xref: /freebsd/sys/vm/vm_kern.c (revision 78adacd4eab39a3508bd8c65f0aba94fc6b907ce)
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_pagequeue.h>
95 #include <vm/vm_phys.h>
96 #include <vm/vm_radix.h>
97 #include <vm/vm_extern.h>
98 #include <vm/uma.h>
99 
100 struct vm_map kernel_map_store;
101 struct vm_map exec_map_store;
102 struct vm_map pipe_map_store;
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__)
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 #define	KVA_NUMA_IMPORT_QUANTUM	(KVA_QUANTUM * 128)
132 
133 extern void     uma_startup2(void);
134 
135 /*
136  *	kva_alloc:
137  *
138  *	Allocate a virtual address range with no underlying object and
139  *	no initial mapping to physical memory.  Any mapping from this
140  *	range to physical memory must be explicitly created prior to
141  *	its use, typically with pmap_qenter().  Any attempt to create
142  *	a mapping on demand through vm_fault() will result in a panic.
143  */
144 vm_offset_t
145 kva_alloc(vm_size_t size)
146 {
147 	vm_offset_t addr;
148 
149 	size = round_page(size);
150 	if (vmem_alloc(kernel_arena, size, M_BESTFIT | M_NOWAIT, &addr))
151 		return (0);
152 
153 	return (addr);
154 }
155 
156 /*
157  *	kva_free:
158  *
159  *	Release a region of kernel virtual memory allocated
160  *	with kva_alloc, and return the physical pages
161  *	associated with that region.
162  *
163  *	This routine may not block on kernel maps.
164  */
165 void
166 kva_free(vm_offset_t addr, vm_size_t size)
167 {
168 
169 	size = round_page(size);
170 	vmem_free(kernel_arena, addr, size);
171 }
172 
173 static vm_page_t
174 kmem_alloc_contig_pages(vm_object_t object, vm_pindex_t pindex, int domain,
175     int pflags, u_long npages, vm_paddr_t low, vm_paddr_t high,
176     u_long alignment, vm_paddr_t boundary, vm_memattr_t memattr)
177 {
178 	vm_page_t m;
179 	int tries;
180 	bool wait;
181 
182 	VM_OBJECT_ASSERT_WLOCKED(object);
183 
184 	wait = (pflags & VM_ALLOC_WAITOK) != 0;
185 	pflags &= ~(VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL);
186 	pflags |= VM_ALLOC_NOWAIT;
187 	for (tries = wait ? 3 : 1;; tries--) {
188 		m = vm_page_alloc_contig_domain(object, pindex, domain, pflags,
189 		    npages, low, high, alignment, boundary, memattr);
190 		if (m != NULL || tries == 0)
191 			break;
192 
193 		VM_OBJECT_WUNLOCK(object);
194 		if (!vm_page_reclaim_contig_domain(domain, pflags, npages,
195 		    low, high, alignment, boundary) && wait)
196 			vm_wait_domain(domain);
197 		VM_OBJECT_WLOCK(object);
198 	}
199 	return (m);
200 }
201 
202 /*
203  *	Allocates a region from the kernel address map and physical pages
204  *	within the specified address range to the kernel object.  Creates a
205  *	wired mapping from this region to these pages, and returns the
206  *	region's starting virtual address.  The allocated pages are not
207  *	necessarily physically contiguous.  If M_ZERO is specified through the
208  *	given flags, then the pages are zeroed before they are mapped.
209  */
210 static vm_offset_t
211 kmem_alloc_attr_domain(int domain, vm_size_t size, int flags, vm_paddr_t low,
212     vm_paddr_t high, vm_memattr_t memattr)
213 {
214 	vmem_t *vmem;
215 	vm_object_t object;
216 	vm_offset_t addr, i, offset;
217 	vm_page_t m;
218 	int pflags;
219 	vm_prot_t prot;
220 
221 	object = kernel_object;
222 	size = round_page(size);
223 	vmem = vm_dom[domain].vmd_kernel_arena;
224 	if (vmem_alloc(vmem, size, M_BESTFIT | flags, &addr))
225 		return (0);
226 	offset = addr - VM_MIN_KERNEL_ADDRESS;
227 	pflags = malloc2vm_flags(flags) | VM_ALLOC_WIRED;
228 	prot = (flags & M_EXEC) != 0 ? VM_PROT_ALL : VM_PROT_RW;
229 	VM_OBJECT_WLOCK(object);
230 	for (i = 0; i < size; i += PAGE_SIZE) {
231 		m = kmem_alloc_contig_pages(object, atop(offset + i),
232 		    domain, pflags, 1, low, high, PAGE_SIZE, 0, memattr);
233 		if (m == NULL) {
234 			VM_OBJECT_WUNLOCK(object);
235 			kmem_unback(object, addr, i);
236 			vmem_free(vmem, addr, size);
237 			return (0);
238 		}
239 		KASSERT(vm_page_domain(m) == domain,
240 		    ("kmem_alloc_attr_domain: Domain mismatch %d != %d",
241 		    vm_page_domain(m), domain));
242 		if ((flags & M_ZERO) && (m->flags & PG_ZERO) == 0)
243 			pmap_zero_page(m);
244 		vm_page_valid(m);
245 		pmap_enter(kernel_pmap, addr + i, m, prot,
246 		    prot | PMAP_ENTER_WIRED, 0);
247 	}
248 	VM_OBJECT_WUNLOCK(object);
249 	return (addr);
250 }
251 
252 vm_offset_t
253 kmem_alloc_attr(vm_size_t size, int flags, vm_paddr_t low, vm_paddr_t high,
254     vm_memattr_t memattr)
255 {
256 
257 	return (kmem_alloc_attr_domainset(DOMAINSET_RR(), size, flags, low,
258 	    high, memattr));
259 }
260 
261 vm_offset_t
262 kmem_alloc_attr_domainset(struct domainset *ds, vm_size_t size, int flags,
263     vm_paddr_t low, vm_paddr_t high, vm_memattr_t memattr)
264 {
265 	struct vm_domainset_iter di;
266 	vm_offset_t addr;
267 	int domain, iflags;
268 
269 	/*
270 	 * Do not allow the domainset iterator to override wait flags.  The
271 	 * contiguous memory allocator defines special semantics for M_WAITOK
272 	 * that do not match the iterator's implementation.
273 	 */
274 	iflags = (flags & ~M_WAITOK) | M_NOWAIT;
275 	vm_domainset_iter_policy_init(&di, ds, &domain, &iflags);
276 	do {
277 		addr = kmem_alloc_attr_domain(domain, size, flags, low, high,
278 		    memattr);
279 		if (addr != 0)
280 			break;
281 	} while (vm_domainset_iter_policy(&di, &domain) == 0);
282 
283 	return (addr);
284 }
285 
286 /*
287  *	Allocates a region from the kernel address map and physically
288  *	contiguous pages within the specified address range to the kernel
289  *	object.  Creates a wired mapping from this region to these pages, and
290  *	returns the region's starting virtual address.  If M_ZERO is specified
291  *	through the given flags, then the pages are zeroed before they are
292  *	mapped.
293  */
294 static vm_offset_t
295 kmem_alloc_contig_domain(int domain, vm_size_t size, int flags, vm_paddr_t low,
296     vm_paddr_t high, u_long alignment, vm_paddr_t boundary,
297     vm_memattr_t memattr)
298 {
299 	vmem_t *vmem;
300 	vm_object_t object;
301 	vm_offset_t addr, offset, tmp;
302 	vm_page_t end_m, m;
303 	u_long npages;
304 	int pflags;
305 
306 	object = kernel_object;
307 	size = round_page(size);
308 	vmem = vm_dom[domain].vmd_kernel_arena;
309 	if (vmem_alloc(vmem, size, flags | M_BESTFIT, &addr))
310 		return (0);
311 	offset = addr - VM_MIN_KERNEL_ADDRESS;
312 	pflags = malloc2vm_flags(flags) | VM_ALLOC_WIRED;
313 	npages = atop(size);
314 	VM_OBJECT_WLOCK(object);
315 	m = kmem_alloc_contig_pages(object, atop(offset), domain,
316 	    pflags, npages, low, high, alignment, boundary, memattr);
317 	if (m == NULL) {
318 		VM_OBJECT_WUNLOCK(object);
319 		vmem_free(vmem, addr, size);
320 		return (0);
321 	}
322 	KASSERT(vm_page_domain(m) == domain,
323 	    ("kmem_alloc_contig_domain: Domain mismatch %d != %d",
324 	    vm_page_domain(m), domain));
325 	end_m = m + npages;
326 	tmp = addr;
327 	for (; m < end_m; m++) {
328 		if ((flags & M_ZERO) && (m->flags & PG_ZERO) == 0)
329 			pmap_zero_page(m);
330 		vm_page_valid(m);
331 		pmap_enter(kernel_pmap, tmp, m, VM_PROT_RW,
332 		    VM_PROT_RW | PMAP_ENTER_WIRED, 0);
333 		tmp += PAGE_SIZE;
334 	}
335 	VM_OBJECT_WUNLOCK(object);
336 	return (addr);
337 }
338 
339 vm_offset_t
340 kmem_alloc_contig(vm_size_t size, int flags, vm_paddr_t low, vm_paddr_t high,
341     u_long alignment, vm_paddr_t boundary, vm_memattr_t memattr)
342 {
343 
344 	return (kmem_alloc_contig_domainset(DOMAINSET_RR(), size, flags, low,
345 	    high, alignment, boundary, memattr));
346 }
347 
348 vm_offset_t
349 kmem_alloc_contig_domainset(struct domainset *ds, vm_size_t size, int flags,
350     vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary,
351     vm_memattr_t memattr)
352 {
353 	struct vm_domainset_iter di;
354 	vm_offset_t addr;
355 	int domain, iflags;
356 
357 	/*
358 	 * Do not allow the domainset iterator to override wait flags.  The
359 	 * contiguous memory allocator defines special semantics for M_WAITOK
360 	 * that do not match the iterator's implementation.
361 	 */
362 	iflags = (flags & ~M_WAITOK) | M_NOWAIT;
363 	vm_domainset_iter_policy_init(&di, ds, &domain, &iflags);
364 	do {
365 		addr = kmem_alloc_contig_domain(domain, size, flags, low, high,
366 		    alignment, boundary, memattr);
367 		if (addr != 0)
368 			break;
369 	} while (vm_domainset_iter_policy(&di, &domain) == 0);
370 
371 	return (addr);
372 }
373 
374 /*
375  *	kmem_subinit:
376  *
377  *	Initializes a map to manage a subrange
378  *	of the kernel virtual address space.
379  *
380  *	Arguments are as follows:
381  *
382  *	parent		Map to take range from
383  *	min, max	Returned endpoints of map
384  *	size		Size of range to find
385  *	superpage_align	Request that min is superpage aligned
386  */
387 void
388 kmem_subinit(vm_map_t map, vm_map_t parent, vm_offset_t *min, vm_offset_t *max,
389     vm_size_t size, bool superpage_align)
390 {
391 	int ret;
392 
393 	size = round_page(size);
394 
395 	*min = vm_map_min(parent);
396 	ret = vm_map_find(parent, NULL, 0, min, size, 0, superpage_align ?
397 	    VMFS_SUPER_SPACE : VMFS_ANY_SPACE, VM_PROT_ALL, VM_PROT_ALL,
398 	    MAP_ACC_NO_CHARGE);
399 	if (ret != KERN_SUCCESS)
400 		panic("kmem_subinit: bad status return of %d", ret);
401 	*max = *min + size;
402 	vm_map_init(map, vm_map_pmap(parent), *min, *max);
403 	if (vm_map_submap(parent, *min, *max, map) != KERN_SUCCESS)
404 		panic("kmem_subinit: unable to change range to submap");
405 }
406 
407 /*
408  *	kmem_malloc_domain:
409  *
410  *	Allocate wired-down pages in the kernel's address space.
411  */
412 static vm_offset_t
413 kmem_malloc_domain(int domain, vm_size_t size, int flags)
414 {
415 	vmem_t *arena;
416 	vm_offset_t addr;
417 	int rv;
418 
419 	if (__predict_true((flags & M_EXEC) == 0))
420 		arena = vm_dom[domain].vmd_kernel_arena;
421 	else
422 		arena = vm_dom[domain].vmd_kernel_rwx_arena;
423 	size = round_page(size);
424 	if (vmem_alloc(arena, size, flags | M_BESTFIT, &addr))
425 		return (0);
426 
427 	rv = kmem_back_domain(domain, kernel_object, addr, size, flags);
428 	if (rv != KERN_SUCCESS) {
429 		vmem_free(arena, addr, size);
430 		return (0);
431 	}
432 	return (addr);
433 }
434 
435 vm_offset_t
436 kmem_malloc(vm_size_t size, int flags)
437 {
438 
439 	return (kmem_malloc_domainset(DOMAINSET_RR(), size, flags));
440 }
441 
442 vm_offset_t
443 kmem_malloc_domainset(struct domainset *ds, vm_size_t size, int flags)
444 {
445 	struct vm_domainset_iter di;
446 	vm_offset_t addr;
447 	int domain;
448 
449 	vm_domainset_iter_policy_init(&di, ds, &domain, &flags);
450 	do {
451 		addr = kmem_malloc_domain(domain, size, flags);
452 		if (addr != 0)
453 			break;
454 	} while (vm_domainset_iter_policy(&di, &domain) == 0);
455 
456 	return (addr);
457 }
458 
459 /*
460  *	kmem_back_domain:
461  *
462  *	Allocate physical pages from the specified domain for the specified
463  *	virtual address range.
464  */
465 int
466 kmem_back_domain(int domain, vm_object_t object, vm_offset_t addr,
467     vm_size_t size, int flags)
468 {
469 	vm_offset_t offset, i;
470 	vm_page_t m, mpred;
471 	vm_prot_t prot;
472 	int pflags;
473 
474 	KASSERT(object == kernel_object,
475 	    ("kmem_back_domain: only supports kernel object."));
476 
477 	offset = addr - VM_MIN_KERNEL_ADDRESS;
478 	pflags = malloc2vm_flags(flags) | VM_ALLOC_WIRED;
479 	pflags &= ~(VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL);
480 	if (flags & M_WAITOK)
481 		pflags |= VM_ALLOC_WAITFAIL;
482 	prot = (flags & M_EXEC) != 0 ? VM_PROT_ALL : VM_PROT_RW;
483 
484 	i = 0;
485 	VM_OBJECT_WLOCK(object);
486 retry:
487 	mpred = vm_radix_lookup_le(&object->rtree, atop(offset + i));
488 	for (; i < size; i += PAGE_SIZE, mpred = m) {
489 		m = vm_page_alloc_domain_after(object, atop(offset + i),
490 		    domain, pflags, mpred);
491 
492 		/*
493 		 * Ran out of space, free everything up and return. Don't need
494 		 * to lock page queues here as we know that the pages we got
495 		 * aren't on any queues.
496 		 */
497 		if (m == NULL) {
498 			if ((flags & M_NOWAIT) == 0)
499 				goto retry;
500 			VM_OBJECT_WUNLOCK(object);
501 			kmem_unback(object, addr, i);
502 			return (KERN_NO_SPACE);
503 		}
504 		KASSERT(vm_page_domain(m) == domain,
505 		    ("kmem_back_domain: Domain mismatch %d != %d",
506 		    vm_page_domain(m), domain));
507 		if (flags & M_ZERO && (m->flags & PG_ZERO) == 0)
508 			pmap_zero_page(m);
509 		KASSERT((m->oflags & VPO_UNMANAGED) != 0,
510 		    ("kmem_malloc: page %p is managed", m));
511 		vm_page_valid(m);
512 		pmap_enter(kernel_pmap, addr + i, m, prot,
513 		    prot | PMAP_ENTER_WIRED, 0);
514 		if (__predict_false((prot & VM_PROT_EXECUTE) != 0))
515 			m->oflags |= VPO_KMEM_EXEC;
516 	}
517 	VM_OBJECT_WUNLOCK(object);
518 
519 	return (KERN_SUCCESS);
520 }
521 
522 /*
523  *	kmem_back:
524  *
525  *	Allocate physical pages for the specified virtual address range.
526  */
527 int
528 kmem_back(vm_object_t object, vm_offset_t addr, vm_size_t size, int flags)
529 {
530 	vm_offset_t end, next, start;
531 	int domain, rv;
532 
533 	KASSERT(object == kernel_object,
534 	    ("kmem_back: only supports kernel object."));
535 
536 	for (start = addr, end = addr + size; addr < end; addr = next) {
537 		/*
538 		 * We must ensure that pages backing a given large virtual page
539 		 * all come from the same physical domain.
540 		 */
541 		if (vm_ndomains > 1) {
542 			domain = (addr >> KVA_QUANTUM_SHIFT) % vm_ndomains;
543 			while (VM_DOMAIN_EMPTY(domain))
544 				domain++;
545 			next = roundup2(addr + 1, KVA_QUANTUM);
546 			if (next > end || next < start)
547 				next = end;
548 		} else {
549 			domain = 0;
550 			next = end;
551 		}
552 		rv = kmem_back_domain(domain, object, addr, next - addr, flags);
553 		if (rv != KERN_SUCCESS) {
554 			kmem_unback(object, start, addr - start);
555 			break;
556 		}
557 	}
558 	return (rv);
559 }
560 
561 /*
562  *	kmem_unback:
563  *
564  *	Unmap and free the physical pages underlying the specified virtual
565  *	address range.
566  *
567  *	A physical page must exist within the specified object at each index
568  *	that is being unmapped.
569  */
570 static struct vmem *
571 _kmem_unback(vm_object_t object, vm_offset_t addr, vm_size_t size)
572 {
573 	struct vmem *arena;
574 	vm_page_t m, next;
575 	vm_offset_t end, offset;
576 	int domain;
577 
578 	KASSERT(object == kernel_object,
579 	    ("kmem_unback: only supports kernel object."));
580 
581 	if (size == 0)
582 		return (NULL);
583 	pmap_remove(kernel_pmap, addr, addr + size);
584 	offset = addr - VM_MIN_KERNEL_ADDRESS;
585 	end = offset + size;
586 	VM_OBJECT_WLOCK(object);
587 	m = vm_page_lookup(object, atop(offset));
588 	domain = vm_page_domain(m);
589 	if (__predict_true((m->oflags & VPO_KMEM_EXEC) == 0))
590 		arena = vm_dom[domain].vmd_kernel_arena;
591 	else
592 		arena = vm_dom[domain].vmd_kernel_rwx_arena;
593 	for (; offset < end; offset += PAGE_SIZE, m = next) {
594 		next = vm_page_next(m);
595 		vm_page_xbusy_claim(m);
596 		vm_page_unwire_noq(m);
597 		vm_page_free(m);
598 	}
599 	VM_OBJECT_WUNLOCK(object);
600 
601 	return (arena);
602 }
603 
604 void
605 kmem_unback(vm_object_t object, vm_offset_t addr, vm_size_t size)
606 {
607 
608 	(void)_kmem_unback(object, addr, size);
609 }
610 
611 /*
612  *	kmem_free:
613  *
614  *	Free memory allocated with kmem_malloc.  The size must match the
615  *	original allocation.
616  */
617 void
618 kmem_free(vm_offset_t addr, vm_size_t size)
619 {
620 	struct vmem *arena;
621 
622 	size = round_page(size);
623 	arena = _kmem_unback(kernel_object, addr, size);
624 	if (arena != NULL)
625 		vmem_free(arena, addr, size);
626 }
627 
628 /*
629  *	kmap_alloc_wait:
630  *
631  *	Allocates pageable memory from a sub-map of the kernel.  If the submap
632  *	has no room, the caller sleeps waiting for more memory in the submap.
633  *
634  *	This routine may block.
635  */
636 vm_offset_t
637 kmap_alloc_wait(vm_map_t map, vm_size_t size)
638 {
639 	vm_offset_t addr;
640 
641 	size = round_page(size);
642 	if (!swap_reserve(size))
643 		return (0);
644 
645 	for (;;) {
646 		/*
647 		 * To make this work for more than one map, use the map's lock
648 		 * to lock out sleepers/wakers.
649 		 */
650 		vm_map_lock(map);
651 		addr = vm_map_findspace(map, vm_map_min(map), size);
652 		if (addr + size <= vm_map_max(map))
653 			break;
654 		/* no space now; see if we can ever get space */
655 		if (vm_map_max(map) - vm_map_min(map) < size) {
656 			vm_map_unlock(map);
657 			swap_release(size);
658 			return (0);
659 		}
660 		map->needs_wakeup = TRUE;
661 		vm_map_unlock_and_wait(map, 0);
662 	}
663 	vm_map_insert(map, NULL, 0, addr, addr + size, VM_PROT_RW, VM_PROT_RW,
664 	    MAP_ACC_CHARGED);
665 	vm_map_unlock(map);
666 	return (addr);
667 }
668 
669 /*
670  *	kmap_free_wakeup:
671  *
672  *	Returns memory to a submap of the kernel, and wakes up any processes
673  *	waiting for memory in that map.
674  */
675 void
676 kmap_free_wakeup(vm_map_t map, vm_offset_t addr, vm_size_t size)
677 {
678 
679 	vm_map_lock(map);
680 	(void) vm_map_delete(map, trunc_page(addr), round_page(addr + size));
681 	if (map->needs_wakeup) {
682 		map->needs_wakeup = FALSE;
683 		vm_map_wakeup(map);
684 	}
685 	vm_map_unlock(map);
686 }
687 
688 void
689 kmem_init_zero_region(void)
690 {
691 	vm_offset_t addr, i;
692 	vm_page_t m;
693 
694 	/*
695 	 * Map a single physical page of zeros to a larger virtual range.
696 	 * This requires less looping in places that want large amounts of
697 	 * zeros, while not using much more physical resources.
698 	 */
699 	addr = kva_alloc(ZERO_REGION_SIZE);
700 	m = vm_page_alloc(NULL, 0, VM_ALLOC_NORMAL |
701 	    VM_ALLOC_NOOBJ | VM_ALLOC_WIRED | VM_ALLOC_ZERO);
702 	if ((m->flags & PG_ZERO) == 0)
703 		pmap_zero_page(m);
704 	for (i = 0; i < ZERO_REGION_SIZE; i += PAGE_SIZE)
705 		pmap_qenter(addr + i, &m, 1);
706 	pmap_protect(kernel_pmap, addr, addr + ZERO_REGION_SIZE, VM_PROT_READ);
707 
708 	zero_region = (const void *)addr;
709 }
710 
711 /*
712  * Import KVA from the kernel map into the kernel arena.
713  */
714 static int
715 kva_import(void *unused, vmem_size_t size, int flags, vmem_addr_t *addrp)
716 {
717 	vm_offset_t addr;
718 	int result;
719 
720 	KASSERT((size % KVA_QUANTUM) == 0,
721 	    ("kva_import: Size %jd is not a multiple of %d",
722 	    (intmax_t)size, (int)KVA_QUANTUM));
723 	addr = vm_map_min(kernel_map);
724 	result = vm_map_find(kernel_map, NULL, 0, &addr, size, 0,
725 	    VMFS_SUPER_SPACE, VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT);
726 	if (result != KERN_SUCCESS)
727                 return (ENOMEM);
728 
729 	*addrp = addr;
730 
731 	return (0);
732 }
733 
734 /*
735  * Import KVA from a parent arena into a per-domain arena.  Imports must be
736  * KVA_QUANTUM-aligned and a multiple of KVA_QUANTUM in size.
737  */
738 static int
739 kva_import_domain(void *arena, vmem_size_t size, int flags, vmem_addr_t *addrp)
740 {
741 
742 	KASSERT((size % KVA_QUANTUM) == 0,
743 	    ("kva_import_domain: Size %jd is not a multiple of %d",
744 	    (intmax_t)size, (int)KVA_QUANTUM));
745 	return (vmem_xalloc(arena, size, KVA_QUANTUM, 0, 0, VMEM_ADDR_MIN,
746 	    VMEM_ADDR_MAX, flags, addrp));
747 }
748 
749 /*
750  * 	kmem_init:
751  *
752  *	Create the kernel map; insert a mapping covering kernel text,
753  *	data, bss, and all space allocated thus far (`boostrap' data).  The
754  *	new map will thus map the range between VM_MIN_KERNEL_ADDRESS and
755  *	`start' as allocated, and the range between `start' and `end' as free.
756  *	Create the kernel vmem arena and its per-domain children.
757  */
758 void
759 kmem_init(vm_offset_t start, vm_offset_t end)
760 {
761 	vm_size_t quantum;
762 	int domain;
763 
764 	vm_map_init(kernel_map, kernel_pmap, VM_MIN_KERNEL_ADDRESS, end);
765 	kernel_map->system_map = 1;
766 	vm_map_lock(kernel_map);
767 	/* N.B.: cannot use kgdb to debug, starting with this assignment ... */
768 	(void)vm_map_insert(kernel_map, 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(kernel_map, 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(kernel_map);
789 
790 	/*
791 	 * Use a large import quantum on NUMA systems.  This helps minimize
792 	 * interleaving of superpages, reducing internal fragmentation within
793 	 * the per-domain arenas.
794 	 */
795 	if (vm_ndomains > 1 && PMAP_HAS_DMAP)
796 		quantum = KVA_NUMA_IMPORT_QUANTUM;
797 	else
798 		quantum = KVA_QUANTUM;
799 
800 	/*
801 	 * Initialize the kernel_arena.  This can grow on demand.
802 	 */
803 	vmem_init(kernel_arena, "kernel arena", 0, 0, PAGE_SIZE, 0, 0);
804 	vmem_set_import(kernel_arena, kva_import, NULL, NULL, quantum);
805 
806 	for (domain = 0; domain < vm_ndomains; domain++) {
807 		/*
808 		 * Initialize the per-domain arenas.  These are used to color
809 		 * the KVA space in a way that ensures that virtual large pages
810 		 * are backed by memory from the same physical domain,
811 		 * maximizing the potential for superpage promotion.
812 		 */
813 		vm_dom[domain].vmd_kernel_arena = vmem_create(
814 		    "kernel arena domain", 0, 0, PAGE_SIZE, 0, M_WAITOK);
815 		vmem_set_import(vm_dom[domain].vmd_kernel_arena,
816 		    kva_import_domain, NULL, kernel_arena, quantum);
817 
818 		/*
819 		 * In architectures with superpages, maintain separate arenas
820 		 * for allocations with permissions that differ from the
821 		 * "standard" read/write permissions used for kernel memory,
822 		 * so as not to inhibit superpage promotion.
823 		 *
824 		 * Use the base import quantum since this arena is rarely used.
825 		 */
826 #if VM_NRESERVLEVEL > 0
827 		vm_dom[domain].vmd_kernel_rwx_arena = vmem_create(
828 		    "kernel rwx arena domain", 0, 0, PAGE_SIZE, 0, M_WAITOK);
829 		vmem_set_import(vm_dom[domain].vmd_kernel_rwx_arena,
830 		    kva_import_domain, (vmem_release_t *)vmem_xfree,
831 		    kernel_arena, KVA_QUANTUM);
832 #else
833 		vm_dom[domain].vmd_kernel_rwx_arena =
834 		    vm_dom[domain].vmd_kernel_arena;
835 #endif
836 	}
837 
838 	/*
839 	 * This must be the very first call so that the virtual address
840 	 * space used for early allocations is properly marked used in
841 	 * the map.
842 	 */
843 	uma_startup2();
844 }
845 
846 /*
847  *	kmem_bootstrap_free:
848  *
849  *	Free pages backing preloaded data (e.g., kernel modules) to the
850  *	system.  Currently only supported on platforms that create a
851  *	vm_phys segment for preloaded data.
852  */
853 void
854 kmem_bootstrap_free(vm_offset_t start, vm_size_t size)
855 {
856 #if defined(__i386__) || defined(__amd64__)
857 	struct vm_domain *vmd;
858 	vm_offset_t end, va;
859 	vm_paddr_t pa;
860 	vm_page_t m;
861 
862 	end = trunc_page(start + size);
863 	start = round_page(start);
864 
865 #ifdef __amd64__
866 	/*
867 	 * Preloaded files do not have execute permissions by default on amd64.
868 	 * Restore the default permissions to ensure that the direct map alias
869 	 * is updated.
870 	 */
871 	pmap_change_prot(start, end - start, VM_PROT_RW);
872 #endif
873 	for (va = start; va < end; va += PAGE_SIZE) {
874 		pa = pmap_kextract(va);
875 		m = PHYS_TO_VM_PAGE(pa);
876 
877 		vmd = vm_pagequeue_domain(m);
878 		vm_domain_free_lock(vmd);
879 		vm_phys_free_pages(m, 0);
880 		vm_domain_free_unlock(vmd);
881 
882 		vm_domain_freecnt_inc(vmd, 1);
883 		vm_cnt.v_page_count++;
884 	}
885 	pmap_remove(kernel_pmap, start, end);
886 	(void)vmem_add(kernel_arena, start, end - start, M_WAITOK);
887 #endif
888 }
889 
890 /*
891  * Allow userspace to directly trigger the VM drain routine for testing
892  * purposes.
893  */
894 static int
895 debug_vm_lowmem(SYSCTL_HANDLER_ARGS)
896 {
897 	int error, i;
898 
899 	i = 0;
900 	error = sysctl_handle_int(oidp, &i, 0, req);
901 	if (error)
902 		return (error);
903 	if ((i & ~(VM_LOW_KMEM | VM_LOW_PAGES)) != 0)
904 		return (EINVAL);
905 	if (i != 0)
906 		EVENTHANDLER_INVOKE(vm_lowmem, i);
907 	return (0);
908 }
909 
910 SYSCTL_PROC(_debug, OID_AUTO, vm_lowmem, CTLTYPE_INT | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, 0,
911     debug_vm_lowmem, "I", "set to trigger vm_lowmem event with given flags");
912