2 * SPDX-License-Identifier: (BSD-3-Clause AND MIT-CMU)
4 * Copyright (c) 1991, 1993
5 * The Regents of the University of California. All rights reserved.
7 * This code is derived from software contributed to Berkeley by
8 * The Mach Operating System project at Carnegie-Mellon University.
10 * Redistribution and use in source and binary forms, with or without
11 * modification, are permitted provided that the following conditions
13 * 1. Redistributions of source code must retain the above copyright
14 * notice, this list of conditions and the following disclaimer.
15 * 2. Redistributions in binary form must reproduce the above copyright
16 * notice, this list of conditions and the following disclaimer in the
17 * documentation and/or other materials provided with the distribution.
18 * 3. Neither the name of the University nor the names of its contributors
19 * may be used to endorse or promote products derived from this software
20 * without specific prior written permission.
22 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
23 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
24 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
25 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
26 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
27 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
28 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
29 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
30 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
31 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
34 * from: @(#)vm_map.c 8.3 (Berkeley) 1/12/94
37 * Copyright (c) 1987, 1990 Carnegie-Mellon University.
38 * All rights reserved.
40 * Authors: Avadis Tevanian, Jr., Michael Wayne Young
42 * Permission to use, copy, modify and distribute this software and
43 * its documentation is hereby granted, provided that both the copyright
44 * notice and this permission notice appear in all copies of the
45 * software, derivative works or modified versions, and any portions
46 * thereof, and that both notices appear in supporting documentation.
48 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
49 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
50 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
52 * Carnegie Mellon requests users of this software to return to
54 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU
55 * School of Computer Science
56 * Carnegie Mellon University
57 * Pittsburgh PA 15213-3890
59 * any improvements or extensions that they make and grant Carnegie the
60 * rights to redistribute these changes.
64 * Virtual memory mapping module.
67 #include <sys/cdefs.h>
68 __FBSDID("$FreeBSD$");
70 #include <sys/param.h>
71 #include <sys/systm.h>
73 #include <sys/kernel.h>
76 #include <sys/mutex.h>
78 #include <sys/vmmeter.h>
80 #include <sys/vnode.h>
81 #include <sys/racct.h>
82 #include <sys/resourcevar.h>
83 #include <sys/rwlock.h>
85 #include <sys/sysctl.h>
86 #include <sys/sysent.h>
90 #include <vm/vm_param.h>
92 #include <vm/vm_map.h>
93 #include <vm/vm_page.h>
94 #include <vm/vm_pageout.h>
95 #include <vm/vm_object.h>
96 #include <vm/vm_pager.h>
97 #include <vm/vm_kern.h>
98 #include <vm/vm_extern.h>
99 #include <vm/vnode_pager.h>
100 #include <vm/swap_pager.h>
104 * Virtual memory maps provide for the mapping, protection,
105 * and sharing of virtual memory objects. In addition,
106 * this module provides for an efficient virtual copy of
107 * memory from one map to another.
109 * Synchronization is required prior to most operations.
111 * Maps consist of an ordered doubly-linked list of simple
112 * entries; a self-adjusting binary search tree of these
113 * entries is used to speed up lookups.
115 * Since portions of maps are specified by start/end addresses,
116 * which may not align with existing map entries, all
117 * routines merely "clip" entries to these start/end values.
118 * [That is, an entry is split into two, bordering at a
119 * start or end value.] Note that these clippings may not
120 * always be necessary (as the two resulting entries are then
121 * not changed); however, the clipping is done for convenience.
123 * As mentioned above, virtual copy operations are performed
124 * by copying VM object references from one map to
125 * another, and then marking both regions as copy-on-write.
128 static struct mtx map_sleep_mtx;
129 static uma_zone_t mapentzone;
130 static uma_zone_t kmapentzone;
131 static uma_zone_t vmspace_zone;
132 static int vmspace_zinit(void *mem, int size, int flags);
133 static void _vm_map_init(vm_map_t map, pmap_t pmap, vm_offset_t min,
135 static void vm_map_entry_deallocate(vm_map_entry_t entry, boolean_t system_map);
136 static void vm_map_entry_dispose(vm_map_t map, vm_map_entry_t entry);
137 static void vm_map_entry_unwire(vm_map_t map, vm_map_entry_t entry);
138 static int vm_map_growstack(vm_map_t map, vm_offset_t addr,
139 vm_map_entry_t gap_entry);
140 static void vm_map_pmap_enter(vm_map_t map, vm_offset_t addr, vm_prot_t prot,
141 vm_object_t object, vm_pindex_t pindex, vm_size_t size, int flags);
143 static void vmspace_zdtor(void *mem, int size, void *arg);
145 static int vm_map_stack_locked(vm_map_t map, vm_offset_t addrbos,
146 vm_size_t max_ssize, vm_size_t growsize, vm_prot_t prot, vm_prot_t max,
148 static void vm_map_wire_entry_failure(vm_map_t map, vm_map_entry_t entry,
149 vm_offset_t failed_addr);
151 #define ENTRY_CHARGED(e) ((e)->cred != NULL || \
152 ((e)->object.vm_object != NULL && (e)->object.vm_object->cred != NULL && \
153 !((e)->eflags & MAP_ENTRY_NEEDS_COPY)))
156 * PROC_VMSPACE_{UN,}LOCK() can be a noop as long as vmspaces are type
159 #define PROC_VMSPACE_LOCK(p) do { } while (0)
160 #define PROC_VMSPACE_UNLOCK(p) do { } while (0)
163 * VM_MAP_RANGE_CHECK: [ internal use only ]
165 * Asserts that the starting and ending region
166 * addresses fall within the valid range of the map.
168 #define VM_MAP_RANGE_CHECK(map, start, end) \
170 if (start < vm_map_min(map)) \
171 start = vm_map_min(map); \
172 if (end > vm_map_max(map)) \
173 end = vm_map_max(map); \
178 #ifndef UMA_MD_SMALL_ALLOC
181 * Allocate a new slab for kernel map entries. The kernel map may be locked or
182 * unlocked, depending on whether the request is coming from the kernel map or a
183 * submap. This function allocates a virtual address range directly from the
184 * kernel map instead of the kmem_* layer to avoid recursion on the kernel map
185 * lock and also to avoid triggering allocator recursion in the vmem boundary
189 kmapent_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
195 *pflag = UMA_SLAB_PRIV;
197 if (!(locked = vm_map_locked(kernel_map)))
198 vm_map_lock(kernel_map);
199 addr = vm_map_findspace(kernel_map, vm_map_min(kernel_map), bytes);
200 if (addr + bytes < addr || addr + bytes > vm_map_max(kernel_map))
201 panic("%s: kernel map is exhausted", __func__);
202 error = vm_map_insert(kernel_map, NULL, 0, addr, addr + bytes,
203 VM_PROT_RW, VM_PROT_RW, MAP_NOFAULT);
204 if (error != KERN_SUCCESS)
205 panic("%s: vm_map_insert() failed: %d", __func__, error);
207 vm_map_unlock(kernel_map);
208 error = kmem_back_domain(domain, kernel_object, addr, bytes, M_NOWAIT |
209 M_USE_RESERVE | (wait & M_ZERO));
210 if (error == KERN_SUCCESS) {
211 return ((void *)addr);
214 vm_map_lock(kernel_map);
215 vm_map_delete(kernel_map, addr, bytes);
217 vm_map_unlock(kernel_map);
223 kmapent_free(void *item, vm_size_t size, uint8_t pflag)
228 if ((pflag & UMA_SLAB_PRIV) == 0)
232 addr = (vm_offset_t)item;
233 kmem_unback(kernel_object, addr, size);
234 error = vm_map_remove(kernel_map, addr, addr + size);
235 KASSERT(error == KERN_SUCCESS,
236 ("%s: vm_map_remove failed: %d", __func__, error));
240 * The worst-case upper bound on the number of kernel map entries that may be
241 * created before the zone must be replenished in _vm_map_unlock().
243 #define KMAPENT_RESERVE 1
245 #endif /* !UMD_MD_SMALL_ALLOC */
250 * Initialize the vm_map module. Must be called before any other vm_map
253 * User map and entry structures are allocated from the general purpose
254 * memory pool. Kernel maps are statically defined. Kernel map entries
255 * require special handling to avoid recursion; see the comments above
256 * kmapent_alloc() and in vm_map_entry_create().
261 mtx_init(&map_sleep_mtx, "vm map sleep mutex", NULL, MTX_DEF);
264 * Disable the use of per-CPU buckets: map entry allocation is
265 * serialized by the kernel map lock.
267 kmapentzone = uma_zcreate("KMAP ENTRY", sizeof(struct vm_map_entry),
268 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR,
269 UMA_ZONE_VM | UMA_ZONE_NOBUCKET);
270 #ifndef UMA_MD_SMALL_ALLOC
271 /* Reserve an extra map entry for use when replenishing the reserve. */
272 uma_zone_reserve(kmapentzone, KMAPENT_RESERVE + 1);
273 uma_prealloc(kmapentzone, KMAPENT_RESERVE + 1);
274 uma_zone_set_allocf(kmapentzone, kmapent_alloc);
275 uma_zone_set_freef(kmapentzone, kmapent_free);
278 mapentzone = uma_zcreate("MAP ENTRY", sizeof(struct vm_map_entry),
279 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
280 vmspace_zone = uma_zcreate("VMSPACE", sizeof(struct vmspace), NULL,
286 vmspace_zinit, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
290 vmspace_zinit(void *mem, int size, int flags)
295 vm = (struct vmspace *)mem;
298 memset(map, 0, sizeof(*map));
299 mtx_init(&map->system_mtx, "vm map (system)", NULL,
300 MTX_DEF | MTX_DUPOK);
301 sx_init(&map->lock, "vm map (user)");
302 PMAP_LOCK_INIT(vmspace_pmap(vm));
308 vmspace_zdtor(void *mem, int size, void *arg)
312 vm = (struct vmspace *)mem;
313 KASSERT(vm->vm_map.nentries == 0,
314 ("vmspace %p nentries == %d on free", vm, vm->vm_map.nentries));
315 KASSERT(vm->vm_map.size == 0,
316 ("vmspace %p size == %ju on free", vm, (uintmax_t)vm->vm_map.size));
318 #endif /* INVARIANTS */
321 * Allocate a vmspace structure, including a vm_map and pmap,
322 * and initialize those structures. The refcnt is set to 1.
325 vmspace_alloc(vm_offset_t min, vm_offset_t max, pmap_pinit_t pinit)
329 vm = uma_zalloc(vmspace_zone, M_WAITOK);
330 KASSERT(vm->vm_map.pmap == NULL, ("vm_map.pmap must be NULL"));
331 if (!pinit(vmspace_pmap(vm))) {
332 uma_zfree(vmspace_zone, vm);
335 CTR1(KTR_VM, "vmspace_alloc: %p", vm);
336 _vm_map_init(&vm->vm_map, vmspace_pmap(vm), min, max);
337 refcount_init(&vm->vm_refcnt, 1);
351 vmspace_container_reset(struct proc *p)
355 racct_set(p, RACCT_DATA, 0);
356 racct_set(p, RACCT_STACK, 0);
357 racct_set(p, RACCT_RSS, 0);
358 racct_set(p, RACCT_MEMLOCK, 0);
359 racct_set(p, RACCT_VMEM, 0);
365 vmspace_dofree(struct vmspace *vm)
368 CTR1(KTR_VM, "vmspace_free: %p", vm);
371 * Make sure any SysV shm is freed, it might not have been in
377 * Lock the map, to wait out all other references to it.
378 * Delete all of the mappings and pages they hold, then call
379 * the pmap module to reclaim anything left.
381 (void)vm_map_remove(&vm->vm_map, vm_map_min(&vm->vm_map),
382 vm_map_max(&vm->vm_map));
384 pmap_release(vmspace_pmap(vm));
385 vm->vm_map.pmap = NULL;
386 uma_zfree(vmspace_zone, vm);
390 vmspace_free(struct vmspace *vm)
393 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
394 "vmspace_free() called");
396 if (refcount_release(&vm->vm_refcnt))
401 vmspace_exitfree(struct proc *p)
405 PROC_VMSPACE_LOCK(p);
408 PROC_VMSPACE_UNLOCK(p);
409 KASSERT(vm == &vmspace0, ("vmspace_exitfree: wrong vmspace"));
414 vmspace_exit(struct thread *td)
424 * Prepare to release the vmspace reference. The thread that releases
425 * the last reference is responsible for tearing down the vmspace.
426 * However, threads not releasing the final reference must switch to the
427 * kernel's vmspace0 before the decrement so that the subsequent pmap
428 * deactivation does not modify a freed vmspace.
430 refcount_acquire(&vmspace0.vm_refcnt);
431 if (!(released = refcount_release_if_last(&vm->vm_refcnt))) {
432 if (p->p_vmspace != &vmspace0) {
433 PROC_VMSPACE_LOCK(p);
434 p->p_vmspace = &vmspace0;
435 PROC_VMSPACE_UNLOCK(p);
438 released = refcount_release(&vm->vm_refcnt);
442 * pmap_remove_pages() expects the pmap to be active, so switch
443 * back first if necessary.
445 if (p->p_vmspace != vm) {
446 PROC_VMSPACE_LOCK(p);
448 PROC_VMSPACE_UNLOCK(p);
451 pmap_remove_pages(vmspace_pmap(vm));
452 PROC_VMSPACE_LOCK(p);
453 p->p_vmspace = &vmspace0;
454 PROC_VMSPACE_UNLOCK(p);
460 vmspace_container_reset(p);
464 /* Acquire reference to vmspace owned by another process. */
467 vmspace_acquire_ref(struct proc *p)
471 PROC_VMSPACE_LOCK(p);
473 if (vm == NULL || !refcount_acquire_if_not_zero(&vm->vm_refcnt)) {
474 PROC_VMSPACE_UNLOCK(p);
477 if (vm != p->p_vmspace) {
478 PROC_VMSPACE_UNLOCK(p);
482 PROC_VMSPACE_UNLOCK(p);
487 * Switch between vmspaces in an AIO kernel process.
489 * The new vmspace is either the vmspace of a user process obtained
490 * from an active AIO request or the initial vmspace of the AIO kernel
491 * process (when it is idling). Because user processes will block to
492 * drain any active AIO requests before proceeding in exit() or
493 * execve(), the reference count for vmspaces from AIO requests can
494 * never be 0. Similarly, AIO kernel processes hold an extra
495 * reference on their initial vmspace for the life of the process. As
496 * a result, the 'newvm' vmspace always has a non-zero reference
497 * count. This permits an additional reference on 'newvm' to be
498 * acquired via a simple atomic increment rather than the loop in
499 * vmspace_acquire_ref() above.
502 vmspace_switch_aio(struct vmspace *newvm)
504 struct vmspace *oldvm;
506 /* XXX: Need some way to assert that this is an aio daemon. */
508 KASSERT(refcount_load(&newvm->vm_refcnt) > 0,
509 ("vmspace_switch_aio: newvm unreferenced"));
511 oldvm = curproc->p_vmspace;
516 * Point to the new address space and refer to it.
518 curproc->p_vmspace = newvm;
519 refcount_acquire(&newvm->vm_refcnt);
521 /* Activate the new mapping. */
522 pmap_activate(curthread);
528 _vm_map_lock(vm_map_t map, const char *file, int line)
532 mtx_lock_flags_(&map->system_mtx, 0, file, line);
534 sx_xlock_(&map->lock, file, line);
539 vm_map_entry_set_vnode_text(vm_map_entry_t entry, bool add)
545 if ((entry->eflags & MAP_ENTRY_VN_EXEC) == 0)
547 KASSERT((entry->eflags & MAP_ENTRY_IS_SUB_MAP) == 0,
548 ("Submap with execs"));
549 object = entry->object.vm_object;
550 KASSERT(object != NULL, ("No object for text, entry %p", entry));
551 if ((object->flags & OBJ_ANON) != 0)
552 object = object->handle;
554 KASSERT(object->backing_object == NULL,
555 ("non-anon object %p shadows", object));
556 KASSERT(object != NULL, ("No content object for text, entry %p obj %p",
557 entry, entry->object.vm_object));
560 * Mostly, we do not lock the backing object. It is
561 * referenced by the entry we are processing, so it cannot go
566 if (object->type == OBJT_DEAD) {
568 * For OBJT_DEAD objects, v_writecount was handled in
569 * vnode_pager_dealloc().
571 } else if (object->type == OBJT_VNODE) {
573 } else if (object->type == OBJT_SWAP) {
574 KASSERT((object->flags & OBJ_TMPFS_NODE) != 0,
575 ("vm_map_entry_set_vnode_text: swap and !TMPFS "
576 "entry %p, object %p, add %d", entry, object, add));
578 * Tmpfs VREG node, which was reclaimed, has
579 * OBJ_TMPFS_NODE flag set, but not OBJ_TMPFS. In
580 * this case there is no v_writecount to adjust.
582 VM_OBJECT_RLOCK(object);
583 if ((object->flags & OBJ_TMPFS) != 0) {
584 vp = object->un_pager.swp.swp_tmpfs;
590 VM_OBJECT_RUNLOCK(object);
593 ("vm_map_entry_set_vnode_text: wrong object type, "
594 "entry %p, object %p, add %d", entry, object, add));
598 VOP_SET_TEXT_CHECKED(vp);
600 vn_lock(vp, LK_SHARED | LK_RETRY);
601 VOP_UNSET_TEXT_CHECKED(vp);
610 * Use a different name for this vm_map_entry field when it's use
611 * is not consistent with its use as part of an ordered search tree.
613 #define defer_next right
616 vm_map_process_deferred(void)
619 vm_map_entry_t entry, next;
623 entry = td->td_map_def_user;
624 td->td_map_def_user = NULL;
625 while (entry != NULL) {
626 next = entry->defer_next;
627 MPASS((entry->eflags & (MAP_ENTRY_WRITECNT |
628 MAP_ENTRY_VN_EXEC)) != (MAP_ENTRY_WRITECNT |
630 if ((entry->eflags & MAP_ENTRY_WRITECNT) != 0) {
632 * Decrement the object's writemappings and
633 * possibly the vnode's v_writecount.
635 KASSERT((entry->eflags & MAP_ENTRY_IS_SUB_MAP) == 0,
636 ("Submap with writecount"));
637 object = entry->object.vm_object;
638 KASSERT(object != NULL, ("No object for writecount"));
639 vm_pager_release_writecount(object, entry->start,
642 vm_map_entry_set_vnode_text(entry, false);
643 vm_map_entry_deallocate(entry, FALSE);
650 _vm_map_assert_locked(vm_map_t map, const char *file, int line)
654 mtx_assert_(&map->system_mtx, MA_OWNED, file, line);
656 sx_assert_(&map->lock, SA_XLOCKED, file, line);
659 #define VM_MAP_ASSERT_LOCKED(map) \
660 _vm_map_assert_locked(map, LOCK_FILE, LOCK_LINE)
662 enum { VMMAP_CHECK_NONE, VMMAP_CHECK_UNLOCK, VMMAP_CHECK_ALL };
664 static int enable_vmmap_check = VMMAP_CHECK_UNLOCK;
666 static int enable_vmmap_check = VMMAP_CHECK_NONE;
668 SYSCTL_INT(_debug, OID_AUTO, vmmap_check, CTLFLAG_RWTUN,
669 &enable_vmmap_check, 0, "Enable vm map consistency checking");
671 static void _vm_map_assert_consistent(vm_map_t map, int check);
673 #define VM_MAP_ASSERT_CONSISTENT(map) \
674 _vm_map_assert_consistent(map, VMMAP_CHECK_ALL)
676 #define VM_MAP_UNLOCK_CONSISTENT(map) do { \
677 if (map->nupdates > map->nentries) { \
678 _vm_map_assert_consistent(map, VMMAP_CHECK_UNLOCK); \
683 #define VM_MAP_UNLOCK_CONSISTENT(map)
686 #define VM_MAP_ASSERT_LOCKED(map)
687 #define VM_MAP_ASSERT_CONSISTENT(map)
688 #define VM_MAP_UNLOCK_CONSISTENT(map)
689 #endif /* INVARIANTS */
692 _vm_map_unlock(vm_map_t map, const char *file, int line)
695 VM_MAP_UNLOCK_CONSISTENT(map);
696 if (map->system_map) {
697 #ifndef UMA_MD_SMALL_ALLOC
698 if (map == kernel_map && (map->flags & MAP_REPLENISH) != 0) {
699 uma_prealloc(kmapentzone, 1);
700 map->flags &= ~MAP_REPLENISH;
703 mtx_unlock_flags_(&map->system_mtx, 0, file, line);
705 sx_xunlock_(&map->lock, file, line);
706 vm_map_process_deferred();
711 _vm_map_lock_read(vm_map_t map, const char *file, int line)
715 mtx_lock_flags_(&map->system_mtx, 0, file, line);
717 sx_slock_(&map->lock, file, line);
721 _vm_map_unlock_read(vm_map_t map, const char *file, int line)
724 if (map->system_map) {
725 KASSERT((map->flags & MAP_REPLENISH) == 0,
726 ("%s: MAP_REPLENISH leaked", __func__));
727 mtx_unlock_flags_(&map->system_mtx, 0, file, line);
729 sx_sunlock_(&map->lock, file, line);
730 vm_map_process_deferred();
735 _vm_map_trylock(vm_map_t map, const char *file, int line)
739 error = map->system_map ?
740 !mtx_trylock_flags_(&map->system_mtx, 0, file, line) :
741 !sx_try_xlock_(&map->lock, file, line);
748 _vm_map_trylock_read(vm_map_t map, const char *file, int line)
752 error = map->system_map ?
753 !mtx_trylock_flags_(&map->system_mtx, 0, file, line) :
754 !sx_try_slock_(&map->lock, file, line);
759 * _vm_map_lock_upgrade: [ internal use only ]
761 * Tries to upgrade a read (shared) lock on the specified map to a write
762 * (exclusive) lock. Returns the value "0" if the upgrade succeeds and a
763 * non-zero value if the upgrade fails. If the upgrade fails, the map is
764 * returned without a read or write lock held.
766 * Requires that the map be read locked.
769 _vm_map_lock_upgrade(vm_map_t map, const char *file, int line)
771 unsigned int last_timestamp;
773 if (map->system_map) {
774 mtx_assert_(&map->system_mtx, MA_OWNED, file, line);
776 if (!sx_try_upgrade_(&map->lock, file, line)) {
777 last_timestamp = map->timestamp;
778 sx_sunlock_(&map->lock, file, line);
779 vm_map_process_deferred();
781 * If the map's timestamp does not change while the
782 * map is unlocked, then the upgrade succeeds.
784 sx_xlock_(&map->lock, file, line);
785 if (last_timestamp != map->timestamp) {
786 sx_xunlock_(&map->lock, file, line);
796 _vm_map_lock_downgrade(vm_map_t map, const char *file, int line)
799 if (map->system_map) {
800 KASSERT((map->flags & MAP_REPLENISH) == 0,
801 ("%s: MAP_REPLENISH leaked", __func__));
802 mtx_assert_(&map->system_mtx, MA_OWNED, file, line);
804 VM_MAP_UNLOCK_CONSISTENT(map);
805 sx_downgrade_(&map->lock, file, line);
812 * Returns a non-zero value if the caller holds a write (exclusive) lock
813 * on the specified map and the value "0" otherwise.
816 vm_map_locked(vm_map_t map)
820 return (mtx_owned(&map->system_mtx));
822 return (sx_xlocked(&map->lock));
826 * _vm_map_unlock_and_wait:
828 * Atomically releases the lock on the specified map and puts the calling
829 * thread to sleep. The calling thread will remain asleep until either
830 * vm_map_wakeup() is performed on the map or the specified timeout is
833 * WARNING! This function does not perform deferred deallocations of
834 * objects and map entries. Therefore, the calling thread is expected to
835 * reacquire the map lock after reawakening and later perform an ordinary
836 * unlock operation, such as vm_map_unlock(), before completing its
837 * operation on the map.
840 _vm_map_unlock_and_wait(vm_map_t map, int timo, const char *file, int line)
843 VM_MAP_UNLOCK_CONSISTENT(map);
844 mtx_lock(&map_sleep_mtx);
845 if (map->system_map) {
846 KASSERT((map->flags & MAP_REPLENISH) == 0,
847 ("%s: MAP_REPLENISH leaked", __func__));
848 mtx_unlock_flags_(&map->system_mtx, 0, file, line);
850 sx_xunlock_(&map->lock, file, line);
852 return (msleep(&map->root, &map_sleep_mtx, PDROP | PVM, "vmmaps",
859 * Awaken any threads that have slept on the map using
860 * vm_map_unlock_and_wait().
863 vm_map_wakeup(vm_map_t map)
867 * Acquire and release map_sleep_mtx to prevent a wakeup()
868 * from being performed (and lost) between the map unlock
869 * and the msleep() in _vm_map_unlock_and_wait().
871 mtx_lock(&map_sleep_mtx);
872 mtx_unlock(&map_sleep_mtx);
877 vm_map_busy(vm_map_t map)
880 VM_MAP_ASSERT_LOCKED(map);
885 vm_map_unbusy(vm_map_t map)
888 VM_MAP_ASSERT_LOCKED(map);
889 KASSERT(map->busy, ("vm_map_unbusy: not busy"));
890 if (--map->busy == 0 && (map->flags & MAP_BUSY_WAKEUP)) {
891 vm_map_modflags(map, 0, MAP_BUSY_WAKEUP);
897 vm_map_wait_busy(vm_map_t map)
900 VM_MAP_ASSERT_LOCKED(map);
902 vm_map_modflags(map, MAP_BUSY_WAKEUP, 0);
904 msleep(&map->busy, &map->system_mtx, 0, "mbusy", 0);
906 sx_sleep(&map->busy, &map->lock, 0, "mbusy", 0);
912 vmspace_resident_count(struct vmspace *vmspace)
914 return pmap_resident_count(vmspace_pmap(vmspace));
918 * Initialize an existing vm_map structure
919 * such as that in the vmspace structure.
922 _vm_map_init(vm_map_t map, pmap_t pmap, vm_offset_t min, vm_offset_t max)
925 map->header.eflags = MAP_ENTRY_HEADER;
926 map->needs_wakeup = FALSE;
929 map->header.end = min;
930 map->header.start = max;
932 map->header.left = map->header.right = &map->header;
943 vm_map_init(vm_map_t map, pmap_t pmap, vm_offset_t min, vm_offset_t max)
946 _vm_map_init(map, pmap, min, max);
947 mtx_init(&map->system_mtx, "vm map (system)", NULL,
948 MTX_DEF | MTX_DUPOK);
949 sx_init(&map->lock, "vm map (user)");
953 * vm_map_entry_dispose: [ internal use only ]
955 * Inverse of vm_map_entry_create.
958 vm_map_entry_dispose(vm_map_t map, vm_map_entry_t entry)
960 uma_zfree(map->system_map ? kmapentzone : mapentzone, entry);
964 * vm_map_entry_create: [ internal use only ]
966 * Allocates a VM map entry for insertion.
967 * No entry fields are filled in.
969 static vm_map_entry_t
970 vm_map_entry_create(vm_map_t map)
972 vm_map_entry_t new_entry;
974 #ifndef UMA_MD_SMALL_ALLOC
975 if (map == kernel_map) {
976 VM_MAP_ASSERT_LOCKED(map);
979 * A new slab of kernel map entries cannot be allocated at this
980 * point because the kernel map has not yet been updated to
981 * reflect the caller's request. Therefore, we allocate a new
982 * map entry, dipping into the reserve if necessary, and set a
983 * flag indicating that the reserve must be replenished before
984 * the map is unlocked.
986 new_entry = uma_zalloc(kmapentzone, M_NOWAIT | M_NOVM);
987 if (new_entry == NULL) {
988 new_entry = uma_zalloc(kmapentzone,
989 M_NOWAIT | M_NOVM | M_USE_RESERVE);
990 kernel_map->flags |= MAP_REPLENISH;
994 if (map->system_map) {
995 new_entry = uma_zalloc(kmapentzone, M_NOWAIT);
997 new_entry = uma_zalloc(mapentzone, M_WAITOK);
999 KASSERT(new_entry != NULL,
1000 ("vm_map_entry_create: kernel resources exhausted"));
1005 * vm_map_entry_set_behavior:
1007 * Set the expected access behavior, either normal, random, or
1011 vm_map_entry_set_behavior(vm_map_entry_t entry, u_char behavior)
1013 entry->eflags = (entry->eflags & ~MAP_ENTRY_BEHAV_MASK) |
1014 (behavior & MAP_ENTRY_BEHAV_MASK);
1018 * vm_map_entry_max_free_{left,right}:
1020 * Compute the size of the largest free gap between two entries,
1021 * one the root of a tree and the other the ancestor of that root
1022 * that is the least or greatest ancestor found on the search path.
1024 static inline vm_size_t
1025 vm_map_entry_max_free_left(vm_map_entry_t root, vm_map_entry_t left_ancestor)
1028 return (root->left != left_ancestor ?
1029 root->left->max_free : root->start - left_ancestor->end);
1032 static inline vm_size_t
1033 vm_map_entry_max_free_right(vm_map_entry_t root, vm_map_entry_t right_ancestor)
1036 return (root->right != right_ancestor ?
1037 root->right->max_free : right_ancestor->start - root->end);
1041 * vm_map_entry_{pred,succ}:
1043 * Find the {predecessor, successor} of the entry by taking one step
1044 * in the appropriate direction and backtracking as much as necessary.
1045 * vm_map_entry_succ is defined in vm_map.h.
1047 static inline vm_map_entry_t
1048 vm_map_entry_pred(vm_map_entry_t entry)
1050 vm_map_entry_t prior;
1052 prior = entry->left;
1053 if (prior->right->start < entry->start) {
1055 prior = prior->right;
1056 while (prior->right != entry);
1061 static inline vm_size_t
1062 vm_size_max(vm_size_t a, vm_size_t b)
1065 return (a > b ? a : b);
1068 #define SPLAY_LEFT_STEP(root, y, llist, rlist, test) do { \
1070 vm_size_t max_free; \
1073 * Infer root->right->max_free == root->max_free when \
1074 * y->max_free < root->max_free || root->max_free == 0. \
1075 * Otherwise, look right to find it. \
1078 max_free = root->max_free; \
1079 KASSERT(max_free == vm_size_max( \
1080 vm_map_entry_max_free_left(root, llist), \
1081 vm_map_entry_max_free_right(root, rlist)), \
1082 ("%s: max_free invariant fails", __func__)); \
1083 if (max_free - 1 < vm_map_entry_max_free_left(root, llist)) \
1084 max_free = vm_map_entry_max_free_right(root, rlist); \
1085 if (y != llist && (test)) { \
1086 /* Rotate right and make y root. */ \
1091 if (max_free < y->max_free) \
1092 root->max_free = max_free = \
1093 vm_size_max(max_free, z->max_free); \
1094 } else if (max_free < y->max_free) \
1095 root->max_free = max_free = \
1096 vm_size_max(max_free, root->start - y->end);\
1100 /* Copy right->max_free. Put root on rlist. */ \
1101 root->max_free = max_free; \
1102 KASSERT(max_free == vm_map_entry_max_free_right(root, rlist), \
1103 ("%s: max_free not copied from right", __func__)); \
1104 root->left = rlist; \
1106 root = y != llist ? y : NULL; \
1109 #define SPLAY_RIGHT_STEP(root, y, llist, rlist, test) do { \
1111 vm_size_t max_free; \
1114 * Infer root->left->max_free == root->max_free when \
1115 * y->max_free < root->max_free || root->max_free == 0. \
1116 * Otherwise, look left to find it. \
1119 max_free = root->max_free; \
1120 KASSERT(max_free == vm_size_max( \
1121 vm_map_entry_max_free_left(root, llist), \
1122 vm_map_entry_max_free_right(root, rlist)), \
1123 ("%s: max_free invariant fails", __func__)); \
1124 if (max_free - 1 < vm_map_entry_max_free_right(root, rlist)) \
1125 max_free = vm_map_entry_max_free_left(root, llist); \
1126 if (y != rlist && (test)) { \
1127 /* Rotate left and make y root. */ \
1132 if (max_free < y->max_free) \
1133 root->max_free = max_free = \
1134 vm_size_max(max_free, z->max_free); \
1135 } else if (max_free < y->max_free) \
1136 root->max_free = max_free = \
1137 vm_size_max(max_free, y->start - root->end);\
1141 /* Copy left->max_free. Put root on llist. */ \
1142 root->max_free = max_free; \
1143 KASSERT(max_free == vm_map_entry_max_free_left(root, llist), \
1144 ("%s: max_free not copied from left", __func__)); \
1145 root->right = llist; \
1147 root = y != rlist ? y : NULL; \
1151 * Walk down the tree until we find addr or a gap where addr would go, breaking
1152 * off left and right subtrees of nodes less than, or greater than addr. Treat
1153 * subtrees with root->max_free < length as empty trees. llist and rlist are
1154 * the two sides in reverse order (bottom-up), with llist linked by the right
1155 * pointer and rlist linked by the left pointer in the vm_map_entry, and both
1156 * lists terminated by &map->header. This function, and the subsequent call to
1157 * vm_map_splay_merge_{left,right,pred,succ}, rely on the start and end address
1158 * values in &map->header.
1160 static __always_inline vm_map_entry_t
1161 vm_map_splay_split(vm_map_t map, vm_offset_t addr, vm_size_t length,
1162 vm_map_entry_t *llist, vm_map_entry_t *rlist)
1164 vm_map_entry_t left, right, root, y;
1166 left = right = &map->header;
1168 while (root != NULL && root->max_free >= length) {
1169 KASSERT(left->end <= root->start &&
1170 root->end <= right->start,
1171 ("%s: root not within tree bounds", __func__));
1172 if (addr < root->start) {
1173 SPLAY_LEFT_STEP(root, y, left, right,
1174 y->max_free >= length && addr < y->start);
1175 } else if (addr >= root->end) {
1176 SPLAY_RIGHT_STEP(root, y, left, right,
1177 y->max_free >= length && addr >= y->end);
1186 static __always_inline void
1187 vm_map_splay_findnext(vm_map_entry_t root, vm_map_entry_t *rlist)
1189 vm_map_entry_t hi, right, y;
1192 hi = root->right == right ? NULL : root->right;
1196 SPLAY_LEFT_STEP(hi, y, root, right, true);
1201 static __always_inline void
1202 vm_map_splay_findprev(vm_map_entry_t root, vm_map_entry_t *llist)
1204 vm_map_entry_t left, lo, y;
1207 lo = root->left == left ? NULL : root->left;
1211 SPLAY_RIGHT_STEP(lo, y, left, root, true);
1217 vm_map_entry_swap(vm_map_entry_t *a, vm_map_entry_t *b)
1227 * Walk back up the two spines, flip the pointers and set max_free. The
1228 * subtrees of the root go at the bottom of llist and rlist.
1231 vm_map_splay_merge_left_walk(vm_map_entry_t header, vm_map_entry_t root,
1232 vm_map_entry_t tail, vm_size_t max_free, vm_map_entry_t llist)
1236 * The max_free values of the children of llist are in
1237 * llist->max_free and max_free. Update with the
1240 llist->max_free = max_free =
1241 vm_size_max(llist->max_free, max_free);
1242 vm_map_entry_swap(&llist->right, &tail);
1243 vm_map_entry_swap(&tail, &llist);
1244 } while (llist != header);
1250 * When llist is known to be the predecessor of root.
1252 static inline vm_size_t
1253 vm_map_splay_merge_pred(vm_map_entry_t header, vm_map_entry_t root,
1254 vm_map_entry_t llist)
1258 max_free = root->start - llist->end;
1259 if (llist != header) {
1260 max_free = vm_map_splay_merge_left_walk(header, root,
1261 root, max_free, llist);
1263 root->left = header;
1264 header->right = root;
1270 * When llist may or may not be the predecessor of root.
1272 static inline vm_size_t
1273 vm_map_splay_merge_left(vm_map_entry_t header, vm_map_entry_t root,
1274 vm_map_entry_t llist)
1278 max_free = vm_map_entry_max_free_left(root, llist);
1279 if (llist != header) {
1280 max_free = vm_map_splay_merge_left_walk(header, root,
1281 root->left == llist ? root : root->left,
1288 vm_map_splay_merge_right_walk(vm_map_entry_t header, vm_map_entry_t root,
1289 vm_map_entry_t tail, vm_size_t max_free, vm_map_entry_t rlist)
1293 * The max_free values of the children of rlist are in
1294 * rlist->max_free and max_free. Update with the
1297 rlist->max_free = max_free =
1298 vm_size_max(rlist->max_free, max_free);
1299 vm_map_entry_swap(&rlist->left, &tail);
1300 vm_map_entry_swap(&tail, &rlist);
1301 } while (rlist != header);
1307 * When rlist is known to be the succecessor of root.
1309 static inline vm_size_t
1310 vm_map_splay_merge_succ(vm_map_entry_t header, vm_map_entry_t root,
1311 vm_map_entry_t rlist)
1315 max_free = rlist->start - root->end;
1316 if (rlist != header) {
1317 max_free = vm_map_splay_merge_right_walk(header, root,
1318 root, max_free, rlist);
1320 root->right = header;
1321 header->left = root;
1327 * When rlist may or may not be the succecessor of root.
1329 static inline vm_size_t
1330 vm_map_splay_merge_right(vm_map_entry_t header, vm_map_entry_t root,
1331 vm_map_entry_t rlist)
1335 max_free = vm_map_entry_max_free_right(root, rlist);
1336 if (rlist != header) {
1337 max_free = vm_map_splay_merge_right_walk(header, root,
1338 root->right == rlist ? root : root->right,
1347 * The Sleator and Tarjan top-down splay algorithm with the
1348 * following variation. Max_free must be computed bottom-up, so
1349 * on the downward pass, maintain the left and right spines in
1350 * reverse order. Then, make a second pass up each side to fix
1351 * the pointers and compute max_free. The time bound is O(log n)
1354 * The tree is threaded, which means that there are no null pointers.
1355 * When a node has no left child, its left pointer points to its
1356 * predecessor, which the last ancestor on the search path from the root
1357 * where the search branched right. Likewise, when a node has no right
1358 * child, its right pointer points to its successor. The map header node
1359 * is the predecessor of the first map entry, and the successor of the
1362 * The new root is the vm_map_entry containing "addr", or else an
1363 * adjacent entry (lower if possible) if addr is not in the tree.
1365 * The map must be locked, and leaves it so.
1367 * Returns: the new root.
1369 static vm_map_entry_t
1370 vm_map_splay(vm_map_t map, vm_offset_t addr)
1372 vm_map_entry_t header, llist, rlist, root;
1373 vm_size_t max_free_left, max_free_right;
1375 header = &map->header;
1376 root = vm_map_splay_split(map, addr, 0, &llist, &rlist);
1378 max_free_left = vm_map_splay_merge_left(header, root, llist);
1379 max_free_right = vm_map_splay_merge_right(header, root, rlist);
1380 } else if (llist != header) {
1382 * Recover the greatest node in the left
1383 * subtree and make it the root.
1386 llist = root->right;
1387 max_free_left = vm_map_splay_merge_left(header, root, llist);
1388 max_free_right = vm_map_splay_merge_succ(header, root, rlist);
1389 } else if (rlist != header) {
1391 * Recover the least node in the right
1392 * subtree and make it the root.
1396 max_free_left = vm_map_splay_merge_pred(header, root, llist);
1397 max_free_right = vm_map_splay_merge_right(header, root, rlist);
1399 /* There is no root. */
1402 root->max_free = vm_size_max(max_free_left, max_free_right);
1404 VM_MAP_ASSERT_CONSISTENT(map);
1409 * vm_map_entry_{un,}link:
1411 * Insert/remove entries from maps. On linking, if new entry clips
1412 * existing entry, trim existing entry to avoid overlap, and manage
1413 * offsets. On unlinking, merge disappearing entry with neighbor, if
1414 * called for, and manage offsets. Callers should not modify fields in
1415 * entries already mapped.
1418 vm_map_entry_link(vm_map_t map, vm_map_entry_t entry)
1420 vm_map_entry_t header, llist, rlist, root;
1421 vm_size_t max_free_left, max_free_right;
1424 "vm_map_entry_link: map %p, nentries %d, entry %p", map,
1425 map->nentries, entry);
1426 VM_MAP_ASSERT_LOCKED(map);
1428 header = &map->header;
1429 root = vm_map_splay_split(map, entry->start, 0, &llist, &rlist);
1432 * The new entry does not overlap any existing entry in the
1433 * map, so it becomes the new root of the map tree.
1435 max_free_left = vm_map_splay_merge_pred(header, entry, llist);
1436 max_free_right = vm_map_splay_merge_succ(header, entry, rlist);
1437 } else if (entry->start == root->start) {
1439 * The new entry is a clone of root, with only the end field
1440 * changed. The root entry will be shrunk to abut the new
1441 * entry, and will be the right child of the new root entry in
1444 KASSERT(entry->end < root->end,
1445 ("%s: clip_start not within entry", __func__));
1446 vm_map_splay_findprev(root, &llist);
1447 root->offset += entry->end - root->start;
1448 root->start = entry->end;
1449 max_free_left = vm_map_splay_merge_pred(header, entry, llist);
1450 max_free_right = root->max_free = vm_size_max(
1451 vm_map_splay_merge_pred(entry, root, entry),
1452 vm_map_splay_merge_right(header, root, rlist));
1455 * The new entry is a clone of root, with only the start field
1456 * changed. The root entry will be shrunk to abut the new
1457 * entry, and will be the left child of the new root entry in
1460 KASSERT(entry->end == root->end,
1461 ("%s: clip_start not within entry", __func__));
1462 vm_map_splay_findnext(root, &rlist);
1463 entry->offset += entry->start - root->start;
1464 root->end = entry->start;
1465 max_free_left = root->max_free = vm_size_max(
1466 vm_map_splay_merge_left(header, root, llist),
1467 vm_map_splay_merge_succ(entry, root, entry));
1468 max_free_right = vm_map_splay_merge_succ(header, entry, rlist);
1470 entry->max_free = vm_size_max(max_free_left, max_free_right);
1472 VM_MAP_ASSERT_CONSISTENT(map);
1475 enum unlink_merge_type {
1481 vm_map_entry_unlink(vm_map_t map, vm_map_entry_t entry,
1482 enum unlink_merge_type op)
1484 vm_map_entry_t header, llist, rlist, root;
1485 vm_size_t max_free_left, max_free_right;
1487 VM_MAP_ASSERT_LOCKED(map);
1488 header = &map->header;
1489 root = vm_map_splay_split(map, entry->start, 0, &llist, &rlist);
1490 KASSERT(root != NULL,
1491 ("vm_map_entry_unlink: unlink object not mapped"));
1493 vm_map_splay_findprev(root, &llist);
1494 vm_map_splay_findnext(root, &rlist);
1495 if (op == UNLINK_MERGE_NEXT) {
1496 rlist->start = root->start;
1497 rlist->offset = root->offset;
1499 if (llist != header) {
1501 llist = root->right;
1502 max_free_left = vm_map_splay_merge_left(header, root, llist);
1503 max_free_right = vm_map_splay_merge_succ(header, root, rlist);
1504 } else if (rlist != header) {
1507 max_free_left = vm_map_splay_merge_pred(header, root, llist);
1508 max_free_right = vm_map_splay_merge_right(header, root, rlist);
1510 header->left = header->right = header;
1514 root->max_free = vm_size_max(max_free_left, max_free_right);
1516 VM_MAP_ASSERT_CONSISTENT(map);
1518 CTR3(KTR_VM, "vm_map_entry_unlink: map %p, nentries %d, entry %p", map,
1519 map->nentries, entry);
1523 * vm_map_entry_resize:
1525 * Resize a vm_map_entry, recompute the amount of free space that
1526 * follows it and propagate that value up the tree.
1528 * The map must be locked, and leaves it so.
1531 vm_map_entry_resize(vm_map_t map, vm_map_entry_t entry, vm_size_t grow_amount)
1533 vm_map_entry_t header, llist, rlist, root;
1535 VM_MAP_ASSERT_LOCKED(map);
1536 header = &map->header;
1537 root = vm_map_splay_split(map, entry->start, 0, &llist, &rlist);
1538 KASSERT(root != NULL, ("%s: resize object not mapped", __func__));
1539 vm_map_splay_findnext(root, &rlist);
1540 entry->end += grow_amount;
1541 root->max_free = vm_size_max(
1542 vm_map_splay_merge_left(header, root, llist),
1543 vm_map_splay_merge_succ(header, root, rlist));
1545 VM_MAP_ASSERT_CONSISTENT(map);
1546 CTR4(KTR_VM, "%s: map %p, nentries %d, entry %p",
1547 __func__, map, map->nentries, entry);
1551 * vm_map_lookup_entry: [ internal use only ]
1553 * Finds the map entry containing (or
1554 * immediately preceding) the specified address
1555 * in the given map; the entry is returned
1556 * in the "entry" parameter. The boolean
1557 * result indicates whether the address is
1558 * actually contained in the map.
1561 vm_map_lookup_entry(
1563 vm_offset_t address,
1564 vm_map_entry_t *entry) /* OUT */
1566 vm_map_entry_t cur, header, lbound, ubound;
1570 * If the map is empty, then the map entry immediately preceding
1571 * "address" is the map's header.
1573 header = &map->header;
1579 if (address >= cur->start && cur->end > address) {
1583 if ((locked = vm_map_locked(map)) ||
1584 sx_try_upgrade(&map->lock)) {
1586 * Splay requires a write lock on the map. However, it only
1587 * restructures the binary search tree; it does not otherwise
1588 * change the map. Thus, the map's timestamp need not change
1589 * on a temporary upgrade.
1591 cur = vm_map_splay(map, address);
1593 VM_MAP_UNLOCK_CONSISTENT(map);
1594 sx_downgrade(&map->lock);
1598 * If "address" is contained within a map entry, the new root
1599 * is that map entry. Otherwise, the new root is a map entry
1600 * immediately before or after "address".
1602 if (address < cur->start) {
1607 return (address < cur->end);
1610 * Since the map is only locked for read access, perform a
1611 * standard binary search tree lookup for "address".
1613 lbound = ubound = header;
1615 if (address < cur->start) {
1620 } else if (cur->end <= address) {
1637 * Inserts the given whole VM object into the target
1638 * map at the specified address range. The object's
1639 * size should match that of the address range.
1641 * Requires that the map be locked, and leaves it so.
1643 * If object is non-NULL, ref count must be bumped by caller
1644 * prior to making call to account for the new entry.
1647 vm_map_insert(vm_map_t map, vm_object_t object, vm_ooffset_t offset,
1648 vm_offset_t start, vm_offset_t end, vm_prot_t prot, vm_prot_t max, int cow)
1650 vm_map_entry_t new_entry, next_entry, prev_entry;
1652 vm_eflags_t protoeflags;
1653 vm_inherit_t inheritance;
1657 VM_MAP_ASSERT_LOCKED(map);
1658 KASSERT(object != kernel_object ||
1659 (cow & MAP_COPY_ON_WRITE) == 0,
1660 ("vm_map_insert: kernel object and COW"));
1661 KASSERT(object == NULL || (cow & MAP_NOFAULT) == 0 ||
1662 (cow & MAP_SPLIT_BOUNDARY_MASK) != 0,
1663 ("vm_map_insert: paradoxical MAP_NOFAULT request, obj %p cow %#x",
1665 KASSERT((prot & ~max) == 0,
1666 ("prot %#x is not subset of max_prot %#x", prot, max));
1669 * Check that the start and end points are not bogus.
1671 if (start == end || !vm_map_range_valid(map, start, end))
1672 return (KERN_INVALID_ADDRESS);
1674 if ((map->flags & MAP_WXORX) != 0 && (prot & (VM_PROT_WRITE |
1675 VM_PROT_EXECUTE)) == (VM_PROT_WRITE | VM_PROT_EXECUTE))
1676 return (KERN_PROTECTION_FAILURE);
1679 * Find the entry prior to the proposed starting address; if it's part
1680 * of an existing entry, this range is bogus.
1682 if (vm_map_lookup_entry(map, start, &prev_entry))
1683 return (KERN_NO_SPACE);
1686 * Assert that the next entry doesn't overlap the end point.
1688 next_entry = vm_map_entry_succ(prev_entry);
1689 if (next_entry->start < end)
1690 return (KERN_NO_SPACE);
1692 if ((cow & MAP_CREATE_GUARD) != 0 && (object != NULL ||
1693 max != VM_PROT_NONE))
1694 return (KERN_INVALID_ARGUMENT);
1697 if (cow & MAP_COPY_ON_WRITE)
1698 protoeflags |= MAP_ENTRY_COW | MAP_ENTRY_NEEDS_COPY;
1699 if (cow & MAP_NOFAULT)
1700 protoeflags |= MAP_ENTRY_NOFAULT;
1701 if (cow & MAP_DISABLE_SYNCER)
1702 protoeflags |= MAP_ENTRY_NOSYNC;
1703 if (cow & MAP_DISABLE_COREDUMP)
1704 protoeflags |= MAP_ENTRY_NOCOREDUMP;
1705 if (cow & MAP_STACK_GROWS_DOWN)
1706 protoeflags |= MAP_ENTRY_GROWS_DOWN;
1707 if (cow & MAP_STACK_GROWS_UP)
1708 protoeflags |= MAP_ENTRY_GROWS_UP;
1709 if (cow & MAP_WRITECOUNT)
1710 protoeflags |= MAP_ENTRY_WRITECNT;
1711 if (cow & MAP_VN_EXEC)
1712 protoeflags |= MAP_ENTRY_VN_EXEC;
1713 if ((cow & MAP_CREATE_GUARD) != 0)
1714 protoeflags |= MAP_ENTRY_GUARD;
1715 if ((cow & MAP_CREATE_STACK_GAP_DN) != 0)
1716 protoeflags |= MAP_ENTRY_STACK_GAP_DN;
1717 if ((cow & MAP_CREATE_STACK_GAP_UP) != 0)
1718 protoeflags |= MAP_ENTRY_STACK_GAP_UP;
1719 if (cow & MAP_INHERIT_SHARE)
1720 inheritance = VM_INHERIT_SHARE;
1722 inheritance = VM_INHERIT_DEFAULT;
1723 if ((cow & MAP_SPLIT_BOUNDARY_MASK) != 0) {
1724 /* This magically ignores index 0, for usual page size. */
1725 bidx = (cow & MAP_SPLIT_BOUNDARY_MASK) >>
1726 MAP_SPLIT_BOUNDARY_SHIFT;
1727 if (bidx >= MAXPAGESIZES)
1728 return (KERN_INVALID_ARGUMENT);
1729 bdry = pagesizes[bidx] - 1;
1730 if ((start & bdry) != 0 || (end & bdry) != 0)
1731 return (KERN_INVALID_ARGUMENT);
1732 protoeflags |= bidx << MAP_ENTRY_SPLIT_BOUNDARY_SHIFT;
1736 if ((cow & (MAP_ACC_NO_CHARGE | MAP_NOFAULT | MAP_CREATE_GUARD)) != 0)
1738 if ((cow & MAP_ACC_CHARGED) || ((prot & VM_PROT_WRITE) &&
1739 ((protoeflags & MAP_ENTRY_NEEDS_COPY) || object == NULL))) {
1740 if (!(cow & MAP_ACC_CHARGED) && !swap_reserve(end - start))
1741 return (KERN_RESOURCE_SHORTAGE);
1742 KASSERT(object == NULL ||
1743 (protoeflags & MAP_ENTRY_NEEDS_COPY) != 0 ||
1744 object->cred == NULL,
1745 ("overcommit: vm_map_insert o %p", object));
1746 cred = curthread->td_ucred;
1750 /* Expand the kernel pmap, if necessary. */
1751 if (map == kernel_map && end > kernel_vm_end)
1752 pmap_growkernel(end);
1753 if (object != NULL) {
1755 * OBJ_ONEMAPPING must be cleared unless this mapping
1756 * is trivially proven to be the only mapping for any
1757 * of the object's pages. (Object granularity
1758 * reference counting is insufficient to recognize
1759 * aliases with precision.)
1761 if ((object->flags & OBJ_ANON) != 0) {
1762 VM_OBJECT_WLOCK(object);
1763 if (object->ref_count > 1 || object->shadow_count != 0)
1764 vm_object_clear_flag(object, OBJ_ONEMAPPING);
1765 VM_OBJECT_WUNLOCK(object);
1767 } else if ((prev_entry->eflags & ~MAP_ENTRY_USER_WIRED) ==
1769 (cow & (MAP_STACK_GROWS_DOWN | MAP_STACK_GROWS_UP |
1770 MAP_VN_EXEC)) == 0 &&
1771 prev_entry->end == start && (prev_entry->cred == cred ||
1772 (prev_entry->object.vm_object != NULL &&
1773 prev_entry->object.vm_object->cred == cred)) &&
1774 vm_object_coalesce(prev_entry->object.vm_object,
1776 (vm_size_t)(prev_entry->end - prev_entry->start),
1777 (vm_size_t)(end - prev_entry->end), cred != NULL &&
1778 (protoeflags & MAP_ENTRY_NEEDS_COPY) == 0)) {
1780 * We were able to extend the object. Determine if we
1781 * can extend the previous map entry to include the
1782 * new range as well.
1784 if (prev_entry->inheritance == inheritance &&
1785 prev_entry->protection == prot &&
1786 prev_entry->max_protection == max &&
1787 prev_entry->wired_count == 0) {
1788 KASSERT((prev_entry->eflags & MAP_ENTRY_USER_WIRED) ==
1789 0, ("prev_entry %p has incoherent wiring",
1791 if ((prev_entry->eflags & MAP_ENTRY_GUARD) == 0)
1792 map->size += end - prev_entry->end;
1793 vm_map_entry_resize(map, prev_entry,
1794 end - prev_entry->end);
1795 vm_map_try_merge_entries(map, prev_entry, next_entry);
1796 return (KERN_SUCCESS);
1800 * If we can extend the object but cannot extend the
1801 * map entry, we have to create a new map entry. We
1802 * must bump the ref count on the extended object to
1803 * account for it. object may be NULL.
1805 object = prev_entry->object.vm_object;
1806 offset = prev_entry->offset +
1807 (prev_entry->end - prev_entry->start);
1808 vm_object_reference(object);
1809 if (cred != NULL && object != NULL && object->cred != NULL &&
1810 !(prev_entry->eflags & MAP_ENTRY_NEEDS_COPY)) {
1811 /* Object already accounts for this uid. */
1819 * Create a new entry
1821 new_entry = vm_map_entry_create(map);
1822 new_entry->start = start;
1823 new_entry->end = end;
1824 new_entry->cred = NULL;
1826 new_entry->eflags = protoeflags;
1827 new_entry->object.vm_object = object;
1828 new_entry->offset = offset;
1830 new_entry->inheritance = inheritance;
1831 new_entry->protection = prot;
1832 new_entry->max_protection = max;
1833 new_entry->wired_count = 0;
1834 new_entry->wiring_thread = NULL;
1835 new_entry->read_ahead = VM_FAULT_READ_AHEAD_INIT;
1836 new_entry->next_read = start;
1838 KASSERT(cred == NULL || !ENTRY_CHARGED(new_entry),
1839 ("overcommit: vm_map_insert leaks vm_map %p", new_entry));
1840 new_entry->cred = cred;
1843 * Insert the new entry into the list
1845 vm_map_entry_link(map, new_entry);
1846 if ((new_entry->eflags & MAP_ENTRY_GUARD) == 0)
1847 map->size += new_entry->end - new_entry->start;
1850 * Try to coalesce the new entry with both the previous and next
1851 * entries in the list. Previously, we only attempted to coalesce
1852 * with the previous entry when object is NULL. Here, we handle the
1853 * other cases, which are less common.
1855 vm_map_try_merge_entries(map, prev_entry, new_entry);
1856 vm_map_try_merge_entries(map, new_entry, next_entry);
1858 if ((cow & (MAP_PREFAULT | MAP_PREFAULT_PARTIAL)) != 0) {
1859 vm_map_pmap_enter(map, start, prot, object, OFF_TO_IDX(offset),
1860 end - start, cow & MAP_PREFAULT_PARTIAL);
1863 return (KERN_SUCCESS);
1869 * Find the first fit (lowest VM address) for "length" free bytes
1870 * beginning at address >= start in the given map.
1872 * In a vm_map_entry, "max_free" is the maximum amount of
1873 * contiguous free space between an entry in its subtree and a
1874 * neighbor of that entry. This allows finding a free region in
1875 * one path down the tree, so O(log n) amortized with splay
1878 * The map must be locked, and leaves it so.
1880 * Returns: starting address if sufficient space,
1881 * vm_map_max(map)-length+1 if insufficient space.
1884 vm_map_findspace(vm_map_t map, vm_offset_t start, vm_size_t length)
1886 vm_map_entry_t header, llist, rlist, root, y;
1887 vm_size_t left_length, max_free_left, max_free_right;
1888 vm_offset_t gap_end;
1890 VM_MAP_ASSERT_LOCKED(map);
1893 * Request must fit within min/max VM address and must avoid
1896 start = MAX(start, vm_map_min(map));
1897 if (start >= vm_map_max(map) || length > vm_map_max(map) - start)
1898 return (vm_map_max(map) - length + 1);
1900 /* Empty tree means wide open address space. */
1901 if (map->root == NULL)
1905 * After splay_split, if start is within an entry, push it to the start
1906 * of the following gap. If rlist is at the end of the gap containing
1907 * start, save the end of that gap in gap_end to see if the gap is big
1908 * enough; otherwise set gap_end to start skip gap-checking and move
1909 * directly to a search of the right subtree.
1911 header = &map->header;
1912 root = vm_map_splay_split(map, start, length, &llist, &rlist);
1913 gap_end = rlist->start;
1916 if (root->right != rlist)
1918 max_free_left = vm_map_splay_merge_left(header, root, llist);
1919 max_free_right = vm_map_splay_merge_right(header, root, rlist);
1920 } else if (rlist != header) {
1923 max_free_left = vm_map_splay_merge_pred(header, root, llist);
1924 max_free_right = vm_map_splay_merge_right(header, root, rlist);
1927 llist = root->right;
1928 max_free_left = vm_map_splay_merge_left(header, root, llist);
1929 max_free_right = vm_map_splay_merge_succ(header, root, rlist);
1931 root->max_free = vm_size_max(max_free_left, max_free_right);
1933 VM_MAP_ASSERT_CONSISTENT(map);
1934 if (length <= gap_end - start)
1937 /* With max_free, can immediately tell if no solution. */
1938 if (root->right == header || length > root->right->max_free)
1939 return (vm_map_max(map) - length + 1);
1942 * Splay for the least large-enough gap in the right subtree.
1944 llist = rlist = header;
1945 for (left_length = 0;;
1946 left_length = vm_map_entry_max_free_left(root, llist)) {
1947 if (length <= left_length)
1948 SPLAY_LEFT_STEP(root, y, llist, rlist,
1949 length <= vm_map_entry_max_free_left(y, llist));
1951 SPLAY_RIGHT_STEP(root, y, llist, rlist,
1952 length > vm_map_entry_max_free_left(y, root));
1957 llist = root->right;
1958 max_free_left = vm_map_splay_merge_left(header, root, llist);
1959 if (rlist == header) {
1960 root->max_free = vm_size_max(max_free_left,
1961 vm_map_splay_merge_succ(header, root, rlist));
1965 y->max_free = vm_size_max(
1966 vm_map_splay_merge_pred(root, y, root),
1967 vm_map_splay_merge_right(header, y, rlist));
1968 root->max_free = vm_size_max(max_free_left, y->max_free);
1971 VM_MAP_ASSERT_CONSISTENT(map);
1976 vm_map_fixed(vm_map_t map, vm_object_t object, vm_ooffset_t offset,
1977 vm_offset_t start, vm_size_t length, vm_prot_t prot,
1978 vm_prot_t max, int cow)
1983 end = start + length;
1984 KASSERT((cow & (MAP_STACK_GROWS_DOWN | MAP_STACK_GROWS_UP)) == 0 ||
1986 ("vm_map_fixed: non-NULL backing object for stack"));
1988 VM_MAP_RANGE_CHECK(map, start, end);
1989 if ((cow & MAP_CHECK_EXCL) == 0) {
1990 result = vm_map_delete(map, start, end);
1991 if (result != KERN_SUCCESS)
1994 if ((cow & (MAP_STACK_GROWS_DOWN | MAP_STACK_GROWS_UP)) != 0) {
1995 result = vm_map_stack_locked(map, start, length, sgrowsiz,
1998 result = vm_map_insert(map, object, offset, start, end,
2006 static const int aslr_pages_rnd_64[2] = {0x1000, 0x10};
2007 static const int aslr_pages_rnd_32[2] = {0x100, 0x4};
2009 static int cluster_anon = 1;
2010 SYSCTL_INT(_vm, OID_AUTO, cluster_anon, CTLFLAG_RW,
2012 "Cluster anonymous mappings: 0 = no, 1 = yes if no hint, 2 = always");
2015 clustering_anon_allowed(vm_offset_t addr)
2018 switch (cluster_anon) {
2029 static long aslr_restarts;
2030 SYSCTL_LONG(_vm, OID_AUTO, aslr_restarts, CTLFLAG_RD,
2032 "Number of aslr failures");
2035 * Searches for the specified amount of free space in the given map with the
2036 * specified alignment. Performs an address-ordered, first-fit search from
2037 * the given address "*addr", with an optional upper bound "max_addr". If the
2038 * parameter "alignment" is zero, then the alignment is computed from the
2039 * given (object, offset) pair so as to enable the greatest possible use of
2040 * superpage mappings. Returns KERN_SUCCESS and the address of the free space
2041 * in "*addr" if successful. Otherwise, returns KERN_NO_SPACE.
2043 * The map must be locked. Initially, there must be at least "length" bytes
2044 * of free space at the given address.
2047 vm_map_alignspace(vm_map_t map, vm_object_t object, vm_ooffset_t offset,
2048 vm_offset_t *addr, vm_size_t length, vm_offset_t max_addr,
2049 vm_offset_t alignment)
2051 vm_offset_t aligned_addr, free_addr;
2053 VM_MAP_ASSERT_LOCKED(map);
2055 KASSERT(free_addr == vm_map_findspace(map, free_addr, length),
2056 ("caller failed to provide space %#jx at address %p",
2057 (uintmax_t)length, (void *)free_addr));
2060 * At the start of every iteration, the free space at address
2061 * "*addr" is at least "length" bytes.
2064 pmap_align_superpage(object, offset, addr, length);
2065 else if ((*addr & (alignment - 1)) != 0) {
2066 *addr &= ~(alignment - 1);
2069 aligned_addr = *addr;
2070 if (aligned_addr == free_addr) {
2072 * Alignment did not change "*addr", so "*addr" must
2073 * still provide sufficient free space.
2075 return (KERN_SUCCESS);
2079 * Test for address wrap on "*addr". A wrapped "*addr" could
2080 * be a valid address, in which case vm_map_findspace() cannot
2081 * be relied upon to fail.
2083 if (aligned_addr < free_addr)
2084 return (KERN_NO_SPACE);
2085 *addr = vm_map_findspace(map, aligned_addr, length);
2086 if (*addr + length > vm_map_max(map) ||
2087 (max_addr != 0 && *addr + length > max_addr))
2088 return (KERN_NO_SPACE);
2090 if (free_addr == aligned_addr) {
2092 * If a successful call to vm_map_findspace() did not
2093 * change "*addr", then "*addr" must still be aligned
2094 * and provide sufficient free space.
2096 return (KERN_SUCCESS);
2102 vm_map_find_aligned(vm_map_t map, vm_offset_t *addr, vm_size_t length,
2103 vm_offset_t max_addr, vm_offset_t alignment)
2105 /* XXXKIB ASLR eh ? */
2106 *addr = vm_map_findspace(map, *addr, length);
2107 if (*addr + length > vm_map_max(map) ||
2108 (max_addr != 0 && *addr + length > max_addr))
2109 return (KERN_NO_SPACE);
2110 return (vm_map_alignspace(map, NULL, 0, addr, length, max_addr,
2115 * vm_map_find finds an unallocated region in the target address
2116 * map with the given length. The search is defined to be
2117 * first-fit from the specified address; the region found is
2118 * returned in the same parameter.
2120 * If object is non-NULL, ref count must be bumped by caller
2121 * prior to making call to account for the new entry.
2124 vm_map_find(vm_map_t map, vm_object_t object, vm_ooffset_t offset,
2125 vm_offset_t *addr, /* IN/OUT */
2126 vm_size_t length, vm_offset_t max_addr, int find_space,
2127 vm_prot_t prot, vm_prot_t max, int cow)
2129 vm_offset_t alignment, curr_min_addr, min_addr;
2130 int gap, pidx, rv, try;
2131 bool cluster, en_aslr, update_anon;
2133 KASSERT((cow & (MAP_STACK_GROWS_DOWN | MAP_STACK_GROWS_UP)) == 0 ||
2135 ("vm_map_find: non-NULL backing object for stack"));
2136 MPASS((cow & MAP_REMAP) == 0 || (find_space == VMFS_NO_SPACE &&
2137 (cow & (MAP_STACK_GROWS_DOWN | MAP_STACK_GROWS_UP)) == 0));
2138 if (find_space == VMFS_OPTIMAL_SPACE && (object == NULL ||
2139 (object->flags & OBJ_COLORED) == 0))
2140 find_space = VMFS_ANY_SPACE;
2141 if (find_space >> 8 != 0) {
2142 KASSERT((find_space & 0xff) == 0, ("bad VMFS flags"));
2143 alignment = (vm_offset_t)1 << (find_space >> 8);
2146 en_aslr = (map->flags & MAP_ASLR) != 0;
2147 update_anon = cluster = clustering_anon_allowed(*addr) &&
2148 (map->flags & MAP_IS_SUB_MAP) == 0 && max_addr == 0 &&
2149 find_space != VMFS_NO_SPACE && object == NULL &&
2150 (cow & (MAP_INHERIT_SHARE | MAP_STACK_GROWS_UP |
2151 MAP_STACK_GROWS_DOWN)) == 0 && prot != PROT_NONE;
2152 curr_min_addr = min_addr = *addr;
2153 if (en_aslr && min_addr == 0 && !cluster &&
2154 find_space != VMFS_NO_SPACE &&
2155 (map->flags & MAP_ASLR_IGNSTART) != 0)
2156 curr_min_addr = min_addr = vm_map_min(map);
2160 curr_min_addr = map->anon_loc;
2161 if (curr_min_addr == 0)
2164 if (find_space != VMFS_NO_SPACE) {
2165 KASSERT(find_space == VMFS_ANY_SPACE ||
2166 find_space == VMFS_OPTIMAL_SPACE ||
2167 find_space == VMFS_SUPER_SPACE ||
2168 alignment != 0, ("unexpected VMFS flag"));
2171 * When creating an anonymous mapping, try clustering
2172 * with an existing anonymous mapping first.
2174 * We make up to two attempts to find address space
2175 * for a given find_space value. The first attempt may
2176 * apply randomization or may cluster with an existing
2177 * anonymous mapping. If this first attempt fails,
2178 * perform a first-fit search of the available address
2181 * If all tries failed, and find_space is
2182 * VMFS_OPTIMAL_SPACE, fallback to VMFS_ANY_SPACE.
2183 * Again enable clustering and randomization.
2190 * Second try: we failed either to find a
2191 * suitable region for randomizing the
2192 * allocation, or to cluster with an existing
2193 * mapping. Retry with free run.
2195 curr_min_addr = (map->flags & MAP_ASLR_IGNSTART) != 0 ?
2196 vm_map_min(map) : min_addr;
2197 atomic_add_long(&aslr_restarts, 1);
2200 if (try == 1 && en_aslr && !cluster) {
2202 * Find space for allocation, including
2203 * gap needed for later randomization.
2205 pidx = MAXPAGESIZES > 1 && pagesizes[1] != 0 &&
2206 (find_space == VMFS_SUPER_SPACE || find_space ==
2207 VMFS_OPTIMAL_SPACE) ? 1 : 0;
2208 gap = vm_map_max(map) > MAP_32BIT_MAX_ADDR &&
2209 (max_addr == 0 || max_addr > MAP_32BIT_MAX_ADDR) ?
2210 aslr_pages_rnd_64[pidx] : aslr_pages_rnd_32[pidx];
2211 *addr = vm_map_findspace(map, curr_min_addr,
2212 length + gap * pagesizes[pidx]);
2213 if (*addr + length + gap * pagesizes[pidx] >
2216 /* And randomize the start address. */
2217 *addr += (arc4random() % gap) * pagesizes[pidx];
2218 if (max_addr != 0 && *addr + length > max_addr)
2221 *addr = vm_map_findspace(map, curr_min_addr, length);
2222 if (*addr + length > vm_map_max(map) ||
2223 (max_addr != 0 && *addr + length > max_addr)) {
2234 if (find_space != VMFS_ANY_SPACE &&
2235 (rv = vm_map_alignspace(map, object, offset, addr, length,
2236 max_addr, alignment)) != KERN_SUCCESS) {
2237 if (find_space == VMFS_OPTIMAL_SPACE) {
2238 find_space = VMFS_ANY_SPACE;
2239 curr_min_addr = min_addr;
2240 cluster = update_anon;
2246 } else if ((cow & MAP_REMAP) != 0) {
2247 if (!vm_map_range_valid(map, *addr, *addr + length)) {
2248 rv = KERN_INVALID_ADDRESS;
2251 rv = vm_map_delete(map, *addr, *addr + length);
2252 if (rv != KERN_SUCCESS)
2255 if ((cow & (MAP_STACK_GROWS_DOWN | MAP_STACK_GROWS_UP)) != 0) {
2256 rv = vm_map_stack_locked(map, *addr, length, sgrowsiz, prot,
2259 rv = vm_map_insert(map, object, offset, *addr, *addr + length,
2262 if (rv == KERN_SUCCESS && update_anon)
2263 map->anon_loc = *addr + length;
2270 * vm_map_find_min() is a variant of vm_map_find() that takes an
2271 * additional parameter (min_addr) and treats the given address
2272 * (*addr) differently. Specifically, it treats *addr as a hint
2273 * and not as the minimum address where the mapping is created.
2275 * This function works in two phases. First, it tries to
2276 * allocate above the hint. If that fails and the hint is
2277 * greater than min_addr, it performs a second pass, replacing
2278 * the hint with min_addr as the minimum address for the
2282 vm_map_find_min(vm_map_t map, vm_object_t object, vm_ooffset_t offset,
2283 vm_offset_t *addr, vm_size_t length, vm_offset_t min_addr,
2284 vm_offset_t max_addr, int find_space, vm_prot_t prot, vm_prot_t max,
2292 rv = vm_map_find(map, object, offset, addr, length, max_addr,
2293 find_space, prot, max, cow);
2294 if (rv == KERN_SUCCESS || min_addr >= hint)
2296 *addr = hint = min_addr;
2301 * A map entry with any of the following flags set must not be merged with
2304 #define MAP_ENTRY_NOMERGE_MASK (MAP_ENTRY_GROWS_DOWN | MAP_ENTRY_GROWS_UP | \
2305 MAP_ENTRY_IN_TRANSITION | MAP_ENTRY_IS_SUB_MAP | MAP_ENTRY_VN_EXEC)
2308 vm_map_mergeable_neighbors(vm_map_entry_t prev, vm_map_entry_t entry)
2311 KASSERT((prev->eflags & MAP_ENTRY_NOMERGE_MASK) == 0 ||
2312 (entry->eflags & MAP_ENTRY_NOMERGE_MASK) == 0,
2313 ("vm_map_mergeable_neighbors: neither %p nor %p are mergeable",
2315 return (prev->end == entry->start &&
2316 prev->object.vm_object == entry->object.vm_object &&
2317 (prev->object.vm_object == NULL ||
2318 prev->offset + (prev->end - prev->start) == entry->offset) &&
2319 prev->eflags == entry->eflags &&
2320 prev->protection == entry->protection &&
2321 prev->max_protection == entry->max_protection &&
2322 prev->inheritance == entry->inheritance &&
2323 prev->wired_count == entry->wired_count &&
2324 prev->cred == entry->cred);
2328 vm_map_merged_neighbor_dispose(vm_map_t map, vm_map_entry_t entry)
2332 * If the backing object is a vnode object, vm_object_deallocate()
2333 * calls vrele(). However, vrele() does not lock the vnode because
2334 * the vnode has additional references. Thus, the map lock can be
2335 * kept without causing a lock-order reversal with the vnode lock.
2337 * Since we count the number of virtual page mappings in
2338 * object->un_pager.vnp.writemappings, the writemappings value
2339 * should not be adjusted when the entry is disposed of.
2341 if (entry->object.vm_object != NULL)
2342 vm_object_deallocate(entry->object.vm_object);
2343 if (entry->cred != NULL)
2344 crfree(entry->cred);
2345 vm_map_entry_dispose(map, entry);
2349 * vm_map_try_merge_entries:
2351 * Compare the given map entry to its predecessor, and merge its precessor
2352 * into it if possible. The entry remains valid, and may be extended.
2353 * The predecessor may be deleted.
2355 * The map must be locked.
2358 vm_map_try_merge_entries(vm_map_t map, vm_map_entry_t prev_entry,
2359 vm_map_entry_t entry)
2362 VM_MAP_ASSERT_LOCKED(map);
2363 if ((entry->eflags & MAP_ENTRY_NOMERGE_MASK) == 0 &&
2364 vm_map_mergeable_neighbors(prev_entry, entry)) {
2365 vm_map_entry_unlink(map, prev_entry, UNLINK_MERGE_NEXT);
2366 vm_map_merged_neighbor_dispose(map, prev_entry);
2371 * vm_map_entry_back:
2373 * Allocate an object to back a map entry.
2376 vm_map_entry_back(vm_map_entry_t entry)
2380 KASSERT(entry->object.vm_object == NULL,
2381 ("map entry %p has backing object", entry));
2382 KASSERT((entry->eflags & MAP_ENTRY_IS_SUB_MAP) == 0,
2383 ("map entry %p is a submap", entry));
2384 object = vm_object_allocate_anon(atop(entry->end - entry->start), NULL,
2385 entry->cred, entry->end - entry->start);
2386 entry->object.vm_object = object;
2392 * vm_map_entry_charge_object
2394 * If there is no object backing this entry, create one. Otherwise, if
2395 * the entry has cred, give it to the backing object.
2398 vm_map_entry_charge_object(vm_map_t map, vm_map_entry_t entry)
2401 VM_MAP_ASSERT_LOCKED(map);
2402 KASSERT((entry->eflags & MAP_ENTRY_IS_SUB_MAP) == 0,
2403 ("map entry %p is a submap", entry));
2404 if (entry->object.vm_object == NULL && !map->system_map &&
2405 (entry->eflags & MAP_ENTRY_GUARD) == 0)
2406 vm_map_entry_back(entry);
2407 else if (entry->object.vm_object != NULL &&
2408 ((entry->eflags & MAP_ENTRY_NEEDS_COPY) == 0) &&
2409 entry->cred != NULL) {
2410 VM_OBJECT_WLOCK(entry->object.vm_object);
2411 KASSERT(entry->object.vm_object->cred == NULL,
2412 ("OVERCOMMIT: %s: both cred e %p", __func__, entry));
2413 entry->object.vm_object->cred = entry->cred;
2414 entry->object.vm_object->charge = entry->end - entry->start;
2415 VM_OBJECT_WUNLOCK(entry->object.vm_object);
2421 * vm_map_entry_clone
2423 * Create a duplicate map entry for clipping.
2425 static vm_map_entry_t
2426 vm_map_entry_clone(vm_map_t map, vm_map_entry_t entry)
2428 vm_map_entry_t new_entry;
2430 VM_MAP_ASSERT_LOCKED(map);
2433 * Create a backing object now, if none exists, so that more individual
2434 * objects won't be created after the map entry is split.
2436 vm_map_entry_charge_object(map, entry);
2438 /* Clone the entry. */
2439 new_entry = vm_map_entry_create(map);
2440 *new_entry = *entry;
2441 if (new_entry->cred != NULL)
2442 crhold(entry->cred);
2443 if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
2444 vm_object_reference(new_entry->object.vm_object);
2445 vm_map_entry_set_vnode_text(new_entry, true);
2447 * The object->un_pager.vnp.writemappings for the object of
2448 * MAP_ENTRY_WRITECNT type entry shall be kept as is here. The
2449 * virtual pages are re-distributed among the clipped entries,
2450 * so the sum is left the same.
2457 * vm_map_clip_start: [ internal use only ]
2459 * Asserts that the given entry begins at or after
2460 * the specified address; if necessary,
2461 * it splits the entry into two.
2464 vm_map_clip_start(vm_map_t map, vm_map_entry_t entry, vm_offset_t startaddr)
2466 vm_map_entry_t new_entry;
2469 if (!map->system_map)
2470 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
2471 "%s: map %p entry %p start 0x%jx", __func__, map, entry,
2472 (uintmax_t)startaddr);
2474 if (startaddr <= entry->start)
2475 return (KERN_SUCCESS);
2477 VM_MAP_ASSERT_LOCKED(map);
2478 KASSERT(entry->end > startaddr && entry->start < startaddr,
2479 ("%s: invalid clip of entry %p", __func__, entry));
2481 bdry_idx = (entry->eflags & MAP_ENTRY_SPLIT_BOUNDARY_MASK) >>
2482 MAP_ENTRY_SPLIT_BOUNDARY_SHIFT;
2483 if (bdry_idx != 0) {
2484 if ((startaddr & (pagesizes[bdry_idx] - 1)) != 0)
2485 return (KERN_INVALID_ARGUMENT);
2488 new_entry = vm_map_entry_clone(map, entry);
2491 * Split off the front portion. Insert the new entry BEFORE this one,
2492 * so that this entry has the specified starting address.
2494 new_entry->end = startaddr;
2495 vm_map_entry_link(map, new_entry);
2496 return (KERN_SUCCESS);
2500 * vm_map_lookup_clip_start:
2502 * Find the entry at or just after 'start', and clip it if 'start' is in
2503 * the interior of the entry. Return entry after 'start', and in
2504 * prev_entry set the entry before 'start'.
2507 vm_map_lookup_clip_start(vm_map_t map, vm_offset_t start,
2508 vm_map_entry_t *res_entry, vm_map_entry_t *prev_entry)
2510 vm_map_entry_t entry;
2513 if (!map->system_map)
2514 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
2515 "%s: map %p start 0x%jx prev %p", __func__, map,
2516 (uintmax_t)start, prev_entry);
2518 if (vm_map_lookup_entry(map, start, prev_entry)) {
2519 entry = *prev_entry;
2520 rv = vm_map_clip_start(map, entry, start);
2521 if (rv != KERN_SUCCESS)
2523 *prev_entry = vm_map_entry_pred(entry);
2525 entry = vm_map_entry_succ(*prev_entry);
2527 return (KERN_SUCCESS);
2531 * vm_map_clip_end: [ internal use only ]
2533 * Asserts that the given entry ends at or before
2534 * the specified address; if necessary,
2535 * it splits the entry into two.
2538 vm_map_clip_end(vm_map_t map, vm_map_entry_t entry, vm_offset_t endaddr)
2540 vm_map_entry_t new_entry;
2543 if (!map->system_map)
2544 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
2545 "%s: map %p entry %p end 0x%jx", __func__, map, entry,
2546 (uintmax_t)endaddr);
2548 if (endaddr >= entry->end)
2549 return (KERN_SUCCESS);
2551 VM_MAP_ASSERT_LOCKED(map);
2552 KASSERT(entry->start < endaddr && entry->end > endaddr,
2553 ("%s: invalid clip of entry %p", __func__, entry));
2555 bdry_idx = (entry->eflags & MAP_ENTRY_SPLIT_BOUNDARY_MASK) >>
2556 MAP_ENTRY_SPLIT_BOUNDARY_SHIFT;
2557 if (bdry_idx != 0) {
2558 if ((endaddr & (pagesizes[bdry_idx] - 1)) != 0)
2559 return (KERN_INVALID_ARGUMENT);
2562 new_entry = vm_map_entry_clone(map, entry);
2565 * Split off the back portion. Insert the new entry AFTER this one,
2566 * so that this entry has the specified ending address.
2568 new_entry->start = endaddr;
2569 vm_map_entry_link(map, new_entry);
2571 return (KERN_SUCCESS);
2575 * vm_map_submap: [ kernel use only ]
2577 * Mark the given range as handled by a subordinate map.
2579 * This range must have been created with vm_map_find,
2580 * and no other operations may have been performed on this
2581 * range prior to calling vm_map_submap.
2583 * Only a limited number of operations can be performed
2584 * within this rage after calling vm_map_submap:
2586 * [Don't try vm_map_copy!]
2588 * To remove a submapping, one must first remove the
2589 * range from the superior map, and then destroy the
2590 * submap (if desired). [Better yet, don't try it.]
2599 vm_map_entry_t entry;
2602 result = KERN_INVALID_ARGUMENT;
2604 vm_map_lock(submap);
2605 submap->flags |= MAP_IS_SUB_MAP;
2606 vm_map_unlock(submap);
2609 VM_MAP_RANGE_CHECK(map, start, end);
2610 if (vm_map_lookup_entry(map, start, &entry) && entry->end >= end &&
2611 (entry->eflags & MAP_ENTRY_COW) == 0 &&
2612 entry->object.vm_object == NULL) {
2613 result = vm_map_clip_start(map, entry, start);
2614 if (result != KERN_SUCCESS)
2616 result = vm_map_clip_end(map, entry, end);
2617 if (result != KERN_SUCCESS)
2619 entry->object.sub_map = submap;
2620 entry->eflags |= MAP_ENTRY_IS_SUB_MAP;
2621 result = KERN_SUCCESS;
2626 if (result != KERN_SUCCESS) {
2627 vm_map_lock(submap);
2628 submap->flags &= ~MAP_IS_SUB_MAP;
2629 vm_map_unlock(submap);
2635 * The maximum number of pages to map if MAP_PREFAULT_PARTIAL is specified
2637 #define MAX_INIT_PT 96
2640 * vm_map_pmap_enter:
2642 * Preload the specified map's pmap with mappings to the specified
2643 * object's memory-resident pages. No further physical pages are
2644 * allocated, and no further virtual pages are retrieved from secondary
2645 * storage. If the specified flags include MAP_PREFAULT_PARTIAL, then a
2646 * limited number of page mappings are created at the low-end of the
2647 * specified address range. (For this purpose, a superpage mapping
2648 * counts as one page mapping.) Otherwise, all resident pages within
2649 * the specified address range are mapped.
2652 vm_map_pmap_enter(vm_map_t map, vm_offset_t addr, vm_prot_t prot,
2653 vm_object_t object, vm_pindex_t pindex, vm_size_t size, int flags)
2656 vm_page_t p, p_start;
2657 vm_pindex_t mask, psize, threshold, tmpidx;
2659 if ((prot & (VM_PROT_READ | VM_PROT_EXECUTE)) == 0 || object == NULL)
2661 if (object->type == OBJT_DEVICE || object->type == OBJT_SG) {
2662 VM_OBJECT_WLOCK(object);
2663 if (object->type == OBJT_DEVICE || object->type == OBJT_SG) {
2664 pmap_object_init_pt(map->pmap, addr, object, pindex,
2666 VM_OBJECT_WUNLOCK(object);
2669 VM_OBJECT_LOCK_DOWNGRADE(object);
2671 VM_OBJECT_RLOCK(object);
2674 if (psize + pindex > object->size) {
2675 if (pindex >= object->size) {
2676 VM_OBJECT_RUNLOCK(object);
2679 psize = object->size - pindex;
2684 threshold = MAX_INIT_PT;
2686 p = vm_page_find_least(object, pindex);
2688 * Assert: the variable p is either (1) the page with the
2689 * least pindex greater than or equal to the parameter pindex
2693 p != NULL && (tmpidx = p->pindex - pindex) < psize;
2694 p = TAILQ_NEXT(p, listq)) {
2696 * don't allow an madvise to blow away our really
2697 * free pages allocating pv entries.
2699 if (((flags & MAP_PREFAULT_MADVISE) != 0 &&
2700 vm_page_count_severe()) ||
2701 ((flags & MAP_PREFAULT_PARTIAL) != 0 &&
2702 tmpidx >= threshold)) {
2706 if (vm_page_all_valid(p)) {
2707 if (p_start == NULL) {
2708 start = addr + ptoa(tmpidx);
2711 /* Jump ahead if a superpage mapping is possible. */
2712 if (p->psind > 0 && ((addr + ptoa(tmpidx)) &
2713 (pagesizes[p->psind] - 1)) == 0) {
2714 mask = atop(pagesizes[p->psind]) - 1;
2715 if (tmpidx + mask < psize &&
2716 vm_page_ps_test(p, PS_ALL_VALID, NULL)) {
2721 } else if (p_start != NULL) {
2722 pmap_enter_object(map->pmap, start, addr +
2723 ptoa(tmpidx), p_start, prot);
2727 if (p_start != NULL)
2728 pmap_enter_object(map->pmap, start, addr + ptoa(psize),
2730 VM_OBJECT_RUNLOCK(object);
2736 * Sets the protection of the specified address
2737 * region in the target map. If "set_max" is
2738 * specified, the maximum protection is to be set;
2739 * otherwise, only the current protection is affected.
2742 vm_map_protect(vm_map_t map, vm_offset_t start, vm_offset_t end,
2743 vm_prot_t new_prot, boolean_t set_max)
2745 vm_map_entry_t entry, first_entry, in_tran, prev_entry;
2752 return (KERN_SUCCESS);
2758 if ((map->flags & MAP_WXORX) != 0 && (new_prot &
2759 (VM_PROT_WRITE | VM_PROT_EXECUTE)) == (VM_PROT_WRITE |
2762 return (KERN_PROTECTION_FAILURE);
2766 * Ensure that we are not concurrently wiring pages. vm_map_wire() may
2767 * need to fault pages into the map and will drop the map lock while
2768 * doing so, and the VM object may end up in an inconsistent state if we
2769 * update the protection on the map entry in between faults.
2771 vm_map_wait_busy(map);
2773 VM_MAP_RANGE_CHECK(map, start, end);
2775 if (!vm_map_lookup_entry(map, start, &first_entry))
2776 first_entry = vm_map_entry_succ(first_entry);
2779 * Make a first pass to check for protection violations.
2781 for (entry = first_entry; entry->start < end;
2782 entry = vm_map_entry_succ(entry)) {
2783 if ((entry->eflags & MAP_ENTRY_GUARD) != 0)
2785 if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) != 0) {
2787 return (KERN_INVALID_ARGUMENT);
2789 if ((new_prot & entry->max_protection) != new_prot) {
2791 return (KERN_PROTECTION_FAILURE);
2793 if ((entry->eflags & MAP_ENTRY_IN_TRANSITION) != 0)
2798 * Postpone the operation until all in-transition map entries have
2799 * stabilized. An in-transition entry might already have its pages
2800 * wired and wired_count incremented, but not yet have its
2801 * MAP_ENTRY_USER_WIRED flag set. In which case, we would fail to call
2802 * vm_fault_copy_entry() in the final loop below.
2804 if (in_tran != NULL) {
2805 in_tran->eflags |= MAP_ENTRY_NEEDS_WAKEUP;
2806 vm_map_unlock_and_wait(map, 0);
2811 * Before changing the protections, try to reserve swap space for any
2812 * private (i.e., copy-on-write) mappings that are transitioning from
2813 * read-only to read/write access. If a reservation fails, break out
2814 * of this loop early and let the next loop simplify the entries, since
2815 * some may now be mergeable.
2817 rv = vm_map_clip_start(map, first_entry, start);
2818 if (rv != KERN_SUCCESS) {
2822 for (entry = first_entry; entry->start < end;
2823 entry = vm_map_entry_succ(entry)) {
2824 rv = vm_map_clip_end(map, entry, end);
2825 if (rv != KERN_SUCCESS) {
2831 ((new_prot & ~entry->protection) & VM_PROT_WRITE) == 0 ||
2832 ENTRY_CHARGED(entry) ||
2833 (entry->eflags & MAP_ENTRY_GUARD) != 0) {
2837 cred = curthread->td_ucred;
2838 obj = entry->object.vm_object;
2841 (entry->eflags & MAP_ENTRY_NEEDS_COPY) != 0) {
2842 if (!swap_reserve(entry->end - entry->start)) {
2843 rv = KERN_RESOURCE_SHORTAGE;
2852 if (obj->type != OBJT_DEFAULT && obj->type != OBJT_SWAP)
2854 VM_OBJECT_WLOCK(obj);
2855 if (obj->type != OBJT_DEFAULT && obj->type != OBJT_SWAP) {
2856 VM_OBJECT_WUNLOCK(obj);
2861 * Charge for the whole object allocation now, since
2862 * we cannot distinguish between non-charged and
2863 * charged clipped mapping of the same object later.
2865 KASSERT(obj->charge == 0,
2866 ("vm_map_protect: object %p overcharged (entry %p)",
2868 if (!swap_reserve(ptoa(obj->size))) {
2869 VM_OBJECT_WUNLOCK(obj);
2870 rv = KERN_RESOURCE_SHORTAGE;
2877 obj->charge = ptoa(obj->size);
2878 VM_OBJECT_WUNLOCK(obj);
2882 * If enough swap space was available, go back and fix up protections.
2883 * Otherwise, just simplify entries, since some may have been modified.
2884 * [Note that clipping is not necessary the second time.]
2886 for (prev_entry = vm_map_entry_pred(first_entry), entry = first_entry;
2888 vm_map_try_merge_entries(map, prev_entry, entry),
2889 prev_entry = entry, entry = vm_map_entry_succ(entry)) {
2890 if (rv != KERN_SUCCESS ||
2891 (entry->eflags & MAP_ENTRY_GUARD) != 0)
2894 old_prot = entry->protection;
2898 (entry->max_protection = new_prot) &
2901 entry->protection = new_prot;
2904 * For user wired map entries, the normal lazy evaluation of
2905 * write access upgrades through soft page faults is
2906 * undesirable. Instead, immediately copy any pages that are
2907 * copy-on-write and enable write access in the physical map.
2909 if ((entry->eflags & MAP_ENTRY_USER_WIRED) != 0 &&
2910 (entry->protection & VM_PROT_WRITE) != 0 &&
2911 (old_prot & VM_PROT_WRITE) == 0)
2912 vm_fault_copy_entry(map, map, entry, entry, NULL);
2915 * When restricting access, update the physical map. Worry
2916 * about copy-on-write here.
2918 if ((old_prot & ~entry->protection) != 0) {
2919 #define MASK(entry) (((entry)->eflags & MAP_ENTRY_COW) ? ~VM_PROT_WRITE : \
2921 pmap_protect(map->pmap, entry->start,
2923 entry->protection & MASK(entry));
2927 vm_map_try_merge_entries(map, prev_entry, entry);
2935 * This routine traverses a processes map handling the madvise
2936 * system call. Advisories are classified as either those effecting
2937 * the vm_map_entry structure, or those effecting the underlying
2947 vm_map_entry_t entry, prev_entry;
2952 * Some madvise calls directly modify the vm_map_entry, in which case
2953 * we need to use an exclusive lock on the map and we need to perform
2954 * various clipping operations. Otherwise we only need a read-lock
2959 case MADV_SEQUENTIAL:
2976 vm_map_lock_read(map);
2983 * Locate starting entry and clip if necessary.
2985 VM_MAP_RANGE_CHECK(map, start, end);
2989 * madvise behaviors that are implemented in the vm_map_entry.
2991 * We clip the vm_map_entry so that behavioral changes are
2992 * limited to the specified address range.
2994 rv = vm_map_lookup_clip_start(map, start, &entry, &prev_entry);
2995 if (rv != KERN_SUCCESS) {
2997 return (vm_mmap_to_errno(rv));
3000 for (; entry->start < end; prev_entry = entry,
3001 entry = vm_map_entry_succ(entry)) {
3002 if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) != 0)
3005 rv = vm_map_clip_end(map, entry, end);
3006 if (rv != KERN_SUCCESS) {
3008 return (vm_mmap_to_errno(rv));
3013 vm_map_entry_set_behavior(entry,
3014 MAP_ENTRY_BEHAV_NORMAL);
3016 case MADV_SEQUENTIAL:
3017 vm_map_entry_set_behavior(entry,
3018 MAP_ENTRY_BEHAV_SEQUENTIAL);
3021 vm_map_entry_set_behavior(entry,
3022 MAP_ENTRY_BEHAV_RANDOM);
3025 entry->eflags |= MAP_ENTRY_NOSYNC;
3028 entry->eflags &= ~MAP_ENTRY_NOSYNC;
3031 entry->eflags |= MAP_ENTRY_NOCOREDUMP;
3034 entry->eflags &= ~MAP_ENTRY_NOCOREDUMP;
3039 vm_map_try_merge_entries(map, prev_entry, entry);
3041 vm_map_try_merge_entries(map, prev_entry, entry);
3044 vm_pindex_t pstart, pend;
3047 * madvise behaviors that are implemented in the underlying
3050 * Since we don't clip the vm_map_entry, we have to clip
3051 * the vm_object pindex and count.
3053 if (!vm_map_lookup_entry(map, start, &entry))
3054 entry = vm_map_entry_succ(entry);
3055 for (; entry->start < end;
3056 entry = vm_map_entry_succ(entry)) {
3057 vm_offset_t useEnd, useStart;
3059 if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) != 0)
3063 * MADV_FREE would otherwise rewind time to
3064 * the creation of the shadow object. Because
3065 * we hold the VM map read-locked, neither the
3066 * entry's object nor the presence of a
3067 * backing object can change.
3069 if (behav == MADV_FREE &&
3070 entry->object.vm_object != NULL &&
3071 entry->object.vm_object->backing_object != NULL)
3074 pstart = OFF_TO_IDX(entry->offset);
3075 pend = pstart + atop(entry->end - entry->start);
3076 useStart = entry->start;
3077 useEnd = entry->end;
3079 if (entry->start < start) {
3080 pstart += atop(start - entry->start);
3083 if (entry->end > end) {
3084 pend -= atop(entry->end - end);
3092 * Perform the pmap_advise() before clearing
3093 * PGA_REFERENCED in vm_page_advise(). Otherwise, a
3094 * concurrent pmap operation, such as pmap_remove(),
3095 * could clear a reference in the pmap and set
3096 * PGA_REFERENCED on the page before the pmap_advise()
3097 * had completed. Consequently, the page would appear
3098 * referenced based upon an old reference that
3099 * occurred before this pmap_advise() ran.
3101 if (behav == MADV_DONTNEED || behav == MADV_FREE)
3102 pmap_advise(map->pmap, useStart, useEnd,
3105 vm_object_madvise(entry->object.vm_object, pstart,
3109 * Pre-populate paging structures in the
3110 * WILLNEED case. For wired entries, the
3111 * paging structures are already populated.
3113 if (behav == MADV_WILLNEED &&
3114 entry->wired_count == 0) {
3115 vm_map_pmap_enter(map,
3118 entry->object.vm_object,
3120 ptoa(pend - pstart),
3121 MAP_PREFAULT_MADVISE
3125 vm_map_unlock_read(map);
3133 * Sets the inheritance of the specified address
3134 * range in the target map. Inheritance
3135 * affects how the map will be shared with
3136 * child maps at the time of vmspace_fork.
3139 vm_map_inherit(vm_map_t map, vm_offset_t start, vm_offset_t end,
3140 vm_inherit_t new_inheritance)
3142 vm_map_entry_t entry, lentry, prev_entry, start_entry;
3145 switch (new_inheritance) {
3146 case VM_INHERIT_NONE:
3147 case VM_INHERIT_COPY:
3148 case VM_INHERIT_SHARE:
3149 case VM_INHERIT_ZERO:
3152 return (KERN_INVALID_ARGUMENT);
3155 return (KERN_SUCCESS);
3157 VM_MAP_RANGE_CHECK(map, start, end);
3158 rv = vm_map_lookup_clip_start(map, start, &start_entry, &prev_entry);
3159 if (rv != KERN_SUCCESS)
3161 if (vm_map_lookup_entry(map, end - 1, &lentry)) {
3162 rv = vm_map_clip_end(map, lentry, end);
3163 if (rv != KERN_SUCCESS)
3166 if (new_inheritance == VM_INHERIT_COPY) {
3167 for (entry = start_entry; entry->start < end;
3168 prev_entry = entry, entry = vm_map_entry_succ(entry)) {
3169 if ((entry->eflags & MAP_ENTRY_SPLIT_BOUNDARY_MASK)
3171 rv = KERN_INVALID_ARGUMENT;
3176 for (entry = start_entry; entry->start < end; prev_entry = entry,
3177 entry = vm_map_entry_succ(entry)) {
3178 KASSERT(entry->end <= end, ("non-clipped entry %p end %jx %jx",
3179 entry, (uintmax_t)entry->end, (uintmax_t)end));
3180 if ((entry->eflags & MAP_ENTRY_GUARD) == 0 ||
3181 new_inheritance != VM_INHERIT_ZERO)
3182 entry->inheritance = new_inheritance;
3183 vm_map_try_merge_entries(map, prev_entry, entry);
3185 vm_map_try_merge_entries(map, prev_entry, entry);
3192 * vm_map_entry_in_transition:
3194 * Release the map lock, and sleep until the entry is no longer in
3195 * transition. Awake and acquire the map lock. If the map changed while
3196 * another held the lock, lookup a possibly-changed entry at or after the
3197 * 'start' position of the old entry.
3199 static vm_map_entry_t
3200 vm_map_entry_in_transition(vm_map_t map, vm_offset_t in_start,
3201 vm_offset_t *io_end, bool holes_ok, vm_map_entry_t in_entry)
3203 vm_map_entry_t entry;
3205 u_int last_timestamp;
3207 VM_MAP_ASSERT_LOCKED(map);
3208 KASSERT((in_entry->eflags & MAP_ENTRY_IN_TRANSITION) != 0,
3209 ("not in-tranition map entry %p", in_entry));
3211 * We have not yet clipped the entry.
3213 start = MAX(in_start, in_entry->start);
3214 in_entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP;
3215 last_timestamp = map->timestamp;
3216 if (vm_map_unlock_and_wait(map, 0)) {
3218 * Allow interruption of user wiring/unwiring?
3222 if (last_timestamp + 1 == map->timestamp)
3226 * Look again for the entry because the map was modified while it was
3227 * unlocked. Specifically, the entry may have been clipped, merged, or
3230 if (!vm_map_lookup_entry(map, start, &entry)) {
3235 entry = vm_map_entry_succ(entry);
3243 * Implements both kernel and user unwiring.
3246 vm_map_unwire(vm_map_t map, vm_offset_t start, vm_offset_t end,
3249 vm_map_entry_t entry, first_entry, next_entry, prev_entry;
3251 bool holes_ok, need_wakeup, user_unwire;
3254 return (KERN_SUCCESS);
3255 holes_ok = (flags & VM_MAP_WIRE_HOLESOK) != 0;
3256 user_unwire = (flags & VM_MAP_WIRE_USER) != 0;
3258 VM_MAP_RANGE_CHECK(map, start, end);
3259 if (!vm_map_lookup_entry(map, start, &first_entry)) {
3261 first_entry = vm_map_entry_succ(first_entry);
3264 return (KERN_INVALID_ADDRESS);
3268 for (entry = first_entry; entry->start < end; entry = next_entry) {
3269 if (entry->eflags & MAP_ENTRY_IN_TRANSITION) {
3271 * We have not yet clipped the entry.
3273 next_entry = vm_map_entry_in_transition(map, start,
3274 &end, holes_ok, entry);
3275 if (next_entry == NULL) {
3276 if (entry == first_entry) {
3278 return (KERN_INVALID_ADDRESS);
3280 rv = KERN_INVALID_ADDRESS;
3283 first_entry = (entry == first_entry) ?
3287 rv = vm_map_clip_start(map, entry, start);
3288 if (rv != KERN_SUCCESS)
3290 rv = vm_map_clip_end(map, entry, end);
3291 if (rv != KERN_SUCCESS)
3295 * Mark the entry in case the map lock is released. (See
3298 KASSERT((entry->eflags & MAP_ENTRY_IN_TRANSITION) == 0 &&
3299 entry->wiring_thread == NULL,
3300 ("owned map entry %p", entry));
3301 entry->eflags |= MAP_ENTRY_IN_TRANSITION;
3302 entry->wiring_thread = curthread;
3303 next_entry = vm_map_entry_succ(entry);
3305 * Check the map for holes in the specified region.
3306 * If holes_ok, skip this check.
3309 entry->end < end && next_entry->start > entry->end) {
3311 rv = KERN_INVALID_ADDRESS;
3315 * If system unwiring, require that the entry is system wired.
3318 vm_map_entry_system_wired_count(entry) == 0) {
3320 rv = KERN_INVALID_ARGUMENT;
3324 need_wakeup = false;
3325 if (first_entry == NULL &&
3326 !vm_map_lookup_entry(map, start, &first_entry)) {
3327 KASSERT(holes_ok, ("vm_map_unwire: lookup failed"));
3328 prev_entry = first_entry;
3329 entry = vm_map_entry_succ(first_entry);
3331 prev_entry = vm_map_entry_pred(first_entry);
3332 entry = first_entry;
3334 for (; entry->start < end;
3335 prev_entry = entry, entry = vm_map_entry_succ(entry)) {
3337 * If holes_ok was specified, an empty
3338 * space in the unwired region could have been mapped
3339 * while the map lock was dropped for draining
3340 * MAP_ENTRY_IN_TRANSITION. Moreover, another thread
3341 * could be simultaneously wiring this new mapping
3342 * entry. Detect these cases and skip any entries
3343 * marked as in transition by us.
3345 if ((entry->eflags & MAP_ENTRY_IN_TRANSITION) == 0 ||
3346 entry->wiring_thread != curthread) {
3348 ("vm_map_unwire: !HOLESOK and new/changed entry"));
3352 if (rv == KERN_SUCCESS && (!user_unwire ||
3353 (entry->eflags & MAP_ENTRY_USER_WIRED))) {
3354 if (entry->wired_count == 1)
3355 vm_map_entry_unwire(map, entry);
3357 entry->wired_count--;
3359 entry->eflags &= ~MAP_ENTRY_USER_WIRED;
3361 KASSERT((entry->eflags & MAP_ENTRY_IN_TRANSITION) != 0,
3362 ("vm_map_unwire: in-transition flag missing %p", entry));
3363 KASSERT(entry->wiring_thread == curthread,
3364 ("vm_map_unwire: alien wire %p", entry));
3365 entry->eflags &= ~MAP_ENTRY_IN_TRANSITION;
3366 entry->wiring_thread = NULL;
3367 if (entry->eflags & MAP_ENTRY_NEEDS_WAKEUP) {
3368 entry->eflags &= ~MAP_ENTRY_NEEDS_WAKEUP;
3371 vm_map_try_merge_entries(map, prev_entry, entry);
3373 vm_map_try_merge_entries(map, prev_entry, entry);
3381 vm_map_wire_user_count_sub(u_long npages)
3384 atomic_subtract_long(&vm_user_wire_count, npages);
3388 vm_map_wire_user_count_add(u_long npages)
3392 wired = vm_user_wire_count;
3394 if (npages + wired > vm_page_max_user_wired)
3396 } while (!atomic_fcmpset_long(&vm_user_wire_count, &wired,
3403 * vm_map_wire_entry_failure:
3405 * Handle a wiring failure on the given entry.
3407 * The map should be locked.
3410 vm_map_wire_entry_failure(vm_map_t map, vm_map_entry_t entry,
3411 vm_offset_t failed_addr)
3414 VM_MAP_ASSERT_LOCKED(map);
3415 KASSERT((entry->eflags & MAP_ENTRY_IN_TRANSITION) != 0 &&
3416 entry->wired_count == 1,
3417 ("vm_map_wire_entry_failure: entry %p isn't being wired", entry));
3418 KASSERT(failed_addr < entry->end,
3419 ("vm_map_wire_entry_failure: entry %p was fully wired", entry));
3422 * If any pages at the start of this entry were successfully wired,
3425 if (failed_addr > entry->start) {
3426 pmap_unwire(map->pmap, entry->start, failed_addr);
3427 vm_object_unwire(entry->object.vm_object, entry->offset,
3428 failed_addr - entry->start, PQ_ACTIVE);
3432 * Assign an out-of-range value to represent the failure to wire this
3435 entry->wired_count = -1;
3439 vm_map_wire(vm_map_t map, vm_offset_t start, vm_offset_t end, int flags)
3444 rv = vm_map_wire_locked(map, start, end, flags);
3450 * vm_map_wire_locked:
3452 * Implements both kernel and user wiring. Returns with the map locked,
3453 * the map lock may be dropped.
3456 vm_map_wire_locked(vm_map_t map, vm_offset_t start, vm_offset_t end, int flags)
3458 vm_map_entry_t entry, first_entry, next_entry, prev_entry;
3459 vm_offset_t faddr, saved_end, saved_start;
3460 u_long incr, npages;
3461 u_int bidx, last_timestamp;
3463 bool holes_ok, need_wakeup, user_wire;
3466 VM_MAP_ASSERT_LOCKED(map);
3469 return (KERN_SUCCESS);
3471 if (flags & VM_MAP_WIRE_WRITE)
3472 prot |= VM_PROT_WRITE;
3473 holes_ok = (flags & VM_MAP_WIRE_HOLESOK) != 0;
3474 user_wire = (flags & VM_MAP_WIRE_USER) != 0;
3475 VM_MAP_RANGE_CHECK(map, start, end);
3476 if (!vm_map_lookup_entry(map, start, &first_entry)) {
3478 first_entry = vm_map_entry_succ(first_entry);
3480 return (KERN_INVALID_ADDRESS);
3482 for (entry = first_entry; entry->start < end; entry = next_entry) {
3483 if (entry->eflags & MAP_ENTRY_IN_TRANSITION) {
3485 * We have not yet clipped the entry.
3487 next_entry = vm_map_entry_in_transition(map, start,
3488 &end, holes_ok, entry);
3489 if (next_entry == NULL) {
3490 if (entry == first_entry)
3491 return (KERN_INVALID_ADDRESS);
3492 rv = KERN_INVALID_ADDRESS;
3495 first_entry = (entry == first_entry) ?
3499 rv = vm_map_clip_start(map, entry, start);
3500 if (rv != KERN_SUCCESS)
3502 rv = vm_map_clip_end(map, entry, end);
3503 if (rv != KERN_SUCCESS)
3507 * Mark the entry in case the map lock is released. (See
3510 KASSERT((entry->eflags & MAP_ENTRY_IN_TRANSITION) == 0 &&
3511 entry->wiring_thread == NULL,
3512 ("owned map entry %p", entry));
3513 entry->eflags |= MAP_ENTRY_IN_TRANSITION;
3514 entry->wiring_thread = curthread;
3515 if ((entry->protection & (VM_PROT_READ | VM_PROT_EXECUTE)) == 0
3516 || (entry->protection & prot) != prot) {
3517 entry->eflags |= MAP_ENTRY_WIRE_SKIPPED;
3520 rv = KERN_INVALID_ADDRESS;
3523 } else if (entry->wired_count == 0) {
3524 entry->wired_count++;
3526 npages = atop(entry->end - entry->start);
3527 if (user_wire && !vm_map_wire_user_count_add(npages)) {
3528 vm_map_wire_entry_failure(map, entry,
3531 rv = KERN_RESOURCE_SHORTAGE;
3536 * Release the map lock, relying on the in-transition
3537 * mark. Mark the map busy for fork.
3539 saved_start = entry->start;
3540 saved_end = entry->end;
3541 last_timestamp = map->timestamp;
3542 bidx = (entry->eflags & MAP_ENTRY_SPLIT_BOUNDARY_MASK)
3543 >> MAP_ENTRY_SPLIT_BOUNDARY_SHIFT;
3544 incr = pagesizes[bidx];
3548 for (faddr = saved_start; faddr < saved_end;
3551 * Simulate a fault to get the page and enter
3552 * it into the physical map.
3554 rv = vm_fault(map, faddr, VM_PROT_NONE,
3555 VM_FAULT_WIRE, NULL);
3556 if (rv != KERN_SUCCESS)
3561 if (last_timestamp + 1 != map->timestamp) {
3563 * Look again for the entry because the map was
3564 * modified while it was unlocked. The entry
3565 * may have been clipped, but NOT merged or
3568 if (!vm_map_lookup_entry(map, saved_start,
3571 ("vm_map_wire: lookup failed"));
3572 first_entry = (entry == first_entry) ?
3574 for (entry = next_entry; entry->end < saved_end;
3575 entry = vm_map_entry_succ(entry)) {
3577 * In case of failure, handle entries
3578 * that were not fully wired here;
3579 * fully wired entries are handled
3582 if (rv != KERN_SUCCESS &&
3584 vm_map_wire_entry_failure(map,
3588 if (rv != KERN_SUCCESS) {
3589 vm_map_wire_entry_failure(map, entry, faddr);
3591 vm_map_wire_user_count_sub(npages);
3595 } else if (!user_wire ||
3596 (entry->eflags & MAP_ENTRY_USER_WIRED) == 0) {
3597 entry->wired_count++;
3600 * Check the map for holes in the specified region.
3601 * If holes_ok was specified, skip this check.
3603 next_entry = vm_map_entry_succ(entry);
3605 entry->end < end && next_entry->start > entry->end) {
3607 rv = KERN_INVALID_ADDRESS;
3613 need_wakeup = false;
3614 if (first_entry == NULL &&
3615 !vm_map_lookup_entry(map, start, &first_entry)) {
3616 KASSERT(holes_ok, ("vm_map_wire: lookup failed"));
3617 prev_entry = first_entry;
3618 entry = vm_map_entry_succ(first_entry);
3620 prev_entry = vm_map_entry_pred(first_entry);
3621 entry = first_entry;
3623 for (; entry->start < end;
3624 prev_entry = entry, entry = vm_map_entry_succ(entry)) {
3626 * If holes_ok was specified, an empty
3627 * space in the unwired region could have been mapped
3628 * while the map lock was dropped for faulting in the
3629 * pages or draining MAP_ENTRY_IN_TRANSITION.
3630 * Moreover, another thread could be simultaneously
3631 * wiring this new mapping entry. Detect these cases
3632 * and skip any entries marked as in transition not by us.
3634 * Another way to get an entry not marked with
3635 * MAP_ENTRY_IN_TRANSITION is after failed clipping,
3636 * which set rv to KERN_INVALID_ARGUMENT.
3638 if ((entry->eflags & MAP_ENTRY_IN_TRANSITION) == 0 ||
3639 entry->wiring_thread != curthread) {
3640 KASSERT(holes_ok || rv == KERN_INVALID_ARGUMENT,
3641 ("vm_map_wire: !HOLESOK and new/changed entry"));
3645 if ((entry->eflags & MAP_ENTRY_WIRE_SKIPPED) != 0) {
3647 } else if (rv == KERN_SUCCESS) {
3649 entry->eflags |= MAP_ENTRY_USER_WIRED;
3650 } else if (entry->wired_count == -1) {
3652 * Wiring failed on this entry. Thus, unwiring is
3655 entry->wired_count = 0;
3656 } else if (!user_wire ||
3657 (entry->eflags & MAP_ENTRY_USER_WIRED) == 0) {
3659 * Undo the wiring. Wiring succeeded on this entry
3660 * but failed on a later entry.
3662 if (entry->wired_count == 1) {
3663 vm_map_entry_unwire(map, entry);
3665 vm_map_wire_user_count_sub(
3666 atop(entry->end - entry->start));
3668 entry->wired_count--;
3670 KASSERT((entry->eflags & MAP_ENTRY_IN_TRANSITION) != 0,
3671 ("vm_map_wire: in-transition flag missing %p", entry));
3672 KASSERT(entry->wiring_thread == curthread,
3673 ("vm_map_wire: alien wire %p", entry));
3674 entry->eflags &= ~(MAP_ENTRY_IN_TRANSITION |
3675 MAP_ENTRY_WIRE_SKIPPED);
3676 entry->wiring_thread = NULL;
3677 if (entry->eflags & MAP_ENTRY_NEEDS_WAKEUP) {
3678 entry->eflags &= ~MAP_ENTRY_NEEDS_WAKEUP;
3681 vm_map_try_merge_entries(map, prev_entry, entry);
3683 vm_map_try_merge_entries(map, prev_entry, entry);
3692 * Push any dirty cached pages in the address range to their pager.
3693 * If syncio is TRUE, dirty pages are written synchronously.
3694 * If invalidate is TRUE, any cached pages are freed as well.
3696 * If the size of the region from start to end is zero, we are
3697 * supposed to flush all modified pages within the region containing
3698 * start. Unfortunately, a region can be split or coalesced with
3699 * neighboring regions, making it difficult to determine what the
3700 * original region was. Therefore, we approximate this requirement by
3701 * flushing the current region containing start.
3703 * Returns an error if any part of the specified range is not mapped.
3711 boolean_t invalidate)
3713 vm_map_entry_t entry, first_entry, next_entry;
3716 vm_ooffset_t offset;
3717 unsigned int last_timestamp;
3721 vm_map_lock_read(map);
3722 VM_MAP_RANGE_CHECK(map, start, end);
3723 if (!vm_map_lookup_entry(map, start, &first_entry)) {
3724 vm_map_unlock_read(map);
3725 return (KERN_INVALID_ADDRESS);
3726 } else if (start == end) {
3727 start = first_entry->start;
3728 end = first_entry->end;
3732 * Make a first pass to check for user-wired memory, holes,
3733 * and partial invalidation of largepage mappings.
3735 for (entry = first_entry; entry->start < end; entry = next_entry) {
3737 if ((entry->eflags & MAP_ENTRY_USER_WIRED) != 0) {
3738 vm_map_unlock_read(map);
3739 return (KERN_INVALID_ARGUMENT);
3741 bdry_idx = (entry->eflags &
3742 MAP_ENTRY_SPLIT_BOUNDARY_MASK) >>
3743 MAP_ENTRY_SPLIT_BOUNDARY_SHIFT;
3744 if (bdry_idx != 0 &&
3745 ((start & (pagesizes[bdry_idx] - 1)) != 0 ||
3746 (end & (pagesizes[bdry_idx] - 1)) != 0)) {
3747 vm_map_unlock_read(map);
3748 return (KERN_INVALID_ARGUMENT);
3751 next_entry = vm_map_entry_succ(entry);
3752 if (end > entry->end &&
3753 entry->end != next_entry->start) {
3754 vm_map_unlock_read(map);
3755 return (KERN_INVALID_ADDRESS);
3760 pmap_remove(map->pmap, start, end);
3764 * Make a second pass, cleaning/uncaching pages from the indicated
3767 for (entry = first_entry; entry->start < end;) {
3768 offset = entry->offset + (start - entry->start);
3769 size = (end <= entry->end ? end : entry->end) - start;
3770 if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) != 0) {
3772 vm_map_entry_t tentry;
3775 smap = entry->object.sub_map;
3776 vm_map_lock_read(smap);
3777 (void) vm_map_lookup_entry(smap, offset, &tentry);
3778 tsize = tentry->end - offset;
3781 object = tentry->object.vm_object;
3782 offset = tentry->offset + (offset - tentry->start);
3783 vm_map_unlock_read(smap);
3785 object = entry->object.vm_object;
3787 vm_object_reference(object);
3788 last_timestamp = map->timestamp;
3789 vm_map_unlock_read(map);
3790 if (!vm_object_sync(object, offset, size, syncio, invalidate))
3793 vm_object_deallocate(object);
3794 vm_map_lock_read(map);
3795 if (last_timestamp == map->timestamp ||
3796 !vm_map_lookup_entry(map, start, &entry))
3797 entry = vm_map_entry_succ(entry);
3800 vm_map_unlock_read(map);
3801 return (failed ? KERN_FAILURE : KERN_SUCCESS);
3805 * vm_map_entry_unwire: [ internal use only ]
3807 * Make the region specified by this entry pageable.
3809 * The map in question should be locked.
3810 * [This is the reason for this routine's existence.]
3813 vm_map_entry_unwire(vm_map_t map, vm_map_entry_t entry)
3817 VM_MAP_ASSERT_LOCKED(map);
3818 KASSERT(entry->wired_count > 0,
3819 ("vm_map_entry_unwire: entry %p isn't wired", entry));
3821 size = entry->end - entry->start;
3822 if ((entry->eflags & MAP_ENTRY_USER_WIRED) != 0)
3823 vm_map_wire_user_count_sub(atop(size));
3824 pmap_unwire(map->pmap, entry->start, entry->end);
3825 vm_object_unwire(entry->object.vm_object, entry->offset, size,
3827 entry->wired_count = 0;
3831 vm_map_entry_deallocate(vm_map_entry_t entry, boolean_t system_map)
3834 if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) == 0)
3835 vm_object_deallocate(entry->object.vm_object);
3836 uma_zfree(system_map ? kmapentzone : mapentzone, entry);
3840 * vm_map_entry_delete: [ internal use only ]
3842 * Deallocate the given entry from the target map.
3845 vm_map_entry_delete(vm_map_t map, vm_map_entry_t entry)
3848 vm_pindex_t offidxstart, offidxend, size1;
3851 vm_map_entry_unlink(map, entry, UNLINK_MERGE_NONE);
3852 object = entry->object.vm_object;
3854 if ((entry->eflags & MAP_ENTRY_GUARD) != 0) {
3855 MPASS(entry->cred == NULL);
3856 MPASS((entry->eflags & MAP_ENTRY_IS_SUB_MAP) == 0);
3857 MPASS(object == NULL);
3858 vm_map_entry_deallocate(entry, map->system_map);
3862 size = entry->end - entry->start;
3865 if (entry->cred != NULL) {
3866 swap_release_by_cred(size, entry->cred);
3867 crfree(entry->cred);
3870 if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) != 0 || object == NULL) {
3871 entry->object.vm_object = NULL;
3872 } else if ((object->flags & OBJ_ANON) != 0 ||
3873 object == kernel_object) {
3874 KASSERT(entry->cred == NULL || object->cred == NULL ||
3875 (entry->eflags & MAP_ENTRY_NEEDS_COPY),
3876 ("OVERCOMMIT vm_map_entry_delete: both cred %p", entry));
3877 offidxstart = OFF_TO_IDX(entry->offset);
3878 offidxend = offidxstart + atop(size);
3879 VM_OBJECT_WLOCK(object);
3880 if (object->ref_count != 1 &&
3881 ((object->flags & OBJ_ONEMAPPING) != 0 ||
3882 object == kernel_object)) {
3883 vm_object_collapse(object);
3886 * The option OBJPR_NOTMAPPED can be passed here
3887 * because vm_map_delete() already performed
3888 * pmap_remove() on the only mapping to this range
3891 vm_object_page_remove(object, offidxstart, offidxend,
3893 if (offidxend >= object->size &&
3894 offidxstart < object->size) {
3895 size1 = object->size;
3896 object->size = offidxstart;
3897 if (object->cred != NULL) {
3898 size1 -= object->size;
3899 KASSERT(object->charge >= ptoa(size1),
3900 ("object %p charge < 0", object));
3901 swap_release_by_cred(ptoa(size1),
3903 object->charge -= ptoa(size1);
3907 VM_OBJECT_WUNLOCK(object);
3909 if (map->system_map)
3910 vm_map_entry_deallocate(entry, TRUE);
3912 entry->defer_next = curthread->td_map_def_user;
3913 curthread->td_map_def_user = entry;
3918 * vm_map_delete: [ internal use only ]
3920 * Deallocates the given address range from the target
3924 vm_map_delete(vm_map_t map, vm_offset_t start, vm_offset_t end)
3926 vm_map_entry_t entry, next_entry, scratch_entry;
3929 VM_MAP_ASSERT_LOCKED(map);
3932 return (KERN_SUCCESS);
3935 * Find the start of the region, and clip it.
3936 * Step through all entries in this region.
3938 rv = vm_map_lookup_clip_start(map, start, &entry, &scratch_entry);
3939 if (rv != KERN_SUCCESS)
3941 for (; entry->start < end; entry = next_entry) {
3943 * Wait for wiring or unwiring of an entry to complete.
3944 * Also wait for any system wirings to disappear on
3947 if ((entry->eflags & MAP_ENTRY_IN_TRANSITION) != 0 ||
3948 (vm_map_pmap(map) != kernel_pmap &&
3949 vm_map_entry_system_wired_count(entry) != 0)) {
3950 unsigned int last_timestamp;
3951 vm_offset_t saved_start;
3953 saved_start = entry->start;
3954 entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP;
3955 last_timestamp = map->timestamp;
3956 (void) vm_map_unlock_and_wait(map, 0);
3958 if (last_timestamp + 1 != map->timestamp) {
3960 * Look again for the entry because the map was
3961 * modified while it was unlocked.
3962 * Specifically, the entry may have been
3963 * clipped, merged, or deleted.
3965 rv = vm_map_lookup_clip_start(map, saved_start,
3966 &next_entry, &scratch_entry);
3967 if (rv != KERN_SUCCESS)
3974 /* XXXKIB or delete to the upper superpage boundary ? */
3975 rv = vm_map_clip_end(map, entry, end);
3976 if (rv != KERN_SUCCESS)
3978 next_entry = vm_map_entry_succ(entry);
3981 * Unwire before removing addresses from the pmap; otherwise,
3982 * unwiring will put the entries back in the pmap.
3984 if (entry->wired_count != 0)
3985 vm_map_entry_unwire(map, entry);
3988 * Remove mappings for the pages, but only if the
3989 * mappings could exist. For instance, it does not
3990 * make sense to call pmap_remove() for guard entries.
3992 if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) != 0 ||
3993 entry->object.vm_object != NULL)
3994 pmap_remove(map->pmap, entry->start, entry->end);
3996 if (entry->end == map->anon_loc)
3997 map->anon_loc = entry->start;
4000 * Delete the entry only after removing all pmap
4001 * entries pointing to its pages. (Otherwise, its
4002 * page frames may be reallocated, and any modify bits
4003 * will be set in the wrong object!)
4005 vm_map_entry_delete(map, entry);
4013 * Remove the given address range from the target map.
4014 * This is the exported form of vm_map_delete.
4017 vm_map_remove(vm_map_t map, vm_offset_t start, vm_offset_t end)
4022 VM_MAP_RANGE_CHECK(map, start, end);
4023 result = vm_map_delete(map, start, end);
4029 * vm_map_check_protection:
4031 * Assert that the target map allows the specified privilege on the
4032 * entire address region given. The entire region must be allocated.
4034 * WARNING! This code does not and should not check whether the
4035 * contents of the region is accessible. For example a smaller file
4036 * might be mapped into a larger address space.
4038 * NOTE! This code is also called by munmap().
4040 * The map must be locked. A read lock is sufficient.
4043 vm_map_check_protection(vm_map_t map, vm_offset_t start, vm_offset_t end,
4044 vm_prot_t protection)
4046 vm_map_entry_t entry;
4047 vm_map_entry_t tmp_entry;
4049 if (!vm_map_lookup_entry(map, start, &tmp_entry))
4053 while (start < end) {
4057 if (start < entry->start)
4060 * Check protection associated with entry.
4062 if ((entry->protection & protection) != protection)
4064 /* go to next entry */
4066 entry = vm_map_entry_succ(entry);
4073 * vm_map_copy_swap_object:
4075 * Copies a swap-backed object from an existing map entry to a
4076 * new one. Carries forward the swap charge. May change the
4077 * src object on return.
4080 vm_map_copy_swap_object(vm_map_entry_t src_entry, vm_map_entry_t dst_entry,
4081 vm_offset_t size, vm_ooffset_t *fork_charge)
4083 vm_object_t src_object;
4087 src_object = src_entry->object.vm_object;
4088 charged = ENTRY_CHARGED(src_entry);
4089 if ((src_object->flags & OBJ_ANON) != 0) {
4090 VM_OBJECT_WLOCK(src_object);
4091 vm_object_collapse(src_object);
4092 if ((src_object->flags & OBJ_ONEMAPPING) != 0) {
4093 vm_object_split(src_entry);
4094 src_object = src_entry->object.vm_object;
4096 vm_object_reference_locked(src_object);
4097 vm_object_clear_flag(src_object, OBJ_ONEMAPPING);
4098 VM_OBJECT_WUNLOCK(src_object);
4100 vm_object_reference(src_object);
4101 if (src_entry->cred != NULL &&
4102 !(src_entry->eflags & MAP_ENTRY_NEEDS_COPY)) {
4103 KASSERT(src_object->cred == NULL,
4104 ("OVERCOMMIT: vm_map_copy_anon_entry: cred %p",
4106 src_object->cred = src_entry->cred;
4107 src_object->charge = size;
4109 dst_entry->object.vm_object = src_object;
4111 cred = curthread->td_ucred;
4113 dst_entry->cred = cred;
4114 *fork_charge += size;
4115 if (!(src_entry->eflags & MAP_ENTRY_NEEDS_COPY)) {
4117 src_entry->cred = cred;
4118 *fork_charge += size;
4124 * vm_map_copy_entry:
4126 * Copies the contents of the source entry to the destination
4127 * entry. The entries *must* be aligned properly.
4133 vm_map_entry_t src_entry,
4134 vm_map_entry_t dst_entry,
4135 vm_ooffset_t *fork_charge)
4137 vm_object_t src_object;
4138 vm_map_entry_t fake_entry;
4141 VM_MAP_ASSERT_LOCKED(dst_map);
4143 if ((dst_entry->eflags|src_entry->eflags) & MAP_ENTRY_IS_SUB_MAP)
4146 if (src_entry->wired_count == 0 ||
4147 (src_entry->protection & VM_PROT_WRITE) == 0) {
4149 * If the source entry is marked needs_copy, it is already
4152 if ((src_entry->eflags & MAP_ENTRY_NEEDS_COPY) == 0 &&
4153 (src_entry->protection & VM_PROT_WRITE) != 0) {
4154 pmap_protect(src_map->pmap,
4157 src_entry->protection & ~VM_PROT_WRITE);
4161 * Make a copy of the object.
4163 size = src_entry->end - src_entry->start;
4164 if ((src_object = src_entry->object.vm_object) != NULL) {
4165 if (src_object->type == OBJT_DEFAULT ||
4166 src_object->type == OBJT_SWAP) {
4167 vm_map_copy_swap_object(src_entry, dst_entry,
4169 /* May have split/collapsed, reload obj. */
4170 src_object = src_entry->object.vm_object;
4172 vm_object_reference(src_object);
4173 dst_entry->object.vm_object = src_object;
4175 src_entry->eflags |= MAP_ENTRY_COW |
4176 MAP_ENTRY_NEEDS_COPY;
4177 dst_entry->eflags |= MAP_ENTRY_COW |
4178 MAP_ENTRY_NEEDS_COPY;
4179 dst_entry->offset = src_entry->offset;
4180 if (src_entry->eflags & MAP_ENTRY_WRITECNT) {
4182 * MAP_ENTRY_WRITECNT cannot
4183 * indicate write reference from
4184 * src_entry, since the entry is
4185 * marked as needs copy. Allocate a
4186 * fake entry that is used to
4187 * decrement object->un_pager writecount
4188 * at the appropriate time. Attach
4189 * fake_entry to the deferred list.
4191 fake_entry = vm_map_entry_create(dst_map);
4192 fake_entry->eflags = MAP_ENTRY_WRITECNT;
4193 src_entry->eflags &= ~MAP_ENTRY_WRITECNT;
4194 vm_object_reference(src_object);
4195 fake_entry->object.vm_object = src_object;
4196 fake_entry->start = src_entry->start;
4197 fake_entry->end = src_entry->end;
4198 fake_entry->defer_next =
4199 curthread->td_map_def_user;
4200 curthread->td_map_def_user = fake_entry;
4203 pmap_copy(dst_map->pmap, src_map->pmap,
4204 dst_entry->start, dst_entry->end - dst_entry->start,
4207 dst_entry->object.vm_object = NULL;
4208 dst_entry->offset = 0;
4209 if (src_entry->cred != NULL) {
4210 dst_entry->cred = curthread->td_ucred;
4211 crhold(dst_entry->cred);
4212 *fork_charge += size;
4217 * We don't want to make writeable wired pages copy-on-write.
4218 * Immediately copy these pages into the new map by simulating
4219 * page faults. The new pages are pageable.
4221 vm_fault_copy_entry(dst_map, src_map, dst_entry, src_entry,
4227 * vmspace_map_entry_forked:
4228 * Update the newly-forked vmspace each time a map entry is inherited
4229 * or copied. The values for vm_dsize and vm_tsize are approximate
4230 * (and mostly-obsolete ideas in the face of mmap(2) et al.)
4233 vmspace_map_entry_forked(const struct vmspace *vm1, struct vmspace *vm2,
4234 vm_map_entry_t entry)
4236 vm_size_t entrysize;
4239 if ((entry->eflags & MAP_ENTRY_GUARD) != 0)
4241 entrysize = entry->end - entry->start;
4242 vm2->vm_map.size += entrysize;
4243 if (entry->eflags & (MAP_ENTRY_GROWS_DOWN | MAP_ENTRY_GROWS_UP)) {
4244 vm2->vm_ssize += btoc(entrysize);
4245 } else if (entry->start >= (vm_offset_t)vm1->vm_daddr &&
4246 entry->start < (vm_offset_t)vm1->vm_daddr + ctob(vm1->vm_dsize)) {
4247 newend = MIN(entry->end,
4248 (vm_offset_t)vm1->vm_daddr + ctob(vm1->vm_dsize));
4249 vm2->vm_dsize += btoc(newend - entry->start);
4250 } else if (entry->start >= (vm_offset_t)vm1->vm_taddr &&
4251 entry->start < (vm_offset_t)vm1->vm_taddr + ctob(vm1->vm_tsize)) {
4252 newend = MIN(entry->end,
4253 (vm_offset_t)vm1->vm_taddr + ctob(vm1->vm_tsize));
4254 vm2->vm_tsize += btoc(newend - entry->start);
4260 * Create a new process vmspace structure and vm_map
4261 * based on those of an existing process. The new map
4262 * is based on the old map, according to the inheritance
4263 * values on the regions in that map.
4265 * XXX It might be worth coalescing the entries added to the new vmspace.
4267 * The source map must not be locked.
4270 vmspace_fork(struct vmspace *vm1, vm_ooffset_t *fork_charge)
4272 struct vmspace *vm2;
4273 vm_map_t new_map, old_map;
4274 vm_map_entry_t new_entry, old_entry;
4279 old_map = &vm1->vm_map;
4280 /* Copy immutable fields of vm1 to vm2. */
4281 vm2 = vmspace_alloc(vm_map_min(old_map), vm_map_max(old_map),
4286 vm2->vm_taddr = vm1->vm_taddr;
4287 vm2->vm_daddr = vm1->vm_daddr;
4288 vm2->vm_maxsaddr = vm1->vm_maxsaddr;
4289 vm_map_lock(old_map);
4291 vm_map_wait_busy(old_map);
4292 new_map = &vm2->vm_map;
4293 locked = vm_map_trylock(new_map); /* trylock to silence WITNESS */
4294 KASSERT(locked, ("vmspace_fork: lock failed"));
4296 error = pmap_vmspace_copy(new_map->pmap, old_map->pmap);
4298 sx_xunlock(&old_map->lock);
4299 sx_xunlock(&new_map->lock);
4300 vm_map_process_deferred();
4305 new_map->anon_loc = old_map->anon_loc;
4306 new_map->flags |= old_map->flags & (MAP_ASLR | MAP_ASLR_IGNSTART |
4309 VM_MAP_ENTRY_FOREACH(old_entry, old_map) {
4310 if ((old_entry->eflags & MAP_ENTRY_IS_SUB_MAP) != 0)
4311 panic("vm_map_fork: encountered a submap");
4313 inh = old_entry->inheritance;
4314 if ((old_entry->eflags & MAP_ENTRY_GUARD) != 0 &&
4315 inh != VM_INHERIT_NONE)
4316 inh = VM_INHERIT_COPY;
4319 case VM_INHERIT_NONE:
4322 case VM_INHERIT_SHARE:
4324 * Clone the entry, creating the shared object if
4327 object = old_entry->object.vm_object;
4328 if (object == NULL) {
4329 vm_map_entry_back(old_entry);
4330 object = old_entry->object.vm_object;
4334 * Add the reference before calling vm_object_shadow
4335 * to insure that a shadow object is created.
4337 vm_object_reference(object);
4338 if (old_entry->eflags & MAP_ENTRY_NEEDS_COPY) {
4339 vm_object_shadow(&old_entry->object.vm_object,
4341 old_entry->end - old_entry->start,
4343 /* Transfer the second reference too. */
4345 old_entry->eflags &= ~MAP_ENTRY_NEEDS_COPY;
4346 old_entry->cred = NULL;
4349 * As in vm_map_merged_neighbor_dispose(),
4350 * the vnode lock will not be acquired in
4351 * this call to vm_object_deallocate().
4353 vm_object_deallocate(object);
4354 object = old_entry->object.vm_object;
4356 VM_OBJECT_WLOCK(object);
4357 vm_object_clear_flag(object, OBJ_ONEMAPPING);
4358 if (old_entry->cred != NULL) {
4359 KASSERT(object->cred == NULL,
4360 ("vmspace_fork both cred"));
4361 object->cred = old_entry->cred;
4362 object->charge = old_entry->end -
4364 old_entry->cred = NULL;
4368 * Assert the correct state of the vnode
4369 * v_writecount while the object is locked, to
4370 * not relock it later for the assertion
4373 if (old_entry->eflags & MAP_ENTRY_WRITECNT &&
4374 object->type == OBJT_VNODE) {
4375 KASSERT(((struct vnode *)object->
4376 handle)->v_writecount > 0,
4377 ("vmspace_fork: v_writecount %p",
4379 KASSERT(object->un_pager.vnp.
4381 ("vmspace_fork: vnp.writecount %p",
4384 VM_OBJECT_WUNLOCK(object);
4388 * Clone the entry, referencing the shared object.
4390 new_entry = vm_map_entry_create(new_map);
4391 *new_entry = *old_entry;
4392 new_entry->eflags &= ~(MAP_ENTRY_USER_WIRED |
4393 MAP_ENTRY_IN_TRANSITION);
4394 new_entry->wiring_thread = NULL;
4395 new_entry->wired_count = 0;
4396 if (new_entry->eflags & MAP_ENTRY_WRITECNT) {
4397 vm_pager_update_writecount(object,
4398 new_entry->start, new_entry->end);
4400 vm_map_entry_set_vnode_text(new_entry, true);
4403 * Insert the entry into the new map -- we know we're
4404 * inserting at the end of the new map.
4406 vm_map_entry_link(new_map, new_entry);
4407 vmspace_map_entry_forked(vm1, vm2, new_entry);
4410 * Update the physical map
4412 pmap_copy(new_map->pmap, old_map->pmap,
4414 (old_entry->end - old_entry->start),
4418 case VM_INHERIT_COPY:
4420 * Clone the entry and link into the map.
4422 new_entry = vm_map_entry_create(new_map);
4423 *new_entry = *old_entry;
4425 * Copied entry is COW over the old object.
4427 new_entry->eflags &= ~(MAP_ENTRY_USER_WIRED |
4428 MAP_ENTRY_IN_TRANSITION | MAP_ENTRY_WRITECNT);
4429 new_entry->wiring_thread = NULL;
4430 new_entry->wired_count = 0;
4431 new_entry->object.vm_object = NULL;
4432 new_entry->cred = NULL;
4433 vm_map_entry_link(new_map, new_entry);
4434 vmspace_map_entry_forked(vm1, vm2, new_entry);
4435 vm_map_copy_entry(old_map, new_map, old_entry,
4436 new_entry, fork_charge);
4437 vm_map_entry_set_vnode_text(new_entry, true);
4440 case VM_INHERIT_ZERO:
4442 * Create a new anonymous mapping entry modelled from
4445 new_entry = vm_map_entry_create(new_map);
4446 memset(new_entry, 0, sizeof(*new_entry));
4448 new_entry->start = old_entry->start;
4449 new_entry->end = old_entry->end;
4450 new_entry->eflags = old_entry->eflags &
4451 ~(MAP_ENTRY_USER_WIRED | MAP_ENTRY_IN_TRANSITION |
4452 MAP_ENTRY_WRITECNT | MAP_ENTRY_VN_EXEC |
4453 MAP_ENTRY_SPLIT_BOUNDARY_MASK);
4454 new_entry->protection = old_entry->protection;
4455 new_entry->max_protection = old_entry->max_protection;
4456 new_entry->inheritance = VM_INHERIT_ZERO;
4458 vm_map_entry_link(new_map, new_entry);
4459 vmspace_map_entry_forked(vm1, vm2, new_entry);
4461 new_entry->cred = curthread->td_ucred;
4462 crhold(new_entry->cred);
4463 *fork_charge += (new_entry->end - new_entry->start);
4469 * Use inlined vm_map_unlock() to postpone handling the deferred
4470 * map entries, which cannot be done until both old_map and
4471 * new_map locks are released.
4473 sx_xunlock(&old_map->lock);
4474 sx_xunlock(&new_map->lock);
4475 vm_map_process_deferred();
4481 * Create a process's stack for exec_new_vmspace(). This function is never
4482 * asked to wire the newly created stack.
4485 vm_map_stack(vm_map_t map, vm_offset_t addrbos, vm_size_t max_ssize,
4486 vm_prot_t prot, vm_prot_t max, int cow)
4488 vm_size_t growsize, init_ssize;
4492 MPASS((map->flags & MAP_WIREFUTURE) == 0);
4493 growsize = sgrowsiz;
4494 init_ssize = (max_ssize < growsize) ? max_ssize : growsize;
4496 vmemlim = lim_cur(curthread, RLIMIT_VMEM);
4497 /* If we would blow our VMEM resource limit, no go */
4498 if (map->size + init_ssize > vmemlim) {
4502 rv = vm_map_stack_locked(map, addrbos, max_ssize, growsize, prot,
4509 static int stack_guard_page = 1;
4510 SYSCTL_INT(_security_bsd, OID_AUTO, stack_guard_page, CTLFLAG_RWTUN,
4511 &stack_guard_page, 0,
4512 "Specifies the number of guard pages for a stack that grows");
4515 vm_map_stack_locked(vm_map_t map, vm_offset_t addrbos, vm_size_t max_ssize,
4516 vm_size_t growsize, vm_prot_t prot, vm_prot_t max, int cow)
4518 vm_map_entry_t new_entry, prev_entry;
4519 vm_offset_t bot, gap_bot, gap_top, top;
4520 vm_size_t init_ssize, sgp;
4524 * The stack orientation is piggybacked with the cow argument.
4525 * Extract it into orient and mask the cow argument so that we
4526 * don't pass it around further.
4528 orient = cow & (MAP_STACK_GROWS_DOWN | MAP_STACK_GROWS_UP);
4529 KASSERT(orient != 0, ("No stack grow direction"));
4530 KASSERT(orient != (MAP_STACK_GROWS_DOWN | MAP_STACK_GROWS_UP),
4533 if (max_ssize == 0 ||
4534 !vm_map_range_valid(map, addrbos, addrbos + max_ssize))
4535 return (KERN_INVALID_ADDRESS);
4536 sgp = ((curproc->p_flag2 & P2_STKGAP_DISABLE) != 0 ||
4537 (curproc->p_fctl0 & NT_FREEBSD_FCTL_STKGAP_DISABLE) != 0) ? 0 :
4538 (vm_size_t)stack_guard_page * PAGE_SIZE;
4539 if (sgp >= max_ssize)
4540 return (KERN_INVALID_ARGUMENT);
4542 init_ssize = growsize;
4543 if (max_ssize < init_ssize + sgp)
4544 init_ssize = max_ssize - sgp;
4546 /* If addr is already mapped, no go */
4547 if (vm_map_lookup_entry(map, addrbos, &prev_entry))
4548 return (KERN_NO_SPACE);
4551 * If we can't accommodate max_ssize in the current mapping, no go.
4553 if (vm_map_entry_succ(prev_entry)->start < addrbos + max_ssize)
4554 return (KERN_NO_SPACE);
4557 * We initially map a stack of only init_ssize. We will grow as
4558 * needed later. Depending on the orientation of the stack (i.e.
4559 * the grow direction) we either map at the top of the range, the
4560 * bottom of the range or in the middle.
4562 * Note: we would normally expect prot and max to be VM_PROT_ALL,
4563 * and cow to be 0. Possibly we should eliminate these as input
4564 * parameters, and just pass these values here in the insert call.
4566 if (orient == MAP_STACK_GROWS_DOWN) {
4567 bot = addrbos + max_ssize - init_ssize;
4568 top = bot + init_ssize;
4571 } else /* if (orient == MAP_STACK_GROWS_UP) */ {
4573 top = bot + init_ssize;
4575 gap_top = addrbos + max_ssize;
4577 rv = vm_map_insert(map, NULL, 0, bot, top, prot, max, cow);
4578 if (rv != KERN_SUCCESS)
4580 new_entry = vm_map_entry_succ(prev_entry);
4581 KASSERT(new_entry->end == top || new_entry->start == bot,
4582 ("Bad entry start/end for new stack entry"));
4583 KASSERT((orient & MAP_STACK_GROWS_DOWN) == 0 ||
4584 (new_entry->eflags & MAP_ENTRY_GROWS_DOWN) != 0,
4585 ("new entry lacks MAP_ENTRY_GROWS_DOWN"));
4586 KASSERT((orient & MAP_STACK_GROWS_UP) == 0 ||
4587 (new_entry->eflags & MAP_ENTRY_GROWS_UP) != 0,
4588 ("new entry lacks MAP_ENTRY_GROWS_UP"));
4589 if (gap_bot == gap_top)
4590 return (KERN_SUCCESS);
4591 rv = vm_map_insert(map, NULL, 0, gap_bot, gap_top, VM_PROT_NONE,
4592 VM_PROT_NONE, MAP_CREATE_GUARD | (orient == MAP_STACK_GROWS_DOWN ?
4593 MAP_CREATE_STACK_GAP_DN : MAP_CREATE_STACK_GAP_UP));
4594 if (rv == KERN_SUCCESS) {
4596 * Gap can never successfully handle a fault, so
4597 * read-ahead logic is never used for it. Re-use
4598 * next_read of the gap entry to store
4599 * stack_guard_page for vm_map_growstack().
4601 if (orient == MAP_STACK_GROWS_DOWN)
4602 vm_map_entry_pred(new_entry)->next_read = sgp;
4604 vm_map_entry_succ(new_entry)->next_read = sgp;
4606 (void)vm_map_delete(map, bot, top);
4612 * Attempts to grow a vm stack entry. Returns KERN_SUCCESS if we
4613 * successfully grow the stack.
4616 vm_map_growstack(vm_map_t map, vm_offset_t addr, vm_map_entry_t gap_entry)
4618 vm_map_entry_t stack_entry;
4622 vm_offset_t gap_end, gap_start, grow_start;
4623 vm_size_t grow_amount, guard, max_grow;
4624 rlim_t lmemlim, stacklim, vmemlim;
4626 bool gap_deleted, grow_down, is_procstack;
4638 * Disallow stack growth when the access is performed by a
4639 * debugger or AIO daemon. The reason is that the wrong
4640 * resource limits are applied.
4642 if (p != initproc && (map != &p->p_vmspace->vm_map ||
4643 p->p_textvp == NULL))
4644 return (KERN_FAILURE);
4646 MPASS(!map->system_map);
4648 lmemlim = lim_cur(curthread, RLIMIT_MEMLOCK);
4649 stacklim = lim_cur(curthread, RLIMIT_STACK);
4650 vmemlim = lim_cur(curthread, RLIMIT_VMEM);
4652 /* If addr is not in a hole for a stack grow area, no need to grow. */
4653 if (gap_entry == NULL && !vm_map_lookup_entry(map, addr, &gap_entry))
4654 return (KERN_FAILURE);
4655 if ((gap_entry->eflags & MAP_ENTRY_GUARD) == 0)
4656 return (KERN_SUCCESS);
4657 if ((gap_entry->eflags & MAP_ENTRY_STACK_GAP_DN) != 0) {
4658 stack_entry = vm_map_entry_succ(gap_entry);
4659 if ((stack_entry->eflags & MAP_ENTRY_GROWS_DOWN) == 0 ||
4660 stack_entry->start != gap_entry->end)
4661 return (KERN_FAILURE);
4662 grow_amount = round_page(stack_entry->start - addr);
4664 } else if ((gap_entry->eflags & MAP_ENTRY_STACK_GAP_UP) != 0) {
4665 stack_entry = vm_map_entry_pred(gap_entry);
4666 if ((stack_entry->eflags & MAP_ENTRY_GROWS_UP) == 0 ||
4667 stack_entry->end != gap_entry->start)
4668 return (KERN_FAILURE);
4669 grow_amount = round_page(addr + 1 - stack_entry->end);
4672 return (KERN_FAILURE);
4674 guard = ((curproc->p_flag2 & P2_STKGAP_DISABLE) != 0 ||
4675 (curproc->p_fctl0 & NT_FREEBSD_FCTL_STKGAP_DISABLE) != 0) ? 0 :
4676 gap_entry->next_read;
4677 max_grow = gap_entry->end - gap_entry->start;
4678 if (guard > max_grow)
4679 return (KERN_NO_SPACE);
4681 if (grow_amount > max_grow)
4682 return (KERN_NO_SPACE);
4685 * If this is the main process stack, see if we're over the stack
4688 is_procstack = addr >= (vm_offset_t)vm->vm_maxsaddr &&
4689 addr < (vm_offset_t)p->p_sysent->sv_usrstack;
4690 if (is_procstack && (ctob(vm->vm_ssize) + grow_amount > stacklim))
4691 return (KERN_NO_SPACE);
4696 if (is_procstack && racct_set(p, RACCT_STACK,
4697 ctob(vm->vm_ssize) + grow_amount)) {
4699 return (KERN_NO_SPACE);
4705 grow_amount = roundup(grow_amount, sgrowsiz);
4706 if (grow_amount > max_grow)
4707 grow_amount = max_grow;
4708 if (is_procstack && (ctob(vm->vm_ssize) + grow_amount > stacklim)) {
4709 grow_amount = trunc_page((vm_size_t)stacklim) -
4715 limit = racct_get_available(p, RACCT_STACK);
4717 if (is_procstack && (ctob(vm->vm_ssize) + grow_amount > limit))
4718 grow_amount = limit - ctob(vm->vm_ssize);
4721 if (!old_mlock && (map->flags & MAP_WIREFUTURE) != 0) {
4722 if (ptoa(pmap_wired_count(map->pmap)) + grow_amount > lmemlim) {
4729 if (racct_set(p, RACCT_MEMLOCK,
4730 ptoa(pmap_wired_count(map->pmap)) + grow_amount)) {
4740 /* If we would blow our VMEM resource limit, no go */
4741 if (map->size + grow_amount > vmemlim) {
4748 if (racct_set(p, RACCT_VMEM, map->size + grow_amount)) {
4757 if (vm_map_lock_upgrade(map)) {
4759 vm_map_lock_read(map);
4764 grow_start = gap_entry->end - grow_amount;
4765 if (gap_entry->start + grow_amount == gap_entry->end) {
4766 gap_start = gap_entry->start;
4767 gap_end = gap_entry->end;
4768 vm_map_entry_delete(map, gap_entry);
4771 MPASS(gap_entry->start < gap_entry->end - grow_amount);
4772 vm_map_entry_resize(map, gap_entry, -grow_amount);
4773 gap_deleted = false;
4775 rv = vm_map_insert(map, NULL, 0, grow_start,
4776 grow_start + grow_amount,
4777 stack_entry->protection, stack_entry->max_protection,
4778 MAP_STACK_GROWS_DOWN);
4779 if (rv != KERN_SUCCESS) {
4781 rv1 = vm_map_insert(map, NULL, 0, gap_start,
4782 gap_end, VM_PROT_NONE, VM_PROT_NONE,
4783 MAP_CREATE_GUARD | MAP_CREATE_STACK_GAP_DN);
4784 MPASS(rv1 == KERN_SUCCESS);
4786 vm_map_entry_resize(map, gap_entry,
4790 grow_start = stack_entry->end;
4791 cred = stack_entry->cred;
4792 if (cred == NULL && stack_entry->object.vm_object != NULL)
4793 cred = stack_entry->object.vm_object->cred;
4794 if (cred != NULL && !swap_reserve_by_cred(grow_amount, cred))
4796 /* Grow the underlying object if applicable. */
4797 else if (stack_entry->object.vm_object == NULL ||
4798 vm_object_coalesce(stack_entry->object.vm_object,
4799 stack_entry->offset,
4800 (vm_size_t)(stack_entry->end - stack_entry->start),
4801 grow_amount, cred != NULL)) {
4802 if (gap_entry->start + grow_amount == gap_entry->end) {
4803 vm_map_entry_delete(map, gap_entry);
4804 vm_map_entry_resize(map, stack_entry,
4807 gap_entry->start += grow_amount;
4808 stack_entry->end += grow_amount;
4810 map->size += grow_amount;
4815 if (rv == KERN_SUCCESS && is_procstack)
4816 vm->vm_ssize += btoc(grow_amount);
4819 * Heed the MAP_WIREFUTURE flag if it was set for this process.
4821 if (rv == KERN_SUCCESS && (map->flags & MAP_WIREFUTURE) != 0) {
4822 rv = vm_map_wire_locked(map, grow_start,
4823 grow_start + grow_amount,
4824 VM_MAP_WIRE_USER | VM_MAP_WIRE_NOHOLES);
4826 vm_map_lock_downgrade(map);
4830 if (racct_enable && rv != KERN_SUCCESS) {
4832 error = racct_set(p, RACCT_VMEM, map->size);
4833 KASSERT(error == 0, ("decreasing RACCT_VMEM failed"));
4835 error = racct_set(p, RACCT_MEMLOCK,
4836 ptoa(pmap_wired_count(map->pmap)));
4837 KASSERT(error == 0, ("decreasing RACCT_MEMLOCK failed"));
4839 error = racct_set(p, RACCT_STACK, ctob(vm->vm_ssize));
4840 KASSERT(error == 0, ("decreasing RACCT_STACK failed"));
4849 * Unshare the specified VM space for exec. If other processes are
4850 * mapped to it, then create a new one. The new vmspace is null.
4853 vmspace_exec(struct proc *p, vm_offset_t minuser, vm_offset_t maxuser)
4855 struct vmspace *oldvmspace = p->p_vmspace;
4856 struct vmspace *newvmspace;
4858 KASSERT((curthread->td_pflags & TDP_EXECVMSPC) == 0,
4859 ("vmspace_exec recursed"));
4860 newvmspace = vmspace_alloc(minuser, maxuser, pmap_pinit);
4861 if (newvmspace == NULL)
4863 newvmspace->vm_swrss = oldvmspace->vm_swrss;
4865 * This code is written like this for prototype purposes. The
4866 * goal is to avoid running down the vmspace here, but let the
4867 * other process's that are still using the vmspace to finally
4868 * run it down. Even though there is little or no chance of blocking
4869 * here, it is a good idea to keep this form for future mods.
4871 PROC_VMSPACE_LOCK(p);
4872 p->p_vmspace = newvmspace;
4873 PROC_VMSPACE_UNLOCK(p);
4874 if (p == curthread->td_proc)
4875 pmap_activate(curthread);
4876 curthread->td_pflags |= TDP_EXECVMSPC;
4881 * Unshare the specified VM space for forcing COW. This
4882 * is called by rfork, for the (RFMEM|RFPROC) == 0 case.
4885 vmspace_unshare(struct proc *p)
4887 struct vmspace *oldvmspace = p->p_vmspace;
4888 struct vmspace *newvmspace;
4889 vm_ooffset_t fork_charge;
4891 if (refcount_load(&oldvmspace->vm_refcnt) == 1)
4894 newvmspace = vmspace_fork(oldvmspace, &fork_charge);
4895 if (newvmspace == NULL)
4897 if (!swap_reserve_by_cred(fork_charge, p->p_ucred)) {
4898 vmspace_free(newvmspace);
4901 PROC_VMSPACE_LOCK(p);
4902 p->p_vmspace = newvmspace;
4903 PROC_VMSPACE_UNLOCK(p);
4904 if (p == curthread->td_proc)
4905 pmap_activate(curthread);
4906 vmspace_free(oldvmspace);
4913 * Finds the VM object, offset, and
4914 * protection for a given virtual address in the
4915 * specified map, assuming a page fault of the
4918 * Leaves the map in question locked for read; return
4919 * values are guaranteed until a vm_map_lookup_done
4920 * call is performed. Note that the map argument
4921 * is in/out; the returned map must be used in
4922 * the call to vm_map_lookup_done.
4924 * A handle (out_entry) is returned for use in
4925 * vm_map_lookup_done, to make that fast.
4927 * If a lookup is requested with "write protection"
4928 * specified, the map may be changed to perform virtual
4929 * copying operations, although the data referenced will
4933 vm_map_lookup(vm_map_t *var_map, /* IN/OUT */
4935 vm_prot_t fault_typea,
4936 vm_map_entry_t *out_entry, /* OUT */
4937 vm_object_t *object, /* OUT */
4938 vm_pindex_t *pindex, /* OUT */
4939 vm_prot_t *out_prot, /* OUT */
4940 boolean_t *wired) /* OUT */
4942 vm_map_entry_t entry;
4943 vm_map_t map = *var_map;
4945 vm_prot_t fault_type;
4946 vm_object_t eobject;
4952 vm_map_lock_read(map);
4956 * Lookup the faulting address.
4958 if (!vm_map_lookup_entry(map, vaddr, out_entry)) {
4959 vm_map_unlock_read(map);
4960 return (KERN_INVALID_ADDRESS);
4968 if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) {
4969 vm_map_t old_map = map;
4971 *var_map = map = entry->object.sub_map;
4972 vm_map_unlock_read(old_map);
4977 * Check whether this task is allowed to have this page.
4979 prot = entry->protection;
4980 if ((fault_typea & VM_PROT_FAULT_LOOKUP) != 0) {
4981 fault_typea &= ~VM_PROT_FAULT_LOOKUP;
4982 if (prot == VM_PROT_NONE && map != kernel_map &&
4983 (entry->eflags & MAP_ENTRY_GUARD) != 0 &&
4984 (entry->eflags & (MAP_ENTRY_STACK_GAP_DN |
4985 MAP_ENTRY_STACK_GAP_UP)) != 0 &&
4986 vm_map_growstack(map, vaddr, entry) == KERN_SUCCESS)
4987 goto RetryLookupLocked;
4989 fault_type = fault_typea & VM_PROT_ALL;
4990 if ((fault_type & prot) != fault_type || prot == VM_PROT_NONE) {
4991 vm_map_unlock_read(map);
4992 return (KERN_PROTECTION_FAILURE);
4994 KASSERT((prot & VM_PROT_WRITE) == 0 || (entry->eflags &
4995 (MAP_ENTRY_USER_WIRED | MAP_ENTRY_NEEDS_COPY)) !=
4996 (MAP_ENTRY_USER_WIRED | MAP_ENTRY_NEEDS_COPY),
4997 ("entry %p flags %x", entry, entry->eflags));
4998 if ((fault_typea & VM_PROT_COPY) != 0 &&
4999 (entry->max_protection & VM_PROT_WRITE) == 0 &&
5000 (entry->eflags & MAP_ENTRY_COW) == 0) {
5001 vm_map_unlock_read(map);
5002 return (KERN_PROTECTION_FAILURE);
5006 * If this page is not pageable, we have to get it for all possible
5009 *wired = (entry->wired_count != 0);
5011 fault_type = entry->protection;
5012 size = entry->end - entry->start;
5015 * If the entry was copy-on-write, we either ...
5017 if (entry->eflags & MAP_ENTRY_NEEDS_COPY) {
5019 * If we want to write the page, we may as well handle that
5020 * now since we've got the map locked.
5022 * If we don't need to write the page, we just demote the
5023 * permissions allowed.
5025 if ((fault_type & VM_PROT_WRITE) != 0 ||
5026 (fault_typea & VM_PROT_COPY) != 0) {
5028 * Make a new object, and place it in the object
5029 * chain. Note that no new references have appeared
5030 * -- one just moved from the map to the new
5033 if (vm_map_lock_upgrade(map))
5036 if (entry->cred == NULL) {
5038 * The debugger owner is charged for
5041 cred = curthread->td_ucred;
5043 if (!swap_reserve_by_cred(size, cred)) {
5046 return (KERN_RESOURCE_SHORTAGE);
5050 eobject = entry->object.vm_object;
5051 vm_object_shadow(&entry->object.vm_object,
5052 &entry->offset, size, entry->cred, false);
5053 if (eobject == entry->object.vm_object) {
5055 * The object was not shadowed.
5057 swap_release_by_cred(size, entry->cred);
5058 crfree(entry->cred);
5061 entry->eflags &= ~MAP_ENTRY_NEEDS_COPY;
5063 vm_map_lock_downgrade(map);
5066 * We're attempting to read a copy-on-write page --
5067 * don't allow writes.
5069 prot &= ~VM_PROT_WRITE;
5074 * Create an object if necessary.
5076 if (entry->object.vm_object == NULL && !map->system_map) {
5077 if (vm_map_lock_upgrade(map))
5079 entry->object.vm_object = vm_object_allocate_anon(atop(size),
5080 NULL, entry->cred, entry->cred != NULL ? size : 0);
5083 vm_map_lock_downgrade(map);
5087 * Return the object/offset from this entry. If the entry was
5088 * copy-on-write or empty, it has been fixed up.
5090 *pindex = OFF_TO_IDX((vaddr - entry->start) + entry->offset);
5091 *object = entry->object.vm_object;
5094 return (KERN_SUCCESS);
5098 * vm_map_lookup_locked:
5100 * Lookup the faulting address. A version of vm_map_lookup that returns
5101 * KERN_FAILURE instead of blocking on map lock or memory allocation.
5104 vm_map_lookup_locked(vm_map_t *var_map, /* IN/OUT */
5106 vm_prot_t fault_typea,
5107 vm_map_entry_t *out_entry, /* OUT */
5108 vm_object_t *object, /* OUT */
5109 vm_pindex_t *pindex, /* OUT */
5110 vm_prot_t *out_prot, /* OUT */
5111 boolean_t *wired) /* OUT */
5113 vm_map_entry_t entry;
5114 vm_map_t map = *var_map;
5116 vm_prot_t fault_type = fault_typea;
5119 * Lookup the faulting address.
5121 if (!vm_map_lookup_entry(map, vaddr, out_entry))
5122 return (KERN_INVALID_ADDRESS);
5127 * Fail if the entry refers to a submap.
5129 if (entry->eflags & MAP_ENTRY_IS_SUB_MAP)
5130 return (KERN_FAILURE);
5133 * Check whether this task is allowed to have this page.
5135 prot = entry->protection;
5136 fault_type &= VM_PROT_READ | VM_PROT_WRITE | VM_PROT_EXECUTE;
5137 if ((fault_type & prot) != fault_type)
5138 return (KERN_PROTECTION_FAILURE);
5141 * If this page is not pageable, we have to get it for all possible
5144 *wired = (entry->wired_count != 0);
5146 fault_type = entry->protection;
5148 if (entry->eflags & MAP_ENTRY_NEEDS_COPY) {
5150 * Fail if the entry was copy-on-write for a write fault.
5152 if (fault_type & VM_PROT_WRITE)
5153 return (KERN_FAILURE);
5155 * We're attempting to read a copy-on-write page --
5156 * don't allow writes.
5158 prot &= ~VM_PROT_WRITE;
5162 * Fail if an object should be created.
5164 if (entry->object.vm_object == NULL && !map->system_map)
5165 return (KERN_FAILURE);
5168 * Return the object/offset from this entry. If the entry was
5169 * copy-on-write or empty, it has been fixed up.
5171 *pindex = OFF_TO_IDX((vaddr - entry->start) + entry->offset);
5172 *object = entry->object.vm_object;
5175 return (KERN_SUCCESS);
5179 * vm_map_lookup_done:
5181 * Releases locks acquired by a vm_map_lookup
5182 * (according to the handle returned by that lookup).
5185 vm_map_lookup_done(vm_map_t map, vm_map_entry_t entry)
5188 * Unlock the main-level map
5190 vm_map_unlock_read(map);
5194 vm_map_max_KBI(const struct vm_map *map)
5197 return (vm_map_max(map));
5201 vm_map_min_KBI(const struct vm_map *map)
5204 return (vm_map_min(map));
5208 vm_map_pmap_KBI(vm_map_t map)
5215 vm_map_range_valid_KBI(vm_map_t map, vm_offset_t start, vm_offset_t end)
5218 return (vm_map_range_valid(map, start, end));
5223 _vm_map_assert_consistent(vm_map_t map, int check)
5225 vm_map_entry_t entry, prev;
5226 vm_map_entry_t cur, header, lbound, ubound;
5227 vm_size_t max_left, max_right;
5232 if (enable_vmmap_check != check)
5235 header = prev = &map->header;
5236 VM_MAP_ENTRY_FOREACH(entry, map) {
5237 KASSERT(prev->end <= entry->start,
5238 ("map %p prev->end = %jx, start = %jx", map,
5239 (uintmax_t)prev->end, (uintmax_t)entry->start));
5240 KASSERT(entry->start < entry->end,
5241 ("map %p start = %jx, end = %jx", map,
5242 (uintmax_t)entry->start, (uintmax_t)entry->end));
5243 KASSERT(entry->left == header ||
5244 entry->left->start < entry->start,
5245 ("map %p left->start = %jx, start = %jx", map,
5246 (uintmax_t)entry->left->start, (uintmax_t)entry->start));
5247 KASSERT(entry->right == header ||
5248 entry->start < entry->right->start,
5249 ("map %p start = %jx, right->start = %jx", map,
5250 (uintmax_t)entry->start, (uintmax_t)entry->right->start));
5252 lbound = ubound = header;
5254 if (entry->start < cur->start) {
5257 KASSERT(cur != lbound,
5258 ("map %p cannot find %jx",
5259 map, (uintmax_t)entry->start));
5260 } else if (cur->end <= entry->start) {
5263 KASSERT(cur != ubound,
5264 ("map %p cannot find %jx",
5265 map, (uintmax_t)entry->start));
5267 KASSERT(cur == entry,
5268 ("map %p cannot find %jx",
5269 map, (uintmax_t)entry->start));
5273 max_left = vm_map_entry_max_free_left(entry, lbound);
5274 max_right = vm_map_entry_max_free_right(entry, ubound);
5275 KASSERT(entry->max_free == vm_size_max(max_left, max_right),
5276 ("map %p max = %jx, max_left = %jx, max_right = %jx", map,
5277 (uintmax_t)entry->max_free,
5278 (uintmax_t)max_left, (uintmax_t)max_right));
5281 KASSERT(prev->end <= entry->start,
5282 ("map %p prev->end = %jx, start = %jx", map,
5283 (uintmax_t)prev->end, (uintmax_t)entry->start));
5287 #include "opt_ddb.h"
5289 #include <sys/kernel.h>
5291 #include <ddb/ddb.h>
5294 vm_map_print(vm_map_t map)
5296 vm_map_entry_t entry, prev;
5298 db_iprintf("Task map %p: pmap=%p, nentries=%d, version=%u\n",
5300 (void *)map->pmap, map->nentries, map->timestamp);
5303 prev = &map->header;
5304 VM_MAP_ENTRY_FOREACH(entry, map) {
5305 db_iprintf("map entry %p: start=%p, end=%p, eflags=%#x, \n",
5306 (void *)entry, (void *)entry->start, (void *)entry->end,
5309 static const char * const inheritance_name[4] =
5310 {"share", "copy", "none", "donate_copy"};
5312 db_iprintf(" prot=%x/%x/%s",
5314 entry->max_protection,
5315 inheritance_name[(int)(unsigned char)
5316 entry->inheritance]);
5317 if (entry->wired_count != 0)
5318 db_printf(", wired");
5320 if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) {
5321 db_printf(", share=%p, offset=0x%jx\n",
5322 (void *)entry->object.sub_map,
5323 (uintmax_t)entry->offset);
5324 if (prev == &map->header ||
5325 prev->object.sub_map !=
5326 entry->object.sub_map) {
5328 vm_map_print((vm_map_t)entry->object.sub_map);
5332 if (entry->cred != NULL)
5333 db_printf(", ruid %d", entry->cred->cr_ruid);
5334 db_printf(", object=%p, offset=0x%jx",
5335 (void *)entry->object.vm_object,
5336 (uintmax_t)entry->offset);
5337 if (entry->object.vm_object && entry->object.vm_object->cred)
5338 db_printf(", obj ruid %d charge %jx",
5339 entry->object.vm_object->cred->cr_ruid,
5340 (uintmax_t)entry->object.vm_object->charge);
5341 if (entry->eflags & MAP_ENTRY_COW)
5342 db_printf(", copy (%s)",
5343 (entry->eflags & MAP_ENTRY_NEEDS_COPY) ? "needed" : "done");
5346 if (prev == &map->header ||
5347 prev->object.vm_object !=
5348 entry->object.vm_object) {
5350 vm_object_print((db_expr_t)(intptr_t)
5351 entry->object.vm_object,
5361 DB_SHOW_COMMAND(map, map)
5365 db_printf("usage: show map <addr>\n");
5368 vm_map_print((vm_map_t)addr);
5371 DB_SHOW_COMMAND(procvm, procvm)
5376 p = db_lookup_proc(addr);
5381 db_printf("p = %p, vmspace = %p, map = %p, pmap = %p\n",
5382 (void *)p, (void *)p->p_vmspace, (void *)&p->p_vmspace->vm_map,
5383 (void *)vmspace_pmap(p->p_vmspace));
5385 vm_map_print((vm_map_t)&p->p_vmspace->vm_map);