xref: /freebsd/sys/vm/vm_kern.c (revision 145992504973bd16cf3518af9ba5ce185fefa82a)
1 /*-
2  * Copyright (c) 1991, 1993
3  *	The Regents of the University of California.  All rights reserved.
4  *
5  * This code is derived from software contributed to Berkeley by
6  * The Mach Operating System project at Carnegie-Mellon University.
7  *
8  * Redistribution and use in source and binary forms, with or without
9  * modification, are permitted provided that the following conditions
10  * are met:
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in the
15  *    documentation and/or other materials provided with the distribution.
16  * 4. Neither the name of the University nor the names of its contributors
17  *    may be used to endorse or promote products derived from this software
18  *    without specific prior written permission.
19  *
20  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
21  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
22  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
23  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
24  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
25  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
26  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
27  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
28  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
29  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
30  * SUCH DAMAGE.
31  *
32  *	from: @(#)vm_kern.c	8.3 (Berkeley) 1/12/94
33  *
34  *
35  * Copyright (c) 1987, 1990 Carnegie-Mellon University.
36  * All rights reserved.
37  *
38  * Authors: Avadis Tevanian, Jr., Michael Wayne Young
39  *
40  * Permission to use, copy, modify and distribute this software and
41  * its documentation is hereby granted, provided that both the copyright
42  * notice and this permission notice appear in all copies of the
43  * software, derivative works or modified versions, and any portions
44  * thereof, and that both notices appear in supporting documentation.
45  *
46  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
47  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
48  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
49  *
50  * Carnegie Mellon requests users of this software to return to
51  *
52  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
53  *  School of Computer Science
54  *  Carnegie Mellon University
55  *  Pittsburgh PA 15213-3890
56  *
57  * any improvements or extensions that they make and grant Carnegie the
58  * rights to redistribute these changes.
59  */
60 
61 /*
62  *	Kernel memory management.
63  */
64 
65 #include <sys/cdefs.h>
66 __FBSDID("$FreeBSD$");
67 
68 #include <sys/param.h>
69 #include <sys/systm.h>
70 #include <sys/kernel.h>		/* for ticks and hz */
71 #include <sys/eventhandler.h>
72 #include <sys/lock.h>
73 #include <sys/mutex.h>
74 #include <sys/proc.h>
75 #include <sys/malloc.h>
76 #include <sys/sysctl.h>
77 
78 #include <vm/vm.h>
79 #include <vm/vm_param.h>
80 #include <vm/pmap.h>
81 #include <vm/vm_map.h>
82 #include <vm/vm_object.h>
83 #include <vm/vm_page.h>
84 #include <vm/vm_pageout.h>
85 #include <vm/vm_extern.h>
86 #include <vm/uma.h>
87 
88 vm_map_t kernel_map=0;
89 vm_map_t kmem_map=0;
90 vm_map_t exec_map=0;
91 vm_map_t pipe_map;
92 vm_map_t buffer_map=0;
93 
94 const void *zero_region;
95 CTASSERT((ZERO_REGION_SIZE & PAGE_MASK) == 0);
96 
97 /*
98  *	kmem_alloc_nofault:
99  *
100  *	Allocate a virtual address range with no underlying object and
101  *	no initial mapping to physical memory.  Any mapping from this
102  *	range to physical memory must be explicitly created prior to
103  *	its use, typically with pmap_qenter().  Any attempt to create
104  *	a mapping on demand through vm_fault() will result in a panic.
105  */
106 vm_offset_t
107 kmem_alloc_nofault(map, size)
108 	vm_map_t map;
109 	vm_size_t size;
110 {
111 	vm_offset_t addr;
112 	int result;
113 
114 	size = round_page(size);
115 	addr = vm_map_min(map);
116 	result = vm_map_find(map, NULL, 0, &addr, size, VMFS_ANY_SPACE,
117 	    VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT);
118 	if (result != KERN_SUCCESS) {
119 		return (0);
120 	}
121 	return (addr);
122 }
123 
124 /*
125  *	kmem_alloc_nofault_space:
126  *
127  *	Allocate a virtual address range with no underlying object and
128  *	no initial mapping to physical memory within the specified
129  *	address space.  Any mapping from this range to physical memory
130  *	must be explicitly created prior to its use, typically with
131  *	pmap_qenter().  Any attempt to create a mapping on demand
132  *	through vm_fault() will result in a panic.
133  */
134 vm_offset_t
135 kmem_alloc_nofault_space(map, size, find_space)
136 	vm_map_t map;
137 	vm_size_t size;
138 	int find_space;
139 {
140 	vm_offset_t addr;
141 	int result;
142 
143 	size = round_page(size);
144 	addr = vm_map_min(map);
145 	result = vm_map_find(map, NULL, 0, &addr, size, find_space,
146 	    VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT);
147 	if (result != KERN_SUCCESS) {
148 		return (0);
149 	}
150 	return (addr);
151 }
152 
153 /*
154  *	Allocate wired-down memory in the kernel's address map
155  *	or a submap.
156  */
157 vm_offset_t
158 kmem_alloc(map, size)
159 	vm_map_t map;
160 	vm_size_t size;
161 {
162 	vm_offset_t addr;
163 	vm_offset_t offset;
164 
165 	size = round_page(size);
166 
167 	/*
168 	 * Use the kernel object for wired-down kernel pages. Assume that no
169 	 * region of the kernel object is referenced more than once.
170 	 */
171 
172 	/*
173 	 * Locate sufficient space in the map.  This will give us the final
174 	 * virtual address for the new memory, and thus will tell us the
175 	 * offset within the kernel map.
176 	 */
177 	vm_map_lock(map);
178 	if (vm_map_findspace(map, vm_map_min(map), size, &addr)) {
179 		vm_map_unlock(map);
180 		return (0);
181 	}
182 	offset = addr - VM_MIN_KERNEL_ADDRESS;
183 	vm_object_reference(kernel_object);
184 	vm_map_insert(map, kernel_object, offset, addr, addr + size,
185 		VM_PROT_ALL, VM_PROT_ALL, 0);
186 	vm_map_unlock(map);
187 
188 	/*
189 	 * And finally, mark the data as non-pageable.
190 	 */
191 	(void) vm_map_wire(map, addr, addr + size,
192 	    VM_MAP_WIRE_SYSTEM|VM_MAP_WIRE_NOHOLES);
193 
194 	return (addr);
195 }
196 
197 /*
198  *	Allocates a region from the kernel address map and physical pages
199  *	within the specified address range to the kernel object.  Creates a
200  *	wired mapping from this region to these pages, and returns the
201  *	region's starting virtual address.  The allocated pages are not
202  *	necessarily physically contiguous.  If M_ZERO is specified through the
203  *	given flags, then the pages are zeroed before they are mapped.
204  */
205 vm_offset_t
206 kmem_alloc_attr(vm_map_t map, vm_size_t size, int flags, vm_paddr_t low,
207     vm_paddr_t high, vm_memattr_t memattr)
208 {
209 	vm_object_t object = kernel_object;
210 	vm_offset_t addr;
211 	vm_ooffset_t end_offset, offset;
212 	vm_page_t m;
213 	int pflags, tries;
214 
215 	size = round_page(size);
216 	vm_map_lock(map);
217 	if (vm_map_findspace(map, vm_map_min(map), size, &addr)) {
218 		vm_map_unlock(map);
219 		return (0);
220 	}
221 	offset = addr - VM_MIN_KERNEL_ADDRESS;
222 	vm_object_reference(object);
223 	vm_map_insert(map, object, offset, addr, addr + size, VM_PROT_ALL,
224 	    VM_PROT_ALL, 0);
225 	if ((flags & (M_NOWAIT | M_USE_RESERVE)) == M_NOWAIT)
226 		pflags = VM_ALLOC_INTERRUPT | VM_ALLOC_NOBUSY;
227 	else
228 		pflags = VM_ALLOC_SYSTEM | VM_ALLOC_NOBUSY;
229 	if (flags & M_ZERO)
230 		pflags |= VM_ALLOC_ZERO;
231 	VM_OBJECT_LOCK(object);
232 	end_offset = offset + size;
233 	for (; offset < end_offset; offset += PAGE_SIZE) {
234 		tries = 0;
235 retry:
236 		m = vm_page_alloc_contig(object, OFF_TO_IDX(offset), pflags, 1,
237 		    low, high, PAGE_SIZE, 0, memattr);
238 		if (m == NULL) {
239 			VM_OBJECT_UNLOCK(object);
240 			if (tries < ((flags & M_NOWAIT) != 0 ? 1 : 3)) {
241 				vm_map_unlock(map);
242 				vm_pageout_grow_cache(tries, low, high);
243 				vm_map_lock(map);
244 				VM_OBJECT_LOCK(object);
245 				tries++;
246 				goto retry;
247 			}
248 
249 			/*
250 			 * Since the pages that were allocated by any previous
251 			 * iterations of this loop are not busy, they can be
252 			 * freed by vm_object_page_remove(), which is called
253 			 * by vm_map_delete().
254 			 */
255 			vm_map_delete(map, addr, addr + size);
256 			vm_map_unlock(map);
257 			return (0);
258 		}
259 		if ((flags & M_ZERO) && (m->flags & PG_ZERO) == 0)
260 			pmap_zero_page(m);
261 		m->valid = VM_PAGE_BITS_ALL;
262 	}
263 	VM_OBJECT_UNLOCK(object);
264 	vm_map_unlock(map);
265 	vm_map_wire(map, addr, addr + size, VM_MAP_WIRE_SYSTEM |
266 	    VM_MAP_WIRE_NOHOLES);
267 	return (addr);
268 }
269 
270 /*
271  *	Allocates a region from the kernel address map and physically
272  *	contiguous pages within the specified address range to the kernel
273  *	object.  Creates a wired mapping from this region to these pages, and
274  *	returns the region's starting virtual address.  If M_ZERO is specified
275  *	through the given flags, then the pages are zeroed before they are
276  *	mapped.
277  */
278 vm_offset_t
279 kmem_alloc_contig(vm_map_t map, vm_size_t size, int flags, vm_paddr_t low,
280     vm_paddr_t high, u_long alignment, vm_paddr_t boundary,
281     vm_memattr_t memattr)
282 {
283 	vm_object_t object = kernel_object;
284 	vm_offset_t addr;
285 	vm_ooffset_t offset;
286 	vm_page_t end_m, m;
287 	int pflags, tries;
288 
289 	size = round_page(size);
290 	vm_map_lock(map);
291 	if (vm_map_findspace(map, vm_map_min(map), size, &addr)) {
292 		vm_map_unlock(map);
293 		return (0);
294 	}
295 	offset = addr - VM_MIN_KERNEL_ADDRESS;
296 	vm_object_reference(object);
297 	vm_map_insert(map, object, offset, addr, addr + size, VM_PROT_ALL,
298 	    VM_PROT_ALL, 0);
299 	if ((flags & (M_NOWAIT | M_USE_RESERVE)) == M_NOWAIT)
300 		pflags = VM_ALLOC_INTERRUPT | VM_ALLOC_NOBUSY;
301 	else
302 		pflags = VM_ALLOC_SYSTEM | VM_ALLOC_NOBUSY;
303 	if (flags & M_ZERO)
304 		pflags |= VM_ALLOC_ZERO;
305 	if (flags & M_NODUMP)
306 		pflags |= VM_ALLOC_NODUMP;
307 	VM_OBJECT_LOCK(object);
308 	tries = 0;
309 retry:
310 	m = vm_page_alloc_contig(object, OFF_TO_IDX(offset), pflags,
311 	    atop(size), low, high, alignment, boundary, memattr);
312 	if (m == NULL) {
313 		VM_OBJECT_UNLOCK(object);
314 		if (tries < ((flags & M_NOWAIT) != 0 ? 1 : 3)) {
315 			vm_map_unlock(map);
316 			vm_pageout_grow_cache(tries, low, high);
317 			vm_map_lock(map);
318 			VM_OBJECT_LOCK(object);
319 			tries++;
320 			goto retry;
321 		}
322 		vm_map_delete(map, addr, addr + size);
323 		vm_map_unlock(map);
324 		return (0);
325 	}
326 	end_m = m + atop(size);
327 	for (; m < end_m; m++) {
328 		if ((flags & M_ZERO) && (m->flags & PG_ZERO) == 0)
329 			pmap_zero_page(m);
330 		m->valid = VM_PAGE_BITS_ALL;
331 	}
332 	VM_OBJECT_UNLOCK(object);
333 	vm_map_unlock(map);
334 	vm_map_wire(map, addr, addr + size, VM_MAP_WIRE_SYSTEM |
335 	    VM_MAP_WIRE_NOHOLES);
336 	return (addr);
337 }
338 
339 /*
340  *	kmem_free:
341  *
342  *	Release a region of kernel virtual memory allocated
343  *	with kmem_alloc, and return the physical pages
344  *	associated with that region.
345  *
346  *	This routine may not block on kernel maps.
347  */
348 void
349 kmem_free(map, addr, size)
350 	vm_map_t map;
351 	vm_offset_t addr;
352 	vm_size_t size;
353 {
354 
355 	(void) vm_map_remove(map, trunc_page(addr), round_page(addr + size));
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, superpage_align ?
382 	    VMFS_ALIGNED_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:
397  *
398  * 	Allocate wired-down memory in the kernel's address map for the higher
399  * 	level kernel memory allocator (kern/kern_malloc.c).  We cannot use
400  * 	kmem_alloc() because we may need to allocate memory at interrupt
401  * 	level where we cannot block (canwait == FALSE).
402  *
403  * 	This routine has its own private kernel submap (kmem_map) and object
404  * 	(kmem_object).  This, combined with the fact that only malloc uses
405  * 	this routine, ensures that we will never block in map or object waits.
406  *
407  * 	We don't worry about expanding the map (adding entries) since entries
408  * 	for wired maps are statically allocated.
409  *
410  *	`map' is ONLY allowed to be kmem_map or one of the mbuf submaps to
411  *	which we never free.
412  */
413 vm_offset_t
414 kmem_malloc(map, size, flags)
415 	vm_map_t map;
416 	vm_size_t size;
417 	int flags;
418 {
419 	vm_offset_t addr;
420 	int i, rv;
421 
422 	size = round_page(size);
423 	addr = vm_map_min(map);
424 
425 	/*
426 	 * Locate sufficient space in the map.  This will give us the final
427 	 * virtual address for the new memory, and thus will tell us the
428 	 * offset within the kernel map.
429 	 */
430 	vm_map_lock(map);
431 	if (vm_map_findspace(map, vm_map_min(map), size, &addr)) {
432 		vm_map_unlock(map);
433                 if ((flags & M_NOWAIT) == 0) {
434 			for (i = 0; i < 8; i++) {
435 				EVENTHANDLER_INVOKE(vm_lowmem, 0);
436 				uma_reclaim();
437 				vm_map_lock(map);
438 				if (vm_map_findspace(map, vm_map_min(map),
439 				    size, &addr) == 0) {
440 					break;
441 				}
442 				vm_map_unlock(map);
443 				tsleep(&i, 0, "nokva", (hz / 4) * (i + 1));
444 			}
445 			if (i == 8) {
446 				panic("kmem_malloc(%ld): kmem_map too small: %ld total allocated",
447 				    (long)size, (long)map->size);
448 			}
449 		} else {
450 			return (0);
451 		}
452 	}
453 
454 	rv = kmem_back(map, addr, size, flags);
455 	vm_map_unlock(map);
456 	return (rv == KERN_SUCCESS ? addr : 0);
457 }
458 
459 /*
460  *	kmem_back:
461  *
462  *	Allocate physical pages for the specified virtual address range.
463  */
464 int
465 kmem_back(vm_map_t map, vm_offset_t addr, vm_size_t size, int flags)
466 {
467 	vm_offset_t offset, i;
468 	vm_map_entry_t entry;
469 	vm_page_t m;
470 	int pflags;
471 	boolean_t found;
472 
473 	KASSERT(vm_map_locked(map), ("kmem_back: map %p is not locked", map));
474 	offset = addr - VM_MIN_KERNEL_ADDRESS;
475 	vm_object_reference(kmem_object);
476 	vm_map_insert(map, kmem_object, offset, addr, addr + size,
477 	    VM_PROT_ALL, VM_PROT_ALL, 0);
478 
479 	/*
480 	 * Assert: vm_map_insert() will never be able to extend the
481 	 * previous entry so vm_map_lookup_entry() will find a new
482 	 * entry exactly corresponding to this address range and it
483 	 * will have wired_count == 0.
484 	 */
485 	found = vm_map_lookup_entry(map, addr, &entry);
486 	KASSERT(found && entry->start == addr && entry->end == addr + size &&
487 	    entry->wired_count == 0 && (entry->eflags & MAP_ENTRY_IN_TRANSITION)
488 	    == 0, ("kmem_back: entry not found or misaligned"));
489 
490 	if ((flags & (M_NOWAIT|M_USE_RESERVE)) == M_NOWAIT)
491 		pflags = VM_ALLOC_INTERRUPT | VM_ALLOC_WIRED;
492 	else
493 		pflags = VM_ALLOC_SYSTEM | VM_ALLOC_WIRED;
494 
495 	if (flags & M_ZERO)
496 		pflags |= VM_ALLOC_ZERO;
497 	if (flags & M_NODUMP)
498 		pflags |= VM_ALLOC_NODUMP;
499 
500 	VM_OBJECT_LOCK(kmem_object);
501 	for (i = 0; i < size; i += PAGE_SIZE) {
502 retry:
503 		m = vm_page_alloc(kmem_object, OFF_TO_IDX(offset + i), pflags);
504 
505 		/*
506 		 * Ran out of space, free everything up and return. Don't need
507 		 * to lock page queues here as we know that the pages we got
508 		 * aren't on any queues.
509 		 */
510 		if (m == NULL) {
511 			if ((flags & M_NOWAIT) == 0) {
512 				VM_OBJECT_UNLOCK(kmem_object);
513 				entry->eflags |= MAP_ENTRY_IN_TRANSITION;
514 				vm_map_unlock(map);
515 				VM_WAIT;
516 				vm_map_lock(map);
517 				KASSERT(
518 (entry->eflags & (MAP_ENTRY_IN_TRANSITION | MAP_ENTRY_NEEDS_WAKEUP)) ==
519 				    MAP_ENTRY_IN_TRANSITION,
520 				    ("kmem_back: volatile entry"));
521 				entry->eflags &= ~MAP_ENTRY_IN_TRANSITION;
522 				VM_OBJECT_LOCK(kmem_object);
523 				goto retry;
524 			}
525 			/*
526 			 * Free the pages before removing the map entry.
527 			 * They are already marked busy.  Calling
528 			 * vm_map_delete before the pages has been freed or
529 			 * unbusied will cause a deadlock.
530 			 */
531 			while (i != 0) {
532 				i -= PAGE_SIZE;
533 				m = vm_page_lookup(kmem_object,
534 						   OFF_TO_IDX(offset + i));
535 				vm_page_unwire(m, 0);
536 				vm_page_free(m);
537 			}
538 			VM_OBJECT_UNLOCK(kmem_object);
539 			vm_map_delete(map, addr, addr + size);
540 			return (KERN_NO_SPACE);
541 		}
542 		if (flags & M_ZERO && (m->flags & PG_ZERO) == 0)
543 			pmap_zero_page(m);
544 		m->valid = VM_PAGE_BITS_ALL;
545 		KASSERT((m->oflags & VPO_UNMANAGED) != 0,
546 		    ("kmem_malloc: page %p is managed", m));
547 	}
548 	VM_OBJECT_UNLOCK(kmem_object);
549 
550 	/*
551 	 * Mark map entry as non-pageable.  Repeat the assert.
552 	 */
553 	KASSERT(entry->start == addr && entry->end == addr + size &&
554 	    entry->wired_count == 0,
555 	    ("kmem_back: entry not found or misaligned after allocation"));
556 	entry->wired_count = 1;
557 
558 	/*
559 	 * At this point, the kmem_object must be unlocked because
560 	 * vm_map_simplify_entry() calls vm_object_deallocate(), which
561 	 * locks the kmem_object.
562 	 */
563 	vm_map_simplify_entry(map, entry);
564 
565 	/*
566 	 * Loop thru pages, entering them in the pmap.
567 	 */
568 	VM_OBJECT_LOCK(kmem_object);
569 	for (i = 0; i < size; i += PAGE_SIZE) {
570 		m = vm_page_lookup(kmem_object, OFF_TO_IDX(offset + i));
571 		/*
572 		 * Because this is kernel_pmap, this call will not block.
573 		 */
574 		pmap_enter(kernel_pmap, addr + i, VM_PROT_ALL, m, VM_PROT_ALL,
575 		    TRUE);
576 		vm_page_wakeup(m);
577 	}
578 	VM_OBJECT_UNLOCK(kmem_object);
579 
580 	return (KERN_SUCCESS);
581 }
582 
583 /*
584  *	kmem_alloc_wait:
585  *
586  *	Allocates pageable memory from a sub-map of the kernel.  If the submap
587  *	has no room, the caller sleeps waiting for more memory in the submap.
588  *
589  *	This routine may block.
590  */
591 vm_offset_t
592 kmem_alloc_wait(map, size)
593 	vm_map_t map;
594 	vm_size_t size;
595 {
596 	vm_offset_t addr;
597 
598 	size = round_page(size);
599 	if (!swap_reserve(size))
600 		return (0);
601 
602 	for (;;) {
603 		/*
604 		 * To make this work for more than one map, use the map's lock
605 		 * to lock out sleepers/wakers.
606 		 */
607 		vm_map_lock(map);
608 		if (vm_map_findspace(map, vm_map_min(map), size, &addr) == 0)
609 			break;
610 		/* no space now; see if we can ever get space */
611 		if (vm_map_max(map) - vm_map_min(map) < size) {
612 			vm_map_unlock(map);
613 			swap_release(size);
614 			return (0);
615 		}
616 		map->needs_wakeup = TRUE;
617 		vm_map_unlock_and_wait(map, 0);
618 	}
619 	vm_map_insert(map, NULL, 0, addr, addr + size, VM_PROT_ALL,
620 	    VM_PROT_ALL, MAP_ACC_CHARGED);
621 	vm_map_unlock(map);
622 	return (addr);
623 }
624 
625 /*
626  *	kmem_free_wakeup:
627  *
628  *	Returns memory to a submap of the kernel, and wakes up any processes
629  *	waiting for memory in that map.
630  */
631 void
632 kmem_free_wakeup(map, addr, size)
633 	vm_map_t map;
634 	vm_offset_t addr;
635 	vm_size_t size;
636 {
637 
638 	vm_map_lock(map);
639 	(void) vm_map_delete(map, trunc_page(addr), round_page(addr + size));
640 	if (map->needs_wakeup) {
641 		map->needs_wakeup = FALSE;
642 		vm_map_wakeup(map);
643 	}
644 	vm_map_unlock(map);
645 }
646 
647 static void
648 kmem_init_zero_region(void)
649 {
650 	vm_offset_t addr, i;
651 	vm_page_t m;
652 	int error;
653 
654 	/*
655 	 * Map a single physical page of zeros to a larger virtual range.
656 	 * This requires less looping in places that want large amounts of
657 	 * zeros, while not using much more physical resources.
658 	 */
659 	addr = kmem_alloc_nofault(kernel_map, ZERO_REGION_SIZE);
660 	m = vm_page_alloc(NULL, 0, VM_ALLOC_NORMAL |
661 	    VM_ALLOC_NOOBJ | VM_ALLOC_WIRED | VM_ALLOC_ZERO);
662 	if ((m->flags & PG_ZERO) == 0)
663 		pmap_zero_page(m);
664 	for (i = 0; i < ZERO_REGION_SIZE; i += PAGE_SIZE)
665 		pmap_qenter(addr + i, &m, 1);
666 	error = vm_map_protect(kernel_map, addr, addr + ZERO_REGION_SIZE,
667 	    VM_PROT_READ, TRUE);
668 	KASSERT(error == 0, ("error=%d", error));
669 
670 	zero_region = (const void *)addr;
671 }
672 
673 /*
674  * 	kmem_init:
675  *
676  *	Create the kernel map; insert a mapping covering kernel text,
677  *	data, bss, and all space allocated thus far (`boostrap' data).  The
678  *	new map will thus map the range between VM_MIN_KERNEL_ADDRESS and
679  *	`start' as allocated, and the range between `start' and `end' as free.
680  */
681 void
682 kmem_init(start, end)
683 	vm_offset_t start, end;
684 {
685 	vm_map_t m;
686 
687 	m = vm_map_create(kernel_pmap, VM_MIN_KERNEL_ADDRESS, end);
688 	m->system_map = 1;
689 	vm_map_lock(m);
690 	/* N.B.: cannot use kgdb to debug, starting with this assignment ... */
691 	kernel_map = m;
692 	(void) vm_map_insert(m, NULL, (vm_ooffset_t) 0,
693 #ifdef __amd64__
694 	    KERNBASE,
695 #else
696 	    VM_MIN_KERNEL_ADDRESS,
697 #endif
698 	    start, VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT);
699 	/* ... and ending with the completion of the above `insert' */
700 	vm_map_unlock(m);
701 
702 	kmem_init_zero_region();
703 }
704 
705 #ifdef DIAGNOSTIC
706 /*
707  * Allow userspace to directly trigger the VM drain routine for testing
708  * purposes.
709  */
710 static int
711 debug_vm_lowmem(SYSCTL_HANDLER_ARGS)
712 {
713 	int error, i;
714 
715 	i = 0;
716 	error = sysctl_handle_int(oidp, &i, 0, req);
717 	if (error)
718 		return (error);
719 	if (i)
720 		EVENTHANDLER_INVOKE(vm_lowmem, 0);
721 	return (0);
722 }
723 
724 SYSCTL_PROC(_debug, OID_AUTO, vm_lowmem, CTLTYPE_INT | CTLFLAG_RW, 0, 0,
725     debug_vm_lowmem, "I", "set to trigger vm_lowmem event");
726 #endif
727