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;
132 static vm_zone_t mapentzone;
133 static struct vm_object mapentobj;
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];
139 static int randomize_mmap;
140 SYSCTL_INT(_vm, OID_AUTO, randomize_mmap, CTLFLAG_RW, &randomize_mmap, 0,
141 "Randomize mmap offsets");
142 static int vm_map_relock_enable = 1;
143 SYSCTL_INT(_vm, OID_AUTO, map_relock_enable, CTLFLAG_RW,
144 &vm_map_relock_enable, 0, "Randomize mmap offsets");
146 static void vm_map_entry_shadow(vm_map_entry_t entry, int addref);
147 static vm_map_entry_t vm_map_entry_create(vm_map_t map, int *);
148 static void vm_map_entry_dispose (vm_map_t map, vm_map_entry_t entry, int *);
149 static void _vm_map_clip_end (vm_map_t, vm_map_entry_t, vm_offset_t, int *);
150 static void _vm_map_clip_start (vm_map_t, vm_map_entry_t, vm_offset_t, int *);
151 static void vm_map_entry_delete (vm_map_t, vm_map_entry_t, int *);
152 static void vm_map_entry_unwire (vm_map_t, vm_map_entry_t);
153 static void vm_map_copy_entry (vm_map_t, vm_map_t, vm_map_entry_t,
155 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);
158 * Initialize the vm_map module. Must be called before any other vm_map
161 * Map and entry structures are allocated from the general purpose
162 * memory pool with some exceptions:
164 * - The kernel map is allocated statically.
165 * - Initial kernel map entries are allocated out of a static pool.
166 * - We must set ZONE_SPECIAL here or the early boot code can get
167 * stuck if there are >63 cores.
169 * These restrictions are necessary since malloc() uses the
170 * maps and requires map entries.
172 * Called from the low level boot code only.
177 mapentzone = &mapentzone_store;
178 zbootinit(mapentzone, "MAP ENTRY", sizeof (struct vm_map_entry),
179 map_entry_init, MAX_MAPENT);
180 mapentzone_store.zflags |= ZONE_SPECIAL;
184 * Called prior to any vmspace allocations.
186 * Called from the low level boot code only.
191 vmspace_cache = objcache_create_mbacked(M_VMSPACE,
192 sizeof(struct vmspace),
194 vmspace_ctor, vmspace_dtor,
196 zinitna(mapentzone, &mapentobj, NULL, 0, 0,
197 ZONE_USE_RESERVE | ZONE_SPECIAL);
203 * objcache support. We leave the pmap root cached as long as possible
204 * for performance reasons.
208 vmspace_ctor(void *obj, void *privdata, int ocflags)
210 struct vmspace *vm = obj;
212 bzero(vm, sizeof(*vm));
213 vm->vm_refcnt = (u_int)-1;
220 vmspace_dtor(void *obj, void *privdata)
222 struct vmspace *vm = obj;
224 KKASSERT(vm->vm_refcnt == (u_int)-1);
225 pmap_puninit(vmspace_pmap(vm));
229 * Red black tree functions
231 * The caller must hold the related map lock.
233 static int rb_vm_map_compare(vm_map_entry_t a, vm_map_entry_t b);
234 RB_GENERATE(vm_map_rb_tree, vm_map_entry, rb_entry, rb_vm_map_compare);
236 /* a->start is address, and the only field has to be initialized */
238 rb_vm_map_compare(vm_map_entry_t a, vm_map_entry_t b)
240 if (a->start < b->start)
242 else if (a->start > b->start)
248 * Initialize vmspace ref/hold counts vmspace0. There is a holdcnt for
252 vmspace_initrefs(struct vmspace *vm)
259 * Allocate a vmspace structure, including a vm_map and pmap.
260 * Initialize numerous fields. While the initial allocation is zerod,
261 * subsequence reuse from the objcache leaves elements of the structure
262 * intact (particularly the pmap), so portions must be zerod.
264 * Returns a referenced vmspace.
269 vmspace_alloc(vm_offset_t min, vm_offset_t max)
273 vm = objcache_get(vmspace_cache, M_WAITOK);
275 bzero(&vm->vm_startcopy,
276 (char *)&vm->vm_endcopy - (char *)&vm->vm_startcopy);
277 vm_map_init(&vm->vm_map, min, max, NULL); /* initializes token */
280 * NOTE: hold to acquires token for safety.
282 * On return vmspace is referenced (refs=1, hold=1). That is,
283 * each refcnt also has a holdcnt. There can be additional holds
284 * (holdcnt) above and beyond the refcnt. Finalization is handled in
285 * two stages, one on refs 1->0, and the the second on hold 1->0.
287 KKASSERT(vm->vm_holdcnt == 0);
288 KKASSERT(vm->vm_refcnt == (u_int)-1);
289 vmspace_initrefs(vm);
291 pmap_pinit(vmspace_pmap(vm)); /* (some fields reused) */
292 vm->vm_map.pmap = vmspace_pmap(vm); /* XXX */
295 cpu_vmspace_alloc(vm);
302 * NOTE: Can return -1 if the vmspace is exiting.
305 vmspace_getrefs(struct vmspace *vm)
307 return ((int)vm->vm_refcnt);
311 * A vmspace object must already have a non-zero hold to be able to gain
312 * further holds on it.
315 vmspace_hold_notoken(struct vmspace *vm)
317 KKASSERT(vm->vm_holdcnt != 0);
318 refcount_acquire(&vm->vm_holdcnt);
322 vmspace_drop_notoken(struct vmspace *vm)
324 if (refcount_release(&vm->vm_holdcnt)) {
325 if (vm->vm_refcnt == (u_int)-1) {
326 vmspace_terminate(vm, 1);
332 vmspace_hold(struct vmspace *vm)
334 vmspace_hold_notoken(vm);
335 lwkt_gettoken(&vm->vm_map.token);
339 vmspace_drop(struct vmspace *vm)
341 lwkt_reltoken(&vm->vm_map.token);
342 vmspace_drop_notoken(vm);
346 * A vmspace object must not be in a terminated state to be able to obtain
347 * additional refs on it.
349 * Ref'ing a vmspace object also increments its hold count.
352 vmspace_ref(struct vmspace *vm)
354 KKASSERT((int)vm->vm_refcnt >= 0);
355 vmspace_hold_notoken(vm);
356 refcount_acquire(&vm->vm_refcnt);
360 * Release a ref on the vmspace. On the 1->0 transition we do stage-1
361 * termination of the vmspace. Then, on the final drop of the hold we
362 * will do stage-2 final termination.
365 vmspace_rel(struct vmspace *vm)
367 if (refcount_release(&vm->vm_refcnt)) {
368 vm->vm_refcnt = (u_int)-1; /* no other refs possible */
369 vmspace_terminate(vm, 0);
371 vmspace_drop_notoken(vm);
375 * This is called during exit indicating that the vmspace is no
376 * longer in used by an exiting process, but the process has not yet
379 * We release the refcnt but not the associated holdcnt.
384 vmspace_relexit(struct vmspace *vm)
386 if (refcount_release(&vm->vm_refcnt)) {
387 vm->vm_refcnt = (u_int)-1; /* no other refs possible */
388 vmspace_terminate(vm, 0);
393 * Called during reap to disconnect the remainder of the vmspace from
394 * the process. On the hold drop the vmspace termination is finalized.
399 vmspace_exitfree(struct proc *p)
405 vmspace_drop_notoken(vm);
409 * Called in two cases:
411 * (1) When the last refcnt is dropped and the vmspace becomes inactive,
412 * called with final == 0. refcnt will be (u_int)-1 at this point,
413 * and holdcnt will still be non-zero.
415 * (2) When holdcnt becomes 0, called with final == 1. There should no
416 * longer be anyone with access to the vmspace.
418 * VMSPACE_EXIT1 flags the primary deactivation
419 * VMSPACE_EXIT2 flags the last reap
422 vmspace_terminate(struct vmspace *vm, int final)
426 lwkt_gettoken(&vm->vm_map.token);
428 KKASSERT((vm->vm_flags & VMSPACE_EXIT1) == 0);
431 * Get rid of most of the resources. Leave the kernel pmap
434 * If the pmap does not contain wired pages we can bulk-delete
435 * the pmap as a performance optimization before removing the related mappings.
437 * If the pmap contains wired pages we cannot do this pre-optimization
438 * because currently vm_fault_unwire() expects the pmap pages to exist
439 * and will not decrement p->wire_count if they do not.
441 vm->vm_flags |= VMSPACE_EXIT1;
443 if (vmspace_pmap(vm)->pm_stats.wired_count) {
444 vm_map_remove(&vm->vm_map, VM_MIN_USER_ADDRESS,
445 VM_MAX_USER_ADDRESS);
446 pmap_remove_pages(vmspace_pmap(vm), VM_MIN_USER_ADDRESS,
447 VM_MAX_USER_ADDRESS);
449 pmap_remove_pages(vmspace_pmap(vm), VM_MIN_USER_ADDRESS,
450 VM_MAX_USER_ADDRESS);
451 vm_map_remove(&vm->vm_map, VM_MIN_USER_ADDRESS,
452 VM_MAX_USER_ADDRESS);
454 lwkt_reltoken(&vm->vm_map.token);
456 KKASSERT((vm->vm_flags & VMSPACE_EXIT1) != 0);
457 KKASSERT((vm->vm_flags & VMSPACE_EXIT2) == 0);
460 * Get rid of remaining basic resources.
462 vm->vm_flags |= VMSPACE_EXIT2;
463 cpu_vmspace_free(vm);
467 * Lock the map, to wait out all other references to it.
468 * Delete all of the mappings and pages they hold, then call
469 * the pmap module to reclaim anything left.
471 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
472 vm_map_lock(&vm->vm_map);
473 vm_map_delete(&vm->vm_map, vm->vm_map.min_offset,
474 vm->vm_map.max_offset, &count);
475 vm_map_unlock(&vm->vm_map);
476 vm_map_entry_release(count);
478 lwkt_gettoken(&vmspace_pmap(vm)->pm_token);
479 pmap_release(vmspace_pmap(vm));
480 lwkt_reltoken(&vmspace_pmap(vm)->pm_token);
481 lwkt_reltoken(&vm->vm_map.token);
482 objcache_put(vmspace_cache, vm);
487 * Swap useage is determined by taking the proportional swap used by
488 * VM objects backing the VM map. To make up for fractional losses,
489 * if the VM object has any swap use at all the associated map entries
490 * count for at least 1 swap page.
495 vmspace_swap_count(struct vmspace *vm)
497 vm_map_t map = &vm->vm_map;
500 vm_offset_t count = 0;
504 for (cur = map->header.next; cur != &map->header; cur = cur->next) {
505 switch(cur->maptype) {
506 case VM_MAPTYPE_NORMAL:
507 case VM_MAPTYPE_VPAGETABLE:
508 if ((object = cur->object.vm_object) == NULL)
510 if (object->swblock_count) {
511 n = (cur->end - cur->start) / PAGE_SIZE;
512 count += object->swblock_count *
513 SWAP_META_PAGES * n / object->size + 1;
526 * Calculate the approximate number of anonymous pages in use by
527 * this vmspace. To make up for fractional losses, we count each
528 * VM object as having at least 1 anonymous page.
533 vmspace_anonymous_count(struct vmspace *vm)
535 vm_map_t map = &vm->vm_map;
538 vm_offset_t count = 0;
541 for (cur = map->header.next; cur != &map->header; cur = cur->next) {
542 switch(cur->maptype) {
543 case VM_MAPTYPE_NORMAL:
544 case VM_MAPTYPE_VPAGETABLE:
545 if ((object = cur->object.vm_object) == NULL)
547 if (object->type != OBJT_DEFAULT &&
548 object->type != OBJT_SWAP) {
551 count += object->resident_page_count;
563 * Initialize an existing vm_map structure such as that in the vmspace
564 * structure. The pmap is initialized elsewhere.
569 vm_map_init(struct vm_map *map, vm_offset_t min, vm_offset_t max, pmap_t pmap)
571 map->header.next = map->header.prev = &map->header;
572 RB_INIT(&map->rb_root);
576 map->min_offset = min;
577 map->max_offset = max;
579 map->first_free = &map->header;
580 map->hint = &map->header;
583 lwkt_token_init(&map->token, "vm_map");
584 lockinit(&map->lock, "vm_maplk", (hz + 9) / 10, 0);
588 * Shadow the vm_map_entry's object. This typically needs to be done when
589 * a write fault is taken on an entry which had previously been cloned by
590 * fork(). The shared object (which might be NULL) must become private so
591 * we add a shadow layer above it.
593 * Object allocation for anonymous mappings is defered as long as possible.
594 * When creating a shadow, however, the underlying object must be instantiated
595 * so it can be shared.
597 * If the map segment is governed by a virtual page table then it is
598 * possible to address offsets beyond the mapped area. Just allocate
599 * a maximally sized object for this case.
601 * If addref is non-zero an additional reference is added to the returned
602 * entry. This mechanic exists because the additional reference might have
603 * to be added atomically and not after return to prevent a premature
606 * The vm_map must be exclusively locked.
607 * No other requirements.
611 vm_map_entry_shadow(vm_map_entry_t entry, int addref)
613 if (entry->maptype == VM_MAPTYPE_VPAGETABLE) {
614 vm_object_shadow(&entry->object.vm_object, &entry->offset,
615 0x7FFFFFFF, addref); /* XXX */
617 vm_object_shadow(&entry->object.vm_object, &entry->offset,
618 atop(entry->end - entry->start), addref);
620 entry->eflags &= ~MAP_ENTRY_NEEDS_COPY;
624 * Allocate an object for a vm_map_entry.
626 * Object allocation for anonymous mappings is defered as long as possible.
627 * This function is called when we can defer no longer, generally when a map
628 * entry might be split or forked or takes a page fault.
630 * If the map segment is governed by a virtual page table then it is
631 * possible to address offsets beyond the mapped area. Just allocate
632 * a maximally sized object for this case.
634 * The vm_map must be exclusively locked.
635 * No other requirements.
638 vm_map_entry_allocate_object(vm_map_entry_t entry)
642 if (entry->maptype == VM_MAPTYPE_VPAGETABLE) {
643 obj = vm_object_allocate(OBJT_DEFAULT, 0x7FFFFFFF); /* XXX */
645 obj = vm_object_allocate(OBJT_DEFAULT,
646 atop(entry->end - entry->start));
648 entry->object.vm_object = obj;
653 * Set an initial negative count so the first attempt to reserve
654 * space preloads a bunch of vm_map_entry's for this cpu. Also
655 * pre-allocate 2 vm_map_entries which will be needed by zalloc() to
656 * map a new page for vm_map_entry structures. SMP systems are
657 * particularly sensitive.
659 * This routine is called in early boot so we cannot just call
660 * vm_map_entry_reserve().
662 * Called from the low level boot code only (for each cpu)
664 * WARNING! Take care not to have too-big a static/BSS structure here
665 * as MAXCPU can be 256+, otherwise the loader's 64MB heap
666 * can get blown out by the kernel plus the initrd image.
669 vm_map_entry_reserve_cpu_init(globaldata_t gd)
671 vm_map_entry_t entry;
675 gd->gd_vme_avail -= MAP_RESERVE_COUNT * 2;
676 if (gd->gd_cpuid == 0) {
677 entry = &cpu_map_entry_init_bsp[0];
678 count = MAPENTRYBSP_CACHE;
680 entry = &cpu_map_entry_init_ap[gd->gd_cpuid][0];
681 count = MAPENTRYAP_CACHE;
683 for (i = 0; i < count; ++i, ++entry) {
684 entry->next = gd->gd_vme_base;
685 gd->gd_vme_base = entry;
690 * Reserves vm_map_entry structures so code later on can manipulate
691 * map_entry structures within a locked map without blocking trying
692 * to allocate a new vm_map_entry.
697 vm_map_entry_reserve(int count)
699 struct globaldata *gd = mycpu;
700 vm_map_entry_t entry;
703 * Make sure we have enough structures in gd_vme_base to handle
704 * the reservation request.
706 * The critical section protects access to the per-cpu gd.
709 while (gd->gd_vme_avail < count) {
710 entry = zalloc(mapentzone);
711 entry->next = gd->gd_vme_base;
712 gd->gd_vme_base = entry;
715 gd->gd_vme_avail -= count;
722 * Releases previously reserved vm_map_entry structures that were not
723 * used. If we have too much junk in our per-cpu cache clean some of
729 vm_map_entry_release(int count)
731 struct globaldata *gd = mycpu;
732 vm_map_entry_t entry;
735 gd->gd_vme_avail += count;
736 while (gd->gd_vme_avail > MAP_RESERVE_SLOP) {
737 entry = gd->gd_vme_base;
738 KKASSERT(entry != NULL);
739 gd->gd_vme_base = entry->next;
742 zfree(mapentzone, entry);
749 * Reserve map entry structures for use in kernel_map itself. These
750 * entries have *ALREADY* been reserved on a per-cpu basis when the map
751 * was inited. This function is used by zalloc() to avoid a recursion
752 * when zalloc() itself needs to allocate additional kernel memory.
754 * This function works like the normal reserve but does not load the
755 * vm_map_entry cache (because that would result in an infinite
756 * recursion). Note that gd_vme_avail may go negative. This is expected.
758 * Any caller of this function must be sure to renormalize after
759 * potentially eating entries to ensure that the reserve supply
765 vm_map_entry_kreserve(int count)
767 struct globaldata *gd = mycpu;
770 gd->gd_vme_avail -= count;
772 KASSERT(gd->gd_vme_base != NULL,
773 ("no reserved entries left, gd_vme_avail = %d",
779 * Release previously reserved map entries for kernel_map. We do not
780 * attempt to clean up like the normal release function as this would
781 * cause an unnecessary (but probably not fatal) deep procedure call.
786 vm_map_entry_krelease(int count)
788 struct globaldata *gd = mycpu;
791 gd->gd_vme_avail += count;
796 * Allocates a VM map entry for insertion. No entry fields are filled in.
798 * The entries should have previously been reserved. The reservation count
799 * is tracked in (*countp).
803 static vm_map_entry_t
804 vm_map_entry_create(vm_map_t map, int *countp)
806 struct globaldata *gd = mycpu;
807 vm_map_entry_t entry;
809 KKASSERT(*countp > 0);
812 entry = gd->gd_vme_base;
813 KASSERT(entry != NULL, ("gd_vme_base NULL! count %d", *countp));
814 gd->gd_vme_base = entry->next;
821 * Dispose of a vm_map_entry that is no longer being referenced.
826 vm_map_entry_dispose(vm_map_t map, vm_map_entry_t entry, int *countp)
828 struct globaldata *gd = mycpu;
830 KKASSERT(map->hint != entry);
831 KKASSERT(map->first_free != entry);
835 entry->next = gd->gd_vme_base;
836 gd->gd_vme_base = entry;
842 * Insert/remove entries from maps.
844 * The related map must be exclusively locked.
845 * The caller must hold map->token
846 * No other requirements.
849 vm_map_entry_link(vm_map_t map,
850 vm_map_entry_t after_where,
851 vm_map_entry_t entry)
853 ASSERT_VM_MAP_LOCKED(map);
856 entry->prev = after_where;
857 entry->next = after_where->next;
858 entry->next->prev = entry;
859 after_where->next = entry;
860 if (vm_map_rb_tree_RB_INSERT(&map->rb_root, entry))
861 panic("vm_map_entry_link: dup addr map %p ent %p", map, entry);
865 vm_map_entry_unlink(vm_map_t map,
866 vm_map_entry_t entry)
871 ASSERT_VM_MAP_LOCKED(map);
873 if (entry->eflags & MAP_ENTRY_IN_TRANSITION) {
874 panic("vm_map_entry_unlink: attempt to mess with "
875 "locked entry! %p", entry);
881 vm_map_rb_tree_RB_REMOVE(&map->rb_root, entry);
886 * Finds the map entry containing (or immediately preceding) the specified
887 * address in the given map. The entry is returned in (*entry).
889 * The boolean result indicates whether the address is actually contained
892 * The related map must be locked.
893 * No other requirements.
896 vm_map_lookup_entry(vm_map_t map, vm_offset_t address, vm_map_entry_t *entry)
901 ASSERT_VM_MAP_LOCKED(map);
904 * XXX TEMPORARILY DISABLED. For some reason our attempt to revive
905 * the hint code with the red-black lookup meets with system crashes
906 * and lockups. We do not yet know why.
908 * It is possible that the problem is related to the setting
909 * of the hint during map_entry deletion, in the code specified
910 * at the GGG comment later on in this file.
912 * YYY More likely it's because this function can be called with
913 * a shared lock on the map, resulting in map->hint updates possibly
914 * racing. Fixed now but untested.
917 * Quickly check the cached hint, there's a good chance of a match.
921 if (tmp != &map->header) {
922 if (address >= tmp->start && address < tmp->end) {
930 * Locate the record from the top of the tree. 'last' tracks the
931 * closest prior record and is returned if no match is found, which
932 * in binary tree terms means tracking the most recent right-branch
933 * taken. If there is no prior record, &map->header is returned.
936 tmp = RB_ROOT(&map->rb_root);
939 if (address >= tmp->start) {
940 if (address < tmp->end) {
946 tmp = RB_RIGHT(tmp, rb_entry);
948 tmp = RB_LEFT(tmp, rb_entry);
956 * Inserts the given whole VM object into the target map at the specified
957 * address range. The object's size should match that of the address range.
959 * The map must be exclusively locked.
960 * The object must be held.
961 * The caller must have reserved sufficient vm_map_entry structures.
963 * If object is non-NULL, ref count must be bumped by caller prior to
964 * making call to account for the new entry.
967 vm_map_insert(vm_map_t map, int *countp, void *map_object, void *map_aux,
968 vm_ooffset_t offset, vm_offset_t start, vm_offset_t end,
969 vm_maptype_t maptype, vm_subsys_t id,
970 vm_prot_t prot, vm_prot_t max, int cow)
972 vm_map_entry_t new_entry;
973 vm_map_entry_t prev_entry;
974 vm_map_entry_t temp_entry;
975 vm_eflags_t protoeflags;
979 if (maptype == VM_MAPTYPE_UKSMAP)
984 ASSERT_VM_MAP_LOCKED(map);
986 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
989 * Check that the start and end points are not bogus.
991 if ((start < map->min_offset) || (end > map->max_offset) ||
993 return (KERN_INVALID_ADDRESS);
996 * Find the entry prior to the proposed starting address; if it's part
997 * of an existing entry, this range is bogus.
999 if (vm_map_lookup_entry(map, start, &temp_entry))
1000 return (KERN_NO_SPACE);
1002 prev_entry = temp_entry;
1005 * Assert that the next entry doesn't overlap the end point.
1008 if ((prev_entry->next != &map->header) &&
1009 (prev_entry->next->start < end))
1010 return (KERN_NO_SPACE);
1014 if (cow & MAP_COPY_ON_WRITE)
1015 protoeflags |= MAP_ENTRY_COW|MAP_ENTRY_NEEDS_COPY;
1017 if (cow & MAP_NOFAULT) {
1018 protoeflags |= MAP_ENTRY_NOFAULT;
1020 KASSERT(object == NULL,
1021 ("vm_map_insert: paradoxical MAP_NOFAULT request"));
1023 if (cow & MAP_DISABLE_SYNCER)
1024 protoeflags |= MAP_ENTRY_NOSYNC;
1025 if (cow & MAP_DISABLE_COREDUMP)
1026 protoeflags |= MAP_ENTRY_NOCOREDUMP;
1027 if (cow & MAP_IS_STACK)
1028 protoeflags |= MAP_ENTRY_STACK;
1029 if (cow & MAP_IS_KSTACK)
1030 protoeflags |= MAP_ENTRY_KSTACK;
1032 lwkt_gettoken(&map->token);
1036 * When object is non-NULL, it could be shared with another
1037 * process. We have to set or clear OBJ_ONEMAPPING
1040 * NOTE: This flag is only applicable to DEFAULT and SWAP
1041 * objects and will already be clear in other types
1042 * of objects, so a shared object lock is ok for
1045 if ((object->ref_count > 1) || (object->shadow_count != 0)) {
1046 vm_object_clear_flag(object, OBJ_ONEMAPPING);
1049 else if ((prev_entry != &map->header) &&
1050 (prev_entry->eflags == protoeflags) &&
1051 (prev_entry->end == start) &&
1052 (prev_entry->wired_count == 0) &&
1053 (prev_entry->id == id) &&
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 * XXX if object is NULL should we set offset to 0 here ?
1084 object = prev_entry->object.vm_object;
1085 offset = prev_entry->offset +
1086 (prev_entry->end - prev_entry->start);
1088 vm_object_hold(object);
1089 vm_object_chain_wait(object, 0);
1090 vm_object_reference_locked(object);
1092 map_object = object;
1097 * NOTE: if conditionals fail, object can be NULL here. This occurs
1098 * in things like the buffer map where we manage kva but do not manage
1103 * Create a new entry
1106 new_entry = vm_map_entry_create(map, countp);
1107 new_entry->start = start;
1108 new_entry->end = end;
1111 new_entry->maptype = maptype;
1112 new_entry->eflags = protoeflags;
1113 new_entry->object.map_object = map_object;
1114 new_entry->aux.master_pde = 0; /* in case size is different */
1115 new_entry->aux.map_aux = map_aux;
1116 new_entry->offset = offset;
1118 new_entry->inheritance = VM_INHERIT_DEFAULT;
1119 new_entry->protection = prot;
1120 new_entry->max_protection = max;
1121 new_entry->wired_count = 0;
1124 * Insert the new entry into the list
1127 vm_map_entry_link(map, prev_entry, new_entry);
1128 map->size += new_entry->end - new_entry->start;
1131 * Update the free space hint. Entries cannot overlap.
1132 * An exact comparison is needed to avoid matching
1133 * against the map->header.
1135 if ((map->first_free == prev_entry) &&
1136 (prev_entry->end == new_entry->start)) {
1137 map->first_free = new_entry;
1142 * Temporarily removed to avoid MAP_STACK panic, due to
1143 * MAP_STACK being a huge hack. Will be added back in
1144 * when MAP_STACK (and the user stack mapping) is fixed.
1147 * It may be possible to simplify the entry
1149 vm_map_simplify_entry(map, new_entry, countp);
1153 * Try to pre-populate the page table. Mappings governed by virtual
1154 * page tables cannot be prepopulated without a lot of work, so
1157 if ((cow & (MAP_PREFAULT|MAP_PREFAULT_PARTIAL)) &&
1158 maptype != VM_MAPTYPE_VPAGETABLE &&
1159 maptype != VM_MAPTYPE_UKSMAP) {
1161 if (vm_map_relock_enable && (cow & MAP_PREFAULT_RELOCK)) {
1163 vm_object_lock_swap();
1164 vm_object_drop(object);
1166 pmap_object_init_pt(map->pmap, start, prot,
1167 object, OFF_TO_IDX(offset), end - start,
1168 cow & MAP_PREFAULT_PARTIAL);
1170 vm_object_hold(object);
1171 vm_object_lock_swap();
1175 vm_object_drop(object);
1177 lwkt_reltoken(&map->token);
1178 return (KERN_SUCCESS);
1182 * Find sufficient space for `length' bytes in the given map, starting at
1183 * `start'. Returns 0 on success, 1 on no space.
1185 * This function will returned an arbitrarily aligned pointer. If no
1186 * particular alignment is required you should pass align as 1. Note that
1187 * the map may return PAGE_SIZE aligned pointers if all the lengths used in
1188 * the map are a multiple of PAGE_SIZE, even if you pass a smaller align
1191 * 'align' should be a power of 2 but is not required to be.
1193 * The map must be exclusively locked.
1194 * No other requirements.
1197 vm_map_findspace(vm_map_t map, vm_offset_t start, vm_size_t length,
1198 vm_size_t align, int flags, vm_offset_t *addr)
1200 vm_map_entry_t entry, next;
1202 vm_offset_t align_mask;
1204 if (start < map->min_offset)
1205 start = map->min_offset;
1206 if (start > map->max_offset)
1210 * If the alignment is not a power of 2 we will have to use
1211 * a mod/division, set align_mask to a special value.
1213 if ((align | (align - 1)) + 1 != (align << 1))
1214 align_mask = (vm_offset_t)-1;
1216 align_mask = align - 1;
1219 * Look for the first possible address; if there's already something
1220 * at this address, we have to start after it.
1222 if (start == map->min_offset) {
1223 if ((entry = map->first_free) != &map->header)
1228 if (vm_map_lookup_entry(map, start, &tmp))
1234 * Look through the rest of the map, trying to fit a new region in the
1235 * gap between existing regions, or after the very last region.
1237 for (;; start = (entry = next)->end) {
1239 * Adjust the proposed start by the requested alignment,
1240 * be sure that we didn't wrap the address.
1242 if (align_mask == (vm_offset_t)-1)
1243 end = roundup(start, align);
1245 end = (start + align_mask) & ~align_mask;
1250 * Find the end of the proposed new region. Be sure we didn't
1251 * go beyond the end of the map, or wrap around the address.
1252 * Then check to see if this is the last entry or if the
1253 * proposed end fits in the gap between this and the next
1256 end = start + length;
1257 if (end > map->max_offset || end < start)
1262 * If the next entry's start address is beyond the desired
1263 * end address we may have found a good entry.
1265 * If the next entry is a stack mapping we do not map into
1266 * the stack's reserved space.
1268 * XXX continue to allow mapping into the stack's reserved
1269 * space if doing a MAP_STACK mapping inside a MAP_STACK
1270 * mapping, for backwards compatibility. But the caller
1271 * really should use MAP_STACK | MAP_TRYFIXED if they
1274 if (next == &map->header)
1276 if (next->start >= end) {
1277 if ((next->eflags & MAP_ENTRY_STACK) == 0)
1279 if (flags & MAP_STACK)
1281 if (next->start - next->aux.avail_ssize >= end)
1288 * Grow the kernel_map if necessary. pmap_growkernel() will panic
1289 * if it fails. The kernel_map is locked and nothing can steal
1290 * our address space if pmap_growkernel() blocks.
1292 * NOTE: This may be unconditionally called for kldload areas on
1293 * x86_64 because these do not bump kernel_vm_end (which would
1294 * fill 128G worth of page tables!). Therefore we must not
1297 if (map == &kernel_map) {
1300 kstop = round_page(start + length);
1301 if (kstop > kernel_vm_end)
1302 pmap_growkernel(start, kstop);
1309 * vm_map_find finds an unallocated region in the target address map with
1310 * the given length and allocates it. The search is defined to be first-fit
1311 * from the specified address; the region found is returned in the same
1314 * If object is non-NULL, ref count must be bumped by caller
1315 * prior to making call to account for the new entry.
1317 * No requirements. This function will lock the map temporarily.
1320 vm_map_find(vm_map_t map, void *map_object, void *map_aux,
1321 vm_ooffset_t offset, vm_offset_t *addr,
1322 vm_size_t length, vm_size_t align, boolean_t fitit,
1323 vm_maptype_t maptype, vm_subsys_t id,
1324 vm_prot_t prot, vm_prot_t max, int cow)
1331 if (maptype == VM_MAPTYPE_UKSMAP)
1334 object = map_object;
1338 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
1341 vm_object_hold_shared(object);
1343 if (vm_map_findspace(map, start, length, align, 0, addr)) {
1345 vm_object_drop(object);
1347 vm_map_entry_release(count);
1348 return (KERN_NO_SPACE);
1352 result = vm_map_insert(map, &count, map_object, map_aux,
1353 offset, start, start + length,
1354 maptype, id, prot, max, cow);
1356 vm_object_drop(object);
1358 vm_map_entry_release(count);
1364 * Simplify the given map entry by merging with either neighbor. This
1365 * routine also has the ability to merge with both neighbors.
1367 * This routine guarentees that the passed entry remains valid (though
1368 * possibly extended). When merging, this routine may delete one or
1369 * both neighbors. No action is taken on entries which have their
1370 * in-transition flag set.
1372 * The map must be exclusively locked.
1375 vm_map_simplify_entry(vm_map_t map, vm_map_entry_t entry, int *countp)
1377 vm_map_entry_t next, prev;
1378 vm_size_t prevsize, esize;
1380 if (entry->eflags & MAP_ENTRY_IN_TRANSITION) {
1381 ++mycpu->gd_cnt.v_intrans_coll;
1385 if (entry->maptype == VM_MAPTYPE_SUBMAP)
1387 if (entry->maptype == VM_MAPTYPE_UKSMAP)
1391 if (prev != &map->header) {
1392 prevsize = prev->end - prev->start;
1393 if ( (prev->end == entry->start) &&
1394 (prev->maptype == entry->maptype) &&
1395 (prev->object.vm_object == entry->object.vm_object) &&
1396 (!prev->object.vm_object ||
1397 (prev->offset + prevsize == entry->offset)) &&
1398 (prev->eflags == entry->eflags) &&
1399 (prev->protection == entry->protection) &&
1400 (prev->max_protection == entry->max_protection) &&
1401 (prev->inheritance == entry->inheritance) &&
1402 (prev->id == entry->id) &&
1403 (prev->wired_count == entry->wired_count)) {
1404 if (map->first_free == prev)
1405 map->first_free = entry;
1406 if (map->hint == prev)
1408 vm_map_entry_unlink(map, prev);
1409 entry->start = prev->start;
1410 entry->offset = prev->offset;
1411 if (prev->object.vm_object)
1412 vm_object_deallocate(prev->object.vm_object);
1413 vm_map_entry_dispose(map, prev, countp);
1418 if (next != &map->header) {
1419 esize = entry->end - entry->start;
1420 if ((entry->end == next->start) &&
1421 (next->maptype == entry->maptype) &&
1422 (next->object.vm_object == entry->object.vm_object) &&
1423 (!entry->object.vm_object ||
1424 (entry->offset + esize == next->offset)) &&
1425 (next->eflags == entry->eflags) &&
1426 (next->protection == entry->protection) &&
1427 (next->max_protection == entry->max_protection) &&
1428 (next->inheritance == entry->inheritance) &&
1429 (next->id == entry->id) &&
1430 (next->wired_count == entry->wired_count)) {
1431 if (map->first_free == next)
1432 map->first_free = entry;
1433 if (map->hint == next)
1435 vm_map_entry_unlink(map, next);
1436 entry->end = next->end;
1437 if (next->object.vm_object)
1438 vm_object_deallocate(next->object.vm_object);
1439 vm_map_entry_dispose(map, next, countp);
1445 * Asserts that the given entry begins at or after the specified address.
1446 * If necessary, it splits the entry into two.
1448 #define vm_map_clip_start(map, entry, startaddr, countp) \
1450 if (startaddr > entry->start) \
1451 _vm_map_clip_start(map, entry, startaddr, countp); \
1455 * This routine is called only when it is known that the entry must be split.
1457 * The map must be exclusively locked.
1460 _vm_map_clip_start(vm_map_t map, vm_map_entry_t entry, vm_offset_t start,
1463 vm_map_entry_t new_entry;
1466 * Split off the front portion -- note that we must insert the new
1467 * entry BEFORE this one, so that this entry has the specified
1471 vm_map_simplify_entry(map, entry, countp);
1474 * If there is no object backing this entry, we might as well create
1475 * one now. If we defer it, an object can get created after the map
1476 * is clipped, and individual objects will be created for the split-up
1477 * map. This is a bit of a hack, but is also about the best place to
1478 * put this improvement.
1480 if (entry->object.vm_object == NULL && !map->system_map) {
1481 vm_map_entry_allocate_object(entry);
1484 new_entry = vm_map_entry_create(map, countp);
1485 *new_entry = *entry;
1487 new_entry->end = start;
1488 entry->offset += (start - entry->start);
1489 entry->start = start;
1491 vm_map_entry_link(map, entry->prev, new_entry);
1493 switch(entry->maptype) {
1494 case VM_MAPTYPE_NORMAL:
1495 case VM_MAPTYPE_VPAGETABLE:
1496 if (new_entry->object.vm_object) {
1497 vm_object_hold(new_entry->object.vm_object);
1498 vm_object_chain_wait(new_entry->object.vm_object, 0);
1499 vm_object_reference_locked(new_entry->object.vm_object);
1500 vm_object_drop(new_entry->object.vm_object);
1509 * Asserts that the given entry ends at or before the specified address.
1510 * If necessary, it splits the entry into two.
1512 * The map must be exclusively locked.
1514 #define vm_map_clip_end(map, entry, endaddr, countp) \
1516 if (endaddr < entry->end) \
1517 _vm_map_clip_end(map, entry, endaddr, countp); \
1521 * This routine is called only when it is known that the entry must be split.
1523 * The map must be exclusively locked.
1526 _vm_map_clip_end(vm_map_t map, vm_map_entry_t entry, vm_offset_t end,
1529 vm_map_entry_t new_entry;
1532 * If there is no object backing this entry, we might as well create
1533 * one now. If we defer it, an object can get created after the map
1534 * is clipped, and individual objects will be created for the split-up
1535 * map. This is a bit of a hack, but is also about the best place to
1536 * put this improvement.
1539 if (entry->object.vm_object == NULL && !map->system_map) {
1540 vm_map_entry_allocate_object(entry);
1544 * Create a new entry and insert it AFTER the specified entry
1547 new_entry = vm_map_entry_create(map, countp);
1548 *new_entry = *entry;
1550 new_entry->start = entry->end = end;
1551 new_entry->offset += (end - entry->start);
1553 vm_map_entry_link(map, entry, new_entry);
1555 switch(entry->maptype) {
1556 case VM_MAPTYPE_NORMAL:
1557 case VM_MAPTYPE_VPAGETABLE:
1558 if (new_entry->object.vm_object) {
1559 vm_object_hold(new_entry->object.vm_object);
1560 vm_object_chain_wait(new_entry->object.vm_object, 0);
1561 vm_object_reference_locked(new_entry->object.vm_object);
1562 vm_object_drop(new_entry->object.vm_object);
1571 * Asserts that the starting and ending region addresses fall within the
1572 * valid range for the map.
1574 #define VM_MAP_RANGE_CHECK(map, start, end) \
1576 if (start < vm_map_min(map)) \
1577 start = vm_map_min(map); \
1578 if (end > vm_map_max(map)) \
1579 end = vm_map_max(map); \
1585 * Used to block when an in-transition collison occurs. The map
1586 * is unlocked for the sleep and relocked before the return.
1589 vm_map_transition_wait(vm_map_t map)
1591 tsleep_interlock(map, 0);
1593 tsleep(map, PINTERLOCKED, "vment", 0);
1598 * When we do blocking operations with the map lock held it is
1599 * possible that a clip might have occured on our in-transit entry,
1600 * requiring an adjustment to the entry in our loop. These macros
1601 * help the pageable and clip_range code deal with the case. The
1602 * conditional costs virtually nothing if no clipping has occured.
1605 #define CLIP_CHECK_BACK(entry, save_start) \
1607 while (entry->start != save_start) { \
1608 entry = entry->prev; \
1609 KASSERT(entry != &map->header, ("bad entry clip")); \
1613 #define CLIP_CHECK_FWD(entry, save_end) \
1615 while (entry->end != save_end) { \
1616 entry = entry->next; \
1617 KASSERT(entry != &map->header, ("bad entry clip")); \
1623 * Clip the specified range and return the base entry. The
1624 * range may cover several entries starting at the returned base
1625 * and the first and last entry in the covering sequence will be
1626 * properly clipped to the requested start and end address.
1628 * If no holes are allowed you should pass the MAP_CLIP_NO_HOLES
1631 * The MAP_ENTRY_IN_TRANSITION flag will be set for the entries
1632 * covered by the requested range.
1634 * The map must be exclusively locked on entry and will remain locked
1635 * on return. If no range exists or the range contains holes and you
1636 * specified that no holes were allowed, NULL will be returned. This
1637 * routine may temporarily unlock the map in order avoid a deadlock when
1642 vm_map_clip_range(vm_map_t map, vm_offset_t start, vm_offset_t end,
1643 int *countp, int flags)
1645 vm_map_entry_t start_entry;
1646 vm_map_entry_t entry;
1649 * Locate the entry and effect initial clipping. The in-transition
1650 * case does not occur very often so do not try to optimize it.
1653 if (vm_map_lookup_entry(map, start, &start_entry) == FALSE)
1655 entry = start_entry;
1656 if (entry->eflags & MAP_ENTRY_IN_TRANSITION) {
1657 entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP;
1658 ++mycpu->gd_cnt.v_intrans_coll;
1659 ++mycpu->gd_cnt.v_intrans_wait;
1660 vm_map_transition_wait(map);
1662 * entry and/or start_entry may have been clipped while
1663 * we slept, or may have gone away entirely. We have
1664 * to restart from the lookup.
1670 * Since we hold an exclusive map lock we do not have to restart
1671 * after clipping, even though clipping may block in zalloc.
1673 vm_map_clip_start(map, entry, start, countp);
1674 vm_map_clip_end(map, entry, end, countp);
1675 entry->eflags |= MAP_ENTRY_IN_TRANSITION;
1678 * Scan entries covered by the range. When working on the next
1679 * entry a restart need only re-loop on the current entry which
1680 * we have already locked, since 'next' may have changed. Also,
1681 * even though entry is safe, it may have been clipped so we
1682 * have to iterate forwards through the clip after sleeping.
1684 while (entry->next != &map->header && entry->next->start < end) {
1685 vm_map_entry_t next = entry->next;
1687 if (flags & MAP_CLIP_NO_HOLES) {
1688 if (next->start > entry->end) {
1689 vm_map_unclip_range(map, start_entry,
1690 start, entry->end, countp, flags);
1695 if (next->eflags & MAP_ENTRY_IN_TRANSITION) {
1696 vm_offset_t save_end = entry->end;
1697 next->eflags |= MAP_ENTRY_NEEDS_WAKEUP;
1698 ++mycpu->gd_cnt.v_intrans_coll;
1699 ++mycpu->gd_cnt.v_intrans_wait;
1700 vm_map_transition_wait(map);
1703 * clips might have occured while we blocked.
1705 CLIP_CHECK_FWD(entry, save_end);
1706 CLIP_CHECK_BACK(start_entry, start);
1710 * No restart necessary even though clip_end may block, we
1711 * are holding the map lock.
1713 vm_map_clip_end(map, next, end, countp);
1714 next->eflags |= MAP_ENTRY_IN_TRANSITION;
1717 if (flags & MAP_CLIP_NO_HOLES) {
1718 if (entry->end != end) {
1719 vm_map_unclip_range(map, start_entry,
1720 start, entry->end, countp, flags);
1724 return(start_entry);
1728 * Undo the effect of vm_map_clip_range(). You should pass the same
1729 * flags and the same range that you passed to vm_map_clip_range().
1730 * This code will clear the in-transition flag on the entries and
1731 * wake up anyone waiting. This code will also simplify the sequence
1732 * and attempt to merge it with entries before and after the sequence.
1734 * The map must be locked on entry and will remain locked on return.
1736 * Note that you should also pass the start_entry returned by
1737 * vm_map_clip_range(). However, if you block between the two calls
1738 * with the map unlocked please be aware that the start_entry may
1739 * have been clipped and you may need to scan it backwards to find
1740 * the entry corresponding with the original start address. You are
1741 * responsible for this, vm_map_unclip_range() expects the correct
1742 * start_entry to be passed to it and will KASSERT otherwise.
1746 vm_map_unclip_range(vm_map_t map, vm_map_entry_t start_entry,
1747 vm_offset_t start, vm_offset_t end,
1748 int *countp, int flags)
1750 vm_map_entry_t entry;
1752 entry = start_entry;
1754 KASSERT(entry->start == start, ("unclip_range: illegal base entry"));
1755 while (entry != &map->header && entry->start < end) {
1756 KASSERT(entry->eflags & MAP_ENTRY_IN_TRANSITION,
1757 ("in-transition flag not set during unclip on: %p",
1759 KASSERT(entry->end <= end,
1760 ("unclip_range: tail wasn't clipped"));
1761 entry->eflags &= ~MAP_ENTRY_IN_TRANSITION;
1762 if (entry->eflags & MAP_ENTRY_NEEDS_WAKEUP) {
1763 entry->eflags &= ~MAP_ENTRY_NEEDS_WAKEUP;
1766 entry = entry->next;
1770 * Simplification does not block so there is no restart case.
1772 entry = start_entry;
1773 while (entry != &map->header && entry->start < end) {
1774 vm_map_simplify_entry(map, entry, countp);
1775 entry = entry->next;
1780 * Mark the given range as handled by a subordinate map.
1782 * This range must have been created with vm_map_find(), and no other
1783 * operations may have been performed on this range prior to calling
1786 * Submappings cannot be removed.
1791 vm_map_submap(vm_map_t map, vm_offset_t start, vm_offset_t end, vm_map_t submap)
1793 vm_map_entry_t entry;
1794 int result = KERN_INVALID_ARGUMENT;
1797 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
1800 VM_MAP_RANGE_CHECK(map, start, end);
1802 if (vm_map_lookup_entry(map, start, &entry)) {
1803 vm_map_clip_start(map, entry, start, &count);
1805 entry = entry->next;
1808 vm_map_clip_end(map, entry, end, &count);
1810 if ((entry->start == start) && (entry->end == end) &&
1811 ((entry->eflags & MAP_ENTRY_COW) == 0) &&
1812 (entry->object.vm_object == NULL)) {
1813 entry->object.sub_map = submap;
1814 entry->maptype = VM_MAPTYPE_SUBMAP;
1815 result = KERN_SUCCESS;
1818 vm_map_entry_release(count);
1824 * Sets the protection of the specified address region in the target map.
1825 * If "set_max" is specified, the maximum protection is to be set;
1826 * otherwise, only the current protection is affected.
1828 * The protection is not applicable to submaps, but is applicable to normal
1829 * maps and maps governed by virtual page tables. For example, when operating
1830 * on a virtual page table our protection basically controls how COW occurs
1831 * on the backing object, whereas the virtual page table abstraction itself
1832 * is an abstraction for userland.
1837 vm_map_protect(vm_map_t map, vm_offset_t start, vm_offset_t end,
1838 vm_prot_t new_prot, boolean_t set_max)
1840 vm_map_entry_t current;
1841 vm_map_entry_t entry;
1844 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
1847 VM_MAP_RANGE_CHECK(map, start, end);
1849 if (vm_map_lookup_entry(map, start, &entry)) {
1850 vm_map_clip_start(map, entry, start, &count);
1852 entry = entry->next;
1856 * Make a first pass to check for protection violations.
1859 while ((current != &map->header) && (current->start < end)) {
1860 if (current->maptype == VM_MAPTYPE_SUBMAP) {
1862 vm_map_entry_release(count);
1863 return (KERN_INVALID_ARGUMENT);
1865 if ((new_prot & current->max_protection) != new_prot) {
1867 vm_map_entry_release(count);
1868 return (KERN_PROTECTION_FAILURE);
1870 current = current->next;
1874 * Go back and fix up protections. [Note that clipping is not
1875 * necessary the second time.]
1879 while ((current != &map->header) && (current->start < end)) {
1882 vm_map_clip_end(map, current, end, &count);
1884 old_prot = current->protection;
1886 current->protection =
1887 (current->max_protection = new_prot) &
1890 current->protection = new_prot;
1894 * Update physical map if necessary. Worry about copy-on-write
1895 * here -- CHECK THIS XXX
1898 if (current->protection != old_prot) {
1899 #define MASK(entry) (((entry)->eflags & MAP_ENTRY_COW) ? ~VM_PROT_WRITE : \
1902 pmap_protect(map->pmap, current->start,
1904 current->protection & MASK(current));
1908 vm_map_simplify_entry(map, current, &count);
1910 current = current->next;
1914 vm_map_entry_release(count);
1915 return (KERN_SUCCESS);
1919 * This routine traverses a processes map handling the madvise
1920 * system call. Advisories are classified as either those effecting
1921 * the vm_map_entry structure, or those effecting the underlying
1924 * The <value> argument is used for extended madvise calls.
1929 vm_map_madvise(vm_map_t map, vm_offset_t start, vm_offset_t end,
1930 int behav, off_t value)
1932 vm_map_entry_t current, entry;
1938 * Some madvise calls directly modify the vm_map_entry, in which case
1939 * we need to use an exclusive lock on the map and we need to perform
1940 * various clipping operations. Otherwise we only need a read-lock
1944 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
1948 case MADV_SEQUENTIAL:
1962 vm_map_lock_read(map);
1965 vm_map_entry_release(count);
1970 * Locate starting entry and clip if necessary.
1973 VM_MAP_RANGE_CHECK(map, start, end);
1975 if (vm_map_lookup_entry(map, start, &entry)) {
1977 vm_map_clip_start(map, entry, start, &count);
1979 entry = entry->next;
1984 * madvise behaviors that are implemented in the vm_map_entry.
1986 * We clip the vm_map_entry so that behavioral changes are
1987 * limited to the specified address range.
1989 for (current = entry;
1990 (current != &map->header) && (current->start < end);
1991 current = current->next
1993 if (current->maptype == VM_MAPTYPE_SUBMAP)
1996 vm_map_clip_end(map, current, end, &count);
2000 vm_map_entry_set_behavior(current, MAP_ENTRY_BEHAV_NORMAL);
2002 case MADV_SEQUENTIAL:
2003 vm_map_entry_set_behavior(current, MAP_ENTRY_BEHAV_SEQUENTIAL);
2006 vm_map_entry_set_behavior(current, MAP_ENTRY_BEHAV_RANDOM);
2009 current->eflags |= MAP_ENTRY_NOSYNC;
2012 current->eflags &= ~MAP_ENTRY_NOSYNC;
2015 current->eflags |= MAP_ENTRY_NOCOREDUMP;
2018 current->eflags &= ~MAP_ENTRY_NOCOREDUMP;
2022 * Invalidate the related pmap entries, used
2023 * to flush portions of the real kernel's
2024 * pmap when the caller has removed or
2025 * modified existing mappings in a virtual
2028 pmap_remove(map->pmap,
2029 current->start, current->end);
2033 * Set the page directory page for a map
2034 * governed by a virtual page table. Mark
2035 * the entry as being governed by a virtual
2036 * page table if it is not.
2038 * XXX the page directory page is stored
2039 * in the avail_ssize field if the map_entry.
2041 * XXX the map simplification code does not
2042 * compare this field so weird things may
2043 * happen if you do not apply this function
2044 * to the entire mapping governed by the
2045 * virtual page table.
2047 if (current->maptype != VM_MAPTYPE_VPAGETABLE) {
2051 current->aux.master_pde = value;
2052 pmap_remove(map->pmap,
2053 current->start, current->end);
2059 vm_map_simplify_entry(map, current, &count);
2067 * madvise behaviors that are implemented in the underlying
2070 * Since we don't clip the vm_map_entry, we have to clip
2071 * the vm_object pindex and count.
2073 * NOTE! We currently do not support these functions on
2074 * virtual page tables.
2076 for (current = entry;
2077 (current != &map->header) && (current->start < end);
2078 current = current->next
2080 vm_offset_t useStart;
2082 if (current->maptype != VM_MAPTYPE_NORMAL)
2085 pindex = OFF_TO_IDX(current->offset);
2086 count = atop(current->end - current->start);
2087 useStart = current->start;
2089 if (current->start < start) {
2090 pindex += atop(start - current->start);
2091 count -= atop(start - current->start);
2094 if (current->end > end)
2095 count -= atop(current->end - end);
2100 vm_object_madvise(current->object.vm_object,
2101 pindex, count, behav);
2104 * Try to populate the page table. Mappings governed
2105 * by virtual page tables cannot be pre-populated
2106 * without a lot of work so don't try.
2108 if (behav == MADV_WILLNEED &&
2109 current->maptype != VM_MAPTYPE_VPAGETABLE) {
2110 pmap_object_init_pt(
2113 current->protection,
2114 current->object.vm_object,
2116 (count << PAGE_SHIFT),
2117 MAP_PREFAULT_MADVISE
2121 vm_map_unlock_read(map);
2123 vm_map_entry_release(count);
2129 * Sets the inheritance of the specified address range in the target map.
2130 * Inheritance affects how the map will be shared with child maps at the
2131 * time of vm_map_fork.
2134 vm_map_inherit(vm_map_t map, vm_offset_t start, vm_offset_t end,
2135 vm_inherit_t new_inheritance)
2137 vm_map_entry_t entry;
2138 vm_map_entry_t temp_entry;
2141 switch (new_inheritance) {
2142 case VM_INHERIT_NONE:
2143 case VM_INHERIT_COPY:
2144 case VM_INHERIT_SHARE:
2147 return (KERN_INVALID_ARGUMENT);
2150 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
2153 VM_MAP_RANGE_CHECK(map, start, end);
2155 if (vm_map_lookup_entry(map, start, &temp_entry)) {
2157 vm_map_clip_start(map, entry, start, &count);
2159 entry = temp_entry->next;
2161 while ((entry != &map->header) && (entry->start < end)) {
2162 vm_map_clip_end(map, entry, end, &count);
2164 entry->inheritance = new_inheritance;
2166 vm_map_simplify_entry(map, entry, &count);
2168 entry = entry->next;
2171 vm_map_entry_release(count);
2172 return (KERN_SUCCESS);
2176 * Implement the semantics of mlock
2179 vm_map_unwire(vm_map_t map, vm_offset_t start, vm_offset_t real_end,
2180 boolean_t new_pageable)
2182 vm_map_entry_t entry;
2183 vm_map_entry_t start_entry;
2185 int rv = KERN_SUCCESS;
2188 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
2190 VM_MAP_RANGE_CHECK(map, start, real_end);
2193 start_entry = vm_map_clip_range(map, start, end, &count,
2195 if (start_entry == NULL) {
2197 vm_map_entry_release(count);
2198 return (KERN_INVALID_ADDRESS);
2201 if (new_pageable == 0) {
2202 entry = start_entry;
2203 while ((entry != &map->header) && (entry->start < end)) {
2204 vm_offset_t save_start;
2205 vm_offset_t save_end;
2208 * Already user wired or hard wired (trivial cases)
2210 if (entry->eflags & MAP_ENTRY_USER_WIRED) {
2211 entry = entry->next;
2214 if (entry->wired_count != 0) {
2215 entry->wired_count++;
2216 entry->eflags |= MAP_ENTRY_USER_WIRED;
2217 entry = entry->next;
2222 * A new wiring requires instantiation of appropriate
2223 * management structures and the faulting in of the
2226 if (entry->maptype == VM_MAPTYPE_NORMAL ||
2227 entry->maptype == VM_MAPTYPE_VPAGETABLE) {
2228 int copyflag = entry->eflags &
2229 MAP_ENTRY_NEEDS_COPY;
2230 if (copyflag && ((entry->protection &
2231 VM_PROT_WRITE) != 0)) {
2232 vm_map_entry_shadow(entry, 0);
2233 } else if (entry->object.vm_object == NULL &&
2235 vm_map_entry_allocate_object(entry);
2238 entry->wired_count++;
2239 entry->eflags |= MAP_ENTRY_USER_WIRED;
2242 * Now fault in the area. Note that vm_fault_wire()
2243 * may release the map lock temporarily, it will be
2244 * relocked on return. The in-transition
2245 * flag protects the entries.
2247 save_start = entry->start;
2248 save_end = entry->end;
2249 rv = vm_fault_wire(map, entry, TRUE, 0);
2251 CLIP_CHECK_BACK(entry, save_start);
2253 KASSERT(entry->wired_count == 1, ("bad wired_count on entry"));
2254 entry->eflags &= ~MAP_ENTRY_USER_WIRED;
2255 entry->wired_count = 0;
2256 if (entry->end == save_end)
2258 entry = entry->next;
2259 KASSERT(entry != &map->header, ("bad entry clip during backout"));
2261 end = save_start; /* unwire the rest */
2265 * note that even though the entry might have been
2266 * clipped, the USER_WIRED flag we set prevents
2267 * duplication so we do not have to do a
2270 entry = entry->next;
2274 * If we failed fall through to the unwiring section to
2275 * unwire what we had wired so far. 'end' has already
2282 * start_entry might have been clipped if we unlocked the
2283 * map and blocked. No matter how clipped it has gotten
2284 * there should be a fragment that is on our start boundary.
2286 CLIP_CHECK_BACK(start_entry, start);
2290 * Deal with the unwiring case.
2294 * This is the unwiring case. We must first ensure that the
2295 * range to be unwired is really wired down. We know there
2298 entry = start_entry;
2299 while ((entry != &map->header) && (entry->start < end)) {
2300 if ((entry->eflags & MAP_ENTRY_USER_WIRED) == 0) {
2301 rv = KERN_INVALID_ARGUMENT;
2304 KASSERT(entry->wired_count != 0, ("wired count was 0 with USER_WIRED set! %p", entry));
2305 entry = entry->next;
2309 * Now decrement the wiring count for each region. If a region
2310 * becomes completely unwired, unwire its physical pages and
2314 * The map entries are processed in a loop, checking to
2315 * make sure the entry is wired and asserting it has a wired
2316 * count. However, another loop was inserted more-or-less in
2317 * the middle of the unwiring path. This loop picks up the
2318 * "entry" loop variable from the first loop without first
2319 * setting it to start_entry. Naturally, the secound loop
2320 * is never entered and the pages backing the entries are
2321 * never unwired. This can lead to a leak of wired pages.
2323 entry = start_entry;
2324 while ((entry != &map->header) && (entry->start < end)) {
2325 KASSERT(entry->eflags & MAP_ENTRY_USER_WIRED,
2326 ("expected USER_WIRED on entry %p", entry));
2327 entry->eflags &= ~MAP_ENTRY_USER_WIRED;
2328 entry->wired_count--;
2329 if (entry->wired_count == 0)
2330 vm_fault_unwire(map, entry);
2331 entry = entry->next;
2335 vm_map_unclip_range(map, start_entry, start, real_end, &count,
2339 vm_map_entry_release(count);
2344 * Sets the pageability of the specified address range in the target map.
2345 * Regions specified as not pageable require locked-down physical
2346 * memory and physical page maps.
2348 * The map must not be locked, but a reference must remain to the map
2349 * throughout the call.
2351 * This function may be called via the zalloc path and must properly
2352 * reserve map entries for kernel_map.
2357 vm_map_wire(vm_map_t map, vm_offset_t start, vm_offset_t real_end, int kmflags)
2359 vm_map_entry_t entry;
2360 vm_map_entry_t start_entry;
2362 int rv = KERN_SUCCESS;
2365 if (kmflags & KM_KRESERVE)
2366 count = vm_map_entry_kreserve(MAP_RESERVE_COUNT);
2368 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
2370 VM_MAP_RANGE_CHECK(map, start, real_end);
2373 start_entry = vm_map_clip_range(map, start, end, &count,
2375 if (start_entry == NULL) {
2377 rv = KERN_INVALID_ADDRESS;
2380 if ((kmflags & KM_PAGEABLE) == 0) {
2384 * 1. Holding the write lock, we create any shadow or zero-fill
2385 * objects that need to be created. Then we clip each map
2386 * entry to the region to be wired and increment its wiring
2387 * count. We create objects before clipping the map entries
2388 * to avoid object proliferation.
2390 * 2. We downgrade to a read lock, and call vm_fault_wire to
2391 * fault in the pages for any newly wired area (wired_count is
2394 * Downgrading to a read lock for vm_fault_wire avoids a
2395 * possible deadlock with another process that may have faulted
2396 * on one of the pages to be wired (it would mark the page busy,
2397 * blocking us, then in turn block on the map lock that we
2398 * hold). Because of problems in the recursive lock package,
2399 * we cannot upgrade to a write lock in vm_map_lookup. Thus,
2400 * any actions that require the write lock must be done
2401 * beforehand. Because we keep the read lock on the map, the
2402 * copy-on-write status of the entries we modify here cannot
2405 entry = start_entry;
2406 while ((entry != &map->header) && (entry->start < end)) {
2408 * Trivial case if the entry is already wired
2410 if (entry->wired_count) {
2411 entry->wired_count++;
2412 entry = entry->next;
2417 * The entry is being newly wired, we have to setup
2418 * appropriate management structures. A shadow
2419 * object is required for a copy-on-write region,
2420 * or a normal object for a zero-fill region. We
2421 * do not have to do this for entries that point to sub
2422 * maps because we won't hold the lock on the sub map.
2424 if (entry->maptype == VM_MAPTYPE_NORMAL ||
2425 entry->maptype == VM_MAPTYPE_VPAGETABLE) {
2426 int copyflag = entry->eflags &
2427 MAP_ENTRY_NEEDS_COPY;
2428 if (copyflag && ((entry->protection &
2429 VM_PROT_WRITE) != 0)) {
2430 vm_map_entry_shadow(entry, 0);
2431 } else if (entry->object.vm_object == NULL &&
2433 vm_map_entry_allocate_object(entry);
2437 entry->wired_count++;
2438 entry = entry->next;
2446 * HACK HACK HACK HACK
2448 * vm_fault_wire() temporarily unlocks the map to avoid
2449 * deadlocks. The in-transition flag from vm_map_clip_range
2450 * call should protect us from changes while the map is
2453 * NOTE: Previously this comment stated that clipping might
2454 * still occur while the entry is unlocked, but from
2455 * what I can tell it actually cannot.
2457 * It is unclear whether the CLIP_CHECK_*() calls
2458 * are still needed but we keep them in anyway.
2460 * HACK HACK HACK HACK
2463 entry = start_entry;
2464 while (entry != &map->header && entry->start < end) {
2466 * If vm_fault_wire fails for any page we need to undo
2467 * what has been done. We decrement the wiring count
2468 * for those pages which have not yet been wired (now)
2469 * and unwire those that have (later).
2471 vm_offset_t save_start = entry->start;
2472 vm_offset_t save_end = entry->end;
2474 if (entry->wired_count == 1)
2475 rv = vm_fault_wire(map, entry, FALSE, kmflags);
2477 CLIP_CHECK_BACK(entry, save_start);
2479 KASSERT(entry->wired_count == 1, ("wired_count changed unexpectedly"));
2480 entry->wired_count = 0;
2481 if (entry->end == save_end)
2483 entry = entry->next;
2484 KASSERT(entry != &map->header, ("bad entry clip during backout"));
2489 CLIP_CHECK_FWD(entry, save_end);
2490 entry = entry->next;
2494 * If a failure occured undo everything by falling through
2495 * to the unwiring code. 'end' has already been adjusted
2499 kmflags |= KM_PAGEABLE;
2502 * start_entry is still IN_TRANSITION but may have been
2503 * clipped since vm_fault_wire() unlocks and relocks the
2504 * map. No matter how clipped it has gotten there should
2505 * be a fragment that is on our start boundary.
2507 CLIP_CHECK_BACK(start_entry, start);
2510 if (kmflags & KM_PAGEABLE) {
2512 * This is the unwiring case. We must first ensure that the
2513 * range to be unwired is really wired down. We know there
2516 entry = start_entry;
2517 while ((entry != &map->header) && (entry->start < end)) {
2518 if (entry->wired_count == 0) {
2519 rv = KERN_INVALID_ARGUMENT;
2522 entry = entry->next;
2526 * Now decrement the wiring count for each region. If a region
2527 * becomes completely unwired, unwire its physical pages and
2530 entry = start_entry;
2531 while ((entry != &map->header) && (entry->start < end)) {
2532 entry->wired_count--;
2533 if (entry->wired_count == 0)
2534 vm_fault_unwire(map, entry);
2535 entry = entry->next;
2539 vm_map_unclip_range(map, start_entry, start, real_end,
2540 &count, MAP_CLIP_NO_HOLES);
2544 if (kmflags & KM_KRESERVE)
2545 vm_map_entry_krelease(count);
2547 vm_map_entry_release(count);
2552 * Mark a newly allocated address range as wired but do not fault in
2553 * the pages. The caller is expected to load the pages into the object.
2555 * The map must be locked on entry and will remain locked on return.
2556 * No other requirements.
2559 vm_map_set_wired_quick(vm_map_t map, vm_offset_t addr, vm_size_t size,
2562 vm_map_entry_t scan;
2563 vm_map_entry_t entry;
2565 entry = vm_map_clip_range(map, addr, addr + size,
2566 countp, MAP_CLIP_NO_HOLES);
2568 scan != &map->header && scan->start < addr + size;
2569 scan = scan->next) {
2570 KKASSERT(scan->wired_count == 0);
2571 scan->wired_count = 1;
2573 vm_map_unclip_range(map, entry, addr, addr + size,
2574 countp, MAP_CLIP_NO_HOLES);
2578 * Push any dirty cached pages in the address range to their pager.
2579 * If syncio is TRUE, dirty pages are written synchronously.
2580 * If invalidate is TRUE, any cached pages are freed as well.
2582 * This routine is called by sys_msync()
2584 * Returns an error if any part of the specified range is not mapped.
2589 vm_map_clean(vm_map_t map, vm_offset_t start, vm_offset_t end,
2590 boolean_t syncio, boolean_t invalidate)
2592 vm_map_entry_t current;
2593 vm_map_entry_t entry;
2597 vm_ooffset_t offset;
2599 vm_map_lock_read(map);
2600 VM_MAP_RANGE_CHECK(map, start, end);
2601 if (!vm_map_lookup_entry(map, start, &entry)) {
2602 vm_map_unlock_read(map);
2603 return (KERN_INVALID_ADDRESS);
2605 lwkt_gettoken(&map->token);
2608 * Make a first pass to check for holes.
2610 for (current = entry; current->start < end; current = current->next) {
2611 if (current->maptype == VM_MAPTYPE_SUBMAP) {
2612 lwkt_reltoken(&map->token);
2613 vm_map_unlock_read(map);
2614 return (KERN_INVALID_ARGUMENT);
2616 if (end > current->end &&
2617 (current->next == &map->header ||
2618 current->end != current->next->start)) {
2619 lwkt_reltoken(&map->token);
2620 vm_map_unlock_read(map);
2621 return (KERN_INVALID_ADDRESS);
2626 pmap_remove(vm_map_pmap(map), start, end);
2629 * Make a second pass, cleaning/uncaching pages from the indicated
2632 for (current = entry; current->start < end; current = current->next) {
2633 offset = current->offset + (start - current->start);
2634 size = (end <= current->end ? end : current->end) - start;
2636 switch(current->maptype) {
2637 case VM_MAPTYPE_SUBMAP:
2640 vm_map_entry_t tentry;
2643 smap = current->object.sub_map;
2644 vm_map_lock_read(smap);
2645 vm_map_lookup_entry(smap, offset, &tentry);
2646 tsize = tentry->end - offset;
2649 object = tentry->object.vm_object;
2650 offset = tentry->offset + (offset - tentry->start);
2651 vm_map_unlock_read(smap);
2654 case VM_MAPTYPE_NORMAL:
2655 case VM_MAPTYPE_VPAGETABLE:
2656 object = current->object.vm_object;
2664 vm_object_hold(object);
2667 * Note that there is absolutely no sense in writing out
2668 * anonymous objects, so we track down the vnode object
2670 * We invalidate (remove) all pages from the address space
2671 * anyway, for semantic correctness.
2673 * note: certain anonymous maps, such as MAP_NOSYNC maps,
2674 * may start out with a NULL object.
2676 while (object && (tobj = object->backing_object) != NULL) {
2677 vm_object_hold(tobj);
2678 if (tobj == object->backing_object) {
2679 vm_object_lock_swap();
2680 offset += object->backing_object_offset;
2681 vm_object_drop(object);
2683 if (object->size < OFF_TO_IDX(offset + size))
2684 size = IDX_TO_OFF(object->size) -
2688 vm_object_drop(tobj);
2690 if (object && (object->type == OBJT_VNODE) &&
2691 (current->protection & VM_PROT_WRITE) &&
2692 (object->flags & OBJ_NOMSYNC) == 0) {
2694 * Flush pages if writing is allowed, invalidate them
2695 * if invalidation requested. Pages undergoing I/O
2696 * will be ignored by vm_object_page_remove().
2698 * We cannot lock the vnode and then wait for paging
2699 * to complete without deadlocking against vm_fault.
2700 * Instead we simply call vm_object_page_remove() and
2701 * allow it to block internally on a page-by-page
2702 * basis when it encounters pages undergoing async
2707 /* no chain wait needed for vnode objects */
2708 vm_object_reference_locked(object);
2709 vn_lock(object->handle, LK_EXCLUSIVE | LK_RETRY);
2710 flags = (syncio || invalidate) ? OBJPC_SYNC : 0;
2711 flags |= invalidate ? OBJPC_INVAL : 0;
2714 * When operating on a virtual page table just
2715 * flush the whole object. XXX we probably ought
2718 switch(current->maptype) {
2719 case VM_MAPTYPE_NORMAL:
2720 vm_object_page_clean(object,
2722 OFF_TO_IDX(offset + size + PAGE_MASK),
2725 case VM_MAPTYPE_VPAGETABLE:
2726 vm_object_page_clean(object, 0, 0, flags);
2729 vn_unlock(((struct vnode *)object->handle));
2730 vm_object_deallocate_locked(object);
2732 if (object && invalidate &&
2733 ((object->type == OBJT_VNODE) ||
2734 (object->type == OBJT_DEVICE) ||
2735 (object->type == OBJT_MGTDEVICE))) {
2737 ((object->type == OBJT_DEVICE) ||
2738 (object->type == OBJT_MGTDEVICE)) ? FALSE : TRUE;
2739 /* no chain wait needed for vnode/device objects */
2740 vm_object_reference_locked(object);
2741 switch(current->maptype) {
2742 case VM_MAPTYPE_NORMAL:
2743 vm_object_page_remove(object,
2745 OFF_TO_IDX(offset + size + PAGE_MASK),
2748 case VM_MAPTYPE_VPAGETABLE:
2749 vm_object_page_remove(object, 0, 0, clean_only);
2752 vm_object_deallocate_locked(object);
2756 vm_object_drop(object);
2759 lwkt_reltoken(&map->token);
2760 vm_map_unlock_read(map);
2762 return (KERN_SUCCESS);
2766 * Make the region specified by this entry pageable.
2768 * The vm_map must be exclusively locked.
2771 vm_map_entry_unwire(vm_map_t map, vm_map_entry_t entry)
2773 entry->eflags &= ~MAP_ENTRY_USER_WIRED;
2774 entry->wired_count = 0;
2775 vm_fault_unwire(map, entry);
2779 * Deallocate the given entry from the target map.
2781 * The vm_map must be exclusively locked.
2784 vm_map_entry_delete(vm_map_t map, vm_map_entry_t entry, int *countp)
2786 vm_map_entry_unlink(map, entry);
2787 map->size -= entry->end - entry->start;
2789 switch(entry->maptype) {
2790 case VM_MAPTYPE_NORMAL:
2791 case VM_MAPTYPE_VPAGETABLE:
2792 case VM_MAPTYPE_SUBMAP:
2793 vm_object_deallocate(entry->object.vm_object);
2795 case VM_MAPTYPE_UKSMAP:
2802 vm_map_entry_dispose(map, entry, countp);
2806 * Deallocates the given address range from the target map.
2808 * The vm_map must be exclusively locked.
2811 vm_map_delete(vm_map_t map, vm_offset_t start, vm_offset_t end, int *countp)
2814 vm_map_entry_t entry;
2815 vm_map_entry_t first_entry;
2817 ASSERT_VM_MAP_LOCKED(map);
2818 lwkt_gettoken(&map->token);
2821 * Find the start of the region, and clip it. Set entry to point
2822 * at the first record containing the requested address or, if no
2823 * such record exists, the next record with a greater address. The
2824 * loop will run from this point until a record beyond the termination
2825 * address is encountered.
2827 * map->hint must be adjusted to not point to anything we delete,
2828 * so set it to the entry prior to the one being deleted.
2830 * GGG see other GGG comment.
2832 if (vm_map_lookup_entry(map, start, &first_entry)) {
2833 entry = first_entry;
2834 vm_map_clip_start(map, entry, start, countp);
2835 map->hint = entry->prev; /* possible problem XXX */
2837 map->hint = first_entry; /* possible problem XXX */
2838 entry = first_entry->next;
2842 * If a hole opens up prior to the current first_free then
2843 * adjust first_free. As with map->hint, map->first_free
2844 * cannot be left set to anything we might delete.
2846 if (entry == &map->header) {
2847 map->first_free = &map->header;
2848 } else if (map->first_free->start >= start) {
2849 map->first_free = entry->prev;
2853 * Step through all entries in this region
2855 while ((entry != &map->header) && (entry->start < end)) {
2856 vm_map_entry_t next;
2858 vm_pindex_t offidxstart, offidxend, count;
2861 * If we hit an in-transition entry we have to sleep and
2862 * retry. It's easier (and not really slower) to just retry
2863 * since this case occurs so rarely and the hint is already
2864 * pointing at the right place. We have to reset the
2865 * start offset so as not to accidently delete an entry
2866 * another process just created in vacated space.
2868 if (entry->eflags & MAP_ENTRY_IN_TRANSITION) {
2869 entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP;
2870 start = entry->start;
2871 ++mycpu->gd_cnt.v_intrans_coll;
2872 ++mycpu->gd_cnt.v_intrans_wait;
2873 vm_map_transition_wait(map);
2876 vm_map_clip_end(map, entry, end, countp);
2882 offidxstart = OFF_TO_IDX(entry->offset);
2883 count = OFF_TO_IDX(e - s);
2885 switch(entry->maptype) {
2886 case VM_MAPTYPE_NORMAL:
2887 case VM_MAPTYPE_VPAGETABLE:
2888 case VM_MAPTYPE_SUBMAP:
2889 object = entry->object.vm_object;
2897 * Unwire before removing addresses from the pmap; otherwise,
2898 * unwiring will put the entries back in the pmap.
2900 if (entry->wired_count != 0)
2901 vm_map_entry_unwire(map, entry);
2903 offidxend = offidxstart + count;
2905 if (object == &kernel_object) {
2906 vm_object_hold(object);
2907 vm_object_page_remove(object, offidxstart,
2909 vm_object_drop(object);
2910 } else if (object && object->type != OBJT_DEFAULT &&
2911 object->type != OBJT_SWAP) {
2913 * vnode object routines cannot be chain-locked,
2914 * but since we aren't removing pages from the
2915 * object here we can use a shared hold.
2917 vm_object_hold_shared(object);
2918 pmap_remove(map->pmap, s, e);
2919 vm_object_drop(object);
2920 } else if (object) {
2921 vm_object_hold(object);
2922 vm_object_chain_acquire(object, 0);
2923 pmap_remove(map->pmap, s, e);
2925 if (object != NULL &&
2926 object->ref_count != 1 &&
2927 (object->flags & (OBJ_NOSPLIT|OBJ_ONEMAPPING)) ==
2929 (object->type == OBJT_DEFAULT ||
2930 object->type == OBJT_SWAP)) {
2931 vm_object_collapse(object, NULL);
2932 vm_object_page_remove(object, offidxstart,
2934 if (object->type == OBJT_SWAP) {
2935 swap_pager_freespace(object,
2939 if (offidxend >= object->size &&
2940 offidxstart < object->size) {
2941 object->size = offidxstart;
2944 vm_object_chain_release(object);
2945 vm_object_drop(object);
2946 } else if (entry->maptype == VM_MAPTYPE_UKSMAP) {
2947 pmap_remove(map->pmap, s, e);
2951 * Delete the entry (which may delete the object) only after
2952 * removing all pmap entries pointing to its pages.
2953 * (Otherwise, its page frames may be reallocated, and any
2954 * modify bits will be set in the wrong object!)
2956 vm_map_entry_delete(map, entry, countp);
2959 lwkt_reltoken(&map->token);
2960 return (KERN_SUCCESS);
2964 * Remove the given address range from the target map.
2965 * This is the exported form of vm_map_delete.
2970 vm_map_remove(vm_map_t map, vm_offset_t start, vm_offset_t end)
2975 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
2977 VM_MAP_RANGE_CHECK(map, start, end);
2978 result = vm_map_delete(map, start, end, &count);
2980 vm_map_entry_release(count);
2986 * Assert that the target map allows the specified privilege on the
2987 * entire address region given. The entire region must be allocated.
2989 * The caller must specify whether the vm_map is already locked or not.
2992 vm_map_check_protection(vm_map_t map, vm_offset_t start, vm_offset_t end,
2993 vm_prot_t protection, boolean_t have_lock)
2995 vm_map_entry_t entry;
2996 vm_map_entry_t tmp_entry;
2999 if (have_lock == FALSE)
3000 vm_map_lock_read(map);
3002 if (!vm_map_lookup_entry(map, start, &tmp_entry)) {
3003 if (have_lock == FALSE)
3004 vm_map_unlock_read(map);
3010 while (start < end) {
3011 if (entry == &map->header) {
3019 if (start < entry->start) {
3024 * Check protection associated with entry.
3027 if ((entry->protection & protection) != protection) {
3031 /* go to next entry */
3034 entry = entry->next;
3036 if (have_lock == FALSE)
3037 vm_map_unlock_read(map);
3042 * If appropriate this function shadows the original object with a new object
3043 * and moves the VM pages from the original object to the new object.
3044 * The original object will also be collapsed, if possible.
3046 * We can only do this for normal memory objects with a single mapping, and
3047 * it only makes sense to do it if there are 2 or more refs on the original
3048 * object. i.e. typically a memory object that has been extended into
3049 * multiple vm_map_entry's with non-overlapping ranges.
3051 * This makes it easier to remove unused pages and keeps object inheritance
3052 * from being a negative impact on memory usage.
3054 * On return the (possibly new) entry->object.vm_object will have an
3055 * additional ref on it for the caller to dispose of (usually by cloning
3056 * the vm_map_entry). The additional ref had to be done in this routine
3057 * to avoid racing a collapse. The object's ONEMAPPING flag will also be
3060 * The vm_map must be locked and its token held.
3063 vm_map_split(vm_map_entry_t entry)
3066 vm_object_t oobject, nobject, bobject;
3069 vm_pindex_t offidxstart, offidxend, idx;
3071 vm_ooffset_t offset;
3075 * Optimize away object locks for vnode objects. Important exit/exec
3078 * OBJ_ONEMAPPING doesn't apply to vnode objects but clear the flag
3081 oobject = entry->object.vm_object;
3082 if (oobject->type != OBJT_DEFAULT && oobject->type != OBJT_SWAP) {
3083 vm_object_reference_quick(oobject);
3084 vm_object_clear_flag(oobject, OBJ_ONEMAPPING);
3089 * Setup. Chain lock the original object throughout the entire
3090 * routine to prevent new page faults from occuring.
3092 * XXX can madvise WILLNEED interfere with us too?
3094 vm_object_hold(oobject);
3095 vm_object_chain_acquire(oobject, 0);
3098 * Original object cannot be split? Might have also changed state.
3100 if (oobject->handle == NULL || (oobject->type != OBJT_DEFAULT &&
3101 oobject->type != OBJT_SWAP)) {
3102 vm_object_chain_release(oobject);
3103 vm_object_reference_locked(oobject);
3104 vm_object_clear_flag(oobject, OBJ_ONEMAPPING);
3105 vm_object_drop(oobject);
3110 * Collapse original object with its backing store as an
3111 * optimization to reduce chain lengths when possible.
3113 * If ref_count <= 1 there aren't other non-overlapping vm_map_entry's
3114 * for oobject, so there's no point collapsing it.
3116 * Then re-check whether the object can be split.
3118 vm_object_collapse(oobject, NULL);
3120 if (oobject->ref_count <= 1 ||
3121 (oobject->type != OBJT_DEFAULT && oobject->type != OBJT_SWAP) ||
3122 (oobject->flags & (OBJ_NOSPLIT|OBJ_ONEMAPPING)) != OBJ_ONEMAPPING) {
3123 vm_object_chain_release(oobject);
3124 vm_object_reference_locked(oobject);
3125 vm_object_clear_flag(oobject, OBJ_ONEMAPPING);
3126 vm_object_drop(oobject);
3131 * Acquire the chain lock on the backing object.
3133 * Give bobject an additional ref count for when it will be shadowed
3137 if ((bobject = oobject->backing_object) != NULL) {
3138 if (bobject->type != OBJT_VNODE) {
3140 vm_object_hold(bobject);
3141 vm_object_chain_wait(bobject, 0);
3142 /* ref for shadowing below */
3143 vm_object_reference_locked(bobject);
3144 vm_object_chain_acquire(bobject, 0);
3145 KKASSERT(bobject->backing_object == bobject);
3146 KKASSERT((bobject->flags & OBJ_DEAD) == 0);
3149 * vnodes are not placed on the shadow list but
3150 * they still get another ref for the backing_object
3153 vm_object_reference_quick(bobject);
3158 * Calculate the object page range and allocate the new object.
3160 offset = entry->offset;
3164 offidxstart = OFF_TO_IDX(offset);
3165 offidxend = offidxstart + OFF_TO_IDX(e - s);
3166 size = offidxend - offidxstart;
3168 switch(oobject->type) {
3170 nobject = default_pager_alloc(NULL, IDX_TO_OFF(size),
3174 nobject = swap_pager_alloc(NULL, IDX_TO_OFF(size),
3183 if (nobject == NULL) {
3185 if (useshadowlist) {
3186 vm_object_chain_release(bobject);
3187 vm_object_deallocate(bobject);
3188 vm_object_drop(bobject);
3190 vm_object_deallocate(bobject);
3193 vm_object_chain_release(oobject);
3194 vm_object_reference_locked(oobject);
3195 vm_object_clear_flag(oobject, OBJ_ONEMAPPING);
3196 vm_object_drop(oobject);
3201 * The new object will replace entry->object.vm_object so it needs
3202 * a second reference (the caller expects an additional ref).
3204 vm_object_hold(nobject);
3205 vm_object_reference_locked(nobject);
3206 vm_object_chain_acquire(nobject, 0);
3209 * nobject shadows bobject (oobject already shadows bobject).
3211 * Adding an object to bobject's shadow list requires refing bobject
3212 * which we did above in the useshadowlist case.
3215 nobject->backing_object_offset =
3216 oobject->backing_object_offset + IDX_TO_OFF(offidxstart);
3217 nobject->backing_object = bobject;
3218 if (useshadowlist) {
3219 bobject->shadow_count++;
3220 bobject->generation++;
3221 LIST_INSERT_HEAD(&bobject->shadow_head,
3222 nobject, shadow_list);
3223 vm_object_clear_flag(bobject, OBJ_ONEMAPPING); /*XXX*/
3224 vm_object_chain_release(bobject);
3225 vm_object_drop(bobject);
3226 vm_object_set_flag(nobject, OBJ_ONSHADOW);
3231 * Move the VM pages from oobject to nobject
3233 for (idx = 0; idx < size; idx++) {
3236 m = vm_page_lookup_busy_wait(oobject, offidxstart + idx,
3242 * We must wait for pending I/O to complete before we can
3245 * We do not have to VM_PROT_NONE the page as mappings should
3246 * not be changed by this operation.
3248 * NOTE: The act of renaming a page updates chaingen for both
3251 vm_page_rename(m, nobject, idx);
3252 /* page automatically made dirty by rename and cache handled */
3253 /* page remains busy */
3256 if (oobject->type == OBJT_SWAP) {
3257 vm_object_pip_add(oobject, 1);
3259 * copy oobject pages into nobject and destroy unneeded
3260 * pages in shadow object.
3262 swap_pager_copy(oobject, nobject, offidxstart, 0);
3263 vm_object_pip_wakeup(oobject);
3267 * Wakeup the pages we played with. No spl protection is needed
3268 * for a simple wakeup.
3270 for (idx = 0; idx < size; idx++) {
3271 m = vm_page_lookup(nobject, idx);
3273 KKASSERT(m->flags & PG_BUSY);
3277 entry->object.vm_object = nobject;
3278 entry->offset = 0LL;
3283 * NOTE: There is no need to remove OBJ_ONEMAPPING from oobject, the
3284 * related pages were moved and are no longer applicable to the
3287 * NOTE: Deallocate oobject (due to its entry->object.vm_object being
3288 * replaced by nobject).
3290 vm_object_chain_release(nobject);
3291 vm_object_drop(nobject);
3292 if (bobject && useshadowlist) {
3293 vm_object_chain_release(bobject);
3294 vm_object_drop(bobject);
3296 vm_object_chain_release(oobject);
3297 /*vm_object_clear_flag(oobject, OBJ_ONEMAPPING);*/
3298 vm_object_deallocate_locked(oobject);
3299 vm_object_drop(oobject);
3303 * Copies the contents of the source entry to the destination
3304 * entry. The entries *must* be aligned properly.
3306 * The vm_maps must be exclusively locked.
3307 * The vm_map's token must be held.
3309 * Because the maps are locked no faults can be in progress during the
3313 vm_map_copy_entry(vm_map_t src_map, vm_map_t dst_map,
3314 vm_map_entry_t src_entry, vm_map_entry_t dst_entry)
3316 vm_object_t src_object;
3318 if (dst_entry->maptype == VM_MAPTYPE_SUBMAP ||
3319 dst_entry->maptype == VM_MAPTYPE_UKSMAP)
3321 if (src_entry->maptype == VM_MAPTYPE_SUBMAP ||
3322 src_entry->maptype == VM_MAPTYPE_UKSMAP)
3325 if (src_entry->wired_count == 0) {
3327 * If the source entry is marked needs_copy, it is already
3330 if ((src_entry->eflags & MAP_ENTRY_NEEDS_COPY) == 0) {
3331 pmap_protect(src_map->pmap,
3334 src_entry->protection & ~VM_PROT_WRITE);
3338 * Make a copy of the object.
3340 * The object must be locked prior to checking the object type
3341 * and for the call to vm_object_collapse() and vm_map_split().
3342 * We cannot use *_hold() here because the split code will
3343 * probably try to destroy the object. The lock is a pool
3344 * token and doesn't care.
3346 * We must bump src_map->timestamp when setting
3347 * MAP_ENTRY_NEEDS_COPY to force any concurrent fault
3348 * to retry, otherwise the concurrent fault might improperly
3349 * install a RW pte when its supposed to be a RO(COW) pte.
3350 * This race can occur because a vnode-backed fault may have
3351 * to temporarily release the map lock.
3353 if (src_entry->object.vm_object != NULL) {
3354 vm_map_split(src_entry);
3355 src_object = src_entry->object.vm_object;
3356 dst_entry->object.vm_object = src_object;
3357 src_entry->eflags |= (MAP_ENTRY_COW |
3358 MAP_ENTRY_NEEDS_COPY);
3359 dst_entry->eflags |= (MAP_ENTRY_COW |
3360 MAP_ENTRY_NEEDS_COPY);
3361 dst_entry->offset = src_entry->offset;
3362 ++src_map->timestamp;
3364 dst_entry->object.vm_object = NULL;
3365 dst_entry->offset = 0;
3368 pmap_copy(dst_map->pmap, src_map->pmap, dst_entry->start,
3369 dst_entry->end - dst_entry->start, src_entry->start);
3372 * Of course, wired down pages can't be set copy-on-write.
3373 * Cause wired pages to be copied into the new map by
3374 * simulating faults (the new pages are pageable)
3376 vm_fault_copy_entry(dst_map, src_map, dst_entry, src_entry);
3382 * Create a new process vmspace structure and vm_map
3383 * based on those of an existing process. The new map
3384 * is based on the old map, according to the inheritance
3385 * values on the regions in that map.
3387 * The source map must not be locked.
3390 static void vmspace_fork_normal_entry(vm_map_t old_map, vm_map_t new_map,
3391 vm_map_entry_t old_entry, int *countp);
3392 static void vmspace_fork_uksmap_entry(vm_map_t old_map, vm_map_t new_map,
3393 vm_map_entry_t old_entry, int *countp);
3396 vmspace_fork(struct vmspace *vm1)
3398 struct vmspace *vm2;
3399 vm_map_t old_map = &vm1->vm_map;
3401 vm_map_entry_t old_entry;
3404 lwkt_gettoken(&vm1->vm_map.token);
3405 vm_map_lock(old_map);
3407 vm2 = vmspace_alloc(old_map->min_offset, old_map->max_offset);
3408 lwkt_gettoken(&vm2->vm_map.token);
3409 bcopy(&vm1->vm_startcopy, &vm2->vm_startcopy,
3410 (caddr_t)&vm1->vm_endcopy - (caddr_t)&vm1->vm_startcopy);
3411 new_map = &vm2->vm_map; /* XXX */
3412 new_map->timestamp = 1;
3414 vm_map_lock(new_map);
3417 old_entry = old_map->header.next;
3418 while (old_entry != &old_map->header) {
3420 old_entry = old_entry->next;
3423 count = vm_map_entry_reserve(count + MAP_RESERVE_COUNT);
3425 old_entry = old_map->header.next;
3426 while (old_entry != &old_map->header) {
3427 switch(old_entry->maptype) {
3428 case VM_MAPTYPE_SUBMAP:
3429 panic("vm_map_fork: encountered a submap");
3431 case VM_MAPTYPE_UKSMAP:
3432 vmspace_fork_uksmap_entry(old_map, new_map,
3435 case VM_MAPTYPE_NORMAL:
3436 case VM_MAPTYPE_VPAGETABLE:
3437 vmspace_fork_normal_entry(old_map, new_map,
3441 old_entry = old_entry->next;
3444 new_map->size = old_map->size;
3445 vm_map_unlock(old_map);
3446 vm_map_unlock(new_map);
3447 vm_map_entry_release(count);
3449 lwkt_reltoken(&vm2->vm_map.token);
3450 lwkt_reltoken(&vm1->vm_map.token);
3457 vmspace_fork_normal_entry(vm_map_t old_map, vm_map_t new_map,
3458 vm_map_entry_t old_entry, int *countp)
3460 vm_map_entry_t new_entry;
3463 switch (old_entry->inheritance) {
3464 case VM_INHERIT_NONE:
3466 case VM_INHERIT_SHARE:
3468 * Clone the entry, creating the shared object if
3471 if (old_entry->object.vm_object == NULL)
3472 vm_map_entry_allocate_object(old_entry);
3474 if (old_entry->eflags & MAP_ENTRY_NEEDS_COPY) {
3476 * Shadow a map_entry which needs a copy,
3477 * replacing its object with a new object
3478 * that points to the old one. Ask the
3479 * shadow code to automatically add an
3480 * additional ref. We can't do it afterwords
3481 * because we might race a collapse. The call
3482 * to vm_map_entry_shadow() will also clear
3485 vm_map_entry_shadow(old_entry, 1);
3486 } else if (old_entry->object.vm_object) {
3488 * We will make a shared copy of the object,
3489 * and must clear OBJ_ONEMAPPING.
3491 * Optimize vnode objects. OBJ_ONEMAPPING
3492 * is non-applicable but clear it anyway,
3493 * and its terminal so we don'th ave to deal
3494 * with chains. Reduces SMP conflicts.
3496 * XXX assert that object.vm_object != NULL
3497 * since we allocate it above.
3499 object = old_entry->object.vm_object;
3500 if (object->type == OBJT_VNODE) {
3501 vm_object_reference_quick(object);
3502 vm_object_clear_flag(object,
3505 vm_object_hold(object);
3506 vm_object_chain_wait(object, 0);
3507 vm_object_reference_locked(object);
3508 vm_object_clear_flag(object,
3510 vm_object_drop(object);
3515 * Clone the entry. We've already bumped the ref on
3518 new_entry = vm_map_entry_create(new_map, countp);
3519 *new_entry = *old_entry;
3520 new_entry->eflags &= ~MAP_ENTRY_USER_WIRED;
3521 new_entry->wired_count = 0;
3524 * Insert the entry into the new map -- we know we're
3525 * inserting at the end of the new map.
3528 vm_map_entry_link(new_map, new_map->header.prev,
3532 * Update the physical map
3534 pmap_copy(new_map->pmap, old_map->pmap,
3536 (old_entry->end - old_entry->start),
3539 case VM_INHERIT_COPY:
3541 * Clone the entry and link into the map.
3543 new_entry = vm_map_entry_create(new_map, countp);
3544 *new_entry = *old_entry;
3545 new_entry->eflags &= ~MAP_ENTRY_USER_WIRED;
3546 new_entry->wired_count = 0;
3547 new_entry->object.vm_object = NULL;
3548 vm_map_entry_link(new_map, new_map->header.prev,
3550 vm_map_copy_entry(old_map, new_map, old_entry,
3557 * When forking user-kernel shared maps, the map might change in the
3558 * child so do not try to copy the underlying pmap entries.
3562 vmspace_fork_uksmap_entry(vm_map_t old_map, vm_map_t new_map,
3563 vm_map_entry_t old_entry, int *countp)
3565 vm_map_entry_t new_entry;
3567 new_entry = vm_map_entry_create(new_map, countp);
3568 *new_entry = *old_entry;
3569 new_entry->eflags &= ~MAP_ENTRY_USER_WIRED;
3570 new_entry->wired_count = 0;
3571 vm_map_entry_link(new_map, new_map->header.prev,
3576 * Create an auto-grow stack entry
3581 vm_map_stack (vm_map_t map, vm_offset_t addrbos, vm_size_t max_ssize,
3582 int flags, vm_prot_t prot, vm_prot_t max, int cow)
3584 vm_map_entry_t prev_entry;
3585 vm_map_entry_t new_stack_entry;
3586 vm_size_t init_ssize;
3589 vm_offset_t tmpaddr;
3591 cow |= MAP_IS_STACK;
3593 if (max_ssize < sgrowsiz)
3594 init_ssize = max_ssize;
3596 init_ssize = sgrowsiz;
3598 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
3602 * Find space for the mapping
3604 if ((flags & (MAP_FIXED | MAP_TRYFIXED)) == 0) {
3605 if (vm_map_findspace(map, addrbos, max_ssize, 1,
3608 vm_map_entry_release(count);
3609 return (KERN_NO_SPACE);
3614 /* If addr is already mapped, no go */
3615 if (vm_map_lookup_entry(map, addrbos, &prev_entry)) {
3617 vm_map_entry_release(count);
3618 return (KERN_NO_SPACE);
3622 /* XXX already handled by kern_mmap() */
3623 /* If we would blow our VMEM resource limit, no go */
3624 if (map->size + init_ssize >
3625 curproc->p_rlimit[RLIMIT_VMEM].rlim_cur) {
3627 vm_map_entry_release(count);
3628 return (KERN_NO_SPACE);
3633 * If we can't accomodate max_ssize in the current mapping,
3634 * no go. However, we need to be aware that subsequent user
3635 * mappings might map into the space we have reserved for
3636 * stack, and currently this space is not protected.
3638 * Hopefully we will at least detect this condition
3639 * when we try to grow the stack.
3641 if ((prev_entry->next != &map->header) &&
3642 (prev_entry->next->start < addrbos + max_ssize)) {
3644 vm_map_entry_release(count);
3645 return (KERN_NO_SPACE);
3649 * We initially map a stack of only init_ssize. We will
3650 * grow as needed later. Since this is to be a grow
3651 * down stack, we map at the top of the range.
3653 * Note: we would normally expect prot and max to be
3654 * VM_PROT_ALL, and cow to be 0. Possibly we should
3655 * eliminate these as input parameters, and just
3656 * pass these values here in the insert call.
3658 rv = vm_map_insert(map, &count, NULL, NULL,
3659 0, addrbos + max_ssize - init_ssize,
3660 addrbos + max_ssize,
3662 VM_SUBSYS_STACK, prot, max, cow);
3664 /* Now set the avail_ssize amount */
3665 if (rv == KERN_SUCCESS) {
3666 if (prev_entry != &map->header)
3667 vm_map_clip_end(map, prev_entry, addrbos + max_ssize - init_ssize, &count);
3668 new_stack_entry = prev_entry->next;
3669 if (new_stack_entry->end != addrbos + max_ssize ||
3670 new_stack_entry->start != addrbos + max_ssize - init_ssize)
3671 panic ("Bad entry start/end for new stack entry");
3673 new_stack_entry->aux.avail_ssize = max_ssize - init_ssize;
3677 vm_map_entry_release(count);
3682 * Attempts to grow a vm stack entry. Returns KERN_SUCCESS if the
3683 * desired address is already mapped, or if we successfully grow
3684 * the stack. Also returns KERN_SUCCESS if addr is outside the
3685 * stack range (this is strange, but preserves compatibility with
3686 * the grow function in vm_machdep.c).
3691 vm_map_growstack (struct proc *p, vm_offset_t addr)
3693 vm_map_entry_t prev_entry;
3694 vm_map_entry_t stack_entry;
3695 vm_map_entry_t new_stack_entry;
3696 struct vmspace *vm = p->p_vmspace;
3697 vm_map_t map = &vm->vm_map;
3700 int rv = KERN_SUCCESS;
3702 int use_read_lock = 1;
3705 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
3708 vm_map_lock_read(map);
3712 /* If addr is already in the entry range, no need to grow.*/
3713 if (vm_map_lookup_entry(map, addr, &prev_entry))
3716 if ((stack_entry = prev_entry->next) == &map->header)
3718 if (prev_entry == &map->header)
3719 end = stack_entry->start - stack_entry->aux.avail_ssize;
3721 end = prev_entry->end;
3724 * This next test mimics the old grow function in vm_machdep.c.
3725 * It really doesn't quite make sense, but we do it anyway
3726 * for compatibility.
3728 * If not growable stack, return success. This signals the
3729 * caller to proceed as he would normally with normal vm.
3731 if (stack_entry->aux.avail_ssize < 1 ||
3732 addr >= stack_entry->start ||
3733 addr < stack_entry->start - stack_entry->aux.avail_ssize) {
3737 /* Find the minimum grow amount */
3738 grow_amount = roundup (stack_entry->start - addr, PAGE_SIZE);
3739 if (grow_amount > stack_entry->aux.avail_ssize) {
3745 * If there is no longer enough space between the entries
3746 * nogo, and adjust the available space. Note: this
3747 * should only happen if the user has mapped into the
3748 * stack area after the stack was created, and is
3749 * probably an error.
3751 * This also effectively destroys any guard page the user
3752 * might have intended by limiting the stack size.
3754 if (grow_amount > stack_entry->start - end) {
3755 if (use_read_lock && vm_map_lock_upgrade(map)) {
3761 stack_entry->aux.avail_ssize = stack_entry->start - end;
3766 is_procstack = addr >= (vm_offset_t)vm->vm_maxsaddr;
3768 /* If this is the main process stack, see if we're over the
3771 if (is_procstack && (ctob(vm->vm_ssize) + grow_amount >
3772 p->p_rlimit[RLIMIT_STACK].rlim_cur)) {
3777 /* Round up the grow amount modulo SGROWSIZ */
3778 grow_amount = roundup (grow_amount, sgrowsiz);
3779 if (grow_amount > stack_entry->aux.avail_ssize) {
3780 grow_amount = stack_entry->aux.avail_ssize;
3782 if (is_procstack && (ctob(vm->vm_ssize) + grow_amount >
3783 p->p_rlimit[RLIMIT_STACK].rlim_cur)) {
3784 grow_amount = p->p_rlimit[RLIMIT_STACK].rlim_cur -
3788 /* If we would blow our VMEM resource limit, no go */
3789 if (map->size + grow_amount > p->p_rlimit[RLIMIT_VMEM].rlim_cur) {
3794 if (use_read_lock && vm_map_lock_upgrade(map)) {
3801 /* Get the preliminary new entry start value */
3802 addr = stack_entry->start - grow_amount;
3804 /* If this puts us into the previous entry, cut back our growth
3805 * to the available space. Also, see the note above.
3808 stack_entry->aux.avail_ssize = stack_entry->start - end;
3812 rv = vm_map_insert(map, &count, NULL, NULL,
3813 0, addr, stack_entry->start,
3815 VM_SUBSYS_STACK, VM_PROT_ALL, VM_PROT_ALL, 0);
3817 /* Adjust the available stack space by the amount we grew. */
3818 if (rv == KERN_SUCCESS) {
3819 if (prev_entry != &map->header)
3820 vm_map_clip_end(map, prev_entry, addr, &count);
3821 new_stack_entry = prev_entry->next;
3822 if (new_stack_entry->end != stack_entry->start ||
3823 new_stack_entry->start != addr)
3824 panic ("Bad stack grow start/end in new stack entry");
3826 new_stack_entry->aux.avail_ssize =
3827 stack_entry->aux.avail_ssize -
3828 (new_stack_entry->end - new_stack_entry->start);
3830 vm->vm_ssize += btoc(new_stack_entry->end -
3831 new_stack_entry->start);
3834 if (map->flags & MAP_WIREFUTURE)
3835 vm_map_unwire(map, new_stack_entry->start,
3836 new_stack_entry->end, FALSE);
3841 vm_map_unlock_read(map);
3844 vm_map_entry_release(count);
3849 * Unshare the specified VM space for exec. If other processes are
3850 * mapped to it, then create a new one. The new vmspace is null.
3855 vmspace_exec(struct proc *p, struct vmspace *vmcopy)
3857 struct vmspace *oldvmspace = p->p_vmspace;
3858 struct vmspace *newvmspace;
3859 vm_map_t map = &p->p_vmspace->vm_map;
3862 * If we are execing a resident vmspace we fork it, otherwise
3863 * we create a new vmspace. Note that exitingcnt is not
3864 * copied to the new vmspace.
3866 lwkt_gettoken(&oldvmspace->vm_map.token);
3868 newvmspace = vmspace_fork(vmcopy);
3869 lwkt_gettoken(&newvmspace->vm_map.token);
3871 newvmspace = vmspace_alloc(map->min_offset, map->max_offset);
3872 lwkt_gettoken(&newvmspace->vm_map.token);
3873 bcopy(&oldvmspace->vm_startcopy, &newvmspace->vm_startcopy,
3874 (caddr_t)&oldvmspace->vm_endcopy -
3875 (caddr_t)&oldvmspace->vm_startcopy);
3879 * Finish initializing the vmspace before assigning it
3880 * to the process. The vmspace will become the current vmspace
3883 pmap_pinit2(vmspace_pmap(newvmspace));
3884 pmap_replacevm(p, newvmspace, 0);
3885 lwkt_reltoken(&newvmspace->vm_map.token);
3886 lwkt_reltoken(&oldvmspace->vm_map.token);
3887 vmspace_rel(oldvmspace);
3891 * Unshare the specified VM space for forcing COW. This
3892 * is called by rfork, for the (RFMEM|RFPROC) == 0 case.
3895 vmspace_unshare(struct proc *p)
3897 struct vmspace *oldvmspace = p->p_vmspace;
3898 struct vmspace *newvmspace;
3900 lwkt_gettoken(&oldvmspace->vm_map.token);
3901 if (vmspace_getrefs(oldvmspace) == 1) {
3902 lwkt_reltoken(&oldvmspace->vm_map.token);
3905 newvmspace = vmspace_fork(oldvmspace);
3906 lwkt_gettoken(&newvmspace->vm_map.token);
3907 pmap_pinit2(vmspace_pmap(newvmspace));
3908 pmap_replacevm(p, newvmspace, 0);
3909 lwkt_reltoken(&newvmspace->vm_map.token);
3910 lwkt_reltoken(&oldvmspace->vm_map.token);
3911 vmspace_rel(oldvmspace);
3915 * vm_map_hint: return the beginning of the best area suitable for
3916 * creating a new mapping with "prot" protection.
3921 vm_map_hint(struct proc *p, vm_offset_t addr, vm_prot_t prot)
3923 struct vmspace *vms = p->p_vmspace;
3925 if (!randomize_mmap || addr != 0) {
3927 * Set a reasonable start point for the hint if it was
3928 * not specified or if it falls within the heap space.
3929 * Hinted mmap()s do not allocate out of the heap space.
3932 (addr >= round_page((vm_offset_t)vms->vm_taddr) &&
3933 addr < round_page((vm_offset_t)vms->vm_daddr + maxdsiz))) {
3934 addr = round_page((vm_offset_t)vms->vm_daddr + maxdsiz);
3939 addr = (vm_offset_t)vms->vm_daddr + MAXDSIZ;
3940 addr += karc4random() & (MIN((256 * 1024 * 1024), MAXDSIZ) - 1);
3942 return (round_page(addr));
3946 * Finds the VM object, offset, and protection for a given virtual address
3947 * in the specified map, assuming a page fault of the type specified.
3949 * Leaves the map in question locked for read; return values are guaranteed
3950 * until a vm_map_lookup_done call is performed. Note that the map argument
3951 * is in/out; the returned map must be used in the call to vm_map_lookup_done.
3953 * A handle (out_entry) is returned for use in vm_map_lookup_done, to make
3956 * If a lookup is requested with "write protection" specified, the map may
3957 * be changed to perform virtual copying operations, although the data
3958 * referenced will remain the same.
3963 vm_map_lookup(vm_map_t *var_map, /* IN/OUT */
3965 vm_prot_t fault_typea,
3966 vm_map_entry_t *out_entry, /* OUT */
3967 vm_object_t *object, /* OUT */
3968 vm_pindex_t *pindex, /* OUT */
3969 vm_prot_t *out_prot, /* OUT */
3970 boolean_t *wired) /* OUT */
3972 vm_map_entry_t entry;
3973 vm_map_t map = *var_map;
3975 vm_prot_t fault_type = fault_typea;
3976 int use_read_lock = 1;
3977 int rv = KERN_SUCCESS;
3981 vm_map_lock_read(map);
3986 * If the map has an interesting hint, try it before calling full
3987 * blown lookup routine.
3994 if ((entry == &map->header) ||
3995 (vaddr < entry->start) || (vaddr >= entry->end)) {
3996 vm_map_entry_t tmp_entry;
3999 * Entry was either not a valid hint, or the vaddr was not
4000 * contained in the entry, so do a full lookup.
4002 if (!vm_map_lookup_entry(map, vaddr, &tmp_entry)) {
4003 rv = KERN_INVALID_ADDRESS;
4014 if (entry->maptype == VM_MAPTYPE_SUBMAP) {
4015 vm_map_t old_map = map;
4017 *var_map = map = entry->object.sub_map;
4019 vm_map_unlock_read(old_map);
4021 vm_map_unlock(old_map);
4027 * Check whether this task is allowed to have this page.
4028 * Note the special case for MAP_ENTRY_COW
4029 * pages with an override. This is to implement a forced
4030 * COW for debuggers.
4033 if (fault_type & VM_PROT_OVERRIDE_WRITE)
4034 prot = entry->max_protection;
4036 prot = entry->protection;
4038 fault_type &= (VM_PROT_READ|VM_PROT_WRITE|VM_PROT_EXECUTE);
4039 if ((fault_type & prot) != fault_type) {
4040 rv = KERN_PROTECTION_FAILURE;
4044 if ((entry->eflags & MAP_ENTRY_USER_WIRED) &&
4045 (entry->eflags & MAP_ENTRY_COW) &&
4046 (fault_type & VM_PROT_WRITE) &&
4047 (fault_typea & VM_PROT_OVERRIDE_WRITE) == 0) {
4048 rv = KERN_PROTECTION_FAILURE;
4053 * If this page is not pageable, we have to get it for all possible
4056 *wired = (entry->wired_count != 0);
4058 prot = fault_type = entry->protection;
4061 * Virtual page tables may need to update the accessed (A) bit
4062 * in a page table entry. Upgrade the fault to a write fault for
4063 * that case if the map will support it. If the map does not support
4064 * it the page table entry simply will not be updated.
4066 if (entry->maptype == VM_MAPTYPE_VPAGETABLE) {
4067 if (prot & VM_PROT_WRITE)
4068 fault_type |= VM_PROT_WRITE;
4071 if (curthread->td_lwp && curthread->td_lwp->lwp_vmspace &&
4072 pmap_emulate_ad_bits(&curthread->td_lwp->lwp_vmspace->vm_pmap)) {
4073 if ((prot & VM_PROT_WRITE) == 0)
4074 fault_type |= VM_PROT_WRITE;
4078 * Only NORMAL and VPAGETABLE maps are object-based. UKSMAPs are not.
4080 if (entry->maptype != VM_MAPTYPE_NORMAL &&
4081 entry->maptype != VM_MAPTYPE_VPAGETABLE) {
4087 * If the entry was copy-on-write, we either ...
4089 if (entry->eflags & MAP_ENTRY_NEEDS_COPY) {
4091 * If we want to write the page, we may as well handle that
4092 * now since we've got the map locked.
4094 * If we don't need to write the page, we just demote the
4095 * permissions allowed.
4098 if (fault_type & VM_PROT_WRITE) {
4100 * Not allowed if TDF_NOFAULT is set as the shadowing
4101 * operation can deadlock against the faulting
4102 * function due to the copy-on-write.
4104 if (curthread->td_flags & TDF_NOFAULT) {
4105 rv = KERN_FAILURE_NOFAULT;
4110 * Make a new object, and place it in the object
4111 * chain. Note that no new references have appeared
4112 * -- one just moved from the map to the new
4116 if (use_read_lock && vm_map_lock_upgrade(map)) {
4123 vm_map_entry_shadow(entry, 0);
4126 * We're attempting to read a copy-on-write page --
4127 * don't allow writes.
4130 prot &= ~VM_PROT_WRITE;
4135 * Create an object if necessary.
4137 if (entry->object.vm_object == NULL && !map->system_map) {
4138 if (use_read_lock && vm_map_lock_upgrade(map)) {
4144 vm_map_entry_allocate_object(entry);
4148 * Return the object/offset from this entry. If the entry was
4149 * copy-on-write or empty, it has been fixed up.
4151 *object = entry->object.vm_object;
4154 *pindex = OFF_TO_IDX((vaddr - entry->start) + entry->offset);
4157 * Return whether this is the only map sharing this data. On
4158 * success we return with a read lock held on the map. On failure
4159 * we return with the map unlocked.
4163 if (rv == KERN_SUCCESS) {
4164 if (use_read_lock == 0)
4165 vm_map_lock_downgrade(map);
4166 } else if (use_read_lock) {
4167 vm_map_unlock_read(map);
4175 * Releases locks acquired by a vm_map_lookup()
4176 * (according to the handle returned by that lookup).
4178 * No other requirements.
4181 vm_map_lookup_done(vm_map_t map, vm_map_entry_t entry, int count)
4184 * Unlock the main-level map
4186 vm_map_unlock_read(map);
4188 vm_map_entry_release(count);
4191 #include "opt_ddb.h"
4193 #include <sys/kernel.h>
4195 #include <ddb/ddb.h>
4200 DB_SHOW_COMMAND(map, vm_map_print)
4203 /* XXX convert args. */
4204 vm_map_t map = (vm_map_t)addr;
4205 boolean_t full = have_addr;
4207 vm_map_entry_t entry;
4209 db_iprintf("Task map %p: pmap=%p, nentries=%d, version=%u\n",
4211 (void *)map->pmap, map->nentries, map->timestamp);
4214 if (!full && db_indent)
4218 for (entry = map->header.next; entry != &map->header;
4219 entry = entry->next) {
4220 db_iprintf("map entry %p: start=%p, end=%p\n",
4221 (void *)entry, (void *)entry->start, (void *)entry->end);
4224 static char *inheritance_name[4] =
4225 {"share", "copy", "none", "donate_copy"};
4227 db_iprintf(" prot=%x/%x/%s",
4229 entry->max_protection,
4230 inheritance_name[(int)(unsigned char)entry->inheritance]);
4231 if (entry->wired_count != 0)
4232 db_printf(", wired");
4234 switch(entry->maptype) {
4235 case VM_MAPTYPE_SUBMAP:
4236 /* XXX no %qd in kernel. Truncate entry->offset. */
4237 db_printf(", share=%p, offset=0x%lx\n",
4238 (void *)entry->object.sub_map,
4239 (long)entry->offset);
4241 if ((entry->prev == &map->header) ||
4242 (entry->prev->object.sub_map !=
4243 entry->object.sub_map)) {
4245 vm_map_print((db_expr_t)(intptr_t)
4246 entry->object.sub_map,
4251 case VM_MAPTYPE_NORMAL:
4252 case VM_MAPTYPE_VPAGETABLE:
4253 /* XXX no %qd in kernel. Truncate entry->offset. */
4254 db_printf(", object=%p, offset=0x%lx",
4255 (void *)entry->object.vm_object,
4256 (long)entry->offset);
4257 if (entry->eflags & MAP_ENTRY_COW)
4258 db_printf(", copy (%s)",
4259 (entry->eflags & MAP_ENTRY_NEEDS_COPY) ? "needed" : "done");
4263 if ((entry->prev == &map->header) ||
4264 (entry->prev->object.vm_object !=
4265 entry->object.vm_object)) {
4267 vm_object_print((db_expr_t)(intptr_t)
4268 entry->object.vm_object,
4274 case VM_MAPTYPE_UKSMAP:
4275 db_printf(", uksmap=%p, offset=0x%lx",
4276 (void *)entry->object.uksmap,
4277 (long)entry->offset);
4278 if (entry->eflags & MAP_ENTRY_COW)
4279 db_printf(", copy (%s)",
4280 (entry->eflags & MAP_ENTRY_NEEDS_COPY) ? "needed" : "done");
4296 DB_SHOW_COMMAND(procvm, procvm)
4301 p = (struct proc *) addr;
4306 db_printf("p = %p, vmspace = %p, map = %p, pmap = %p\n",
4307 (void *)p, (void *)p->p_vmspace, (void *)&p->p_vmspace->vm_map,
4308 (void *)vmspace_pmap(p->p_vmspace));
4310 vm_map_print((db_expr_t)(intptr_t)&p->p_vmspace->vm_map, 1, 0, NULL);