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 * 4. Neither the name of the University nor the names of its contributors
17 * may be used to endorse or promote products derived from this software
18 * without specific prior written permission.
20 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
21 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
22 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
23 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
24 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
25 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
26 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
27 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
28 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
29 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32 * from: @(#)vm_map.c 8.3 (Berkeley) 1/12/94
35 * Copyright (c) 1987, 1990 Carnegie-Mellon University.
36 * All rights reserved.
38 * Authors: Avadis Tevanian, Jr., Michael Wayne Young
40 * Permission to use, copy, modify and distribute this software and
41 * its documentation is hereby granted, provided that both the copyright
42 * notice and this permission notice appear in all copies of the
43 * software, derivative works or modified versions, and any portions
44 * thereof, and that both notices appear in supporting documentation.
46 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
47 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
48 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
50 * Carnegie Mellon requests users of this software to return to
52 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU
53 * School of Computer Science
54 * Carnegie Mellon University
55 * Pittsburgh PA 15213-3890
57 * any improvements or extensions that they make and grant Carnegie the
58 * rights to redistribute these changes.
62 * Virtual memory mapping module.
65 #include <sys/cdefs.h>
66 __FBSDID("$FreeBSD$");
68 #include <sys/param.h>
69 #include <sys/systm.h>
70 #include <sys/kernel.h>
73 #include <sys/mutex.h>
75 #include <sys/vmmeter.h>
77 #include <sys/vnode.h>
78 #include <sys/racct.h>
79 #include <sys/resourcevar.h>
80 #include <sys/rwlock.h>
82 #include <sys/sysctl.h>
83 #include <sys/sysent.h>
87 #include <vm/vm_param.h>
89 #include <vm/vm_map.h>
90 #include <vm/vm_page.h>
91 #include <vm/vm_object.h>
92 #include <vm/vm_pager.h>
93 #include <vm/vm_kern.h>
94 #include <vm/vm_extern.h>
95 #include <vm/vnode_pager.h>
96 #include <vm/swap_pager.h>
100 * Virtual memory maps provide for the mapping, protection,
101 * and sharing of virtual memory objects. In addition,
102 * this module provides for an efficient virtual copy of
103 * memory from one map to another.
105 * Synchronization is required prior to most operations.
107 * Maps consist of an ordered doubly-linked list of simple
108 * entries; a self-adjusting binary search tree of these
109 * entries is used to speed up lookups.
111 * Since portions of maps are specified by start/end addresses,
112 * which may not align with existing map entries, all
113 * routines merely "clip" entries to these start/end values.
114 * [That is, an entry is split into two, bordering at a
115 * start or end value.] Note that these clippings may not
116 * always be necessary (as the two resulting entries are then
117 * not changed); however, the clipping is done for convenience.
119 * As mentioned above, virtual copy operations are performed
120 * by copying VM object references from one map to
121 * another, and then marking both regions as copy-on-write.
124 static struct mtx map_sleep_mtx;
125 static uma_zone_t mapentzone;
126 static uma_zone_t kmapentzone;
127 static uma_zone_t mapzone;
128 static uma_zone_t vmspace_zone;
129 static int vmspace_zinit(void *mem, int size, int flags);
130 static int vm_map_zinit(void *mem, int ize, int flags);
131 static void _vm_map_init(vm_map_t map, pmap_t pmap, vm_offset_t min,
133 static void vm_map_entry_deallocate(vm_map_entry_t entry, boolean_t system_map);
134 static void vm_map_entry_dispose(vm_map_t map, vm_map_entry_t entry);
136 static void vm_map_zdtor(void *mem, int size, void *arg);
137 static void vmspace_zdtor(void *mem, int size, void *arg);
140 #define ENTRY_CHARGED(e) ((e)->cred != NULL || \
141 ((e)->object.vm_object != NULL && (e)->object.vm_object->cred != NULL && \
142 !((e)->eflags & MAP_ENTRY_NEEDS_COPY)))
145 * PROC_VMSPACE_{UN,}LOCK() can be a noop as long as vmspaces are type
148 #define PROC_VMSPACE_LOCK(p) do { } while (0)
149 #define PROC_VMSPACE_UNLOCK(p) do { } while (0)
152 * VM_MAP_RANGE_CHECK: [ internal use only ]
154 * Asserts that the starting and ending region
155 * addresses fall within the valid range of the map.
157 #define VM_MAP_RANGE_CHECK(map, start, end) \
159 if (start < vm_map_min(map)) \
160 start = vm_map_min(map); \
161 if (end > vm_map_max(map)) \
162 end = vm_map_max(map); \
170 * Initialize the vm_map module. Must be called before
171 * any other vm_map routines.
173 * Map and entry structures are allocated from the general
174 * purpose memory pool with some exceptions:
176 * - The kernel map and kmem submap are allocated statically.
177 * - Kernel map entries are allocated out of a static pool.
179 * These restrictions are necessary since malloc() uses the
180 * maps and requires map entries.
186 mtx_init(&map_sleep_mtx, "vm map sleep mutex", NULL, MTX_DEF);
187 mapzone = uma_zcreate("MAP", sizeof(struct vm_map), NULL,
193 vm_map_zinit, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
194 uma_prealloc(mapzone, MAX_KMAP);
195 kmapentzone = uma_zcreate("KMAP ENTRY", sizeof(struct vm_map_entry),
196 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR,
197 UMA_ZONE_MTXCLASS | UMA_ZONE_VM);
198 mapentzone = uma_zcreate("MAP ENTRY", sizeof(struct vm_map_entry),
199 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
200 vmspace_zone = uma_zcreate("VMSPACE", sizeof(struct vmspace), NULL,
206 vmspace_zinit, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
210 vmspace_zinit(void *mem, int size, int flags)
214 vm = (struct vmspace *)mem;
216 vm->vm_map.pmap = NULL;
217 (void)vm_map_zinit(&vm->vm_map, sizeof(vm->vm_map), flags);
218 PMAP_LOCK_INIT(vmspace_pmap(vm));
223 vm_map_zinit(void *mem, int size, int flags)
228 memset(map, 0, sizeof(*map));
229 mtx_init(&map->system_mtx, "vm map (system)", NULL, MTX_DEF | MTX_DUPOK);
230 sx_init(&map->lock, "vm map (user)");
236 vmspace_zdtor(void *mem, int size, void *arg)
240 vm = (struct vmspace *)mem;
242 vm_map_zdtor(&vm->vm_map, sizeof(vm->vm_map), arg);
245 vm_map_zdtor(void *mem, int size, void *arg)
250 KASSERT(map->nentries == 0,
251 ("map %p nentries == %d on free.",
252 map, map->nentries));
253 KASSERT(map->size == 0,
254 ("map %p size == %lu on free.",
255 map, (unsigned long)map->size));
257 #endif /* INVARIANTS */
260 * Allocate a vmspace structure, including a vm_map and pmap,
261 * and initialize those structures. The refcnt is set to 1.
263 * If 'pinit' is NULL then the embedded pmap is initialized via pmap_pinit().
266 vmspace_alloc(vm_offset_t min, vm_offset_t max, pmap_pinit_t pinit)
270 vm = uma_zalloc(vmspace_zone, M_WAITOK);
272 KASSERT(vm->vm_map.pmap == NULL, ("vm_map.pmap must be NULL"));
277 if (!pinit(vmspace_pmap(vm))) {
278 uma_zfree(vmspace_zone, vm);
281 CTR1(KTR_VM, "vmspace_alloc: %p", vm);
282 _vm_map_init(&vm->vm_map, vmspace_pmap(vm), min, max);
296 vmspace_container_reset(struct proc *p)
301 racct_set(p, RACCT_DATA, 0);
302 racct_set(p, RACCT_STACK, 0);
303 racct_set(p, RACCT_RSS, 0);
304 racct_set(p, RACCT_MEMLOCK, 0);
305 racct_set(p, RACCT_VMEM, 0);
311 vmspace_dofree(struct vmspace *vm)
314 CTR1(KTR_VM, "vmspace_free: %p", vm);
317 * Make sure any SysV shm is freed, it might not have been in
323 * Lock the map, to wait out all other references to it.
324 * Delete all of the mappings and pages they hold, then call
325 * the pmap module to reclaim anything left.
327 (void)vm_map_remove(&vm->vm_map, vm->vm_map.min_offset,
328 vm->vm_map.max_offset);
330 pmap_release(vmspace_pmap(vm));
331 vm->vm_map.pmap = NULL;
332 uma_zfree(vmspace_zone, vm);
336 vmspace_free(struct vmspace *vm)
339 if (vm->vm_refcnt == 0)
340 panic("vmspace_free: attempt to free already freed vmspace");
342 if (atomic_fetchadd_int(&vm->vm_refcnt, -1) == 1)
347 vmspace_exitfree(struct proc *p)
351 PROC_VMSPACE_LOCK(p);
354 PROC_VMSPACE_UNLOCK(p);
355 KASSERT(vm == &vmspace0, ("vmspace_exitfree: wrong vmspace"));
360 vmspace_exit(struct thread *td)
367 * Release user portion of address space.
368 * This releases references to vnodes,
369 * which could cause I/O if the file has been unlinked.
370 * Need to do this early enough that we can still sleep.
372 * The last exiting process to reach this point releases as
373 * much of the environment as it can. vmspace_dofree() is the
374 * slower fallback in case another process had a temporary
375 * reference to the vmspace.
380 atomic_add_int(&vmspace0.vm_refcnt, 1);
382 refcnt = vm->vm_refcnt;
383 if (refcnt > 1 && p->p_vmspace != &vmspace0) {
384 /* Switch now since other proc might free vmspace */
385 PROC_VMSPACE_LOCK(p);
386 p->p_vmspace = &vmspace0;
387 PROC_VMSPACE_UNLOCK(p);
390 } while (!atomic_cmpset_int(&vm->vm_refcnt, refcnt, refcnt - 1));
392 if (p->p_vmspace != vm) {
393 /* vmspace not yet freed, switch back */
394 PROC_VMSPACE_LOCK(p);
396 PROC_VMSPACE_UNLOCK(p);
399 pmap_remove_pages(vmspace_pmap(vm));
400 /* Switch now since this proc will free vmspace */
401 PROC_VMSPACE_LOCK(p);
402 p->p_vmspace = &vmspace0;
403 PROC_VMSPACE_UNLOCK(p);
407 vmspace_container_reset(p);
410 /* Acquire reference to vmspace owned by another process. */
413 vmspace_acquire_ref(struct proc *p)
418 PROC_VMSPACE_LOCK(p);
421 PROC_VMSPACE_UNLOCK(p);
425 refcnt = vm->vm_refcnt;
426 if (refcnt <= 0) { /* Avoid 0->1 transition */
427 PROC_VMSPACE_UNLOCK(p);
430 } while (!atomic_cmpset_int(&vm->vm_refcnt, refcnt, refcnt + 1));
431 if (vm != p->p_vmspace) {
432 PROC_VMSPACE_UNLOCK(p);
436 PROC_VMSPACE_UNLOCK(p);
441 _vm_map_lock(vm_map_t map, const char *file, int line)
445 mtx_lock_flags_(&map->system_mtx, 0, file, line);
447 sx_xlock_(&map->lock, file, line);
452 vm_map_process_deferred(void)
455 vm_map_entry_t entry, next;
459 entry = td->td_map_def_user;
460 td->td_map_def_user = NULL;
461 while (entry != NULL) {
463 if ((entry->eflags & MAP_ENTRY_VN_WRITECNT) != 0) {
465 * Decrement the object's writemappings and
466 * possibly the vnode's v_writecount.
468 KASSERT((entry->eflags & MAP_ENTRY_IS_SUB_MAP) == 0,
469 ("Submap with writecount"));
470 object = entry->object.vm_object;
471 KASSERT(object != NULL, ("No object for writecount"));
472 vnode_pager_release_writecount(object, entry->start,
475 vm_map_entry_deallocate(entry, FALSE);
481 _vm_map_unlock(vm_map_t map, const char *file, int line)
485 mtx_unlock_flags_(&map->system_mtx, 0, file, line);
487 sx_xunlock_(&map->lock, file, line);
488 vm_map_process_deferred();
493 _vm_map_lock_read(vm_map_t map, const char *file, int line)
497 mtx_lock_flags_(&map->system_mtx, 0, file, line);
499 sx_slock_(&map->lock, file, line);
503 _vm_map_unlock_read(vm_map_t map, const char *file, int line)
507 mtx_unlock_flags_(&map->system_mtx, 0, file, line);
509 sx_sunlock_(&map->lock, file, line);
510 vm_map_process_deferred();
515 _vm_map_trylock(vm_map_t map, const char *file, int line)
519 error = map->system_map ?
520 !mtx_trylock_flags_(&map->system_mtx, 0, file, line) :
521 !sx_try_xlock_(&map->lock, file, line);
528 _vm_map_trylock_read(vm_map_t map, const char *file, int line)
532 error = map->system_map ?
533 !mtx_trylock_flags_(&map->system_mtx, 0, file, line) :
534 !sx_try_slock_(&map->lock, file, line);
539 * _vm_map_lock_upgrade: [ internal use only ]
541 * Tries to upgrade a read (shared) lock on the specified map to a write
542 * (exclusive) lock. Returns the value "0" if the upgrade succeeds and a
543 * non-zero value if the upgrade fails. If the upgrade fails, the map is
544 * returned without a read or write lock held.
546 * Requires that the map be read locked.
549 _vm_map_lock_upgrade(vm_map_t map, const char *file, int line)
551 unsigned int last_timestamp;
553 if (map->system_map) {
554 mtx_assert_(&map->system_mtx, MA_OWNED, file, line);
556 if (!sx_try_upgrade_(&map->lock, file, line)) {
557 last_timestamp = map->timestamp;
558 sx_sunlock_(&map->lock, file, line);
559 vm_map_process_deferred();
561 * If the map's timestamp does not change while the
562 * map is unlocked, then the upgrade succeeds.
564 sx_xlock_(&map->lock, file, line);
565 if (last_timestamp != map->timestamp) {
566 sx_xunlock_(&map->lock, file, line);
576 _vm_map_lock_downgrade(vm_map_t map, const char *file, int line)
579 if (map->system_map) {
580 mtx_assert_(&map->system_mtx, MA_OWNED, file, line);
582 sx_downgrade_(&map->lock, file, line);
588 * Returns a non-zero value if the caller holds a write (exclusive) lock
589 * on the specified map and the value "0" otherwise.
592 vm_map_locked(vm_map_t map)
596 return (mtx_owned(&map->system_mtx));
598 return (sx_xlocked(&map->lock));
603 _vm_map_assert_locked(vm_map_t map, const char *file, int line)
607 mtx_assert_(&map->system_mtx, MA_OWNED, file, line);
609 sx_assert_(&map->lock, SA_XLOCKED, file, line);
612 #define VM_MAP_ASSERT_LOCKED(map) \
613 _vm_map_assert_locked(map, LOCK_FILE, LOCK_LINE)
615 #define VM_MAP_ASSERT_LOCKED(map)
619 * _vm_map_unlock_and_wait:
621 * Atomically releases the lock on the specified map and puts the calling
622 * thread to sleep. The calling thread will remain asleep until either
623 * vm_map_wakeup() is performed on the map or the specified timeout is
626 * WARNING! This function does not perform deferred deallocations of
627 * objects and map entries. Therefore, the calling thread is expected to
628 * reacquire the map lock after reawakening and later perform an ordinary
629 * unlock operation, such as vm_map_unlock(), before completing its
630 * operation on the map.
633 _vm_map_unlock_and_wait(vm_map_t map, int timo, const char *file, int line)
636 mtx_lock(&map_sleep_mtx);
638 mtx_unlock_flags_(&map->system_mtx, 0, file, line);
640 sx_xunlock_(&map->lock, file, line);
641 return (msleep(&map->root, &map_sleep_mtx, PDROP | PVM, "vmmaps",
648 * Awaken any threads that have slept on the map using
649 * vm_map_unlock_and_wait().
652 vm_map_wakeup(vm_map_t map)
656 * Acquire and release map_sleep_mtx to prevent a wakeup()
657 * from being performed (and lost) between the map unlock
658 * and the msleep() in _vm_map_unlock_and_wait().
660 mtx_lock(&map_sleep_mtx);
661 mtx_unlock(&map_sleep_mtx);
666 vm_map_busy(vm_map_t map)
669 VM_MAP_ASSERT_LOCKED(map);
674 vm_map_unbusy(vm_map_t map)
677 VM_MAP_ASSERT_LOCKED(map);
678 KASSERT(map->busy, ("vm_map_unbusy: not busy"));
679 if (--map->busy == 0 && (map->flags & MAP_BUSY_WAKEUP)) {
680 vm_map_modflags(map, 0, MAP_BUSY_WAKEUP);
686 vm_map_wait_busy(vm_map_t map)
689 VM_MAP_ASSERT_LOCKED(map);
691 vm_map_modflags(map, MAP_BUSY_WAKEUP, 0);
693 msleep(&map->busy, &map->system_mtx, 0, "mbusy", 0);
695 sx_sleep(&map->busy, &map->lock, 0, "mbusy", 0);
701 vmspace_resident_count(struct vmspace *vmspace)
703 return pmap_resident_count(vmspace_pmap(vmspace));
709 * Creates and returns a new empty VM map with
710 * the given physical map structure, and having
711 * the given lower and upper address bounds.
714 vm_map_create(pmap_t pmap, vm_offset_t min, vm_offset_t max)
718 result = uma_zalloc(mapzone, M_WAITOK);
719 CTR1(KTR_VM, "vm_map_create: %p", result);
720 _vm_map_init(result, pmap, min, max);
725 * Initialize an existing vm_map structure
726 * such as that in the vmspace structure.
729 _vm_map_init(vm_map_t map, pmap_t pmap, vm_offset_t min, vm_offset_t max)
732 map->header.next = map->header.prev = &map->header;
733 map->needs_wakeup = FALSE;
736 map->min_offset = min;
737 map->max_offset = max;
745 vm_map_init(vm_map_t map, pmap_t pmap, vm_offset_t min, vm_offset_t max)
748 _vm_map_init(map, pmap, min, max);
749 mtx_init(&map->system_mtx, "system map", NULL, MTX_DEF | MTX_DUPOK);
750 sx_init(&map->lock, "user map");
754 * vm_map_entry_dispose: [ internal use only ]
756 * Inverse of vm_map_entry_create.
759 vm_map_entry_dispose(vm_map_t map, vm_map_entry_t entry)
761 uma_zfree(map->system_map ? kmapentzone : mapentzone, entry);
765 * vm_map_entry_create: [ internal use only ]
767 * Allocates a VM map entry for insertion.
768 * No entry fields are filled in.
770 static vm_map_entry_t
771 vm_map_entry_create(vm_map_t map)
773 vm_map_entry_t new_entry;
776 new_entry = uma_zalloc(kmapentzone, M_NOWAIT);
778 new_entry = uma_zalloc(mapentzone, M_WAITOK);
779 if (new_entry == NULL)
780 panic("vm_map_entry_create: kernel resources exhausted");
785 * vm_map_entry_set_behavior:
787 * Set the expected access behavior, either normal, random, or
791 vm_map_entry_set_behavior(vm_map_entry_t entry, u_char behavior)
793 entry->eflags = (entry->eflags & ~MAP_ENTRY_BEHAV_MASK) |
794 (behavior & MAP_ENTRY_BEHAV_MASK);
798 * vm_map_entry_set_max_free:
800 * Set the max_free field in a vm_map_entry.
803 vm_map_entry_set_max_free(vm_map_entry_t entry)
806 entry->max_free = entry->adj_free;
807 if (entry->left != NULL && entry->left->max_free > entry->max_free)
808 entry->max_free = entry->left->max_free;
809 if (entry->right != NULL && entry->right->max_free > entry->max_free)
810 entry->max_free = entry->right->max_free;
814 * vm_map_entry_splay:
816 * The Sleator and Tarjan top-down splay algorithm with the
817 * following variation. Max_free must be computed bottom-up, so
818 * on the downward pass, maintain the left and right spines in
819 * reverse order. Then, make a second pass up each side to fix
820 * the pointers and compute max_free. The time bound is O(log n)
823 * The new root is the vm_map_entry containing "addr", or else an
824 * adjacent entry (lower or higher) if addr is not in the tree.
826 * The map must be locked, and leaves it so.
828 * Returns: the new root.
830 static vm_map_entry_t
831 vm_map_entry_splay(vm_offset_t addr, vm_map_entry_t root)
833 vm_map_entry_t llist, rlist;
834 vm_map_entry_t ltree, rtree;
837 /* Special case of empty tree. */
842 * Pass One: Splay down the tree until we find addr or a NULL
843 * pointer where addr would go. llist and rlist are the two
844 * sides in reverse order (bottom-up), with llist linked by
845 * the right pointer and rlist linked by the left pointer in
846 * the vm_map_entry. Wait until Pass Two to set max_free on
852 /* root is never NULL in here. */
853 if (addr < root->start) {
857 if (addr < y->start && y->left != NULL) {
858 /* Rotate right and put y on rlist. */
859 root->left = y->right;
861 vm_map_entry_set_max_free(root);
866 /* Put root on rlist. */
871 } else if (addr >= root->end) {
875 if (addr >= y->end && y->right != NULL) {
876 /* Rotate left and put y on llist. */
877 root->right = y->left;
879 vm_map_entry_set_max_free(root);
884 /* Put root on llist. */
894 * Pass Two: Walk back up the two spines, flip the pointers
895 * and set max_free. The subtrees of the root go at the
896 * bottom of llist and rlist.
899 while (llist != NULL) {
901 llist->right = ltree;
902 vm_map_entry_set_max_free(llist);
907 while (rlist != NULL) {
910 vm_map_entry_set_max_free(rlist);
916 * Final assembly: add ltree and rtree as subtrees of root.
920 vm_map_entry_set_max_free(root);
926 * vm_map_entry_{un,}link:
928 * Insert/remove entries from maps.
931 vm_map_entry_link(vm_map_t map,
932 vm_map_entry_t after_where,
933 vm_map_entry_t entry)
937 "vm_map_entry_link: map %p, nentries %d, entry %p, after %p", map,
938 map->nentries, entry, after_where);
939 VM_MAP_ASSERT_LOCKED(map);
941 entry->prev = after_where;
942 entry->next = after_where->next;
943 entry->next->prev = entry;
944 after_where->next = entry;
946 if (after_where != &map->header) {
947 if (after_where != map->root)
948 vm_map_entry_splay(after_where->start, map->root);
949 entry->right = after_where->right;
950 entry->left = after_where;
951 after_where->right = NULL;
952 after_where->adj_free = entry->start - after_where->end;
953 vm_map_entry_set_max_free(after_where);
955 entry->right = map->root;
958 entry->adj_free = (entry->next == &map->header ? map->max_offset :
959 entry->next->start) - entry->end;
960 vm_map_entry_set_max_free(entry);
965 vm_map_entry_unlink(vm_map_t map,
966 vm_map_entry_t entry)
968 vm_map_entry_t next, prev, root;
970 VM_MAP_ASSERT_LOCKED(map);
971 if (entry != map->root)
972 vm_map_entry_splay(entry->start, map->root);
973 if (entry->left == NULL)
976 root = vm_map_entry_splay(entry->start, entry->left);
977 root->right = entry->right;
978 root->adj_free = (entry->next == &map->header ? map->max_offset :
979 entry->next->start) - root->end;
980 vm_map_entry_set_max_free(root);
989 CTR3(KTR_VM, "vm_map_entry_unlink: map %p, nentries %d, entry %p", map,
990 map->nentries, entry);
994 * vm_map_entry_resize_free:
996 * Recompute the amount of free space following a vm_map_entry
997 * and propagate that value up the tree. Call this function after
998 * resizing a map entry in-place, that is, without a call to
999 * vm_map_entry_link() or _unlink().
1001 * The map must be locked, and leaves it so.
1004 vm_map_entry_resize_free(vm_map_t map, vm_map_entry_t entry)
1008 * Using splay trees without parent pointers, propagating
1009 * max_free up the tree is done by moving the entry to the
1010 * root and making the change there.
1012 if (entry != map->root)
1013 map->root = vm_map_entry_splay(entry->start, map->root);
1015 entry->adj_free = (entry->next == &map->header ? map->max_offset :
1016 entry->next->start) - entry->end;
1017 vm_map_entry_set_max_free(entry);
1021 * vm_map_lookup_entry: [ internal use only ]
1023 * Finds the map entry containing (or
1024 * immediately preceding) the specified address
1025 * in the given map; the entry is returned
1026 * in the "entry" parameter. The boolean
1027 * result indicates whether the address is
1028 * actually contained in the map.
1031 vm_map_lookup_entry(
1033 vm_offset_t address,
1034 vm_map_entry_t *entry) /* OUT */
1040 * If the map is empty, then the map entry immediately preceding
1041 * "address" is the map's header.
1045 *entry = &map->header;
1046 else if (address >= cur->start && cur->end > address) {
1049 } else if ((locked = vm_map_locked(map)) ||
1050 sx_try_upgrade(&map->lock)) {
1052 * Splay requires a write lock on the map. However, it only
1053 * restructures the binary search tree; it does not otherwise
1054 * change the map. Thus, the map's timestamp need not change
1055 * on a temporary upgrade.
1057 map->root = cur = vm_map_entry_splay(address, cur);
1059 sx_downgrade(&map->lock);
1062 * If "address" is contained within a map entry, the new root
1063 * is that map entry. Otherwise, the new root is a map entry
1064 * immediately before or after "address".
1066 if (address >= cur->start) {
1068 if (cur->end > address)
1074 * Since the map is only locked for read access, perform a
1075 * standard binary search tree lookup for "address".
1078 if (address < cur->start) {
1079 if (cur->left == NULL) {
1084 } else if (cur->end > address) {
1088 if (cur->right == NULL) {
1101 * Inserts the given whole VM object into the target
1102 * map at the specified address range. The object's
1103 * size should match that of the address range.
1105 * Requires that the map be locked, and leaves it so.
1107 * If object is non-NULL, ref count must be bumped by caller
1108 * prior to making call to account for the new entry.
1111 vm_map_insert(vm_map_t map, vm_object_t object, vm_ooffset_t offset,
1112 vm_offset_t start, vm_offset_t end, vm_prot_t prot, vm_prot_t max,
1115 vm_map_entry_t new_entry;
1116 vm_map_entry_t prev_entry;
1117 vm_map_entry_t temp_entry;
1118 vm_eflags_t protoeflags;
1120 vm_inherit_t inheritance;
1121 boolean_t charge_prev_obj;
1123 VM_MAP_ASSERT_LOCKED(map);
1126 * Check that the start and end points are not bogus.
1128 if ((start < map->min_offset) || (end > map->max_offset) ||
1130 return (KERN_INVALID_ADDRESS);
1133 * Find the entry prior to the proposed starting address; if it's part
1134 * of an existing entry, this range is bogus.
1136 if (vm_map_lookup_entry(map, start, &temp_entry))
1137 return (KERN_NO_SPACE);
1139 prev_entry = temp_entry;
1142 * Assert that the next entry doesn't overlap the end point.
1144 if ((prev_entry->next != &map->header) &&
1145 (prev_entry->next->start < end))
1146 return (KERN_NO_SPACE);
1149 charge_prev_obj = FALSE;
1151 if (cow & MAP_COPY_ON_WRITE)
1152 protoeflags |= MAP_ENTRY_COW|MAP_ENTRY_NEEDS_COPY;
1154 if (cow & MAP_NOFAULT) {
1155 protoeflags |= MAP_ENTRY_NOFAULT;
1157 KASSERT(object == NULL,
1158 ("vm_map_insert: paradoxical MAP_NOFAULT request"));
1160 if (cow & MAP_DISABLE_SYNCER)
1161 protoeflags |= MAP_ENTRY_NOSYNC;
1162 if (cow & MAP_DISABLE_COREDUMP)
1163 protoeflags |= MAP_ENTRY_NOCOREDUMP;
1164 if (cow & MAP_VN_WRITECOUNT)
1165 protoeflags |= MAP_ENTRY_VN_WRITECNT;
1166 if (cow & MAP_INHERIT_SHARE)
1167 inheritance = VM_INHERIT_SHARE;
1169 inheritance = VM_INHERIT_DEFAULT;
1172 KASSERT((object != kmem_object && object != kernel_object) ||
1173 ((object == kmem_object || object == kernel_object) &&
1174 !(protoeflags & MAP_ENTRY_NEEDS_COPY)),
1175 ("kmem or kernel object and cow"));
1176 if (cow & (MAP_ACC_NO_CHARGE | MAP_NOFAULT))
1178 if ((cow & MAP_ACC_CHARGED) || ((prot & VM_PROT_WRITE) &&
1179 ((protoeflags & MAP_ENTRY_NEEDS_COPY) || object == NULL))) {
1180 if (!(cow & MAP_ACC_CHARGED) && !swap_reserve(end - start))
1181 return (KERN_RESOURCE_SHORTAGE);
1182 KASSERT(object == NULL || (protoeflags & MAP_ENTRY_NEEDS_COPY) ||
1183 object->cred == NULL,
1184 ("OVERCOMMIT: vm_map_insert o %p", object));
1185 cred = curthread->td_ucred;
1187 if (object == NULL && !(protoeflags & MAP_ENTRY_NEEDS_COPY))
1188 charge_prev_obj = TRUE;
1192 /* Expand the kernel pmap, if necessary. */
1193 if (map == kernel_map && end > kernel_vm_end)
1194 pmap_growkernel(end);
1195 if (object != NULL) {
1197 * OBJ_ONEMAPPING must be cleared unless this mapping
1198 * is trivially proven to be the only mapping for any
1199 * of the object's pages. (Object granularity
1200 * reference counting is insufficient to recognize
1201 * aliases with precision.)
1203 VM_OBJECT_WLOCK(object);
1204 if (object->ref_count > 1 || object->shadow_count != 0)
1205 vm_object_clear_flag(object, OBJ_ONEMAPPING);
1206 VM_OBJECT_WUNLOCK(object);
1208 else if ((prev_entry != &map->header) &&
1209 (prev_entry->eflags == protoeflags) &&
1210 (cow & (MAP_ENTRY_GROWS_DOWN | MAP_ENTRY_GROWS_UP)) == 0 &&
1211 (prev_entry->end == start) &&
1212 (prev_entry->wired_count == 0) &&
1213 (prev_entry->cred == cred ||
1214 (prev_entry->object.vm_object != NULL &&
1215 (prev_entry->object.vm_object->cred == cred))) &&
1216 vm_object_coalesce(prev_entry->object.vm_object,
1218 (vm_size_t)(prev_entry->end - prev_entry->start),
1219 (vm_size_t)(end - prev_entry->end), charge_prev_obj)) {
1221 * We were able to extend the object. Determine if we
1222 * can extend the previous map entry to include the
1223 * new range as well.
1225 if ((prev_entry->inheritance == inheritance) &&
1226 (prev_entry->protection == prot) &&
1227 (prev_entry->max_protection == max)) {
1228 map->size += (end - prev_entry->end);
1229 prev_entry->end = end;
1230 vm_map_entry_resize_free(map, prev_entry);
1231 vm_map_simplify_entry(map, prev_entry);
1234 return (KERN_SUCCESS);
1238 * If we can extend the object but cannot extend the
1239 * map entry, we have to create a new map entry. We
1240 * must bump the ref count on the extended object to
1241 * account for it. object may be NULL.
1243 object = prev_entry->object.vm_object;
1244 offset = prev_entry->offset +
1245 (prev_entry->end - prev_entry->start);
1246 vm_object_reference(object);
1247 if (cred != NULL && object != NULL && object->cred != NULL &&
1248 !(prev_entry->eflags & MAP_ENTRY_NEEDS_COPY)) {
1249 /* Object already accounts for this uid. */
1256 * NOTE: if conditionals fail, object can be NULL here. This occurs
1257 * in things like the buffer map where we manage kva but do not manage
1262 * Create a new entry
1264 new_entry = vm_map_entry_create(map);
1265 new_entry->start = start;
1266 new_entry->end = end;
1267 new_entry->cred = NULL;
1269 new_entry->eflags = protoeflags;
1270 new_entry->object.vm_object = object;
1271 new_entry->offset = offset;
1272 new_entry->avail_ssize = 0;
1274 new_entry->inheritance = inheritance;
1275 new_entry->protection = prot;
1276 new_entry->max_protection = max;
1277 new_entry->wired_count = 0;
1278 new_entry->read_ahead = VM_FAULT_READ_AHEAD_INIT;
1279 new_entry->next_read = OFF_TO_IDX(offset);
1281 KASSERT(cred == NULL || !ENTRY_CHARGED(new_entry),
1282 ("OVERCOMMIT: vm_map_insert leaks vm_map %p", new_entry));
1283 new_entry->cred = cred;
1286 * Insert the new entry into the list
1288 vm_map_entry_link(map, prev_entry, new_entry);
1289 map->size += new_entry->end - new_entry->start;
1292 * It may be possible to merge the new entry with the next and/or
1293 * previous entries. However, due to MAP_STACK_* being a hack, a
1294 * panic can result from merging such entries.
1296 if ((cow & (MAP_STACK_GROWS_DOWN | MAP_STACK_GROWS_UP)) == 0)
1297 vm_map_simplify_entry(map, new_entry);
1299 if (cow & (MAP_PREFAULT|MAP_PREFAULT_PARTIAL)) {
1300 vm_map_pmap_enter(map, start, prot,
1301 object, OFF_TO_IDX(offset), end - start,
1302 cow & MAP_PREFAULT_PARTIAL);
1305 return (KERN_SUCCESS);
1311 * Find the first fit (lowest VM address) for "length" free bytes
1312 * beginning at address >= start in the given map.
1314 * In a vm_map_entry, "adj_free" is the amount of free space
1315 * adjacent (higher address) to this entry, and "max_free" is the
1316 * maximum amount of contiguous free space in its subtree. This
1317 * allows finding a free region in one path down the tree, so
1318 * O(log n) amortized with splay trees.
1320 * The map must be locked, and leaves it so.
1322 * Returns: 0 on success, and starting address in *addr,
1323 * 1 if insufficient space.
1326 vm_map_findspace(vm_map_t map, vm_offset_t start, vm_size_t length,
1327 vm_offset_t *addr) /* OUT */
1329 vm_map_entry_t entry;
1333 * Request must fit within min/max VM address and must avoid
1336 if (start < map->min_offset)
1337 start = map->min_offset;
1338 if (start + length > map->max_offset || start + length < start)
1341 /* Empty tree means wide open address space. */
1342 if (map->root == NULL) {
1348 * After splay, if start comes before root node, then there
1349 * must be a gap from start to the root.
1351 map->root = vm_map_entry_splay(start, map->root);
1352 if (start + length <= map->root->start) {
1358 * Root is the last node that might begin its gap before
1359 * start, and this is the last comparison where address
1360 * wrap might be a problem.
1362 st = (start > map->root->end) ? start : map->root->end;
1363 if (length <= map->root->end + map->root->adj_free - st) {
1368 /* With max_free, can immediately tell if no solution. */
1369 entry = map->root->right;
1370 if (entry == NULL || length > entry->max_free)
1374 * Search the right subtree in the order: left subtree, root,
1375 * right subtree (first fit). The previous splay implies that
1376 * all regions in the right subtree have addresses > start.
1378 while (entry != NULL) {
1379 if (entry->left != NULL && entry->left->max_free >= length)
1380 entry = entry->left;
1381 else if (entry->adj_free >= length) {
1385 entry = entry->right;
1388 /* Can't get here, so panic if we do. */
1389 panic("vm_map_findspace: max_free corrupt");
1393 vm_map_fixed(vm_map_t map, vm_object_t object, vm_ooffset_t offset,
1394 vm_offset_t start, vm_size_t length, vm_prot_t prot,
1395 vm_prot_t max, int cow)
1400 end = start + length;
1402 VM_MAP_RANGE_CHECK(map, start, end);
1403 (void) vm_map_delete(map, start, end);
1404 result = vm_map_insert(map, object, offset, start, end, prot,
1411 * vm_map_find finds an unallocated region in the target address
1412 * map with the given length. The search is defined to be
1413 * first-fit from the specified address; the region found is
1414 * returned in the same parameter.
1416 * If object is non-NULL, ref count must be bumped by caller
1417 * prior to making call to account for the new entry.
1420 vm_map_find(vm_map_t map, vm_object_t object, vm_ooffset_t offset,
1421 vm_offset_t *addr, /* IN/OUT */
1422 vm_size_t length, vm_offset_t max_addr, int find_space,
1423 vm_prot_t prot, vm_prot_t max, int cow)
1425 vm_offset_t alignment, initial_addr, start;
1428 if (find_space == VMFS_OPTIMAL_SPACE && (object == NULL ||
1429 (object->flags & OBJ_COLORED) == 0))
1430 find_space = VMFS_ANY_SPACE;
1431 if (find_space >> 8 != 0) {
1432 KASSERT((find_space & 0xff) == 0, ("bad VMFS flags"));
1433 alignment = (vm_offset_t)1 << (find_space >> 8);
1436 initial_addr = *addr;
1438 start = initial_addr;
1441 if (find_space != VMFS_NO_SPACE) {
1442 if (vm_map_findspace(map, start, length, addr) ||
1443 (max_addr != 0 && *addr + length > max_addr)) {
1445 if (find_space == VMFS_OPTIMAL_SPACE) {
1446 find_space = VMFS_ANY_SPACE;
1449 return (KERN_NO_SPACE);
1451 switch (find_space) {
1452 case VMFS_SUPER_SPACE:
1453 case VMFS_OPTIMAL_SPACE:
1454 pmap_align_superpage(object, offset, addr,
1457 case VMFS_ANY_SPACE:
1460 if ((*addr & (alignment - 1)) != 0) {
1461 *addr &= ~(alignment - 1);
1469 result = vm_map_insert(map, object, offset, start, start +
1470 length, prot, max, cow);
1471 } while (result == KERN_NO_SPACE && find_space != VMFS_NO_SPACE &&
1472 find_space != VMFS_ANY_SPACE);
1478 * vm_map_simplify_entry:
1480 * Simplify the given map entry by merging with either neighbor. This
1481 * routine also has the ability to merge with both neighbors.
1483 * The map must be locked.
1485 * This routine guarentees that the passed entry remains valid (though
1486 * possibly extended). When merging, this routine may delete one or
1490 vm_map_simplify_entry(vm_map_t map, vm_map_entry_t entry)
1492 vm_map_entry_t next, prev;
1493 vm_size_t prevsize, esize;
1495 if (entry->eflags & (MAP_ENTRY_IN_TRANSITION | MAP_ENTRY_IS_SUB_MAP))
1499 if (prev != &map->header) {
1500 prevsize = prev->end - prev->start;
1501 if ( (prev->end == entry->start) &&
1502 (prev->object.vm_object == entry->object.vm_object) &&
1503 (!prev->object.vm_object ||
1504 (prev->offset + prevsize == entry->offset)) &&
1505 (prev->eflags == entry->eflags) &&
1506 (prev->protection == entry->protection) &&
1507 (prev->max_protection == entry->max_protection) &&
1508 (prev->inheritance == entry->inheritance) &&
1509 (prev->wired_count == entry->wired_count) &&
1510 (prev->cred == entry->cred)) {
1511 vm_map_entry_unlink(map, prev);
1512 entry->start = prev->start;
1513 entry->offset = prev->offset;
1514 if (entry->prev != &map->header)
1515 vm_map_entry_resize_free(map, entry->prev);
1518 * If the backing object is a vnode object,
1519 * vm_object_deallocate() calls vrele().
1520 * However, vrele() does not lock the vnode
1521 * because the vnode has additional
1522 * references. Thus, the map lock can be kept
1523 * without causing a lock-order reversal with
1526 * Since we count the number of virtual page
1527 * mappings in object->un_pager.vnp.writemappings,
1528 * the writemappings value should not be adjusted
1529 * when the entry is disposed of.
1531 if (prev->object.vm_object)
1532 vm_object_deallocate(prev->object.vm_object);
1533 if (prev->cred != NULL)
1535 vm_map_entry_dispose(map, prev);
1540 if (next != &map->header) {
1541 esize = entry->end - entry->start;
1542 if ((entry->end == next->start) &&
1543 (next->object.vm_object == entry->object.vm_object) &&
1544 (!entry->object.vm_object ||
1545 (entry->offset + esize == next->offset)) &&
1546 (next->eflags == entry->eflags) &&
1547 (next->protection == entry->protection) &&
1548 (next->max_protection == entry->max_protection) &&
1549 (next->inheritance == entry->inheritance) &&
1550 (next->wired_count == entry->wired_count) &&
1551 (next->cred == entry->cred)) {
1552 vm_map_entry_unlink(map, next);
1553 entry->end = next->end;
1554 vm_map_entry_resize_free(map, entry);
1557 * See comment above.
1559 if (next->object.vm_object)
1560 vm_object_deallocate(next->object.vm_object);
1561 if (next->cred != NULL)
1563 vm_map_entry_dispose(map, next);
1568 * vm_map_clip_start: [ internal use only ]
1570 * Asserts that the given entry begins at or after
1571 * the specified address; if necessary,
1572 * it splits the entry into two.
1574 #define vm_map_clip_start(map, entry, startaddr) \
1576 if (startaddr > entry->start) \
1577 _vm_map_clip_start(map, entry, startaddr); \
1581 * This routine is called only when it is known that
1582 * the entry must be split.
1585 _vm_map_clip_start(vm_map_t map, vm_map_entry_t entry, vm_offset_t start)
1587 vm_map_entry_t new_entry;
1589 VM_MAP_ASSERT_LOCKED(map);
1592 * Split off the front portion -- note that we must insert the new
1593 * entry BEFORE this one, so that this entry has the specified
1596 vm_map_simplify_entry(map, entry);
1599 * If there is no object backing this entry, we might as well create
1600 * one now. If we defer it, an object can get created after the map
1601 * is clipped, and individual objects will be created for the split-up
1602 * map. This is a bit of a hack, but is also about the best place to
1603 * put this improvement.
1605 if (entry->object.vm_object == NULL && !map->system_map) {
1607 object = vm_object_allocate(OBJT_DEFAULT,
1608 atop(entry->end - entry->start));
1609 entry->object.vm_object = object;
1611 if (entry->cred != NULL) {
1612 object->cred = entry->cred;
1613 object->charge = entry->end - entry->start;
1616 } else if (entry->object.vm_object != NULL &&
1617 ((entry->eflags & MAP_ENTRY_NEEDS_COPY) == 0) &&
1618 entry->cred != NULL) {
1619 VM_OBJECT_WLOCK(entry->object.vm_object);
1620 KASSERT(entry->object.vm_object->cred == NULL,
1621 ("OVERCOMMIT: vm_entry_clip_start: both cred e %p", entry));
1622 entry->object.vm_object->cred = entry->cred;
1623 entry->object.vm_object->charge = entry->end - entry->start;
1624 VM_OBJECT_WUNLOCK(entry->object.vm_object);
1628 new_entry = vm_map_entry_create(map);
1629 *new_entry = *entry;
1631 new_entry->end = start;
1632 entry->offset += (start - entry->start);
1633 entry->start = start;
1634 if (new_entry->cred != NULL)
1635 crhold(entry->cred);
1637 vm_map_entry_link(map, entry->prev, new_entry);
1639 if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
1640 vm_object_reference(new_entry->object.vm_object);
1642 * The object->un_pager.vnp.writemappings for the
1643 * object of MAP_ENTRY_VN_WRITECNT type entry shall be
1644 * kept as is here. The virtual pages are
1645 * re-distributed among the clipped entries, so the sum is
1652 * vm_map_clip_end: [ internal use only ]
1654 * Asserts that the given entry ends at or before
1655 * the specified address; if necessary,
1656 * it splits the entry into two.
1658 #define vm_map_clip_end(map, entry, endaddr) \
1660 if ((endaddr) < (entry->end)) \
1661 _vm_map_clip_end((map), (entry), (endaddr)); \
1665 * This routine is called only when it is known that
1666 * the entry must be split.
1669 _vm_map_clip_end(vm_map_t map, vm_map_entry_t entry, vm_offset_t end)
1671 vm_map_entry_t new_entry;
1673 VM_MAP_ASSERT_LOCKED(map);
1676 * If there is no object backing this entry, we might as well create
1677 * one now. If we defer it, an object can get created after the map
1678 * is clipped, and individual objects will be created for the split-up
1679 * map. This is a bit of a hack, but is also about the best place to
1680 * put this improvement.
1682 if (entry->object.vm_object == NULL && !map->system_map) {
1684 object = vm_object_allocate(OBJT_DEFAULT,
1685 atop(entry->end - entry->start));
1686 entry->object.vm_object = object;
1688 if (entry->cred != NULL) {
1689 object->cred = entry->cred;
1690 object->charge = entry->end - entry->start;
1693 } else if (entry->object.vm_object != NULL &&
1694 ((entry->eflags & MAP_ENTRY_NEEDS_COPY) == 0) &&
1695 entry->cred != NULL) {
1696 VM_OBJECT_WLOCK(entry->object.vm_object);
1697 KASSERT(entry->object.vm_object->cred == NULL,
1698 ("OVERCOMMIT: vm_entry_clip_end: both cred e %p", entry));
1699 entry->object.vm_object->cred = entry->cred;
1700 entry->object.vm_object->charge = entry->end - entry->start;
1701 VM_OBJECT_WUNLOCK(entry->object.vm_object);
1706 * Create a new entry and insert it AFTER the specified entry
1708 new_entry = vm_map_entry_create(map);
1709 *new_entry = *entry;
1711 new_entry->start = entry->end = end;
1712 new_entry->offset += (end - entry->start);
1713 if (new_entry->cred != NULL)
1714 crhold(entry->cred);
1716 vm_map_entry_link(map, entry, new_entry);
1718 if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
1719 vm_object_reference(new_entry->object.vm_object);
1724 * vm_map_submap: [ kernel use only ]
1726 * Mark the given range as handled by a subordinate map.
1728 * This range must have been created with vm_map_find,
1729 * and no other operations may have been performed on this
1730 * range prior to calling vm_map_submap.
1732 * Only a limited number of operations can be performed
1733 * within this rage after calling vm_map_submap:
1735 * [Don't try vm_map_copy!]
1737 * To remove a submapping, one must first remove the
1738 * range from the superior map, and then destroy the
1739 * submap (if desired). [Better yet, don't try it.]
1748 vm_map_entry_t entry;
1749 int result = KERN_INVALID_ARGUMENT;
1753 VM_MAP_RANGE_CHECK(map, start, end);
1755 if (vm_map_lookup_entry(map, start, &entry)) {
1756 vm_map_clip_start(map, entry, start);
1758 entry = entry->next;
1760 vm_map_clip_end(map, entry, end);
1762 if ((entry->start == start) && (entry->end == end) &&
1763 ((entry->eflags & MAP_ENTRY_COW) == 0) &&
1764 (entry->object.vm_object == NULL)) {
1765 entry->object.sub_map = submap;
1766 entry->eflags |= MAP_ENTRY_IS_SUB_MAP;
1767 result = KERN_SUCCESS;
1775 * The maximum number of pages to map
1777 #define MAX_INIT_PT 96
1780 * vm_map_pmap_enter:
1782 * Preload read-only mappings for the specified object's resident pages
1783 * into the target map. If "flags" is MAP_PREFAULT_PARTIAL, then only
1784 * the resident pages within the address range [addr, addr + ulmin(size,
1785 * ptoa(MAX_INIT_PT))) are mapped. Otherwise, all resident pages within
1786 * the specified address range are mapped. This eliminates many soft
1787 * faults on process startup and immediately after an mmap(2). Because
1788 * these are speculative mappings, cached pages are not reactivated and
1792 vm_map_pmap_enter(vm_map_t map, vm_offset_t addr, vm_prot_t prot,
1793 vm_object_t object, vm_pindex_t pindex, vm_size_t size, int flags)
1796 vm_page_t p, p_start;
1797 vm_pindex_t psize, tmpidx;
1799 if ((prot & (VM_PROT_READ | VM_PROT_EXECUTE)) == 0 || object == NULL)
1801 VM_OBJECT_RLOCK(object);
1802 if (object->type == OBJT_DEVICE || object->type == OBJT_SG) {
1803 VM_OBJECT_RUNLOCK(object);
1804 VM_OBJECT_WLOCK(object);
1805 if (object->type == OBJT_DEVICE || object->type == OBJT_SG) {
1806 pmap_object_init_pt(map->pmap, addr, object, pindex,
1808 VM_OBJECT_WUNLOCK(object);
1811 VM_OBJECT_LOCK_DOWNGRADE(object);
1815 if (psize > MAX_INIT_PT && (flags & MAP_PREFAULT_PARTIAL) != 0)
1816 psize = MAX_INIT_PT;
1817 if (psize + pindex > object->size) {
1818 if (object->size < pindex) {
1819 VM_OBJECT_RUNLOCK(object);
1822 psize = object->size - pindex;
1828 p = vm_page_find_least(object, pindex);
1830 * Assert: the variable p is either (1) the page with the
1831 * least pindex greater than or equal to the parameter pindex
1835 p != NULL && (tmpidx = p->pindex - pindex) < psize;
1836 p = TAILQ_NEXT(p, listq)) {
1838 * don't allow an madvise to blow away our really
1839 * free pages allocating pv entries.
1841 if ((flags & MAP_PREFAULT_MADVISE) &&
1842 cnt.v_free_count < cnt.v_free_reserved) {
1846 if (p->valid == VM_PAGE_BITS_ALL) {
1847 if (p_start == NULL) {
1848 start = addr + ptoa(tmpidx);
1851 } else if (p_start != NULL) {
1852 pmap_enter_object(map->pmap, start, addr +
1853 ptoa(tmpidx), p_start, prot);
1857 if (p_start != NULL)
1858 pmap_enter_object(map->pmap, start, addr + ptoa(psize),
1860 VM_OBJECT_RUNLOCK(object);
1866 * Sets the protection of the specified address
1867 * region in the target map. If "set_max" is
1868 * specified, the maximum protection is to be set;
1869 * otherwise, only the current protection is affected.
1872 vm_map_protect(vm_map_t map, vm_offset_t start, vm_offset_t end,
1873 vm_prot_t new_prot, boolean_t set_max)
1875 vm_map_entry_t current, entry;
1881 return (KERN_SUCCESS);
1885 VM_MAP_RANGE_CHECK(map, start, end);
1887 if (vm_map_lookup_entry(map, start, &entry)) {
1888 vm_map_clip_start(map, entry, start);
1890 entry = entry->next;
1894 * Make a first pass to check for protection violations.
1897 while ((current != &map->header) && (current->start < end)) {
1898 if (current->eflags & MAP_ENTRY_IS_SUB_MAP) {
1900 return (KERN_INVALID_ARGUMENT);
1902 if ((new_prot & current->max_protection) != new_prot) {
1904 return (KERN_PROTECTION_FAILURE);
1906 current = current->next;
1911 * Do an accounting pass for private read-only mappings that
1912 * now will do cow due to allowed write (e.g. debugger sets
1913 * breakpoint on text segment)
1915 for (current = entry; (current != &map->header) &&
1916 (current->start < end); current = current->next) {
1918 vm_map_clip_end(map, current, end);
1921 ((new_prot & ~(current->protection)) & VM_PROT_WRITE) == 0 ||
1922 ENTRY_CHARGED(current)) {
1926 cred = curthread->td_ucred;
1927 obj = current->object.vm_object;
1929 if (obj == NULL || (current->eflags & MAP_ENTRY_NEEDS_COPY)) {
1930 if (!swap_reserve(current->end - current->start)) {
1932 return (KERN_RESOURCE_SHORTAGE);
1935 current->cred = cred;
1939 VM_OBJECT_WLOCK(obj);
1940 if (obj->type != OBJT_DEFAULT && obj->type != OBJT_SWAP) {
1941 VM_OBJECT_WUNLOCK(obj);
1946 * Charge for the whole object allocation now, since
1947 * we cannot distinguish between non-charged and
1948 * charged clipped mapping of the same object later.
1950 KASSERT(obj->charge == 0,
1951 ("vm_map_protect: object %p overcharged\n", obj));
1952 if (!swap_reserve(ptoa(obj->size))) {
1953 VM_OBJECT_WUNLOCK(obj);
1955 return (KERN_RESOURCE_SHORTAGE);
1960 obj->charge = ptoa(obj->size);
1961 VM_OBJECT_WUNLOCK(obj);
1965 * Go back and fix up protections. [Note that clipping is not
1966 * necessary the second time.]
1969 while ((current != &map->header) && (current->start < end)) {
1970 old_prot = current->protection;
1973 current->protection =
1974 (current->max_protection = new_prot) &
1977 current->protection = new_prot;
1979 if ((current->eflags & (MAP_ENTRY_COW | MAP_ENTRY_USER_WIRED))
1980 == (MAP_ENTRY_COW | MAP_ENTRY_USER_WIRED) &&
1981 (current->protection & VM_PROT_WRITE) != 0 &&
1982 (old_prot & VM_PROT_WRITE) == 0) {
1983 vm_fault_copy_entry(map, map, current, current, NULL);
1987 * When restricting access, update the physical map. Worry
1988 * about copy-on-write here.
1990 if ((old_prot & ~current->protection) != 0) {
1991 #define MASK(entry) (((entry)->eflags & MAP_ENTRY_COW) ? ~VM_PROT_WRITE : \
1993 pmap_protect(map->pmap, current->start,
1995 current->protection & MASK(current));
1998 vm_map_simplify_entry(map, current);
1999 current = current->next;
2002 return (KERN_SUCCESS);
2008 * This routine traverses a processes map handling the madvise
2009 * system call. Advisories are classified as either those effecting
2010 * the vm_map_entry structure, or those effecting the underlying
2020 vm_map_entry_t current, entry;
2024 * Some madvise calls directly modify the vm_map_entry, in which case
2025 * we need to use an exclusive lock on the map and we need to perform
2026 * various clipping operations. Otherwise we only need a read-lock
2031 case MADV_SEQUENTIAL:
2038 return (KERN_SUCCESS);
2046 return (KERN_SUCCESS);
2047 vm_map_lock_read(map);
2050 return (KERN_INVALID_ARGUMENT);
2054 * Locate starting entry and clip if necessary.
2056 VM_MAP_RANGE_CHECK(map, start, end);
2058 if (vm_map_lookup_entry(map, start, &entry)) {
2060 vm_map_clip_start(map, entry, start);
2062 entry = entry->next;
2067 * madvise behaviors that are implemented in the vm_map_entry.
2069 * We clip the vm_map_entry so that behavioral changes are
2070 * limited to the specified address range.
2072 for (current = entry;
2073 (current != &map->header) && (current->start < end);
2074 current = current->next
2076 if (current->eflags & MAP_ENTRY_IS_SUB_MAP)
2079 vm_map_clip_end(map, current, end);
2083 vm_map_entry_set_behavior(current, MAP_ENTRY_BEHAV_NORMAL);
2085 case MADV_SEQUENTIAL:
2086 vm_map_entry_set_behavior(current, MAP_ENTRY_BEHAV_SEQUENTIAL);
2089 vm_map_entry_set_behavior(current, MAP_ENTRY_BEHAV_RANDOM);
2092 current->eflags |= MAP_ENTRY_NOSYNC;
2095 current->eflags &= ~MAP_ENTRY_NOSYNC;
2098 current->eflags |= MAP_ENTRY_NOCOREDUMP;
2101 current->eflags &= ~MAP_ENTRY_NOCOREDUMP;
2106 vm_map_simplify_entry(map, current);
2110 vm_pindex_t pstart, pend;
2113 * madvise behaviors that are implemented in the underlying
2116 * Since we don't clip the vm_map_entry, we have to clip
2117 * the vm_object pindex and count.
2119 for (current = entry;
2120 (current != &map->header) && (current->start < end);
2121 current = current->next
2123 vm_offset_t useEnd, useStart;
2125 if (current->eflags & MAP_ENTRY_IS_SUB_MAP)
2128 pstart = OFF_TO_IDX(current->offset);
2129 pend = pstart + atop(current->end - current->start);
2130 useStart = current->start;
2131 useEnd = current->end;
2133 if (current->start < start) {
2134 pstart += atop(start - current->start);
2137 if (current->end > end) {
2138 pend -= atop(current->end - end);
2146 * Perform the pmap_advise() before clearing
2147 * PGA_REFERENCED in vm_page_advise(). Otherwise, a
2148 * concurrent pmap operation, such as pmap_remove(),
2149 * could clear a reference in the pmap and set
2150 * PGA_REFERENCED on the page before the pmap_advise()
2151 * had completed. Consequently, the page would appear
2152 * referenced based upon an old reference that
2153 * occurred before this pmap_advise() ran.
2155 if (behav == MADV_DONTNEED || behav == MADV_FREE)
2156 pmap_advise(map->pmap, useStart, useEnd,
2159 vm_object_madvise(current->object.vm_object, pstart,
2161 if (behav == MADV_WILLNEED) {
2162 vm_map_pmap_enter(map,
2164 current->protection,
2165 current->object.vm_object,
2167 ptoa(pend - pstart),
2168 MAP_PREFAULT_MADVISE
2172 vm_map_unlock_read(map);
2181 * Sets the inheritance of the specified address
2182 * range in the target map. Inheritance
2183 * affects how the map will be shared with
2184 * child maps at the time of vmspace_fork.
2187 vm_map_inherit(vm_map_t map, vm_offset_t start, vm_offset_t end,
2188 vm_inherit_t new_inheritance)
2190 vm_map_entry_t entry;
2191 vm_map_entry_t temp_entry;
2193 switch (new_inheritance) {
2194 case VM_INHERIT_NONE:
2195 case VM_INHERIT_COPY:
2196 case VM_INHERIT_SHARE:
2199 return (KERN_INVALID_ARGUMENT);
2202 return (KERN_SUCCESS);
2204 VM_MAP_RANGE_CHECK(map, start, end);
2205 if (vm_map_lookup_entry(map, start, &temp_entry)) {
2207 vm_map_clip_start(map, entry, start);
2209 entry = temp_entry->next;
2210 while ((entry != &map->header) && (entry->start < end)) {
2211 vm_map_clip_end(map, entry, end);
2212 entry->inheritance = new_inheritance;
2213 vm_map_simplify_entry(map, entry);
2214 entry = entry->next;
2217 return (KERN_SUCCESS);
2223 * Implements both kernel and user unwiring.
2226 vm_map_unwire(vm_map_t map, vm_offset_t start, vm_offset_t end,
2229 vm_map_entry_t entry, first_entry, tmp_entry;
2230 vm_offset_t saved_start;
2231 unsigned int last_timestamp;
2233 boolean_t need_wakeup, result, user_unwire;
2236 return (KERN_SUCCESS);
2237 user_unwire = (flags & VM_MAP_WIRE_USER) ? TRUE : FALSE;
2239 VM_MAP_RANGE_CHECK(map, start, end);
2240 if (!vm_map_lookup_entry(map, start, &first_entry)) {
2241 if (flags & VM_MAP_WIRE_HOLESOK)
2242 first_entry = first_entry->next;
2245 return (KERN_INVALID_ADDRESS);
2248 last_timestamp = map->timestamp;
2249 entry = first_entry;
2250 while (entry != &map->header && entry->start < end) {
2251 if (entry->eflags & MAP_ENTRY_IN_TRANSITION) {
2253 * We have not yet clipped the entry.
2255 saved_start = (start >= entry->start) ? start :
2257 entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP;
2258 if (vm_map_unlock_and_wait(map, 0)) {
2260 * Allow interruption of user unwiring?
2264 if (last_timestamp+1 != map->timestamp) {
2266 * Look again for the entry because the map was
2267 * modified while it was unlocked.
2268 * Specifically, the entry may have been
2269 * clipped, merged, or deleted.
2271 if (!vm_map_lookup_entry(map, saved_start,
2273 if (flags & VM_MAP_WIRE_HOLESOK)
2274 tmp_entry = tmp_entry->next;
2276 if (saved_start == start) {
2278 * First_entry has been deleted.
2281 return (KERN_INVALID_ADDRESS);
2284 rv = KERN_INVALID_ADDRESS;
2288 if (entry == first_entry)
2289 first_entry = tmp_entry;
2294 last_timestamp = map->timestamp;
2297 vm_map_clip_start(map, entry, start);
2298 vm_map_clip_end(map, entry, end);
2300 * Mark the entry in case the map lock is released. (See
2303 KASSERT((entry->eflags & MAP_ENTRY_IN_TRANSITION) == 0 &&
2304 entry->wiring_thread == NULL,
2305 ("owned map entry %p", entry));
2306 entry->eflags |= MAP_ENTRY_IN_TRANSITION;
2307 entry->wiring_thread = curthread;
2309 * Check the map for holes in the specified region.
2310 * If VM_MAP_WIRE_HOLESOK was specified, skip this check.
2312 if (((flags & VM_MAP_WIRE_HOLESOK) == 0) &&
2313 (entry->end < end && (entry->next == &map->header ||
2314 entry->next->start > entry->end))) {
2316 rv = KERN_INVALID_ADDRESS;
2320 * If system unwiring, require that the entry is system wired.
2323 vm_map_entry_system_wired_count(entry) == 0) {
2325 rv = KERN_INVALID_ARGUMENT;
2328 entry = entry->next;
2332 need_wakeup = FALSE;
2333 if (first_entry == NULL) {
2334 result = vm_map_lookup_entry(map, start, &first_entry);
2335 if (!result && (flags & VM_MAP_WIRE_HOLESOK))
2336 first_entry = first_entry->next;
2338 KASSERT(result, ("vm_map_unwire: lookup failed"));
2340 for (entry = first_entry; entry != &map->header && entry->start < end;
2341 entry = entry->next) {
2343 * If VM_MAP_WIRE_HOLESOK was specified, an empty
2344 * space in the unwired region could have been mapped
2345 * while the map lock was dropped for draining
2346 * MAP_ENTRY_IN_TRANSITION. Moreover, another thread
2347 * could be simultaneously wiring this new mapping
2348 * entry. Detect these cases and skip any entries
2349 * marked as in transition by us.
2351 if ((entry->eflags & MAP_ENTRY_IN_TRANSITION) == 0 ||
2352 entry->wiring_thread != curthread) {
2353 KASSERT((flags & VM_MAP_WIRE_HOLESOK) != 0,
2354 ("vm_map_unwire: !HOLESOK and new/changed entry"));
2358 if (rv == KERN_SUCCESS && (!user_unwire ||
2359 (entry->eflags & MAP_ENTRY_USER_WIRED))) {
2361 entry->eflags &= ~MAP_ENTRY_USER_WIRED;
2362 entry->wired_count--;
2363 if (entry->wired_count == 0) {
2365 * Retain the map lock.
2367 vm_fault_unwire(map, entry->start, entry->end,
2368 entry->object.vm_object != NULL &&
2369 (entry->object.vm_object->flags &
2370 OBJ_FICTITIOUS) != 0);
2373 KASSERT((entry->eflags & MAP_ENTRY_IN_TRANSITION) != 0,
2374 ("vm_map_unwire: in-transition flag missing %p", entry));
2375 KASSERT(entry->wiring_thread == curthread,
2376 ("vm_map_unwire: alien wire %p", entry));
2377 entry->eflags &= ~MAP_ENTRY_IN_TRANSITION;
2378 entry->wiring_thread = NULL;
2379 if (entry->eflags & MAP_ENTRY_NEEDS_WAKEUP) {
2380 entry->eflags &= ~MAP_ENTRY_NEEDS_WAKEUP;
2383 vm_map_simplify_entry(map, entry);
2394 * Implements both kernel and user wiring.
2397 vm_map_wire(vm_map_t map, vm_offset_t start, vm_offset_t end,
2400 vm_map_entry_t entry, first_entry, tmp_entry;
2401 vm_offset_t saved_end, saved_start;
2402 unsigned int last_timestamp;
2404 boolean_t fictitious, need_wakeup, result, user_wire;
2408 return (KERN_SUCCESS);
2410 if (flags & VM_MAP_WIRE_WRITE)
2411 prot |= VM_PROT_WRITE;
2412 user_wire = (flags & VM_MAP_WIRE_USER) ? TRUE : FALSE;
2414 VM_MAP_RANGE_CHECK(map, start, end);
2415 if (!vm_map_lookup_entry(map, start, &first_entry)) {
2416 if (flags & VM_MAP_WIRE_HOLESOK)
2417 first_entry = first_entry->next;
2420 return (KERN_INVALID_ADDRESS);
2423 last_timestamp = map->timestamp;
2424 entry = first_entry;
2425 while (entry != &map->header && entry->start < end) {
2426 if (entry->eflags & MAP_ENTRY_IN_TRANSITION) {
2428 * We have not yet clipped the entry.
2430 saved_start = (start >= entry->start) ? start :
2432 entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP;
2433 if (vm_map_unlock_and_wait(map, 0)) {
2435 * Allow interruption of user wiring?
2439 if (last_timestamp + 1 != map->timestamp) {
2441 * Look again for the entry because the map was
2442 * modified while it was unlocked.
2443 * Specifically, the entry may have been
2444 * clipped, merged, or deleted.
2446 if (!vm_map_lookup_entry(map, saved_start,
2448 if (flags & VM_MAP_WIRE_HOLESOK)
2449 tmp_entry = tmp_entry->next;
2451 if (saved_start == start) {
2453 * first_entry has been deleted.
2456 return (KERN_INVALID_ADDRESS);
2459 rv = KERN_INVALID_ADDRESS;
2463 if (entry == first_entry)
2464 first_entry = tmp_entry;
2469 last_timestamp = map->timestamp;
2472 vm_map_clip_start(map, entry, start);
2473 vm_map_clip_end(map, entry, end);
2475 * Mark the entry in case the map lock is released. (See
2478 KASSERT((entry->eflags & MAP_ENTRY_IN_TRANSITION) == 0 &&
2479 entry->wiring_thread == NULL,
2480 ("owned map entry %p", entry));
2481 entry->eflags |= MAP_ENTRY_IN_TRANSITION;
2482 entry->wiring_thread = curthread;
2483 if ((entry->protection & (VM_PROT_READ | VM_PROT_EXECUTE)) == 0
2484 || (entry->protection & prot) != prot) {
2485 entry->eflags |= MAP_ENTRY_WIRE_SKIPPED;
2486 if ((flags & VM_MAP_WIRE_HOLESOK) == 0) {
2488 rv = KERN_INVALID_ADDRESS;
2493 if (entry->wired_count == 0) {
2494 entry->wired_count++;
2495 saved_start = entry->start;
2496 saved_end = entry->end;
2497 fictitious = entry->object.vm_object != NULL &&
2498 (entry->object.vm_object->flags &
2499 OBJ_FICTITIOUS) != 0;
2501 * Release the map lock, relying on the in-transition
2502 * mark. Mark the map busy for fork.
2506 rv = vm_fault_wire(map, saved_start, saved_end,
2510 if (last_timestamp + 1 != map->timestamp) {
2512 * Look again for the entry because the map was
2513 * modified while it was unlocked. The entry
2514 * may have been clipped, but NOT merged or
2517 result = vm_map_lookup_entry(map, saved_start,
2519 KASSERT(result, ("vm_map_wire: lookup failed"));
2520 if (entry == first_entry)
2521 first_entry = tmp_entry;
2525 while (entry->end < saved_end) {
2526 if (rv != KERN_SUCCESS) {
2527 KASSERT(entry->wired_count == 1,
2528 ("vm_map_wire: bad count"));
2529 entry->wired_count = -1;
2531 entry = entry->next;
2534 last_timestamp = map->timestamp;
2535 if (rv != KERN_SUCCESS) {
2536 KASSERT(entry->wired_count == 1,
2537 ("vm_map_wire: bad count"));
2539 * Assign an out-of-range value to represent
2540 * the failure to wire this entry.
2542 entry->wired_count = -1;
2546 } else if (!user_wire ||
2547 (entry->eflags & MAP_ENTRY_USER_WIRED) == 0) {
2548 entry->wired_count++;
2551 * Check the map for holes in the specified region.
2552 * If VM_MAP_WIRE_HOLESOK was specified, skip this check.
2555 if (((flags & VM_MAP_WIRE_HOLESOK) == 0) &&
2556 (entry->end < end && (entry->next == &map->header ||
2557 entry->next->start > entry->end))) {
2559 rv = KERN_INVALID_ADDRESS;
2562 entry = entry->next;
2566 need_wakeup = FALSE;
2567 if (first_entry == NULL) {
2568 result = vm_map_lookup_entry(map, start, &first_entry);
2569 if (!result && (flags & VM_MAP_WIRE_HOLESOK))
2570 first_entry = first_entry->next;
2572 KASSERT(result, ("vm_map_wire: lookup failed"));
2574 for (entry = first_entry; entry != &map->header && entry->start < end;
2575 entry = entry->next) {
2576 if ((entry->eflags & MAP_ENTRY_WIRE_SKIPPED) != 0)
2577 goto next_entry_done;
2580 * If VM_MAP_WIRE_HOLESOK was specified, an empty
2581 * space in the unwired region could have been mapped
2582 * while the map lock was dropped for faulting in the
2583 * pages or draining MAP_ENTRY_IN_TRANSITION.
2584 * Moreover, another thread could be simultaneously
2585 * wiring this new mapping entry. Detect these cases
2586 * and skip any entries marked as in transition by us.
2588 if ((entry->eflags & MAP_ENTRY_IN_TRANSITION) == 0 ||
2589 entry->wiring_thread != curthread) {
2590 KASSERT((flags & VM_MAP_WIRE_HOLESOK) != 0,
2591 ("vm_map_wire: !HOLESOK and new/changed entry"));
2595 if (rv == KERN_SUCCESS) {
2597 entry->eflags |= MAP_ENTRY_USER_WIRED;
2598 } else if (entry->wired_count == -1) {
2600 * Wiring failed on this entry. Thus, unwiring is
2603 entry->wired_count = 0;
2606 (entry->eflags & MAP_ENTRY_USER_WIRED) == 0)
2607 entry->wired_count--;
2608 if (entry->wired_count == 0) {
2610 * Retain the map lock.
2612 vm_fault_unwire(map, entry->start, entry->end,
2613 entry->object.vm_object != NULL &&
2614 (entry->object.vm_object->flags &
2615 OBJ_FICTITIOUS) != 0);
2619 KASSERT((entry->eflags & MAP_ENTRY_IN_TRANSITION) != 0,
2620 ("vm_map_wire: in-transition flag missing %p", entry));
2621 KASSERT(entry->wiring_thread == curthread,
2622 ("vm_map_wire: alien wire %p", entry));
2623 entry->eflags &= ~(MAP_ENTRY_IN_TRANSITION |
2624 MAP_ENTRY_WIRE_SKIPPED);
2625 entry->wiring_thread = NULL;
2626 if (entry->eflags & MAP_ENTRY_NEEDS_WAKEUP) {
2627 entry->eflags &= ~MAP_ENTRY_NEEDS_WAKEUP;
2630 vm_map_simplify_entry(map, entry);
2641 * Push any dirty cached pages in the address range to their pager.
2642 * If syncio is TRUE, dirty pages are written synchronously.
2643 * If invalidate is TRUE, any cached pages are freed as well.
2645 * If the size of the region from start to end is zero, we are
2646 * supposed to flush all modified pages within the region containing
2647 * start. Unfortunately, a region can be split or coalesced with
2648 * neighboring regions, making it difficult to determine what the
2649 * original region was. Therefore, we approximate this requirement by
2650 * flushing the current region containing start.
2652 * Returns an error if any part of the specified range is not mapped.
2660 boolean_t invalidate)
2662 vm_map_entry_t current;
2663 vm_map_entry_t entry;
2666 vm_ooffset_t offset;
2667 unsigned int last_timestamp;
2670 vm_map_lock_read(map);
2671 VM_MAP_RANGE_CHECK(map, start, end);
2672 if (!vm_map_lookup_entry(map, start, &entry)) {
2673 vm_map_unlock_read(map);
2674 return (KERN_INVALID_ADDRESS);
2675 } else if (start == end) {
2676 start = entry->start;
2680 * Make a first pass to check for user-wired memory and holes.
2682 for (current = entry; current != &map->header && current->start < end;
2683 current = current->next) {
2684 if (invalidate && (current->eflags & MAP_ENTRY_USER_WIRED)) {
2685 vm_map_unlock_read(map);
2686 return (KERN_INVALID_ARGUMENT);
2688 if (end > current->end &&
2689 (current->next == &map->header ||
2690 current->end != current->next->start)) {
2691 vm_map_unlock_read(map);
2692 return (KERN_INVALID_ADDRESS);
2697 pmap_remove(map->pmap, start, end);
2701 * Make a second pass, cleaning/uncaching pages from the indicated
2704 for (current = entry; current != &map->header && current->start < end;) {
2705 offset = current->offset + (start - current->start);
2706 size = (end <= current->end ? end : current->end) - start;
2707 if (current->eflags & MAP_ENTRY_IS_SUB_MAP) {
2709 vm_map_entry_t tentry;
2712 smap = current->object.sub_map;
2713 vm_map_lock_read(smap);
2714 (void) vm_map_lookup_entry(smap, offset, &tentry);
2715 tsize = tentry->end - offset;
2718 object = tentry->object.vm_object;
2719 offset = tentry->offset + (offset - tentry->start);
2720 vm_map_unlock_read(smap);
2722 object = current->object.vm_object;
2724 vm_object_reference(object);
2725 last_timestamp = map->timestamp;
2726 vm_map_unlock_read(map);
2727 if (!vm_object_sync(object, offset, size, syncio, invalidate))
2730 vm_object_deallocate(object);
2731 vm_map_lock_read(map);
2732 if (last_timestamp == map->timestamp ||
2733 !vm_map_lookup_entry(map, start, ¤t))
2734 current = current->next;
2737 vm_map_unlock_read(map);
2738 return (failed ? KERN_FAILURE : KERN_SUCCESS);
2742 * vm_map_entry_unwire: [ internal use only ]
2744 * Make the region specified by this entry pageable.
2746 * The map in question should be locked.
2747 * [This is the reason for this routine's existence.]
2750 vm_map_entry_unwire(vm_map_t map, vm_map_entry_t entry)
2752 vm_fault_unwire(map, entry->start, entry->end,
2753 entry->object.vm_object != NULL &&
2754 (entry->object.vm_object->flags & OBJ_FICTITIOUS) != 0);
2755 entry->wired_count = 0;
2759 vm_map_entry_deallocate(vm_map_entry_t entry, boolean_t system_map)
2762 if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) == 0)
2763 vm_object_deallocate(entry->object.vm_object);
2764 uma_zfree(system_map ? kmapentzone : mapentzone, entry);
2768 * vm_map_entry_delete: [ internal use only ]
2770 * Deallocate the given entry from the target map.
2773 vm_map_entry_delete(vm_map_t map, vm_map_entry_t entry)
2776 vm_pindex_t offidxstart, offidxend, count, size1;
2779 vm_map_entry_unlink(map, entry);
2780 object = entry->object.vm_object;
2781 size = entry->end - entry->start;
2784 if (entry->cred != NULL) {
2785 swap_release_by_cred(size, entry->cred);
2786 crfree(entry->cred);
2789 if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) == 0 &&
2791 KASSERT(entry->cred == NULL || object->cred == NULL ||
2792 (entry->eflags & MAP_ENTRY_NEEDS_COPY),
2793 ("OVERCOMMIT vm_map_entry_delete: both cred %p", entry));
2794 count = OFF_TO_IDX(size);
2795 offidxstart = OFF_TO_IDX(entry->offset);
2796 offidxend = offidxstart + count;
2797 VM_OBJECT_WLOCK(object);
2798 if (object->ref_count != 1 &&
2799 ((object->flags & (OBJ_NOSPLIT|OBJ_ONEMAPPING)) == OBJ_ONEMAPPING ||
2800 object == kernel_object || object == kmem_object)) {
2801 vm_object_collapse(object);
2804 * The option OBJPR_NOTMAPPED can be passed here
2805 * because vm_map_delete() already performed
2806 * pmap_remove() on the only mapping to this range
2809 vm_object_page_remove(object, offidxstart, offidxend,
2811 if (object->type == OBJT_SWAP)
2812 swap_pager_freespace(object, offidxstart, count);
2813 if (offidxend >= object->size &&
2814 offidxstart < object->size) {
2815 size1 = object->size;
2816 object->size = offidxstart;
2817 if (object->cred != NULL) {
2818 size1 -= object->size;
2819 KASSERT(object->charge >= ptoa(size1),
2820 ("vm_map_entry_delete: object->charge < 0"));
2821 swap_release_by_cred(ptoa(size1), object->cred);
2822 object->charge -= ptoa(size1);
2826 VM_OBJECT_WUNLOCK(object);
2828 entry->object.vm_object = NULL;
2829 if (map->system_map)
2830 vm_map_entry_deallocate(entry, TRUE);
2832 entry->next = curthread->td_map_def_user;
2833 curthread->td_map_def_user = entry;
2838 * vm_map_delete: [ internal use only ]
2840 * Deallocates the given address range from the target
2844 vm_map_delete(vm_map_t map, vm_offset_t start, vm_offset_t end)
2846 vm_map_entry_t entry;
2847 vm_map_entry_t first_entry;
2849 VM_MAP_ASSERT_LOCKED(map);
2851 return (KERN_SUCCESS);
2854 * Find the start of the region, and clip it
2856 if (!vm_map_lookup_entry(map, start, &first_entry))
2857 entry = first_entry->next;
2859 entry = first_entry;
2860 vm_map_clip_start(map, entry, start);
2864 * Step through all entries in this region
2866 while ((entry != &map->header) && (entry->start < end)) {
2867 vm_map_entry_t next;
2870 * Wait for wiring or unwiring of an entry to complete.
2871 * Also wait for any system wirings to disappear on
2874 if ((entry->eflags & MAP_ENTRY_IN_TRANSITION) != 0 ||
2875 (vm_map_pmap(map) != kernel_pmap &&
2876 vm_map_entry_system_wired_count(entry) != 0)) {
2877 unsigned int last_timestamp;
2878 vm_offset_t saved_start;
2879 vm_map_entry_t tmp_entry;
2881 saved_start = entry->start;
2882 entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP;
2883 last_timestamp = map->timestamp;
2884 (void) vm_map_unlock_and_wait(map, 0);
2886 if (last_timestamp + 1 != map->timestamp) {
2888 * Look again for the entry because the map was
2889 * modified while it was unlocked.
2890 * Specifically, the entry may have been
2891 * clipped, merged, or deleted.
2893 if (!vm_map_lookup_entry(map, saved_start,
2895 entry = tmp_entry->next;
2898 vm_map_clip_start(map, entry,
2904 vm_map_clip_end(map, entry, end);
2909 * Unwire before removing addresses from the pmap; otherwise,
2910 * unwiring will put the entries back in the pmap.
2912 if (entry->wired_count != 0) {
2913 vm_map_entry_unwire(map, entry);
2916 pmap_remove(map->pmap, entry->start, entry->end);
2919 * Delete the entry only after removing all pmap
2920 * entries pointing to its pages. (Otherwise, its
2921 * page frames may be reallocated, and any modify bits
2922 * will be set in the wrong object!)
2924 vm_map_entry_delete(map, entry);
2927 return (KERN_SUCCESS);
2933 * Remove the given address range from the target map.
2934 * This is the exported form of vm_map_delete.
2937 vm_map_remove(vm_map_t map, vm_offset_t start, vm_offset_t end)
2942 VM_MAP_RANGE_CHECK(map, start, end);
2943 result = vm_map_delete(map, start, end);
2949 * vm_map_check_protection:
2951 * Assert that the target map allows the specified privilege on the
2952 * entire address region given. The entire region must be allocated.
2954 * WARNING! This code does not and should not check whether the
2955 * contents of the region is accessible. For example a smaller file
2956 * might be mapped into a larger address space.
2958 * NOTE! This code is also called by munmap().
2960 * The map must be locked. A read lock is sufficient.
2963 vm_map_check_protection(vm_map_t map, vm_offset_t start, vm_offset_t end,
2964 vm_prot_t protection)
2966 vm_map_entry_t entry;
2967 vm_map_entry_t tmp_entry;
2969 if (!vm_map_lookup_entry(map, start, &tmp_entry))
2973 while (start < end) {
2974 if (entry == &map->header)
2979 if (start < entry->start)
2982 * Check protection associated with entry.
2984 if ((entry->protection & protection) != protection)
2986 /* go to next entry */
2988 entry = entry->next;
2994 * vm_map_copy_entry:
2996 * Copies the contents of the source entry to the destination
2997 * entry. The entries *must* be aligned properly.
3003 vm_map_entry_t src_entry,
3004 vm_map_entry_t dst_entry,
3005 vm_ooffset_t *fork_charge)
3007 vm_object_t src_object;
3008 vm_map_entry_t fake_entry;
3013 VM_MAP_ASSERT_LOCKED(dst_map);
3015 if ((dst_entry->eflags|src_entry->eflags) & MAP_ENTRY_IS_SUB_MAP)
3018 if (src_entry->wired_count == 0) {
3021 * If the source entry is marked needs_copy, it is already
3024 if ((src_entry->eflags & MAP_ENTRY_NEEDS_COPY) == 0) {
3025 pmap_protect(src_map->pmap,
3028 src_entry->protection & ~VM_PROT_WRITE);
3032 * Make a copy of the object.
3034 size = src_entry->end - src_entry->start;
3035 if ((src_object = src_entry->object.vm_object) != NULL) {
3036 VM_OBJECT_WLOCK(src_object);
3037 charged = ENTRY_CHARGED(src_entry);
3038 if ((src_object->handle == NULL) &&
3039 (src_object->type == OBJT_DEFAULT ||
3040 src_object->type == OBJT_SWAP)) {
3041 vm_object_collapse(src_object);
3042 if ((src_object->flags & (OBJ_NOSPLIT|OBJ_ONEMAPPING)) == OBJ_ONEMAPPING) {
3043 vm_object_split(src_entry);
3044 src_object = src_entry->object.vm_object;
3047 vm_object_reference_locked(src_object);
3048 vm_object_clear_flag(src_object, OBJ_ONEMAPPING);
3049 if (src_entry->cred != NULL &&
3050 !(src_entry->eflags & MAP_ENTRY_NEEDS_COPY)) {
3051 KASSERT(src_object->cred == NULL,
3052 ("OVERCOMMIT: vm_map_copy_entry: cred %p",
3054 src_object->cred = src_entry->cred;
3055 src_object->charge = size;
3057 VM_OBJECT_WUNLOCK(src_object);
3058 dst_entry->object.vm_object = src_object;
3060 cred = curthread->td_ucred;
3062 dst_entry->cred = cred;
3063 *fork_charge += size;
3064 if (!(src_entry->eflags &
3065 MAP_ENTRY_NEEDS_COPY)) {
3067 src_entry->cred = cred;
3068 *fork_charge += size;
3071 src_entry->eflags |= (MAP_ENTRY_COW|MAP_ENTRY_NEEDS_COPY);
3072 dst_entry->eflags |= (MAP_ENTRY_COW|MAP_ENTRY_NEEDS_COPY);
3073 dst_entry->offset = src_entry->offset;
3074 if (src_entry->eflags & MAP_ENTRY_VN_WRITECNT) {
3076 * MAP_ENTRY_VN_WRITECNT cannot
3077 * indicate write reference from
3078 * src_entry, since the entry is
3079 * marked as needs copy. Allocate a
3080 * fake entry that is used to
3081 * decrement object->un_pager.vnp.writecount
3082 * at the appropriate time. Attach
3083 * fake_entry to the deferred list.
3085 fake_entry = vm_map_entry_create(dst_map);
3086 fake_entry->eflags = MAP_ENTRY_VN_WRITECNT;
3087 src_entry->eflags &= ~MAP_ENTRY_VN_WRITECNT;
3088 vm_object_reference(src_object);
3089 fake_entry->object.vm_object = src_object;
3090 fake_entry->start = src_entry->start;
3091 fake_entry->end = src_entry->end;
3092 fake_entry->next = curthread->td_map_def_user;
3093 curthread->td_map_def_user = fake_entry;
3096 dst_entry->object.vm_object = NULL;
3097 dst_entry->offset = 0;
3098 if (src_entry->cred != NULL) {
3099 dst_entry->cred = curthread->td_ucred;
3100 crhold(dst_entry->cred);
3101 *fork_charge += size;
3105 pmap_copy(dst_map->pmap, src_map->pmap, dst_entry->start,
3106 dst_entry->end - dst_entry->start, src_entry->start);
3109 * Of course, wired down pages can't be set copy-on-write.
3110 * Cause wired pages to be copied into the new map by
3111 * simulating faults (the new pages are pageable)
3113 vm_fault_copy_entry(dst_map, src_map, dst_entry, src_entry,
3119 * vmspace_map_entry_forked:
3120 * Update the newly-forked vmspace each time a map entry is inherited
3121 * or copied. The values for vm_dsize and vm_tsize are approximate
3122 * (and mostly-obsolete ideas in the face of mmap(2) et al.)
3125 vmspace_map_entry_forked(const struct vmspace *vm1, struct vmspace *vm2,
3126 vm_map_entry_t entry)
3128 vm_size_t entrysize;
3131 entrysize = entry->end - entry->start;
3132 vm2->vm_map.size += entrysize;
3133 if (entry->eflags & (MAP_ENTRY_GROWS_DOWN | MAP_ENTRY_GROWS_UP)) {
3134 vm2->vm_ssize += btoc(entrysize);
3135 } else if (entry->start >= (vm_offset_t)vm1->vm_daddr &&
3136 entry->start < (vm_offset_t)vm1->vm_daddr + ctob(vm1->vm_dsize)) {
3137 newend = MIN(entry->end,
3138 (vm_offset_t)vm1->vm_daddr + ctob(vm1->vm_dsize));
3139 vm2->vm_dsize += btoc(newend - entry->start);
3140 } else if (entry->start >= (vm_offset_t)vm1->vm_taddr &&
3141 entry->start < (vm_offset_t)vm1->vm_taddr + ctob(vm1->vm_tsize)) {
3142 newend = MIN(entry->end,
3143 (vm_offset_t)vm1->vm_taddr + ctob(vm1->vm_tsize));
3144 vm2->vm_tsize += btoc(newend - entry->start);
3150 * Create a new process vmspace structure and vm_map
3151 * based on those of an existing process. The new map
3152 * is based on the old map, according to the inheritance
3153 * values on the regions in that map.
3155 * XXX It might be worth coalescing the entries added to the new vmspace.
3157 * The source map must not be locked.
3160 vmspace_fork(struct vmspace *vm1, vm_ooffset_t *fork_charge)
3162 struct vmspace *vm2;
3163 vm_map_t new_map, old_map;
3164 vm_map_entry_t new_entry, old_entry;
3168 old_map = &vm1->vm_map;
3169 /* Copy immutable fields of vm1 to vm2. */
3170 vm2 = vmspace_alloc(old_map->min_offset, old_map->max_offset, NULL);
3173 vm2->vm_taddr = vm1->vm_taddr;
3174 vm2->vm_daddr = vm1->vm_daddr;
3175 vm2->vm_maxsaddr = vm1->vm_maxsaddr;
3176 vm_map_lock(old_map);
3178 vm_map_wait_busy(old_map);
3179 new_map = &vm2->vm_map;
3180 locked = vm_map_trylock(new_map); /* trylock to silence WITNESS */
3181 KASSERT(locked, ("vmspace_fork: lock failed"));
3183 old_entry = old_map->header.next;
3185 while (old_entry != &old_map->header) {
3186 if (old_entry->eflags & MAP_ENTRY_IS_SUB_MAP)
3187 panic("vm_map_fork: encountered a submap");
3189 switch (old_entry->inheritance) {
3190 case VM_INHERIT_NONE:
3193 case VM_INHERIT_SHARE:
3195 * Clone the entry, creating the shared object if necessary.
3197 object = old_entry->object.vm_object;
3198 if (object == NULL) {
3199 object = vm_object_allocate(OBJT_DEFAULT,
3200 atop(old_entry->end - old_entry->start));
3201 old_entry->object.vm_object = object;
3202 old_entry->offset = 0;
3203 if (old_entry->cred != NULL) {
3204 object->cred = old_entry->cred;
3205 object->charge = old_entry->end -
3207 old_entry->cred = NULL;
3212 * Add the reference before calling vm_object_shadow
3213 * to insure that a shadow object is created.
3215 vm_object_reference(object);
3216 if (old_entry->eflags & MAP_ENTRY_NEEDS_COPY) {
3217 vm_object_shadow(&old_entry->object.vm_object,
3219 old_entry->end - old_entry->start);
3220 old_entry->eflags &= ~MAP_ENTRY_NEEDS_COPY;
3221 /* Transfer the second reference too. */
3222 vm_object_reference(
3223 old_entry->object.vm_object);
3226 * As in vm_map_simplify_entry(), the
3227 * vnode lock will not be acquired in
3228 * this call to vm_object_deallocate().
3230 vm_object_deallocate(object);
3231 object = old_entry->object.vm_object;
3233 VM_OBJECT_WLOCK(object);
3234 vm_object_clear_flag(object, OBJ_ONEMAPPING);
3235 if (old_entry->cred != NULL) {
3236 KASSERT(object->cred == NULL, ("vmspace_fork both cred"));
3237 object->cred = old_entry->cred;
3238 object->charge = old_entry->end - old_entry->start;
3239 old_entry->cred = NULL;
3243 * Assert the correct state of the vnode
3244 * v_writecount while the object is locked, to
3245 * not relock it later for the assertion
3248 if (old_entry->eflags & MAP_ENTRY_VN_WRITECNT &&
3249 object->type == OBJT_VNODE) {
3250 KASSERT(((struct vnode *)object->handle)->
3252 ("vmspace_fork: v_writecount %p", object));
3253 KASSERT(object->un_pager.vnp.writemappings > 0,
3254 ("vmspace_fork: vnp.writecount %p",
3257 VM_OBJECT_WUNLOCK(object);
3260 * Clone the entry, referencing the shared object.
3262 new_entry = vm_map_entry_create(new_map);
3263 *new_entry = *old_entry;
3264 new_entry->eflags &= ~(MAP_ENTRY_USER_WIRED |
3265 MAP_ENTRY_IN_TRANSITION);
3266 new_entry->wiring_thread = NULL;
3267 new_entry->wired_count = 0;
3268 if (new_entry->eflags & MAP_ENTRY_VN_WRITECNT) {
3269 vnode_pager_update_writecount(object,
3270 new_entry->start, new_entry->end);
3274 * Insert the entry into the new map -- we know we're
3275 * inserting at the end of the new map.
3277 vm_map_entry_link(new_map, new_map->header.prev,
3279 vmspace_map_entry_forked(vm1, vm2, new_entry);
3282 * Update the physical map
3284 pmap_copy(new_map->pmap, old_map->pmap,
3286 (old_entry->end - old_entry->start),
3290 case VM_INHERIT_COPY:
3292 * Clone the entry and link into the map.
3294 new_entry = vm_map_entry_create(new_map);
3295 *new_entry = *old_entry;
3297 * Copied entry is COW over the old object.
3299 new_entry->eflags &= ~(MAP_ENTRY_USER_WIRED |
3300 MAP_ENTRY_IN_TRANSITION | MAP_ENTRY_VN_WRITECNT);
3301 new_entry->wiring_thread = NULL;
3302 new_entry->wired_count = 0;
3303 new_entry->object.vm_object = NULL;
3304 new_entry->cred = NULL;
3305 vm_map_entry_link(new_map, new_map->header.prev,
3307 vmspace_map_entry_forked(vm1, vm2, new_entry);
3308 vm_map_copy_entry(old_map, new_map, old_entry,
3309 new_entry, fork_charge);
3312 old_entry = old_entry->next;
3315 * Use inlined vm_map_unlock() to postpone handling the deferred
3316 * map entries, which cannot be done until both old_map and
3317 * new_map locks are released.
3319 sx_xunlock(&old_map->lock);
3320 sx_xunlock(&new_map->lock);
3321 vm_map_process_deferred();
3327 vm_map_stack(vm_map_t map, vm_offset_t addrbos, vm_size_t max_ssize,
3328 vm_prot_t prot, vm_prot_t max, int cow)
3330 vm_map_entry_t new_entry, prev_entry;
3331 vm_offset_t bot, top;
3332 vm_size_t growsize, init_ssize;
3334 rlim_t lmemlim, vmemlim;
3337 * The stack orientation is piggybacked with the cow argument.
3338 * Extract it into orient and mask the cow argument so that we
3339 * don't pass it around further.
3340 * NOTE: We explicitly allow bi-directional stacks.
3342 orient = cow & (MAP_STACK_GROWS_DOWN|MAP_STACK_GROWS_UP);
3343 KASSERT(orient != 0, ("No stack grow direction"));
3345 if (addrbos < vm_map_min(map) ||
3346 addrbos > vm_map_max(map) ||
3347 addrbos + max_ssize < addrbos)
3348 return (KERN_NO_SPACE);
3350 growsize = sgrowsiz;
3351 init_ssize = (max_ssize < growsize) ? max_ssize : growsize;
3354 lmemlim = lim_cur(curproc, RLIMIT_MEMLOCK);
3355 vmemlim = lim_cur(curproc, RLIMIT_VMEM);
3356 PROC_UNLOCK(curproc);
3360 /* If addr is already mapped, no go */
3361 if (vm_map_lookup_entry(map, addrbos, &prev_entry)) {
3363 return (KERN_NO_SPACE);
3366 if (!old_mlock && map->flags & MAP_WIREFUTURE) {
3367 if (ptoa(pmap_wired_count(map->pmap)) + init_ssize > lmemlim) {
3369 return (KERN_NO_SPACE);
3373 /* If we would blow our VMEM resource limit, no go */
3374 if (map->size + init_ssize > vmemlim) {
3376 return (KERN_NO_SPACE);
3380 * If we can't accomodate max_ssize in the current mapping, no go.
3381 * However, we need to be aware that subsequent user mappings might
3382 * map into the space we have reserved for stack, and currently this
3383 * space is not protected.
3385 * Hopefully we will at least detect this condition when we try to
3388 if ((prev_entry->next != &map->header) &&
3389 (prev_entry->next->start < addrbos + max_ssize)) {
3391 return (KERN_NO_SPACE);
3395 * We initially map a stack of only init_ssize. We will grow as
3396 * needed later. Depending on the orientation of the stack (i.e.
3397 * the grow direction) we either map at the top of the range, the
3398 * bottom of the range or in the middle.
3400 * Note: we would normally expect prot and max to be VM_PROT_ALL,
3401 * and cow to be 0. Possibly we should eliminate these as input
3402 * parameters, and just pass these values here in the insert call.
3404 if (orient == MAP_STACK_GROWS_DOWN)
3405 bot = addrbos + max_ssize - init_ssize;
3406 else if (orient == MAP_STACK_GROWS_UP)
3409 bot = round_page(addrbos + max_ssize/2 - init_ssize/2);
3410 top = bot + init_ssize;
3411 rv = vm_map_insert(map, NULL, 0, bot, top, prot, max, cow);
3413 /* Now set the avail_ssize amount. */
3414 if (rv == KERN_SUCCESS) {
3415 if (prev_entry != &map->header)
3416 vm_map_clip_end(map, prev_entry, bot);
3417 new_entry = prev_entry->next;
3418 if (new_entry->end != top || new_entry->start != bot)
3419 panic("Bad entry start/end for new stack entry");
3421 new_entry->avail_ssize = max_ssize - init_ssize;
3422 if (orient & MAP_STACK_GROWS_DOWN)
3423 new_entry->eflags |= MAP_ENTRY_GROWS_DOWN;
3424 if (orient & MAP_STACK_GROWS_UP)
3425 new_entry->eflags |= MAP_ENTRY_GROWS_UP;
3432 static int stack_guard_page = 0;
3433 TUNABLE_INT("security.bsd.stack_guard_page", &stack_guard_page);
3434 SYSCTL_INT(_security_bsd, OID_AUTO, stack_guard_page, CTLFLAG_RW,
3435 &stack_guard_page, 0,
3436 "Insert stack guard page ahead of the growable segments.");
3438 /* Attempts to grow a vm stack entry. Returns KERN_SUCCESS if the
3439 * desired address is already mapped, or if we successfully grow
3440 * the stack. Also returns KERN_SUCCESS if addr is outside the
3441 * stack range (this is strange, but preserves compatibility with
3442 * the grow function in vm_machdep.c).
3445 vm_map_growstack(struct proc *p, vm_offset_t addr)
3447 vm_map_entry_t next_entry, prev_entry;
3448 vm_map_entry_t new_entry, stack_entry;
3449 struct vmspace *vm = p->p_vmspace;
3450 vm_map_t map = &vm->vm_map;
3453 size_t grow_amount, max_grow;
3454 rlim_t lmemlim, stacklim, vmemlim;
3455 int is_procstack, rv;
3466 lmemlim = lim_cur(p, RLIMIT_MEMLOCK);
3467 stacklim = lim_cur(p, RLIMIT_STACK);
3468 vmemlim = lim_cur(p, RLIMIT_VMEM);
3471 vm_map_lock_read(map);
3473 /* If addr is already in the entry range, no need to grow.*/
3474 if (vm_map_lookup_entry(map, addr, &prev_entry)) {
3475 vm_map_unlock_read(map);
3476 return (KERN_SUCCESS);
3479 next_entry = prev_entry->next;
3480 if (!(prev_entry->eflags & MAP_ENTRY_GROWS_UP)) {
3482 * This entry does not grow upwards. Since the address lies
3483 * beyond this entry, the next entry (if one exists) has to
3484 * be a downward growable entry. The entry list header is
3485 * never a growable entry, so it suffices to check the flags.
3487 if (!(next_entry->eflags & MAP_ENTRY_GROWS_DOWN)) {
3488 vm_map_unlock_read(map);
3489 return (KERN_SUCCESS);
3491 stack_entry = next_entry;
3494 * This entry grows upward. If the next entry does not at
3495 * least grow downwards, this is the entry we need to grow.
3496 * otherwise we have two possible choices and we have to
3499 if (next_entry->eflags & MAP_ENTRY_GROWS_DOWN) {
3501 * We have two choices; grow the entry closest to
3502 * the address to minimize the amount of growth.
3504 if (addr - prev_entry->end <= next_entry->start - addr)
3505 stack_entry = prev_entry;
3507 stack_entry = next_entry;
3509 stack_entry = prev_entry;
3512 if (stack_entry == next_entry) {
3513 KASSERT(stack_entry->eflags & MAP_ENTRY_GROWS_DOWN, ("foo"));
3514 KASSERT(addr < stack_entry->start, ("foo"));
3515 end = (prev_entry != &map->header) ? prev_entry->end :
3516 stack_entry->start - stack_entry->avail_ssize;
3517 grow_amount = roundup(stack_entry->start - addr, PAGE_SIZE);
3518 max_grow = stack_entry->start - end;
3520 KASSERT(stack_entry->eflags & MAP_ENTRY_GROWS_UP, ("foo"));
3521 KASSERT(addr >= stack_entry->end, ("foo"));
3522 end = (next_entry != &map->header) ? next_entry->start :
3523 stack_entry->end + stack_entry->avail_ssize;
3524 grow_amount = roundup(addr + 1 - stack_entry->end, PAGE_SIZE);
3525 max_grow = end - stack_entry->end;
3528 if (grow_amount > stack_entry->avail_ssize) {
3529 vm_map_unlock_read(map);
3530 return (KERN_NO_SPACE);
3534 * If there is no longer enough space between the entries nogo, and
3535 * adjust the available space. Note: this should only happen if the
3536 * user has mapped into the stack area after the stack was created,
3537 * and is probably an error.
3539 * This also effectively destroys any guard page the user might have
3540 * intended by limiting the stack size.
3542 if (grow_amount + (stack_guard_page ? PAGE_SIZE : 0) > max_grow) {
3543 if (vm_map_lock_upgrade(map))
3546 stack_entry->avail_ssize = max_grow;
3549 return (KERN_NO_SPACE);
3552 is_procstack = (addr >= (vm_offset_t)vm->vm_maxsaddr) ? 1 : 0;
3555 * If this is the main process stack, see if we're over the stack
3558 if (is_procstack && (ctob(vm->vm_ssize) + grow_amount > stacklim)) {
3559 vm_map_unlock_read(map);
3560 return (KERN_NO_SPACE);
3565 racct_set(p, RACCT_STACK, ctob(vm->vm_ssize) + grow_amount)) {
3567 vm_map_unlock_read(map);
3568 return (KERN_NO_SPACE);
3573 /* Round up the grow amount modulo sgrowsiz */
3574 growsize = sgrowsiz;
3575 grow_amount = roundup(grow_amount, growsize);
3576 if (grow_amount > stack_entry->avail_ssize)
3577 grow_amount = stack_entry->avail_ssize;
3578 if (is_procstack && (ctob(vm->vm_ssize) + grow_amount > stacklim)) {
3579 grow_amount = trunc_page((vm_size_t)stacklim) -
3584 limit = racct_get_available(p, RACCT_STACK);
3586 if (is_procstack && (ctob(vm->vm_ssize) + grow_amount > limit))
3587 grow_amount = limit - ctob(vm->vm_ssize);
3589 if (!old_mlock && map->flags & MAP_WIREFUTURE) {
3590 if (ptoa(pmap_wired_count(map->pmap)) + grow_amount > lmemlim) {
3591 vm_map_unlock_read(map);
3597 if (racct_set(p, RACCT_MEMLOCK,
3598 ptoa(pmap_wired_count(map->pmap)) + grow_amount)) {
3600 vm_map_unlock_read(map);
3607 /* If we would blow our VMEM resource limit, no go */
3608 if (map->size + grow_amount > vmemlim) {
3609 vm_map_unlock_read(map);
3615 if (racct_set(p, RACCT_VMEM, map->size + grow_amount)) {
3617 vm_map_unlock_read(map);
3624 if (vm_map_lock_upgrade(map))
3627 if (stack_entry == next_entry) {
3631 /* Get the preliminary new entry start value */
3632 addr = stack_entry->start - grow_amount;
3635 * If this puts us into the previous entry, cut back our
3636 * growth to the available space. Also, see the note above.
3639 stack_entry->avail_ssize = max_grow;
3641 if (stack_guard_page)
3645 rv = vm_map_insert(map, NULL, 0, addr, stack_entry->start,
3646 next_entry->protection, next_entry->max_protection, 0);
3648 /* Adjust the available stack space by the amount we grew. */
3649 if (rv == KERN_SUCCESS) {
3650 if (prev_entry != &map->header)
3651 vm_map_clip_end(map, prev_entry, addr);
3652 new_entry = prev_entry->next;
3653 KASSERT(new_entry == stack_entry->prev, ("foo"));
3654 KASSERT(new_entry->end == stack_entry->start, ("foo"));
3655 KASSERT(new_entry->start == addr, ("foo"));
3656 grow_amount = new_entry->end - new_entry->start;
3657 new_entry->avail_ssize = stack_entry->avail_ssize -
3659 stack_entry->eflags &= ~MAP_ENTRY_GROWS_DOWN;
3660 new_entry->eflags |= MAP_ENTRY_GROWS_DOWN;
3666 addr = stack_entry->end + grow_amount;
3669 * If this puts us into the next entry, cut back our growth
3670 * to the available space. Also, see the note above.
3673 stack_entry->avail_ssize = end - stack_entry->end;
3675 if (stack_guard_page)
3679 grow_amount = addr - stack_entry->end;
3680 cred = stack_entry->cred;
3681 if (cred == NULL && stack_entry->object.vm_object != NULL)
3682 cred = stack_entry->object.vm_object->cred;
3683 if (cred != NULL && !swap_reserve_by_cred(grow_amount, cred))
3685 /* Grow the underlying object if applicable. */
3686 else if (stack_entry->object.vm_object == NULL ||
3687 vm_object_coalesce(stack_entry->object.vm_object,
3688 stack_entry->offset,
3689 (vm_size_t)(stack_entry->end - stack_entry->start),
3690 (vm_size_t)grow_amount, cred != NULL)) {
3691 map->size += (addr - stack_entry->end);
3692 /* Update the current entry. */
3693 stack_entry->end = addr;
3694 stack_entry->avail_ssize -= grow_amount;
3695 vm_map_entry_resize_free(map, stack_entry);
3698 if (next_entry != &map->header)
3699 vm_map_clip_start(map, next_entry, addr);
3704 if (rv == KERN_SUCCESS && is_procstack)
3705 vm->vm_ssize += btoc(grow_amount);
3710 * Heed the MAP_WIREFUTURE flag if it was set for this process.
3712 if (rv == KERN_SUCCESS && (map->flags & MAP_WIREFUTURE)) {
3714 (stack_entry == next_entry) ? addr : addr - grow_amount,
3715 (stack_entry == next_entry) ? stack_entry->start : addr,
3716 (p->p_flag & P_SYSTEM)
3717 ? VM_MAP_WIRE_SYSTEM|VM_MAP_WIRE_NOHOLES
3718 : VM_MAP_WIRE_USER|VM_MAP_WIRE_NOHOLES);
3723 if (rv != KERN_SUCCESS) {
3725 error = racct_set(p, RACCT_VMEM, map->size);
3726 KASSERT(error == 0, ("decreasing RACCT_VMEM failed"));
3728 error = racct_set(p, RACCT_MEMLOCK,
3729 ptoa(pmap_wired_count(map->pmap)));
3730 KASSERT(error == 0, ("decreasing RACCT_MEMLOCK failed"));
3732 error = racct_set(p, RACCT_STACK, ctob(vm->vm_ssize));
3733 KASSERT(error == 0, ("decreasing RACCT_STACK failed"));
3742 * Unshare the specified VM space for exec. If other processes are
3743 * mapped to it, then create a new one. The new vmspace is null.
3746 vmspace_exec(struct proc *p, vm_offset_t minuser, vm_offset_t maxuser)
3748 struct vmspace *oldvmspace = p->p_vmspace;
3749 struct vmspace *newvmspace;
3751 newvmspace = vmspace_alloc(minuser, maxuser, NULL);
3752 if (newvmspace == NULL)
3754 newvmspace->vm_swrss = oldvmspace->vm_swrss;
3756 * This code is written like this for prototype purposes. The
3757 * goal is to avoid running down the vmspace here, but let the
3758 * other process's that are still using the vmspace to finally
3759 * run it down. Even though there is little or no chance of blocking
3760 * here, it is a good idea to keep this form for future mods.
3762 PROC_VMSPACE_LOCK(p);
3763 p->p_vmspace = newvmspace;
3764 PROC_VMSPACE_UNLOCK(p);
3765 if (p == curthread->td_proc)
3766 pmap_activate(curthread);
3767 vmspace_free(oldvmspace);
3772 * Unshare the specified VM space for forcing COW. This
3773 * is called by rfork, for the (RFMEM|RFPROC) == 0 case.
3776 vmspace_unshare(struct proc *p)
3778 struct vmspace *oldvmspace = p->p_vmspace;
3779 struct vmspace *newvmspace;
3780 vm_ooffset_t fork_charge;
3782 if (oldvmspace->vm_refcnt == 1)
3785 newvmspace = vmspace_fork(oldvmspace, &fork_charge);
3786 if (newvmspace == NULL)
3788 if (!swap_reserve_by_cred(fork_charge, p->p_ucred)) {
3789 vmspace_free(newvmspace);
3792 PROC_VMSPACE_LOCK(p);
3793 p->p_vmspace = newvmspace;
3794 PROC_VMSPACE_UNLOCK(p);
3795 if (p == curthread->td_proc)
3796 pmap_activate(curthread);
3797 vmspace_free(oldvmspace);
3804 * Finds the VM object, offset, and
3805 * protection for a given virtual address in the
3806 * specified map, assuming a page fault of the
3809 * Leaves the map in question locked for read; return
3810 * values are guaranteed until a vm_map_lookup_done
3811 * call is performed. Note that the map argument
3812 * is in/out; the returned map must be used in
3813 * the call to vm_map_lookup_done.
3815 * A handle (out_entry) is returned for use in
3816 * vm_map_lookup_done, to make that fast.
3818 * If a lookup is requested with "write protection"
3819 * specified, the map may be changed to perform virtual
3820 * copying operations, although the data referenced will
3824 vm_map_lookup(vm_map_t *var_map, /* IN/OUT */
3826 vm_prot_t fault_typea,
3827 vm_map_entry_t *out_entry, /* OUT */
3828 vm_object_t *object, /* OUT */
3829 vm_pindex_t *pindex, /* OUT */
3830 vm_prot_t *out_prot, /* OUT */
3831 boolean_t *wired) /* OUT */
3833 vm_map_entry_t entry;
3834 vm_map_t map = *var_map;
3836 vm_prot_t fault_type = fault_typea;
3837 vm_object_t eobject;
3843 vm_map_lock_read(map);
3846 * Lookup the faulting address.
3848 if (!vm_map_lookup_entry(map, vaddr, out_entry)) {
3849 vm_map_unlock_read(map);
3850 return (KERN_INVALID_ADDRESS);
3858 if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) {
3859 vm_map_t old_map = map;
3861 *var_map = map = entry->object.sub_map;
3862 vm_map_unlock_read(old_map);
3867 * Check whether this task is allowed to have this page.
3869 prot = entry->protection;
3870 fault_type &= (VM_PROT_READ|VM_PROT_WRITE|VM_PROT_EXECUTE);
3871 if ((fault_type & prot) != fault_type || prot == VM_PROT_NONE) {
3872 vm_map_unlock_read(map);
3873 return (KERN_PROTECTION_FAILURE);
3875 if ((entry->eflags & MAP_ENTRY_USER_WIRED) &&
3876 (entry->eflags & MAP_ENTRY_COW) &&
3877 (fault_type & VM_PROT_WRITE)) {
3878 vm_map_unlock_read(map);
3879 return (KERN_PROTECTION_FAILURE);
3881 if ((fault_typea & VM_PROT_COPY) != 0 &&
3882 (entry->max_protection & VM_PROT_WRITE) == 0 &&
3883 (entry->eflags & MAP_ENTRY_COW) == 0) {
3884 vm_map_unlock_read(map);
3885 return (KERN_PROTECTION_FAILURE);
3889 * If this page is not pageable, we have to get it for all possible
3892 *wired = (entry->wired_count != 0);
3894 fault_type = entry->protection;
3895 size = entry->end - entry->start;
3897 * If the entry was copy-on-write, we either ...
3899 if (entry->eflags & MAP_ENTRY_NEEDS_COPY) {
3901 * If we want to write the page, we may as well handle that
3902 * now since we've got the map locked.
3904 * If we don't need to write the page, we just demote the
3905 * permissions allowed.
3907 if ((fault_type & VM_PROT_WRITE) != 0 ||
3908 (fault_typea & VM_PROT_COPY) != 0) {
3910 * Make a new object, and place it in the object
3911 * chain. Note that no new references have appeared
3912 * -- one just moved from the map to the new
3915 if (vm_map_lock_upgrade(map))
3918 if (entry->cred == NULL) {
3920 * The debugger owner is charged for
3923 cred = curthread->td_ucred;
3925 if (!swap_reserve_by_cred(size, cred)) {
3928 return (KERN_RESOURCE_SHORTAGE);
3932 vm_object_shadow(&entry->object.vm_object,
3933 &entry->offset, size);
3934 entry->eflags &= ~MAP_ENTRY_NEEDS_COPY;
3935 eobject = entry->object.vm_object;
3936 if (eobject->cred != NULL) {
3938 * The object was not shadowed.
3940 swap_release_by_cred(size, entry->cred);
3941 crfree(entry->cred);
3943 } else if (entry->cred != NULL) {
3944 VM_OBJECT_WLOCK(eobject);
3945 eobject->cred = entry->cred;
3946 eobject->charge = size;
3947 VM_OBJECT_WUNLOCK(eobject);
3951 vm_map_lock_downgrade(map);
3954 * We're attempting to read a copy-on-write page --
3955 * don't allow writes.
3957 prot &= ~VM_PROT_WRITE;
3962 * Create an object if necessary.
3964 if (entry->object.vm_object == NULL &&
3966 if (vm_map_lock_upgrade(map))
3968 entry->object.vm_object = vm_object_allocate(OBJT_DEFAULT,
3971 if (entry->cred != NULL) {
3972 VM_OBJECT_WLOCK(entry->object.vm_object);
3973 entry->object.vm_object->cred = entry->cred;
3974 entry->object.vm_object->charge = size;
3975 VM_OBJECT_WUNLOCK(entry->object.vm_object);
3978 vm_map_lock_downgrade(map);
3982 * Return the object/offset from this entry. If the entry was
3983 * copy-on-write or empty, it has been fixed up.
3985 *pindex = OFF_TO_IDX((vaddr - entry->start) + entry->offset);
3986 *object = entry->object.vm_object;
3989 return (KERN_SUCCESS);
3993 * vm_map_lookup_locked:
3995 * Lookup the faulting address. A version of vm_map_lookup that returns
3996 * KERN_FAILURE instead of blocking on map lock or memory allocation.
3999 vm_map_lookup_locked(vm_map_t *var_map, /* IN/OUT */
4001 vm_prot_t fault_typea,
4002 vm_map_entry_t *out_entry, /* OUT */
4003 vm_object_t *object, /* OUT */
4004 vm_pindex_t *pindex, /* OUT */
4005 vm_prot_t *out_prot, /* OUT */
4006 boolean_t *wired) /* OUT */
4008 vm_map_entry_t entry;
4009 vm_map_t map = *var_map;
4011 vm_prot_t fault_type = fault_typea;
4014 * Lookup the faulting address.
4016 if (!vm_map_lookup_entry(map, vaddr, out_entry))
4017 return (KERN_INVALID_ADDRESS);
4022 * Fail if the entry refers to a submap.
4024 if (entry->eflags & MAP_ENTRY_IS_SUB_MAP)
4025 return (KERN_FAILURE);
4028 * Check whether this task is allowed to have this page.
4030 prot = entry->protection;
4031 fault_type &= VM_PROT_READ | VM_PROT_WRITE | VM_PROT_EXECUTE;
4032 if ((fault_type & prot) != fault_type)
4033 return (KERN_PROTECTION_FAILURE);
4034 if ((entry->eflags & MAP_ENTRY_USER_WIRED) &&
4035 (entry->eflags & MAP_ENTRY_COW) &&
4036 (fault_type & VM_PROT_WRITE))
4037 return (KERN_PROTECTION_FAILURE);
4040 * If this page is not pageable, we have to get it for all possible
4043 *wired = (entry->wired_count != 0);
4045 fault_type = entry->protection;
4047 if (entry->eflags & MAP_ENTRY_NEEDS_COPY) {
4049 * Fail if the entry was copy-on-write for a write fault.
4051 if (fault_type & VM_PROT_WRITE)
4052 return (KERN_FAILURE);
4054 * We're attempting to read a copy-on-write page --
4055 * don't allow writes.
4057 prot &= ~VM_PROT_WRITE;
4061 * Fail if an object should be created.
4063 if (entry->object.vm_object == NULL && !map->system_map)
4064 return (KERN_FAILURE);
4067 * Return the object/offset from this entry. If the entry was
4068 * copy-on-write or empty, it has been fixed up.
4070 *pindex = OFF_TO_IDX((vaddr - entry->start) + entry->offset);
4071 *object = entry->object.vm_object;
4074 return (KERN_SUCCESS);
4078 * vm_map_lookup_done:
4080 * Releases locks acquired by a vm_map_lookup
4081 * (according to the handle returned by that lookup).
4084 vm_map_lookup_done(vm_map_t map, vm_map_entry_t entry)
4087 * Unlock the main-level map
4089 vm_map_unlock_read(map);
4092 #include "opt_ddb.h"
4094 #include <sys/kernel.h>
4096 #include <ddb/ddb.h>
4099 vm_map_print(vm_map_t map)
4101 vm_map_entry_t entry;
4103 db_iprintf("Task map %p: pmap=%p, nentries=%d, version=%u\n",
4105 (void *)map->pmap, map->nentries, map->timestamp);
4108 for (entry = map->header.next; entry != &map->header;
4109 entry = entry->next) {
4110 db_iprintf("map entry %p: start=%p, end=%p\n",
4111 (void *)entry, (void *)entry->start, (void *)entry->end);
4113 static char *inheritance_name[4] =
4114 {"share", "copy", "none", "donate_copy"};
4116 db_iprintf(" prot=%x/%x/%s",
4118 entry->max_protection,
4119 inheritance_name[(int)(unsigned char)entry->inheritance]);
4120 if (entry->wired_count != 0)
4121 db_printf(", wired");
4123 if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) {
4124 db_printf(", share=%p, offset=0x%jx\n",
4125 (void *)entry->object.sub_map,
4126 (uintmax_t)entry->offset);
4127 if ((entry->prev == &map->header) ||
4128 (entry->prev->object.sub_map !=
4129 entry->object.sub_map)) {
4131 vm_map_print((vm_map_t)entry->object.sub_map);
4135 if (entry->cred != NULL)
4136 db_printf(", ruid %d", entry->cred->cr_ruid);
4137 db_printf(", object=%p, offset=0x%jx",
4138 (void *)entry->object.vm_object,
4139 (uintmax_t)entry->offset);
4140 if (entry->object.vm_object && entry->object.vm_object->cred)
4141 db_printf(", obj ruid %d charge %jx",
4142 entry->object.vm_object->cred->cr_ruid,
4143 (uintmax_t)entry->object.vm_object->charge);
4144 if (entry->eflags & MAP_ENTRY_COW)
4145 db_printf(", copy (%s)",
4146 (entry->eflags & MAP_ENTRY_NEEDS_COPY) ? "needed" : "done");
4149 if ((entry->prev == &map->header) ||
4150 (entry->prev->object.vm_object !=
4151 entry->object.vm_object)) {
4153 vm_object_print((db_expr_t)(intptr_t)
4154 entry->object.vm_object,
4163 DB_SHOW_COMMAND(map, map)
4167 db_printf("usage: show map <addr>\n");
4170 vm_map_print((vm_map_t)addr);
4173 DB_SHOW_COMMAND(procvm, procvm)
4178 p = (struct proc *) addr;
4183 db_printf("p = %p, vmspace = %p, map = %p, pmap = %p\n",
4184 (void *)p, (void *)p->p_vmspace, (void *)&p->p_vmspace->vm_map,
4185 (void *)vmspace_pmap(p->p_vmspace));
4187 vm_map_print((vm_map_t)&p->p_vmspace->vm_map);