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. All advertising materials mentioning features or use of this software
19 * must display the following acknowledgement:
20 * This product includes software developed by the University of
21 * California, Berkeley and its contributors.
22 * 4. Neither the name of the University nor the names of its contributors
23 * may be used to endorse or promote products derived from this software
24 * without specific prior written permission.
26 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
27 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
28 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
29 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
30 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
31 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
32 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
33 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
34 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
35 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
38 * from: @(#)vm_map.c 8.3 (Berkeley) 1/12/94
41 * Copyright (c) 1987, 1990 Carnegie-Mellon University.
42 * All rights reserved.
44 * Authors: Avadis Tevanian, Jr., Michael Wayne Young
46 * Permission to use, copy, modify and distribute this software and
47 * its documentation is hereby granted, provided that both the copyright
48 * notice and this permission notice appear in all copies of the
49 * software, derivative works or modified versions, and any portions
50 * thereof, and that both notices appear in supporting documentation.
52 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
53 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
54 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
56 * Carnegie Mellon requests users of this software to return to
58 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU
59 * School of Computer Science
60 * Carnegie Mellon University
61 * Pittsburgh PA 15213-3890
63 * any improvements or extensions that they make and grant Carnegie the
64 * rights to redistribute these changes.
66 * $FreeBSD: src/sys/vm/vm_map.c,v 1.187.2.19 2003/05/27 00:47:02 alc Exp $
67 * $DragonFly: src/sys/vm/vm_map.c,v 1.56 2007/04/29 18:25:41 dillon Exp $
71 * Virtual memory mapping module.
74 #include <sys/param.h>
75 #include <sys/systm.h>
76 #include <sys/kernel.h>
78 #include <sys/serialize.h>
80 #include <sys/vmmeter.h>
82 #include <sys/vnode.h>
83 #include <sys/resourcevar.h>
86 #include <sys/malloc.h>
89 #include <vm/vm_param.h>
91 #include <vm/vm_map.h>
92 #include <vm/vm_page.h>
93 #include <vm/vm_object.h>
94 #include <vm/vm_pager.h>
95 #include <vm/vm_kern.h>
96 #include <vm/vm_extern.h>
97 #include <vm/swap_pager.h>
98 #include <vm/vm_zone.h>
100 #include <sys/thread2.h>
101 #include <sys/sysref2.h>
102 #include <sys/random.h>
103 #include <sys/sysctl.h>
106 * Virtual memory maps provide for the mapping, protection, and sharing
107 * of virtual memory objects. In addition, this module provides for an
108 * efficient virtual copy of memory from one map to another.
110 * Synchronization is required prior to most operations.
112 * Maps consist of an ordered doubly-linked list of simple entries.
113 * A hint and a RB tree is used to speed-up lookups.
115 * Callers looking to modify maps specify start/end addresses which cause
116 * the related map entry to be clipped if necessary, and then later
117 * recombined if the pieces remained compatible.
119 * Virtual copy operations are performed by copying VM object references
120 * from one map to another, and then marking both regions as copy-on-write.
122 static void vmspace_terminate(struct vmspace *vm);
123 static void vmspace_lock(struct vmspace *vm);
124 static void vmspace_unlock(struct vmspace *vm);
125 static void vmspace_dtor(void *obj, void *private);
127 MALLOC_DEFINE(M_VMSPACE, "vmspace", "vmspace objcache backingstore");
129 struct sysref_class vmspace_sysref_class = {
132 .proto = SYSREF_PROTO_VMSPACE,
133 .offset = offsetof(struct vmspace, vm_sysref),
134 .objsize = sizeof(struct vmspace),
136 .flags = SRC_MANAGEDINIT,
137 .dtor = vmspace_dtor,
139 .terminate = (sysref_terminate_func_t)vmspace_terminate,
140 .lock = (sysref_lock_func_t)vmspace_lock,
141 .unlock = (sysref_lock_func_t)vmspace_unlock
146 * per-cpu page table cross mappings are initialized in early boot
147 * and might require a considerable number of vm_map_entry structures.
149 #define VMEPERCPU (MAXCPU+1)
151 static struct vm_zone mapentzone_store, mapzone_store;
152 static vm_zone_t mapentzone, mapzone;
153 static struct vm_object mapentobj, mapobj;
155 static struct vm_map_entry map_entry_init[MAX_MAPENT];
156 static struct vm_map_entry cpu_map_entry_init[MAXCPU][VMEPERCPU];
157 static struct vm_map map_init[MAX_KMAP];
159 static int randomize_mmap;
160 SYSCTL_INT(_vm, OID_AUTO, randomize_mmap, CTLFLAG_RW, &randomize_mmap, 0,
161 "Randomize mmap offsets");
163 static void vm_map_entry_shadow(vm_map_entry_t entry);
164 static vm_map_entry_t vm_map_entry_create(vm_map_t map, int *);
165 static void vm_map_entry_dispose (vm_map_t map, vm_map_entry_t entry, int *);
166 static void _vm_map_clip_end (vm_map_t, vm_map_entry_t, vm_offset_t, int *);
167 static void _vm_map_clip_start (vm_map_t, vm_map_entry_t, vm_offset_t, int *);
168 static void vm_map_entry_delete (vm_map_t, vm_map_entry_t, int *);
169 static void vm_map_entry_unwire (vm_map_t, vm_map_entry_t);
170 static void vm_map_copy_entry (vm_map_t, vm_map_t, vm_map_entry_t,
172 static void vm_map_split (vm_map_entry_t);
173 static void vm_map_unclip_range (vm_map_t map, vm_map_entry_t start_entry, vm_offset_t start, vm_offset_t end, int *count, int flags);
176 * Initialize the vm_map module. Must be called before any other vm_map
179 * Map and entry structures are allocated from the general purpose
180 * memory pool with some exceptions:
182 * - The kernel map is allocated statically.
183 * - Initial kernel map entries are allocated out of a static pool.
185 * These restrictions are necessary since malloc() uses the
186 * maps and requires map entries.
188 * Called from the low level boot code only.
193 mapzone = &mapzone_store;
194 zbootinit(mapzone, "MAP", sizeof (struct vm_map),
196 mapentzone = &mapentzone_store;
197 zbootinit(mapentzone, "MAP ENTRY", sizeof (struct vm_map_entry),
198 map_entry_init, MAX_MAPENT);
202 * Called prior to any vmspace allocations.
204 * Called from the low level boot code only.
209 zinitna(mapentzone, &mapentobj, NULL, 0, 0,
210 ZONE_USE_RESERVE | ZONE_SPECIAL, 1);
211 zinitna(mapzone, &mapobj, NULL, 0, 0, 0, 1);
218 * Red black tree functions
220 * The caller must hold the related map lock.
222 static int rb_vm_map_compare(vm_map_entry_t a, vm_map_entry_t b);
223 RB_GENERATE(vm_map_rb_tree, vm_map_entry, rb_entry, rb_vm_map_compare);
225 /* a->start is address, and the only field has to be initialized */
227 rb_vm_map_compare(vm_map_entry_t a, vm_map_entry_t b)
229 if (a->start < b->start)
231 else if (a->start > b->start)
237 * Allocate a vmspace structure, including a vm_map and pmap.
238 * Initialize numerous fields. While the initial allocation is zerod,
239 * subsequence reuse from the objcache leaves elements of the structure
240 * intact (particularly the pmap), so portions must be zerod.
242 * The structure is not considered activated until we call sysref_activate().
247 vmspace_alloc(vm_offset_t min, vm_offset_t max)
251 lwkt_gettoken(&vmspace_token);
252 vm = sysref_alloc(&vmspace_sysref_class);
253 bzero(&vm->vm_startcopy,
254 (char *)&vm->vm_endcopy - (char *)&vm->vm_startcopy);
255 vm_map_init(&vm->vm_map, min, max, NULL);
256 pmap_pinit(vmspace_pmap(vm)); /* (some fields reused) */
257 vm->vm_map.pmap = vmspace_pmap(vm); /* XXX */
259 vm->vm_exitingcnt = 0;
260 cpu_vmspace_alloc(vm);
261 sysref_activate(&vm->vm_sysref);
262 lwkt_reltoken(&vmspace_token);
268 * dtor function - Some elements of the pmap are retained in the
269 * free-cached vmspaces to improve performance. We have to clean them up
270 * here before returning the vmspace to the memory pool.
275 vmspace_dtor(void *obj, void *private)
277 struct vmspace *vm = obj;
279 pmap_puninit(vmspace_pmap(vm));
283 * Called in two cases:
285 * (1) When the last sysref is dropped, but exitingcnt might still be
288 * (2) When there are no sysrefs (i.e. refcnt is negative) left and the
289 * exitingcnt becomes zero
291 * sysref will not scrap the object until we call sysref_put() once more
292 * after the last ref has been dropped.
294 * Interlocked by the sysref API.
297 vmspace_terminate(struct vmspace *vm)
302 * If exitingcnt is non-zero we can't get rid of the entire vmspace
303 * yet, but we can scrap user memory.
305 lwkt_gettoken(&vmspace_token);
306 if (vm->vm_exitingcnt) {
308 pmap_remove_pages(vmspace_pmap(vm), VM_MIN_USER_ADDRESS,
309 VM_MAX_USER_ADDRESS);
310 vm_map_remove(&vm->vm_map, VM_MIN_USER_ADDRESS,
311 VM_MAX_USER_ADDRESS);
312 lwkt_reltoken(&vmspace_token);
315 cpu_vmspace_free(vm);
318 * Make sure any SysV shm is freed, it might not have in
323 KKASSERT(vm->vm_upcalls == NULL);
326 * Lock the map, to wait out all other references to it.
327 * Delete all of the mappings and pages they hold, then call
328 * the pmap module to reclaim anything left.
330 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
331 vm_map_lock(&vm->vm_map);
332 vm_map_delete(&vm->vm_map, vm->vm_map.min_offset,
333 vm->vm_map.max_offset, &count);
334 vm_map_unlock(&vm->vm_map);
335 vm_map_entry_release(count);
337 pmap_release(vmspace_pmap(vm));
338 sysref_put(&vm->vm_sysref);
339 lwkt_reltoken(&vmspace_token);
343 * vmspaces are not currently locked.
346 vmspace_lock(struct vmspace *vm __unused)
351 vmspace_unlock(struct vmspace *vm __unused)
356 * This is called during exit indicating that the vmspace is no
357 * longer in used by an exiting process, but the process has not yet
363 vmspace_exitbump(struct vmspace *vm)
365 lwkt_gettoken(&vmspace_token);
367 lwkt_reltoken(&vmspace_token);
371 * This is called in the wait*() handling code. The vmspace can be terminated
372 * after the last wait is finished using it.
377 vmspace_exitfree(struct proc *p)
381 lwkt_gettoken(&vmspace_token);
385 if (--vm->vm_exitingcnt == 0 && sysref_isinactive(&vm->vm_sysref))
386 vmspace_terminate(vm);
387 lwkt_reltoken(&vmspace_token);
391 * Swap useage is determined by taking the proportional swap used by
392 * VM objects backing the VM map. To make up for fractional losses,
393 * if the VM object has any swap use at all the associated map entries
394 * count for at least 1 swap page.
399 vmspace_swap_count(struct vmspace *vmspace)
401 vm_map_t map = &vmspace->vm_map;
407 lwkt_gettoken(&vmspace_token);
408 for (cur = map->header.next; cur != &map->header; cur = cur->next) {
409 switch(cur->maptype) {
410 case VM_MAPTYPE_NORMAL:
411 case VM_MAPTYPE_VPAGETABLE:
412 if ((object = cur->object.vm_object) == NULL)
414 if (object->swblock_count) {
415 n = (cur->end - cur->start) / PAGE_SIZE;
416 count += object->swblock_count *
417 SWAP_META_PAGES * n / object->size + 1;
424 lwkt_reltoken(&vmspace_token);
429 * Calculate the approximate number of anonymous pages in use by
430 * this vmspace. To make up for fractional losses, we count each
431 * VM object as having at least 1 anonymous page.
436 vmspace_anonymous_count(struct vmspace *vmspace)
438 vm_map_t map = &vmspace->vm_map;
443 lwkt_gettoken(&vmspace_token);
444 for (cur = map->header.next; cur != &map->header; cur = cur->next) {
445 switch(cur->maptype) {
446 case VM_MAPTYPE_NORMAL:
447 case VM_MAPTYPE_VPAGETABLE:
448 if ((object = cur->object.vm_object) == NULL)
450 if (object->type != OBJT_DEFAULT &&
451 object->type != OBJT_SWAP) {
454 count += object->resident_page_count;
460 lwkt_reltoken(&vmspace_token);
465 * Creates and returns a new empty VM map with the given physical map
466 * structure, and having the given lower and upper address bounds.
471 vm_map_create(vm_map_t result, pmap_t pmap, vm_offset_t min, vm_offset_t max)
474 result = zalloc(mapzone);
475 vm_map_init(result, min, max, pmap);
480 * Initialize an existing vm_map structure such as that in the vmspace
481 * structure. The pmap is initialized elsewhere.
486 vm_map_init(struct vm_map *map, vm_offset_t min, vm_offset_t max, pmap_t pmap)
488 map->header.next = map->header.prev = &map->header;
489 RB_INIT(&map->rb_root);
493 map->min_offset = min;
494 map->max_offset = max;
496 map->first_free = &map->header;
497 map->hint = &map->header;
500 lockinit(&map->lock, "thrd_sleep", (hz + 9) / 10, 0);
501 TUNABLE_INT("vm.cache_vmspaces", &vmspace_sysref_class.nom_cache);
505 * Shadow the vm_map_entry's object. This typically needs to be done when
506 * a write fault is taken on an entry which had previously been cloned by
507 * fork(). The shared object (which might be NULL) must become private so
508 * we add a shadow layer above it.
510 * Object allocation for anonymous mappings is defered as long as possible.
511 * When creating a shadow, however, the underlying object must be instantiated
512 * so it can be shared.
514 * If the map segment is governed by a virtual page table then it is
515 * possible to address offsets beyond the mapped area. Just allocate
516 * a maximally sized object for this case.
518 * The vm_map must be exclusively locked.
519 * No other requirements.
523 vm_map_entry_shadow(vm_map_entry_t entry)
525 if (entry->maptype == VM_MAPTYPE_VPAGETABLE) {
526 vm_object_shadow(&entry->object.vm_object, &entry->offset,
527 0x7FFFFFFF); /* XXX */
529 vm_object_shadow(&entry->object.vm_object, &entry->offset,
530 atop(entry->end - entry->start));
532 entry->eflags &= ~MAP_ENTRY_NEEDS_COPY;
536 * Allocate an object for a vm_map_entry.
538 * Object allocation for anonymous mappings is defered as long as possible.
539 * This function is called when we can defer no longer, generally when a map
540 * entry might be split or forked or takes a page fault.
542 * If the map segment is governed by a virtual page table then it is
543 * possible to address offsets beyond the mapped area. Just allocate
544 * a maximally sized object for this case.
546 * The vm_map must be exclusively locked.
547 * No other requirements.
550 vm_map_entry_allocate_object(vm_map_entry_t entry)
554 if (entry->maptype == VM_MAPTYPE_VPAGETABLE) {
555 obj = vm_object_allocate(OBJT_DEFAULT, 0x7FFFFFFF); /* XXX */
557 obj = vm_object_allocate(OBJT_DEFAULT,
558 atop(entry->end - entry->start));
560 entry->object.vm_object = obj;
565 * Set an initial negative count so the first attempt to reserve
566 * space preloads a bunch of vm_map_entry's for this cpu. Also
567 * pre-allocate 2 vm_map_entries which will be needed by zalloc() to
568 * map a new page for vm_map_entry structures. SMP systems are
569 * particularly sensitive.
571 * This routine is called in early boot so we cannot just call
572 * vm_map_entry_reserve().
574 * Called from the low level boot code only (for each cpu)
577 vm_map_entry_reserve_cpu_init(globaldata_t gd)
579 vm_map_entry_t entry;
582 gd->gd_vme_avail -= MAP_RESERVE_COUNT * 2;
583 entry = &cpu_map_entry_init[gd->gd_cpuid][0];
584 for (i = 0; i < VMEPERCPU; ++i, ++entry) {
585 entry->next = gd->gd_vme_base;
586 gd->gd_vme_base = entry;
591 * Reserves vm_map_entry structures so code later on can manipulate
592 * map_entry structures within a locked map without blocking trying
593 * to allocate a new vm_map_entry.
598 vm_map_entry_reserve(int count)
600 struct globaldata *gd = mycpu;
601 vm_map_entry_t entry;
604 * Make sure we have enough structures in gd_vme_base to handle
605 * the reservation request.
607 * The critical section protects access to the per-cpu gd.
610 while (gd->gd_vme_avail < count) {
611 entry = zalloc(mapentzone);
612 entry->next = gd->gd_vme_base;
613 gd->gd_vme_base = entry;
616 gd->gd_vme_avail -= count;
623 * Releases previously reserved vm_map_entry structures that were not
624 * used. If we have too much junk in our per-cpu cache clean some of
630 vm_map_entry_release(int count)
632 struct globaldata *gd = mycpu;
633 vm_map_entry_t entry;
636 gd->gd_vme_avail += count;
637 while (gd->gd_vme_avail > MAP_RESERVE_SLOP) {
638 entry = gd->gd_vme_base;
639 KKASSERT(entry != NULL);
640 gd->gd_vme_base = entry->next;
643 zfree(mapentzone, entry);
650 * Reserve map entry structures for use in kernel_map itself. These
651 * entries have *ALREADY* been reserved on a per-cpu basis when the map
652 * was inited. This function is used by zalloc() to avoid a recursion
653 * when zalloc() itself needs to allocate additional kernel memory.
655 * This function works like the normal reserve but does not load the
656 * vm_map_entry cache (because that would result in an infinite
657 * recursion). Note that gd_vme_avail may go negative. This is expected.
659 * Any caller of this function must be sure to renormalize after
660 * potentially eating entries to ensure that the reserve supply
666 vm_map_entry_kreserve(int count)
668 struct globaldata *gd = mycpu;
671 gd->gd_vme_avail -= count;
673 KASSERT(gd->gd_vme_base != NULL,
674 ("no reserved entries left, gd_vme_avail = %d\n",
680 * Release previously reserved map entries for kernel_map. We do not
681 * attempt to clean up like the normal release function as this would
682 * cause an unnecessary (but probably not fatal) deep procedure call.
687 vm_map_entry_krelease(int count)
689 struct globaldata *gd = mycpu;
692 gd->gd_vme_avail += count;
697 * Allocates a VM map entry for insertion. No entry fields are filled in.
699 * The entries should have previously been reserved. The reservation count
700 * is tracked in (*countp).
704 static vm_map_entry_t
705 vm_map_entry_create(vm_map_t map, int *countp)
707 struct globaldata *gd = mycpu;
708 vm_map_entry_t entry;
710 KKASSERT(*countp > 0);
713 entry = gd->gd_vme_base;
714 KASSERT(entry != NULL, ("gd_vme_base NULL! count %d", *countp));
715 gd->gd_vme_base = entry->next;
722 * Dispose of a vm_map_entry that is no longer being referenced.
727 vm_map_entry_dispose(vm_map_t map, vm_map_entry_t entry, int *countp)
729 struct globaldata *gd = mycpu;
731 KKASSERT(map->hint != entry);
732 KKASSERT(map->first_free != entry);
736 entry->next = gd->gd_vme_base;
737 gd->gd_vme_base = entry;
743 * Insert/remove entries from maps.
745 * The related map must be exclusively locked.
746 * No other requirements.
748 * NOTE! We currently acquire the vmspace_token only to avoid races
749 * against the pageout daemon's calls to vmspace_*_count(), which
750 * are unable to safely lock the vm_map without potentially
754 vm_map_entry_link(vm_map_t map,
755 vm_map_entry_t after_where,
756 vm_map_entry_t entry)
758 ASSERT_VM_MAP_LOCKED(map);
760 lwkt_gettoken(&vmspace_token);
762 entry->prev = after_where;
763 entry->next = after_where->next;
764 entry->next->prev = entry;
765 after_where->next = entry;
766 if (vm_map_rb_tree_RB_INSERT(&map->rb_root, entry))
767 panic("vm_map_entry_link: dup addr map %p ent %p", map, entry);
768 lwkt_reltoken(&vmspace_token);
772 vm_map_entry_unlink(vm_map_t map,
773 vm_map_entry_t entry)
778 ASSERT_VM_MAP_LOCKED(map);
780 if (entry->eflags & MAP_ENTRY_IN_TRANSITION) {
781 panic("vm_map_entry_unlink: attempt to mess with "
782 "locked entry! %p", entry);
784 lwkt_gettoken(&vmspace_token);
789 vm_map_rb_tree_RB_REMOVE(&map->rb_root, entry);
791 lwkt_reltoken(&vmspace_token);
795 * Finds the map entry containing (or immediately preceding) the specified
796 * address in the given map. The entry is returned in (*entry).
798 * The boolean result indicates whether the address is actually contained
801 * The related map must be locked.
802 * No other requirements.
805 vm_map_lookup_entry(vm_map_t map, vm_offset_t address, vm_map_entry_t *entry)
810 ASSERT_VM_MAP_LOCKED(map);
813 * XXX TEMPORARILY DISABLED. For some reason our attempt to revive
814 * the hint code with the red-black lookup meets with system crashes
815 * and lockups. We do not yet know why.
817 * It is possible that the problem is related to the setting
818 * of the hint during map_entry deletion, in the code specified
819 * at the GGG comment later on in this file.
822 * Quickly check the cached hint, there's a good chance of a match.
824 if (map->hint != &map->header) {
826 if (address >= tmp->start && address < tmp->end) {
834 * Locate the record from the top of the tree. 'last' tracks the
835 * closest prior record and is returned if no match is found, which
836 * in binary tree terms means tracking the most recent right-branch
837 * taken. If there is no prior record, &map->header is returned.
840 tmp = RB_ROOT(&map->rb_root);
843 if (address >= tmp->start) {
844 if (address < tmp->end) {
850 tmp = RB_RIGHT(tmp, rb_entry);
852 tmp = RB_LEFT(tmp, rb_entry);
860 * Inserts the given whole VM object into the target map at the specified
861 * address range. The object's size should match that of the address range.
863 * The map must be exclusively locked.
864 * The caller must have reserved sufficient vm_map_entry structures.
866 * If object is non-NULL, ref count must be bumped by caller
867 * prior to making call to account for the new entry.
870 vm_map_insert(vm_map_t map, int *countp,
871 vm_object_t object, vm_ooffset_t offset,
872 vm_offset_t start, vm_offset_t end,
873 vm_maptype_t maptype,
874 vm_prot_t prot, vm_prot_t max,
877 vm_map_entry_t new_entry;
878 vm_map_entry_t prev_entry;
879 vm_map_entry_t temp_entry;
880 vm_eflags_t protoeflags;
882 ASSERT_VM_MAP_LOCKED(map);
885 * Check that the start and end points are not bogus.
887 if ((start < map->min_offset) || (end > map->max_offset) ||
889 return (KERN_INVALID_ADDRESS);
892 * Find the entry prior to the proposed starting address; if it's part
893 * of an existing entry, this range is bogus.
895 if (vm_map_lookup_entry(map, start, &temp_entry))
896 return (KERN_NO_SPACE);
898 prev_entry = temp_entry;
901 * Assert that the next entry doesn't overlap the end point.
904 if ((prev_entry->next != &map->header) &&
905 (prev_entry->next->start < end))
906 return (KERN_NO_SPACE);
910 if (cow & MAP_COPY_ON_WRITE)
911 protoeflags |= MAP_ENTRY_COW|MAP_ENTRY_NEEDS_COPY;
913 if (cow & MAP_NOFAULT) {
914 protoeflags |= MAP_ENTRY_NOFAULT;
916 KASSERT(object == NULL,
917 ("vm_map_insert: paradoxical MAP_NOFAULT request"));
919 if (cow & MAP_DISABLE_SYNCER)
920 protoeflags |= MAP_ENTRY_NOSYNC;
921 if (cow & MAP_DISABLE_COREDUMP)
922 protoeflags |= MAP_ENTRY_NOCOREDUMP;
923 if (cow & MAP_IS_STACK)
924 protoeflags |= MAP_ENTRY_STACK;
925 if (cow & MAP_IS_KSTACK)
926 protoeflags |= MAP_ENTRY_KSTACK;
928 lwkt_gettoken(&vm_token);
929 lwkt_gettoken(&vmobj_token);
933 * When object is non-NULL, it could be shared with another
934 * process. We have to set or clear OBJ_ONEMAPPING
938 if ((object->ref_count > 1) || (object->shadow_count != 0)) {
939 vm_object_clear_flag(object, OBJ_ONEMAPPING);
942 else if ((prev_entry != &map->header) &&
943 (prev_entry->eflags == protoeflags) &&
944 (prev_entry->end == start) &&
945 (prev_entry->wired_count == 0) &&
946 prev_entry->maptype == maptype &&
947 ((prev_entry->object.vm_object == NULL) ||
948 vm_object_coalesce(prev_entry->object.vm_object,
949 OFF_TO_IDX(prev_entry->offset),
950 (vm_size_t)(prev_entry->end - prev_entry->start),
951 (vm_size_t)(end - prev_entry->end)))) {
953 * We were able to extend the object. Determine if we
954 * can extend the previous map entry to include the
957 if ((prev_entry->inheritance == VM_INHERIT_DEFAULT) &&
958 (prev_entry->protection == prot) &&
959 (prev_entry->max_protection == max)) {
960 lwkt_reltoken(&vmobj_token);
961 lwkt_reltoken(&vm_token);
962 map->size += (end - prev_entry->end);
963 prev_entry->end = end;
964 vm_map_simplify_entry(map, prev_entry, countp);
965 return (KERN_SUCCESS);
969 * If we can extend the object but cannot extend the
970 * map entry, we have to create a new map entry. We
971 * must bump the ref count on the extended object to
972 * account for it. object may be NULL.
974 object = prev_entry->object.vm_object;
975 offset = prev_entry->offset +
976 (prev_entry->end - prev_entry->start);
977 vm_object_reference_locked(object);
980 lwkt_reltoken(&vmobj_token);
981 lwkt_reltoken(&vm_token);
984 * NOTE: if conditionals fail, object can be NULL here. This occurs
985 * in things like the buffer map where we manage kva but do not manage
993 new_entry = vm_map_entry_create(map, countp);
994 new_entry->start = start;
995 new_entry->end = end;
997 new_entry->maptype = maptype;
998 new_entry->eflags = protoeflags;
999 new_entry->object.vm_object = object;
1000 new_entry->offset = offset;
1001 new_entry->aux.master_pde = 0;
1003 new_entry->inheritance = VM_INHERIT_DEFAULT;
1004 new_entry->protection = prot;
1005 new_entry->max_protection = max;
1006 new_entry->wired_count = 0;
1009 * Insert the new entry into the list
1012 vm_map_entry_link(map, prev_entry, new_entry);
1013 map->size += new_entry->end - new_entry->start;
1016 * Update the free space hint. Entries cannot overlap.
1017 * An exact comparison is needed to avoid matching
1018 * against the map->header.
1020 if ((map->first_free == prev_entry) &&
1021 (prev_entry->end == new_entry->start)) {
1022 map->first_free = new_entry;
1027 * Temporarily removed to avoid MAP_STACK panic, due to
1028 * MAP_STACK being a huge hack. Will be added back in
1029 * when MAP_STACK (and the user stack mapping) is fixed.
1032 * It may be possible to simplify the entry
1034 vm_map_simplify_entry(map, new_entry, countp);
1038 * Try to pre-populate the page table. Mappings governed by virtual
1039 * page tables cannot be prepopulated without a lot of work, so
1042 if ((cow & (MAP_PREFAULT|MAP_PREFAULT_PARTIAL)) &&
1043 maptype != VM_MAPTYPE_VPAGETABLE) {
1044 pmap_object_init_pt(map->pmap, start, prot,
1045 object, OFF_TO_IDX(offset), end - start,
1046 cow & MAP_PREFAULT_PARTIAL);
1049 return (KERN_SUCCESS);
1053 * Find sufficient space for `length' bytes in the given map, starting at
1054 * `start'. Returns 0 on success, 1 on no space.
1056 * This function will returned an arbitrarily aligned pointer. If no
1057 * particular alignment is required you should pass align as 1. Note that
1058 * the map may return PAGE_SIZE aligned pointers if all the lengths used in
1059 * the map are a multiple of PAGE_SIZE, even if you pass a smaller align
1062 * 'align' should be a power of 2 but is not required to be.
1064 * The map must be exclusively locked.
1065 * No other requirements.
1068 vm_map_findspace(vm_map_t map, vm_offset_t start, vm_size_t length,
1069 vm_size_t align, int flags, vm_offset_t *addr)
1071 vm_map_entry_t entry, next;
1073 vm_offset_t align_mask;
1075 if (start < map->min_offset)
1076 start = map->min_offset;
1077 if (start > map->max_offset)
1081 * If the alignment is not a power of 2 we will have to use
1082 * a mod/division, set align_mask to a special value.
1084 if ((align | (align - 1)) + 1 != (align << 1))
1085 align_mask = (vm_offset_t)-1;
1087 align_mask = align - 1;
1090 * Look for the first possible address; if there's already something
1091 * at this address, we have to start after it.
1093 if (start == map->min_offset) {
1094 if ((entry = map->first_free) != &map->header)
1099 if (vm_map_lookup_entry(map, start, &tmp))
1105 * Look through the rest of the map, trying to fit a new region in the
1106 * gap between existing regions, or after the very last region.
1108 for (;; start = (entry = next)->end) {
1110 * Adjust the proposed start by the requested alignment,
1111 * be sure that we didn't wrap the address.
1113 if (align_mask == (vm_offset_t)-1)
1114 end = ((start + align - 1) / align) * align;
1116 end = (start + align_mask) & ~align_mask;
1121 * Find the end of the proposed new region. Be sure we didn't
1122 * go beyond the end of the map, or wrap around the address.
1123 * Then check to see if this is the last entry or if the
1124 * proposed end fits in the gap between this and the next
1127 end = start + length;
1128 if (end > map->max_offset || end < start)
1133 * If the next entry's start address is beyond the desired
1134 * end address we may have found a good entry.
1136 * If the next entry is a stack mapping we do not map into
1137 * the stack's reserved space.
1139 * XXX continue to allow mapping into the stack's reserved
1140 * space if doing a MAP_STACK mapping inside a MAP_STACK
1141 * mapping, for backwards compatibility. But the caller
1142 * really should use MAP_STACK | MAP_TRYFIXED if they
1145 if (next == &map->header)
1147 if (next->start >= end) {
1148 if ((next->eflags & MAP_ENTRY_STACK) == 0)
1150 if (flags & MAP_STACK)
1152 if (next->start - next->aux.avail_ssize >= end)
1159 * Grow the kernel_map if necessary. pmap_growkernel() will panic
1160 * if it fails. The kernel_map is locked and nothing can steal
1161 * our address space if pmap_growkernel() blocks.
1163 * NOTE: This may be unconditionally called for kldload areas on
1164 * x86_64 because these do not bump kernel_vm_end (which would
1165 * fill 128G worth of page tables!). Therefore we must not
1168 if (map == &kernel_map) {
1171 kstop = round_page(start + length);
1172 if (kstop > kernel_vm_end)
1173 pmap_growkernel(start, kstop);
1180 * vm_map_find finds an unallocated region in the target address map with
1181 * the given length. The search is defined to be first-fit from the
1182 * specified address; the region found is returned in the same parameter.
1184 * If object is non-NULL, ref count must be bumped by caller
1185 * prior to making call to account for the new entry.
1187 * No requirements. This function will lock the map temporarily.
1190 vm_map_find(vm_map_t map, vm_object_t object, vm_ooffset_t offset,
1191 vm_offset_t *addr, vm_size_t length, vm_size_t align,
1193 vm_maptype_t maptype,
1194 vm_prot_t prot, vm_prot_t max,
1203 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
1206 if (vm_map_findspace(map, start, length, align, 0, addr)) {
1208 vm_map_entry_release(count);
1209 return (KERN_NO_SPACE);
1213 result = vm_map_insert(map, &count, object, offset,
1214 start, start + length,
1219 vm_map_entry_release(count);
1225 * Simplify the given map entry by merging with either neighbor. This
1226 * routine also has the ability to merge with both neighbors.
1228 * This routine guarentees that the passed entry remains valid (though
1229 * possibly extended). When merging, this routine may delete one or
1230 * both neighbors. No action is taken on entries which have their
1231 * in-transition flag set.
1233 * The map must be exclusively locked.
1236 vm_map_simplify_entry(vm_map_t map, vm_map_entry_t entry, int *countp)
1238 vm_map_entry_t next, prev;
1239 vm_size_t prevsize, esize;
1241 if (entry->eflags & MAP_ENTRY_IN_TRANSITION) {
1242 ++mycpu->gd_cnt.v_intrans_coll;
1246 if (entry->maptype == VM_MAPTYPE_SUBMAP)
1250 if (prev != &map->header) {
1251 prevsize = prev->end - prev->start;
1252 if ( (prev->end == entry->start) &&
1253 (prev->maptype == entry->maptype) &&
1254 (prev->object.vm_object == entry->object.vm_object) &&
1255 (!prev->object.vm_object ||
1256 (prev->offset + prevsize == entry->offset)) &&
1257 (prev->eflags == entry->eflags) &&
1258 (prev->protection == entry->protection) &&
1259 (prev->max_protection == entry->max_protection) &&
1260 (prev->inheritance == entry->inheritance) &&
1261 (prev->wired_count == entry->wired_count)) {
1262 if (map->first_free == prev)
1263 map->first_free = entry;
1264 if (map->hint == prev)
1266 vm_map_entry_unlink(map, prev);
1267 entry->start = prev->start;
1268 entry->offset = prev->offset;
1269 if (prev->object.vm_object)
1270 vm_object_deallocate(prev->object.vm_object);
1271 vm_map_entry_dispose(map, prev, countp);
1276 if (next != &map->header) {
1277 esize = entry->end - entry->start;
1278 if ((entry->end == next->start) &&
1279 (next->maptype == entry->maptype) &&
1280 (next->object.vm_object == entry->object.vm_object) &&
1281 (!entry->object.vm_object ||
1282 (entry->offset + esize == next->offset)) &&
1283 (next->eflags == entry->eflags) &&
1284 (next->protection == entry->protection) &&
1285 (next->max_protection == entry->max_protection) &&
1286 (next->inheritance == entry->inheritance) &&
1287 (next->wired_count == entry->wired_count)) {
1288 if (map->first_free == next)
1289 map->first_free = entry;
1290 if (map->hint == next)
1292 vm_map_entry_unlink(map, next);
1293 entry->end = next->end;
1294 if (next->object.vm_object)
1295 vm_object_deallocate(next->object.vm_object);
1296 vm_map_entry_dispose(map, next, countp);
1302 * Asserts that the given entry begins at or after the specified address.
1303 * If necessary, it splits the entry into two.
1305 #define vm_map_clip_start(map, entry, startaddr, countp) \
1307 if (startaddr > entry->start) \
1308 _vm_map_clip_start(map, entry, startaddr, countp); \
1312 * This routine is called only when it is known that the entry must be split.
1314 * The map must be exclusively locked.
1317 _vm_map_clip_start(vm_map_t map, vm_map_entry_t entry, vm_offset_t start,
1320 vm_map_entry_t new_entry;
1323 * Split off the front portion -- note that we must insert the new
1324 * entry BEFORE this one, so that this entry has the specified
1328 vm_map_simplify_entry(map, entry, countp);
1331 * If there is no object backing this entry, we might as well create
1332 * one now. If we defer it, an object can get created after the map
1333 * is clipped, and individual objects will be created for the split-up
1334 * map. This is a bit of a hack, but is also about the best place to
1335 * put this improvement.
1337 if (entry->object.vm_object == NULL && !map->system_map) {
1338 vm_map_entry_allocate_object(entry);
1341 new_entry = vm_map_entry_create(map, countp);
1342 *new_entry = *entry;
1344 new_entry->end = start;
1345 entry->offset += (start - entry->start);
1346 entry->start = start;
1348 vm_map_entry_link(map, entry->prev, new_entry);
1350 switch(entry->maptype) {
1351 case VM_MAPTYPE_NORMAL:
1352 case VM_MAPTYPE_VPAGETABLE:
1353 vm_object_reference(new_entry->object.vm_object);
1361 * Asserts that the given entry ends at or before the specified address.
1362 * If necessary, it splits the entry into two.
1364 * The map must be exclusively locked.
1366 #define vm_map_clip_end(map, entry, endaddr, countp) \
1368 if (endaddr < entry->end) \
1369 _vm_map_clip_end(map, entry, endaddr, countp); \
1373 * This routine is called only when it is known that the entry must be split.
1375 * The map must be exclusively locked.
1378 _vm_map_clip_end(vm_map_t map, vm_map_entry_t entry, vm_offset_t end,
1381 vm_map_entry_t new_entry;
1384 * If there is no object backing this entry, we might as well create
1385 * one now. If we defer it, an object can get created after the map
1386 * is clipped, and individual objects will be created for the split-up
1387 * map. This is a bit of a hack, but is also about the best place to
1388 * put this improvement.
1391 if (entry->object.vm_object == NULL && !map->system_map) {
1392 vm_map_entry_allocate_object(entry);
1396 * Create a new entry and insert it AFTER the specified entry
1399 new_entry = vm_map_entry_create(map, countp);
1400 *new_entry = *entry;
1402 new_entry->start = entry->end = end;
1403 new_entry->offset += (end - entry->start);
1405 vm_map_entry_link(map, entry, new_entry);
1407 switch(entry->maptype) {
1408 case VM_MAPTYPE_NORMAL:
1409 case VM_MAPTYPE_VPAGETABLE:
1410 vm_object_reference(new_entry->object.vm_object);
1418 * Asserts that the starting and ending region addresses fall within the
1419 * valid range for the map.
1421 #define VM_MAP_RANGE_CHECK(map, start, end) \
1423 if (start < vm_map_min(map)) \
1424 start = vm_map_min(map); \
1425 if (end > vm_map_max(map)) \
1426 end = vm_map_max(map); \
1432 * Used to block when an in-transition collison occurs. The map
1433 * is unlocked for the sleep and relocked before the return.
1436 vm_map_transition_wait(vm_map_t map)
1438 tsleep_interlock(map, 0);
1440 tsleep(map, PINTERLOCKED, "vment", 0);
1445 * When we do blocking operations with the map lock held it is
1446 * possible that a clip might have occured on our in-transit entry,
1447 * requiring an adjustment to the entry in our loop. These macros
1448 * help the pageable and clip_range code deal with the case. The
1449 * conditional costs virtually nothing if no clipping has occured.
1452 #define CLIP_CHECK_BACK(entry, save_start) \
1454 while (entry->start != save_start) { \
1455 entry = entry->prev; \
1456 KASSERT(entry != &map->header, ("bad entry clip")); \
1460 #define CLIP_CHECK_FWD(entry, save_end) \
1462 while (entry->end != save_end) { \
1463 entry = entry->next; \
1464 KASSERT(entry != &map->header, ("bad entry clip")); \
1470 * Clip the specified range and return the base entry. The
1471 * range may cover several entries starting at the returned base
1472 * and the first and last entry in the covering sequence will be
1473 * properly clipped to the requested start and end address.
1475 * If no holes are allowed you should pass the MAP_CLIP_NO_HOLES
1478 * The MAP_ENTRY_IN_TRANSITION flag will be set for the entries
1479 * covered by the requested range.
1481 * The map must be exclusively locked on entry and will remain locked
1482 * on return. If no range exists or the range contains holes and you
1483 * specified that no holes were allowed, NULL will be returned. This
1484 * routine may temporarily unlock the map in order avoid a deadlock when
1489 vm_map_clip_range(vm_map_t map, vm_offset_t start, vm_offset_t end,
1490 int *countp, int flags)
1492 vm_map_entry_t start_entry;
1493 vm_map_entry_t entry;
1496 * Locate the entry and effect initial clipping. The in-transition
1497 * case does not occur very often so do not try to optimize it.
1500 if (vm_map_lookup_entry(map, start, &start_entry) == FALSE)
1502 entry = start_entry;
1503 if (entry->eflags & MAP_ENTRY_IN_TRANSITION) {
1504 entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP;
1505 ++mycpu->gd_cnt.v_intrans_coll;
1506 ++mycpu->gd_cnt.v_intrans_wait;
1507 vm_map_transition_wait(map);
1509 * entry and/or start_entry may have been clipped while
1510 * we slept, or may have gone away entirely. We have
1511 * to restart from the lookup.
1517 * Since we hold an exclusive map lock we do not have to restart
1518 * after clipping, even though clipping may block in zalloc.
1520 vm_map_clip_start(map, entry, start, countp);
1521 vm_map_clip_end(map, entry, end, countp);
1522 entry->eflags |= MAP_ENTRY_IN_TRANSITION;
1525 * Scan entries covered by the range. When working on the next
1526 * entry a restart need only re-loop on the current entry which
1527 * we have already locked, since 'next' may have changed. Also,
1528 * even though entry is safe, it may have been clipped so we
1529 * have to iterate forwards through the clip after sleeping.
1531 while (entry->next != &map->header && entry->next->start < end) {
1532 vm_map_entry_t next = entry->next;
1534 if (flags & MAP_CLIP_NO_HOLES) {
1535 if (next->start > entry->end) {
1536 vm_map_unclip_range(map, start_entry,
1537 start, entry->end, countp, flags);
1542 if (next->eflags & MAP_ENTRY_IN_TRANSITION) {
1543 vm_offset_t save_end = entry->end;
1544 next->eflags |= MAP_ENTRY_NEEDS_WAKEUP;
1545 ++mycpu->gd_cnt.v_intrans_coll;
1546 ++mycpu->gd_cnt.v_intrans_wait;
1547 vm_map_transition_wait(map);
1550 * clips might have occured while we blocked.
1552 CLIP_CHECK_FWD(entry, save_end);
1553 CLIP_CHECK_BACK(start_entry, start);
1557 * No restart necessary even though clip_end may block, we
1558 * are holding the map lock.
1560 vm_map_clip_end(map, next, end, countp);
1561 next->eflags |= MAP_ENTRY_IN_TRANSITION;
1564 if (flags & MAP_CLIP_NO_HOLES) {
1565 if (entry->end != end) {
1566 vm_map_unclip_range(map, start_entry,
1567 start, entry->end, countp, flags);
1571 return(start_entry);
1575 * Undo the effect of vm_map_clip_range(). You should pass the same
1576 * flags and the same range that you passed to vm_map_clip_range().
1577 * This code will clear the in-transition flag on the entries and
1578 * wake up anyone waiting. This code will also simplify the sequence
1579 * and attempt to merge it with entries before and after the sequence.
1581 * The map must be locked on entry and will remain locked on return.
1583 * Note that you should also pass the start_entry returned by
1584 * vm_map_clip_range(). However, if you block between the two calls
1585 * with the map unlocked please be aware that the start_entry may
1586 * have been clipped and you may need to scan it backwards to find
1587 * the entry corresponding with the original start address. You are
1588 * responsible for this, vm_map_unclip_range() expects the correct
1589 * start_entry to be passed to it and will KASSERT otherwise.
1593 vm_map_unclip_range(vm_map_t map, vm_map_entry_t start_entry,
1594 vm_offset_t start, vm_offset_t end,
1595 int *countp, int flags)
1597 vm_map_entry_t entry;
1599 entry = start_entry;
1601 KASSERT(entry->start == start, ("unclip_range: illegal base entry"));
1602 while (entry != &map->header && entry->start < end) {
1603 KASSERT(entry->eflags & MAP_ENTRY_IN_TRANSITION,
1604 ("in-transition flag not set during unclip on: %p",
1606 KASSERT(entry->end <= end,
1607 ("unclip_range: tail wasn't clipped"));
1608 entry->eflags &= ~MAP_ENTRY_IN_TRANSITION;
1609 if (entry->eflags & MAP_ENTRY_NEEDS_WAKEUP) {
1610 entry->eflags &= ~MAP_ENTRY_NEEDS_WAKEUP;
1613 entry = entry->next;
1617 * Simplification does not block so there is no restart case.
1619 entry = start_entry;
1620 while (entry != &map->header && entry->start < end) {
1621 vm_map_simplify_entry(map, entry, countp);
1622 entry = entry->next;
1627 * Mark the given range as handled by a subordinate map.
1629 * This range must have been created with vm_map_find(), and no other
1630 * operations may have been performed on this range prior to calling
1633 * Submappings cannot be removed.
1638 vm_map_submap(vm_map_t map, vm_offset_t start, vm_offset_t end, vm_map_t submap)
1640 vm_map_entry_t entry;
1641 int result = KERN_INVALID_ARGUMENT;
1644 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
1647 VM_MAP_RANGE_CHECK(map, start, end);
1649 if (vm_map_lookup_entry(map, start, &entry)) {
1650 vm_map_clip_start(map, entry, start, &count);
1652 entry = entry->next;
1655 vm_map_clip_end(map, entry, end, &count);
1657 if ((entry->start == start) && (entry->end == end) &&
1658 ((entry->eflags & MAP_ENTRY_COW) == 0) &&
1659 (entry->object.vm_object == NULL)) {
1660 entry->object.sub_map = submap;
1661 entry->maptype = VM_MAPTYPE_SUBMAP;
1662 result = KERN_SUCCESS;
1665 vm_map_entry_release(count);
1671 * Sets the protection of the specified address region in the target map.
1672 * If "set_max" is specified, the maximum protection is to be set;
1673 * otherwise, only the current protection is affected.
1675 * The protection is not applicable to submaps, but is applicable to normal
1676 * maps and maps governed by virtual page tables. For example, when operating
1677 * on a virtual page table our protection basically controls how COW occurs
1678 * on the backing object, whereas the virtual page table abstraction itself
1679 * is an abstraction for userland.
1684 vm_map_protect(vm_map_t map, vm_offset_t start, vm_offset_t end,
1685 vm_prot_t new_prot, boolean_t set_max)
1687 vm_map_entry_t current;
1688 vm_map_entry_t entry;
1691 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
1694 VM_MAP_RANGE_CHECK(map, start, end);
1696 if (vm_map_lookup_entry(map, start, &entry)) {
1697 vm_map_clip_start(map, entry, start, &count);
1699 entry = entry->next;
1703 * Make a first pass to check for protection violations.
1706 while ((current != &map->header) && (current->start < end)) {
1707 if (current->maptype == VM_MAPTYPE_SUBMAP) {
1709 vm_map_entry_release(count);
1710 return (KERN_INVALID_ARGUMENT);
1712 if ((new_prot & current->max_protection) != new_prot) {
1714 vm_map_entry_release(count);
1715 return (KERN_PROTECTION_FAILURE);
1717 current = current->next;
1721 * Go back and fix up protections. [Note that clipping is not
1722 * necessary the second time.]
1726 while ((current != &map->header) && (current->start < end)) {
1729 vm_map_clip_end(map, current, end, &count);
1731 old_prot = current->protection;
1733 current->protection =
1734 (current->max_protection = new_prot) &
1737 current->protection = new_prot;
1741 * Update physical map if necessary. Worry about copy-on-write
1742 * here -- CHECK THIS XXX
1745 if (current->protection != old_prot) {
1746 #define MASK(entry) (((entry)->eflags & MAP_ENTRY_COW) ? ~VM_PROT_WRITE : \
1749 pmap_protect(map->pmap, current->start,
1751 current->protection & MASK(current));
1755 vm_map_simplify_entry(map, current, &count);
1757 current = current->next;
1761 vm_map_entry_release(count);
1762 return (KERN_SUCCESS);
1766 * This routine traverses a processes map handling the madvise
1767 * system call. Advisories are classified as either those effecting
1768 * the vm_map_entry structure, or those effecting the underlying
1771 * The <value> argument is used for extended madvise calls.
1776 vm_map_madvise(vm_map_t map, vm_offset_t start, vm_offset_t end,
1777 int behav, off_t value)
1779 vm_map_entry_t current, entry;
1785 * Some madvise calls directly modify the vm_map_entry, in which case
1786 * we need to use an exclusive lock on the map and we need to perform
1787 * various clipping operations. Otherwise we only need a read-lock
1791 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
1795 case MADV_SEQUENTIAL:
1809 vm_map_lock_read(map);
1812 vm_map_entry_release(count);
1817 * Locate starting entry and clip if necessary.
1820 VM_MAP_RANGE_CHECK(map, start, end);
1822 if (vm_map_lookup_entry(map, start, &entry)) {
1824 vm_map_clip_start(map, entry, start, &count);
1826 entry = entry->next;
1831 * madvise behaviors that are implemented in the vm_map_entry.
1833 * We clip the vm_map_entry so that behavioral changes are
1834 * limited to the specified address range.
1836 for (current = entry;
1837 (current != &map->header) && (current->start < end);
1838 current = current->next
1840 if (current->maptype == VM_MAPTYPE_SUBMAP)
1843 vm_map_clip_end(map, current, end, &count);
1847 vm_map_entry_set_behavior(current, MAP_ENTRY_BEHAV_NORMAL);
1849 case MADV_SEQUENTIAL:
1850 vm_map_entry_set_behavior(current, MAP_ENTRY_BEHAV_SEQUENTIAL);
1853 vm_map_entry_set_behavior(current, MAP_ENTRY_BEHAV_RANDOM);
1856 current->eflags |= MAP_ENTRY_NOSYNC;
1859 current->eflags &= ~MAP_ENTRY_NOSYNC;
1862 current->eflags |= MAP_ENTRY_NOCOREDUMP;
1865 current->eflags &= ~MAP_ENTRY_NOCOREDUMP;
1869 * Invalidate the related pmap entries, used
1870 * to flush portions of the real kernel's
1871 * pmap when the caller has removed or
1872 * modified existing mappings in a virtual
1875 pmap_remove(map->pmap,
1876 current->start, current->end);
1880 * Set the page directory page for a map
1881 * governed by a virtual page table. Mark
1882 * the entry as being governed by a virtual
1883 * page table if it is not.
1885 * XXX the page directory page is stored
1886 * in the avail_ssize field if the map_entry.
1888 * XXX the map simplification code does not
1889 * compare this field so weird things may
1890 * happen if you do not apply this function
1891 * to the entire mapping governed by the
1892 * virtual page table.
1894 if (current->maptype != VM_MAPTYPE_VPAGETABLE) {
1898 current->aux.master_pde = value;
1899 pmap_remove(map->pmap,
1900 current->start, current->end);
1906 vm_map_simplify_entry(map, current, &count);
1914 * madvise behaviors that are implemented in the underlying
1917 * Since we don't clip the vm_map_entry, we have to clip
1918 * the vm_object pindex and count.
1920 * NOTE! We currently do not support these functions on
1921 * virtual page tables.
1923 for (current = entry;
1924 (current != &map->header) && (current->start < end);
1925 current = current->next
1927 vm_offset_t useStart;
1929 if (current->maptype != VM_MAPTYPE_NORMAL)
1932 pindex = OFF_TO_IDX(current->offset);
1933 count = atop(current->end - current->start);
1934 useStart = current->start;
1936 if (current->start < start) {
1937 pindex += atop(start - current->start);
1938 count -= atop(start - current->start);
1941 if (current->end > end)
1942 count -= atop(current->end - end);
1947 vm_object_madvise(current->object.vm_object,
1948 pindex, count, behav);
1951 * Try to populate the page table. Mappings governed
1952 * by virtual page tables cannot be pre-populated
1953 * without a lot of work so don't try.
1955 if (behav == MADV_WILLNEED &&
1956 current->maptype != VM_MAPTYPE_VPAGETABLE) {
1957 pmap_object_init_pt(
1960 current->protection,
1961 current->object.vm_object,
1963 (count << PAGE_SHIFT),
1964 MAP_PREFAULT_MADVISE
1968 vm_map_unlock_read(map);
1970 vm_map_entry_release(count);
1976 * Sets the inheritance of the specified address range in the target map.
1977 * Inheritance affects how the map will be shared with child maps at the
1978 * time of vm_map_fork.
1981 vm_map_inherit(vm_map_t map, vm_offset_t start, vm_offset_t end,
1982 vm_inherit_t new_inheritance)
1984 vm_map_entry_t entry;
1985 vm_map_entry_t temp_entry;
1988 switch (new_inheritance) {
1989 case VM_INHERIT_NONE:
1990 case VM_INHERIT_COPY:
1991 case VM_INHERIT_SHARE:
1994 return (KERN_INVALID_ARGUMENT);
1997 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
2000 VM_MAP_RANGE_CHECK(map, start, end);
2002 if (vm_map_lookup_entry(map, start, &temp_entry)) {
2004 vm_map_clip_start(map, entry, start, &count);
2006 entry = temp_entry->next;
2008 while ((entry != &map->header) && (entry->start < end)) {
2009 vm_map_clip_end(map, entry, end, &count);
2011 entry->inheritance = new_inheritance;
2013 vm_map_simplify_entry(map, entry, &count);
2015 entry = entry->next;
2018 vm_map_entry_release(count);
2019 return (KERN_SUCCESS);
2023 * Implement the semantics of mlock
2026 vm_map_unwire(vm_map_t map, vm_offset_t start, vm_offset_t real_end,
2027 boolean_t new_pageable)
2029 vm_map_entry_t entry;
2030 vm_map_entry_t start_entry;
2032 int rv = KERN_SUCCESS;
2035 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
2037 VM_MAP_RANGE_CHECK(map, start, real_end);
2040 start_entry = vm_map_clip_range(map, start, end, &count,
2042 if (start_entry == NULL) {
2044 vm_map_entry_release(count);
2045 return (KERN_INVALID_ADDRESS);
2048 if (new_pageable == 0) {
2049 entry = start_entry;
2050 while ((entry != &map->header) && (entry->start < end)) {
2051 vm_offset_t save_start;
2052 vm_offset_t save_end;
2055 * Already user wired or hard wired (trivial cases)
2057 if (entry->eflags & MAP_ENTRY_USER_WIRED) {
2058 entry = entry->next;
2061 if (entry->wired_count != 0) {
2062 entry->wired_count++;
2063 entry->eflags |= MAP_ENTRY_USER_WIRED;
2064 entry = entry->next;
2069 * A new wiring requires instantiation of appropriate
2070 * management structures and the faulting in of the
2073 if (entry->maptype != VM_MAPTYPE_SUBMAP) {
2074 int copyflag = entry->eflags &
2075 MAP_ENTRY_NEEDS_COPY;
2076 if (copyflag && ((entry->protection &
2077 VM_PROT_WRITE) != 0)) {
2078 vm_map_entry_shadow(entry);
2079 } else if (entry->object.vm_object == NULL &&
2081 vm_map_entry_allocate_object(entry);
2084 entry->wired_count++;
2085 entry->eflags |= MAP_ENTRY_USER_WIRED;
2088 * Now fault in the area. Note that vm_fault_wire()
2089 * may release the map lock temporarily, it will be
2090 * relocked on return. The in-transition
2091 * flag protects the entries.
2093 save_start = entry->start;
2094 save_end = entry->end;
2095 rv = vm_fault_wire(map, entry, TRUE);
2097 CLIP_CHECK_BACK(entry, save_start);
2099 KASSERT(entry->wired_count == 1, ("bad wired_count on entry"));
2100 entry->eflags &= ~MAP_ENTRY_USER_WIRED;
2101 entry->wired_count = 0;
2102 if (entry->end == save_end)
2104 entry = entry->next;
2105 KASSERT(entry != &map->header, ("bad entry clip during backout"));
2107 end = save_start; /* unwire the rest */
2111 * note that even though the entry might have been
2112 * clipped, the USER_WIRED flag we set prevents
2113 * duplication so we do not have to do a
2116 entry = entry->next;
2120 * If we failed fall through to the unwiring section to
2121 * unwire what we had wired so far. 'end' has already
2128 * start_entry might have been clipped if we unlocked the
2129 * map and blocked. No matter how clipped it has gotten
2130 * there should be a fragment that is on our start boundary.
2132 CLIP_CHECK_BACK(start_entry, start);
2136 * Deal with the unwiring case.
2140 * This is the unwiring case. We must first ensure that the
2141 * range to be unwired is really wired down. We know there
2144 entry = start_entry;
2145 while ((entry != &map->header) && (entry->start < end)) {
2146 if ((entry->eflags & MAP_ENTRY_USER_WIRED) == 0) {
2147 rv = KERN_INVALID_ARGUMENT;
2150 KASSERT(entry->wired_count != 0, ("wired count was 0 with USER_WIRED set! %p", entry));
2151 entry = entry->next;
2155 * Now decrement the wiring count for each region. If a region
2156 * becomes completely unwired, unwire its physical pages and
2160 * The map entries are processed in a loop, checking to
2161 * make sure the entry is wired and asserting it has a wired
2162 * count. However, another loop was inserted more-or-less in
2163 * the middle of the unwiring path. This loop picks up the
2164 * "entry" loop variable from the first loop without first
2165 * setting it to start_entry. Naturally, the secound loop
2166 * is never entered and the pages backing the entries are
2167 * never unwired. This can lead to a leak of wired pages.
2169 entry = start_entry;
2170 while ((entry != &map->header) && (entry->start < end)) {
2171 KASSERT(entry->eflags & MAP_ENTRY_USER_WIRED,
2172 ("expected USER_WIRED on entry %p", entry));
2173 entry->eflags &= ~MAP_ENTRY_USER_WIRED;
2174 entry->wired_count--;
2175 if (entry->wired_count == 0)
2176 vm_fault_unwire(map, entry);
2177 entry = entry->next;
2181 vm_map_unclip_range(map, start_entry, start, real_end, &count,
2185 vm_map_entry_release(count);
2190 * Sets the pageability of the specified address range in the target map.
2191 * Regions specified as not pageable require locked-down physical
2192 * memory and physical page maps.
2194 * The map must not be locked, but a reference must remain to the map
2195 * throughout the call.
2197 * This function may be called via the zalloc path and must properly
2198 * reserve map entries for kernel_map.
2203 vm_map_wire(vm_map_t map, vm_offset_t start, vm_offset_t real_end, int kmflags)
2205 vm_map_entry_t entry;
2206 vm_map_entry_t start_entry;
2208 int rv = KERN_SUCCESS;
2211 if (kmflags & KM_KRESERVE)
2212 count = vm_map_entry_kreserve(MAP_RESERVE_COUNT);
2214 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
2216 VM_MAP_RANGE_CHECK(map, start, real_end);
2219 start_entry = vm_map_clip_range(map, start, end, &count,
2221 if (start_entry == NULL) {
2223 rv = KERN_INVALID_ADDRESS;
2226 if ((kmflags & KM_PAGEABLE) == 0) {
2230 * 1. Holding the write lock, we create any shadow or zero-fill
2231 * objects that need to be created. Then we clip each map
2232 * entry to the region to be wired and increment its wiring
2233 * count. We create objects before clipping the map entries
2234 * to avoid object proliferation.
2236 * 2. We downgrade to a read lock, and call vm_fault_wire to
2237 * fault in the pages for any newly wired area (wired_count is
2240 * Downgrading to a read lock for vm_fault_wire avoids a
2241 * possible deadlock with another process that may have faulted
2242 * on one of the pages to be wired (it would mark the page busy,
2243 * blocking us, then in turn block on the map lock that we
2244 * hold). Because of problems in the recursive lock package,
2245 * we cannot upgrade to a write lock in vm_map_lookup. Thus,
2246 * any actions that require the write lock must be done
2247 * beforehand. Because we keep the read lock on the map, the
2248 * copy-on-write status of the entries we modify here cannot
2251 entry = start_entry;
2252 while ((entry != &map->header) && (entry->start < end)) {
2254 * Trivial case if the entry is already wired
2256 if (entry->wired_count) {
2257 entry->wired_count++;
2258 entry = entry->next;
2263 * The entry is being newly wired, we have to setup
2264 * appropriate management structures. A shadow
2265 * object is required for a copy-on-write region,
2266 * or a normal object for a zero-fill region. We
2267 * do not have to do this for entries that point to sub
2268 * maps because we won't hold the lock on the sub map.
2270 if (entry->maptype != VM_MAPTYPE_SUBMAP) {
2271 int copyflag = entry->eflags &
2272 MAP_ENTRY_NEEDS_COPY;
2273 if (copyflag && ((entry->protection &
2274 VM_PROT_WRITE) != 0)) {
2275 vm_map_entry_shadow(entry);
2276 } else if (entry->object.vm_object == NULL &&
2278 vm_map_entry_allocate_object(entry);
2282 entry->wired_count++;
2283 entry = entry->next;
2291 * HACK HACK HACK HACK
2293 * vm_fault_wire() temporarily unlocks the map to avoid
2294 * deadlocks. The in-transition flag from vm_map_clip_range
2295 * call should protect us from changes while the map is
2298 * NOTE: Previously this comment stated that clipping might
2299 * still occur while the entry is unlocked, but from
2300 * what I can tell it actually cannot.
2302 * It is unclear whether the CLIP_CHECK_*() calls
2303 * are still needed but we keep them in anyway.
2305 * HACK HACK HACK HACK
2308 entry = start_entry;
2309 while (entry != &map->header && entry->start < end) {
2311 * If vm_fault_wire fails for any page we need to undo
2312 * what has been done. We decrement the wiring count
2313 * for those pages which have not yet been wired (now)
2314 * and unwire those that have (later).
2316 vm_offset_t save_start = entry->start;
2317 vm_offset_t save_end = entry->end;
2319 if (entry->wired_count == 1)
2320 rv = vm_fault_wire(map, entry, FALSE);
2322 CLIP_CHECK_BACK(entry, save_start);
2324 KASSERT(entry->wired_count == 1, ("wired_count changed unexpectedly"));
2325 entry->wired_count = 0;
2326 if (entry->end == save_end)
2328 entry = entry->next;
2329 KASSERT(entry != &map->header, ("bad entry clip during backout"));
2334 CLIP_CHECK_FWD(entry, save_end);
2335 entry = entry->next;
2339 * If a failure occured undo everything by falling through
2340 * to the unwiring code. 'end' has already been adjusted
2344 kmflags |= KM_PAGEABLE;
2347 * start_entry is still IN_TRANSITION but may have been
2348 * clipped since vm_fault_wire() unlocks and relocks the
2349 * map. No matter how clipped it has gotten there should
2350 * be a fragment that is on our start boundary.
2352 CLIP_CHECK_BACK(start_entry, start);
2355 if (kmflags & KM_PAGEABLE) {
2357 * This is the unwiring case. We must first ensure that the
2358 * range to be unwired is really wired down. We know there
2361 entry = start_entry;
2362 while ((entry != &map->header) && (entry->start < end)) {
2363 if (entry->wired_count == 0) {
2364 rv = KERN_INVALID_ARGUMENT;
2367 entry = entry->next;
2371 * Now decrement the wiring count for each region. If a region
2372 * becomes completely unwired, unwire its physical pages and
2375 entry = start_entry;
2376 while ((entry != &map->header) && (entry->start < end)) {
2377 entry->wired_count--;
2378 if (entry->wired_count == 0)
2379 vm_fault_unwire(map, entry);
2380 entry = entry->next;
2384 vm_map_unclip_range(map, start_entry, start, real_end,
2385 &count, MAP_CLIP_NO_HOLES);
2389 if (kmflags & KM_KRESERVE)
2390 vm_map_entry_krelease(count);
2392 vm_map_entry_release(count);
2397 * Mark a newly allocated address range as wired but do not fault in
2398 * the pages. The caller is expected to load the pages into the object.
2400 * The map must be locked on entry and will remain locked on return.
2401 * No other requirements.
2404 vm_map_set_wired_quick(vm_map_t map, vm_offset_t addr, vm_size_t size,
2407 vm_map_entry_t scan;
2408 vm_map_entry_t entry;
2410 entry = vm_map_clip_range(map, addr, addr + size,
2411 countp, MAP_CLIP_NO_HOLES);
2413 scan != &map->header && scan->start < addr + size;
2414 scan = scan->next) {
2415 KKASSERT(entry->wired_count == 0);
2416 entry->wired_count = 1;
2418 vm_map_unclip_range(map, entry, addr, addr + size,
2419 countp, MAP_CLIP_NO_HOLES);
2423 * Push any dirty cached pages in the address range to their pager.
2424 * If syncio is TRUE, dirty pages are written synchronously.
2425 * If invalidate is TRUE, any cached pages are freed as well.
2427 * This routine is called by sys_msync()
2429 * Returns an error if any part of the specified range is not mapped.
2434 vm_map_clean(vm_map_t map, vm_offset_t start, vm_offset_t end,
2435 boolean_t syncio, boolean_t invalidate)
2437 vm_map_entry_t current;
2438 vm_map_entry_t entry;
2441 vm_ooffset_t offset;
2443 vm_map_lock_read(map);
2444 VM_MAP_RANGE_CHECK(map, start, end);
2445 if (!vm_map_lookup_entry(map, start, &entry)) {
2446 vm_map_unlock_read(map);
2447 return (KERN_INVALID_ADDRESS);
2450 * Make a first pass to check for holes.
2452 for (current = entry; current->start < end; current = current->next) {
2453 if (current->maptype == VM_MAPTYPE_SUBMAP) {
2454 vm_map_unlock_read(map);
2455 return (KERN_INVALID_ARGUMENT);
2457 if (end > current->end &&
2458 (current->next == &map->header ||
2459 current->end != current->next->start)) {
2460 vm_map_unlock_read(map);
2461 return (KERN_INVALID_ADDRESS);
2466 pmap_remove(vm_map_pmap(map), start, end);
2469 * Make a second pass, cleaning/uncaching pages from the indicated
2472 * Hold vm_token to avoid blocking in vm_object_reference()
2474 lwkt_gettoken(&vm_token);
2475 lwkt_gettoken(&vmobj_token);
2477 for (current = entry; current->start < end; current = current->next) {
2478 offset = current->offset + (start - current->start);
2479 size = (end <= current->end ? end : current->end) - start;
2480 if (current->maptype == VM_MAPTYPE_SUBMAP) {
2482 vm_map_entry_t tentry;
2485 smap = current->object.sub_map;
2486 vm_map_lock_read(smap);
2487 vm_map_lookup_entry(smap, offset, &tentry);
2488 tsize = tentry->end - offset;
2491 object = tentry->object.vm_object;
2492 offset = tentry->offset + (offset - tentry->start);
2493 vm_map_unlock_read(smap);
2495 object = current->object.vm_object;
2498 * Note that there is absolutely no sense in writing out
2499 * anonymous objects, so we track down the vnode object
2501 * We invalidate (remove) all pages from the address space
2502 * anyway, for semantic correctness.
2504 * note: certain anonymous maps, such as MAP_NOSYNC maps,
2505 * may start out with a NULL object.
2507 while (object && object->backing_object) {
2508 offset += object->backing_object_offset;
2509 object = object->backing_object;
2510 if (object->size < OFF_TO_IDX( offset + size))
2511 size = IDX_TO_OFF(object->size) - offset;
2513 if (object && (object->type == OBJT_VNODE) &&
2514 (current->protection & VM_PROT_WRITE) &&
2515 (object->flags & OBJ_NOMSYNC) == 0) {
2517 * Flush pages if writing is allowed, invalidate them
2518 * if invalidation requested. Pages undergoing I/O
2519 * will be ignored by vm_object_page_remove().
2521 * We cannot lock the vnode and then wait for paging
2522 * to complete without deadlocking against vm_fault.
2523 * Instead we simply call vm_object_page_remove() and
2524 * allow it to block internally on a page-by-page
2525 * basis when it encounters pages undergoing async
2530 vm_object_reference_locked(object);
2531 vn_lock(object->handle, LK_EXCLUSIVE | LK_RETRY);
2532 flags = (syncio || invalidate) ? OBJPC_SYNC : 0;
2533 flags |= invalidate ? OBJPC_INVAL : 0;
2536 * When operating on a virtual page table just
2537 * flush the whole object. XXX we probably ought
2540 switch(current->maptype) {
2541 case VM_MAPTYPE_NORMAL:
2542 vm_object_page_clean(object,
2544 OFF_TO_IDX(offset + size + PAGE_MASK),
2547 case VM_MAPTYPE_VPAGETABLE:
2548 vm_object_page_clean(object, 0, 0, flags);
2551 vn_unlock(((struct vnode *)object->handle));
2552 vm_object_deallocate_locked(object);
2554 if (object && invalidate &&
2555 ((object->type == OBJT_VNODE) ||
2556 (object->type == OBJT_DEVICE))) {
2558 (object->type == OBJT_DEVICE) ? FALSE : TRUE;
2559 vm_object_reference_locked(object);
2560 switch(current->maptype) {
2561 case VM_MAPTYPE_NORMAL:
2562 vm_object_page_remove(object,
2564 OFF_TO_IDX(offset + size + PAGE_MASK),
2567 case VM_MAPTYPE_VPAGETABLE:
2568 vm_object_page_remove(object, 0, 0, clean_only);
2571 vm_object_deallocate_locked(object);
2576 lwkt_reltoken(&vmobj_token);
2577 lwkt_reltoken(&vm_token);
2578 vm_map_unlock_read(map);
2580 return (KERN_SUCCESS);
2584 * Make the region specified by this entry pageable.
2586 * The vm_map must be exclusively locked.
2589 vm_map_entry_unwire(vm_map_t map, vm_map_entry_t entry)
2591 entry->eflags &= ~MAP_ENTRY_USER_WIRED;
2592 entry->wired_count = 0;
2593 vm_fault_unwire(map, entry);
2597 * Deallocate the given entry from the target map.
2599 * The vm_map must be exclusively locked.
2602 vm_map_entry_delete(vm_map_t map, vm_map_entry_t entry, int *countp)
2604 vm_map_entry_unlink(map, entry);
2605 map->size -= entry->end - entry->start;
2607 switch(entry->maptype) {
2608 case VM_MAPTYPE_NORMAL:
2609 case VM_MAPTYPE_VPAGETABLE:
2610 vm_object_deallocate(entry->object.vm_object);
2616 vm_map_entry_dispose(map, entry, countp);
2620 * Deallocates the given address range from the target map.
2622 * The vm_map must be exclusively locked.
2625 vm_map_delete(vm_map_t map, vm_offset_t start, vm_offset_t end, int *countp)
2628 vm_map_entry_t entry;
2629 vm_map_entry_t first_entry;
2631 ASSERT_VM_MAP_LOCKED(map);
2634 * Find the start of the region, and clip it. Set entry to point
2635 * at the first record containing the requested address or, if no
2636 * such record exists, the next record with a greater address. The
2637 * loop will run from this point until a record beyond the termination
2638 * address is encountered.
2640 * map->hint must be adjusted to not point to anything we delete,
2641 * so set it to the entry prior to the one being deleted.
2643 * GGG see other GGG comment.
2645 if (vm_map_lookup_entry(map, start, &first_entry)) {
2646 entry = first_entry;
2647 vm_map_clip_start(map, entry, start, countp);
2648 map->hint = entry->prev; /* possible problem XXX */
2650 map->hint = first_entry; /* possible problem XXX */
2651 entry = first_entry->next;
2655 * If a hole opens up prior to the current first_free then
2656 * adjust first_free. As with map->hint, map->first_free
2657 * cannot be left set to anything we might delete.
2659 if (entry == &map->header) {
2660 map->first_free = &map->header;
2661 } else if (map->first_free->start >= start) {
2662 map->first_free = entry->prev;
2666 * Step through all entries in this region
2668 while ((entry != &map->header) && (entry->start < end)) {
2669 vm_map_entry_t next;
2671 vm_pindex_t offidxstart, offidxend, count;
2674 * If we hit an in-transition entry we have to sleep and
2675 * retry. It's easier (and not really slower) to just retry
2676 * since this case occurs so rarely and the hint is already
2677 * pointing at the right place. We have to reset the
2678 * start offset so as not to accidently delete an entry
2679 * another process just created in vacated space.
2681 if (entry->eflags & MAP_ENTRY_IN_TRANSITION) {
2682 entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP;
2683 start = entry->start;
2684 ++mycpu->gd_cnt.v_intrans_coll;
2685 ++mycpu->gd_cnt.v_intrans_wait;
2686 vm_map_transition_wait(map);
2689 vm_map_clip_end(map, entry, end, countp);
2695 offidxstart = OFF_TO_IDX(entry->offset);
2696 count = OFF_TO_IDX(e - s);
2697 object = entry->object.vm_object;
2700 * Unwire before removing addresses from the pmap; otherwise,
2701 * unwiring will put the entries back in the pmap.
2703 if (entry->wired_count != 0)
2704 vm_map_entry_unwire(map, entry);
2706 offidxend = offidxstart + count;
2709 * Hold vm_token when manipulating vm_objects,
2711 * Hold vmobj_token when potentially adding or removing
2712 * objects (collapse requires both).
2714 lwkt_gettoken(&vm_token);
2715 lwkt_gettoken(&vmobj_token);
2716 vm_object_hold(object);
2718 if (object == &kernel_object) {
2719 vm_object_page_remove(object, offidxstart,
2722 pmap_remove(map->pmap, s, e);
2724 if (object != NULL &&
2725 object->ref_count != 1 &&
2726 (object->flags & (OBJ_NOSPLIT|OBJ_ONEMAPPING)) ==
2728 (object->type == OBJT_DEFAULT ||
2729 object->type == OBJT_SWAP)) {
2730 vm_object_collapse(object);
2731 vm_object_page_remove(object, offidxstart,
2733 if (object->type == OBJT_SWAP) {
2734 swap_pager_freespace(object,
2738 if (offidxend >= object->size &&
2739 offidxstart < object->size) {
2740 object->size = offidxstart;
2745 vm_object_drop(object);
2746 lwkt_reltoken(&vmobj_token);
2747 lwkt_reltoken(&vm_token);
2750 * Delete the entry (which may delete the object) only after
2751 * removing all pmap entries pointing to its pages.
2752 * (Otherwise, its page frames may be reallocated, and any
2753 * modify bits will be set in the wrong object!)
2755 vm_map_entry_delete(map, entry, countp);
2758 return (KERN_SUCCESS);
2762 * Remove the given address range from the target map.
2763 * This is the exported form of vm_map_delete.
2768 vm_map_remove(vm_map_t map, vm_offset_t start, vm_offset_t end)
2773 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
2775 VM_MAP_RANGE_CHECK(map, start, end);
2776 result = vm_map_delete(map, start, end, &count);
2778 vm_map_entry_release(count);
2784 * Assert that the target map allows the specified privilege on the
2785 * entire address region given. The entire region must be allocated.
2787 * The caller must specify whether the vm_map is already locked or not.
2790 vm_map_check_protection(vm_map_t map, vm_offset_t start, vm_offset_t end,
2791 vm_prot_t protection, boolean_t have_lock)
2793 vm_map_entry_t entry;
2794 vm_map_entry_t tmp_entry;
2797 if (have_lock == FALSE)
2798 vm_map_lock_read(map);
2800 if (!vm_map_lookup_entry(map, start, &tmp_entry)) {
2801 if (have_lock == FALSE)
2802 vm_map_unlock_read(map);
2808 while (start < end) {
2809 if (entry == &map->header) {
2817 if (start < entry->start) {
2822 * Check protection associated with entry.
2825 if ((entry->protection & protection) != protection) {
2829 /* go to next entry */
2832 entry = entry->next;
2834 if (have_lock == FALSE)
2835 vm_map_unlock_read(map);
2840 * Split the pages in a map entry into a new object. This affords
2841 * easier removal of unused pages, and keeps object inheritance from
2842 * being a negative impact on memory usage.
2844 * The vm_map must be exclusively locked.
2845 * The orig_object should be held.
2848 vm_map_split(vm_map_entry_t entry)
2851 vm_object_t orig_object, new_object, source;
2853 vm_pindex_t offidxstart, offidxend, idx;
2855 vm_ooffset_t offset;
2857 orig_object = entry->object.vm_object;
2858 if (orig_object->type != OBJT_DEFAULT && orig_object->type != OBJT_SWAP)
2860 if (orig_object->ref_count <= 1)
2863 offset = entry->offset;
2867 offidxstart = OFF_TO_IDX(offset);
2868 offidxend = offidxstart + OFF_TO_IDX(e - s);
2869 size = offidxend - offidxstart;
2871 switch(orig_object->type) {
2873 new_object = default_pager_alloc(NULL, IDX_TO_OFF(size),
2877 new_object = swap_pager_alloc(NULL, IDX_TO_OFF(size),
2885 if (new_object == NULL)
2889 * vm_token required when manipulating vm_objects.
2891 lwkt_gettoken(&vm_token);
2892 lwkt_gettoken(&vmobj_token);
2894 vm_object_hold(new_object);
2896 source = orig_object->backing_object;
2897 if (source != NULL) {
2898 vm_object_hold(source);
2899 /* Referenced by new_object */
2900 vm_object_reference_locked(source);
2901 LIST_INSERT_HEAD(&source->shadow_head,
2902 new_object, shadow_list);
2903 vm_object_clear_flag(source, OBJ_ONEMAPPING);
2904 new_object->backing_object_offset =
2905 orig_object->backing_object_offset +
2906 IDX_TO_OFF(offidxstart);
2907 new_object->backing_object = source;
2908 source->shadow_count++;
2909 source->generation++;
2910 vm_object_drop(source);
2913 for (idx = 0; idx < size; idx++) {
2917 m = vm_page_lookup(orig_object, offidxstart + idx);
2922 * We must wait for pending I/O to complete before we can
2925 * We do not have to VM_PROT_NONE the page as mappings should
2926 * not be changed by this operation.
2928 if (vm_page_sleep_busy(m, TRUE, "spltwt"))
2931 vm_page_rename(m, new_object, idx);
2932 /* page automatically made dirty by rename and cache handled */
2936 if (orig_object->type == OBJT_SWAP) {
2937 vm_object_pip_add(orig_object, 1);
2939 * copy orig_object pages into new_object
2940 * and destroy unneeded pages in
2943 swap_pager_copy(orig_object, new_object, offidxstart, 0);
2944 vm_object_pip_wakeup(orig_object);
2948 * Wakeup the pages we played with. No spl protection is needed
2949 * for a simple wakeup.
2951 for (idx = 0; idx < size; idx++) {
2952 m = vm_page_lookup(new_object, idx);
2957 entry->object.vm_object = new_object;
2958 entry->offset = 0LL;
2959 vm_object_deallocate_locked(orig_object);
2960 vm_object_drop(new_object);
2961 lwkt_reltoken(&vmobj_token);
2962 lwkt_reltoken(&vm_token);
2966 * Copies the contents of the source entry to the destination
2967 * entry. The entries *must* be aligned properly.
2969 * The vm_map must be exclusively locked.
2970 * vm_token must be held
2973 vm_map_copy_entry(vm_map_t src_map, vm_map_t dst_map,
2974 vm_map_entry_t src_entry, vm_map_entry_t dst_entry)
2976 vm_object_t src_object;
2978 if (dst_entry->maptype == VM_MAPTYPE_SUBMAP)
2980 if (src_entry->maptype == VM_MAPTYPE_SUBMAP)
2983 ASSERT_LWKT_TOKEN_HELD(&vm_token);
2984 lwkt_gettoken(&vmobj_token); /* required for collapse */
2986 if (src_entry->wired_count == 0) {
2988 * If the source entry is marked needs_copy, it is already
2991 if ((src_entry->eflags & MAP_ENTRY_NEEDS_COPY) == 0) {
2992 pmap_protect(src_map->pmap,
2995 src_entry->protection & ~VM_PROT_WRITE);
2999 * Make a copy of the object.
3001 * The object must be locked prior to checking the object type
3002 * and for the call to vm_object_collapse() and vm_map_split().
3003 * We cannot use *_hold() here because the split code will
3004 * probably try to destroy the object. The lock is a pool
3005 * token and doesn't care.
3007 if ((src_object = src_entry->object.vm_object) != NULL) {
3008 vm_object_lock(src_object);
3009 if ((src_object->handle == NULL) &&
3010 (src_object->type == OBJT_DEFAULT ||
3011 src_object->type == OBJT_SWAP)) {
3012 vm_object_collapse(src_object);
3013 if ((src_object->flags & (OBJ_NOSPLIT|OBJ_ONEMAPPING)) == OBJ_ONEMAPPING) {
3014 vm_map_split(src_entry);
3015 vm_object_unlock(src_object);
3016 src_object = src_entry->object.vm_object;
3017 vm_object_lock(src_object);
3020 vm_object_reference_locked(src_object);
3021 vm_object_unlock(src_object);
3022 vm_object_clear_flag(src_object, OBJ_ONEMAPPING);
3023 dst_entry->object.vm_object = src_object;
3024 src_entry->eflags |= (MAP_ENTRY_COW|MAP_ENTRY_NEEDS_COPY);
3025 dst_entry->eflags |= (MAP_ENTRY_COW|MAP_ENTRY_NEEDS_COPY);
3026 dst_entry->offset = src_entry->offset;
3028 dst_entry->object.vm_object = NULL;
3029 dst_entry->offset = 0;
3032 pmap_copy(dst_map->pmap, src_map->pmap, dst_entry->start,
3033 dst_entry->end - dst_entry->start, src_entry->start);
3036 * Of course, wired down pages can't be set copy-on-write.
3037 * Cause wired pages to be copied into the new map by
3038 * simulating faults (the new pages are pageable)
3040 vm_fault_copy_entry(dst_map, src_map, dst_entry, src_entry);
3042 lwkt_reltoken(&vmobj_token);
3047 * Create a new process vmspace structure and vm_map
3048 * based on those of an existing process. The new map
3049 * is based on the old map, according to the inheritance
3050 * values on the regions in that map.
3052 * The source map must not be locked.
3056 vmspace_fork(struct vmspace *vm1)
3058 struct vmspace *vm2;
3059 vm_map_t old_map = &vm1->vm_map;
3061 vm_map_entry_t old_entry;
3062 vm_map_entry_t new_entry;
3066 lwkt_gettoken(&vm_token);
3067 lwkt_gettoken(&vmspace_token);
3068 lwkt_gettoken(&vmobj_token);
3069 vm_map_lock(old_map);
3072 * XXX Note: upcalls are not copied.
3074 vm2 = vmspace_alloc(old_map->min_offset, old_map->max_offset);
3075 bcopy(&vm1->vm_startcopy, &vm2->vm_startcopy,
3076 (caddr_t)&vm1->vm_endcopy - (caddr_t)&vm1->vm_startcopy);
3077 new_map = &vm2->vm_map; /* XXX */
3078 new_map->timestamp = 1;
3080 vm_map_lock(new_map);
3083 old_entry = old_map->header.next;
3084 while (old_entry != &old_map->header) {
3086 old_entry = old_entry->next;
3089 count = vm_map_entry_reserve(count + MAP_RESERVE_COUNT);
3091 old_entry = old_map->header.next;
3092 while (old_entry != &old_map->header) {
3093 if (old_entry->maptype == VM_MAPTYPE_SUBMAP)
3094 panic("vm_map_fork: encountered a submap");
3096 switch (old_entry->inheritance) {
3097 case VM_INHERIT_NONE:
3099 case VM_INHERIT_SHARE:
3101 * Clone the entry, creating the shared object if
3104 object = old_entry->object.vm_object;
3105 if (object == NULL) {
3106 vm_map_entry_allocate_object(old_entry);
3107 object = old_entry->object.vm_object;
3111 * Add the reference before calling vm_map_entry_shadow
3112 * to insure that a shadow object is created.
3114 vm_object_reference_locked(object);
3115 if (old_entry->eflags & MAP_ENTRY_NEEDS_COPY) {
3116 vm_map_entry_shadow(old_entry);
3117 /* Transfer the second reference too. */
3118 vm_object_reference_locked(
3119 old_entry->object.vm_object);
3120 vm_object_deallocate_locked(object);
3121 object = old_entry->object.vm_object;
3123 vm_object_clear_flag(object, OBJ_ONEMAPPING);
3126 * Clone the entry, referencing the shared object.
3128 new_entry = vm_map_entry_create(new_map, &count);
3129 *new_entry = *old_entry;
3130 new_entry->eflags &= ~MAP_ENTRY_USER_WIRED;
3131 new_entry->wired_count = 0;
3134 * Insert the entry into the new map -- we know we're
3135 * inserting at the end of the new map.
3138 vm_map_entry_link(new_map, new_map->header.prev,
3142 * Update the physical map
3144 pmap_copy(new_map->pmap, old_map->pmap,
3146 (old_entry->end - old_entry->start),
3149 case VM_INHERIT_COPY:
3151 * Clone the entry and link into the map.
3153 new_entry = vm_map_entry_create(new_map, &count);
3154 *new_entry = *old_entry;
3155 new_entry->eflags &= ~MAP_ENTRY_USER_WIRED;
3156 new_entry->wired_count = 0;
3157 new_entry->object.vm_object = NULL;
3158 vm_map_entry_link(new_map, new_map->header.prev,
3160 vm_map_copy_entry(old_map, new_map, old_entry,
3164 old_entry = old_entry->next;
3167 new_map->size = old_map->size;
3168 vm_map_unlock(old_map);
3169 vm_map_unlock(new_map);
3170 vm_map_entry_release(count);
3172 lwkt_reltoken(&vmobj_token);
3173 lwkt_reltoken(&vmspace_token);
3174 lwkt_reltoken(&vm_token);
3180 * Create an auto-grow stack entry
3185 vm_map_stack (vm_map_t map, vm_offset_t addrbos, vm_size_t max_ssize,
3186 int flags, vm_prot_t prot, vm_prot_t max, int cow)
3188 vm_map_entry_t prev_entry;
3189 vm_map_entry_t new_stack_entry;
3190 vm_size_t init_ssize;
3193 vm_offset_t tmpaddr;
3195 cow |= MAP_IS_STACK;
3197 if (max_ssize < sgrowsiz)
3198 init_ssize = max_ssize;
3200 init_ssize = sgrowsiz;
3202 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
3206 * Find space for the mapping
3208 if ((flags & (MAP_FIXED | MAP_TRYFIXED)) == 0) {
3209 if (vm_map_findspace(map, addrbos, max_ssize, 1,
3212 vm_map_entry_release(count);
3213 return (KERN_NO_SPACE);
3218 /* If addr is already mapped, no go */
3219 if (vm_map_lookup_entry(map, addrbos, &prev_entry)) {
3221 vm_map_entry_release(count);
3222 return (KERN_NO_SPACE);
3226 /* XXX already handled by kern_mmap() */
3227 /* If we would blow our VMEM resource limit, no go */
3228 if (map->size + init_ssize >
3229 curproc->p_rlimit[RLIMIT_VMEM].rlim_cur) {
3231 vm_map_entry_release(count);
3232 return (KERN_NO_SPACE);
3237 * If we can't accomodate max_ssize in the current mapping,
3238 * no go. However, we need to be aware that subsequent user
3239 * mappings might map into the space we have reserved for
3240 * stack, and currently this space is not protected.
3242 * Hopefully we will at least detect this condition
3243 * when we try to grow the stack.
3245 if ((prev_entry->next != &map->header) &&
3246 (prev_entry->next->start < addrbos + max_ssize)) {
3248 vm_map_entry_release(count);
3249 return (KERN_NO_SPACE);
3253 * We initially map a stack of only init_ssize. We will
3254 * grow as needed later. Since this is to be a grow
3255 * down stack, we map at the top of the range.
3257 * Note: we would normally expect prot and max to be
3258 * VM_PROT_ALL, and cow to be 0. Possibly we should
3259 * eliminate these as input parameters, and just
3260 * pass these values here in the insert call.
3262 rv = vm_map_insert(map, &count,
3263 NULL, 0, addrbos + max_ssize - init_ssize,
3264 addrbos + max_ssize,
3269 /* Now set the avail_ssize amount */
3270 if (rv == KERN_SUCCESS) {
3271 if (prev_entry != &map->header)
3272 vm_map_clip_end(map, prev_entry, addrbos + max_ssize - init_ssize, &count);
3273 new_stack_entry = prev_entry->next;
3274 if (new_stack_entry->end != addrbos + max_ssize ||
3275 new_stack_entry->start != addrbos + max_ssize - init_ssize)
3276 panic ("Bad entry start/end for new stack entry");
3278 new_stack_entry->aux.avail_ssize = max_ssize - init_ssize;
3282 vm_map_entry_release(count);
3287 * Attempts to grow a vm stack entry. Returns KERN_SUCCESS if the
3288 * desired address is already mapped, or if we successfully grow
3289 * the stack. Also returns KERN_SUCCESS if addr is outside the
3290 * stack range (this is strange, but preserves compatibility with
3291 * the grow function in vm_machdep.c).
3296 vm_map_growstack (struct proc *p, vm_offset_t addr)
3298 vm_map_entry_t prev_entry;
3299 vm_map_entry_t stack_entry;
3300 vm_map_entry_t new_stack_entry;
3301 struct vmspace *vm = p->p_vmspace;
3302 vm_map_t map = &vm->vm_map;
3305 int rv = KERN_SUCCESS;
3307 int use_read_lock = 1;
3310 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
3313 vm_map_lock_read(map);
3317 /* If addr is already in the entry range, no need to grow.*/
3318 if (vm_map_lookup_entry(map, addr, &prev_entry))
3321 if ((stack_entry = prev_entry->next) == &map->header)
3323 if (prev_entry == &map->header)
3324 end = stack_entry->start - stack_entry->aux.avail_ssize;
3326 end = prev_entry->end;
3329 * This next test mimics the old grow function in vm_machdep.c.
3330 * It really doesn't quite make sense, but we do it anyway
3331 * for compatibility.
3333 * If not growable stack, return success. This signals the
3334 * caller to proceed as he would normally with normal vm.
3336 if (stack_entry->aux.avail_ssize < 1 ||
3337 addr >= stack_entry->start ||
3338 addr < stack_entry->start - stack_entry->aux.avail_ssize) {
3342 /* Find the minimum grow amount */
3343 grow_amount = roundup (stack_entry->start - addr, PAGE_SIZE);
3344 if (grow_amount > stack_entry->aux.avail_ssize) {
3350 * If there is no longer enough space between the entries
3351 * nogo, and adjust the available space. Note: this
3352 * should only happen if the user has mapped into the
3353 * stack area after the stack was created, and is
3354 * probably an error.
3356 * This also effectively destroys any guard page the user
3357 * might have intended by limiting the stack size.
3359 if (grow_amount > stack_entry->start - end) {
3360 if (use_read_lock && vm_map_lock_upgrade(map)) {
3365 stack_entry->aux.avail_ssize = stack_entry->start - end;
3370 is_procstack = addr >= (vm_offset_t)vm->vm_maxsaddr;
3372 /* If this is the main process stack, see if we're over the
3375 if (is_procstack && (ctob(vm->vm_ssize) + grow_amount >
3376 p->p_rlimit[RLIMIT_STACK].rlim_cur)) {
3381 /* Round up the grow amount modulo SGROWSIZ */
3382 grow_amount = roundup (grow_amount, sgrowsiz);
3383 if (grow_amount > stack_entry->aux.avail_ssize) {
3384 grow_amount = stack_entry->aux.avail_ssize;
3386 if (is_procstack && (ctob(vm->vm_ssize) + grow_amount >
3387 p->p_rlimit[RLIMIT_STACK].rlim_cur)) {
3388 grow_amount = p->p_rlimit[RLIMIT_STACK].rlim_cur -
3392 /* If we would blow our VMEM resource limit, no go */
3393 if (map->size + grow_amount > p->p_rlimit[RLIMIT_VMEM].rlim_cur) {
3398 if (use_read_lock && vm_map_lock_upgrade(map)) {
3404 /* Get the preliminary new entry start value */
3405 addr = stack_entry->start - grow_amount;
3407 /* If this puts us into the previous entry, cut back our growth
3408 * to the available space. Also, see the note above.
3411 stack_entry->aux.avail_ssize = stack_entry->start - end;
3415 rv = vm_map_insert(map, &count,
3416 NULL, 0, addr, stack_entry->start,
3418 VM_PROT_ALL, VM_PROT_ALL,
3421 /* Adjust the available stack space by the amount we grew. */
3422 if (rv == KERN_SUCCESS) {
3423 if (prev_entry != &map->header)
3424 vm_map_clip_end(map, prev_entry, addr, &count);
3425 new_stack_entry = prev_entry->next;
3426 if (new_stack_entry->end != stack_entry->start ||
3427 new_stack_entry->start != addr)
3428 panic ("Bad stack grow start/end in new stack entry");
3430 new_stack_entry->aux.avail_ssize =
3431 stack_entry->aux.avail_ssize -
3432 (new_stack_entry->end - new_stack_entry->start);
3434 vm->vm_ssize += btoc(new_stack_entry->end -
3435 new_stack_entry->start);
3438 if (map->flags & MAP_WIREFUTURE)
3439 vm_map_unwire(map, new_stack_entry->start,
3440 new_stack_entry->end, FALSE);
3445 vm_map_unlock_read(map);
3448 vm_map_entry_release(count);
3453 * Unshare the specified VM space for exec. If other processes are
3454 * mapped to it, then create a new one. The new vmspace is null.
3459 vmspace_exec(struct proc *p, struct vmspace *vmcopy)
3461 struct vmspace *oldvmspace = p->p_vmspace;
3462 struct vmspace *newvmspace;
3463 vm_map_t map = &p->p_vmspace->vm_map;
3466 * If we are execing a resident vmspace we fork it, otherwise
3467 * we create a new vmspace. Note that exitingcnt and upcalls
3468 * are not copied to the new vmspace.
3470 lwkt_gettoken(&vmspace_token);
3472 newvmspace = vmspace_fork(vmcopy);
3474 newvmspace = vmspace_alloc(map->min_offset, map->max_offset);
3475 bcopy(&oldvmspace->vm_startcopy, &newvmspace->vm_startcopy,
3476 (caddr_t)&oldvmspace->vm_endcopy -
3477 (caddr_t)&oldvmspace->vm_startcopy);
3481 * Finish initializing the vmspace before assigning it
3482 * to the process. The vmspace will become the current vmspace
3485 pmap_pinit2(vmspace_pmap(newvmspace));
3486 pmap_replacevm(p, newvmspace, 0);
3487 sysref_put(&oldvmspace->vm_sysref);
3488 lwkt_reltoken(&vmspace_token);
3492 * Unshare the specified VM space for forcing COW. This
3493 * is called by rfork, for the (RFMEM|RFPROC) == 0 case.
3495 * The exitingcnt test is not strictly necessary but has been
3496 * included for code sanity (to make the code a bit more deterministic).
3499 vmspace_unshare(struct proc *p)
3501 struct vmspace *oldvmspace = p->p_vmspace;
3502 struct vmspace *newvmspace;
3504 lwkt_gettoken(&vmspace_token);
3505 if (oldvmspace->vm_sysref.refcnt == 1 && oldvmspace->vm_exitingcnt == 0)
3507 newvmspace = vmspace_fork(oldvmspace);
3508 pmap_pinit2(vmspace_pmap(newvmspace));
3509 pmap_replacevm(p, newvmspace, 0);
3510 sysref_put(&oldvmspace->vm_sysref);
3511 lwkt_reltoken(&vmspace_token);
3515 * vm_map_hint: return the beginning of the best area suitable for
3516 * creating a new mapping with "prot" protection.
3521 vm_map_hint(struct proc *p, vm_offset_t addr, vm_prot_t prot)
3523 struct vmspace *vms = p->p_vmspace;
3525 if (!randomize_mmap) {
3527 * Set a reasonable start point for the hint if it was
3528 * not specified or if it falls within the heap space.
3529 * Hinted mmap()s do not allocate out of the heap space.
3532 (addr >= round_page((vm_offset_t)vms->vm_taddr) &&
3533 addr < round_page((vm_offset_t)vms->vm_daddr + maxdsiz))) {
3534 addr = round_page((vm_offset_t)vms->vm_daddr + maxdsiz);
3540 if (addr != 0 && addr >= (vm_offset_t)vms->vm_daddr)
3546 * If executable skip first two pages, otherwise start
3547 * after data + heap region.
3549 if ((prot & VM_PROT_EXECUTE) &&
3550 ((vm_offset_t)vms->vm_daddr >= I386_MAX_EXE_ADDR)) {
3551 addr = (PAGE_SIZE * 2) +
3552 (karc4random() & (I386_MAX_EXE_ADDR / 2 - 1));
3553 return (round_page(addr));
3555 #endif /* __i386__ */
3558 addr = (vm_offset_t)vms->vm_daddr + MAXDSIZ;
3559 addr += karc4random() & (MIN((256 * 1024 * 1024), MAXDSIZ) - 1);
3561 return (round_page(addr));
3565 * Finds the VM object, offset, and protection for a given virtual address
3566 * in the specified map, assuming a page fault of the type specified.
3568 * Leaves the map in question locked for read; return values are guaranteed
3569 * until a vm_map_lookup_done call is performed. Note that the map argument
3570 * is in/out; the returned map must be used in the call to vm_map_lookup_done.
3572 * A handle (out_entry) is returned for use in vm_map_lookup_done, to make
3575 * If a lookup is requested with "write protection" specified, the map may
3576 * be changed to perform virtual copying operations, although the data
3577 * referenced will remain the same.
3582 vm_map_lookup(vm_map_t *var_map, /* IN/OUT */
3584 vm_prot_t fault_typea,
3585 vm_map_entry_t *out_entry, /* OUT */
3586 vm_object_t *object, /* OUT */
3587 vm_pindex_t *pindex, /* OUT */
3588 vm_prot_t *out_prot, /* OUT */
3589 boolean_t *wired) /* OUT */
3591 vm_map_entry_t entry;
3592 vm_map_t map = *var_map;
3594 vm_prot_t fault_type = fault_typea;
3595 int use_read_lock = 1;
3596 int rv = KERN_SUCCESS;
3600 vm_map_lock_read(map);
3605 * If the map has an interesting hint, try it before calling full
3606 * blown lookup routine.
3611 if ((entry == &map->header) ||
3612 (vaddr < entry->start) || (vaddr >= entry->end)) {
3613 vm_map_entry_t tmp_entry;
3616 * Entry was either not a valid hint, or the vaddr was not
3617 * contained in the entry, so do a full lookup.
3619 if (!vm_map_lookup_entry(map, vaddr, &tmp_entry)) {
3620 rv = KERN_INVALID_ADDRESS;
3631 if (entry->maptype == VM_MAPTYPE_SUBMAP) {
3632 vm_map_t old_map = map;
3634 *var_map = map = entry->object.sub_map;
3636 vm_map_unlock_read(old_map);
3638 vm_map_unlock(old_map);
3644 * Check whether this task is allowed to have this page.
3645 * Note the special case for MAP_ENTRY_COW
3646 * pages with an override. This is to implement a forced
3647 * COW for debuggers.
3650 if (fault_type & VM_PROT_OVERRIDE_WRITE)
3651 prot = entry->max_protection;
3653 prot = entry->protection;
3655 fault_type &= (VM_PROT_READ|VM_PROT_WRITE|VM_PROT_EXECUTE);
3656 if ((fault_type & prot) != fault_type) {
3657 rv = KERN_PROTECTION_FAILURE;
3661 if ((entry->eflags & MAP_ENTRY_USER_WIRED) &&
3662 (entry->eflags & MAP_ENTRY_COW) &&
3663 (fault_type & VM_PROT_WRITE) &&
3664 (fault_typea & VM_PROT_OVERRIDE_WRITE) == 0) {
3665 rv = KERN_PROTECTION_FAILURE;
3670 * If this page is not pageable, we have to get it for all possible
3673 *wired = (entry->wired_count != 0);
3675 prot = fault_type = entry->protection;
3678 * Virtual page tables may need to update the accessed (A) bit
3679 * in a page table entry. Upgrade the fault to a write fault for
3680 * that case if the map will support it. If the map does not support
3681 * it the page table entry simply will not be updated.
3683 if (entry->maptype == VM_MAPTYPE_VPAGETABLE) {
3684 if (prot & VM_PROT_WRITE)
3685 fault_type |= VM_PROT_WRITE;
3689 * If the entry was copy-on-write, we either ...
3691 if (entry->eflags & MAP_ENTRY_NEEDS_COPY) {
3693 * If we want to write the page, we may as well handle that
3694 * now since we've got the map locked.
3696 * If we don't need to write the page, we just demote the
3697 * permissions allowed.
3700 if (fault_type & VM_PROT_WRITE) {
3702 * Make a new object, and place it in the object
3703 * chain. Note that no new references have appeared
3704 * -- one just moved from the map to the new
3708 if (use_read_lock && vm_map_lock_upgrade(map)) {
3714 vm_map_entry_shadow(entry);
3717 * We're attempting to read a copy-on-write page --
3718 * don't allow writes.
3721 prot &= ~VM_PROT_WRITE;
3726 * Create an object if necessary.
3728 if (entry->object.vm_object == NULL &&
3730 if (use_read_lock && vm_map_lock_upgrade(map)) {
3735 vm_map_entry_allocate_object(entry);
3739 * Return the object/offset from this entry. If the entry was
3740 * copy-on-write or empty, it has been fixed up.
3743 *pindex = OFF_TO_IDX((vaddr - entry->start) + entry->offset);
3744 *object = entry->object.vm_object;
3747 * Return whether this is the only map sharing this data. On
3748 * success we return with a read lock held on the map. On failure
3749 * we return with the map unlocked.
3753 if (rv == KERN_SUCCESS) {
3754 if (use_read_lock == 0)
3755 vm_map_lock_downgrade(map);
3756 } else if (use_read_lock) {
3757 vm_map_unlock_read(map);
3765 * Releases locks acquired by a vm_map_lookup()
3766 * (according to the handle returned by that lookup).
3768 * No other requirements.
3771 vm_map_lookup_done(vm_map_t map, vm_map_entry_t entry, int count)
3774 * Unlock the main-level map
3776 vm_map_unlock_read(map);
3778 vm_map_entry_release(count);
3781 #include "opt_ddb.h"
3783 #include <sys/kernel.h>
3785 #include <ddb/ddb.h>
3790 DB_SHOW_COMMAND(map, vm_map_print)
3793 /* XXX convert args. */
3794 vm_map_t map = (vm_map_t)addr;
3795 boolean_t full = have_addr;
3797 vm_map_entry_t entry;
3799 db_iprintf("Task map %p: pmap=%p, nentries=%d, version=%u\n",
3801 (void *)map->pmap, map->nentries, map->timestamp);
3804 if (!full && db_indent)
3808 for (entry = map->header.next; entry != &map->header;
3809 entry = entry->next) {
3810 db_iprintf("map entry %p: start=%p, end=%p\n",
3811 (void *)entry, (void *)entry->start, (void *)entry->end);
3814 static char *inheritance_name[4] =
3815 {"share", "copy", "none", "donate_copy"};
3817 db_iprintf(" prot=%x/%x/%s",
3819 entry->max_protection,
3820 inheritance_name[(int)(unsigned char)entry->inheritance]);
3821 if (entry->wired_count != 0)
3822 db_printf(", wired");
3824 if (entry->maptype == VM_MAPTYPE_SUBMAP) {
3825 /* XXX no %qd in kernel. Truncate entry->offset. */
3826 db_printf(", share=%p, offset=0x%lx\n",
3827 (void *)entry->object.sub_map,
3828 (long)entry->offset);
3830 if ((entry->prev == &map->header) ||
3831 (entry->prev->object.sub_map !=
3832 entry->object.sub_map)) {
3834 vm_map_print((db_expr_t)(intptr_t)
3835 entry->object.sub_map,
3840 /* XXX no %qd in kernel. Truncate entry->offset. */
3841 db_printf(", object=%p, offset=0x%lx",
3842 (void *)entry->object.vm_object,
3843 (long)entry->offset);
3844 if (entry->eflags & MAP_ENTRY_COW)
3845 db_printf(", copy (%s)",
3846 (entry->eflags & MAP_ENTRY_NEEDS_COPY) ? "needed" : "done");
3850 if ((entry->prev == &map->header) ||
3851 (entry->prev->object.vm_object !=
3852 entry->object.vm_object)) {
3854 vm_object_print((db_expr_t)(intptr_t)
3855 entry->object.vm_object,
3870 DB_SHOW_COMMAND(procvm, procvm)
3875 p = (struct proc *) addr;
3880 db_printf("p = %p, vmspace = %p, map = %p, pmap = %p\n",
3881 (void *)p, (void *)p->p_vmspace, (void *)&p->p_vmspace->vm_map,
3882 (void *)vmspace_pmap(p->p_vmspace));
3884 vm_map_print((db_expr_t)(intptr_t)&p->p_vmspace->vm_map, 1, 0, NULL);