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
564 vm_pager_getvp(object, &vp, &vp_held);
567 VOP_SET_TEXT_CHECKED(vp);
569 vn_lock(vp, LK_SHARED | LK_RETRY);
570 VOP_UNSET_TEXT_CHECKED(vp);
579 * Use a different name for this vm_map_entry field when it's use
580 * is not consistent with its use as part of an ordered search tree.
582 #define defer_next right
585 vm_map_process_deferred(void)
588 vm_map_entry_t entry, next;
592 entry = td->td_map_def_user;
593 td->td_map_def_user = NULL;
594 while (entry != NULL) {
595 next = entry->defer_next;
596 MPASS((entry->eflags & (MAP_ENTRY_WRITECNT |
597 MAP_ENTRY_VN_EXEC)) != (MAP_ENTRY_WRITECNT |
599 if ((entry->eflags & MAP_ENTRY_WRITECNT) != 0) {
601 * Decrement the object's writemappings and
602 * possibly the vnode's v_writecount.
604 KASSERT((entry->eflags & MAP_ENTRY_IS_SUB_MAP) == 0,
605 ("Submap with writecount"));
606 object = entry->object.vm_object;
607 KASSERT(object != NULL, ("No object for writecount"));
608 vm_pager_release_writecount(object, entry->start,
611 vm_map_entry_set_vnode_text(entry, false);
612 vm_map_entry_deallocate(entry, FALSE);
619 _vm_map_assert_locked(vm_map_t map, const char *file, int line)
623 mtx_assert_(&map->system_mtx, MA_OWNED, file, line);
625 sx_assert_(&map->lock, SA_XLOCKED, file, line);
628 #define VM_MAP_ASSERT_LOCKED(map) \
629 _vm_map_assert_locked(map, LOCK_FILE, LOCK_LINE)
631 enum { VMMAP_CHECK_NONE, VMMAP_CHECK_UNLOCK, VMMAP_CHECK_ALL };
633 static int enable_vmmap_check = VMMAP_CHECK_UNLOCK;
635 static int enable_vmmap_check = VMMAP_CHECK_NONE;
637 SYSCTL_INT(_debug, OID_AUTO, vmmap_check, CTLFLAG_RWTUN,
638 &enable_vmmap_check, 0, "Enable vm map consistency checking");
640 static void _vm_map_assert_consistent(vm_map_t map, int check);
642 #define VM_MAP_ASSERT_CONSISTENT(map) \
643 _vm_map_assert_consistent(map, VMMAP_CHECK_ALL)
645 #define VM_MAP_UNLOCK_CONSISTENT(map) do { \
646 if (map->nupdates > map->nentries) { \
647 _vm_map_assert_consistent(map, VMMAP_CHECK_UNLOCK); \
652 #define VM_MAP_UNLOCK_CONSISTENT(map)
655 #define VM_MAP_ASSERT_LOCKED(map)
656 #define VM_MAP_ASSERT_CONSISTENT(map)
657 #define VM_MAP_UNLOCK_CONSISTENT(map)
658 #endif /* INVARIANTS */
661 _vm_map_unlock(vm_map_t map, const char *file, int line)
664 VM_MAP_UNLOCK_CONSISTENT(map);
665 if (map->system_map) {
666 #ifndef UMA_MD_SMALL_ALLOC
667 if (map == kernel_map && (map->flags & MAP_REPLENISH) != 0) {
668 uma_prealloc(kmapentzone, 1);
669 map->flags &= ~MAP_REPLENISH;
672 mtx_unlock_flags_(&map->system_mtx, 0, file, line);
674 sx_xunlock_(&map->lock, file, line);
675 vm_map_process_deferred();
680 _vm_map_lock_read(vm_map_t map, const char *file, int line)
684 mtx_lock_flags_(&map->system_mtx, 0, file, line);
686 sx_slock_(&map->lock, file, line);
690 _vm_map_unlock_read(vm_map_t map, const char *file, int line)
693 if (map->system_map) {
694 KASSERT((map->flags & MAP_REPLENISH) == 0,
695 ("%s: MAP_REPLENISH leaked", __func__));
696 mtx_unlock_flags_(&map->system_mtx, 0, file, line);
698 sx_sunlock_(&map->lock, file, line);
699 vm_map_process_deferred();
704 _vm_map_trylock(vm_map_t map, const char *file, int line)
708 error = map->system_map ?
709 !mtx_trylock_flags_(&map->system_mtx, 0, file, line) :
710 !sx_try_xlock_(&map->lock, file, line);
717 _vm_map_trylock_read(vm_map_t map, const char *file, int line)
721 error = map->system_map ?
722 !mtx_trylock_flags_(&map->system_mtx, 0, file, line) :
723 !sx_try_slock_(&map->lock, file, line);
728 * _vm_map_lock_upgrade: [ internal use only ]
730 * Tries to upgrade a read (shared) lock on the specified map to a write
731 * (exclusive) lock. Returns the value "0" if the upgrade succeeds and a
732 * non-zero value if the upgrade fails. If the upgrade fails, the map is
733 * returned without a read or write lock held.
735 * Requires that the map be read locked.
738 _vm_map_lock_upgrade(vm_map_t map, const char *file, int line)
740 unsigned int last_timestamp;
742 if (map->system_map) {
743 mtx_assert_(&map->system_mtx, MA_OWNED, file, line);
745 if (!sx_try_upgrade_(&map->lock, file, line)) {
746 last_timestamp = map->timestamp;
747 sx_sunlock_(&map->lock, file, line);
748 vm_map_process_deferred();
750 * If the map's timestamp does not change while the
751 * map is unlocked, then the upgrade succeeds.
753 sx_xlock_(&map->lock, file, line);
754 if (last_timestamp != map->timestamp) {
755 sx_xunlock_(&map->lock, file, line);
765 _vm_map_lock_downgrade(vm_map_t map, const char *file, int line)
768 if (map->system_map) {
769 KASSERT((map->flags & MAP_REPLENISH) == 0,
770 ("%s: MAP_REPLENISH leaked", __func__));
771 mtx_assert_(&map->system_mtx, MA_OWNED, file, line);
773 VM_MAP_UNLOCK_CONSISTENT(map);
774 sx_downgrade_(&map->lock, file, line);
781 * Returns a non-zero value if the caller holds a write (exclusive) lock
782 * on the specified map and the value "0" otherwise.
785 vm_map_locked(vm_map_t map)
789 return (mtx_owned(&map->system_mtx));
791 return (sx_xlocked(&map->lock));
795 * _vm_map_unlock_and_wait:
797 * Atomically releases the lock on the specified map and puts the calling
798 * thread to sleep. The calling thread will remain asleep until either
799 * vm_map_wakeup() is performed on the map or the specified timeout is
802 * WARNING! This function does not perform deferred deallocations of
803 * objects and map entries. Therefore, the calling thread is expected to
804 * reacquire the map lock after reawakening and later perform an ordinary
805 * unlock operation, such as vm_map_unlock(), before completing its
806 * operation on the map.
809 _vm_map_unlock_and_wait(vm_map_t map, int timo, const char *file, int line)
812 VM_MAP_UNLOCK_CONSISTENT(map);
813 mtx_lock(&map_sleep_mtx);
814 if (map->system_map) {
815 KASSERT((map->flags & MAP_REPLENISH) == 0,
816 ("%s: MAP_REPLENISH leaked", __func__));
817 mtx_unlock_flags_(&map->system_mtx, 0, file, line);
819 sx_xunlock_(&map->lock, file, line);
821 return (msleep(&map->root, &map_sleep_mtx, PDROP | PVM, "vmmaps",
828 * Awaken any threads that have slept on the map using
829 * vm_map_unlock_and_wait().
832 vm_map_wakeup(vm_map_t map)
836 * Acquire and release map_sleep_mtx to prevent a wakeup()
837 * from being performed (and lost) between the map unlock
838 * and the msleep() in _vm_map_unlock_and_wait().
840 mtx_lock(&map_sleep_mtx);
841 mtx_unlock(&map_sleep_mtx);
846 vm_map_busy(vm_map_t map)
849 VM_MAP_ASSERT_LOCKED(map);
854 vm_map_unbusy(vm_map_t map)
857 VM_MAP_ASSERT_LOCKED(map);
858 KASSERT(map->busy, ("vm_map_unbusy: not busy"));
859 if (--map->busy == 0 && (map->flags & MAP_BUSY_WAKEUP)) {
860 vm_map_modflags(map, 0, MAP_BUSY_WAKEUP);
866 vm_map_wait_busy(vm_map_t map)
869 VM_MAP_ASSERT_LOCKED(map);
871 vm_map_modflags(map, MAP_BUSY_WAKEUP, 0);
873 msleep(&map->busy, &map->system_mtx, 0, "mbusy", 0);
875 sx_sleep(&map->busy, &map->lock, 0, "mbusy", 0);
881 vmspace_resident_count(struct vmspace *vmspace)
883 return pmap_resident_count(vmspace_pmap(vmspace));
887 * Initialize an existing vm_map structure
888 * such as that in the vmspace structure.
891 _vm_map_init(vm_map_t map, pmap_t pmap, vm_offset_t min, vm_offset_t max)
894 map->header.eflags = MAP_ENTRY_HEADER;
895 map->needs_wakeup = FALSE;
898 map->header.end = min;
899 map->header.start = max;
901 map->header.left = map->header.right = &map->header;
912 vm_map_init(vm_map_t map, pmap_t pmap, vm_offset_t min, vm_offset_t max)
915 _vm_map_init(map, pmap, min, max);
916 mtx_init(&map->system_mtx, "vm map (system)", NULL,
917 MTX_DEF | MTX_DUPOK);
918 sx_init(&map->lock, "vm map (user)");
922 * vm_map_entry_dispose: [ internal use only ]
924 * Inverse of vm_map_entry_create.
927 vm_map_entry_dispose(vm_map_t map, vm_map_entry_t entry)
929 uma_zfree(map->system_map ? kmapentzone : mapentzone, entry);
933 * vm_map_entry_create: [ internal use only ]
935 * Allocates a VM map entry for insertion.
936 * No entry fields are filled in.
938 static vm_map_entry_t
939 vm_map_entry_create(vm_map_t map)
941 vm_map_entry_t new_entry;
943 #ifndef UMA_MD_SMALL_ALLOC
944 if (map == kernel_map) {
945 VM_MAP_ASSERT_LOCKED(map);
948 * A new slab of kernel map entries cannot be allocated at this
949 * point because the kernel map has not yet been updated to
950 * reflect the caller's request. Therefore, we allocate a new
951 * map entry, dipping into the reserve if necessary, and set a
952 * flag indicating that the reserve must be replenished before
953 * the map is unlocked.
955 new_entry = uma_zalloc(kmapentzone, M_NOWAIT | M_NOVM);
956 if (new_entry == NULL) {
957 new_entry = uma_zalloc(kmapentzone,
958 M_NOWAIT | M_NOVM | M_USE_RESERVE);
959 kernel_map->flags |= MAP_REPLENISH;
963 if (map->system_map) {
964 new_entry = uma_zalloc(kmapentzone, M_NOWAIT);
966 new_entry = uma_zalloc(mapentzone, M_WAITOK);
968 KASSERT(new_entry != NULL,
969 ("vm_map_entry_create: kernel resources exhausted"));
974 * vm_map_entry_set_behavior:
976 * Set the expected access behavior, either normal, random, or
980 vm_map_entry_set_behavior(vm_map_entry_t entry, u_char behavior)
982 entry->eflags = (entry->eflags & ~MAP_ENTRY_BEHAV_MASK) |
983 (behavior & MAP_ENTRY_BEHAV_MASK);
987 * vm_map_entry_max_free_{left,right}:
989 * Compute the size of the largest free gap between two entries,
990 * one the root of a tree and the other the ancestor of that root
991 * that is the least or greatest ancestor found on the search path.
993 static inline vm_size_t
994 vm_map_entry_max_free_left(vm_map_entry_t root, vm_map_entry_t left_ancestor)
997 return (root->left != left_ancestor ?
998 root->left->max_free : root->start - left_ancestor->end);
1001 static inline vm_size_t
1002 vm_map_entry_max_free_right(vm_map_entry_t root, vm_map_entry_t right_ancestor)
1005 return (root->right != right_ancestor ?
1006 root->right->max_free : right_ancestor->start - root->end);
1010 * vm_map_entry_{pred,succ}:
1012 * Find the {predecessor, successor} of the entry by taking one step
1013 * in the appropriate direction and backtracking as much as necessary.
1014 * vm_map_entry_succ is defined in vm_map.h.
1016 static inline vm_map_entry_t
1017 vm_map_entry_pred(vm_map_entry_t entry)
1019 vm_map_entry_t prior;
1021 prior = entry->left;
1022 if (prior->right->start < entry->start) {
1024 prior = prior->right;
1025 while (prior->right != entry);
1030 static inline vm_size_t
1031 vm_size_max(vm_size_t a, vm_size_t b)
1034 return (a > b ? a : b);
1037 #define SPLAY_LEFT_STEP(root, y, llist, rlist, test) do { \
1039 vm_size_t max_free; \
1042 * Infer root->right->max_free == root->max_free when \
1043 * y->max_free < root->max_free || root->max_free == 0. \
1044 * Otherwise, look right to find it. \
1047 max_free = root->max_free; \
1048 KASSERT(max_free == vm_size_max( \
1049 vm_map_entry_max_free_left(root, llist), \
1050 vm_map_entry_max_free_right(root, rlist)), \
1051 ("%s: max_free invariant fails", __func__)); \
1052 if (max_free - 1 < vm_map_entry_max_free_left(root, llist)) \
1053 max_free = vm_map_entry_max_free_right(root, rlist); \
1054 if (y != llist && (test)) { \
1055 /* Rotate right and make y root. */ \
1060 if (max_free < y->max_free) \
1061 root->max_free = max_free = \
1062 vm_size_max(max_free, z->max_free); \
1063 } else if (max_free < y->max_free) \
1064 root->max_free = max_free = \
1065 vm_size_max(max_free, root->start - y->end);\
1069 /* Copy right->max_free. Put root on rlist. */ \
1070 root->max_free = max_free; \
1071 KASSERT(max_free == vm_map_entry_max_free_right(root, rlist), \
1072 ("%s: max_free not copied from right", __func__)); \
1073 root->left = rlist; \
1075 root = y != llist ? y : NULL; \
1078 #define SPLAY_RIGHT_STEP(root, y, llist, rlist, test) do { \
1080 vm_size_t max_free; \
1083 * Infer root->left->max_free == root->max_free when \
1084 * y->max_free < root->max_free || root->max_free == 0. \
1085 * Otherwise, look left to find it. \
1088 max_free = root->max_free; \
1089 KASSERT(max_free == vm_size_max( \
1090 vm_map_entry_max_free_left(root, llist), \
1091 vm_map_entry_max_free_right(root, rlist)), \
1092 ("%s: max_free invariant fails", __func__)); \
1093 if (max_free - 1 < vm_map_entry_max_free_right(root, rlist)) \
1094 max_free = vm_map_entry_max_free_left(root, llist); \
1095 if (y != rlist && (test)) { \
1096 /* Rotate left and make y root. */ \
1101 if (max_free < y->max_free) \
1102 root->max_free = max_free = \
1103 vm_size_max(max_free, z->max_free); \
1104 } else if (max_free < y->max_free) \
1105 root->max_free = max_free = \
1106 vm_size_max(max_free, y->start - root->end);\
1110 /* Copy left->max_free. Put root on llist. */ \
1111 root->max_free = max_free; \
1112 KASSERT(max_free == vm_map_entry_max_free_left(root, llist), \
1113 ("%s: max_free not copied from left", __func__)); \
1114 root->right = llist; \
1116 root = y != rlist ? y : NULL; \
1120 * Walk down the tree until we find addr or a gap where addr would go, breaking
1121 * off left and right subtrees of nodes less than, or greater than addr. Treat
1122 * subtrees with root->max_free < length as empty trees. llist and rlist are
1123 * the two sides in reverse order (bottom-up), with llist linked by the right
1124 * pointer and rlist linked by the left pointer in the vm_map_entry, and both
1125 * lists terminated by &map->header. This function, and the subsequent call to
1126 * vm_map_splay_merge_{left,right,pred,succ}, rely on the start and end address
1127 * values in &map->header.
1129 static __always_inline vm_map_entry_t
1130 vm_map_splay_split(vm_map_t map, vm_offset_t addr, vm_size_t length,
1131 vm_map_entry_t *llist, vm_map_entry_t *rlist)
1133 vm_map_entry_t left, right, root, y;
1135 left = right = &map->header;
1137 while (root != NULL && root->max_free >= length) {
1138 KASSERT(left->end <= root->start &&
1139 root->end <= right->start,
1140 ("%s: root not within tree bounds", __func__));
1141 if (addr < root->start) {
1142 SPLAY_LEFT_STEP(root, y, left, right,
1143 y->max_free >= length && addr < y->start);
1144 } else if (addr >= root->end) {
1145 SPLAY_RIGHT_STEP(root, y, left, right,
1146 y->max_free >= length && addr >= y->end);
1155 static __always_inline void
1156 vm_map_splay_findnext(vm_map_entry_t root, vm_map_entry_t *rlist)
1158 vm_map_entry_t hi, right, y;
1161 hi = root->right == right ? NULL : root->right;
1165 SPLAY_LEFT_STEP(hi, y, root, right, true);
1170 static __always_inline void
1171 vm_map_splay_findprev(vm_map_entry_t root, vm_map_entry_t *llist)
1173 vm_map_entry_t left, lo, y;
1176 lo = root->left == left ? NULL : root->left;
1180 SPLAY_RIGHT_STEP(lo, y, left, root, true);
1186 vm_map_entry_swap(vm_map_entry_t *a, vm_map_entry_t *b)
1196 * Walk back up the two spines, flip the pointers and set max_free. The
1197 * subtrees of the root go at the bottom of llist and rlist.
1200 vm_map_splay_merge_left_walk(vm_map_entry_t header, vm_map_entry_t root,
1201 vm_map_entry_t tail, vm_size_t max_free, vm_map_entry_t llist)
1205 * The max_free values of the children of llist are in
1206 * llist->max_free and max_free. Update with the
1209 llist->max_free = max_free =
1210 vm_size_max(llist->max_free, max_free);
1211 vm_map_entry_swap(&llist->right, &tail);
1212 vm_map_entry_swap(&tail, &llist);
1213 } while (llist != header);
1219 * When llist is known to be the predecessor of root.
1221 static inline vm_size_t
1222 vm_map_splay_merge_pred(vm_map_entry_t header, vm_map_entry_t root,
1223 vm_map_entry_t llist)
1227 max_free = root->start - llist->end;
1228 if (llist != header) {
1229 max_free = vm_map_splay_merge_left_walk(header, root,
1230 root, max_free, llist);
1232 root->left = header;
1233 header->right = root;
1239 * When llist may or may not be the predecessor of root.
1241 static inline vm_size_t
1242 vm_map_splay_merge_left(vm_map_entry_t header, vm_map_entry_t root,
1243 vm_map_entry_t llist)
1247 max_free = vm_map_entry_max_free_left(root, llist);
1248 if (llist != header) {
1249 max_free = vm_map_splay_merge_left_walk(header, root,
1250 root->left == llist ? root : root->left,
1257 vm_map_splay_merge_right_walk(vm_map_entry_t header, vm_map_entry_t root,
1258 vm_map_entry_t tail, vm_size_t max_free, vm_map_entry_t rlist)
1262 * The max_free values of the children of rlist are in
1263 * rlist->max_free and max_free. Update with the
1266 rlist->max_free = max_free =
1267 vm_size_max(rlist->max_free, max_free);
1268 vm_map_entry_swap(&rlist->left, &tail);
1269 vm_map_entry_swap(&tail, &rlist);
1270 } while (rlist != header);
1276 * When rlist is known to be the succecessor of root.
1278 static inline vm_size_t
1279 vm_map_splay_merge_succ(vm_map_entry_t header, vm_map_entry_t root,
1280 vm_map_entry_t rlist)
1284 max_free = rlist->start - root->end;
1285 if (rlist != header) {
1286 max_free = vm_map_splay_merge_right_walk(header, root,
1287 root, max_free, rlist);
1289 root->right = header;
1290 header->left = root;
1296 * When rlist may or may not be the succecessor of root.
1298 static inline vm_size_t
1299 vm_map_splay_merge_right(vm_map_entry_t header, vm_map_entry_t root,
1300 vm_map_entry_t rlist)
1304 max_free = vm_map_entry_max_free_right(root, rlist);
1305 if (rlist != header) {
1306 max_free = vm_map_splay_merge_right_walk(header, root,
1307 root->right == rlist ? root : root->right,
1316 * The Sleator and Tarjan top-down splay algorithm with the
1317 * following variation. Max_free must be computed bottom-up, so
1318 * on the downward pass, maintain the left and right spines in
1319 * reverse order. Then, make a second pass up each side to fix
1320 * the pointers and compute max_free. The time bound is O(log n)
1323 * The tree is threaded, which means that there are no null pointers.
1324 * When a node has no left child, its left pointer points to its
1325 * predecessor, which the last ancestor on the search path from the root
1326 * where the search branched right. Likewise, when a node has no right
1327 * child, its right pointer points to its successor. The map header node
1328 * is the predecessor of the first map entry, and the successor of the
1331 * The new root is the vm_map_entry containing "addr", or else an
1332 * adjacent entry (lower if possible) if addr is not in the tree.
1334 * The map must be locked, and leaves it so.
1336 * Returns: the new root.
1338 static vm_map_entry_t
1339 vm_map_splay(vm_map_t map, vm_offset_t addr)
1341 vm_map_entry_t header, llist, rlist, root;
1342 vm_size_t max_free_left, max_free_right;
1344 header = &map->header;
1345 root = vm_map_splay_split(map, addr, 0, &llist, &rlist);
1347 max_free_left = vm_map_splay_merge_left(header, root, llist);
1348 max_free_right = vm_map_splay_merge_right(header, root, rlist);
1349 } else if (llist != header) {
1351 * Recover the greatest node in the left
1352 * subtree and make it the root.
1355 llist = root->right;
1356 max_free_left = vm_map_splay_merge_left(header, root, llist);
1357 max_free_right = vm_map_splay_merge_succ(header, root, rlist);
1358 } else if (rlist != header) {
1360 * Recover the least node in the right
1361 * subtree and make it the root.
1365 max_free_left = vm_map_splay_merge_pred(header, root, llist);
1366 max_free_right = vm_map_splay_merge_right(header, root, rlist);
1368 /* There is no root. */
1371 root->max_free = vm_size_max(max_free_left, max_free_right);
1373 VM_MAP_ASSERT_CONSISTENT(map);
1378 * vm_map_entry_{un,}link:
1380 * Insert/remove entries from maps. On linking, if new entry clips
1381 * existing entry, trim existing entry to avoid overlap, and manage
1382 * offsets. On unlinking, merge disappearing entry with neighbor, if
1383 * called for, and manage offsets. Callers should not modify fields in
1384 * entries already mapped.
1387 vm_map_entry_link(vm_map_t map, vm_map_entry_t entry)
1389 vm_map_entry_t header, llist, rlist, root;
1390 vm_size_t max_free_left, max_free_right;
1393 "vm_map_entry_link: map %p, nentries %d, entry %p", map,
1394 map->nentries, entry);
1395 VM_MAP_ASSERT_LOCKED(map);
1397 header = &map->header;
1398 root = vm_map_splay_split(map, entry->start, 0, &llist, &rlist);
1401 * The new entry does not overlap any existing entry in the
1402 * map, so it becomes the new root of the map tree.
1404 max_free_left = vm_map_splay_merge_pred(header, entry, llist);
1405 max_free_right = vm_map_splay_merge_succ(header, entry, rlist);
1406 } else if (entry->start == root->start) {
1408 * The new entry is a clone of root, with only the end field
1409 * changed. The root entry will be shrunk to abut the new
1410 * entry, and will be the right child of the new root entry in
1413 KASSERT(entry->end < root->end,
1414 ("%s: clip_start not within entry", __func__));
1415 vm_map_splay_findprev(root, &llist);
1416 root->offset += entry->end - root->start;
1417 root->start = entry->end;
1418 max_free_left = vm_map_splay_merge_pred(header, entry, llist);
1419 max_free_right = root->max_free = vm_size_max(
1420 vm_map_splay_merge_pred(entry, root, entry),
1421 vm_map_splay_merge_right(header, root, rlist));
1424 * The new entry is a clone of root, with only the start field
1425 * changed. The root entry will be shrunk to abut the new
1426 * entry, and will be the left child of the new root entry in
1429 KASSERT(entry->end == root->end,
1430 ("%s: clip_start not within entry", __func__));
1431 vm_map_splay_findnext(root, &rlist);
1432 entry->offset += entry->start - root->start;
1433 root->end = entry->start;
1434 max_free_left = root->max_free = vm_size_max(
1435 vm_map_splay_merge_left(header, root, llist),
1436 vm_map_splay_merge_succ(entry, root, entry));
1437 max_free_right = vm_map_splay_merge_succ(header, entry, rlist);
1439 entry->max_free = vm_size_max(max_free_left, max_free_right);
1441 VM_MAP_ASSERT_CONSISTENT(map);
1444 enum unlink_merge_type {
1450 vm_map_entry_unlink(vm_map_t map, vm_map_entry_t entry,
1451 enum unlink_merge_type op)
1453 vm_map_entry_t header, llist, rlist, root;
1454 vm_size_t max_free_left, max_free_right;
1456 VM_MAP_ASSERT_LOCKED(map);
1457 header = &map->header;
1458 root = vm_map_splay_split(map, entry->start, 0, &llist, &rlist);
1459 KASSERT(root != NULL,
1460 ("vm_map_entry_unlink: unlink object not mapped"));
1462 vm_map_splay_findprev(root, &llist);
1463 vm_map_splay_findnext(root, &rlist);
1464 if (op == UNLINK_MERGE_NEXT) {
1465 rlist->start = root->start;
1466 rlist->offset = root->offset;
1468 if (llist != header) {
1470 llist = root->right;
1471 max_free_left = vm_map_splay_merge_left(header, root, llist);
1472 max_free_right = vm_map_splay_merge_succ(header, root, rlist);
1473 } else if (rlist != header) {
1476 max_free_left = vm_map_splay_merge_pred(header, root, llist);
1477 max_free_right = vm_map_splay_merge_right(header, root, rlist);
1479 header->left = header->right = header;
1483 root->max_free = vm_size_max(max_free_left, max_free_right);
1485 VM_MAP_ASSERT_CONSISTENT(map);
1487 CTR3(KTR_VM, "vm_map_entry_unlink: map %p, nentries %d, entry %p", map,
1488 map->nentries, entry);
1492 * vm_map_entry_resize:
1494 * Resize a vm_map_entry, recompute the amount of free space that
1495 * follows it and propagate that value up the tree.
1497 * The map must be locked, and leaves it so.
1500 vm_map_entry_resize(vm_map_t map, vm_map_entry_t entry, vm_size_t grow_amount)
1502 vm_map_entry_t header, llist, rlist, root;
1504 VM_MAP_ASSERT_LOCKED(map);
1505 header = &map->header;
1506 root = vm_map_splay_split(map, entry->start, 0, &llist, &rlist);
1507 KASSERT(root != NULL, ("%s: resize object not mapped", __func__));
1508 vm_map_splay_findnext(root, &rlist);
1509 entry->end += grow_amount;
1510 root->max_free = vm_size_max(
1511 vm_map_splay_merge_left(header, root, llist),
1512 vm_map_splay_merge_succ(header, root, rlist));
1514 VM_MAP_ASSERT_CONSISTENT(map);
1515 CTR4(KTR_VM, "%s: map %p, nentries %d, entry %p",
1516 __func__, map, map->nentries, entry);
1520 * vm_map_lookup_entry: [ internal use only ]
1522 * Finds the map entry containing (or
1523 * immediately preceding) the specified address
1524 * in the given map; the entry is returned
1525 * in the "entry" parameter. The boolean
1526 * result indicates whether the address is
1527 * actually contained in the map.
1530 vm_map_lookup_entry(
1532 vm_offset_t address,
1533 vm_map_entry_t *entry) /* OUT */
1535 vm_map_entry_t cur, header, lbound, ubound;
1539 * If the map is empty, then the map entry immediately preceding
1540 * "address" is the map's header.
1542 header = &map->header;
1548 if (address >= cur->start && cur->end > address) {
1552 if ((locked = vm_map_locked(map)) ||
1553 sx_try_upgrade(&map->lock)) {
1555 * Splay requires a write lock on the map. However, it only
1556 * restructures the binary search tree; it does not otherwise
1557 * change the map. Thus, the map's timestamp need not change
1558 * on a temporary upgrade.
1560 cur = vm_map_splay(map, address);
1562 VM_MAP_UNLOCK_CONSISTENT(map);
1563 sx_downgrade(&map->lock);
1567 * If "address" is contained within a map entry, the new root
1568 * is that map entry. Otherwise, the new root is a map entry
1569 * immediately before or after "address".
1571 if (address < cur->start) {
1576 return (address < cur->end);
1579 * Since the map is only locked for read access, perform a
1580 * standard binary search tree lookup for "address".
1582 lbound = ubound = header;
1584 if (address < cur->start) {
1589 } else if (cur->end <= address) {
1606 * Inserts the given whole VM object into the target
1607 * map at the specified address range. The object's
1608 * size should match that of the address range.
1610 * Requires that the map be locked, and leaves it so.
1612 * If object is non-NULL, ref count must be bumped by caller
1613 * prior to making call to account for the new entry.
1616 vm_map_insert(vm_map_t map, vm_object_t object, vm_ooffset_t offset,
1617 vm_offset_t start, vm_offset_t end, vm_prot_t prot, vm_prot_t max, int cow)
1619 vm_map_entry_t new_entry, next_entry, prev_entry;
1621 vm_eflags_t protoeflags;
1622 vm_inherit_t inheritance;
1626 VM_MAP_ASSERT_LOCKED(map);
1627 KASSERT(object != kernel_object ||
1628 (cow & MAP_COPY_ON_WRITE) == 0,
1629 ("vm_map_insert: kernel object and COW"));
1630 KASSERT(object == NULL || (cow & MAP_NOFAULT) == 0 ||
1631 (cow & MAP_SPLIT_BOUNDARY_MASK) != 0,
1632 ("vm_map_insert: paradoxical MAP_NOFAULT request, obj %p cow %#x",
1634 KASSERT((prot & ~max) == 0,
1635 ("prot %#x is not subset of max_prot %#x", prot, max));
1638 * Check that the start and end points are not bogus.
1640 if (start == end || !vm_map_range_valid(map, start, end))
1641 return (KERN_INVALID_ADDRESS);
1643 if ((map->flags & MAP_WXORX) != 0 && (prot & (VM_PROT_WRITE |
1644 VM_PROT_EXECUTE)) == (VM_PROT_WRITE | VM_PROT_EXECUTE))
1645 return (KERN_PROTECTION_FAILURE);
1648 * Find the entry prior to the proposed starting address; if it's part
1649 * of an existing entry, this range is bogus.
1651 if (vm_map_lookup_entry(map, start, &prev_entry))
1652 return (KERN_NO_SPACE);
1655 * Assert that the next entry doesn't overlap the end point.
1657 next_entry = vm_map_entry_succ(prev_entry);
1658 if (next_entry->start < end)
1659 return (KERN_NO_SPACE);
1661 if ((cow & MAP_CREATE_GUARD) != 0 && (object != NULL ||
1662 max != VM_PROT_NONE))
1663 return (KERN_INVALID_ARGUMENT);
1666 if (cow & MAP_COPY_ON_WRITE)
1667 protoeflags |= MAP_ENTRY_COW | MAP_ENTRY_NEEDS_COPY;
1668 if (cow & MAP_NOFAULT)
1669 protoeflags |= MAP_ENTRY_NOFAULT;
1670 if (cow & MAP_DISABLE_SYNCER)
1671 protoeflags |= MAP_ENTRY_NOSYNC;
1672 if (cow & MAP_DISABLE_COREDUMP)
1673 protoeflags |= MAP_ENTRY_NOCOREDUMP;
1674 if (cow & MAP_STACK_GROWS_DOWN)
1675 protoeflags |= MAP_ENTRY_GROWS_DOWN;
1676 if (cow & MAP_STACK_GROWS_UP)
1677 protoeflags |= MAP_ENTRY_GROWS_UP;
1678 if (cow & MAP_WRITECOUNT)
1679 protoeflags |= MAP_ENTRY_WRITECNT;
1680 if (cow & MAP_VN_EXEC)
1681 protoeflags |= MAP_ENTRY_VN_EXEC;
1682 if ((cow & MAP_CREATE_GUARD) != 0)
1683 protoeflags |= MAP_ENTRY_GUARD;
1684 if ((cow & MAP_CREATE_STACK_GAP_DN) != 0)
1685 protoeflags |= MAP_ENTRY_STACK_GAP_DN;
1686 if ((cow & MAP_CREATE_STACK_GAP_UP) != 0)
1687 protoeflags |= MAP_ENTRY_STACK_GAP_UP;
1688 if (cow & MAP_INHERIT_SHARE)
1689 inheritance = VM_INHERIT_SHARE;
1691 inheritance = VM_INHERIT_DEFAULT;
1692 if ((cow & MAP_SPLIT_BOUNDARY_MASK) != 0) {
1693 /* This magically ignores index 0, for usual page size. */
1694 bidx = (cow & MAP_SPLIT_BOUNDARY_MASK) >>
1695 MAP_SPLIT_BOUNDARY_SHIFT;
1696 if (bidx >= MAXPAGESIZES)
1697 return (KERN_INVALID_ARGUMENT);
1698 bdry = pagesizes[bidx] - 1;
1699 if ((start & bdry) != 0 || (end & bdry) != 0)
1700 return (KERN_INVALID_ARGUMENT);
1701 protoeflags |= bidx << MAP_ENTRY_SPLIT_BOUNDARY_SHIFT;
1705 if ((cow & (MAP_ACC_NO_CHARGE | MAP_NOFAULT | MAP_CREATE_GUARD)) != 0)
1707 if ((cow & MAP_ACC_CHARGED) || ((prot & VM_PROT_WRITE) &&
1708 ((protoeflags & MAP_ENTRY_NEEDS_COPY) || object == NULL))) {
1709 if (!(cow & MAP_ACC_CHARGED) && !swap_reserve(end - start))
1710 return (KERN_RESOURCE_SHORTAGE);
1711 KASSERT(object == NULL ||
1712 (protoeflags & MAP_ENTRY_NEEDS_COPY) != 0 ||
1713 object->cred == NULL,
1714 ("overcommit: vm_map_insert o %p", object));
1715 cred = curthread->td_ucred;
1719 /* Expand the kernel pmap, if necessary. */
1720 if (map == kernel_map && end > kernel_vm_end)
1721 pmap_growkernel(end);
1722 if (object != NULL) {
1724 * OBJ_ONEMAPPING must be cleared unless this mapping
1725 * is trivially proven to be the only mapping for any
1726 * of the object's pages. (Object granularity
1727 * reference counting is insufficient to recognize
1728 * aliases with precision.)
1730 if ((object->flags & OBJ_ANON) != 0) {
1731 VM_OBJECT_WLOCK(object);
1732 if (object->ref_count > 1 || object->shadow_count != 0)
1733 vm_object_clear_flag(object, OBJ_ONEMAPPING);
1734 VM_OBJECT_WUNLOCK(object);
1736 } else if ((prev_entry->eflags & ~MAP_ENTRY_USER_WIRED) ==
1738 (cow & (MAP_STACK_GROWS_DOWN | MAP_STACK_GROWS_UP |
1739 MAP_VN_EXEC)) == 0 &&
1740 prev_entry->end == start && (prev_entry->cred == cred ||
1741 (prev_entry->object.vm_object != NULL &&
1742 prev_entry->object.vm_object->cred == cred)) &&
1743 vm_object_coalesce(prev_entry->object.vm_object,
1745 (vm_size_t)(prev_entry->end - prev_entry->start),
1746 (vm_size_t)(end - prev_entry->end), cred != NULL &&
1747 (protoeflags & MAP_ENTRY_NEEDS_COPY) == 0)) {
1749 * We were able to extend the object. Determine if we
1750 * can extend the previous map entry to include the
1751 * new range as well.
1753 if (prev_entry->inheritance == inheritance &&
1754 prev_entry->protection == prot &&
1755 prev_entry->max_protection == max &&
1756 prev_entry->wired_count == 0) {
1757 KASSERT((prev_entry->eflags & MAP_ENTRY_USER_WIRED) ==
1758 0, ("prev_entry %p has incoherent wiring",
1760 if ((prev_entry->eflags & MAP_ENTRY_GUARD) == 0)
1761 map->size += end - prev_entry->end;
1762 vm_map_entry_resize(map, prev_entry,
1763 end - prev_entry->end);
1764 vm_map_try_merge_entries(map, prev_entry, next_entry);
1765 return (KERN_SUCCESS);
1769 * If we can extend the object but cannot extend the
1770 * map entry, we have to create a new map entry. We
1771 * must bump the ref count on the extended object to
1772 * account for it. object may be NULL.
1774 object = prev_entry->object.vm_object;
1775 offset = prev_entry->offset +
1776 (prev_entry->end - prev_entry->start);
1777 vm_object_reference(object);
1778 if (cred != NULL && object != NULL && object->cred != NULL &&
1779 !(prev_entry->eflags & MAP_ENTRY_NEEDS_COPY)) {
1780 /* Object already accounts for this uid. */
1788 * Create a new entry
1790 new_entry = vm_map_entry_create(map);
1791 new_entry->start = start;
1792 new_entry->end = end;
1793 new_entry->cred = NULL;
1795 new_entry->eflags = protoeflags;
1796 new_entry->object.vm_object = object;
1797 new_entry->offset = offset;
1799 new_entry->inheritance = inheritance;
1800 new_entry->protection = prot;
1801 new_entry->max_protection = max;
1802 new_entry->wired_count = 0;
1803 new_entry->wiring_thread = NULL;
1804 new_entry->read_ahead = VM_FAULT_READ_AHEAD_INIT;
1805 new_entry->next_read = start;
1807 KASSERT(cred == NULL || !ENTRY_CHARGED(new_entry),
1808 ("overcommit: vm_map_insert leaks vm_map %p", new_entry));
1809 new_entry->cred = cred;
1812 * Insert the new entry into the list
1814 vm_map_entry_link(map, new_entry);
1815 if ((new_entry->eflags & MAP_ENTRY_GUARD) == 0)
1816 map->size += new_entry->end - new_entry->start;
1819 * Try to coalesce the new entry with both the previous and next
1820 * entries in the list. Previously, we only attempted to coalesce
1821 * with the previous entry when object is NULL. Here, we handle the
1822 * other cases, which are less common.
1824 vm_map_try_merge_entries(map, prev_entry, new_entry);
1825 vm_map_try_merge_entries(map, new_entry, next_entry);
1827 if ((cow & (MAP_PREFAULT | MAP_PREFAULT_PARTIAL)) != 0) {
1828 vm_map_pmap_enter(map, start, prot, object, OFF_TO_IDX(offset),
1829 end - start, cow & MAP_PREFAULT_PARTIAL);
1832 return (KERN_SUCCESS);
1838 * Find the first fit (lowest VM address) for "length" free bytes
1839 * beginning at address >= start in the given map.
1841 * In a vm_map_entry, "max_free" is the maximum amount of
1842 * contiguous free space between an entry in its subtree and a
1843 * neighbor of that entry. This allows finding a free region in
1844 * one path down the tree, so O(log n) amortized with splay
1847 * The map must be locked, and leaves it so.
1849 * Returns: starting address if sufficient space,
1850 * vm_map_max(map)-length+1 if insufficient space.
1853 vm_map_findspace(vm_map_t map, vm_offset_t start, vm_size_t length)
1855 vm_map_entry_t header, llist, rlist, root, y;
1856 vm_size_t left_length, max_free_left, max_free_right;
1857 vm_offset_t gap_end;
1859 VM_MAP_ASSERT_LOCKED(map);
1862 * Request must fit within min/max VM address and must avoid
1865 start = MAX(start, vm_map_min(map));
1866 if (start >= vm_map_max(map) || length > vm_map_max(map) - start)
1867 return (vm_map_max(map) - length + 1);
1869 /* Empty tree means wide open address space. */
1870 if (map->root == NULL)
1874 * After splay_split, if start is within an entry, push it to the start
1875 * of the following gap. If rlist is at the end of the gap containing
1876 * start, save the end of that gap in gap_end to see if the gap is big
1877 * enough; otherwise set gap_end to start skip gap-checking and move
1878 * directly to a search of the right subtree.
1880 header = &map->header;
1881 root = vm_map_splay_split(map, start, length, &llist, &rlist);
1882 gap_end = rlist->start;
1885 if (root->right != rlist)
1887 max_free_left = vm_map_splay_merge_left(header, root, llist);
1888 max_free_right = vm_map_splay_merge_right(header, root, rlist);
1889 } else if (rlist != header) {
1892 max_free_left = vm_map_splay_merge_pred(header, root, llist);
1893 max_free_right = vm_map_splay_merge_right(header, root, rlist);
1896 llist = root->right;
1897 max_free_left = vm_map_splay_merge_left(header, root, llist);
1898 max_free_right = vm_map_splay_merge_succ(header, root, rlist);
1900 root->max_free = vm_size_max(max_free_left, max_free_right);
1902 VM_MAP_ASSERT_CONSISTENT(map);
1903 if (length <= gap_end - start)
1906 /* With max_free, can immediately tell if no solution. */
1907 if (root->right == header || length > root->right->max_free)
1908 return (vm_map_max(map) - length + 1);
1911 * Splay for the least large-enough gap in the right subtree.
1913 llist = rlist = header;
1914 for (left_length = 0;;
1915 left_length = vm_map_entry_max_free_left(root, llist)) {
1916 if (length <= left_length)
1917 SPLAY_LEFT_STEP(root, y, llist, rlist,
1918 length <= vm_map_entry_max_free_left(y, llist));
1920 SPLAY_RIGHT_STEP(root, y, llist, rlist,
1921 length > vm_map_entry_max_free_left(y, root));
1926 llist = root->right;
1927 max_free_left = vm_map_splay_merge_left(header, root, llist);
1928 if (rlist == header) {
1929 root->max_free = vm_size_max(max_free_left,
1930 vm_map_splay_merge_succ(header, root, rlist));
1934 y->max_free = vm_size_max(
1935 vm_map_splay_merge_pred(root, y, root),
1936 vm_map_splay_merge_right(header, y, rlist));
1937 root->max_free = vm_size_max(max_free_left, y->max_free);
1940 VM_MAP_ASSERT_CONSISTENT(map);
1945 vm_map_fixed(vm_map_t map, vm_object_t object, vm_ooffset_t offset,
1946 vm_offset_t start, vm_size_t length, vm_prot_t prot,
1947 vm_prot_t max, int cow)
1952 end = start + length;
1953 KASSERT((cow & (MAP_STACK_GROWS_DOWN | MAP_STACK_GROWS_UP)) == 0 ||
1955 ("vm_map_fixed: non-NULL backing object for stack"));
1957 VM_MAP_RANGE_CHECK(map, start, end);
1958 if ((cow & MAP_CHECK_EXCL) == 0) {
1959 result = vm_map_delete(map, start, end);
1960 if (result != KERN_SUCCESS)
1963 if ((cow & (MAP_STACK_GROWS_DOWN | MAP_STACK_GROWS_UP)) != 0) {
1964 result = vm_map_stack_locked(map, start, length, sgrowsiz,
1967 result = vm_map_insert(map, object, offset, start, end,
1975 static const int aslr_pages_rnd_64[2] = {0x1000, 0x10};
1976 static const int aslr_pages_rnd_32[2] = {0x100, 0x4};
1978 static int cluster_anon = 1;
1979 SYSCTL_INT(_vm, OID_AUTO, cluster_anon, CTLFLAG_RW,
1981 "Cluster anonymous mappings: 0 = no, 1 = yes if no hint, 2 = always");
1984 clustering_anon_allowed(vm_offset_t addr)
1987 switch (cluster_anon) {
1998 static long aslr_restarts;
1999 SYSCTL_LONG(_vm, OID_AUTO, aslr_restarts, CTLFLAG_RD,
2001 "Number of aslr failures");
2004 * Searches for the specified amount of free space in the given map with the
2005 * specified alignment. Performs an address-ordered, first-fit search from
2006 * the given address "*addr", with an optional upper bound "max_addr". If the
2007 * parameter "alignment" is zero, then the alignment is computed from the
2008 * given (object, offset) pair so as to enable the greatest possible use of
2009 * superpage mappings. Returns KERN_SUCCESS and the address of the free space
2010 * in "*addr" if successful. Otherwise, returns KERN_NO_SPACE.
2012 * The map must be locked. Initially, there must be at least "length" bytes
2013 * of free space at the given address.
2016 vm_map_alignspace(vm_map_t map, vm_object_t object, vm_ooffset_t offset,
2017 vm_offset_t *addr, vm_size_t length, vm_offset_t max_addr,
2018 vm_offset_t alignment)
2020 vm_offset_t aligned_addr, free_addr;
2022 VM_MAP_ASSERT_LOCKED(map);
2024 KASSERT(free_addr == vm_map_findspace(map, free_addr, length),
2025 ("caller failed to provide space %#jx at address %p",
2026 (uintmax_t)length, (void *)free_addr));
2029 * At the start of every iteration, the free space at address
2030 * "*addr" is at least "length" bytes.
2033 pmap_align_superpage(object, offset, addr, length);
2034 else if ((*addr & (alignment - 1)) != 0) {
2035 *addr &= ~(alignment - 1);
2038 aligned_addr = *addr;
2039 if (aligned_addr == free_addr) {
2041 * Alignment did not change "*addr", so "*addr" must
2042 * still provide sufficient free space.
2044 return (KERN_SUCCESS);
2048 * Test for address wrap on "*addr". A wrapped "*addr" could
2049 * be a valid address, in which case vm_map_findspace() cannot
2050 * be relied upon to fail.
2052 if (aligned_addr < free_addr)
2053 return (KERN_NO_SPACE);
2054 *addr = vm_map_findspace(map, aligned_addr, length);
2055 if (*addr + length > vm_map_max(map) ||
2056 (max_addr != 0 && *addr + length > max_addr))
2057 return (KERN_NO_SPACE);
2059 if (free_addr == aligned_addr) {
2061 * If a successful call to vm_map_findspace() did not
2062 * change "*addr", then "*addr" must still be aligned
2063 * and provide sufficient free space.
2065 return (KERN_SUCCESS);
2071 vm_map_find_aligned(vm_map_t map, vm_offset_t *addr, vm_size_t length,
2072 vm_offset_t max_addr, vm_offset_t alignment)
2074 /* XXXKIB ASLR eh ? */
2075 *addr = vm_map_findspace(map, *addr, length);
2076 if (*addr + length > vm_map_max(map) ||
2077 (max_addr != 0 && *addr + length > max_addr))
2078 return (KERN_NO_SPACE);
2079 return (vm_map_alignspace(map, NULL, 0, addr, length, max_addr,
2084 * vm_map_find finds an unallocated region in the target address
2085 * map with the given length. The search is defined to be
2086 * first-fit from the specified address; the region found is
2087 * returned in the same parameter.
2089 * If object is non-NULL, ref count must be bumped by caller
2090 * prior to making call to account for the new entry.
2093 vm_map_find(vm_map_t map, vm_object_t object, vm_ooffset_t offset,
2094 vm_offset_t *addr, /* IN/OUT */
2095 vm_size_t length, vm_offset_t max_addr, int find_space,
2096 vm_prot_t prot, vm_prot_t max, int cow)
2098 vm_offset_t alignment, curr_min_addr, min_addr;
2099 int gap, pidx, rv, try;
2100 bool cluster, en_aslr, update_anon;
2102 KASSERT((cow & (MAP_STACK_GROWS_DOWN | MAP_STACK_GROWS_UP)) == 0 ||
2104 ("vm_map_find: non-NULL backing object for stack"));
2105 MPASS((cow & MAP_REMAP) == 0 || (find_space == VMFS_NO_SPACE &&
2106 (cow & (MAP_STACK_GROWS_DOWN | MAP_STACK_GROWS_UP)) == 0));
2107 if (find_space == VMFS_OPTIMAL_SPACE && (object == NULL ||
2108 (object->flags & OBJ_COLORED) == 0))
2109 find_space = VMFS_ANY_SPACE;
2110 if (find_space >> 8 != 0) {
2111 KASSERT((find_space & 0xff) == 0, ("bad VMFS flags"));
2112 alignment = (vm_offset_t)1 << (find_space >> 8);
2115 en_aslr = (map->flags & MAP_ASLR) != 0;
2116 update_anon = cluster = clustering_anon_allowed(*addr) &&
2117 (map->flags & MAP_IS_SUB_MAP) == 0 && max_addr == 0 &&
2118 find_space != VMFS_NO_SPACE && object == NULL &&
2119 (cow & (MAP_INHERIT_SHARE | MAP_STACK_GROWS_UP |
2120 MAP_STACK_GROWS_DOWN)) == 0 && prot != PROT_NONE;
2121 curr_min_addr = min_addr = *addr;
2122 if (en_aslr && min_addr == 0 && !cluster &&
2123 find_space != VMFS_NO_SPACE &&
2124 (map->flags & MAP_ASLR_IGNSTART) != 0)
2125 curr_min_addr = min_addr = vm_map_min(map);
2129 curr_min_addr = map->anon_loc;
2130 if (curr_min_addr == 0)
2133 if (find_space != VMFS_NO_SPACE) {
2134 KASSERT(find_space == VMFS_ANY_SPACE ||
2135 find_space == VMFS_OPTIMAL_SPACE ||
2136 find_space == VMFS_SUPER_SPACE ||
2137 alignment != 0, ("unexpected VMFS flag"));
2140 * When creating an anonymous mapping, try clustering
2141 * with an existing anonymous mapping first.
2143 * We make up to two attempts to find address space
2144 * for a given find_space value. The first attempt may
2145 * apply randomization or may cluster with an existing
2146 * anonymous mapping. If this first attempt fails,
2147 * perform a first-fit search of the available address
2150 * If all tries failed, and find_space is
2151 * VMFS_OPTIMAL_SPACE, fallback to VMFS_ANY_SPACE.
2152 * Again enable clustering and randomization.
2159 * Second try: we failed either to find a
2160 * suitable region for randomizing the
2161 * allocation, or to cluster with an existing
2162 * mapping. Retry with free run.
2164 curr_min_addr = (map->flags & MAP_ASLR_IGNSTART) != 0 ?
2165 vm_map_min(map) : min_addr;
2166 atomic_add_long(&aslr_restarts, 1);
2169 if (try == 1 && en_aslr && !cluster) {
2171 * Find space for allocation, including
2172 * gap needed for later randomization.
2174 pidx = MAXPAGESIZES > 1 && pagesizes[1] != 0 &&
2175 (find_space == VMFS_SUPER_SPACE || find_space ==
2176 VMFS_OPTIMAL_SPACE) ? 1 : 0;
2177 gap = vm_map_max(map) > MAP_32BIT_MAX_ADDR &&
2178 (max_addr == 0 || max_addr > MAP_32BIT_MAX_ADDR) ?
2179 aslr_pages_rnd_64[pidx] : aslr_pages_rnd_32[pidx];
2180 *addr = vm_map_findspace(map, curr_min_addr,
2181 length + gap * pagesizes[pidx]);
2182 if (*addr + length + gap * pagesizes[pidx] >
2185 /* And randomize the start address. */
2186 *addr += (arc4random() % gap) * pagesizes[pidx];
2187 if (max_addr != 0 && *addr + length > max_addr)
2190 *addr = vm_map_findspace(map, curr_min_addr, length);
2191 if (*addr + length > vm_map_max(map) ||
2192 (max_addr != 0 && *addr + length > max_addr)) {
2203 if (find_space != VMFS_ANY_SPACE &&
2204 (rv = vm_map_alignspace(map, object, offset, addr, length,
2205 max_addr, alignment)) != KERN_SUCCESS) {
2206 if (find_space == VMFS_OPTIMAL_SPACE) {
2207 find_space = VMFS_ANY_SPACE;
2208 curr_min_addr = min_addr;
2209 cluster = update_anon;
2215 } else if ((cow & MAP_REMAP) != 0) {
2216 if (!vm_map_range_valid(map, *addr, *addr + length)) {
2217 rv = KERN_INVALID_ADDRESS;
2220 rv = vm_map_delete(map, *addr, *addr + length);
2221 if (rv != KERN_SUCCESS)
2224 if ((cow & (MAP_STACK_GROWS_DOWN | MAP_STACK_GROWS_UP)) != 0) {
2225 rv = vm_map_stack_locked(map, *addr, length, sgrowsiz, prot,
2228 rv = vm_map_insert(map, object, offset, *addr, *addr + length,
2231 if (rv == KERN_SUCCESS && update_anon)
2232 map->anon_loc = *addr + length;
2239 * vm_map_find_min() is a variant of vm_map_find() that takes an
2240 * additional parameter (min_addr) and treats the given address
2241 * (*addr) differently. Specifically, it treats *addr as a hint
2242 * and not as the minimum address where the mapping is created.
2244 * This function works in two phases. First, it tries to
2245 * allocate above the hint. If that fails and the hint is
2246 * greater than min_addr, it performs a second pass, replacing
2247 * the hint with min_addr as the minimum address for the
2251 vm_map_find_min(vm_map_t map, vm_object_t object, vm_ooffset_t offset,
2252 vm_offset_t *addr, vm_size_t length, vm_offset_t min_addr,
2253 vm_offset_t max_addr, int find_space, vm_prot_t prot, vm_prot_t max,
2261 rv = vm_map_find(map, object, offset, addr, length, max_addr,
2262 find_space, prot, max, cow);
2263 if (rv == KERN_SUCCESS || min_addr >= hint)
2265 *addr = hint = min_addr;
2270 * A map entry with any of the following flags set must not be merged with
2273 #define MAP_ENTRY_NOMERGE_MASK (MAP_ENTRY_GROWS_DOWN | MAP_ENTRY_GROWS_UP | \
2274 MAP_ENTRY_IN_TRANSITION | MAP_ENTRY_IS_SUB_MAP | MAP_ENTRY_VN_EXEC)
2277 vm_map_mergeable_neighbors(vm_map_entry_t prev, vm_map_entry_t entry)
2280 KASSERT((prev->eflags & MAP_ENTRY_NOMERGE_MASK) == 0 ||
2281 (entry->eflags & MAP_ENTRY_NOMERGE_MASK) == 0,
2282 ("vm_map_mergeable_neighbors: neither %p nor %p are mergeable",
2284 return (prev->end == entry->start &&
2285 prev->object.vm_object == entry->object.vm_object &&
2286 (prev->object.vm_object == NULL ||
2287 prev->offset + (prev->end - prev->start) == entry->offset) &&
2288 prev->eflags == entry->eflags &&
2289 prev->protection == entry->protection &&
2290 prev->max_protection == entry->max_protection &&
2291 prev->inheritance == entry->inheritance &&
2292 prev->wired_count == entry->wired_count &&
2293 prev->cred == entry->cred);
2297 vm_map_merged_neighbor_dispose(vm_map_t map, vm_map_entry_t entry)
2301 * If the backing object is a vnode object, vm_object_deallocate()
2302 * calls vrele(). However, vrele() does not lock the vnode because
2303 * the vnode has additional references. Thus, the map lock can be
2304 * kept without causing a lock-order reversal with the vnode lock.
2306 * Since we count the number of virtual page mappings in
2307 * object->un_pager.vnp.writemappings, the writemappings value
2308 * should not be adjusted when the entry is disposed of.
2310 if (entry->object.vm_object != NULL)
2311 vm_object_deallocate(entry->object.vm_object);
2312 if (entry->cred != NULL)
2313 crfree(entry->cred);
2314 vm_map_entry_dispose(map, entry);
2318 * vm_map_try_merge_entries:
2320 * Compare the given map entry to its predecessor, and merge its precessor
2321 * into it if possible. The entry remains valid, and may be extended.
2322 * The predecessor may be deleted.
2324 * The map must be locked.
2327 vm_map_try_merge_entries(vm_map_t map, vm_map_entry_t prev_entry,
2328 vm_map_entry_t entry)
2331 VM_MAP_ASSERT_LOCKED(map);
2332 if ((entry->eflags & MAP_ENTRY_NOMERGE_MASK) == 0 &&
2333 vm_map_mergeable_neighbors(prev_entry, entry)) {
2334 vm_map_entry_unlink(map, prev_entry, UNLINK_MERGE_NEXT);
2335 vm_map_merged_neighbor_dispose(map, prev_entry);
2340 * vm_map_entry_back:
2342 * Allocate an object to back a map entry.
2345 vm_map_entry_back(vm_map_entry_t entry)
2349 KASSERT(entry->object.vm_object == NULL,
2350 ("map entry %p has backing object", entry));
2351 KASSERT((entry->eflags & MAP_ENTRY_IS_SUB_MAP) == 0,
2352 ("map entry %p is a submap", entry));
2353 object = vm_object_allocate_anon(atop(entry->end - entry->start), NULL,
2354 entry->cred, entry->end - entry->start);
2355 entry->object.vm_object = object;
2361 * vm_map_entry_charge_object
2363 * If there is no object backing this entry, create one. Otherwise, if
2364 * the entry has cred, give it to the backing object.
2367 vm_map_entry_charge_object(vm_map_t map, vm_map_entry_t entry)
2370 VM_MAP_ASSERT_LOCKED(map);
2371 KASSERT((entry->eflags & MAP_ENTRY_IS_SUB_MAP) == 0,
2372 ("map entry %p is a submap", entry));
2373 if (entry->object.vm_object == NULL && !map->system_map &&
2374 (entry->eflags & MAP_ENTRY_GUARD) == 0)
2375 vm_map_entry_back(entry);
2376 else if (entry->object.vm_object != NULL &&
2377 ((entry->eflags & MAP_ENTRY_NEEDS_COPY) == 0) &&
2378 entry->cred != NULL) {
2379 VM_OBJECT_WLOCK(entry->object.vm_object);
2380 KASSERT(entry->object.vm_object->cred == NULL,
2381 ("OVERCOMMIT: %s: both cred e %p", __func__, entry));
2382 entry->object.vm_object->cred = entry->cred;
2383 entry->object.vm_object->charge = entry->end - entry->start;
2384 VM_OBJECT_WUNLOCK(entry->object.vm_object);
2390 * vm_map_entry_clone
2392 * Create a duplicate map entry for clipping.
2394 static vm_map_entry_t
2395 vm_map_entry_clone(vm_map_t map, vm_map_entry_t entry)
2397 vm_map_entry_t new_entry;
2399 VM_MAP_ASSERT_LOCKED(map);
2402 * Create a backing object now, if none exists, so that more individual
2403 * objects won't be created after the map entry is split.
2405 vm_map_entry_charge_object(map, entry);
2407 /* Clone the entry. */
2408 new_entry = vm_map_entry_create(map);
2409 *new_entry = *entry;
2410 if (new_entry->cred != NULL)
2411 crhold(entry->cred);
2412 if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
2413 vm_object_reference(new_entry->object.vm_object);
2414 vm_map_entry_set_vnode_text(new_entry, true);
2416 * The object->un_pager.vnp.writemappings for the object of
2417 * MAP_ENTRY_WRITECNT type entry shall be kept as is here. The
2418 * virtual pages are re-distributed among the clipped entries,
2419 * so the sum is left the same.
2426 * vm_map_clip_start: [ internal use only ]
2428 * Asserts that the given entry begins at or after
2429 * the specified address; if necessary,
2430 * it splits the entry into two.
2433 vm_map_clip_start(vm_map_t map, vm_map_entry_t entry, vm_offset_t startaddr)
2435 vm_map_entry_t new_entry;
2438 if (!map->system_map)
2439 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
2440 "%s: map %p entry %p start 0x%jx", __func__, map, entry,
2441 (uintmax_t)startaddr);
2443 if (startaddr <= entry->start)
2444 return (KERN_SUCCESS);
2446 VM_MAP_ASSERT_LOCKED(map);
2447 KASSERT(entry->end > startaddr && entry->start < startaddr,
2448 ("%s: invalid clip of entry %p", __func__, entry));
2450 bdry_idx = (entry->eflags & MAP_ENTRY_SPLIT_BOUNDARY_MASK) >>
2451 MAP_ENTRY_SPLIT_BOUNDARY_SHIFT;
2452 if (bdry_idx != 0) {
2453 if ((startaddr & (pagesizes[bdry_idx] - 1)) != 0)
2454 return (KERN_INVALID_ARGUMENT);
2457 new_entry = vm_map_entry_clone(map, entry);
2460 * Split off the front portion. Insert the new entry BEFORE this one,
2461 * so that this entry has the specified starting address.
2463 new_entry->end = startaddr;
2464 vm_map_entry_link(map, new_entry);
2465 return (KERN_SUCCESS);
2469 * vm_map_lookup_clip_start:
2471 * Find the entry at or just after 'start', and clip it if 'start' is in
2472 * the interior of the entry. Return entry after 'start', and in
2473 * prev_entry set the entry before 'start'.
2476 vm_map_lookup_clip_start(vm_map_t map, vm_offset_t start,
2477 vm_map_entry_t *res_entry, vm_map_entry_t *prev_entry)
2479 vm_map_entry_t entry;
2482 if (!map->system_map)
2483 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
2484 "%s: map %p start 0x%jx prev %p", __func__, map,
2485 (uintmax_t)start, prev_entry);
2487 if (vm_map_lookup_entry(map, start, prev_entry)) {
2488 entry = *prev_entry;
2489 rv = vm_map_clip_start(map, entry, start);
2490 if (rv != KERN_SUCCESS)
2492 *prev_entry = vm_map_entry_pred(entry);
2494 entry = vm_map_entry_succ(*prev_entry);
2496 return (KERN_SUCCESS);
2500 * vm_map_clip_end: [ internal use only ]
2502 * Asserts that the given entry ends at or before
2503 * the specified address; if necessary,
2504 * it splits the entry into two.
2507 vm_map_clip_end(vm_map_t map, vm_map_entry_t entry, vm_offset_t endaddr)
2509 vm_map_entry_t new_entry;
2512 if (!map->system_map)
2513 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
2514 "%s: map %p entry %p end 0x%jx", __func__, map, entry,
2515 (uintmax_t)endaddr);
2517 if (endaddr >= entry->end)
2518 return (KERN_SUCCESS);
2520 VM_MAP_ASSERT_LOCKED(map);
2521 KASSERT(entry->start < endaddr && entry->end > endaddr,
2522 ("%s: invalid clip of entry %p", __func__, entry));
2524 bdry_idx = (entry->eflags & MAP_ENTRY_SPLIT_BOUNDARY_MASK) >>
2525 MAP_ENTRY_SPLIT_BOUNDARY_SHIFT;
2526 if (bdry_idx != 0) {
2527 if ((endaddr & (pagesizes[bdry_idx] - 1)) != 0)
2528 return (KERN_INVALID_ARGUMENT);
2531 new_entry = vm_map_entry_clone(map, entry);
2534 * Split off the back portion. Insert the new entry AFTER this one,
2535 * so that this entry has the specified ending address.
2537 new_entry->start = endaddr;
2538 vm_map_entry_link(map, new_entry);
2540 return (KERN_SUCCESS);
2544 * vm_map_submap: [ kernel use only ]
2546 * Mark the given range as handled by a subordinate map.
2548 * This range must have been created with vm_map_find,
2549 * and no other operations may have been performed on this
2550 * range prior to calling vm_map_submap.
2552 * Only a limited number of operations can be performed
2553 * within this rage after calling vm_map_submap:
2555 * [Don't try vm_map_copy!]
2557 * To remove a submapping, one must first remove the
2558 * range from the superior map, and then destroy the
2559 * submap (if desired). [Better yet, don't try it.]
2568 vm_map_entry_t entry;
2571 result = KERN_INVALID_ARGUMENT;
2573 vm_map_lock(submap);
2574 submap->flags |= MAP_IS_SUB_MAP;
2575 vm_map_unlock(submap);
2578 VM_MAP_RANGE_CHECK(map, start, end);
2579 if (vm_map_lookup_entry(map, start, &entry) && entry->end >= end &&
2580 (entry->eflags & MAP_ENTRY_COW) == 0 &&
2581 entry->object.vm_object == NULL) {
2582 result = vm_map_clip_start(map, entry, start);
2583 if (result != KERN_SUCCESS)
2585 result = vm_map_clip_end(map, entry, end);
2586 if (result != KERN_SUCCESS)
2588 entry->object.sub_map = submap;
2589 entry->eflags |= MAP_ENTRY_IS_SUB_MAP;
2590 result = KERN_SUCCESS;
2595 if (result != KERN_SUCCESS) {
2596 vm_map_lock(submap);
2597 submap->flags &= ~MAP_IS_SUB_MAP;
2598 vm_map_unlock(submap);
2604 * The maximum number of pages to map if MAP_PREFAULT_PARTIAL is specified
2606 #define MAX_INIT_PT 96
2609 * vm_map_pmap_enter:
2611 * Preload the specified map's pmap with mappings to the specified
2612 * object's memory-resident pages. No further physical pages are
2613 * allocated, and no further virtual pages are retrieved from secondary
2614 * storage. If the specified flags include MAP_PREFAULT_PARTIAL, then a
2615 * limited number of page mappings are created at the low-end of the
2616 * specified address range. (For this purpose, a superpage mapping
2617 * counts as one page mapping.) Otherwise, all resident pages within
2618 * the specified address range are mapped.
2621 vm_map_pmap_enter(vm_map_t map, vm_offset_t addr, vm_prot_t prot,
2622 vm_object_t object, vm_pindex_t pindex, vm_size_t size, int flags)
2625 vm_page_t p, p_start;
2626 vm_pindex_t mask, psize, threshold, tmpidx;
2628 if ((prot & (VM_PROT_READ | VM_PROT_EXECUTE)) == 0 || object == NULL)
2630 if (object->type == OBJT_DEVICE || object->type == OBJT_SG) {
2631 VM_OBJECT_WLOCK(object);
2632 if (object->type == OBJT_DEVICE || object->type == OBJT_SG) {
2633 pmap_object_init_pt(map->pmap, addr, object, pindex,
2635 VM_OBJECT_WUNLOCK(object);
2638 VM_OBJECT_LOCK_DOWNGRADE(object);
2640 VM_OBJECT_RLOCK(object);
2643 if (psize + pindex > object->size) {
2644 if (pindex >= object->size) {
2645 VM_OBJECT_RUNLOCK(object);
2648 psize = object->size - pindex;
2653 threshold = MAX_INIT_PT;
2655 p = vm_page_find_least(object, pindex);
2657 * Assert: the variable p is either (1) the page with the
2658 * least pindex greater than or equal to the parameter pindex
2662 p != NULL && (tmpidx = p->pindex - pindex) < psize;
2663 p = TAILQ_NEXT(p, listq)) {
2665 * don't allow an madvise to blow away our really
2666 * free pages allocating pv entries.
2668 if (((flags & MAP_PREFAULT_MADVISE) != 0 &&
2669 vm_page_count_severe()) ||
2670 ((flags & MAP_PREFAULT_PARTIAL) != 0 &&
2671 tmpidx >= threshold)) {
2675 if (vm_page_all_valid(p)) {
2676 if (p_start == NULL) {
2677 start = addr + ptoa(tmpidx);
2680 /* Jump ahead if a superpage mapping is possible. */
2681 if (p->psind > 0 && ((addr + ptoa(tmpidx)) &
2682 (pagesizes[p->psind] - 1)) == 0) {
2683 mask = atop(pagesizes[p->psind]) - 1;
2684 if (tmpidx + mask < psize &&
2685 vm_page_ps_test(p, PS_ALL_VALID, NULL)) {
2690 } else if (p_start != NULL) {
2691 pmap_enter_object(map->pmap, start, addr +
2692 ptoa(tmpidx), p_start, prot);
2696 if (p_start != NULL)
2697 pmap_enter_object(map->pmap, start, addr + ptoa(psize),
2699 VM_OBJECT_RUNLOCK(object);
2705 * Sets the protection and/or the maximum protection of the
2706 * specified address region in the target map.
2709 vm_map_protect(vm_map_t map, vm_offset_t start, vm_offset_t end,
2710 vm_prot_t new_prot, vm_prot_t new_maxprot, int flags)
2712 vm_map_entry_t entry, first_entry, in_tran, prev_entry;
2719 return (KERN_SUCCESS);
2721 if ((flags & (VM_MAP_PROTECT_SET_PROT | VM_MAP_PROTECT_SET_MAXPROT)) ==
2722 (VM_MAP_PROTECT_SET_PROT | VM_MAP_PROTECT_SET_MAXPROT) &&
2723 (new_prot & new_maxprot) != new_prot)
2724 return (KERN_OUT_OF_BOUNDS);
2730 if ((map->flags & MAP_WXORX) != 0 &&
2731 (flags & VM_MAP_PROTECT_SET_PROT) != 0 &&
2732 (new_prot & (VM_PROT_WRITE | VM_PROT_EXECUTE)) == (VM_PROT_WRITE |
2735 return (KERN_PROTECTION_FAILURE);
2739 * Ensure that we are not concurrently wiring pages. vm_map_wire() may
2740 * need to fault pages into the map and will drop the map lock while
2741 * doing so, and the VM object may end up in an inconsistent state if we
2742 * update the protection on the map entry in between faults.
2744 vm_map_wait_busy(map);
2746 VM_MAP_RANGE_CHECK(map, start, end);
2748 if (!vm_map_lookup_entry(map, start, &first_entry))
2749 first_entry = vm_map_entry_succ(first_entry);
2752 * Make a first pass to check for protection violations.
2754 for (entry = first_entry; entry->start < end;
2755 entry = vm_map_entry_succ(entry)) {
2756 if ((entry->eflags & MAP_ENTRY_GUARD) != 0)
2758 if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) != 0) {
2760 return (KERN_INVALID_ARGUMENT);
2762 if ((flags & VM_MAP_PROTECT_SET_PROT) == 0)
2763 new_prot = entry->protection;
2764 if ((flags & VM_MAP_PROTECT_SET_MAXPROT) == 0)
2765 new_maxprot = entry->max_protection;
2766 if ((new_prot & entry->max_protection) != new_prot ||
2767 (new_maxprot & entry->max_protection) != new_maxprot) {
2769 return (KERN_PROTECTION_FAILURE);
2771 if ((entry->eflags & MAP_ENTRY_IN_TRANSITION) != 0)
2776 * Postpone the operation until all in-transition map entries have
2777 * stabilized. An in-transition entry might already have its pages
2778 * wired and wired_count incremented, but not yet have its
2779 * MAP_ENTRY_USER_WIRED flag set. In which case, we would fail to call
2780 * vm_fault_copy_entry() in the final loop below.
2782 if (in_tran != NULL) {
2783 in_tran->eflags |= MAP_ENTRY_NEEDS_WAKEUP;
2784 vm_map_unlock_and_wait(map, 0);
2789 * Before changing the protections, try to reserve swap space for any
2790 * private (i.e., copy-on-write) mappings that are transitioning from
2791 * read-only to read/write access. If a reservation fails, break out
2792 * of this loop early and let the next loop simplify the entries, since
2793 * some may now be mergeable.
2795 rv = vm_map_clip_start(map, first_entry, start);
2796 if (rv != KERN_SUCCESS) {
2800 for (entry = first_entry; entry->start < end;
2801 entry = vm_map_entry_succ(entry)) {
2802 rv = vm_map_clip_end(map, entry, end);
2803 if (rv != KERN_SUCCESS) {
2808 if ((flags & VM_MAP_PROTECT_SET_PROT) == 0 ||
2809 ((new_prot & ~entry->protection) & VM_PROT_WRITE) == 0 ||
2810 ENTRY_CHARGED(entry) ||
2811 (entry->eflags & MAP_ENTRY_GUARD) != 0)
2814 cred = curthread->td_ucred;
2815 obj = entry->object.vm_object;
2818 (entry->eflags & MAP_ENTRY_NEEDS_COPY) != 0) {
2819 if (!swap_reserve(entry->end - entry->start)) {
2820 rv = KERN_RESOURCE_SHORTAGE;
2829 if (obj->type != OBJT_DEFAULT &&
2830 (obj->flags & OBJ_SWAP) == 0)
2832 VM_OBJECT_WLOCK(obj);
2833 if (obj->type != OBJT_DEFAULT &&
2834 (obj->flags & OBJ_SWAP) == 0) {
2835 VM_OBJECT_WUNLOCK(obj);
2840 * Charge for the whole object allocation now, since
2841 * we cannot distinguish between non-charged and
2842 * charged clipped mapping of the same object later.
2844 KASSERT(obj->charge == 0,
2845 ("vm_map_protect: object %p overcharged (entry %p)",
2847 if (!swap_reserve(ptoa(obj->size))) {
2848 VM_OBJECT_WUNLOCK(obj);
2849 rv = KERN_RESOURCE_SHORTAGE;
2856 obj->charge = ptoa(obj->size);
2857 VM_OBJECT_WUNLOCK(obj);
2861 * If enough swap space was available, go back and fix up protections.
2862 * Otherwise, just simplify entries, since some may have been modified.
2863 * [Note that clipping is not necessary the second time.]
2865 for (prev_entry = vm_map_entry_pred(first_entry), entry = first_entry;
2867 vm_map_try_merge_entries(map, prev_entry, entry),
2868 prev_entry = entry, entry = vm_map_entry_succ(entry)) {
2869 if (rv != KERN_SUCCESS ||
2870 (entry->eflags & MAP_ENTRY_GUARD) != 0)
2873 old_prot = entry->protection;
2875 if ((flags & VM_MAP_PROTECT_SET_MAXPROT) != 0) {
2876 entry->max_protection = new_maxprot;
2877 entry->protection = new_maxprot & old_prot;
2879 if ((flags & VM_MAP_PROTECT_SET_PROT) != 0)
2880 entry->protection = new_prot;
2883 * For user wired map entries, the normal lazy evaluation of
2884 * write access upgrades through soft page faults is
2885 * undesirable. Instead, immediately copy any pages that are
2886 * copy-on-write and enable write access in the physical map.
2888 if ((entry->eflags & MAP_ENTRY_USER_WIRED) != 0 &&
2889 (entry->protection & VM_PROT_WRITE) != 0 &&
2890 (old_prot & VM_PROT_WRITE) == 0)
2891 vm_fault_copy_entry(map, map, entry, entry, NULL);
2894 * When restricting access, update the physical map. Worry
2895 * about copy-on-write here.
2897 if ((old_prot & ~entry->protection) != 0) {
2898 #define MASK(entry) (((entry)->eflags & MAP_ENTRY_COW) ? ~VM_PROT_WRITE : \
2900 pmap_protect(map->pmap, entry->start,
2902 entry->protection & MASK(entry));
2906 vm_map_try_merge_entries(map, prev_entry, entry);
2914 * This routine traverses a processes map handling the madvise
2915 * system call. Advisories are classified as either those effecting
2916 * the vm_map_entry structure, or those effecting the underlying
2926 vm_map_entry_t entry, prev_entry;
2931 * Some madvise calls directly modify the vm_map_entry, in which case
2932 * we need to use an exclusive lock on the map and we need to perform
2933 * various clipping operations. Otherwise we only need a read-lock
2938 case MADV_SEQUENTIAL:
2955 vm_map_lock_read(map);
2962 * Locate starting entry and clip if necessary.
2964 VM_MAP_RANGE_CHECK(map, start, end);
2968 * madvise behaviors that are implemented in the vm_map_entry.
2970 * We clip the vm_map_entry so that behavioral changes are
2971 * limited to the specified address range.
2973 rv = vm_map_lookup_clip_start(map, start, &entry, &prev_entry);
2974 if (rv != KERN_SUCCESS) {
2976 return (vm_mmap_to_errno(rv));
2979 for (; entry->start < end; prev_entry = entry,
2980 entry = vm_map_entry_succ(entry)) {
2981 if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) != 0)
2984 rv = vm_map_clip_end(map, entry, end);
2985 if (rv != KERN_SUCCESS) {
2987 return (vm_mmap_to_errno(rv));
2992 vm_map_entry_set_behavior(entry,
2993 MAP_ENTRY_BEHAV_NORMAL);
2995 case MADV_SEQUENTIAL:
2996 vm_map_entry_set_behavior(entry,
2997 MAP_ENTRY_BEHAV_SEQUENTIAL);
3000 vm_map_entry_set_behavior(entry,
3001 MAP_ENTRY_BEHAV_RANDOM);
3004 entry->eflags |= MAP_ENTRY_NOSYNC;
3007 entry->eflags &= ~MAP_ENTRY_NOSYNC;
3010 entry->eflags |= MAP_ENTRY_NOCOREDUMP;
3013 entry->eflags &= ~MAP_ENTRY_NOCOREDUMP;
3018 vm_map_try_merge_entries(map, prev_entry, entry);
3020 vm_map_try_merge_entries(map, prev_entry, entry);
3023 vm_pindex_t pstart, pend;
3026 * madvise behaviors that are implemented in the underlying
3029 * Since we don't clip the vm_map_entry, we have to clip
3030 * the vm_object pindex and count.
3032 if (!vm_map_lookup_entry(map, start, &entry))
3033 entry = vm_map_entry_succ(entry);
3034 for (; entry->start < end;
3035 entry = vm_map_entry_succ(entry)) {
3036 vm_offset_t useEnd, useStart;
3038 if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) != 0)
3042 * MADV_FREE would otherwise rewind time to
3043 * the creation of the shadow object. Because
3044 * we hold the VM map read-locked, neither the
3045 * entry's object nor the presence of a
3046 * backing object can change.
3048 if (behav == MADV_FREE &&
3049 entry->object.vm_object != NULL &&
3050 entry->object.vm_object->backing_object != NULL)
3053 pstart = OFF_TO_IDX(entry->offset);
3054 pend = pstart + atop(entry->end - entry->start);
3055 useStart = entry->start;
3056 useEnd = entry->end;
3058 if (entry->start < start) {
3059 pstart += atop(start - entry->start);
3062 if (entry->end > end) {
3063 pend -= atop(entry->end - end);
3071 * Perform the pmap_advise() before clearing
3072 * PGA_REFERENCED in vm_page_advise(). Otherwise, a
3073 * concurrent pmap operation, such as pmap_remove(),
3074 * could clear a reference in the pmap and set
3075 * PGA_REFERENCED on the page before the pmap_advise()
3076 * had completed. Consequently, the page would appear
3077 * referenced based upon an old reference that
3078 * occurred before this pmap_advise() ran.
3080 if (behav == MADV_DONTNEED || behav == MADV_FREE)
3081 pmap_advise(map->pmap, useStart, useEnd,
3084 vm_object_madvise(entry->object.vm_object, pstart,
3088 * Pre-populate paging structures in the
3089 * WILLNEED case. For wired entries, the
3090 * paging structures are already populated.
3092 if (behav == MADV_WILLNEED &&
3093 entry->wired_count == 0) {
3094 vm_map_pmap_enter(map,
3097 entry->object.vm_object,
3099 ptoa(pend - pstart),
3100 MAP_PREFAULT_MADVISE
3104 vm_map_unlock_read(map);
3112 * Sets the inheritance of the specified address
3113 * range in the target map. Inheritance
3114 * affects how the map will be shared with
3115 * child maps at the time of vmspace_fork.
3118 vm_map_inherit(vm_map_t map, vm_offset_t start, vm_offset_t end,
3119 vm_inherit_t new_inheritance)
3121 vm_map_entry_t entry, lentry, prev_entry, start_entry;
3124 switch (new_inheritance) {
3125 case VM_INHERIT_NONE:
3126 case VM_INHERIT_COPY:
3127 case VM_INHERIT_SHARE:
3128 case VM_INHERIT_ZERO:
3131 return (KERN_INVALID_ARGUMENT);
3134 return (KERN_SUCCESS);
3136 VM_MAP_RANGE_CHECK(map, start, end);
3137 rv = vm_map_lookup_clip_start(map, start, &start_entry, &prev_entry);
3138 if (rv != KERN_SUCCESS)
3140 if (vm_map_lookup_entry(map, end - 1, &lentry)) {
3141 rv = vm_map_clip_end(map, lentry, end);
3142 if (rv != KERN_SUCCESS)
3145 if (new_inheritance == VM_INHERIT_COPY) {
3146 for (entry = start_entry; entry->start < end;
3147 prev_entry = entry, entry = vm_map_entry_succ(entry)) {
3148 if ((entry->eflags & MAP_ENTRY_SPLIT_BOUNDARY_MASK)
3150 rv = KERN_INVALID_ARGUMENT;
3155 for (entry = start_entry; entry->start < end; prev_entry = entry,
3156 entry = vm_map_entry_succ(entry)) {
3157 KASSERT(entry->end <= end, ("non-clipped entry %p end %jx %jx",
3158 entry, (uintmax_t)entry->end, (uintmax_t)end));
3159 if ((entry->eflags & MAP_ENTRY_GUARD) == 0 ||
3160 new_inheritance != VM_INHERIT_ZERO)
3161 entry->inheritance = new_inheritance;
3162 vm_map_try_merge_entries(map, prev_entry, entry);
3164 vm_map_try_merge_entries(map, prev_entry, entry);
3171 * vm_map_entry_in_transition:
3173 * Release the map lock, and sleep until the entry is no longer in
3174 * transition. Awake and acquire the map lock. If the map changed while
3175 * another held the lock, lookup a possibly-changed entry at or after the
3176 * 'start' position of the old entry.
3178 static vm_map_entry_t
3179 vm_map_entry_in_transition(vm_map_t map, vm_offset_t in_start,
3180 vm_offset_t *io_end, bool holes_ok, vm_map_entry_t in_entry)
3182 vm_map_entry_t entry;
3184 u_int last_timestamp;
3186 VM_MAP_ASSERT_LOCKED(map);
3187 KASSERT((in_entry->eflags & MAP_ENTRY_IN_TRANSITION) != 0,
3188 ("not in-tranition map entry %p", in_entry));
3190 * We have not yet clipped the entry.
3192 start = MAX(in_start, in_entry->start);
3193 in_entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP;
3194 last_timestamp = map->timestamp;
3195 if (vm_map_unlock_and_wait(map, 0)) {
3197 * Allow interruption of user wiring/unwiring?
3201 if (last_timestamp + 1 == map->timestamp)
3205 * Look again for the entry because the map was modified while it was
3206 * unlocked. Specifically, the entry may have been clipped, merged, or
3209 if (!vm_map_lookup_entry(map, start, &entry)) {
3214 entry = vm_map_entry_succ(entry);
3222 * Implements both kernel and user unwiring.
3225 vm_map_unwire(vm_map_t map, vm_offset_t start, vm_offset_t end,
3228 vm_map_entry_t entry, first_entry, next_entry, prev_entry;
3230 bool holes_ok, need_wakeup, user_unwire;
3233 return (KERN_SUCCESS);
3234 holes_ok = (flags & VM_MAP_WIRE_HOLESOK) != 0;
3235 user_unwire = (flags & VM_MAP_WIRE_USER) != 0;
3237 VM_MAP_RANGE_CHECK(map, start, end);
3238 if (!vm_map_lookup_entry(map, start, &first_entry)) {
3240 first_entry = vm_map_entry_succ(first_entry);
3243 return (KERN_INVALID_ADDRESS);
3247 for (entry = first_entry; entry->start < end; entry = next_entry) {
3248 if (entry->eflags & MAP_ENTRY_IN_TRANSITION) {
3250 * We have not yet clipped the entry.
3252 next_entry = vm_map_entry_in_transition(map, start,
3253 &end, holes_ok, entry);
3254 if (next_entry == NULL) {
3255 if (entry == first_entry) {
3257 return (KERN_INVALID_ADDRESS);
3259 rv = KERN_INVALID_ADDRESS;
3262 first_entry = (entry == first_entry) ?
3266 rv = vm_map_clip_start(map, entry, start);
3267 if (rv != KERN_SUCCESS)
3269 rv = vm_map_clip_end(map, entry, end);
3270 if (rv != KERN_SUCCESS)
3274 * Mark the entry in case the map lock is released. (See
3277 KASSERT((entry->eflags & MAP_ENTRY_IN_TRANSITION) == 0 &&
3278 entry->wiring_thread == NULL,
3279 ("owned map entry %p", entry));
3280 entry->eflags |= MAP_ENTRY_IN_TRANSITION;
3281 entry->wiring_thread = curthread;
3282 next_entry = vm_map_entry_succ(entry);
3284 * Check the map for holes in the specified region.
3285 * If holes_ok, skip this check.
3288 entry->end < end && next_entry->start > entry->end) {
3290 rv = KERN_INVALID_ADDRESS;
3294 * If system unwiring, require that the entry is system wired.
3297 vm_map_entry_system_wired_count(entry) == 0) {
3299 rv = KERN_INVALID_ARGUMENT;
3303 need_wakeup = false;
3304 if (first_entry == NULL &&
3305 !vm_map_lookup_entry(map, start, &first_entry)) {
3306 KASSERT(holes_ok, ("vm_map_unwire: lookup failed"));
3307 prev_entry = first_entry;
3308 entry = vm_map_entry_succ(first_entry);
3310 prev_entry = vm_map_entry_pred(first_entry);
3311 entry = first_entry;
3313 for (; entry->start < end;
3314 prev_entry = entry, entry = vm_map_entry_succ(entry)) {
3316 * If holes_ok was specified, an empty
3317 * space in the unwired region could have been mapped
3318 * while the map lock was dropped for draining
3319 * MAP_ENTRY_IN_TRANSITION. Moreover, another thread
3320 * could be simultaneously wiring this new mapping
3321 * entry. Detect these cases and skip any entries
3322 * marked as in transition by us.
3324 if ((entry->eflags & MAP_ENTRY_IN_TRANSITION) == 0 ||
3325 entry->wiring_thread != curthread) {
3327 ("vm_map_unwire: !HOLESOK and new/changed entry"));
3331 if (rv == KERN_SUCCESS && (!user_unwire ||
3332 (entry->eflags & MAP_ENTRY_USER_WIRED))) {
3333 if (entry->wired_count == 1)
3334 vm_map_entry_unwire(map, entry);
3336 entry->wired_count--;
3338 entry->eflags &= ~MAP_ENTRY_USER_WIRED;
3340 KASSERT((entry->eflags & MAP_ENTRY_IN_TRANSITION) != 0,
3341 ("vm_map_unwire: in-transition flag missing %p", entry));
3342 KASSERT(entry->wiring_thread == curthread,
3343 ("vm_map_unwire: alien wire %p", entry));
3344 entry->eflags &= ~MAP_ENTRY_IN_TRANSITION;
3345 entry->wiring_thread = NULL;
3346 if (entry->eflags & MAP_ENTRY_NEEDS_WAKEUP) {
3347 entry->eflags &= ~MAP_ENTRY_NEEDS_WAKEUP;
3350 vm_map_try_merge_entries(map, prev_entry, entry);
3352 vm_map_try_merge_entries(map, prev_entry, entry);
3360 vm_map_wire_user_count_sub(u_long npages)
3363 atomic_subtract_long(&vm_user_wire_count, npages);
3367 vm_map_wire_user_count_add(u_long npages)
3371 wired = vm_user_wire_count;
3373 if (npages + wired > vm_page_max_user_wired)
3375 } while (!atomic_fcmpset_long(&vm_user_wire_count, &wired,
3382 * vm_map_wire_entry_failure:
3384 * Handle a wiring failure on the given entry.
3386 * The map should be locked.
3389 vm_map_wire_entry_failure(vm_map_t map, vm_map_entry_t entry,
3390 vm_offset_t failed_addr)
3393 VM_MAP_ASSERT_LOCKED(map);
3394 KASSERT((entry->eflags & MAP_ENTRY_IN_TRANSITION) != 0 &&
3395 entry->wired_count == 1,
3396 ("vm_map_wire_entry_failure: entry %p isn't being wired", entry));
3397 KASSERT(failed_addr < entry->end,
3398 ("vm_map_wire_entry_failure: entry %p was fully wired", entry));
3401 * If any pages at the start of this entry were successfully wired,
3404 if (failed_addr > entry->start) {
3405 pmap_unwire(map->pmap, entry->start, failed_addr);
3406 vm_object_unwire(entry->object.vm_object, entry->offset,
3407 failed_addr - entry->start, PQ_ACTIVE);
3411 * Assign an out-of-range value to represent the failure to wire this
3414 entry->wired_count = -1;
3418 vm_map_wire(vm_map_t map, vm_offset_t start, vm_offset_t end, int flags)
3423 rv = vm_map_wire_locked(map, start, end, flags);
3429 * vm_map_wire_locked:
3431 * Implements both kernel and user wiring. Returns with the map locked,
3432 * the map lock may be dropped.
3435 vm_map_wire_locked(vm_map_t map, vm_offset_t start, vm_offset_t end, int flags)
3437 vm_map_entry_t entry, first_entry, next_entry, prev_entry;
3438 vm_offset_t faddr, saved_end, saved_start;
3439 u_long incr, npages;
3440 u_int bidx, last_timestamp;
3442 bool holes_ok, need_wakeup, user_wire;
3445 VM_MAP_ASSERT_LOCKED(map);
3448 return (KERN_SUCCESS);
3450 if (flags & VM_MAP_WIRE_WRITE)
3451 prot |= VM_PROT_WRITE;
3452 holes_ok = (flags & VM_MAP_WIRE_HOLESOK) != 0;
3453 user_wire = (flags & VM_MAP_WIRE_USER) != 0;
3454 VM_MAP_RANGE_CHECK(map, start, end);
3455 if (!vm_map_lookup_entry(map, start, &first_entry)) {
3457 first_entry = vm_map_entry_succ(first_entry);
3459 return (KERN_INVALID_ADDRESS);
3461 for (entry = first_entry; entry->start < end; entry = next_entry) {
3462 if (entry->eflags & MAP_ENTRY_IN_TRANSITION) {
3464 * We have not yet clipped the entry.
3466 next_entry = vm_map_entry_in_transition(map, start,
3467 &end, holes_ok, entry);
3468 if (next_entry == NULL) {
3469 if (entry == first_entry)
3470 return (KERN_INVALID_ADDRESS);
3471 rv = KERN_INVALID_ADDRESS;
3474 first_entry = (entry == first_entry) ?
3478 rv = vm_map_clip_start(map, entry, start);
3479 if (rv != KERN_SUCCESS)
3481 rv = vm_map_clip_end(map, entry, end);
3482 if (rv != KERN_SUCCESS)
3486 * Mark the entry in case the map lock is released. (See
3489 KASSERT((entry->eflags & MAP_ENTRY_IN_TRANSITION) == 0 &&
3490 entry->wiring_thread == NULL,
3491 ("owned map entry %p", entry));
3492 entry->eflags |= MAP_ENTRY_IN_TRANSITION;
3493 entry->wiring_thread = curthread;
3494 if ((entry->protection & (VM_PROT_READ | VM_PROT_EXECUTE)) == 0
3495 || (entry->protection & prot) != prot) {
3496 entry->eflags |= MAP_ENTRY_WIRE_SKIPPED;
3499 rv = KERN_INVALID_ADDRESS;
3502 } else if (entry->wired_count == 0) {
3503 entry->wired_count++;
3505 npages = atop(entry->end - entry->start);
3506 if (user_wire && !vm_map_wire_user_count_add(npages)) {
3507 vm_map_wire_entry_failure(map, entry,
3510 rv = KERN_RESOURCE_SHORTAGE;
3515 * Release the map lock, relying on the in-transition
3516 * mark. Mark the map busy for fork.
3518 saved_start = entry->start;
3519 saved_end = entry->end;
3520 last_timestamp = map->timestamp;
3521 bidx = (entry->eflags & MAP_ENTRY_SPLIT_BOUNDARY_MASK)
3522 >> MAP_ENTRY_SPLIT_BOUNDARY_SHIFT;
3523 incr = pagesizes[bidx];
3527 for (faddr = saved_start; faddr < saved_end;
3530 * Simulate a fault to get the page and enter
3531 * it into the physical map.
3533 rv = vm_fault(map, faddr, VM_PROT_NONE,
3534 VM_FAULT_WIRE, NULL);
3535 if (rv != KERN_SUCCESS)
3540 if (last_timestamp + 1 != map->timestamp) {
3542 * Look again for the entry because the map was
3543 * modified while it was unlocked. The entry
3544 * may have been clipped, but NOT merged or
3547 if (!vm_map_lookup_entry(map, saved_start,
3550 ("vm_map_wire: lookup failed"));
3551 first_entry = (entry == first_entry) ?
3553 for (entry = next_entry; entry->end < saved_end;
3554 entry = vm_map_entry_succ(entry)) {
3556 * In case of failure, handle entries
3557 * that were not fully wired here;
3558 * fully wired entries are handled
3561 if (rv != KERN_SUCCESS &&
3563 vm_map_wire_entry_failure(map,
3567 if (rv != KERN_SUCCESS) {
3568 vm_map_wire_entry_failure(map, entry, faddr);
3570 vm_map_wire_user_count_sub(npages);
3574 } else if (!user_wire ||
3575 (entry->eflags & MAP_ENTRY_USER_WIRED) == 0) {
3576 entry->wired_count++;
3579 * Check the map for holes in the specified region.
3580 * If holes_ok was specified, skip this check.
3582 next_entry = vm_map_entry_succ(entry);
3584 entry->end < end && next_entry->start > entry->end) {
3586 rv = KERN_INVALID_ADDRESS;
3592 need_wakeup = false;
3593 if (first_entry == NULL &&
3594 !vm_map_lookup_entry(map, start, &first_entry)) {
3595 KASSERT(holes_ok, ("vm_map_wire: lookup failed"));
3596 prev_entry = first_entry;
3597 entry = vm_map_entry_succ(first_entry);
3599 prev_entry = vm_map_entry_pred(first_entry);
3600 entry = first_entry;
3602 for (; entry->start < end;
3603 prev_entry = entry, entry = vm_map_entry_succ(entry)) {
3605 * If holes_ok was specified, an empty
3606 * space in the unwired region could have been mapped
3607 * while the map lock was dropped for faulting in the
3608 * pages or draining MAP_ENTRY_IN_TRANSITION.
3609 * Moreover, another thread could be simultaneously
3610 * wiring this new mapping entry. Detect these cases
3611 * and skip any entries marked as in transition not by us.
3613 * Another way to get an entry not marked with
3614 * MAP_ENTRY_IN_TRANSITION is after failed clipping,
3615 * which set rv to KERN_INVALID_ARGUMENT.
3617 if ((entry->eflags & MAP_ENTRY_IN_TRANSITION) == 0 ||
3618 entry->wiring_thread != curthread) {
3619 KASSERT(holes_ok || rv == KERN_INVALID_ARGUMENT,
3620 ("vm_map_wire: !HOLESOK and new/changed entry"));
3624 if ((entry->eflags & MAP_ENTRY_WIRE_SKIPPED) != 0) {
3626 } else if (rv == KERN_SUCCESS) {
3628 entry->eflags |= MAP_ENTRY_USER_WIRED;
3629 } else if (entry->wired_count == -1) {
3631 * Wiring failed on this entry. Thus, unwiring is
3634 entry->wired_count = 0;
3635 } else if (!user_wire ||
3636 (entry->eflags & MAP_ENTRY_USER_WIRED) == 0) {
3638 * Undo the wiring. Wiring succeeded on this entry
3639 * but failed on a later entry.
3641 if (entry->wired_count == 1) {
3642 vm_map_entry_unwire(map, entry);
3644 vm_map_wire_user_count_sub(
3645 atop(entry->end - entry->start));
3647 entry->wired_count--;
3649 KASSERT((entry->eflags & MAP_ENTRY_IN_TRANSITION) != 0,
3650 ("vm_map_wire: in-transition flag missing %p", entry));
3651 KASSERT(entry->wiring_thread == curthread,
3652 ("vm_map_wire: alien wire %p", entry));
3653 entry->eflags &= ~(MAP_ENTRY_IN_TRANSITION |
3654 MAP_ENTRY_WIRE_SKIPPED);
3655 entry->wiring_thread = NULL;
3656 if (entry->eflags & MAP_ENTRY_NEEDS_WAKEUP) {
3657 entry->eflags &= ~MAP_ENTRY_NEEDS_WAKEUP;
3660 vm_map_try_merge_entries(map, prev_entry, entry);
3662 vm_map_try_merge_entries(map, prev_entry, entry);
3671 * Push any dirty cached pages in the address range to their pager.
3672 * If syncio is TRUE, dirty pages are written synchronously.
3673 * If invalidate is TRUE, any cached pages are freed as well.
3675 * If the size of the region from start to end is zero, we are
3676 * supposed to flush all modified pages within the region containing
3677 * start. Unfortunately, a region can be split or coalesced with
3678 * neighboring regions, making it difficult to determine what the
3679 * original region was. Therefore, we approximate this requirement by
3680 * flushing the current region containing start.
3682 * Returns an error if any part of the specified range is not mapped.
3690 boolean_t invalidate)
3692 vm_map_entry_t entry, first_entry, next_entry;
3695 vm_ooffset_t offset;
3696 unsigned int last_timestamp;
3700 vm_map_lock_read(map);
3701 VM_MAP_RANGE_CHECK(map, start, end);
3702 if (!vm_map_lookup_entry(map, start, &first_entry)) {
3703 vm_map_unlock_read(map);
3704 return (KERN_INVALID_ADDRESS);
3705 } else if (start == end) {
3706 start = first_entry->start;
3707 end = first_entry->end;
3711 * Make a first pass to check for user-wired memory, holes,
3712 * and partial invalidation of largepage mappings.
3714 for (entry = first_entry; entry->start < end; entry = next_entry) {
3716 if ((entry->eflags & MAP_ENTRY_USER_WIRED) != 0) {
3717 vm_map_unlock_read(map);
3718 return (KERN_INVALID_ARGUMENT);
3720 bdry_idx = (entry->eflags &
3721 MAP_ENTRY_SPLIT_BOUNDARY_MASK) >>
3722 MAP_ENTRY_SPLIT_BOUNDARY_SHIFT;
3723 if (bdry_idx != 0 &&
3724 ((start & (pagesizes[bdry_idx] - 1)) != 0 ||
3725 (end & (pagesizes[bdry_idx] - 1)) != 0)) {
3726 vm_map_unlock_read(map);
3727 return (KERN_INVALID_ARGUMENT);
3730 next_entry = vm_map_entry_succ(entry);
3731 if (end > entry->end &&
3732 entry->end != next_entry->start) {
3733 vm_map_unlock_read(map);
3734 return (KERN_INVALID_ADDRESS);
3739 pmap_remove(map->pmap, start, end);
3743 * Make a second pass, cleaning/uncaching pages from the indicated
3746 for (entry = first_entry; entry->start < end;) {
3747 offset = entry->offset + (start - entry->start);
3748 size = (end <= entry->end ? end : entry->end) - start;
3749 if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) != 0) {
3751 vm_map_entry_t tentry;
3754 smap = entry->object.sub_map;
3755 vm_map_lock_read(smap);
3756 (void) vm_map_lookup_entry(smap, offset, &tentry);
3757 tsize = tentry->end - offset;
3760 object = tentry->object.vm_object;
3761 offset = tentry->offset + (offset - tentry->start);
3762 vm_map_unlock_read(smap);
3764 object = entry->object.vm_object;
3766 vm_object_reference(object);
3767 last_timestamp = map->timestamp;
3768 vm_map_unlock_read(map);
3769 if (!vm_object_sync(object, offset, size, syncio, invalidate))
3772 vm_object_deallocate(object);
3773 vm_map_lock_read(map);
3774 if (last_timestamp == map->timestamp ||
3775 !vm_map_lookup_entry(map, start, &entry))
3776 entry = vm_map_entry_succ(entry);
3779 vm_map_unlock_read(map);
3780 return (failed ? KERN_FAILURE : KERN_SUCCESS);
3784 * vm_map_entry_unwire: [ internal use only ]
3786 * Make the region specified by this entry pageable.
3788 * The map in question should be locked.
3789 * [This is the reason for this routine's existence.]
3792 vm_map_entry_unwire(vm_map_t map, vm_map_entry_t entry)
3796 VM_MAP_ASSERT_LOCKED(map);
3797 KASSERT(entry->wired_count > 0,
3798 ("vm_map_entry_unwire: entry %p isn't wired", entry));
3800 size = entry->end - entry->start;
3801 if ((entry->eflags & MAP_ENTRY_USER_WIRED) != 0)
3802 vm_map_wire_user_count_sub(atop(size));
3803 pmap_unwire(map->pmap, entry->start, entry->end);
3804 vm_object_unwire(entry->object.vm_object, entry->offset, size,
3806 entry->wired_count = 0;
3810 vm_map_entry_deallocate(vm_map_entry_t entry, boolean_t system_map)
3813 if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) == 0)
3814 vm_object_deallocate(entry->object.vm_object);
3815 uma_zfree(system_map ? kmapentzone : mapentzone, entry);
3819 * vm_map_entry_delete: [ internal use only ]
3821 * Deallocate the given entry from the target map.
3824 vm_map_entry_delete(vm_map_t map, vm_map_entry_t entry)
3827 vm_pindex_t offidxstart, offidxend, size1;
3830 vm_map_entry_unlink(map, entry, UNLINK_MERGE_NONE);
3831 object = entry->object.vm_object;
3833 if ((entry->eflags & MAP_ENTRY_GUARD) != 0) {
3834 MPASS(entry->cred == NULL);
3835 MPASS((entry->eflags & MAP_ENTRY_IS_SUB_MAP) == 0);
3836 MPASS(object == NULL);
3837 vm_map_entry_deallocate(entry, map->system_map);
3841 size = entry->end - entry->start;
3844 if (entry->cred != NULL) {
3845 swap_release_by_cred(size, entry->cred);
3846 crfree(entry->cred);
3849 if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) != 0 || object == NULL) {
3850 entry->object.vm_object = NULL;
3851 } else if ((object->flags & OBJ_ANON) != 0 ||
3852 object == kernel_object) {
3853 KASSERT(entry->cred == NULL || object->cred == NULL ||
3854 (entry->eflags & MAP_ENTRY_NEEDS_COPY),
3855 ("OVERCOMMIT vm_map_entry_delete: both cred %p", entry));
3856 offidxstart = OFF_TO_IDX(entry->offset);
3857 offidxend = offidxstart + atop(size);
3858 VM_OBJECT_WLOCK(object);
3859 if (object->ref_count != 1 &&
3860 ((object->flags & OBJ_ONEMAPPING) != 0 ||
3861 object == kernel_object)) {
3862 vm_object_collapse(object);
3865 * The option OBJPR_NOTMAPPED can be passed here
3866 * because vm_map_delete() already performed
3867 * pmap_remove() on the only mapping to this range
3870 vm_object_page_remove(object, offidxstart, offidxend,
3872 if (offidxend >= object->size &&
3873 offidxstart < object->size) {
3874 size1 = object->size;
3875 object->size = offidxstart;
3876 if (object->cred != NULL) {
3877 size1 -= object->size;
3878 KASSERT(object->charge >= ptoa(size1),
3879 ("object %p charge < 0", object));
3880 swap_release_by_cred(ptoa(size1),
3882 object->charge -= ptoa(size1);
3886 VM_OBJECT_WUNLOCK(object);
3888 if (map->system_map)
3889 vm_map_entry_deallocate(entry, TRUE);
3891 entry->defer_next = curthread->td_map_def_user;
3892 curthread->td_map_def_user = entry;
3897 * vm_map_delete: [ internal use only ]
3899 * Deallocates the given address range from the target
3903 vm_map_delete(vm_map_t map, vm_offset_t start, vm_offset_t end)
3905 vm_map_entry_t entry, next_entry, scratch_entry;
3908 VM_MAP_ASSERT_LOCKED(map);
3911 return (KERN_SUCCESS);
3914 * Find the start of the region, and clip it.
3915 * Step through all entries in this region.
3917 rv = vm_map_lookup_clip_start(map, start, &entry, &scratch_entry);
3918 if (rv != KERN_SUCCESS)
3920 for (; entry->start < end; entry = next_entry) {
3922 * Wait for wiring or unwiring of an entry to complete.
3923 * Also wait for any system wirings to disappear on
3926 if ((entry->eflags & MAP_ENTRY_IN_TRANSITION) != 0 ||
3927 (vm_map_pmap(map) != kernel_pmap &&
3928 vm_map_entry_system_wired_count(entry) != 0)) {
3929 unsigned int last_timestamp;
3930 vm_offset_t saved_start;
3932 saved_start = entry->start;
3933 entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP;
3934 last_timestamp = map->timestamp;
3935 (void) vm_map_unlock_and_wait(map, 0);
3937 if (last_timestamp + 1 != map->timestamp) {
3939 * Look again for the entry because the map was
3940 * modified while it was unlocked.
3941 * Specifically, the entry may have been
3942 * clipped, merged, or deleted.
3944 rv = vm_map_lookup_clip_start(map, saved_start,
3945 &next_entry, &scratch_entry);
3946 if (rv != KERN_SUCCESS)
3953 /* XXXKIB or delete to the upper superpage boundary ? */
3954 rv = vm_map_clip_end(map, entry, end);
3955 if (rv != KERN_SUCCESS)
3957 next_entry = vm_map_entry_succ(entry);
3960 * Unwire before removing addresses from the pmap; otherwise,
3961 * unwiring will put the entries back in the pmap.
3963 if (entry->wired_count != 0)
3964 vm_map_entry_unwire(map, entry);
3967 * Remove mappings for the pages, but only if the
3968 * mappings could exist. For instance, it does not
3969 * make sense to call pmap_remove() for guard entries.
3971 if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) != 0 ||
3972 entry->object.vm_object != NULL)
3973 pmap_remove(map->pmap, entry->start, entry->end);
3975 if (entry->end == map->anon_loc)
3976 map->anon_loc = entry->start;
3979 * Delete the entry only after removing all pmap
3980 * entries pointing to its pages. (Otherwise, its
3981 * page frames may be reallocated, and any modify bits
3982 * will be set in the wrong object!)
3984 vm_map_entry_delete(map, entry);
3992 * Remove the given address range from the target map.
3993 * This is the exported form of vm_map_delete.
3996 vm_map_remove(vm_map_t map, vm_offset_t start, vm_offset_t end)
4001 VM_MAP_RANGE_CHECK(map, start, end);
4002 result = vm_map_delete(map, start, end);
4008 * vm_map_check_protection:
4010 * Assert that the target map allows the specified privilege on the
4011 * entire address region given. The entire region must be allocated.
4013 * WARNING! This code does not and should not check whether the
4014 * contents of the region is accessible. For example a smaller file
4015 * might be mapped into a larger address space.
4017 * NOTE! This code is also called by munmap().
4019 * The map must be locked. A read lock is sufficient.
4022 vm_map_check_protection(vm_map_t map, vm_offset_t start, vm_offset_t end,
4023 vm_prot_t protection)
4025 vm_map_entry_t entry;
4026 vm_map_entry_t tmp_entry;
4028 if (!vm_map_lookup_entry(map, start, &tmp_entry))
4032 while (start < end) {
4036 if (start < entry->start)
4039 * Check protection associated with entry.
4041 if ((entry->protection & protection) != protection)
4043 /* go to next entry */
4045 entry = vm_map_entry_succ(entry);
4052 * vm_map_copy_swap_object:
4054 * Copies a swap-backed object from an existing map entry to a
4055 * new one. Carries forward the swap charge. May change the
4056 * src object on return.
4059 vm_map_copy_swap_object(vm_map_entry_t src_entry, vm_map_entry_t dst_entry,
4060 vm_offset_t size, vm_ooffset_t *fork_charge)
4062 vm_object_t src_object;
4066 src_object = src_entry->object.vm_object;
4067 charged = ENTRY_CHARGED(src_entry);
4068 if ((src_object->flags & OBJ_ANON) != 0) {
4069 VM_OBJECT_WLOCK(src_object);
4070 vm_object_collapse(src_object);
4071 if ((src_object->flags & OBJ_ONEMAPPING) != 0) {
4072 vm_object_split(src_entry);
4073 src_object = src_entry->object.vm_object;
4075 vm_object_reference_locked(src_object);
4076 vm_object_clear_flag(src_object, OBJ_ONEMAPPING);
4077 VM_OBJECT_WUNLOCK(src_object);
4079 vm_object_reference(src_object);
4080 if (src_entry->cred != NULL &&
4081 !(src_entry->eflags & MAP_ENTRY_NEEDS_COPY)) {
4082 KASSERT(src_object->cred == NULL,
4083 ("OVERCOMMIT: vm_map_copy_anon_entry: cred %p",
4085 src_object->cred = src_entry->cred;
4086 src_object->charge = size;
4088 dst_entry->object.vm_object = src_object;
4090 cred = curthread->td_ucred;
4092 dst_entry->cred = cred;
4093 *fork_charge += size;
4094 if (!(src_entry->eflags & MAP_ENTRY_NEEDS_COPY)) {
4096 src_entry->cred = cred;
4097 *fork_charge += size;
4103 * vm_map_copy_entry:
4105 * Copies the contents of the source entry to the destination
4106 * entry. The entries *must* be aligned properly.
4112 vm_map_entry_t src_entry,
4113 vm_map_entry_t dst_entry,
4114 vm_ooffset_t *fork_charge)
4116 vm_object_t src_object;
4117 vm_map_entry_t fake_entry;
4120 VM_MAP_ASSERT_LOCKED(dst_map);
4122 if ((dst_entry->eflags|src_entry->eflags) & MAP_ENTRY_IS_SUB_MAP)
4125 if (src_entry->wired_count == 0 ||
4126 (src_entry->protection & VM_PROT_WRITE) == 0) {
4128 * If the source entry is marked needs_copy, it is already
4131 if ((src_entry->eflags & MAP_ENTRY_NEEDS_COPY) == 0 &&
4132 (src_entry->protection & VM_PROT_WRITE) != 0) {
4133 pmap_protect(src_map->pmap,
4136 src_entry->protection & ~VM_PROT_WRITE);
4140 * Make a copy of the object.
4142 size = src_entry->end - src_entry->start;
4143 if ((src_object = src_entry->object.vm_object) != NULL) {
4144 if (src_object->type == OBJT_DEFAULT ||
4145 (src_object->flags & OBJ_SWAP) != 0) {
4146 vm_map_copy_swap_object(src_entry, dst_entry,
4148 /* May have split/collapsed, reload obj. */
4149 src_object = src_entry->object.vm_object;
4151 vm_object_reference(src_object);
4152 dst_entry->object.vm_object = src_object;
4154 src_entry->eflags |= MAP_ENTRY_COW |
4155 MAP_ENTRY_NEEDS_COPY;
4156 dst_entry->eflags |= MAP_ENTRY_COW |
4157 MAP_ENTRY_NEEDS_COPY;
4158 dst_entry->offset = src_entry->offset;
4159 if (src_entry->eflags & MAP_ENTRY_WRITECNT) {
4161 * MAP_ENTRY_WRITECNT cannot
4162 * indicate write reference from
4163 * src_entry, since the entry is
4164 * marked as needs copy. Allocate a
4165 * fake entry that is used to
4166 * decrement object->un_pager writecount
4167 * at the appropriate time. Attach
4168 * fake_entry to the deferred list.
4170 fake_entry = vm_map_entry_create(dst_map);
4171 fake_entry->eflags = MAP_ENTRY_WRITECNT;
4172 src_entry->eflags &= ~MAP_ENTRY_WRITECNT;
4173 vm_object_reference(src_object);
4174 fake_entry->object.vm_object = src_object;
4175 fake_entry->start = src_entry->start;
4176 fake_entry->end = src_entry->end;
4177 fake_entry->defer_next =
4178 curthread->td_map_def_user;
4179 curthread->td_map_def_user = fake_entry;
4182 pmap_copy(dst_map->pmap, src_map->pmap,
4183 dst_entry->start, dst_entry->end - dst_entry->start,
4186 dst_entry->object.vm_object = NULL;
4187 dst_entry->offset = 0;
4188 if (src_entry->cred != NULL) {
4189 dst_entry->cred = curthread->td_ucred;
4190 crhold(dst_entry->cred);
4191 *fork_charge += size;
4196 * We don't want to make writeable wired pages copy-on-write.
4197 * Immediately copy these pages into the new map by simulating
4198 * page faults. The new pages are pageable.
4200 vm_fault_copy_entry(dst_map, src_map, dst_entry, src_entry,
4206 * vmspace_map_entry_forked:
4207 * Update the newly-forked vmspace each time a map entry is inherited
4208 * or copied. The values for vm_dsize and vm_tsize are approximate
4209 * (and mostly-obsolete ideas in the face of mmap(2) et al.)
4212 vmspace_map_entry_forked(const struct vmspace *vm1, struct vmspace *vm2,
4213 vm_map_entry_t entry)
4215 vm_size_t entrysize;
4218 if ((entry->eflags & MAP_ENTRY_GUARD) != 0)
4220 entrysize = entry->end - entry->start;
4221 vm2->vm_map.size += entrysize;
4222 if (entry->eflags & (MAP_ENTRY_GROWS_DOWN | MAP_ENTRY_GROWS_UP)) {
4223 vm2->vm_ssize += btoc(entrysize);
4224 } else if (entry->start >= (vm_offset_t)vm1->vm_daddr &&
4225 entry->start < (vm_offset_t)vm1->vm_daddr + ctob(vm1->vm_dsize)) {
4226 newend = MIN(entry->end,
4227 (vm_offset_t)vm1->vm_daddr + ctob(vm1->vm_dsize));
4228 vm2->vm_dsize += btoc(newend - entry->start);
4229 } else if (entry->start >= (vm_offset_t)vm1->vm_taddr &&
4230 entry->start < (vm_offset_t)vm1->vm_taddr + ctob(vm1->vm_tsize)) {
4231 newend = MIN(entry->end,
4232 (vm_offset_t)vm1->vm_taddr + ctob(vm1->vm_tsize));
4233 vm2->vm_tsize += btoc(newend - entry->start);
4239 * Create a new process vmspace structure and vm_map
4240 * based on those of an existing process. The new map
4241 * is based on the old map, according to the inheritance
4242 * values on the regions in that map.
4244 * XXX It might be worth coalescing the entries added to the new vmspace.
4246 * The source map must not be locked.
4249 vmspace_fork(struct vmspace *vm1, vm_ooffset_t *fork_charge)
4251 struct vmspace *vm2;
4252 vm_map_t new_map, old_map;
4253 vm_map_entry_t new_entry, old_entry;
4258 old_map = &vm1->vm_map;
4259 /* Copy immutable fields of vm1 to vm2. */
4260 vm2 = vmspace_alloc(vm_map_min(old_map), vm_map_max(old_map),
4265 vm2->vm_taddr = vm1->vm_taddr;
4266 vm2->vm_daddr = vm1->vm_daddr;
4267 vm2->vm_maxsaddr = vm1->vm_maxsaddr;
4268 vm2->vm_stacktop = vm1->vm_stacktop;
4269 vm_map_lock(old_map);
4271 vm_map_wait_busy(old_map);
4272 new_map = &vm2->vm_map;
4273 locked = vm_map_trylock(new_map); /* trylock to silence WITNESS */
4274 KASSERT(locked, ("vmspace_fork: lock failed"));
4276 error = pmap_vmspace_copy(new_map->pmap, old_map->pmap);
4278 sx_xunlock(&old_map->lock);
4279 sx_xunlock(&new_map->lock);
4280 vm_map_process_deferred();
4285 new_map->anon_loc = old_map->anon_loc;
4286 new_map->flags |= old_map->flags & (MAP_ASLR | MAP_ASLR_IGNSTART |
4287 MAP_ASLR_STACK | MAP_WXORX);
4289 VM_MAP_ENTRY_FOREACH(old_entry, old_map) {
4290 if ((old_entry->eflags & MAP_ENTRY_IS_SUB_MAP) != 0)
4291 panic("vm_map_fork: encountered a submap");
4293 inh = old_entry->inheritance;
4294 if ((old_entry->eflags & MAP_ENTRY_GUARD) != 0 &&
4295 inh != VM_INHERIT_NONE)
4296 inh = VM_INHERIT_COPY;
4299 case VM_INHERIT_NONE:
4302 case VM_INHERIT_SHARE:
4304 * Clone the entry, creating the shared object if
4307 object = old_entry->object.vm_object;
4308 if (object == NULL) {
4309 vm_map_entry_back(old_entry);
4310 object = old_entry->object.vm_object;
4314 * Add the reference before calling vm_object_shadow
4315 * to insure that a shadow object is created.
4317 vm_object_reference(object);
4318 if (old_entry->eflags & MAP_ENTRY_NEEDS_COPY) {
4319 vm_object_shadow(&old_entry->object.vm_object,
4321 old_entry->end - old_entry->start,
4323 /* Transfer the second reference too. */
4325 old_entry->eflags &= ~MAP_ENTRY_NEEDS_COPY;
4326 old_entry->cred = NULL;
4329 * As in vm_map_merged_neighbor_dispose(),
4330 * the vnode lock will not be acquired in
4331 * this call to vm_object_deallocate().
4333 vm_object_deallocate(object);
4334 object = old_entry->object.vm_object;
4336 VM_OBJECT_WLOCK(object);
4337 vm_object_clear_flag(object, OBJ_ONEMAPPING);
4338 if (old_entry->cred != NULL) {
4339 KASSERT(object->cred == NULL,
4340 ("vmspace_fork both cred"));
4341 object->cred = old_entry->cred;
4342 object->charge = old_entry->end -
4344 old_entry->cred = NULL;
4348 * Assert the correct state of the vnode
4349 * v_writecount while the object is locked, to
4350 * not relock it later for the assertion
4353 if (old_entry->eflags & MAP_ENTRY_WRITECNT &&
4354 object->type == OBJT_VNODE) {
4355 KASSERT(((struct vnode *)object->
4356 handle)->v_writecount > 0,
4357 ("vmspace_fork: v_writecount %p",
4359 KASSERT(object->un_pager.vnp.
4361 ("vmspace_fork: vnp.writecount %p",
4364 VM_OBJECT_WUNLOCK(object);
4368 * Clone the entry, referencing the shared object.
4370 new_entry = vm_map_entry_create(new_map);
4371 *new_entry = *old_entry;
4372 new_entry->eflags &= ~(MAP_ENTRY_USER_WIRED |
4373 MAP_ENTRY_IN_TRANSITION);
4374 new_entry->wiring_thread = NULL;
4375 new_entry->wired_count = 0;
4376 if (new_entry->eflags & MAP_ENTRY_WRITECNT) {
4377 vm_pager_update_writecount(object,
4378 new_entry->start, new_entry->end);
4380 vm_map_entry_set_vnode_text(new_entry, true);
4383 * Insert the entry into the new map -- we know we're
4384 * inserting at the end of the new map.
4386 vm_map_entry_link(new_map, new_entry);
4387 vmspace_map_entry_forked(vm1, vm2, new_entry);
4390 * Update the physical map
4392 pmap_copy(new_map->pmap, old_map->pmap,
4394 (old_entry->end - old_entry->start),
4398 case VM_INHERIT_COPY:
4400 * Clone the entry and link into the map.
4402 new_entry = vm_map_entry_create(new_map);
4403 *new_entry = *old_entry;
4405 * Copied entry is COW over the old object.
4407 new_entry->eflags &= ~(MAP_ENTRY_USER_WIRED |
4408 MAP_ENTRY_IN_TRANSITION | MAP_ENTRY_WRITECNT);
4409 new_entry->wiring_thread = NULL;
4410 new_entry->wired_count = 0;
4411 new_entry->object.vm_object = NULL;
4412 new_entry->cred = NULL;
4413 vm_map_entry_link(new_map, new_entry);
4414 vmspace_map_entry_forked(vm1, vm2, new_entry);
4415 vm_map_copy_entry(old_map, new_map, old_entry,
4416 new_entry, fork_charge);
4417 vm_map_entry_set_vnode_text(new_entry, true);
4420 case VM_INHERIT_ZERO:
4422 * Create a new anonymous mapping entry modelled from
4425 new_entry = vm_map_entry_create(new_map);
4426 memset(new_entry, 0, sizeof(*new_entry));
4428 new_entry->start = old_entry->start;
4429 new_entry->end = old_entry->end;
4430 new_entry->eflags = old_entry->eflags &
4431 ~(MAP_ENTRY_USER_WIRED | MAP_ENTRY_IN_TRANSITION |
4432 MAP_ENTRY_WRITECNT | MAP_ENTRY_VN_EXEC |
4433 MAP_ENTRY_SPLIT_BOUNDARY_MASK);
4434 new_entry->protection = old_entry->protection;
4435 new_entry->max_protection = old_entry->max_protection;
4436 new_entry->inheritance = VM_INHERIT_ZERO;
4438 vm_map_entry_link(new_map, new_entry);
4439 vmspace_map_entry_forked(vm1, vm2, new_entry);
4441 new_entry->cred = curthread->td_ucred;
4442 crhold(new_entry->cred);
4443 *fork_charge += (new_entry->end - new_entry->start);
4449 * Use inlined vm_map_unlock() to postpone handling the deferred
4450 * map entries, which cannot be done until both old_map and
4451 * new_map locks are released.
4453 sx_xunlock(&old_map->lock);
4454 sx_xunlock(&new_map->lock);
4455 vm_map_process_deferred();
4461 * Create a process's stack for exec_new_vmspace(). This function is never
4462 * asked to wire the newly created stack.
4465 vm_map_stack(vm_map_t map, vm_offset_t addrbos, vm_size_t max_ssize,
4466 vm_prot_t prot, vm_prot_t max, int cow)
4468 vm_size_t growsize, init_ssize;
4472 MPASS((map->flags & MAP_WIREFUTURE) == 0);
4473 growsize = sgrowsiz;
4474 init_ssize = (max_ssize < growsize) ? max_ssize : growsize;
4476 vmemlim = lim_cur(curthread, RLIMIT_VMEM);
4477 /* If we would blow our VMEM resource limit, no go */
4478 if (map->size + init_ssize > vmemlim) {
4482 rv = vm_map_stack_locked(map, addrbos, max_ssize, growsize, prot,
4489 static int stack_guard_page = 1;
4490 SYSCTL_INT(_security_bsd, OID_AUTO, stack_guard_page, CTLFLAG_RWTUN,
4491 &stack_guard_page, 0,
4492 "Specifies the number of guard pages for a stack that grows");
4495 vm_map_stack_locked(vm_map_t map, vm_offset_t addrbos, vm_size_t max_ssize,
4496 vm_size_t growsize, vm_prot_t prot, vm_prot_t max, int cow)
4498 vm_map_entry_t new_entry, prev_entry;
4499 vm_offset_t bot, gap_bot, gap_top, top;
4500 vm_size_t init_ssize, sgp;
4504 * The stack orientation is piggybacked with the cow argument.
4505 * Extract it into orient and mask the cow argument so that we
4506 * don't pass it around further.
4508 orient = cow & (MAP_STACK_GROWS_DOWN | MAP_STACK_GROWS_UP);
4509 KASSERT(orient != 0, ("No stack grow direction"));
4510 KASSERT(orient != (MAP_STACK_GROWS_DOWN | MAP_STACK_GROWS_UP),
4513 if (max_ssize == 0 ||
4514 !vm_map_range_valid(map, addrbos, addrbos + max_ssize))
4515 return (KERN_INVALID_ADDRESS);
4516 sgp = ((curproc->p_flag2 & P2_STKGAP_DISABLE) != 0 ||
4517 (curproc->p_fctl0 & NT_FREEBSD_FCTL_STKGAP_DISABLE) != 0) ? 0 :
4518 (vm_size_t)stack_guard_page * PAGE_SIZE;
4519 if (sgp >= max_ssize)
4520 return (KERN_INVALID_ARGUMENT);
4522 init_ssize = growsize;
4523 if (max_ssize < init_ssize + sgp)
4524 init_ssize = max_ssize - sgp;
4526 /* If addr is already mapped, no go */
4527 if (vm_map_lookup_entry(map, addrbos, &prev_entry))
4528 return (KERN_NO_SPACE);
4531 * If we can't accommodate max_ssize in the current mapping, no go.
4533 if (vm_map_entry_succ(prev_entry)->start < addrbos + max_ssize)
4534 return (KERN_NO_SPACE);
4537 * We initially map a stack of only init_ssize. We will grow as
4538 * needed later. Depending on the orientation of the stack (i.e.
4539 * the grow direction) we either map at the top of the range, the
4540 * bottom of the range or in the middle.
4542 * Note: we would normally expect prot and max to be VM_PROT_ALL,
4543 * and cow to be 0. Possibly we should eliminate these as input
4544 * parameters, and just pass these values here in the insert call.
4546 if (orient == MAP_STACK_GROWS_DOWN) {
4547 bot = addrbos + max_ssize - init_ssize;
4548 top = bot + init_ssize;
4551 } else /* if (orient == MAP_STACK_GROWS_UP) */ {
4553 top = bot + init_ssize;
4555 gap_top = addrbos + max_ssize;
4557 rv = vm_map_insert(map, NULL, 0, bot, top, prot, max, cow);
4558 if (rv != KERN_SUCCESS)
4560 new_entry = vm_map_entry_succ(prev_entry);
4561 KASSERT(new_entry->end == top || new_entry->start == bot,
4562 ("Bad entry start/end for new stack entry"));
4563 KASSERT((orient & MAP_STACK_GROWS_DOWN) == 0 ||
4564 (new_entry->eflags & MAP_ENTRY_GROWS_DOWN) != 0,
4565 ("new entry lacks MAP_ENTRY_GROWS_DOWN"));
4566 KASSERT((orient & MAP_STACK_GROWS_UP) == 0 ||
4567 (new_entry->eflags & MAP_ENTRY_GROWS_UP) != 0,
4568 ("new entry lacks MAP_ENTRY_GROWS_UP"));
4569 if (gap_bot == gap_top)
4570 return (KERN_SUCCESS);
4571 rv = vm_map_insert(map, NULL, 0, gap_bot, gap_top, VM_PROT_NONE,
4572 VM_PROT_NONE, MAP_CREATE_GUARD | (orient == MAP_STACK_GROWS_DOWN ?
4573 MAP_CREATE_STACK_GAP_DN : MAP_CREATE_STACK_GAP_UP));
4574 if (rv == KERN_SUCCESS) {
4576 * Gap can never successfully handle a fault, so
4577 * read-ahead logic is never used for it. Re-use
4578 * next_read of the gap entry to store
4579 * stack_guard_page for vm_map_growstack().
4581 if (orient == MAP_STACK_GROWS_DOWN)
4582 vm_map_entry_pred(new_entry)->next_read = sgp;
4584 vm_map_entry_succ(new_entry)->next_read = sgp;
4586 (void)vm_map_delete(map, bot, top);
4592 * Attempts to grow a vm stack entry. Returns KERN_SUCCESS if we
4593 * successfully grow the stack.
4596 vm_map_growstack(vm_map_t map, vm_offset_t addr, vm_map_entry_t gap_entry)
4598 vm_map_entry_t stack_entry;
4602 vm_offset_t gap_end, gap_start, grow_start;
4603 vm_size_t grow_amount, guard, max_grow;
4604 rlim_t lmemlim, stacklim, vmemlim;
4606 bool gap_deleted, grow_down, is_procstack;
4618 * Disallow stack growth when the access is performed by a
4619 * debugger or AIO daemon. The reason is that the wrong
4620 * resource limits are applied.
4622 if (p != initproc && (map != &p->p_vmspace->vm_map ||
4623 p->p_textvp == NULL))
4624 return (KERN_FAILURE);
4626 MPASS(!map->system_map);
4628 lmemlim = lim_cur(curthread, RLIMIT_MEMLOCK);
4629 stacklim = lim_cur(curthread, RLIMIT_STACK);
4630 vmemlim = lim_cur(curthread, RLIMIT_VMEM);
4632 /* If addr is not in a hole for a stack grow area, no need to grow. */
4633 if (gap_entry == NULL && !vm_map_lookup_entry(map, addr, &gap_entry))
4634 return (KERN_FAILURE);
4635 if ((gap_entry->eflags & MAP_ENTRY_GUARD) == 0)
4636 return (KERN_SUCCESS);
4637 if ((gap_entry->eflags & MAP_ENTRY_STACK_GAP_DN) != 0) {
4638 stack_entry = vm_map_entry_succ(gap_entry);
4639 if ((stack_entry->eflags & MAP_ENTRY_GROWS_DOWN) == 0 ||
4640 stack_entry->start != gap_entry->end)
4641 return (KERN_FAILURE);
4642 grow_amount = round_page(stack_entry->start - addr);
4644 } else if ((gap_entry->eflags & MAP_ENTRY_STACK_GAP_UP) != 0) {
4645 stack_entry = vm_map_entry_pred(gap_entry);
4646 if ((stack_entry->eflags & MAP_ENTRY_GROWS_UP) == 0 ||
4647 stack_entry->end != gap_entry->start)
4648 return (KERN_FAILURE);
4649 grow_amount = round_page(addr + 1 - stack_entry->end);
4652 return (KERN_FAILURE);
4654 guard = ((curproc->p_flag2 & P2_STKGAP_DISABLE) != 0 ||
4655 (curproc->p_fctl0 & NT_FREEBSD_FCTL_STKGAP_DISABLE) != 0) ? 0 :
4656 gap_entry->next_read;
4657 max_grow = gap_entry->end - gap_entry->start;
4658 if (guard > max_grow)
4659 return (KERN_NO_SPACE);
4661 if (grow_amount > max_grow)
4662 return (KERN_NO_SPACE);
4665 * If this is the main process stack, see if we're over the stack
4668 is_procstack = addr >= (vm_offset_t)vm->vm_maxsaddr &&
4669 addr < (vm_offset_t)vm->vm_stacktop;
4670 if (is_procstack && (ctob(vm->vm_ssize) + grow_amount > stacklim))
4671 return (KERN_NO_SPACE);
4676 if (is_procstack && racct_set(p, RACCT_STACK,
4677 ctob(vm->vm_ssize) + grow_amount)) {
4679 return (KERN_NO_SPACE);
4685 grow_amount = roundup(grow_amount, sgrowsiz);
4686 if (grow_amount > max_grow)
4687 grow_amount = max_grow;
4688 if (is_procstack && (ctob(vm->vm_ssize) + grow_amount > stacklim)) {
4689 grow_amount = trunc_page((vm_size_t)stacklim) -
4695 limit = racct_get_available(p, RACCT_STACK);
4697 if (is_procstack && (ctob(vm->vm_ssize) + grow_amount > limit))
4698 grow_amount = limit - ctob(vm->vm_ssize);
4701 if (!old_mlock && (map->flags & MAP_WIREFUTURE) != 0) {
4702 if (ptoa(pmap_wired_count(map->pmap)) + grow_amount > lmemlim) {
4709 if (racct_set(p, RACCT_MEMLOCK,
4710 ptoa(pmap_wired_count(map->pmap)) + grow_amount)) {
4720 /* If we would blow our VMEM resource limit, no go */
4721 if (map->size + grow_amount > vmemlim) {
4728 if (racct_set(p, RACCT_VMEM, map->size + grow_amount)) {
4737 if (vm_map_lock_upgrade(map)) {
4739 vm_map_lock_read(map);
4744 grow_start = gap_entry->end - grow_amount;
4745 if (gap_entry->start + grow_amount == gap_entry->end) {
4746 gap_start = gap_entry->start;
4747 gap_end = gap_entry->end;
4748 vm_map_entry_delete(map, gap_entry);
4751 MPASS(gap_entry->start < gap_entry->end - grow_amount);
4752 vm_map_entry_resize(map, gap_entry, -grow_amount);
4753 gap_deleted = false;
4755 rv = vm_map_insert(map, NULL, 0, grow_start,
4756 grow_start + grow_amount,
4757 stack_entry->protection, stack_entry->max_protection,
4758 MAP_STACK_GROWS_DOWN);
4759 if (rv != KERN_SUCCESS) {
4761 rv1 = vm_map_insert(map, NULL, 0, gap_start,
4762 gap_end, VM_PROT_NONE, VM_PROT_NONE,
4763 MAP_CREATE_GUARD | MAP_CREATE_STACK_GAP_DN);
4764 MPASS(rv1 == KERN_SUCCESS);
4766 vm_map_entry_resize(map, gap_entry,
4770 grow_start = stack_entry->end;
4771 cred = stack_entry->cred;
4772 if (cred == NULL && stack_entry->object.vm_object != NULL)
4773 cred = stack_entry->object.vm_object->cred;
4774 if (cred != NULL && !swap_reserve_by_cred(grow_amount, cred))
4776 /* Grow the underlying object if applicable. */
4777 else if (stack_entry->object.vm_object == NULL ||
4778 vm_object_coalesce(stack_entry->object.vm_object,
4779 stack_entry->offset,
4780 (vm_size_t)(stack_entry->end - stack_entry->start),
4781 grow_amount, cred != NULL)) {
4782 if (gap_entry->start + grow_amount == gap_entry->end) {
4783 vm_map_entry_delete(map, gap_entry);
4784 vm_map_entry_resize(map, stack_entry,
4787 gap_entry->start += grow_amount;
4788 stack_entry->end += grow_amount;
4790 map->size += grow_amount;
4795 if (rv == KERN_SUCCESS && is_procstack)
4796 vm->vm_ssize += btoc(grow_amount);
4799 * Heed the MAP_WIREFUTURE flag if it was set for this process.
4801 if (rv == KERN_SUCCESS && (map->flags & MAP_WIREFUTURE) != 0) {
4802 rv = vm_map_wire_locked(map, grow_start,
4803 grow_start + grow_amount,
4804 VM_MAP_WIRE_USER | VM_MAP_WIRE_NOHOLES);
4806 vm_map_lock_downgrade(map);
4810 if (racct_enable && rv != KERN_SUCCESS) {
4812 error = racct_set(p, RACCT_VMEM, map->size);
4813 KASSERT(error == 0, ("decreasing RACCT_VMEM failed"));
4815 error = racct_set(p, RACCT_MEMLOCK,
4816 ptoa(pmap_wired_count(map->pmap)));
4817 KASSERT(error == 0, ("decreasing RACCT_MEMLOCK failed"));
4819 error = racct_set(p, RACCT_STACK, ctob(vm->vm_ssize));
4820 KASSERT(error == 0, ("decreasing RACCT_STACK failed"));
4829 * Unshare the specified VM space for exec. If other processes are
4830 * mapped to it, then create a new one. The new vmspace is null.
4833 vmspace_exec(struct proc *p, vm_offset_t minuser, vm_offset_t maxuser)
4835 struct vmspace *oldvmspace = p->p_vmspace;
4836 struct vmspace *newvmspace;
4838 KASSERT((curthread->td_pflags & TDP_EXECVMSPC) == 0,
4839 ("vmspace_exec recursed"));
4840 newvmspace = vmspace_alloc(minuser, maxuser, pmap_pinit);
4841 if (newvmspace == NULL)
4843 newvmspace->vm_swrss = oldvmspace->vm_swrss;
4845 * This code is written like this for prototype purposes. The
4846 * goal is to avoid running down the vmspace here, but let the
4847 * other process's that are still using the vmspace to finally
4848 * run it down. Even though there is little or no chance of blocking
4849 * here, it is a good idea to keep this form for future mods.
4851 PROC_VMSPACE_LOCK(p);
4852 p->p_vmspace = newvmspace;
4853 PROC_VMSPACE_UNLOCK(p);
4854 if (p == curthread->td_proc)
4855 pmap_activate(curthread);
4856 curthread->td_pflags |= TDP_EXECVMSPC;
4861 * Unshare the specified VM space for forcing COW. This
4862 * is called by rfork, for the (RFMEM|RFPROC) == 0 case.
4865 vmspace_unshare(struct proc *p)
4867 struct vmspace *oldvmspace = p->p_vmspace;
4868 struct vmspace *newvmspace;
4869 vm_ooffset_t fork_charge;
4872 * The caller is responsible for ensuring that the reference count
4873 * cannot concurrently transition 1 -> 2.
4875 if (refcount_load(&oldvmspace->vm_refcnt) == 1)
4878 newvmspace = vmspace_fork(oldvmspace, &fork_charge);
4879 if (newvmspace == NULL)
4881 if (!swap_reserve_by_cred(fork_charge, p->p_ucred)) {
4882 vmspace_free(newvmspace);
4885 PROC_VMSPACE_LOCK(p);
4886 p->p_vmspace = newvmspace;
4887 PROC_VMSPACE_UNLOCK(p);
4888 if (p == curthread->td_proc)
4889 pmap_activate(curthread);
4890 vmspace_free(oldvmspace);
4897 * Finds the VM object, offset, and
4898 * protection for a given virtual address in the
4899 * specified map, assuming a page fault of the
4902 * Leaves the map in question locked for read; return
4903 * values are guaranteed until a vm_map_lookup_done
4904 * call is performed. Note that the map argument
4905 * is in/out; the returned map must be used in
4906 * the call to vm_map_lookup_done.
4908 * A handle (out_entry) is returned for use in
4909 * vm_map_lookup_done, to make that fast.
4911 * If a lookup is requested with "write protection"
4912 * specified, the map may be changed to perform virtual
4913 * copying operations, although the data referenced will
4917 vm_map_lookup(vm_map_t *var_map, /* IN/OUT */
4919 vm_prot_t fault_typea,
4920 vm_map_entry_t *out_entry, /* OUT */
4921 vm_object_t *object, /* OUT */
4922 vm_pindex_t *pindex, /* OUT */
4923 vm_prot_t *out_prot, /* OUT */
4924 boolean_t *wired) /* OUT */
4926 vm_map_entry_t entry;
4927 vm_map_t map = *var_map;
4929 vm_prot_t fault_type;
4930 vm_object_t eobject;
4936 vm_map_lock_read(map);
4940 * Lookup the faulting address.
4942 if (!vm_map_lookup_entry(map, vaddr, out_entry)) {
4943 vm_map_unlock_read(map);
4944 return (KERN_INVALID_ADDRESS);
4952 if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) {
4953 vm_map_t old_map = map;
4955 *var_map = map = entry->object.sub_map;
4956 vm_map_unlock_read(old_map);
4961 * Check whether this task is allowed to have this page.
4963 prot = entry->protection;
4964 if ((fault_typea & VM_PROT_FAULT_LOOKUP) != 0) {
4965 fault_typea &= ~VM_PROT_FAULT_LOOKUP;
4966 if (prot == VM_PROT_NONE && map != kernel_map &&
4967 (entry->eflags & MAP_ENTRY_GUARD) != 0 &&
4968 (entry->eflags & (MAP_ENTRY_STACK_GAP_DN |
4969 MAP_ENTRY_STACK_GAP_UP)) != 0 &&
4970 vm_map_growstack(map, vaddr, entry) == KERN_SUCCESS)
4971 goto RetryLookupLocked;
4973 fault_type = fault_typea & VM_PROT_ALL;
4974 if ((fault_type & prot) != fault_type || prot == VM_PROT_NONE) {
4975 vm_map_unlock_read(map);
4976 return (KERN_PROTECTION_FAILURE);
4978 KASSERT((prot & VM_PROT_WRITE) == 0 || (entry->eflags &
4979 (MAP_ENTRY_USER_WIRED | MAP_ENTRY_NEEDS_COPY)) !=
4980 (MAP_ENTRY_USER_WIRED | MAP_ENTRY_NEEDS_COPY),
4981 ("entry %p flags %x", entry, entry->eflags));
4982 if ((fault_typea & VM_PROT_COPY) != 0 &&
4983 (entry->max_protection & VM_PROT_WRITE) == 0 &&
4984 (entry->eflags & MAP_ENTRY_COW) == 0) {
4985 vm_map_unlock_read(map);
4986 return (KERN_PROTECTION_FAILURE);
4990 * If this page is not pageable, we have to get it for all possible
4993 *wired = (entry->wired_count != 0);
4995 fault_type = entry->protection;
4996 size = entry->end - entry->start;
4999 * If the entry was copy-on-write, we either ...
5001 if (entry->eflags & MAP_ENTRY_NEEDS_COPY) {
5003 * If we want to write the page, we may as well handle that
5004 * now since we've got the map locked.
5006 * If we don't need to write the page, we just demote the
5007 * permissions allowed.
5009 if ((fault_type & VM_PROT_WRITE) != 0 ||
5010 (fault_typea & VM_PROT_COPY) != 0) {
5012 * Make a new object, and place it in the object
5013 * chain. Note that no new references have appeared
5014 * -- one just moved from the map to the new
5017 if (vm_map_lock_upgrade(map))
5020 if (entry->cred == NULL) {
5022 * The debugger owner is charged for
5025 cred = curthread->td_ucred;
5027 if (!swap_reserve_by_cred(size, cred)) {
5030 return (KERN_RESOURCE_SHORTAGE);
5034 eobject = entry->object.vm_object;
5035 vm_object_shadow(&entry->object.vm_object,
5036 &entry->offset, size, entry->cred, false);
5037 if (eobject == entry->object.vm_object) {
5039 * The object was not shadowed.
5041 swap_release_by_cred(size, entry->cred);
5042 crfree(entry->cred);
5045 entry->eflags &= ~MAP_ENTRY_NEEDS_COPY;
5047 vm_map_lock_downgrade(map);
5050 * We're attempting to read a copy-on-write page --
5051 * don't allow writes.
5053 prot &= ~VM_PROT_WRITE;
5058 * Create an object if necessary.
5060 if (entry->object.vm_object == NULL && !map->system_map) {
5061 if (vm_map_lock_upgrade(map))
5063 entry->object.vm_object = vm_object_allocate_anon(atop(size),
5064 NULL, entry->cred, entry->cred != NULL ? size : 0);
5067 vm_map_lock_downgrade(map);
5071 * Return the object/offset from this entry. If the entry was
5072 * copy-on-write or empty, it has been fixed up.
5074 *pindex = OFF_TO_IDX((vaddr - entry->start) + entry->offset);
5075 *object = entry->object.vm_object;
5078 return (KERN_SUCCESS);
5082 * vm_map_lookup_locked:
5084 * Lookup the faulting address. A version of vm_map_lookup that returns
5085 * KERN_FAILURE instead of blocking on map lock or memory allocation.
5088 vm_map_lookup_locked(vm_map_t *var_map, /* IN/OUT */
5090 vm_prot_t fault_typea,
5091 vm_map_entry_t *out_entry, /* OUT */
5092 vm_object_t *object, /* OUT */
5093 vm_pindex_t *pindex, /* OUT */
5094 vm_prot_t *out_prot, /* OUT */
5095 boolean_t *wired) /* OUT */
5097 vm_map_entry_t entry;
5098 vm_map_t map = *var_map;
5100 vm_prot_t fault_type = fault_typea;
5103 * Lookup the faulting address.
5105 if (!vm_map_lookup_entry(map, vaddr, out_entry))
5106 return (KERN_INVALID_ADDRESS);
5111 * Fail if the entry refers to a submap.
5113 if (entry->eflags & MAP_ENTRY_IS_SUB_MAP)
5114 return (KERN_FAILURE);
5117 * Check whether this task is allowed to have this page.
5119 prot = entry->protection;
5120 fault_type &= VM_PROT_READ | VM_PROT_WRITE | VM_PROT_EXECUTE;
5121 if ((fault_type & prot) != fault_type)
5122 return (KERN_PROTECTION_FAILURE);
5125 * If this page is not pageable, we have to get it for all possible
5128 *wired = (entry->wired_count != 0);
5130 fault_type = entry->protection;
5132 if (entry->eflags & MAP_ENTRY_NEEDS_COPY) {
5134 * Fail if the entry was copy-on-write for a write fault.
5136 if (fault_type & VM_PROT_WRITE)
5137 return (KERN_FAILURE);
5139 * We're attempting to read a copy-on-write page --
5140 * don't allow writes.
5142 prot &= ~VM_PROT_WRITE;
5146 * Fail if an object should be created.
5148 if (entry->object.vm_object == NULL && !map->system_map)
5149 return (KERN_FAILURE);
5152 * Return the object/offset from this entry. If the entry was
5153 * copy-on-write or empty, it has been fixed up.
5155 *pindex = OFF_TO_IDX((vaddr - entry->start) + entry->offset);
5156 *object = entry->object.vm_object;
5159 return (KERN_SUCCESS);
5163 * vm_map_lookup_done:
5165 * Releases locks acquired by a vm_map_lookup
5166 * (according to the handle returned by that lookup).
5169 vm_map_lookup_done(vm_map_t map, vm_map_entry_t entry)
5172 * Unlock the main-level map
5174 vm_map_unlock_read(map);
5178 vm_map_max_KBI(const struct vm_map *map)
5181 return (vm_map_max(map));
5185 vm_map_min_KBI(const struct vm_map *map)
5188 return (vm_map_min(map));
5192 vm_map_pmap_KBI(vm_map_t map)
5199 vm_map_range_valid_KBI(vm_map_t map, vm_offset_t start, vm_offset_t end)
5202 return (vm_map_range_valid(map, start, end));
5207 _vm_map_assert_consistent(vm_map_t map, int check)
5209 vm_map_entry_t entry, prev;
5210 vm_map_entry_t cur, header, lbound, ubound;
5211 vm_size_t max_left, max_right;
5216 if (enable_vmmap_check != check)
5219 header = prev = &map->header;
5220 VM_MAP_ENTRY_FOREACH(entry, map) {
5221 KASSERT(prev->end <= entry->start,
5222 ("map %p prev->end = %jx, start = %jx", map,
5223 (uintmax_t)prev->end, (uintmax_t)entry->start));
5224 KASSERT(entry->start < entry->end,
5225 ("map %p start = %jx, end = %jx", map,
5226 (uintmax_t)entry->start, (uintmax_t)entry->end));
5227 KASSERT(entry->left == header ||
5228 entry->left->start < entry->start,
5229 ("map %p left->start = %jx, start = %jx", map,
5230 (uintmax_t)entry->left->start, (uintmax_t)entry->start));
5231 KASSERT(entry->right == header ||
5232 entry->start < entry->right->start,
5233 ("map %p start = %jx, right->start = %jx", map,
5234 (uintmax_t)entry->start, (uintmax_t)entry->right->start));
5236 lbound = ubound = header;
5238 if (entry->start < cur->start) {
5241 KASSERT(cur != lbound,
5242 ("map %p cannot find %jx",
5243 map, (uintmax_t)entry->start));
5244 } else if (cur->end <= entry->start) {
5247 KASSERT(cur != ubound,
5248 ("map %p cannot find %jx",
5249 map, (uintmax_t)entry->start));
5251 KASSERT(cur == entry,
5252 ("map %p cannot find %jx",
5253 map, (uintmax_t)entry->start));
5257 max_left = vm_map_entry_max_free_left(entry, lbound);
5258 max_right = vm_map_entry_max_free_right(entry, ubound);
5259 KASSERT(entry->max_free == vm_size_max(max_left, max_right),
5260 ("map %p max = %jx, max_left = %jx, max_right = %jx", map,
5261 (uintmax_t)entry->max_free,
5262 (uintmax_t)max_left, (uintmax_t)max_right));
5265 KASSERT(prev->end <= entry->start,
5266 ("map %p prev->end = %jx, start = %jx", map,
5267 (uintmax_t)prev->end, (uintmax_t)entry->start));
5271 #include "opt_ddb.h"
5273 #include <sys/kernel.h>
5275 #include <ddb/ddb.h>
5278 vm_map_print(vm_map_t map)
5280 vm_map_entry_t entry, prev;
5282 db_iprintf("Task map %p: pmap=%p, nentries=%d, version=%u\n",
5284 (void *)map->pmap, map->nentries, map->timestamp);
5287 prev = &map->header;
5288 VM_MAP_ENTRY_FOREACH(entry, map) {
5289 db_iprintf("map entry %p: start=%p, end=%p, eflags=%#x, \n",
5290 (void *)entry, (void *)entry->start, (void *)entry->end,
5293 static const char * const inheritance_name[4] =
5294 {"share", "copy", "none", "donate_copy"};
5296 db_iprintf(" prot=%x/%x/%s",
5298 entry->max_protection,
5299 inheritance_name[(int)(unsigned char)
5300 entry->inheritance]);
5301 if (entry->wired_count != 0)
5302 db_printf(", wired");
5304 if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) {
5305 db_printf(", share=%p, offset=0x%jx\n",
5306 (void *)entry->object.sub_map,
5307 (uintmax_t)entry->offset);
5308 if (prev == &map->header ||
5309 prev->object.sub_map !=
5310 entry->object.sub_map) {
5312 vm_map_print((vm_map_t)entry->object.sub_map);
5316 if (entry->cred != NULL)
5317 db_printf(", ruid %d", entry->cred->cr_ruid);
5318 db_printf(", object=%p, offset=0x%jx",
5319 (void *)entry->object.vm_object,
5320 (uintmax_t)entry->offset);
5321 if (entry->object.vm_object && entry->object.vm_object->cred)
5322 db_printf(", obj ruid %d charge %jx",
5323 entry->object.vm_object->cred->cr_ruid,
5324 (uintmax_t)entry->object.vm_object->charge);
5325 if (entry->eflags & MAP_ENTRY_COW)
5326 db_printf(", copy (%s)",
5327 (entry->eflags & MAP_ENTRY_NEEDS_COPY) ? "needed" : "done");
5330 if (prev == &map->header ||
5331 prev->object.vm_object !=
5332 entry->object.vm_object) {
5334 vm_object_print((db_expr_t)(intptr_t)
5335 entry->object.vm_object,
5345 DB_SHOW_COMMAND(map, map)
5349 db_printf("usage: show map <addr>\n");
5352 vm_map_print((vm_map_t)addr);
5355 DB_SHOW_COMMAND(procvm, procvm)
5360 p = db_lookup_proc(addr);
5365 db_printf("p = %p, vmspace = %p, map = %p, pmap = %p\n",
5366 (void *)p, (void *)p->p_vmspace, (void *)&p->p_vmspace->vm_map,
5367 (void *)vmspace_pmap(p->p_vmspace));
5369 vm_map_print((vm_map_t)&p->p_vmspace->vm_map);