i < vm_ndomains; i++)
644 		DOMAINSET_SET(i, &all_domains);
645 #else
646 	(void)ndomains;
647 	(void)affinity;
648 	(void)locality;
649 #endif
650 }
651 
652 /*
653  * Split a contiguous, power of two-sized set of physical pages.
654  *
655  * When this function is called by a page allocation function, the caller
656  * should request insertion at the head unless the order [order, oind) queues
657  * are known to be empty.  The objective being to reduce the likelihood of
658  * long-term fragmentation by promoting contemporaneous allocation and
659  * (hopefully) deallocation.
660  */
661 static __inline void
662 vm_phys_split_pages(vm_page_t m, int oind, struct vm_freelist *fl, int order,
663     int tail)
664 {
665 	vm_page_t m_buddy;
666 
667 	while (oind > order) {
668 		oind--;
669 		m_buddy = &m[1 << oind];
670 		KASSERT(m_buddy->order == VM_NFREEORDER,
671 		    ("vm_phys_split_pages: page %p has unexpected order %d",
672 		    m_buddy, m_buddy->order));
673 		vm_freelist_add(fl, m_buddy, oind, tail);
674         }
675 }
676 
677 /*
678  * Add the physical pages [m, m + npages) at the end of a power-of-two aligned
679  * and sized set to the specified free list.
680  *
681  * When this function is called by a page allocation function, the caller
682  * should request insertion at the head unless the lower-order queues are
683  * known to be empty.  The objective being to reduce the likelihood of long-
684  * term fragmentation by promoting contemporaneous allocation and (hopefully)
685  * deallocation.
686  *
687  * The physical page m's buddy must not be free.
688  */
689 static void
690 vm_phys_enq_range(vm_page_t m, u_int npages, struct vm_freelist *fl, int tail)
691 {
692 	u_int n;
693 	int order;
694 
695 	KASSERT(npages > 0, ("vm_phys_enq_range: npages is 0"));
696 	KASSERT(((VM_PAGE_TO_PHYS(m) + npages * PAGE_SIZE) &
697 	    ((PAGE_SIZE << (fls(npages) - 1)) - 1)) == 0,
698 	    ("vm_phys_enq_range: page %p and npages %u are misaligned",
699 	    m, npages));
700 	do {
701 		KASSERT(m->order == VM_NFREEORDER,
702 		    ("vm_phys_enq_range: page %p has unexpected order %d",
703 		    m, m->order));
704 		order = ffs(npages) - 1;
705 		KASSERT(order < VM_NFREEORDER,
706 		    ("vm_phys_enq_range: order %d is out of range", order));
707 		vm_freelist_add(fl, m, order, tail);
708 		n = 1 << order;
709 		m += n;
710 		npages -= n;
711 	} while (npages > 0);
712 }
713 
714 /*
715  * Tries to allocate the specified number of pages from the specified pool
716  * within the specified domain.  Returns the actual number of allocated pages
717  * and a pointer to each page through the array ma[].
718  *
719  * The returned pages may not be physically contiguous.  However, in contrast
720  * to performing multiple, back-to-back calls to vm_phys_alloc_pages(..., 0),
721  * calling this function once to allocate the desired number of pages will
722  * avoid wasted time in vm_phys_split_pages().
723  *
724  * The free page queues for the specified domain must be locked.
725  */
726 int
727 vm_phys_alloc_npages(int domain, int pool, int npages, vm_page_t ma[])
728 {
729 	struct vm_freelist *alt, *fl;
730 	vm_page_t m;
731 	int avail, end, flind, freelist, i, need, oind, pind;
732 
733 	KASSERT(domain >= 0 && domain < vm_ndomains,
734 	    ("vm_phys_alloc_npages: domain %d is out of range", domain));
735 	KASSERT(pool < VM_NFREEPOOL,
736 	    ("vm_phys_alloc_npages: pool %d is out of range", pool));
737 	KASSERT(npages <= 1 << (VM_NFREEORDER - 1),
738 	    ("vm_phys_alloc_npages: npages %d is out of range", npages));
739 	vm_domain_free_assert_locked(VM_DOMAIN(domain));
740 	i = 0;
741 	for (freelist = 0; freelist < VM_NFREELIST; freelist++) {
742 		flind = vm_freelist_to_flind[freelist];
743 		if (flind < 0)
744 			continue;
745 		fl = vm_phys_free_queues[domain][flind][pool];
746 		for (oind = 0; oind < VM_NFREEORDER; oind++) {
747 			while ((m = TAILQ_FIRST(&fl[oind].pl)) != NULL) {
748 				vm_freelist_rem(fl, m, oind);
749 				avail = 1 << oind;
750 				need = imin(npages - i, avail);
751 				for (end = i + need; i < end;)
752 					ma[i++] = m++;
753 				if (need < avail) {
754 					/*
755 					 * Return excess pages to fl.  Its
756 					 * order [0, oind) queues are empty.
757 					 */
758 					vm_phys_enq_range(m, avail - need, fl,
759 					    1);
760 					return (npages);
761 				} else if (i == npages)
762 					return (npages);
763 			}
764 		}
765 		for (oind = VM_NFREEORDER - 1; oind >= 0; oind--) {
766 			for (pind = 0; pind < VM_NFREEPOOL; pind++) {
767 				alt = vm_phys_free_queues[domain][flind][pind];
768 				while ((m = TAILQ_FIRST(&alt[oind].pl)) !=
769 				    NULL) {
770 					vm_freelist_rem(alt, m, oind);
771 					vm_phys_set_pool(pool, m, oind);
772 					avail = 1 << oind;
773 					need = imin(npages - i, avail);
774 					for (end = i + need; i < end;)
775 						ma[i++] = m++;
776 					if (need < avail) {
777 						/*
778 						 * Return excess pages to fl.
779 						 * Its order [0, oind) queues
780 						 * are empty.
781 						 */
782 						vm_phys_enq_range(m, avail -
783 						    need, fl, 1);
784 						return (npages);
785 					} else if (i == npages)
786 						return (npages);
787 				}
788 			}
789 		}
790 	}
791 	return (i);
792 }
793 
794 /*
795  * Allocate a contiguous, power of two-sized set of physical pages
796  * from the free lists.
797  *
798  * The free page queues must be locked.
799  */
800 vm_page_t
801 vm_phys_alloc_pages(int domain, int pool, int order)
802 {
803 	vm_page_t m;
804 	int freelist;
805 
806 	for (freelist = 0; freelist < VM_NFREELIST; freelist++) {
807 		m = vm_phys_alloc_freelist_pages(domain, freelist, pool, order);
808 		if (m != NULL)
809 			return (m);
810 	}
811 	return (NULL);
812 }
813 
814 /*
815  * Allocate a contiguous, power of two-sized set of physical pages from the
816  * specified free list.  The free list must be specified using one of the
817  * manifest constants VM_FREELIST_*.
818  *
819  * The free page queues must be locked.
820  */
821 vm_page_t
822 vm_phys_alloc_freelist_pages(int domain, int freelist, int pool, int order)
823 {
824 	struct vm_freelist *alt, *fl;
825 	vm_page_t m;
826 	int oind, pind, flind;
827 
828 	KASSERT(domain >= 0 && domain < vm_ndomains,
829 	    ("vm_phys_alloc_freelist_pages: domain %d is out of range",
830 	    domain));
831 	KASSERT(freelist < VM_NFREELIST,
832 	    ("vm_phys_alloc_freelist_pages: freelist %d is out of range",
833 	    freelist));
834 	KASSERT(pool < VM_NFREEPOOL,
835 	    ("vm_phys_alloc_freelist_pages: pool %d is out of range", pool));
836 	KASSERT(order < VM_NFREEORDER,
837 	    ("vm_phys_alloc_freelist_pages: order %d is out of range", order));
838 
839 	flind = vm_freelist_to_flind[freelist];
840 	/* Check if freelist is present */
841 	if (flind < 0)
842 		return (NULL);
843 
844 	vm_domain_free_assert_locked(VM_DOMAIN(domain));
845 	fl = &vm_phys_free_queues[domain][flind][pool][0];
846 	for (oind = order; oind < VM_NFREEORDER; oind++) {
847 		m = TAILQ_FIRST(&fl[oind].pl);
848 		if (m != NULL) {
849 			vm_freelist_rem(fl, m, oind);
850 			/* The order [order, oind) queues are empty. */
851 			vm_phys_split_pages(m, oind, fl, order, 1);
852 			return (m);
853 		}
854 	}
855 
856 	/*
857 	 * The given pool was empty.  Find the largest
858 	 * contiguous, power-of-two-sized set of pages in any
859 	 * pool.  Transfer these pages to the given pool, and
860 	 * use them to satisfy the allocation.
861 	 */
862 	for (oind = VM_NFREEORDER - 1; oind >= order; oind--) {
863 		for (pind = 0; pind < VM_NFREEPOOL; pind++) {
864 			alt = &vm_phys_free_queues[domain][flind][pind][0];
865 			m = TAILQ_FIRST(&alt[oind].pl);
866 			if (m != NULL) {
867 				vm_freelist_rem(alt, m, oind);
868 				vm_phys_set_pool(pool, m, oind);
869 				/* The order [order, oind) queues are empty. */
870 				vm_phys_split_pages(m, oind, fl, order, 1);
871 				return (m);
872 			}
873 		}
874 	}
875 	return (NULL);
876 }
877 
878 /*
879  * Find the vm_page corresponding to the given physical address.
880  */
881 vm_page_t
882 vm_phys_paddr_to_vm_page(vm_paddr_t pa)
883 {
884 	struct vm_phys_seg *seg;
885 	int segind;
886 
887 	for (segind = 0; segind < vm_phys_nsegs; segind++) {
888 		seg = &vm_phys_segs[segind];
889 		if (pa >= seg->start && pa < seg->end)
890 			return (&seg->first_page[atop(pa - seg->start)]);
891 	}
892 	return (NULL);
893 }
894 
895 vm_page_t
896 vm_phys_fictitious_to_vm_page(vm_paddr_t pa)
897 {
898 	struct vm_phys_fictitious_seg tmp, *seg;
899 	vm_page_t m;
900 
901 	m = NULL;
902 	tmp.start = pa;
903 	tmp.end = 0;
904 
905 	rw_rlock(&vm_phys_fictitious_reg_lock);
906 	seg = RB_FIND(fict_tree, &vm_phys_fictitious_tree, &tmp);
907 	rw_runlock(&vm_phys_fictitious_reg_lock);
908 	if (seg == NULL)
909 		return (NULL);
910 
911 	m = &seg->first_page[atop(pa - seg->start)];
912 	KASSERT((m->flags & PG_FICTITIOUS) != 0, ("%p not fictitious", m));
913 
914 	return (m);
915 }
916 
917 static inline void
918 vm_phys_fictitious_init_range(vm_page_t range, vm_paddr_t start,
919     long page_count, vm_memattr_t memattr)
920 {
921 	long i;
922 
923 	bzero(range, page_count * sizeof(*range));
924 	for (i = 0; i < page_count; i++) {
925 		vm_page_initfake(&range[i], start + PAGE_SIZE * i, memattr);
926 		range[i].oflags &= ~VPO_UNMANAGED;
927 		range[i].busy_lock = VPB_UNBUSIED;
928 	}
929 }
930 
931 int
932 vm_phys_fictitious_reg_range(vm_paddr_t start, vm_paddr_t end,
933     vm_memattr_t memattr)
934 {
935 	struct vm_phys_fictitious_seg *seg;
936 	vm_page_t fp;
937 	long page_count;
938 #ifdef VM_PHYSSEG_DENSE
939 	long pi, pe;
940 	long dpage_count;
941 #endif
942 
943 	KASSERT(start < end,
944 	    ("Start of segment isn't less than end (start: %jx end: %jx)",
945 	    (uintmax_t)start, (uintmax_t)end));
946 
947 	page_count = (end - start) / PAGE_SIZE;
948 
949 #ifdef VM_PHYSSEG_DENSE
950 	pi = atop(start);
951 	pe = atop(end);
952 	if (pi >= first_page && (pi - first_page) < vm_page_array_size) {
953 		fp = &vm_page_array[pi - first_page];
954 		if ((pe - first_page) > vm_page_array_size) {
955 			/*
956 			 * We have a segment that starts inside
957 			 * of vm_page_array, but ends outside of it.
958 			 *
959 			 * Use vm_page_array pages for those that are
960 			 * inside of the vm_page_array range, and
961 			 * allocate the remaining ones.
962 			 */
963 			dpage_count = vm_page_array_size - (pi - first_page);
964 			vm_phys_fictitious_init_range(fp, start, dpage_count,
965 			    memattr);
966 			page_count -= dpage_count;
967 			start += ptoa(dpage_count);
968 			goto alloc;
969 		}
970 		/*
971 		 * We can allocate the full range from vm_page_array,
972 		 * so there's no need to register the range in the tree.
973 		 */
974 		vm_phys_fictitious_init_range(fp, start, page_count, memattr);
975 		return (0);
976 	} else if (pe > first_page && (pe - first_page) < vm_page_array_size) {
977 		/*
978 		 * We have a segment that ends inside of vm_page_array,
979 		 * but starts outside of it.
980 		 */
981 		fp = &vm_page_array[0];
982 		dpage_count = pe - first_page;
983 		vm_phys_fictitious_init_range(fp, ptoa(first_page), dpage_count,
984 		    memattr);
985 		end -= ptoa(dpage_count);
986 		page_count -= dpage_count;
987 		goto alloc;
988 	} else if (pi < first_page && pe > (first_page + vm_page_array_size)) {
989 		/*
990 		 * Trying to register a fictitious range that expands before
991 		 * and after vm_page_array.
992 		 */
993 		return (EINVAL);
994 	} else {
995 alloc:
996 #endif
997 		fp = malloc(page_count * sizeof(struct vm_page), M_FICT_PAGES,
998 		    M_WAITOK);
999 #ifdef VM_PHYSSEG_DENSE
1000 	}
1001 #endif
1002 	vm_phys_fictitious_init_range(fp, start, page_count, memattr);
1003 
1004 	seg = malloc(sizeof(*seg), M_FICT_PAGES, M_WAITOK | M_ZERO);
1005 	seg->start = start;
1006 	seg->end = end;
1007 	seg->first_page = fp;
1008 
1009 	rw_wlock(&vm_phys_fictitious_reg_lock);
1010 	RB_INSERT(fict_tree, &vm_phys_fictitious_tree, seg);
1011 	rw_wunlock(&vm_phys_fictitious_reg_lock);
1012 
1013 	return (0);
1014 }
1015 
1016 void
1017 vm_phys_fictitious_unreg_range(vm_paddr_t start, vm_paddr_t end)
1018 {
1019 	struct vm_phys_fictitious_seg *seg, tmp;
1020 #ifdef VM_PHYSSEG_DENSE
1021 	long pi, pe;
1022 #endif
1023 
1024 	KASSERT(start < end,
1025 	    ("Start of segment isn't less than end (start: %jx end: %jx)",
1026 	    (uintmax_t)start, (uintmax_t)end));
1027 
1028 #ifdef VM_PHYSSEG_DENSE
1029 	pi = atop(start);
1030 	pe = atop(end);
1031 	if (pi >= first_page && (pi - first_page) < vm_page_array_size) {
1032 		if ((pe - first_page) <= vm_page_array_size) {
1033 			/*
1034 			 * This segment was allocated using vm_page_array
1035 			 * only, there's nothing to do since those pages
1036 			 * were never added to the tree.
1037 			 */
1038 			return;
1039 		}
1040 		/*
1041 		 * We have a segment that starts inside
1042 		 * of vm_page_array, but ends outside of it.
1043 		 *
1044 		 * Calculate how many pages were added to the
1045 		 * tree and free them.
1046 		 */
1047 		start = ptoa(first_page + vm_page_array_size);
1048 	} else if (pe > first_page && (pe - first_page) < vm_page_array_size) {
1049 		/*
1050 		 * We have a segment that ends inside of vm_page_array,
1051 		 * but starts outside of it.
1052 		 */
1053 		end = ptoa(first_page);
1054 	} else if (pi < first_page && pe > (first_page + vm_page_array_size)) {
1055 		/* Since it's not possible to register such a range, panic. */
1056 		panic(
1057 		    "Unregistering not registered fictitious range [%#jx:%#jx]",
1058 		    (uintmax_t)start, (uintmax_t)end);
1059 	}
1060 #endif
1061 	tmp.start = start;
1062 	tmp.end = 0;
1063 
1064 	rw_wlock(&vm_phys_fictitious_reg_lock);
1065 	seg = RB_FIND(fict_tree, &vm_phys_fictitious_tree, &tmp);
1066 	if (seg->start != start || seg->end != end) {
1067 		rw_wunlock(&vm_phys_fictitious_reg_lock);
1068 		panic(
1069 		    "Unregistering not registered fictitious range [%#jx:%#jx]",
1070 		    (uintmax_t)start, (uintmax_t)end);
1071 	}
1072 	RB_REMOVE(fict_tree, &vm_phys_fictitious_tree, seg);
1073 	rw_wunlock(&vm_phys_fictitious_reg_lock);
1074 	free(seg->first_page, M_FICT_PAGES);
1075 	free(seg, M_FICT_PAGES);
1076 }
1077 
1078 /*
1079  * Free a contiguous, power of two-sized set of physical pages.
1080  *
1081  * The free page queues must be locked.
1082  */
1083 void
1084 vm_phys_free_pages(vm_page_t m, int order)
1085 {
1086 	struct vm_freelist *fl;
1087 	struct vm_phys_seg *seg;
1088 	vm_paddr_t pa;
1089 	vm_page_t m_buddy;
1090 
1091 	KASSERT(m->order == VM_NFREEORDER,
1092 	    ("vm_phys_free_pages: page %p has unexpected order %d",
1093 	    m, m->order));
1094 	KASSERT(m->pool < VM_NFREEPOOL,
1095 	    ("vm_phys_free_pages: page %p has unexpected pool %d",
1096 	    m, m->pool));
1097 	KASSERT(order < VM_NFREEORDER,
1098 	    ("vm_phys_free_pages: order %d is out of range", order));
1099 	seg = &vm_phys_segs[m->segind];
1100 	vm_domain_free_assert_locked(VM_DOMAIN(seg->domain));
1101 	if (order < VM_NFREEORDER - 1) {
1102 		pa = VM_PAGE_TO_PHYS(m);
1103 		do {
1104 			pa ^= ((vm_paddr_t)1 << (PAGE_SHIFT + order));
1105 			if (pa < seg->start || pa >= seg->end)
1106 				break;
1107 			m_buddy = &seg->first_page[atop(pa - seg->start)];
1108 			if (m_buddy->order != order)
1109 				break;
1110 			fl = (*seg->free_queues)[m_buddy->pool];
1111 			vm_freelist_rem(fl, m_buddy, order);
1112 			if (m_buddy->pool != m->pool)
1113 				vm_phys_set_pool(m->pool, m_buddy, order);
1114 			order++;
1115 			pa &= ~(((vm_paddr_t)1 << (PAGE_SHIFT + order)) - 1);
1116 			m = &seg->first_page[atop(pa - seg->start)];
1117 		} while (order < VM_NFREEORDER - 1);
1118 	}
1119 	fl = (*seg->free_queues)[m->pool];
1120 	vm_freelist_add(fl, m, order, 1);
1121 }
1122 
1123 /*
1124  * Return the largest possible order of a set of pages starting at m.
1125  */
1126 static int
1127 max_order(vm_page_t m)
1128 {
1129 
1130 	/*
1131 	 * Unsigned "min" is used here so that "order" is assigned
1132 	 * "VM_NFREEORDER - 1" when "m"'s physical address is zero
1133 	 * or the low-order bits of its physical address are zero
1134 	 * because the size of a physical address exceeds the size of
1135 	 * a long.
1136 	 */
1137 	return (min(ffsl(VM_PAGE_TO_PHYS(m) >> PAGE_SHIFT) - 1,
1138 	    VM_NFREEORDER - 1));
1139 }
1140 
1141 /*
1142  * Free a contiguous, arbitrarily sized set of physical pages, without
1143  * merging across set boundaries.
1144  *
1145  * The free page queues must be locked.
1146  */
1147 void
1148 vm_phys_enqueue_contig(vm_page_t m, u_long npages)
1149 {
1150 	struct vm_freelist *fl;
1151 	struct vm_phys_seg *seg;
1152 	vm_page_t m_end;
1153 	int order;
1154 
1155 	/*
1156 	 * Avoid unnecessary coalescing by freeing the pages in the largest
1157 	 * possible power-of-two-sized subsets.
1158 	 */
1159 	vm_domain_free_assert_locked(vm_pagequeue_domain(m));
1160 	seg = &vm_phys_segs[m->segind];
1161 	fl = (*seg->free_queues)[m->pool];
1162 	m_end = m + npages;
1163 	/* Free blocks of increasing size. */
1164 	while ((order = max_order(m)) < VM_NFREEORDER - 1 &&
1165 	    m + (1 << order) <= m_end) {
1166 		KASSERT(seg == &vm_phys_segs[m->segind],
1167 		    ("%s: page range [%p,%p) spans multiple segments",
1168 		    __func__, m_end - npages, m));
1169 		vm_freelist_add(fl, m, order, 1);
1170 		m += 1 << order;
1171 	}
1172 	/* Free blocks of maximum size. */
1173 	while (m + (1 << order) <= m_end) {
1174 		KASSERT(seg == &vm_phys_segs[m->segind],
1175 		    ("%s: page range [%p,%p) spans multiple segments",
1176 		    __func__, m_end - npages, m));
1177 		vm_freelist_add(fl, m, order, 1);
1178 		m += 1 << order;
1179 	}
1180 	/* Free blocks of diminishing size. */
1181 	while (m < m_end) {
1182 		KASSERT(seg == &vm_phys_segs[m->segind],
1183 		    ("%s: page range [%p,%p) spans multiple segments",
1184 		    __func__, m_end - npages, m));
1185 		order = flsl(m_end - m) - 1;
1186 		vm_freelist_add(fl, m, order, 1);
1187 		m += 1 << order;
1188 	}
1189 }
1190 
1191 /*
1192  * Free a contiguous, arbitrarily sized set of physical pages.
1193  *
1194  * The free page queues must be locked.
1195  */
1196 void
1197 vm_phys_free_contig(vm_page_t m, u_long npages)
1198 {
1199 	int order_start, order_end;
1200 	vm_page_t m_start, m_end;
1201 
1202 	vm_domain_free_assert_locked(vm_pagequeue_domain(m));
1203 
1204 	m_start = m;
1205 	order_start = max_order(m_start);
1206 	if (order_start < VM_NFREEORDER - 1)
1207 		m_start += 1 << order_start;
1208 	m_end = m + npages;
1209 	order_end = max_order(m_end);
1210 	if (order_end < VM_NFREEORDER - 1)
1211 		m_end -= 1 << order_end;
1212 	/*
1213 	 * Avoid unnecessary coalescing by freeing the pages at the start and
1214 	 * end of the range last.
1215 	 */
1216 	if (m_start < m_end)
1217 		vm_phys_enqueue_contig(m_start, m_end - m_start);
1218 	if (order_start < VM_NFREEORDER - 1)
1219 		vm_phys_free_pages(m, order_start);
1220 	if (order_end < VM_NFREEORDER - 1)
1221 		vm_phys_free_pages(m_end, order_end);
1222 }
1223 
1224 /*
1225  * Scan physical memory between the specified addresses "low" and "high" for a
1226  * run of contiguous physical pages that satisfy the specified conditions, and
1227  * return the lowest page in the run.  The specified "alignment" determines
1228  * the alignment of the lowest physical page in the run.  If the specified
1229  * "boundary" is non-zero, then the run of physical pages cannot span a
1230  * physical address that is a multiple of "boundary".
1231  *
1232  * "npages" must be greater than zero.  Both "alignment" and "boundary" must
1233  * be a power of two.
1234  */
1235 vm_page_t
1236 vm_phys_scan_contig(int domain, u_long npages, vm_paddr_t low, vm_paddr_t high,
1237     u_long alignment, vm_paddr_t boundary, int options)
1238 {
1239 	vm_paddr_t pa_end;
1240 	vm_page_t m_end, m_run, m_start;
1241 	struct vm_phys_seg *seg;
1242 	int segind;
1243 
1244 	KASSERT(npages > 0, ("npages is 0"));
1245 	KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
1246 	KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
1247 	if (low >= high)
1248 		return (NULL);
1249 	for (segind = 0; segind < vm_phys_nsegs; segind++) {
1250 		seg = &vm_phys_segs[segind];
1251 		if (seg->domain != domain)
1252 			continue;
1253 		if (seg->start >= high)
1254 			break;
1255 		if (low >= seg->end)
1256 			continue;
1257 		if (low <= seg->start)
1258 			m_start = seg->first_page;
1259 		else
1260 			m_start = &seg->first_page[atop(low - seg->start)];
1261 		if (high < seg->end)
1262 			pa_end = high;
1263 		else
1264 			pa_end = seg->end;
1265 		if (pa_end - VM_PAGE_TO_PHYS(m_start) < ptoa(npages))
1266 			continue;
1267 		m_end = &seg->first_page[atop(pa_end - seg->start)];
1268 		m_run = vm_page_scan_contig(npages, m_start, m_end,
1269 		    alignment, boundary, options);
1270 		if (m_run != NULL)
1271 			return (m_run);
1272 	}
1273 	return (NULL);
1274 }
1275 
1276 /*
1277  * Set the pool for a contiguous, power of two-sized set of physical pages.
1278  */
1279 void
1280 vm_phys_set_pool(int pool, vm_page_t m, int order)
1281 {
1282 	vm_page_t m_tmp;
1283 
1284 	for (m_tmp = m; m_tmp < &m[1 << order]; m_tmp++)
1285 		m_tmp->pool = pool;
1286 }
1287 
1288 /*
1289  * Search for the given physical page "m" in the free lists.  If the search
1290  * succeeds, remove "m" from the free lists and return TRUE.  Otherwise, return
1291  * FALSE, indicating that "m" is not in the free lists.
1292  *
1293  * The free page queues must be locked.
1294  */
1295 boolean_t
1296 vm_phys_unfree_page(vm_page_t m)
1297 {
1298 	struct vm_freelist *fl;
1299 	struct vm_phys_seg *seg;
1300 	vm_paddr_t pa, pa_half;
1301 	vm_page_t m_set, m_tmp;
1302 	int order;
1303 
1304 	/*
1305 	 * First, find the contiguous, power of two-sized set of free
1306 	 * physical pages containing the given physical page "m" and
1307 	 * assign it to "m_set".
1308 	 */
1309 	seg = &vm_phys_segs[m->segind];
1310 	vm_domain_free_assert_locked(VM_DOMAIN(seg->domain));
1311 	for (m_set = m, order = 0; m_set->order == VM_NFREEORDER &&
1312 	    order < VM_NFREEORDER - 1; ) {
1313 		order++;
1314 		pa = m->phys_addr & (~(vm_paddr_t)0 << (PAGE_SHIFT + order));
1315 		if (pa >= seg->start)
1316 			m_set = &seg->first_page[atop(pa - seg->start)];
1317 		else
1318 			return (FALSE);
1319 	}
1320 	if (m_set->order < order)
1321 		return (FALSE);
1322 	if (m_set->order == VM_NFREEORDER)
1323 		return (FALSE);
1324 	KASSERT(m_set->order < VM_NFREEORDER,
1325 	    ("vm_phys_unfree_page: page %p has unexpected order %d",
1326 	    m_set, m_set->order));
1327 
1328 	/*
1329 	 * Next, remove "m_set" from the free lists.  Finally, extract
1330 	 * "m" from "m_set" using an iterative algorithm: While "m_set"
1331 	 * is larger than a page, shrink "m_set" by returning the half
1332 	 * of "m_set" that does not contain "m" to the free lists.
1333 	 */
1334 	fl = (*seg->free_queues)[m_set->pool];
1335 	order = m_set->order;
1336 	vm_freelist_rem(fl, m_set, order);
1337 	while (order > 0) {
1338 		order--;
1339 		pa_half = m_set->phys_addr ^ (1 << (PAGE_SHIFT + order));
1340 		if (m->phys_addr < pa_half)
1341 			m_tmp = &seg->first_page[atop(pa_half - seg->start)];
1342 		else {
1343 			m_tmp = m_set;
1344 			m_set = &seg->first_page[atop(pa_half - seg->start)];
1345 		}
1346 		vm_freelist_add(fl, m_tmp, order, 0);
1347 	}
1348 	KASSERT(m_set == m, ("vm_phys_unfree_page: fatal inconsistency"));
1349 	return (TRUE);
1350 }
1351 
1352 /*
1353  * Allocate a contiguous set of physical pages of the given size
1354  * "npages" from the free lists.  All of the physical pages must be at
1355  * or above the given physical address "low" and below the given
1356  * physical address "high".  The given value "alignment" determines the
1357  * alignment of the first physical page in the set.  If the given value
1358  * "boundary" is non-zero, then the set of physical pages cannot cross
1359  * any physical address boundary that is a multiple of that value.  Both
1360  * "alignment" and "boundary" must be a power of two.
1361  */
1362 vm_page_t
1363 vm_phys_alloc_contig(int domain, u_long npages, vm_paddr_t low, vm_paddr_t high,
1364     u_long alignment, vm_paddr_t boundary)
1365 {
1366 	vm_paddr_t pa_end, pa_start;
1367 	vm_page_t m_run;
1368 	struct vm_phys_seg *seg;
1369 	int segind;
1370 
1371 	KASSERT(npages > 0, ("npages is 0"));
1372 	KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
1373 	KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
1374 	vm_domain_free_assert_locked(VM_DOMAIN(domain));
1375 	if (low >= high)
1376 		return (NULL);
1377 	m_run = NULL;
1378 	for (segind = vm_phys_nsegs - 1; segind >= 0; segind--) {
1379 		seg = &vm_phys_segs[segind];
1380 		if (seg->start >= high || seg->domain != domain)
1381 			continue;
1382 		if (low >= seg->end)
1383 			break;
1384 		if (low <= seg->start)
1385 			pa_start = seg->start;
1386 		else
1387 			pa_start = low;
1388 		if (high < seg->end)
1389 			pa_end = high;
1390 		else
1391 			pa_end = seg->end;
1392 		if (pa_end - pa_start < ptoa(npages))
1393 			continue;
1394 		m_run = vm_phys_alloc_seg_contig(seg, npages, low, high,
1395 		    alignment, boundary);
1396 		if (m_run != NULL)
1397 			break;
1398 	}
1399 	return (m_run);
1400 }
1401 
1402 /*
1403  * Allocate a run of contiguous physical pages from the free list for the
1404  * specified segment.
1405  */
1406 static vm_page_t
1407 vm_phys_alloc_seg_contig(struct vm_phys_seg *seg, u_long npages,
1408     vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary)
1409 {
1410 	struct vm_freelist *fl;
1411 	vm_paddr_t pa, pa_end, size;
1412 	vm_page_t m, m_ret;
1413 	u_long npages_end;
1414 	int oind, order, pind;
1415 
1416 	KASSERT(npages > 0, ("npages is 0"));
1417 	KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
1418 	KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
1419 	vm_domain_free_assert_locked(VM_DOMAIN(seg->domain));
1420 	/* Compute the queue that is the best fit for npages. */
1421 	order = flsl(npages - 1);
1422 	/* Search for a run satisfying the specified conditions. */
1423 	size = npages << PAGE_SHIFT;
1424 	for (oind = min(order, VM_NFREEORDER - 1); oind < VM_NFREEORDER;
1425 	    oind++) {
1426 		for (pind = 0; pind < VM_NFREEPOOL; pind++) {
1427 			fl = (*seg->free_queues)[pind];
1428 			TAILQ_FOREACH(m_ret, &fl[oind].pl, listq) {
1429 				/*
1430 				 * Is the size of this allocation request
1431 				 * larger than the largest block size?
1432 				 */
1433 				if (order >= VM_NFREEORDER) {
1434 					/*
1435 					 * Determine if a sufficient number of
1436 					 * subsequent blocks to satisfy the
1437 					 * allocation request are free.
1438 					 */
1439 					pa = VM_PAGE_TO_PHYS(m_ret);
1440 					pa_end = pa + size;
1441 					if (pa_end < pa)
1442 						continue;
1443 					for (;;) {
1444 						pa += 1 << (PAGE_SHIFT +
1445 						    VM_NFREEORDER - 1);
1446 						if (pa >= pa_end ||
1447 						    pa < seg->start ||
1448 						    pa >= seg->end)
1449 							break;
1450 						m = &seg->first_page[atop(pa -
1451 						    seg->start)];
1452 						if (m->order != VM_NFREEORDER -
1453 						    1)
1454 							break;
1455 					}
1456 					/* If not, go to the next block. */
1457 					if (pa < pa_end)
1458 						continue;
1459 				}
1460 
1461 				/*
1462 				 * Determine if the blocks are within the
1463 				 * given range, satisfy the given alignment,
1464 				 * and do not cross the given boundary.
1465 				 */
1466 				pa = VM_PAGE_TO_PHYS(m_ret);
1467 				pa_end = pa + size;
1468 				if (pa >= low && pa_end <= high &&
1469 				    (pa & (alignment - 1)) == 0 &&
1470 				    rounddown2(pa ^ (pa_end - 1), boundary) == 0)
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_index = 0;
1650 	mem_start = 0;
1651 	mem_end = -1;
1652 #ifdef NUMA
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