4 * Copyright (c) 1991, 1993
5 * The Regents of the University of California. All rights reserved.
7 * This code is derived from software contributed to Berkeley by
8 * The Mach Operating System project at Carnegie-Mellon University.
10 * Redistribution and use in source and binary forms, with or without
11 * modification, are permitted provided that the following conditions
13 * 1. Redistributions of source code must retain the above copyright
14 * notice, this list of conditions and the following disclaimer.
15 * 2. Redistributions in binary form must reproduce the above copyright
16 * notice, this list of conditions and the following disclaimer in the
17 * documentation and/or other materials provided with the distribution.
18 * 3. Neither the name of the University nor the names of its contributors
19 * may be used to endorse or promote products derived from this software
20 * without specific prior written permission.
22 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
23 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
24 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
25 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
26 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
27 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
28 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
29 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
30 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
31 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
34 * from: @(#)vm_map.c 8.3 (Berkeley) 1/12/94
37 * Copyright (c) 1987, 1990 Carnegie-Mellon University.
38 * All rights reserved.
40 * Authors: Avadis Tevanian, Jr., Michael Wayne Young
42 * Permission to use, copy, modify and distribute this software and
43 * its documentation is hereby granted, provided that both the copyright
44 * notice and this permission notice appear in all copies of the
45 * software, derivative works or modified versions, and any portions
46 * thereof, and that both notices appear in supporting documentation.
48 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
49 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
50 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
52 * Carnegie Mellon requests users of this software to return to
54 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU
55 * School of Computer Science
56 * Carnegie Mellon University
57 * Pittsburgh PA 15213-3890
59 * any improvements or extensions that they make and grant Carnegie the
60 * rights to redistribute these changes.
62 * $FreeBSD: src/sys/vm/vm_map.c,v 1.187.2.19 2003/05/27 00:47:02 alc Exp $
66 * Virtual memory mapping module.
69 #include <sys/param.h>
70 #include <sys/systm.h>
71 #include <sys/kernel.h>
73 #include <sys/serialize.h>
75 #include <sys/vmmeter.h>
77 #include <sys/vnode.h>
78 #include <sys/resourcevar.h>
81 #include <sys/malloc.h>
82 #include <sys/objcache.h>
85 #include <vm/vm_param.h>
87 #include <vm/vm_map.h>
88 #include <vm/vm_page.h>
89 #include <vm/vm_object.h>
90 #include <vm/vm_pager.h>
91 #include <vm/vm_kern.h>
92 #include <vm/vm_extern.h>
93 #include <vm/swap_pager.h>
94 #include <vm/vm_zone.h>
96 #include <sys/thread2.h>
97 #include <sys/random.h>
98 #include <sys/sysctl.h>
101 * Virtual memory maps provide for the mapping, protection, and sharing
102 * of virtual memory objects. In addition, this module provides for an
103 * efficient virtual copy of memory from one map to another.
105 * Synchronization is required prior to most operations.
107 * Maps consist of an ordered doubly-linked list of simple entries.
108 * A hint and a RB tree is used to speed-up lookups.
110 * Callers looking to modify maps specify start/end addresses which cause
111 * the related map entry to be clipped if necessary, and then later
112 * recombined if the pieces remained compatible.
114 * Virtual copy operations are performed by copying VM object references
115 * from one map to another, and then marking both regions as copy-on-write.
117 static __boolean_t vmspace_ctor(void *obj, void *privdata, int ocflags);
118 static void vmspace_dtor(void *obj, void *privdata);
119 static void vmspace_terminate(struct vmspace *vm, int final);
121 MALLOC_DEFINE(M_VMSPACE, "vmspace", "vmspace objcache backingstore");
122 static struct objcache *vmspace_cache;
125 * per-cpu page table cross mappings are initialized in early boot
126 * and might require a considerable number of vm_map_entry structures.
128 #define MAPENTRYBSP_CACHE (MAXCPU+1)
129 #define MAPENTRYAP_CACHE 8
131 static struct vm_zone mapentzone_store, mapzone_store;
132 static vm_zone_t mapentzone, mapzone;
133 static struct vm_object mapentobj, mapobj;
135 static struct vm_map_entry map_entry_init[MAX_MAPENT];
136 static struct vm_map_entry cpu_map_entry_init_bsp[MAPENTRYBSP_CACHE];
137 static struct vm_map_entry cpu_map_entry_init_ap[MAXCPU][MAPENTRYAP_CACHE];
138 static struct vm_map map_init[MAX_KMAP];
140 static int randomize_mmap;
141 SYSCTL_INT(_vm, OID_AUTO, randomize_mmap, CTLFLAG_RW, &randomize_mmap, 0,
142 "Randomize mmap offsets");
143 static int vm_map_relock_enable = 1;
144 SYSCTL_INT(_vm, OID_AUTO, map_relock_enable, CTLFLAG_RW,
145 &vm_map_relock_enable, 0, "Randomize mmap offsets");
147 static void vm_map_entry_shadow(vm_map_entry_t entry, int addref);
148 static vm_map_entry_t vm_map_entry_create(vm_map_t map, int *);
149 static void vm_map_entry_dispose (vm_map_t map, vm_map_entry_t entry, int *);
150 static void _vm_map_clip_end (vm_map_t, vm_map_entry_t, vm_offset_t, int *);
151 static void _vm_map_clip_start (vm_map_t, vm_map_entry_t, vm_offset_t, int *);
152 static void vm_map_entry_delete (vm_map_t, vm_map_entry_t, int *);
153 static void vm_map_entry_unwire (vm_map_t, vm_map_entry_t);
154 static void vm_map_copy_entry (vm_map_t, vm_map_t, vm_map_entry_t,
156 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);
159 * Initialize the vm_map module. Must be called before any other vm_map
162 * Map and entry structures are allocated from the general purpose
163 * memory pool with some exceptions:
165 * - The kernel map is allocated statically.
166 * - Initial kernel map entries are allocated out of a static pool.
167 * - We must set ZONE_SPECIAL here or the early boot code can get
168 * stuck if there are >63 cores.
170 * These restrictions are necessary since malloc() uses the
171 * maps and requires map entries.
173 * Called from the low level boot code only.
178 mapzone = &mapzone_store;
179 zbootinit(mapzone, "MAP", sizeof (struct vm_map),
181 mapentzone = &mapentzone_store;
182 zbootinit(mapentzone, "MAP ENTRY", sizeof (struct vm_map_entry),
183 map_entry_init, MAX_MAPENT);
184 mapentzone_store.zflags |= ZONE_SPECIAL;
188 * Called prior to any vmspace allocations.
190 * Called from the low level boot code only.
195 vmspace_cache = objcache_create_mbacked(M_VMSPACE,
196 sizeof(struct vmspace),
198 vmspace_ctor, vmspace_dtor,
200 zinitna(mapentzone, &mapentobj, NULL, 0, 0,
201 ZONE_USE_RESERVE | ZONE_SPECIAL, 1);
202 zinitna(mapzone, &mapobj, NULL, 0, 0, 0, 1);
208 * objcache support. We leave the pmap root cached as long as possible
209 * for performance reasons.
213 vmspace_ctor(void *obj, void *privdata, int ocflags)
215 struct vmspace *vm = obj;
217 bzero(vm, sizeof(*vm));
218 vm->vm_refcnt = (u_int)-1;
225 vmspace_dtor(void *obj, void *privdata)
227 struct vmspace *vm = obj;
229 KKASSERT(vm->vm_refcnt == (u_int)-1);
230 pmap_puninit(vmspace_pmap(vm));
234 * Red black tree functions
236 * The caller must hold the related map lock.
238 static int rb_vm_map_compare(vm_map_entry_t a, vm_map_entry_t b);
239 RB_GENERATE(vm_map_rb_tree, vm_map_entry, rb_entry, rb_vm_map_compare);
241 /* a->start is address, and the only field has to be initialized */
243 rb_vm_map_compare(vm_map_entry_t a, vm_map_entry_t b)
245 if (a->start < b->start)
247 else if (a->start > b->start)
253 * Initialize vmspace ref/hold counts vmspace0. There is a holdcnt for
257 vmspace_initrefs(struct vmspace *vm)
264 * Allocate a vmspace structure, including a vm_map and pmap.
265 * Initialize numerous fields. While the initial allocation is zerod,
266 * subsequence reuse from the objcache leaves elements of the structure
267 * intact (particularly the pmap), so portions must be zerod.
269 * Returns a referenced vmspace.
274 vmspace_alloc(vm_offset_t min, vm_offset_t max)
278 vm = objcache_get(vmspace_cache, M_WAITOK);
280 bzero(&vm->vm_startcopy,
281 (char *)&vm->vm_endcopy - (char *)&vm->vm_startcopy);
282 vm_map_init(&vm->vm_map, min, max, NULL); /* initializes token */
285 * NOTE: hold to acquires token for safety.
287 * On return vmspace is referenced (refs=1, hold=1). That is,
288 * each refcnt also has a holdcnt. There can be additional holds
289 * (holdcnt) above and beyond the refcnt. Finalization is handled in
290 * two stages, one on refs 1->0, and the the second on hold 1->0.
292 KKASSERT(vm->vm_holdcnt == 0);
293 KKASSERT(vm->vm_refcnt == (u_int)-1);
294 vmspace_initrefs(vm);
296 pmap_pinit(vmspace_pmap(vm)); /* (some fields reused) */
297 vm->vm_map.pmap = vmspace_pmap(vm); /* XXX */
300 cpu_vmspace_alloc(vm);
307 * NOTE: Can return -1 if the vmspace is exiting.
310 vmspace_getrefs(struct vmspace *vm)
312 return ((int)vm->vm_refcnt);
316 * A vmspace object must already have a non-zero hold to be able to gain
317 * further holds on it.
320 vmspace_hold_notoken(struct vmspace *vm)
322 KKASSERT(vm->vm_holdcnt != 0);
323 refcount_acquire(&vm->vm_holdcnt);
327 vmspace_drop_notoken(struct vmspace *vm)
329 if (refcount_release(&vm->vm_holdcnt)) {
330 if (vm->vm_refcnt == (u_int)-1) {
331 vmspace_terminate(vm, 1);
337 vmspace_hold(struct vmspace *vm)
339 vmspace_hold_notoken(vm);
340 lwkt_gettoken(&vm->vm_map.token);
344 vmspace_drop(struct vmspace *vm)
346 lwkt_reltoken(&vm->vm_map.token);
347 vmspace_drop_notoken(vm);
351 * A vmspace object must not be in a terminated state to be able to obtain
352 * additional refs on it.
354 * Ref'ing a vmspace object also increments its hold count.
357 vmspace_ref(struct vmspace *vm)
359 KKASSERT((int)vm->vm_refcnt >= 0);
360 vmspace_hold_notoken(vm);
361 refcount_acquire(&vm->vm_refcnt);
365 * Release a ref on the vmspace. On the 1->0 transition we do stage-1
366 * termination of the vmspace. Then, on the final drop of the hold we
367 * will do stage-2 final termination.
370 vmspace_rel(struct vmspace *vm)
372 if (refcount_release(&vm->vm_refcnt)) {
373 vm->vm_refcnt = (u_int)-1; /* no other refs possible */
374 vmspace_terminate(vm, 0);
376 vmspace_drop_notoken(vm);
380 * This is called during exit indicating that the vmspace is no
381 * longer in used by an exiting process, but the process has not yet
384 * We release the refcnt but not the associated holdcnt.
389 vmspace_relexit(struct vmspace *vm)
391 if (refcount_release(&vm->vm_refcnt)) {
392 vm->vm_refcnt = (u_int)-1; /* no other refs possible */
393 vmspace_terminate(vm, 0);
398 * Called during reap to disconnect the remainder of the vmspace from
399 * the process. On the hold drop the vmspace termination is finalized.
404 vmspace_exitfree(struct proc *p)
410 vmspace_drop_notoken(vm);
414 * Called in two cases:
416 * (1) When the last refcnt is dropped and the vmspace becomes inactive,
417 * called with final == 0. refcnt will be (u_int)-1 at this point,
418 * and holdcnt will still be non-zero.
420 * (2) When holdcnt becomes 0, called with final == 1. There should no
421 * longer be anyone with access to the vmspace.
423 * VMSPACE_EXIT1 flags the primary deactivation
424 * VMSPACE_EXIT2 flags the last reap
427 vmspace_terminate(struct vmspace *vm, int final)
431 lwkt_gettoken(&vm->vm_map.token);
433 KKASSERT((vm->vm_flags & VMSPACE_EXIT1) == 0);
436 * Get rid of most of the resources. Leave the kernel pmap
439 vm->vm_flags |= VMSPACE_EXIT1;
441 pmap_remove_pages(vmspace_pmap(vm), VM_MIN_USER_ADDRESS,
442 VM_MAX_USER_ADDRESS);
443 vm_map_remove(&vm->vm_map, VM_MIN_USER_ADDRESS,
444 VM_MAX_USER_ADDRESS);
445 lwkt_reltoken(&vm->vm_map.token);
447 KKASSERT((vm->vm_flags & VMSPACE_EXIT1) != 0);
448 KKASSERT((vm->vm_flags & VMSPACE_EXIT2) == 0);
451 * Get rid of remaining basic resources.
453 vm->vm_flags |= VMSPACE_EXIT2;
454 cpu_vmspace_free(vm);
458 * Lock the map, to wait out all other references to it.
459 * Delete all of the mappings and pages they hold, then call
460 * the pmap module to reclaim anything left.
462 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
463 vm_map_lock(&vm->vm_map);
464 vm_map_delete(&vm->vm_map, vm->vm_map.min_offset,
465 vm->vm_map.max_offset, &count);
466 vm_map_unlock(&vm->vm_map);
467 vm_map_entry_release(count);
469 lwkt_gettoken(&vmspace_pmap(vm)->pm_token);
470 pmap_release(vmspace_pmap(vm));
471 lwkt_reltoken(&vmspace_pmap(vm)->pm_token);
472 lwkt_reltoken(&vm->vm_map.token);
473 objcache_put(vmspace_cache, vm);
478 * Swap useage is determined by taking the proportional swap used by
479 * VM objects backing the VM map. To make up for fractional losses,
480 * if the VM object has any swap use at all the associated map entries
481 * count for at least 1 swap page.
486 vmspace_swap_count(struct vmspace *vm)
488 vm_map_t map = &vm->vm_map;
495 for (cur = map->header.next; cur != &map->header; cur = cur->next) {
496 switch(cur->maptype) {
497 case VM_MAPTYPE_NORMAL:
498 case VM_MAPTYPE_VPAGETABLE:
499 if ((object = cur->object.vm_object) == NULL)
501 if (object->swblock_count) {
502 n = (cur->end - cur->start) / PAGE_SIZE;
503 count += object->swblock_count *
504 SWAP_META_PAGES * n / object->size + 1;
517 * Calculate the approximate number of anonymous pages in use by
518 * this vmspace. To make up for fractional losses, we count each
519 * VM object as having at least 1 anonymous page.
524 vmspace_anonymous_count(struct vmspace *vm)
526 vm_map_t map = &vm->vm_map;
532 for (cur = map->header.next; cur != &map->header; cur = cur->next) {
533 switch(cur->maptype) {
534 case VM_MAPTYPE_NORMAL:
535 case VM_MAPTYPE_VPAGETABLE:
536 if ((object = cur->object.vm_object) == NULL)
538 if (object->type != OBJT_DEFAULT &&
539 object->type != OBJT_SWAP) {
542 count += object->resident_page_count;
554 * Creates and returns a new empty VM map with the given physical map
555 * structure, and having the given lower and upper address bounds.
560 vm_map_create(vm_map_t result, pmap_t pmap, vm_offset_t min, vm_offset_t max)
563 result = zalloc(mapzone);
564 vm_map_init(result, min, max, pmap);
569 * Initialize an existing vm_map structure such as that in the vmspace
570 * structure. The pmap is initialized elsewhere.
575 vm_map_init(struct vm_map *map, vm_offset_t min, vm_offset_t max, pmap_t pmap)
577 map->header.next = map->header.prev = &map->header;
578 RB_INIT(&map->rb_root);
582 map->min_offset = min;
583 map->max_offset = max;
585 map->first_free = &map->header;
586 map->hint = &map->header;
589 lwkt_token_init(&map->token, "vm_map");
590 lockinit(&map->lock, "vm_maplk", (hz + 9) / 10, 0);
594 * Shadow the vm_map_entry's object. This typically needs to be done when
595 * a write fault is taken on an entry which had previously been cloned by
596 * fork(). The shared object (which might be NULL) must become private so
597 * we add a shadow layer above it.
599 * Object allocation for anonymous mappings is defered as long as possible.
600 * When creating a shadow, however, the underlying object must be instantiated
601 * so it can be shared.
603 * If the map segment is governed by a virtual page table then it is
604 * possible to address offsets beyond the mapped area. Just allocate
605 * a maximally sized object for this case.
607 * The vm_map must be exclusively locked.
608 * No other requirements.
612 vm_map_entry_shadow(vm_map_entry_t entry, int addref)
614 if (entry->maptype == VM_MAPTYPE_VPAGETABLE) {
615 vm_object_shadow(&entry->object.vm_object, &entry->offset,
616 0x7FFFFFFF, addref); /* XXX */
618 vm_object_shadow(&entry->object.vm_object, &entry->offset,
619 atop(entry->end - entry->start), addref);
621 entry->eflags &= ~MAP_ENTRY_NEEDS_COPY;
625 * Allocate an object for a vm_map_entry.
627 * Object allocation for anonymous mappings is defered as long as possible.
628 * This function is called when we can defer no longer, generally when a map
629 * entry might be split or forked or takes a page fault.
631 * If the map segment is governed by a virtual page table then it is
632 * possible to address offsets beyond the mapped area. Just allocate
633 * a maximally sized object for this case.
635 * The vm_map must be exclusively locked.
636 * No other requirements.
639 vm_map_entry_allocate_object(vm_map_entry_t entry)
643 if (entry->maptype == VM_MAPTYPE_VPAGETABLE) {
644 obj = vm_object_allocate(OBJT_DEFAULT, 0x7FFFFFFF); /* XXX */
646 obj = vm_object_allocate(OBJT_DEFAULT,
647 atop(entry->end - entry->start));
649 entry->object.vm_object = obj;
654 * Set an initial negative count so the first attempt to reserve
655 * space preloads a bunch of vm_map_entry's for this cpu. Also
656 * pre-allocate 2 vm_map_entries which will be needed by zalloc() to
657 * map a new page for vm_map_entry structures. SMP systems are
658 * particularly sensitive.
660 * This routine is called in early boot so we cannot just call
661 * vm_map_entry_reserve().
663 * Called from the low level boot code only (for each cpu)
665 * WARNING! Take care not to have too-big a static/BSS structure here
666 * as MAXCPU can be 256+, otherwise the loader's 64MB heap
667 * can get blown out by the kernel plus the initrd image.
670 vm_map_entry_reserve_cpu_init(globaldata_t gd)
672 vm_map_entry_t entry;
676 gd->gd_vme_avail -= MAP_RESERVE_COUNT * 2;
677 if (gd->gd_cpuid == 0) {
678 entry = &cpu_map_entry_init_bsp[0];
679 count = MAPENTRYBSP_CACHE;
681 entry = &cpu_map_entry_init_ap[gd->gd_cpuid][0];
682 count = MAPENTRYAP_CACHE;
684 for (i = 0; i < count; ++i, ++entry) {
685 entry->next = gd->gd_vme_base;
686 gd->gd_vme_base = entry;
691 * Reserves vm_map_entry structures so code later on can manipulate
692 * map_entry structures within a locked map without blocking trying
693 * to allocate a new vm_map_entry.
698 vm_map_entry_reserve(int count)
700 struct globaldata *gd = mycpu;
701 vm_map_entry_t entry;
704 * Make sure we have enough structures in gd_vme_base to handle
705 * the reservation request.
707 * The critical section protects access to the per-cpu gd.
710 while (gd->gd_vme_avail < count) {
711 entry = zalloc(mapentzone);
712 entry->next = gd->gd_vme_base;
713 gd->gd_vme_base = entry;
716 gd->gd_vme_avail -= count;
723 * Releases previously reserved vm_map_entry structures that were not
724 * used. If we have too much junk in our per-cpu cache clean some of
730 vm_map_entry_release(int count)
732 struct globaldata *gd = mycpu;
733 vm_map_entry_t entry;
736 gd->gd_vme_avail += count;
737 while (gd->gd_vme_avail > MAP_RESERVE_SLOP) {
738 entry = gd->gd_vme_base;
739 KKASSERT(entry != NULL);
740 gd->gd_vme_base = entry->next;
743 zfree(mapentzone, entry);
750 * Reserve map entry structures for use in kernel_map itself. These
751 * entries have *ALREADY* been reserved on a per-cpu basis when the map
752 * was inited. This function is used by zalloc() to avoid a recursion
753 * when zalloc() itself needs to allocate additional kernel memory.
755 * This function works like the normal reserve but does not load the
756 * vm_map_entry cache (because that would result in an infinite
757 * recursion). Note that gd_vme_avail may go negative. This is expected.
759 * Any caller of this function must be sure to renormalize after
760 * potentially eating entries to ensure that the reserve supply
766 vm_map_entry_kreserve(int count)
768 struct globaldata *gd = mycpu;
771 gd->gd_vme_avail -= count;
773 KASSERT(gd->gd_vme_base != NULL,
774 ("no reserved entries left, gd_vme_avail = %d",
780 * Release previously reserved map entries for kernel_map. We do not
781 * attempt to clean up like the normal release function as this would
782 * cause an unnecessary (but probably not fatal) deep procedure call.
787 vm_map_entry_krelease(int count)
789 struct globaldata *gd = mycpu;
792 gd->gd_vme_avail += count;
797 * Allocates a VM map entry for insertion. No entry fields are filled in.
799 * The entries should have previously been reserved. The reservation count
800 * is tracked in (*countp).
804 static vm_map_entry_t
805 vm_map_entry_create(vm_map_t map, int *countp)
807 struct globaldata *gd = mycpu;
808 vm_map_entry_t entry;
810 KKASSERT(*countp > 0);
813 entry = gd->gd_vme_base;
814 KASSERT(entry != NULL, ("gd_vme_base NULL! count %d", *countp));
815 gd->gd_vme_base = entry->next;
822 * Dispose of a vm_map_entry that is no longer being referenced.
827 vm_map_entry_dispose(vm_map_t map, vm_map_entry_t entry, int *countp)
829 struct globaldata *gd = mycpu;
831 KKASSERT(map->hint != entry);
832 KKASSERT(map->first_free != entry);
836 entry->next = gd->gd_vme_base;
837 gd->gd_vme_base = entry;
843 * Insert/remove entries from maps.
845 * The related map must be exclusively locked.
846 * The caller must hold map->token
847 * No other requirements.
850 vm_map_entry_link(vm_map_t map,
851 vm_map_entry_t after_where,
852 vm_map_entry_t entry)
854 ASSERT_VM_MAP_LOCKED(map);
857 entry->prev = after_where;
858 entry->next = after_where->next;
859 entry->next->prev = entry;
860 after_where->next = entry;
861 if (vm_map_rb_tree_RB_INSERT(&map->rb_root, entry))
862 panic("vm_map_entry_link: dup addr map %p ent %p", map, entry);
866 vm_map_entry_unlink(vm_map_t map,
867 vm_map_entry_t entry)
872 ASSERT_VM_MAP_LOCKED(map);
874 if (entry->eflags & MAP_ENTRY_IN_TRANSITION) {
875 panic("vm_map_entry_unlink: attempt to mess with "
876 "locked entry! %p", entry);
882 vm_map_rb_tree_RB_REMOVE(&map->rb_root, entry);
887 * Finds the map entry containing (or immediately preceding) the specified
888 * address in the given map. The entry is returned in (*entry).
890 * The boolean result indicates whether the address is actually contained
893 * The related map must be locked.
894 * No other requirements.
897 vm_map_lookup_entry(vm_map_t map, vm_offset_t address, vm_map_entry_t *entry)
902 ASSERT_VM_MAP_LOCKED(map);
905 * XXX TEMPORARILY DISABLED. For some reason our attempt to revive
906 * the hint code with the red-black lookup meets with system crashes
907 * and lockups. We do not yet know why.
909 * It is possible that the problem is related to the setting
910 * of the hint during map_entry deletion, in the code specified
911 * at the GGG comment later on in this file.
913 * YYY More likely it's because this function can be called with
914 * a shared lock on the map, resulting in map->hint updates possibly
915 * racing. Fixed now but untested.
918 * Quickly check the cached hint, there's a good chance of a match.
922 if (tmp != &map->header) {
923 if (address >= tmp->start && address < tmp->end) {
931 * Locate the record from the top of the tree. 'last' tracks the
932 * closest prior record and is returned if no match is found, which
933 * in binary tree terms means tracking the most recent right-branch
934 * taken. If there is no prior record, &map->header is returned.
937 tmp = RB_ROOT(&map->rb_root);
940 if (address >= tmp->start) {
941 if (address < tmp->end) {
947 tmp = RB_RIGHT(tmp, rb_entry);
949 tmp = RB_LEFT(tmp, rb_entry);
957 * Inserts the given whole VM object into the target map at the specified
958 * address range. The object's size should match that of the address range.
960 * The map must be exclusively locked.
961 * The object must be held.
962 * The caller must have reserved sufficient vm_map_entry structures.
964 * If object is non-NULL, ref count must be bumped by caller prior to
965 * making call to account for the new entry.
968 vm_map_insert(vm_map_t map, int *countp, void *map_object, void *map_aux,
969 vm_ooffset_t offset, vm_offset_t start, vm_offset_t end,
970 vm_maptype_t maptype,
971 vm_prot_t prot, vm_prot_t max, int cow)
973 vm_map_entry_t new_entry;
974 vm_map_entry_t prev_entry;
975 vm_map_entry_t temp_entry;
976 vm_eflags_t protoeflags;
980 if (maptype == VM_MAPTYPE_UKSMAP)
985 ASSERT_VM_MAP_LOCKED(map);
987 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
990 * Check that the start and end points are not bogus.
992 if ((start < map->min_offset) || (end > map->max_offset) ||
994 return (KERN_INVALID_ADDRESS);
997 * Find the entry prior to the proposed starting address; if it's part
998 * of an existing entry, this range is bogus.
1000 if (vm_map_lookup_entry(map, start, &temp_entry))
1001 return (KERN_NO_SPACE);
1003 prev_entry = temp_entry;
1006 * Assert that the next entry doesn't overlap the end point.
1009 if ((prev_entry->next != &map->header) &&
1010 (prev_entry->next->start < end))
1011 return (KERN_NO_SPACE);
1015 if (cow & MAP_COPY_ON_WRITE)
1016 protoeflags |= MAP_ENTRY_COW|MAP_ENTRY_NEEDS_COPY;
1018 if (cow & MAP_NOFAULT) {
1019 protoeflags |= MAP_ENTRY_NOFAULT;
1021 KASSERT(object == NULL,
1022 ("vm_map_insert: paradoxical MAP_NOFAULT request"));
1024 if (cow & MAP_DISABLE_SYNCER)
1025 protoeflags |= MAP_ENTRY_NOSYNC;
1026 if (cow & MAP_DISABLE_COREDUMP)
1027 protoeflags |= MAP_ENTRY_NOCOREDUMP;
1028 if (cow & MAP_IS_STACK)
1029 protoeflags |= MAP_ENTRY_STACK;
1030 if (cow & MAP_IS_KSTACK)
1031 protoeflags |= MAP_ENTRY_KSTACK;
1033 lwkt_gettoken(&map->token);
1037 * When object is non-NULL, it could be shared with another
1038 * process. We have to set or clear OBJ_ONEMAPPING
1041 * NOTE: This flag is only applicable to DEFAULT and SWAP
1042 * objects and will already be clear in other types
1043 * of objects, so a shared object lock is ok for
1046 if ((object->ref_count > 1) || (object->shadow_count != 0)) {
1047 vm_object_clear_flag(object, OBJ_ONEMAPPING);
1050 else if ((prev_entry != &map->header) &&
1051 (prev_entry->eflags == protoeflags) &&
1052 (prev_entry->end == start) &&
1053 (prev_entry->wired_count == 0) &&
1054 prev_entry->maptype == maptype &&
1055 maptype == VM_MAPTYPE_NORMAL &&
1056 ((prev_entry->object.vm_object == NULL) ||
1057 vm_object_coalesce(prev_entry->object.vm_object,
1058 OFF_TO_IDX(prev_entry->offset),
1059 (vm_size_t)(prev_entry->end - prev_entry->start),
1060 (vm_size_t)(end - prev_entry->end)))) {
1062 * We were able to extend the object. Determine if we
1063 * can extend the previous map entry to include the
1064 * new range as well.
1066 if ((prev_entry->inheritance == VM_INHERIT_DEFAULT) &&
1067 (prev_entry->protection == prot) &&
1068 (prev_entry->max_protection == max)) {
1069 map->size += (end - prev_entry->end);
1070 prev_entry->end = end;
1071 vm_map_simplify_entry(map, prev_entry, countp);
1072 lwkt_reltoken(&map->token);
1073 return (KERN_SUCCESS);
1077 * If we can extend the object but cannot extend the
1078 * map entry, we have to create a new map entry. We
1079 * must bump the ref count on the extended object to
1080 * account for it. object may be NULL.
1082 object = prev_entry->object.vm_object;
1083 offset = prev_entry->offset +
1084 (prev_entry->end - prev_entry->start);
1086 vm_object_hold(object);
1087 vm_object_chain_wait(object, 0);
1088 vm_object_reference_locked(object);
1094 * NOTE: if conditionals fail, object can be NULL here. This occurs
1095 * in things like the buffer map where we manage kva but do not manage
1100 * Create a new entry
1103 new_entry = vm_map_entry_create(map, countp);
1104 new_entry->start = start;
1105 new_entry->end = end;
1107 new_entry->maptype = maptype;
1108 new_entry->eflags = protoeflags;
1109 new_entry->object.map_object = map_object;
1110 new_entry->aux.master_pde = 0; /* in case size is different */
1111 new_entry->aux.map_aux = map_aux;
1112 new_entry->offset = offset;
1114 new_entry->inheritance = VM_INHERIT_DEFAULT;
1115 new_entry->protection = prot;
1116 new_entry->max_protection = max;
1117 new_entry->wired_count = 0;
1120 * Insert the new entry into the list
1123 vm_map_entry_link(map, prev_entry, new_entry);
1124 map->size += new_entry->end - new_entry->start;
1127 * Update the free space hint. Entries cannot overlap.
1128 * An exact comparison is needed to avoid matching
1129 * against the map->header.
1131 if ((map->first_free == prev_entry) &&
1132 (prev_entry->end == new_entry->start)) {
1133 map->first_free = new_entry;
1138 * Temporarily removed to avoid MAP_STACK panic, due to
1139 * MAP_STACK being a huge hack. Will be added back in
1140 * when MAP_STACK (and the user stack mapping) is fixed.
1143 * It may be possible to simplify the entry
1145 vm_map_simplify_entry(map, new_entry, countp);
1149 * Try to pre-populate the page table. Mappings governed by virtual
1150 * page tables cannot be prepopulated without a lot of work, so
1153 if ((cow & (MAP_PREFAULT|MAP_PREFAULT_PARTIAL)) &&
1154 maptype != VM_MAPTYPE_VPAGETABLE &&
1155 maptype != VM_MAPTYPE_UKSMAP) {
1157 if (vm_map_relock_enable && (cow & MAP_PREFAULT_RELOCK)) {
1159 vm_object_lock_swap();
1160 vm_object_drop(object);
1162 pmap_object_init_pt(map->pmap, start, prot,
1163 object, OFF_TO_IDX(offset), end - start,
1164 cow & MAP_PREFAULT_PARTIAL);
1166 vm_object_hold(object);
1167 vm_object_lock_swap();
1171 vm_object_drop(object);
1173 lwkt_reltoken(&map->token);
1174 return (KERN_SUCCESS);
1178 * Find sufficient space for `length' bytes in the given map, starting at
1179 * `start'. Returns 0 on success, 1 on no space.
1181 * This function will returned an arbitrarily aligned pointer. If no
1182 * particular alignment is required you should pass align as 1. Note that
1183 * the map may return PAGE_SIZE aligned pointers if all the lengths used in
1184 * the map are a multiple of PAGE_SIZE, even if you pass a smaller align
1187 * 'align' should be a power of 2 but is not required to be.
1189 * The map must be exclusively locked.
1190 * No other requirements.
1193 vm_map_findspace(vm_map_t map, vm_offset_t start, vm_size_t length,
1194 vm_size_t align, int flags, vm_offset_t *addr)
1196 vm_map_entry_t entry, next;
1198 vm_offset_t align_mask;
1200 if (start < map->min_offset)
1201 start = map->min_offset;
1202 if (start > map->max_offset)
1206 * If the alignment is not a power of 2 we will have to use
1207 * a mod/division, set align_mask to a special value.
1209 if ((align | (align - 1)) + 1 != (align << 1))
1210 align_mask = (vm_offset_t)-1;
1212 align_mask = align - 1;
1215 * Look for the first possible address; if there's already something
1216 * at this address, we have to start after it.
1218 if (start == map->min_offset) {
1219 if ((entry = map->first_free) != &map->header)
1224 if (vm_map_lookup_entry(map, start, &tmp))
1230 * Look through the rest of the map, trying to fit a new region in the
1231 * gap between existing regions, or after the very last region.
1233 for (;; start = (entry = next)->end) {
1235 * Adjust the proposed start by the requested alignment,
1236 * be sure that we didn't wrap the address.
1238 if (align_mask == (vm_offset_t)-1)
1239 end = ((start + align - 1) / align) * align;
1241 end = (start + align_mask) & ~align_mask;
1246 * Find the end of the proposed new region. Be sure we didn't
1247 * go beyond the end of the map, or wrap around the address.
1248 * Then check to see if this is the last entry or if the
1249 * proposed end fits in the gap between this and the next
1252 end = start + length;
1253 if (end > map->max_offset || end < start)
1258 * If the next entry's start address is beyond the desired
1259 * end address we may have found a good entry.
1261 * If the next entry is a stack mapping we do not map into
1262 * the stack's reserved space.
1264 * XXX continue to allow mapping into the stack's reserved
1265 * space if doing a MAP_STACK mapping inside a MAP_STACK
1266 * mapping, for backwards compatibility. But the caller
1267 * really should use MAP_STACK | MAP_TRYFIXED if they
1270 if (next == &map->header)
1272 if (next->start >= end) {
1273 if ((next->eflags & MAP_ENTRY_STACK) == 0)
1275 if (flags & MAP_STACK)
1277 if (next->start - next->aux.avail_ssize >= end)
1284 * Grow the kernel_map if necessary. pmap_growkernel() will panic
1285 * if it fails. The kernel_map is locked and nothing can steal
1286 * our address space if pmap_growkernel() blocks.
1288 * NOTE: This may be unconditionally called for kldload areas on
1289 * x86_64 because these do not bump kernel_vm_end (which would
1290 * fill 128G worth of page tables!). Therefore we must not
1293 if (map == &kernel_map) {
1296 kstop = round_page(start + length);
1297 if (kstop > kernel_vm_end)
1298 pmap_growkernel(start, kstop);
1305 * vm_map_find finds an unallocated region in the target address map with
1306 * the given length and allocates it. The search is defined to be first-fit
1307 * from the specified address; the region found is returned in the same
1310 * If object is non-NULL, ref count must be bumped by caller
1311 * prior to making call to account for the new entry.
1313 * No requirements. This function will lock the map temporarily.
1316 vm_map_find(vm_map_t map, void *map_object, void *map_aux,
1317 vm_ooffset_t offset, vm_offset_t *addr,
1318 vm_size_t length, vm_size_t align,
1320 vm_maptype_t maptype,
1321 vm_prot_t prot, vm_prot_t max,
1329 if (maptype == VM_MAPTYPE_UKSMAP)
1332 object = map_object;
1336 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
1339 vm_object_hold_shared(object);
1341 if (vm_map_findspace(map, start, length, align, 0, addr)) {
1343 vm_object_drop(object);
1345 vm_map_entry_release(count);
1346 return (KERN_NO_SPACE);
1350 result = vm_map_insert(map, &count, map_object, map_aux,
1351 offset, start, start + length,
1352 maptype, prot, max, cow);
1354 vm_object_drop(object);
1356 vm_map_entry_release(count);
1362 * Simplify the given map entry by merging with either neighbor. This
1363 * routine also has the ability to merge with both neighbors.
1365 * This routine guarentees that the passed entry remains valid (though
1366 * possibly extended). When merging, this routine may delete one or
1367 * both neighbors. No action is taken on entries which have their
1368 * in-transition flag set.
1370 * The map must be exclusively locked.
1373 vm_map_simplify_entry(vm_map_t map, vm_map_entry_t entry, int *countp)
1375 vm_map_entry_t next, prev;
1376 vm_size_t prevsize, esize;
1378 if (entry->eflags & MAP_ENTRY_IN_TRANSITION) {
1379 ++mycpu->gd_cnt.v_intrans_coll;
1383 if (entry->maptype == VM_MAPTYPE_SUBMAP)
1385 if (entry->maptype == VM_MAPTYPE_UKSMAP)
1389 if (prev != &map->header) {
1390 prevsize = prev->end - prev->start;
1391 if ( (prev->end == entry->start) &&
1392 (prev->maptype == entry->maptype) &&
1393 (prev->object.vm_object == entry->object.vm_object) &&
1394 (!prev->object.vm_object ||
1395 (prev->offset + prevsize == entry->offset)) &&
1396 (prev->eflags == entry->eflags) &&
1397 (prev->protection == entry->protection) &&
1398 (prev->max_protection == entry->max_protection) &&
1399 (prev->inheritance == entry->inheritance) &&
1400 (prev->wired_count == entry->wired_count)) {
1401 if (map->first_free == prev)
1402 map->first_free = entry;
1403 if (map->hint == prev)
1405 vm_map_entry_unlink(map, prev);
1406 entry->start = prev->start;
1407 entry->offset = prev->offset;
1408 if (prev->object.vm_object)
1409 vm_object_deallocate(prev->object.vm_object);
1410 vm_map_entry_dispose(map, prev, countp);
1415 if (next != &map->header) {
1416 esize = entry->end - entry->start;
1417 if ((entry->end == next->start) &&
1418 (next->maptype == entry->maptype) &&
1419 (next->object.vm_object == entry->object.vm_object) &&
1420 (!entry->object.vm_object ||
1421 (entry->offset + esize == next->offset)) &&
1422 (next->eflags == entry->eflags) &&
1423 (next->protection == entry->protection) &&
1424 (next->max_protection == entry->max_protection) &&
1425 (next->inheritance == entry->inheritance) &&
1426 (next->wired_count == entry->wired_count)) {
1427 if (map->first_free == next)
1428 map->first_free = entry;
1429 if (map->hint == next)
1431 vm_map_entry_unlink(map, next);
1432 entry->end = next->end;
1433 if (next->object.vm_object)
1434 vm_object_deallocate(next->object.vm_object);
1435 vm_map_entry_dispose(map, next, countp);
1441 * Asserts that the given entry begins at or after the specified address.
1442 * If necessary, it splits the entry into two.
1444 #define vm_map_clip_start(map, entry, startaddr, countp) \
1446 if (startaddr > entry->start) \
1447 _vm_map_clip_start(map, entry, startaddr, countp); \
1451 * This routine is called only when it is known that the entry must be split.
1453 * The map must be exclusively locked.
1456 _vm_map_clip_start(vm_map_t map, vm_map_entry_t entry, vm_offset_t start,
1459 vm_map_entry_t new_entry;
1462 * Split off the front portion -- note that we must insert the new
1463 * entry BEFORE this one, so that this entry has the specified
1467 vm_map_simplify_entry(map, entry, countp);
1470 * If there is no object backing this entry, we might as well create
1471 * one now. If we defer it, an object can get created after the map
1472 * is clipped, and individual objects will be created for the split-up
1473 * map. This is a bit of a hack, but is also about the best place to
1474 * put this improvement.
1476 if (entry->object.vm_object == NULL && !map->system_map) {
1477 vm_map_entry_allocate_object(entry);
1480 new_entry = vm_map_entry_create(map, countp);
1481 *new_entry = *entry;
1483 new_entry->end = start;
1484 entry->offset += (start - entry->start);
1485 entry->start = start;
1487 vm_map_entry_link(map, entry->prev, new_entry);
1489 switch(entry->maptype) {
1490 case VM_MAPTYPE_NORMAL:
1491 case VM_MAPTYPE_VPAGETABLE:
1492 if (new_entry->object.vm_object) {
1493 vm_object_hold(new_entry->object.vm_object);
1494 vm_object_chain_wait(new_entry->object.vm_object, 0);
1495 vm_object_reference_locked(new_entry->object.vm_object);
1496 vm_object_drop(new_entry->object.vm_object);
1505 * Asserts that the given entry ends at or before the specified address.
1506 * If necessary, it splits the entry into two.
1508 * The map must be exclusively locked.
1510 #define vm_map_clip_end(map, entry, endaddr, countp) \
1512 if (endaddr < entry->end) \
1513 _vm_map_clip_end(map, entry, endaddr, countp); \
1517 * This routine is called only when it is known that the entry must be split.
1519 * The map must be exclusively locked.
1522 _vm_map_clip_end(vm_map_t map, vm_map_entry_t entry, vm_offset_t end,
1525 vm_map_entry_t new_entry;
1528 * If there is no object backing this entry, we might as well create
1529 * one now. If we defer it, an object can get created after the map
1530 * is clipped, and individual objects will be created for the split-up
1531 * map. This is a bit of a hack, but is also about the best place to
1532 * put this improvement.
1535 if (entry->object.vm_object == NULL && !map->system_map) {
1536 vm_map_entry_allocate_object(entry);
1540 * Create a new entry and insert it AFTER the specified entry
1543 new_entry = vm_map_entry_create(map, countp);
1544 *new_entry = *entry;
1546 new_entry->start = entry->end = end;
1547 new_entry->offset += (end - entry->start);
1549 vm_map_entry_link(map, entry, new_entry);
1551 switch(entry->maptype) {
1552 case VM_MAPTYPE_NORMAL:
1553 case VM_MAPTYPE_VPAGETABLE:
1554 if (new_entry->object.vm_object) {
1555 vm_object_hold(new_entry->object.vm_object);
1556 vm_object_chain_wait(new_entry->object.vm_object, 0);
1557 vm_object_reference_locked(new_entry->object.vm_object);
1558 vm_object_drop(new_entry->object.vm_object);
1567 * Asserts that the starting and ending region addresses fall within the
1568 * valid range for the map.
1570 #define VM_MAP_RANGE_CHECK(map, start, end) \
1572 if (start < vm_map_min(map)) \
1573 start = vm_map_min(map); \
1574 if (end > vm_map_max(map)) \
1575 end = vm_map_max(map); \
1581 * Used to block when an in-transition collison occurs. The map
1582 * is unlocked for the sleep and relocked before the return.
1585 vm_map_transition_wait(vm_map_t map)
1587 tsleep_interlock(map, 0);
1589 tsleep(map, PINTERLOCKED, "vment", 0);
1594 * When we do blocking operations with the map lock held it is
1595 * possible that a clip might have occured on our in-transit entry,
1596 * requiring an adjustment to the entry in our loop. These macros
1597 * help the pageable and clip_range code deal with the case. The
1598 * conditional costs virtually nothing if no clipping has occured.
1601 #define CLIP_CHECK_BACK(entry, save_start) \
1603 while (entry->start != save_start) { \
1604 entry = entry->prev; \
1605 KASSERT(entry != &map->header, ("bad entry clip")); \
1609 #define CLIP_CHECK_FWD(entry, save_end) \
1611 while (entry->end != save_end) { \
1612 entry = entry->next; \
1613 KASSERT(entry != &map->header, ("bad entry clip")); \
1619 * Clip the specified range and return the base entry. The
1620 * range may cover several entries starting at the returned base
1621 * and the first and last entry in the covering sequence will be
1622 * properly clipped to the requested start and end address.
1624 * If no holes are allowed you should pass the MAP_CLIP_NO_HOLES
1627 * The MAP_ENTRY_IN_TRANSITION flag will be set for the entries
1628 * covered by the requested range.
1630 * The map must be exclusively locked on entry and will remain locked
1631 * on return. If no range exists or the range contains holes and you
1632 * specified that no holes were allowed, NULL will be returned. This
1633 * routine may temporarily unlock the map in order avoid a deadlock when
1638 vm_map_clip_range(vm_map_t map, vm_offset_t start, vm_offset_t end,
1639 int *countp, int flags)
1641 vm_map_entry_t start_entry;
1642 vm_map_entry_t entry;
1645 * Locate the entry and effect initial clipping. The in-transition
1646 * case does not occur very often so do not try to optimize it.
1649 if (vm_map_lookup_entry(map, start, &start_entry) == FALSE)
1651 entry = start_entry;
1652 if (entry->eflags & MAP_ENTRY_IN_TRANSITION) {
1653 entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP;
1654 ++mycpu->gd_cnt.v_intrans_coll;
1655 ++mycpu->gd_cnt.v_intrans_wait;
1656 vm_map_transition_wait(map);
1658 * entry and/or start_entry may have been clipped while
1659 * we slept, or may have gone away entirely. We have
1660 * to restart from the lookup.
1666 * Since we hold an exclusive map lock we do not have to restart
1667 * after clipping, even though clipping may block in zalloc.
1669 vm_map_clip_start(map, entry, start, countp);
1670 vm_map_clip_end(map, entry, end, countp);
1671 entry->eflags |= MAP_ENTRY_IN_TRANSITION;
1674 * Scan entries covered by the range. When working on the next
1675 * entry a restart need only re-loop on the current entry which
1676 * we have already locked, since 'next' may have changed. Also,
1677 * even though entry is safe, it may have been clipped so we
1678 * have to iterate forwards through the clip after sleeping.
1680 while (entry->next != &map->header && entry->next->start < end) {
1681 vm_map_entry_t next = entry->next;
1683 if (flags & MAP_CLIP_NO_HOLES) {
1684 if (next->start > entry->end) {
1685 vm_map_unclip_range(map, start_entry,
1686 start, entry->end, countp, flags);
1691 if (next->eflags & MAP_ENTRY_IN_TRANSITION) {
1692 vm_offset_t save_end = entry->end;
1693 next->eflags |= MAP_ENTRY_NEEDS_WAKEUP;
1694 ++mycpu->gd_cnt.v_intrans_coll;
1695 ++mycpu->gd_cnt.v_intrans_wait;
1696 vm_map_transition_wait(map);
1699 * clips might have occured while we blocked.
1701 CLIP_CHECK_FWD(entry, save_end);
1702 CLIP_CHECK_BACK(start_entry, start);
1706 * No restart necessary even though clip_end may block, we
1707 * are holding the map lock.
1709 vm_map_clip_end(map, next, end, countp);
1710 next->eflags |= MAP_ENTRY_IN_TRANSITION;
1713 if (flags & MAP_CLIP_NO_HOLES) {
1714 if (entry->end != end) {
1715 vm_map_unclip_range(map, start_entry,
1716 start, entry->end, countp, flags);
1720 return(start_entry);
1724 * Undo the effect of vm_map_clip_range(). You should pass the same
1725 * flags and the same range that you passed to vm_map_clip_range().
1726 * This code will clear the in-transition flag on the entries and
1727 * wake up anyone waiting. This code will also simplify the sequence
1728 * and attempt to merge it with entries before and after the sequence.
1730 * The map must be locked on entry and will remain locked on return.
1732 * Note that you should also pass the start_entry returned by
1733 * vm_map_clip_range(). However, if you block between the two calls
1734 * with the map unlocked please be aware that the start_entry may
1735 * have been clipped and you may need to scan it backwards to find
1736 * the entry corresponding with the original start address. You are
1737 * responsible for this, vm_map_unclip_range() expects the correct
1738 * start_entry to be passed to it and will KASSERT otherwise.
1742 vm_map_unclip_range(vm_map_t map, vm_map_entry_t start_entry,
1743 vm_offset_t start, vm_offset_t end,
1744 int *countp, int flags)
1746 vm_map_entry_t entry;
1748 entry = start_entry;
1750 KASSERT(entry->start == start, ("unclip_range: illegal base entry"));
1751 while (entry != &map->header && entry->start < end) {
1752 KASSERT(entry->eflags & MAP_ENTRY_IN_TRANSITION,
1753 ("in-transition flag not set during unclip on: %p",
1755 KASSERT(entry->end <= end,
1756 ("unclip_range: tail wasn't clipped"));
1757 entry->eflags &= ~MAP_ENTRY_IN_TRANSITION;
1758 if (entry->eflags & MAP_ENTRY_NEEDS_WAKEUP) {
1759 entry->eflags &= ~MAP_ENTRY_NEEDS_WAKEUP;
1762 entry = entry->next;
1766 * Simplification does not block so there is no restart case.
1768 entry = start_entry;
1769 while (entry != &map->header && entry->start < end) {
1770 vm_map_simplify_entry(map, entry, countp);
1771 entry = entry->next;
1776 * Mark the given range as handled by a subordinate map.
1778 * This range must have been created with vm_map_find(), and no other
1779 * operations may have been performed on this range prior to calling
1782 * Submappings cannot be removed.
1787 vm_map_submap(vm_map_t map, vm_offset_t start, vm_offset_t end, vm_map_t submap)
1789 vm_map_entry_t entry;
1790 int result = KERN_INVALID_ARGUMENT;
1793 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
1796 VM_MAP_RANGE_CHECK(map, start, end);
1798 if (vm_map_lookup_entry(map, start, &entry)) {
1799 vm_map_clip_start(map, entry, start, &count);
1801 entry = entry->next;
1804 vm_map_clip_end(map, entry, end, &count);
1806 if ((entry->start == start) && (entry->end == end) &&
1807 ((entry->eflags & MAP_ENTRY_COW) == 0) &&
1808 (entry->object.vm_object == NULL)) {
1809 entry->object.sub_map = submap;
1810 entry->maptype = VM_MAPTYPE_SUBMAP;
1811 result = KERN_SUCCESS;
1814 vm_map_entry_release(count);
1820 * Sets the protection of the specified address region in the target map.
1821 * If "set_max" is specified, the maximum protection is to be set;
1822 * otherwise, only the current protection is affected.
1824 * The protection is not applicable to submaps, but is applicable to normal
1825 * maps and maps governed by virtual page tables. For example, when operating
1826 * on a virtual page table our protection basically controls how COW occurs
1827 * on the backing object, whereas the virtual page table abstraction itself
1828 * is an abstraction for userland.
1833 vm_map_protect(vm_map_t map, vm_offset_t start, vm_offset_t end,
1834 vm_prot_t new_prot, boolean_t set_max)
1836 vm_map_entry_t current;
1837 vm_map_entry_t entry;
1840 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
1843 VM_MAP_RANGE_CHECK(map, start, end);
1845 if (vm_map_lookup_entry(map, start, &entry)) {
1846 vm_map_clip_start(map, entry, start, &count);
1848 entry = entry->next;
1852 * Make a first pass to check for protection violations.
1855 while ((current != &map->header) && (current->start < end)) {
1856 if (current->maptype == VM_MAPTYPE_SUBMAP) {
1858 vm_map_entry_release(count);
1859 return (KERN_INVALID_ARGUMENT);
1861 if ((new_prot & current->max_protection) != new_prot) {
1863 vm_map_entry_release(count);
1864 return (KERN_PROTECTION_FAILURE);
1866 current = current->next;
1870 * Go back and fix up protections. [Note that clipping is not
1871 * necessary the second time.]
1875 while ((current != &map->header) && (current->start < end)) {
1878 vm_map_clip_end(map, current, end, &count);
1880 old_prot = current->protection;
1882 current->protection =
1883 (current->max_protection = new_prot) &
1886 current->protection = new_prot;
1890 * Update physical map if necessary. Worry about copy-on-write
1891 * here -- CHECK THIS XXX
1894 if (current->protection != old_prot) {
1895 #define MASK(entry) (((entry)->eflags & MAP_ENTRY_COW) ? ~VM_PROT_WRITE : \
1898 pmap_protect(map->pmap, current->start,
1900 current->protection & MASK(current));
1904 vm_map_simplify_entry(map, current, &count);
1906 current = current->next;
1910 vm_map_entry_release(count);
1911 return (KERN_SUCCESS);
1915 * This routine traverses a processes map handling the madvise
1916 * system call. Advisories are classified as either those effecting
1917 * the vm_map_entry structure, or those effecting the underlying
1920 * The <value> argument is used for extended madvise calls.
1925 vm_map_madvise(vm_map_t map, vm_offset_t start, vm_offset_t end,
1926 int behav, off_t value)
1928 vm_map_entry_t current, entry;
1934 * Some madvise calls directly modify the vm_map_entry, in which case
1935 * we need to use an exclusive lock on the map and we need to perform
1936 * various clipping operations. Otherwise we only need a read-lock
1940 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
1944 case MADV_SEQUENTIAL:
1958 vm_map_lock_read(map);
1961 vm_map_entry_release(count);
1966 * Locate starting entry and clip if necessary.
1969 VM_MAP_RANGE_CHECK(map, start, end);
1971 if (vm_map_lookup_entry(map, start, &entry)) {
1973 vm_map_clip_start(map, entry, start, &count);
1975 entry = entry->next;
1980 * madvise behaviors that are implemented in the vm_map_entry.
1982 * We clip the vm_map_entry so that behavioral changes are
1983 * limited to the specified address range.
1985 for (current = entry;
1986 (current != &map->header) && (current->start < end);
1987 current = current->next
1989 if (current->maptype == VM_MAPTYPE_SUBMAP)
1992 vm_map_clip_end(map, current, end, &count);
1996 vm_map_entry_set_behavior(current, MAP_ENTRY_BEHAV_NORMAL);
1998 case MADV_SEQUENTIAL:
1999 vm_map_entry_set_behavior(current, MAP_ENTRY_BEHAV_SEQUENTIAL);
2002 vm_map_entry_set_behavior(current, MAP_ENTRY_BEHAV_RANDOM);
2005 current->eflags |= MAP_ENTRY_NOSYNC;
2008 current->eflags &= ~MAP_ENTRY_NOSYNC;
2011 current->eflags |= MAP_ENTRY_NOCOREDUMP;
2014 current->eflags &= ~MAP_ENTRY_NOCOREDUMP;
2018 * Invalidate the related pmap entries, used
2019 * to flush portions of the real kernel's
2020 * pmap when the caller has removed or
2021 * modified existing mappings in a virtual
2024 pmap_remove(map->pmap,
2025 current->start, current->end);
2029 * Set the page directory page for a map
2030 * governed by a virtual page table. Mark
2031 * the entry as being governed by a virtual
2032 * page table if it is not.
2034 * XXX the page directory page is stored
2035 * in the avail_ssize field if the map_entry.
2037 * XXX the map simplification code does not
2038 * compare this field so weird things may
2039 * happen if you do not apply this function
2040 * to the entire mapping governed by the
2041 * virtual page table.
2043 if (current->maptype != VM_MAPTYPE_VPAGETABLE) {
2047 current->aux.master_pde = value;
2048 pmap_remove(map->pmap,
2049 current->start, current->end);
2055 vm_map_simplify_entry(map, current, &count);
2063 * madvise behaviors that are implemented in the underlying
2066 * Since we don't clip the vm_map_entry, we have to clip
2067 * the vm_object pindex and count.
2069 * NOTE! We currently do not support these functions on
2070 * virtual page tables.
2072 for (current = entry;
2073 (current != &map->header) && (current->start < end);
2074 current = current->next
2076 vm_offset_t useStart;
2078 if (current->maptype != VM_MAPTYPE_NORMAL)
2081 pindex = OFF_TO_IDX(current->offset);
2082 count = atop(current->end - current->start);
2083 useStart = current->start;
2085 if (current->start < start) {
2086 pindex += atop(start - current->start);
2087 count -= atop(start - current->start);
2090 if (current->end > end)
2091 count -= atop(current->end - end);
2096 vm_object_madvise(current->object.vm_object,
2097 pindex, count, behav);
2100 * Try to populate the page table. Mappings governed
2101 * by virtual page tables cannot be pre-populated
2102 * without a lot of work so don't try.
2104 if (behav == MADV_WILLNEED &&
2105 current->maptype != VM_MAPTYPE_VPAGETABLE) {
2106 pmap_object_init_pt(
2109 current->protection,
2110 current->object.vm_object,
2112 (count << PAGE_SHIFT),
2113 MAP_PREFAULT_MADVISE
2117 vm_map_unlock_read(map);
2119 vm_map_entry_release(count);
2125 * Sets the inheritance of the specified address range in the target map.
2126 * Inheritance affects how the map will be shared with child maps at the
2127 * time of vm_map_fork.
2130 vm_map_inherit(vm_map_t map, vm_offset_t start, vm_offset_t end,
2131 vm_inherit_t new_inheritance)
2133 vm_map_entry_t entry;
2134 vm_map_entry_t temp_entry;
2137 switch (new_inheritance) {
2138 case VM_INHERIT_NONE:
2139 case VM_INHERIT_COPY:
2140 case VM_INHERIT_SHARE:
2143 return (KERN_INVALID_ARGUMENT);
2146 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
2149 VM_MAP_RANGE_CHECK(map, start, end);
2151 if (vm_map_lookup_entry(map, start, &temp_entry)) {
2153 vm_map_clip_start(map, entry, start, &count);
2155 entry = temp_entry->next;
2157 while ((entry != &map->header) && (entry->start < end)) {
2158 vm_map_clip_end(map, entry, end, &count);
2160 entry->inheritance = new_inheritance;
2162 vm_map_simplify_entry(map, entry, &count);
2164 entry = entry->next;
2167 vm_map_entry_release(count);
2168 return (KERN_SUCCESS);
2172 * Implement the semantics of mlock
2175 vm_map_unwire(vm_map_t map, vm_offset_t start, vm_offset_t real_end,
2176 boolean_t new_pageable)
2178 vm_map_entry_t entry;
2179 vm_map_entry_t start_entry;
2181 int rv = KERN_SUCCESS;
2184 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
2186 VM_MAP_RANGE_CHECK(map, start, real_end);
2189 start_entry = vm_map_clip_range(map, start, end, &count,
2191 if (start_entry == NULL) {
2193 vm_map_entry_release(count);
2194 return (KERN_INVALID_ADDRESS);
2197 if (new_pageable == 0) {
2198 entry = start_entry;
2199 while ((entry != &map->header) && (entry->start < end)) {
2200 vm_offset_t save_start;
2201 vm_offset_t save_end;
2204 * Already user wired or hard wired (trivial cases)
2206 if (entry->eflags & MAP_ENTRY_USER_WIRED) {
2207 entry = entry->next;
2210 if (entry->wired_count != 0) {
2211 entry->wired_count++;
2212 entry->eflags |= MAP_ENTRY_USER_WIRED;
2213 entry = entry->next;
2218 * A new wiring requires instantiation of appropriate
2219 * management structures and the faulting in of the
2222 if (entry->maptype == VM_MAPTYPE_NORMAL ||
2223 entry->maptype == VM_MAPTYPE_VPAGETABLE) {
2224 int copyflag = entry->eflags &
2225 MAP_ENTRY_NEEDS_COPY;
2226 if (copyflag && ((entry->protection &
2227 VM_PROT_WRITE) != 0)) {
2228 vm_map_entry_shadow(entry, 0);
2229 } else if (entry->object.vm_object == NULL &&
2231 vm_map_entry_allocate_object(entry);
2234 entry->wired_count++;
2235 entry->eflags |= MAP_ENTRY_USER_WIRED;
2238 * Now fault in the area. Note that vm_fault_wire()
2239 * may release the map lock temporarily, it will be
2240 * relocked on return. The in-transition
2241 * flag protects the entries.
2243 save_start = entry->start;
2244 save_end = entry->end;
2245 rv = vm_fault_wire(map, entry, TRUE, 0);
2247 CLIP_CHECK_BACK(entry, save_start);
2249 KASSERT(entry->wired_count == 1, ("bad wired_count on entry"));
2250 entry->eflags &= ~MAP_ENTRY_USER_WIRED;
2251 entry->wired_count = 0;
2252 if (entry->end == save_end)
2254 entry = entry->next;
2255 KASSERT(entry != &map->header, ("bad entry clip during backout"));
2257 end = save_start; /* unwire the rest */
2261 * note that even though the entry might have been
2262 * clipped, the USER_WIRED flag we set prevents
2263 * duplication so we do not have to do a
2266 entry = entry->next;
2270 * If we failed fall through to the unwiring section to
2271 * unwire what we had wired so far. 'end' has already
2278 * start_entry might have been clipped if we unlocked the
2279 * map and blocked. No matter how clipped it has gotten
2280 * there should be a fragment that is on our start boundary.
2282 CLIP_CHECK_BACK(start_entry, start);
2286 * Deal with the unwiring case.
2290 * This is the unwiring case. We must first ensure that the
2291 * range to be unwired is really wired down. We know there
2294 entry = start_entry;
2295 while ((entry != &map->header) && (entry->start < end)) {
2296 if ((entry->eflags & MAP_ENTRY_USER_WIRED) == 0) {
2297 rv = KERN_INVALID_ARGUMENT;
2300 KASSERT(entry->wired_count != 0, ("wired count was 0 with USER_WIRED set! %p", entry));
2301 entry = entry->next;
2305 * Now decrement the wiring count for each region. If a region
2306 * becomes completely unwired, unwire its physical pages and
2310 * The map entries are processed in a loop, checking to
2311 * make sure the entry is wired and asserting it has a wired
2312 * count. However, another loop was inserted more-or-less in
2313 * the middle of the unwiring path. This loop picks up the
2314 * "entry" loop variable from the first loop without first
2315 * setting it to start_entry. Naturally, the secound loop
2316 * is never entered and the pages backing the entries are
2317 * never unwired. This can lead to a leak of wired pages.
2319 entry = start_entry;
2320 while ((entry != &map->header) && (entry->start < end)) {
2321 KASSERT(entry->eflags & MAP_ENTRY_USER_WIRED,
2322 ("expected USER_WIRED on entry %p", entry));
2323 entry->eflags &= ~MAP_ENTRY_USER_WIRED;
2324 entry->wired_count--;
2325 if (entry->wired_count == 0)
2326 vm_fault_unwire(map, entry);
2327 entry = entry->next;
2331 vm_map_unclip_range(map, start_entry, start, real_end, &count,
2335 vm_map_entry_release(count);
2340 * Sets the pageability of the specified address range in the target map.
2341 * Regions specified as not pageable require locked-down physical
2342 * memory and physical page maps.
2344 * The map must not be locked, but a reference must remain to the map
2345 * throughout the call.
2347 * This function may be called via the zalloc path and must properly
2348 * reserve map entries for kernel_map.
2353 vm_map_wire(vm_map_t map, vm_offset_t start, vm_offset_t real_end, int kmflags)
2355 vm_map_entry_t entry;
2356 vm_map_entry_t start_entry;
2358 int rv = KERN_SUCCESS;
2361 if (kmflags & KM_KRESERVE)
2362 count = vm_map_entry_kreserve(MAP_RESERVE_COUNT);
2364 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
2366 VM_MAP_RANGE_CHECK(map, start, real_end);
2369 start_entry = vm_map_clip_range(map, start, end, &count,
2371 if (start_entry == NULL) {
2373 rv = KERN_INVALID_ADDRESS;
2376 if ((kmflags & KM_PAGEABLE) == 0) {
2380 * 1. Holding the write lock, we create any shadow or zero-fill
2381 * objects that need to be created. Then we clip each map
2382 * entry to the region to be wired and increment its wiring
2383 * count. We create objects before clipping the map entries
2384 * to avoid object proliferation.
2386 * 2. We downgrade to a read lock, and call vm_fault_wire to
2387 * fault in the pages for any newly wired area (wired_count is
2390 * Downgrading to a read lock for vm_fault_wire avoids a
2391 * possible deadlock with another process that may have faulted
2392 * on one of the pages to be wired (it would mark the page busy,
2393 * blocking us, then in turn block on the map lock that we
2394 * hold). Because of problems in the recursive lock package,
2395 * we cannot upgrade to a write lock in vm_map_lookup. Thus,
2396 * any actions that require the write lock must be done
2397 * beforehand. Because we keep the read lock on the map, the
2398 * copy-on-write status of the entries we modify here cannot
2401 entry = start_entry;
2402 while ((entry != &map->header) && (entry->start < end)) {
2404 * Trivial case if the entry is already wired
2406 if (entry->wired_count) {
2407 entry->wired_count++;
2408 entry = entry->next;
2413 * The entry is being newly wired, we have to setup
2414 * appropriate management structures. A shadow
2415 * object is required for a copy-on-write region,
2416 * or a normal object for a zero-fill region. We
2417 * do not have to do this for entries that point to sub
2418 * maps because we won't hold the lock on the sub map.
2420 if (entry->maptype == VM_MAPTYPE_NORMAL ||
2421 entry->maptype == VM_MAPTYPE_VPAGETABLE) {
2422 int copyflag = entry->eflags &
2423 MAP_ENTRY_NEEDS_COPY;
2424 if (copyflag && ((entry->protection &
2425 VM_PROT_WRITE) != 0)) {
2426 vm_map_entry_shadow(entry, 0);
2427 } else if (entry->object.vm_object == NULL &&
2429 vm_map_entry_allocate_object(entry);
2433 entry->wired_count++;
2434 entry = entry->next;
2442 * HACK HACK HACK HACK
2444 * vm_fault_wire() temporarily unlocks the map to avoid
2445 * deadlocks. The in-transition flag from vm_map_clip_range
2446 * call should protect us from changes while the map is
2449 * NOTE: Previously this comment stated that clipping might
2450 * still occur while the entry is unlocked, but from
2451 * what I can tell it actually cannot.
2453 * It is unclear whether the CLIP_CHECK_*() calls
2454 * are still needed but we keep them in anyway.
2456 * HACK HACK HACK HACK
2459 entry = start_entry;
2460 while (entry != &map->header && entry->start < end) {
2462 * If vm_fault_wire fails for any page we need to undo
2463 * what has been done. We decrement the wiring count
2464 * for those pages which have not yet been wired (now)
2465 * and unwire those that have (later).
2467 vm_offset_t save_start = entry->start;
2468 vm_offset_t save_end = entry->end;
2470 if (entry->wired_count == 1)
2471 rv = vm_fault_wire(map, entry, FALSE, kmflags);
2473 CLIP_CHECK_BACK(entry, save_start);
2475 KASSERT(entry->wired_count == 1, ("wired_count changed unexpectedly"));
2476 entry->wired_count = 0;
2477 if (entry->end == save_end)
2479 entry = entry->next;
2480 KASSERT(entry != &map->header, ("bad entry clip during backout"));
2485 CLIP_CHECK_FWD(entry, save_end);
2486 entry = entry->next;
2490 * If a failure occured undo everything by falling through
2491 * to the unwiring code. 'end' has already been adjusted
2495 kmflags |= KM_PAGEABLE;
2498 * start_entry is still IN_TRANSITION but may have been
2499 * clipped since vm_fault_wire() unlocks and relocks the
2500 * map. No matter how clipped it has gotten there should
2501 * be a fragment that is on our start boundary.
2503 CLIP_CHECK_BACK(start_entry, start);
2506 if (kmflags & KM_PAGEABLE) {
2508 * This is the unwiring case. We must first ensure that the
2509 * range to be unwired is really wired down. We know there
2512 entry = start_entry;
2513 while ((entry != &map->header) && (entry->start < end)) {
2514 if (entry->wired_count == 0) {
2515 rv = KERN_INVALID_ARGUMENT;
2518 entry = entry->next;
2522 * Now decrement the wiring count for each region. If a region
2523 * becomes completely unwired, unwire its physical pages and
2526 entry = start_entry;
2527 while ((entry != &map->header) && (entry->start < end)) {
2528 entry->wired_count--;
2529 if (entry->wired_count == 0)
2530 vm_fault_unwire(map, entry);
2531 entry = entry->next;
2535 vm_map_unclip_range(map, start_entry, start, real_end,
2536 &count, MAP_CLIP_NO_HOLES);
2540 if (kmflags & KM_KRESERVE)
2541 vm_map_entry_krelease(count);
2543 vm_map_entry_release(count);
2548 * Mark a newly allocated address range as wired but do not fault in
2549 * the pages. The caller is expected to load the pages into the object.
2551 * The map must be locked on entry and will remain locked on return.
2552 * No other requirements.
2555 vm_map_set_wired_quick(vm_map_t map, vm_offset_t addr, vm_size_t size,
2558 vm_map_entry_t scan;
2559 vm_map_entry_t entry;
2561 entry = vm_map_clip_range(map, addr, addr + size,
2562 countp, MAP_CLIP_NO_HOLES);
2564 scan != &map->header && scan->start < addr + size;
2565 scan = scan->next) {
2566 KKASSERT(scan->wired_count == 0);
2567 scan->wired_count = 1;
2569 vm_map_unclip_range(map, entry, addr, addr + size,
2570 countp, MAP_CLIP_NO_HOLES);
2574 * Push any dirty cached pages in the address range to their pager.
2575 * If syncio is TRUE, dirty pages are written synchronously.
2576 * If invalidate is TRUE, any cached pages are freed as well.
2578 * This routine is called by sys_msync()
2580 * Returns an error if any part of the specified range is not mapped.
2585 vm_map_clean(vm_map_t map, vm_offset_t start, vm_offset_t end,
2586 boolean_t syncio, boolean_t invalidate)
2588 vm_map_entry_t current;
2589 vm_map_entry_t entry;
2593 vm_ooffset_t offset;
2595 vm_map_lock_read(map);
2596 VM_MAP_RANGE_CHECK(map, start, end);
2597 if (!vm_map_lookup_entry(map, start, &entry)) {
2598 vm_map_unlock_read(map);
2599 return (KERN_INVALID_ADDRESS);
2601 lwkt_gettoken(&map->token);
2604 * Make a first pass to check for holes.
2606 for (current = entry; current->start < end; current = current->next) {
2607 if (current->maptype == VM_MAPTYPE_SUBMAP) {
2608 lwkt_reltoken(&map->token);
2609 vm_map_unlock_read(map);
2610 return (KERN_INVALID_ARGUMENT);
2612 if (end > current->end &&
2613 (current->next == &map->header ||
2614 current->end != current->next->start)) {
2615 lwkt_reltoken(&map->token);
2616 vm_map_unlock_read(map);
2617 return (KERN_INVALID_ADDRESS);
2622 pmap_remove(vm_map_pmap(map), start, end);
2625 * Make a second pass, cleaning/uncaching pages from the indicated
2628 for (current = entry; current->start < end; current = current->next) {
2629 offset = current->offset + (start - current->start);
2630 size = (end <= current->end ? end : current->end) - start;
2632 switch(current->maptype) {
2633 case VM_MAPTYPE_SUBMAP:
2636 vm_map_entry_t tentry;
2639 smap = current->object.sub_map;
2640 vm_map_lock_read(smap);
2641 vm_map_lookup_entry(smap, offset, &tentry);
2642 tsize = tentry->end - offset;
2645 object = tentry->object.vm_object;
2646 offset = tentry->offset + (offset - tentry->start);
2647 vm_map_unlock_read(smap);
2650 case VM_MAPTYPE_NORMAL:
2651 case VM_MAPTYPE_VPAGETABLE:
2652 object = current->object.vm_object;
2660 vm_object_hold(object);
2663 * Note that there is absolutely no sense in writing out
2664 * anonymous objects, so we track down the vnode object
2666 * We invalidate (remove) all pages from the address space
2667 * anyway, for semantic correctness.
2669 * note: certain anonymous maps, such as MAP_NOSYNC maps,
2670 * may start out with a NULL object.
2672 while (object && (tobj = object->backing_object) != NULL) {
2673 vm_object_hold(tobj);
2674 if (tobj == object->backing_object) {
2675 vm_object_lock_swap();
2676 offset += object->backing_object_offset;
2677 vm_object_drop(object);
2679 if (object->size < OFF_TO_IDX(offset + size))
2680 size = IDX_TO_OFF(object->size) -
2684 vm_object_drop(tobj);
2686 if (object && (object->type == OBJT_VNODE) &&
2687 (current->protection & VM_PROT_WRITE) &&
2688 (object->flags & OBJ_NOMSYNC) == 0) {
2690 * Flush pages if writing is allowed, invalidate them
2691 * if invalidation requested. Pages undergoing I/O
2692 * will be ignored by vm_object_page_remove().
2694 * We cannot lock the vnode and then wait for paging
2695 * to complete without deadlocking against vm_fault.
2696 * Instead we simply call vm_object_page_remove() and
2697 * allow it to block internally on a page-by-page
2698 * basis when it encounters pages undergoing async
2703 /* no chain wait needed for vnode objects */
2704 vm_object_reference_locked(object);
2705 vn_lock(object->handle, LK_EXCLUSIVE | LK_RETRY);
2706 flags = (syncio || invalidate) ? OBJPC_SYNC : 0;
2707 flags |= invalidate ? OBJPC_INVAL : 0;
2710 * When operating on a virtual page table just
2711 * flush the whole object. XXX we probably ought
2714 switch(current->maptype) {
2715 case VM_MAPTYPE_NORMAL:
2716 vm_object_page_clean(object,
2718 OFF_TO_IDX(offset + size + PAGE_MASK),
2721 case VM_MAPTYPE_VPAGETABLE:
2722 vm_object_page_clean(object, 0, 0, flags);
2725 vn_unlock(((struct vnode *)object->handle));
2726 vm_object_deallocate_locked(object);
2728 if (object && invalidate &&
2729 ((object->type == OBJT_VNODE) ||
2730 (object->type == OBJT_DEVICE) ||
2731 (object->type == OBJT_MGTDEVICE))) {
2733 ((object->type == OBJT_DEVICE) ||
2734 (object->type == OBJT_MGTDEVICE)) ? FALSE : TRUE;
2735 /* no chain wait needed for vnode/device objects */
2736 vm_object_reference_locked(object);
2737 switch(current->maptype) {
2738 case VM_MAPTYPE_NORMAL:
2739 vm_object_page_remove(object,
2741 OFF_TO_IDX(offset + size + PAGE_MASK),
2744 case VM_MAPTYPE_VPAGETABLE:
2745 vm_object_page_remove(object, 0, 0, clean_only);
2748 vm_object_deallocate_locked(object);
2752 vm_object_drop(object);
2755 lwkt_reltoken(&map->token);
2756 vm_map_unlock_read(map);
2758 return (KERN_SUCCESS);
2762 * Make the region specified by this entry pageable.
2764 * The vm_map must be exclusively locked.
2767 vm_map_entry_unwire(vm_map_t map, vm_map_entry_t entry)
2769 entry->eflags &= ~MAP_ENTRY_USER_WIRED;
2770 entry->wired_count = 0;
2771 vm_fault_unwire(map, entry);
2775 * Deallocate the given entry from the target map.
2777 * The vm_map must be exclusively locked.
2780 vm_map_entry_delete(vm_map_t map, vm_map_entry_t entry, int *countp)
2782 vm_map_entry_unlink(map, entry);
2783 map->size -= entry->end - entry->start;
2785 switch(entry->maptype) {
2786 case VM_MAPTYPE_NORMAL:
2787 case VM_MAPTYPE_VPAGETABLE:
2788 case VM_MAPTYPE_SUBMAP:
2789 vm_object_deallocate(entry->object.vm_object);
2791 case VM_MAPTYPE_UKSMAP:
2798 vm_map_entry_dispose(map, entry, countp);
2802 * Deallocates the given address range from the target map.
2804 * The vm_map must be exclusively locked.
2807 vm_map_delete(vm_map_t map, vm_offset_t start, vm_offset_t end, int *countp)
2810 vm_map_entry_t entry;
2811 vm_map_entry_t first_entry;
2813 ASSERT_VM_MAP_LOCKED(map);
2814 lwkt_gettoken(&map->token);
2817 * Find the start of the region, and clip it. Set entry to point
2818 * at the first record containing the requested address or, if no
2819 * such record exists, the next record with a greater address. The
2820 * loop will run from this point until a record beyond the termination
2821 * address is encountered.
2823 * map->hint must be adjusted to not point to anything we delete,
2824 * so set it to the entry prior to the one being deleted.
2826 * GGG see other GGG comment.
2828 if (vm_map_lookup_entry(map, start, &first_entry)) {
2829 entry = first_entry;
2830 vm_map_clip_start(map, entry, start, countp);
2831 map->hint = entry->prev; /* possible problem XXX */
2833 map->hint = first_entry; /* possible problem XXX */
2834 entry = first_entry->next;
2838 * If a hole opens up prior to the current first_free then
2839 * adjust first_free. As with map->hint, map->first_free
2840 * cannot be left set to anything we might delete.
2842 if (entry == &map->header) {
2843 map->first_free = &map->header;
2844 } else if (map->first_free->start >= start) {
2845 map->first_free = entry->prev;
2849 * Step through all entries in this region
2851 while ((entry != &map->header) && (entry->start < end)) {
2852 vm_map_entry_t next;
2854 vm_pindex_t offidxstart, offidxend, count;
2857 * If we hit an in-transition entry we have to sleep and
2858 * retry. It's easier (and not really slower) to just retry
2859 * since this case occurs so rarely and the hint is already
2860 * pointing at the right place. We have to reset the
2861 * start offset so as not to accidently delete an entry
2862 * another process just created in vacated space.
2864 if (entry->eflags & MAP_ENTRY_IN_TRANSITION) {
2865 entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP;
2866 start = entry->start;
2867 ++mycpu->gd_cnt.v_intrans_coll;
2868 ++mycpu->gd_cnt.v_intrans_wait;
2869 vm_map_transition_wait(map);
2872 vm_map_clip_end(map, entry, end, countp);
2878 offidxstart = OFF_TO_IDX(entry->offset);
2879 count = OFF_TO_IDX(e - s);
2881 switch(entry->maptype) {
2882 case VM_MAPTYPE_NORMAL:
2883 case VM_MAPTYPE_VPAGETABLE:
2884 case VM_MAPTYPE_SUBMAP:
2885 object = entry->object.vm_object;
2893 * Unwire before removing addresses from the pmap; otherwise,
2894 * unwiring will put the entries back in the pmap.
2896 if (entry->wired_count != 0)
2897 vm_map_entry_unwire(map, entry);
2899 offidxend = offidxstart + count;
2901 if (object == &kernel_object) {
2902 vm_object_hold(object);
2903 vm_object_page_remove(object, offidxstart,
2905 vm_object_drop(object);
2906 } else if (object && object->type != OBJT_DEFAULT &&
2907 object->type != OBJT_SWAP) {
2909 * vnode object routines cannot be chain-locked,
2910 * but since we aren't removing pages from the
2911 * object here we can use a shared hold.
2913 vm_object_hold_shared(object);
2914 pmap_remove(map->pmap, s, e);
2915 vm_object_drop(object);
2916 } else if (object) {
2917 vm_object_hold(object);
2918 vm_object_chain_acquire(object, 0);
2919 pmap_remove(map->pmap, s, e);
2921 if (object != NULL &&
2922 object->ref_count != 1 &&
2923 (object->flags & (OBJ_NOSPLIT|OBJ_ONEMAPPING)) ==
2925 (object->type == OBJT_DEFAULT ||
2926 object->type == OBJT_SWAP)) {
2927 vm_object_collapse(object, NULL);
2928 vm_object_page_remove(object, offidxstart,
2930 if (object->type == OBJT_SWAP) {
2931 swap_pager_freespace(object,
2935 if (offidxend >= object->size &&
2936 offidxstart < object->size) {
2937 object->size = offidxstart;
2940 vm_object_chain_release(object);
2941 vm_object_drop(object);
2942 } else if (entry->maptype == VM_MAPTYPE_UKSMAP) {
2943 pmap_remove(map->pmap, s, e);
2947 * Delete the entry (which may delete the object) only after
2948 * removing all pmap entries pointing to its pages.
2949 * (Otherwise, its page frames may be reallocated, and any
2950 * modify bits will be set in the wrong object!)
2952 vm_map_entry_delete(map, entry, countp);
2955 lwkt_reltoken(&map->token);
2956 return (KERN_SUCCESS);
2960 * Remove the given address range from the target map.
2961 * This is the exported form of vm_map_delete.
2966 vm_map_remove(vm_map_t map, vm_offset_t start, vm_offset_t end)
2971 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
2973 VM_MAP_RANGE_CHECK(map, start, end);
2974 result = vm_map_delete(map, start, end, &count);
2976 vm_map_entry_release(count);
2982 * Assert that the target map allows the specified privilege on the
2983 * entire address region given. The entire region must be allocated.
2985 * The caller must specify whether the vm_map is already locked or not.
2988 vm_map_check_protection(vm_map_t map, vm_offset_t start, vm_offset_t end,
2989 vm_prot_t protection, boolean_t have_lock)
2991 vm_map_entry_t entry;
2992 vm_map_entry_t tmp_entry;
2995 if (have_lock == FALSE)
2996 vm_map_lock_read(map);
2998 if (!vm_map_lookup_entry(map, start, &tmp_entry)) {
2999 if (have_lock == FALSE)
3000 vm_map_unlock_read(map);
3006 while (start < end) {
3007 if (entry == &map->header) {
3015 if (start < entry->start) {
3020 * Check protection associated with entry.
3023 if ((entry->protection & protection) != protection) {
3027 /* go to next entry */
3030 entry = entry->next;
3032 if (have_lock == FALSE)
3033 vm_map_unlock_read(map);
3038 * If appropriate this function shadows the original object with a new object
3039 * and moves the VM pages from the original object to the new object.
3040 * The original object will also be collapsed, if possible.
3042 * We can only do this for normal memory objects with a single mapping, and
3043 * it only makes sense to do it if there are 2 or more refs on the original
3044 * object. i.e. typically a memory object that has been extended into
3045 * multiple vm_map_entry's with non-overlapping ranges.
3047 * This makes it easier to remove unused pages and keeps object inheritance
3048 * from being a negative impact on memory usage.
3050 * On return the (possibly new) entry->object.vm_object will have an
3051 * additional ref on it for the caller to dispose of (usually by cloning
3052 * the vm_map_entry). The additional ref had to be done in this routine
3053 * to avoid racing a collapse. The object's ONEMAPPING flag will also be
3056 * The vm_map must be locked and its token held.
3059 vm_map_split(vm_map_entry_t entry)
3062 vm_object_t oobject, nobject, bobject;
3065 vm_pindex_t offidxstart, offidxend, idx;
3067 vm_ooffset_t offset;
3071 * Optimize away object locks for vnode objects. Important exit/exec
3074 * OBJ_ONEMAPPING doesn't apply to vnode objects but clear the flag
3077 oobject = entry->object.vm_object;
3078 if (oobject->type != OBJT_DEFAULT && oobject->type != OBJT_SWAP) {
3079 vm_object_reference_quick(oobject);
3080 vm_object_clear_flag(oobject, OBJ_ONEMAPPING);
3085 * Setup. Chain lock the original object throughout the entire
3086 * routine to prevent new page faults from occuring.
3088 * XXX can madvise WILLNEED interfere with us too?
3090 vm_object_hold(oobject);
3091 vm_object_chain_acquire(oobject, 0);
3094 * Original object cannot be split? Might have also changed state.
3096 if (oobject->handle == NULL || (oobject->type != OBJT_DEFAULT &&
3097 oobject->type != OBJT_SWAP)) {
3098 vm_object_chain_release(oobject);
3099 vm_object_reference_locked(oobject);
3100 vm_object_clear_flag(oobject, OBJ_ONEMAPPING);
3101 vm_object_drop(oobject);
3106 * Collapse original object with its backing store as an
3107 * optimization to reduce chain lengths when possible.
3109 * If ref_count <= 1 there aren't other non-overlapping vm_map_entry's
3110 * for oobject, so there's no point collapsing it.
3112 * Then re-check whether the object can be split.
3114 vm_object_collapse(oobject, NULL);
3116 if (oobject->ref_count <= 1 ||
3117 (oobject->type != OBJT_DEFAULT && oobject->type != OBJT_SWAP) ||
3118 (oobject->flags & (OBJ_NOSPLIT|OBJ_ONEMAPPING)) != OBJ_ONEMAPPING) {
3119 vm_object_chain_release(oobject);
3120 vm_object_reference_locked(oobject);
3121 vm_object_clear_flag(oobject, OBJ_ONEMAPPING);
3122 vm_object_drop(oobject);
3127 * Acquire the chain lock on the backing object.
3129 * Give bobject an additional ref count for when it will be shadowed
3133 if ((bobject = oobject->backing_object) != NULL) {
3134 if (bobject->type != OBJT_VNODE) {
3136 vm_object_hold(bobject);
3137 vm_object_chain_wait(bobject, 0);
3138 vm_object_reference_locked(bobject);
3139 vm_object_chain_acquire(bobject, 0);
3140 KKASSERT(bobject->backing_object == bobject);
3141 KKASSERT((bobject->flags & OBJ_DEAD) == 0);
3143 vm_object_reference_quick(bobject);
3148 * Calculate the object page range and allocate the new object.
3150 offset = entry->offset;
3154 offidxstart = OFF_TO_IDX(offset);
3155 offidxend = offidxstart + OFF_TO_IDX(e - s);
3156 size = offidxend - offidxstart;
3158 switch(oobject->type) {
3160 nobject = default_pager_alloc(NULL, IDX_TO_OFF(size),
3164 nobject = swap_pager_alloc(NULL, IDX_TO_OFF(size),
3173 if (nobject == NULL) {
3175 if (useshadowlist) {
3176 vm_object_chain_release(bobject);
3177 vm_object_deallocate(bobject);
3178 vm_object_drop(bobject);
3180 vm_object_deallocate(bobject);
3183 vm_object_chain_release(oobject);
3184 vm_object_reference_locked(oobject);
3185 vm_object_clear_flag(oobject, OBJ_ONEMAPPING);
3186 vm_object_drop(oobject);
3191 * The new object will replace entry->object.vm_object so it needs
3192 * a second reference (the caller expects an additional ref).
3194 vm_object_hold(nobject);
3195 vm_object_reference_locked(nobject);
3196 vm_object_chain_acquire(nobject, 0);
3199 * nobject shadows bobject (oobject already shadows bobject).
3202 nobject->backing_object_offset =
3203 oobject->backing_object_offset + IDX_TO_OFF(offidxstart);
3204 nobject->backing_object = bobject;
3205 if (useshadowlist) {
3206 bobject->shadow_count++;
3207 bobject->generation++;
3208 LIST_INSERT_HEAD(&bobject->shadow_head,
3209 nobject, shadow_list);
3210 vm_object_clear_flag(bobject, OBJ_ONEMAPPING); /*XXX*/
3211 vm_object_chain_release(bobject);
3212 vm_object_drop(bobject);
3213 vm_object_set_flag(nobject, OBJ_ONSHADOW);
3218 * Move the VM pages from oobject to nobject
3220 for (idx = 0; idx < size; idx++) {
3223 m = vm_page_lookup_busy_wait(oobject, offidxstart + idx,
3229 * We must wait for pending I/O to complete before we can
3232 * We do not have to VM_PROT_NONE the page as mappings should
3233 * not be changed by this operation.
3235 * NOTE: The act of renaming a page updates chaingen for both
3238 vm_page_rename(m, nobject, idx);
3239 /* page automatically made dirty by rename and cache handled */
3240 /* page remains busy */
3243 if (oobject->type == OBJT_SWAP) {
3244 vm_object_pip_add(oobject, 1);
3246 * copy oobject pages into nobject and destroy unneeded
3247 * pages in shadow object.
3249 swap_pager_copy(oobject, nobject, offidxstart, 0);
3250 vm_object_pip_wakeup(oobject);
3254 * Wakeup the pages we played with. No spl protection is needed
3255 * for a simple wakeup.
3257 for (idx = 0; idx < size; idx++) {
3258 m = vm_page_lookup(nobject, idx);
3260 KKASSERT(m->flags & PG_BUSY);
3264 entry->object.vm_object = nobject;
3265 entry->offset = 0LL;
3270 * NOTE: There is no need to remove OBJ_ONEMAPPING from oobject, the
3271 * related pages were moved and are no longer applicable to the
3274 * NOTE: Deallocate oobject (due to its entry->object.vm_object being
3275 * replaced by nobject).
3277 vm_object_chain_release(nobject);
3278 vm_object_drop(nobject);
3279 if (bobject && useshadowlist) {
3280 vm_object_chain_release(bobject);
3281 vm_object_drop(bobject);
3283 vm_object_chain_release(oobject);
3284 /*vm_object_clear_flag(oobject, OBJ_ONEMAPPING);*/
3285 vm_object_deallocate_locked(oobject);
3286 vm_object_drop(oobject);
3290 * Copies the contents of the source entry to the destination
3291 * entry. The entries *must* be aligned properly.
3293 * The vm_maps must be exclusively locked.
3294 * The vm_map's token must be held.
3296 * Because the maps are locked no faults can be in progress during the
3300 vm_map_copy_entry(vm_map_t src_map, vm_map_t dst_map,
3301 vm_map_entry_t src_entry, vm_map_entry_t dst_entry)
3303 vm_object_t src_object;
3305 if (dst_entry->maptype == VM_MAPTYPE_SUBMAP ||
3306 dst_entry->maptype == VM_MAPTYPE_UKSMAP)
3308 if (src_entry->maptype == VM_MAPTYPE_SUBMAP ||
3309 src_entry->maptype == VM_MAPTYPE_UKSMAP)
3312 if (src_entry->wired_count == 0) {
3314 * If the source entry is marked needs_copy, it is already
3317 if ((src_entry->eflags & MAP_ENTRY_NEEDS_COPY) == 0) {
3318 pmap_protect(src_map->pmap,
3321 src_entry->protection & ~VM_PROT_WRITE);
3325 * Make a copy of the object.
3327 * The object must be locked prior to checking the object type
3328 * and for the call to vm_object_collapse() and vm_map_split().
3329 * We cannot use *_hold() here because the split code will
3330 * probably try to destroy the object. The lock is a pool
3331 * token and doesn't care.
3333 * We must bump src_map->timestamp when setting
3334 * MAP_ENTRY_NEEDS_COPY to force any concurrent fault
3335 * to retry, otherwise the concurrent fault might improperly
3336 * install a RW pte when its supposed to be a RO(COW) pte.
3337 * This race can occur because a vnode-backed fault may have
3338 * to temporarily release the map lock.
3340 if (src_entry->object.vm_object != NULL) {
3341 vm_map_split(src_entry);
3342 src_object = src_entry->object.vm_object;
3343 dst_entry->object.vm_object = src_object;
3344 src_entry->eflags |= (MAP_ENTRY_COW |
3345 MAP_ENTRY_NEEDS_COPY);
3346 dst_entry->eflags |= (MAP_ENTRY_COW |
3347 MAP_ENTRY_NEEDS_COPY);
3348 dst_entry->offset = src_entry->offset;
3349 ++src_map->timestamp;
3351 dst_entry->object.vm_object = NULL;
3352 dst_entry->offset = 0;
3355 pmap_copy(dst_map->pmap, src_map->pmap, dst_entry->start,
3356 dst_entry->end - dst_entry->start, src_entry->start);
3359 * Of course, wired down pages can't be set copy-on-write.
3360 * Cause wired pages to be copied into the new map by
3361 * simulating faults (the new pages are pageable)
3363 vm_fault_copy_entry(dst_map, src_map, dst_entry, src_entry);
3369 * Create a new process vmspace structure and vm_map
3370 * based on those of an existing process. The new map
3371 * is based on the old map, according to the inheritance
3372 * values on the regions in that map.
3374 * The source map must not be locked.
3377 static void vmspace_fork_normal_entry(vm_map_t old_map, vm_map_t new_map,
3378 vm_map_entry_t old_entry, int *countp);
3379 static void vmspace_fork_uksmap_entry(vm_map_t old_map, vm_map_t new_map,
3380 vm_map_entry_t old_entry, int *countp);
3383 vmspace_fork(struct vmspace *vm1)
3385 struct vmspace *vm2;
3386 vm_map_t old_map = &vm1->vm_map;
3388 vm_map_entry_t old_entry;
3391 lwkt_gettoken(&vm1->vm_map.token);
3392 vm_map_lock(old_map);
3394 vm2 = vmspace_alloc(old_map->min_offset, old_map->max_offset);
3395 lwkt_gettoken(&vm2->vm_map.token);
3396 bcopy(&vm1->vm_startcopy, &vm2->vm_startcopy,
3397 (caddr_t)&vm1->vm_endcopy - (caddr_t)&vm1->vm_startcopy);
3398 new_map = &vm2->vm_map; /* XXX */
3399 new_map->timestamp = 1;
3401 vm_map_lock(new_map);
3404 old_entry = old_map->header.next;
3405 while (old_entry != &old_map->header) {
3407 old_entry = old_entry->next;
3410 count = vm_map_entry_reserve(count + MAP_RESERVE_COUNT);
3412 old_entry = old_map->header.next;
3413 while (old_entry != &old_map->header) {
3414 switch(old_entry->maptype) {
3415 case VM_MAPTYPE_SUBMAP:
3416 panic("vm_map_fork: encountered a submap");
3418 case VM_MAPTYPE_UKSMAP:
3419 vmspace_fork_uksmap_entry(old_map, new_map,
3422 case VM_MAPTYPE_NORMAL:
3423 case VM_MAPTYPE_VPAGETABLE:
3424 vmspace_fork_normal_entry(old_map, new_map,
3428 old_entry = old_entry->next;
3431 new_map->size = old_map->size;
3432 vm_map_unlock(old_map);
3433 vm_map_unlock(new_map);
3434 vm_map_entry_release(count);
3436 lwkt_reltoken(&vm2->vm_map.token);
3437 lwkt_reltoken(&vm1->vm_map.token);
3444 vmspace_fork_normal_entry(vm_map_t old_map, vm_map_t new_map,
3445 vm_map_entry_t old_entry, int *countp)
3447 vm_map_entry_t new_entry;
3450 switch (old_entry->inheritance) {
3451 case VM_INHERIT_NONE:
3453 case VM_INHERIT_SHARE:
3455 * Clone the entry, creating the shared object if
3458 if (old_entry->object.vm_object == NULL)
3459 vm_map_entry_allocate_object(old_entry);
3461 if (old_entry->eflags & MAP_ENTRY_NEEDS_COPY) {
3463 * Shadow a map_entry which needs a copy,
3464 * replacing its object with a new object
3465 * that points to the old one. Ask the
3466 * shadow code to automatically add an
3467 * additional ref. We can't do it afterwords
3468 * because we might race a collapse. The call
3469 * to vm_map_entry_shadow() will also clear
3472 vm_map_entry_shadow(old_entry, 1);
3473 } else if (old_entry->object.vm_object) {
3475 * We will make a shared copy of the object,
3476 * and must clear OBJ_ONEMAPPING.
3478 * Optimize vnode objects. OBJ_ONEMAPPING
3479 * is non-applicable but clear it anyway,
3480 * and its terminal so we don'th ave to deal
3481 * with chains. Reduces SMP conflicts.
3483 * XXX assert that object.vm_object != NULL
3484 * since we allocate it above.
3486 object = old_entry->object.vm_object;
3487 if (object->type == OBJT_VNODE) {
3488 vm_object_reference_quick(object);
3489 vm_object_clear_flag(object,
3492 vm_object_hold(object);
3493 vm_object_chain_wait(object, 0);
3494 vm_object_reference_locked(object);
3495 vm_object_clear_flag(object,
3497 vm_object_drop(object);
3502 * Clone the entry. We've already bumped the ref on
3505 new_entry = vm_map_entry_create(new_map, countp);
3506 *new_entry = *old_entry;
3507 new_entry->eflags &= ~MAP_ENTRY_USER_WIRED;
3508 new_entry->wired_count = 0;
3511 * Insert the entry into the new map -- we know we're
3512 * inserting at the end of the new map.
3515 vm_map_entry_link(new_map, new_map->header.prev,
3519 * Update the physical map
3521 pmap_copy(new_map->pmap, old_map->pmap,
3523 (old_entry->end - old_entry->start),
3526 case VM_INHERIT_COPY:
3528 * Clone the entry and link into the map.
3530 new_entry = vm_map_entry_create(new_map, countp);
3531 *new_entry = *old_entry;
3532 new_entry->eflags &= ~MAP_ENTRY_USER_WIRED;
3533 new_entry->wired_count = 0;
3534 new_entry->object.vm_object = NULL;
3535 vm_map_entry_link(new_map, new_map->header.prev,
3537 vm_map_copy_entry(old_map, new_map, old_entry,
3544 * When forking user-kernel shared maps, the map might change in the
3545 * child so do not try to copy the underlying pmap entries.
3549 vmspace_fork_uksmap_entry(vm_map_t old_map, vm_map_t new_map,
3550 vm_map_entry_t old_entry, int *countp)
3552 vm_map_entry_t new_entry;
3554 new_entry = vm_map_entry_create(new_map, countp);
3555 *new_entry = *old_entry;
3556 new_entry->eflags &= ~MAP_ENTRY_USER_WIRED;
3557 new_entry->wired_count = 0;
3558 vm_map_entry_link(new_map, new_map->header.prev,
3563 * Create an auto-grow stack entry
3568 vm_map_stack (vm_map_t map, vm_offset_t addrbos, vm_size_t max_ssize,
3569 int flags, vm_prot_t prot, vm_prot_t max, int cow)
3571 vm_map_entry_t prev_entry;
3572 vm_map_entry_t new_stack_entry;
3573 vm_size_t init_ssize;
3576 vm_offset_t tmpaddr;
3578 cow |= MAP_IS_STACK;
3580 if (max_ssize < sgrowsiz)
3581 init_ssize = max_ssize;
3583 init_ssize = sgrowsiz;
3585 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
3589 * Find space for the mapping
3591 if ((flags & (MAP_FIXED | MAP_TRYFIXED)) == 0) {
3592 if (vm_map_findspace(map, addrbos, max_ssize, 1,
3595 vm_map_entry_release(count);
3596 return (KERN_NO_SPACE);
3601 /* If addr is already mapped, no go */
3602 if (vm_map_lookup_entry(map, addrbos, &prev_entry)) {
3604 vm_map_entry_release(count);
3605 return (KERN_NO_SPACE);
3609 /* XXX already handled by kern_mmap() */
3610 /* If we would blow our VMEM resource limit, no go */
3611 if (map->size + init_ssize >
3612 curproc->p_rlimit[RLIMIT_VMEM].rlim_cur) {
3614 vm_map_entry_release(count);
3615 return (KERN_NO_SPACE);
3620 * If we can't accomodate max_ssize in the current mapping,
3621 * no go. However, we need to be aware that subsequent user
3622 * mappings might map into the space we have reserved for
3623 * stack, and currently this space is not protected.
3625 * Hopefully we will at least detect this condition
3626 * when we try to grow the stack.
3628 if ((prev_entry->next != &map->header) &&
3629 (prev_entry->next->start < addrbos + max_ssize)) {
3631 vm_map_entry_release(count);
3632 return (KERN_NO_SPACE);
3636 * We initially map a stack of only init_ssize. We will
3637 * grow as needed later. Since this is to be a grow
3638 * down stack, we map at the top of the range.
3640 * Note: we would normally expect prot and max to be
3641 * VM_PROT_ALL, and cow to be 0. Possibly we should
3642 * eliminate these as input parameters, and just
3643 * pass these values here in the insert call.
3645 rv = vm_map_insert(map, &count, NULL, NULL,
3646 0, addrbos + max_ssize - init_ssize,
3647 addrbos + max_ssize,
3651 /* Now set the avail_ssize amount */
3652 if (rv == KERN_SUCCESS) {
3653 if (prev_entry != &map->header)
3654 vm_map_clip_end(map, prev_entry, addrbos + max_ssize - init_ssize, &count);
3655 new_stack_entry = prev_entry->next;
3656 if (new_stack_entry->end != addrbos + max_ssize ||
3657 new_stack_entry->start != addrbos + max_ssize - init_ssize)
3658 panic ("Bad entry start/end for new stack entry");
3660 new_stack_entry->aux.avail_ssize = max_ssize - init_ssize;
3664 vm_map_entry_release(count);
3669 * Attempts to grow a vm stack entry. Returns KERN_SUCCESS if the
3670 * desired address is already mapped, or if we successfully grow
3671 * the stack. Also returns KERN_SUCCESS if addr is outside the
3672 * stack range (this is strange, but preserves compatibility with
3673 * the grow function in vm_machdep.c).
3678 vm_map_growstack (struct proc *p, vm_offset_t addr)
3680 vm_map_entry_t prev_entry;
3681 vm_map_entry_t stack_entry;
3682 vm_map_entry_t new_stack_entry;
3683 struct vmspace *vm = p->p_vmspace;
3684 vm_map_t map = &vm->vm_map;
3687 int rv = KERN_SUCCESS;
3689 int use_read_lock = 1;
3692 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
3695 vm_map_lock_read(map);
3699 /* If addr is already in the entry range, no need to grow.*/
3700 if (vm_map_lookup_entry(map, addr, &prev_entry))
3703 if ((stack_entry = prev_entry->next) == &map->header)
3705 if (prev_entry == &map->header)
3706 end = stack_entry->start - stack_entry->aux.avail_ssize;
3708 end = prev_entry->end;
3711 * This next test mimics the old grow function in vm_machdep.c.
3712 * It really doesn't quite make sense, but we do it anyway
3713 * for compatibility.
3715 * If not growable stack, return success. This signals the
3716 * caller to proceed as he would normally with normal vm.
3718 if (stack_entry->aux.avail_ssize < 1 ||
3719 addr >= stack_entry->start ||
3720 addr < stack_entry->start - stack_entry->aux.avail_ssize) {
3724 /* Find the minimum grow amount */
3725 grow_amount = roundup (stack_entry->start - addr, PAGE_SIZE);
3726 if (grow_amount > stack_entry->aux.avail_ssize) {
3732 * If there is no longer enough space between the entries
3733 * nogo, and adjust the available space. Note: this
3734 * should only happen if the user has mapped into the
3735 * stack area after the stack was created, and is
3736 * probably an error.
3738 * This also effectively destroys any guard page the user
3739 * might have intended by limiting the stack size.
3741 if (grow_amount > stack_entry->start - end) {
3742 if (use_read_lock && vm_map_lock_upgrade(map)) {
3748 stack_entry->aux.avail_ssize = stack_entry->start - end;
3753 is_procstack = addr >= (vm_offset_t)vm->vm_maxsaddr;
3755 /* If this is the main process stack, see if we're over the
3758 if (is_procstack && (ctob(vm->vm_ssize) + grow_amount >
3759 p->p_rlimit[RLIMIT_STACK].rlim_cur)) {
3764 /* Round up the grow amount modulo SGROWSIZ */
3765 grow_amount = roundup (grow_amount, sgrowsiz);
3766 if (grow_amount > stack_entry->aux.avail_ssize) {
3767 grow_amount = stack_entry->aux.avail_ssize;
3769 if (is_procstack && (ctob(vm->vm_ssize) + grow_amount >
3770 p->p_rlimit[RLIMIT_STACK].rlim_cur)) {
3771 grow_amount = p->p_rlimit[RLIMIT_STACK].rlim_cur -
3775 /* If we would blow our VMEM resource limit, no go */
3776 if (map->size + grow_amount > p->p_rlimit[RLIMIT_VMEM].rlim_cur) {
3781 if (use_read_lock && vm_map_lock_upgrade(map)) {
3788 /* Get the preliminary new entry start value */
3789 addr = stack_entry->start - grow_amount;
3791 /* If this puts us into the previous entry, cut back our growth
3792 * to the available space. Also, see the note above.
3795 stack_entry->aux.avail_ssize = stack_entry->start - end;
3799 rv = vm_map_insert(map, &count, NULL, NULL,
3800 0, addr, stack_entry->start,
3802 VM_PROT_ALL, VM_PROT_ALL, 0);
3804 /* Adjust the available stack space by the amount we grew. */
3805 if (rv == KERN_SUCCESS) {
3806 if (prev_entry != &map->header)
3807 vm_map_clip_end(map, prev_entry, addr, &count);
3808 new_stack_entry = prev_entry->next;
3809 if (new_stack_entry->end != stack_entry->start ||
3810 new_stack_entry->start != addr)
3811 panic ("Bad stack grow start/end in new stack entry");
3813 new_stack_entry->aux.avail_ssize =
3814 stack_entry->aux.avail_ssize -
3815 (new_stack_entry->end - new_stack_entry->start);
3817 vm->vm_ssize += btoc(new_stack_entry->end -
3818 new_stack_entry->start);
3821 if (map->flags & MAP_WIREFUTURE)
3822 vm_map_unwire(map, new_stack_entry->start,
3823 new_stack_entry->end, FALSE);
3828 vm_map_unlock_read(map);
3831 vm_map_entry_release(count);
3836 * Unshare the specified VM space for exec. If other processes are
3837 * mapped to it, then create a new one. The new vmspace is null.
3842 vmspace_exec(struct proc *p, struct vmspace *vmcopy)
3844 struct vmspace *oldvmspace = p->p_vmspace;
3845 struct vmspace *newvmspace;
3846 vm_map_t map = &p->p_vmspace->vm_map;
3849 * If we are execing a resident vmspace we fork it, otherwise
3850 * we create a new vmspace. Note that exitingcnt is not
3851 * copied to the new vmspace.
3853 lwkt_gettoken(&oldvmspace->vm_map.token);
3855 newvmspace = vmspace_fork(vmcopy);
3856 lwkt_gettoken(&newvmspace->vm_map.token);
3858 newvmspace = vmspace_alloc(map->min_offset, map->max_offset);
3859 lwkt_gettoken(&newvmspace->vm_map.token);
3860 bcopy(&oldvmspace->vm_startcopy, &newvmspace->vm_startcopy,
3861 (caddr_t)&oldvmspace->vm_endcopy -
3862 (caddr_t)&oldvmspace->vm_startcopy);
3866 * Finish initializing the vmspace before assigning it
3867 * to the process. The vmspace will become the current vmspace
3870 pmap_pinit2(vmspace_pmap(newvmspace));
3871 pmap_replacevm(p, newvmspace, 0);
3872 lwkt_reltoken(&newvmspace->vm_map.token);
3873 lwkt_reltoken(&oldvmspace->vm_map.token);
3874 vmspace_rel(oldvmspace);
3878 * Unshare the specified VM space for forcing COW. This
3879 * is called by rfork, for the (RFMEM|RFPROC) == 0 case.
3882 vmspace_unshare(struct proc *p)
3884 struct vmspace *oldvmspace = p->p_vmspace;
3885 struct vmspace *newvmspace;
3887 lwkt_gettoken(&oldvmspace->vm_map.token);
3888 if (vmspace_getrefs(oldvmspace) == 1) {
3889 lwkt_reltoken(&oldvmspace->vm_map.token);
3892 newvmspace = vmspace_fork(oldvmspace);
3893 lwkt_gettoken(&newvmspace->vm_map.token);
3894 pmap_pinit2(vmspace_pmap(newvmspace));
3895 pmap_replacevm(p, newvmspace, 0);
3896 lwkt_reltoken(&newvmspace->vm_map.token);
3897 lwkt_reltoken(&oldvmspace->vm_map.token);
3898 vmspace_rel(oldvmspace);
3902 * vm_map_hint: return the beginning of the best area suitable for
3903 * creating a new mapping with "prot" protection.
3908 vm_map_hint(struct proc *p, vm_offset_t addr, vm_prot_t prot)
3910 struct vmspace *vms = p->p_vmspace;
3912 if (!randomize_mmap || addr != 0) {
3914 * Set a reasonable start point for the hint if it was
3915 * not specified or if it falls within the heap space.
3916 * Hinted mmap()s do not allocate out of the heap space.
3919 (addr >= round_page((vm_offset_t)vms->vm_taddr) &&
3920 addr < round_page((vm_offset_t)vms->vm_daddr + maxdsiz))) {
3921 addr = round_page((vm_offset_t)vms->vm_daddr + maxdsiz);
3930 * If executable skip first two pages, otherwise start
3931 * after data + heap region.
3933 if ((prot & VM_PROT_EXECUTE) &&
3934 ((vm_offset_t)vms->vm_daddr >= I386_MAX_EXE_ADDR)) {
3935 addr = (PAGE_SIZE * 2) +
3936 (karc4random() & (I386_MAX_EXE_ADDR / 2 - 1));
3937 return (round_page(addr));
3939 #endif /* __i386__ */
3942 addr = (vm_offset_t)vms->vm_daddr + MAXDSIZ;
3943 addr += karc4random() & (MIN((256 * 1024 * 1024), MAXDSIZ) - 1);
3945 return (round_page(addr));
3949 * Finds the VM object, offset, and protection for a given virtual address
3950 * in the specified map, assuming a page fault of the type specified.
3952 * Leaves the map in question locked for read; return values are guaranteed
3953 * until a vm_map_lookup_done call is performed. Note that the map argument
3954 * is in/out; the returned map must be used in the call to vm_map_lookup_done.
3956 * A handle (out_entry) is returned for use in vm_map_lookup_done, to make
3959 * If a lookup is requested with "write protection" specified, the map may
3960 * be changed to perform virtual copying operations, although the data
3961 * referenced will remain the same.
3966 vm_map_lookup(vm_map_t *var_map, /* IN/OUT */
3968 vm_prot_t fault_typea,
3969 vm_map_entry_t *out_entry, /* OUT */
3970 vm_object_t *object, /* OUT */
3971 vm_pindex_t *pindex, /* OUT */
3972 vm_prot_t *out_prot, /* OUT */
3973 boolean_t *wired) /* OUT */
3975 vm_map_entry_t entry;
3976 vm_map_t map = *var_map;
3978 vm_prot_t fault_type = fault_typea;
3979 int use_read_lock = 1;
3980 int rv = KERN_SUCCESS;
3984 vm_map_lock_read(map);
3989 * If the map has an interesting hint, try it before calling full
3990 * blown lookup routine.
3997 if ((entry == &map->header) ||
3998 (vaddr < entry->start) || (vaddr >= entry->end)) {
3999 vm_map_entry_t tmp_entry;
4002 * Entry was either not a valid hint, or the vaddr was not
4003 * contained in the entry, so do a full lookup.
4005 if (!vm_map_lookup_entry(map, vaddr, &tmp_entry)) {
4006 rv = KERN_INVALID_ADDRESS;
4017 if (entry->maptype == VM_MAPTYPE_SUBMAP) {
4018 vm_map_t old_map = map;
4020 *var_map = map = entry->object.sub_map;
4022 vm_map_unlock_read(old_map);
4024 vm_map_unlock(old_map);
4030 * Check whether this task is allowed to have this page.
4031 * Note the special case for MAP_ENTRY_COW
4032 * pages with an override. This is to implement a forced
4033 * COW for debuggers.
4036 if (fault_type & VM_PROT_OVERRIDE_WRITE)
4037 prot = entry->max_protection;
4039 prot = entry->protection;
4041 fault_type &= (VM_PROT_READ|VM_PROT_WRITE|VM_PROT_EXECUTE);
4042 if ((fault_type & prot) != fault_type) {
4043 rv = KERN_PROTECTION_FAILURE;
4047 if ((entry->eflags & MAP_ENTRY_USER_WIRED) &&
4048 (entry->eflags & MAP_ENTRY_COW) &&
4049 (fault_type & VM_PROT_WRITE) &&
4050 (fault_typea & VM_PROT_OVERRIDE_WRITE) == 0) {
4051 rv = KERN_PROTECTION_FAILURE;
4056 * If this page is not pageable, we have to get it for all possible
4059 *wired = (entry->wired_count != 0);
4061 prot = fault_type = entry->protection;
4064 * Virtual page tables may need to update the accessed (A) bit
4065 * in a page table entry. Upgrade the fault to a write fault for
4066 * that case if the map will support it. If the map does not support
4067 * it the page table entry simply will not be updated.
4069 if (entry->maptype == VM_MAPTYPE_VPAGETABLE) {
4070 if (prot & VM_PROT_WRITE)
4071 fault_type |= VM_PROT_WRITE;
4074 if (curthread->td_lwp && curthread->td_lwp->lwp_vmspace &&
4075 pmap_emulate_ad_bits(&curthread->td_lwp->lwp_vmspace->vm_pmap)) {
4076 if ((prot & VM_PROT_WRITE) == 0)
4077 fault_type |= VM_PROT_WRITE;
4081 * Only NORMAL and VPAGETABLE maps are object-based. UKSMAPs are not.
4083 if (entry->maptype != VM_MAPTYPE_NORMAL &&
4084 entry->maptype != VM_MAPTYPE_VPAGETABLE) {
4090 * If the entry was copy-on-write, we either ...
4092 if (entry->eflags & MAP_ENTRY_NEEDS_COPY) {
4094 * If we want to write the page, we may as well handle that
4095 * now since we've got the map locked.
4097 * If we don't need to write the page, we just demote the
4098 * permissions allowed.
4101 if (fault_type & VM_PROT_WRITE) {
4103 * Make a new object, and place it in the object
4104 * chain. Note that no new references have appeared
4105 * -- one just moved from the map to the new
4109 if (use_read_lock && vm_map_lock_upgrade(map)) {
4116 vm_map_entry_shadow(entry, 0);
4119 * We're attempting to read a copy-on-write page --
4120 * don't allow writes.
4123 prot &= ~VM_PROT_WRITE;
4128 * Create an object if necessary.
4130 if (entry->object.vm_object == NULL && !map->system_map) {
4131 if (use_read_lock && vm_map_lock_upgrade(map)) {
4137 vm_map_entry_allocate_object(entry);
4141 * Return the object/offset from this entry. If the entry was
4142 * copy-on-write or empty, it has been fixed up.
4144 *object = entry->object.vm_object;
4147 *pindex = OFF_TO_IDX((vaddr - entry->start) + entry->offset);
4150 * Return whether this is the only map sharing this data. On
4151 * success we return with a read lock held on the map. On failure
4152 * we return with the map unlocked.
4156 if (rv == KERN_SUCCESS) {
4157 if (use_read_lock == 0)
4158 vm_map_lock_downgrade(map);
4159 } else if (use_read_lock) {
4160 vm_map_unlock_read(map);
4168 * Releases locks acquired by a vm_map_lookup()
4169 * (according to the handle returned by that lookup).
4171 * No other requirements.
4174 vm_map_lookup_done(vm_map_t map, vm_map_entry_t entry, int count)
4177 * Unlock the main-level map
4179 vm_map_unlock_read(map);
4181 vm_map_entry_release(count);
4184 #include "opt_ddb.h"
4186 #include <sys/kernel.h>
4188 #include <ddb/ddb.h>
4193 DB_SHOW_COMMAND(map, vm_map_print)
4196 /* XXX convert args. */
4197 vm_map_t map = (vm_map_t)addr;
4198 boolean_t full = have_addr;
4200 vm_map_entry_t entry;
4202 db_iprintf("Task map %p: pmap=%p, nentries=%d, version=%u\n",
4204 (void *)map->pmap, map->nentries, map->timestamp);
4207 if (!full && db_indent)
4211 for (entry = map->header.next; entry != &map->header;
4212 entry = entry->next) {
4213 db_iprintf("map entry %p: start=%p, end=%p\n",
4214 (void *)entry, (void *)entry->start, (void *)entry->end);
4217 static char *inheritance_name[4] =
4218 {"share", "copy", "none", "donate_copy"};
4220 db_iprintf(" prot=%x/%x/%s",
4222 entry->max_protection,
4223 inheritance_name[(int)(unsigned char)entry->inheritance]);
4224 if (entry->wired_count != 0)
4225 db_printf(", wired");
4227 switch(entry->maptype) {
4228 case VM_MAPTYPE_SUBMAP:
4229 /* XXX no %qd in kernel. Truncate entry->offset. */
4230 db_printf(", share=%p, offset=0x%lx\n",
4231 (void *)entry->object.sub_map,
4232 (long)entry->offset);
4234 if ((entry->prev == &map->header) ||
4235 (entry->prev->object.sub_map !=
4236 entry->object.sub_map)) {
4238 vm_map_print((db_expr_t)(intptr_t)
4239 entry->object.sub_map,
4244 case VM_MAPTYPE_NORMAL:
4245 case VM_MAPTYPE_VPAGETABLE:
4246 /* XXX no %qd in kernel. Truncate entry->offset. */
4247 db_printf(", object=%p, offset=0x%lx",
4248 (void *)entry->object.vm_object,
4249 (long)entry->offset);
4250 if (entry->eflags & MAP_ENTRY_COW)
4251 db_printf(", copy (%s)",
4252 (entry->eflags & MAP_ENTRY_NEEDS_COPY) ? "needed" : "done");
4256 if ((entry->prev == &map->header) ||
4257 (entry->prev->object.vm_object !=
4258 entry->object.vm_object)) {
4260 vm_object_print((db_expr_t)(intptr_t)
4261 entry->object.vm_object,
4267 case VM_MAPTYPE_UKSMAP:
4268 db_printf(", uksmap=%p, offset=0x%lx",
4269 (void *)entry->object.uksmap,
4270 (long)entry->offset);
4271 if (entry->eflags & MAP_ENTRY_COW)
4272 db_printf(", copy (%s)",
4273 (entry->eflags & MAP_ENTRY_NEEDS_COPY) ? "needed" : "done");
4289 DB_SHOW_COMMAND(procvm, procvm)
4294 p = (struct proc *) addr;
4299 db_printf("p = %p, vmspace = %p, map = %p, pmap = %p\n",
4300 (void *)p, (void *)p->p_vmspace, (void *)&p->p_vmspace->vm_map,
4301 (void *)vmspace_pmap(p->p_vmspace));
4303 vm_map_print((db_expr_t)(intptr_t)&p->p_vmspace->vm_map, 1, 0, NULL);