xref: /freebsd/sys/vm/vm_kern.c (revision c9504239e8e891292b751f9c498dc3b03234004b)
1 /*-
2  * SPDX-License-Identifier: (BSD-3-Clause AND MIT-CMU)
3  *
4  * Copyright (c) 1991, 1993
5  *	The Regents of the University of California.  All rights reserved.
6  *
7  * This code is derived from software contributed to Berkeley by
8  * The Mach Operating System project at Carnegie-Mellon University.
9  *
10  * Redistribution and use in source and binary forms, with or without
11  * modification, are permitted provided that the following conditions
12  * are met:
13  * 1. Redistributions of source code must retain the above copyright
14  *    notice, this list of conditions and the following disclaimer.
15  * 2. Redistributions in binary form must reproduce the above copyright
16  *    notice, this list of conditions and the following disclaimer in the
17  *    documentation and/or other materials provided with the distribution.
18  * 3. Neither the name of the University nor the names of its contributors
19  *    may be used to endorse or promote products derived from this software
20  *    without specific prior written permission.
21  *
22  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
23  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
24  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
25  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
26  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
27  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
28  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
29  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
30  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
31  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32  * SUCH DAMAGE.
33  *
34  *	from: @(#)vm_kern.c	8.3 (Berkeley) 1/12/94
35  *
36  *
37  * Copyright (c) 1987, 1990 Carnegie-Mellon University.
38  * All rights reserved.
39  *
40  * Authors: Avadis Tevanian, Jr., Michael Wayne Young
41  *
42  * Permission to use, copy, modify and distribute this software and
43  * its documentation is hereby granted, provided that both the copyright
44  * notice and this permission notice appear in all copies of the
45  * software, derivative works or modified versions, and any portions
46  * thereof, and that both notices appear in supporting documentation.
47  *
48  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
49  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
50  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
51  *
52  * Carnegie Mellon requests users of this software to return to
53  *
54  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
55  *  School of Computer Science
56  *  Carnegie Mellon University
57  *  Pittsburgh PA 15213-3890
58  *
59  * any improvements or extensions that they make and grant Carnegie the
60  * rights to redistribute these changes.
61  */
62 
63 /*
64  *	Kernel memory management.
65  */
66 
67 #include <sys/cdefs.h>
68 __FBSDID("$FreeBSD$");
69 
70 #include "opt_vm.h"
71 
72 #include <sys/param.h>
73 #include <sys/systm.h>
74 #include <sys/kernel.h>		/* for ticks and hz */
75 #include <sys/domainset.h>
76 #include <sys/eventhandler.h>
77 #include <sys/lock.h>
78 #include <sys/proc.h>
79 #include <sys/malloc.h>
80 #include <sys/rwlock.h>
81 #include <sys/sysctl.h>
82 #include <sys/vmem.h>
83 #include <sys/vmmeter.h>
84 
85 #include <vm/vm.h>
86 #include <vm/vm_param.h>
87 #include <vm/vm_domainset.h>
88 #include <vm/vm_kern.h>
89 #include <vm/pmap.h>
90 #include <vm/vm_map.h>
91 #include <vm/vm_object.h>
92 #include <vm/vm_page.h>
93 #include <vm/vm_pageout.h>
94 #include <vm/vm_phys.h>
95 #include <vm/vm_pagequeue.h>
96 #include <vm/vm_radix.h>
97 #include <vm/vm_extern.h>
98 #include <vm/uma.h>
99 
100 vm_map_t kernel_map;
101 vm_map_t exec_map;
102 vm_map_t pipe_map;
103 
104 const void *zero_region;
105 CTASSERT((ZERO_REGION_SIZE & PAGE_MASK) == 0);
106 
107 /* NB: Used by kernel debuggers. */
108 const u_long vm_maxuser_address = VM_MAXUSER_ADDRESS;
109 
110 u_int exec_map_entry_size;
111 u_int exec_map_entries;
112 
113 SYSCTL_ULONG(_vm, OID_AUTO, min_kernel_address, CTLFLAG_RD,
114     SYSCTL_NULL_ULONG_PTR, VM_MIN_KERNEL_ADDRESS, "Min kernel address");
115 
116 SYSCTL_ULONG(_vm, OID_AUTO, max_kernel_address, CTLFLAG_RD,
117 #if defined(__arm__) || defined(__sparc64__)
118     &vm_max_kernel_address, 0,
119 #else
120     SYSCTL_NULL_ULONG_PTR, VM_MAX_KERNEL_ADDRESS,
121 #endif
122     "Max kernel address");
123 
124 #if VM_NRESERVLEVEL > 0
125 #define	KVA_QUANTUM_SHIFT	(VM_LEVEL_0_ORDER + PAGE_SHIFT)
126 #else
127 /* On non-superpage architectures want large import sizes. */
128 #define	KVA_QUANTUM_SHIFT	(10 + PAGE_SHIFT)
129 #endif
130 #define	KVA_QUANTUM		(1 << KVA_QUANTUM_SHIFT)
131 
132 /*
133  *	kva_alloc:
134  *
135  *	Allocate a virtual address range with no underlying object and
136  *	no initial mapping to physical memory.  Any mapping from this
137  *	range to physical memory must be explicitly created prior to
138  *	its use, typically with pmap_qenter().  Any attempt to create
139  *	a mapping on demand through vm_fault() will result in a panic.
140  */
141 vm_offset_t
142 kva_alloc(vm_size_t size)
143 {
144 	vm_offset_t addr;
145 
146 	size = round_page(size);
147 	if (vmem_alloc(kernel_arena, size, M_BESTFIT | M_NOWAIT, &addr))
148 		return (0);
149 
150 	return (addr);
151 }
152 
153 /*
154  *	kva_free:
155  *
156  *	Release a region of kernel virtual memory allocated
157  *	with kva_alloc, and return the physical pages
158  *	associated with that region.
159  *
160  *	This routine may not block on kernel maps.
161  */
162 void
163 kva_free(vm_offset_t addr, vm_size_t size)
164 {
165 
166 	size = round_page(size);
167 	vmem_free(kernel_arena, addr, size);
168 }
169 
170 /*
171  *	Allocates a region from the kernel address map and physical pages
172  *	within the specified address range to the kernel object.  Creates a
173  *	wired mapping from this region to these pages, and returns the
174  *	region's starting virtual address.  The allocated pages are not
175  *	necessarily physically contiguous.  If M_ZERO is specified through the
176  *	given flags, then the pages are zeroed before they are mapped.
177  */
178 vm_offset_t
179 kmem_alloc_attr_domain(int domain, vm_size_t size, int flags, vm_paddr_t low,
180     vm_paddr_t high, vm_memattr_t memattr)
181 {
182 	vmem_t *vmem;
183 	vm_object_t object = kernel_object;
184 	vm_offset_t addr, i, offset;
185 	vm_page_t m;
186 	int pflags, tries;
187 
188 	size = round_page(size);
189 	vmem = vm_dom[domain].vmd_kernel_arena;
190 	if (vmem_alloc(vmem, size, M_BESTFIT | flags, &addr))
191 		return (0);
192 	offset = addr - VM_MIN_KERNEL_ADDRESS;
193 	pflags = malloc2vm_flags(flags) | VM_ALLOC_NOBUSY | VM_ALLOC_WIRED;
194 	pflags &= ~(VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL);
195 	pflags |= VM_ALLOC_NOWAIT;
196 	VM_OBJECT_WLOCK(object);
197 	for (i = 0; i < size; i += PAGE_SIZE) {
198 		tries = 0;
199 retry:
200 		m = vm_page_alloc_contig_domain(object, atop(offset + i),
201 		    domain, pflags, 1, low, high, PAGE_SIZE, 0, memattr);
202 		if (m == NULL) {
203 			VM_OBJECT_WUNLOCK(object);
204 			if (tries < ((flags & M_NOWAIT) != 0 ? 1 : 3)) {
205 				if (!vm_page_reclaim_contig_domain(domain,
206 				    pflags, 1, low, high, PAGE_SIZE, 0) &&
207 				    (flags & M_WAITOK) != 0)
208 					vm_wait_domain(domain);
209 				VM_OBJECT_WLOCK(object);
210 				tries++;
211 				goto retry;
212 			}
213 			kmem_unback(object, addr, i);
214 			vmem_free(vmem, addr, size);
215 			return (0);
216 		}
217 		KASSERT(vm_phys_domain(m) == domain,
218 		    ("kmem_alloc_attr_domain: Domain mismatch %d != %d",
219 		    vm_phys_domain(m), domain));
220 		if ((flags & M_ZERO) && (m->flags & PG_ZERO) == 0)
221 			pmap_zero_page(m);
222 		m->valid = VM_PAGE_BITS_ALL;
223 		pmap_enter(kernel_pmap, addr + i, m, VM_PROT_RW,
224 		    VM_PROT_RW | PMAP_ENTER_WIRED, 0);
225 	}
226 	VM_OBJECT_WUNLOCK(object);
227 	return (addr);
228 }
229 
230 vm_offset_t
231 kmem_alloc_attr(vm_size_t size, int flags, vm_paddr_t low, vm_paddr_t high,
232     vm_memattr_t memattr)
233 {
234 	struct vm_domainset_iter di;
235 	vm_offset_t addr;
236 	int domain;
237 
238 	vm_domainset_iter_malloc_init(&di, kernel_object, &domain, &flags);
239 	do {
240 		addr = kmem_alloc_attr_domain(domain, size, flags, low, high,
241 		    memattr);
242 		if (addr != 0)
243 			break;
244 	} while (vm_domainset_iter_malloc(&di, &domain, &flags) == 0);
245 
246 	return (addr);
247 }
248 
249 /*
250  *	Allocates a region from the kernel address map and physically
251  *	contiguous pages within the specified address range to the kernel
252  *	object.  Creates a wired mapping from this region to these pages, and
253  *	returns the region's starting virtual address.  If M_ZERO is specified
254  *	through the given flags, then the pages are zeroed before they are
255  *	mapped.
256  */
257 vm_offset_t
258 kmem_alloc_contig_domain(int domain, vm_size_t size, int flags, vm_paddr_t low,
259     vm_paddr_t high, u_long alignment, vm_paddr_t boundary,
260     vm_memattr_t memattr)
261 {
262 	vmem_t *vmem;
263 	vm_object_t object = kernel_object;
264 	vm_offset_t addr, offset, tmp;
265 	vm_page_t end_m, m;
266 	u_long npages;
267 	int pflags, tries;
268 
269 	size = round_page(size);
270 	vmem = vm_dom[domain].vmd_kernel_arena;
271 	if (vmem_alloc(vmem, size, flags | M_BESTFIT, &addr))
272 		return (0);
273 	offset = addr - VM_MIN_KERNEL_ADDRESS;
274 	pflags = malloc2vm_flags(flags) | VM_ALLOC_NOBUSY | VM_ALLOC_WIRED;
275 	pflags &= ~(VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL);
276 	pflags |= VM_ALLOC_NOWAIT;
277 	npages = atop(size);
278 	VM_OBJECT_WLOCK(object);
279 	tries = 0;
280 retry:
281 	m = vm_page_alloc_contig_domain(object, atop(offset), domain, pflags,
282 	    npages, low, high, alignment, boundary, memattr);
283 	if (m == NULL) {
284 		VM_OBJECT_WUNLOCK(object);
285 		if (tries < ((flags & M_NOWAIT) != 0 ? 1 : 3)) {
286 			if (!vm_page_reclaim_contig_domain(domain, pflags,
287 			    npages, low, high, alignment, boundary) &&
288 			    (flags & M_WAITOK) != 0)
289 				vm_wait_domain(domain);
290 			VM_OBJECT_WLOCK(object);
291 			tries++;
292 			goto retry;
293 		}
294 		vmem_free(vmem, addr, size);
295 		return (0);
296 	}
297 	KASSERT(vm_phys_domain(m) == domain,
298 	    ("kmem_alloc_contig_domain: Domain mismatch %d != %d",
299 	    vm_phys_domain(m), domain));
300 	end_m = m + npages;
301 	tmp = addr;
302 	for (; m < end_m; m++) {
303 		if ((flags & M_ZERO) && (m->flags & PG_ZERO) == 0)
304 			pmap_zero_page(m);
305 		m->valid = VM_PAGE_BITS_ALL;
306 		pmap_enter(kernel_pmap, tmp, m, VM_PROT_RW,
307 		    VM_PROT_RW | PMAP_ENTER_WIRED, 0);
308 		tmp += PAGE_SIZE;
309 	}
310 	VM_OBJECT_WUNLOCK(object);
311 	return (addr);
312 }
313 
314 vm_offset_t
315 kmem_alloc_contig(vm_size_t size, int flags, vm_paddr_t low, vm_paddr_t high,
316     u_long alignment, vm_paddr_t boundary, vm_memattr_t memattr)
317 {
318 	struct vm_domainset_iter di;
319 	vm_offset_t addr;
320 	int domain;
321 
322 	vm_domainset_iter_malloc_init(&di, kernel_object, &domain, &flags);
323 	do {
324 		addr = kmem_alloc_contig_domain(domain, size, flags, low, high,
325 		    alignment, boundary, memattr);
326 		if (addr != 0)
327 			break;
328 	} while (vm_domainset_iter_malloc(&di, &domain, &flags) == 0);
329 
330 	return (addr);
331 }
332 
333 /*
334  *	kmem_suballoc:
335  *
336  *	Allocates a map to manage a subrange
337  *	of the kernel virtual address space.
338  *
339  *	Arguments are as follows:
340  *
341  *	parent		Map to take range from
342  *	min, max	Returned endpoints of map
343  *	size		Size of range to find
344  *	superpage_align	Request that min is superpage aligned
345  */
346 vm_map_t
347 kmem_suballoc(vm_map_t parent, vm_offset_t *min, vm_offset_t *max,
348     vm_size_t size, boolean_t superpage_align)
349 {
350 	int ret;
351 	vm_map_t result;
352 
353 	size = round_page(size);
354 
355 	*min = vm_map_min(parent);
356 	ret = vm_map_find(parent, NULL, 0, min, size, 0, superpage_align ?
357 	    VMFS_SUPER_SPACE : VMFS_ANY_SPACE, VM_PROT_ALL, VM_PROT_ALL,
358 	    MAP_ACC_NO_CHARGE);
359 	if (ret != KERN_SUCCESS)
360 		panic("kmem_suballoc: bad status return of %d", ret);
361 	*max = *min + size;
362 	result = vm_map_create(vm_map_pmap(parent), *min, *max);
363 	if (result == NULL)
364 		panic("kmem_suballoc: cannot create submap");
365 	if (vm_map_submap(parent, *min, *max, result) != KERN_SUCCESS)
366 		panic("kmem_suballoc: unable to change range to submap");
367 	return (result);
368 }
369 
370 /*
371  *	kmem_malloc:
372  *
373  *	Allocate wired-down pages in the kernel's address space.
374  */
375 vm_offset_t
376 kmem_malloc_domain(int domain, vm_size_t size, int flags)
377 {
378 	vmem_t *arena;
379 	vm_offset_t addr;
380 	int rv;
381 
382 #if VM_NRESERVLEVEL > 0
383 	if (__predict_true((flags & M_EXEC) == 0))
384 		arena = vm_dom[domain].vmd_kernel_arena;
385 	else
386 		arena = vm_dom[domain].vmd_kernel_rwx_arena;
387 #else
388 	arena = vm_dom[domain].vmd_kernel_arena;
389 #endif
390 	size = round_page(size);
391 	if (vmem_alloc(arena, size, flags | M_BESTFIT, &addr))
392 		return (0);
393 
394 	rv = kmem_back_domain(domain, kernel_object, addr, size, flags);
395 	if (rv != KERN_SUCCESS) {
396 		vmem_free(arena, addr, size);
397 		return (0);
398 	}
399 	return (addr);
400 }
401 
402 vm_offset_t
403 kmem_malloc(vm_size_t size, int flags)
404 {
405 	struct vm_domainset_iter di;
406 	vm_offset_t addr;
407 	int domain;
408 
409 	vm_domainset_iter_malloc_init(&di, kernel_object, &domain, &flags);
410 	do {
411 		addr = kmem_malloc_domain(domain, size, flags);
412 		if (addr != 0)
413 			break;
414 	} while (vm_domainset_iter_malloc(&di, &domain, &flags) == 0);
415 
416 	return (addr);
417 }
418 
419 /*
420  *	kmem_back_domain:
421  *
422  *	Allocate physical pages from the specified domain for the specified
423  *	virtual address range.
424  */
425 int
426 kmem_back_domain(int domain, vm_object_t object, vm_offset_t addr,
427     vm_size_t size, int flags)
428 {
429 	vm_offset_t offset, i;
430 	vm_page_t m, mpred;
431 	vm_prot_t prot;
432 	int pflags;
433 
434 	KASSERT(object == kernel_object,
435 	    ("kmem_back_domain: only supports kernel object."));
436 
437 	offset = addr - VM_MIN_KERNEL_ADDRESS;
438 	pflags = malloc2vm_flags(flags) | VM_ALLOC_NOBUSY | VM_ALLOC_WIRED;
439 	pflags &= ~(VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL);
440 	if (flags & M_WAITOK)
441 		pflags |= VM_ALLOC_WAITFAIL;
442 	prot = (flags & M_EXEC) != 0 ? VM_PROT_ALL : VM_PROT_RW;
443 
444 	i = 0;
445 	VM_OBJECT_WLOCK(object);
446 retry:
447 	mpred = vm_radix_lookup_le(&object->rtree, atop(offset + i));
448 	for (; i < size; i += PAGE_SIZE, mpred = m) {
449 		m = vm_page_alloc_domain_after(object, atop(offset + i),
450 		    domain, pflags, mpred);
451 
452 		/*
453 		 * Ran out of space, free everything up and return. Don't need
454 		 * to lock page queues here as we know that the pages we got
455 		 * aren't on any queues.
456 		 */
457 		if (m == NULL) {
458 			if ((flags & M_NOWAIT) == 0)
459 				goto retry;
460 			VM_OBJECT_WUNLOCK(object);
461 			kmem_unback(object, addr, i);
462 			return (KERN_NO_SPACE);
463 		}
464 		KASSERT(vm_phys_domain(m) == domain,
465 		    ("kmem_back_domain: Domain mismatch %d != %d",
466 		    vm_phys_domain(m), domain));
467 		if (flags & M_ZERO && (m->flags & PG_ZERO) == 0)
468 			pmap_zero_page(m);
469 		KASSERT((m->oflags & VPO_UNMANAGED) != 0,
470 		    ("kmem_malloc: page %p is managed", m));
471 		m->valid = VM_PAGE_BITS_ALL;
472 		pmap_enter(kernel_pmap, addr + i, m, prot,
473 		    prot | PMAP_ENTER_WIRED, 0);
474 #if VM_NRESERVLEVEL > 0
475 		if (__predict_false((prot & VM_PROT_EXECUTE) != 0))
476 			m->oflags |= VPO_KMEM_EXEC;
477 #endif
478 	}
479 	VM_OBJECT_WUNLOCK(object);
480 
481 	return (KERN_SUCCESS);
482 }
483 
484 /*
485  *	kmem_back:
486  *
487  *	Allocate physical pages for the specified virtual address range.
488  */
489 int
490 kmem_back(vm_object_t object, vm_offset_t addr, vm_size_t size, int flags)
491 {
492 	vm_offset_t end, next, start;
493 	int domain, rv;
494 
495 	KASSERT(object == kernel_object,
496 	    ("kmem_back: only supports kernel object."));
497 
498 	for (start = addr, end = addr + size; addr < end; addr = next) {
499 		/*
500 		 * We must ensure that pages backing a given large virtual page
501 		 * all come from the same physical domain.
502 		 */
503 		if (vm_ndomains > 1) {
504 			domain = (addr >> KVA_QUANTUM_SHIFT) % vm_ndomains;
505 			while (VM_DOMAIN_EMPTY(domain))
506 				domain++;
507 			next = roundup2(addr + 1, KVA_QUANTUM);
508 			if (next > end || next < start)
509 				next = end;
510 		} else {
511 			domain = 0;
512 			next = end;
513 		}
514 		rv = kmem_back_domain(domain, object, addr, next - addr, flags);
515 		if (rv != KERN_SUCCESS) {
516 			kmem_unback(object, start, addr - start);
517 			break;
518 		}
519 	}
520 	return (rv);
521 }
522 
523 /*
524  *	kmem_unback:
525  *
526  *	Unmap and free the physical pages underlying the specified virtual
527  *	address range.
528  *
529  *	A physical page must exist within the specified object at each index
530  *	that is being unmapped.
531  */
532 static struct vmem *
533 _kmem_unback(vm_object_t object, vm_offset_t addr, vm_size_t size)
534 {
535 	struct vmem *arena;
536 	vm_page_t m, next;
537 	vm_offset_t end, offset;
538 	int domain;
539 
540 	KASSERT(object == kernel_object,
541 	    ("kmem_unback: only supports kernel object."));
542 
543 	if (size == 0)
544 		return (NULL);
545 	pmap_remove(kernel_pmap, addr, addr + size);
546 	offset = addr - VM_MIN_KERNEL_ADDRESS;
547 	end = offset + size;
548 	VM_OBJECT_WLOCK(object);
549 	m = vm_page_lookup(object, atop(offset));
550 	domain = vm_phys_domain(m);
551 #if VM_NRESERVLEVEL > 0
552 	if (__predict_true((m->oflags & VPO_KMEM_EXEC) == 0))
553 		arena = vm_dom[domain].vmd_kernel_arena;
554 	else
555 		arena = vm_dom[domain].vmd_kernel_rwx_arena;
556 #else
557 	arena = vm_dom[domain].vmd_kernel_arena;
558 #endif
559 	for (; offset < end; offset += PAGE_SIZE, m = next) {
560 		next = vm_page_next(m);
561 		vm_page_unwire(m, PQ_NONE);
562 		vm_page_free(m);
563 	}
564 	VM_OBJECT_WUNLOCK(object);
565 
566 	return (arena);
567 }
568 
569 void
570 kmem_unback(vm_object_t object, vm_offset_t addr, vm_size_t size)
571 {
572 
573 	(void)_kmem_unback(object, addr, size);
574 }
575 
576 /*
577  *	kmem_free:
578  *
579  *	Free memory allocated with kmem_malloc.  The size must match the
580  *	original allocation.
581  */
582 void
583 kmem_free(vm_offset_t addr, vm_size_t size)
584 {
585 	struct vmem *arena;
586 
587 	size = round_page(size);
588 	arena = _kmem_unback(kernel_object, addr, size);
589 	if (arena != NULL)
590 		vmem_free(arena, addr, size);
591 }
592 
593 /*
594  *	kmap_alloc_wait:
595  *
596  *	Allocates pageable memory from a sub-map of the kernel.  If the submap
597  *	has no room, the caller sleeps waiting for more memory in the submap.
598  *
599  *	This routine may block.
600  */
601 vm_offset_t
602 kmap_alloc_wait(vm_map_t map, vm_size_t size)
603 {
604 	vm_offset_t addr;
605 
606 	size = round_page(size);
607 	if (!swap_reserve(size))
608 		return (0);
609 
610 	for (;;) {
611 		/*
612 		 * To make this work for more than one map, use the map's lock
613 		 * to lock out sleepers/wakers.
614 		 */
615 		vm_map_lock(map);
616 		if (vm_map_findspace(map, vm_map_min(map), size, &addr) == 0)
617 			break;
618 		/* no space now; see if we can ever get space */
619 		if (vm_map_max(map) - vm_map_min(map) < size) {
620 			vm_map_unlock(map);
621 			swap_release(size);
622 			return (0);
623 		}
624 		map->needs_wakeup = TRUE;
625 		vm_map_unlock_and_wait(map, 0);
626 	}
627 	vm_map_insert(map, NULL, 0, addr, addr + size, VM_PROT_ALL,
628 	    VM_PROT_ALL, MAP_ACC_CHARGED);
629 	vm_map_unlock(map);
630 	return (addr);
631 }
632 
633 /*
634  *	kmap_free_wakeup:
635  *
636  *	Returns memory to a submap of the kernel, and wakes up any processes
637  *	waiting for memory in that map.
638  */
639 void
640 kmap_free_wakeup(vm_map_t map, vm_offset_t addr, vm_size_t size)
641 {
642 
643 	vm_map_lock(map);
644 	(void) vm_map_delete(map, trunc_page(addr), round_page(addr + size));
645 	if (map->needs_wakeup) {
646 		map->needs_wakeup = FALSE;
647 		vm_map_wakeup(map);
648 	}
649 	vm_map_unlock(map);
650 }
651 
652 void
653 kmem_init_zero_region(void)
654 {
655 	vm_offset_t addr, i;
656 	vm_page_t m;
657 
658 	/*
659 	 * Map a single physical page of zeros to a larger virtual range.
660 	 * This requires less looping in places that want large amounts of
661 	 * zeros, while not using much more physical resources.
662 	 */
663 	addr = kva_alloc(ZERO_REGION_SIZE);
664 	m = vm_page_alloc(NULL, 0, VM_ALLOC_NORMAL |
665 	    VM_ALLOC_NOOBJ | VM_ALLOC_WIRED | VM_ALLOC_ZERO);
666 	if ((m->flags & PG_ZERO) == 0)
667 		pmap_zero_page(m);
668 	for (i = 0; i < ZERO_REGION_SIZE; i += PAGE_SIZE)
669 		pmap_qenter(addr + i, &m, 1);
670 	pmap_protect(kernel_pmap, addr, addr + ZERO_REGION_SIZE, VM_PROT_READ);
671 
672 	zero_region = (const void *)addr;
673 }
674 
675 /*
676  * Import KVA from the kernel map into the kernel arena.
677  */
678 static int
679 kva_import(void *unused, vmem_size_t size, int flags, vmem_addr_t *addrp)
680 {
681 	vm_offset_t addr;
682 	int result;
683 
684 	KASSERT((size % KVA_QUANTUM) == 0,
685 	    ("kva_import: Size %jd is not a multiple of %d",
686 	    (intmax_t)size, (int)KVA_QUANTUM));
687 	addr = vm_map_min(kernel_map);
688 	result = vm_map_find(kernel_map, NULL, 0, &addr, size, 0,
689 	    VMFS_SUPER_SPACE, VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT);
690 	if (result != KERN_SUCCESS)
691                 return (ENOMEM);
692 
693 	*addrp = addr;
694 
695 	return (0);
696 }
697 
698 /*
699  * Import KVA from a parent arena into a per-domain arena.  Imports must be
700  * KVA_QUANTUM-aligned and a multiple of KVA_QUANTUM in size.
701  */
702 static int
703 kva_import_domain(void *arena, vmem_size_t size, int flags, vmem_addr_t *addrp)
704 {
705 
706 	KASSERT((size % KVA_QUANTUM) == 0,
707 	    ("kva_import_domain: Size %jd is not a multiple of %d",
708 	    (intmax_t)size, (int)KVA_QUANTUM));
709 	return (vmem_xalloc(arena, size, KVA_QUANTUM, 0, 0, VMEM_ADDR_MIN,
710 	    VMEM_ADDR_MAX, flags, addrp));
711 }
712 
713 /*
714  * 	kmem_init:
715  *
716  *	Create the kernel map; insert a mapping covering kernel text,
717  *	data, bss, and all space allocated thus far (`boostrap' data).  The
718  *	new map will thus map the range between VM_MIN_KERNEL_ADDRESS and
719  *	`start' as allocated, and the range between `start' and `end' as free.
720  *	Create the kernel vmem arena and its per-domain children.
721  */
722 void
723 kmem_init(vm_offset_t start, vm_offset_t end)
724 {
725 	vm_map_t m;
726 	int domain;
727 
728 	m = vm_map_create(kernel_pmap, VM_MIN_KERNEL_ADDRESS, end);
729 	m->system_map = 1;
730 	vm_map_lock(m);
731 	/* N.B.: cannot use kgdb to debug, starting with this assignment ... */
732 	kernel_map = m;
733 	(void) vm_map_insert(m, NULL, (vm_ooffset_t) 0,
734 #ifdef __amd64__
735 	    KERNBASE,
736 #else
737 	    VM_MIN_KERNEL_ADDRESS,
738 #endif
739 	    start, VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT);
740 	/* ... and ending with the completion of the above `insert' */
741 	vm_map_unlock(m);
742 
743 	/*
744 	 * Initialize the kernel_arena.  This can grow on demand.
745 	 */
746 	vmem_init(kernel_arena, "kernel arena", 0, 0, PAGE_SIZE, 0, 0);
747 	vmem_set_import(kernel_arena, kva_import, NULL, NULL, KVA_QUANTUM);
748 
749 	for (domain = 0; domain < vm_ndomains; domain++) {
750 		/*
751 		 * Initialize the per-domain arenas.  These are used to color
752 		 * the KVA space in a way that ensures that virtual large pages
753 		 * are backed by memory from the same physical domain,
754 		 * maximizing the potential for superpage promotion.
755 		 */
756 		vm_dom[domain].vmd_kernel_arena = vmem_create(
757 		    "kernel arena domain", 0, 0, PAGE_SIZE, 0, M_WAITOK);
758 		vmem_set_import(vm_dom[domain].vmd_kernel_arena,
759 		    kva_import_domain, NULL, kernel_arena, KVA_QUANTUM);
760 
761 		/*
762 		 * In architectures with superpages, maintain separate arenas
763 		 * for allocations with permissions that differ from the
764 		 * "standard" read/write permissions used for kernel memory,
765 		 * so as not to inhibit superpage promotion.
766 		 */
767 #if VM_NRESERVLEVEL > 0
768 		vm_dom[domain].vmd_kernel_rwx_arena = vmem_create(
769 		    "kernel rwx arena domain", 0, 0, PAGE_SIZE, 0, M_WAITOK);
770 		vmem_set_import(vm_dom[domain].vmd_kernel_rwx_arena,
771 		    kva_import_domain, (vmem_release_t *)vmem_xfree,
772 		    kernel_arena, KVA_QUANTUM);
773 #endif
774 	}
775 }
776 
777 /*
778  *	kmem_bootstrap_free:
779  *
780  *	Free pages backing preloaded data (e.g., kernel modules) to the
781  *	system.  Currently only supported on platforms that create a
782  *	vm_phys segment for preloaded data.
783  */
784 void
785 kmem_bootstrap_free(vm_offset_t start, vm_size_t size)
786 {
787 #if defined(__i386__) || defined(__amd64__)
788 	struct vm_domain *vmd;
789 	vm_offset_t end, va;
790 	vm_paddr_t pa;
791 	vm_page_t m;
792 
793 	end = trunc_page(start + size);
794 	start = round_page(start);
795 
796 	for (va = start; va < end; va += PAGE_SIZE) {
797 		pa = pmap_kextract(va);
798 		m = PHYS_TO_VM_PAGE(pa);
799 
800 		vmd = vm_pagequeue_domain(m);
801 		vm_domain_free_lock(vmd);
802 		vm_phys_free_pages(m, 0);
803 		vmd->vmd_page_count++;
804 		vm_domain_free_unlock(vmd);
805 
806 		vm_domain_freecnt_inc(vmd, 1);
807 		vm_cnt.v_page_count++;
808 	}
809 	pmap_remove(kernel_pmap, start, end);
810 	(void)vmem_add(kernel_arena, start, end - start, M_WAITOK);
811 #endif
812 }
813 
814 #ifdef DIAGNOSTIC
815 /*
816  * Allow userspace to directly trigger the VM drain routine for testing
817  * purposes.
818  */
819 static int
820 debug_vm_lowmem(SYSCTL_HANDLER_ARGS)
821 {
822 	int error, i;
823 
824 	i = 0;
825 	error = sysctl_handle_int(oidp, &i, 0, req);
826 	if (error)
827 		return (error);
828 	if ((i & ~(VM_LOW_KMEM | VM_LOW_PAGES)) != 0)
829 		return (EINVAL);
830 	if (i != 0)
831 		EVENTHANDLER_INVOKE(vm_lowmem, i);
832 	return (0);
833 }
834 
835 SYSCTL_PROC(_debug, OID_AUTO, vm_lowmem, CTLTYPE_INT | CTLFLAG_RW, 0, 0,
836     debug_vm_lowmem, "I", "set to trigger vm_lowmem event with given flags");
837 #endif
838