xref: /freebsd/sys/vm/vm_kern.c (revision 43e8403953cb97ffa6306dd13d2e92fdec91e47c)
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 we want large import sizes. */
128 #define	KVA_QUANTUM_SHIFT	(8 + PAGE_SHIFT)
129 #endif
130 #define	KVA_QUANTUM		(1 << KVA_QUANTUM_SHIFT)
131 
132 extern void     uma_startup2(void);
133 
134 /*
135  *	kva_alloc:
136  *
137  *	Allocate a virtual address range with no underlying object and
138  *	no initial mapping to physical memory.  Any mapping from this
139  *	range to physical memory must be explicitly created prior to
140  *	its use, typically with pmap_qenter().  Any attempt to create
141  *	a mapping on demand through vm_fault() will result in a panic.
142  */
143 vm_offset_t
144 kva_alloc(vm_size_t size)
145 {
146 	vm_offset_t addr;
147 
148 	size = round_page(size);
149 	if (vmem_alloc(kernel_arena, size, M_BESTFIT | M_NOWAIT, &addr))
150 		return (0);
151 
152 	return (addr);
153 }
154 
155 /*
156  *	kva_free:
157  *
158  *	Release a region of kernel virtual memory allocated
159  *	with kva_alloc, and return the physical pages
160  *	associated with that region.
161  *
162  *	This routine may not block on kernel maps.
163  */
164 void
165 kva_free(vm_offset_t addr, vm_size_t size)
166 {
167 
168 	size = round_page(size);
169 	vmem_free(kernel_arena, addr, size);
170 }
171 
172 /*
173  *	Allocates a region from the kernel address map and physical pages
174  *	within the specified address range to the kernel object.  Creates a
175  *	wired mapping from this region to these pages, and returns the
176  *	region's starting virtual address.  The allocated pages are not
177  *	necessarily physically contiguous.  If M_ZERO is specified through the
178  *	given flags, then the pages are zeroed before they are mapped.
179  */
180 static vm_offset_t
181 kmem_alloc_attr_domain(int domain, vm_size_t size, int flags, vm_paddr_t low,
182     vm_paddr_t high, vm_memattr_t memattr)
183 {
184 	vmem_t *vmem;
185 	vm_object_t object = kernel_object;
186 	vm_offset_t addr, i, offset;
187 	vm_page_t m;
188 	int pflags, tries;
189 	vm_prot_t prot;
190 
191 	size = round_page(size);
192 	vmem = vm_dom[domain].vmd_kernel_arena;
193 	if (vmem_alloc(vmem, size, M_BESTFIT | flags, &addr))
194 		return (0);
195 	offset = addr - VM_MIN_KERNEL_ADDRESS;
196 	pflags = malloc2vm_flags(flags) | VM_ALLOC_WIRED;
197 	pflags &= ~(VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL);
198 	pflags |= VM_ALLOC_NOWAIT;
199 	prot = (flags & M_EXEC) != 0 ? VM_PROT_ALL : VM_PROT_RW;
200 	VM_OBJECT_WLOCK(object);
201 	for (i = 0; i < size; i += PAGE_SIZE) {
202 		tries = 0;
203 retry:
204 		m = vm_page_alloc_contig_domain(object, atop(offset + i),
205 		    domain, pflags, 1, low, high, PAGE_SIZE, 0, memattr);
206 		if (m == NULL) {
207 			VM_OBJECT_WUNLOCK(object);
208 			if (tries < ((flags & M_NOWAIT) != 0 ? 1 : 3)) {
209 				if (!vm_page_reclaim_contig_domain(domain,
210 				    pflags, 1, low, high, PAGE_SIZE, 0) &&
211 				    (flags & M_WAITOK) != 0)
212 					vm_wait_domain(domain);
213 				VM_OBJECT_WLOCK(object);
214 				tries++;
215 				goto retry;
216 			}
217 			kmem_unback(object, addr, i);
218 			vmem_free(vmem, addr, size);
219 			return (0);
220 		}
221 		KASSERT(vm_phys_domain(m) == domain,
222 		    ("kmem_alloc_attr_domain: Domain mismatch %d != %d",
223 		    vm_phys_domain(m), domain));
224 		if ((flags & M_ZERO) && (m->flags & PG_ZERO) == 0)
225 			pmap_zero_page(m);
226 		vm_page_valid(m);
227 		vm_page_xunbusy(m);
228 		pmap_enter(kernel_pmap, addr + i, m, prot,
229 		    prot | PMAP_ENTER_WIRED, 0);
230 	}
231 	VM_OBJECT_WUNLOCK(object);
232 	return (addr);
233 }
234 
235 vm_offset_t
236 kmem_alloc_attr(vm_size_t size, int flags, vm_paddr_t low, vm_paddr_t high,
237     vm_memattr_t memattr)
238 {
239 
240 	return (kmem_alloc_attr_domainset(DOMAINSET_RR(), size, flags, low,
241 	    high, memattr));
242 }
243 
244 vm_offset_t
245 kmem_alloc_attr_domainset(struct domainset *ds, vm_size_t size, int flags,
246     vm_paddr_t low, vm_paddr_t high, vm_memattr_t memattr)
247 {
248 	struct vm_domainset_iter di;
249 	vm_offset_t addr;
250 	int domain;
251 
252 	vm_domainset_iter_policy_init(&di, ds, &domain, &flags);
253 	do {
254 		addr = kmem_alloc_attr_domain(domain, size, flags, low, high,
255 		    memattr);
256 		if (addr != 0)
257 			break;
258 	} while (vm_domainset_iter_policy(&di, &domain) == 0);
259 
260 	return (addr);
261 }
262 
263 /*
264  *	Allocates a region from the kernel address map and physically
265  *	contiguous pages within the specified address range to the kernel
266  *	object.  Creates a wired mapping from this region to these pages, and
267  *	returns the region's starting virtual address.  If M_ZERO is specified
268  *	through the given flags, then the pages are zeroed before they are
269  *	mapped.
270  */
271 static vm_offset_t
272 kmem_alloc_contig_domain(int domain, vm_size_t size, int flags, vm_paddr_t low,
273     vm_paddr_t high, u_long alignment, vm_paddr_t boundary,
274     vm_memattr_t memattr)
275 {
276 	vmem_t *vmem;
277 	vm_object_t object = kernel_object;
278 	vm_offset_t addr, offset, tmp;
279 	vm_page_t end_m, m;
280 	u_long npages;
281 	int pflags, tries;
282 
283 	size = round_page(size);
284 	vmem = vm_dom[domain].vmd_kernel_arena;
285 	if (vmem_alloc(vmem, size, flags | M_BESTFIT, &addr))
286 		return (0);
287 	offset = addr - VM_MIN_KERNEL_ADDRESS;
288 	pflags = malloc2vm_flags(flags) | VM_ALLOC_WIRED;
289 	pflags &= ~(VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL);
290 	pflags |= VM_ALLOC_NOWAIT;
291 	npages = atop(size);
292 	VM_OBJECT_WLOCK(object);
293 	tries = 0;
294 retry:
295 	m = vm_page_alloc_contig_domain(object, atop(offset), domain, pflags,
296 	    npages, low, high, alignment, boundary, memattr);
297 	if (m == NULL) {
298 		VM_OBJECT_WUNLOCK(object);
299 		if (tries < ((flags & M_NOWAIT) != 0 ? 1 : 3)) {
300 			if (!vm_page_reclaim_contig_domain(domain, pflags,
301 			    npages, low, high, alignment, boundary) &&
302 			    (flags & M_WAITOK) != 0)
303 				vm_wait_domain(domain);
304 			VM_OBJECT_WLOCK(object);
305 			tries++;
306 			goto retry;
307 		}
308 		vmem_free(vmem, addr, size);
309 		return (0);
310 	}
311 	KASSERT(vm_phys_domain(m) == domain,
312 	    ("kmem_alloc_contig_domain: Domain mismatch %d != %d",
313 	    vm_phys_domain(m), domain));
314 	end_m = m + npages;
315 	tmp = addr;
316 	for (; m < end_m; m++) {
317 		if ((flags & M_ZERO) && (m->flags & PG_ZERO) == 0)
318 			pmap_zero_page(m);
319 		vm_page_valid(m);
320 		vm_page_xunbusy(m);
321 		pmap_enter(kernel_pmap, tmp, m, VM_PROT_RW,
322 		    VM_PROT_RW | PMAP_ENTER_WIRED, 0);
323 		tmp += PAGE_SIZE;
324 	}
325 	VM_OBJECT_WUNLOCK(object);
326 	return (addr);
327 }
328 
329 vm_offset_t
330 kmem_alloc_contig(vm_size_t size, int flags, vm_paddr_t low, vm_paddr_t high,
331     u_long alignment, vm_paddr_t boundary, vm_memattr_t memattr)
332 {
333 
334 	return (kmem_alloc_contig_domainset(DOMAINSET_RR(), size, flags, low,
335 	    high, alignment, boundary, memattr));
336 }
337 
338 vm_offset_t
339 kmem_alloc_contig_domainset(struct domainset *ds, vm_size_t size, int flags,
340     vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary,
341     vm_memattr_t memattr)
342 {
343 	struct vm_domainset_iter di;
344 	vm_offset_t addr;
345 	int domain;
346 
347 	vm_domainset_iter_policy_init(&di, ds, &domain, &flags);
348 	do {
349 		addr = kmem_alloc_contig_domain(domain, size, flags, low, high,
350 		    alignment, boundary, memattr);
351 		if (addr != 0)
352 			break;
353 	} while (vm_domainset_iter_policy(&di, &domain) == 0);
354 
355 	return (addr);
356 }
357 
358 /*
359  *	kmem_suballoc:
360  *
361  *	Allocates a map to manage a subrange
362  *	of the kernel virtual address space.
363  *
364  *	Arguments are as follows:
365  *
366  *	parent		Map to take range from
367  *	min, max	Returned endpoints of map
368  *	size		Size of range to find
369  *	superpage_align	Request that min is superpage aligned
370  */
371 vm_map_t
372 kmem_suballoc(vm_map_t parent, vm_offset_t *min, vm_offset_t *max,
373     vm_size_t size, boolean_t superpage_align)
374 {
375 	int ret;
376 	vm_map_t result;
377 
378 	size = round_page(size);
379 
380 	*min = vm_map_min(parent);
381 	ret = vm_map_find(parent, NULL, 0, min, size, 0, superpage_align ?
382 	    VMFS_SUPER_SPACE : VMFS_ANY_SPACE, VM_PROT_ALL, VM_PROT_ALL,
383 	    MAP_ACC_NO_CHARGE);
384 	if (ret != KERN_SUCCESS)
385 		panic("kmem_suballoc: bad status return of %d", ret);
386 	*max = *min + size;
387 	result = vm_map_create(vm_map_pmap(parent), *min, *max);
388 	if (result == NULL)
389 		panic("kmem_suballoc: cannot create submap");
390 	if (vm_map_submap(parent, *min, *max, result) != KERN_SUCCESS)
391 		panic("kmem_suballoc: unable to change range to submap");
392 	return (result);
393 }
394 
395 /*
396  *	kmem_malloc_domain:
397  *
398  *	Allocate wired-down pages in the kernel's address space.
399  */
400 static vm_offset_t
401 kmem_malloc_domain(int domain, vm_size_t size, int flags)
402 {
403 	vmem_t *arena;
404 	vm_offset_t addr;
405 	int rv;
406 
407 #if VM_NRESERVLEVEL > 0
408 	if (__predict_true((flags & M_EXEC) == 0))
409 		arena = vm_dom[domain].vmd_kernel_arena;
410 	else
411 		arena = vm_dom[domain].vmd_kernel_rwx_arena;
412 #else
413 	arena = vm_dom[domain].vmd_kernel_arena;
414 #endif
415 	size = round_page(size);
416 	if (vmem_alloc(arena, size, flags | M_BESTFIT, &addr))
417 		return (0);
418 
419 	rv = kmem_back_domain(domain, kernel_object, addr, size, flags);
420 	if (rv != KERN_SUCCESS) {
421 		vmem_free(arena, addr, size);
422 		return (0);
423 	}
424 	return (addr);
425 }
426 
427 vm_offset_t
428 kmem_malloc(vm_size_t size, int flags)
429 {
430 
431 	return (kmem_malloc_domainset(DOMAINSET_RR(), size, flags));
432 }
433 
434 vm_offset_t
435 kmem_malloc_domainset(struct domainset *ds, vm_size_t size, int flags)
436 {
437 	struct vm_domainset_iter di;
438 	vm_offset_t addr;
439 	int domain;
440 
441 	vm_domainset_iter_policy_init(&di, ds, &domain, &flags);
442 	do {
443 		addr = kmem_malloc_domain(domain, size, flags);
444 		if (addr != 0)
445 			break;
446 	} while (vm_domainset_iter_policy(&di, &domain) == 0);
447 
448 	return (addr);
449 }
450 
451 /*
452  *	kmem_back_domain:
453  *
454  *	Allocate physical pages from the specified domain for the specified
455  *	virtual address range.
456  */
457 int
458 kmem_back_domain(int domain, vm_object_t object, vm_offset_t addr,
459     vm_size_t size, int flags)
460 {
461 	vm_offset_t offset, i;
462 	vm_page_t m, mpred;
463 	vm_prot_t prot;
464 	int pflags;
465 
466 	KASSERT(object == kernel_object,
467 	    ("kmem_back_domain: only supports kernel object."));
468 
469 	offset = addr - VM_MIN_KERNEL_ADDRESS;
470 	pflags = malloc2vm_flags(flags) | VM_ALLOC_WIRED;
471 	pflags &= ~(VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL);
472 	if (flags & M_WAITOK)
473 		pflags |= VM_ALLOC_WAITFAIL;
474 	prot = (flags & M_EXEC) != 0 ? VM_PROT_ALL : VM_PROT_RW;
475 
476 	i = 0;
477 	VM_OBJECT_WLOCK(object);
478 retry:
479 	mpred = vm_radix_lookup_le(&object->rtree, atop(offset + i));
480 	for (; i < size; i += PAGE_SIZE, mpred = m) {
481 		m = vm_page_alloc_domain_after(object, atop(offset + i),
482 		    domain, pflags, mpred);
483 
484 		/*
485 		 * Ran out of space, free everything up and return. Don't need
486 		 * to lock page queues here as we know that the pages we got
487 		 * aren't on any queues.
488 		 */
489 		if (m == NULL) {
490 			if ((flags & M_NOWAIT) == 0)
491 				goto retry;
492 			VM_OBJECT_WUNLOCK(object);
493 			kmem_unback(object, addr, i);
494 			return (KERN_NO_SPACE);
495 		}
496 		KASSERT(vm_phys_domain(m) == domain,
497 		    ("kmem_back_domain: Domain mismatch %d != %d",
498 		    vm_phys_domain(m), domain));
499 		if (flags & M_ZERO && (m->flags & PG_ZERO) == 0)
500 			pmap_zero_page(m);
501 		KASSERT((m->oflags & VPO_UNMANAGED) != 0,
502 		    ("kmem_malloc: page %p is managed", m));
503 		vm_page_valid(m);
504 		vm_page_xunbusy(m);
505 		pmap_enter(kernel_pmap, addr + i, m, prot,
506 		    prot | PMAP_ENTER_WIRED, 0);
507 #if VM_NRESERVLEVEL > 0
508 		if (__predict_false((prot & VM_PROT_EXECUTE) != 0))
509 			m->oflags |= VPO_KMEM_EXEC;
510 #endif
511 	}
512 	VM_OBJECT_WUNLOCK(object);
513 
514 	return (KERN_SUCCESS);
515 }
516 
517 /*
518  *	kmem_back:
519  *
520  *	Allocate physical pages for the specified virtual address range.
521  */
522 int
523 kmem_back(vm_object_t object, vm_offset_t addr, vm_size_t size, int flags)
524 {
525 	vm_offset_t end, next, start;
526 	int domain, rv;
527 
528 	KASSERT(object == kernel_object,
529 	    ("kmem_back: only supports kernel object."));
530 
531 	for (start = addr, end = addr + size; addr < end; addr = next) {
532 		/*
533 		 * We must ensure that pages backing a given large virtual page
534 		 * all come from the same physical domain.
535 		 */
536 		if (vm_ndomains > 1) {
537 			domain = (addr >> KVA_QUANTUM_SHIFT) % vm_ndomains;
538 			while (VM_DOMAIN_EMPTY(domain))
539 				domain++;
540 			next = roundup2(addr + 1, KVA_QUANTUM);
541 			if (next > end || next < start)
542 				next = end;
543 		} else {
544 			domain = 0;
545 			next = end;
546 		}
547 		rv = kmem_back_domain(domain, object, addr, next - addr, flags);
548 		if (rv != KERN_SUCCESS) {
549 			kmem_unback(object, start, addr - start);
550 			break;
551 		}
552 	}
553 	return (rv);
554 }
555 
556 /*
557  *	kmem_unback:
558  *
559  *	Unmap and free the physical pages underlying the specified virtual
560  *	address range.
561  *
562  *	A physical page must exist within the specified object at each index
563  *	that is being unmapped.
564  */
565 static struct vmem *
566 _kmem_unback(vm_object_t object, vm_offset_t addr, vm_size_t size)
567 {
568 	struct vmem *arena;
569 	vm_page_t m, next;
570 	vm_offset_t end, offset;
571 	int domain;
572 
573 	KASSERT(object == kernel_object,
574 	    ("kmem_unback: only supports kernel object."));
575 
576 	if (size == 0)
577 		return (NULL);
578 	pmap_remove(kernel_pmap, addr, addr + size);
579 	offset = addr - VM_MIN_KERNEL_ADDRESS;
580 	end = offset + size;
581 	VM_OBJECT_WLOCK(object);
582 	m = vm_page_lookup(object, atop(offset));
583 	domain = vm_phys_domain(m);
584 #if VM_NRESERVLEVEL > 0
585 	if (__predict_true((m->oflags & VPO_KMEM_EXEC) == 0))
586 		arena = vm_dom[domain].vmd_kernel_arena;
587 	else
588 		arena = vm_dom[domain].vmd_kernel_rwx_arena;
589 #else
590 	arena = vm_dom[domain].vmd_kernel_arena;
591 #endif
592 	for (; offset < end; offset += PAGE_SIZE, m = next) {
593 		next = vm_page_next(m);
594 		vm_page_busy_acquire(m, 0);
595 		vm_page_unwire_noq(m);
596 		vm_page_free(m);
597 	}
598 	VM_OBJECT_WUNLOCK(object);
599 
600 	return (arena);
601 }
602 
603 void
604 kmem_unback(vm_object_t object, vm_offset_t addr, vm_size_t size)
605 {
606 
607 	(void)_kmem_unback(object, addr, size);
608 }
609 
610 /*
611  *	kmem_free:
612  *
613  *	Free memory allocated with kmem_malloc.  The size must match the
614  *	original allocation.
615  */
616 void
617 kmem_free(vm_offset_t addr, vm_size_t size)
618 {
619 	struct vmem *arena;
620 
621 	size = round_page(size);
622 	arena = _kmem_unback(kernel_object, addr, size);
623 	if (arena != NULL)
624 		vmem_free(arena, addr, size);
625 }
626 
627 /*
628  *	kmap_alloc_wait:
629  *
630  *	Allocates pageable memory from a sub-map of the kernel.  If the submap
631  *	has no room, the caller sleeps waiting for more memory in the submap.
632  *
633  *	This routine may block.
634  */
635 vm_offset_t
636 kmap_alloc_wait(vm_map_t map, vm_size_t size)
637 {
638 	vm_offset_t addr;
639 
640 	size = round_page(size);
641 	if (!swap_reserve(size))
642 		return (0);
643 
644 	for (;;) {
645 		/*
646 		 * To make this work for more than one map, use the map's lock
647 		 * to lock out sleepers/wakers.
648 		 */
649 		vm_map_lock(map);
650 		addr = vm_map_findspace(map, vm_map_min(map), size);
651 		if (addr + size <= vm_map_max(map))
652 			break;
653 		/* no space now; see if we can ever get space */
654 		if (vm_map_max(map) - vm_map_min(map) < size) {
655 			vm_map_unlock(map);
656 			swap_release(size);
657 			return (0);
658 		}
659 		map->needs_wakeup = TRUE;
660 		vm_map_unlock_and_wait(map, 0);
661 	}
662 	vm_map_insert(map, NULL, 0, addr, addr + size, VM_PROT_RW, VM_PROT_RW,
663 	    MAP_ACC_CHARGED);
664 	vm_map_unlock(map);
665 	return (addr);
666 }
667 
668 /*
669  *	kmap_free_wakeup:
670  *
671  *	Returns memory to a submap of the kernel, and wakes up any processes
672  *	waiting for memory in that map.
673  */
674 void
675 kmap_free_wakeup(vm_map_t map, vm_offset_t addr, vm_size_t size)
676 {
677 
678 	vm_map_lock(map);
679 	(void) vm_map_delete(map, trunc_page(addr), round_page(addr + size));
680 	if (map->needs_wakeup) {
681 		map->needs_wakeup = FALSE;
682 		vm_map_wakeup(map);
683 	}
684 	vm_map_unlock(map);
685 }
686 
687 void
688 kmem_init_zero_region(void)
689 {
690 	vm_offset_t addr, i;
691 	vm_page_t m;
692 
693 	/*
694 	 * Map a single physical page of zeros to a larger virtual range.
695 	 * This requires less looping in places that want large amounts of
696 	 * zeros, while not using much more physical resources.
697 	 */
698 	addr = kva_alloc(ZERO_REGION_SIZE);
699 	m = vm_page_alloc(NULL, 0, VM_ALLOC_NORMAL |
700 	    VM_ALLOC_NOOBJ | VM_ALLOC_WIRED | VM_ALLOC_ZERO);
701 	if ((m->flags & PG_ZERO) == 0)
702 		pmap_zero_page(m);
703 	for (i = 0; i < ZERO_REGION_SIZE; i += PAGE_SIZE)
704 		pmap_qenter(addr + i, &m, 1);
705 	pmap_protect(kernel_pmap, addr, addr + ZERO_REGION_SIZE, VM_PROT_READ);
706 
707 	zero_region = (const void *)addr;
708 }
709 
710 /*
711  * Import KVA from the kernel map into the kernel arena.
712  */
713 static int
714 kva_import(void *unused, vmem_size_t size, int flags, vmem_addr_t *addrp)
715 {
716 	vm_offset_t addr;
717 	int result;
718 
719 	KASSERT((size % KVA_QUANTUM) == 0,
720 	    ("kva_import: Size %jd is not a multiple of %d",
721 	    (intmax_t)size, (int)KVA_QUANTUM));
722 	addr = vm_map_min(kernel_map);
723 	result = vm_map_find(kernel_map, NULL, 0, &addr, size, 0,
724 	    VMFS_SUPER_SPACE, VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT);
725 	if (result != KERN_SUCCESS)
726                 return (ENOMEM);
727 
728 	*addrp = addr;
729 
730 	return (0);
731 }
732 
733 /*
734  * Import KVA from a parent arena into a per-domain arena.  Imports must be
735  * KVA_QUANTUM-aligned and a multiple of KVA_QUANTUM in size.
736  */
737 static int
738 kva_import_domain(void *arena, vmem_size_t size, int flags, vmem_addr_t *addrp)
739 {
740 
741 	KASSERT((size % KVA_QUANTUM) == 0,
742 	    ("kva_import_domain: Size %jd is not a multiple of %d",
743 	    (intmax_t)size, (int)KVA_QUANTUM));
744 	return (vmem_xalloc(arena, size, KVA_QUANTUM, 0, 0, VMEM_ADDR_MIN,
745 	    VMEM_ADDR_MAX, flags, addrp));
746 }
747 
748 /*
749  * 	kmem_init:
750  *
751  *	Create the kernel map; insert a mapping covering kernel text,
752  *	data, bss, and all space allocated thus far (`boostrap' data).  The
753  *	new map will thus map the range between VM_MIN_KERNEL_ADDRESS and
754  *	`start' as allocated, and the range between `start' and `end' as free.
755  *	Create the kernel vmem arena and its per-domain children.
756  */
757 void
758 kmem_init(vm_offset_t start, vm_offset_t end)
759 {
760 	vm_map_t m;
761 	int domain;
762 
763 	m = vm_map_create(kernel_pmap, VM_MIN_KERNEL_ADDRESS, end);
764 	m->system_map = 1;
765 	vm_map_lock(m);
766 	/* N.B.: cannot use kgdb to debug, starting with this assignment ... */
767 	kernel_map = m;
768 	(void)vm_map_insert(m, NULL, 0,
769 #ifdef __amd64__
770 	    KERNBASE,
771 #else
772 	    VM_MIN_KERNEL_ADDRESS,
773 #endif
774 	    start, VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT);
775 	/* ... and ending with the completion of the above `insert' */
776 
777 #ifdef __amd64__
778 	/*
779 	 * Mark KVA used for the page array as allocated.  Other platforms
780 	 * that handle vm_page_array allocation can simply adjust virtual_avail
781 	 * instead.
782 	 */
783 	(void)vm_map_insert(m, NULL, 0, (vm_offset_t)vm_page_array,
784 	    (vm_offset_t)vm_page_array + round_2mpage(vm_page_array_size *
785 	    sizeof(struct vm_page)),
786 	    VM_PROT_RW, VM_PROT_RW, MAP_NOFAULT);
787 #endif
788 	vm_map_unlock(m);
789 
790 	/*
791 	 * Initialize the kernel_arena.  This can grow on demand.
792 	 */
793 	vmem_init(kernel_arena, "kernel arena", 0, 0, PAGE_SIZE, 0, 0);
794 	vmem_set_import(kernel_arena, kva_import, NULL, NULL, KVA_QUANTUM);
795 
796 	for (domain = 0; domain < vm_ndomains; domain++) {
797 		/*
798 		 * Initialize the per-domain arenas.  These are used to color
799 		 * the KVA space in a way that ensures that virtual large pages
800 		 * are backed by memory from the same physical domain,
801 		 * maximizing the potential for superpage promotion.
802 		 */
803 		vm_dom[domain].vmd_kernel_arena = vmem_create(
804 		    "kernel arena domain", 0, 0, PAGE_SIZE, 0, M_WAITOK);
805 		vmem_set_import(vm_dom[domain].vmd_kernel_arena,
806 		    kva_import_domain, NULL, kernel_arena, KVA_QUANTUM);
807 
808 		/*
809 		 * In architectures with superpages, maintain separate arenas
810 		 * for allocations with permissions that differ from the
811 		 * "standard" read/write permissions used for kernel memory,
812 		 * so as not to inhibit superpage promotion.
813 		 */
814 #if VM_NRESERVLEVEL > 0
815 		vm_dom[domain].vmd_kernel_rwx_arena = vmem_create(
816 		    "kernel rwx arena domain", 0, 0, PAGE_SIZE, 0, M_WAITOK);
817 		vmem_set_import(vm_dom[domain].vmd_kernel_rwx_arena,
818 		    kva_import_domain, (vmem_release_t *)vmem_xfree,
819 		    kernel_arena, KVA_QUANTUM);
820 #endif
821 	}
822 
823 	/*
824 	 * This must be the very first call so that the virtual address
825 	 * space used for early allocations is properly marked used in
826 	 * the map.
827 	 */
828 	uma_startup2();
829 }
830 
831 /*
832  *	kmem_bootstrap_free:
833  *
834  *	Free pages backing preloaded data (e.g., kernel modules) to the
835  *	system.  Currently only supported on platforms that create a
836  *	vm_phys segment for preloaded data.
837  */
838 void
839 kmem_bootstrap_free(vm_offset_t start, vm_size_t size)
840 {
841 #if defined(__i386__) || defined(__amd64__)
842 	struct vm_domain *vmd;
843 	vm_offset_t end, va;
844 	vm_paddr_t pa;
845 	vm_page_t m;
846 
847 	end = trunc_page(start + size);
848 	start = round_page(start);
849 
850 #ifdef __amd64__
851 	/*
852 	 * Preloaded files do not have execute permissions by default on amd64.
853 	 * Restore the default permissions to ensure that the direct map alias
854 	 * is updated.
855 	 */
856 	pmap_change_prot(start, end - start, VM_PROT_RW);
857 #endif
858 	for (va = start; va < end; va += PAGE_SIZE) {
859 		pa = pmap_kextract(va);
860 		m = PHYS_TO_VM_PAGE(pa);
861 
862 		vmd = vm_pagequeue_domain(m);
863 		vm_domain_free_lock(vmd);
864 		vm_phys_free_pages(m, 0);
865 		vm_domain_free_unlock(vmd);
866 
867 		vm_domain_freecnt_inc(vmd, 1);
868 		vm_cnt.v_page_count++;
869 	}
870 	pmap_remove(kernel_pmap, start, end);
871 	(void)vmem_add(kernel_arena, start, end - start, M_WAITOK);
872 #endif
873 }
874 
875 /*
876  * Allow userspace to directly trigger the VM drain routine for testing
877  * purposes.
878  */
879 static int
880 debug_vm_lowmem(SYSCTL_HANDLER_ARGS)
881 {
882 	int error, i;
883 
884 	i = 0;
885 	error = sysctl_handle_int(oidp, &i, 0, req);
886 	if (error)
887 		return (error);
888 	if ((i & ~(VM_LOW_KMEM | VM_LOW_PAGES)) != 0)
889 		return (EINVAL);
890 	if (i != 0)
891 		EVENTHANDLER_INVOKE(vm_lowmem, i);
892 	return (0);
893 }
894 
895 SYSCTL_PROC(_debug, OID_AUTO, vm_lowmem, CTLTYPE_INT | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, 0,
896     debug_vm_lowmem, "I", "set to trigger vm_lowmem event with given flags");
897