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
146 * per-cpu page table cross mappings are initialized in early boot
147 * and might require a considerable number of vm_map_entry structures.
149 #define VMEPERCPU (MAXCPU+1)
151 static struct vm_zone mapentzone_store, mapzone_store;
152 static vm_zone_t mapentzone, mapzone;
153 static struct vm_object mapentobj, mapobj;
155 static struct vm_map_entry map_entry_init[MAX_MAPENT];
156 static struct vm_map_entry cpu_map_entry_init[MAXCPU][VMEPERCPU];
157 static struct vm_map map_init[MAX_KMAP];
159 static int randomize_mmap;
160 SYSCTL_INT(_vm, OID_AUTO, randomize_mmap, CTLFLAG_RW, &randomize_mmap, 0,
161 "Randomize mmap offsets");
163 static void vm_map_entry_shadow(vm_map_entry_t entry, int addref);
164 static vm_map_entry_t vm_map_entry_create(vm_map_t map, int *);
165 static void vm_map_entry_dispose (vm_map_t map, vm_map_entry_t entry, int *);
166 static void _vm_map_clip_end (vm_map_t, vm_map_entry_t, vm_offset_t, int *);
167 static void _vm_map_clip_start (vm_map_t, vm_map_entry_t, vm_offset_t, int *);
168 static void vm_map_entry_delete (vm_map_t, vm_map_entry_t, int *);
169 static void vm_map_entry_unwire (vm_map_t, vm_map_entry_t);
170 static void vm_map_copy_entry (vm_map_t, vm_map_t, vm_map_entry_t,
172 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);
175 * Initialize the vm_map module. Must be called before any other vm_map
178 * Map and entry structures are allocated from the general purpose
179 * memory pool with some exceptions:
181 * - The kernel map is allocated statically.
182 * - Initial kernel map entries are allocated out of a static pool.
184 * These restrictions are necessary since malloc() uses the
185 * maps and requires map entries.
187 * Called from the low level boot code only.
192 mapzone = &mapzone_store;
193 zbootinit(mapzone, "MAP", sizeof (struct vm_map),
195 mapentzone = &mapentzone_store;
196 zbootinit(mapentzone, "MAP ENTRY", sizeof (struct vm_map_entry),
197 map_entry_init, MAX_MAPENT);
201 * Called prior to any vmspace allocations.
203 * Called from the low level boot code only.
208 zinitna(mapentzone, &mapentobj, NULL, 0, 0,
209 ZONE_USE_RESERVE | ZONE_SPECIAL, 1);
210 zinitna(mapzone, &mapobj, NULL, 0, 0, 0, 1);
217 * Red black tree functions
219 * The caller must hold the related map lock.
221 static int rb_vm_map_compare(vm_map_entry_t a, vm_map_entry_t b);
222 RB_GENERATE(vm_map_rb_tree, vm_map_entry, rb_entry, rb_vm_map_compare);
224 /* a->start is address, and the only field has to be initialized */
226 rb_vm_map_compare(vm_map_entry_t a, vm_map_entry_t b)
228 if (a->start < b->start)
230 else if (a->start > b->start)
236 * Allocate a vmspace structure, including a vm_map and pmap.
237 * Initialize numerous fields. While the initial allocation is zerod,
238 * subsequence reuse from the objcache leaves elements of the structure
239 * intact (particularly the pmap), so portions must be zerod.
241 * The structure is not considered activated until we call sysref_activate().
246 vmspace_alloc(vm_offset_t min, vm_offset_t max)
250 vm = sysref_alloc(&vmspace_sysref_class);
251 bzero(&vm->vm_startcopy,
252 (char *)&vm->vm_endcopy - (char *)&vm->vm_startcopy);
253 vm_map_init(&vm->vm_map, min, max, NULL);
254 pmap_pinit(vmspace_pmap(vm)); /* (some fields reused) */
255 lwkt_gettoken(&vm->vm_map.token);
256 vm->vm_map.pmap = vmspace_pmap(vm); /* XXX */
258 vm->vm_exitingcnt = 0;
259 cpu_vmspace_alloc(vm);
260 sysref_activate(&vm->vm_sysref);
261 lwkt_reltoken(&vm->vm_map.token);
267 * dtor function - Some elements of the pmap are retained in the
268 * free-cached vmspaces to improve performance. We have to clean them up
269 * here before returning the vmspace to the memory pool.
274 vmspace_dtor(void *obj, void *private)
276 struct vmspace *vm = obj;
278 pmap_puninit(vmspace_pmap(vm));
282 * Called in two cases:
284 * (1) When the last sysref is dropped, but exitingcnt might still be
287 * (2) When there are no sysrefs (i.e. refcnt is negative) left and the
288 * exitingcnt becomes zero
290 * sysref will not scrap the object until we call sysref_put() once more
291 * after the last ref has been dropped.
293 * Interlocked by the sysref API.
296 vmspace_terminate(struct vmspace *vm)
301 * If exitingcnt is non-zero we can't get rid of the entire vmspace
302 * yet, but we can scrap user memory.
304 lwkt_gettoken(&vm->vm_map.token);
305 if (vm->vm_exitingcnt) {
307 pmap_remove_pages(vmspace_pmap(vm), VM_MIN_USER_ADDRESS,
308 VM_MAX_USER_ADDRESS);
309 vm_map_remove(&vm->vm_map, VM_MIN_USER_ADDRESS,
310 VM_MAX_USER_ADDRESS);
311 lwkt_reltoken(&vm->vm_map.token);
314 cpu_vmspace_free(vm);
317 * Make sure any SysV shm is freed, it might not have in
322 KKASSERT(vm->vm_upcalls == NULL);
325 * Lock the map, to wait out all other references to it.
326 * Delete all of the mappings and pages they hold, then call
327 * the pmap module to reclaim anything left.
329 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
330 vm_map_lock(&vm->vm_map);
331 vm_map_delete(&vm->vm_map, vm->vm_map.min_offset,
332 vm->vm_map.max_offset, &count);
333 vm_map_unlock(&vm->vm_map);
334 vm_map_entry_release(count);
336 lwkt_gettoken(&vmspace_pmap(vm)->pm_token);
337 pmap_release(vmspace_pmap(vm));
338 lwkt_reltoken(&vmspace_pmap(vm)->pm_token);
339 lwkt_reltoken(&vm->vm_map.token);
340 sysref_put(&vm->vm_sysref);
344 * vmspaces are not currently locked.
347 vmspace_lock(struct vmspace *vm __unused)
352 vmspace_unlock(struct vmspace *vm __unused)
357 * This is called during exit indicating that the vmspace is no
358 * longer in used by an exiting process, but the process has not yet
364 vmspace_exitbump(struct vmspace *vm)
366 lwkt_gettoken(&vm->vm_map.token);
368 lwkt_reltoken(&vm->vm_map.token);
372 * This is called in the wait*() handling code. The vmspace can be terminated
373 * after the last wait is finished using it.
378 vmspace_exitfree(struct proc *p)
383 lwkt_gettoken(&vm->vm_map.token);
386 if (--vm->vm_exitingcnt == 0 && sysref_isinactive(&vm->vm_sysref)) {
387 lwkt_reltoken(&vm->vm_map.token);
388 vmspace_terminate(vm);
390 lwkt_reltoken(&vm->vm_map.token);
395 * Swap useage is determined by taking the proportional swap used by
396 * VM objects backing the VM map. To make up for fractional losses,
397 * if the VM object has any swap use at all the associated map entries
398 * count for at least 1 swap page.
403 vmspace_swap_count(struct vmspace *vm)
405 vm_map_t map = &vm->vm_map;
411 lwkt_gettoken(&vm->vm_map.token);
412 for (cur = map->header.next; cur != &map->header; cur = cur->next) {
413 switch(cur->maptype) {
414 case VM_MAPTYPE_NORMAL:
415 case VM_MAPTYPE_VPAGETABLE:
416 if ((object = cur->object.vm_object) == NULL)
418 if (object->swblock_count) {
419 n = (cur->end - cur->start) / PAGE_SIZE;
420 count += object->swblock_count *
421 SWAP_META_PAGES * n / object->size + 1;
428 lwkt_reltoken(&vm->vm_map.token);
433 * Calculate the approximate number of anonymous pages in use by
434 * this vmspace. To make up for fractional losses, we count each
435 * VM object as having at least 1 anonymous page.
440 vmspace_anonymous_count(struct vmspace *vm)
442 vm_map_t map = &vm->vm_map;
447 lwkt_gettoken(&vm->vm_map.token);
448 for (cur = map->header.next; cur != &map->header; cur = cur->next) {
449 switch(cur->maptype) {
450 case VM_MAPTYPE_NORMAL:
451 case VM_MAPTYPE_VPAGETABLE:
452 if ((object = cur->object.vm_object) == NULL)
454 if (object->type != OBJT_DEFAULT &&
455 object->type != OBJT_SWAP) {
458 count += object->resident_page_count;
464 lwkt_reltoken(&vm->vm_map.token);
469 * Creates and returns a new empty VM map with the given physical map
470 * structure, and having the given lower and upper address bounds.
475 vm_map_create(vm_map_t result, pmap_t pmap, vm_offset_t min, vm_offset_t max)
478 result = zalloc(mapzone);
479 vm_map_init(result, min, max, pmap);
484 * Initialize an existing vm_map structure such as that in the vmspace
485 * structure. The pmap is initialized elsewhere.
490 vm_map_init(struct vm_map *map, vm_offset_t min, vm_offset_t max, pmap_t pmap)
492 map->header.next = map->header.prev = &map->header;
493 RB_INIT(&map->rb_root);
497 map->min_offset = min;
498 map->max_offset = max;
500 map->first_free = &map->header;
501 map->hint = &map->header;
504 lwkt_token_init(&map->token, "vm_map");
505 lockinit(&map->lock, "thrd_sleep", (hz + 9) / 10, 0);
506 TUNABLE_INT("vm.cache_vmspaces", &vmspace_sysref_class.nom_cache);
510 * Shadow the vm_map_entry's object. This typically needs to be done when
511 * a write fault is taken on an entry which had previously been cloned by
512 * fork(). The shared object (which might be NULL) must become private so
513 * we add a shadow layer above it.
515 * Object allocation for anonymous mappings is defered as long as possible.
516 * When creating a shadow, however, the underlying object must be instantiated
517 * so it can be shared.
519 * If the map segment is governed by a virtual page table then it is
520 * possible to address offsets beyond the mapped area. Just allocate
521 * a maximally sized object for this case.
523 * The vm_map must be exclusively locked.
524 * No other requirements.
528 vm_map_entry_shadow(vm_map_entry_t entry, int addref)
530 if (entry->maptype == VM_MAPTYPE_VPAGETABLE) {
531 vm_object_shadow(&entry->object.vm_object, &entry->offset,
532 0x7FFFFFFF, addref); /* XXX */
534 vm_object_shadow(&entry->object.vm_object, &entry->offset,
535 atop(entry->end - entry->start), addref);
537 entry->eflags &= ~MAP_ENTRY_NEEDS_COPY;
541 * Allocate an object for a vm_map_entry.
543 * Object allocation for anonymous mappings is defered as long as possible.
544 * This function is called when we can defer no longer, generally when a map
545 * entry might be split or forked or takes a page fault.
547 * If the map segment is governed by a virtual page table then it is
548 * possible to address offsets beyond the mapped area. Just allocate
549 * a maximally sized object for this case.
551 * The vm_map must be exclusively locked.
552 * No other requirements.
555 vm_map_entry_allocate_object(vm_map_entry_t entry)
559 if (entry->maptype == VM_MAPTYPE_VPAGETABLE) {
560 obj = vm_object_allocate(OBJT_DEFAULT, 0x7FFFFFFF); /* XXX */
562 obj = vm_object_allocate(OBJT_DEFAULT,
563 atop(entry->end - entry->start));
565 entry->object.vm_object = obj;
570 * Set an initial negative count so the first attempt to reserve
571 * space preloads a bunch of vm_map_entry's for this cpu. Also
572 * pre-allocate 2 vm_map_entries which will be needed by zalloc() to
573 * map a new page for vm_map_entry structures. SMP systems are
574 * particularly sensitive.
576 * This routine is called in early boot so we cannot just call
577 * vm_map_entry_reserve().
579 * Called from the low level boot code only (for each cpu)
582 vm_map_entry_reserve_cpu_init(globaldata_t gd)
584 vm_map_entry_t entry;
587 gd->gd_vme_avail -= MAP_RESERVE_COUNT * 2;
588 entry = &cpu_map_entry_init[gd->gd_cpuid][0];
589 for (i = 0; i < VMEPERCPU; ++i, ++entry) {
590 entry->next = gd->gd_vme_base;
591 gd->gd_vme_base = entry;
596 * Reserves vm_map_entry structures so code later on can manipulate
597 * map_entry structures within a locked map without blocking trying
598 * to allocate a new vm_map_entry.
603 vm_map_entry_reserve(int count)
605 struct globaldata *gd = mycpu;
606 vm_map_entry_t entry;
609 * Make sure we have enough structures in gd_vme_base to handle
610 * the reservation request.
612 * The critical section protects access to the per-cpu gd.
615 while (gd->gd_vme_avail < count) {
616 entry = zalloc(mapentzone);
617 entry->next = gd->gd_vme_base;
618 gd->gd_vme_base = entry;
621 gd->gd_vme_avail -= count;
628 * Releases previously reserved vm_map_entry structures that were not
629 * used. If we have too much junk in our per-cpu cache clean some of
635 vm_map_entry_release(int count)
637 struct globaldata *gd = mycpu;
638 vm_map_entry_t entry;
641 gd->gd_vme_avail += count;
642 while (gd->gd_vme_avail > MAP_RESERVE_SLOP) {
643 entry = gd->gd_vme_base;
644 KKASSERT(entry != NULL);
645 gd->gd_vme_base = entry->next;
648 zfree(mapentzone, entry);
655 * Reserve map entry structures for use in kernel_map itself. These
656 * entries have *ALREADY* been reserved on a per-cpu basis when the map
657 * was inited. This function is used by zalloc() to avoid a recursion
658 * when zalloc() itself needs to allocate additional kernel memory.
660 * This function works like the normal reserve but does not load the
661 * vm_map_entry cache (because that would result in an infinite
662 * recursion). Note that gd_vme_avail may go negative. This is expected.
664 * Any caller of this function must be sure to renormalize after
665 * potentially eating entries to ensure that the reserve supply
671 vm_map_entry_kreserve(int count)
673 struct globaldata *gd = mycpu;
676 gd->gd_vme_avail -= count;
678 KASSERT(gd->gd_vme_base != NULL,
679 ("no reserved entries left, gd_vme_avail = %d\n",
685 * Release previously reserved map entries for kernel_map. We do not
686 * attempt to clean up like the normal release function as this would
687 * cause an unnecessary (but probably not fatal) deep procedure call.
692 vm_map_entry_krelease(int count)
694 struct globaldata *gd = mycpu;
697 gd->gd_vme_avail += count;
702 * Allocates a VM map entry for insertion. No entry fields are filled in.
704 * The entries should have previously been reserved. The reservation count
705 * is tracked in (*countp).
709 static vm_map_entry_t
710 vm_map_entry_create(vm_map_t map, int *countp)
712 struct globaldata *gd = mycpu;
713 vm_map_entry_t entry;
715 KKASSERT(*countp > 0);
718 entry = gd->gd_vme_base;
719 KASSERT(entry != NULL, ("gd_vme_base NULL! count %d", *countp));
720 gd->gd_vme_base = entry->next;
727 * Dispose of a vm_map_entry that is no longer being referenced.
732 vm_map_entry_dispose(vm_map_t map, vm_map_entry_t entry, int *countp)
734 struct globaldata *gd = mycpu;
736 KKASSERT(map->hint != entry);
737 KKASSERT(map->first_free != entry);
741 entry->next = gd->gd_vme_base;
742 gd->gd_vme_base = entry;
748 * Insert/remove entries from maps.
750 * The related map must be exclusively locked.
751 * The caller must hold map->token
752 * No other requirements.
755 vm_map_entry_link(vm_map_t map,
756 vm_map_entry_t after_where,
757 vm_map_entry_t entry)
759 ASSERT_VM_MAP_LOCKED(map);
762 entry->prev = after_where;
763 entry->next = after_where->next;
764 entry->next->prev = entry;
765 after_where->next = entry;
766 if (vm_map_rb_tree_RB_INSERT(&map->rb_root, entry))
767 panic("vm_map_entry_link: dup addr map %p ent %p", map, entry);
771 vm_map_entry_unlink(vm_map_t map,
772 vm_map_entry_t entry)
777 ASSERT_VM_MAP_LOCKED(map);
779 if (entry->eflags & MAP_ENTRY_IN_TRANSITION) {
780 panic("vm_map_entry_unlink: attempt to mess with "
781 "locked entry! %p", entry);
787 vm_map_rb_tree_RB_REMOVE(&map->rb_root, entry);
792 * Finds the map entry containing (or immediately preceding) the specified
793 * address in the given map. The entry is returned in (*entry).
795 * The boolean result indicates whether the address is actually contained
798 * The related map must be locked.
799 * No other requirements.
802 vm_map_lookup_entry(vm_map_t map, vm_offset_t address, vm_map_entry_t *entry)
807 ASSERT_VM_MAP_LOCKED(map);
810 * XXX TEMPORARILY DISABLED. For some reason our attempt to revive
811 * the hint code with the red-black lookup meets with system crashes
812 * and lockups. We do not yet know why.
814 * It is possible that the problem is related to the setting
815 * of the hint during map_entry deletion, in the code specified
816 * at the GGG comment later on in this file.
818 * YYY More likely it's because this function can be called with
819 * a shared lock on the map, resulting in map->hint updates possibly
820 * racing. Fixed now but untested.
823 * Quickly check the cached hint, there's a good chance of a match.
827 if (tmp != &map->header) {
828 if (address >= tmp->start && address < tmp->end) {
836 * Locate the record from the top of the tree. 'last' tracks the
837 * closest prior record and is returned if no match is found, which
838 * in binary tree terms means tracking the most recent right-branch
839 * taken. If there is no prior record, &map->header is returned.
842 tmp = RB_ROOT(&map->rb_root);
845 if (address >= tmp->start) {
846 if (address < tmp->end) {
852 tmp = RB_RIGHT(tmp, rb_entry);
854 tmp = RB_LEFT(tmp, rb_entry);
862 * Inserts the given whole VM object into the target map at the specified
863 * address range. The object's size should match that of the address range.
865 * The map must be exclusively locked.
866 * The object must be held.
867 * The caller must have reserved sufficient vm_map_entry structures.
869 * If object is non-NULL, ref count must be bumped by caller prior to
870 * making call to account for the new entry.
873 vm_map_insert(vm_map_t map, int *countp,
874 vm_object_t object, vm_ooffset_t offset,
875 vm_offset_t start, vm_offset_t end,
876 vm_maptype_t maptype,
877 vm_prot_t prot, vm_prot_t max,
880 vm_map_entry_t new_entry;
881 vm_map_entry_t prev_entry;
882 vm_map_entry_t temp_entry;
883 vm_eflags_t protoeflags;
886 ASSERT_VM_MAP_LOCKED(map);
888 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
891 * Check that the start and end points are not bogus.
893 if ((start < map->min_offset) || (end > map->max_offset) ||
895 return (KERN_INVALID_ADDRESS);
898 * Find the entry prior to the proposed starting address; if it's part
899 * of an existing entry, this range is bogus.
901 if (vm_map_lookup_entry(map, start, &temp_entry))
902 return (KERN_NO_SPACE);
904 prev_entry = temp_entry;
907 * Assert that the next entry doesn't overlap the end point.
910 if ((prev_entry->next != &map->header) &&
911 (prev_entry->next->start < end))
912 return (KERN_NO_SPACE);
916 if (cow & MAP_COPY_ON_WRITE)
917 protoeflags |= MAP_ENTRY_COW|MAP_ENTRY_NEEDS_COPY;
919 if (cow & MAP_NOFAULT) {
920 protoeflags |= MAP_ENTRY_NOFAULT;
922 KASSERT(object == NULL,
923 ("vm_map_insert: paradoxical MAP_NOFAULT request"));
925 if (cow & MAP_DISABLE_SYNCER)
926 protoeflags |= MAP_ENTRY_NOSYNC;
927 if (cow & MAP_DISABLE_COREDUMP)
928 protoeflags |= MAP_ENTRY_NOCOREDUMP;
929 if (cow & MAP_IS_STACK)
930 protoeflags |= MAP_ENTRY_STACK;
931 if (cow & MAP_IS_KSTACK)
932 protoeflags |= MAP_ENTRY_KSTACK;
934 lwkt_gettoken(&map->token);
938 * When object is non-NULL, it could be shared with another
939 * process. We have to set or clear OBJ_ONEMAPPING
942 if ((object->ref_count > 1) || (object->shadow_count != 0)) {
943 vm_object_clear_flag(object, OBJ_ONEMAPPING);
946 else if ((prev_entry != &map->header) &&
947 (prev_entry->eflags == protoeflags) &&
948 (prev_entry->end == start) &&
949 (prev_entry->wired_count == 0) &&
950 prev_entry->maptype == maptype &&
951 ((prev_entry->object.vm_object == NULL) ||
952 vm_object_coalesce(prev_entry->object.vm_object,
953 OFF_TO_IDX(prev_entry->offset),
954 (vm_size_t)(prev_entry->end - prev_entry->start),
955 (vm_size_t)(end - prev_entry->end)))) {
957 * We were able to extend the object. Determine if we
958 * can extend the previous map entry to include the
961 if ((prev_entry->inheritance == VM_INHERIT_DEFAULT) &&
962 (prev_entry->protection == prot) &&
963 (prev_entry->max_protection == max)) {
964 map->size += (end - prev_entry->end);
965 prev_entry->end = end;
966 vm_map_simplify_entry(map, prev_entry, countp);
967 lwkt_reltoken(&map->token);
968 return (KERN_SUCCESS);
972 * If we can extend the object but cannot extend the
973 * map entry, we have to create a new map entry. We
974 * must bump the ref count on the extended object to
975 * account for it. object may be NULL.
977 object = prev_entry->object.vm_object;
978 offset = prev_entry->offset +
979 (prev_entry->end - prev_entry->start);
981 vm_object_hold(object);
982 vm_object_chain_wait(object);
983 vm_object_reference_locked(object);
989 * NOTE: if conditionals fail, object can be NULL here. This occurs
990 * in things like the buffer map where we manage kva but do not manage
998 new_entry = vm_map_entry_create(map, countp);
999 new_entry->start = start;
1000 new_entry->end = end;
1002 new_entry->maptype = maptype;
1003 new_entry->eflags = protoeflags;
1004 new_entry->object.vm_object = object;
1005 new_entry->offset = offset;
1006 new_entry->aux.master_pde = 0;
1008 new_entry->inheritance = VM_INHERIT_DEFAULT;
1009 new_entry->protection = prot;
1010 new_entry->max_protection = max;
1011 new_entry->wired_count = 0;
1014 * Insert the new entry into the list
1017 vm_map_entry_link(map, prev_entry, new_entry);
1018 map->size += new_entry->end - new_entry->start;
1021 * Update the free space hint. Entries cannot overlap.
1022 * An exact comparison is needed to avoid matching
1023 * against the map->header.
1025 if ((map->first_free == prev_entry) &&
1026 (prev_entry->end == new_entry->start)) {
1027 map->first_free = new_entry;
1032 * Temporarily removed to avoid MAP_STACK panic, due to
1033 * MAP_STACK being a huge hack. Will be added back in
1034 * when MAP_STACK (and the user stack mapping) is fixed.
1037 * It may be possible to simplify the entry
1039 vm_map_simplify_entry(map, new_entry, countp);
1043 * Try to pre-populate the page table. Mappings governed by virtual
1044 * page tables cannot be prepopulated without a lot of work, so
1047 if ((cow & (MAP_PREFAULT|MAP_PREFAULT_PARTIAL)) &&
1048 maptype != VM_MAPTYPE_VPAGETABLE) {
1049 pmap_object_init_pt(map->pmap, start, prot,
1050 object, OFF_TO_IDX(offset), end - start,
1051 cow & MAP_PREFAULT_PARTIAL);
1054 vm_object_drop(object);
1056 lwkt_reltoken(&map->token);
1057 return (KERN_SUCCESS);
1061 * Find sufficient space for `length' bytes in the given map, starting at
1062 * `start'. Returns 0 on success, 1 on no space.
1064 * This function will returned an arbitrarily aligned pointer. If no
1065 * particular alignment is required you should pass align as 1. Note that
1066 * the map may return PAGE_SIZE aligned pointers if all the lengths used in
1067 * the map are a multiple of PAGE_SIZE, even if you pass a smaller align
1070 * 'align' should be a power of 2 but is not required to be.
1072 * The map must be exclusively locked.
1073 * No other requirements.
1076 vm_map_findspace(vm_map_t map, vm_offset_t start, vm_size_t length,
1077 vm_size_t align, int flags, vm_offset_t *addr)
1079 vm_map_entry_t entry, next;
1081 vm_offset_t align_mask;
1083 if (start < map->min_offset)
1084 start = map->min_offset;
1085 if (start > map->max_offset)
1089 * If the alignment is not a power of 2 we will have to use
1090 * a mod/division, set align_mask to a special value.
1092 if ((align | (align - 1)) + 1 != (align << 1))
1093 align_mask = (vm_offset_t)-1;
1095 align_mask = align - 1;
1098 * Look for the first possible address; if there's already something
1099 * at this address, we have to start after it.
1101 if (start == map->min_offset) {
1102 if ((entry = map->first_free) != &map->header)
1107 if (vm_map_lookup_entry(map, start, &tmp))
1113 * Look through the rest of the map, trying to fit a new region in the
1114 * gap between existing regions, or after the very last region.
1116 for (;; start = (entry = next)->end) {
1118 * Adjust the proposed start by the requested alignment,
1119 * be sure that we didn't wrap the address.
1121 if (align_mask == (vm_offset_t)-1)
1122 end = ((start + align - 1) / align) * align;
1124 end = (start + align_mask) & ~align_mask;
1129 * Find the end of the proposed new region. Be sure we didn't
1130 * go beyond the end of the map, or wrap around the address.
1131 * Then check to see if this is the last entry or if the
1132 * proposed end fits in the gap between this and the next
1135 end = start + length;
1136 if (end > map->max_offset || end < start)
1141 * If the next entry's start address is beyond the desired
1142 * end address we may have found a good entry.
1144 * If the next entry is a stack mapping we do not map into
1145 * the stack's reserved space.
1147 * XXX continue to allow mapping into the stack's reserved
1148 * space if doing a MAP_STACK mapping inside a MAP_STACK
1149 * mapping, for backwards compatibility. But the caller
1150 * really should use MAP_STACK | MAP_TRYFIXED if they
1153 if (next == &map->header)
1155 if (next->start >= end) {
1156 if ((next->eflags & MAP_ENTRY_STACK) == 0)
1158 if (flags & MAP_STACK)
1160 if (next->start - next->aux.avail_ssize >= end)
1167 * Grow the kernel_map if necessary. pmap_growkernel() will panic
1168 * if it fails. The kernel_map is locked and nothing can steal
1169 * our address space if pmap_growkernel() blocks.
1171 * NOTE: This may be unconditionally called for kldload areas on
1172 * x86_64 because these do not bump kernel_vm_end (which would
1173 * fill 128G worth of page tables!). Therefore we must not
1176 if (map == &kernel_map) {
1179 kstop = round_page(start + length);
1180 if (kstop > kernel_vm_end)
1181 pmap_growkernel(start, kstop);
1188 * vm_map_find finds an unallocated region in the target address map with
1189 * the given length and allocates it. The search is defined to be first-fit
1190 * from the specified address; the region found is returned in the same
1193 * If object is non-NULL, ref count must be bumped by caller
1194 * prior to making call to account for the new entry.
1196 * No requirements. This function will lock the map temporarily.
1199 vm_map_find(vm_map_t map, vm_object_t object, vm_ooffset_t offset,
1200 vm_offset_t *addr, vm_size_t length, vm_size_t align,
1202 vm_maptype_t maptype,
1203 vm_prot_t prot, vm_prot_t max,
1212 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
1215 vm_object_hold(object);
1217 if (vm_map_findspace(map, start, length, align, 0, addr)) {
1219 vm_map_entry_release(count);
1220 return (KERN_NO_SPACE);
1224 result = vm_map_insert(map, &count, object, offset,
1225 start, start + length,
1230 vm_object_drop(object);
1232 vm_map_entry_release(count);
1238 * Simplify the given map entry by merging with either neighbor. This
1239 * routine also has the ability to merge with both neighbors.
1241 * This routine guarentees that the passed entry remains valid (though
1242 * possibly extended). When merging, this routine may delete one or
1243 * both neighbors. No action is taken on entries which have their
1244 * in-transition flag set.
1246 * The map must be exclusively locked.
1249 vm_map_simplify_entry(vm_map_t map, vm_map_entry_t entry, int *countp)
1251 vm_map_entry_t next, prev;
1252 vm_size_t prevsize, esize;
1254 if (entry->eflags & MAP_ENTRY_IN_TRANSITION) {
1255 ++mycpu->gd_cnt.v_intrans_coll;
1259 if (entry->maptype == VM_MAPTYPE_SUBMAP)
1263 if (prev != &map->header) {
1264 prevsize = prev->end - prev->start;
1265 if ( (prev->end == entry->start) &&
1266 (prev->maptype == entry->maptype) &&
1267 (prev->object.vm_object == entry->object.vm_object) &&
1268 (!prev->object.vm_object ||
1269 (prev->offset + prevsize == entry->offset)) &&
1270 (prev->eflags == entry->eflags) &&
1271 (prev->protection == entry->protection) &&
1272 (prev->max_protection == entry->max_protection) &&
1273 (prev->inheritance == entry->inheritance) &&
1274 (prev->wired_count == entry->wired_count)) {
1275 if (map->first_free == prev)
1276 map->first_free = entry;
1277 if (map->hint == prev)
1279 vm_map_entry_unlink(map, prev);
1280 entry->start = prev->start;
1281 entry->offset = prev->offset;
1282 if (prev->object.vm_object)
1283 vm_object_deallocate(prev->object.vm_object);
1284 vm_map_entry_dispose(map, prev, countp);
1289 if (next != &map->header) {
1290 esize = entry->end - entry->start;
1291 if ((entry->end == next->start) &&
1292 (next->maptype == entry->maptype) &&
1293 (next->object.vm_object == entry->object.vm_object) &&
1294 (!entry->object.vm_object ||
1295 (entry->offset + esize == next->offset)) &&
1296 (next->eflags == entry->eflags) &&
1297 (next->protection == entry->protection) &&
1298 (next->max_protection == entry->max_protection) &&
1299 (next->inheritance == entry->inheritance) &&
1300 (next->wired_count == entry->wired_count)) {
1301 if (map->first_free == next)
1302 map->first_free = entry;
1303 if (map->hint == next)
1305 vm_map_entry_unlink(map, next);
1306 entry->end = next->end;
1307 if (next->object.vm_object)
1308 vm_object_deallocate(next->object.vm_object);
1309 vm_map_entry_dispose(map, next, countp);
1315 * Asserts that the given entry begins at or after the specified address.
1316 * If necessary, it splits the entry into two.
1318 #define vm_map_clip_start(map, entry, startaddr, countp) \
1320 if (startaddr > entry->start) \
1321 _vm_map_clip_start(map, entry, startaddr, countp); \
1325 * This routine is called only when it is known that the entry must be split.
1327 * The map must be exclusively locked.
1330 _vm_map_clip_start(vm_map_t map, vm_map_entry_t entry, vm_offset_t start,
1333 vm_map_entry_t new_entry;
1336 * Split off the front portion -- note that we must insert the new
1337 * entry BEFORE this one, so that this entry has the specified
1341 vm_map_simplify_entry(map, entry, countp);
1344 * If there is no object backing this entry, we might as well create
1345 * one now. If we defer it, an object can get created after the map
1346 * is clipped, and individual objects will be created for the split-up
1347 * map. This is a bit of a hack, but is also about the best place to
1348 * put this improvement.
1350 if (entry->object.vm_object == NULL && !map->system_map) {
1351 vm_map_entry_allocate_object(entry);
1354 new_entry = vm_map_entry_create(map, countp);
1355 *new_entry = *entry;
1357 new_entry->end = start;
1358 entry->offset += (start - entry->start);
1359 entry->start = start;
1361 vm_map_entry_link(map, entry->prev, new_entry);
1363 switch(entry->maptype) {
1364 case VM_MAPTYPE_NORMAL:
1365 case VM_MAPTYPE_VPAGETABLE:
1366 if (new_entry->object.vm_object) {
1367 vm_object_hold(new_entry->object.vm_object);
1368 vm_object_chain_wait(new_entry->object.vm_object);
1369 vm_object_reference_locked(new_entry->object.vm_object);
1370 vm_object_drop(new_entry->object.vm_object);
1379 * Asserts that the given entry ends at or before the specified address.
1380 * If necessary, it splits the entry into two.
1382 * The map must be exclusively locked.
1384 #define vm_map_clip_end(map, entry, endaddr, countp) \
1386 if (endaddr < entry->end) \
1387 _vm_map_clip_end(map, entry, endaddr, countp); \
1391 * This routine is called only when it is known that the entry must be split.
1393 * The map must be exclusively locked.
1396 _vm_map_clip_end(vm_map_t map, vm_map_entry_t entry, vm_offset_t end,
1399 vm_map_entry_t new_entry;
1402 * If there is no object backing this entry, we might as well create
1403 * one now. If we defer it, an object can get created after the map
1404 * is clipped, and individual objects will be created for the split-up
1405 * map. This is a bit of a hack, but is also about the best place to
1406 * put this improvement.
1409 if (entry->object.vm_object == NULL && !map->system_map) {
1410 vm_map_entry_allocate_object(entry);
1414 * Create a new entry and insert it AFTER the specified entry
1417 new_entry = vm_map_entry_create(map, countp);
1418 *new_entry = *entry;
1420 new_entry->start = entry->end = end;
1421 new_entry->offset += (end - entry->start);
1423 vm_map_entry_link(map, entry, new_entry);
1425 switch(entry->maptype) {
1426 case VM_MAPTYPE_NORMAL:
1427 case VM_MAPTYPE_VPAGETABLE:
1428 if (new_entry->object.vm_object) {
1429 vm_object_hold(new_entry->object.vm_object);
1430 vm_object_chain_wait(new_entry->object.vm_object);
1431 vm_object_reference_locked(new_entry->object.vm_object);
1432 vm_object_drop(new_entry->object.vm_object);
1441 * Asserts that the starting and ending region addresses fall within the
1442 * valid range for the map.
1444 #define VM_MAP_RANGE_CHECK(map, start, end) \
1446 if (start < vm_map_min(map)) \
1447 start = vm_map_min(map); \
1448 if (end > vm_map_max(map)) \
1449 end = vm_map_max(map); \
1455 * Used to block when an in-transition collison occurs. The map
1456 * is unlocked for the sleep and relocked before the return.
1459 vm_map_transition_wait(vm_map_t map)
1461 tsleep_interlock(map, 0);
1463 tsleep(map, PINTERLOCKED, "vment", 0);
1468 * When we do blocking operations with the map lock held it is
1469 * possible that a clip might have occured on our in-transit entry,
1470 * requiring an adjustment to the entry in our loop. These macros
1471 * help the pageable and clip_range code deal with the case. The
1472 * conditional costs virtually nothing if no clipping has occured.
1475 #define CLIP_CHECK_BACK(entry, save_start) \
1477 while (entry->start != save_start) { \
1478 entry = entry->prev; \
1479 KASSERT(entry != &map->header, ("bad entry clip")); \
1483 #define CLIP_CHECK_FWD(entry, save_end) \
1485 while (entry->end != save_end) { \
1486 entry = entry->next; \
1487 KASSERT(entry != &map->header, ("bad entry clip")); \
1493 * Clip the specified range and return the base entry. The
1494 * range may cover several entries starting at the returned base
1495 * and the first and last entry in the covering sequence will be
1496 * properly clipped to the requested start and end address.
1498 * If no holes are allowed you should pass the MAP_CLIP_NO_HOLES
1501 * The MAP_ENTRY_IN_TRANSITION flag will be set for the entries
1502 * covered by the requested range.
1504 * The map must be exclusively locked on entry and will remain locked
1505 * on return. If no range exists or the range contains holes and you
1506 * specified that no holes were allowed, NULL will be returned. This
1507 * routine may temporarily unlock the map in order avoid a deadlock when
1512 vm_map_clip_range(vm_map_t map, vm_offset_t start, vm_offset_t end,
1513 int *countp, int flags)
1515 vm_map_entry_t start_entry;
1516 vm_map_entry_t entry;
1519 * Locate the entry and effect initial clipping. The in-transition
1520 * case does not occur very often so do not try to optimize it.
1523 if (vm_map_lookup_entry(map, start, &start_entry) == FALSE)
1525 entry = start_entry;
1526 if (entry->eflags & MAP_ENTRY_IN_TRANSITION) {
1527 entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP;
1528 ++mycpu->gd_cnt.v_intrans_coll;
1529 ++mycpu->gd_cnt.v_intrans_wait;
1530 vm_map_transition_wait(map);
1532 * entry and/or start_entry may have been clipped while
1533 * we slept, or may have gone away entirely. We have
1534 * to restart from the lookup.
1540 * Since we hold an exclusive map lock we do not have to restart
1541 * after clipping, even though clipping may block in zalloc.
1543 vm_map_clip_start(map, entry, start, countp);
1544 vm_map_clip_end(map, entry, end, countp);
1545 entry->eflags |= MAP_ENTRY_IN_TRANSITION;
1548 * Scan entries covered by the range. When working on the next
1549 * entry a restart need only re-loop on the current entry which
1550 * we have already locked, since 'next' may have changed. Also,
1551 * even though entry is safe, it may have been clipped so we
1552 * have to iterate forwards through the clip after sleeping.
1554 while (entry->next != &map->header && entry->next->start < end) {
1555 vm_map_entry_t next = entry->next;
1557 if (flags & MAP_CLIP_NO_HOLES) {
1558 if (next->start > entry->end) {
1559 vm_map_unclip_range(map, start_entry,
1560 start, entry->end, countp, flags);
1565 if (next->eflags & MAP_ENTRY_IN_TRANSITION) {
1566 vm_offset_t save_end = entry->end;
1567 next->eflags |= MAP_ENTRY_NEEDS_WAKEUP;
1568 ++mycpu->gd_cnt.v_intrans_coll;
1569 ++mycpu->gd_cnt.v_intrans_wait;
1570 vm_map_transition_wait(map);
1573 * clips might have occured while we blocked.
1575 CLIP_CHECK_FWD(entry, save_end);
1576 CLIP_CHECK_BACK(start_entry, start);
1580 * No restart necessary even though clip_end may block, we
1581 * are holding the map lock.
1583 vm_map_clip_end(map, next, end, countp);
1584 next->eflags |= MAP_ENTRY_IN_TRANSITION;
1587 if (flags & MAP_CLIP_NO_HOLES) {
1588 if (entry->end != end) {
1589 vm_map_unclip_range(map, start_entry,
1590 start, entry->end, countp, flags);
1594 return(start_entry);
1598 * Undo the effect of vm_map_clip_range(). You should pass the same
1599 * flags and the same range that you passed to vm_map_clip_range().
1600 * This code will clear the in-transition flag on the entries and
1601 * wake up anyone waiting. This code will also simplify the sequence
1602 * and attempt to merge it with entries before and after the sequence.
1604 * The map must be locked on entry and will remain locked on return.
1606 * Note that you should also pass the start_entry returned by
1607 * vm_map_clip_range(). However, if you block between the two calls
1608 * with the map unlocked please be aware that the start_entry may
1609 * have been clipped and you may need to scan it backwards to find
1610 * the entry corresponding with the original start address. You are
1611 * responsible for this, vm_map_unclip_range() expects the correct
1612 * start_entry to be passed to it and will KASSERT otherwise.
1616 vm_map_unclip_range(vm_map_t map, vm_map_entry_t start_entry,
1617 vm_offset_t start, vm_offset_t end,
1618 int *countp, int flags)
1620 vm_map_entry_t entry;
1622 entry = start_entry;
1624 KASSERT(entry->start == start, ("unclip_range: illegal base entry"));
1625 while (entry != &map->header && entry->start < end) {
1626 KASSERT(entry->eflags & MAP_ENTRY_IN_TRANSITION,
1627 ("in-transition flag not set during unclip on: %p",
1629 KASSERT(entry->end <= end,
1630 ("unclip_range: tail wasn't clipped"));
1631 entry->eflags &= ~MAP_ENTRY_IN_TRANSITION;
1632 if (entry->eflags & MAP_ENTRY_NEEDS_WAKEUP) {
1633 entry->eflags &= ~MAP_ENTRY_NEEDS_WAKEUP;
1636 entry = entry->next;
1640 * Simplification does not block so there is no restart case.
1642 entry = start_entry;
1643 while (entry != &map->header && entry->start < end) {
1644 vm_map_simplify_entry(map, entry, countp);
1645 entry = entry->next;
1650 * Mark the given range as handled by a subordinate map.
1652 * This range must have been created with vm_map_find(), and no other
1653 * operations may have been performed on this range prior to calling
1656 * Submappings cannot be removed.
1661 vm_map_submap(vm_map_t map, vm_offset_t start, vm_offset_t end, vm_map_t submap)
1663 vm_map_entry_t entry;
1664 int result = KERN_INVALID_ARGUMENT;
1667 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
1670 VM_MAP_RANGE_CHECK(map, start, end);
1672 if (vm_map_lookup_entry(map, start, &entry)) {
1673 vm_map_clip_start(map, entry, start, &count);
1675 entry = entry->next;
1678 vm_map_clip_end(map, entry, end, &count);
1680 if ((entry->start == start) && (entry->end == end) &&
1681 ((entry->eflags & MAP_ENTRY_COW) == 0) &&
1682 (entry->object.vm_object == NULL)) {
1683 entry->object.sub_map = submap;
1684 entry->maptype = VM_MAPTYPE_SUBMAP;
1685 result = KERN_SUCCESS;
1688 vm_map_entry_release(count);
1694 * Sets the protection of the specified address region in the target map.
1695 * If "set_max" is specified, the maximum protection is to be set;
1696 * otherwise, only the current protection is affected.
1698 * The protection is not applicable to submaps, but is applicable to normal
1699 * maps and maps governed by virtual page tables. For example, when operating
1700 * on a virtual page table our protection basically controls how COW occurs
1701 * on the backing object, whereas the virtual page table abstraction itself
1702 * is an abstraction for userland.
1707 vm_map_protect(vm_map_t map, vm_offset_t start, vm_offset_t end,
1708 vm_prot_t new_prot, boolean_t set_max)
1710 vm_map_entry_t current;
1711 vm_map_entry_t entry;
1714 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
1717 VM_MAP_RANGE_CHECK(map, start, end);
1719 if (vm_map_lookup_entry(map, start, &entry)) {
1720 vm_map_clip_start(map, entry, start, &count);
1722 entry = entry->next;
1726 * Make a first pass to check for protection violations.
1729 while ((current != &map->header) && (current->start < end)) {
1730 if (current->maptype == VM_MAPTYPE_SUBMAP) {
1732 vm_map_entry_release(count);
1733 return (KERN_INVALID_ARGUMENT);
1735 if ((new_prot & current->max_protection) != new_prot) {
1737 vm_map_entry_release(count);
1738 return (KERN_PROTECTION_FAILURE);
1740 current = current->next;
1744 * Go back and fix up protections. [Note that clipping is not
1745 * necessary the second time.]
1749 while ((current != &map->header) && (current->start < end)) {
1752 vm_map_clip_end(map, current, end, &count);
1754 old_prot = current->protection;
1756 current->protection =
1757 (current->max_protection = new_prot) &
1760 current->protection = new_prot;
1764 * Update physical map if necessary. Worry about copy-on-write
1765 * here -- CHECK THIS XXX
1768 if (current->protection != old_prot) {
1769 #define MASK(entry) (((entry)->eflags & MAP_ENTRY_COW) ? ~VM_PROT_WRITE : \
1772 pmap_protect(map->pmap, current->start,
1774 current->protection & MASK(current));
1778 vm_map_simplify_entry(map, current, &count);
1780 current = current->next;
1784 vm_map_entry_release(count);
1785 return (KERN_SUCCESS);
1789 * This routine traverses a processes map handling the madvise
1790 * system call. Advisories are classified as either those effecting
1791 * the vm_map_entry structure, or those effecting the underlying
1794 * The <value> argument is used for extended madvise calls.
1799 vm_map_madvise(vm_map_t map, vm_offset_t start, vm_offset_t end,
1800 int behav, off_t value)
1802 vm_map_entry_t current, entry;
1808 * Some madvise calls directly modify the vm_map_entry, in which case
1809 * we need to use an exclusive lock on the map and we need to perform
1810 * various clipping operations. Otherwise we only need a read-lock
1814 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
1818 case MADV_SEQUENTIAL:
1832 vm_map_lock_read(map);
1835 vm_map_entry_release(count);
1840 * Locate starting entry and clip if necessary.
1843 VM_MAP_RANGE_CHECK(map, start, end);
1845 if (vm_map_lookup_entry(map, start, &entry)) {
1847 vm_map_clip_start(map, entry, start, &count);
1849 entry = entry->next;
1854 * madvise behaviors that are implemented in the vm_map_entry.
1856 * We clip the vm_map_entry so that behavioral changes are
1857 * limited to the specified address range.
1859 for (current = entry;
1860 (current != &map->header) && (current->start < end);
1861 current = current->next
1863 if (current->maptype == VM_MAPTYPE_SUBMAP)
1866 vm_map_clip_end(map, current, end, &count);
1870 vm_map_entry_set_behavior(current, MAP_ENTRY_BEHAV_NORMAL);
1872 case MADV_SEQUENTIAL:
1873 vm_map_entry_set_behavior(current, MAP_ENTRY_BEHAV_SEQUENTIAL);
1876 vm_map_entry_set_behavior(current, MAP_ENTRY_BEHAV_RANDOM);
1879 current->eflags |= MAP_ENTRY_NOSYNC;
1882 current->eflags &= ~MAP_ENTRY_NOSYNC;
1885 current->eflags |= MAP_ENTRY_NOCOREDUMP;
1888 current->eflags &= ~MAP_ENTRY_NOCOREDUMP;
1892 * Invalidate the related pmap entries, used
1893 * to flush portions of the real kernel's
1894 * pmap when the caller has removed or
1895 * modified existing mappings in a virtual
1898 pmap_remove(map->pmap,
1899 current->start, current->end);
1903 * Set the page directory page for a map
1904 * governed by a virtual page table. Mark
1905 * the entry as being governed by a virtual
1906 * page table if it is not.
1908 * XXX the page directory page is stored
1909 * in the avail_ssize field if the map_entry.
1911 * XXX the map simplification code does not
1912 * compare this field so weird things may
1913 * happen if you do not apply this function
1914 * to the entire mapping governed by the
1915 * virtual page table.
1917 if (current->maptype != VM_MAPTYPE_VPAGETABLE) {
1921 current->aux.master_pde = value;
1922 pmap_remove(map->pmap,
1923 current->start, current->end);
1929 vm_map_simplify_entry(map, current, &count);
1937 * madvise behaviors that are implemented in the underlying
1940 * Since we don't clip the vm_map_entry, we have to clip
1941 * the vm_object pindex and count.
1943 * NOTE! We currently do not support these functions on
1944 * virtual page tables.
1946 for (current = entry;
1947 (current != &map->header) && (current->start < end);
1948 current = current->next
1950 vm_offset_t useStart;
1952 if (current->maptype != VM_MAPTYPE_NORMAL)
1955 pindex = OFF_TO_IDX(current->offset);
1956 count = atop(current->end - current->start);
1957 useStart = current->start;
1959 if (current->start < start) {
1960 pindex += atop(start - current->start);
1961 count -= atop(start - current->start);
1964 if (current->end > end)
1965 count -= atop(current->end - end);
1970 vm_object_madvise(current->object.vm_object,
1971 pindex, count, behav);
1974 * Try to populate the page table. Mappings governed
1975 * by virtual page tables cannot be pre-populated
1976 * without a lot of work so don't try.
1978 if (behav == MADV_WILLNEED &&
1979 current->maptype != VM_MAPTYPE_VPAGETABLE) {
1980 pmap_object_init_pt(
1983 current->protection,
1984 current->object.vm_object,
1986 (count << PAGE_SHIFT),
1987 MAP_PREFAULT_MADVISE
1991 vm_map_unlock_read(map);
1993 vm_map_entry_release(count);
1999 * Sets the inheritance of the specified address range in the target map.
2000 * Inheritance affects how the map will be shared with child maps at the
2001 * time of vm_map_fork.
2004 vm_map_inherit(vm_map_t map, vm_offset_t start, vm_offset_t end,
2005 vm_inherit_t new_inheritance)
2007 vm_map_entry_t entry;
2008 vm_map_entry_t temp_entry;
2011 switch (new_inheritance) {
2012 case VM_INHERIT_NONE:
2013 case VM_INHERIT_COPY:
2014 case VM_INHERIT_SHARE:
2017 return (KERN_INVALID_ARGUMENT);
2020 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
2023 VM_MAP_RANGE_CHECK(map, start, end);
2025 if (vm_map_lookup_entry(map, start, &temp_entry)) {
2027 vm_map_clip_start(map, entry, start, &count);
2029 entry = temp_entry->next;
2031 while ((entry != &map->header) && (entry->start < end)) {
2032 vm_map_clip_end(map, entry, end, &count);
2034 entry->inheritance = new_inheritance;
2036 vm_map_simplify_entry(map, entry, &count);
2038 entry = entry->next;
2041 vm_map_entry_release(count);
2042 return (KERN_SUCCESS);
2046 * Implement the semantics of mlock
2049 vm_map_unwire(vm_map_t map, vm_offset_t start, vm_offset_t real_end,
2050 boolean_t new_pageable)
2052 vm_map_entry_t entry;
2053 vm_map_entry_t start_entry;
2055 int rv = KERN_SUCCESS;
2058 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
2060 VM_MAP_RANGE_CHECK(map, start, real_end);
2063 start_entry = vm_map_clip_range(map, start, end, &count,
2065 if (start_entry == NULL) {
2067 vm_map_entry_release(count);
2068 return (KERN_INVALID_ADDRESS);
2071 if (new_pageable == 0) {
2072 entry = start_entry;
2073 while ((entry != &map->header) && (entry->start < end)) {
2074 vm_offset_t save_start;
2075 vm_offset_t save_end;
2078 * Already user wired or hard wired (trivial cases)
2080 if (entry->eflags & MAP_ENTRY_USER_WIRED) {
2081 entry = entry->next;
2084 if (entry->wired_count != 0) {
2085 entry->wired_count++;
2086 entry->eflags |= MAP_ENTRY_USER_WIRED;
2087 entry = entry->next;
2092 * A new wiring requires instantiation of appropriate
2093 * management structures and the faulting in of the
2096 if (entry->maptype != VM_MAPTYPE_SUBMAP) {
2097 int copyflag = entry->eflags &
2098 MAP_ENTRY_NEEDS_COPY;
2099 if (copyflag && ((entry->protection &
2100 VM_PROT_WRITE) != 0)) {
2101 vm_map_entry_shadow(entry, 0);
2102 } else if (entry->object.vm_object == NULL &&
2104 vm_map_entry_allocate_object(entry);
2107 entry->wired_count++;
2108 entry->eflags |= MAP_ENTRY_USER_WIRED;
2111 * Now fault in the area. Note that vm_fault_wire()
2112 * may release the map lock temporarily, it will be
2113 * relocked on return. The in-transition
2114 * flag protects the entries.
2116 save_start = entry->start;
2117 save_end = entry->end;
2118 rv = vm_fault_wire(map, entry, TRUE);
2120 CLIP_CHECK_BACK(entry, save_start);
2122 KASSERT(entry->wired_count == 1, ("bad wired_count on entry"));
2123 entry->eflags &= ~MAP_ENTRY_USER_WIRED;
2124 entry->wired_count = 0;
2125 if (entry->end == save_end)
2127 entry = entry->next;
2128 KASSERT(entry != &map->header, ("bad entry clip during backout"));
2130 end = save_start; /* unwire the rest */
2134 * note that even though the entry might have been
2135 * clipped, the USER_WIRED flag we set prevents
2136 * duplication so we do not have to do a
2139 entry = entry->next;
2143 * If we failed fall through to the unwiring section to
2144 * unwire what we had wired so far. 'end' has already
2151 * start_entry might have been clipped if we unlocked the
2152 * map and blocked. No matter how clipped it has gotten
2153 * there should be a fragment that is on our start boundary.
2155 CLIP_CHECK_BACK(start_entry, start);
2159 * Deal with the unwiring case.
2163 * This is the unwiring case. We must first ensure that the
2164 * range to be unwired is really wired down. We know there
2167 entry = start_entry;
2168 while ((entry != &map->header) && (entry->start < end)) {
2169 if ((entry->eflags & MAP_ENTRY_USER_WIRED) == 0) {
2170 rv = KERN_INVALID_ARGUMENT;
2173 KASSERT(entry->wired_count != 0, ("wired count was 0 with USER_WIRED set! %p", entry));
2174 entry = entry->next;
2178 * Now decrement the wiring count for each region. If a region
2179 * becomes completely unwired, unwire its physical pages and
2183 * The map entries are processed in a loop, checking to
2184 * make sure the entry is wired and asserting it has a wired
2185 * count. However, another loop was inserted more-or-less in
2186 * the middle of the unwiring path. This loop picks up the
2187 * "entry" loop variable from the first loop without first
2188 * setting it to start_entry. Naturally, the secound loop
2189 * is never entered and the pages backing the entries are
2190 * never unwired. This can lead to a leak of wired pages.
2192 entry = start_entry;
2193 while ((entry != &map->header) && (entry->start < end)) {
2194 KASSERT(entry->eflags & MAP_ENTRY_USER_WIRED,
2195 ("expected USER_WIRED on entry %p", entry));
2196 entry->eflags &= ~MAP_ENTRY_USER_WIRED;
2197 entry->wired_count--;
2198 if (entry->wired_count == 0)
2199 vm_fault_unwire(map, entry);
2200 entry = entry->next;
2204 vm_map_unclip_range(map, start_entry, start, real_end, &count,
2208 vm_map_entry_release(count);
2213 * Sets the pageability of the specified address range in the target map.
2214 * Regions specified as not pageable require locked-down physical
2215 * memory and physical page maps.
2217 * The map must not be locked, but a reference must remain to the map
2218 * throughout the call.
2220 * This function may be called via the zalloc path and must properly
2221 * reserve map entries for kernel_map.
2226 vm_map_wire(vm_map_t map, vm_offset_t start, vm_offset_t real_end, int kmflags)
2228 vm_map_entry_t entry;
2229 vm_map_entry_t start_entry;
2231 int rv = KERN_SUCCESS;
2234 if (kmflags & KM_KRESERVE)
2235 count = vm_map_entry_kreserve(MAP_RESERVE_COUNT);
2237 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
2239 VM_MAP_RANGE_CHECK(map, start, real_end);
2242 start_entry = vm_map_clip_range(map, start, end, &count,
2244 if (start_entry == NULL) {
2246 rv = KERN_INVALID_ADDRESS;
2249 if ((kmflags & KM_PAGEABLE) == 0) {
2253 * 1. Holding the write lock, we create any shadow or zero-fill
2254 * objects that need to be created. Then we clip each map
2255 * entry to the region to be wired and increment its wiring
2256 * count. We create objects before clipping the map entries
2257 * to avoid object proliferation.
2259 * 2. We downgrade to a read lock, and call vm_fault_wire to
2260 * fault in the pages for any newly wired area (wired_count is
2263 * Downgrading to a read lock for vm_fault_wire avoids a
2264 * possible deadlock with another process that may have faulted
2265 * on one of the pages to be wired (it would mark the page busy,
2266 * blocking us, then in turn block on the map lock that we
2267 * hold). Because of problems in the recursive lock package,
2268 * we cannot upgrade to a write lock in vm_map_lookup. Thus,
2269 * any actions that require the write lock must be done
2270 * beforehand. Because we keep the read lock on the map, the
2271 * copy-on-write status of the entries we modify here cannot
2274 entry = start_entry;
2275 while ((entry != &map->header) && (entry->start < end)) {
2277 * Trivial case if the entry is already wired
2279 if (entry->wired_count) {
2280 entry->wired_count++;
2281 entry = entry->next;
2286 * The entry is being newly wired, we have to setup
2287 * appropriate management structures. A shadow
2288 * object is required for a copy-on-write region,
2289 * or a normal object for a zero-fill region. We
2290 * do not have to do this for entries that point to sub
2291 * maps because we won't hold the lock on the sub map.
2293 if (entry->maptype != VM_MAPTYPE_SUBMAP) {
2294 int copyflag = entry->eflags &
2295 MAP_ENTRY_NEEDS_COPY;
2296 if (copyflag && ((entry->protection &
2297 VM_PROT_WRITE) != 0)) {
2298 vm_map_entry_shadow(entry, 0);
2299 } else if (entry->object.vm_object == NULL &&
2301 vm_map_entry_allocate_object(entry);
2305 entry->wired_count++;
2306 entry = entry->next;
2314 * HACK HACK HACK HACK
2316 * vm_fault_wire() temporarily unlocks the map to avoid
2317 * deadlocks. The in-transition flag from vm_map_clip_range
2318 * call should protect us from changes while the map is
2321 * NOTE: Previously this comment stated that clipping might
2322 * still occur while the entry is unlocked, but from
2323 * what I can tell it actually cannot.
2325 * It is unclear whether the CLIP_CHECK_*() calls
2326 * are still needed but we keep them in anyway.
2328 * HACK HACK HACK HACK
2331 entry = start_entry;
2332 while (entry != &map->header && entry->start < end) {
2334 * If vm_fault_wire fails for any page we need to undo
2335 * what has been done. We decrement the wiring count
2336 * for those pages which have not yet been wired (now)
2337 * and unwire those that have (later).
2339 vm_offset_t save_start = entry->start;
2340 vm_offset_t save_end = entry->end;
2342 if (entry->wired_count == 1)
2343 rv = vm_fault_wire(map, entry, FALSE);
2345 CLIP_CHECK_BACK(entry, save_start);
2347 KASSERT(entry->wired_count == 1, ("wired_count changed unexpectedly"));
2348 entry->wired_count = 0;
2349 if (entry->end == save_end)
2351 entry = entry->next;
2352 KASSERT(entry != &map->header, ("bad entry clip during backout"));
2357 CLIP_CHECK_FWD(entry, save_end);
2358 entry = entry->next;
2362 * If a failure occured undo everything by falling through
2363 * to the unwiring code. 'end' has already been adjusted
2367 kmflags |= KM_PAGEABLE;
2370 * start_entry is still IN_TRANSITION but may have been
2371 * clipped since vm_fault_wire() unlocks and relocks the
2372 * map. No matter how clipped it has gotten there should
2373 * be a fragment that is on our start boundary.
2375 CLIP_CHECK_BACK(start_entry, start);
2378 if (kmflags & KM_PAGEABLE) {
2380 * This is the unwiring case. We must first ensure that the
2381 * range to be unwired is really wired down. We know there
2384 entry = start_entry;
2385 while ((entry != &map->header) && (entry->start < end)) {
2386 if (entry->wired_count == 0) {
2387 rv = KERN_INVALID_ARGUMENT;
2390 entry = entry->next;
2394 * Now decrement the wiring count for each region. If a region
2395 * becomes completely unwired, unwire its physical pages and
2398 entry = start_entry;
2399 while ((entry != &map->header) && (entry->start < end)) {
2400 entry->wired_count--;
2401 if (entry->wired_count == 0)
2402 vm_fault_unwire(map, entry);
2403 entry = entry->next;
2407 vm_map_unclip_range(map, start_entry, start, real_end,
2408 &count, MAP_CLIP_NO_HOLES);
2412 if (kmflags & KM_KRESERVE)
2413 vm_map_entry_krelease(count);
2415 vm_map_entry_release(count);
2420 * Mark a newly allocated address range as wired but do not fault in
2421 * the pages. The caller is expected to load the pages into the object.
2423 * The map must be locked on entry and will remain locked on return.
2424 * No other requirements.
2427 vm_map_set_wired_quick(vm_map_t map, vm_offset_t addr, vm_size_t size,
2430 vm_map_entry_t scan;
2431 vm_map_entry_t entry;
2433 entry = vm_map_clip_range(map, addr, addr + size,
2434 countp, MAP_CLIP_NO_HOLES);
2436 scan != &map->header && scan->start < addr + size;
2437 scan = scan->next) {
2438 KKASSERT(entry->wired_count == 0);
2439 entry->wired_count = 1;
2441 vm_map_unclip_range(map, entry, addr, addr + size,
2442 countp, MAP_CLIP_NO_HOLES);
2446 * Push any dirty cached pages in the address range to their pager.
2447 * If syncio is TRUE, dirty pages are written synchronously.
2448 * If invalidate is TRUE, any cached pages are freed as well.
2450 * This routine is called by sys_msync()
2452 * Returns an error if any part of the specified range is not mapped.
2457 vm_map_clean(vm_map_t map, vm_offset_t start, vm_offset_t end,
2458 boolean_t syncio, boolean_t invalidate)
2460 vm_map_entry_t current;
2461 vm_map_entry_t entry;
2465 vm_ooffset_t offset;
2467 vm_map_lock_read(map);
2468 VM_MAP_RANGE_CHECK(map, start, end);
2469 if (!vm_map_lookup_entry(map, start, &entry)) {
2470 vm_map_unlock_read(map);
2471 return (KERN_INVALID_ADDRESS);
2473 lwkt_gettoken(&map->token);
2476 * Make a first pass to check for holes.
2478 for (current = entry; current->start < end; current = current->next) {
2479 if (current->maptype == VM_MAPTYPE_SUBMAP) {
2480 lwkt_reltoken(&map->token);
2481 vm_map_unlock_read(map);
2482 return (KERN_INVALID_ARGUMENT);
2484 if (end > current->end &&
2485 (current->next == &map->header ||
2486 current->end != current->next->start)) {
2487 lwkt_reltoken(&map->token);
2488 vm_map_unlock_read(map);
2489 return (KERN_INVALID_ADDRESS);
2494 pmap_remove(vm_map_pmap(map), start, end);
2497 * Make a second pass, cleaning/uncaching pages from the indicated
2500 for (current = entry; current->start < end; current = current->next) {
2501 offset = current->offset + (start - current->start);
2502 size = (end <= current->end ? end : current->end) - start;
2503 if (current->maptype == VM_MAPTYPE_SUBMAP) {
2505 vm_map_entry_t tentry;
2508 smap = current->object.sub_map;
2509 vm_map_lock_read(smap);
2510 vm_map_lookup_entry(smap, offset, &tentry);
2511 tsize = tentry->end - offset;
2514 object = tentry->object.vm_object;
2515 offset = tentry->offset + (offset - tentry->start);
2516 vm_map_unlock_read(smap);
2518 object = current->object.vm_object;
2522 vm_object_hold(object);
2525 * Note that there is absolutely no sense in writing out
2526 * anonymous objects, so we track down the vnode object
2528 * We invalidate (remove) all pages from the address space
2529 * anyway, for semantic correctness.
2531 * note: certain anonymous maps, such as MAP_NOSYNC maps,
2532 * may start out with a NULL object.
2534 while (object && (tobj = object->backing_object) != NULL) {
2535 vm_object_hold(tobj);
2536 if (tobj == object->backing_object) {
2537 vm_object_lock_swap();
2538 offset += object->backing_object_offset;
2539 vm_object_drop(object);
2541 if (object->size < OFF_TO_IDX(offset + size))
2542 size = IDX_TO_OFF(object->size) -
2546 vm_object_drop(tobj);
2548 if (object && (object->type == OBJT_VNODE) &&
2549 (current->protection & VM_PROT_WRITE) &&
2550 (object->flags & OBJ_NOMSYNC) == 0) {
2552 * Flush pages if writing is allowed, invalidate them
2553 * if invalidation requested. Pages undergoing I/O
2554 * will be ignored by vm_object_page_remove().
2556 * We cannot lock the vnode and then wait for paging
2557 * to complete without deadlocking against vm_fault.
2558 * Instead we simply call vm_object_page_remove() and
2559 * allow it to block internally on a page-by-page
2560 * basis when it encounters pages undergoing async
2565 /* no chain wait needed for vnode objects */
2566 vm_object_reference_locked(object);
2567 vn_lock(object->handle, LK_EXCLUSIVE | LK_RETRY);
2568 flags = (syncio || invalidate) ? OBJPC_SYNC : 0;
2569 flags |= invalidate ? OBJPC_INVAL : 0;
2572 * When operating on a virtual page table just
2573 * flush the whole object. XXX we probably ought
2576 switch(current->maptype) {
2577 case VM_MAPTYPE_NORMAL:
2578 vm_object_page_clean(object,
2580 OFF_TO_IDX(offset + size + PAGE_MASK),
2583 case VM_MAPTYPE_VPAGETABLE:
2584 vm_object_page_clean(object, 0, 0, flags);
2587 vn_unlock(((struct vnode *)object->handle));
2588 vm_object_deallocate_locked(object);
2590 if (object && invalidate &&
2591 ((object->type == OBJT_VNODE) ||
2592 (object->type == OBJT_DEVICE))) {
2594 (object->type == OBJT_DEVICE) ? FALSE : TRUE;
2595 /* no chain wait needed for vnode/device objects */
2596 vm_object_reference_locked(object);
2597 switch(current->maptype) {
2598 case VM_MAPTYPE_NORMAL:
2599 vm_object_page_remove(object,
2601 OFF_TO_IDX(offset + size + PAGE_MASK),
2604 case VM_MAPTYPE_VPAGETABLE:
2605 vm_object_page_remove(object, 0, 0, clean_only);
2608 vm_object_deallocate_locked(object);
2612 vm_object_drop(object);
2615 lwkt_reltoken(&map->token);
2616 vm_map_unlock_read(map);
2618 return (KERN_SUCCESS);
2622 * Make the region specified by this entry pageable.
2624 * The vm_map must be exclusively locked.
2627 vm_map_entry_unwire(vm_map_t map, vm_map_entry_t entry)
2629 entry->eflags &= ~MAP_ENTRY_USER_WIRED;
2630 entry->wired_count = 0;
2631 vm_fault_unwire(map, entry);
2635 * Deallocate the given entry from the target map.
2637 * The vm_map must be exclusively locked.
2640 vm_map_entry_delete(vm_map_t map, vm_map_entry_t entry, int *countp)
2642 vm_map_entry_unlink(map, entry);
2643 map->size -= entry->end - entry->start;
2645 switch(entry->maptype) {
2646 case VM_MAPTYPE_NORMAL:
2647 case VM_MAPTYPE_VPAGETABLE:
2648 vm_object_deallocate(entry->object.vm_object);
2654 vm_map_entry_dispose(map, entry, countp);
2658 * Deallocates the given address range from the target map.
2660 * The vm_map must be exclusively locked.
2663 vm_map_delete(vm_map_t map, vm_offset_t start, vm_offset_t end, int *countp)
2666 vm_map_entry_t entry;
2667 vm_map_entry_t first_entry;
2669 ASSERT_VM_MAP_LOCKED(map);
2670 lwkt_gettoken(&map->token);
2673 * Find the start of the region, and clip it. Set entry to point
2674 * at the first record containing the requested address or, if no
2675 * such record exists, the next record with a greater address. The
2676 * loop will run from this point until a record beyond the termination
2677 * address is encountered.
2679 * map->hint must be adjusted to not point to anything we delete,
2680 * so set it to the entry prior to the one being deleted.
2682 * GGG see other GGG comment.
2684 if (vm_map_lookup_entry(map, start, &first_entry)) {
2685 entry = first_entry;
2686 vm_map_clip_start(map, entry, start, countp);
2687 map->hint = entry->prev; /* possible problem XXX */
2689 map->hint = first_entry; /* possible problem XXX */
2690 entry = first_entry->next;
2694 * If a hole opens up prior to the current first_free then
2695 * adjust first_free. As with map->hint, map->first_free
2696 * cannot be left set to anything we might delete.
2698 if (entry == &map->header) {
2699 map->first_free = &map->header;
2700 } else if (map->first_free->start >= start) {
2701 map->first_free = entry->prev;
2705 * Step through all entries in this region
2707 while ((entry != &map->header) && (entry->start < end)) {
2708 vm_map_entry_t next;
2710 vm_pindex_t offidxstart, offidxend, count;
2713 * If we hit an in-transition entry we have to sleep and
2714 * retry. It's easier (and not really slower) to just retry
2715 * since this case occurs so rarely and the hint is already
2716 * pointing at the right place. We have to reset the
2717 * start offset so as not to accidently delete an entry
2718 * another process just created in vacated space.
2720 if (entry->eflags & MAP_ENTRY_IN_TRANSITION) {
2721 entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP;
2722 start = entry->start;
2723 ++mycpu->gd_cnt.v_intrans_coll;
2724 ++mycpu->gd_cnt.v_intrans_wait;
2725 vm_map_transition_wait(map);
2728 vm_map_clip_end(map, entry, end, countp);
2734 offidxstart = OFF_TO_IDX(entry->offset);
2735 count = OFF_TO_IDX(e - s);
2736 object = entry->object.vm_object;
2739 * Unwire before removing addresses from the pmap; otherwise,
2740 * unwiring will put the entries back in the pmap.
2742 if (entry->wired_count != 0)
2743 vm_map_entry_unwire(map, entry);
2745 offidxend = offidxstart + count;
2747 if (object == &kernel_object) {
2748 vm_object_hold(object);
2749 vm_object_page_remove(object, offidxstart,
2751 vm_object_drop(object);
2752 } else if (object && object->type != OBJT_DEFAULT &&
2753 object->type != OBJT_SWAP) {
2755 * vnode object routines cannot be chain-locked
2757 vm_object_hold(object);
2758 pmap_remove(map->pmap, s, e);
2759 vm_object_drop(object);
2760 } else if (object) {
2761 vm_object_hold(object);
2762 vm_object_chain_acquire(object);
2763 pmap_remove(map->pmap, s, e);
2765 if (object != NULL &&
2766 object->ref_count != 1 &&
2767 (object->flags & (OBJ_NOSPLIT|OBJ_ONEMAPPING)) ==
2769 (object->type == OBJT_DEFAULT ||
2770 object->type == OBJT_SWAP)) {
2771 vm_object_collapse(object, NULL);
2772 vm_object_page_remove(object, offidxstart,
2774 if (object->type == OBJT_SWAP) {
2775 swap_pager_freespace(object,
2779 if (offidxend >= object->size &&
2780 offidxstart < object->size) {
2781 object->size = offidxstart;
2784 vm_object_chain_release(object);
2785 vm_object_drop(object);
2789 * Delete the entry (which may delete the object) only after
2790 * removing all pmap entries pointing to its pages.
2791 * (Otherwise, its page frames may be reallocated, and any
2792 * modify bits will be set in the wrong object!)
2794 vm_map_entry_delete(map, entry, countp);
2797 lwkt_reltoken(&map->token);
2798 return (KERN_SUCCESS);
2802 * Remove the given address range from the target map.
2803 * This is the exported form of vm_map_delete.
2808 vm_map_remove(vm_map_t map, vm_offset_t start, vm_offset_t end)
2813 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
2815 VM_MAP_RANGE_CHECK(map, start, end);
2816 result = vm_map_delete(map, start, end, &count);
2818 vm_map_entry_release(count);
2824 * Assert that the target map allows the specified privilege on the
2825 * entire address region given. The entire region must be allocated.
2827 * The caller must specify whether the vm_map is already locked or not.
2830 vm_map_check_protection(vm_map_t map, vm_offset_t start, vm_offset_t end,
2831 vm_prot_t protection, boolean_t have_lock)
2833 vm_map_entry_t entry;
2834 vm_map_entry_t tmp_entry;
2837 if (have_lock == FALSE)
2838 vm_map_lock_read(map);
2840 if (!vm_map_lookup_entry(map, start, &tmp_entry)) {
2841 if (have_lock == FALSE)
2842 vm_map_unlock_read(map);
2848 while (start < end) {
2849 if (entry == &map->header) {
2857 if (start < entry->start) {
2862 * Check protection associated with entry.
2865 if ((entry->protection & protection) != protection) {
2869 /* go to next entry */
2872 entry = entry->next;
2874 if (have_lock == FALSE)
2875 vm_map_unlock_read(map);
2880 * If appropriate this function shadows the original object with a new object
2881 * and moves the VM pages from the original object to the new object.
2882 * The original object will also be collapsed, if possible.
2884 * We can only do this for normal memory objects with a single mapping, and
2885 * it only makes sense to do it if there are 2 or more refs on the original
2886 * object. i.e. typically a memory object that has been extended into
2887 * multiple vm_map_entry's with non-overlapping ranges.
2889 * This makes it easier to remove unused pages and keeps object inheritance
2890 * from being a negative impact on memory usage.
2892 * On return the (possibly new) entry->object.vm_object will have an
2893 * additional ref on it for the caller to dispose of (usually by cloning
2894 * the vm_map_entry). The additional ref had to be done in this routine
2895 * to avoid racing a collapse. The object's ONEMAPPING flag will also be
2898 * The vm_map must be locked and its token held.
2901 vm_map_split(vm_map_entry_t entry)
2905 vm_object_t oobject;
2907 oobject = entry->object.vm_object;
2908 vm_object_hold(oobject);
2909 vm_object_chain_wait(oobject);
2910 vm_object_reference_locked(oobject);
2911 vm_object_clear_flag(oobject, OBJ_ONEMAPPING);
2912 vm_object_drop(oobject);
2915 vm_object_t oobject, nobject, bobject;
2918 vm_pindex_t offidxstart, offidxend, idx;
2920 vm_ooffset_t offset;
2923 * Setup. Chain lock the original object throughout the entire
2924 * routine to prevent new page faults from occuring.
2926 * XXX can madvise WILLNEED interfere with us too?
2928 oobject = entry->object.vm_object;
2929 vm_object_hold(oobject);
2930 vm_object_chain_acquire(oobject);
2933 * Original object cannot be split?
2935 if (oobject->handle == NULL || (oobject->type != OBJT_DEFAULT &&
2936 oobject->type != OBJT_SWAP)) {
2937 vm_object_chain_release(oobject);
2938 vm_object_reference_locked(oobject);
2939 vm_object_clear_flag(oobject, OBJ_ONEMAPPING);
2940 vm_object_drop(oobject);
2945 * Collapse original object with its backing store as an
2946 * optimization to reduce chain lengths when possible.
2948 * If ref_count <= 1 there aren't other non-overlapping vm_map_entry's
2949 * for oobject, so there's no point collapsing it.
2951 * Then re-check whether the object can be split.
2953 vm_object_collapse(oobject, NULL);
2955 if (oobject->ref_count <= 1 ||
2956 (oobject->type != OBJT_DEFAULT && oobject->type != OBJT_SWAP) ||
2957 (oobject->flags & (OBJ_NOSPLIT|OBJ_ONEMAPPING)) != OBJ_ONEMAPPING) {
2958 vm_object_chain_release(oobject);
2959 vm_object_reference_locked(oobject);
2960 vm_object_clear_flag(oobject, OBJ_ONEMAPPING);
2961 vm_object_drop(oobject);
2966 * Acquire the chain lock on the backing object.
2968 * Give bobject an additional ref count for when it will be shadowed
2971 if ((bobject = oobject->backing_object) != NULL) {
2972 vm_object_hold(bobject);
2973 vm_object_chain_wait(bobject);
2974 vm_object_reference_locked(bobject);
2975 vm_object_chain_acquire(bobject);
2976 KKASSERT(bobject->backing_object == bobject);
2977 KKASSERT((bobject->flags & OBJ_DEAD) == 0);
2981 * Calculate the object page range and allocate the new object.
2983 offset = entry->offset;
2987 offidxstart = OFF_TO_IDX(offset);
2988 offidxend = offidxstart + OFF_TO_IDX(e - s);
2989 size = offidxend - offidxstart;
2991 switch(oobject->type) {
2993 nobject = default_pager_alloc(NULL, IDX_TO_OFF(size),
2997 nobject = swap_pager_alloc(NULL, IDX_TO_OFF(size),
3006 if (nobject == NULL) {
3008 vm_object_chain_release(bobject);
3009 vm_object_deallocate(bobject);
3010 vm_object_drop(bobject);
3012 vm_object_chain_release(oobject);
3013 vm_object_reference_locked(oobject);
3014 vm_object_clear_flag(oobject, OBJ_ONEMAPPING);
3015 vm_object_drop(oobject);
3020 * The new object will replace entry->object.vm_object so it needs
3021 * a second reference (the caller expects an additional ref).
3023 vm_object_hold(nobject);
3024 vm_object_reference_locked(nobject);
3025 vm_object_chain_acquire(nobject);
3028 * nobject shadows bobject (oobject already shadows bobject).
3031 nobject->backing_object_offset =
3032 oobject->backing_object_offset + IDX_TO_OFF(offidxstart);
3033 nobject->backing_object = bobject;
3034 bobject->shadow_count++;
3035 bobject->generation++;
3036 LIST_INSERT_HEAD(&bobject->shadow_head, nobject, shadow_list);
3037 vm_object_clear_flag(bobject, OBJ_ONEMAPPING); /* XXX? */
3038 vm_object_chain_release(bobject);
3039 vm_object_drop(bobject);
3043 * Move the VM pages from oobject to nobject
3045 for (idx = 0; idx < size; idx++) {
3048 m = vm_page_lookup_busy_wait(oobject, offidxstart + idx,
3054 * We must wait for pending I/O to complete before we can
3057 * We do not have to VM_PROT_NONE the page as mappings should
3058 * not be changed by this operation.
3060 * NOTE: The act of renaming a page updates chaingen for both
3063 vm_page_rename(m, nobject, idx);
3064 /* page automatically made dirty by rename and cache handled */
3065 /* page remains busy */
3068 if (oobject->type == OBJT_SWAP) {
3069 vm_object_pip_add(oobject, 1);
3071 * copy oobject pages into nobject and destroy unneeded
3072 * pages in shadow object.
3074 swap_pager_copy(oobject, nobject, offidxstart, 0);
3075 vm_object_pip_wakeup(oobject);
3079 * Wakeup the pages we played with. No spl protection is needed
3080 * for a simple wakeup.
3082 for (idx = 0; idx < size; idx++) {
3083 m = vm_page_lookup(nobject, idx);
3085 KKASSERT(m->flags & PG_BUSY);
3089 entry->object.vm_object = nobject;
3090 entry->offset = 0LL;
3095 * NOTE: There is no need to remove OBJ_ONEMAPPING from oobject, the
3096 * related pages were moved and are no longer applicable to the
3099 * NOTE: Deallocate oobject (due to its entry->object.vm_object being
3100 * replaced by nobject).
3102 vm_object_chain_release(nobject);
3103 vm_object_drop(nobject);
3105 vm_object_chain_release(bobject);
3106 vm_object_drop(bobject);
3108 vm_object_chain_release(oobject);
3109 /*vm_object_clear_flag(oobject, OBJ_ONEMAPPING);*/
3110 vm_object_deallocate_locked(oobject);
3111 vm_object_drop(oobject);
3116 * Copies the contents of the source entry to the destination
3117 * entry. The entries *must* be aligned properly.
3119 * The vm_maps must be exclusively locked.
3120 * The vm_map's token must be held.
3122 * Because the maps are locked no faults can be in progress during the
3126 vm_map_copy_entry(vm_map_t src_map, vm_map_t dst_map,
3127 vm_map_entry_t src_entry, vm_map_entry_t dst_entry)
3129 vm_object_t src_object;
3131 if (dst_entry->maptype == VM_MAPTYPE_SUBMAP)
3133 if (src_entry->maptype == VM_MAPTYPE_SUBMAP)
3136 if (src_entry->wired_count == 0) {
3138 * If the source entry is marked needs_copy, it is already
3141 if ((src_entry->eflags & MAP_ENTRY_NEEDS_COPY) == 0) {
3142 pmap_protect(src_map->pmap,
3145 src_entry->protection & ~VM_PROT_WRITE);
3149 * Make a copy of the object.
3151 * The object must be locked prior to checking the object type
3152 * and for the call to vm_object_collapse() and vm_map_split().
3153 * We cannot use *_hold() here because the split code will
3154 * probably try to destroy the object. The lock is a pool
3155 * token and doesn't care.
3157 if (src_entry->object.vm_object != NULL) {
3158 vm_map_split(src_entry);
3159 src_object = src_entry->object.vm_object;
3160 dst_entry->object.vm_object = src_object;
3161 src_entry->eflags |= (MAP_ENTRY_COW |
3162 MAP_ENTRY_NEEDS_COPY);
3163 dst_entry->eflags |= (MAP_ENTRY_COW |
3164 MAP_ENTRY_NEEDS_COPY);
3165 dst_entry->offset = src_entry->offset;
3167 dst_entry->object.vm_object = NULL;
3168 dst_entry->offset = 0;
3171 pmap_copy(dst_map->pmap, src_map->pmap, dst_entry->start,
3172 dst_entry->end - dst_entry->start, src_entry->start);
3175 * Of course, wired down pages can't be set copy-on-write.
3176 * Cause wired pages to be copied into the new map by
3177 * simulating faults (the new pages are pageable)
3179 vm_fault_copy_entry(dst_map, src_map, dst_entry, src_entry);
3185 * Create a new process vmspace structure and vm_map
3186 * based on those of an existing process. The new map
3187 * is based on the old map, according to the inheritance
3188 * values on the regions in that map.
3190 * The source map must not be locked.
3194 vmspace_fork(struct vmspace *vm1)
3196 struct vmspace *vm2;
3197 vm_map_t old_map = &vm1->vm_map;
3199 vm_map_entry_t old_entry;
3200 vm_map_entry_t new_entry;
3204 lwkt_gettoken(&vm1->vm_map.token);
3205 vm_map_lock(old_map);
3208 * XXX Note: upcalls are not copied.
3210 vm2 = vmspace_alloc(old_map->min_offset, old_map->max_offset);
3211 lwkt_gettoken(&vm2->vm_map.token);
3212 bcopy(&vm1->vm_startcopy, &vm2->vm_startcopy,
3213 (caddr_t)&vm1->vm_endcopy - (caddr_t)&vm1->vm_startcopy);
3214 new_map = &vm2->vm_map; /* XXX */
3215 new_map->timestamp = 1;
3217 vm_map_lock(new_map);
3220 old_entry = old_map->header.next;
3221 while (old_entry != &old_map->header) {
3223 old_entry = old_entry->next;
3226 count = vm_map_entry_reserve(count + MAP_RESERVE_COUNT);
3228 old_entry = old_map->header.next;
3229 while (old_entry != &old_map->header) {
3230 if (old_entry->maptype == VM_MAPTYPE_SUBMAP)
3231 panic("vm_map_fork: encountered a submap");
3233 switch (old_entry->inheritance) {
3234 case VM_INHERIT_NONE:
3236 case VM_INHERIT_SHARE:
3238 * Clone the entry, creating the shared object if
3241 if (old_entry->object.vm_object == NULL)
3242 vm_map_entry_allocate_object(old_entry);
3244 if (old_entry->eflags & MAP_ENTRY_NEEDS_COPY) {
3246 * Shadow a map_entry which needs a copy,
3247 * replacing its object with a new object
3248 * that points to the old one. Ask the
3249 * shadow code to automatically add an
3250 * additional ref. We can't do it afterwords
3251 * because we might race a collapse. The call
3252 * to vm_map_entry_shadow() will also clear
3255 vm_map_entry_shadow(old_entry, 1);
3258 * We will make a shared copy of the object,
3259 * and must clear OBJ_ONEMAPPING.
3261 * XXX assert that object.vm_object != NULL
3262 * since we allocate it above.
3264 if (old_entry->object.vm_object) {
3265 object = old_entry->object.vm_object;
3266 vm_object_hold(object);
3267 vm_object_chain_wait(object);
3268 vm_object_reference_locked(object);
3269 vm_object_clear_flag(object,
3271 vm_object_drop(object);
3276 * Clone the entry. We've already bumped the ref on
3279 new_entry = vm_map_entry_create(new_map, &count);
3280 *new_entry = *old_entry;
3281 new_entry->eflags &= ~MAP_ENTRY_USER_WIRED;
3282 new_entry->wired_count = 0;
3285 * Insert the entry into the new map -- we know we're
3286 * inserting at the end of the new map.
3289 vm_map_entry_link(new_map, new_map->header.prev,
3293 * Update the physical map
3295 pmap_copy(new_map->pmap, old_map->pmap,
3297 (old_entry->end - old_entry->start),
3300 case VM_INHERIT_COPY:
3302 * Clone the entry and link into the map.
3304 new_entry = vm_map_entry_create(new_map, &count);
3305 *new_entry = *old_entry;
3306 new_entry->eflags &= ~MAP_ENTRY_USER_WIRED;
3307 new_entry->wired_count = 0;
3308 new_entry->object.vm_object = NULL;
3309 vm_map_entry_link(new_map, new_map->header.prev,
3311 vm_map_copy_entry(old_map, new_map, old_entry,
3315 old_entry = old_entry->next;
3318 new_map->size = old_map->size;
3319 vm_map_unlock(old_map);
3320 vm_map_unlock(new_map);
3321 vm_map_entry_release(count);
3323 lwkt_reltoken(&vm2->vm_map.token);
3324 lwkt_reltoken(&vm1->vm_map.token);
3330 * Create an auto-grow stack entry
3335 vm_map_stack (vm_map_t map, vm_offset_t addrbos, vm_size_t max_ssize,
3336 int flags, vm_prot_t prot, vm_prot_t max, int cow)
3338 vm_map_entry_t prev_entry;
3339 vm_map_entry_t new_stack_entry;
3340 vm_size_t init_ssize;
3343 vm_offset_t tmpaddr;
3345 cow |= MAP_IS_STACK;
3347 if (max_ssize < sgrowsiz)
3348 init_ssize = max_ssize;
3350 init_ssize = sgrowsiz;
3352 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
3356 * Find space for the mapping
3358 if ((flags & (MAP_FIXED | MAP_TRYFIXED)) == 0) {
3359 if (vm_map_findspace(map, addrbos, max_ssize, 1,
3362 vm_map_entry_release(count);
3363 return (KERN_NO_SPACE);
3368 /* If addr is already mapped, no go */
3369 if (vm_map_lookup_entry(map, addrbos, &prev_entry)) {
3371 vm_map_entry_release(count);
3372 return (KERN_NO_SPACE);
3376 /* XXX already handled by kern_mmap() */
3377 /* If we would blow our VMEM resource limit, no go */
3378 if (map->size + init_ssize >
3379 curproc->p_rlimit[RLIMIT_VMEM].rlim_cur) {
3381 vm_map_entry_release(count);
3382 return (KERN_NO_SPACE);
3387 * If we can't accomodate max_ssize in the current mapping,
3388 * no go. However, we need to be aware that subsequent user
3389 * mappings might map into the space we have reserved for
3390 * stack, and currently this space is not protected.
3392 * Hopefully we will at least detect this condition
3393 * when we try to grow the stack.
3395 if ((prev_entry->next != &map->header) &&
3396 (prev_entry->next->start < addrbos + max_ssize)) {
3398 vm_map_entry_release(count);
3399 return (KERN_NO_SPACE);
3403 * We initially map a stack of only init_ssize. We will
3404 * grow as needed later. Since this is to be a grow
3405 * down stack, we map at the top of the range.
3407 * Note: we would normally expect prot and max to be
3408 * VM_PROT_ALL, and cow to be 0. Possibly we should
3409 * eliminate these as input parameters, and just
3410 * pass these values here in the insert call.
3412 rv = vm_map_insert(map, &count,
3413 NULL, 0, addrbos + max_ssize - init_ssize,
3414 addrbos + max_ssize,
3419 /* Now set the avail_ssize amount */
3420 if (rv == KERN_SUCCESS) {
3421 if (prev_entry != &map->header)
3422 vm_map_clip_end(map, prev_entry, addrbos + max_ssize - init_ssize, &count);
3423 new_stack_entry = prev_entry->next;
3424 if (new_stack_entry->end != addrbos + max_ssize ||
3425 new_stack_entry->start != addrbos + max_ssize - init_ssize)
3426 panic ("Bad entry start/end for new stack entry");
3428 new_stack_entry->aux.avail_ssize = max_ssize - init_ssize;
3432 vm_map_entry_release(count);
3437 * Attempts to grow a vm stack entry. Returns KERN_SUCCESS if the
3438 * desired address is already mapped, or if we successfully grow
3439 * the stack. Also returns KERN_SUCCESS if addr is outside the
3440 * stack range (this is strange, but preserves compatibility with
3441 * the grow function in vm_machdep.c).
3446 vm_map_growstack (struct proc *p, vm_offset_t addr)
3448 vm_map_entry_t prev_entry;
3449 vm_map_entry_t stack_entry;
3450 vm_map_entry_t new_stack_entry;
3451 struct vmspace *vm = p->p_vmspace;
3452 vm_map_t map = &vm->vm_map;
3455 int rv = KERN_SUCCESS;
3457 int use_read_lock = 1;
3460 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
3463 vm_map_lock_read(map);
3467 /* If addr is already in the entry range, no need to grow.*/
3468 if (vm_map_lookup_entry(map, addr, &prev_entry))
3471 if ((stack_entry = prev_entry->next) == &map->header)
3473 if (prev_entry == &map->header)
3474 end = stack_entry->start - stack_entry->aux.avail_ssize;
3476 end = prev_entry->end;
3479 * This next test mimics the old grow function in vm_machdep.c.
3480 * It really doesn't quite make sense, but we do it anyway
3481 * for compatibility.
3483 * If not growable stack, return success. This signals the
3484 * caller to proceed as he would normally with normal vm.
3486 if (stack_entry->aux.avail_ssize < 1 ||
3487 addr >= stack_entry->start ||
3488 addr < stack_entry->start - stack_entry->aux.avail_ssize) {
3492 /* Find the minimum grow amount */
3493 grow_amount = roundup (stack_entry->start - addr, PAGE_SIZE);
3494 if (grow_amount > stack_entry->aux.avail_ssize) {
3500 * If there is no longer enough space between the entries
3501 * nogo, and adjust the available space. Note: this
3502 * should only happen if the user has mapped into the
3503 * stack area after the stack was created, and is
3504 * probably an error.
3506 * This also effectively destroys any guard page the user
3507 * might have intended by limiting the stack size.
3509 if (grow_amount > stack_entry->start - end) {
3510 if (use_read_lock && vm_map_lock_upgrade(map)) {
3516 stack_entry->aux.avail_ssize = stack_entry->start - end;
3521 is_procstack = addr >= (vm_offset_t)vm->vm_maxsaddr;
3523 /* If this is the main process stack, see if we're over the
3526 if (is_procstack && (ctob(vm->vm_ssize) + grow_amount >
3527 p->p_rlimit[RLIMIT_STACK].rlim_cur)) {
3532 /* Round up the grow amount modulo SGROWSIZ */
3533 grow_amount = roundup (grow_amount, sgrowsiz);
3534 if (grow_amount > stack_entry->aux.avail_ssize) {
3535 grow_amount = stack_entry->aux.avail_ssize;
3537 if (is_procstack && (ctob(vm->vm_ssize) + grow_amount >
3538 p->p_rlimit[RLIMIT_STACK].rlim_cur)) {
3539 grow_amount = p->p_rlimit[RLIMIT_STACK].rlim_cur -
3543 /* If we would blow our VMEM resource limit, no go */
3544 if (map->size + grow_amount > p->p_rlimit[RLIMIT_VMEM].rlim_cur) {
3549 if (use_read_lock && vm_map_lock_upgrade(map)) {
3556 /* Get the preliminary new entry start value */
3557 addr = stack_entry->start - grow_amount;
3559 /* If this puts us into the previous entry, cut back our growth
3560 * to the available space. Also, see the note above.
3563 stack_entry->aux.avail_ssize = stack_entry->start - end;
3567 rv = vm_map_insert(map, &count,
3568 NULL, 0, addr, stack_entry->start,
3570 VM_PROT_ALL, VM_PROT_ALL,
3573 /* Adjust the available stack space by the amount we grew. */
3574 if (rv == KERN_SUCCESS) {
3575 if (prev_entry != &map->header)
3576 vm_map_clip_end(map, prev_entry, addr, &count);
3577 new_stack_entry = prev_entry->next;
3578 if (new_stack_entry->end != stack_entry->start ||
3579 new_stack_entry->start != addr)
3580 panic ("Bad stack grow start/end in new stack entry");
3582 new_stack_entry->aux.avail_ssize =
3583 stack_entry->aux.avail_ssize -
3584 (new_stack_entry->end - new_stack_entry->start);
3586 vm->vm_ssize += btoc(new_stack_entry->end -
3587 new_stack_entry->start);
3590 if (map->flags & MAP_WIREFUTURE)
3591 vm_map_unwire(map, new_stack_entry->start,
3592 new_stack_entry->end, FALSE);
3597 vm_map_unlock_read(map);
3600 vm_map_entry_release(count);
3605 * Unshare the specified VM space for exec. If other processes are
3606 * mapped to it, then create a new one. The new vmspace is null.
3611 vmspace_exec(struct proc *p, struct vmspace *vmcopy)
3613 struct vmspace *oldvmspace = p->p_vmspace;
3614 struct vmspace *newvmspace;
3615 vm_map_t map = &p->p_vmspace->vm_map;
3618 * If we are execing a resident vmspace we fork it, otherwise
3619 * we create a new vmspace. Note that exitingcnt and upcalls
3620 * are not copied to the new vmspace.
3622 lwkt_gettoken(&oldvmspace->vm_map.token);
3624 newvmspace = vmspace_fork(vmcopy);
3625 lwkt_gettoken(&newvmspace->vm_map.token);
3627 newvmspace = vmspace_alloc(map->min_offset, map->max_offset);
3628 lwkt_gettoken(&newvmspace->vm_map.token);
3629 bcopy(&oldvmspace->vm_startcopy, &newvmspace->vm_startcopy,
3630 (caddr_t)&oldvmspace->vm_endcopy -
3631 (caddr_t)&oldvmspace->vm_startcopy);
3635 * Finish initializing the vmspace before assigning it
3636 * to the process. The vmspace will become the current vmspace
3639 pmap_pinit2(vmspace_pmap(newvmspace));
3640 pmap_replacevm(p, newvmspace, 0);
3641 lwkt_reltoken(&newvmspace->vm_map.token);
3642 lwkt_reltoken(&oldvmspace->vm_map.token);
3643 sysref_put(&oldvmspace->vm_sysref);
3647 * Unshare the specified VM space for forcing COW. This
3648 * is called by rfork, for the (RFMEM|RFPROC) == 0 case.
3650 * The exitingcnt test is not strictly necessary but has been
3651 * included for code sanity (to make the code a bit more deterministic).
3654 vmspace_unshare(struct proc *p)
3656 struct vmspace *oldvmspace = p->p_vmspace;
3657 struct vmspace *newvmspace;
3659 lwkt_gettoken(&oldvmspace->vm_map.token);
3660 if (oldvmspace->vm_sysref.refcnt == 1 && oldvmspace->vm_exitingcnt == 0) {
3661 lwkt_reltoken(&oldvmspace->vm_map.token);
3664 newvmspace = vmspace_fork(oldvmspace);
3665 lwkt_gettoken(&newvmspace->vm_map.token);
3666 pmap_pinit2(vmspace_pmap(newvmspace));
3667 pmap_replacevm(p, newvmspace, 0);
3668 lwkt_reltoken(&newvmspace->vm_map.token);
3669 lwkt_reltoken(&oldvmspace->vm_map.token);
3670 sysref_put(&oldvmspace->vm_sysref);
3674 * vm_map_hint: return the beginning of the best area suitable for
3675 * creating a new mapping with "prot" protection.
3680 vm_map_hint(struct proc *p, vm_offset_t addr, vm_prot_t prot)
3682 struct vmspace *vms = p->p_vmspace;
3684 if (!randomize_mmap) {
3686 * Set a reasonable start point for the hint if it was
3687 * not specified or if it falls within the heap space.
3688 * Hinted mmap()s do not allocate out of the heap space.
3691 (addr >= round_page((vm_offset_t)vms->vm_taddr) &&
3692 addr < round_page((vm_offset_t)vms->vm_daddr + maxdsiz))) {
3693 addr = round_page((vm_offset_t)vms->vm_daddr + maxdsiz);
3699 if (addr != 0 && addr >= (vm_offset_t)vms->vm_daddr)
3705 * If executable skip first two pages, otherwise start
3706 * after data + heap region.
3708 if ((prot & VM_PROT_EXECUTE) &&
3709 ((vm_offset_t)vms->vm_daddr >= I386_MAX_EXE_ADDR)) {
3710 addr = (PAGE_SIZE * 2) +
3711 (karc4random() & (I386_MAX_EXE_ADDR / 2 - 1));
3712 return (round_page(addr));
3714 #endif /* __i386__ */
3717 addr = (vm_offset_t)vms->vm_daddr + MAXDSIZ;
3718 addr += karc4random() & (MIN((256 * 1024 * 1024), MAXDSIZ) - 1);
3720 return (round_page(addr));
3724 * Finds the VM object, offset, and protection for a given virtual address
3725 * in the specified map, assuming a page fault of the type specified.
3727 * Leaves the map in question locked for read; return values are guaranteed
3728 * until a vm_map_lookup_done call is performed. Note that the map argument
3729 * is in/out; the returned map must be used in the call to vm_map_lookup_done.
3731 * A handle (out_entry) is returned for use in vm_map_lookup_done, to make
3734 * If a lookup is requested with "write protection" specified, the map may
3735 * be changed to perform virtual copying operations, although the data
3736 * referenced will remain the same.
3741 vm_map_lookup(vm_map_t *var_map, /* IN/OUT */
3743 vm_prot_t fault_typea,
3744 vm_map_entry_t *out_entry, /* OUT */
3745 vm_object_t *object, /* OUT */
3746 vm_pindex_t *pindex, /* OUT */
3747 vm_prot_t *out_prot, /* OUT */
3748 boolean_t *wired) /* OUT */
3750 vm_map_entry_t entry;
3751 vm_map_t map = *var_map;
3753 vm_prot_t fault_type = fault_typea;
3754 int use_read_lock = 1;
3755 int rv = KERN_SUCCESS;
3759 vm_map_lock_read(map);
3764 * If the map has an interesting hint, try it before calling full
3765 * blown lookup routine.
3772 if ((entry == &map->header) ||
3773 (vaddr < entry->start) || (vaddr >= entry->end)) {
3774 vm_map_entry_t tmp_entry;
3777 * Entry was either not a valid hint, or the vaddr was not
3778 * contained in the entry, so do a full lookup.
3780 if (!vm_map_lookup_entry(map, vaddr, &tmp_entry)) {
3781 rv = KERN_INVALID_ADDRESS;
3792 if (entry->maptype == VM_MAPTYPE_SUBMAP) {
3793 vm_map_t old_map = map;
3795 *var_map = map = entry->object.sub_map;
3797 vm_map_unlock_read(old_map);
3799 vm_map_unlock(old_map);
3805 * Check whether this task is allowed to have this page.
3806 * Note the special case for MAP_ENTRY_COW
3807 * pages with an override. This is to implement a forced
3808 * COW for debuggers.
3811 if (fault_type & VM_PROT_OVERRIDE_WRITE)
3812 prot = entry->max_protection;
3814 prot = entry->protection;
3816 fault_type &= (VM_PROT_READ|VM_PROT_WRITE|VM_PROT_EXECUTE);
3817 if ((fault_type & prot) != fault_type) {
3818 rv = KERN_PROTECTION_FAILURE;
3822 if ((entry->eflags & MAP_ENTRY_USER_WIRED) &&
3823 (entry->eflags & MAP_ENTRY_COW) &&
3824 (fault_type & VM_PROT_WRITE) &&
3825 (fault_typea & VM_PROT_OVERRIDE_WRITE) == 0) {
3826 rv = KERN_PROTECTION_FAILURE;
3831 * If this page is not pageable, we have to get it for all possible
3834 *wired = (entry->wired_count != 0);
3836 prot = fault_type = entry->protection;
3839 * Virtual page tables may need to update the accessed (A) bit
3840 * in a page table entry. Upgrade the fault to a write fault for
3841 * that case if the map will support it. If the map does not support
3842 * it the page table entry simply will not be updated.
3844 if (entry->maptype == VM_MAPTYPE_VPAGETABLE) {
3845 if (prot & VM_PROT_WRITE)
3846 fault_type |= VM_PROT_WRITE;
3850 * If the entry was copy-on-write, we either ...
3852 if (entry->eflags & MAP_ENTRY_NEEDS_COPY) {
3854 * If we want to write the page, we may as well handle that
3855 * now since we've got the map locked.
3857 * If we don't need to write the page, we just demote the
3858 * permissions allowed.
3861 if (fault_type & VM_PROT_WRITE) {
3863 * Make a new object, and place it in the object
3864 * chain. Note that no new references have appeared
3865 * -- one just moved from the map to the new
3869 if (use_read_lock && vm_map_lock_upgrade(map)) {
3876 vm_map_entry_shadow(entry, 0);
3879 * We're attempting to read a copy-on-write page --
3880 * don't allow writes.
3883 prot &= ~VM_PROT_WRITE;
3888 * Create an object if necessary.
3890 if (entry->object.vm_object == NULL && !map->system_map) {
3891 if (use_read_lock && vm_map_lock_upgrade(map)) {
3897 vm_map_entry_allocate_object(entry);
3901 * Return the object/offset from this entry. If the entry was
3902 * copy-on-write or empty, it has been fixed up.
3905 *pindex = OFF_TO_IDX((vaddr - entry->start) + entry->offset);
3906 *object = entry->object.vm_object;
3909 * Return whether this is the only map sharing this data. On
3910 * success we return with a read lock held on the map. On failure
3911 * we return with the map unlocked.
3915 if (rv == KERN_SUCCESS) {
3916 if (use_read_lock == 0)
3917 vm_map_lock_downgrade(map);
3918 } else if (use_read_lock) {
3919 vm_map_unlock_read(map);
3927 * Releases locks acquired by a vm_map_lookup()
3928 * (according to the handle returned by that lookup).
3930 * No other requirements.
3933 vm_map_lookup_done(vm_map_t map, vm_map_entry_t entry, int count)
3936 * Unlock the main-level map
3938 vm_map_unlock_read(map);
3940 vm_map_entry_release(count);
3943 #include "opt_ddb.h"
3945 #include <sys/kernel.h>
3947 #include <ddb/ddb.h>
3952 DB_SHOW_COMMAND(map, vm_map_print)
3955 /* XXX convert args. */
3956 vm_map_t map = (vm_map_t)addr;
3957 boolean_t full = have_addr;
3959 vm_map_entry_t entry;
3961 db_iprintf("Task map %p: pmap=%p, nentries=%d, version=%u\n",
3963 (void *)map->pmap, map->nentries, map->timestamp);
3966 if (!full && db_indent)
3970 for (entry = map->header.next; entry != &map->header;
3971 entry = entry->next) {
3972 db_iprintf("map entry %p: start=%p, end=%p\n",
3973 (void *)entry, (void *)entry->start, (void *)entry->end);
3976 static char *inheritance_name[4] =
3977 {"share", "copy", "none", "donate_copy"};
3979 db_iprintf(" prot=%x/%x/%s",
3981 entry->max_protection,
3982 inheritance_name[(int)(unsigned char)entry->inheritance]);
3983 if (entry->wired_count != 0)
3984 db_printf(", wired");
3986 if (entry->maptype == VM_MAPTYPE_SUBMAP) {
3987 /* XXX no %qd in kernel. Truncate entry->offset. */
3988 db_printf(", share=%p, offset=0x%lx\n",
3989 (void *)entry->object.sub_map,
3990 (long)entry->offset);
3992 if ((entry->prev == &map->header) ||
3993 (entry->prev->object.sub_map !=
3994 entry->object.sub_map)) {
3996 vm_map_print((db_expr_t)(intptr_t)
3997 entry->object.sub_map,
4002 /* XXX no %qd in kernel. Truncate entry->offset. */
4003 db_printf(", object=%p, offset=0x%lx",
4004 (void *)entry->object.vm_object,
4005 (long)entry->offset);
4006 if (entry->eflags & MAP_ENTRY_COW)
4007 db_printf(", copy (%s)",
4008 (entry->eflags & MAP_ENTRY_NEEDS_COPY) ? "needed" : "done");
4012 if ((entry->prev == &map->header) ||
4013 (entry->prev->object.vm_object !=
4014 entry->object.vm_object)) {
4016 vm_object_print((db_expr_t)(intptr_t)
4017 entry->object.vm_object,
4032 DB_SHOW_COMMAND(procvm, procvm)
4037 p = (struct proc *) addr;
4042 db_printf("p = %p, vmspace = %p, map = %p, pmap = %p\n",
4043 (void *)p, (void *)p->p_vmspace, (void *)&p->p_vmspace->vm_map,
4044 (void *)vmspace_pmap(p->p_vmspace));
4046 vm_map_print((db_expr_t)(intptr_t)&p->p_vmspace->vm_map, 1, 0, NULL);