2 * Copyright (c) 1991, 1993
3 * The Regents of the University of California. All rights reserved.
5 * This code is derived from software contributed to Berkeley by
6 * The Mach Operating System project at Carnegie-Mellon University.
8 * Redistribution and use in source and binary forms, with or without
9 * modification, are permitted provided that the following conditions
11 * 1. Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * 2. Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in the
15 * documentation and/or other materials provided with the distribution.
16 * 3. All advertising materials mentioning features or use of this software
17 * must display the following acknowledgement:
18 * This product includes software developed by the University of
19 * California, Berkeley and its contributors.
20 * 4. Neither the name of the University nor the names of its contributors
21 * may be used to endorse or promote products derived from this software
22 * without specific prior written permission.
24 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
25 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
26 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
27 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
28 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
29 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
30 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
31 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
32 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
33 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
36 * from: @(#)vm_map.c 8.3 (Berkeley) 1/12/94
39 * Copyright (c) 1987, 1990 Carnegie-Mellon University.
40 * All rights reserved.
42 * Authors: Avadis Tevanian, Jr., Michael Wayne Young
44 * Permission to use, copy, modify and distribute this software and
45 * its documentation is hereby granted, provided that both the copyright
46 * notice and this permission notice appear in all copies of the
47 * software, derivative works or modified versions, and any portions
48 * thereof, and that both notices appear in supporting documentation.
50 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
51 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
52 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
54 * Carnegie Mellon requests users of this software to return to
56 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU
57 * School of Computer Science
58 * Carnegie Mellon University
59 * Pittsburgh PA 15213-3890
61 * any improvements or extensions that they make and grant Carnegie the
62 * rights to redistribute these changes.
64 * $FreeBSD: src/sys/vm/vm_map.c,v 1.187.2.19 2003/05/27 00:47:02 alc Exp $
65 * $DragonFly: src/sys/vm/vm_map.c,v 1.44 2006/05/17 17:47:58 dillon Exp $
69 * Virtual memory mapping module.
72 #include <sys/param.h>
73 #include <sys/systm.h>
76 #include <sys/vmmeter.h>
78 #include <sys/vnode.h>
79 #include <sys/resourcevar.h>
84 #include <vm/vm_param.h>
86 #include <vm/vm_map.h>
87 #include <vm/vm_page.h>
88 #include <vm/vm_object.h>
89 #include <vm/vm_pager.h>
90 #include <vm/vm_kern.h>
91 #include <vm/vm_extern.h>
92 #include <vm/swap_pager.h>
93 #include <vm/vm_zone.h>
95 #include <sys/thread2.h>
98 * Virtual memory maps provide for the mapping, protection,
99 * and sharing of virtual memory objects. In addition,
100 * this module provides for an efficient virtual copy of
101 * memory from one map to another.
103 * Synchronization is required prior to most operations.
105 * Maps consist of an ordered doubly-linked list of simple
106 * entries; a single hint is used to speed up lookups.
108 * Since portions of maps are specified by start/end addresses,
109 * which may not align with existing map entries, all
110 * routines merely "clip" entries to these start/end values.
111 * [That is, an entry is split into two, bordering at a
112 * start or end value.] Note that these clippings may not
113 * always be necessary (as the two resulting entries are then
114 * not changed); however, the clipping is done for convenience.
116 * As mentioned above, virtual copy operations are performed
117 * by copying VM object references from one map to
118 * another, and then marking both regions as copy-on-write.
124 * Initialize the vm_map module. Must be called before
125 * any other vm_map routines.
127 * Map and entry structures are allocated from the general
128 * purpose memory pool with some exceptions:
130 * - The kernel map and kmem submap are allocated statically.
131 * - Kernel map entries are allocated out of a static pool.
133 * These restrictions are necessary since malloc() uses the
134 * maps and requires map entries.
139 static struct vm_zone mapentzone_store, mapzone_store;
140 static vm_zone_t mapentzone, mapzone, vmspace_zone;
141 static struct vm_object mapentobj, mapobj;
143 static struct vm_map_entry map_entry_init[MAX_MAPENT];
144 static struct vm_map_entry cpu_map_entry_init[MAXCPU][VMEPERCPU];
145 static struct vm_map map_init[MAX_KMAP];
147 static vm_map_entry_t vm_map_entry_create(vm_map_t map, int *);
148 static void vm_map_entry_dispose (vm_map_t map, vm_map_entry_t entry, int *);
149 static void _vm_map_clip_end (vm_map_t, vm_map_entry_t, vm_offset_t, int *);
150 static void _vm_map_clip_start (vm_map_t, vm_map_entry_t, vm_offset_t, int *);
151 static void vm_map_entry_delete (vm_map_t, vm_map_entry_t, int *);
152 static void vm_map_entry_unwire (vm_map_t, vm_map_entry_t);
153 static void vm_map_copy_entry (vm_map_t, vm_map_t, vm_map_entry_t,
155 static void vm_map_split (vm_map_entry_t);
156 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);
161 mapzone = &mapzone_store;
162 zbootinit(mapzone, "MAP", sizeof (struct vm_map),
164 mapentzone = &mapentzone_store;
165 zbootinit(mapentzone, "MAP ENTRY", sizeof (struct vm_map_entry),
166 map_entry_init, MAX_MAPENT);
170 * Red black tree functions
172 static int rb_vm_map_compare(vm_map_entry_t a, vm_map_entry_t b);
173 RB_GENERATE(vm_map_rb_tree, vm_map_entry, rb_entry, rb_vm_map_compare);
175 /* a->start is address, and the only field has to be initialized */
177 rb_vm_map_compare(vm_map_entry_t a, vm_map_entry_t b)
179 if (a->start < b->start)
181 else if (a->start > b->start)
187 * Allocate a vmspace structure, including a vm_map and pmap,
188 * and initialize those structures. The refcnt is set to 1.
189 * The remaining fields must be initialized by the caller.
192 vmspace_alloc(vm_offset_t min, vm_offset_t max)
196 vm = zalloc(vmspace_zone);
197 vm_map_init(&vm->vm_map, min, max);
198 pmap_pinit(vmspace_pmap(vm));
199 vm->vm_map.pmap = vmspace_pmap(vm); /* XXX */
202 vm->vm_exitingcnt = 0;
209 zinitna(mapentzone, &mapentobj, NULL, 0, 0,
210 ZONE_USE_RESERVE | ZONE_SPECIAL, 1);
211 zinitna(mapzone, &mapobj, NULL, 0, 0, 0, 1);
212 vmspace_zone = zinit("VMSPACE", sizeof (struct vmspace), 0, 0, 3);
218 vmspace_dofree(struct vmspace *vm)
223 * Make sure any SysV shm is freed, it might not have in
228 KKASSERT(vm->vm_upcalls == NULL);
231 * Lock the map, to wait out all other references to it.
232 * Delete all of the mappings and pages they hold, then call
233 * the pmap module to reclaim anything left.
235 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
236 vm_map_lock(&vm->vm_map);
237 vm_map_delete(&vm->vm_map, vm->vm_map.min_offset,
238 vm->vm_map.max_offset, &count);
239 vm_map_unlock(&vm->vm_map);
240 vm_map_entry_release(count);
242 pmap_release(vmspace_pmap(vm));
243 zfree(vmspace_zone, vm);
247 vmspace_free(struct vmspace *vm)
249 if (vm->vm_refcnt == 0)
250 panic("vmspace_free: attempt to free already freed vmspace");
252 if (--vm->vm_refcnt == 0 && vm->vm_exitingcnt == 0)
257 vmspace_exitfree(struct proc *p)
265 * cleanup by parent process wait()ing on exiting child. vm_refcnt
266 * may not be 0 (e.g. fork() and child exits without exec()ing).
267 * exitingcnt may increment above 0 and drop back down to zero
268 * several times while vm_refcnt is held non-zero. vm_refcnt
269 * may also increment above 0 and drop back down to zero several
270 * times while vm_exitingcnt is held non-zero.
272 * The last wait on the exiting child's vmspace will clean up
273 * the remainder of the vmspace.
275 if (--vm->vm_exitingcnt == 0 && vm->vm_refcnt == 0)
280 * vmspace_swap_count() - count the approximate swap useage in pages for a
283 * Swap useage is determined by taking the proportional swap used by
284 * VM objects backing the VM map. To make up for fractional losses,
285 * if the VM object has any swap use at all the associated map entries
286 * count for at least 1 swap page.
289 vmspace_swap_count(struct vmspace *vmspace)
291 vm_map_t map = &vmspace->vm_map;
295 for (cur = map->header.next; cur != &map->header; cur = cur->next) {
298 if ((cur->eflags & MAP_ENTRY_IS_SUB_MAP) == 0 &&
299 (object = cur->object.vm_object) != NULL &&
300 object->type == OBJT_SWAP
302 int n = (cur->end - cur->start) / PAGE_SIZE;
304 if (object->un_pager.swp.swp_bcount) {
305 count += object->un_pager.swp.swp_bcount *
306 SWAP_META_PAGES * n / object->size + 1;
317 * Creates and returns a new empty VM map with
318 * the given physical map structure, and having
319 * the given lower and upper address bounds.
322 vm_map_create(pmap_t pmap, vm_offset_t min, vm_offset_t max)
326 result = zalloc(mapzone);
327 vm_map_init(result, min, max);
333 * Initialize an existing vm_map structure
334 * such as that in the vmspace structure.
335 * The pmap is set elsewhere.
338 vm_map_init(struct vm_map *map, vm_offset_t min, vm_offset_t max)
340 map->header.next = map->header.prev = &map->header;
341 RB_INIT(&map->rb_root);
346 map->min_offset = min;
347 map->max_offset = max;
348 map->first_free = &map->header;
349 map->hint = &map->header;
351 lockinit(&map->lock, "thrd_sleep", 0, LK_NOPAUSE);
355 * vm_map_entry_reserve_cpu_init:
357 * Set an initial negative count so the first attempt to reserve
358 * space preloads a bunch of vm_map_entry's for this cpu. Also
359 * pre-allocate 2 vm_map_entries which will be needed by zalloc() to
360 * map a new page for vm_map_entry structures. SMP systems are
361 * particularly sensitive.
363 * This routine is called in early boot so we cannot just call
364 * vm_map_entry_reserve().
366 * May be called for a gd other then mycpu, but may only be called
370 vm_map_entry_reserve_cpu_init(globaldata_t gd)
372 vm_map_entry_t entry;
375 gd->gd_vme_avail -= MAP_RESERVE_COUNT * 2;
376 entry = &cpu_map_entry_init[gd->gd_cpuid][0];
377 for (i = 0; i < VMEPERCPU; ++i, ++entry) {
378 entry->next = gd->gd_vme_base;
379 gd->gd_vme_base = entry;
384 * vm_map_entry_reserve:
386 * Reserves vm_map_entry structures so code later on can manipulate
387 * map_entry structures within a locked map without blocking trying
388 * to allocate a new vm_map_entry.
391 vm_map_entry_reserve(int count)
393 struct globaldata *gd = mycpu;
394 vm_map_entry_t entry;
399 * Make sure we have enough structures in gd_vme_base to handle
400 * the reservation request.
402 while (gd->gd_vme_avail < count) {
403 entry = zalloc(mapentzone);
404 entry->next = gd->gd_vme_base;
405 gd->gd_vme_base = entry;
408 gd->gd_vme_avail -= count;
414 * vm_map_entry_release:
416 * Releases previously reserved vm_map_entry structures that were not
417 * used. If we have too much junk in our per-cpu cache clean some of
421 vm_map_entry_release(int count)
423 struct globaldata *gd = mycpu;
424 vm_map_entry_t entry;
427 gd->gd_vme_avail += count;
428 while (gd->gd_vme_avail > MAP_RESERVE_SLOP) {
429 entry = gd->gd_vme_base;
430 KKASSERT(entry != NULL);
431 gd->gd_vme_base = entry->next;
434 zfree(mapentzone, entry);
441 * vm_map_entry_kreserve:
443 * Reserve map entry structures for use in kernel_map itself. These
444 * entries have *ALREADY* been reserved on a per-cpu basis when the map
445 * was inited. This function is used by zalloc() to avoid a recursion
446 * when zalloc() itself needs to allocate additional kernel memory.
448 * This function works like the normal reserve but does not load the
449 * vm_map_entry cache (because that would result in an infinite
450 * recursion). Note that gd_vme_avail may go negative. This is expected.
452 * Any caller of this function must be sure to renormalize after
453 * potentially eating entries to ensure that the reserve supply
457 vm_map_entry_kreserve(int count)
459 struct globaldata *gd = mycpu;
462 gd->gd_vme_avail -= count;
464 KASSERT(gd->gd_vme_base != NULL, ("no reserved entries left, gd_vme_avail = %d\n", gd->gd_vme_avail));
469 * vm_map_entry_krelease:
471 * Release previously reserved map entries for kernel_map. We do not
472 * attempt to clean up like the normal release function as this would
473 * cause an unnecessary (but probably not fatal) deep procedure call.
476 vm_map_entry_krelease(int count)
478 struct globaldata *gd = mycpu;
481 gd->gd_vme_avail += count;
486 * vm_map_entry_create: [ internal use only ]
488 * Allocates a VM map entry for insertion. No entry fields are filled
491 * This routine may be called from an interrupt thread but not a FAST
492 * interrupt. This routine may recurse the map lock.
494 static vm_map_entry_t
495 vm_map_entry_create(vm_map_t map, int *countp)
497 struct globaldata *gd = mycpu;
498 vm_map_entry_t entry;
500 KKASSERT(*countp > 0);
503 entry = gd->gd_vme_base;
504 KASSERT(entry != NULL, ("gd_vme_base NULL! count %d", *countp));
505 gd->gd_vme_base = entry->next;
511 * vm_map_entry_dispose: [ internal use only ]
513 * Dispose of a vm_map_entry that is no longer being referenced. This
514 * function may be called from an interrupt.
517 vm_map_entry_dispose(vm_map_t map, vm_map_entry_t entry, int *countp)
519 struct globaldata *gd = mycpu;
521 KKASSERT(map->hint != entry);
522 KKASSERT(map->first_free != entry);
526 entry->next = gd->gd_vme_base;
527 gd->gd_vme_base = entry;
533 * vm_map_entry_{un,}link:
535 * Insert/remove entries from maps.
538 vm_map_entry_link(vm_map_t map,
539 vm_map_entry_t after_where,
540 vm_map_entry_t entry)
543 entry->prev = after_where;
544 entry->next = after_where->next;
545 entry->next->prev = entry;
546 after_where->next = entry;
547 if (vm_map_rb_tree_RB_INSERT(&map->rb_root, entry))
548 panic("vm_map_entry_link: dup addr map %p ent %p", map, entry);
552 vm_map_entry_unlink(vm_map_t map,
553 vm_map_entry_t entry)
558 if (entry->eflags & MAP_ENTRY_IN_TRANSITION)
559 panic("vm_map_entry_unlink: attempt to mess with locked entry! %p", entry);
564 vm_map_rb_tree_RB_REMOVE(&map->rb_root, entry);
569 * vm_map_lookup_entry: [ internal use only ]
571 * Finds the map entry containing (or
572 * immediately preceding) the specified address
573 * in the given map; the entry is returned
574 * in the "entry" parameter. The boolean
575 * result indicates whether the address is
576 * actually contained in the map.
579 vm_map_lookup_entry(vm_map_t map, vm_offset_t address,
580 vm_map_entry_t *entry /* OUT */)
587 * XXX TEMPORARILY DISABLED. For some reason our attempt to revive
588 * the hint code with the red-black lookup meets with system crashes
589 * and lockups. We do not yet know why.
591 * It is possible that the problem is related to the setting
592 * of the hint during map_entry deletion, in the code specified
593 * at the GGG comment later on in this file.
596 * Quickly check the cached hint, there's a good chance of a match.
598 if (map->hint != &map->header) {
600 if (address >= tmp->start && address < tmp->end) {
608 * Locate the record from the top of the tree. 'last' tracks the
609 * closest prior record and is returned if no match is found, which
610 * in binary tree terms means tracking the most recent right-branch
611 * taken. If there is no prior record, &map->header is returned.
614 tmp = RB_ROOT(&map->rb_root);
617 if (address >= tmp->start) {
618 if (address < tmp->end) {
624 tmp = RB_RIGHT(tmp, rb_entry);
626 tmp = RB_LEFT(tmp, rb_entry);
636 * Inserts the given whole VM object into the target
637 * map at the specified address range. The object's
638 * size should match that of the address range.
640 * Requires that the map be locked, and leaves it so. Requires that
641 * sufficient vm_map_entry structures have been reserved and tracks
642 * the use via countp.
644 * If object is non-NULL, ref count must be bumped by caller
645 * prior to making call to account for the new entry.
648 vm_map_insert(vm_map_t map, int *countp,
649 vm_object_t object, vm_ooffset_t offset,
650 vm_offset_t start, vm_offset_t end, vm_prot_t prot, vm_prot_t max,
653 vm_map_entry_t new_entry;
654 vm_map_entry_t prev_entry;
655 vm_map_entry_t temp_entry;
656 vm_eflags_t protoeflags;
659 * Check that the start and end points are not bogus.
662 if ((start < map->min_offset) || (end > map->max_offset) ||
664 return (KERN_INVALID_ADDRESS);
667 * Find the entry prior to the proposed starting address; if it's part
668 * of an existing entry, this range is bogus.
671 if (vm_map_lookup_entry(map, start, &temp_entry))
672 return (KERN_NO_SPACE);
674 prev_entry = temp_entry;
677 * Assert that the next entry doesn't overlap the end point.
680 if ((prev_entry->next != &map->header) &&
681 (prev_entry->next->start < end))
682 return (KERN_NO_SPACE);
686 if (cow & MAP_COPY_ON_WRITE)
687 protoeflags |= MAP_ENTRY_COW|MAP_ENTRY_NEEDS_COPY;
689 if (cow & MAP_NOFAULT) {
690 protoeflags |= MAP_ENTRY_NOFAULT;
692 KASSERT(object == NULL,
693 ("vm_map_insert: paradoxical MAP_NOFAULT request"));
695 if (cow & MAP_DISABLE_SYNCER)
696 protoeflags |= MAP_ENTRY_NOSYNC;
697 if (cow & MAP_DISABLE_COREDUMP)
698 protoeflags |= MAP_ENTRY_NOCOREDUMP;
702 * When object is non-NULL, it could be shared with another
703 * process. We have to set or clear OBJ_ONEMAPPING
706 if ((object->ref_count > 1) || (object->shadow_count != 0)) {
707 vm_object_clear_flag(object, OBJ_ONEMAPPING);
710 else if ((prev_entry != &map->header) &&
711 (prev_entry->eflags == protoeflags) &&
712 (prev_entry->end == start) &&
713 (prev_entry->wired_count == 0) &&
714 ((prev_entry->object.vm_object == NULL) ||
715 vm_object_coalesce(prev_entry->object.vm_object,
716 OFF_TO_IDX(prev_entry->offset),
717 (vm_size_t)(prev_entry->end - prev_entry->start),
718 (vm_size_t)(end - prev_entry->end)))) {
720 * We were able to extend the object. Determine if we
721 * can extend the previous map entry to include the
724 if ((prev_entry->inheritance == VM_INHERIT_DEFAULT) &&
725 (prev_entry->protection == prot) &&
726 (prev_entry->max_protection == max)) {
727 map->size += (end - prev_entry->end);
728 prev_entry->end = end;
729 vm_map_simplify_entry(map, prev_entry, countp);
730 return (KERN_SUCCESS);
734 * If we can extend the object but cannot extend the
735 * map entry, we have to create a new map entry. We
736 * must bump the ref count on the extended object to
737 * account for it. object may be NULL.
739 object = prev_entry->object.vm_object;
740 offset = prev_entry->offset +
741 (prev_entry->end - prev_entry->start);
742 vm_object_reference(object);
746 * NOTE: if conditionals fail, object can be NULL here. This occurs
747 * in things like the buffer map where we manage kva but do not manage
755 new_entry = vm_map_entry_create(map, countp);
756 new_entry->start = start;
757 new_entry->end = end;
759 new_entry->eflags = protoeflags;
760 new_entry->object.vm_object = object;
761 new_entry->offset = offset;
762 new_entry->avail_ssize = 0;
764 new_entry->inheritance = VM_INHERIT_DEFAULT;
765 new_entry->protection = prot;
766 new_entry->max_protection = max;
767 new_entry->wired_count = 0;
770 * Insert the new entry into the list
773 vm_map_entry_link(map, prev_entry, new_entry);
774 map->size += new_entry->end - new_entry->start;
777 * Update the free space hint
779 if ((map->first_free == prev_entry) &&
780 (prev_entry->end >= new_entry->start)) {
781 map->first_free = new_entry;
786 * Temporarily removed to avoid MAP_STACK panic, due to
787 * MAP_STACK being a huge hack. Will be added back in
788 * when MAP_STACK (and the user stack mapping) is fixed.
791 * It may be possible to simplify the entry
793 vm_map_simplify_entry(map, new_entry, countp);
796 if (cow & (MAP_PREFAULT|MAP_PREFAULT_PARTIAL)) {
797 pmap_object_init_pt(map->pmap, start, prot,
798 object, OFF_TO_IDX(offset), end - start,
799 cow & MAP_PREFAULT_PARTIAL);
802 return (KERN_SUCCESS);
806 * Find sufficient space for `length' bytes in the given map, starting at
807 * `start'. The map must be locked. Returns 0 on success, 1 on no space.
809 * This function will returned an arbitrarily aligned pointer. If no
810 * particular alignment is required you should pass align as 1. Note that
811 * the map may return PAGE_SIZE aligned pointers if all the lengths used in
812 * the map are a multiple of PAGE_SIZE, even if you pass a smaller align
815 * 'align' should be a power of 2 but is not required to be.
825 vm_map_entry_t entry, next;
827 vm_offset_t align_mask;
829 if (start < map->min_offset)
830 start = map->min_offset;
831 if (start > map->max_offset)
835 * If the alignment is not a power of 2 we will have to use
836 * a mod/division, set align_mask to a special value.
838 if ((align | (align - 1)) + 1 != (align << 1))
839 align_mask = (vm_offset_t)-1;
841 align_mask = align - 1;
845 * Look for the first possible address; if there's already something
846 * at this address, we have to start after it.
848 if (start == map->min_offset) {
849 if ((entry = map->first_free) != &map->header)
854 if (vm_map_lookup_entry(map, start, &tmp))
860 * Look through the rest of the map, trying to fit a new region in the
861 * gap between existing regions, or after the very last region.
863 for (;; start = (entry = next)->end) {
865 * Adjust the proposed start by the requested alignment,
866 * be sure that we didn't wrap the address.
868 if (align_mask == (vm_offset_t)-1)
869 end = ((start + align - 1) / align) * align;
871 end = (start + align_mask) & ~align_mask;
876 * Find the end of the proposed new region. Be sure we didn't
877 * go beyond the end of the map, or wrap around the address.
878 * Then check to see if this is the last entry or if the
879 * proposed end fits in the gap between this and the next
882 end = start + length;
883 if (end > map->max_offset || end < start)
886 if (next == &map->header || next->start >= end)
890 if (map == kernel_map) {
892 if ((ksize = round_page(start + length)) > kernel_vm_end) {
893 pmap_growkernel(ksize);
902 * vm_map_find finds an unallocated region in the target address
903 * map with the given length. The search is defined to be
904 * first-fit from the specified address; the region found is
905 * returned in the same parameter.
907 * If object is non-NULL, ref count must be bumped by caller
908 * prior to making call to account for the new entry.
911 vm_map_find(vm_map_t map, vm_object_t object, vm_ooffset_t offset,
912 vm_offset_t *addr, /* IN/OUT */
913 vm_size_t length, boolean_t find_space, vm_prot_t prot,
914 vm_prot_t max, int cow)
922 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
925 if (vm_map_findspace(map, start, length, 1, addr)) {
927 vm_map_entry_release(count);
928 return (KERN_NO_SPACE);
932 result = vm_map_insert(map, &count, object, offset,
933 start, start + length, prot, max, cow);
935 vm_map_entry_release(count);
941 * vm_map_simplify_entry:
943 * Simplify the given map entry by merging with either neighbor. This
944 * routine also has the ability to merge with both neighbors.
946 * The map must be locked.
948 * This routine guarentees that the passed entry remains valid (though
949 * possibly extended). When merging, this routine may delete one or
950 * both neighbors. No action is taken on entries which have their
951 * in-transition flag set.
954 vm_map_simplify_entry(vm_map_t map, vm_map_entry_t entry, int *countp)
956 vm_map_entry_t next, prev;
957 vm_size_t prevsize, esize;
959 if (entry->eflags & (MAP_ENTRY_IN_TRANSITION | MAP_ENTRY_IS_SUB_MAP)) {
960 ++mycpu->gd_cnt.v_intrans_coll;
965 if (prev != &map->header) {
966 prevsize = prev->end - prev->start;
967 if ( (prev->end == entry->start) &&
968 (prev->object.vm_object == entry->object.vm_object) &&
969 (!prev->object.vm_object ||
970 (prev->offset + prevsize == entry->offset)) &&
971 (prev->eflags == entry->eflags) &&
972 (prev->protection == entry->protection) &&
973 (prev->max_protection == entry->max_protection) &&
974 (prev->inheritance == entry->inheritance) &&
975 (prev->wired_count == entry->wired_count)) {
976 if (map->first_free == prev)
977 map->first_free = entry;
978 if (map->hint == prev)
980 vm_map_entry_unlink(map, prev);
981 entry->start = prev->start;
982 entry->offset = prev->offset;
983 if (prev->object.vm_object)
984 vm_object_deallocate(prev->object.vm_object);
985 vm_map_entry_dispose(map, prev, countp);
990 if (next != &map->header) {
991 esize = entry->end - entry->start;
992 if ((entry->end == next->start) &&
993 (next->object.vm_object == entry->object.vm_object) &&
994 (!entry->object.vm_object ||
995 (entry->offset + esize == next->offset)) &&
996 (next->eflags == entry->eflags) &&
997 (next->protection == entry->protection) &&
998 (next->max_protection == entry->max_protection) &&
999 (next->inheritance == entry->inheritance) &&
1000 (next->wired_count == entry->wired_count)) {
1001 if (map->first_free == next)
1002 map->first_free = entry;
1003 if (map->hint == next)
1005 vm_map_entry_unlink(map, next);
1006 entry->end = next->end;
1007 if (next->object.vm_object)
1008 vm_object_deallocate(next->object.vm_object);
1009 vm_map_entry_dispose(map, next, countp);
1014 * vm_map_clip_start: [ internal use only ]
1016 * Asserts that the given entry begins at or after
1017 * the specified address; if necessary,
1018 * it splits the entry into two.
1020 #define vm_map_clip_start(map, entry, startaddr, countp) \
1022 if (startaddr > entry->start) \
1023 _vm_map_clip_start(map, entry, startaddr, countp); \
1027 * This routine is called only when it is known that
1028 * the entry must be split.
1031 _vm_map_clip_start(vm_map_t map, vm_map_entry_t entry, vm_offset_t start, int *countp)
1033 vm_map_entry_t new_entry;
1036 * Split off the front portion -- note that we must insert the new
1037 * entry BEFORE this one, so that this entry has the specified
1041 vm_map_simplify_entry(map, entry, countp);
1044 * If there is no object backing this entry, we might as well create
1045 * one now. If we defer it, an object can get created after the map
1046 * is clipped, and individual objects will be created for the split-up
1047 * map. This is a bit of a hack, but is also about the best place to
1048 * put this improvement.
1051 if (entry->object.vm_object == NULL && !map->system_map) {
1053 object = vm_object_allocate(OBJT_DEFAULT,
1054 atop(entry->end - entry->start));
1055 entry->object.vm_object = object;
1059 new_entry = vm_map_entry_create(map, countp);
1060 *new_entry = *entry;
1062 new_entry->end = start;
1063 entry->offset += (start - entry->start);
1064 entry->start = start;
1066 vm_map_entry_link(map, entry->prev, new_entry);
1068 if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
1069 vm_object_reference(new_entry->object.vm_object);
1074 * vm_map_clip_end: [ internal use only ]
1076 * Asserts that the given entry ends at or before
1077 * the specified address; if necessary,
1078 * it splits the entry into two.
1081 #define vm_map_clip_end(map, entry, endaddr, countp) \
1083 if (endaddr < entry->end) \
1084 _vm_map_clip_end(map, entry, endaddr, countp); \
1088 * This routine is called only when it is known that
1089 * the entry must be split.
1092 _vm_map_clip_end(vm_map_t map, vm_map_entry_t entry, vm_offset_t end, int *countp)
1094 vm_map_entry_t new_entry;
1097 * If there is no object backing this entry, we might as well create
1098 * one now. If we defer it, an object can get created after the map
1099 * is clipped, and individual objects will be created for the split-up
1100 * map. This is a bit of a hack, but is also about the best place to
1101 * put this improvement.
1104 if (entry->object.vm_object == NULL && !map->system_map) {
1106 object = vm_object_allocate(OBJT_DEFAULT,
1107 atop(entry->end - entry->start));
1108 entry->object.vm_object = object;
1113 * Create a new entry and insert it AFTER the specified entry
1116 new_entry = vm_map_entry_create(map, countp);
1117 *new_entry = *entry;
1119 new_entry->start = entry->end = end;
1120 new_entry->offset += (end - entry->start);
1122 vm_map_entry_link(map, entry, new_entry);
1124 if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
1125 vm_object_reference(new_entry->object.vm_object);
1130 * VM_MAP_RANGE_CHECK: [ internal use only ]
1132 * Asserts that the starting and ending region
1133 * addresses fall within the valid range of the map.
1135 #define VM_MAP_RANGE_CHECK(map, start, end) \
1137 if (start < vm_map_min(map)) \
1138 start = vm_map_min(map); \
1139 if (end > vm_map_max(map)) \
1140 end = vm_map_max(map); \
1146 * vm_map_transition_wait: [ kernel use only ]
1148 * Used to block when an in-transition collison occurs. The map
1149 * is unlocked for the sleep and relocked before the return.
1153 vm_map_transition_wait(vm_map_t map)
1156 tsleep(map, 0, "vment", 0);
1164 * When we do blocking operations with the map lock held it is
1165 * possible that a clip might have occured on our in-transit entry,
1166 * requiring an adjustment to the entry in our loop. These macros
1167 * help the pageable and clip_range code deal with the case. The
1168 * conditional costs virtually nothing if no clipping has occured.
1171 #define CLIP_CHECK_BACK(entry, save_start) \
1173 while (entry->start != save_start) { \
1174 entry = entry->prev; \
1175 KASSERT(entry != &map->header, ("bad entry clip")); \
1179 #define CLIP_CHECK_FWD(entry, save_end) \
1181 while (entry->end != save_end) { \
1182 entry = entry->next; \
1183 KASSERT(entry != &map->header, ("bad entry clip")); \
1189 * vm_map_clip_range: [ kernel use only ]
1191 * Clip the specified range and return the base entry. The
1192 * range may cover several entries starting at the returned base
1193 * and the first and last entry in the covering sequence will be
1194 * properly clipped to the requested start and end address.
1196 * If no holes are allowed you should pass the MAP_CLIP_NO_HOLES
1199 * The MAP_ENTRY_IN_TRANSITION flag will be set for the entries
1200 * covered by the requested range.
1202 * The map must be exclusively locked on entry and will remain locked
1203 * on return. If no range exists or the range contains holes and you
1204 * specified that no holes were allowed, NULL will be returned. This
1205 * routine may temporarily unlock the map in order avoid a deadlock when
1210 vm_map_clip_range(vm_map_t map, vm_offset_t start, vm_offset_t end,
1211 int *countp, int flags)
1213 vm_map_entry_t start_entry;
1214 vm_map_entry_t entry;
1217 * Locate the entry and effect initial clipping. The in-transition
1218 * case does not occur very often so do not try to optimize it.
1221 if (vm_map_lookup_entry(map, start, &start_entry) == FALSE)
1223 entry = start_entry;
1224 if (entry->eflags & MAP_ENTRY_IN_TRANSITION) {
1225 entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP;
1226 ++mycpu->gd_cnt.v_intrans_coll;
1227 ++mycpu->gd_cnt.v_intrans_wait;
1228 vm_map_transition_wait(map);
1230 * entry and/or start_entry may have been clipped while
1231 * we slept, or may have gone away entirely. We have
1232 * to restart from the lookup.
1237 * Since we hold an exclusive map lock we do not have to restart
1238 * after clipping, even though clipping may block in zalloc.
1240 vm_map_clip_start(map, entry, start, countp);
1241 vm_map_clip_end(map, entry, end, countp);
1242 entry->eflags |= MAP_ENTRY_IN_TRANSITION;
1245 * Scan entries covered by the range. When working on the next
1246 * entry a restart need only re-loop on the current entry which
1247 * we have already locked, since 'next' may have changed. Also,
1248 * even though entry is safe, it may have been clipped so we
1249 * have to iterate forwards through the clip after sleeping.
1251 while (entry->next != &map->header && entry->next->start < end) {
1252 vm_map_entry_t next = entry->next;
1254 if (flags & MAP_CLIP_NO_HOLES) {
1255 if (next->start > entry->end) {
1256 vm_map_unclip_range(map, start_entry,
1257 start, entry->end, countp, flags);
1262 if (next->eflags & MAP_ENTRY_IN_TRANSITION) {
1263 vm_offset_t save_end = entry->end;
1264 next->eflags |= MAP_ENTRY_NEEDS_WAKEUP;
1265 ++mycpu->gd_cnt.v_intrans_coll;
1266 ++mycpu->gd_cnt.v_intrans_wait;
1267 vm_map_transition_wait(map);
1270 * clips might have occured while we blocked.
1272 CLIP_CHECK_FWD(entry, save_end);
1273 CLIP_CHECK_BACK(start_entry, start);
1277 * No restart necessary even though clip_end may block, we
1278 * are holding the map lock.
1280 vm_map_clip_end(map, next, end, countp);
1281 next->eflags |= MAP_ENTRY_IN_TRANSITION;
1284 if (flags & MAP_CLIP_NO_HOLES) {
1285 if (entry->end != end) {
1286 vm_map_unclip_range(map, start_entry,
1287 start, entry->end, countp, flags);
1291 return(start_entry);
1295 * vm_map_unclip_range: [ kernel use only ]
1297 * Undo the effect of vm_map_clip_range(). You should pass the same
1298 * flags and the same range that you passed to vm_map_clip_range().
1299 * This code will clear the in-transition flag on the entries and
1300 * wake up anyone waiting. This code will also simplify the sequence
1301 * and attempt to merge it with entries before and after the sequence.
1303 * The map must be locked on entry and will remain locked on return.
1305 * Note that you should also pass the start_entry returned by
1306 * vm_map_clip_range(). However, if you block between the two calls
1307 * with the map unlocked please be aware that the start_entry may
1308 * have been clipped and you may need to scan it backwards to find
1309 * the entry corresponding with the original start address. You are
1310 * responsible for this, vm_map_unclip_range() expects the correct
1311 * start_entry to be passed to it and will KASSERT otherwise.
1315 vm_map_unclip_range(
1317 vm_map_entry_t start_entry,
1323 vm_map_entry_t entry;
1325 entry = start_entry;
1327 KASSERT(entry->start == start, ("unclip_range: illegal base entry"));
1328 while (entry != &map->header && entry->start < end) {
1329 KASSERT(entry->eflags & MAP_ENTRY_IN_TRANSITION, ("in-transition flag not set during unclip on: %p", entry));
1330 KASSERT(entry->end <= end, ("unclip_range: tail wasn't clipped"));
1331 entry->eflags &= ~MAP_ENTRY_IN_TRANSITION;
1332 if (entry->eflags & MAP_ENTRY_NEEDS_WAKEUP) {
1333 entry->eflags &= ~MAP_ENTRY_NEEDS_WAKEUP;
1336 entry = entry->next;
1340 * Simplification does not block so there is no restart case.
1342 entry = start_entry;
1343 while (entry != &map->header && entry->start < end) {
1344 vm_map_simplify_entry(map, entry, countp);
1345 entry = entry->next;
1350 * vm_map_submap: [ kernel use only ]
1352 * Mark the given range as handled by a subordinate map.
1354 * This range must have been created with vm_map_find,
1355 * and no other operations may have been performed on this
1356 * range prior to calling vm_map_submap.
1358 * Only a limited number of operations can be performed
1359 * within this rage after calling vm_map_submap:
1361 * [Don't try vm_map_copy!]
1363 * To remove a submapping, one must first remove the
1364 * range from the superior map, and then destroy the
1365 * submap (if desired). [Better yet, don't try it.]
1368 vm_map_submap(vm_map_t map, vm_offset_t start, vm_offset_t end, vm_map_t submap)
1370 vm_map_entry_t entry;
1371 int result = KERN_INVALID_ARGUMENT;
1374 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
1377 VM_MAP_RANGE_CHECK(map, start, end);
1379 if (vm_map_lookup_entry(map, start, &entry)) {
1380 vm_map_clip_start(map, entry, start, &count);
1382 entry = entry->next;
1385 vm_map_clip_end(map, entry, end, &count);
1387 if ((entry->start == start) && (entry->end == end) &&
1388 ((entry->eflags & MAP_ENTRY_COW) == 0) &&
1389 (entry->object.vm_object == NULL)) {
1390 entry->object.sub_map = submap;
1391 entry->eflags |= MAP_ENTRY_IS_SUB_MAP;
1392 result = KERN_SUCCESS;
1395 vm_map_entry_release(count);
1403 * Sets the protection of the specified address
1404 * region in the target map. If "set_max" is
1405 * specified, the maximum protection is to be set;
1406 * otherwise, only the current protection is affected.
1409 vm_map_protect(vm_map_t map, vm_offset_t start, vm_offset_t end,
1410 vm_prot_t new_prot, boolean_t set_max)
1412 vm_map_entry_t current;
1413 vm_map_entry_t entry;
1416 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
1419 VM_MAP_RANGE_CHECK(map, start, end);
1421 if (vm_map_lookup_entry(map, start, &entry)) {
1422 vm_map_clip_start(map, entry, start, &count);
1424 entry = entry->next;
1428 * Make a first pass to check for protection violations.
1432 while ((current != &map->header) && (current->start < end)) {
1433 if (current->eflags & MAP_ENTRY_IS_SUB_MAP) {
1435 vm_map_entry_release(count);
1436 return (KERN_INVALID_ARGUMENT);
1438 if ((new_prot & current->max_protection) != new_prot) {
1440 vm_map_entry_release(count);
1441 return (KERN_PROTECTION_FAILURE);
1443 current = current->next;
1447 * Go back and fix up protections. [Note that clipping is not
1448 * necessary the second time.]
1452 while ((current != &map->header) && (current->start < end)) {
1455 vm_map_clip_end(map, current, end, &count);
1457 old_prot = current->protection;
1459 current->protection =
1460 (current->max_protection = new_prot) &
1463 current->protection = new_prot;
1466 * Update physical map if necessary. Worry about copy-on-write
1467 * here -- CHECK THIS XXX
1470 if (current->protection != old_prot) {
1471 #define MASK(entry) (((entry)->eflags & MAP_ENTRY_COW) ? ~VM_PROT_WRITE : \
1474 pmap_protect(map->pmap, current->start,
1476 current->protection & MASK(current));
1480 vm_map_simplify_entry(map, current, &count);
1482 current = current->next;
1486 vm_map_entry_release(count);
1487 return (KERN_SUCCESS);
1493 * This routine traverses a processes map handling the madvise
1494 * system call. Advisories are classified as either those effecting
1495 * the vm_map_entry structure, or those effecting the underlying
1500 vm_map_madvise(vm_map_t map, vm_offset_t start, vm_offset_t end, int behav)
1502 vm_map_entry_t current, entry;
1507 * Some madvise calls directly modify the vm_map_entry, in which case
1508 * we need to use an exclusive lock on the map and we need to perform
1509 * various clipping operations. Otherwise we only need a read-lock
1513 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
1517 case MADV_SEQUENTIAL:
1529 vm_map_lock_read(map);
1532 vm_map_entry_release(count);
1533 return (KERN_INVALID_ARGUMENT);
1537 * Locate starting entry and clip if necessary.
1540 VM_MAP_RANGE_CHECK(map, start, end);
1542 if (vm_map_lookup_entry(map, start, &entry)) {
1544 vm_map_clip_start(map, entry, start, &count);
1546 entry = entry->next;
1551 * madvise behaviors that are implemented in the vm_map_entry.
1553 * We clip the vm_map_entry so that behavioral changes are
1554 * limited to the specified address range.
1556 for (current = entry;
1557 (current != &map->header) && (current->start < end);
1558 current = current->next
1560 if (current->eflags & MAP_ENTRY_IS_SUB_MAP)
1563 vm_map_clip_end(map, current, end, &count);
1567 vm_map_entry_set_behavior(current, MAP_ENTRY_BEHAV_NORMAL);
1569 case MADV_SEQUENTIAL:
1570 vm_map_entry_set_behavior(current, MAP_ENTRY_BEHAV_SEQUENTIAL);
1573 vm_map_entry_set_behavior(current, MAP_ENTRY_BEHAV_RANDOM);
1576 current->eflags |= MAP_ENTRY_NOSYNC;
1579 current->eflags &= ~MAP_ENTRY_NOSYNC;
1582 current->eflags |= MAP_ENTRY_NOCOREDUMP;
1585 current->eflags &= ~MAP_ENTRY_NOCOREDUMP;
1590 vm_map_simplify_entry(map, current, &count);
1598 * madvise behaviors that are implemented in the underlying
1601 * Since we don't clip the vm_map_entry, we have to clip
1602 * the vm_object pindex and count.
1604 for (current = entry;
1605 (current != &map->header) && (current->start < end);
1606 current = current->next
1608 vm_offset_t useStart;
1610 if (current->eflags & MAP_ENTRY_IS_SUB_MAP)
1613 pindex = OFF_TO_IDX(current->offset);
1614 count = atop(current->end - current->start);
1615 useStart = current->start;
1617 if (current->start < start) {
1618 pindex += atop(start - current->start);
1619 count -= atop(start - current->start);
1622 if (current->end > end)
1623 count -= atop(current->end - end);
1628 vm_object_madvise(current->object.vm_object,
1629 pindex, count, behav);
1630 if (behav == MADV_WILLNEED) {
1631 pmap_object_init_pt(
1634 current->protection,
1635 current->object.vm_object,
1637 (count << PAGE_SHIFT),
1638 MAP_PREFAULT_MADVISE
1642 vm_map_unlock_read(map);
1644 vm_map_entry_release(count);
1652 * Sets the inheritance of the specified address
1653 * range in the target map. Inheritance
1654 * affects how the map will be shared with
1655 * child maps at the time of vm_map_fork.
1658 vm_map_inherit(vm_map_t map, vm_offset_t start, vm_offset_t end,
1659 vm_inherit_t new_inheritance)
1661 vm_map_entry_t entry;
1662 vm_map_entry_t temp_entry;
1665 switch (new_inheritance) {
1666 case VM_INHERIT_NONE:
1667 case VM_INHERIT_COPY:
1668 case VM_INHERIT_SHARE:
1671 return (KERN_INVALID_ARGUMENT);
1674 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
1677 VM_MAP_RANGE_CHECK(map, start, end);
1679 if (vm_map_lookup_entry(map, start, &temp_entry)) {
1681 vm_map_clip_start(map, entry, start, &count);
1683 entry = temp_entry->next;
1685 while ((entry != &map->header) && (entry->start < end)) {
1686 vm_map_clip_end(map, entry, end, &count);
1688 entry->inheritance = new_inheritance;
1690 vm_map_simplify_entry(map, entry, &count);
1692 entry = entry->next;
1695 vm_map_entry_release(count);
1696 return (KERN_SUCCESS);
1700 * Implement the semantics of mlock
1703 vm_map_unwire(vm_map_t map, vm_offset_t start, vm_offset_t real_end,
1704 boolean_t new_pageable)
1706 vm_map_entry_t entry;
1707 vm_map_entry_t start_entry;
1709 int rv = KERN_SUCCESS;
1712 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
1714 VM_MAP_RANGE_CHECK(map, start, real_end);
1717 start_entry = vm_map_clip_range(map, start, end, &count, MAP_CLIP_NO_HOLES);
1718 if (start_entry == NULL) {
1720 vm_map_entry_release(count);
1721 return (KERN_INVALID_ADDRESS);
1724 if (new_pageable == 0) {
1725 entry = start_entry;
1726 while ((entry != &map->header) && (entry->start < end)) {
1727 vm_offset_t save_start;
1728 vm_offset_t save_end;
1731 * Already user wired or hard wired (trivial cases)
1733 if (entry->eflags & MAP_ENTRY_USER_WIRED) {
1734 entry = entry->next;
1737 if (entry->wired_count != 0) {
1738 entry->wired_count++;
1739 entry->eflags |= MAP_ENTRY_USER_WIRED;
1740 entry = entry->next;
1745 * A new wiring requires instantiation of appropriate
1746 * management structures and the faulting in of the
1749 if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
1750 int copyflag = entry->eflags & MAP_ENTRY_NEEDS_COPY;
1751 if (copyflag && ((entry->protection & VM_PROT_WRITE) != 0)) {
1753 vm_object_shadow(&entry->object.vm_object,
1755 atop(entry->end - entry->start));
1756 entry->eflags &= ~MAP_ENTRY_NEEDS_COPY;
1758 } else if (entry->object.vm_object == NULL &&
1761 entry->object.vm_object =
1762 vm_object_allocate(OBJT_DEFAULT,
1763 atop(entry->end - entry->start));
1764 entry->offset = (vm_offset_t) 0;
1768 entry->wired_count++;
1769 entry->eflags |= MAP_ENTRY_USER_WIRED;
1772 * Now fault in the area. Note that vm_fault_wire()
1773 * may release the map lock temporarily, it will be
1774 * relocked on return. The in-transition
1775 * flag protects the entries.
1777 save_start = entry->start;
1778 save_end = entry->end;
1779 rv = vm_fault_wire(map, entry, TRUE);
1781 CLIP_CHECK_BACK(entry, save_start);
1783 KASSERT(entry->wired_count == 1, ("bad wired_count on entry"));
1784 entry->eflags &= ~MAP_ENTRY_USER_WIRED;
1785 entry->wired_count = 0;
1786 if (entry->end == save_end)
1788 entry = entry->next;
1789 KASSERT(entry != &map->header, ("bad entry clip during backout"));
1791 end = save_start; /* unwire the rest */
1795 * note that even though the entry might have been
1796 * clipped, the USER_WIRED flag we set prevents
1797 * duplication so we do not have to do a
1800 entry = entry->next;
1804 * If we failed fall through to the unwiring section to
1805 * unwire what we had wired so far. 'end' has already
1812 * start_entry might have been clipped if we unlocked the
1813 * map and blocked. No matter how clipped it has gotten
1814 * there should be a fragment that is on our start boundary.
1816 CLIP_CHECK_BACK(start_entry, start);
1820 * Deal with the unwiring case.
1824 * This is the unwiring case. We must first ensure that the
1825 * range to be unwired is really wired down. We know there
1828 entry = start_entry;
1829 while ((entry != &map->header) && (entry->start < end)) {
1830 if ((entry->eflags & MAP_ENTRY_USER_WIRED) == 0) {
1831 rv = KERN_INVALID_ARGUMENT;
1834 KASSERT(entry->wired_count != 0, ("wired count was 0 with USER_WIRED set! %p", entry));
1835 entry = entry->next;
1839 * Now decrement the wiring count for each region. If a region
1840 * becomes completely unwired, unwire its physical pages and
1844 * The map entries are processed in a loop, checking to
1845 * make sure the entry is wired and asserting it has a wired
1846 * count. However, another loop was inserted more-or-less in
1847 * the middle of the unwiring path. This loop picks up the
1848 * "entry" loop variable from the first loop without first
1849 * setting it to start_entry. Naturally, the secound loop
1850 * is never entered and the pages backing the entries are
1851 * never unwired. This can lead to a leak of wired pages.
1853 entry = start_entry;
1854 while ((entry != &map->header) && (entry->start < end)) {
1855 KASSERT(entry->eflags & MAP_ENTRY_USER_WIRED,
1856 ("expected USER_WIRED on entry %p", entry));
1857 entry->eflags &= ~MAP_ENTRY_USER_WIRED;
1858 entry->wired_count--;
1859 if (entry->wired_count == 0)
1860 vm_fault_unwire(map, entry);
1861 entry = entry->next;
1865 vm_map_unclip_range(map, start_entry, start, real_end, &count,
1869 vm_map_entry_release(count);
1876 * Sets the pageability of the specified address
1877 * range in the target map. Regions specified
1878 * as not pageable require locked-down physical
1879 * memory and physical page maps.
1881 * The map must not be locked, but a reference
1882 * must remain to the map throughout the call.
1884 * This function may be called via the zalloc path and must properly
1885 * reserve map entries for kernel_map.
1888 vm_map_wire(vm_map_t map, vm_offset_t start, vm_offset_t real_end, int kmflags)
1890 vm_map_entry_t entry;
1891 vm_map_entry_t start_entry;
1893 int rv = KERN_SUCCESS;
1896 if (kmflags & KM_KRESERVE)
1897 count = vm_map_entry_kreserve(MAP_RESERVE_COUNT);
1899 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
1901 VM_MAP_RANGE_CHECK(map, start, real_end);
1904 start_entry = vm_map_clip_range(map, start, end, &count, MAP_CLIP_NO_HOLES);
1905 if (start_entry == NULL) {
1907 rv = KERN_INVALID_ADDRESS;
1910 if ((kmflags & KM_PAGEABLE) == 0) {
1914 * 1. Holding the write lock, we create any shadow or zero-fill
1915 * objects that need to be created. Then we clip each map
1916 * entry to the region to be wired and increment its wiring
1917 * count. We create objects before clipping the map entries
1918 * to avoid object proliferation.
1920 * 2. We downgrade to a read lock, and call vm_fault_wire to
1921 * fault in the pages for any newly wired area (wired_count is
1924 * Downgrading to a read lock for vm_fault_wire avoids a
1925 * possible deadlock with another process that may have faulted
1926 * on one of the pages to be wired (it would mark the page busy,
1927 * blocking us, then in turn block on the map lock that we
1928 * hold). Because of problems in the recursive lock package,
1929 * we cannot upgrade to a write lock in vm_map_lookup. Thus,
1930 * any actions that require the write lock must be done
1931 * beforehand. Because we keep the read lock on the map, the
1932 * copy-on-write status of the entries we modify here cannot
1936 entry = start_entry;
1937 while ((entry != &map->header) && (entry->start < end)) {
1939 * Trivial case if the entry is already wired
1941 if (entry->wired_count) {
1942 entry->wired_count++;
1943 entry = entry->next;
1948 * The entry is being newly wired, we have to setup
1949 * appropriate management structures. A shadow
1950 * object is required for a copy-on-write region,
1951 * or a normal object for a zero-fill region. We
1952 * do not have to do this for entries that point to sub
1953 * maps because we won't hold the lock on the sub map.
1955 if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
1956 int copyflag = entry->eflags & MAP_ENTRY_NEEDS_COPY;
1958 ((entry->protection & VM_PROT_WRITE) != 0)) {
1960 vm_object_shadow(&entry->object.vm_object,
1962 atop(entry->end - entry->start));
1963 entry->eflags &= ~MAP_ENTRY_NEEDS_COPY;
1964 } else if (entry->object.vm_object == NULL &&
1966 entry->object.vm_object =
1967 vm_object_allocate(OBJT_DEFAULT,
1968 atop(entry->end - entry->start));
1969 entry->offset = (vm_offset_t) 0;
1973 entry->wired_count++;
1974 entry = entry->next;
1982 * HACK HACK HACK HACK
1984 * Unlock the map to avoid deadlocks. The in-transit flag
1985 * protects us from most changes but note that
1986 * clipping may still occur. To prevent clipping from
1987 * occuring after the unlock, except for when we are
1988 * blocking in vm_fault_wire, we must run in a critical
1989 * section, otherwise our accesses to entry->start and
1990 * entry->end could be corrupted. We have to enter the
1991 * critical section prior to unlocking so start_entry does
1992 * not change out from under us at the very beginning of the
1995 * HACK HACK HACK HACK
2000 entry = start_entry;
2001 while (entry != &map->header && entry->start < end) {
2003 * If vm_fault_wire fails for any page we need to undo
2004 * what has been done. We decrement the wiring count
2005 * for those pages which have not yet been wired (now)
2006 * and unwire those that have (later).
2008 vm_offset_t save_start = entry->start;
2009 vm_offset_t save_end = entry->end;
2011 if (entry->wired_count == 1)
2012 rv = vm_fault_wire(map, entry, FALSE);
2014 CLIP_CHECK_BACK(entry, save_start);
2016 KASSERT(entry->wired_count == 1, ("wired_count changed unexpectedly"));
2017 entry->wired_count = 0;
2018 if (entry->end == save_end)
2020 entry = entry->next;
2021 KASSERT(entry != &map->header, ("bad entry clip during backout"));
2026 CLIP_CHECK_FWD(entry, save_end);
2027 entry = entry->next;
2032 * If a failure occured undo everything by falling through
2033 * to the unwiring code. 'end' has already been adjusted
2037 kmflags |= KM_PAGEABLE;
2040 * start_entry is still IN_TRANSITION but may have been
2041 * clipped since vm_fault_wire() unlocks and relocks the
2042 * map. No matter how clipped it has gotten there should
2043 * be a fragment that is on our start boundary.
2045 CLIP_CHECK_BACK(start_entry, start);
2048 if (kmflags & KM_PAGEABLE) {
2050 * This is the unwiring case. We must first ensure that the
2051 * range to be unwired is really wired down. We know there
2054 entry = start_entry;
2055 while ((entry != &map->header) && (entry->start < end)) {
2056 if (entry->wired_count == 0) {
2057 rv = KERN_INVALID_ARGUMENT;
2060 entry = entry->next;
2064 * Now decrement the wiring count for each region. If a region
2065 * becomes completely unwired, unwire its physical pages and
2068 entry = start_entry;
2069 while ((entry != &map->header) && (entry->start < end)) {
2070 entry->wired_count--;
2071 if (entry->wired_count == 0)
2072 vm_fault_unwire(map, entry);
2073 entry = entry->next;
2077 vm_map_unclip_range(map, start_entry, start, real_end, &count,
2082 if (kmflags & KM_KRESERVE)
2083 vm_map_entry_krelease(count);
2085 vm_map_entry_release(count);
2090 * vm_map_set_wired_quick()
2092 * Mark a newly allocated address range as wired but do not fault in
2093 * the pages. The caller is expected to load the pages into the object.
2095 * The map must be locked on entry and will remain locked on return.
2098 vm_map_set_wired_quick(vm_map_t map, vm_offset_t addr, vm_size_t size, int *countp)
2100 vm_map_entry_t scan;
2101 vm_map_entry_t entry;
2103 entry = vm_map_clip_range(map, addr, addr + size, countp, MAP_CLIP_NO_HOLES);
2104 for (scan = entry; scan != &map->header && scan->start < addr + size; scan = scan->next) {
2105 KKASSERT(entry->wired_count == 0);
2106 entry->wired_count = 1;
2108 vm_map_unclip_range(map, entry, addr, addr + size, countp, MAP_CLIP_NO_HOLES);
2114 * Push any dirty cached pages in the address range to their pager.
2115 * If syncio is TRUE, dirty pages are written synchronously.
2116 * If invalidate is TRUE, any cached pages are freed as well.
2118 * Returns an error if any part of the specified range is not mapped.
2121 vm_map_clean(vm_map_t map, vm_offset_t start, vm_offset_t end, boolean_t syncio,
2122 boolean_t invalidate)
2124 vm_map_entry_t current;
2125 vm_map_entry_t entry;
2128 vm_ooffset_t offset;
2130 vm_map_lock_read(map);
2131 VM_MAP_RANGE_CHECK(map, start, end);
2132 if (!vm_map_lookup_entry(map, start, &entry)) {
2133 vm_map_unlock_read(map);
2134 return (KERN_INVALID_ADDRESS);
2137 * Make a first pass to check for holes.
2139 for (current = entry; current->start < end; current = current->next) {
2140 if (current->eflags & MAP_ENTRY_IS_SUB_MAP) {
2141 vm_map_unlock_read(map);
2142 return (KERN_INVALID_ARGUMENT);
2144 if (end > current->end &&
2145 (current->next == &map->header ||
2146 current->end != current->next->start)) {
2147 vm_map_unlock_read(map);
2148 return (KERN_INVALID_ADDRESS);
2153 pmap_remove(vm_map_pmap(map), start, end);
2155 * Make a second pass, cleaning/uncaching pages from the indicated
2158 for (current = entry; current->start < end; current = current->next) {
2159 offset = current->offset + (start - current->start);
2160 size = (end <= current->end ? end : current->end) - start;
2161 if (current->eflags & MAP_ENTRY_IS_SUB_MAP) {
2163 vm_map_entry_t tentry;
2166 smap = current->object.sub_map;
2167 vm_map_lock_read(smap);
2168 vm_map_lookup_entry(smap, offset, &tentry);
2169 tsize = tentry->end - offset;
2172 object = tentry->object.vm_object;
2173 offset = tentry->offset + (offset - tentry->start);
2174 vm_map_unlock_read(smap);
2176 object = current->object.vm_object;
2179 * Note that there is absolutely no sense in writing out
2180 * anonymous objects, so we track down the vnode object
2182 * We invalidate (remove) all pages from the address space
2183 * anyway, for semantic correctness.
2185 * note: certain anonymous maps, such as MAP_NOSYNC maps,
2186 * may start out with a NULL object.
2188 while (object && object->backing_object) {
2189 offset += object->backing_object_offset;
2190 object = object->backing_object;
2191 if (object->size < OFF_TO_IDX( offset + size))
2192 size = IDX_TO_OFF(object->size) - offset;
2194 if (object && (object->type == OBJT_VNODE) &&
2195 (current->protection & VM_PROT_WRITE)) {
2197 * Flush pages if writing is allowed, invalidate them
2198 * if invalidation requested. Pages undergoing I/O
2199 * will be ignored by vm_object_page_remove().
2201 * We cannot lock the vnode and then wait for paging
2202 * to complete without deadlocking against vm_fault.
2203 * Instead we simply call vm_object_page_remove() and
2204 * allow it to block internally on a page-by-page
2205 * basis when it encounters pages undergoing async
2210 vm_object_reference(object);
2211 vn_lock(object->handle, LK_EXCLUSIVE | LK_RETRY);
2212 flags = (syncio || invalidate) ? OBJPC_SYNC : 0;
2213 flags |= invalidate ? OBJPC_INVAL : 0;
2214 vm_object_page_clean(object,
2216 OFF_TO_IDX(offset + size + PAGE_MASK),
2218 VOP_UNLOCK(((struct vnode *)object->handle), 0);
2219 vm_object_deallocate(object);
2221 if (object && invalidate &&
2222 ((object->type == OBJT_VNODE) ||
2223 (object->type == OBJT_DEVICE))) {
2225 (object->type == OBJT_DEVICE) ? FALSE : TRUE;
2226 vm_object_reference(object);
2227 vm_object_page_remove(object,
2229 OFF_TO_IDX(offset + size + PAGE_MASK),
2231 vm_object_deallocate(object);
2236 vm_map_unlock_read(map);
2237 return (KERN_SUCCESS);
2241 * vm_map_entry_unwire: [ internal use only ]
2243 * Make the region specified by this entry pageable.
2245 * The map in question should be locked.
2246 * [This is the reason for this routine's existence.]
2249 vm_map_entry_unwire(vm_map_t map, vm_map_entry_t entry)
2251 entry->eflags &= ~MAP_ENTRY_USER_WIRED;
2252 entry->wired_count = 0;
2253 vm_fault_unwire(map, entry);
2257 * vm_map_entry_delete: [ internal use only ]
2259 * Deallocate the given entry from the target map.
2262 vm_map_entry_delete(vm_map_t map, vm_map_entry_t entry, int *countp)
2264 vm_map_entry_unlink(map, entry);
2265 map->size -= entry->end - entry->start;
2267 if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
2268 vm_object_deallocate(entry->object.vm_object);
2271 vm_map_entry_dispose(map, entry, countp);
2275 * vm_map_delete: [ internal use only ]
2277 * Deallocates the given address range from the target
2281 vm_map_delete(vm_map_t map, vm_offset_t start, vm_offset_t end, int *countp)
2284 vm_map_entry_t entry;
2285 vm_map_entry_t first_entry;
2289 * Find the start of the region, and clip it. Set entry to point
2290 * at the first record containing the requested address or, if no
2291 * such record exists, the next record with a greater address. The
2292 * loop will run from this point until a record beyond the termination
2293 * address is encountered.
2295 * map->hint must be adjusted to not point to anything we delete,
2296 * so set it to the entry prior to the one being deleted.
2298 * GGG see other GGG comment.
2300 if (vm_map_lookup_entry(map, start, &first_entry)) {
2301 entry = first_entry;
2302 vm_map_clip_start(map, entry, start, countp);
2303 map->hint = entry->prev; /* possible problem XXX */
2305 map->hint = first_entry; /* possible problem XXX */
2306 entry = first_entry->next;
2310 * If a hole opens up prior to the current first_free then
2311 * adjust first_free. As with map->hint, map->first_free
2312 * cannot be left set to anything we might delete.
2314 if (entry == &map->header) {
2315 map->first_free = &map->header;
2316 } else if (map->first_free->start >= start) {
2317 map->first_free = entry->prev;
2321 * Step through all entries in this region
2324 while ((entry != &map->header) && (entry->start < end)) {
2325 vm_map_entry_t next;
2327 vm_pindex_t offidxstart, offidxend, count;
2330 * If we hit an in-transition entry we have to sleep and
2331 * retry. It's easier (and not really slower) to just retry
2332 * since this case occurs so rarely and the hint is already
2333 * pointing at the right place. We have to reset the
2334 * start offset so as not to accidently delete an entry
2335 * another process just created in vacated space.
2337 if (entry->eflags & MAP_ENTRY_IN_TRANSITION) {
2338 entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP;
2339 start = entry->start;
2340 ++mycpu->gd_cnt.v_intrans_coll;
2341 ++mycpu->gd_cnt.v_intrans_wait;
2342 vm_map_transition_wait(map);
2345 vm_map_clip_end(map, entry, end, countp);
2351 offidxstart = OFF_TO_IDX(entry->offset);
2352 count = OFF_TO_IDX(e - s);
2353 object = entry->object.vm_object;
2356 * Unwire before removing addresses from the pmap; otherwise,
2357 * unwiring will put the entries back in the pmap.
2359 if (entry->wired_count != 0)
2360 vm_map_entry_unwire(map, entry);
2362 offidxend = offidxstart + count;
2364 if ((object == kernel_object) || (object == kmem_object)) {
2365 vm_object_page_remove(object, offidxstart, offidxend, FALSE);
2367 pmap_remove(map->pmap, s, e);
2368 if (object != NULL &&
2369 object->ref_count != 1 &&
2370 (object->flags & (OBJ_NOSPLIT|OBJ_ONEMAPPING)) == OBJ_ONEMAPPING &&
2371 (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP)) {
2372 vm_object_collapse(object);
2373 vm_object_page_remove(object, offidxstart, offidxend, FALSE);
2374 if (object->type == OBJT_SWAP) {
2375 swap_pager_freespace(object, offidxstart, count);
2377 if (offidxend >= object->size &&
2378 offidxstart < object->size) {
2379 object->size = offidxstart;
2385 * Delete the entry (which may delete the object) only after
2386 * removing all pmap entries pointing to its pages.
2387 * (Otherwise, its page frames may be reallocated, and any
2388 * modify bits will be set in the wrong object!)
2390 vm_map_entry_delete(map, entry, countp);
2393 return (KERN_SUCCESS);
2399 * Remove the given address range from the target map.
2400 * This is the exported form of vm_map_delete.
2403 vm_map_remove(vm_map_t map, vm_offset_t start, vm_offset_t end)
2408 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
2410 VM_MAP_RANGE_CHECK(map, start, end);
2411 result = vm_map_delete(map, start, end, &count);
2413 vm_map_entry_release(count);
2419 * vm_map_check_protection:
2421 * Assert that the target map allows the specified
2422 * privilege on the entire address region given.
2423 * The entire region must be allocated.
2426 vm_map_check_protection(vm_map_t map, vm_offset_t start, vm_offset_t end,
2427 vm_prot_t protection)
2429 vm_map_entry_t entry;
2430 vm_map_entry_t tmp_entry;
2432 if (!vm_map_lookup_entry(map, start, &tmp_entry)) {
2437 while (start < end) {
2438 if (entry == &map->header) {
2445 if (start < entry->start) {
2449 * Check protection associated with entry.
2452 if ((entry->protection & protection) != protection) {
2455 /* go to next entry */
2458 entry = entry->next;
2464 * Split the pages in a map entry into a new object. This affords
2465 * easier removal of unused pages, and keeps object inheritance from
2466 * being a negative impact on memory usage.
2469 vm_map_split(vm_map_entry_t entry)
2472 vm_object_t orig_object, new_object, source;
2474 vm_pindex_t offidxstart, offidxend, idx;
2476 vm_ooffset_t offset;
2478 orig_object = entry->object.vm_object;
2479 if (orig_object->type != OBJT_DEFAULT && orig_object->type != OBJT_SWAP)
2481 if (orig_object->ref_count <= 1)
2484 offset = entry->offset;
2488 offidxstart = OFF_TO_IDX(offset);
2489 offidxend = offidxstart + OFF_TO_IDX(e - s);
2490 size = offidxend - offidxstart;
2492 new_object = vm_pager_allocate(orig_object->type, NULL,
2493 IDX_TO_OFF(size), VM_PROT_ALL, 0);
2494 if (new_object == NULL)
2497 source = orig_object->backing_object;
2498 if (source != NULL) {
2499 vm_object_reference(source); /* Referenced by new_object */
2500 LIST_INSERT_HEAD(&source->shadow_head,
2501 new_object, shadow_list);
2502 vm_object_clear_flag(source, OBJ_ONEMAPPING);
2503 new_object->backing_object_offset =
2504 orig_object->backing_object_offset + IDX_TO_OFF(offidxstart);
2505 new_object->backing_object = source;
2506 source->shadow_count++;
2507 source->generation++;
2510 for (idx = 0; idx < size; idx++) {
2514 * A critical section is required to avoid a race between
2515 * the lookup and an interrupt/unbusy/free and our busy
2520 m = vm_page_lookup(orig_object, offidxstart + idx);
2527 * We must wait for pending I/O to complete before we can
2530 * We do not have to VM_PROT_NONE the page as mappings should
2531 * not be changed by this operation.
2533 if (vm_page_sleep_busy(m, TRUE, "spltwt"))
2536 vm_page_rename(m, new_object, idx);
2537 /* page automatically made dirty by rename and cache handled */
2542 if (orig_object->type == OBJT_SWAP) {
2543 vm_object_pip_add(orig_object, 1);
2545 * copy orig_object pages into new_object
2546 * and destroy unneeded pages in
2549 swap_pager_copy(orig_object, new_object, offidxstart, 0);
2550 vm_object_pip_wakeup(orig_object);
2554 * Wakeup the pages we played with. No spl protection is needed
2555 * for a simple wakeup.
2557 for (idx = 0; idx < size; idx++) {
2558 m = vm_page_lookup(new_object, idx);
2563 entry->object.vm_object = new_object;
2564 entry->offset = 0LL;
2565 vm_object_deallocate(orig_object);
2569 * vm_map_copy_entry:
2571 * Copies the contents of the source entry to the destination
2572 * entry. The entries *must* be aligned properly.
2575 vm_map_copy_entry(vm_map_t src_map, vm_map_t dst_map,
2576 vm_map_entry_t src_entry, vm_map_entry_t dst_entry)
2578 vm_object_t src_object;
2580 if ((dst_entry->eflags|src_entry->eflags) & MAP_ENTRY_IS_SUB_MAP)
2583 if (src_entry->wired_count == 0) {
2586 * If the source entry is marked needs_copy, it is already
2589 if ((src_entry->eflags & MAP_ENTRY_NEEDS_COPY) == 0) {
2590 pmap_protect(src_map->pmap,
2593 src_entry->protection & ~VM_PROT_WRITE);
2597 * Make a copy of the object.
2599 if ((src_object = src_entry->object.vm_object) != NULL) {
2601 if ((src_object->handle == NULL) &&
2602 (src_object->type == OBJT_DEFAULT ||
2603 src_object->type == OBJT_SWAP)) {
2604 vm_object_collapse(src_object);
2605 if ((src_object->flags & (OBJ_NOSPLIT|OBJ_ONEMAPPING)) == OBJ_ONEMAPPING) {
2606 vm_map_split(src_entry);
2607 src_object = src_entry->object.vm_object;
2611 vm_object_reference(src_object);
2612 vm_object_clear_flag(src_object, OBJ_ONEMAPPING);
2613 dst_entry->object.vm_object = src_object;
2614 src_entry->eflags |= (MAP_ENTRY_COW|MAP_ENTRY_NEEDS_COPY);
2615 dst_entry->eflags |= (MAP_ENTRY_COW|MAP_ENTRY_NEEDS_COPY);
2616 dst_entry->offset = src_entry->offset;
2618 dst_entry->object.vm_object = NULL;
2619 dst_entry->offset = 0;
2622 pmap_copy(dst_map->pmap, src_map->pmap, dst_entry->start,
2623 dst_entry->end - dst_entry->start, src_entry->start);
2626 * Of course, wired down pages can't be set copy-on-write.
2627 * Cause wired pages to be copied into the new map by
2628 * simulating faults (the new pages are pageable)
2630 vm_fault_copy_entry(dst_map, src_map, dst_entry, src_entry);
2636 * Create a new process vmspace structure and vm_map
2637 * based on those of an existing process. The new map
2638 * is based on the old map, according to the inheritance
2639 * values on the regions in that map.
2641 * The source map must not be locked.
2644 vmspace_fork(struct vmspace *vm1)
2646 struct vmspace *vm2;
2647 vm_map_t old_map = &vm1->vm_map;
2649 vm_map_entry_t old_entry;
2650 vm_map_entry_t new_entry;
2654 vm_map_lock(old_map);
2655 old_map->infork = 1;
2658 * XXX Note: upcalls are not copied.
2660 vm2 = vmspace_alloc(old_map->min_offset, old_map->max_offset);
2661 bcopy(&vm1->vm_startcopy, &vm2->vm_startcopy,
2662 (caddr_t)&vm1->vm_endcopy - (caddr_t)&vm1->vm_startcopy);
2663 new_map = &vm2->vm_map; /* XXX */
2664 new_map->timestamp = 1;
2667 old_entry = old_map->header.next;
2668 while (old_entry != &old_map->header) {
2670 old_entry = old_entry->next;
2673 count = vm_map_entry_reserve(count + MAP_RESERVE_COUNT);
2675 old_entry = old_map->header.next;
2676 while (old_entry != &old_map->header) {
2677 if (old_entry->eflags & MAP_ENTRY_IS_SUB_MAP)
2678 panic("vm_map_fork: encountered a submap");
2680 switch (old_entry->inheritance) {
2681 case VM_INHERIT_NONE:
2684 case VM_INHERIT_SHARE:
2686 * Clone the entry, creating the shared object if necessary.
2688 object = old_entry->object.vm_object;
2689 if (object == NULL) {
2690 object = vm_object_allocate(OBJT_DEFAULT,
2691 atop(old_entry->end - old_entry->start));
2692 old_entry->object.vm_object = object;
2693 old_entry->offset = (vm_offset_t) 0;
2697 * Add the reference before calling vm_object_shadow
2698 * to insure that a shadow object is created.
2700 vm_object_reference(object);
2701 if (old_entry->eflags & MAP_ENTRY_NEEDS_COPY) {
2702 vm_object_shadow(&old_entry->object.vm_object,
2704 atop(old_entry->end - old_entry->start));
2705 old_entry->eflags &= ~MAP_ENTRY_NEEDS_COPY;
2706 /* Transfer the second reference too. */
2707 vm_object_reference(
2708 old_entry->object.vm_object);
2709 vm_object_deallocate(object);
2710 object = old_entry->object.vm_object;
2712 vm_object_clear_flag(object, OBJ_ONEMAPPING);
2715 * Clone the entry, referencing the shared object.
2717 new_entry = vm_map_entry_create(new_map, &count);
2718 *new_entry = *old_entry;
2719 new_entry->eflags &= ~MAP_ENTRY_USER_WIRED;
2720 new_entry->wired_count = 0;
2723 * Insert the entry into the new map -- we know we're
2724 * inserting at the end of the new map.
2727 vm_map_entry_link(new_map, new_map->header.prev,
2731 * Update the physical map
2734 pmap_copy(new_map->pmap, old_map->pmap,
2736 (old_entry->end - old_entry->start),
2740 case VM_INHERIT_COPY:
2742 * Clone the entry and link into the map.
2744 new_entry = vm_map_entry_create(new_map, &count);
2745 *new_entry = *old_entry;
2746 new_entry->eflags &= ~MAP_ENTRY_USER_WIRED;
2747 new_entry->wired_count = 0;
2748 new_entry->object.vm_object = NULL;
2749 vm_map_entry_link(new_map, new_map->header.prev,
2751 vm_map_copy_entry(old_map, new_map, old_entry,
2755 old_entry = old_entry->next;
2758 new_map->size = old_map->size;
2759 old_map->infork = 0;
2760 vm_map_unlock(old_map);
2761 vm_map_entry_release(count);
2767 vm_map_stack (vm_map_t map, vm_offset_t addrbos, vm_size_t max_ssize,
2768 vm_prot_t prot, vm_prot_t max, int cow)
2770 vm_map_entry_t prev_entry;
2771 vm_map_entry_t new_stack_entry;
2772 vm_size_t init_ssize;
2776 if (VM_MIN_ADDRESS > 0 && addrbos < VM_MIN_ADDRESS)
2777 return (KERN_NO_SPACE);
2779 if (max_ssize < sgrowsiz)
2780 init_ssize = max_ssize;
2782 init_ssize = sgrowsiz;
2784 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
2787 /* If addr is already mapped, no go */
2788 if (vm_map_lookup_entry(map, addrbos, &prev_entry)) {
2790 vm_map_entry_release(count);
2791 return (KERN_NO_SPACE);
2794 /* If we would blow our VMEM resource limit, no go */
2795 if (map->size + init_ssize >
2796 curproc->p_rlimit[RLIMIT_VMEM].rlim_cur) {
2798 vm_map_entry_release(count);
2799 return (KERN_NO_SPACE);
2802 /* If we can't accomodate max_ssize in the current mapping,
2803 * no go. However, we need to be aware that subsequent user
2804 * mappings might map into the space we have reserved for
2805 * stack, and currently this space is not protected.
2807 * Hopefully we will at least detect this condition
2808 * when we try to grow the stack.
2810 if ((prev_entry->next != &map->header) &&
2811 (prev_entry->next->start < addrbos + max_ssize)) {
2813 vm_map_entry_release(count);
2814 return (KERN_NO_SPACE);
2817 /* We initially map a stack of only init_ssize. We will
2818 * grow as needed later. Since this is to be a grow
2819 * down stack, we map at the top of the range.
2821 * Note: we would normally expect prot and max to be
2822 * VM_PROT_ALL, and cow to be 0. Possibly we should
2823 * eliminate these as input parameters, and just
2824 * pass these values here in the insert call.
2826 rv = vm_map_insert(map, &count,
2827 NULL, 0, addrbos + max_ssize - init_ssize,
2828 addrbos + max_ssize, prot, max, cow);
2830 /* Now set the avail_ssize amount */
2831 if (rv == KERN_SUCCESS) {
2832 if (prev_entry != &map->header)
2833 vm_map_clip_end(map, prev_entry, addrbos + max_ssize - init_ssize, &count);
2834 new_stack_entry = prev_entry->next;
2835 if (new_stack_entry->end != addrbos + max_ssize ||
2836 new_stack_entry->start != addrbos + max_ssize - init_ssize)
2837 panic ("Bad entry start/end for new stack entry");
2839 new_stack_entry->avail_ssize = max_ssize - init_ssize;
2843 vm_map_entry_release(count);
2847 /* Attempts to grow a vm stack entry. Returns KERN_SUCCESS if the
2848 * desired address is already mapped, or if we successfully grow
2849 * the stack. Also returns KERN_SUCCESS if addr is outside the
2850 * stack range (this is strange, but preserves compatibility with
2851 * the grow function in vm_machdep.c).
2854 vm_map_growstack (struct proc *p, vm_offset_t addr)
2856 vm_map_entry_t prev_entry;
2857 vm_map_entry_t stack_entry;
2858 vm_map_entry_t new_stack_entry;
2859 struct vmspace *vm = p->p_vmspace;
2860 vm_map_t map = &vm->vm_map;
2863 int rv = KERN_SUCCESS;
2865 int use_read_lock = 1;
2868 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
2871 vm_map_lock_read(map);
2875 /* If addr is already in the entry range, no need to grow.*/
2876 if (vm_map_lookup_entry(map, addr, &prev_entry))
2879 if ((stack_entry = prev_entry->next) == &map->header)
2881 if (prev_entry == &map->header)
2882 end = stack_entry->start - stack_entry->avail_ssize;
2884 end = prev_entry->end;
2886 /* This next test mimics the old grow function in vm_machdep.c.
2887 * It really doesn't quite make sense, but we do it anyway
2888 * for compatibility.
2890 * If not growable stack, return success. This signals the
2891 * caller to proceed as he would normally with normal vm.
2893 if (stack_entry->avail_ssize < 1 ||
2894 addr >= stack_entry->start ||
2895 addr < stack_entry->start - stack_entry->avail_ssize) {
2899 /* Find the minimum grow amount */
2900 grow_amount = roundup (stack_entry->start - addr, PAGE_SIZE);
2901 if (grow_amount > stack_entry->avail_ssize) {
2906 /* If there is no longer enough space between the entries
2907 * nogo, and adjust the available space. Note: this
2908 * should only happen if the user has mapped into the
2909 * stack area after the stack was created, and is
2910 * probably an error.
2912 * This also effectively destroys any guard page the user
2913 * might have intended by limiting the stack size.
2915 if (grow_amount > stack_entry->start - end) {
2916 if (use_read_lock && vm_map_lock_upgrade(map)) {
2921 stack_entry->avail_ssize = stack_entry->start - end;
2926 is_procstack = addr >= (vm_offset_t)vm->vm_maxsaddr;
2928 /* If this is the main process stack, see if we're over the
2931 if (is_procstack && (ctob(vm->vm_ssize) + grow_amount >
2932 p->p_rlimit[RLIMIT_STACK].rlim_cur)) {
2937 /* Round up the grow amount modulo SGROWSIZ */
2938 grow_amount = roundup (grow_amount, sgrowsiz);
2939 if (grow_amount > stack_entry->avail_ssize) {
2940 grow_amount = stack_entry->avail_ssize;
2942 if (is_procstack && (ctob(vm->vm_ssize) + grow_amount >
2943 p->p_rlimit[RLIMIT_STACK].rlim_cur)) {
2944 grow_amount = p->p_rlimit[RLIMIT_STACK].rlim_cur -
2948 /* If we would blow our VMEM resource limit, no go */
2949 if (map->size + grow_amount > p->p_rlimit[RLIMIT_VMEM].rlim_cur) {
2954 if (use_read_lock && vm_map_lock_upgrade(map)) {
2960 /* Get the preliminary new entry start value */
2961 addr = stack_entry->start - grow_amount;
2963 /* If this puts us into the previous entry, cut back our growth
2964 * to the available space. Also, see the note above.
2967 stack_entry->avail_ssize = stack_entry->start - end;
2971 rv = vm_map_insert(map, &count,
2972 NULL, 0, addr, stack_entry->start,
2977 /* Adjust the available stack space by the amount we grew. */
2978 if (rv == KERN_SUCCESS) {
2979 if (prev_entry != &map->header)
2980 vm_map_clip_end(map, prev_entry, addr, &count);
2981 new_stack_entry = prev_entry->next;
2982 if (new_stack_entry->end != stack_entry->start ||
2983 new_stack_entry->start != addr)
2984 panic ("Bad stack grow start/end in new stack entry");
2986 new_stack_entry->avail_ssize = stack_entry->avail_ssize -
2987 (new_stack_entry->end -
2988 new_stack_entry->start);
2990 vm->vm_ssize += btoc(new_stack_entry->end -
2991 new_stack_entry->start);
2997 vm_map_unlock_read(map);
3000 vm_map_entry_release(count);
3005 * Unshare the specified VM space for exec. If other processes are
3006 * mapped to it, then create a new one. The new vmspace is null.
3010 vmspace_exec(struct proc *p, struct vmspace *vmcopy)
3012 struct vmspace *oldvmspace = p->p_vmspace;
3013 struct vmspace *newvmspace;
3014 vm_map_t map = &p->p_vmspace->vm_map;
3017 * If we are execing a resident vmspace we fork it, otherwise
3018 * we create a new vmspace. Note that exitingcnt and upcalls
3019 * are not copied to the new vmspace.
3022 newvmspace = vmspace_fork(vmcopy);
3024 newvmspace = vmspace_alloc(map->min_offset, map->max_offset);
3025 bcopy(&oldvmspace->vm_startcopy, &newvmspace->vm_startcopy,
3026 (caddr_t)&oldvmspace->vm_endcopy -
3027 (caddr_t)&oldvmspace->vm_startcopy);
3031 * This code is written like this for prototype purposes. The
3032 * goal is to avoid running down the vmspace here, but let the
3033 * other process's that are still using the vmspace to finally
3034 * run it down. Even though there is little or no chance of blocking
3035 * here, it is a good idea to keep this form for future mods.
3037 p->p_vmspace = newvmspace;
3038 pmap_pinit2(vmspace_pmap(newvmspace));
3041 vmspace_free(oldvmspace);
3045 * Unshare the specified VM space for forcing COW. This
3046 * is called by rfork, for the (RFMEM|RFPROC) == 0 case.
3048 * The exitingcnt test is not strictly necessary but has been
3049 * included for code sanity (to make the code a bit more deterministic).
3053 vmspace_unshare(struct proc *p)
3055 struct vmspace *oldvmspace = p->p_vmspace;
3056 struct vmspace *newvmspace;
3058 if (oldvmspace->vm_refcnt == 1 && oldvmspace->vm_exitingcnt == 0)
3060 newvmspace = vmspace_fork(oldvmspace);
3061 p->p_vmspace = newvmspace;
3062 pmap_pinit2(vmspace_pmap(newvmspace));
3065 vmspace_free(oldvmspace);
3071 * Finds the VM object, offset, and
3072 * protection for a given virtual address in the
3073 * specified map, assuming a page fault of the
3076 * Leaves the map in question locked for read; return
3077 * values are guaranteed until a vm_map_lookup_done
3078 * call is performed. Note that the map argument
3079 * is in/out; the returned map must be used in
3080 * the call to vm_map_lookup_done.
3082 * A handle (out_entry) is returned for use in
3083 * vm_map_lookup_done, to make that fast.
3085 * If a lookup is requested with "write protection"
3086 * specified, the map may be changed to perform virtual
3087 * copying operations, although the data referenced will
3091 vm_map_lookup(vm_map_t *var_map, /* IN/OUT */
3093 vm_prot_t fault_typea,
3094 vm_map_entry_t *out_entry, /* OUT */
3095 vm_object_t *object, /* OUT */
3096 vm_pindex_t *pindex, /* OUT */
3097 vm_prot_t *out_prot, /* OUT */
3098 boolean_t *wired) /* OUT */
3100 vm_map_entry_t entry;
3101 vm_map_t map = *var_map;
3103 vm_prot_t fault_type = fault_typea;
3104 int use_read_lock = 1;
3105 int rv = KERN_SUCCESS;
3109 vm_map_lock_read(map);
3114 * If the map has an interesting hint, try it before calling full
3115 * blown lookup routine.
3120 if ((entry == &map->header) ||
3121 (vaddr < entry->start) || (vaddr >= entry->end)) {
3122 vm_map_entry_t tmp_entry;
3125 * Entry was either not a valid hint, or the vaddr was not
3126 * contained in the entry, so do a full lookup.
3128 if (!vm_map_lookup_entry(map, vaddr, &tmp_entry)) {
3129 rv = KERN_INVALID_ADDRESS;
3141 if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) {
3142 vm_map_t old_map = map;
3144 *var_map = map = entry->object.sub_map;
3146 vm_map_unlock_read(old_map);
3148 vm_map_unlock(old_map);
3154 * Check whether this task is allowed to have this page.
3155 * Note the special case for MAP_ENTRY_COW
3156 * pages with an override. This is to implement a forced
3157 * COW for debuggers.
3160 if (fault_type & VM_PROT_OVERRIDE_WRITE)
3161 prot = entry->max_protection;
3163 prot = entry->protection;
3165 fault_type &= (VM_PROT_READ|VM_PROT_WRITE|VM_PROT_EXECUTE);
3166 if ((fault_type & prot) != fault_type) {
3167 rv = KERN_PROTECTION_FAILURE;
3171 if ((entry->eflags & MAP_ENTRY_USER_WIRED) &&
3172 (entry->eflags & MAP_ENTRY_COW) &&
3173 (fault_type & VM_PROT_WRITE) &&
3174 (fault_typea & VM_PROT_OVERRIDE_WRITE) == 0) {
3175 rv = KERN_PROTECTION_FAILURE;
3180 * If this page is not pageable, we have to get it for all possible
3184 *wired = (entry->wired_count != 0);
3186 prot = fault_type = entry->protection;
3189 * If the entry was copy-on-write, we either ...
3192 if (entry->eflags & MAP_ENTRY_NEEDS_COPY) {
3194 * If we want to write the page, we may as well handle that
3195 * now since we've got the map locked.
3197 * If we don't need to write the page, we just demote the
3198 * permissions allowed.
3201 if (fault_type & VM_PROT_WRITE) {
3203 * Make a new object, and place it in the object
3204 * chain. Note that no new references have appeared
3205 * -- one just moved from the map to the new
3209 if (use_read_lock && vm_map_lock_upgrade(map)) {
3216 &entry->object.vm_object,
3218 atop(entry->end - entry->start));
3220 entry->eflags &= ~MAP_ENTRY_NEEDS_COPY;
3223 * We're attempting to read a copy-on-write page --
3224 * don't allow writes.
3227 prot &= ~VM_PROT_WRITE;
3232 * Create an object if necessary.
3234 if (entry->object.vm_object == NULL &&
3236 if (use_read_lock && vm_map_lock_upgrade(map)) {
3241 entry->object.vm_object = vm_object_allocate(OBJT_DEFAULT,
3242 atop(entry->end - entry->start));
3247 * Return the object/offset from this entry. If the entry was
3248 * copy-on-write or empty, it has been fixed up.
3251 *pindex = OFF_TO_IDX((vaddr - entry->start) + entry->offset);
3252 *object = entry->object.vm_object;
3255 * Return whether this is the only map sharing this data. On
3256 * success we return with a read lock held on the map. On failure
3257 * we return with the map unlocked.
3261 if (rv == KERN_SUCCESS) {
3262 if (use_read_lock == 0)
3263 vm_map_lock_downgrade(map);
3264 } else if (use_read_lock) {
3265 vm_map_unlock_read(map);
3273 * vm_map_lookup_done:
3275 * Releases locks acquired by a vm_map_lookup
3276 * (according to the handle returned by that lookup).
3280 vm_map_lookup_done(vm_map_t map, vm_map_entry_t entry, int count)
3283 * Unlock the main-level map
3285 vm_map_unlock_read(map);
3287 vm_map_entry_release(count);
3290 #include "opt_ddb.h"
3292 #include <sys/kernel.h>
3294 #include <ddb/ddb.h>
3297 * vm_map_print: [ debug ]
3299 DB_SHOW_COMMAND(map, vm_map_print)
3302 /* XXX convert args. */
3303 vm_map_t map = (vm_map_t)addr;
3304 boolean_t full = have_addr;
3306 vm_map_entry_t entry;
3308 db_iprintf("Task map %p: pmap=%p, nentries=%d, version=%u\n",
3310 (void *)map->pmap, map->nentries, map->timestamp);
3313 if (!full && db_indent)
3317 for (entry = map->header.next; entry != &map->header;
3318 entry = entry->next) {
3319 db_iprintf("map entry %p: start=%p, end=%p\n",
3320 (void *)entry, (void *)entry->start, (void *)entry->end);
3323 static char *inheritance_name[4] =
3324 {"share", "copy", "none", "donate_copy"};
3326 db_iprintf(" prot=%x/%x/%s",
3328 entry->max_protection,
3329 inheritance_name[(int)(unsigned char)entry->inheritance]);
3330 if (entry->wired_count != 0)
3331 db_printf(", wired");
3333 if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) {
3334 /* XXX no %qd in kernel. Truncate entry->offset. */
3335 db_printf(", share=%p, offset=0x%lx\n",
3336 (void *)entry->object.sub_map,
3337 (long)entry->offset);
3339 if ((entry->prev == &map->header) ||
3340 (entry->prev->object.sub_map !=
3341 entry->object.sub_map)) {
3343 vm_map_print((db_expr_t)(intptr_t)
3344 entry->object.sub_map,
3345 full, 0, (char *)0);
3349 /* XXX no %qd in kernel. Truncate entry->offset. */
3350 db_printf(", object=%p, offset=0x%lx",
3351 (void *)entry->object.vm_object,
3352 (long)entry->offset);
3353 if (entry->eflags & MAP_ENTRY_COW)
3354 db_printf(", copy (%s)",
3355 (entry->eflags & MAP_ENTRY_NEEDS_COPY) ? "needed" : "done");
3359 if ((entry->prev == &map->header) ||
3360 (entry->prev->object.vm_object !=
3361 entry->object.vm_object)) {
3363 vm_object_print((db_expr_t)(intptr_t)
3364 entry->object.vm_object,
3365 full, 0, (char *)0);
3377 DB_SHOW_COMMAND(procvm, procvm)
3382 p = (struct proc *) addr;
3387 db_printf("p = %p, vmspace = %p, map = %p, pmap = %p\n",
3388 (void *)p, (void *)p->p_vmspace, (void *)&p->p_vmspace->vm_map,
3389 (void *)vmspace_pmap(p->p_vmspace));
3391 vm_map_print((db_expr_t)(intptr_t)&p->p_vmspace->vm_map, 1, 0, NULL);