xref: /freebsd/sys/vm/vm_kern.c (revision 5956d97f4b3204318ceb6aa9c77bd0bc6ea87a41)
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/asan.h>
75 #include <sys/domainset.h>
76 #include <sys/eventhandler.h>
77 #include <sys/kernel.h>
78 #include <sys/lock.h>
79 #include <sys/malloc.h>
80 #include <sys/msan.h>
81 #include <sys/proc.h>
82 #include <sys/rwlock.h>
83 #include <sys/sysctl.h>
84 #include <sys/vmem.h>
85 #include <sys/vmmeter.h>
86 
87 #include <vm/vm.h>
88 #include <vm/vm_param.h>
89 #include <vm/vm_domainset.h>
90 #include <vm/vm_kern.h>
91 #include <vm/pmap.h>
92 #include <vm/vm_map.h>
93 #include <vm/vm_object.h>
94 #include <vm/vm_page.h>
95 #include <vm/vm_pageout.h>
96 #include <vm/vm_pagequeue.h>
97 #include <vm/vm_phys.h>
98 #include <vm/vm_radix.h>
99 #include <vm/vm_extern.h>
100 #include <vm/uma.h>
101 
102 struct vm_map kernel_map_store;
103 struct vm_map exec_map_store;
104 struct vm_map pipe_map_store;
105 
106 const void *zero_region;
107 CTASSERT((ZERO_REGION_SIZE & PAGE_MASK) == 0);
108 
109 /* NB: Used by kernel debuggers. */
110 const u_long vm_maxuser_address = VM_MAXUSER_ADDRESS;
111 
112 u_int exec_map_entry_size;
113 u_int exec_map_entries;
114 
115 SYSCTL_ULONG(_vm, OID_AUTO, min_kernel_address, CTLFLAG_RD,
116     SYSCTL_NULL_ULONG_PTR, VM_MIN_KERNEL_ADDRESS, "Min kernel address");
117 
118 SYSCTL_ULONG(_vm, OID_AUTO, max_kernel_address, CTLFLAG_RD,
119 #if defined(__arm__)
120     &vm_max_kernel_address, 0,
121 #else
122     SYSCTL_NULL_ULONG_PTR, VM_MAX_KERNEL_ADDRESS,
123 #endif
124     "Max kernel address");
125 
126 #if VM_NRESERVLEVEL > 0
127 #define	KVA_QUANTUM_SHIFT	(VM_LEVEL_0_ORDER + PAGE_SHIFT)
128 #else
129 /* On non-superpage architectures we want large import sizes. */
130 #define	KVA_QUANTUM_SHIFT	(8 + PAGE_SHIFT)
131 #endif
132 #define	KVA_QUANTUM		(1ul << KVA_QUANTUM_SHIFT)
133 #define	KVA_NUMA_IMPORT_QUANTUM	(KVA_QUANTUM * 128)
134 
135 extern void     uma_startup2(void);
136 
137 /*
138  *	kva_alloc:
139  *
140  *	Allocate a virtual address range with no underlying object and
141  *	no initial mapping to physical memory.  Any mapping from this
142  *	range to physical memory must be explicitly created prior to
143  *	its use, typically with pmap_qenter().  Any attempt to create
144  *	a mapping on demand through vm_fault() will result in a panic.
145  */
146 vm_offset_t
147 kva_alloc(vm_size_t size)
148 {
149 	vm_offset_t addr;
150 
151 	size = round_page(size);
152 	if (vmem_alloc(kernel_arena, size, M_BESTFIT | M_NOWAIT, &addr))
153 		return (0);
154 
155 	return (addr);
156 }
157 
158 /*
159  *	kva_free:
160  *
161  *	Release a region of kernel virtual memory allocated
162  *	with kva_alloc, and return the physical pages
163  *	associated with that region.
164  *
165  *	This routine may not block on kernel maps.
166  */
167 void
168 kva_free(vm_offset_t addr, vm_size_t size)
169 {
170 
171 	size = round_page(size);
172 	vmem_free(kernel_arena, addr, size);
173 }
174 
175 /*
176  * Update sanitizer shadow state to reflect a new allocation.  Force inlining to
177  * help make KMSAN origin tracking more precise.
178  */
179 static __always_inline void
180 kmem_alloc_san(vm_offset_t addr, vm_size_t size, vm_size_t asize, int flags)
181 {
182 	if ((flags & M_ZERO) == 0) {
183 		kmsan_mark((void *)addr, asize, KMSAN_STATE_UNINIT);
184 		kmsan_orig((void *)addr, asize, KMSAN_TYPE_KMEM,
185 		    KMSAN_RET_ADDR);
186 	} else {
187 		kmsan_mark((void *)addr, asize, KMSAN_STATE_INITED);
188 	}
189 	kasan_mark((void *)addr, size, asize, KASAN_KMEM_REDZONE);
190 }
191 
192 static vm_page_t
193 kmem_alloc_contig_pages(vm_object_t object, vm_pindex_t pindex, int domain,
194     int pflags, u_long npages, vm_paddr_t low, vm_paddr_t high,
195     u_long alignment, vm_paddr_t boundary, vm_memattr_t memattr)
196 {
197 	vm_page_t m;
198 	int tries;
199 	bool wait, reclaim;
200 
201 	VM_OBJECT_ASSERT_WLOCKED(object);
202 
203 	wait = (pflags & VM_ALLOC_WAITOK) != 0;
204 	reclaim = (pflags & VM_ALLOC_NORECLAIM) == 0;
205 	pflags &= ~(VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL);
206 	pflags |= VM_ALLOC_NOWAIT;
207 	for (tries = wait ? 3 : 1;; tries--) {
208 		m = vm_page_alloc_contig_domain(object, pindex, domain, pflags,
209 		    npages, low, high, alignment, boundary, memattr);
210 		if (m != NULL || tries == 0 || !reclaim)
211 			break;
212 
213 		VM_OBJECT_WUNLOCK(object);
214 		if (!vm_page_reclaim_contig_domain(domain, pflags, npages,
215 		    low, high, alignment, boundary) && wait)
216 			vm_wait_domain(domain);
217 		VM_OBJECT_WLOCK(object);
218 	}
219 	return (m);
220 }
221 
222 /*
223  *	Allocates a region from the kernel address map and physical pages
224  *	within the specified address range to the kernel object.  Creates a
225  *	wired mapping from this region to these pages, and returns the
226  *	region's starting virtual address.  The allocated pages are not
227  *	necessarily physically contiguous.  If M_ZERO is specified through the
228  *	given flags, then the pages are zeroed before they are mapped.
229  */
230 static vm_offset_t
231 kmem_alloc_attr_domain(int domain, vm_size_t size, int flags, vm_paddr_t low,
232     vm_paddr_t high, vm_memattr_t memattr)
233 {
234 	vmem_t *vmem;
235 	vm_object_t object;
236 	vm_offset_t addr, i, offset;
237 	vm_page_t m;
238 	vm_size_t asize;
239 	int pflags;
240 	vm_prot_t prot;
241 
242 	object = kernel_object;
243 	asize = round_page(size);
244 	vmem = vm_dom[domain].vmd_kernel_arena;
245 	if (vmem_alloc(vmem, asize, M_BESTFIT | flags, &addr))
246 		return (0);
247 	offset = addr - VM_MIN_KERNEL_ADDRESS;
248 	pflags = malloc2vm_flags(flags) | VM_ALLOC_WIRED;
249 	prot = (flags & M_EXEC) != 0 ? VM_PROT_ALL : VM_PROT_RW;
250 	VM_OBJECT_WLOCK(object);
251 	for (i = 0; i < asize; i += PAGE_SIZE) {
252 		m = kmem_alloc_contig_pages(object, atop(offset + i),
253 		    domain, pflags, 1, low, high, PAGE_SIZE, 0, memattr);
254 		if (m == NULL) {
255 			VM_OBJECT_WUNLOCK(object);
256 			kmem_unback(object, addr, i);
257 			vmem_free(vmem, addr, asize);
258 			return (0);
259 		}
260 		KASSERT(vm_page_domain(m) == domain,
261 		    ("kmem_alloc_attr_domain: Domain mismatch %d != %d",
262 		    vm_page_domain(m), domain));
263 		if ((flags & M_ZERO) && (m->flags & PG_ZERO) == 0)
264 			pmap_zero_page(m);
265 		vm_page_valid(m);
266 		pmap_enter(kernel_pmap, addr + i, m, prot,
267 		    prot | PMAP_ENTER_WIRED, 0);
268 	}
269 	VM_OBJECT_WUNLOCK(object);
270 	kmem_alloc_san(addr, size, asize, flags);
271 	return (addr);
272 }
273 
274 vm_offset_t
275 kmem_alloc_attr(vm_size_t size, int flags, vm_paddr_t low, vm_paddr_t high,
276     vm_memattr_t memattr)
277 {
278 
279 	return (kmem_alloc_attr_domainset(DOMAINSET_RR(), size, flags, low,
280 	    high, memattr));
281 }
282 
283 vm_offset_t
284 kmem_alloc_attr_domainset(struct domainset *ds, vm_size_t size, int flags,
285     vm_paddr_t low, vm_paddr_t high, vm_memattr_t memattr)
286 {
287 	struct vm_domainset_iter di;
288 	vm_offset_t addr;
289 	int domain;
290 
291 	vm_domainset_iter_policy_init(&di, ds, &domain, &flags);
292 	do {
293 		addr = kmem_alloc_attr_domain(domain, size, flags, low, high,
294 		    memattr);
295 		if (addr != 0)
296 			break;
297 	} while (vm_domainset_iter_policy(&di, &domain) == 0);
298 
299 	return (addr);
300 }
301 
302 /*
303  *	Allocates a region from the kernel address map and physically
304  *	contiguous pages within the specified address range to the kernel
305  *	object.  Creates a wired mapping from this region to these pages, and
306  *	returns the region's starting virtual address.  If M_ZERO is specified
307  *	through the given flags, then the pages are zeroed before they are
308  *	mapped.
309  */
310 static vm_offset_t
311 kmem_alloc_contig_domain(int domain, vm_size_t size, int flags, vm_paddr_t low,
312     vm_paddr_t high, u_long alignment, vm_paddr_t boundary,
313     vm_memattr_t memattr)
314 {
315 	vmem_t *vmem;
316 	vm_object_t object;
317 	vm_offset_t addr, offset, tmp;
318 	vm_page_t end_m, m;
319 	vm_size_t asize;
320 	u_long npages;
321 	int pflags;
322 
323 	object = kernel_object;
324 	asize = round_page(size);
325 	vmem = vm_dom[domain].vmd_kernel_arena;
326 	if (vmem_alloc(vmem, asize, flags | M_BESTFIT, &addr))
327 		return (0);
328 	offset = addr - VM_MIN_KERNEL_ADDRESS;
329 	pflags = malloc2vm_flags(flags) | VM_ALLOC_WIRED;
330 	npages = atop(asize);
331 	VM_OBJECT_WLOCK(object);
332 	m = kmem_alloc_contig_pages(object, atop(offset), domain,
333 	    pflags, npages, low, high, alignment, boundary, memattr);
334 	if (m == NULL) {
335 		VM_OBJECT_WUNLOCK(object);
336 		vmem_free(vmem, addr, asize);
337 		return (0);
338 	}
339 	KASSERT(vm_page_domain(m) == domain,
340 	    ("kmem_alloc_contig_domain: Domain mismatch %d != %d",
341 	    vm_page_domain(m), domain));
342 	end_m = m + npages;
343 	tmp = addr;
344 	for (; m < end_m; m++) {
345 		if ((flags & M_ZERO) && (m->flags & PG_ZERO) == 0)
346 			pmap_zero_page(m);
347 		vm_page_valid(m);
348 		pmap_enter(kernel_pmap, tmp, m, VM_PROT_RW,
349 		    VM_PROT_RW | PMAP_ENTER_WIRED, 0);
350 		tmp += PAGE_SIZE;
351 	}
352 	VM_OBJECT_WUNLOCK(object);
353 	kmem_alloc_san(addr, size, asize, flags);
354 	return (addr);
355 }
356 
357 vm_offset_t
358 kmem_alloc_contig(vm_size_t size, int flags, vm_paddr_t low, vm_paddr_t high,
359     u_long alignment, vm_paddr_t boundary, vm_memattr_t memattr)
360 {
361 
362 	return (kmem_alloc_contig_domainset(DOMAINSET_RR(), size, flags, low,
363 	    high, alignment, boundary, memattr));
364 }
365 
366 vm_offset_t
367 kmem_alloc_contig_domainset(struct domainset *ds, vm_size_t size, int flags,
368     vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary,
369     vm_memattr_t memattr)
370 {
371 	struct vm_domainset_iter di;
372 	vm_offset_t addr;
373 	int domain;
374 
375 	vm_domainset_iter_policy_init(&di, ds, &domain, &flags);
376 	do {
377 		addr = kmem_alloc_contig_domain(domain, size, flags, low, high,
378 		    alignment, boundary, memattr);
379 		if (addr != 0)
380 			break;
381 	} while (vm_domainset_iter_policy(&di, &domain) == 0);
382 
383 	return (addr);
384 }
385 
386 /*
387  *	kmem_subinit:
388  *
389  *	Initializes a map to manage a subrange
390  *	of the kernel virtual address space.
391  *
392  *	Arguments are as follows:
393  *
394  *	parent		Map to take range from
395  *	min, max	Returned endpoints of map
396  *	size		Size of range to find
397  *	superpage_align	Request that min is superpage aligned
398  */
399 void
400 kmem_subinit(vm_map_t map, vm_map_t parent, vm_offset_t *min, vm_offset_t *max,
401     vm_size_t size, bool superpage_align)
402 {
403 	int ret;
404 
405 	size = round_page(size);
406 
407 	*min = vm_map_min(parent);
408 	ret = vm_map_find(parent, NULL, 0, min, size, 0, superpage_align ?
409 	    VMFS_SUPER_SPACE : VMFS_ANY_SPACE, VM_PROT_ALL, VM_PROT_ALL,
410 	    MAP_ACC_NO_CHARGE);
411 	if (ret != KERN_SUCCESS)
412 		panic("kmem_subinit: bad status return of %d", ret);
413 	*max = *min + size;
414 	vm_map_init(map, vm_map_pmap(parent), *min, *max);
415 	if (vm_map_submap(parent, *min, *max, map) != KERN_SUCCESS)
416 		panic("kmem_subinit: unable to change range to submap");
417 }
418 
419 /*
420  *	kmem_malloc_domain:
421  *
422  *	Allocate wired-down pages in the kernel's address space.
423  */
424 static vm_offset_t
425 kmem_malloc_domain(int domain, vm_size_t size, int flags)
426 {
427 	vmem_t *arena;
428 	vm_offset_t addr;
429 	vm_size_t asize;
430 	int rv;
431 
432 	if (__predict_true((flags & M_EXEC) == 0))
433 		arena = vm_dom[domain].vmd_kernel_arena;
434 	else
435 		arena = vm_dom[domain].vmd_kernel_rwx_arena;
436 	asize = round_page(size);
437 	if (vmem_alloc(arena, asize, flags | M_BESTFIT, &addr))
438 		return (0);
439 
440 	rv = kmem_back_domain(domain, kernel_object, addr, asize, flags);
441 	if (rv != KERN_SUCCESS) {
442 		vmem_free(arena, addr, asize);
443 		return (0);
444 	}
445 	kasan_mark((void *)addr, size, asize, KASAN_KMEM_REDZONE);
446 	return (addr);
447 }
448 
449 vm_offset_t
450 kmem_malloc(vm_size_t size, int flags)
451 {
452 
453 	return (kmem_malloc_domainset(DOMAINSET_RR(), size, flags));
454 }
455 
456 vm_offset_t
457 kmem_malloc_domainset(struct domainset *ds, vm_size_t size, int flags)
458 {
459 	struct vm_domainset_iter di;
460 	vm_offset_t addr;
461 	int domain;
462 
463 	vm_domainset_iter_policy_init(&di, ds, &domain, &flags);
464 	do {
465 		addr = kmem_malloc_domain(domain, size, flags);
466 		if (addr != 0)
467 			break;
468 	} while (vm_domainset_iter_policy(&di, &domain) == 0);
469 
470 	return (addr);
471 }
472 
473 /*
474  *	kmem_back_domain:
475  *
476  *	Allocate physical pages from the specified domain for the specified
477  *	virtual address range.
478  */
479 int
480 kmem_back_domain(int domain, vm_object_t object, vm_offset_t addr,
481     vm_size_t size, int flags)
482 {
483 	vm_offset_t offset, i;
484 	vm_page_t m, mpred;
485 	vm_prot_t prot;
486 	int pflags;
487 
488 	KASSERT(object == kernel_object,
489 	    ("kmem_back_domain: only supports kernel object."));
490 
491 	offset = addr - VM_MIN_KERNEL_ADDRESS;
492 	pflags = malloc2vm_flags(flags) | VM_ALLOC_WIRED;
493 	pflags &= ~(VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL);
494 	if (flags & M_WAITOK)
495 		pflags |= VM_ALLOC_WAITFAIL;
496 	prot = (flags & M_EXEC) != 0 ? VM_PROT_ALL : VM_PROT_RW;
497 
498 	i = 0;
499 	VM_OBJECT_WLOCK(object);
500 retry:
501 	mpred = vm_radix_lookup_le(&object->rtree, atop(offset + i));
502 	for (; i < size; i += PAGE_SIZE, mpred = m) {
503 		m = vm_page_alloc_domain_after(object, atop(offset + i),
504 		    domain, pflags, mpred);
505 
506 		/*
507 		 * Ran out of space, free everything up and return. Don't need
508 		 * to lock page queues here as we know that the pages we got
509 		 * aren't on any queues.
510 		 */
511 		if (m == NULL) {
512 			if ((flags & M_NOWAIT) == 0)
513 				goto retry;
514 			VM_OBJECT_WUNLOCK(object);
515 			kmem_unback(object, addr, i);
516 			return (KERN_NO_SPACE);
517 		}
518 		KASSERT(vm_page_domain(m) == domain,
519 		    ("kmem_back_domain: Domain mismatch %d != %d",
520 		    vm_page_domain(m), domain));
521 		if (flags & M_ZERO && (m->flags & PG_ZERO) == 0)
522 			pmap_zero_page(m);
523 		KASSERT((m->oflags & VPO_UNMANAGED) != 0,
524 		    ("kmem_malloc: page %p is managed", m));
525 		vm_page_valid(m);
526 		pmap_enter(kernel_pmap, addr + i, m, prot,
527 		    prot | PMAP_ENTER_WIRED, 0);
528 		if (__predict_false((prot & VM_PROT_EXECUTE) != 0))
529 			m->oflags |= VPO_KMEM_EXEC;
530 	}
531 	VM_OBJECT_WUNLOCK(object);
532 	kmem_alloc_san(addr, size, size, flags);
533 	return (KERN_SUCCESS);
534 }
535 
536 /*
537  *	kmem_back:
538  *
539  *	Allocate physical pages for the specified virtual address range.
540  */
541 int
542 kmem_back(vm_object_t object, vm_offset_t addr, vm_size_t size, int flags)
543 {
544 	vm_offset_t end, next, start;
545 	int domain, rv;
546 
547 	KASSERT(object == kernel_object,
548 	    ("kmem_back: only supports kernel object."));
549 
550 	for (start = addr, end = addr + size; addr < end; addr = next) {
551 		/*
552 		 * We must ensure that pages backing a given large virtual page
553 		 * all come from the same physical domain.
554 		 */
555 		if (vm_ndomains > 1) {
556 			domain = (addr >> KVA_QUANTUM_SHIFT) % vm_ndomains;
557 			while (VM_DOMAIN_EMPTY(domain))
558 				domain++;
559 			next = roundup2(addr + 1, KVA_QUANTUM);
560 			if (next > end || next < start)
561 				next = end;
562 		} else {
563 			domain = 0;
564 			next = end;
565 		}
566 		rv = kmem_back_domain(domain, object, addr, next - addr, flags);
567 		if (rv != KERN_SUCCESS) {
568 			kmem_unback(object, start, addr - start);
569 			break;
570 		}
571 	}
572 	return (rv);
573 }
574 
575 /*
576  *	kmem_unback:
577  *
578  *	Unmap and free the physical pages underlying the specified virtual
579  *	address range.
580  *
581  *	A physical page must exist within the specified object at each index
582  *	that is being unmapped.
583  */
584 static struct vmem *
585 _kmem_unback(vm_object_t object, vm_offset_t addr, vm_size_t size)
586 {
587 	struct vmem *arena;
588 	vm_page_t m, next;
589 	vm_offset_t end, offset;
590 	int domain;
591 
592 	KASSERT(object == kernel_object,
593 	    ("kmem_unback: only supports kernel object."));
594 
595 	if (size == 0)
596 		return (NULL);
597 	pmap_remove(kernel_pmap, addr, addr + size);
598 	offset = addr - VM_MIN_KERNEL_ADDRESS;
599 	end = offset + size;
600 	VM_OBJECT_WLOCK(object);
601 	m = vm_page_lookup(object, atop(offset));
602 	domain = vm_page_domain(m);
603 	if (__predict_true((m->oflags & VPO_KMEM_EXEC) == 0))
604 		arena = vm_dom[domain].vmd_kernel_arena;
605 	else
606 		arena = vm_dom[domain].vmd_kernel_rwx_arena;
607 	for (; offset < end; offset += PAGE_SIZE, m = next) {
608 		next = vm_page_next(m);
609 		vm_page_xbusy_claim(m);
610 		vm_page_unwire_noq(m);
611 		vm_page_free(m);
612 	}
613 	VM_OBJECT_WUNLOCK(object);
614 
615 	return (arena);
616 }
617 
618 void
619 kmem_unback(vm_object_t object, vm_offset_t addr, vm_size_t size)
620 {
621 
622 	(void)_kmem_unback(object, addr, size);
623 }
624 
625 /*
626  *	kmem_free:
627  *
628  *	Free memory allocated with kmem_malloc.  The size must match the
629  *	original allocation.
630  */
631 void
632 kmem_free(vm_offset_t addr, vm_size_t size)
633 {
634 	struct vmem *arena;
635 
636 	size = round_page(size);
637 	kasan_mark((void *)addr, size, size, 0);
638 	arena = _kmem_unback(kernel_object, addr, size);
639 	if (arena != NULL)
640 		vmem_free(arena, addr, size);
641 }
642 
643 /*
644  *	kmap_alloc_wait:
645  *
646  *	Allocates pageable memory from a sub-map of the kernel.  If the submap
647  *	has no room, the caller sleeps waiting for more memory in the submap.
648  *
649  *	This routine may block.
650  */
651 vm_offset_t
652 kmap_alloc_wait(vm_map_t map, vm_size_t size)
653 {
654 	vm_offset_t addr;
655 
656 	size = round_page(size);
657 	if (!swap_reserve(size))
658 		return (0);
659 
660 	for (;;) {
661 		/*
662 		 * To make this work for more than one map, use the map's lock
663 		 * to lock out sleepers/wakers.
664 		 */
665 		vm_map_lock(map);
666 		addr = vm_map_findspace(map, vm_map_min(map), size);
667 		if (addr + size <= vm_map_max(map))
668 			break;
669 		/* no space now; see if we can ever get space */
670 		if (vm_map_max(map) - vm_map_min(map) < size) {
671 			vm_map_unlock(map);
672 			swap_release(size);
673 			return (0);
674 		}
675 		map->needs_wakeup = TRUE;
676 		vm_map_unlock_and_wait(map, 0);
677 	}
678 	vm_map_insert(map, NULL, 0, addr, addr + size, VM_PROT_RW, VM_PROT_RW,
679 	    MAP_ACC_CHARGED);
680 	vm_map_unlock(map);
681 	return (addr);
682 }
683 
684 /*
685  *	kmap_free_wakeup:
686  *
687  *	Returns memory to a submap of the kernel, and wakes up any processes
688  *	waiting for memory in that map.
689  */
690 void
691 kmap_free_wakeup(vm_map_t map, vm_offset_t addr, vm_size_t size)
692 {
693 
694 	vm_map_lock(map);
695 	(void) vm_map_delete(map, trunc_page(addr), round_page(addr + size));
696 	if (map->needs_wakeup) {
697 		map->needs_wakeup = FALSE;
698 		vm_map_wakeup(map);
699 	}
700 	vm_map_unlock(map);
701 }
702 
703 void
704 kmem_init_zero_region(void)
705 {
706 	vm_offset_t addr, i;
707 	vm_page_t m;
708 
709 	/*
710 	 * Map a single physical page of zeros to a larger virtual range.
711 	 * This requires less looping in places that want large amounts of
712 	 * zeros, while not using much more physical resources.
713 	 */
714 	addr = kva_alloc(ZERO_REGION_SIZE);
715 	m = vm_page_alloc_noobj(VM_ALLOC_WIRED | VM_ALLOC_ZERO);
716 	for (i = 0; i < ZERO_REGION_SIZE; i += PAGE_SIZE)
717 		pmap_qenter(addr + i, &m, 1);
718 	pmap_protect(kernel_pmap, addr, addr + ZERO_REGION_SIZE, VM_PROT_READ);
719 
720 	zero_region = (const void *)addr;
721 }
722 
723 /*
724  * Import KVA from the kernel map into the kernel arena.
725  */
726 static int
727 kva_import(void *unused, vmem_size_t size, int flags, vmem_addr_t *addrp)
728 {
729 	vm_offset_t addr;
730 	int result;
731 
732 	KASSERT((size % KVA_QUANTUM) == 0,
733 	    ("kva_import: Size %jd is not a multiple of %d",
734 	    (intmax_t)size, (int)KVA_QUANTUM));
735 	addr = vm_map_min(kernel_map);
736 	result = vm_map_find(kernel_map, NULL, 0, &addr, size, 0,
737 	    VMFS_SUPER_SPACE, VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT);
738 	if (result != KERN_SUCCESS)
739                 return (ENOMEM);
740 
741 	*addrp = addr;
742 
743 	return (0);
744 }
745 
746 /*
747  * Import KVA from a parent arena into a per-domain arena.  Imports must be
748  * KVA_QUANTUM-aligned and a multiple of KVA_QUANTUM in size.
749  */
750 static int
751 kva_import_domain(void *arena, vmem_size_t size, int flags, vmem_addr_t *addrp)
752 {
753 
754 	KASSERT((size % KVA_QUANTUM) == 0,
755 	    ("kva_import_domain: Size %jd is not a multiple of %d",
756 	    (intmax_t)size, (int)KVA_QUANTUM));
757 	return (vmem_xalloc(arena, size, KVA_QUANTUM, 0, 0, VMEM_ADDR_MIN,
758 	    VMEM_ADDR_MAX, flags, addrp));
759 }
760 
761 /*
762  * 	kmem_init:
763  *
764  *	Create the kernel map; insert a mapping covering kernel text,
765  *	data, bss, and all space allocated thus far (`boostrap' data).  The
766  *	new map will thus map the range between VM_MIN_KERNEL_ADDRESS and
767  *	`start' as allocated, and the range between `start' and `end' as free.
768  *	Create the kernel vmem arena and its per-domain children.
769  */
770 void
771 kmem_init(vm_offset_t start, vm_offset_t end)
772 {
773 	vm_size_t quantum;
774 	int domain;
775 
776 	vm_map_init(kernel_map, kernel_pmap, VM_MIN_KERNEL_ADDRESS, end);
777 	kernel_map->system_map = 1;
778 	vm_map_lock(kernel_map);
779 	/* N.B.: cannot use kgdb to debug, starting with this assignment ... */
780 	(void)vm_map_insert(kernel_map, NULL, 0,
781 #ifdef __amd64__
782 	    KERNBASE,
783 #else
784 	    VM_MIN_KERNEL_ADDRESS,
785 #endif
786 	    start, VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT);
787 	/* ... and ending with the completion of the above `insert' */
788 
789 #ifdef __amd64__
790 	/*
791 	 * Mark KVA used for the page array as allocated.  Other platforms
792 	 * that handle vm_page_array allocation can simply adjust virtual_avail
793 	 * instead.
794 	 */
795 	(void)vm_map_insert(kernel_map, NULL, 0, (vm_offset_t)vm_page_array,
796 	    (vm_offset_t)vm_page_array + round_2mpage(vm_page_array_size *
797 	    sizeof(struct vm_page)),
798 	    VM_PROT_RW, VM_PROT_RW, MAP_NOFAULT);
799 #endif
800 	vm_map_unlock(kernel_map);
801 
802 	/*
803 	 * Use a large import quantum on NUMA systems.  This helps minimize
804 	 * interleaving of superpages, reducing internal fragmentation within
805 	 * the per-domain arenas.
806 	 */
807 	if (vm_ndomains > 1 && PMAP_HAS_DMAP)
808 		quantum = KVA_NUMA_IMPORT_QUANTUM;
809 	else
810 		quantum = KVA_QUANTUM;
811 
812 	/*
813 	 * Initialize the kernel_arena.  This can grow on demand.
814 	 */
815 	vmem_init(kernel_arena, "kernel arena", 0, 0, PAGE_SIZE, 0, 0);
816 	vmem_set_import(kernel_arena, kva_import, NULL, NULL, quantum);
817 
818 	for (domain = 0; domain < vm_ndomains; domain++) {
819 		/*
820 		 * Initialize the per-domain arenas.  These are used to color
821 		 * the KVA space in a way that ensures that virtual large pages
822 		 * are backed by memory from the same physical domain,
823 		 * maximizing the potential for superpage promotion.
824 		 */
825 		vm_dom[domain].vmd_kernel_arena = vmem_create(
826 		    "kernel arena domain", 0, 0, PAGE_SIZE, 0, M_WAITOK);
827 		vmem_set_import(vm_dom[domain].vmd_kernel_arena,
828 		    kva_import_domain, NULL, kernel_arena, quantum);
829 
830 		/*
831 		 * In architectures with superpages, maintain separate arenas
832 		 * for allocations with permissions that differ from the
833 		 * "standard" read/write permissions used for kernel memory,
834 		 * so as not to inhibit superpage promotion.
835 		 *
836 		 * Use the base import quantum since this arena is rarely used.
837 		 */
838 #if VM_NRESERVLEVEL > 0
839 		vm_dom[domain].vmd_kernel_rwx_arena = vmem_create(
840 		    "kernel rwx arena domain", 0, 0, PAGE_SIZE, 0, M_WAITOK);
841 		vmem_set_import(vm_dom[domain].vmd_kernel_rwx_arena,
842 		    kva_import_domain, (vmem_release_t *)vmem_xfree,
843 		    kernel_arena, KVA_QUANTUM);
844 #else
845 		vm_dom[domain].vmd_kernel_rwx_arena =
846 		    vm_dom[domain].vmd_kernel_arena;
847 #endif
848 	}
849 
850 	/*
851 	 * This must be the very first call so that the virtual address
852 	 * space used for early allocations is properly marked used in
853 	 * the map.
854 	 */
855 	uma_startup2();
856 }
857 
858 /*
859  *	kmem_bootstrap_free:
860  *
861  *	Free pages backing preloaded data (e.g., kernel modules) to the
862  *	system.  Currently only supported on platforms that create a
863  *	vm_phys segment for preloaded data.
864  */
865 void
866 kmem_bootstrap_free(vm_offset_t start, vm_size_t size)
867 {
868 #if defined(__i386__) || defined(__amd64__)
869 	struct vm_domain *vmd;
870 	vm_offset_t end, va;
871 	vm_paddr_t pa;
872 	vm_page_t m;
873 
874 	end = trunc_page(start + size);
875 	start = round_page(start);
876 
877 #ifdef __amd64__
878 	/*
879 	 * Preloaded files do not have execute permissions by default on amd64.
880 	 * Restore the default permissions to ensure that the direct map alias
881 	 * is updated.
882 	 */
883 	pmap_change_prot(start, end - start, VM_PROT_RW);
884 #endif
885 	for (va = start; va < end; va += PAGE_SIZE) {
886 		pa = pmap_kextract(va);
887 		m = PHYS_TO_VM_PAGE(pa);
888 
889 		vmd = vm_pagequeue_domain(m);
890 		vm_domain_free_lock(vmd);
891 		vm_phys_free_pages(m, 0);
892 		vm_domain_free_unlock(vmd);
893 
894 		vm_domain_freecnt_inc(vmd, 1);
895 		vm_cnt.v_page_count++;
896 	}
897 	pmap_remove(kernel_pmap, start, end);
898 	(void)vmem_add(kernel_arena, start, end - start, M_WAITOK);
899 #endif
900 }
901 
902 /*
903  * Allow userspace to directly trigger the VM drain routine for testing
904  * purposes.
905  */
906 static int
907 debug_vm_lowmem(SYSCTL_HANDLER_ARGS)
908 {
909 	int error, i;
910 
911 	i = 0;
912 	error = sysctl_handle_int(oidp, &i, 0, req);
913 	if (error != 0)
914 		return (error);
915 	if ((i & ~(VM_LOW_KMEM | VM_LOW_PAGES)) != 0)
916 		return (EINVAL);
917 	if (i != 0)
918 		EVENTHANDLER_INVOKE(vm_lowmem, i);
919 	return (0);
920 }
921 SYSCTL_PROC(_debug, OID_AUTO, vm_lowmem,
922     CTLTYPE_INT | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, 0, debug_vm_lowmem, "I",
923     "set to trigger vm_lowmem event with given flags");
924 
925 static int
926 debug_uma_reclaim(SYSCTL_HANDLER_ARGS)
927 {
928 	int error, i;
929 
930 	i = 0;
931 	error = sysctl_handle_int(oidp, &i, 0, req);
932 	if (error != 0 || req->newptr == NULL)
933 		return (error);
934 	if (i != UMA_RECLAIM_TRIM && i != UMA_RECLAIM_DRAIN &&
935 	    i != UMA_RECLAIM_DRAIN_CPU)
936 		return (EINVAL);
937 	uma_reclaim(i);
938 	return (0);
939 }
940 SYSCTL_PROC(_debug, OID_AUTO, uma_reclaim,
941     CTLTYPE_INT | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, 0, debug_uma_reclaim, "I",
942     "set to generate request to reclaim uma caches");
943 
944 static int
945 debug_uma_reclaim_domain(SYSCTL_HANDLER_ARGS)
946 {
947 	int domain, error, request;
948 
949 	request = 0;
950 	error = sysctl_handle_int(oidp, &request, 0, req);
951 	if (error != 0 || req->newptr == NULL)
952 		return (error);
953 
954 	domain = request >> 4;
955 	request &= 0xf;
956 	if (request != UMA_RECLAIM_TRIM && request != UMA_RECLAIM_DRAIN &&
957 	    request != UMA_RECLAIM_DRAIN_CPU)
958 		return (EINVAL);
959 	if (domain < 0 || domain >= vm_ndomains)
960 		return (EINVAL);
961 	uma_reclaim_domain(request, domain);
962 	return (0);
963 }
964 SYSCTL_PROC(_debug, OID_AUTO, uma_reclaim_domain,
965     CTLTYPE_INT | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, 0,
966     debug_uma_reclaim_domain, "I",
967     "");
968