2 * SPDX-License-Identifier: (BSD-3-Clause AND MIT-CMU)
4 * Copyright (c) 1991, 1993
5 * The Regents of the University of California. All rights reserved.
7 * This code is derived from software contributed to Berkeley by
8 * The Mach Operating System project at Carnegie-Mellon University.
10 * Redistribution and use in source and binary forms, with or without
11 * modification, are permitted provided that the following conditions
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.
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
34 * from: @(#)vm_kern.c 8.3 (Berkeley) 1/12/94
37 * Copyright (c) 1987, 1990 Carnegie-Mellon University.
38 * All rights reserved.
40 * Authors: Avadis Tevanian, Jr., Michael Wayne Young
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.
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.
52 * Carnegie Mellon requests users of this software to return to
54 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU
55 * School of Computer Science
56 * Carnegie Mellon University
57 * Pittsburgh PA 15213-3890
59 * any improvements or extensions that they make and grant Carnegie the
60 * rights to redistribute these changes.
64 * Kernel memory management.
67 #include <sys/cdefs.h>
68 __FBSDID("$FreeBSD$");
72 #include <sys/param.h>
73 #include <sys/systm.h>
75 #include <sys/domainset.h>
76 #include <sys/eventhandler.h>
77 #include <sys/kernel.h>
79 #include <sys/malloc.h>
82 #include <sys/rwlock.h>
83 #include <sys/sysctl.h>
85 #include <sys/vmmeter.h>
88 #include <vm/vm_param.h>
89 #include <vm/vm_domainset.h>
90 #include <vm/vm_kern.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>
102 struct vm_map kernel_map_store;
103 struct vm_map exec_map_store;
104 struct vm_map pipe_map_store;
106 const void *zero_region;
107 CTASSERT((ZERO_REGION_SIZE & PAGE_MASK) == 0);
109 /* NB: Used by kernel debuggers. */
110 const u_long vm_maxuser_address = VM_MAXUSER_ADDRESS;
112 u_int exec_map_entry_size;
113 u_int exec_map_entries;
115 SYSCTL_ULONG(_vm, OID_AUTO, min_kernel_address, CTLFLAG_RD,
116 SYSCTL_NULL_ULONG_PTR, VM_MIN_KERNEL_ADDRESS, "Min kernel address");
118 SYSCTL_ULONG(_vm, OID_AUTO, max_kernel_address, CTLFLAG_RD,
120 &vm_max_kernel_address, 0,
122 SYSCTL_NULL_ULONG_PTR, VM_MAX_KERNEL_ADDRESS,
124 "Max kernel address");
126 #if VM_NRESERVLEVEL > 0
127 #define KVA_QUANTUM_SHIFT (VM_LEVEL_0_ORDER + PAGE_SHIFT)
129 /* On non-superpage architectures we want large import sizes. */
130 #define KVA_QUANTUM_SHIFT (8 + PAGE_SHIFT)
132 #define KVA_QUANTUM (1ul << KVA_QUANTUM_SHIFT)
133 #define KVA_NUMA_IMPORT_QUANTUM (KVA_QUANTUM * 128)
135 extern void uma_startup2(void);
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.
147 kva_alloc(vm_size_t size)
151 size = round_page(size);
152 if (vmem_alloc(kernel_arena, size, M_BESTFIT | M_NOWAIT, &addr))
161 * Release a region of kernel virtual memory allocated
162 * with kva_alloc, and return the physical pages
163 * associated with that region.
165 * This routine may not block on kernel maps.
168 kva_free(vm_offset_t addr, vm_size_t size)
171 size = round_page(size);
172 vmem_free(kernel_arena, addr, size);
176 * Update sanitizer shadow state to reflect a new allocation. Force inlining to
177 * help make KMSAN origin tracking more precise.
179 static __always_inline void
180 kmem_alloc_san(vm_offset_t addr, vm_size_t size, vm_size_t asize, int flags)
182 if ((flags & M_ZERO) == 0) {
183 kmsan_mark((void *)addr, asize, KMSAN_STATE_UNINIT);
184 kmsan_orig((void *)addr, asize, KMSAN_TYPE_KMEM,
187 kmsan_mark((void *)addr, asize, KMSAN_STATE_INITED);
189 kasan_mark((void *)addr, size, asize, KASAN_KMEM_REDZONE);
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)
201 VM_OBJECT_ASSERT_WLOCKED(object);
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)
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);
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.
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)
236 vm_offset_t addr, i, offset;
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))
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);
255 VM_OBJECT_WUNLOCK(object);
256 kmem_unback(object, addr, i);
257 vmem_free(vmem, addr, asize);
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)
266 pmap_enter(kernel_pmap, addr + i, m, prot,
267 prot | PMAP_ENTER_WIRED, 0);
269 VM_OBJECT_WUNLOCK(object);
270 kmem_alloc_san(addr, size, asize, flags);
275 kmem_alloc_attr(vm_size_t size, int flags, vm_paddr_t low, vm_paddr_t high,
276 vm_memattr_t memattr)
279 return (kmem_alloc_attr_domainset(DOMAINSET_RR(), size, flags, low,
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)
287 struct vm_domainset_iter di;
291 vm_domainset_iter_policy_init(&di, ds, &domain, &flags);
293 addr = kmem_alloc_attr_domain(domain, size, flags, low, high,
297 } while (vm_domainset_iter_policy(&di, &domain) == 0);
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
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)
317 vm_offset_t addr, offset, tmp;
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))
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);
335 VM_OBJECT_WUNLOCK(object);
336 vmem_free(vmem, addr, asize);
339 KASSERT(vm_page_domain(m) == domain,
340 ("kmem_alloc_contig_domain: Domain mismatch %d != %d",
341 vm_page_domain(m), domain));
344 for (; m < end_m; m++) {
345 if ((flags & M_ZERO) && (m->flags & PG_ZERO) == 0)
348 pmap_enter(kernel_pmap, tmp, m, VM_PROT_RW,
349 VM_PROT_RW | PMAP_ENTER_WIRED, 0);
352 VM_OBJECT_WUNLOCK(object);
353 kmem_alloc_san(addr, size, asize, flags);
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)
362 return (kmem_alloc_contig_domainset(DOMAINSET_RR(), size, flags, low,
363 high, alignment, boundary, memattr));
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)
371 struct vm_domainset_iter di;
375 vm_domainset_iter_policy_init(&di, ds, &domain, &flags);
377 addr = kmem_alloc_contig_domain(domain, size, flags, low, high,
378 alignment, boundary, memattr);
381 } while (vm_domainset_iter_policy(&di, &domain) == 0);
389 * Initializes a map to manage a subrange
390 * of the kernel virtual address space.
392 * Arguments are as follows:
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
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)
405 size = round_page(size);
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,
411 if (ret != KERN_SUCCESS)
412 panic("kmem_subinit: bad status return of %d", ret);
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");
420 * kmem_malloc_domain:
422 * Allocate wired-down pages in the kernel's address space.
425 kmem_malloc_domain(int domain, vm_size_t size, int flags)
432 if (__predict_true((flags & M_EXEC) == 0))
433 arena = vm_dom[domain].vmd_kernel_arena;
435 arena = vm_dom[domain].vmd_kernel_rwx_arena;
436 asize = round_page(size);
437 if (vmem_alloc(arena, asize, flags | M_BESTFIT, &addr))
440 rv = kmem_back_domain(domain, kernel_object, addr, asize, flags);
441 if (rv != KERN_SUCCESS) {
442 vmem_free(arena, addr, asize);
445 kasan_mark((void *)addr, size, asize, KASAN_KMEM_REDZONE);
450 kmem_malloc(vm_size_t size, int flags)
453 return (kmem_malloc_domainset(DOMAINSET_RR(), size, flags));
457 kmem_malloc_domainset(struct domainset *ds, vm_size_t size, int flags)
459 struct vm_domainset_iter di;
463 vm_domainset_iter_policy_init(&di, ds, &domain, &flags);
465 addr = kmem_malloc_domain(domain, size, flags);
468 } while (vm_domainset_iter_policy(&di, &domain) == 0);
476 * Allocate physical pages from the specified domain for the specified
477 * virtual address range.
480 kmem_back_domain(int domain, vm_object_t object, vm_offset_t addr,
481 vm_size_t size, int flags)
483 vm_offset_t offset, i;
488 KASSERT(object == kernel_object,
489 ("kmem_back_domain: only supports kernel object."));
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;
499 VM_OBJECT_WLOCK(object);
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);
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.
512 if ((flags & M_NOWAIT) == 0)
514 VM_OBJECT_WUNLOCK(object);
515 kmem_unback(object, addr, i);
516 return (KERN_NO_SPACE);
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)
523 KASSERT((m->oflags & VPO_UNMANAGED) != 0,
524 ("kmem_malloc: page %p is managed", 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;
531 VM_OBJECT_WUNLOCK(object);
532 kmem_alloc_san(addr, size, size, flags);
533 return (KERN_SUCCESS);
539 * Allocate physical pages for the specified virtual address range.
542 kmem_back(vm_object_t object, vm_offset_t addr, vm_size_t size, int flags)
544 vm_offset_t end, next, start;
547 KASSERT(object == kernel_object,
548 ("kmem_back: only supports kernel object."));
550 for (start = addr, end = addr + size; addr < end; addr = next) {
552 * We must ensure that pages backing a given large virtual page
553 * all come from the same physical domain.
555 if (vm_ndomains > 1) {
556 domain = (addr >> KVA_QUANTUM_SHIFT) % vm_ndomains;
557 while (VM_DOMAIN_EMPTY(domain))
559 next = roundup2(addr + 1, KVA_QUANTUM);
560 if (next > end || next < start)
566 rv = kmem_back_domain(domain, object, addr, next - addr, flags);
567 if (rv != KERN_SUCCESS) {
568 kmem_unback(object, start, addr - start);
578 * Unmap and free the physical pages underlying the specified virtual
581 * A physical page must exist within the specified object at each index
582 * that is being unmapped.
585 _kmem_unback(vm_object_t object, vm_offset_t addr, vm_size_t size)
589 vm_offset_t end, offset;
592 KASSERT(object == kernel_object,
593 ("kmem_unback: only supports kernel object."));
597 pmap_remove(kernel_pmap, addr, addr + size);
598 offset = addr - VM_MIN_KERNEL_ADDRESS;
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;
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);
613 VM_OBJECT_WUNLOCK(object);
619 kmem_unback(vm_object_t object, vm_offset_t addr, vm_size_t size)
622 (void)_kmem_unback(object, addr, size);
628 * Free memory allocated with kmem_malloc. The size must match the
629 * original allocation.
632 kmem_free(vm_offset_t addr, vm_size_t size)
636 size = round_page(size);
637 kasan_mark((void *)addr, size, size, 0);
638 arena = _kmem_unback(kernel_object, addr, size);
640 vmem_free(arena, addr, size);
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.
649 * This routine may block.
652 kmap_alloc_wait(vm_map_t map, vm_size_t size)
656 size = round_page(size);
657 if (!swap_reserve(size))
662 * To make this work for more than one map, use the map's lock
663 * to lock out sleepers/wakers.
666 addr = vm_map_findspace(map, vm_map_min(map), size);
667 if (addr + size <= vm_map_max(map))
669 /* no space now; see if we can ever get space */
670 if (vm_map_max(map) - vm_map_min(map) < size) {
675 map->needs_wakeup = TRUE;
676 vm_map_unlock_and_wait(map, 0);
678 vm_map_insert(map, NULL, 0, addr, addr + size, VM_PROT_RW, VM_PROT_RW,
687 * Returns memory to a submap of the kernel, and wakes up any processes
688 * waiting for memory in that map.
691 kmap_free_wakeup(vm_map_t map, vm_offset_t addr, vm_size_t size)
695 (void) vm_map_delete(map, trunc_page(addr), round_page(addr + size));
696 if (map->needs_wakeup) {
697 map->needs_wakeup = FALSE;
704 kmem_init_zero_region(void)
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.
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);
720 zero_region = (const void *)addr;
724 * Import KVA from the kernel map into the kernel arena.
727 kva_import(void *unused, vmem_size_t size, int flags, vmem_addr_t *addrp)
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)
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.
751 kva_import_domain(void *arena, vmem_size_t size, int flags, vmem_addr_t *addrp)
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));
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.
771 kmem_init(vm_offset_t start, vm_offset_t end)
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,
784 VM_MIN_KERNEL_ADDRESS,
786 start, VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT);
787 /* ... and ending with the completion of the above `insert' */
791 * Mark KVA used for the page array as allocated. Other platforms
792 * that handle vm_page_array allocation can simply adjust virtual_avail
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);
800 vm_map_unlock(kernel_map);
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.
807 if (vm_ndomains > 1 && PMAP_HAS_DMAP)
808 quantum = KVA_NUMA_IMPORT_QUANTUM;
810 quantum = KVA_QUANTUM;
813 * Initialize the kernel_arena. This can grow on demand.
815 vmem_init(kernel_arena, "kernel arena", 0, 0, PAGE_SIZE, 0, 0);
816 vmem_set_import(kernel_arena, kva_import, NULL, NULL, quantum);
818 for (domain = 0; domain < vm_ndomains; domain++) {
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.
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);
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.
836 * Use the base import quantum since this arena is rarely used.
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);
845 vm_dom[domain].vmd_kernel_rwx_arena =
846 vm_dom[domain].vmd_kernel_arena;
851 * This must be the very first call so that the virtual address
852 * space used for early allocations is properly marked used in
859 * kmem_bootstrap_free:
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.
866 kmem_bootstrap_free(vm_offset_t start, vm_size_t size)
868 #if defined(__i386__) || defined(__amd64__)
869 struct vm_domain *vmd;
874 end = trunc_page(start + size);
875 start = round_page(start);
879 * Preloaded files do not have execute permissions by default on amd64.
880 * Restore the default permissions to ensure that the direct map alias
883 pmap_change_prot(start, end - start, VM_PROT_RW);
885 for (va = start; va < end; va += PAGE_SIZE) {
886 pa = pmap_kextract(va);
887 m = PHYS_TO_VM_PAGE(pa);
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);
894 vm_domain_freecnt_inc(vmd, 1);
895 vm_cnt.v_page_count++;
897 pmap_remove(kernel_pmap, start, end);
898 (void)vmem_add(kernel_arena, start, end - start, M_WAITOK);
903 * Allow userspace to directly trigger the VM drain routine for testing
907 debug_vm_lowmem(SYSCTL_HANDLER_ARGS)
912 error = sysctl_handle_int(oidp, &i, 0, req);
915 if ((i & ~(VM_LOW_KMEM | VM_LOW_PAGES)) != 0)
918 EVENTHANDLER_INVOKE(vm_lowmem, i);
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");
926 debug_uma_reclaim(SYSCTL_HANDLER_ARGS)
931 error = sysctl_handle_int(oidp, &i, 0, req);
932 if (error != 0 || req->newptr == NULL)
934 if (i != UMA_RECLAIM_TRIM && i != UMA_RECLAIM_DRAIN &&
935 i != UMA_RECLAIM_DRAIN_CPU)
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");
945 debug_uma_reclaim_domain(SYSCTL_HANDLER_ARGS)
947 int domain, error, request;
950 error = sysctl_handle_int(oidp, &request, 0, req);
951 if (error != 0 || req->newptr == NULL)
954 domain = request >> 4;
956 if (request != UMA_RECLAIM_TRIM && request != UMA_RECLAIM_DRAIN &&
957 request != UMA_RECLAIM_DRAIN_CPU)
959 if (domain < 0 || domain >= vm_ndomains)
961 uma_reclaim_domain(request, domain);
964 SYSCTL_PROC(_debug, OID_AUTO, uma_reclaim_domain,
965 CTLTYPE_INT | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, 0,
966 debug_uma_reclaim_domain, "I",