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 vmspace_drop_notoken(struct vmspace *vm);
147 static void vm_map_entry_shadow(vm_map_entry_t entry, int addref);
148 static vm_map_entry_t vm_map_entry_create(vm_map_t map, int *);
149 static void vm_map_entry_dispose (vm_map_t map, vm_map_entry_t entry, int *);
150 static void _vm_map_clip_end (vm_map_t, vm_map_entry_t, vm_offset_t, int *);
151 static void _vm_map_clip_start (vm_map_t, vm_map_entry_t, vm_offset_t, int *);
152 static void vm_map_entry_delete (vm_map_t, vm_map_entry_t, int *);
153 static void vm_map_entry_unwire (vm_map_t, vm_map_entry_t);
154 static void vm_map_copy_entry (vm_map_t, vm_map_t, vm_map_entry_t,
156 static void vm_map_unclip_range (vm_map_t map, vm_map_entry_t start_entry, vm_offset_t start, vm_offset_t end, int *count, int flags);
159 * Initialize the vm_map module. Must be called before any other vm_map
162 * Map and entry structures are allocated from the general purpose
163 * memory pool with some exceptions:
165 * - The kernel map is allocated statically.
166 * - Initial kernel map entries are allocated out of a static pool.
167 * - We must set ZONE_SPECIAL here or the early boot code can get
168 * stuck if there are >63 cores.
170 * These restrictions are necessary since malloc() uses the
171 * maps and requires map entries.
173 * Called from the low level boot code only.
178 mapentzone = &mapentzone_store;
179 zbootinit(mapentzone, "MAP ENTRY", sizeof (struct vm_map_entry),
180 map_entry_init, MAX_MAPENT);
181 mapentzone_store.zflags |= ZONE_SPECIAL;
185 * Called prior to any vmspace allocations.
187 * Called from the low level boot code only.
192 vmspace_cache = objcache_create_mbacked(M_VMSPACE,
193 sizeof(struct vmspace),
195 vmspace_ctor, vmspace_dtor,
197 zinitna(mapentzone, &mapentobj, NULL, 0, 0,
198 ZONE_USE_RESERVE | ZONE_SPECIAL);
204 * objcache support. We leave the pmap root cached as long as possible
205 * for performance reasons.
209 vmspace_ctor(void *obj, void *privdata, int ocflags)
211 struct vmspace *vm = obj;
213 bzero(vm, sizeof(*vm));
214 vm->vm_refcnt = VM_REF_DELETED;
221 vmspace_dtor(void *obj, void *privdata)
223 struct vmspace *vm = obj;
225 KKASSERT(vm->vm_refcnt == VM_REF_DELETED);
226 pmap_puninit(vmspace_pmap(vm));
230 * Red black tree functions
232 * The caller must hold the related map lock.
234 static int rb_vm_map_compare(vm_map_entry_t a, vm_map_entry_t b);
235 RB_GENERATE(vm_map_rb_tree, vm_map_entry, rb_entry, rb_vm_map_compare);
237 /* a->start is address, and the only field has to be initialized */
239 rb_vm_map_compare(vm_map_entry_t a, vm_map_entry_t b)
241 if (a->start < b->start)
243 else if (a->start > b->start)
249 * Initialize vmspace ref/hold counts vmspace0. There is a holdcnt for
253 vmspace_initrefs(struct vmspace *vm)
260 * Allocate a vmspace structure, including a vm_map and pmap.
261 * Initialize numerous fields. While the initial allocation is zerod,
262 * subsequence reuse from the objcache leaves elements of the structure
263 * intact (particularly the pmap), so portions must be zerod.
265 * Returns a referenced vmspace.
270 vmspace_alloc(vm_offset_t min, vm_offset_t max)
274 vm = objcache_get(vmspace_cache, M_WAITOK);
276 bzero(&vm->vm_startcopy,
277 (char *)&vm->vm_endcopy - (char *)&vm->vm_startcopy);
278 vm_map_init(&vm->vm_map, min, max, NULL); /* initializes token */
281 * NOTE: hold to acquires token for safety.
283 * On return vmspace is referenced (refs=1, hold=1). That is,
284 * each refcnt also has a holdcnt. There can be additional holds
285 * (holdcnt) above and beyond the refcnt. Finalization is handled in
286 * two stages, one on refs 1->0, and the the second on hold 1->0.
288 KKASSERT(vm->vm_holdcnt == 0);
289 KKASSERT(vm->vm_refcnt == VM_REF_DELETED);
290 vmspace_initrefs(vm);
292 pmap_pinit(vmspace_pmap(vm)); /* (some fields reused) */
293 vm->vm_map.pmap = vmspace_pmap(vm); /* XXX */
296 cpu_vmspace_alloc(vm);
303 * NOTE: Can return 0 if the vmspace is exiting.
306 vmspace_getrefs(struct vmspace *vm)
308 return ((int)(vm->vm_refcnt & ~VM_REF_DELETED));
312 vmspace_hold(struct vmspace *vm)
314 atomic_add_int(&vm->vm_holdcnt, 1);
315 lwkt_gettoken(&vm->vm_map.token);
319 vmspace_drop(struct vmspace *vm)
321 lwkt_reltoken(&vm->vm_map.token);
322 vmspace_drop_notoken(vm);
326 vmspace_drop_notoken(struct vmspace *vm)
328 if (atomic_fetchadd_int(&vm->vm_holdcnt, -1) == 1) {
329 if (vm->vm_refcnt & VM_REF_DELETED)
330 vmspace_terminate(vm, 1);
335 * A vmspace object must not be in a terminated state to be able to obtain
336 * additional refs on it.
338 * These are official references to the vmspace, the count is used to check
339 * for vmspace sharing. Foreign accessors should use 'hold' and not 'ref'.
341 * XXX we need to combine hold & ref together into one 64-bit field to allow
342 * holds to prevent stage-1 termination.
345 vmspace_ref(struct vmspace *vm)
349 n = atomic_fetchadd_int(&vm->vm_refcnt, 1);
350 KKASSERT((n & VM_REF_DELETED) == 0);
354 * Release a ref on the vmspace. On the 1->0 transition we do stage-1
355 * termination of the vmspace. Then, on the final drop of the hold we
356 * will do stage-2 final termination.
359 vmspace_rel(struct vmspace *vm)
366 KKASSERT((int)n > 0); /* at least one ref & not deleted */
370 * We must have a hold first to interlock the
371 * VM_REF_DELETED check that the drop tests.
373 atomic_add_int(&vm->vm_holdcnt, 1);
374 if (atomic_cmpset_int(&vm->vm_refcnt, n,
376 vmspace_terminate(vm, 0);
377 vmspace_drop_notoken(vm);
380 vmspace_drop_notoken(vm);
381 } else if (atomic_cmpset_int(&vm->vm_refcnt, n, n - 1)) {
388 * This is called during exit indicating that the vmspace is no
389 * longer in used by an exiting process, but the process has not yet
392 * We drop refs, allowing for stage-1 termination, but maintain a holdcnt
393 * to prevent stage-2 until the process is reaped. Note hte order of
394 * operation, we must hold first.
399 vmspace_relexit(struct vmspace *vm)
401 atomic_add_int(&vm->vm_holdcnt, 1);
406 * Called during reap to disconnect the remainder of the vmspace from
407 * the process. On the hold drop the vmspace termination is finalized.
412 vmspace_exitfree(struct proc *p)
418 vmspace_drop_notoken(vm);
422 * Called in two cases:
424 * (1) When the last refcnt is dropped and the vmspace becomes inactive,
425 * called with final == 0. refcnt will be (u_int)-1 at this point,
426 * and holdcnt will still be non-zero.
428 * (2) When holdcnt becomes 0, called with final == 1. There should no
429 * longer be anyone with access to the vmspace.
431 * VMSPACE_EXIT1 flags the primary deactivation
432 * VMSPACE_EXIT2 flags the last reap
435 vmspace_terminate(struct vmspace *vm, int final)
439 lwkt_gettoken(&vm->vm_map.token);
441 KKASSERT((vm->vm_flags & VMSPACE_EXIT1) == 0);
442 vm->vm_flags |= VMSPACE_EXIT1;
445 * Get rid of most of the resources. Leave the kernel pmap
448 * If the pmap does not contain wired pages we can bulk-delete
449 * the pmap as a performance optimization before removing the
452 * If the pmap contains wired pages we cannot do this
453 * pre-optimization because currently vm_fault_unwire()
454 * expects the pmap pages to exist and will not decrement
455 * p->wire_count if they do not.
458 if (vmspace_pmap(vm)->pm_stats.wired_count) {
459 vm_map_remove(&vm->vm_map, VM_MIN_USER_ADDRESS,
460 VM_MAX_USER_ADDRESS);
461 pmap_remove_pages(vmspace_pmap(vm), VM_MIN_USER_ADDRESS,
462 VM_MAX_USER_ADDRESS);
464 pmap_remove_pages(vmspace_pmap(vm), VM_MIN_USER_ADDRESS,
465 VM_MAX_USER_ADDRESS);
466 vm_map_remove(&vm->vm_map, VM_MIN_USER_ADDRESS,
467 VM_MAX_USER_ADDRESS);
469 lwkt_reltoken(&vm->vm_map.token);
471 KKASSERT((vm->vm_flags & VMSPACE_EXIT1) != 0);
472 KKASSERT((vm->vm_flags & VMSPACE_EXIT2) == 0);
475 * Get rid of remaining basic resources.
477 vm->vm_flags |= VMSPACE_EXIT2;
480 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
481 vm_map_lock(&vm->vm_map);
482 cpu_vmspace_free(vm);
485 * Lock the map, to wait out all other references to it.
486 * Delete all of the mappings and pages they hold, then call
487 * the pmap module to reclaim anything left.
489 vm_map_delete(&vm->vm_map, vm->vm_map.min_offset,
490 vm->vm_map.max_offset, &count);
491 vm_map_unlock(&vm->vm_map);
492 vm_map_entry_release(count);
494 pmap_release(vmspace_pmap(vm));
495 lwkt_reltoken(&vm->vm_map.token);
496 objcache_put(vmspace_cache, vm);
501 * Swap useage is determined by taking the proportional swap used by
502 * VM objects backing the VM map. To make up for fractional losses,
503 * if the VM object has any swap use at all the associated map entries
504 * count for at least 1 swap page.
509 vmspace_swap_count(struct vmspace *vm)
511 vm_map_t map = &vm->vm_map;
514 vm_offset_t count = 0;
518 for (cur = map->header.next; cur != &map->header; cur = cur->next) {
519 switch(cur->maptype) {
520 case VM_MAPTYPE_NORMAL:
521 case VM_MAPTYPE_VPAGETABLE:
522 if ((object = cur->object.vm_object) == NULL)
524 if (object->swblock_count) {
525 n = (cur->end - cur->start) / PAGE_SIZE;
526 count += object->swblock_count *
527 SWAP_META_PAGES * n / object->size + 1;
540 * Calculate the approximate number of anonymous pages in use by
541 * this vmspace. To make up for fractional losses, we count each
542 * VM object as having at least 1 anonymous page.
547 vmspace_anonymous_count(struct vmspace *vm)
549 vm_map_t map = &vm->vm_map;
552 vm_offset_t count = 0;
555 for (cur = map->header.next; cur != &map->header; cur = cur->next) {
556 switch(cur->maptype) {
557 case VM_MAPTYPE_NORMAL:
558 case VM_MAPTYPE_VPAGETABLE:
559 if ((object = cur->object.vm_object) == NULL)
561 if (object->type != OBJT_DEFAULT &&
562 object->type != OBJT_SWAP) {
565 count += object->resident_page_count;
577 * Initialize an existing vm_map structure such as that in the vmspace
578 * structure. The pmap is initialized elsewhere.
583 vm_map_init(struct vm_map *map, vm_offset_t min, vm_offset_t max, pmap_t pmap)
585 map->header.next = map->header.prev = &map->header;
586 RB_INIT(&map->rb_root);
590 map->min_offset = min;
591 map->max_offset = max;
593 map->first_free = &map->header;
594 map->hint = &map->header;
597 lwkt_token_init(&map->token, "vm_map");
598 lockinit(&map->lock, "vm_maplk", (hz + 9) / 10, 0);
602 * Shadow the vm_map_entry's object. This typically needs to be done when
603 * a write fault is taken on an entry which had previously been cloned by
604 * fork(). The shared object (which might be NULL) must become private so
605 * we add a shadow layer above it.
607 * Object allocation for anonymous mappings is defered as long as possible.
608 * When creating a shadow, however, the underlying object must be instantiated
609 * so it can be shared.
611 * If the map segment is governed by a virtual page table then it is
612 * possible to address offsets beyond the mapped area. Just allocate
613 * a maximally sized object for this case.
615 * If addref is non-zero an additional reference is added to the returned
616 * entry. This mechanic exists because the additional reference might have
617 * to be added atomically and not after return to prevent a premature
620 * The vm_map must be exclusively locked.
621 * No other requirements.
625 vm_map_entry_shadow(vm_map_entry_t entry, int addref)
627 if (entry->maptype == VM_MAPTYPE_VPAGETABLE) {
628 vm_object_shadow(&entry->object.vm_object, &entry->offset,
629 0x7FFFFFFF, addref); /* XXX */
631 vm_object_shadow(&entry->object.vm_object, &entry->offset,
632 atop(entry->end - entry->start), addref);
634 entry->eflags &= ~MAP_ENTRY_NEEDS_COPY;
638 * Allocate an object for a vm_map_entry.
640 * Object allocation for anonymous mappings is defered as long as possible.
641 * This function is called when we can defer no longer, generally when a map
642 * entry might be split or forked or takes a page fault.
644 * If the map segment is governed by a virtual page table then it is
645 * possible to address offsets beyond the mapped area. Just allocate
646 * a maximally sized object for this case.
648 * The vm_map must be exclusively locked.
649 * No other requirements.
652 vm_map_entry_allocate_object(vm_map_entry_t entry)
656 if (entry->maptype == VM_MAPTYPE_VPAGETABLE) {
657 obj = vm_object_allocate(OBJT_DEFAULT, 0x7FFFFFFF); /* XXX */
659 obj = vm_object_allocate(OBJT_DEFAULT,
660 atop(entry->end - entry->start));
662 entry->object.vm_object = obj;
667 * Set an initial negative count so the first attempt to reserve
668 * space preloads a bunch of vm_map_entry's for this cpu. Also
669 * pre-allocate 2 vm_map_entries which will be needed by zalloc() to
670 * map a new page for vm_map_entry structures. SMP systems are
671 * particularly sensitive.
673 * This routine is called in early boot so we cannot just call
674 * vm_map_entry_reserve().
676 * Called from the low level boot code only (for each cpu)
678 * WARNING! Take care not to have too-big a static/BSS structure here
679 * as MAXCPU can be 256+, otherwise the loader's 64MB heap
680 * can get blown out by the kernel plus the initrd image.
683 vm_map_entry_reserve_cpu_init(globaldata_t gd)
685 vm_map_entry_t entry;
689 gd->gd_vme_avail -= MAP_RESERVE_COUNT * 2;
690 if (gd->gd_cpuid == 0) {
691 entry = &cpu_map_entry_init_bsp[0];
692 count = MAPENTRYBSP_CACHE;
694 entry = &cpu_map_entry_init_ap[gd->gd_cpuid][0];
695 count = MAPENTRYAP_CACHE;
697 for (i = 0; i < count; ++i, ++entry) {
698 entry->next = gd->gd_vme_base;
699 gd->gd_vme_base = entry;
704 * Reserves vm_map_entry structures so code later on can manipulate
705 * map_entry structures within a locked map without blocking trying
706 * to allocate a new vm_map_entry.
711 vm_map_entry_reserve(int count)
713 struct globaldata *gd = mycpu;
714 vm_map_entry_t entry;
717 * Make sure we have enough structures in gd_vme_base to handle
718 * the reservation request.
720 * The critical section protects access to the per-cpu gd.
723 while (gd->gd_vme_avail < count) {
724 entry = zalloc(mapentzone);
725 entry->next = gd->gd_vme_base;
726 gd->gd_vme_base = entry;
729 gd->gd_vme_avail -= count;
736 * Releases previously reserved vm_map_entry structures that were not
737 * used. If we have too much junk in our per-cpu cache clean some of
743 vm_map_entry_release(int count)
745 struct globaldata *gd = mycpu;
746 vm_map_entry_t entry;
749 gd->gd_vme_avail += count;
750 while (gd->gd_vme_avail > MAP_RESERVE_SLOP) {
751 entry = gd->gd_vme_base;
752 KKASSERT(entry != NULL);
753 gd->gd_vme_base = entry->next;
756 zfree(mapentzone, entry);
763 * Reserve map entry structures for use in kernel_map itself. These
764 * entries have *ALREADY* been reserved on a per-cpu basis when the map
765 * was inited. This function is used by zalloc() to avoid a recursion
766 * when zalloc() itself needs to allocate additional kernel memory.
768 * This function works like the normal reserve but does not load the
769 * vm_map_entry cache (because that would result in an infinite
770 * recursion). Note that gd_vme_avail may go negative. This is expected.
772 * Any caller of this function must be sure to renormalize after
773 * potentially eating entries to ensure that the reserve supply
779 vm_map_entry_kreserve(int count)
781 struct globaldata *gd = mycpu;
784 gd->gd_vme_avail -= count;
786 KASSERT(gd->gd_vme_base != NULL,
787 ("no reserved entries left, gd_vme_avail = %d",
793 * Release previously reserved map entries for kernel_map. We do not
794 * attempt to clean up like the normal release function as this would
795 * cause an unnecessary (but probably not fatal) deep procedure call.
800 vm_map_entry_krelease(int count)
802 struct globaldata *gd = mycpu;
805 gd->gd_vme_avail += count;
810 * Allocates a VM map entry for insertion. No entry fields are filled in.
812 * The entries should have previously been reserved. The reservation count
813 * is tracked in (*countp).
817 static vm_map_entry_t
818 vm_map_entry_create(vm_map_t map, int *countp)
820 struct globaldata *gd = mycpu;
821 vm_map_entry_t entry;
823 KKASSERT(*countp > 0);
826 entry = gd->gd_vme_base;
827 KASSERT(entry != NULL, ("gd_vme_base NULL! count %d", *countp));
828 gd->gd_vme_base = entry->next;
835 * Dispose of a vm_map_entry that is no longer being referenced.
840 vm_map_entry_dispose(vm_map_t map, vm_map_entry_t entry, int *countp)
842 struct globaldata *gd = mycpu;
844 KKASSERT(map->hint != entry);
845 KKASSERT(map->first_free != entry);
849 entry->next = gd->gd_vme_base;
850 gd->gd_vme_base = entry;
856 * Insert/remove entries from maps.
858 * The related map must be exclusively locked.
859 * The caller must hold map->token
860 * No other requirements.
863 vm_map_entry_link(vm_map_t map,
864 vm_map_entry_t after_where,
865 vm_map_entry_t entry)
867 ASSERT_VM_MAP_LOCKED(map);
870 entry->prev = after_where;
871 entry->next = after_where->next;
872 entry->next->prev = entry;
873 after_where->next = entry;
874 if (vm_map_rb_tree_RB_INSERT(&map->rb_root, entry))
875 panic("vm_map_entry_link: dup addr map %p ent %p", map, entry);
879 vm_map_entry_unlink(vm_map_t map,
880 vm_map_entry_t entry)
885 ASSERT_VM_MAP_LOCKED(map);
887 if (entry->eflags & MAP_ENTRY_IN_TRANSITION) {
888 panic("vm_map_entry_unlink: attempt to mess with "
889 "locked entry! %p", entry);
895 vm_map_rb_tree_RB_REMOVE(&map->rb_root, entry);
900 * Finds the map entry containing (or immediately preceding) the specified
901 * address in the given map. The entry is returned in (*entry).
903 * The boolean result indicates whether the address is actually contained
906 * The related map must be locked.
907 * No other requirements.
910 vm_map_lookup_entry(vm_map_t map, vm_offset_t address, vm_map_entry_t *entry)
915 ASSERT_VM_MAP_LOCKED(map);
918 * XXX TEMPORARILY DISABLED. For some reason our attempt to revive
919 * the hint code with the red-black lookup meets with system crashes
920 * and lockups. We do not yet know why.
922 * It is possible that the problem is related to the setting
923 * of the hint during map_entry deletion, in the code specified
924 * at the GGG comment later on in this file.
926 * YYY More likely it's because this function can be called with
927 * a shared lock on the map, resulting in map->hint updates possibly
928 * racing. Fixed now but untested.
931 * Quickly check the cached hint, there's a good chance of a match.
935 if (tmp != &map->header) {
936 if (address >= tmp->start && address < tmp->end) {
944 * Locate the record from the top of the tree. 'last' tracks the
945 * closest prior record and is returned if no match is found, which
946 * in binary tree terms means tracking the most recent right-branch
947 * taken. If there is no prior record, &map->header is returned.
950 tmp = RB_ROOT(&map->rb_root);
953 if (address >= tmp->start) {
954 if (address < tmp->end) {
960 tmp = RB_RIGHT(tmp, rb_entry);
962 tmp = RB_LEFT(tmp, rb_entry);
970 * Inserts the given whole VM object into the target map at the specified
971 * address range. The object's size should match that of the address range.
973 * The map must be exclusively locked.
974 * The object must be held.
975 * The caller must have reserved sufficient vm_map_entry structures.
977 * If object is non-NULL, ref count must be bumped by caller prior to
978 * making call to account for the new entry.
981 vm_map_insert(vm_map_t map, int *countp, void *map_object, void *map_aux,
982 vm_ooffset_t offset, vm_offset_t start, vm_offset_t end,
983 vm_maptype_t maptype, vm_subsys_t id,
984 vm_prot_t prot, vm_prot_t max, int cow)
986 vm_map_entry_t new_entry;
987 vm_map_entry_t prev_entry;
988 vm_map_entry_t temp_entry;
989 vm_eflags_t protoeflags;
993 if (maptype == VM_MAPTYPE_UKSMAP)
998 ASSERT_VM_MAP_LOCKED(map);
1000 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
1003 * Check that the start and end points are not bogus.
1005 if ((start < map->min_offset) || (end > map->max_offset) ||
1007 return (KERN_INVALID_ADDRESS);
1010 * Find the entry prior to the proposed starting address; if it's part
1011 * of an existing entry, this range is bogus.
1013 if (vm_map_lookup_entry(map, start, &temp_entry))
1014 return (KERN_NO_SPACE);
1016 prev_entry = temp_entry;
1019 * Assert that the next entry doesn't overlap the end point.
1022 if ((prev_entry->next != &map->header) &&
1023 (prev_entry->next->start < end))
1024 return (KERN_NO_SPACE);
1028 if (cow & MAP_COPY_ON_WRITE)
1029 protoeflags |= MAP_ENTRY_COW|MAP_ENTRY_NEEDS_COPY;
1031 if (cow & MAP_NOFAULT) {
1032 protoeflags |= MAP_ENTRY_NOFAULT;
1034 KASSERT(object == NULL,
1035 ("vm_map_insert: paradoxical MAP_NOFAULT request"));
1037 if (cow & MAP_DISABLE_SYNCER)
1038 protoeflags |= MAP_ENTRY_NOSYNC;
1039 if (cow & MAP_DISABLE_COREDUMP)
1040 protoeflags |= MAP_ENTRY_NOCOREDUMP;
1041 if (cow & MAP_IS_STACK)
1042 protoeflags |= MAP_ENTRY_STACK;
1043 if (cow & MAP_IS_KSTACK)
1044 protoeflags |= MAP_ENTRY_KSTACK;
1046 lwkt_gettoken(&map->token);
1050 * When object is non-NULL, it could be shared with another
1051 * process. We have to set or clear OBJ_ONEMAPPING
1054 * NOTE: This flag is only applicable to DEFAULT and SWAP
1055 * objects and will already be clear in other types
1056 * of objects, so a shared object lock is ok for
1059 if ((object->ref_count > 1) || (object->shadow_count != 0)) {
1060 vm_object_clear_flag(object, OBJ_ONEMAPPING);
1063 else if ((prev_entry != &map->header) &&
1064 (prev_entry->eflags == protoeflags) &&
1065 (prev_entry->end == start) &&
1066 (prev_entry->wired_count == 0) &&
1067 (prev_entry->id == id) &&
1068 prev_entry->maptype == maptype &&
1069 maptype == VM_MAPTYPE_NORMAL &&
1070 ((prev_entry->object.vm_object == NULL) ||
1071 vm_object_coalesce(prev_entry->object.vm_object,
1072 OFF_TO_IDX(prev_entry->offset),
1073 (vm_size_t)(prev_entry->end - prev_entry->start),
1074 (vm_size_t)(end - prev_entry->end)))) {
1076 * We were able to extend the object. Determine if we
1077 * can extend the previous map entry to include the
1078 * new range as well.
1080 if ((prev_entry->inheritance == VM_INHERIT_DEFAULT) &&
1081 (prev_entry->protection == prot) &&
1082 (prev_entry->max_protection == max)) {
1083 map->size += (end - prev_entry->end);
1084 prev_entry->end = end;
1085 vm_map_simplify_entry(map, prev_entry, countp);
1086 lwkt_reltoken(&map->token);
1087 return (KERN_SUCCESS);
1091 * If we can extend the object but cannot extend the
1092 * map entry, we have to create a new map entry. We
1093 * must bump the ref count on the extended object to
1094 * account for it. object may be NULL.
1096 * XXX if object is NULL should we set offset to 0 here ?
1098 object = prev_entry->object.vm_object;
1099 offset = prev_entry->offset +
1100 (prev_entry->end - prev_entry->start);
1102 vm_object_hold(object);
1103 vm_object_chain_wait(object, 0);
1104 vm_object_reference_locked(object);
1106 map_object = object;
1111 * NOTE: if conditionals fail, object can be NULL here. This occurs
1112 * in things like the buffer map where we manage kva but do not manage
1117 * Create a new entry
1120 new_entry = vm_map_entry_create(map, countp);
1121 new_entry->start = start;
1122 new_entry->end = end;
1125 new_entry->maptype = maptype;
1126 new_entry->eflags = protoeflags;
1127 new_entry->object.map_object = map_object;
1128 new_entry->aux.master_pde = 0; /* in case size is different */
1129 new_entry->aux.map_aux = map_aux;
1130 new_entry->offset = offset;
1132 new_entry->inheritance = VM_INHERIT_DEFAULT;
1133 new_entry->protection = prot;
1134 new_entry->max_protection = max;
1135 new_entry->wired_count = 0;
1138 * Insert the new entry into the list
1141 vm_map_entry_link(map, prev_entry, new_entry);
1142 map->size += new_entry->end - new_entry->start;
1145 * Update the free space hint. Entries cannot overlap.
1146 * An exact comparison is needed to avoid matching
1147 * against the map->header.
1149 if ((map->first_free == prev_entry) &&
1150 (prev_entry->end == new_entry->start)) {
1151 map->first_free = new_entry;
1156 * Temporarily removed to avoid MAP_STACK panic, due to
1157 * MAP_STACK being a huge hack. Will be added back in
1158 * when MAP_STACK (and the user stack mapping) is fixed.
1161 * It may be possible to simplify the entry
1163 vm_map_simplify_entry(map, new_entry, countp);
1167 * Try to pre-populate the page table. Mappings governed by virtual
1168 * page tables cannot be prepopulated without a lot of work, so
1171 if ((cow & (MAP_PREFAULT|MAP_PREFAULT_PARTIAL)) &&
1172 maptype != VM_MAPTYPE_VPAGETABLE &&
1173 maptype != VM_MAPTYPE_UKSMAP) {
1175 if (vm_map_relock_enable && (cow & MAP_PREFAULT_RELOCK)) {
1177 vm_object_lock_swap();
1178 vm_object_drop(object);
1180 pmap_object_init_pt(map->pmap, start, prot,
1181 object, OFF_TO_IDX(offset), end - start,
1182 cow & MAP_PREFAULT_PARTIAL);
1184 vm_object_hold(object);
1185 vm_object_lock_swap();
1189 vm_object_drop(object);
1191 lwkt_reltoken(&map->token);
1192 return (KERN_SUCCESS);
1196 * Find sufficient space for `length' bytes in the given map, starting at
1197 * `start'. Returns 0 on success, 1 on no space.
1199 * This function will returned an arbitrarily aligned pointer. If no
1200 * particular alignment is required you should pass align as 1. Note that
1201 * the map may return PAGE_SIZE aligned pointers if all the lengths used in
1202 * the map are a multiple of PAGE_SIZE, even if you pass a smaller align
1205 * 'align' should be a power of 2 but is not required to be.
1207 * The map must be exclusively locked.
1208 * No other requirements.
1211 vm_map_findspace(vm_map_t map, vm_offset_t start, vm_size_t length,
1212 vm_size_t align, int flags, vm_offset_t *addr)
1214 vm_map_entry_t entry, next;
1216 vm_offset_t align_mask;
1218 if (start < map->min_offset)
1219 start = map->min_offset;
1220 if (start > map->max_offset)
1224 * If the alignment is not a power of 2 we will have to use
1225 * a mod/division, set align_mask to a special value.
1227 if ((align | (align - 1)) + 1 != (align << 1))
1228 align_mask = (vm_offset_t)-1;
1230 align_mask = align - 1;
1233 * Look for the first possible address; if there's already something
1234 * at this address, we have to start after it.
1236 if (start == map->min_offset) {
1237 if ((entry = map->first_free) != &map->header)
1242 if (vm_map_lookup_entry(map, start, &tmp))
1248 * Look through the rest of the map, trying to fit a new region in the
1249 * gap between existing regions, or after the very last region.
1251 for (;; start = (entry = next)->end) {
1253 * Adjust the proposed start by the requested alignment,
1254 * be sure that we didn't wrap the address.
1256 if (align_mask == (vm_offset_t)-1)
1257 end = roundup(start, align);
1259 end = (start + align_mask) & ~align_mask;
1264 * Find the end of the proposed new region. Be sure we didn't
1265 * go beyond the end of the map, or wrap around the address.
1266 * Then check to see if this is the last entry or if the
1267 * proposed end fits in the gap between this and the next
1270 end = start + length;
1271 if (end > map->max_offset || end < start)
1276 * If the next entry's start address is beyond the desired
1277 * end address we may have found a good entry.
1279 * If the next entry is a stack mapping we do not map into
1280 * the stack's reserved space.
1282 * XXX continue to allow mapping into the stack's reserved
1283 * space if doing a MAP_STACK mapping inside a MAP_STACK
1284 * mapping, for backwards compatibility. But the caller
1285 * really should use MAP_STACK | MAP_TRYFIXED if they
1288 if (next == &map->header)
1290 if (next->start >= end) {
1291 if ((next->eflags & MAP_ENTRY_STACK) == 0)
1293 if (flags & MAP_STACK)
1295 if (next->start - next->aux.avail_ssize >= end)
1302 * Grow the kernel_map if necessary. pmap_growkernel() will panic
1303 * if it fails. The kernel_map is locked and nothing can steal
1304 * our address space if pmap_growkernel() blocks.
1306 * NOTE: This may be unconditionally called for kldload areas on
1307 * x86_64 because these do not bump kernel_vm_end (which would
1308 * fill 128G worth of page tables!). Therefore we must not
1311 if (map == &kernel_map) {
1314 kstop = round_page(start + length);
1315 if (kstop > kernel_vm_end)
1316 pmap_growkernel(start, kstop);
1323 * vm_map_find finds an unallocated region in the target address map with
1324 * the given length and allocates it. The search is defined to be first-fit
1325 * from the specified address; the region found is returned in the same
1328 * If object is non-NULL, ref count must be bumped by caller
1329 * prior to making call to account for the new entry.
1331 * No requirements. This function will lock the map temporarily.
1334 vm_map_find(vm_map_t map, void *map_object, void *map_aux,
1335 vm_ooffset_t offset, vm_offset_t *addr,
1336 vm_size_t length, vm_size_t align, boolean_t fitit,
1337 vm_maptype_t maptype, vm_subsys_t id,
1338 vm_prot_t prot, vm_prot_t max, int cow)
1345 if (maptype == VM_MAPTYPE_UKSMAP)
1348 object = map_object;
1352 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
1355 vm_object_hold_shared(object);
1357 if (vm_map_findspace(map, start, length, align, 0, addr)) {
1359 vm_object_drop(object);
1361 vm_map_entry_release(count);
1362 return (KERN_NO_SPACE);
1366 result = vm_map_insert(map, &count, map_object, map_aux,
1367 offset, start, start + length,
1368 maptype, id, prot, max, cow);
1370 vm_object_drop(object);
1372 vm_map_entry_release(count);
1378 * Simplify the given map entry by merging with either neighbor. This
1379 * routine also has the ability to merge with both neighbors.
1381 * This routine guarentees that the passed entry remains valid (though
1382 * possibly extended). When merging, this routine may delete one or
1383 * both neighbors. No action is taken on entries which have their
1384 * in-transition flag set.
1386 * The map must be exclusively locked.
1389 vm_map_simplify_entry(vm_map_t map, vm_map_entry_t entry, int *countp)
1391 vm_map_entry_t next, prev;
1392 vm_size_t prevsize, esize;
1394 if (entry->eflags & MAP_ENTRY_IN_TRANSITION) {
1395 ++mycpu->gd_cnt.v_intrans_coll;
1399 if (entry->maptype == VM_MAPTYPE_SUBMAP)
1401 if (entry->maptype == VM_MAPTYPE_UKSMAP)
1405 if (prev != &map->header) {
1406 prevsize = prev->end - prev->start;
1407 if ( (prev->end == entry->start) &&
1408 (prev->maptype == entry->maptype) &&
1409 (prev->object.vm_object == entry->object.vm_object) &&
1410 (!prev->object.vm_object ||
1411 (prev->offset + prevsize == entry->offset)) &&
1412 (prev->eflags == entry->eflags) &&
1413 (prev->protection == entry->protection) &&
1414 (prev->max_protection == entry->max_protection) &&
1415 (prev->inheritance == entry->inheritance) &&
1416 (prev->id == entry->id) &&
1417 (prev->wired_count == entry->wired_count)) {
1418 if (map->first_free == prev)
1419 map->first_free = entry;
1420 if (map->hint == prev)
1422 vm_map_entry_unlink(map, prev);
1423 entry->start = prev->start;
1424 entry->offset = prev->offset;
1425 if (prev->object.vm_object)
1426 vm_object_deallocate(prev->object.vm_object);
1427 vm_map_entry_dispose(map, prev, countp);
1432 if (next != &map->header) {
1433 esize = entry->end - entry->start;
1434 if ((entry->end == next->start) &&
1435 (next->maptype == entry->maptype) &&
1436 (next->object.vm_object == entry->object.vm_object) &&
1437 (!entry->object.vm_object ||
1438 (entry->offset + esize == next->offset)) &&
1439 (next->eflags == entry->eflags) &&
1440 (next->protection == entry->protection) &&
1441 (next->max_protection == entry->max_protection) &&
1442 (next->inheritance == entry->inheritance) &&
1443 (next->id == entry->id) &&
1444 (next->wired_count == entry->wired_count)) {
1445 if (map->first_free == next)
1446 map->first_free = entry;
1447 if (map->hint == next)
1449 vm_map_entry_unlink(map, next);
1450 entry->end = next->end;
1451 if (next->object.vm_object)
1452 vm_object_deallocate(next->object.vm_object);
1453 vm_map_entry_dispose(map, next, countp);
1459 * Asserts that the given entry begins at or after the specified address.
1460 * If necessary, it splits the entry into two.
1462 #define vm_map_clip_start(map, entry, startaddr, countp) \
1464 if (startaddr > entry->start) \
1465 _vm_map_clip_start(map, entry, startaddr, countp); \
1469 * This routine is called only when it is known that the entry must be split.
1471 * The map must be exclusively locked.
1474 _vm_map_clip_start(vm_map_t map, vm_map_entry_t entry, vm_offset_t start,
1477 vm_map_entry_t new_entry;
1480 * Split off the front portion -- note that we must insert the new
1481 * entry BEFORE this one, so that this entry has the specified
1485 vm_map_simplify_entry(map, entry, countp);
1488 * If there is no object backing this entry, we might as well create
1489 * one now. If we defer it, an object can get created after the map
1490 * is clipped, and individual objects will be created for the split-up
1491 * map. This is a bit of a hack, but is also about the best place to
1492 * put this improvement.
1494 if (entry->object.vm_object == NULL && !map->system_map) {
1495 vm_map_entry_allocate_object(entry);
1498 new_entry = vm_map_entry_create(map, countp);
1499 *new_entry = *entry;
1501 new_entry->end = start;
1502 entry->offset += (start - entry->start);
1503 entry->start = start;
1505 vm_map_entry_link(map, entry->prev, new_entry);
1507 switch(entry->maptype) {
1508 case VM_MAPTYPE_NORMAL:
1509 case VM_MAPTYPE_VPAGETABLE:
1510 if (new_entry->object.vm_object) {
1511 vm_object_hold(new_entry->object.vm_object);
1512 vm_object_chain_wait(new_entry->object.vm_object, 0);
1513 vm_object_reference_locked(new_entry->object.vm_object);
1514 vm_object_drop(new_entry->object.vm_object);
1523 * Asserts that the given entry ends at or before the specified address.
1524 * If necessary, it splits the entry into two.
1526 * The map must be exclusively locked.
1528 #define vm_map_clip_end(map, entry, endaddr, countp) \
1530 if (endaddr < entry->end) \
1531 _vm_map_clip_end(map, entry, endaddr, countp); \
1535 * This routine is called only when it is known that the entry must be split.
1537 * The map must be exclusively locked.
1540 _vm_map_clip_end(vm_map_t map, vm_map_entry_t entry, vm_offset_t end,
1543 vm_map_entry_t new_entry;
1546 * If there is no object backing this entry, we might as well create
1547 * one now. If we defer it, an object can get created after the map
1548 * is clipped, and individual objects will be created for the split-up
1549 * map. This is a bit of a hack, but is also about the best place to
1550 * put this improvement.
1553 if (entry->object.vm_object == NULL && !map->system_map) {
1554 vm_map_entry_allocate_object(entry);
1558 * Create a new entry and insert it AFTER the specified entry
1561 new_entry = vm_map_entry_create(map, countp);
1562 *new_entry = *entry;
1564 new_entry->start = entry->end = end;
1565 new_entry->offset += (end - entry->start);
1567 vm_map_entry_link(map, entry, new_entry);
1569 switch(entry->maptype) {
1570 case VM_MAPTYPE_NORMAL:
1571 case VM_MAPTYPE_VPAGETABLE:
1572 if (new_entry->object.vm_object) {
1573 vm_object_hold(new_entry->object.vm_object);
1574 vm_object_chain_wait(new_entry->object.vm_object, 0);
1575 vm_object_reference_locked(new_entry->object.vm_object);
1576 vm_object_drop(new_entry->object.vm_object);
1585 * Asserts that the starting and ending region addresses fall within the
1586 * valid range for the map.
1588 #define VM_MAP_RANGE_CHECK(map, start, end) \
1590 if (start < vm_map_min(map)) \
1591 start = vm_map_min(map); \
1592 if (end > vm_map_max(map)) \
1593 end = vm_map_max(map); \
1599 * Used to block when an in-transition collison occurs. The map
1600 * is unlocked for the sleep and relocked before the return.
1603 vm_map_transition_wait(vm_map_t map)
1605 tsleep_interlock(map, 0);
1607 tsleep(map, PINTERLOCKED, "vment", 0);
1612 * When we do blocking operations with the map lock held it is
1613 * possible that a clip might have occured on our in-transit entry,
1614 * requiring an adjustment to the entry in our loop. These macros
1615 * help the pageable and clip_range code deal with the case. The
1616 * conditional costs virtually nothing if no clipping has occured.
1619 #define CLIP_CHECK_BACK(entry, save_start) \
1621 while (entry->start != save_start) { \
1622 entry = entry->prev; \
1623 KASSERT(entry != &map->header, ("bad entry clip")); \
1627 #define CLIP_CHECK_FWD(entry, save_end) \
1629 while (entry->end != save_end) { \
1630 entry = entry->next; \
1631 KASSERT(entry != &map->header, ("bad entry clip")); \
1637 * Clip the specified range and return the base entry. The
1638 * range may cover several entries starting at the returned base
1639 * and the first and last entry in the covering sequence will be
1640 * properly clipped to the requested start and end address.
1642 * If no holes are allowed you should pass the MAP_CLIP_NO_HOLES
1645 * The MAP_ENTRY_IN_TRANSITION flag will be set for the entries
1646 * covered by the requested range.
1648 * The map must be exclusively locked on entry and will remain locked
1649 * on return. If no range exists or the range contains holes and you
1650 * specified that no holes were allowed, NULL will be returned. This
1651 * routine may temporarily unlock the map in order avoid a deadlock when
1656 vm_map_clip_range(vm_map_t map, vm_offset_t start, vm_offset_t end,
1657 int *countp, int flags)
1659 vm_map_entry_t start_entry;
1660 vm_map_entry_t entry;
1663 * Locate the entry and effect initial clipping. The in-transition
1664 * case does not occur very often so do not try to optimize it.
1667 if (vm_map_lookup_entry(map, start, &start_entry) == FALSE)
1669 entry = start_entry;
1670 if (entry->eflags & MAP_ENTRY_IN_TRANSITION) {
1671 entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP;
1672 ++mycpu->gd_cnt.v_intrans_coll;
1673 ++mycpu->gd_cnt.v_intrans_wait;
1674 vm_map_transition_wait(map);
1676 * entry and/or start_entry may have been clipped while
1677 * we slept, or may have gone away entirely. We have
1678 * to restart from the lookup.
1684 * Since we hold an exclusive map lock we do not have to restart
1685 * after clipping, even though clipping may block in zalloc.
1687 vm_map_clip_start(map, entry, start, countp);
1688 vm_map_clip_end(map, entry, end, countp);
1689 entry->eflags |= MAP_ENTRY_IN_TRANSITION;
1692 * Scan entries covered by the range. When working on the next
1693 * entry a restart need only re-loop on the current entry which
1694 * we have already locked, since 'next' may have changed. Also,
1695 * even though entry is safe, it may have been clipped so we
1696 * have to iterate forwards through the clip after sleeping.
1698 while (entry->next != &map->header && entry->next->start < end) {
1699 vm_map_entry_t next = entry->next;
1701 if (flags & MAP_CLIP_NO_HOLES) {
1702 if (next->start > entry->end) {
1703 vm_map_unclip_range(map, start_entry,
1704 start, entry->end, countp, flags);
1709 if (next->eflags & MAP_ENTRY_IN_TRANSITION) {
1710 vm_offset_t save_end = entry->end;
1711 next->eflags |= MAP_ENTRY_NEEDS_WAKEUP;
1712 ++mycpu->gd_cnt.v_intrans_coll;
1713 ++mycpu->gd_cnt.v_intrans_wait;
1714 vm_map_transition_wait(map);
1717 * clips might have occured while we blocked.
1719 CLIP_CHECK_FWD(entry, save_end);
1720 CLIP_CHECK_BACK(start_entry, start);
1724 * No restart necessary even though clip_end may block, we
1725 * are holding the map lock.
1727 vm_map_clip_end(map, next, end, countp);
1728 next->eflags |= MAP_ENTRY_IN_TRANSITION;
1731 if (flags & MAP_CLIP_NO_HOLES) {
1732 if (entry->end != end) {
1733 vm_map_unclip_range(map, start_entry,
1734 start, entry->end, countp, flags);
1738 return(start_entry);
1742 * Undo the effect of vm_map_clip_range(). You should pass the same
1743 * flags and the same range that you passed to vm_map_clip_range().
1744 * This code will clear the in-transition flag on the entries and
1745 * wake up anyone waiting. This code will also simplify the sequence
1746 * and attempt to merge it with entries before and after the sequence.
1748 * The map must be locked on entry and will remain locked on return.
1750 * Note that you should also pass the start_entry returned by
1751 * vm_map_clip_range(). However, if you block between the two calls
1752 * with the map unlocked please be aware that the start_entry may
1753 * have been clipped and you may need to scan it backwards to find
1754 * the entry corresponding with the original start address. You are
1755 * responsible for this, vm_map_unclip_range() expects the correct
1756 * start_entry to be passed to it and will KASSERT otherwise.
1760 vm_map_unclip_range(vm_map_t map, vm_map_entry_t start_entry,
1761 vm_offset_t start, vm_offset_t end,
1762 int *countp, int flags)
1764 vm_map_entry_t entry;
1766 entry = start_entry;
1768 KASSERT(entry->start == start, ("unclip_range: illegal base entry"));
1769 while (entry != &map->header && entry->start < end) {
1770 KASSERT(entry->eflags & MAP_ENTRY_IN_TRANSITION,
1771 ("in-transition flag not set during unclip on: %p",
1773 KASSERT(entry->end <= end,
1774 ("unclip_range: tail wasn't clipped"));
1775 entry->eflags &= ~MAP_ENTRY_IN_TRANSITION;
1776 if (entry->eflags & MAP_ENTRY_NEEDS_WAKEUP) {
1777 entry->eflags &= ~MAP_ENTRY_NEEDS_WAKEUP;
1780 entry = entry->next;
1784 * Simplification does not block so there is no restart case.
1786 entry = start_entry;
1787 while (entry != &map->header && entry->start < end) {
1788 vm_map_simplify_entry(map, entry, countp);
1789 entry = entry->next;
1794 * Mark the given range as handled by a subordinate map.
1796 * This range must have been created with vm_map_find(), and no other
1797 * operations may have been performed on this range prior to calling
1800 * Submappings cannot be removed.
1805 vm_map_submap(vm_map_t map, vm_offset_t start, vm_offset_t end, vm_map_t submap)
1807 vm_map_entry_t entry;
1808 int result = KERN_INVALID_ARGUMENT;
1811 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
1814 VM_MAP_RANGE_CHECK(map, start, end);
1816 if (vm_map_lookup_entry(map, start, &entry)) {
1817 vm_map_clip_start(map, entry, start, &count);
1819 entry = entry->next;
1822 vm_map_clip_end(map, entry, end, &count);
1824 if ((entry->start == start) && (entry->end == end) &&
1825 ((entry->eflags & MAP_ENTRY_COW) == 0) &&
1826 (entry->object.vm_object == NULL)) {
1827 entry->object.sub_map = submap;
1828 entry->maptype = VM_MAPTYPE_SUBMAP;
1829 result = KERN_SUCCESS;
1832 vm_map_entry_release(count);
1838 * Sets the protection of the specified address region in the target map.
1839 * If "set_max" is specified, the maximum protection is to be set;
1840 * otherwise, only the current protection is affected.
1842 * The protection is not applicable to submaps, but is applicable to normal
1843 * maps and maps governed by virtual page tables. For example, when operating
1844 * on a virtual page table our protection basically controls how COW occurs
1845 * on the backing object, whereas the virtual page table abstraction itself
1846 * is an abstraction for userland.
1851 vm_map_protect(vm_map_t map, vm_offset_t start, vm_offset_t end,
1852 vm_prot_t new_prot, boolean_t set_max)
1854 vm_map_entry_t current;
1855 vm_map_entry_t entry;
1858 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
1861 VM_MAP_RANGE_CHECK(map, start, end);
1863 if (vm_map_lookup_entry(map, start, &entry)) {
1864 vm_map_clip_start(map, entry, start, &count);
1866 entry = entry->next;
1870 * Make a first pass to check for protection violations.
1873 while ((current != &map->header) && (current->start < end)) {
1874 if (current->maptype == VM_MAPTYPE_SUBMAP) {
1876 vm_map_entry_release(count);
1877 return (KERN_INVALID_ARGUMENT);
1879 if ((new_prot & current->max_protection) != new_prot) {
1881 vm_map_entry_release(count);
1882 return (KERN_PROTECTION_FAILURE);
1884 current = current->next;
1888 * Go back and fix up protections. [Note that clipping is not
1889 * necessary the second time.]
1893 while ((current != &map->header) && (current->start < end)) {
1896 vm_map_clip_end(map, current, end, &count);
1898 old_prot = current->protection;
1900 current->max_protection = new_prot;
1901 current->protection = new_prot & old_prot;
1903 current->protection = new_prot;
1907 * Update physical map if necessary. Worry about copy-on-write
1908 * here -- CHECK THIS XXX
1911 if (current->protection != old_prot) {
1912 #define MASK(entry) (((entry)->eflags & MAP_ENTRY_COW) ? ~VM_PROT_WRITE : \
1915 pmap_protect(map->pmap, current->start,
1917 current->protection & MASK(current));
1921 vm_map_simplify_entry(map, current, &count);
1923 current = current->next;
1927 vm_map_entry_release(count);
1928 return (KERN_SUCCESS);
1932 * This routine traverses a processes map handling the madvise
1933 * system call. Advisories are classified as either those effecting
1934 * the vm_map_entry structure, or those effecting the underlying
1937 * The <value> argument is used for extended madvise calls.
1942 vm_map_madvise(vm_map_t map, vm_offset_t start, vm_offset_t end,
1943 int behav, off_t value)
1945 vm_map_entry_t current, entry;
1951 * Some madvise calls directly modify the vm_map_entry, in which case
1952 * we need to use an exclusive lock on the map and we need to perform
1953 * various clipping operations. Otherwise we only need a read-lock
1957 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
1961 case MADV_SEQUENTIAL:
1975 vm_map_lock_read(map);
1978 vm_map_entry_release(count);
1983 * Locate starting entry and clip if necessary.
1986 VM_MAP_RANGE_CHECK(map, start, end);
1988 if (vm_map_lookup_entry(map, start, &entry)) {
1990 vm_map_clip_start(map, entry, start, &count);
1992 entry = entry->next;
1997 * madvise behaviors that are implemented in the vm_map_entry.
1999 * We clip the vm_map_entry so that behavioral changes are
2000 * limited to the specified address range.
2002 for (current = entry;
2003 (current != &map->header) && (current->start < end);
2004 current = current->next
2006 if (current->maptype == VM_MAPTYPE_SUBMAP)
2009 vm_map_clip_end(map, current, end, &count);
2013 vm_map_entry_set_behavior(current, MAP_ENTRY_BEHAV_NORMAL);
2015 case MADV_SEQUENTIAL:
2016 vm_map_entry_set_behavior(current, MAP_ENTRY_BEHAV_SEQUENTIAL);
2019 vm_map_entry_set_behavior(current, MAP_ENTRY_BEHAV_RANDOM);
2022 current->eflags |= MAP_ENTRY_NOSYNC;
2025 current->eflags &= ~MAP_ENTRY_NOSYNC;
2028 current->eflags |= MAP_ENTRY_NOCOREDUMP;
2031 current->eflags &= ~MAP_ENTRY_NOCOREDUMP;
2035 * Set the page directory page for a map
2036 * governed by a virtual page table. Mark
2037 * the entry as being governed by a virtual
2038 * page table if it is not.
2040 * XXX the page directory page is stored
2041 * in the avail_ssize field if the map_entry.
2043 * XXX the map simplification code does not
2044 * compare this field so weird things may
2045 * happen if you do not apply this function
2046 * to the entire mapping governed by the
2047 * virtual page table.
2049 if (current->maptype != VM_MAPTYPE_VPAGETABLE) {
2053 current->aux.master_pde = value;
2054 pmap_remove(map->pmap,
2055 current->start, current->end);
2059 * Invalidate the related pmap entries, used
2060 * to flush portions of the real kernel's
2061 * pmap when the caller has removed or
2062 * modified existing mappings in a virtual
2065 * (exclusive locked map version)
2067 pmap_remove(map->pmap,
2068 current->start, current->end);
2074 vm_map_simplify_entry(map, current, &count);
2082 * madvise behaviors that are implemented in the underlying
2085 * Since we don't clip the vm_map_entry, we have to clip
2086 * the vm_object pindex and count.
2088 * NOTE! These functions are only supported on normal maps,
2089 * except MADV_INVAL which is also supported on
2090 * virtual page tables.
2092 for (current = entry;
2093 (current != &map->header) && (current->start < end);
2094 current = current->next
2096 vm_offset_t useStart;
2098 if (current->maptype != VM_MAPTYPE_NORMAL &&
2099 (current->maptype != VM_MAPTYPE_VPAGETABLE ||
2100 behav != MADV_INVAL)) {
2104 pindex = OFF_TO_IDX(current->offset);
2105 delta = atop(current->end - current->start);
2106 useStart = current->start;
2108 if (current->start < start) {
2109 pindex += atop(start - current->start);
2110 delta -= atop(start - current->start);
2113 if (current->end > end)
2114 delta -= atop(current->end - end);
2116 if ((vm_spindex_t)delta <= 0)
2119 if (behav == MADV_INVAL) {
2121 * Invalidate the related pmap entries, used
2122 * to flush portions of the real kernel's
2123 * pmap when the caller has removed or
2124 * modified existing mappings in a virtual
2127 * (shared locked map version)
2129 KASSERT(useStart >= VM_MIN_USER_ADDRESS &&
2130 useStart + ptoa(delta) <=
2131 VM_MAX_USER_ADDRESS,
2132 ("Bad range %016jx-%016jx (%016jx)",
2133 useStart, useStart + ptoa(delta),
2135 pmap_remove(map->pmap,
2137 useStart + ptoa(delta));
2139 vm_object_madvise(current->object.vm_object,
2140 pindex, delta, behav);
2144 * Try to populate the page table. Mappings governed
2145 * by virtual page tables cannot be pre-populated
2146 * without a lot of work so don't try.
2148 if (behav == MADV_WILLNEED &&
2149 current->maptype != VM_MAPTYPE_VPAGETABLE) {
2150 pmap_object_init_pt(
2153 current->protection,
2154 current->object.vm_object,
2156 (count << PAGE_SHIFT),
2157 MAP_PREFAULT_MADVISE
2161 vm_map_unlock_read(map);
2163 vm_map_entry_release(count);
2169 * Sets the inheritance of the specified address range in the target map.
2170 * Inheritance affects how the map will be shared with child maps at the
2171 * time of vm_map_fork.
2174 vm_map_inherit(vm_map_t map, vm_offset_t start, vm_offset_t end,
2175 vm_inherit_t new_inheritance)
2177 vm_map_entry_t entry;
2178 vm_map_entry_t temp_entry;
2181 switch (new_inheritance) {
2182 case VM_INHERIT_NONE:
2183 case VM_INHERIT_COPY:
2184 case VM_INHERIT_SHARE:
2187 return (KERN_INVALID_ARGUMENT);
2190 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
2193 VM_MAP_RANGE_CHECK(map, start, end);
2195 if (vm_map_lookup_entry(map, start, &temp_entry)) {
2197 vm_map_clip_start(map, entry, start, &count);
2199 entry = temp_entry->next;
2201 while ((entry != &map->header) && (entry->start < end)) {
2202 vm_map_clip_end(map, entry, end, &count);
2204 entry->inheritance = new_inheritance;
2206 vm_map_simplify_entry(map, entry, &count);
2208 entry = entry->next;
2211 vm_map_entry_release(count);
2212 return (KERN_SUCCESS);
2216 * Implement the semantics of mlock
2219 vm_map_unwire(vm_map_t map, vm_offset_t start, vm_offset_t real_end,
2220 boolean_t new_pageable)
2222 vm_map_entry_t entry;
2223 vm_map_entry_t start_entry;
2225 int rv = KERN_SUCCESS;
2228 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
2230 VM_MAP_RANGE_CHECK(map, start, real_end);
2233 start_entry = vm_map_clip_range(map, start, end, &count,
2235 if (start_entry == NULL) {
2237 vm_map_entry_release(count);
2238 return (KERN_INVALID_ADDRESS);
2241 if (new_pageable == 0) {
2242 entry = start_entry;
2243 while ((entry != &map->header) && (entry->start < end)) {
2244 vm_offset_t save_start;
2245 vm_offset_t save_end;
2248 * Already user wired or hard wired (trivial cases)
2250 if (entry->eflags & MAP_ENTRY_USER_WIRED) {
2251 entry = entry->next;
2254 if (entry->wired_count != 0) {
2255 entry->wired_count++;
2256 entry->eflags |= MAP_ENTRY_USER_WIRED;
2257 entry = entry->next;
2262 * A new wiring requires instantiation of appropriate
2263 * management structures and the faulting in of the
2266 if (entry->maptype == VM_MAPTYPE_NORMAL ||
2267 entry->maptype == VM_MAPTYPE_VPAGETABLE) {
2268 int copyflag = entry->eflags &
2269 MAP_ENTRY_NEEDS_COPY;
2270 if (copyflag && ((entry->protection &
2271 VM_PROT_WRITE) != 0)) {
2272 vm_map_entry_shadow(entry, 0);
2273 } else if (entry->object.vm_object == NULL &&
2275 vm_map_entry_allocate_object(entry);
2278 entry->wired_count++;
2279 entry->eflags |= MAP_ENTRY_USER_WIRED;
2282 * Now fault in the area. Note that vm_fault_wire()
2283 * may release the map lock temporarily, it will be
2284 * relocked on return. The in-transition
2285 * flag protects the entries.
2287 save_start = entry->start;
2288 save_end = entry->end;
2289 rv = vm_fault_wire(map, entry, TRUE, 0);
2291 CLIP_CHECK_BACK(entry, save_start);
2293 KASSERT(entry->wired_count == 1, ("bad wired_count on entry"));
2294 entry->eflags &= ~MAP_ENTRY_USER_WIRED;
2295 entry->wired_count = 0;
2296 if (entry->end == save_end)
2298 entry = entry->next;
2299 KASSERT(entry != &map->header, ("bad entry clip during backout"));
2301 end = save_start; /* unwire the rest */
2305 * note that even though the entry might have been
2306 * clipped, the USER_WIRED flag we set prevents
2307 * duplication so we do not have to do a
2310 entry = entry->next;
2314 * If we failed fall through to the unwiring section to
2315 * unwire what we had wired so far. 'end' has already
2322 * start_entry might have been clipped if we unlocked the
2323 * map and blocked. No matter how clipped it has gotten
2324 * there should be a fragment that is on our start boundary.
2326 CLIP_CHECK_BACK(start_entry, start);
2330 * Deal with the unwiring case.
2334 * This is the unwiring case. We must first ensure that the
2335 * range to be unwired is really wired down. We know there
2338 entry = start_entry;
2339 while ((entry != &map->header) && (entry->start < end)) {
2340 if ((entry->eflags & MAP_ENTRY_USER_WIRED) == 0) {
2341 rv = KERN_INVALID_ARGUMENT;
2344 KASSERT(entry->wired_count != 0, ("wired count was 0 with USER_WIRED set! %p", entry));
2345 entry = entry->next;
2349 * Now decrement the wiring count for each region. If a region
2350 * becomes completely unwired, unwire its physical pages and
2354 * The map entries are processed in a loop, checking to
2355 * make sure the entry is wired and asserting it has a wired
2356 * count. However, another loop was inserted more-or-less in
2357 * the middle of the unwiring path. This loop picks up the
2358 * "entry" loop variable from the first loop without first
2359 * setting it to start_entry. Naturally, the secound loop
2360 * is never entered and the pages backing the entries are
2361 * never unwired. This can lead to a leak of wired pages.
2363 entry = start_entry;
2364 while ((entry != &map->header) && (entry->start < end)) {
2365 KASSERT(entry->eflags & MAP_ENTRY_USER_WIRED,
2366 ("expected USER_WIRED on entry %p", entry));
2367 entry->eflags &= ~MAP_ENTRY_USER_WIRED;
2368 entry->wired_count--;
2369 if (entry->wired_count == 0)
2370 vm_fault_unwire(map, entry);
2371 entry = entry->next;
2375 vm_map_unclip_range(map, start_entry, start, real_end, &count,
2379 vm_map_entry_release(count);
2384 * Sets the pageability of the specified address range in the target map.
2385 * Regions specified as not pageable require locked-down physical
2386 * memory and physical page maps.
2388 * The map must not be locked, but a reference must remain to the map
2389 * throughout the call.
2391 * This function may be called via the zalloc path and must properly
2392 * reserve map entries for kernel_map.
2397 vm_map_wire(vm_map_t map, vm_offset_t start, vm_offset_t real_end, int kmflags)
2399 vm_map_entry_t entry;
2400 vm_map_entry_t start_entry;
2402 int rv = KERN_SUCCESS;
2405 if (kmflags & KM_KRESERVE)
2406 count = vm_map_entry_kreserve(MAP_RESERVE_COUNT);
2408 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
2410 VM_MAP_RANGE_CHECK(map, start, real_end);
2413 start_entry = vm_map_clip_range(map, start, end, &count,
2415 if (start_entry == NULL) {
2417 rv = KERN_INVALID_ADDRESS;
2420 if ((kmflags & KM_PAGEABLE) == 0) {
2424 * 1. Holding the write lock, we create any shadow or zero-fill
2425 * objects that need to be created. Then we clip each map
2426 * entry to the region to be wired and increment its wiring
2427 * count. We create objects before clipping the map entries
2428 * to avoid object proliferation.
2430 * 2. We downgrade to a read lock, and call vm_fault_wire to
2431 * fault in the pages for any newly wired area (wired_count is
2434 * Downgrading to a read lock for vm_fault_wire avoids a
2435 * possible deadlock with another process that may have faulted
2436 * on one of the pages to be wired (it would mark the page busy,
2437 * blocking us, then in turn block on the map lock that we
2438 * hold). Because of problems in the recursive lock package,
2439 * we cannot upgrade to a write lock in vm_map_lookup. Thus,
2440 * any actions that require the write lock must be done
2441 * beforehand. Because we keep the read lock on the map, the
2442 * copy-on-write status of the entries we modify here cannot
2445 entry = start_entry;
2446 while ((entry != &map->header) && (entry->start < end)) {
2448 * Trivial case if the entry is already wired
2450 if (entry->wired_count) {
2451 entry->wired_count++;
2452 entry = entry->next;
2457 * The entry is being newly wired, we have to setup
2458 * appropriate management structures. A shadow
2459 * object is required for a copy-on-write region,
2460 * or a normal object for a zero-fill region. We
2461 * do not have to do this for entries that point to sub
2462 * maps because we won't hold the lock on the sub map.
2464 if (entry->maptype == VM_MAPTYPE_NORMAL ||
2465 entry->maptype == VM_MAPTYPE_VPAGETABLE) {
2466 int copyflag = entry->eflags &
2467 MAP_ENTRY_NEEDS_COPY;
2468 if (copyflag && ((entry->protection &
2469 VM_PROT_WRITE) != 0)) {
2470 vm_map_entry_shadow(entry, 0);
2471 } else if (entry->object.vm_object == NULL &&
2473 vm_map_entry_allocate_object(entry);
2477 entry->wired_count++;
2478 entry = entry->next;
2486 * HACK HACK HACK HACK
2488 * vm_fault_wire() temporarily unlocks the map to avoid
2489 * deadlocks. The in-transition flag from vm_map_clip_range
2490 * call should protect us from changes while the map is
2493 * NOTE: Previously this comment stated that clipping might
2494 * still occur while the entry is unlocked, but from
2495 * what I can tell it actually cannot.
2497 * It is unclear whether the CLIP_CHECK_*() calls
2498 * are still needed but we keep them in anyway.
2500 * HACK HACK HACK HACK
2503 entry = start_entry;
2504 while (entry != &map->header && entry->start < end) {
2506 * If vm_fault_wire fails for any page we need to undo
2507 * what has been done. We decrement the wiring count
2508 * for those pages which have not yet been wired (now)
2509 * and unwire those that have (later).
2511 vm_offset_t save_start = entry->start;
2512 vm_offset_t save_end = entry->end;
2514 if (entry->wired_count == 1)
2515 rv = vm_fault_wire(map, entry, FALSE, kmflags);
2517 CLIP_CHECK_BACK(entry, save_start);
2519 KASSERT(entry->wired_count == 1, ("wired_count changed unexpectedly"));
2520 entry->wired_count = 0;
2521 if (entry->end == save_end)
2523 entry = entry->next;
2524 KASSERT(entry != &map->header, ("bad entry clip during backout"));
2529 CLIP_CHECK_FWD(entry, save_end);
2530 entry = entry->next;
2534 * If a failure occured undo everything by falling through
2535 * to the unwiring code. 'end' has already been adjusted
2539 kmflags |= KM_PAGEABLE;
2542 * start_entry is still IN_TRANSITION but may have been
2543 * clipped since vm_fault_wire() unlocks and relocks the
2544 * map. No matter how clipped it has gotten there should
2545 * be a fragment that is on our start boundary.
2547 CLIP_CHECK_BACK(start_entry, start);
2550 if (kmflags & KM_PAGEABLE) {
2552 * This is the unwiring case. We must first ensure that the
2553 * range to be unwired is really wired down. We know there
2556 entry = start_entry;
2557 while ((entry != &map->header) && (entry->start < end)) {
2558 if (entry->wired_count == 0) {
2559 rv = KERN_INVALID_ARGUMENT;
2562 entry = entry->next;
2566 * Now decrement the wiring count for each region. If a region
2567 * becomes completely unwired, unwire its physical pages and
2570 entry = start_entry;
2571 while ((entry != &map->header) && (entry->start < end)) {
2572 entry->wired_count--;
2573 if (entry->wired_count == 0)
2574 vm_fault_unwire(map, entry);
2575 entry = entry->next;
2579 vm_map_unclip_range(map, start_entry, start, real_end,
2580 &count, MAP_CLIP_NO_HOLES);
2584 if (kmflags & KM_KRESERVE)
2585 vm_map_entry_krelease(count);
2587 vm_map_entry_release(count);
2592 * Mark a newly allocated address range as wired but do not fault in
2593 * the pages. The caller is expected to load the pages into the object.
2595 * The map must be locked on entry and will remain locked on return.
2596 * No other requirements.
2599 vm_map_set_wired_quick(vm_map_t map, vm_offset_t addr, vm_size_t size,
2602 vm_map_entry_t scan;
2603 vm_map_entry_t entry;
2605 entry = vm_map_clip_range(map, addr, addr + size,
2606 countp, MAP_CLIP_NO_HOLES);
2608 scan != &map->header && scan->start < addr + size;
2609 scan = scan->next) {
2610 KKASSERT(scan->wired_count == 0);
2611 scan->wired_count = 1;
2613 vm_map_unclip_range(map, entry, addr, addr + size,
2614 countp, MAP_CLIP_NO_HOLES);
2618 * Push any dirty cached pages in the address range to their pager.
2619 * If syncio is TRUE, dirty pages are written synchronously.
2620 * If invalidate is TRUE, any cached pages are freed as well.
2622 * This routine is called by sys_msync()
2624 * Returns an error if any part of the specified range is not mapped.
2629 vm_map_clean(vm_map_t map, vm_offset_t start, vm_offset_t end,
2630 boolean_t syncio, boolean_t invalidate)
2632 vm_map_entry_t current;
2633 vm_map_entry_t entry;
2637 vm_ooffset_t offset;
2639 vm_map_lock_read(map);
2640 VM_MAP_RANGE_CHECK(map, start, end);
2641 if (!vm_map_lookup_entry(map, start, &entry)) {
2642 vm_map_unlock_read(map);
2643 return (KERN_INVALID_ADDRESS);
2645 lwkt_gettoken(&map->token);
2648 * Make a first pass to check for holes.
2650 for (current = entry; current->start < end; current = current->next) {
2651 if (current->maptype == VM_MAPTYPE_SUBMAP) {
2652 lwkt_reltoken(&map->token);
2653 vm_map_unlock_read(map);
2654 return (KERN_INVALID_ARGUMENT);
2656 if (end > current->end &&
2657 (current->next == &map->header ||
2658 current->end != current->next->start)) {
2659 lwkt_reltoken(&map->token);
2660 vm_map_unlock_read(map);
2661 return (KERN_INVALID_ADDRESS);
2666 pmap_remove(vm_map_pmap(map), start, end);
2669 * Make a second pass, cleaning/uncaching pages from the indicated
2672 for (current = entry; current->start < end; current = current->next) {
2673 offset = current->offset + (start - current->start);
2674 size = (end <= current->end ? end : current->end) - start;
2676 switch(current->maptype) {
2677 case VM_MAPTYPE_SUBMAP:
2680 vm_map_entry_t tentry;
2683 smap = current->object.sub_map;
2684 vm_map_lock_read(smap);
2685 vm_map_lookup_entry(smap, offset, &tentry);
2686 tsize = tentry->end - offset;
2689 object = tentry->object.vm_object;
2690 offset = tentry->offset + (offset - tentry->start);
2691 vm_map_unlock_read(smap);
2694 case VM_MAPTYPE_NORMAL:
2695 case VM_MAPTYPE_VPAGETABLE:
2696 object = current->object.vm_object;
2704 vm_object_hold(object);
2707 * Note that there is absolutely no sense in writing out
2708 * anonymous objects, so we track down the vnode object
2710 * We invalidate (remove) all pages from the address space
2711 * anyway, for semantic correctness.
2713 * note: certain anonymous maps, such as MAP_NOSYNC maps,
2714 * may start out with a NULL object.
2716 while (object && (tobj = object->backing_object) != NULL) {
2717 vm_object_hold(tobj);
2718 if (tobj == object->backing_object) {
2719 vm_object_lock_swap();
2720 offset += object->backing_object_offset;
2721 vm_object_drop(object);
2723 if (object->size < OFF_TO_IDX(offset + size))
2724 size = IDX_TO_OFF(object->size) -
2728 vm_object_drop(tobj);
2730 if (object && (object->type == OBJT_VNODE) &&
2731 (current->protection & VM_PROT_WRITE) &&
2732 (object->flags & OBJ_NOMSYNC) == 0) {
2734 * Flush pages if writing is allowed, invalidate them
2735 * if invalidation requested. Pages undergoing I/O
2736 * will be ignored by vm_object_page_remove().
2738 * We cannot lock the vnode and then wait for paging
2739 * to complete without deadlocking against vm_fault.
2740 * Instead we simply call vm_object_page_remove() and
2741 * allow it to block internally on a page-by-page
2742 * basis when it encounters pages undergoing async
2747 /* no chain wait needed for vnode objects */
2748 vm_object_reference_locked(object);
2749 vn_lock(object->handle, LK_EXCLUSIVE | LK_RETRY);
2750 flags = (syncio || invalidate) ? OBJPC_SYNC : 0;
2751 flags |= invalidate ? OBJPC_INVAL : 0;
2754 * When operating on a virtual page table just
2755 * flush the whole object. XXX we probably ought
2758 switch(current->maptype) {
2759 case VM_MAPTYPE_NORMAL:
2760 vm_object_page_clean(object,
2762 OFF_TO_IDX(offset + size + PAGE_MASK),
2765 case VM_MAPTYPE_VPAGETABLE:
2766 vm_object_page_clean(object, 0, 0, flags);
2769 vn_unlock(((struct vnode *)object->handle));
2770 vm_object_deallocate_locked(object);
2772 if (object && invalidate &&
2773 ((object->type == OBJT_VNODE) ||
2774 (object->type == OBJT_DEVICE) ||
2775 (object->type == OBJT_MGTDEVICE))) {
2777 ((object->type == OBJT_DEVICE) ||
2778 (object->type == OBJT_MGTDEVICE)) ? FALSE : TRUE;
2779 /* no chain wait needed for vnode/device objects */
2780 vm_object_reference_locked(object);
2781 switch(current->maptype) {
2782 case VM_MAPTYPE_NORMAL:
2783 vm_object_page_remove(object,
2785 OFF_TO_IDX(offset + size + PAGE_MASK),
2788 case VM_MAPTYPE_VPAGETABLE:
2789 vm_object_page_remove(object, 0, 0, clean_only);
2792 vm_object_deallocate_locked(object);
2796 vm_object_drop(object);
2799 lwkt_reltoken(&map->token);
2800 vm_map_unlock_read(map);
2802 return (KERN_SUCCESS);
2806 * Make the region specified by this entry pageable.
2808 * The vm_map must be exclusively locked.
2811 vm_map_entry_unwire(vm_map_t map, vm_map_entry_t entry)
2813 entry->eflags &= ~MAP_ENTRY_USER_WIRED;
2814 entry->wired_count = 0;
2815 vm_fault_unwire(map, entry);
2819 * Deallocate the given entry from the target map.
2821 * The vm_map must be exclusively locked.
2824 vm_map_entry_delete(vm_map_t map, vm_map_entry_t entry, int *countp)
2826 vm_map_entry_unlink(map, entry);
2827 map->size -= entry->end - entry->start;
2829 switch(entry->maptype) {
2830 case VM_MAPTYPE_NORMAL:
2831 case VM_MAPTYPE_VPAGETABLE:
2832 case VM_MAPTYPE_SUBMAP:
2833 vm_object_deallocate(entry->object.vm_object);
2835 case VM_MAPTYPE_UKSMAP:
2842 vm_map_entry_dispose(map, entry, countp);
2846 * Deallocates the given address range from the target map.
2848 * The vm_map must be exclusively locked.
2851 vm_map_delete(vm_map_t map, vm_offset_t start, vm_offset_t end, int *countp)
2854 vm_map_entry_t entry;
2855 vm_map_entry_t first_entry;
2857 ASSERT_VM_MAP_LOCKED(map);
2858 lwkt_gettoken(&map->token);
2861 * Find the start of the region, and clip it. Set entry to point
2862 * at the first record containing the requested address or, if no
2863 * such record exists, the next record with a greater address. The
2864 * loop will run from this point until a record beyond the termination
2865 * address is encountered.
2867 * map->hint must be adjusted to not point to anything we delete,
2868 * so set it to the entry prior to the one being deleted.
2870 * GGG see other GGG comment.
2872 if (vm_map_lookup_entry(map, start, &first_entry)) {
2873 entry = first_entry;
2874 vm_map_clip_start(map, entry, start, countp);
2875 map->hint = entry->prev; /* possible problem XXX */
2877 map->hint = first_entry; /* possible problem XXX */
2878 entry = first_entry->next;
2882 * If a hole opens up prior to the current first_free then
2883 * adjust first_free. As with map->hint, map->first_free
2884 * cannot be left set to anything we might delete.
2886 if (entry == &map->header) {
2887 map->first_free = &map->header;
2888 } else if (map->first_free->start >= start) {
2889 map->first_free = entry->prev;
2893 * Step through all entries in this region
2895 while ((entry != &map->header) && (entry->start < end)) {
2896 vm_map_entry_t next;
2898 vm_pindex_t offidxstart, offidxend, count;
2901 * If we hit an in-transition entry we have to sleep and
2902 * retry. It's easier (and not really slower) to just retry
2903 * since this case occurs so rarely and the hint is already
2904 * pointing at the right place. We have to reset the
2905 * start offset so as not to accidently delete an entry
2906 * another process just created in vacated space.
2908 if (entry->eflags & MAP_ENTRY_IN_TRANSITION) {
2909 entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP;
2910 start = entry->start;
2911 ++mycpu->gd_cnt.v_intrans_coll;
2912 ++mycpu->gd_cnt.v_intrans_wait;
2913 vm_map_transition_wait(map);
2916 vm_map_clip_end(map, entry, end, countp);
2922 offidxstart = OFF_TO_IDX(entry->offset);
2923 count = OFF_TO_IDX(e - s);
2925 switch(entry->maptype) {
2926 case VM_MAPTYPE_NORMAL:
2927 case VM_MAPTYPE_VPAGETABLE:
2928 case VM_MAPTYPE_SUBMAP:
2929 object = entry->object.vm_object;
2937 * Unwire before removing addresses from the pmap; otherwise,
2938 * unwiring will put the entries back in the pmap.
2940 if (entry->wired_count != 0)
2941 vm_map_entry_unwire(map, entry);
2943 offidxend = offidxstart + count;
2945 if (object == &kernel_object) {
2946 vm_object_hold(object);
2947 vm_object_page_remove(object, offidxstart,
2949 vm_object_drop(object);
2950 } else if (object && object->type != OBJT_DEFAULT &&
2951 object->type != OBJT_SWAP) {
2953 * vnode object routines cannot be chain-locked,
2954 * but since we aren't removing pages from the
2955 * object here we can use a shared hold.
2957 vm_object_hold_shared(object);
2958 pmap_remove(map->pmap, s, e);
2959 vm_object_drop(object);
2960 } else if (object) {
2961 vm_object_hold(object);
2962 vm_object_chain_acquire(object, 0);
2963 pmap_remove(map->pmap, s, e);
2965 if (object != NULL &&
2966 object->ref_count != 1 &&
2967 (object->flags & (OBJ_NOSPLIT|OBJ_ONEMAPPING)) ==
2969 (object->type == OBJT_DEFAULT ||
2970 object->type == OBJT_SWAP)) {
2971 vm_object_collapse(object, NULL);
2972 vm_object_page_remove(object, offidxstart,
2974 if (object->type == OBJT_SWAP) {
2975 swap_pager_freespace(object,
2979 if (offidxend >= object->size &&
2980 offidxstart < object->size) {
2981 object->size = offidxstart;
2984 vm_object_chain_release(object);
2985 vm_object_drop(object);
2986 } else if (entry->maptype == VM_MAPTYPE_UKSMAP) {
2987 pmap_remove(map->pmap, s, e);
2991 * Delete the entry (which may delete the object) only after
2992 * removing all pmap entries pointing to its pages.
2993 * (Otherwise, its page frames may be reallocated, and any
2994 * modify bits will be set in the wrong object!)
2996 vm_map_entry_delete(map, entry, countp);
2999 lwkt_reltoken(&map->token);
3000 return (KERN_SUCCESS);
3004 * Remove the given address range from the target map.
3005 * This is the exported form of vm_map_delete.
3010 vm_map_remove(vm_map_t map, vm_offset_t start, vm_offset_t end)
3015 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
3017 VM_MAP_RANGE_CHECK(map, start, end);
3018 result = vm_map_delete(map, start, end, &count);
3020 vm_map_entry_release(count);
3026 * Assert that the target map allows the specified privilege on the
3027 * entire address region given. The entire region must be allocated.
3029 * The caller must specify whether the vm_map is already locked or not.
3032 vm_map_check_protection(vm_map_t map, vm_offset_t start, vm_offset_t end,
3033 vm_prot_t protection, boolean_t have_lock)
3035 vm_map_entry_t entry;
3036 vm_map_entry_t tmp_entry;
3039 if (have_lock == FALSE)
3040 vm_map_lock_read(map);
3042 if (!vm_map_lookup_entry(map, start, &tmp_entry)) {
3043 if (have_lock == FALSE)
3044 vm_map_unlock_read(map);
3050 while (start < end) {
3051 if (entry == &map->header) {
3059 if (start < entry->start) {
3064 * Check protection associated with entry.
3067 if ((entry->protection & protection) != protection) {
3071 /* go to next entry */
3074 entry = entry->next;
3076 if (have_lock == FALSE)
3077 vm_map_unlock_read(map);
3082 * If appropriate this function shadows the original object with a new object
3083 * and moves the VM pages from the original object to the new object.
3084 * The original object will also be collapsed, if possible.
3086 * We can only do this for normal memory objects with a single mapping, and
3087 * it only makes sense to do it if there are 2 or more refs on the original
3088 * object. i.e. typically a memory object that has been extended into
3089 * multiple vm_map_entry's with non-overlapping ranges.
3091 * This makes it easier to remove unused pages and keeps object inheritance
3092 * from being a negative impact on memory usage.
3094 * On return the (possibly new) entry->object.vm_object will have an
3095 * additional ref on it for the caller to dispose of (usually by cloning
3096 * the vm_map_entry). The additional ref had to be done in this routine
3097 * to avoid racing a collapse. The object's ONEMAPPING flag will also be
3100 * The vm_map must be locked and its token held.
3103 vm_map_split(vm_map_entry_t entry)
3106 vm_object_t oobject, nobject, bobject;
3109 vm_pindex_t offidxstart, offidxend, idx;
3111 vm_ooffset_t offset;
3115 * Optimize away object locks for vnode objects. Important exit/exec
3118 * OBJ_ONEMAPPING doesn't apply to vnode objects but clear the flag
3121 oobject = entry->object.vm_object;
3122 if (oobject->type != OBJT_DEFAULT && oobject->type != OBJT_SWAP) {
3123 vm_object_reference_quick(oobject);
3124 vm_object_clear_flag(oobject, OBJ_ONEMAPPING);
3129 * Setup. Chain lock the original object throughout the entire
3130 * routine to prevent new page faults from occuring.
3132 * XXX can madvise WILLNEED interfere with us too?
3134 vm_object_hold(oobject);
3135 vm_object_chain_acquire(oobject, 0);
3138 * Original object cannot be split? Might have also changed state.
3140 if (oobject->handle == NULL || (oobject->type != OBJT_DEFAULT &&
3141 oobject->type != OBJT_SWAP)) {
3142 vm_object_chain_release(oobject);
3143 vm_object_reference_locked(oobject);
3144 vm_object_clear_flag(oobject, OBJ_ONEMAPPING);
3145 vm_object_drop(oobject);
3150 * Collapse original object with its backing store as an
3151 * optimization to reduce chain lengths when possible.
3153 * If ref_count <= 1 there aren't other non-overlapping vm_map_entry's
3154 * for oobject, so there's no point collapsing it.
3156 * Then re-check whether the object can be split.
3158 vm_object_collapse(oobject, NULL);
3160 if (oobject->ref_count <= 1 ||
3161 (oobject->type != OBJT_DEFAULT && oobject->type != OBJT_SWAP) ||
3162 (oobject->flags & (OBJ_NOSPLIT|OBJ_ONEMAPPING)) != OBJ_ONEMAPPING) {
3163 vm_object_chain_release(oobject);
3164 vm_object_reference_locked(oobject);
3165 vm_object_clear_flag(oobject, OBJ_ONEMAPPING);
3166 vm_object_drop(oobject);
3171 * Acquire the chain lock on the backing object.
3173 * Give bobject an additional ref count for when it will be shadowed
3177 if ((bobject = oobject->backing_object) != NULL) {
3178 if (bobject->type != OBJT_VNODE) {
3180 vm_object_hold(bobject);
3181 vm_object_chain_wait(bobject, 0);
3182 /* ref for shadowing below */
3183 vm_object_reference_locked(bobject);
3184 vm_object_chain_acquire(bobject, 0);
3185 KKASSERT(bobject->backing_object == bobject);
3186 KKASSERT((bobject->flags & OBJ_DEAD) == 0);
3189 * vnodes are not placed on the shadow list but
3190 * they still get another ref for the backing_object
3193 vm_object_reference_quick(bobject);
3198 * Calculate the object page range and allocate the new object.
3200 offset = entry->offset;
3204 offidxstart = OFF_TO_IDX(offset);
3205 offidxend = offidxstart + OFF_TO_IDX(e - s);
3206 size = offidxend - offidxstart;
3208 switch(oobject->type) {
3210 nobject = default_pager_alloc(NULL, IDX_TO_OFF(size),
3214 nobject = swap_pager_alloc(NULL, IDX_TO_OFF(size),
3223 if (nobject == NULL) {
3225 if (useshadowlist) {
3226 vm_object_chain_release(bobject);
3227 vm_object_deallocate(bobject);
3228 vm_object_drop(bobject);
3230 vm_object_deallocate(bobject);
3233 vm_object_chain_release(oobject);
3234 vm_object_reference_locked(oobject);
3235 vm_object_clear_flag(oobject, OBJ_ONEMAPPING);
3236 vm_object_drop(oobject);
3241 * The new object will replace entry->object.vm_object so it needs
3242 * a second reference (the caller expects an additional ref).
3244 vm_object_hold(nobject);
3245 vm_object_reference_locked(nobject);
3246 vm_object_chain_acquire(nobject, 0);
3249 * nobject shadows bobject (oobject already shadows bobject).
3251 * Adding an object to bobject's shadow list requires refing bobject
3252 * which we did above in the useshadowlist case.
3255 nobject->backing_object_offset =
3256 oobject->backing_object_offset + IDX_TO_OFF(offidxstart);
3257 nobject->backing_object = bobject;
3258 if (useshadowlist) {
3259 bobject->shadow_count++;
3260 bobject->generation++;
3261 LIST_INSERT_HEAD(&bobject->shadow_head,
3262 nobject, shadow_list);
3263 vm_object_clear_flag(bobject, OBJ_ONEMAPPING); /*XXX*/
3264 vm_object_chain_release(bobject);
3265 vm_object_drop(bobject);
3266 vm_object_set_flag(nobject, OBJ_ONSHADOW);
3271 * Move the VM pages from oobject to nobject
3273 for (idx = 0; idx < size; idx++) {
3276 m = vm_page_lookup_busy_wait(oobject, offidxstart + idx,
3282 * We must wait for pending I/O to complete before we can
3285 * We do not have to VM_PROT_NONE the page as mappings should
3286 * not be changed by this operation.
3288 * NOTE: The act of renaming a page updates chaingen for both
3291 vm_page_rename(m, nobject, idx);
3292 /* page automatically made dirty by rename and cache handled */
3293 /* page remains busy */
3296 if (oobject->type == OBJT_SWAP) {
3297 vm_object_pip_add(oobject, 1);
3299 * copy oobject pages into nobject and destroy unneeded
3300 * pages in shadow object.
3302 swap_pager_copy(oobject, nobject, offidxstart, 0);
3303 vm_object_pip_wakeup(oobject);
3307 * Wakeup the pages we played with. No spl protection is needed
3308 * for a simple wakeup.
3310 for (idx = 0; idx < size; idx++) {
3311 m = vm_page_lookup(nobject, idx);
3313 KKASSERT(m->flags & PG_BUSY);
3317 entry->object.vm_object = nobject;
3318 entry->offset = 0LL;
3323 * NOTE: There is no need to remove OBJ_ONEMAPPING from oobject, the
3324 * related pages were moved and are no longer applicable to the
3327 * NOTE: Deallocate oobject (due to its entry->object.vm_object being
3328 * replaced by nobject).
3330 vm_object_chain_release(nobject);
3331 vm_object_drop(nobject);
3332 if (bobject && useshadowlist) {
3333 vm_object_chain_release(bobject);
3334 vm_object_drop(bobject);
3336 vm_object_chain_release(oobject);
3337 /*vm_object_clear_flag(oobject, OBJ_ONEMAPPING);*/
3338 vm_object_deallocate_locked(oobject);
3339 vm_object_drop(oobject);
3343 * Copies the contents of the source entry to the destination
3344 * entry. The entries *must* be aligned properly.
3346 * The vm_maps must be exclusively locked.
3347 * The vm_map's token must be held.
3349 * Because the maps are locked no faults can be in progress during the
3353 vm_map_copy_entry(vm_map_t src_map, vm_map_t dst_map,
3354 vm_map_entry_t src_entry, vm_map_entry_t dst_entry)
3356 vm_object_t src_object;
3358 if (dst_entry->maptype == VM_MAPTYPE_SUBMAP ||
3359 dst_entry->maptype == VM_MAPTYPE_UKSMAP)
3361 if (src_entry->maptype == VM_MAPTYPE_SUBMAP ||
3362 src_entry->maptype == VM_MAPTYPE_UKSMAP)
3365 if (src_entry->wired_count == 0) {
3367 * If the source entry is marked needs_copy, it is already
3370 if ((src_entry->eflags & MAP_ENTRY_NEEDS_COPY) == 0) {
3371 pmap_protect(src_map->pmap,
3374 src_entry->protection & ~VM_PROT_WRITE);
3378 * Make a copy of the object.
3380 * The object must be locked prior to checking the object type
3381 * and for the call to vm_object_collapse() and vm_map_split().
3382 * We cannot use *_hold() here because the split code will
3383 * probably try to destroy the object. The lock is a pool
3384 * token and doesn't care.
3386 * We must bump src_map->timestamp when setting
3387 * MAP_ENTRY_NEEDS_COPY to force any concurrent fault
3388 * to retry, otherwise the concurrent fault might improperly
3389 * install a RW pte when its supposed to be a RO(COW) pte.
3390 * This race can occur because a vnode-backed fault may have
3391 * to temporarily release the map lock.
3393 if (src_entry->object.vm_object != NULL) {
3394 vm_map_split(src_entry);
3395 src_object = src_entry->object.vm_object;
3396 dst_entry->object.vm_object = src_object;
3397 src_entry->eflags |= (MAP_ENTRY_COW |
3398 MAP_ENTRY_NEEDS_COPY);
3399 dst_entry->eflags |= (MAP_ENTRY_COW |
3400 MAP_ENTRY_NEEDS_COPY);
3401 dst_entry->offset = src_entry->offset;
3402 ++src_map->timestamp;
3404 dst_entry->object.vm_object = NULL;
3405 dst_entry->offset = 0;
3408 pmap_copy(dst_map->pmap, src_map->pmap, dst_entry->start,
3409 dst_entry->end - dst_entry->start, src_entry->start);
3412 * Of course, wired down pages can't be set copy-on-write.
3413 * Cause wired pages to be copied into the new map by
3414 * simulating faults (the new pages are pageable)
3416 vm_fault_copy_entry(dst_map, src_map, dst_entry, src_entry);
3422 * Create a new process vmspace structure and vm_map
3423 * based on those of an existing process. The new map
3424 * is based on the old map, according to the inheritance
3425 * values on the regions in that map.
3427 * The source map must not be locked.
3430 static void vmspace_fork_normal_entry(vm_map_t old_map, vm_map_t new_map,
3431 vm_map_entry_t old_entry, int *countp);
3432 static void vmspace_fork_uksmap_entry(vm_map_t old_map, vm_map_t new_map,
3433 vm_map_entry_t old_entry, int *countp);
3436 vmspace_fork(struct vmspace *vm1)
3438 struct vmspace *vm2;
3439 vm_map_t old_map = &vm1->vm_map;
3441 vm_map_entry_t old_entry;
3444 lwkt_gettoken(&vm1->vm_map.token);
3445 vm_map_lock(old_map);
3447 vm2 = vmspace_alloc(old_map->min_offset, old_map->max_offset);
3448 lwkt_gettoken(&vm2->vm_map.token);
3449 bcopy(&vm1->vm_startcopy, &vm2->vm_startcopy,
3450 (caddr_t)&vm1->vm_endcopy - (caddr_t)&vm1->vm_startcopy);
3451 new_map = &vm2->vm_map; /* XXX */
3452 new_map->timestamp = 1;
3454 vm_map_lock(new_map);
3457 old_entry = old_map->header.next;
3458 while (old_entry != &old_map->header) {
3460 old_entry = old_entry->next;
3463 count = vm_map_entry_reserve(count + MAP_RESERVE_COUNT);
3465 old_entry = old_map->header.next;
3466 while (old_entry != &old_map->header) {
3467 switch(old_entry->maptype) {
3468 case VM_MAPTYPE_SUBMAP:
3469 panic("vm_map_fork: encountered a submap");
3471 case VM_MAPTYPE_UKSMAP:
3472 vmspace_fork_uksmap_entry(old_map, new_map,
3475 case VM_MAPTYPE_NORMAL:
3476 case VM_MAPTYPE_VPAGETABLE:
3477 vmspace_fork_normal_entry(old_map, new_map,
3481 old_entry = old_entry->next;
3484 new_map->size = old_map->size;
3485 vm_map_unlock(old_map);
3486 vm_map_unlock(new_map);
3487 vm_map_entry_release(count);
3489 lwkt_reltoken(&vm2->vm_map.token);
3490 lwkt_reltoken(&vm1->vm_map.token);
3497 vmspace_fork_normal_entry(vm_map_t old_map, vm_map_t new_map,
3498 vm_map_entry_t old_entry, int *countp)
3500 vm_map_entry_t new_entry;
3503 switch (old_entry->inheritance) {
3504 case VM_INHERIT_NONE:
3506 case VM_INHERIT_SHARE:
3508 * Clone the entry, creating the shared object if
3511 if (old_entry->object.vm_object == NULL)
3512 vm_map_entry_allocate_object(old_entry);
3514 if (old_entry->eflags & MAP_ENTRY_NEEDS_COPY) {
3516 * Shadow a map_entry which needs a copy,
3517 * replacing its object with a new object
3518 * that points to the old one. Ask the
3519 * shadow code to automatically add an
3520 * additional ref. We can't do it afterwords
3521 * because we might race a collapse. The call
3522 * to vm_map_entry_shadow() will also clear
3525 vm_map_entry_shadow(old_entry, 1);
3526 } else if (old_entry->object.vm_object) {
3528 * We will make a shared copy of the object,
3529 * and must clear OBJ_ONEMAPPING.
3531 * Optimize vnode objects. OBJ_ONEMAPPING
3532 * is non-applicable but clear it anyway,
3533 * and its terminal so we don'th ave to deal
3534 * with chains. Reduces SMP conflicts.
3536 * XXX assert that object.vm_object != NULL
3537 * since we allocate it above.
3539 object = old_entry->object.vm_object;
3540 if (object->type == OBJT_VNODE) {
3541 vm_object_reference_quick(object);
3542 vm_object_clear_flag(object,
3545 vm_object_hold(object);
3546 vm_object_chain_wait(object, 0);
3547 vm_object_reference_locked(object);
3548 vm_object_clear_flag(object,
3550 vm_object_drop(object);
3555 * Clone the entry. We've already bumped the ref on
3558 new_entry = vm_map_entry_create(new_map, countp);
3559 *new_entry = *old_entry;
3560 new_entry->eflags &= ~MAP_ENTRY_USER_WIRED;
3561 new_entry->wired_count = 0;
3564 * Insert the entry into the new map -- we know we're
3565 * inserting at the end of the new map.
3568 vm_map_entry_link(new_map, new_map->header.prev,
3572 * Update the physical map
3574 pmap_copy(new_map->pmap, old_map->pmap,
3576 (old_entry->end - old_entry->start),
3579 case VM_INHERIT_COPY:
3581 * Clone the entry and link into the map.
3583 new_entry = vm_map_entry_create(new_map, countp);
3584 *new_entry = *old_entry;
3585 new_entry->eflags &= ~MAP_ENTRY_USER_WIRED;
3586 new_entry->wired_count = 0;
3587 new_entry->object.vm_object = NULL;
3588 vm_map_entry_link(new_map, new_map->header.prev,
3590 vm_map_copy_entry(old_map, new_map, old_entry,
3597 * When forking user-kernel shared maps, the map might change in the
3598 * child so do not try to copy the underlying pmap entries.
3602 vmspace_fork_uksmap_entry(vm_map_t old_map, vm_map_t new_map,
3603 vm_map_entry_t old_entry, int *countp)
3605 vm_map_entry_t new_entry;
3607 new_entry = vm_map_entry_create(new_map, countp);
3608 *new_entry = *old_entry;
3609 new_entry->eflags &= ~MAP_ENTRY_USER_WIRED;
3610 new_entry->wired_count = 0;
3611 vm_map_entry_link(new_map, new_map->header.prev,
3616 * Create an auto-grow stack entry
3621 vm_map_stack (vm_map_t map, vm_offset_t addrbos, vm_size_t max_ssize,
3622 int flags, vm_prot_t prot, vm_prot_t max, int cow)
3624 vm_map_entry_t prev_entry;
3625 vm_map_entry_t new_stack_entry;
3626 vm_size_t init_ssize;
3629 vm_offset_t tmpaddr;
3631 cow |= MAP_IS_STACK;
3633 if (max_ssize < sgrowsiz)
3634 init_ssize = max_ssize;
3636 init_ssize = sgrowsiz;
3638 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
3642 * Find space for the mapping
3644 if ((flags & (MAP_FIXED | MAP_TRYFIXED)) == 0) {
3645 if (vm_map_findspace(map, addrbos, max_ssize, 1,
3648 vm_map_entry_release(count);
3649 return (KERN_NO_SPACE);
3654 /* If addr is already mapped, no go */
3655 if (vm_map_lookup_entry(map, addrbos, &prev_entry)) {
3657 vm_map_entry_release(count);
3658 return (KERN_NO_SPACE);
3662 /* XXX already handled by kern_mmap() */
3663 /* If we would blow our VMEM resource limit, no go */
3664 if (map->size + init_ssize >
3665 curproc->p_rlimit[RLIMIT_VMEM].rlim_cur) {
3667 vm_map_entry_release(count);
3668 return (KERN_NO_SPACE);
3673 * If we can't accomodate max_ssize in the current mapping,
3674 * no go. However, we need to be aware that subsequent user
3675 * mappings might map into the space we have reserved for
3676 * stack, and currently this space is not protected.
3678 * Hopefully we will at least detect this condition
3679 * when we try to grow the stack.
3681 if ((prev_entry->next != &map->header) &&
3682 (prev_entry->next->start < addrbos + max_ssize)) {
3684 vm_map_entry_release(count);
3685 return (KERN_NO_SPACE);
3689 * We initially map a stack of only init_ssize. We will
3690 * grow as needed later. Since this is to be a grow
3691 * down stack, we map at the top of the range.
3693 * Note: we would normally expect prot and max to be
3694 * VM_PROT_ALL, and cow to be 0. Possibly we should
3695 * eliminate these as input parameters, and just
3696 * pass these values here in the insert call.
3698 rv = vm_map_insert(map, &count, NULL, NULL,
3699 0, addrbos + max_ssize - init_ssize,
3700 addrbos + max_ssize,
3702 VM_SUBSYS_STACK, prot, max, cow);
3704 /* Now set the avail_ssize amount */
3705 if (rv == KERN_SUCCESS) {
3706 if (prev_entry != &map->header)
3707 vm_map_clip_end(map, prev_entry, addrbos + max_ssize - init_ssize, &count);
3708 new_stack_entry = prev_entry->next;
3709 if (new_stack_entry->end != addrbos + max_ssize ||
3710 new_stack_entry->start != addrbos + max_ssize - init_ssize)
3711 panic ("Bad entry start/end for new stack entry");
3713 new_stack_entry->aux.avail_ssize = max_ssize - init_ssize;
3717 vm_map_entry_release(count);
3722 * Attempts to grow a vm stack entry. Returns KERN_SUCCESS if the
3723 * desired address is already mapped, or if we successfully grow
3724 * the stack. Also returns KERN_SUCCESS if addr is outside the
3725 * stack range (this is strange, but preserves compatibility with
3726 * the grow function in vm_machdep.c).
3731 vm_map_growstack (struct proc *p, vm_offset_t addr)
3733 vm_map_entry_t prev_entry;
3734 vm_map_entry_t stack_entry;
3735 vm_map_entry_t new_stack_entry;
3736 struct vmspace *vm = p->p_vmspace;
3737 vm_map_t map = &vm->vm_map;
3740 int rv = KERN_SUCCESS;
3742 int use_read_lock = 1;
3745 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
3748 vm_map_lock_read(map);
3752 /* If addr is already in the entry range, no need to grow.*/
3753 if (vm_map_lookup_entry(map, addr, &prev_entry))
3756 if ((stack_entry = prev_entry->next) == &map->header)
3758 if (prev_entry == &map->header)
3759 end = stack_entry->start - stack_entry->aux.avail_ssize;
3761 end = prev_entry->end;
3764 * This next test mimics the old grow function in vm_machdep.c.
3765 * It really doesn't quite make sense, but we do it anyway
3766 * for compatibility.
3768 * If not growable stack, return success. This signals the
3769 * caller to proceed as he would normally with normal vm.
3771 if (stack_entry->aux.avail_ssize < 1 ||
3772 addr >= stack_entry->start ||
3773 addr < stack_entry->start - stack_entry->aux.avail_ssize) {
3777 /* Find the minimum grow amount */
3778 grow_amount = roundup (stack_entry->start - addr, PAGE_SIZE);
3779 if (grow_amount > stack_entry->aux.avail_ssize) {
3785 * If there is no longer enough space between the entries
3786 * nogo, and adjust the available space. Note: this
3787 * should only happen if the user has mapped into the
3788 * stack area after the stack was created, and is
3789 * probably an error.
3791 * This also effectively destroys any guard page the user
3792 * might have intended by limiting the stack size.
3794 if (grow_amount > stack_entry->start - end) {
3795 if (use_read_lock && vm_map_lock_upgrade(map)) {
3801 stack_entry->aux.avail_ssize = stack_entry->start - end;
3806 is_procstack = addr >= (vm_offset_t)vm->vm_maxsaddr;
3808 /* If this is the main process stack, see if we're over the
3811 if (is_procstack && (ctob(vm->vm_ssize) + grow_amount >
3812 p->p_rlimit[RLIMIT_STACK].rlim_cur)) {
3817 /* Round up the grow amount modulo SGROWSIZ */
3818 grow_amount = roundup (grow_amount, sgrowsiz);
3819 if (grow_amount > stack_entry->aux.avail_ssize) {
3820 grow_amount = stack_entry->aux.avail_ssize;
3822 if (is_procstack && (ctob(vm->vm_ssize) + grow_amount >
3823 p->p_rlimit[RLIMIT_STACK].rlim_cur)) {
3824 grow_amount = p->p_rlimit[RLIMIT_STACK].rlim_cur -
3828 /* If we would blow our VMEM resource limit, no go */
3829 if (map->size + grow_amount > p->p_rlimit[RLIMIT_VMEM].rlim_cur) {
3834 if (use_read_lock && vm_map_lock_upgrade(map)) {
3841 /* Get the preliminary new entry start value */
3842 addr = stack_entry->start - grow_amount;
3844 /* If this puts us into the previous entry, cut back our growth
3845 * to the available space. Also, see the note above.
3848 stack_entry->aux.avail_ssize = stack_entry->start - end;
3852 rv = vm_map_insert(map, &count, NULL, NULL,
3853 0, addr, stack_entry->start,
3855 VM_SUBSYS_STACK, VM_PROT_ALL, VM_PROT_ALL, 0);
3857 /* Adjust the available stack space by the amount we grew. */
3858 if (rv == KERN_SUCCESS) {
3859 if (prev_entry != &map->header)
3860 vm_map_clip_end(map, prev_entry, addr, &count);
3861 new_stack_entry = prev_entry->next;
3862 if (new_stack_entry->end != stack_entry->start ||
3863 new_stack_entry->start != addr)
3864 panic ("Bad stack grow start/end in new stack entry");
3866 new_stack_entry->aux.avail_ssize =
3867 stack_entry->aux.avail_ssize -
3868 (new_stack_entry->end - new_stack_entry->start);
3870 vm->vm_ssize += btoc(new_stack_entry->end -
3871 new_stack_entry->start);
3874 if (map->flags & MAP_WIREFUTURE)
3875 vm_map_unwire(map, new_stack_entry->start,
3876 new_stack_entry->end, FALSE);
3881 vm_map_unlock_read(map);
3884 vm_map_entry_release(count);
3889 * Unshare the specified VM space for exec. If other processes are
3890 * mapped to it, then create a new one. The new vmspace is null.
3895 vmspace_exec(struct proc *p, struct vmspace *vmcopy)
3897 struct vmspace *oldvmspace = p->p_vmspace;
3898 struct vmspace *newvmspace;
3899 vm_map_t map = &p->p_vmspace->vm_map;
3902 * If we are execing a resident vmspace we fork it, otherwise
3903 * we create a new vmspace. Note that exitingcnt is not
3904 * copied to the new vmspace.
3906 lwkt_gettoken(&oldvmspace->vm_map.token);
3908 newvmspace = vmspace_fork(vmcopy);
3909 lwkt_gettoken(&newvmspace->vm_map.token);
3911 newvmspace = vmspace_alloc(map->min_offset, map->max_offset);
3912 lwkt_gettoken(&newvmspace->vm_map.token);
3913 bcopy(&oldvmspace->vm_startcopy, &newvmspace->vm_startcopy,
3914 (caddr_t)&oldvmspace->vm_endcopy -
3915 (caddr_t)&oldvmspace->vm_startcopy);
3919 * Finish initializing the vmspace before assigning it
3920 * to the process. The vmspace will become the current vmspace
3923 pmap_pinit2(vmspace_pmap(newvmspace));
3924 pmap_replacevm(p, newvmspace, 0);
3925 lwkt_reltoken(&newvmspace->vm_map.token);
3926 lwkt_reltoken(&oldvmspace->vm_map.token);
3927 vmspace_rel(oldvmspace);
3931 * Unshare the specified VM space for forcing COW. This
3932 * is called by rfork, for the (RFMEM|RFPROC) == 0 case.
3935 vmspace_unshare(struct proc *p)
3937 struct vmspace *oldvmspace = p->p_vmspace;
3938 struct vmspace *newvmspace;
3940 lwkt_gettoken(&oldvmspace->vm_map.token);
3941 if (vmspace_getrefs(oldvmspace) == 1) {
3942 lwkt_reltoken(&oldvmspace->vm_map.token);
3945 newvmspace = vmspace_fork(oldvmspace);
3946 lwkt_gettoken(&newvmspace->vm_map.token);
3947 pmap_pinit2(vmspace_pmap(newvmspace));
3948 pmap_replacevm(p, newvmspace, 0);
3949 lwkt_reltoken(&newvmspace->vm_map.token);
3950 lwkt_reltoken(&oldvmspace->vm_map.token);
3951 vmspace_rel(oldvmspace);
3955 * vm_map_hint: return the beginning of the best area suitable for
3956 * creating a new mapping with "prot" protection.
3961 vm_map_hint(struct proc *p, vm_offset_t addr, vm_prot_t prot)
3963 struct vmspace *vms = p->p_vmspace;
3965 if (!randomize_mmap || addr != 0) {
3967 * Set a reasonable start point for the hint if it was
3968 * not specified or if it falls within the heap space.
3969 * Hinted mmap()s do not allocate out of the heap space.
3972 (addr >= round_page((vm_offset_t)vms->vm_taddr) &&
3973 addr < round_page((vm_offset_t)vms->vm_daddr + maxdsiz))) {
3974 addr = round_page((vm_offset_t)vms->vm_daddr + maxdsiz);
3979 addr = (vm_offset_t)vms->vm_daddr + MAXDSIZ;
3980 addr += karc4random() & (MIN((256 * 1024 * 1024), MAXDSIZ) - 1);
3982 return (round_page(addr));
3986 * Finds the VM object, offset, and protection for a given virtual address
3987 * in the specified map, assuming a page fault of the type specified.
3989 * Leaves the map in question locked for read; return values are guaranteed
3990 * until a vm_map_lookup_done call is performed. Note that the map argument
3991 * is in/out; the returned map must be used in the call to vm_map_lookup_done.
3993 * A handle (out_entry) is returned for use in vm_map_lookup_done, to make
3996 * If a lookup is requested with "write protection" specified, the map may
3997 * be changed to perform virtual copying operations, although the data
3998 * referenced will remain the same.
4003 vm_map_lookup(vm_map_t *var_map, /* IN/OUT */
4005 vm_prot_t fault_typea,
4006 vm_map_entry_t *out_entry, /* OUT */
4007 vm_object_t *object, /* OUT */
4008 vm_pindex_t *pindex, /* OUT */
4009 vm_prot_t *out_prot, /* OUT */
4010 boolean_t *wired) /* OUT */
4012 vm_map_entry_t entry;
4013 vm_map_t map = *var_map;
4015 vm_prot_t fault_type = fault_typea;
4016 int use_read_lock = 1;
4017 int rv = KERN_SUCCESS;
4021 vm_map_lock_read(map);
4026 * If the map has an interesting hint, try it before calling full
4027 * blown lookup routine.
4034 if ((entry == &map->header) ||
4035 (vaddr < entry->start) || (vaddr >= entry->end)) {
4036 vm_map_entry_t tmp_entry;
4039 * Entry was either not a valid hint, or the vaddr was not
4040 * contained in the entry, so do a full lookup.
4042 if (!vm_map_lookup_entry(map, vaddr, &tmp_entry)) {
4043 rv = KERN_INVALID_ADDRESS;
4054 if (entry->maptype == VM_MAPTYPE_SUBMAP) {
4055 vm_map_t old_map = map;
4057 *var_map = map = entry->object.sub_map;
4059 vm_map_unlock_read(old_map);
4061 vm_map_unlock(old_map);
4067 * Check whether this task is allowed to have this page.
4068 * Note the special case for MAP_ENTRY_COW pages with an override.
4069 * This is to implement a forced COW for debuggers.
4071 if (fault_type & VM_PROT_OVERRIDE_WRITE)
4072 prot = entry->max_protection;
4074 prot = entry->protection;
4076 fault_type &= (VM_PROT_READ|VM_PROT_WRITE|VM_PROT_EXECUTE);
4077 if ((fault_type & prot) != fault_type) {
4078 rv = KERN_PROTECTION_FAILURE;
4082 if ((entry->eflags & MAP_ENTRY_USER_WIRED) &&
4083 (entry->eflags & MAP_ENTRY_COW) &&
4084 (fault_type & VM_PROT_WRITE) &&
4085 (fault_typea & VM_PROT_OVERRIDE_WRITE) == 0) {
4086 rv = KERN_PROTECTION_FAILURE;
4091 * If this page is not pageable, we have to get it for all possible
4094 *wired = (entry->wired_count != 0);
4096 prot = fault_type = entry->protection;
4099 * Virtual page tables may need to update the accessed (A) bit
4100 * in a page table entry. Upgrade the fault to a write fault for
4101 * that case if the map will support it. If the map does not support
4102 * it the page table entry simply will not be updated.
4104 if (entry->maptype == VM_MAPTYPE_VPAGETABLE) {
4105 if (prot & VM_PROT_WRITE)
4106 fault_type |= VM_PROT_WRITE;
4109 if (curthread->td_lwp && curthread->td_lwp->lwp_vmspace &&
4110 pmap_emulate_ad_bits(&curthread->td_lwp->lwp_vmspace->vm_pmap)) {
4111 if ((prot & VM_PROT_WRITE) == 0)
4112 fault_type |= VM_PROT_WRITE;
4116 * Only NORMAL and VPAGETABLE maps are object-based. UKSMAPs are not.
4118 if (entry->maptype != VM_MAPTYPE_NORMAL &&
4119 entry->maptype != VM_MAPTYPE_VPAGETABLE) {
4125 * If the entry was copy-on-write, we either ...
4127 if (entry->eflags & MAP_ENTRY_NEEDS_COPY) {
4129 * If we want to write the page, we may as well handle that
4130 * now since we've got the map locked.
4132 * If we don't need to write the page, we just demote the
4133 * permissions allowed.
4136 if (fault_type & VM_PROT_WRITE) {
4138 * Not allowed if TDF_NOFAULT is set as the shadowing
4139 * operation can deadlock against the faulting
4140 * function due to the copy-on-write.
4142 if (curthread->td_flags & TDF_NOFAULT) {
4143 rv = KERN_FAILURE_NOFAULT;
4148 * Make a new object, and place it in the object
4149 * chain. Note that no new references have appeared
4150 * -- one just moved from the map to the new
4154 if (use_read_lock && vm_map_lock_upgrade(map)) {
4161 vm_map_entry_shadow(entry, 0);
4164 * We're attempting to read a copy-on-write page --
4165 * don't allow writes.
4168 prot &= ~VM_PROT_WRITE;
4173 * Create an object if necessary.
4175 if (entry->object.vm_object == NULL && !map->system_map) {
4176 if (use_read_lock && vm_map_lock_upgrade(map)) {
4182 vm_map_entry_allocate_object(entry);
4186 * Return the object/offset from this entry. If the entry was
4187 * copy-on-write or empty, it has been fixed up.
4189 *object = entry->object.vm_object;
4192 *pindex = OFF_TO_IDX((vaddr - entry->start) + entry->offset);
4195 * Return whether this is the only map sharing this data. On
4196 * success we return with a read lock held on the map. On failure
4197 * we return with the map unlocked.
4201 if (rv == KERN_SUCCESS) {
4202 if (use_read_lock == 0)
4203 vm_map_lock_downgrade(map);
4204 } else if (use_read_lock) {
4205 vm_map_unlock_read(map);
4213 * Releases locks acquired by a vm_map_lookup()
4214 * (according to the handle returned by that lookup).
4216 * No other requirements.
4219 vm_map_lookup_done(vm_map_t map, vm_map_entry_t entry, int count)
4222 * Unlock the main-level map
4224 vm_map_unlock_read(map);
4226 vm_map_entry_release(count);
4229 #include "opt_ddb.h"
4231 #include <sys/kernel.h>
4233 #include <ddb/ddb.h>
4238 DB_SHOW_COMMAND(map, vm_map_print)
4241 /* XXX convert args. */
4242 vm_map_t map = (vm_map_t)addr;
4243 boolean_t full = have_addr;
4245 vm_map_entry_t entry;
4247 db_iprintf("Task map %p: pmap=%p, nentries=%d, version=%u\n",
4249 (void *)map->pmap, map->nentries, map->timestamp);
4252 if (!full && db_indent)
4256 for (entry = map->header.next; entry != &map->header;
4257 entry = entry->next) {
4258 db_iprintf("map entry %p: start=%p, end=%p\n",
4259 (void *)entry, (void *)entry->start, (void *)entry->end);
4262 static char *inheritance_name[4] =
4263 {"share", "copy", "none", "donate_copy"};
4265 db_iprintf(" prot=%x/%x/%s",
4267 entry->max_protection,
4268 inheritance_name[(int)(unsigned char)
4269 entry->inheritance]);
4270 if (entry->wired_count != 0)
4271 db_printf(", wired");
4273 switch(entry->maptype) {
4274 case VM_MAPTYPE_SUBMAP:
4275 /* XXX no %qd in kernel. Truncate entry->offset. */
4276 db_printf(", share=%p, offset=0x%lx\n",
4277 (void *)entry->object.sub_map,
4278 (long)entry->offset);
4280 if ((entry->prev == &map->header) ||
4281 (entry->prev->object.sub_map !=
4282 entry->object.sub_map)) {
4284 vm_map_print((db_expr_t)(intptr_t)
4285 entry->object.sub_map,
4290 case VM_MAPTYPE_NORMAL:
4291 case VM_MAPTYPE_VPAGETABLE:
4292 /* XXX no %qd in kernel. Truncate entry->offset. */
4293 db_printf(", object=%p, offset=0x%lx",
4294 (void *)entry->object.vm_object,
4295 (long)entry->offset);
4296 if (entry->eflags & MAP_ENTRY_COW)
4297 db_printf(", copy (%s)",
4298 (entry->eflags & MAP_ENTRY_NEEDS_COPY) ? "needed" : "done");
4302 if ((entry->prev == &map->header) ||
4303 (entry->prev->object.vm_object !=
4304 entry->object.vm_object)) {
4306 vm_object_print((db_expr_t)(intptr_t)
4307 entry->object.vm_object,
4313 case VM_MAPTYPE_UKSMAP:
4314 db_printf(", uksmap=%p, offset=0x%lx",
4315 (void *)entry->object.uksmap,
4316 (long)entry->offset);
4317 if (entry->eflags & MAP_ENTRY_COW)
4318 db_printf(", copy (%s)",
4319 (entry->eflags & MAP_ENTRY_NEEDS_COPY) ? "needed" : "done");
4335 DB_SHOW_COMMAND(procvm, procvm)
4340 p = (struct proc *) addr;
4345 db_printf("p = %p, vmspace = %p, map = %p, pmap = %p\n",
4346 (void *)p, (void *)p->p_vmspace, (void *)&p->p_vmspace->vm_map,
4347 (void *)vmspace_pmap(p->p_vmspace));
4349 vm_map_print((db_expr_t)(intptr_t)&p->p_vmspace->vm_map, 1, 0, NULL);