4 * Copyright (c) 1991, 1993
5 * The Regents of the University of California. All rights reserved.
7 * This code is derived from software contributed to Berkeley by
8 * The Mach Operating System project at Carnegie-Mellon University.
10 * Redistribution and use in source and binary forms, with or without
11 * modification, are permitted provided that the following conditions
13 * 1. Redistributions of source code must retain the above copyright
14 * notice, this list of conditions and the following disclaimer.
15 * 2. Redistributions in binary form must reproduce the above copyright
16 * notice, this list of conditions and the following disclaimer in the
17 * documentation and/or other materials provided with the distribution.
18 * 3. All advertising materials mentioning features or use of this software
19 * must display the following acknowledgement:
20 * This product includes software developed by the University of
21 * California, Berkeley and its contributors.
22 * 4. Neither the name of the University nor the names of its contributors
23 * may be used to endorse or promote products derived from this software
24 * without specific prior written permission.
26 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
27 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
28 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
29 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
30 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
31 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
32 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
33 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
34 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
35 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
38 * from: @(#)vm_map.c 8.3 (Berkeley) 1/12/94
41 * Copyright (c) 1987, 1990 Carnegie-Mellon University.
42 * All rights reserved.
44 * Authors: Avadis Tevanian, Jr., Michael Wayne Young
46 * Permission to use, copy, modify and distribute this software and
47 * its documentation is hereby granted, provided that both the copyright
48 * notice and this permission notice appear in all copies of the
49 * software, derivative works or modified versions, and any portions
50 * thereof, and that both notices appear in supporting documentation.
52 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
53 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
54 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
56 * Carnegie Mellon requests users of this software to return to
58 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU
59 * School of Computer Science
60 * Carnegie Mellon University
61 * Pittsburgh PA 15213-3890
63 * any improvements or extensions that they make and grant Carnegie the
64 * rights to redistribute these changes.
66 * $FreeBSD: src/sys/vm/vm_map.c,v 1.187.2.19 2003/05/27 00:47:02 alc Exp $
67 * $DragonFly: src/sys/vm/vm_map.c,v 1.56 2007/04/29 18:25:41 dillon Exp $
71 * Virtual memory mapping module.
74 #include <sys/param.h>
75 #include <sys/systm.h>
76 #include <sys/kernel.h>
78 #include <sys/serialize.h>
80 #include <sys/vmmeter.h>
82 #include <sys/vnode.h>
83 #include <sys/resourcevar.h>
86 #include <sys/malloc.h>
89 #include <vm/vm_param.h>
91 #include <vm/vm_map.h>
92 #include <vm/vm_page.h>
93 #include <vm/vm_object.h>
94 #include <vm/vm_pager.h>
95 #include <vm/vm_kern.h>
96 #include <vm/vm_extern.h>
97 #include <vm/swap_pager.h>
98 #include <vm/vm_zone.h>
100 #include <sys/thread2.h>
101 #include <sys/sysref2.h>
102 #include <sys/random.h>
103 #include <sys/sysctl.h>
106 * Virtual memory maps provide for the mapping, protection, and sharing
107 * of virtual memory objects. In addition, this module provides for an
108 * efficient virtual copy of memory from one map to another.
110 * Synchronization is required prior to most operations.
112 * Maps consist of an ordered doubly-linked list of simple entries.
113 * A hint and a RB tree is used to speed-up lookups.
115 * Callers looking to modify maps specify start/end addresses which cause
116 * the related map entry to be clipped if necessary, and then later
117 * recombined if the pieces remained compatible.
119 * Virtual copy operations are performed by copying VM object references
120 * from one map to another, and then marking both regions as copy-on-write.
122 static void vmspace_terminate(struct vmspace *vm);
123 static void vmspace_lock(struct vmspace *vm);
124 static void vmspace_unlock(struct vmspace *vm);
125 static void vmspace_dtor(void *obj, void *private);
127 MALLOC_DEFINE(M_VMSPACE, "vmspace", "vmspace objcache backingstore");
129 struct sysref_class vmspace_sysref_class = {
132 .proto = SYSREF_PROTO_VMSPACE,
133 .offset = offsetof(struct vmspace, vm_sysref),
134 .objsize = sizeof(struct vmspace),
136 .flags = SRC_MANAGEDINIT,
137 .dtor = vmspace_dtor,
139 .terminate = (sysref_terminate_func_t)vmspace_terminate,
140 .lock = (sysref_lock_func_t)vmspace_lock,
141 .unlock = (sysref_lock_func_t)vmspace_unlock
147 static struct vm_zone mapentzone_store, mapzone_store;
148 static vm_zone_t mapentzone, mapzone;
149 static struct vm_object mapentobj, mapobj;
151 static struct vm_map_entry map_entry_init[MAX_MAPENT];
152 static struct vm_map_entry cpu_map_entry_init[MAXCPU][VMEPERCPU];
153 static struct vm_map map_init[MAX_KMAP];
155 static int randomize_mmap;
156 SYSCTL_INT(_vm, OID_AUTO, randomize_mmap, CTLFLAG_RW, &randomize_mmap, 0,
157 "Randomize mmap offsets");
159 static void vm_map_entry_shadow(vm_map_entry_t entry);
160 static vm_map_entry_t vm_map_entry_create(vm_map_t map, int *);
161 static void vm_map_entry_dispose (vm_map_t map, vm_map_entry_t entry, int *);
162 static void _vm_map_clip_end (vm_map_t, vm_map_entry_t, vm_offset_t, int *);
163 static void _vm_map_clip_start (vm_map_t, vm_map_entry_t, vm_offset_t, int *);
164 static void vm_map_entry_delete (vm_map_t, vm_map_entry_t, int *);
165 static void vm_map_entry_unwire (vm_map_t, vm_map_entry_t);
166 static void vm_map_copy_entry (vm_map_t, vm_map_t, vm_map_entry_t,
168 static void vm_map_split (vm_map_entry_t);
169 static void vm_map_unclip_range (vm_map_t map, vm_map_entry_t start_entry, vm_offset_t start, vm_offset_t end, int *count, int flags);
172 * Initialize the vm_map module. Must be called before any other vm_map
175 * Map and entry structures are allocated from the general purpose
176 * memory pool with some exceptions:
178 * - The kernel map is allocated statically.
179 * - Initial kernel map entries are allocated out of a static pool.
181 * These restrictions are necessary since malloc() uses the
182 * maps and requires map entries.
184 * Called from the low level boot code only.
189 mapzone = &mapzone_store;
190 zbootinit(mapzone, "MAP", sizeof (struct vm_map),
192 mapentzone = &mapentzone_store;
193 zbootinit(mapentzone, "MAP ENTRY", sizeof (struct vm_map_entry),
194 map_entry_init, MAX_MAPENT);
198 * Called prior to any vmspace allocations.
200 * Called from the low level boot code only.
205 zinitna(mapentzone, &mapentobj, NULL, 0, 0,
206 ZONE_USE_RESERVE | ZONE_SPECIAL, 1);
207 zinitna(mapzone, &mapobj, NULL, 0, 0, 0, 1);
214 * Red black tree functions
216 * The caller must hold the related map lock.
218 static int rb_vm_map_compare(vm_map_entry_t a, vm_map_entry_t b);
219 RB_GENERATE(vm_map_rb_tree, vm_map_entry, rb_entry, rb_vm_map_compare);
221 /* a->start is address, and the only field has to be initialized */
223 rb_vm_map_compare(vm_map_entry_t a, vm_map_entry_t b)
225 if (a->start < b->start)
227 else if (a->start > b->start)
233 * Allocate a vmspace structure, including a vm_map and pmap.
234 * Initialize numerous fields. While the initial allocation is zerod,
235 * subsequence reuse from the objcache leaves elements of the structure
236 * intact (particularly the pmap), so portions must be zerod.
238 * The structure is not considered activated until we call sysref_activate().
243 vmspace_alloc(vm_offset_t min, vm_offset_t max)
247 lwkt_gettoken(&vmspace_token);
248 vm = sysref_alloc(&vmspace_sysref_class);
249 bzero(&vm->vm_startcopy,
250 (char *)&vm->vm_endcopy - (char *)&vm->vm_startcopy);
251 vm_map_init(&vm->vm_map, min, max, NULL);
252 pmap_pinit(vmspace_pmap(vm)); /* (some fields reused) */
253 vm->vm_map.pmap = vmspace_pmap(vm); /* XXX */
255 vm->vm_exitingcnt = 0;
256 cpu_vmspace_alloc(vm);
257 sysref_activate(&vm->vm_sysref);
258 lwkt_reltoken(&vmspace_token);
264 * dtor function - Some elements of the pmap are retained in the
265 * free-cached vmspaces to improve performance. We have to clean them up
266 * here before returning the vmspace to the memory pool.
271 vmspace_dtor(void *obj, void *private)
273 struct vmspace *vm = obj;
275 pmap_puninit(vmspace_pmap(vm));
279 * Called in two cases:
281 * (1) When the last sysref is dropped, but exitingcnt might still be
284 * (2) When there are no sysrefs (i.e. refcnt is negative) left and the
285 * exitingcnt becomes zero
287 * sysref will not scrap the object until we call sysref_put() once more
288 * after the last ref has been dropped.
290 * Interlocked by the sysref API.
293 vmspace_terminate(struct vmspace *vm)
298 * If exitingcnt is non-zero we can't get rid of the entire vmspace
299 * yet, but we can scrap user memory.
301 lwkt_gettoken(&vmspace_token);
302 if (vm->vm_exitingcnt) {
304 pmap_remove_pages(vmspace_pmap(vm), VM_MIN_USER_ADDRESS,
305 VM_MAX_USER_ADDRESS);
306 vm_map_remove(&vm->vm_map, VM_MIN_USER_ADDRESS,
307 VM_MAX_USER_ADDRESS);
308 lwkt_reltoken(&vmspace_token);
311 cpu_vmspace_free(vm);
314 * Make sure any SysV shm is freed, it might not have in
319 KKASSERT(vm->vm_upcalls == NULL);
322 * Lock the map, to wait out all other references to it.
323 * Delete all of the mappings and pages they hold, then call
324 * the pmap module to reclaim anything left.
326 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
327 vm_map_lock(&vm->vm_map);
328 vm_map_delete(&vm->vm_map, vm->vm_map.min_offset,
329 vm->vm_map.max_offset, &count);
330 vm_map_unlock(&vm->vm_map);
331 vm_map_entry_release(count);
333 pmap_release(vmspace_pmap(vm));
334 sysref_put(&vm->vm_sysref);
335 lwkt_reltoken(&vmspace_token);
339 * vmspaces are not currently locked.
342 vmspace_lock(struct vmspace *vm __unused)
347 vmspace_unlock(struct vmspace *vm __unused)
352 * This is called during exit indicating that the vmspace is no
353 * longer in used by an exiting process, but the process has not yet
359 vmspace_exitbump(struct vmspace *vm)
361 lwkt_gettoken(&vmspace_token);
363 lwkt_reltoken(&vmspace_token);
367 * This is called in the wait*() handling code. The vmspace can be terminated
368 * after the last wait is finished using it.
373 vmspace_exitfree(struct proc *p)
377 lwkt_gettoken(&vmspace_token);
381 if (--vm->vm_exitingcnt == 0 && sysref_isinactive(&vm->vm_sysref))
382 vmspace_terminate(vm);
383 lwkt_reltoken(&vmspace_token);
387 * Swap useage is determined by taking the proportional swap used by
388 * VM objects backing the VM map. To make up for fractional losses,
389 * if the VM object has any swap use at all the associated map entries
390 * count for at least 1 swap page.
395 vmspace_swap_count(struct vmspace *vmspace)
397 vm_map_t map = &vmspace->vm_map;
403 lwkt_gettoken(&vmspace_token);
404 for (cur = map->header.next; cur != &map->header; cur = cur->next) {
405 switch(cur->maptype) {
406 case VM_MAPTYPE_NORMAL:
407 case VM_MAPTYPE_VPAGETABLE:
408 if ((object = cur->object.vm_object) == NULL)
410 if (object->swblock_count) {
411 n = (cur->end - cur->start) / PAGE_SIZE;
412 count += object->swblock_count *
413 SWAP_META_PAGES * n / object->size + 1;
420 lwkt_reltoken(&vmspace_token);
425 * Calculate the approximate number of anonymous pages in use by
426 * this vmspace. To make up for fractional losses, we count each
427 * VM object as having at least 1 anonymous page.
432 vmspace_anonymous_count(struct vmspace *vmspace)
434 vm_map_t map = &vmspace->vm_map;
439 lwkt_gettoken(&vmspace_token);
440 for (cur = map->header.next; cur != &map->header; cur = cur->next) {
441 switch(cur->maptype) {
442 case VM_MAPTYPE_NORMAL:
443 case VM_MAPTYPE_VPAGETABLE:
444 if ((object = cur->object.vm_object) == NULL)
446 if (object->type != OBJT_DEFAULT &&
447 object->type != OBJT_SWAP) {
450 count += object->resident_page_count;
456 lwkt_reltoken(&vmspace_token);
461 * Creates and returns a new empty VM map with the given physical map
462 * structure, and having the given lower and upper address bounds.
467 vm_map_create(vm_map_t result, pmap_t pmap, vm_offset_t min, vm_offset_t max)
470 result = zalloc(mapzone);
471 vm_map_init(result, min, max, pmap);
476 * Initialize an existing vm_map structure such as that in the vmspace
477 * structure. The pmap is initialized elsewhere.
482 vm_map_init(struct vm_map *map, vm_offset_t min, vm_offset_t max, pmap_t pmap)
484 map->header.next = map->header.prev = &map->header;
485 RB_INIT(&map->rb_root);
490 map->min_offset = min;
491 map->max_offset = max;
493 map->first_free = &map->header;
494 map->hint = &map->header;
497 lockinit(&map->lock, "thrd_sleep", 0, 0);
501 * Shadow the vm_map_entry's object. This typically needs to be done when
502 * a write fault is taken on an entry which had previously been cloned by
503 * fork(). The shared object (which might be NULL) must become private so
504 * we add a shadow layer above it.
506 * Object allocation for anonymous mappings is defered as long as possible.
507 * When creating a shadow, however, the underlying object must be instantiated
508 * so it can be shared.
510 * If the map segment is governed by a virtual page table then it is
511 * possible to address offsets beyond the mapped area. Just allocate
512 * a maximally sized object for this case.
514 * The vm_map must be exclusively locked.
515 * No other requirements.
519 vm_map_entry_shadow(vm_map_entry_t entry)
521 if (entry->maptype == VM_MAPTYPE_VPAGETABLE) {
522 vm_object_shadow(&entry->object.vm_object, &entry->offset,
523 0x7FFFFFFF); /* XXX */
525 vm_object_shadow(&entry->object.vm_object, &entry->offset,
526 atop(entry->end - entry->start));
528 entry->eflags &= ~MAP_ENTRY_NEEDS_COPY;
532 * Allocate an object for a vm_map_entry.
534 * Object allocation for anonymous mappings is defered as long as possible.
535 * This function is called when we can defer no longer, generally when a map
536 * entry might be split or forked or takes a page fault.
538 * If the map segment is governed by a virtual page table then it is
539 * possible to address offsets beyond the mapped area. Just allocate
540 * a maximally sized object for this case.
542 * The vm_map must be exclusively locked.
543 * No other requirements.
546 vm_map_entry_allocate_object(vm_map_entry_t entry)
550 if (entry->maptype == VM_MAPTYPE_VPAGETABLE) {
551 obj = vm_object_allocate(OBJT_DEFAULT, 0x7FFFFFFF); /* XXX */
553 obj = vm_object_allocate(OBJT_DEFAULT,
554 atop(entry->end - entry->start));
556 entry->object.vm_object = obj;
561 * Set an initial negative count so the first attempt to reserve
562 * space preloads a bunch of vm_map_entry's for this cpu. Also
563 * pre-allocate 2 vm_map_entries which will be needed by zalloc() to
564 * map a new page for vm_map_entry structures. SMP systems are
565 * particularly sensitive.
567 * This routine is called in early boot so we cannot just call
568 * vm_map_entry_reserve().
570 * Called from the low level boot code only (for each cpu)
573 vm_map_entry_reserve_cpu_init(globaldata_t gd)
575 vm_map_entry_t entry;
578 gd->gd_vme_avail -= MAP_RESERVE_COUNT * 2;
579 entry = &cpu_map_entry_init[gd->gd_cpuid][0];
580 for (i = 0; i < VMEPERCPU; ++i, ++entry) {
581 entry->next = gd->gd_vme_base;
582 gd->gd_vme_base = entry;
587 * Reserves vm_map_entry structures so code later on can manipulate
588 * map_entry structures within a locked map without blocking trying
589 * to allocate a new vm_map_entry.
594 vm_map_entry_reserve(int count)
596 struct globaldata *gd = mycpu;
597 vm_map_entry_t entry;
600 * Make sure we have enough structures in gd_vme_base to handle
601 * the reservation request.
604 while (gd->gd_vme_avail < count) {
605 entry = zalloc(mapentzone);
606 entry->next = gd->gd_vme_base;
607 gd->gd_vme_base = entry;
610 gd->gd_vme_avail -= count;
617 * Releases previously reserved vm_map_entry structures that were not
618 * used. If we have too much junk in our per-cpu cache clean some of
624 vm_map_entry_release(int count)
626 struct globaldata *gd = mycpu;
627 vm_map_entry_t entry;
630 gd->gd_vme_avail += count;
631 while (gd->gd_vme_avail > MAP_RESERVE_SLOP) {
632 entry = gd->gd_vme_base;
633 KKASSERT(entry != NULL);
634 gd->gd_vme_base = entry->next;
637 zfree(mapentzone, entry);
644 * Reserve map entry structures for use in kernel_map itself. These
645 * entries have *ALREADY* been reserved on a per-cpu basis when the map
646 * was inited. This function is used by zalloc() to avoid a recursion
647 * when zalloc() itself needs to allocate additional kernel memory.
649 * This function works like the normal reserve but does not load the
650 * vm_map_entry cache (because that would result in an infinite
651 * recursion). Note that gd_vme_avail may go negative. This is expected.
653 * Any caller of this function must be sure to renormalize after
654 * potentially eating entries to ensure that the reserve supply
660 vm_map_entry_kreserve(int count)
662 struct globaldata *gd = mycpu;
665 gd->gd_vme_avail -= count;
667 KASSERT(gd->gd_vme_base != NULL,
668 ("no reserved entries left, gd_vme_avail = %d\n",
674 * Release previously reserved map entries for kernel_map. We do not
675 * attempt to clean up like the normal release function as this would
676 * cause an unnecessary (but probably not fatal) deep procedure call.
681 vm_map_entry_krelease(int count)
683 struct globaldata *gd = mycpu;
686 gd->gd_vme_avail += count;
691 * Allocates a VM map entry for insertion. No entry fields are filled in.
693 * The entries should have previously been reserved. The reservation count
694 * is tracked in (*countp).
698 static vm_map_entry_t
699 vm_map_entry_create(vm_map_t map, int *countp)
701 struct globaldata *gd = mycpu;
702 vm_map_entry_t entry;
704 KKASSERT(*countp > 0);
707 entry = gd->gd_vme_base;
708 KASSERT(entry != NULL, ("gd_vme_base NULL! count %d", *countp));
709 gd->gd_vme_base = entry->next;
716 * Dispose of a vm_map_entry that is no longer being referenced.
721 vm_map_entry_dispose(vm_map_t map, vm_map_entry_t entry, int *countp)
723 struct globaldata *gd = mycpu;
725 KKASSERT(map->hint != entry);
726 KKASSERT(map->first_free != entry);
730 entry->next = gd->gd_vme_base;
731 gd->gd_vme_base = entry;
737 * Insert/remove entries from maps.
739 * The related map must be exclusively locked.
740 * No other requirements.
742 * NOTE! We currently acquire the vmspace_token only to avoid races
743 * against the pageout daemon's calls to vmspace_*_count(), which
744 * are unable to safely lock the vm_map without potentially
748 vm_map_entry_link(vm_map_t map,
749 vm_map_entry_t after_where,
750 vm_map_entry_t entry)
752 ASSERT_VM_MAP_LOCKED(map);
754 lwkt_gettoken(&vmspace_token);
756 entry->prev = after_where;
757 entry->next = after_where->next;
758 entry->next->prev = entry;
759 after_where->next = entry;
760 if (vm_map_rb_tree_RB_INSERT(&map->rb_root, entry))
761 panic("vm_map_entry_link: dup addr map %p ent %p", map, entry);
762 lwkt_reltoken(&vmspace_token);
766 vm_map_entry_unlink(vm_map_t map,
767 vm_map_entry_t entry)
772 ASSERT_VM_MAP_LOCKED(map);
774 if (entry->eflags & MAP_ENTRY_IN_TRANSITION) {
775 panic("vm_map_entry_unlink: attempt to mess with "
776 "locked entry! %p", entry);
778 lwkt_gettoken(&vmspace_token);
783 vm_map_rb_tree_RB_REMOVE(&map->rb_root, entry);
785 lwkt_reltoken(&vmspace_token);
789 * Finds the map entry containing (or immediately preceding) the specified
790 * address in the given map. The entry is returned in (*entry).
792 * The boolean result indicates whether the address is actually contained
795 * The related map must be locked.
796 * No other requirements.
799 vm_map_lookup_entry(vm_map_t map, vm_offset_t address, vm_map_entry_t *entry)
804 ASSERT_VM_MAP_LOCKED(map);
807 * XXX TEMPORARILY DISABLED. For some reason our attempt to revive
808 * the hint code with the red-black lookup meets with system crashes
809 * and lockups. We do not yet know why.
811 * It is possible that the problem is related to the setting
812 * of the hint during map_entry deletion, in the code specified
813 * at the GGG comment later on in this file.
816 * Quickly check the cached hint, there's a good chance of a match.
818 if (map->hint != &map->header) {
820 if (address >= tmp->start && address < tmp->end) {
828 * Locate the record from the top of the tree. 'last' tracks the
829 * closest prior record and is returned if no match is found, which
830 * in binary tree terms means tracking the most recent right-branch
831 * taken. If there is no prior record, &map->header is returned.
834 tmp = RB_ROOT(&map->rb_root);
837 if (address >= tmp->start) {
838 if (address < tmp->end) {
844 tmp = RB_RIGHT(tmp, rb_entry);
846 tmp = RB_LEFT(tmp, rb_entry);
854 * Inserts the given whole VM object into the target map at the specified
855 * address range. The object's size should match that of the address range.
857 * The map must be exclusively locked.
858 * The caller must have reserved sufficient vm_map_entry structures.
860 * If object is non-NULL, ref count must be bumped by caller
861 * prior to making call to account for the new entry.
864 vm_map_insert(vm_map_t map, int *countp,
865 vm_object_t object, vm_ooffset_t offset,
866 vm_offset_t start, vm_offset_t end,
867 vm_maptype_t maptype,
868 vm_prot_t prot, vm_prot_t max,
871 vm_map_entry_t new_entry;
872 vm_map_entry_t prev_entry;
873 vm_map_entry_t temp_entry;
874 vm_eflags_t protoeflags;
876 ASSERT_VM_MAP_LOCKED(map);
879 * Check that the start and end points are not bogus.
881 if ((start < map->min_offset) || (end > map->max_offset) ||
883 return (KERN_INVALID_ADDRESS);
886 * Find the entry prior to the proposed starting address; if it's part
887 * of an existing entry, this range is bogus.
889 if (vm_map_lookup_entry(map, start, &temp_entry))
890 return (KERN_NO_SPACE);
892 prev_entry = temp_entry;
895 * Assert that the next entry doesn't overlap the end point.
898 if ((prev_entry->next != &map->header) &&
899 (prev_entry->next->start < end))
900 return (KERN_NO_SPACE);
904 if (cow & MAP_COPY_ON_WRITE)
905 protoeflags |= MAP_ENTRY_COW|MAP_ENTRY_NEEDS_COPY;
907 if (cow & MAP_NOFAULT) {
908 protoeflags |= MAP_ENTRY_NOFAULT;
910 KASSERT(object == NULL,
911 ("vm_map_insert: paradoxical MAP_NOFAULT request"));
913 if (cow & MAP_DISABLE_SYNCER)
914 protoeflags |= MAP_ENTRY_NOSYNC;
915 if (cow & MAP_DISABLE_COREDUMP)
916 protoeflags |= MAP_ENTRY_NOCOREDUMP;
917 if (cow & MAP_IS_STACK)
918 protoeflags |= MAP_ENTRY_STACK;
919 if (cow & MAP_IS_KSTACK)
920 protoeflags |= MAP_ENTRY_KSTACK;
922 lwkt_gettoken(&vm_token);
923 lwkt_gettoken(&vmobj_token);
927 * When object is non-NULL, it could be shared with another
928 * process. We have to set or clear OBJ_ONEMAPPING
931 if ((object->ref_count > 1) || (object->shadow_count != 0)) {
932 vm_object_clear_flag(object, OBJ_ONEMAPPING);
935 else if ((prev_entry != &map->header) &&
936 (prev_entry->eflags == protoeflags) &&
937 (prev_entry->end == start) &&
938 (prev_entry->wired_count == 0) &&
939 prev_entry->maptype == maptype &&
940 ((prev_entry->object.vm_object == NULL) ||
941 vm_object_coalesce(prev_entry->object.vm_object,
942 OFF_TO_IDX(prev_entry->offset),
943 (vm_size_t)(prev_entry->end - prev_entry->start),
944 (vm_size_t)(end - prev_entry->end)))) {
946 * We were able to extend the object. Determine if we
947 * can extend the previous map entry to include the
950 if ((prev_entry->inheritance == VM_INHERIT_DEFAULT) &&
951 (prev_entry->protection == prot) &&
952 (prev_entry->max_protection == max)) {
953 lwkt_reltoken(&vmobj_token);
954 lwkt_reltoken(&vm_token);
955 map->size += (end - prev_entry->end);
956 prev_entry->end = end;
957 vm_map_simplify_entry(map, prev_entry, countp);
958 return (KERN_SUCCESS);
962 * If we can extend the object but cannot extend the
963 * map entry, we have to create a new map entry. We
964 * must bump the ref count on the extended object to
965 * account for it. object may be NULL.
967 object = prev_entry->object.vm_object;
968 offset = prev_entry->offset +
969 (prev_entry->end - prev_entry->start);
970 vm_object_reference_locked(object);
973 lwkt_reltoken(&vmobj_token);
974 lwkt_reltoken(&vm_token);
977 * NOTE: if conditionals fail, object can be NULL here. This occurs
978 * in things like the buffer map where we manage kva but do not manage
986 new_entry = vm_map_entry_create(map, countp);
987 new_entry->start = start;
988 new_entry->end = end;
990 new_entry->maptype = maptype;
991 new_entry->eflags = protoeflags;
992 new_entry->object.vm_object = object;
993 new_entry->offset = offset;
994 new_entry->aux.master_pde = 0;
996 new_entry->inheritance = VM_INHERIT_DEFAULT;
997 new_entry->protection = prot;
998 new_entry->max_protection = max;
999 new_entry->wired_count = 0;
1002 * Insert the new entry into the list
1005 vm_map_entry_link(map, prev_entry, new_entry);
1006 map->size += new_entry->end - new_entry->start;
1009 * Update the free space hint. Entries cannot overlap.
1010 * An exact comparison is needed to avoid matching
1011 * against the map->header.
1013 if ((map->first_free == prev_entry) &&
1014 (prev_entry->end == new_entry->start)) {
1015 map->first_free = new_entry;
1020 * Temporarily removed to avoid MAP_STACK panic, due to
1021 * MAP_STACK being a huge hack. Will be added back in
1022 * when MAP_STACK (and the user stack mapping) is fixed.
1025 * It may be possible to simplify the entry
1027 vm_map_simplify_entry(map, new_entry, countp);
1031 * Try to pre-populate the page table. Mappings governed by virtual
1032 * page tables cannot be prepopulated without a lot of work, so
1035 if ((cow & (MAP_PREFAULT|MAP_PREFAULT_PARTIAL)) &&
1036 maptype != VM_MAPTYPE_VPAGETABLE) {
1037 pmap_object_init_pt(map->pmap, start, prot,
1038 object, OFF_TO_IDX(offset), end - start,
1039 cow & MAP_PREFAULT_PARTIAL);
1042 return (KERN_SUCCESS);
1046 * Find sufficient space for `length' bytes in the given map, starting at
1047 * `start'. Returns 0 on success, 1 on no space.
1049 * This function will returned an arbitrarily aligned pointer. If no
1050 * particular alignment is required you should pass align as 1. Note that
1051 * the map may return PAGE_SIZE aligned pointers if all the lengths used in
1052 * the map are a multiple of PAGE_SIZE, even if you pass a smaller align
1055 * 'align' should be a power of 2 but is not required to be.
1057 * The map must be exclusively locked.
1058 * No other requirements.
1061 vm_map_findspace(vm_map_t map, vm_offset_t start, vm_size_t length,
1062 vm_size_t align, int flags, vm_offset_t *addr)
1064 vm_map_entry_t entry, next;
1066 vm_offset_t align_mask;
1068 if (start < map->min_offset)
1069 start = map->min_offset;
1070 if (start > map->max_offset)
1074 * If the alignment is not a power of 2 we will have to use
1075 * a mod/division, set align_mask to a special value.
1077 if ((align | (align - 1)) + 1 != (align << 1))
1078 align_mask = (vm_offset_t)-1;
1080 align_mask = align - 1;
1083 * Look for the first possible address; if there's already something
1084 * at this address, we have to start after it.
1086 if (start == map->min_offset) {
1087 if ((entry = map->first_free) != &map->header)
1092 if (vm_map_lookup_entry(map, start, &tmp))
1098 * Look through the rest of the map, trying to fit a new region in the
1099 * gap between existing regions, or after the very last region.
1101 for (;; start = (entry = next)->end) {
1103 * Adjust the proposed start by the requested alignment,
1104 * be sure that we didn't wrap the address.
1106 if (align_mask == (vm_offset_t)-1)
1107 end = ((start + align - 1) / align) * align;
1109 end = (start + align_mask) & ~align_mask;
1114 * Find the end of the proposed new region. Be sure we didn't
1115 * go beyond the end of the map, or wrap around the address.
1116 * Then check to see if this is the last entry or if the
1117 * proposed end fits in the gap between this and the next
1120 end = start + length;
1121 if (end > map->max_offset || end < start)
1126 * If the next entry's start address is beyond the desired
1127 * end address we may have found a good entry.
1129 * If the next entry is a stack mapping we do not map into
1130 * the stack's reserved space.
1132 * XXX continue to allow mapping into the stack's reserved
1133 * space if doing a MAP_STACK mapping inside a MAP_STACK
1134 * mapping, for backwards compatibility. But the caller
1135 * really should use MAP_STACK | MAP_TRYFIXED if they
1138 if (next == &map->header)
1140 if (next->start >= end) {
1141 if ((next->eflags & MAP_ENTRY_STACK) == 0)
1143 if (flags & MAP_STACK)
1145 if (next->start - next->aux.avail_ssize >= end)
1152 * Grow the kernel_map if necessary. pmap_growkernel() will panic
1153 * if it fails. The kernel_map is locked and nothing can steal
1154 * our address space if pmap_growkernel() blocks.
1156 * NOTE: This may be unconditionally called for kldload areas on
1157 * x86_64 because these do not bump kernel_vm_end (which would
1158 * fill 128G worth of page tables!). Therefore we must not
1161 if (map == &kernel_map) {
1164 kstop = round_page(start + length);
1165 if (kstop > kernel_vm_end)
1166 pmap_growkernel(start, kstop);
1173 * vm_map_find finds an unallocated region in the target address map with
1174 * the given length. The search is defined to be first-fit from the
1175 * specified address; the region found is returned in the same parameter.
1177 * If object is non-NULL, ref count must be bumped by caller
1178 * prior to making call to account for the new entry.
1180 * No requirements. This function will lock the map temporarily.
1183 vm_map_find(vm_map_t map, vm_object_t object, vm_ooffset_t offset,
1184 vm_offset_t *addr, vm_size_t length, vm_size_t align,
1186 vm_maptype_t maptype,
1187 vm_prot_t prot, vm_prot_t max,
1196 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
1199 if (vm_map_findspace(map, start, length, align, 0, addr)) {
1201 vm_map_entry_release(count);
1202 return (KERN_NO_SPACE);
1206 result = vm_map_insert(map, &count, object, offset,
1207 start, start + length,
1212 vm_map_entry_release(count);
1218 * Simplify the given map entry by merging with either neighbor. This
1219 * routine also has the ability to merge with both neighbors.
1221 * This routine guarentees that the passed entry remains valid (though
1222 * possibly extended). When merging, this routine may delete one or
1223 * both neighbors. No action is taken on entries which have their
1224 * in-transition flag set.
1226 * The map must be exclusively locked.
1229 vm_map_simplify_entry(vm_map_t map, vm_map_entry_t entry, int *countp)
1231 vm_map_entry_t next, prev;
1232 vm_size_t prevsize, esize;
1234 if (entry->eflags & MAP_ENTRY_IN_TRANSITION) {
1235 ++mycpu->gd_cnt.v_intrans_coll;
1239 if (entry->maptype == VM_MAPTYPE_SUBMAP)
1243 if (prev != &map->header) {
1244 prevsize = prev->end - prev->start;
1245 if ( (prev->end == entry->start) &&
1246 (prev->maptype == entry->maptype) &&
1247 (prev->object.vm_object == entry->object.vm_object) &&
1248 (!prev->object.vm_object ||
1249 (prev->offset + prevsize == entry->offset)) &&
1250 (prev->eflags == entry->eflags) &&
1251 (prev->protection == entry->protection) &&
1252 (prev->max_protection == entry->max_protection) &&
1253 (prev->inheritance == entry->inheritance) &&
1254 (prev->wired_count == entry->wired_count)) {
1255 if (map->first_free == prev)
1256 map->first_free = entry;
1257 if (map->hint == prev)
1259 vm_map_entry_unlink(map, prev);
1260 entry->start = prev->start;
1261 entry->offset = prev->offset;
1262 if (prev->object.vm_object)
1263 vm_object_deallocate(prev->object.vm_object);
1264 vm_map_entry_dispose(map, prev, countp);
1269 if (next != &map->header) {
1270 esize = entry->end - entry->start;
1271 if ((entry->end == next->start) &&
1272 (next->maptype == entry->maptype) &&
1273 (next->object.vm_object == entry->object.vm_object) &&
1274 (!entry->object.vm_object ||
1275 (entry->offset + esize == next->offset)) &&
1276 (next->eflags == entry->eflags) &&
1277 (next->protection == entry->protection) &&
1278 (next->max_protection == entry->max_protection) &&
1279 (next->inheritance == entry->inheritance) &&
1280 (next->wired_count == entry->wired_count)) {
1281 if (map->first_free == next)
1282 map->first_free = entry;
1283 if (map->hint == next)
1285 vm_map_entry_unlink(map, next);
1286 entry->end = next->end;
1287 if (next->object.vm_object)
1288 vm_object_deallocate(next->object.vm_object);
1289 vm_map_entry_dispose(map, next, countp);
1295 * Asserts that the given entry begins at or after the specified address.
1296 * If necessary, it splits the entry into two.
1298 #define vm_map_clip_start(map, entry, startaddr, countp) \
1300 if (startaddr > entry->start) \
1301 _vm_map_clip_start(map, entry, startaddr, countp); \
1305 * This routine is called only when it is known that the entry must be split.
1307 * The map must be exclusively locked.
1310 _vm_map_clip_start(vm_map_t map, vm_map_entry_t entry, vm_offset_t start,
1313 vm_map_entry_t new_entry;
1316 * Split off the front portion -- note that we must insert the new
1317 * entry BEFORE this one, so that this entry has the specified
1321 vm_map_simplify_entry(map, entry, countp);
1324 * If there is no object backing this entry, we might as well create
1325 * one now. If we defer it, an object can get created after the map
1326 * is clipped, and individual objects will be created for the split-up
1327 * map. This is a bit of a hack, but is also about the best place to
1328 * put this improvement.
1330 if (entry->object.vm_object == NULL && !map->system_map) {
1331 vm_map_entry_allocate_object(entry);
1334 new_entry = vm_map_entry_create(map, countp);
1335 *new_entry = *entry;
1337 new_entry->end = start;
1338 entry->offset += (start - entry->start);
1339 entry->start = start;
1341 vm_map_entry_link(map, entry->prev, new_entry);
1343 switch(entry->maptype) {
1344 case VM_MAPTYPE_NORMAL:
1345 case VM_MAPTYPE_VPAGETABLE:
1346 vm_object_reference(new_entry->object.vm_object);
1354 * Asserts that the given entry ends at or before the specified address.
1355 * If necessary, it splits the entry into two.
1357 * The map must be exclusively locked.
1359 #define vm_map_clip_end(map, entry, endaddr, countp) \
1361 if (endaddr < entry->end) \
1362 _vm_map_clip_end(map, entry, endaddr, countp); \
1366 * This routine is called only when it is known that the entry must be split.
1368 * The map must be exclusively locked.
1371 _vm_map_clip_end(vm_map_t map, vm_map_entry_t entry, vm_offset_t end,
1374 vm_map_entry_t new_entry;
1377 * If there is no object backing this entry, we might as well create
1378 * one now. If we defer it, an object can get created after the map
1379 * is clipped, and individual objects will be created for the split-up
1380 * map. This is a bit of a hack, but is also about the best place to
1381 * put this improvement.
1384 if (entry->object.vm_object == NULL && !map->system_map) {
1385 vm_map_entry_allocate_object(entry);
1389 * Create a new entry and insert it AFTER the specified entry
1392 new_entry = vm_map_entry_create(map, countp);
1393 *new_entry = *entry;
1395 new_entry->start = entry->end = end;
1396 new_entry->offset += (end - entry->start);
1398 vm_map_entry_link(map, entry, new_entry);
1400 switch(entry->maptype) {
1401 case VM_MAPTYPE_NORMAL:
1402 case VM_MAPTYPE_VPAGETABLE:
1403 vm_object_reference(new_entry->object.vm_object);
1411 * Asserts that the starting and ending region addresses fall within the
1412 * valid range for the map.
1414 #define VM_MAP_RANGE_CHECK(map, start, end) \
1416 if (start < vm_map_min(map)) \
1417 start = vm_map_min(map); \
1418 if (end > vm_map_max(map)) \
1419 end = vm_map_max(map); \
1425 * Used to block when an in-transition collison occurs. The map
1426 * is unlocked for the sleep and relocked before the return.
1429 vm_map_transition_wait(vm_map_t map)
1431 tsleep_interlock(map, 0);
1433 tsleep(map, PINTERLOCKED, "vment", 0);
1438 * When we do blocking operations with the map lock held it is
1439 * possible that a clip might have occured on our in-transit entry,
1440 * requiring an adjustment to the entry in our loop. These macros
1441 * help the pageable and clip_range code deal with the case. The
1442 * conditional costs virtually nothing if no clipping has occured.
1445 #define CLIP_CHECK_BACK(entry, save_start) \
1447 while (entry->start != save_start) { \
1448 entry = entry->prev; \
1449 KASSERT(entry != &map->header, ("bad entry clip")); \
1453 #define CLIP_CHECK_FWD(entry, save_end) \
1455 while (entry->end != save_end) { \
1456 entry = entry->next; \
1457 KASSERT(entry != &map->header, ("bad entry clip")); \
1463 * Clip the specified range and return the base entry. The
1464 * range may cover several entries starting at the returned base
1465 * and the first and last entry in the covering sequence will be
1466 * properly clipped to the requested start and end address.
1468 * If no holes are allowed you should pass the MAP_CLIP_NO_HOLES
1471 * The MAP_ENTRY_IN_TRANSITION flag will be set for the entries
1472 * covered by the requested range.
1474 * The map must be exclusively locked on entry and will remain locked
1475 * on return. If no range exists or the range contains holes and you
1476 * specified that no holes were allowed, NULL will be returned. This
1477 * routine may temporarily unlock the map in order avoid a deadlock when
1482 vm_map_clip_range(vm_map_t map, vm_offset_t start, vm_offset_t end,
1483 int *countp, int flags)
1485 vm_map_entry_t start_entry;
1486 vm_map_entry_t entry;
1489 * Locate the entry and effect initial clipping. The in-transition
1490 * case does not occur very often so do not try to optimize it.
1493 if (vm_map_lookup_entry(map, start, &start_entry) == FALSE)
1495 entry = start_entry;
1496 if (entry->eflags & MAP_ENTRY_IN_TRANSITION) {
1497 entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP;
1498 ++mycpu->gd_cnt.v_intrans_coll;
1499 ++mycpu->gd_cnt.v_intrans_wait;
1500 vm_map_transition_wait(map);
1502 * entry and/or start_entry may have been clipped while
1503 * we slept, or may have gone away entirely. We have
1504 * to restart from the lookup.
1510 * Since we hold an exclusive map lock we do not have to restart
1511 * after clipping, even though clipping may block in zalloc.
1513 vm_map_clip_start(map, entry, start, countp);
1514 vm_map_clip_end(map, entry, end, countp);
1515 entry->eflags |= MAP_ENTRY_IN_TRANSITION;
1518 * Scan entries covered by the range. When working on the next
1519 * entry a restart need only re-loop on the current entry which
1520 * we have already locked, since 'next' may have changed. Also,
1521 * even though entry is safe, it may have been clipped so we
1522 * have to iterate forwards through the clip after sleeping.
1524 while (entry->next != &map->header && entry->next->start < end) {
1525 vm_map_entry_t next = entry->next;
1527 if (flags & MAP_CLIP_NO_HOLES) {
1528 if (next->start > entry->end) {
1529 vm_map_unclip_range(map, start_entry,
1530 start, entry->end, countp, flags);
1535 if (next->eflags & MAP_ENTRY_IN_TRANSITION) {
1536 vm_offset_t save_end = entry->end;
1537 next->eflags |= MAP_ENTRY_NEEDS_WAKEUP;
1538 ++mycpu->gd_cnt.v_intrans_coll;
1539 ++mycpu->gd_cnt.v_intrans_wait;
1540 vm_map_transition_wait(map);
1543 * clips might have occured while we blocked.
1545 CLIP_CHECK_FWD(entry, save_end);
1546 CLIP_CHECK_BACK(start_entry, start);
1550 * No restart necessary even though clip_end may block, we
1551 * are holding the map lock.
1553 vm_map_clip_end(map, next, end, countp);
1554 next->eflags |= MAP_ENTRY_IN_TRANSITION;
1557 if (flags & MAP_CLIP_NO_HOLES) {
1558 if (entry->end != end) {
1559 vm_map_unclip_range(map, start_entry,
1560 start, entry->end, countp, flags);
1564 return(start_entry);
1568 * Undo the effect of vm_map_clip_range(). You should pass the same
1569 * flags and the same range that you passed to vm_map_clip_range().
1570 * This code will clear the in-transition flag on the entries and
1571 * wake up anyone waiting. This code will also simplify the sequence
1572 * and attempt to merge it with entries before and after the sequence.
1574 * The map must be locked on entry and will remain locked on return.
1576 * Note that you should also pass the start_entry returned by
1577 * vm_map_clip_range(). However, if you block between the two calls
1578 * with the map unlocked please be aware that the start_entry may
1579 * have been clipped and you may need to scan it backwards to find
1580 * the entry corresponding with the original start address. You are
1581 * responsible for this, vm_map_unclip_range() expects the correct
1582 * start_entry to be passed to it and will KASSERT otherwise.
1586 vm_map_unclip_range(vm_map_t map, vm_map_entry_t start_entry,
1587 vm_offset_t start, vm_offset_t end,
1588 int *countp, int flags)
1590 vm_map_entry_t entry;
1592 entry = start_entry;
1594 KASSERT(entry->start == start, ("unclip_range: illegal base entry"));
1595 while (entry != &map->header && entry->start < end) {
1596 KASSERT(entry->eflags & MAP_ENTRY_IN_TRANSITION,
1597 ("in-transition flag not set during unclip on: %p",
1599 KASSERT(entry->end <= end,
1600 ("unclip_range: tail wasn't clipped"));
1601 entry->eflags &= ~MAP_ENTRY_IN_TRANSITION;
1602 if (entry->eflags & MAP_ENTRY_NEEDS_WAKEUP) {
1603 entry->eflags &= ~MAP_ENTRY_NEEDS_WAKEUP;
1606 entry = entry->next;
1610 * Simplification does not block so there is no restart case.
1612 entry = start_entry;
1613 while (entry != &map->header && entry->start < end) {
1614 vm_map_simplify_entry(map, entry, countp);
1615 entry = entry->next;
1620 * Mark the given range as handled by a subordinate map.
1622 * This range must have been created with vm_map_find(), and no other
1623 * operations may have been performed on this range prior to calling
1626 * Submappings cannot be removed.
1631 vm_map_submap(vm_map_t map, vm_offset_t start, vm_offset_t end, vm_map_t submap)
1633 vm_map_entry_t entry;
1634 int result = KERN_INVALID_ARGUMENT;
1637 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
1640 VM_MAP_RANGE_CHECK(map, start, end);
1642 if (vm_map_lookup_entry(map, start, &entry)) {
1643 vm_map_clip_start(map, entry, start, &count);
1645 entry = entry->next;
1648 vm_map_clip_end(map, entry, end, &count);
1650 if ((entry->start == start) && (entry->end == end) &&
1651 ((entry->eflags & MAP_ENTRY_COW) == 0) &&
1652 (entry->object.vm_object == NULL)) {
1653 entry->object.sub_map = submap;
1654 entry->maptype = VM_MAPTYPE_SUBMAP;
1655 result = KERN_SUCCESS;
1658 vm_map_entry_release(count);
1664 * Sets the protection of the specified address region in the target map.
1665 * If "set_max" is specified, the maximum protection is to be set;
1666 * otherwise, only the current protection is affected.
1668 * The protection is not applicable to submaps, but is applicable to normal
1669 * maps and maps governed by virtual page tables. For example, when operating
1670 * on a virtual page table our protection basically controls how COW occurs
1671 * on the backing object, whereas the virtual page table abstraction itself
1672 * is an abstraction for userland.
1677 vm_map_protect(vm_map_t map, vm_offset_t start, vm_offset_t end,
1678 vm_prot_t new_prot, boolean_t set_max)
1680 vm_map_entry_t current;
1681 vm_map_entry_t entry;
1684 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
1687 VM_MAP_RANGE_CHECK(map, start, end);
1689 if (vm_map_lookup_entry(map, start, &entry)) {
1690 vm_map_clip_start(map, entry, start, &count);
1692 entry = entry->next;
1696 * Make a first pass to check for protection violations.
1699 while ((current != &map->header) && (current->start < end)) {
1700 if (current->maptype == VM_MAPTYPE_SUBMAP) {
1702 vm_map_entry_release(count);
1703 return (KERN_INVALID_ARGUMENT);
1705 if ((new_prot & current->max_protection) != new_prot) {
1707 vm_map_entry_release(count);
1708 return (KERN_PROTECTION_FAILURE);
1710 current = current->next;
1714 * Go back and fix up protections. [Note that clipping is not
1715 * necessary the second time.]
1719 while ((current != &map->header) && (current->start < end)) {
1722 vm_map_clip_end(map, current, end, &count);
1724 old_prot = current->protection;
1726 current->protection =
1727 (current->max_protection = new_prot) &
1730 current->protection = new_prot;
1734 * Update physical map if necessary. Worry about copy-on-write
1735 * here -- CHECK THIS XXX
1738 if (current->protection != old_prot) {
1739 #define MASK(entry) (((entry)->eflags & MAP_ENTRY_COW) ? ~VM_PROT_WRITE : \
1742 pmap_protect(map->pmap, current->start,
1744 current->protection & MASK(current));
1748 vm_map_simplify_entry(map, current, &count);
1750 current = current->next;
1754 vm_map_entry_release(count);
1755 return (KERN_SUCCESS);
1759 * This routine traverses a processes map handling the madvise
1760 * system call. Advisories are classified as either those effecting
1761 * the vm_map_entry structure, or those effecting the underlying
1764 * The <value> argument is used for extended madvise calls.
1769 vm_map_madvise(vm_map_t map, vm_offset_t start, vm_offset_t end,
1770 int behav, off_t value)
1772 vm_map_entry_t current, entry;
1778 * Some madvise calls directly modify the vm_map_entry, in which case
1779 * we need to use an exclusive lock on the map and we need to perform
1780 * various clipping operations. Otherwise we only need a read-lock
1784 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
1788 case MADV_SEQUENTIAL:
1802 vm_map_lock_read(map);
1805 vm_map_entry_release(count);
1810 * Locate starting entry and clip if necessary.
1813 VM_MAP_RANGE_CHECK(map, start, end);
1815 if (vm_map_lookup_entry(map, start, &entry)) {
1817 vm_map_clip_start(map, entry, start, &count);
1819 entry = entry->next;
1824 * madvise behaviors that are implemented in the vm_map_entry.
1826 * We clip the vm_map_entry so that behavioral changes are
1827 * limited to the specified address range.
1829 for (current = entry;
1830 (current != &map->header) && (current->start < end);
1831 current = current->next
1833 if (current->maptype == VM_MAPTYPE_SUBMAP)
1836 vm_map_clip_end(map, current, end, &count);
1840 vm_map_entry_set_behavior(current, MAP_ENTRY_BEHAV_NORMAL);
1842 case MADV_SEQUENTIAL:
1843 vm_map_entry_set_behavior(current, MAP_ENTRY_BEHAV_SEQUENTIAL);
1846 vm_map_entry_set_behavior(current, MAP_ENTRY_BEHAV_RANDOM);
1849 current->eflags |= MAP_ENTRY_NOSYNC;
1852 current->eflags &= ~MAP_ENTRY_NOSYNC;
1855 current->eflags |= MAP_ENTRY_NOCOREDUMP;
1858 current->eflags &= ~MAP_ENTRY_NOCOREDUMP;
1862 * Invalidate the related pmap entries, used
1863 * to flush portions of the real kernel's
1864 * pmap when the caller has removed or
1865 * modified existing mappings in a virtual
1868 pmap_remove(map->pmap,
1869 current->start, current->end);
1873 * Set the page directory page for a map
1874 * governed by a virtual page table. Mark
1875 * the entry as being governed by a virtual
1876 * page table if it is not.
1878 * XXX the page directory page is stored
1879 * in the avail_ssize field if the map_entry.
1881 * XXX the map simplification code does not
1882 * compare this field so weird things may
1883 * happen if you do not apply this function
1884 * to the entire mapping governed by the
1885 * virtual page table.
1887 if (current->maptype != VM_MAPTYPE_VPAGETABLE) {
1891 current->aux.master_pde = value;
1892 pmap_remove(map->pmap,
1893 current->start, current->end);
1899 vm_map_simplify_entry(map, current, &count);
1907 * madvise behaviors that are implemented in the underlying
1910 * Since we don't clip the vm_map_entry, we have to clip
1911 * the vm_object pindex and count.
1913 * NOTE! We currently do not support these functions on
1914 * virtual page tables.
1916 for (current = entry;
1917 (current != &map->header) && (current->start < end);
1918 current = current->next
1920 vm_offset_t useStart;
1922 if (current->maptype != VM_MAPTYPE_NORMAL)
1925 pindex = OFF_TO_IDX(current->offset);
1926 count = atop(current->end - current->start);
1927 useStart = current->start;
1929 if (current->start < start) {
1930 pindex += atop(start - current->start);
1931 count -= atop(start - current->start);
1934 if (current->end > end)
1935 count -= atop(current->end - end);
1940 vm_object_madvise(current->object.vm_object,
1941 pindex, count, behav);
1944 * Try to populate the page table. Mappings governed
1945 * by virtual page tables cannot be pre-populated
1946 * without a lot of work so don't try.
1948 if (behav == MADV_WILLNEED &&
1949 current->maptype != VM_MAPTYPE_VPAGETABLE) {
1950 pmap_object_init_pt(
1953 current->protection,
1954 current->object.vm_object,
1956 (count << PAGE_SHIFT),
1957 MAP_PREFAULT_MADVISE
1961 vm_map_unlock_read(map);
1963 vm_map_entry_release(count);
1969 * Sets the inheritance of the specified address range in the target map.
1970 * Inheritance affects how the map will be shared with child maps at the
1971 * time of vm_map_fork.
1974 vm_map_inherit(vm_map_t map, vm_offset_t start, vm_offset_t end,
1975 vm_inherit_t new_inheritance)
1977 vm_map_entry_t entry;
1978 vm_map_entry_t temp_entry;
1981 switch (new_inheritance) {
1982 case VM_INHERIT_NONE:
1983 case VM_INHERIT_COPY:
1984 case VM_INHERIT_SHARE:
1987 return (KERN_INVALID_ARGUMENT);
1990 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
1993 VM_MAP_RANGE_CHECK(map, start, end);
1995 if (vm_map_lookup_entry(map, start, &temp_entry)) {
1997 vm_map_clip_start(map, entry, start, &count);
1999 entry = temp_entry->next;
2001 while ((entry != &map->header) && (entry->start < end)) {
2002 vm_map_clip_end(map, entry, end, &count);
2004 entry->inheritance = new_inheritance;
2006 vm_map_simplify_entry(map, entry, &count);
2008 entry = entry->next;
2011 vm_map_entry_release(count);
2012 return (KERN_SUCCESS);
2016 * Implement the semantics of mlock
2019 vm_map_unwire(vm_map_t map, vm_offset_t start, vm_offset_t real_end,
2020 boolean_t new_pageable)
2022 vm_map_entry_t entry;
2023 vm_map_entry_t start_entry;
2025 int rv = KERN_SUCCESS;
2028 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
2030 VM_MAP_RANGE_CHECK(map, start, real_end);
2033 start_entry = vm_map_clip_range(map, start, end, &count,
2035 if (start_entry == NULL) {
2037 vm_map_entry_release(count);
2038 return (KERN_INVALID_ADDRESS);
2041 if (new_pageable == 0) {
2042 entry = start_entry;
2043 while ((entry != &map->header) && (entry->start < end)) {
2044 vm_offset_t save_start;
2045 vm_offset_t save_end;
2048 * Already user wired or hard wired (trivial cases)
2050 if (entry->eflags & MAP_ENTRY_USER_WIRED) {
2051 entry = entry->next;
2054 if (entry->wired_count != 0) {
2055 entry->wired_count++;
2056 entry->eflags |= MAP_ENTRY_USER_WIRED;
2057 entry = entry->next;
2062 * A new wiring requires instantiation of appropriate
2063 * management structures and the faulting in of the
2066 if (entry->maptype != VM_MAPTYPE_SUBMAP) {
2067 int copyflag = entry->eflags &
2068 MAP_ENTRY_NEEDS_COPY;
2069 if (copyflag && ((entry->protection &
2070 VM_PROT_WRITE) != 0)) {
2071 vm_map_entry_shadow(entry);
2072 } else if (entry->object.vm_object == NULL &&
2074 vm_map_entry_allocate_object(entry);
2077 entry->wired_count++;
2078 entry->eflags |= MAP_ENTRY_USER_WIRED;
2081 * Now fault in the area. Note that vm_fault_wire()
2082 * may release the map lock temporarily, it will be
2083 * relocked on return. The in-transition
2084 * flag protects the entries.
2086 save_start = entry->start;
2087 save_end = entry->end;
2088 rv = vm_fault_wire(map, entry, TRUE);
2090 CLIP_CHECK_BACK(entry, save_start);
2092 KASSERT(entry->wired_count == 1, ("bad wired_count on entry"));
2093 entry->eflags &= ~MAP_ENTRY_USER_WIRED;
2094 entry->wired_count = 0;
2095 if (entry->end == save_end)
2097 entry = entry->next;
2098 KASSERT(entry != &map->header, ("bad entry clip during backout"));
2100 end = save_start; /* unwire the rest */
2104 * note that even though the entry might have been
2105 * clipped, the USER_WIRED flag we set prevents
2106 * duplication so we do not have to do a
2109 entry = entry->next;
2113 * If we failed fall through to the unwiring section to
2114 * unwire what we had wired so far. 'end' has already
2121 * start_entry might have been clipped if we unlocked the
2122 * map and blocked. No matter how clipped it has gotten
2123 * there should be a fragment that is on our start boundary.
2125 CLIP_CHECK_BACK(start_entry, start);
2129 * Deal with the unwiring case.
2133 * This is the unwiring case. We must first ensure that the
2134 * range to be unwired is really wired down. We know there
2137 entry = start_entry;
2138 while ((entry != &map->header) && (entry->start < end)) {
2139 if ((entry->eflags & MAP_ENTRY_USER_WIRED) == 0) {
2140 rv = KERN_INVALID_ARGUMENT;
2143 KASSERT(entry->wired_count != 0, ("wired count was 0 with USER_WIRED set! %p", entry));
2144 entry = entry->next;
2148 * Now decrement the wiring count for each region. If a region
2149 * becomes completely unwired, unwire its physical pages and
2153 * The map entries are processed in a loop, checking to
2154 * make sure the entry is wired and asserting it has a wired
2155 * count. However, another loop was inserted more-or-less in
2156 * the middle of the unwiring path. This loop picks up the
2157 * "entry" loop variable from the first loop without first
2158 * setting it to start_entry. Naturally, the secound loop
2159 * is never entered and the pages backing the entries are
2160 * never unwired. This can lead to a leak of wired pages.
2162 entry = start_entry;
2163 while ((entry != &map->header) && (entry->start < end)) {
2164 KASSERT(entry->eflags & MAP_ENTRY_USER_WIRED,
2165 ("expected USER_WIRED on entry %p", entry));
2166 entry->eflags &= ~MAP_ENTRY_USER_WIRED;
2167 entry->wired_count--;
2168 if (entry->wired_count == 0)
2169 vm_fault_unwire(map, entry);
2170 entry = entry->next;
2174 vm_map_unclip_range(map, start_entry, start, real_end, &count,
2178 vm_map_entry_release(count);
2183 * Sets the pageability of the specified address range in the target map.
2184 * Regions specified as not pageable require locked-down physical
2185 * memory and physical page maps.
2187 * The map must not be locked, but a reference must remain to the map
2188 * throughout the call.
2190 * This function may be called via the zalloc path and must properly
2191 * reserve map entries for kernel_map.
2196 vm_map_wire(vm_map_t map, vm_offset_t start, vm_offset_t real_end, int kmflags)
2198 vm_map_entry_t entry;
2199 vm_map_entry_t start_entry;
2201 int rv = KERN_SUCCESS;
2204 if (kmflags & KM_KRESERVE)
2205 count = vm_map_entry_kreserve(MAP_RESERVE_COUNT);
2207 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
2209 VM_MAP_RANGE_CHECK(map, start, real_end);
2212 start_entry = vm_map_clip_range(map, start, end, &count,
2214 if (start_entry == NULL) {
2216 rv = KERN_INVALID_ADDRESS;
2219 if ((kmflags & KM_PAGEABLE) == 0) {
2223 * 1. Holding the write lock, we create any shadow or zero-fill
2224 * objects that need to be created. Then we clip each map
2225 * entry to the region to be wired and increment its wiring
2226 * count. We create objects before clipping the map entries
2227 * to avoid object proliferation.
2229 * 2. We downgrade to a read lock, and call vm_fault_wire to
2230 * fault in the pages for any newly wired area (wired_count is
2233 * Downgrading to a read lock for vm_fault_wire avoids a
2234 * possible deadlock with another process that may have faulted
2235 * on one of the pages to be wired (it would mark the page busy,
2236 * blocking us, then in turn block on the map lock that we
2237 * hold). Because of problems in the recursive lock package,
2238 * we cannot upgrade to a write lock in vm_map_lookup. Thus,
2239 * any actions that require the write lock must be done
2240 * beforehand. Because we keep the read lock on the map, the
2241 * copy-on-write status of the entries we modify here cannot
2244 entry = start_entry;
2245 while ((entry != &map->header) && (entry->start < end)) {
2247 * Trivial case if the entry is already wired
2249 if (entry->wired_count) {
2250 entry->wired_count++;
2251 entry = entry->next;
2256 * The entry is being newly wired, we have to setup
2257 * appropriate management structures. A shadow
2258 * object is required for a copy-on-write region,
2259 * or a normal object for a zero-fill region. We
2260 * do not have to do this for entries that point to sub
2261 * maps because we won't hold the lock on the sub map.
2263 if (entry->maptype != VM_MAPTYPE_SUBMAP) {
2264 int copyflag = entry->eflags &
2265 MAP_ENTRY_NEEDS_COPY;
2266 if (copyflag && ((entry->protection &
2267 VM_PROT_WRITE) != 0)) {
2268 vm_map_entry_shadow(entry);
2269 } else if (entry->object.vm_object == NULL &&
2271 vm_map_entry_allocate_object(entry);
2275 entry->wired_count++;
2276 entry = entry->next;
2284 * HACK HACK HACK HACK
2286 * vm_fault_wire() temporarily unlocks the map to avoid
2287 * deadlocks. The in-transition flag from vm_map_clip_range
2288 * call should protect us from changes while the map is
2291 * NOTE: Previously this comment stated that clipping might
2292 * still occur while the entry is unlocked, but from
2293 * what I can tell it actually cannot.
2295 * It is unclear whether the CLIP_CHECK_*() calls
2296 * are still needed but we keep them in anyway.
2298 * HACK HACK HACK HACK
2301 entry = start_entry;
2302 while (entry != &map->header && entry->start < end) {
2304 * If vm_fault_wire fails for any page we need to undo
2305 * what has been done. We decrement the wiring count
2306 * for those pages which have not yet been wired (now)
2307 * and unwire those that have (later).
2309 vm_offset_t save_start = entry->start;
2310 vm_offset_t save_end = entry->end;
2312 if (entry->wired_count == 1)
2313 rv = vm_fault_wire(map, entry, FALSE);
2315 CLIP_CHECK_BACK(entry, save_start);
2317 KASSERT(entry->wired_count == 1, ("wired_count changed unexpectedly"));
2318 entry->wired_count = 0;
2319 if (entry->end == save_end)
2321 entry = entry->next;
2322 KASSERT(entry != &map->header, ("bad entry clip during backout"));
2327 CLIP_CHECK_FWD(entry, save_end);
2328 entry = entry->next;
2332 * If a failure occured undo everything by falling through
2333 * to the unwiring code. 'end' has already been adjusted
2337 kmflags |= KM_PAGEABLE;
2340 * start_entry is still IN_TRANSITION but may have been
2341 * clipped since vm_fault_wire() unlocks and relocks the
2342 * map. No matter how clipped it has gotten there should
2343 * be a fragment that is on our start boundary.
2345 CLIP_CHECK_BACK(start_entry, start);
2348 if (kmflags & KM_PAGEABLE) {
2350 * This is the unwiring case. We must first ensure that the
2351 * range to be unwired is really wired down. We know there
2354 entry = start_entry;
2355 while ((entry != &map->header) && (entry->start < end)) {
2356 if (entry->wired_count == 0) {
2357 rv = KERN_INVALID_ARGUMENT;
2360 entry = entry->next;
2364 * Now decrement the wiring count for each region. If a region
2365 * becomes completely unwired, unwire its physical pages and
2368 entry = start_entry;
2369 while ((entry != &map->header) && (entry->start < end)) {
2370 entry->wired_count--;
2371 if (entry->wired_count == 0)
2372 vm_fault_unwire(map, entry);
2373 entry = entry->next;
2377 vm_map_unclip_range(map, start_entry, start, real_end,
2378 &count, MAP_CLIP_NO_HOLES);
2382 if (kmflags & KM_KRESERVE)
2383 vm_map_entry_krelease(count);
2385 vm_map_entry_release(count);
2390 * Mark a newly allocated address range as wired but do not fault in
2391 * the pages. The caller is expected to load the pages into the object.
2393 * The map must be locked on entry and will remain locked on return.
2394 * No other requirements.
2397 vm_map_set_wired_quick(vm_map_t map, vm_offset_t addr, vm_size_t size,
2400 vm_map_entry_t scan;
2401 vm_map_entry_t entry;
2403 entry = vm_map_clip_range(map, addr, addr + size,
2404 countp, MAP_CLIP_NO_HOLES);
2406 scan != &map->header && scan->start < addr + size;
2407 scan = scan->next) {
2408 KKASSERT(entry->wired_count == 0);
2409 entry->wired_count = 1;
2411 vm_map_unclip_range(map, entry, addr, addr + size,
2412 countp, MAP_CLIP_NO_HOLES);
2416 * Push any dirty cached pages in the address range to their pager.
2417 * If syncio is TRUE, dirty pages are written synchronously.
2418 * If invalidate is TRUE, any cached pages are freed as well.
2420 * This routine is called by sys_msync()
2422 * Returns an error if any part of the specified range is not mapped.
2427 vm_map_clean(vm_map_t map, vm_offset_t start, vm_offset_t end,
2428 boolean_t syncio, boolean_t invalidate)
2430 vm_map_entry_t current;
2431 vm_map_entry_t entry;
2434 vm_ooffset_t offset;
2436 vm_map_lock_read(map);
2437 VM_MAP_RANGE_CHECK(map, start, end);
2438 if (!vm_map_lookup_entry(map, start, &entry)) {
2439 vm_map_unlock_read(map);
2440 return (KERN_INVALID_ADDRESS);
2443 * Make a first pass to check for holes.
2445 for (current = entry; current->start < end; current = current->next) {
2446 if (current->maptype == VM_MAPTYPE_SUBMAP) {
2447 vm_map_unlock_read(map);
2448 return (KERN_INVALID_ARGUMENT);
2450 if (end > current->end &&
2451 (current->next == &map->header ||
2452 current->end != current->next->start)) {
2453 vm_map_unlock_read(map);
2454 return (KERN_INVALID_ADDRESS);
2459 pmap_remove(vm_map_pmap(map), start, end);
2462 * Make a second pass, cleaning/uncaching pages from the indicated
2465 * Hold vm_token to avoid blocking in vm_object_reference()
2467 lwkt_gettoken(&vm_token);
2468 lwkt_gettoken(&vmobj_token);
2470 for (current = entry; current->start < end; current = current->next) {
2471 offset = current->offset + (start - current->start);
2472 size = (end <= current->end ? end : current->end) - start;
2473 if (current->maptype == VM_MAPTYPE_SUBMAP) {
2475 vm_map_entry_t tentry;
2478 smap = current->object.sub_map;
2479 vm_map_lock_read(smap);
2480 vm_map_lookup_entry(smap, offset, &tentry);
2481 tsize = tentry->end - offset;
2484 object = tentry->object.vm_object;
2485 offset = tentry->offset + (offset - tentry->start);
2486 vm_map_unlock_read(smap);
2488 object = current->object.vm_object;
2491 * Note that there is absolutely no sense in writing out
2492 * anonymous objects, so we track down the vnode object
2494 * We invalidate (remove) all pages from the address space
2495 * anyway, for semantic correctness.
2497 * note: certain anonymous maps, such as MAP_NOSYNC maps,
2498 * may start out with a NULL object.
2500 while (object && object->backing_object) {
2501 offset += object->backing_object_offset;
2502 object = object->backing_object;
2503 if (object->size < OFF_TO_IDX( offset + size))
2504 size = IDX_TO_OFF(object->size) - offset;
2506 if (object && (object->type == OBJT_VNODE) &&
2507 (current->protection & VM_PROT_WRITE) &&
2508 (object->flags & OBJ_NOMSYNC) == 0) {
2510 * Flush pages if writing is allowed, invalidate them
2511 * if invalidation requested. Pages undergoing I/O
2512 * will be ignored by vm_object_page_remove().
2514 * We cannot lock the vnode and then wait for paging
2515 * to complete without deadlocking against vm_fault.
2516 * Instead we simply call vm_object_page_remove() and
2517 * allow it to block internally on a page-by-page
2518 * basis when it encounters pages undergoing async
2523 vm_object_reference_locked(object);
2524 vn_lock(object->handle, LK_EXCLUSIVE | LK_RETRY);
2525 flags = (syncio || invalidate) ? OBJPC_SYNC : 0;
2526 flags |= invalidate ? OBJPC_INVAL : 0;
2529 * When operating on a virtual page table just
2530 * flush the whole object. XXX we probably ought
2533 switch(current->maptype) {
2534 case VM_MAPTYPE_NORMAL:
2535 vm_object_page_clean(object,
2537 OFF_TO_IDX(offset + size + PAGE_MASK),
2540 case VM_MAPTYPE_VPAGETABLE:
2541 vm_object_page_clean(object, 0, 0, flags);
2544 vn_unlock(((struct vnode *)object->handle));
2545 vm_object_deallocate_locked(object);
2547 if (object && invalidate &&
2548 ((object->type == OBJT_VNODE) ||
2549 (object->type == OBJT_DEVICE))) {
2551 (object->type == OBJT_DEVICE) ? FALSE : TRUE;
2552 vm_object_reference_locked(object);
2553 switch(current->maptype) {
2554 case VM_MAPTYPE_NORMAL:
2555 vm_object_page_remove(object,
2557 OFF_TO_IDX(offset + size + PAGE_MASK),
2560 case VM_MAPTYPE_VPAGETABLE:
2561 vm_object_page_remove(object, 0, 0, clean_only);
2564 vm_object_deallocate_locked(object);
2569 lwkt_reltoken(&vmobj_token);
2570 lwkt_reltoken(&vm_token);
2571 vm_map_unlock_read(map);
2573 return (KERN_SUCCESS);
2577 * Make the region specified by this entry pageable.
2579 * The vm_map must be exclusively locked.
2582 vm_map_entry_unwire(vm_map_t map, vm_map_entry_t entry)
2584 entry->eflags &= ~MAP_ENTRY_USER_WIRED;
2585 entry->wired_count = 0;
2586 vm_fault_unwire(map, entry);
2590 * Deallocate the given entry from the target map.
2592 * The vm_map must be exclusively locked.
2595 vm_map_entry_delete(vm_map_t map, vm_map_entry_t entry, int *countp)
2597 vm_map_entry_unlink(map, entry);
2598 map->size -= entry->end - entry->start;
2600 switch(entry->maptype) {
2601 case VM_MAPTYPE_NORMAL:
2602 case VM_MAPTYPE_VPAGETABLE:
2603 vm_object_deallocate(entry->object.vm_object);
2609 vm_map_entry_dispose(map, entry, countp);
2613 * Deallocates the given address range from the target map.
2615 * The vm_map must be exclusively locked.
2618 vm_map_delete(vm_map_t map, vm_offset_t start, vm_offset_t end, int *countp)
2621 vm_map_entry_t entry;
2622 vm_map_entry_t first_entry;
2624 ASSERT_VM_MAP_LOCKED(map);
2627 * Find the start of the region, and clip it. Set entry to point
2628 * at the first record containing the requested address or, if no
2629 * such record exists, the next record with a greater address. The
2630 * loop will run from this point until a record beyond the termination
2631 * address is encountered.
2633 * map->hint must be adjusted to not point to anything we delete,
2634 * so set it to the entry prior to the one being deleted.
2636 * GGG see other GGG comment.
2638 if (vm_map_lookup_entry(map, start, &first_entry)) {
2639 entry = first_entry;
2640 vm_map_clip_start(map, entry, start, countp);
2641 map->hint = entry->prev; /* possible problem XXX */
2643 map->hint = first_entry; /* possible problem XXX */
2644 entry = first_entry->next;
2648 * If a hole opens up prior to the current first_free then
2649 * adjust first_free. As with map->hint, map->first_free
2650 * cannot be left set to anything we might delete.
2652 if (entry == &map->header) {
2653 map->first_free = &map->header;
2654 } else if (map->first_free->start >= start) {
2655 map->first_free = entry->prev;
2659 * Step through all entries in this region
2661 while ((entry != &map->header) && (entry->start < end)) {
2662 vm_map_entry_t next;
2664 vm_pindex_t offidxstart, offidxend, count;
2667 * If we hit an in-transition entry we have to sleep and
2668 * retry. It's easier (and not really slower) to just retry
2669 * since this case occurs so rarely and the hint is already
2670 * pointing at the right place. We have to reset the
2671 * start offset so as not to accidently delete an entry
2672 * another process just created in vacated space.
2674 if (entry->eflags & MAP_ENTRY_IN_TRANSITION) {
2675 entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP;
2676 start = entry->start;
2677 ++mycpu->gd_cnt.v_intrans_coll;
2678 ++mycpu->gd_cnt.v_intrans_wait;
2679 vm_map_transition_wait(map);
2682 vm_map_clip_end(map, entry, end, countp);
2688 offidxstart = OFF_TO_IDX(entry->offset);
2689 count = OFF_TO_IDX(e - s);
2690 object = entry->object.vm_object;
2693 * Unwire before removing addresses from the pmap; otherwise,
2694 * unwiring will put the entries back in the pmap.
2696 if (entry->wired_count != 0)
2697 vm_map_entry_unwire(map, entry);
2699 offidxend = offidxstart + count;
2702 * Hold vm_token when manipulating vm_objects,
2704 * Hold vmobj_token when potentially adding or removing
2705 * objects (collapse requires both).
2707 lwkt_gettoken(&vm_token);
2708 lwkt_gettoken(&vmobj_token);
2710 if (object == &kernel_object) {
2711 vm_object_page_remove(object, offidxstart,
2714 pmap_remove(map->pmap, s, e);
2716 if (object != NULL &&
2717 object->ref_count != 1 &&
2718 (object->flags & (OBJ_NOSPLIT|OBJ_ONEMAPPING)) ==
2720 (object->type == OBJT_DEFAULT ||
2721 object->type == OBJT_SWAP)) {
2722 vm_object_collapse(object);
2723 vm_object_page_remove(object, offidxstart,
2725 if (object->type == OBJT_SWAP) {
2726 swap_pager_freespace(object,
2730 if (offidxend >= object->size &&
2731 offidxstart < object->size) {
2732 object->size = offidxstart;
2736 lwkt_reltoken(&vmobj_token);
2737 lwkt_reltoken(&vm_token);
2740 * Delete the entry (which may delete the object) only after
2741 * removing all pmap entries pointing to its pages.
2742 * (Otherwise, its page frames may be reallocated, and any
2743 * modify bits will be set in the wrong object!)
2745 vm_map_entry_delete(map, entry, countp);
2748 return (KERN_SUCCESS);
2752 * Remove the given address range from the target map.
2753 * This is the exported form of vm_map_delete.
2758 vm_map_remove(vm_map_t map, vm_offset_t start, vm_offset_t end)
2763 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
2765 VM_MAP_RANGE_CHECK(map, start, end);
2766 result = vm_map_delete(map, start, end, &count);
2768 vm_map_entry_release(count);
2774 * Assert that the target map allows the specified privilege on the
2775 * entire address region given. The entire region must be allocated.
2777 * The caller must specify whether the vm_map is already locked or not.
2780 vm_map_check_protection(vm_map_t map, vm_offset_t start, vm_offset_t end,
2781 vm_prot_t protection, boolean_t have_lock)
2783 vm_map_entry_t entry;
2784 vm_map_entry_t tmp_entry;
2787 if (have_lock == FALSE)
2788 vm_map_lock_read(map);
2790 if (!vm_map_lookup_entry(map, start, &tmp_entry)) {
2791 if (have_lock == FALSE)
2792 vm_map_unlock_read(map);
2798 while (start < end) {
2799 if (entry == &map->header) {
2807 if (start < entry->start) {
2812 * Check protection associated with entry.
2815 if ((entry->protection & protection) != protection) {
2819 /* go to next entry */
2822 entry = entry->next;
2824 if (have_lock == FALSE)
2825 vm_map_unlock_read(map);
2830 * Split the pages in a map entry into a new object. This affords
2831 * easier removal of unused pages, and keeps object inheritance from
2832 * being a negative impact on memory usage.
2834 * The vm_map must be exclusively locked.
2837 vm_map_split(vm_map_entry_t entry)
2840 vm_object_t orig_object, new_object, source;
2842 vm_pindex_t offidxstart, offidxend, idx;
2844 vm_ooffset_t offset;
2846 orig_object = entry->object.vm_object;
2847 if (orig_object->type != OBJT_DEFAULT && orig_object->type != OBJT_SWAP)
2849 if (orig_object->ref_count <= 1)
2852 offset = entry->offset;
2856 offidxstart = OFF_TO_IDX(offset);
2857 offidxend = offidxstart + OFF_TO_IDX(e - s);
2858 size = offidxend - offidxstart;
2860 switch(orig_object->type) {
2862 new_object = default_pager_alloc(NULL, IDX_TO_OFF(size),
2866 new_object = swap_pager_alloc(NULL, IDX_TO_OFF(size),
2874 if (new_object == NULL)
2878 * vm_token required when manipulating vm_objects.
2880 lwkt_gettoken(&vm_token);
2881 lwkt_gettoken(&vmobj_token);
2883 source = orig_object->backing_object;
2884 if (source != NULL) {
2885 /* Referenced by new_object */
2886 vm_object_reference_locked(source);
2887 LIST_INSERT_HEAD(&source->shadow_head,
2888 new_object, shadow_list);
2889 vm_object_clear_flag(source, OBJ_ONEMAPPING);
2890 new_object->backing_object_offset =
2891 orig_object->backing_object_offset +
2892 IDX_TO_OFF(offidxstart);
2893 new_object->backing_object = source;
2894 source->shadow_count++;
2895 source->generation++;
2898 for (idx = 0; idx < size; idx++) {
2902 m = vm_page_lookup(orig_object, offidxstart + idx);
2907 * We must wait for pending I/O to complete before we can
2910 * We do not have to VM_PROT_NONE the page as mappings should
2911 * not be changed by this operation.
2913 if (vm_page_sleep_busy(m, TRUE, "spltwt"))
2916 vm_page_rename(m, new_object, idx);
2917 /* page automatically made dirty by rename and cache handled */
2921 if (orig_object->type == OBJT_SWAP) {
2922 vm_object_pip_add(orig_object, 1);
2924 * copy orig_object pages into new_object
2925 * and destroy unneeded pages in
2928 swap_pager_copy(orig_object, new_object, offidxstart, 0);
2929 vm_object_pip_wakeup(orig_object);
2933 * Wakeup the pages we played with. No spl protection is needed
2934 * for a simple wakeup.
2936 for (idx = 0; idx < size; idx++) {
2937 m = vm_page_lookup(new_object, idx);
2942 entry->object.vm_object = new_object;
2943 entry->offset = 0LL;
2944 vm_object_deallocate_locked(orig_object);
2945 lwkt_reltoken(&vmobj_token);
2946 lwkt_reltoken(&vm_token);
2950 * Copies the contents of the source entry to the destination
2951 * entry. The entries *must* be aligned properly.
2953 * The vm_map must be exclusively locked.
2954 * vm_token must be held
2957 vm_map_copy_entry(vm_map_t src_map, vm_map_t dst_map,
2958 vm_map_entry_t src_entry, vm_map_entry_t dst_entry)
2960 vm_object_t src_object;
2962 if (dst_entry->maptype == VM_MAPTYPE_SUBMAP)
2964 if (src_entry->maptype == VM_MAPTYPE_SUBMAP)
2967 ASSERT_LWKT_TOKEN_HELD(&vm_token);
2968 lwkt_gettoken(&vmobj_token); /* required for collapse */
2970 if (src_entry->wired_count == 0) {
2972 * If the source entry is marked needs_copy, it is already
2975 if ((src_entry->eflags & MAP_ENTRY_NEEDS_COPY) == 0) {
2976 pmap_protect(src_map->pmap,
2979 src_entry->protection & ~VM_PROT_WRITE);
2983 * Make a copy of the object.
2985 if ((src_object = src_entry->object.vm_object) != NULL) {
2986 if ((src_object->handle == NULL) &&
2987 (src_object->type == OBJT_DEFAULT ||
2988 src_object->type == OBJT_SWAP)) {
2989 vm_object_collapse(src_object);
2990 if ((src_object->flags & (OBJ_NOSPLIT|OBJ_ONEMAPPING)) == OBJ_ONEMAPPING) {
2991 vm_map_split(src_entry);
2992 src_object = src_entry->object.vm_object;
2996 vm_object_reference_locked(src_object);
2997 vm_object_clear_flag(src_object, OBJ_ONEMAPPING);
2998 dst_entry->object.vm_object = src_object;
2999 src_entry->eflags |= (MAP_ENTRY_COW|MAP_ENTRY_NEEDS_COPY);
3000 dst_entry->eflags |= (MAP_ENTRY_COW|MAP_ENTRY_NEEDS_COPY);
3001 dst_entry->offset = src_entry->offset;
3003 dst_entry->object.vm_object = NULL;
3004 dst_entry->offset = 0;
3007 pmap_copy(dst_map->pmap, src_map->pmap, dst_entry->start,
3008 dst_entry->end - dst_entry->start, src_entry->start);
3011 * Of course, wired down pages can't be set copy-on-write.
3012 * Cause wired pages to be copied into the new map by
3013 * simulating faults (the new pages are pageable)
3015 vm_fault_copy_entry(dst_map, src_map, dst_entry, src_entry);
3017 lwkt_reltoken(&vmobj_token);
3022 * Create a new process vmspace structure and vm_map
3023 * based on those of an existing process. The new map
3024 * is based on the old map, according to the inheritance
3025 * values on the regions in that map.
3027 * The source map must not be locked.
3031 vmspace_fork(struct vmspace *vm1)
3033 struct vmspace *vm2;
3034 vm_map_t old_map = &vm1->vm_map;
3036 vm_map_entry_t old_entry;
3037 vm_map_entry_t new_entry;
3041 lwkt_gettoken(&vm_token);
3042 lwkt_gettoken(&vmspace_token);
3043 lwkt_gettoken(&vmobj_token);
3044 vm_map_lock(old_map);
3045 old_map->infork = 1;
3048 * XXX Note: upcalls are not copied.
3050 vm2 = vmspace_alloc(old_map->min_offset, old_map->max_offset);
3051 bcopy(&vm1->vm_startcopy, &vm2->vm_startcopy,
3052 (caddr_t)&vm1->vm_endcopy - (caddr_t)&vm1->vm_startcopy);
3053 new_map = &vm2->vm_map; /* XXX */
3054 new_map->timestamp = 1;
3056 vm_map_lock(new_map);
3059 old_entry = old_map->header.next;
3060 while (old_entry != &old_map->header) {
3062 old_entry = old_entry->next;
3065 count = vm_map_entry_reserve(count + MAP_RESERVE_COUNT);
3067 old_entry = old_map->header.next;
3068 while (old_entry != &old_map->header) {
3069 if (old_entry->maptype == VM_MAPTYPE_SUBMAP)
3070 panic("vm_map_fork: encountered a submap");
3072 switch (old_entry->inheritance) {
3073 case VM_INHERIT_NONE:
3075 case VM_INHERIT_SHARE:
3077 * Clone the entry, creating the shared object if
3080 object = old_entry->object.vm_object;
3081 if (object == NULL) {
3082 vm_map_entry_allocate_object(old_entry);
3083 object = old_entry->object.vm_object;
3087 * Add the reference before calling vm_map_entry_shadow
3088 * to insure that a shadow object is created.
3090 vm_object_reference_locked(object);
3091 if (old_entry->eflags & MAP_ENTRY_NEEDS_COPY) {
3092 vm_map_entry_shadow(old_entry);
3093 /* Transfer the second reference too. */
3094 vm_object_reference_locked(
3095 old_entry->object.vm_object);
3096 vm_object_deallocate_locked(object);
3097 object = old_entry->object.vm_object;
3099 vm_object_clear_flag(object, OBJ_ONEMAPPING);
3102 * Clone the entry, referencing the shared object.
3104 new_entry = vm_map_entry_create(new_map, &count);
3105 *new_entry = *old_entry;
3106 new_entry->eflags &= ~MAP_ENTRY_USER_WIRED;
3107 new_entry->wired_count = 0;
3110 * Insert the entry into the new map -- we know we're
3111 * inserting at the end of the new map.
3114 vm_map_entry_link(new_map, new_map->header.prev,
3118 * Update the physical map
3120 pmap_copy(new_map->pmap, old_map->pmap,
3122 (old_entry->end - old_entry->start),
3125 case VM_INHERIT_COPY:
3127 * Clone the entry and link into the map.
3129 new_entry = vm_map_entry_create(new_map, &count);
3130 *new_entry = *old_entry;
3131 new_entry->eflags &= ~MAP_ENTRY_USER_WIRED;
3132 new_entry->wired_count = 0;
3133 new_entry->object.vm_object = NULL;
3134 vm_map_entry_link(new_map, new_map->header.prev,
3136 vm_map_copy_entry(old_map, new_map, old_entry,
3140 old_entry = old_entry->next;
3143 new_map->size = old_map->size;
3144 old_map->infork = 0;
3145 vm_map_unlock(old_map);
3146 vm_map_unlock(new_map);
3147 vm_map_entry_release(count);
3149 lwkt_reltoken(&vmobj_token);
3150 lwkt_reltoken(&vmspace_token);
3151 lwkt_reltoken(&vm_token);
3157 * Create an auto-grow stack entry
3162 vm_map_stack (vm_map_t map, vm_offset_t addrbos, vm_size_t max_ssize,
3163 int flags, vm_prot_t prot, vm_prot_t max, int cow)
3165 vm_map_entry_t prev_entry;
3166 vm_map_entry_t new_stack_entry;
3167 vm_size_t init_ssize;
3170 vm_offset_t tmpaddr;
3172 cow |= MAP_IS_STACK;
3174 if (max_ssize < sgrowsiz)
3175 init_ssize = max_ssize;
3177 init_ssize = sgrowsiz;
3179 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
3183 * Find space for the mapping
3185 if ((flags & (MAP_FIXED | MAP_TRYFIXED)) == 0) {
3186 if (vm_map_findspace(map, addrbos, max_ssize, 1,
3189 vm_map_entry_release(count);
3190 return (KERN_NO_SPACE);
3195 /* If addr is already mapped, no go */
3196 if (vm_map_lookup_entry(map, addrbos, &prev_entry)) {
3198 vm_map_entry_release(count);
3199 return (KERN_NO_SPACE);
3203 /* XXX already handled by kern_mmap() */
3204 /* If we would blow our VMEM resource limit, no go */
3205 if (map->size + init_ssize >
3206 curproc->p_rlimit[RLIMIT_VMEM].rlim_cur) {
3208 vm_map_entry_release(count);
3209 return (KERN_NO_SPACE);
3214 * If we can't accomodate max_ssize in the current mapping,
3215 * no go. However, we need to be aware that subsequent user
3216 * mappings might map into the space we have reserved for
3217 * stack, and currently this space is not protected.
3219 * Hopefully we will at least detect this condition
3220 * when we try to grow the stack.
3222 if ((prev_entry->next != &map->header) &&
3223 (prev_entry->next->start < addrbos + max_ssize)) {
3225 vm_map_entry_release(count);
3226 return (KERN_NO_SPACE);
3230 * We initially map a stack of only init_ssize. We will
3231 * grow as needed later. Since this is to be a grow
3232 * down stack, we map at the top of the range.
3234 * Note: we would normally expect prot and max to be
3235 * VM_PROT_ALL, and cow to be 0. Possibly we should
3236 * eliminate these as input parameters, and just
3237 * pass these values here in the insert call.
3239 rv = vm_map_insert(map, &count,
3240 NULL, 0, addrbos + max_ssize - init_ssize,
3241 addrbos + max_ssize,
3246 /* Now set the avail_ssize amount */
3247 if (rv == KERN_SUCCESS) {
3248 if (prev_entry != &map->header)
3249 vm_map_clip_end(map, prev_entry, addrbos + max_ssize - init_ssize, &count);
3250 new_stack_entry = prev_entry->next;
3251 if (new_stack_entry->end != addrbos + max_ssize ||
3252 new_stack_entry->start != addrbos + max_ssize - init_ssize)
3253 panic ("Bad entry start/end for new stack entry");
3255 new_stack_entry->aux.avail_ssize = max_ssize - init_ssize;
3259 vm_map_entry_release(count);
3264 * Attempts to grow a vm stack entry. Returns KERN_SUCCESS if the
3265 * desired address is already mapped, or if we successfully grow
3266 * the stack. Also returns KERN_SUCCESS if addr is outside the
3267 * stack range (this is strange, but preserves compatibility with
3268 * the grow function in vm_machdep.c).
3273 vm_map_growstack (struct proc *p, vm_offset_t addr)
3275 vm_map_entry_t prev_entry;
3276 vm_map_entry_t stack_entry;
3277 vm_map_entry_t new_stack_entry;
3278 struct vmspace *vm = p->p_vmspace;
3279 vm_map_t map = &vm->vm_map;
3282 int rv = KERN_SUCCESS;
3284 int use_read_lock = 1;
3287 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
3290 vm_map_lock_read(map);
3294 /* If addr is already in the entry range, no need to grow.*/
3295 if (vm_map_lookup_entry(map, addr, &prev_entry))
3298 if ((stack_entry = prev_entry->next) == &map->header)
3300 if (prev_entry == &map->header)
3301 end = stack_entry->start - stack_entry->aux.avail_ssize;
3303 end = prev_entry->end;
3306 * This next test mimics the old grow function in vm_machdep.c.
3307 * It really doesn't quite make sense, but we do it anyway
3308 * for compatibility.
3310 * If not growable stack, return success. This signals the
3311 * caller to proceed as he would normally with normal vm.
3313 if (stack_entry->aux.avail_ssize < 1 ||
3314 addr >= stack_entry->start ||
3315 addr < stack_entry->start - stack_entry->aux.avail_ssize) {
3319 /* Find the minimum grow amount */
3320 grow_amount = roundup (stack_entry->start - addr, PAGE_SIZE);
3321 if (grow_amount > stack_entry->aux.avail_ssize) {
3327 * If there is no longer enough space between the entries
3328 * nogo, and adjust the available space. Note: this
3329 * should only happen if the user has mapped into the
3330 * stack area after the stack was created, and is
3331 * probably an error.
3333 * This also effectively destroys any guard page the user
3334 * might have intended by limiting the stack size.
3336 if (grow_amount > stack_entry->start - end) {
3337 if (use_read_lock && vm_map_lock_upgrade(map)) {
3342 stack_entry->aux.avail_ssize = stack_entry->start - end;
3347 is_procstack = addr >= (vm_offset_t)vm->vm_maxsaddr;
3349 /* If this is the main process stack, see if we're over the
3352 if (is_procstack && (ctob(vm->vm_ssize) + grow_amount >
3353 p->p_rlimit[RLIMIT_STACK].rlim_cur)) {
3358 /* Round up the grow amount modulo SGROWSIZ */
3359 grow_amount = roundup (grow_amount, sgrowsiz);
3360 if (grow_amount > stack_entry->aux.avail_ssize) {
3361 grow_amount = stack_entry->aux.avail_ssize;
3363 if (is_procstack && (ctob(vm->vm_ssize) + grow_amount >
3364 p->p_rlimit[RLIMIT_STACK].rlim_cur)) {
3365 grow_amount = p->p_rlimit[RLIMIT_STACK].rlim_cur -
3369 /* If we would blow our VMEM resource limit, no go */
3370 if (map->size + grow_amount > p->p_rlimit[RLIMIT_VMEM].rlim_cur) {
3375 if (use_read_lock && vm_map_lock_upgrade(map)) {
3381 /* Get the preliminary new entry start value */
3382 addr = stack_entry->start - grow_amount;
3384 /* If this puts us into the previous entry, cut back our growth
3385 * to the available space. Also, see the note above.
3388 stack_entry->aux.avail_ssize = stack_entry->start - end;
3392 rv = vm_map_insert(map, &count,
3393 NULL, 0, addr, stack_entry->start,
3395 VM_PROT_ALL, VM_PROT_ALL,
3398 /* Adjust the available stack space by the amount we grew. */
3399 if (rv == KERN_SUCCESS) {
3400 if (prev_entry != &map->header)
3401 vm_map_clip_end(map, prev_entry, addr, &count);
3402 new_stack_entry = prev_entry->next;
3403 if (new_stack_entry->end != stack_entry->start ||
3404 new_stack_entry->start != addr)
3405 panic ("Bad stack grow start/end in new stack entry");
3407 new_stack_entry->aux.avail_ssize =
3408 stack_entry->aux.avail_ssize -
3409 (new_stack_entry->end - new_stack_entry->start);
3411 vm->vm_ssize += btoc(new_stack_entry->end -
3412 new_stack_entry->start);
3415 if (map->flags & MAP_WIREFUTURE)
3416 vm_map_unwire(map, new_stack_entry->start,
3417 new_stack_entry->end, FALSE);
3422 vm_map_unlock_read(map);
3425 vm_map_entry_release(count);
3430 * Unshare the specified VM space for exec. If other processes are
3431 * mapped to it, then create a new one. The new vmspace is null.
3436 vmspace_exec(struct proc *p, struct vmspace *vmcopy)
3438 struct vmspace *oldvmspace = p->p_vmspace;
3439 struct vmspace *newvmspace;
3440 vm_map_t map = &p->p_vmspace->vm_map;
3443 * If we are execing a resident vmspace we fork it, otherwise
3444 * we create a new vmspace. Note that exitingcnt and upcalls
3445 * are not copied to the new vmspace.
3447 lwkt_gettoken(&vmspace_token);
3449 newvmspace = vmspace_fork(vmcopy);
3451 newvmspace = vmspace_alloc(map->min_offset, map->max_offset);
3452 bcopy(&oldvmspace->vm_startcopy, &newvmspace->vm_startcopy,
3453 (caddr_t)&oldvmspace->vm_endcopy -
3454 (caddr_t)&oldvmspace->vm_startcopy);
3458 * Finish initializing the vmspace before assigning it
3459 * to the process. The vmspace will become the current vmspace
3462 pmap_pinit2(vmspace_pmap(newvmspace));
3463 pmap_replacevm(p, newvmspace, 0);
3464 sysref_put(&oldvmspace->vm_sysref);
3465 lwkt_reltoken(&vmspace_token);
3469 * Unshare the specified VM space for forcing COW. This
3470 * is called by rfork, for the (RFMEM|RFPROC) == 0 case.
3472 * The exitingcnt test is not strictly necessary but has been
3473 * included for code sanity (to make the code a bit more deterministic).
3476 vmspace_unshare(struct proc *p)
3478 struct vmspace *oldvmspace = p->p_vmspace;
3479 struct vmspace *newvmspace;
3481 lwkt_gettoken(&vmspace_token);
3482 if (oldvmspace->vm_sysref.refcnt == 1 && oldvmspace->vm_exitingcnt == 0)
3484 newvmspace = vmspace_fork(oldvmspace);
3485 pmap_pinit2(vmspace_pmap(newvmspace));
3486 pmap_replacevm(p, newvmspace, 0);
3487 sysref_put(&oldvmspace->vm_sysref);
3488 lwkt_reltoken(&vmspace_token);
3492 * vm_map_hint: return the beginning of the best area suitable for
3493 * creating a new mapping with "prot" protection.
3498 vm_map_hint(struct proc *p, vm_offset_t addr, vm_prot_t prot)
3500 struct vmspace *vms = p->p_vmspace;
3502 if (!randomize_mmap) {
3504 * Set a reasonable start point for the hint if it was
3505 * not specified or if it falls within the heap space.
3506 * Hinted mmap()s do not allocate out of the heap space.
3509 (addr >= round_page((vm_offset_t)vms->vm_taddr) &&
3510 addr < round_page((vm_offset_t)vms->vm_daddr + maxdsiz))) {
3511 addr = round_page((vm_offset_t)vms->vm_daddr + maxdsiz);
3517 if (addr != 0 && addr >= (vm_offset_t)vms->vm_daddr)
3523 * If executable skip first two pages, otherwise start
3524 * after data + heap region.
3526 if ((prot & VM_PROT_EXECUTE) &&
3527 ((vm_offset_t)vms->vm_daddr >= I386_MAX_EXE_ADDR)) {
3528 addr = (PAGE_SIZE * 2) +
3529 (karc4random() & (I386_MAX_EXE_ADDR / 2 - 1));
3530 return (round_page(addr));
3532 #endif /* __i386__ */
3535 addr = (vm_offset_t)vms->vm_daddr + MAXDSIZ;
3536 addr += karc4random() & (MIN((256 * 1024 * 1024), MAXDSIZ) - 1);
3538 return (round_page(addr));
3542 * Finds the VM object, offset, and protection for a given virtual address
3543 * in the specified map, assuming a page fault of the type specified.
3545 * Leaves the map in question locked for read; return values are guaranteed
3546 * until a vm_map_lookup_done call is performed. Note that the map argument
3547 * is in/out; the returned map must be used in the call to vm_map_lookup_done.
3549 * A handle (out_entry) is returned for use in vm_map_lookup_done, to make
3552 * If a lookup is requested with "write protection" specified, the map may
3553 * be changed to perform virtual copying operations, although the data
3554 * referenced will remain the same.
3559 vm_map_lookup(vm_map_t *var_map, /* IN/OUT */
3561 vm_prot_t fault_typea,
3562 vm_map_entry_t *out_entry, /* OUT */
3563 vm_object_t *object, /* OUT */
3564 vm_pindex_t *pindex, /* OUT */
3565 vm_prot_t *out_prot, /* OUT */
3566 boolean_t *wired) /* OUT */
3568 vm_map_entry_t entry;
3569 vm_map_t map = *var_map;
3571 vm_prot_t fault_type = fault_typea;
3572 int use_read_lock = 1;
3573 int rv = KERN_SUCCESS;
3577 vm_map_lock_read(map);
3582 * If the map has an interesting hint, try it before calling full
3583 * blown lookup routine.
3588 if ((entry == &map->header) ||
3589 (vaddr < entry->start) || (vaddr >= entry->end)) {
3590 vm_map_entry_t tmp_entry;
3593 * Entry was either not a valid hint, or the vaddr was not
3594 * contained in the entry, so do a full lookup.
3596 if (!vm_map_lookup_entry(map, vaddr, &tmp_entry)) {
3597 rv = KERN_INVALID_ADDRESS;
3608 if (entry->maptype == VM_MAPTYPE_SUBMAP) {
3609 vm_map_t old_map = map;
3611 *var_map = map = entry->object.sub_map;
3613 vm_map_unlock_read(old_map);
3615 vm_map_unlock(old_map);
3621 * Check whether this task is allowed to have this page.
3622 * Note the special case for MAP_ENTRY_COW
3623 * pages with an override. This is to implement a forced
3624 * COW for debuggers.
3627 if (fault_type & VM_PROT_OVERRIDE_WRITE)
3628 prot = entry->max_protection;
3630 prot = entry->protection;
3632 fault_type &= (VM_PROT_READ|VM_PROT_WRITE|VM_PROT_EXECUTE);
3633 if ((fault_type & prot) != fault_type) {
3634 rv = KERN_PROTECTION_FAILURE;
3638 if ((entry->eflags & MAP_ENTRY_USER_WIRED) &&
3639 (entry->eflags & MAP_ENTRY_COW) &&
3640 (fault_type & VM_PROT_WRITE) &&
3641 (fault_typea & VM_PROT_OVERRIDE_WRITE) == 0) {
3642 rv = KERN_PROTECTION_FAILURE;
3647 * If this page is not pageable, we have to get it for all possible
3650 *wired = (entry->wired_count != 0);
3652 prot = fault_type = entry->protection;
3655 * Virtual page tables may need to update the accessed (A) bit
3656 * in a page table entry. Upgrade the fault to a write fault for
3657 * that case if the map will support it. If the map does not support
3658 * it the page table entry simply will not be updated.
3660 if (entry->maptype == VM_MAPTYPE_VPAGETABLE) {
3661 if (prot & VM_PROT_WRITE)
3662 fault_type |= VM_PROT_WRITE;
3666 * If the entry was copy-on-write, we either ...
3668 if (entry->eflags & MAP_ENTRY_NEEDS_COPY) {
3670 * If we want to write the page, we may as well handle that
3671 * now since we've got the map locked.
3673 * If we don't need to write the page, we just demote the
3674 * permissions allowed.
3677 if (fault_type & VM_PROT_WRITE) {
3679 * Make a new object, and place it in the object
3680 * chain. Note that no new references have appeared
3681 * -- one just moved from the map to the new
3685 if (use_read_lock && vm_map_lock_upgrade(map)) {
3691 vm_map_entry_shadow(entry);
3694 * We're attempting to read a copy-on-write page --
3695 * don't allow writes.
3698 prot &= ~VM_PROT_WRITE;
3703 * Create an object if necessary.
3705 if (entry->object.vm_object == NULL &&
3707 if (use_read_lock && vm_map_lock_upgrade(map)) {
3712 vm_map_entry_allocate_object(entry);
3716 * Return the object/offset from this entry. If the entry was
3717 * copy-on-write or empty, it has been fixed up.
3720 *pindex = OFF_TO_IDX((vaddr - entry->start) + entry->offset);
3721 *object = entry->object.vm_object;
3724 * Return whether this is the only map sharing this data. On
3725 * success we return with a read lock held on the map. On failure
3726 * we return with the map unlocked.
3730 if (rv == KERN_SUCCESS) {
3731 if (use_read_lock == 0)
3732 vm_map_lock_downgrade(map);
3733 } else if (use_read_lock) {
3734 vm_map_unlock_read(map);
3742 * Releases locks acquired by a vm_map_lookup()
3743 * (according to the handle returned by that lookup).
3745 * No other requirements.
3748 vm_map_lookup_done(vm_map_t map, vm_map_entry_t entry, int count)
3751 * Unlock the main-level map
3753 vm_map_unlock_read(map);
3755 vm_map_entry_release(count);
3758 #include "opt_ddb.h"
3760 #include <sys/kernel.h>
3762 #include <ddb/ddb.h>
3767 DB_SHOW_COMMAND(map, vm_map_print)
3770 /* XXX convert args. */
3771 vm_map_t map = (vm_map_t)addr;
3772 boolean_t full = have_addr;
3774 vm_map_entry_t entry;
3776 db_iprintf("Task map %p: pmap=%p, nentries=%d, version=%u\n",
3778 (void *)map->pmap, map->nentries, map->timestamp);
3781 if (!full && db_indent)
3785 for (entry = map->header.next; entry != &map->header;
3786 entry = entry->next) {
3787 db_iprintf("map entry %p: start=%p, end=%p\n",
3788 (void *)entry, (void *)entry->start, (void *)entry->end);
3791 static char *inheritance_name[4] =
3792 {"share", "copy", "none", "donate_copy"};
3794 db_iprintf(" prot=%x/%x/%s",
3796 entry->max_protection,
3797 inheritance_name[(int)(unsigned char)entry->inheritance]);
3798 if (entry->wired_count != 0)
3799 db_printf(", wired");
3801 if (entry->maptype == VM_MAPTYPE_SUBMAP) {
3802 /* XXX no %qd in kernel. Truncate entry->offset. */
3803 db_printf(", share=%p, offset=0x%lx\n",
3804 (void *)entry->object.sub_map,
3805 (long)entry->offset);
3807 if ((entry->prev == &map->header) ||
3808 (entry->prev->object.sub_map !=
3809 entry->object.sub_map)) {
3811 vm_map_print((db_expr_t)(intptr_t)
3812 entry->object.sub_map,
3817 /* XXX no %qd in kernel. Truncate entry->offset. */
3818 db_printf(", object=%p, offset=0x%lx",
3819 (void *)entry->object.vm_object,
3820 (long)entry->offset);
3821 if (entry->eflags & MAP_ENTRY_COW)
3822 db_printf(", copy (%s)",
3823 (entry->eflags & MAP_ENTRY_NEEDS_COPY) ? "needed" : "done");
3827 if ((entry->prev == &map->header) ||
3828 (entry->prev->object.vm_object !=
3829 entry->object.vm_object)) {
3831 vm_object_print((db_expr_t)(intptr_t)
3832 entry->object.vm_object,
3847 DB_SHOW_COMMAND(procvm, procvm)
3852 p = (struct proc *) addr;
3857 db_printf("p = %p, vmspace = %p, map = %p, pmap = %p\n",
3858 (void *)p, (void *)p->p_vmspace, (void *)&p->p_vmspace->vm_map,
3859 (void *)vmspace_pmap(p->p_vmspace));
3861 vm_map_print((db_expr_t)(intptr_t)&p->p_vmspace->vm_map, 1, 0, NULL);