2 * Copyright (c) 1991, 1993
3 * The Regents of the University of California. All rights reserved.
4 * Copyright (c) 2003-2017 The DragonFly Project. All rights reserved.
6 * This code is derived from software contributed to Berkeley by
7 * The Mach Operating System project at Carnegie-Mellon University.
9 * This code is derived from software contributed to The DragonFly Project
10 * by Matthew Dillon <dillon@backplane.com>
12 * Redistribution and use in source and binary forms, with or without
13 * modification, are permitted provided that the following conditions
15 * 1. Redistributions of source code must retain the above copyright
16 * notice, this list of conditions and the following disclaimer.
17 * 2. Redistributions in binary form must reproduce the above copyright
18 * notice, this list of conditions and the following disclaimer in the
19 * documentation and/or other materials provided with the distribution.
20 * 3. Neither the name of the University nor the names of its contributors
21 * may be used to endorse or promote products derived from this software
22 * without specific prior written permission.
24 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
25 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
26 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
27 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
28 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
29 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
30 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
31 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
32 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
33 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
36 * from: @(#)vm_map.c 8.3 (Berkeley) 1/12/94
39 * Copyright (c) 1987, 1990 Carnegie-Mellon University.
40 * All rights reserved.
42 * Authors: Avadis Tevanian, Jr., Michael Wayne Young
44 * Permission to use, copy, modify and distribute this software and
45 * its documentation is hereby granted, provided that both the copyright
46 * notice and this permission notice appear in all copies of the
47 * software, derivative works or modified versions, and any portions
48 * thereof, and that both notices appear in supporting documentation.
50 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
51 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
52 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
54 * Carnegie Mellon requests users of this software to return to
56 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU
57 * School of Computer Science
58 * Carnegie Mellon University
59 * Pittsburgh PA 15213-3890
61 * any improvements or extensions that they make and grant Carnegie the
62 * rights to redistribute these changes.
64 * $FreeBSD: src/sys/vm/vm_map.c,v 1.187.2.19 2003/05/27 00:47:02 alc Exp $
68 * Virtual memory mapping module.
71 #include <sys/param.h>
72 #include <sys/systm.h>
73 #include <sys/kernel.h>
75 #include <sys/serialize.h>
77 #include <sys/vmmeter.h>
79 #include <sys/vnode.h>
80 #include <sys/resourcevar.h>
83 #include <sys/malloc.h>
84 #include <sys/objcache.h>
85 #include <sys/kern_syscall.h>
88 #include <vm/vm_param.h>
90 #include <vm/vm_map.h>
91 #include <vm/vm_page.h>
92 #include <vm/vm_object.h>
93 #include <vm/vm_pager.h>
94 #include <vm/vm_kern.h>
95 #include <vm/vm_extern.h>
96 #include <vm/swap_pager.h>
97 #include <vm/vm_zone.h>
99 #include <sys/random.h>
100 #include <sys/sysctl.h>
101 #include <sys/spinlock.h>
103 #include <sys/thread2.h>
104 #include <sys/spinlock2.h>
107 * Virtual memory maps provide for the mapping, protection, and sharing
108 * of virtual memory objects. In addition, this module provides for an
109 * efficient virtual copy of memory from one map to another.
111 * Synchronization is required prior to most operations.
113 * Maps consist of an ordered doubly-linked list of simple entries.
114 * A hint and a RB tree is used to speed-up lookups.
116 * Callers looking to modify maps specify start/end addresses which cause
117 * the related map entry to be clipped if necessary, and then later
118 * recombined if the pieces remained compatible.
120 * Virtual copy operations are performed by copying VM object references
121 * from one map to another, and then marking both regions as copy-on-write.
123 static boolean_t vmspace_ctor(void *obj, void *privdata, int ocflags);
124 static void vmspace_dtor(void *obj, void *privdata);
125 static void vmspace_terminate(struct vmspace *vm, int final);
127 MALLOC_DEFINE(M_VMSPACE, "vmspace", "vmspace objcache backingstore");
128 static struct objcache *vmspace_cache;
131 * per-cpu page table cross mappings are initialized in early boot
132 * and might require a considerable number of vm_map_entry structures.
134 #define MAPENTRYBSP_CACHE (MAXCPU+1)
135 #define MAPENTRYAP_CACHE 8
138 * Partioning threaded programs with large anonymous memory areas can
139 * improve concurrent fault performance.
141 #define MAP_ENTRY_PARTITION_SIZE ((vm_offset_t)(32 * 1024 * 1024))
142 #define MAP_ENTRY_PARTITION_MASK (MAP_ENTRY_PARTITION_SIZE - 1)
144 #define VM_MAP_ENTRY_WITHIN_PARTITION(entry) \
145 ((((entry)->start ^ (entry)->end) & ~MAP_ENTRY_PARTITION_MASK) == 0)
147 static struct vm_zone mapentzone_store;
148 static vm_zone_t mapentzone;
150 static struct vm_map_entry map_entry_init[MAX_MAPENT];
151 static struct vm_map_entry cpu_map_entry_init_bsp[MAPENTRYBSP_CACHE];
152 static struct vm_map_entry cpu_map_entry_init_ap[MAXCPU][MAPENTRYAP_CACHE];
154 static int randomize_mmap;
155 SYSCTL_INT(_vm, OID_AUTO, randomize_mmap, CTLFLAG_RW, &randomize_mmap, 0,
156 "Randomize mmap offsets");
157 static int vm_map_relock_enable = 1;
158 SYSCTL_INT(_vm, OID_AUTO, map_relock_enable, CTLFLAG_RW,
159 &vm_map_relock_enable, 0, "insert pop pgtable optimization");
160 static int vm_map_partition_enable = 1;
161 SYSCTL_INT(_vm, OID_AUTO, map_partition_enable, CTLFLAG_RW,
162 &vm_map_partition_enable, 0, "Break up larger vm_map_entry's");
164 static void vmspace_drop_notoken(struct vmspace *vm);
165 static void vm_map_entry_shadow(vm_map_entry_t entry, int addref);
166 static vm_map_entry_t vm_map_entry_create(vm_map_t map, int *);
167 static void vm_map_entry_dispose (vm_map_t map, vm_map_entry_t entry, int *);
168 static void _vm_map_clip_end (vm_map_t, vm_map_entry_t, vm_offset_t, int *);
169 static void _vm_map_clip_start (vm_map_t, vm_map_entry_t, vm_offset_t, int *);
170 static void vm_map_entry_delete (vm_map_t, vm_map_entry_t, int *);
171 static void vm_map_entry_unwire (vm_map_t, vm_map_entry_t);
172 static void vm_map_copy_entry (vm_map_t, vm_map_t, vm_map_entry_t,
174 static void vm_map_unclip_range (vm_map_t map, vm_map_entry_t start_entry,
175 vm_offset_t start, vm_offset_t end, int *countp, int flags);
176 static void vm_map_entry_partition(vm_map_t map, vm_map_entry_t entry,
177 vm_offset_t vaddr, int *countp);
180 * Initialize the vm_map module. Must be called before any other vm_map
183 * Map and entry structures are allocated from the general purpose
184 * memory pool with some exceptions:
186 * - The kernel map is allocated statically.
187 * - Initial kernel map entries are allocated out of a static pool.
188 * - We must set ZONE_SPECIAL here or the early boot code can get
189 * stuck if there are >63 cores.
191 * These restrictions are necessary since malloc() uses the
192 * maps and requires map entries.
194 * Called from the low level boot code only.
199 mapentzone = &mapentzone_store;
200 zbootinit(mapentzone, "MAP ENTRY", sizeof (struct vm_map_entry),
201 map_entry_init, MAX_MAPENT);
202 mapentzone_store.zflags |= ZONE_SPECIAL;
206 * Called prior to any vmspace allocations.
208 * Called from the low level boot code only.
213 vmspace_cache = objcache_create_mbacked(M_VMSPACE,
214 sizeof(struct vmspace),
216 vmspace_ctor, vmspace_dtor,
218 zinitna(mapentzone, NULL, 0, 0, ZONE_USE_RESERVE | ZONE_SPECIAL);
224 * objcache support. We leave the pmap root cached as long as possible
225 * for performance reasons.
229 vmspace_ctor(void *obj, void *privdata, int ocflags)
231 struct vmspace *vm = obj;
233 bzero(vm, sizeof(*vm));
234 vm->vm_refcnt = VM_REF_DELETED;
241 vmspace_dtor(void *obj, void *privdata)
243 struct vmspace *vm = obj;
245 KKASSERT(vm->vm_refcnt == VM_REF_DELETED);
246 pmap_puninit(vmspace_pmap(vm));
250 * Red black tree functions
252 * The caller must hold the related map lock.
254 static int rb_vm_map_compare(vm_map_entry_t a, vm_map_entry_t b);
255 RB_GENERATE(vm_map_rb_tree, vm_map_entry, rb_entry, rb_vm_map_compare);
257 /* a->start is address, and the only field has to be initialized */
259 rb_vm_map_compare(vm_map_entry_t a, vm_map_entry_t b)
261 if (a->start < b->start)
263 else if (a->start > b->start)
269 * Initialize vmspace ref/hold counts vmspace0. There is a holdcnt for
273 vmspace_initrefs(struct vmspace *vm)
280 * Allocate a vmspace structure, including a vm_map and pmap.
281 * Initialize numerous fields. While the initial allocation is zerod,
282 * subsequence reuse from the objcache leaves elements of the structure
283 * intact (particularly the pmap), so portions must be zerod.
285 * Returns a referenced vmspace.
290 vmspace_alloc(vm_offset_t min, vm_offset_t max)
294 vm = objcache_get(vmspace_cache, M_WAITOK);
296 bzero(&vm->vm_startcopy,
297 (char *)&vm->vm_endcopy - (char *)&vm->vm_startcopy);
298 vm_map_init(&vm->vm_map, min, max, NULL); /* initializes token */
301 * NOTE: hold to acquires token for safety.
303 * On return vmspace is referenced (refs=1, hold=1). That is,
304 * each refcnt also has a holdcnt. There can be additional holds
305 * (holdcnt) above and beyond the refcnt. Finalization is handled in
306 * two stages, one on refs 1->0, and the the second on hold 1->0.
308 KKASSERT(vm->vm_holdcnt == 0);
309 KKASSERT(vm->vm_refcnt == VM_REF_DELETED);
310 vmspace_initrefs(vm);
312 pmap_pinit(vmspace_pmap(vm)); /* (some fields reused) */
313 vm->vm_map.pmap = vmspace_pmap(vm); /* XXX */
316 cpu_vmspace_alloc(vm);
323 * NOTE: Can return 0 if the vmspace is exiting.
326 vmspace_getrefs(struct vmspace *vm)
332 if (n & VM_REF_DELETED)
338 vmspace_hold(struct vmspace *vm)
340 atomic_add_int(&vm->vm_holdcnt, 1);
341 lwkt_gettoken(&vm->vm_map.token);
345 * Drop with final termination interlock.
348 vmspace_drop(struct vmspace *vm)
350 lwkt_reltoken(&vm->vm_map.token);
351 vmspace_drop_notoken(vm);
355 vmspace_drop_notoken(struct vmspace *vm)
357 if (atomic_fetchadd_int(&vm->vm_holdcnt, -1) == 1) {
358 if (vm->vm_refcnt & VM_REF_DELETED)
359 vmspace_terminate(vm, 1);
364 * A vmspace object must not be in a terminated state to be able to obtain
365 * additional refs on it.
367 * These are official references to the vmspace, the count is used to check
368 * for vmspace sharing. Foreign accessors should use 'hold' and not 'ref'.
370 * XXX we need to combine hold & ref together into one 64-bit field to allow
371 * holds to prevent stage-1 termination.
374 vmspace_ref(struct vmspace *vm)
378 atomic_add_int(&vm->vm_holdcnt, 1);
379 n = atomic_fetchadd_int(&vm->vm_refcnt, 1);
380 KKASSERT((n & VM_REF_DELETED) == 0);
384 * Release a ref on the vmspace. On the 1->0 transition we do stage-1
385 * termination of the vmspace. Then, on the final drop of the hold we
386 * will do stage-2 final termination.
389 vmspace_rel(struct vmspace *vm)
394 * Drop refs. Each ref also has a hold which is also dropped.
396 * When refs hits 0 compete to get the VM_REF_DELETED flag (hold
397 * prevent finalization) to start termination processing.
398 * Finalization occurs when the last hold count drops to 0.
400 n = atomic_fetchadd_int(&vm->vm_refcnt, -1) - 1;
402 if (atomic_cmpset_int(&vm->vm_refcnt, 0, VM_REF_DELETED)) {
403 vmspace_terminate(vm, 0);
409 vmspace_drop_notoken(vm);
413 * This is called during exit indicating that the vmspace is no
414 * longer in used by an exiting process, but the process has not yet
417 * We drop refs, allowing for stage-1 termination, but maintain a holdcnt
418 * to prevent stage-2 until the process is reaped. Note hte order of
419 * operation, we must hold first.
424 vmspace_relexit(struct vmspace *vm)
426 atomic_add_int(&vm->vm_holdcnt, 1);
431 * Called during reap to disconnect the remainder of the vmspace from
432 * the process. On the hold drop the vmspace termination is finalized.
437 vmspace_exitfree(struct proc *p)
443 vmspace_drop_notoken(vm);
447 * Called in two cases:
449 * (1) When the last refcnt is dropped and the vmspace becomes inactive,
450 * called with final == 0. refcnt will be (u_int)-1 at this point,
451 * and holdcnt will still be non-zero.
453 * (2) When holdcnt becomes 0, called with final == 1. There should no
454 * longer be anyone with access to the vmspace.
456 * VMSPACE_EXIT1 flags the primary deactivation
457 * VMSPACE_EXIT2 flags the last reap
460 vmspace_terminate(struct vmspace *vm, int final)
464 lwkt_gettoken(&vm->vm_map.token);
466 KKASSERT((vm->vm_flags & VMSPACE_EXIT1) == 0);
467 vm->vm_flags |= VMSPACE_EXIT1;
470 * Get rid of most of the resources. Leave the kernel pmap
473 * If the pmap does not contain wired pages we can bulk-delete
474 * the pmap as a performance optimization before removing the
477 * If the pmap contains wired pages we cannot do this
478 * pre-optimization because currently vm_fault_unwire()
479 * expects the pmap pages to exist and will not decrement
480 * p->wire_count if they do not.
483 if (vmspace_pmap(vm)->pm_stats.wired_count) {
484 vm_map_remove(&vm->vm_map, VM_MIN_USER_ADDRESS,
485 VM_MAX_USER_ADDRESS);
486 pmap_remove_pages(vmspace_pmap(vm), VM_MIN_USER_ADDRESS,
487 VM_MAX_USER_ADDRESS);
489 pmap_remove_pages(vmspace_pmap(vm), VM_MIN_USER_ADDRESS,
490 VM_MAX_USER_ADDRESS);
491 vm_map_remove(&vm->vm_map, VM_MIN_USER_ADDRESS,
492 VM_MAX_USER_ADDRESS);
494 lwkt_reltoken(&vm->vm_map.token);
496 KKASSERT((vm->vm_flags & VMSPACE_EXIT1) != 0);
497 KKASSERT((vm->vm_flags & VMSPACE_EXIT2) == 0);
500 * Get rid of remaining basic resources.
502 vm->vm_flags |= VMSPACE_EXIT2;
505 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
506 vm_map_lock(&vm->vm_map);
507 cpu_vmspace_free(vm);
510 * Lock the map, to wait out all other references to it.
511 * Delete all of the mappings and pages they hold, then call
512 * the pmap module to reclaim anything left.
514 vm_map_delete(&vm->vm_map,
515 vm_map_min(&vm->vm_map),
516 vm_map_max(&vm->vm_map),
518 vm_map_unlock(&vm->vm_map);
519 vm_map_entry_release(count);
521 pmap_release(vmspace_pmap(vm));
522 lwkt_reltoken(&vm->vm_map.token);
523 objcache_put(vmspace_cache, vm);
528 * Swap useage is determined by taking the proportional swap used by
529 * VM objects backing the VM map. To make up for fractional losses,
530 * if the VM object has any swap use at all the associated map entries
531 * count for at least 1 swap page.
536 vmspace_swap_count(struct vmspace *vm)
538 vm_map_t map = &vm->vm_map;
541 vm_offset_t count = 0;
546 RB_FOREACH(cur, vm_map_rb_tree, &map->rb_root) {
547 switch(cur->maptype) {
548 case VM_MAPTYPE_NORMAL:
549 case VM_MAPTYPE_VPAGETABLE:
550 if ((object = cur->object.vm_object) == NULL)
552 if (object->swblock_count) {
553 n = (cur->end - cur->start) / PAGE_SIZE;
554 count += object->swblock_count *
555 SWAP_META_PAGES * n / object->size + 1;
568 * Calculate the approximate number of anonymous pages in use by
569 * this vmspace. To make up for fractional losses, we count each
570 * VM object as having at least 1 anonymous page.
575 vmspace_anonymous_count(struct vmspace *vm)
577 vm_map_t map = &vm->vm_map;
580 vm_offset_t count = 0;
583 RB_FOREACH(cur, vm_map_rb_tree, &map->rb_root) {
584 switch(cur->maptype) {
585 case VM_MAPTYPE_NORMAL:
586 case VM_MAPTYPE_VPAGETABLE:
587 if ((object = cur->object.vm_object) == NULL)
589 if (object->type != OBJT_DEFAULT &&
590 object->type != OBJT_SWAP) {
593 count += object->resident_page_count;
605 * Initialize an existing vm_map structure such as that in the vmspace
606 * structure. The pmap is initialized elsewhere.
611 vm_map_init(struct vm_map *map, vm_offset_t min_addr, vm_offset_t max_addr,
614 RB_INIT(&map->rb_root);
615 spin_init(&map->ilock_spin, "ilock");
616 map->ilock_base = NULL;
620 vm_map_min(map) = min_addr;
621 vm_map_max(map) = max_addr;
625 bzero(&map->freehint, sizeof(map->freehint));
626 lwkt_token_init(&map->token, "vm_map");
627 lockinit(&map->lock, "vm_maplk", (hz + 9) / 10, 0);
631 * Find the first possible free address for the specified request length.
632 * Returns 0 if we don't have one cached.
636 vm_map_freehint_find(vm_map_t map, vm_size_t length, vm_size_t align)
638 vm_map_freehint_t *scan;
640 scan = &map->freehint[0];
641 while (scan < &map->freehint[VM_MAP_FFCOUNT]) {
642 if (scan->length == length && scan->align == align)
650 * Unconditionally set the freehint. Called by vm_map_findspace() after
651 * it finds an address. This will help us iterate optimally on the next
656 vm_map_freehint_update(vm_map_t map, vm_offset_t start,
657 vm_size_t length, vm_size_t align)
659 vm_map_freehint_t *scan;
661 scan = &map->freehint[0];
662 while (scan < &map->freehint[VM_MAP_FFCOUNT]) {
663 if (scan->length == length && scan->align == align) {
669 scan = &map->freehint[map->freehint_newindex & VM_MAP_FFMASK];
672 scan->length = length;
673 ++map->freehint_newindex;
677 * Update any existing freehints (for any alignment), for the hole we just
682 vm_map_freehint_hole(vm_map_t map, vm_offset_t start, vm_size_t length)
684 vm_map_freehint_t *scan;
686 scan = &map->freehint[0];
687 while (scan < &map->freehint[VM_MAP_FFCOUNT]) {
688 if (scan->length <= length && scan->start > start)
695 * Shadow the vm_map_entry's object. This typically needs to be done when
696 * a write fault is taken on an entry which had previously been cloned by
697 * fork(). The shared object (which might be NULL) must become private so
698 * we add a shadow layer above it.
700 * Object allocation for anonymous mappings is defered as long as possible.
701 * When creating a shadow, however, the underlying object must be instantiated
702 * so it can be shared.
704 * If the map segment is governed by a virtual page table then it is
705 * possible to address offsets beyond the mapped area. Just allocate
706 * a maximally sized object for this case.
708 * If addref is non-zero an additional reference is added to the returned
709 * entry. This mechanic exists because the additional reference might have
710 * to be added atomically and not after return to prevent a premature
713 * The vm_map must be exclusively locked.
714 * No other requirements.
718 vm_map_entry_shadow(vm_map_entry_t entry, int addref)
720 if (entry->maptype == VM_MAPTYPE_VPAGETABLE) {
721 vm_object_shadow(&entry->object.vm_object, &entry->offset,
722 0x7FFFFFFF, addref); /* XXX */
724 vm_object_shadow(&entry->object.vm_object, &entry->offset,
725 atop(entry->end - entry->start), addref);
727 entry->eflags &= ~MAP_ENTRY_NEEDS_COPY;
731 * Allocate an object for a vm_map_entry.
733 * Object allocation for anonymous mappings is defered as long as possible.
734 * This function is called when we can defer no longer, generally when a map
735 * entry might be split or forked or takes a page fault.
737 * If the map segment is governed by a virtual page table then it is
738 * possible to address offsets beyond the mapped area. Just allocate
739 * a maximally sized object for this case.
741 * The vm_map must be exclusively locked.
742 * No other requirements.
745 vm_map_entry_allocate_object(vm_map_entry_t entry)
749 if (entry->maptype == VM_MAPTYPE_VPAGETABLE) {
750 obj = vm_object_allocate(OBJT_DEFAULT, 0x7FFFFFFF); /* XXX */
752 obj = vm_object_allocate(OBJT_DEFAULT,
753 atop(entry->end - entry->start));
755 entry->object.vm_object = obj;
760 * Set an initial negative count so the first attempt to reserve
761 * space preloads a bunch of vm_map_entry's for this cpu. Also
762 * pre-allocate 2 vm_map_entries which will be needed by zalloc() to
763 * map a new page for vm_map_entry structures. SMP systems are
764 * particularly sensitive.
766 * This routine is called in early boot so we cannot just call
767 * vm_map_entry_reserve().
769 * Called from the low level boot code only (for each cpu)
771 * WARNING! Take care not to have too-big a static/BSS structure here
772 * as MAXCPU can be 256+, otherwise the loader's 64MB heap
773 * can get blown out by the kernel plus the initrd image.
776 vm_map_entry_reserve_cpu_init(globaldata_t gd)
778 vm_map_entry_t entry;
782 atomic_add_int(&gd->gd_vme_avail, -MAP_RESERVE_COUNT * 2);
783 if (gd->gd_cpuid == 0) {
784 entry = &cpu_map_entry_init_bsp[0];
785 count = MAPENTRYBSP_CACHE;
787 entry = &cpu_map_entry_init_ap[gd->gd_cpuid][0];
788 count = MAPENTRYAP_CACHE;
790 for (i = 0; i < count; ++i, ++entry) {
791 MAPENT_FREELIST(entry) = gd->gd_vme_base;
792 gd->gd_vme_base = entry;
797 * Reserves vm_map_entry structures so code later-on can manipulate
798 * map_entry structures within a locked map without blocking trying
799 * to allocate a new vm_map_entry.
803 * WARNING! We must not decrement gd_vme_avail until after we have
804 * ensured that sufficient entries exist, otherwise we can
805 * get into an endless call recursion in the zalloc code
809 vm_map_entry_reserve(int count)
811 struct globaldata *gd = mycpu;
812 vm_map_entry_t entry;
815 * Make sure we have enough structures in gd_vme_base to handle
816 * the reservation request.
818 * Use a critical section to protect against VM faults. It might
819 * not be needed, but we have to be careful here.
821 if (gd->gd_vme_avail < count) {
823 while (gd->gd_vme_avail < count) {
824 entry = zalloc(mapentzone);
825 MAPENT_FREELIST(entry) = gd->gd_vme_base;
826 gd->gd_vme_base = entry;
827 atomic_add_int(&gd->gd_vme_avail, 1);
831 atomic_add_int(&gd->gd_vme_avail, -count);
837 * Releases previously reserved vm_map_entry structures that were not
838 * used. If we have too much junk in our per-cpu cache clean some of
844 vm_map_entry_release(int count)
846 struct globaldata *gd = mycpu;
847 vm_map_entry_t entry;
848 vm_map_entry_t efree;
850 count = atomic_fetchadd_int(&gd->gd_vme_avail, count) + count;
851 if (gd->gd_vme_avail > MAP_RESERVE_SLOP) {
854 while (gd->gd_vme_avail > MAP_RESERVE_HYST) {
855 entry = gd->gd_vme_base;
856 KKASSERT(entry != NULL);
857 gd->gd_vme_base = MAPENT_FREELIST(entry);
858 atomic_add_int(&gd->gd_vme_avail, -1);
859 MAPENT_FREELIST(entry) = efree;
863 while ((entry = efree) != NULL) {
864 efree = MAPENT_FREELIST(efree);
865 zfree(mapentzone, entry);
871 * Reserve map entry structures for use in kernel_map itself. These
872 * entries have *ALREADY* been reserved on a per-cpu basis when the map
873 * was inited. This function is used by zalloc() to avoid a recursion
874 * when zalloc() itself needs to allocate additional kernel memory.
876 * This function works like the normal reserve but does not load the
877 * vm_map_entry cache (because that would result in an infinite
878 * recursion). Note that gd_vme_avail may go negative. This is expected.
880 * Any caller of this function must be sure to renormalize after
881 * potentially eating entries to ensure that the reserve supply
887 vm_map_entry_kreserve(int count)
889 struct globaldata *gd = mycpu;
891 atomic_add_int(&gd->gd_vme_avail, -count);
892 KASSERT(gd->gd_vme_base != NULL,
893 ("no reserved entries left, gd_vme_avail = %d",
899 * Release previously reserved map entries for kernel_map. We do not
900 * attempt to clean up like the normal release function as this would
901 * cause an unnecessary (but probably not fatal) deep procedure call.
906 vm_map_entry_krelease(int count)
908 struct globaldata *gd = mycpu;
910 atomic_add_int(&gd->gd_vme_avail, count);
914 * Allocates a VM map entry for insertion. No entry fields are filled in.
916 * The entries should have previously been reserved. The reservation count
917 * is tracked in (*countp).
921 static vm_map_entry_t
922 vm_map_entry_create(vm_map_t map, int *countp)
924 struct globaldata *gd = mycpu;
925 vm_map_entry_t entry;
927 KKASSERT(*countp > 0);
930 entry = gd->gd_vme_base;
931 KASSERT(entry != NULL, ("gd_vme_base NULL! count %d", *countp));
932 gd->gd_vme_base = MAPENT_FREELIST(entry);
939 * Dispose of a vm_map_entry that is no longer being referenced.
944 vm_map_entry_dispose(vm_map_t map, vm_map_entry_t entry, int *countp)
946 struct globaldata *gd = mycpu;
950 MAPENT_FREELIST(entry) = gd->gd_vme_base;
951 gd->gd_vme_base = entry;
957 * Insert/remove entries from maps.
959 * The related map must be exclusively locked.
960 * The caller must hold map->token
961 * No other requirements.
964 vm_map_entry_link(vm_map_t map, vm_map_entry_t entry)
966 ASSERT_VM_MAP_LOCKED(map);
969 if (vm_map_rb_tree_RB_INSERT(&map->rb_root, entry))
970 panic("vm_map_entry_link: dup addr map %p ent %p", map, entry);
974 vm_map_entry_unlink(vm_map_t map,
975 vm_map_entry_t entry)
977 ASSERT_VM_MAP_LOCKED(map);
979 if (entry->eflags & MAP_ENTRY_IN_TRANSITION) {
980 panic("vm_map_entry_unlink: attempt to mess with "
981 "locked entry! %p", entry);
983 vm_map_rb_tree_RB_REMOVE(&map->rb_root, entry);
988 * Finds the map entry containing (or immediately preceding) the specified
989 * address in the given map. The entry is returned in (*entry).
991 * The boolean result indicates whether the address is actually contained
994 * The related map must be locked.
995 * No other requirements.
998 vm_map_lookup_entry(vm_map_t map, vm_offset_t address, vm_map_entry_t *entry)
1001 vm_map_entry_t last;
1003 ASSERT_VM_MAP_LOCKED(map);
1006 * Locate the record from the top of the tree. 'last' tracks the
1007 * closest prior record and is returned if no match is found, which
1008 * in binary tree terms means tracking the most recent right-branch
1009 * taken. If there is no prior record, *entry is set to NULL.
1012 tmp = RB_ROOT(&map->rb_root);
1015 if (address >= tmp->start) {
1016 if (address < tmp->end) {
1021 tmp = RB_RIGHT(tmp, rb_entry);
1023 tmp = RB_LEFT(tmp, rb_entry);
1031 * Inserts the given whole VM object into the target map at the specified
1032 * address range. The object's size should match that of the address range.
1034 * The map must be exclusively locked.
1035 * The object must be held.
1036 * The caller must have reserved sufficient vm_map_entry structures.
1038 * If object is non-NULL, ref count must be bumped by caller prior to
1039 * making call to account for the new entry.
1042 vm_map_insert(vm_map_t map, int *countp, void *map_object, void *map_aux,
1043 vm_ooffset_t offset, vm_offset_t start, vm_offset_t end,
1044 vm_maptype_t maptype, vm_subsys_t id,
1045 vm_prot_t prot, vm_prot_t max, int cow)
1047 vm_map_entry_t new_entry;
1048 vm_map_entry_t prev_entry;
1049 vm_map_entry_t next;
1050 vm_map_entry_t temp_entry;
1051 vm_eflags_t protoeflags;
1055 if (maptype == VM_MAPTYPE_UKSMAP)
1058 object = map_object;
1060 ASSERT_VM_MAP_LOCKED(map);
1062 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
1065 * Check that the start and end points are not bogus.
1067 if ((start < vm_map_min(map)) || (end > vm_map_max(map)) ||
1069 return (KERN_INVALID_ADDRESS);
1073 * Find the entry prior to the proposed starting address; if it's part
1074 * of an existing entry, this range is bogus.
1076 if (vm_map_lookup_entry(map, start, &temp_entry))
1077 return (KERN_NO_SPACE);
1078 prev_entry = temp_entry;
1081 * Assert that the next entry doesn't overlap the end point.
1084 next = vm_map_rb_tree_RB_NEXT(prev_entry);
1086 next = RB_MIN(vm_map_rb_tree, &map->rb_root);
1087 if (next && next->start < end)
1088 return (KERN_NO_SPACE);
1092 if (cow & MAP_COPY_ON_WRITE)
1093 protoeflags |= MAP_ENTRY_COW|MAP_ENTRY_NEEDS_COPY;
1095 if (cow & MAP_NOFAULT) {
1096 protoeflags |= MAP_ENTRY_NOFAULT;
1098 KASSERT(object == NULL,
1099 ("vm_map_insert: paradoxical MAP_NOFAULT request"));
1101 if (cow & MAP_DISABLE_SYNCER)
1102 protoeflags |= MAP_ENTRY_NOSYNC;
1103 if (cow & MAP_DISABLE_COREDUMP)
1104 protoeflags |= MAP_ENTRY_NOCOREDUMP;
1105 if (cow & MAP_IS_STACK)
1106 protoeflags |= MAP_ENTRY_STACK;
1107 if (cow & MAP_IS_KSTACK)
1108 protoeflags |= MAP_ENTRY_KSTACK;
1110 lwkt_gettoken(&map->token);
1114 * When object is non-NULL, it could be shared with another
1115 * process. We have to set or clear OBJ_ONEMAPPING
1118 * NOTE: This flag is only applicable to DEFAULT and SWAP
1119 * objects and will already be clear in other types
1120 * of objects, so a shared object lock is ok for
1123 if ((object->ref_count > 1) || (object->shadow_count != 0)) {
1124 vm_object_clear_flag(object, OBJ_ONEMAPPING);
1127 else if (prev_entry &&
1128 (prev_entry->eflags == protoeflags) &&
1129 (prev_entry->end == start) &&
1130 (prev_entry->wired_count == 0) &&
1131 (prev_entry->id == id) &&
1132 prev_entry->maptype == maptype &&
1133 maptype == VM_MAPTYPE_NORMAL &&
1134 ((prev_entry->object.vm_object == NULL) ||
1135 vm_object_coalesce(prev_entry->object.vm_object,
1136 OFF_TO_IDX(prev_entry->offset),
1137 (vm_size_t)(prev_entry->end - prev_entry->start),
1138 (vm_size_t)(end - prev_entry->end)))) {
1140 * We were able to extend the object. Determine if we
1141 * can extend the previous map entry to include the
1142 * new range as well.
1144 if ((prev_entry->inheritance == VM_INHERIT_DEFAULT) &&
1145 (prev_entry->protection == prot) &&
1146 (prev_entry->max_protection == max)) {
1147 map->size += (end - prev_entry->end);
1148 prev_entry->end = end;
1149 vm_map_simplify_entry(map, prev_entry, countp);
1150 lwkt_reltoken(&map->token);
1151 return (KERN_SUCCESS);
1155 * If we can extend the object but cannot extend the
1156 * map entry, we have to create a new map entry. We
1157 * must bump the ref count on the extended object to
1158 * account for it. object may be NULL.
1160 * XXX if object is NULL should we set offset to 0 here ?
1162 object = prev_entry->object.vm_object;
1163 offset = prev_entry->offset +
1164 (prev_entry->end - prev_entry->start);
1166 vm_object_hold(object);
1167 vm_object_chain_wait(object, 0);
1168 vm_object_reference_locked(object);
1170 map_object = object;
1175 * NOTE: if conditionals fail, object can be NULL here. This occurs
1176 * in things like the buffer map where we manage kva but do not manage
1181 * Create a new entry
1184 new_entry = vm_map_entry_create(map, countp);
1185 new_entry->start = start;
1186 new_entry->end = end;
1189 new_entry->maptype = maptype;
1190 new_entry->eflags = protoeflags;
1191 new_entry->object.map_object = map_object;
1192 new_entry->aux.master_pde = 0; /* in case size is different */
1193 new_entry->aux.map_aux = map_aux;
1194 new_entry->offset = offset;
1196 new_entry->inheritance = VM_INHERIT_DEFAULT;
1197 new_entry->protection = prot;
1198 new_entry->max_protection = max;
1199 new_entry->wired_count = 0;
1202 * Insert the new entry into the list
1205 vm_map_entry_link(map, new_entry);
1206 map->size += new_entry->end - new_entry->start;
1209 * Don't worry about updating freehint[] when inserting, allow
1210 * addresses to be lower than the actual first free spot.
1214 * Temporarily removed to avoid MAP_STACK panic, due to
1215 * MAP_STACK being a huge hack. Will be added back in
1216 * when MAP_STACK (and the user stack mapping) is fixed.
1219 * It may be possible to simplify the entry
1221 vm_map_simplify_entry(map, new_entry, countp);
1225 * Try to pre-populate the page table. Mappings governed by virtual
1226 * page tables cannot be prepopulated without a lot of work, so
1229 if ((cow & (MAP_PREFAULT|MAP_PREFAULT_PARTIAL)) &&
1230 maptype != VM_MAPTYPE_VPAGETABLE &&
1231 maptype != VM_MAPTYPE_UKSMAP) {
1233 if (vm_map_relock_enable && (cow & MAP_PREFAULT_RELOCK)) {
1235 vm_object_lock_swap();
1236 vm_object_drop(object);
1238 pmap_object_init_pt(map->pmap, start, prot,
1239 object, OFF_TO_IDX(offset), end - start,
1240 cow & MAP_PREFAULT_PARTIAL);
1242 vm_object_hold(object);
1243 vm_object_lock_swap();
1247 vm_object_drop(object);
1249 lwkt_reltoken(&map->token);
1250 return (KERN_SUCCESS);
1254 * Find sufficient space for `length' bytes in the given map, starting at
1255 * `start'. Returns 0 on success, 1 on no space.
1257 * This function will returned an arbitrarily aligned pointer. If no
1258 * particular alignment is required you should pass align as 1. Note that
1259 * the map may return PAGE_SIZE aligned pointers if all the lengths used in
1260 * the map are a multiple of PAGE_SIZE, even if you pass a smaller align
1263 * 'align' should be a power of 2 but is not required to be.
1265 * The map must be exclusively locked.
1266 * No other requirements.
1269 vm_map_findspace(vm_map_t map, vm_offset_t start, vm_size_t length,
1270 vm_size_t align, int flags, vm_offset_t *addr)
1272 vm_map_entry_t entry;
1274 vm_offset_t hole_start;
1276 vm_offset_t align_mask;
1278 if (start < vm_map_min(map))
1279 start = vm_map_min(map);
1280 if (start > vm_map_max(map))
1284 * If the alignment is not a power of 2 we will have to use
1285 * a mod/division, set align_mask to a special value.
1287 if ((align | (align - 1)) + 1 != (align << 1))
1288 align_mask = (vm_offset_t)-1;
1290 align_mask = align - 1;
1293 * Use freehint to adjust the start point, hopefully reducing
1294 * the iteration to O(1).
1296 hole_start = vm_map_freehint_find(map, length, align);
1297 if (start < hole_start)
1299 if (vm_map_lookup_entry(map, start, &tmp))
1301 entry = tmp; /* may be NULL */
1304 * Look through the rest of the map, trying to fit a new region in the
1305 * gap between existing regions, or after the very last region.
1309 * Adjust the proposed start by the requested alignment,
1310 * be sure that we didn't wrap the address.
1312 if (align_mask == (vm_offset_t)-1)
1313 end = roundup(start, align);
1315 end = (start + align_mask) & ~align_mask;
1321 * Find the end of the proposed new region. Be sure we didn't
1322 * go beyond the end of the map, or wrap around the address.
1323 * Then check to see if this is the last entry or if the
1324 * proposed end fits in the gap between this and the next
1327 end = start + length;
1328 if (end > vm_map_max(map) || end < start)
1332 * Locate the next entry, we can stop if this is the
1333 * last entry (we know we are in-bounds so that would
1337 entry = vm_map_rb_tree_RB_NEXT(entry);
1339 entry = RB_MIN(vm_map_rb_tree, &map->rb_root);
1344 * Determine if the proposed area would overlap the
1347 if (entry->start >= end) {
1348 if ((entry->eflags & MAP_ENTRY_STACK) == 0)
1350 if (flags & MAP_STACK)
1352 if (entry->start - entry->aux.avail_ssize >= end)
1359 * Update the freehint
1361 vm_map_freehint_update(map, start, length, align);
1364 * Grow the kernel_map if necessary. pmap_growkernel() will panic
1365 * if it fails. The kernel_map is locked and nothing can steal
1366 * our address space if pmap_growkernel() blocks.
1368 * NOTE: This may be unconditionally called for kldload areas on
1369 * x86_64 because these do not bump kernel_vm_end (which would
1370 * fill 128G worth of page tables!). Therefore we must not
1373 if (map == &kernel_map) {
1376 kstop = round_page(start + length);
1377 if (kstop > kernel_vm_end)
1378 pmap_growkernel(start, kstop);
1385 * vm_map_find finds an unallocated region in the target address map with
1386 * the given length and allocates it. The search is defined to be first-fit
1387 * from the specified address; the region found is returned in the same
1390 * If object is non-NULL, ref count must be bumped by caller
1391 * prior to making call to account for the new entry.
1393 * No requirements. This function will lock the map temporarily.
1396 vm_map_find(vm_map_t map, void *map_object, void *map_aux,
1397 vm_ooffset_t offset, vm_offset_t *addr,
1398 vm_size_t length, vm_size_t align, boolean_t fitit,
1399 vm_maptype_t maptype, vm_subsys_t id,
1400 vm_prot_t prot, vm_prot_t max, int cow)
1407 if (maptype == VM_MAPTYPE_UKSMAP)
1410 object = map_object;
1414 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
1417 vm_object_hold_shared(object);
1419 if (vm_map_findspace(map, start, length, align, 0, addr)) {
1421 vm_object_drop(object);
1423 vm_map_entry_release(count);
1424 return (KERN_NO_SPACE);
1428 result = vm_map_insert(map, &count, map_object, map_aux,
1429 offset, start, start + length,
1430 maptype, id, prot, max, cow);
1432 vm_object_drop(object);
1434 vm_map_entry_release(count);
1440 * Simplify the given map entry by merging with either neighbor. This
1441 * routine also has the ability to merge with both neighbors.
1443 * This routine guarentees that the passed entry remains valid (though
1444 * possibly extended). When merging, this routine may delete one or
1445 * both neighbors. No action is taken on entries which have their
1446 * in-transition flag set.
1448 * The map must be exclusively locked.
1451 vm_map_simplify_entry(vm_map_t map, vm_map_entry_t entry, int *countp)
1453 vm_map_entry_t next, prev;
1454 vm_size_t prevsize, esize;
1456 if (entry->eflags & MAP_ENTRY_IN_TRANSITION) {
1457 ++mycpu->gd_cnt.v_intrans_coll;
1461 if (entry->maptype == VM_MAPTYPE_SUBMAP)
1463 if (entry->maptype == VM_MAPTYPE_UKSMAP)
1466 prev = vm_map_rb_tree_RB_PREV(entry);
1468 prevsize = prev->end - prev->start;
1469 if ( (prev->end == entry->start) &&
1470 (prev->maptype == entry->maptype) &&
1471 (prev->object.vm_object == entry->object.vm_object) &&
1472 (!prev->object.vm_object ||
1473 (prev->offset + prevsize == entry->offset)) &&
1474 (prev->eflags == entry->eflags) &&
1475 (prev->protection == entry->protection) &&
1476 (prev->max_protection == entry->max_protection) &&
1477 (prev->inheritance == entry->inheritance) &&
1478 (prev->id == entry->id) &&
1479 (prev->wired_count == entry->wired_count)) {
1480 vm_map_entry_unlink(map, prev);
1481 entry->start = prev->start;
1482 entry->offset = prev->offset;
1483 if (prev->object.vm_object)
1484 vm_object_deallocate(prev->object.vm_object);
1485 vm_map_entry_dispose(map, prev, countp);
1489 next = vm_map_rb_tree_RB_NEXT(entry);
1491 esize = entry->end - entry->start;
1492 if ((entry->end == next->start) &&
1493 (next->maptype == entry->maptype) &&
1494 (next->object.vm_object == entry->object.vm_object) &&
1495 (!entry->object.vm_object ||
1496 (entry->offset + esize == next->offset)) &&
1497 (next->eflags == entry->eflags) &&
1498 (next->protection == entry->protection) &&
1499 (next->max_protection == entry->max_protection) &&
1500 (next->inheritance == entry->inheritance) &&
1501 (next->id == entry->id) &&
1502 (next->wired_count == entry->wired_count)) {
1503 vm_map_entry_unlink(map, next);
1504 entry->end = next->end;
1505 if (next->object.vm_object)
1506 vm_object_deallocate(next->object.vm_object);
1507 vm_map_entry_dispose(map, next, countp);
1513 * Asserts that the given entry begins at or after the specified address.
1514 * If necessary, it splits the entry into two.
1516 #define vm_map_clip_start(map, entry, startaddr, countp) \
1518 if (startaddr > entry->start) \
1519 _vm_map_clip_start(map, entry, startaddr, countp); \
1523 * This routine is called only when it is known that the entry must be split.
1525 * The map must be exclusively locked.
1528 _vm_map_clip_start(vm_map_t map, vm_map_entry_t entry, vm_offset_t start,
1531 vm_map_entry_t new_entry;
1534 * Split off the front portion -- note that we must insert the new
1535 * entry BEFORE this one, so that this entry has the specified
1539 vm_map_simplify_entry(map, entry, countp);
1542 * If there is no object backing this entry, we might as well create
1543 * one now. If we defer it, an object can get created after the map
1544 * is clipped, and individual objects will be created for the split-up
1545 * map. This is a bit of a hack, but is also about the best place to
1546 * put this improvement.
1548 if (entry->object.vm_object == NULL && !map->system_map &&
1549 VM_MAP_ENTRY_WITHIN_PARTITION(entry)) {
1550 vm_map_entry_allocate_object(entry);
1553 new_entry = vm_map_entry_create(map, countp);
1554 *new_entry = *entry;
1556 new_entry->end = start;
1557 entry->offset += (start - entry->start);
1558 entry->start = start;
1560 vm_map_entry_link(map, new_entry);
1562 switch(entry->maptype) {
1563 case VM_MAPTYPE_NORMAL:
1564 case VM_MAPTYPE_VPAGETABLE:
1565 if (new_entry->object.vm_object) {
1566 vm_object_hold(new_entry->object.vm_object);
1567 vm_object_chain_wait(new_entry->object.vm_object, 0);
1568 vm_object_reference_locked(new_entry->object.vm_object);
1569 vm_object_drop(new_entry->object.vm_object);
1578 * Asserts that the given entry ends at or before the specified address.
1579 * If necessary, it splits the entry into two.
1581 * The map must be exclusively locked.
1583 #define vm_map_clip_end(map, entry, endaddr, countp) \
1585 if (endaddr < entry->end) \
1586 _vm_map_clip_end(map, entry, endaddr, countp); \
1590 * This routine is called only when it is known that the entry must be split.
1592 * The map must be exclusively locked.
1595 _vm_map_clip_end(vm_map_t map, vm_map_entry_t entry, vm_offset_t end,
1598 vm_map_entry_t new_entry;
1601 * If there is no object backing this entry, we might as well create
1602 * one now. If we defer it, an object can get created after the map
1603 * is clipped, and individual objects will be created for the split-up
1604 * map. This is a bit of a hack, but is also about the best place to
1605 * put this improvement.
1608 if (entry->object.vm_object == NULL && !map->system_map &&
1609 VM_MAP_ENTRY_WITHIN_PARTITION(entry)) {
1610 vm_map_entry_allocate_object(entry);
1614 * Create a new entry and insert it AFTER the specified entry
1616 new_entry = vm_map_entry_create(map, countp);
1617 *new_entry = *entry;
1619 new_entry->start = entry->end = end;
1620 new_entry->offset += (end - entry->start);
1622 vm_map_entry_link(map, new_entry);
1624 switch(entry->maptype) {
1625 case VM_MAPTYPE_NORMAL:
1626 case VM_MAPTYPE_VPAGETABLE:
1627 if (new_entry->object.vm_object) {
1628 vm_object_hold(new_entry->object.vm_object);
1629 vm_object_chain_wait(new_entry->object.vm_object, 0);
1630 vm_object_reference_locked(new_entry->object.vm_object);
1631 vm_object_drop(new_entry->object.vm_object);
1640 * Asserts that the starting and ending region addresses fall within the
1641 * valid range for the map.
1643 #define VM_MAP_RANGE_CHECK(map, start, end) \
1645 if (start < vm_map_min(map)) \
1646 start = vm_map_min(map); \
1647 if (end > vm_map_max(map)) \
1648 end = vm_map_max(map); \
1654 * Used to block when an in-transition collison occurs. The map
1655 * is unlocked for the sleep and relocked before the return.
1658 vm_map_transition_wait(vm_map_t map, int relock)
1660 tsleep_interlock(map, 0);
1662 tsleep(map, PINTERLOCKED, "vment", 0);
1668 * When we do blocking operations with the map lock held it is
1669 * possible that a clip might have occured on our in-transit entry,
1670 * requiring an adjustment to the entry in our loop. These macros
1671 * help the pageable and clip_range code deal with the case. The
1672 * conditional costs virtually nothing if no clipping has occured.
1675 #define CLIP_CHECK_BACK(entry, save_start) \
1677 while (entry->start != save_start) { \
1678 entry = vm_map_rb_tree_RB_PREV(entry); \
1679 KASSERT(entry, ("bad entry clip")); \
1683 #define CLIP_CHECK_FWD(entry, save_end) \
1685 while (entry->end != save_end) { \
1686 entry = vm_map_rb_tree_RB_NEXT(entry); \
1687 KASSERT(entry, ("bad entry clip")); \
1693 * Clip the specified range and return the base entry. The
1694 * range may cover several entries starting at the returned base
1695 * and the first and last entry in the covering sequence will be
1696 * properly clipped to the requested start and end address.
1698 * If no holes are allowed you should pass the MAP_CLIP_NO_HOLES
1701 * The MAP_ENTRY_IN_TRANSITION flag will be set for the entries
1702 * covered by the requested range.
1704 * The map must be exclusively locked on entry and will remain locked
1705 * on return. If no range exists or the range contains holes and you
1706 * specified that no holes were allowed, NULL will be returned. This
1707 * routine may temporarily unlock the map in order avoid a deadlock when
1712 vm_map_clip_range(vm_map_t map, vm_offset_t start, vm_offset_t end,
1713 int *countp, int flags)
1715 vm_map_entry_t start_entry;
1716 vm_map_entry_t entry;
1717 vm_map_entry_t next;
1720 * Locate the entry and effect initial clipping. The in-transition
1721 * case does not occur very often so do not try to optimize it.
1724 if (vm_map_lookup_entry(map, start, &start_entry) == FALSE)
1726 entry = start_entry;
1727 if (entry->eflags & MAP_ENTRY_IN_TRANSITION) {
1728 entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP;
1729 ++mycpu->gd_cnt.v_intrans_coll;
1730 ++mycpu->gd_cnt.v_intrans_wait;
1731 vm_map_transition_wait(map, 1);
1733 * entry and/or start_entry may have been clipped while
1734 * we slept, or may have gone away entirely. We have
1735 * to restart from the lookup.
1741 * Since we hold an exclusive map lock we do not have to restart
1742 * after clipping, even though clipping may block in zalloc.
1744 vm_map_clip_start(map, entry, start, countp);
1745 vm_map_clip_end(map, entry, end, countp);
1746 entry->eflags |= MAP_ENTRY_IN_TRANSITION;
1749 * Scan entries covered by the range. When working on the next
1750 * entry a restart need only re-loop on the current entry which
1751 * we have already locked, since 'next' may have changed. Also,
1752 * even though entry is safe, it may have been clipped so we
1753 * have to iterate forwards through the clip after sleeping.
1756 next = vm_map_rb_tree_RB_NEXT(entry);
1757 if (next == NULL || next->start >= end)
1759 if (flags & MAP_CLIP_NO_HOLES) {
1760 if (next->start > entry->end) {
1761 vm_map_unclip_range(map, start_entry,
1762 start, entry->end, countp, flags);
1767 if (next->eflags & MAP_ENTRY_IN_TRANSITION) {
1768 vm_offset_t save_end = entry->end;
1769 next->eflags |= MAP_ENTRY_NEEDS_WAKEUP;
1770 ++mycpu->gd_cnt.v_intrans_coll;
1771 ++mycpu->gd_cnt.v_intrans_wait;
1772 vm_map_transition_wait(map, 1);
1775 * clips might have occured while we blocked.
1777 CLIP_CHECK_FWD(entry, save_end);
1778 CLIP_CHECK_BACK(start_entry, start);
1783 * No restart necessary even though clip_end may block, we
1784 * are holding the map lock.
1786 vm_map_clip_end(map, next, end, countp);
1787 next->eflags |= MAP_ENTRY_IN_TRANSITION;
1790 if (flags & MAP_CLIP_NO_HOLES) {
1791 if (entry->end != end) {
1792 vm_map_unclip_range(map, start_entry,
1793 start, entry->end, countp, flags);
1797 return(start_entry);
1801 * Undo the effect of vm_map_clip_range(). You should pass the same
1802 * flags and the same range that you passed to vm_map_clip_range().
1803 * This code will clear the in-transition flag on the entries and
1804 * wake up anyone waiting. This code will also simplify the sequence
1805 * and attempt to merge it with entries before and after the sequence.
1807 * The map must be locked on entry and will remain locked on return.
1809 * Note that you should also pass the start_entry returned by
1810 * vm_map_clip_range(). However, if you block between the two calls
1811 * with the map unlocked please be aware that the start_entry may
1812 * have been clipped and you may need to scan it backwards to find
1813 * the entry corresponding with the original start address. You are
1814 * responsible for this, vm_map_unclip_range() expects the correct
1815 * start_entry to be passed to it and will KASSERT otherwise.
1819 vm_map_unclip_range(vm_map_t map, vm_map_entry_t start_entry,
1820 vm_offset_t start, vm_offset_t end,
1821 int *countp, int flags)
1823 vm_map_entry_t entry;
1825 entry = start_entry;
1827 KASSERT(entry->start == start, ("unclip_range: illegal base entry"));
1828 while (entry && entry->start < end) {
1829 KASSERT(entry->eflags & MAP_ENTRY_IN_TRANSITION,
1830 ("in-transition flag not set during unclip on: %p",
1832 KASSERT(entry->end <= end,
1833 ("unclip_range: tail wasn't clipped"));
1834 entry->eflags &= ~MAP_ENTRY_IN_TRANSITION;
1835 if (entry->eflags & MAP_ENTRY_NEEDS_WAKEUP) {
1836 entry->eflags &= ~MAP_ENTRY_NEEDS_WAKEUP;
1839 entry = vm_map_rb_tree_RB_NEXT(entry);
1843 * Simplification does not block so there is no restart case.
1845 entry = start_entry;
1846 while (entry && entry->start < end) {
1847 vm_map_simplify_entry(map, entry, countp);
1848 entry = vm_map_rb_tree_RB_NEXT(entry);
1853 * Mark the given range as handled by a subordinate map.
1855 * This range must have been created with vm_map_find(), and no other
1856 * operations may have been performed on this range prior to calling
1859 * Submappings cannot be removed.
1864 vm_map_submap(vm_map_t map, vm_offset_t start, vm_offset_t end, vm_map_t submap)
1866 vm_map_entry_t entry;
1867 int result = KERN_INVALID_ARGUMENT;
1870 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
1873 VM_MAP_RANGE_CHECK(map, start, end);
1875 if (vm_map_lookup_entry(map, start, &entry)) {
1876 vm_map_clip_start(map, entry, start, &count);
1878 entry = vm_map_rb_tree_RB_NEXT(entry);
1880 entry = RB_MIN(vm_map_rb_tree, &map->rb_root);
1883 vm_map_clip_end(map, entry, end, &count);
1885 if ((entry->start == start) && (entry->end == end) &&
1886 ((entry->eflags & MAP_ENTRY_COW) == 0) &&
1887 (entry->object.vm_object == NULL)) {
1888 entry->object.sub_map = submap;
1889 entry->maptype = VM_MAPTYPE_SUBMAP;
1890 result = KERN_SUCCESS;
1893 vm_map_entry_release(count);
1899 * Sets the protection of the specified address region in the target map.
1900 * If "set_max" is specified, the maximum protection is to be set;
1901 * otherwise, only the current protection is affected.
1903 * The protection is not applicable to submaps, but is applicable to normal
1904 * maps and maps governed by virtual page tables. For example, when operating
1905 * on a virtual page table our protection basically controls how COW occurs
1906 * on the backing object, whereas the virtual page table abstraction itself
1907 * is an abstraction for userland.
1912 vm_map_protect(vm_map_t map, vm_offset_t start, vm_offset_t end,
1913 vm_prot_t new_prot, boolean_t set_max)
1915 vm_map_entry_t current;
1916 vm_map_entry_t entry;
1919 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
1922 VM_MAP_RANGE_CHECK(map, start, end);
1924 if (vm_map_lookup_entry(map, start, &entry)) {
1925 vm_map_clip_start(map, entry, start, &count);
1927 entry = vm_map_rb_tree_RB_NEXT(entry);
1929 entry = RB_MIN(vm_map_rb_tree, &map->rb_root);
1933 * Make a first pass to check for protection violations.
1936 while (current && current->start < end) {
1937 if (current->maptype == VM_MAPTYPE_SUBMAP) {
1939 vm_map_entry_release(count);
1940 return (KERN_INVALID_ARGUMENT);
1942 if ((new_prot & current->max_protection) != new_prot) {
1944 vm_map_entry_release(count);
1945 return (KERN_PROTECTION_FAILURE);
1949 * When making a SHARED+RW file mmap writable, update
1952 if (new_prot & PROT_WRITE &&
1953 (current->eflags & MAP_ENTRY_NEEDS_COPY) == 0 &&
1954 (current->maptype == VM_MAPTYPE_NORMAL ||
1955 current->maptype == VM_MAPTYPE_VPAGETABLE) &&
1956 current->object.vm_object &&
1957 current->object.vm_object->type == OBJT_VNODE) {
1960 vp = current->object.vm_object->handle;
1961 if (vp && vn_lock(vp, LK_EXCLUSIVE | LK_RETRY | LK_NOWAIT) == 0) {
1962 vfs_timestamp(&vp->v_lastwrite_ts);
1963 vsetflags(vp, VLASTWRITETS);
1967 current = vm_map_rb_tree_RB_NEXT(current);
1971 * Go back and fix up protections. [Note that clipping is not
1972 * necessary the second time.]
1976 while (current && current->start < end) {
1979 vm_map_clip_end(map, current, end, &count);
1981 old_prot = current->protection;
1983 current->max_protection = new_prot;
1984 current->protection = new_prot & old_prot;
1986 current->protection = new_prot;
1990 * Update physical map if necessary. Worry about copy-on-write
1991 * here -- CHECK THIS XXX
1993 if (current->protection != old_prot) {
1994 #define MASK(entry) (((entry)->eflags & MAP_ENTRY_COW) ? ~VM_PROT_WRITE : \
1997 pmap_protect(map->pmap, current->start,
1999 current->protection & MASK(current));
2003 vm_map_simplify_entry(map, current, &count);
2005 current = vm_map_rb_tree_RB_NEXT(current);
2008 vm_map_entry_release(count);
2009 return (KERN_SUCCESS);
2013 * This routine traverses a processes map handling the madvise
2014 * system call. Advisories are classified as either those effecting
2015 * the vm_map_entry structure, or those effecting the underlying
2018 * The <value> argument is used for extended madvise calls.
2023 vm_map_madvise(vm_map_t map, vm_offset_t start, vm_offset_t end,
2024 int behav, off_t value)
2026 vm_map_entry_t current, entry;
2032 * Some madvise calls directly modify the vm_map_entry, in which case
2033 * we need to use an exclusive lock on the map and we need to perform
2034 * various clipping operations. Otherwise we only need a read-lock
2037 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
2041 case MADV_SEQUENTIAL:
2055 vm_map_lock_read(map);
2058 vm_map_entry_release(count);
2063 * Locate starting entry and clip if necessary.
2066 VM_MAP_RANGE_CHECK(map, start, end);
2068 if (vm_map_lookup_entry(map, start, &entry)) {
2070 vm_map_clip_start(map, entry, start, &count);
2072 entry = vm_map_rb_tree_RB_NEXT(entry);
2074 entry = RB_MIN(vm_map_rb_tree, &map->rb_root);
2079 * madvise behaviors that are implemented in the vm_map_entry.
2081 * We clip the vm_map_entry so that behavioral changes are
2082 * limited to the specified address range.
2084 for (current = entry;
2085 current && current->start < end;
2086 current = vm_map_rb_tree_RB_NEXT(current)) {
2090 if (current->maptype == VM_MAPTYPE_SUBMAP)
2093 vm_map_clip_end(map, current, end, &count);
2097 vm_map_entry_set_behavior(current, MAP_ENTRY_BEHAV_NORMAL);
2099 case MADV_SEQUENTIAL:
2100 vm_map_entry_set_behavior(current, MAP_ENTRY_BEHAV_SEQUENTIAL);
2103 vm_map_entry_set_behavior(current, MAP_ENTRY_BEHAV_RANDOM);
2106 current->eflags |= MAP_ENTRY_NOSYNC;
2109 current->eflags &= ~MAP_ENTRY_NOSYNC;
2112 current->eflags |= MAP_ENTRY_NOCOREDUMP;
2115 current->eflags &= ~MAP_ENTRY_NOCOREDUMP;
2119 * Set the page directory page for a map
2120 * governed by a virtual page table. Mark
2121 * the entry as being governed by a virtual
2122 * page table if it is not.
2124 * XXX the page directory page is stored
2125 * in the avail_ssize field if the map_entry.
2127 * XXX the map simplification code does not
2128 * compare this field so weird things may
2129 * happen if you do not apply this function
2130 * to the entire mapping governed by the
2131 * virtual page table.
2133 if (current->maptype != VM_MAPTYPE_VPAGETABLE) {
2137 current->aux.master_pde = value;
2138 pmap_remove(map->pmap,
2139 current->start, current->end);
2143 * Invalidate the related pmap entries, used
2144 * to flush portions of the real kernel's
2145 * pmap when the caller has removed or
2146 * modified existing mappings in a virtual
2149 * (exclusive locked map version does not
2150 * need the range interlock).
2152 pmap_remove(map->pmap,
2153 current->start, current->end);
2159 vm_map_simplify_entry(map, current, &count);
2167 * madvise behaviors that are implemented in the underlying
2170 * Since we don't clip the vm_map_entry, we have to clip
2171 * the vm_object pindex and count.
2173 * NOTE! These functions are only supported on normal maps,
2174 * except MADV_INVAL which is also supported on
2175 * virtual page tables.
2177 for (current = entry;
2178 current && current->start < end;
2179 current = vm_map_rb_tree_RB_NEXT(current)) {
2180 vm_offset_t useStart;
2182 if (current->maptype != VM_MAPTYPE_NORMAL &&
2183 (current->maptype != VM_MAPTYPE_VPAGETABLE ||
2184 behav != MADV_INVAL)) {
2188 pindex = OFF_TO_IDX(current->offset);
2189 delta = atop(current->end - current->start);
2190 useStart = current->start;
2192 if (current->start < start) {
2193 pindex += atop(start - current->start);
2194 delta -= atop(start - current->start);
2197 if (current->end > end)
2198 delta -= atop(current->end - end);
2200 if ((vm_spindex_t)delta <= 0)
2203 if (behav == MADV_INVAL) {
2205 * Invalidate the related pmap entries, used
2206 * to flush portions of the real kernel's
2207 * pmap when the caller has removed or
2208 * modified existing mappings in a virtual
2211 * (shared locked map version needs the
2212 * interlock, see vm_fault()).
2214 struct vm_map_ilock ilock;
2216 KASSERT(useStart >= VM_MIN_USER_ADDRESS &&
2217 useStart + ptoa(delta) <=
2218 VM_MAX_USER_ADDRESS,
2219 ("Bad range %016jx-%016jx (%016jx)",
2220 useStart, useStart + ptoa(delta),
2222 vm_map_interlock(map, &ilock,
2224 useStart + ptoa(delta));
2225 pmap_remove(map->pmap,
2227 useStart + ptoa(delta));
2228 vm_map_deinterlock(map, &ilock);
2230 vm_object_madvise(current->object.vm_object,
2231 pindex, delta, behav);
2235 * Try to populate the page table. Mappings governed
2236 * by virtual page tables cannot be pre-populated
2237 * without a lot of work so don't try.
2239 if (behav == MADV_WILLNEED &&
2240 current->maptype != VM_MAPTYPE_VPAGETABLE) {
2241 pmap_object_init_pt(
2244 current->protection,
2245 current->object.vm_object,
2247 (count << PAGE_SHIFT),
2248 MAP_PREFAULT_MADVISE
2252 vm_map_unlock_read(map);
2254 vm_map_entry_release(count);
2260 * Sets the inheritance of the specified address range in the target map.
2261 * Inheritance affects how the map will be shared with child maps at the
2262 * time of vm_map_fork.
2265 vm_map_inherit(vm_map_t map, vm_offset_t start, vm_offset_t end,
2266 vm_inherit_t new_inheritance)
2268 vm_map_entry_t entry;
2269 vm_map_entry_t temp_entry;
2272 switch (new_inheritance) {
2273 case VM_INHERIT_NONE:
2274 case VM_INHERIT_COPY:
2275 case VM_INHERIT_SHARE:
2278 return (KERN_INVALID_ARGUMENT);
2281 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
2284 VM_MAP_RANGE_CHECK(map, start, end);
2286 if (vm_map_lookup_entry(map, start, &temp_entry)) {
2288 vm_map_clip_start(map, entry, start, &count);
2289 } else if (temp_entry) {
2290 entry = vm_map_rb_tree_RB_NEXT(temp_entry);
2292 entry = RB_MIN(vm_map_rb_tree, &map->rb_root);
2295 while (entry && entry->start < end) {
2296 vm_map_clip_end(map, entry, end, &count);
2298 entry->inheritance = new_inheritance;
2300 vm_map_simplify_entry(map, entry, &count);
2302 entry = vm_map_rb_tree_RB_NEXT(entry);
2305 vm_map_entry_release(count);
2306 return (KERN_SUCCESS);
2310 * Implement the semantics of mlock
2313 vm_map_unwire(vm_map_t map, vm_offset_t start, vm_offset_t real_end,
2314 boolean_t new_pageable)
2316 vm_map_entry_t entry;
2317 vm_map_entry_t start_entry;
2319 int rv = KERN_SUCCESS;
2322 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
2324 VM_MAP_RANGE_CHECK(map, start, real_end);
2327 start_entry = vm_map_clip_range(map, start, end, &count,
2329 if (start_entry == NULL) {
2331 vm_map_entry_release(count);
2332 return (KERN_INVALID_ADDRESS);
2335 if (new_pageable == 0) {
2336 entry = start_entry;
2337 while (entry && entry->start < end) {
2338 vm_offset_t save_start;
2339 vm_offset_t save_end;
2342 * Already user wired or hard wired (trivial cases)
2344 if (entry->eflags & MAP_ENTRY_USER_WIRED) {
2345 entry = vm_map_rb_tree_RB_NEXT(entry);
2348 if (entry->wired_count != 0) {
2349 entry->wired_count++;
2350 entry->eflags |= MAP_ENTRY_USER_WIRED;
2351 entry = vm_map_rb_tree_RB_NEXT(entry);
2356 * A new wiring requires instantiation of appropriate
2357 * management structures and the faulting in of the
2360 if (entry->maptype == VM_MAPTYPE_NORMAL ||
2361 entry->maptype == VM_MAPTYPE_VPAGETABLE) {
2362 int copyflag = entry->eflags &
2363 MAP_ENTRY_NEEDS_COPY;
2364 if (copyflag && ((entry->protection &
2365 VM_PROT_WRITE) != 0)) {
2366 vm_map_entry_shadow(entry, 0);
2367 } else if (entry->object.vm_object == NULL &&
2369 vm_map_entry_allocate_object(entry);
2372 entry->wired_count++;
2373 entry->eflags |= MAP_ENTRY_USER_WIRED;
2376 * Now fault in the area. Note that vm_fault_wire()
2377 * may release the map lock temporarily, it will be
2378 * relocked on return. The in-transition
2379 * flag protects the entries.
2381 save_start = entry->start;
2382 save_end = entry->end;
2383 rv = vm_fault_wire(map, entry, TRUE, 0);
2385 CLIP_CHECK_BACK(entry, save_start);
2387 KASSERT(entry->wired_count == 1, ("bad wired_count on entry"));
2388 entry->eflags &= ~MAP_ENTRY_USER_WIRED;
2389 entry->wired_count = 0;
2390 if (entry->end == save_end)
2392 entry = vm_map_rb_tree_RB_NEXT(entry);
2394 ("bad entry clip during backout"));
2396 end = save_start; /* unwire the rest */
2400 * note that even though the entry might have been
2401 * clipped, the USER_WIRED flag we set prevents
2402 * duplication so we do not have to do a
2405 entry = vm_map_rb_tree_RB_NEXT(entry);
2409 * If we failed fall through to the unwiring section to
2410 * unwire what we had wired so far. 'end' has already
2417 * start_entry might have been clipped if we unlocked the
2418 * map and blocked. No matter how clipped it has gotten
2419 * there should be a fragment that is on our start boundary.
2421 CLIP_CHECK_BACK(start_entry, start);
2425 * Deal with the unwiring case.
2429 * This is the unwiring case. We must first ensure that the
2430 * range to be unwired is really wired down. We know there
2433 entry = start_entry;
2434 while (entry && entry->start < end) {
2435 if ((entry->eflags & MAP_ENTRY_USER_WIRED) == 0) {
2436 rv = KERN_INVALID_ARGUMENT;
2439 KASSERT(entry->wired_count != 0,
2440 ("wired count was 0 with USER_WIRED set! %p",
2442 entry = vm_map_rb_tree_RB_NEXT(entry);
2446 * Now decrement the wiring count for each region. If a region
2447 * becomes completely unwired, unwire its physical pages and
2451 * The map entries are processed in a loop, checking to
2452 * make sure the entry is wired and asserting it has a wired
2453 * count. However, another loop was inserted more-or-less in
2454 * the middle of the unwiring path. This loop picks up the
2455 * "entry" loop variable from the first loop without first
2456 * setting it to start_entry. Naturally, the secound loop
2457 * is never entered and the pages backing the entries are
2458 * never unwired. This can lead to a leak of wired pages.
2460 entry = start_entry;
2461 while (entry && entry->start < end) {
2462 KASSERT(entry->eflags & MAP_ENTRY_USER_WIRED,
2463 ("expected USER_WIRED on entry %p", entry));
2464 entry->eflags &= ~MAP_ENTRY_USER_WIRED;
2465 entry->wired_count--;
2466 if (entry->wired_count == 0)
2467 vm_fault_unwire(map, entry);
2468 entry = vm_map_rb_tree_RB_NEXT(entry);
2472 vm_map_unclip_range(map, start_entry, start, real_end, &count,
2475 vm_map_entry_release(count);
2481 * Sets the pageability of the specified address range in the target map.
2482 * Regions specified as not pageable require locked-down physical
2483 * memory and physical page maps.
2485 * The map must not be locked, but a reference must remain to the map
2486 * throughout the call.
2488 * This function may be called via the zalloc path and must properly
2489 * reserve map entries for kernel_map.
2494 vm_map_wire(vm_map_t map, vm_offset_t start, vm_offset_t real_end, int kmflags)
2496 vm_map_entry_t entry;
2497 vm_map_entry_t start_entry;
2499 int rv = KERN_SUCCESS;
2502 if (kmflags & KM_KRESERVE)
2503 count = vm_map_entry_kreserve(MAP_RESERVE_COUNT);
2505 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
2507 VM_MAP_RANGE_CHECK(map, start, real_end);
2510 start_entry = vm_map_clip_range(map, start, end, &count,
2512 if (start_entry == NULL) {
2514 rv = KERN_INVALID_ADDRESS;
2517 if ((kmflags & KM_PAGEABLE) == 0) {
2521 * 1. Holding the write lock, we create any shadow or zero-fill
2522 * objects that need to be created. Then we clip each map
2523 * entry to the region to be wired and increment its wiring
2524 * count. We create objects before clipping the map entries
2525 * to avoid object proliferation.
2527 * 2. We downgrade to a read lock, and call vm_fault_wire to
2528 * fault in the pages for any newly wired area (wired_count is
2531 * Downgrading to a read lock for vm_fault_wire avoids a
2532 * possible deadlock with another process that may have faulted
2533 * on one of the pages to be wired (it would mark the page busy,
2534 * blocking us, then in turn block on the map lock that we
2535 * hold). Because of problems in the recursive lock package,
2536 * we cannot upgrade to a write lock in vm_map_lookup. Thus,
2537 * any actions that require the write lock must be done
2538 * beforehand. Because we keep the read lock on the map, the
2539 * copy-on-write status of the entries we modify here cannot
2542 entry = start_entry;
2543 while (entry && entry->start < end) {
2545 * Trivial case if the entry is already wired
2547 if (entry->wired_count) {
2548 entry->wired_count++;
2549 entry = vm_map_rb_tree_RB_NEXT(entry);
2554 * The entry is being newly wired, we have to setup
2555 * appropriate management structures. A shadow
2556 * object is required for a copy-on-write region,
2557 * or a normal object for a zero-fill region. We
2558 * do not have to do this for entries that point to sub
2559 * maps because we won't hold the lock on the sub map.
2561 if (entry->maptype == VM_MAPTYPE_NORMAL ||
2562 entry->maptype == VM_MAPTYPE_VPAGETABLE) {
2563 int copyflag = entry->eflags &
2564 MAP_ENTRY_NEEDS_COPY;
2565 if (copyflag && ((entry->protection &
2566 VM_PROT_WRITE) != 0)) {
2567 vm_map_entry_shadow(entry, 0);
2568 } else if (entry->object.vm_object == NULL &&
2570 vm_map_entry_allocate_object(entry);
2573 entry->wired_count++;
2574 entry = vm_map_rb_tree_RB_NEXT(entry);
2582 * HACK HACK HACK HACK
2584 * vm_fault_wire() temporarily unlocks the map to avoid
2585 * deadlocks. The in-transition flag from vm_map_clip_range
2586 * call should protect us from changes while the map is
2589 * NOTE: Previously this comment stated that clipping might
2590 * still occur while the entry is unlocked, but from
2591 * what I can tell it actually cannot.
2593 * It is unclear whether the CLIP_CHECK_*() calls
2594 * are still needed but we keep them in anyway.
2596 * HACK HACK HACK HACK
2599 entry = start_entry;
2600 while (entry && entry->start < end) {
2602 * If vm_fault_wire fails for any page we need to undo
2603 * what has been done. We decrement the wiring count
2604 * for those pages which have not yet been wired (now)
2605 * and unwire those that have (later).
2607 vm_offset_t save_start = entry->start;
2608 vm_offset_t save_end = entry->end;
2610 if (entry->wired_count == 1)
2611 rv = vm_fault_wire(map, entry, FALSE, kmflags);
2613 CLIP_CHECK_BACK(entry, save_start);
2615 KASSERT(entry->wired_count == 1,
2616 ("wired_count changed unexpectedly"));
2617 entry->wired_count = 0;
2618 if (entry->end == save_end)
2620 entry = vm_map_rb_tree_RB_NEXT(entry);
2622 ("bad entry clip during backout"));
2627 CLIP_CHECK_FWD(entry, save_end);
2628 entry = vm_map_rb_tree_RB_NEXT(entry);
2632 * If a failure occured undo everything by falling through
2633 * to the unwiring code. 'end' has already been adjusted
2637 kmflags |= KM_PAGEABLE;
2640 * start_entry is still IN_TRANSITION but may have been
2641 * clipped since vm_fault_wire() unlocks and relocks the
2642 * map. No matter how clipped it has gotten there should
2643 * be a fragment that is on our start boundary.
2645 CLIP_CHECK_BACK(start_entry, start);
2648 if (kmflags & KM_PAGEABLE) {
2650 * This is the unwiring case. We must first ensure that the
2651 * range to be unwired is really wired down. We know there
2654 entry = start_entry;
2655 while (entry && entry->start < end) {
2656 if (entry->wired_count == 0) {
2657 rv = KERN_INVALID_ARGUMENT;
2660 entry = vm_map_rb_tree_RB_NEXT(entry);
2664 * Now decrement the wiring count for each region. If a region
2665 * becomes completely unwired, unwire its physical pages and
2668 entry = start_entry;
2669 while (entry && entry->start < end) {
2670 entry->wired_count--;
2671 if (entry->wired_count == 0)
2672 vm_fault_unwire(map, entry);
2673 entry = vm_map_rb_tree_RB_NEXT(entry);
2677 vm_map_unclip_range(map, start_entry, start, real_end,
2678 &count, MAP_CLIP_NO_HOLES);
2681 if (kmflags & KM_KRESERVE)
2682 vm_map_entry_krelease(count);
2684 vm_map_entry_release(count);
2689 * Mark a newly allocated address range as wired but do not fault in
2690 * the pages. The caller is expected to load the pages into the object.
2692 * The map must be locked on entry and will remain locked on return.
2693 * No other requirements.
2696 vm_map_set_wired_quick(vm_map_t map, vm_offset_t addr, vm_size_t size,
2699 vm_map_entry_t scan;
2700 vm_map_entry_t entry;
2702 entry = vm_map_clip_range(map, addr, addr + size,
2703 countp, MAP_CLIP_NO_HOLES);
2705 while (scan && scan->start < addr + size) {
2706 KKASSERT(scan->wired_count == 0);
2707 scan->wired_count = 1;
2708 scan = vm_map_rb_tree_RB_NEXT(scan);
2710 vm_map_unclip_range(map, entry, addr, addr + size,
2711 countp, MAP_CLIP_NO_HOLES);
2715 * Push any dirty cached pages in the address range to their pager.
2716 * If syncio is TRUE, dirty pages are written synchronously.
2717 * If invalidate is TRUE, any cached pages are freed as well.
2719 * This routine is called by sys_msync()
2721 * Returns an error if any part of the specified range is not mapped.
2726 vm_map_clean(vm_map_t map, vm_offset_t start, vm_offset_t end,
2727 boolean_t syncio, boolean_t invalidate)
2729 vm_map_entry_t current;
2730 vm_map_entry_t next;
2731 vm_map_entry_t entry;
2735 vm_ooffset_t offset;
2737 vm_map_lock_read(map);
2738 VM_MAP_RANGE_CHECK(map, start, end);
2739 if (!vm_map_lookup_entry(map, start, &entry)) {
2740 vm_map_unlock_read(map);
2741 return (KERN_INVALID_ADDRESS);
2743 lwkt_gettoken(&map->token);
2746 * Make a first pass to check for holes.
2749 while (current && current->start < end) {
2750 if (current->maptype == VM_MAPTYPE_SUBMAP) {
2751 lwkt_reltoken(&map->token);
2752 vm_map_unlock_read(map);
2753 return (KERN_INVALID_ARGUMENT);
2755 next = vm_map_rb_tree_RB_NEXT(current);
2756 if (end > current->end &&
2758 current->end != next->start)) {
2759 lwkt_reltoken(&map->token);
2760 vm_map_unlock_read(map);
2761 return (KERN_INVALID_ADDRESS);
2767 pmap_remove(vm_map_pmap(map), start, end);
2770 * Make a second pass, cleaning/uncaching pages from the indicated
2774 while (current && current->start < end) {
2775 offset = current->offset + (start - current->start);
2776 size = (end <= current->end ? end : current->end) - start;
2778 switch(current->maptype) {
2779 case VM_MAPTYPE_SUBMAP:
2782 vm_map_entry_t tentry;
2785 smap = current->object.sub_map;
2786 vm_map_lock_read(smap);
2787 vm_map_lookup_entry(smap, offset, &tentry);
2788 if (tentry == NULL) {
2789 tsize = vm_map_max(smap) - offset;
2791 offset = 0 + (offset - vm_map_min(smap));
2793 tsize = tentry->end - offset;
2794 object = tentry->object.vm_object;
2795 offset = tentry->offset +
2796 (offset - tentry->start);
2798 vm_map_unlock_read(smap);
2803 case VM_MAPTYPE_NORMAL:
2804 case VM_MAPTYPE_VPAGETABLE:
2805 object = current->object.vm_object;
2813 vm_object_hold(object);
2816 * Note that there is absolutely no sense in writing out
2817 * anonymous objects, so we track down the vnode object
2819 * We invalidate (remove) all pages from the address space
2820 * anyway, for semantic correctness.
2822 * note: certain anonymous maps, such as MAP_NOSYNC maps,
2823 * may start out with a NULL object.
2825 while (object && (tobj = object->backing_object) != NULL) {
2826 vm_object_hold(tobj);
2827 if (tobj == object->backing_object) {
2828 vm_object_lock_swap();
2829 offset += object->backing_object_offset;
2830 vm_object_drop(object);
2832 if (object->size < OFF_TO_IDX(offset + size))
2833 size = IDX_TO_OFF(object->size) -
2837 vm_object_drop(tobj);
2839 if (object && (object->type == OBJT_VNODE) &&
2840 (current->protection & VM_PROT_WRITE) &&
2841 (object->flags & OBJ_NOMSYNC) == 0) {
2843 * Flush pages if writing is allowed, invalidate them
2844 * if invalidation requested. Pages undergoing I/O
2845 * will be ignored by vm_object_page_remove().
2847 * We cannot lock the vnode and then wait for paging
2848 * to complete without deadlocking against vm_fault.
2849 * Instead we simply call vm_object_page_remove() and
2850 * allow it to block internally on a page-by-page
2851 * basis when it encounters pages undergoing async
2856 /* no chain wait needed for vnode objects */
2857 vm_object_reference_locked(object);
2858 vn_lock(object->handle, LK_EXCLUSIVE | LK_RETRY);
2859 flags = (syncio || invalidate) ? OBJPC_SYNC : 0;
2860 flags |= invalidate ? OBJPC_INVAL : 0;
2863 * When operating on a virtual page table just
2864 * flush the whole object. XXX we probably ought
2867 switch(current->maptype) {
2868 case VM_MAPTYPE_NORMAL:
2869 vm_object_page_clean(object,
2871 OFF_TO_IDX(offset + size + PAGE_MASK),
2874 case VM_MAPTYPE_VPAGETABLE:
2875 vm_object_page_clean(object, 0, 0, flags);
2878 vn_unlock(((struct vnode *)object->handle));
2879 vm_object_deallocate_locked(object);
2881 if (object && invalidate &&
2882 ((object->type == OBJT_VNODE) ||
2883 (object->type == OBJT_DEVICE) ||
2884 (object->type == OBJT_MGTDEVICE))) {
2886 ((object->type == OBJT_DEVICE) ||
2887 (object->type == OBJT_MGTDEVICE)) ? FALSE : TRUE;
2888 /* no chain wait needed for vnode/device objects */
2889 vm_object_reference_locked(object);
2890 switch(current->maptype) {
2891 case VM_MAPTYPE_NORMAL:
2892 vm_object_page_remove(object,
2894 OFF_TO_IDX(offset + size + PAGE_MASK),
2897 case VM_MAPTYPE_VPAGETABLE:
2898 vm_object_page_remove(object, 0, 0, clean_only);
2901 vm_object_deallocate_locked(object);
2905 vm_object_drop(object);
2906 current = vm_map_rb_tree_RB_NEXT(current);
2909 lwkt_reltoken(&map->token);
2910 vm_map_unlock_read(map);
2912 return (KERN_SUCCESS);
2916 * Make the region specified by this entry pageable.
2918 * The vm_map must be exclusively locked.
2921 vm_map_entry_unwire(vm_map_t map, vm_map_entry_t entry)
2923 entry->eflags &= ~MAP_ENTRY_USER_WIRED;
2924 entry->wired_count = 0;
2925 vm_fault_unwire(map, entry);
2929 * Deallocate the given entry from the target map.
2931 * The vm_map must be exclusively locked.
2934 vm_map_entry_delete(vm_map_t map, vm_map_entry_t entry, int *countp)
2936 vm_map_entry_unlink(map, entry);
2937 map->size -= entry->end - entry->start;
2939 switch(entry->maptype) {
2940 case VM_MAPTYPE_NORMAL:
2941 case VM_MAPTYPE_VPAGETABLE:
2942 case VM_MAPTYPE_SUBMAP:
2943 vm_object_deallocate(entry->object.vm_object);
2945 case VM_MAPTYPE_UKSMAP:
2952 vm_map_entry_dispose(map, entry, countp);
2956 * Deallocates the given address range from the target map.
2958 * The vm_map must be exclusively locked.
2961 vm_map_delete(vm_map_t map, vm_offset_t start, vm_offset_t end, int *countp)
2964 vm_map_entry_t entry;
2965 vm_map_entry_t first_entry;
2966 vm_offset_t hole_start;
2968 ASSERT_VM_MAP_LOCKED(map);
2969 lwkt_gettoken(&map->token);
2972 * Find the start of the region, and clip it. Set entry to point
2973 * at the first record containing the requested address or, if no
2974 * such record exists, the next record with a greater address. The
2975 * loop will run from this point until a record beyond the termination
2976 * address is encountered.
2978 * Adjust freehint[] for either the clip case or the extension case.
2980 * GGG see other GGG comment.
2982 if (vm_map_lookup_entry(map, start, &first_entry)) {
2983 entry = first_entry;
2984 vm_map_clip_start(map, entry, start, countp);
2988 entry = vm_map_rb_tree_RB_NEXT(first_entry);
2990 hole_start = first_entry->start;
2992 hole_start = first_entry->end;
2994 entry = RB_MIN(vm_map_rb_tree, &map->rb_root);
2996 hole_start = vm_map_min(map);
2998 hole_start = vm_map_max(map);
3003 * Step through all entries in this region
3005 while (entry && entry->start < end) {
3006 vm_map_entry_t next;
3008 vm_pindex_t offidxstart, offidxend, count;
3011 * If we hit an in-transition entry we have to sleep and
3012 * retry. It's easier (and not really slower) to just retry
3013 * since this case occurs so rarely and the hint is already
3014 * pointing at the right place. We have to reset the
3015 * start offset so as not to accidently delete an entry
3016 * another process just created in vacated space.
3018 if (entry->eflags & MAP_ENTRY_IN_TRANSITION) {
3019 entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP;
3020 start = entry->start;
3021 ++mycpu->gd_cnt.v_intrans_coll;
3022 ++mycpu->gd_cnt.v_intrans_wait;
3023 vm_map_transition_wait(map, 1);
3026 vm_map_clip_end(map, entry, end, countp);
3030 next = vm_map_rb_tree_RB_NEXT(entry);
3032 offidxstart = OFF_TO_IDX(entry->offset);
3033 count = OFF_TO_IDX(e - s);
3035 switch(entry->maptype) {
3036 case VM_MAPTYPE_NORMAL:
3037 case VM_MAPTYPE_VPAGETABLE:
3038 case VM_MAPTYPE_SUBMAP:
3039 object = entry->object.vm_object;
3047 * Unwire before removing addresses from the pmap; otherwise,
3048 * unwiring will put the entries back in the pmap.
3050 * Generally speaking, doing a bulk pmap_remove() before
3051 * removing the pages from the VM object is better at
3052 * reducing unnecessary IPIs. The pmap code is now optimized
3053 * to not blindly iterate the range when pt and pd pages
3056 if (entry->wired_count != 0)
3057 vm_map_entry_unwire(map, entry);
3059 offidxend = offidxstart + count;
3061 if (object == &kernel_object) {
3062 pmap_remove(map->pmap, s, e);
3063 vm_object_hold(object);
3064 vm_object_page_remove(object, offidxstart,
3066 vm_object_drop(object);
3067 } else if (object && object->type != OBJT_DEFAULT &&
3068 object->type != OBJT_SWAP) {
3070 * vnode object routines cannot be chain-locked,
3071 * but since we aren't removing pages from the
3072 * object here we can use a shared hold.
3074 vm_object_hold_shared(object);
3075 pmap_remove(map->pmap, s, e);
3076 vm_object_drop(object);
3077 } else if (object) {
3078 vm_object_hold(object);
3079 vm_object_chain_acquire(object, 0);
3080 pmap_remove(map->pmap, s, e);
3082 if (object != NULL &&
3083 object->ref_count != 1 &&
3084 (object->flags & (OBJ_NOSPLIT|OBJ_ONEMAPPING)) ==
3086 (object->type == OBJT_DEFAULT ||
3087 object->type == OBJT_SWAP)) {
3089 * When ONEMAPPING is set we can destroy the
3090 * pages underlying the entry's range.
3092 vm_object_collapse(object, NULL);
3093 vm_object_page_remove(object, offidxstart,
3095 if (object->type == OBJT_SWAP) {
3096 swap_pager_freespace(object,
3100 if (offidxend >= object->size &&
3101 offidxstart < object->size) {
3102 object->size = offidxstart;
3105 vm_object_chain_release(object);
3106 vm_object_drop(object);
3107 } else if (entry->maptype == VM_MAPTYPE_UKSMAP) {
3108 pmap_remove(map->pmap, s, e);
3112 * Delete the entry (which may delete the object) only after
3113 * removing all pmap entries pointing to its pages.
3114 * (Otherwise, its page frames may be reallocated, and any
3115 * modify bits will be set in the wrong object!)
3117 vm_map_entry_delete(map, entry, countp);
3122 * We either reached the end and use vm_map_max as the end
3123 * address, or we didn't and we use the next entry as the
3126 if (entry == NULL) {
3127 vm_map_freehint_hole(map, hole_start,
3128 vm_map_max(map) - hole_start);
3130 vm_map_freehint_hole(map, hole_start,
3131 entry->start - hole_start);
3134 lwkt_reltoken(&map->token);
3136 return (KERN_SUCCESS);
3140 * Remove the given address range from the target map.
3141 * This is the exported form of vm_map_delete.
3146 vm_map_remove(vm_map_t map, vm_offset_t start, vm_offset_t end)
3151 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
3153 VM_MAP_RANGE_CHECK(map, start, end);
3154 result = vm_map_delete(map, start, end, &count);
3156 vm_map_entry_release(count);
3162 * Assert that the target map allows the specified privilege on the
3163 * entire address region given. The entire region must be allocated.
3165 * The caller must specify whether the vm_map is already locked or not.
3168 vm_map_check_protection(vm_map_t map, vm_offset_t start, vm_offset_t end,
3169 vm_prot_t protection, boolean_t have_lock)
3171 vm_map_entry_t entry;
3172 vm_map_entry_t tmp_entry;
3175 if (have_lock == FALSE)
3176 vm_map_lock_read(map);
3178 if (!vm_map_lookup_entry(map, start, &tmp_entry)) {
3179 if (have_lock == FALSE)
3180 vm_map_unlock_read(map);
3186 while (start < end) {
3187 if (entry == NULL) {
3196 if (start < entry->start) {
3201 * Check protection associated with entry.
3204 if ((entry->protection & protection) != protection) {
3208 /* go to next entry */
3210 entry = vm_map_rb_tree_RB_NEXT(entry);
3212 if (have_lock == FALSE)
3213 vm_map_unlock_read(map);
3218 * If appropriate this function shadows the original object with a new object
3219 * and moves the VM pages from the original object to the new object.
3220 * The original object will also be collapsed, if possible.
3222 * Caller must supply entry->object.vm_object held and chain_acquired, and
3223 * should chain_release and drop the object upon return.
3225 * We can only do this for normal memory objects with a single mapping, and
3226 * it only makes sense to do it if there are 2 or more refs on the original
3227 * object. i.e. typically a memory object that has been extended into
3228 * multiple vm_map_entry's with non-overlapping ranges.
3230 * This makes it easier to remove unused pages and keeps object inheritance
3231 * from being a negative impact on memory usage.
3233 * On return the (possibly new) entry->object.vm_object will have an
3234 * additional ref on it for the caller to dispose of (usually by cloning
3235 * the vm_map_entry). The additional ref had to be done in this routine
3236 * to avoid racing a collapse. The object's ONEMAPPING flag will also be
3239 * The vm_map must be locked and its token held.
3242 vm_map_split(vm_map_entry_t entry, vm_object_t oobject)
3245 vm_object_t nobject, bobject;
3248 vm_pindex_t offidxstart, offidxend, idx;
3250 vm_ooffset_t offset;
3254 * Optimize away object locks for vnode objects. Important exit/exec
3257 * OBJ_ONEMAPPING doesn't apply to vnode objects but clear the flag
3260 if (oobject->type != OBJT_DEFAULT && oobject->type != OBJT_SWAP) {
3261 vm_object_reference_quick(oobject);
3262 vm_object_clear_flag(oobject, OBJ_ONEMAPPING);
3268 * Original object cannot be split?
3270 if (oobject->handle == NULL) {
3271 vm_object_reference_locked_chain_held(oobject);
3272 vm_object_clear_flag(oobject, OBJ_ONEMAPPING);
3278 * Collapse original object with its backing store as an
3279 * optimization to reduce chain lengths when possible.
3281 * If ref_count <= 1 there aren't other non-overlapping vm_map_entry's
3282 * for oobject, so there's no point collapsing it.
3284 * Then re-check whether the object can be split.
3286 vm_object_collapse(oobject, NULL);
3288 if (oobject->ref_count <= 1 ||
3289 (oobject->type != OBJT_DEFAULT && oobject->type != OBJT_SWAP) ||
3290 (oobject->flags & (OBJ_NOSPLIT|OBJ_ONEMAPPING)) != OBJ_ONEMAPPING) {
3291 vm_object_reference_locked_chain_held(oobject);
3292 vm_object_clear_flag(oobject, OBJ_ONEMAPPING);
3297 * Acquire the chain lock on the backing object.
3299 * Give bobject an additional ref count for when it will be shadowed
3303 if ((bobject = oobject->backing_object) != NULL) {
3304 if (bobject->type != OBJT_VNODE) {
3306 vm_object_hold(bobject);
3307 vm_object_chain_wait(bobject, 0);
3308 /* ref for shadowing below */
3309 vm_object_reference_locked(bobject);
3310 vm_object_chain_acquire(bobject, 0);
3311 KKASSERT(oobject->backing_object == bobject);
3312 KKASSERT((bobject->flags & OBJ_DEAD) == 0);
3315 * vnodes are not placed on the shadow list but
3316 * they still get another ref for the backing_object
3319 vm_object_reference_quick(bobject);
3324 * Calculate the object page range and allocate the new object.
3326 offset = entry->offset;
3330 offidxstart = OFF_TO_IDX(offset);
3331 offidxend = offidxstart + OFF_TO_IDX(e - s);
3332 size = offidxend - offidxstart;
3334 switch(oobject->type) {
3336 nobject = default_pager_alloc(NULL, IDX_TO_OFF(size),
3340 nobject = swap_pager_alloc(NULL, IDX_TO_OFF(size),
3350 * If we could not allocate nobject just clear ONEMAPPING on
3351 * oobject and return.
3353 if (nobject == NULL) {
3355 if (useshadowlist) {
3356 vm_object_chain_release(bobject);
3357 vm_object_deallocate(bobject);
3358 vm_object_drop(bobject);
3360 vm_object_deallocate(bobject);
3363 vm_object_reference_locked_chain_held(oobject);
3364 vm_object_clear_flag(oobject, OBJ_ONEMAPPING);
3369 * The new object will replace entry->object.vm_object so it needs
3370 * a second reference (the caller expects an additional ref).
3372 vm_object_hold(nobject);
3373 vm_object_reference_locked(nobject);
3374 vm_object_chain_acquire(nobject, 0);
3377 * nobject shadows bobject (oobject already shadows bobject).
3379 * Adding an object to bobject's shadow list requires refing bobject
3380 * which we did above in the useshadowlist case.
3382 * XXX it is unclear if we need to clear ONEMAPPING on bobject here
3386 nobject->backing_object_offset =
3387 oobject->backing_object_offset + IDX_TO_OFF(offidxstart);
3388 nobject->backing_object = bobject;
3389 if (useshadowlist) {
3390 bobject->shadow_count++;
3391 atomic_add_int(&bobject->generation, 1);
3392 LIST_INSERT_HEAD(&bobject->shadow_head,
3393 nobject, shadow_list);
3394 vm_object_clear_flag(bobject, OBJ_ONEMAPPING); /*XXX*/
3395 vm_object_set_flag(nobject, OBJ_ONSHADOW);
3400 * Move the VM pages from oobject to nobject
3402 for (idx = 0; idx < size; idx++) {
3405 m = vm_page_lookup_busy_wait(oobject, offidxstart + idx,
3411 * We must wait for pending I/O to complete before we can
3414 * We do not have to VM_PROT_NONE the page as mappings should
3415 * not be changed by this operation.
3417 * NOTE: The act of renaming a page updates chaingen for both
3420 vm_page_rename(m, nobject, idx);
3421 /* page automatically made dirty by rename and cache handled */
3422 /* page remains busy */
3425 if (oobject->type == OBJT_SWAP) {
3426 vm_object_pip_add(oobject, 1);
3428 * copy oobject pages into nobject and destroy unneeded
3429 * pages in shadow object.
3431 swap_pager_copy(oobject, nobject, offidxstart, 0);
3432 vm_object_pip_wakeup(oobject);
3436 * Wakeup the pages we played with. No spl protection is needed
3437 * for a simple wakeup.
3439 for (idx = 0; idx < size; idx++) {
3440 m = vm_page_lookup(nobject, idx);
3442 KKASSERT(m->busy_count & PBUSY_LOCKED);
3446 entry->object.vm_object = nobject;
3447 entry->offset = 0LL;
3450 * The map is being split and nobject is going to wind up on both
3451 * vm_map_entry's, so make sure OBJ_ONEMAPPING is cleared on
3454 vm_object_clear_flag(nobject, OBJ_ONEMAPPING);
3459 * NOTE: There is no need to remove OBJ_ONEMAPPING from oobject, the
3460 * related pages were moved and are no longer applicable to the
3463 * NOTE: Deallocate oobject (due to its entry->object.vm_object being
3464 * replaced by nobject).
3466 vm_object_chain_release(nobject);
3467 vm_object_drop(nobject);
3468 if (bobject && useshadowlist) {
3469 vm_object_chain_release(bobject);
3470 vm_object_drop(bobject);
3474 if (oobject->resident_page_count) {
3475 kprintf("oobject %p still contains %jd pages!\n",
3476 oobject, (intmax_t)oobject->resident_page_count);
3477 for (idx = 0; idx < size; idx++) {
3480 m = vm_page_lookup_busy_wait(oobject, offidxstart + idx,
3483 kprintf("oobject %p idx %jd\n",
3491 /*vm_object_clear_flag(oobject, OBJ_ONEMAPPING);*/
3492 vm_object_deallocate_locked(oobject);
3496 * Copies the contents of the source entry to the destination
3497 * entry. The entries *must* be aligned properly.
3499 * The vm_maps must be exclusively locked.
3500 * The vm_map's token must be held.
3502 * Because the maps are locked no faults can be in progress during the
3506 vm_map_copy_entry(vm_map_t src_map, vm_map_t dst_map,
3507 vm_map_entry_t src_entry, vm_map_entry_t dst_entry)
3509 vm_object_t src_object;
3510 vm_object_t oobject;
3512 if (dst_entry->maptype == VM_MAPTYPE_SUBMAP ||
3513 dst_entry->maptype == VM_MAPTYPE_UKSMAP)
3515 if (src_entry->maptype == VM_MAPTYPE_SUBMAP ||
3516 src_entry->maptype == VM_MAPTYPE_UKSMAP)
3519 if (src_entry->wired_count == 0) {
3521 * If the source entry is marked needs_copy, it is already
3524 * To avoid interacting with a vm_fault that might have
3525 * released its vm_map, we must acquire the fronting
3528 oobject = src_entry->object.vm_object;
3530 vm_object_hold(oobject);
3531 vm_object_chain_acquire(oobject, 0);
3534 if ((src_entry->eflags & MAP_ENTRY_NEEDS_COPY) == 0) {
3535 pmap_protect(src_map->pmap,
3538 src_entry->protection & ~VM_PROT_WRITE);
3542 * Make a copy of the object.
3544 * The object must be locked prior to checking the object type
3545 * and for the call to vm_object_collapse() and vm_map_split().
3546 * We cannot use *_hold() here because the split code will
3547 * probably try to destroy the object. The lock is a pool
3548 * token and doesn't care.
3550 * We must bump src_map->timestamp when setting
3551 * MAP_ENTRY_NEEDS_COPY to force any concurrent fault
3552 * to retry, otherwise the concurrent fault might improperly
3553 * install a RW pte when its supposed to be a RO(COW) pte.
3554 * This race can occur because a vnode-backed fault may have
3555 * to temporarily release the map lock. This was handled
3556 * when the caller locked the map exclusively.
3559 vm_map_split(src_entry, oobject);
3561 src_object = src_entry->object.vm_object;
3562 dst_entry->object.vm_object = src_object;
3563 src_entry->eflags |= (MAP_ENTRY_COW |
3564 MAP_ENTRY_NEEDS_COPY);
3565 dst_entry->eflags |= (MAP_ENTRY_COW |
3566 MAP_ENTRY_NEEDS_COPY);
3567 dst_entry->offset = src_entry->offset;
3569 dst_entry->object.vm_object = NULL;
3570 dst_entry->offset = 0;
3572 pmap_copy(dst_map->pmap, src_map->pmap, dst_entry->start,
3573 dst_entry->end - dst_entry->start,
3576 vm_object_chain_release(oobject);
3577 vm_object_drop(oobject);
3581 * Of course, wired down pages can't be set copy-on-write.
3582 * Cause wired pages to be copied into the new map by
3583 * simulating faults (the new pages are pageable)
3585 vm_fault_copy_entry(dst_map, src_map, dst_entry, src_entry);
3591 * Create a new process vmspace structure and vm_map
3592 * based on those of an existing process. The new map
3593 * is based on the old map, according to the inheritance
3594 * values on the regions in that map.
3596 * The source map must not be locked.
3599 static void vmspace_fork_normal_entry(vm_map_t old_map, vm_map_t new_map,
3600 vm_map_entry_t old_entry, int *countp);
3601 static void vmspace_fork_uksmap_entry(vm_map_t old_map, vm_map_t new_map,
3602 vm_map_entry_t old_entry, int *countp);
3605 vmspace_fork(struct vmspace *vm1)
3607 struct vmspace *vm2;
3608 vm_map_t old_map = &vm1->vm_map;
3610 vm_map_entry_t old_entry;
3613 lwkt_gettoken(&vm1->vm_map.token);
3614 vm_map_lock(old_map);
3616 vm2 = vmspace_alloc(vm_map_min(old_map), vm_map_max(old_map));
3617 lwkt_gettoken(&vm2->vm_map.token);
3620 * We must bump the timestamp to force any concurrent fault
3623 bcopy(&vm1->vm_startcopy, &vm2->vm_startcopy,
3624 (caddr_t)&vm1->vm_endcopy - (caddr_t)&vm1->vm_startcopy);
3625 new_map = &vm2->vm_map; /* XXX */
3626 new_map->timestamp = 1;
3628 vm_map_lock(new_map);
3630 count = old_map->nentries;
3631 count = vm_map_entry_reserve(count + MAP_RESERVE_COUNT);
3633 RB_FOREACH(old_entry, vm_map_rb_tree, &old_map->rb_root) {
3634 switch(old_entry->maptype) {
3635 case VM_MAPTYPE_SUBMAP:
3636 panic("vm_map_fork: encountered a submap");
3638 case VM_MAPTYPE_UKSMAP:
3639 vmspace_fork_uksmap_entry(old_map, new_map,
3642 case VM_MAPTYPE_NORMAL:
3643 case VM_MAPTYPE_VPAGETABLE:
3644 vmspace_fork_normal_entry(old_map, new_map,
3650 new_map->size = old_map->size;
3651 vm_map_unlock(old_map);
3652 vm_map_unlock(new_map);
3653 vm_map_entry_release(count);
3655 lwkt_reltoken(&vm2->vm_map.token);
3656 lwkt_reltoken(&vm1->vm_map.token);
3663 vmspace_fork_normal_entry(vm_map_t old_map, vm_map_t new_map,
3664 vm_map_entry_t old_entry, int *countp)
3666 vm_map_entry_t new_entry;
3669 switch (old_entry->inheritance) {
3670 case VM_INHERIT_NONE:
3672 case VM_INHERIT_SHARE:
3674 * Clone the entry, creating the shared object if
3677 if (old_entry->object.vm_object == NULL)
3678 vm_map_entry_allocate_object(old_entry);
3680 if (old_entry->eflags & MAP_ENTRY_NEEDS_COPY) {
3682 * Shadow a map_entry which needs a copy,
3683 * replacing its object with a new object
3684 * that points to the old one. Ask the
3685 * shadow code to automatically add an
3686 * additional ref. We can't do it afterwords
3687 * because we might race a collapse. The call
3688 * to vm_map_entry_shadow() will also clear
3691 vm_map_entry_shadow(old_entry, 1);
3692 } else if (old_entry->object.vm_object) {
3694 * We will make a shared copy of the object,
3695 * and must clear OBJ_ONEMAPPING.
3697 * Optimize vnode objects. OBJ_ONEMAPPING
3698 * is non-applicable but clear it anyway,
3699 * and its terminal so we don't have to deal
3700 * with chains. Reduces SMP conflicts.
3702 * XXX assert that object.vm_object != NULL
3703 * since we allocate it above.
3705 object = old_entry->object.vm_object;
3706 if (object->type == OBJT_VNODE) {
3707 vm_object_reference_quick(object);
3708 vm_object_clear_flag(object,
3711 vm_object_hold(object);
3712 vm_object_chain_wait(object, 0);
3713 vm_object_reference_locked(object);
3714 vm_object_clear_flag(object,
3716 vm_object_drop(object);
3721 * Clone the entry. We've already bumped the ref on
3724 new_entry = vm_map_entry_create(new_map, countp);
3725 *new_entry = *old_entry;
3726 new_entry->eflags &= ~MAP_ENTRY_USER_WIRED;
3727 new_entry->wired_count = 0;
3730 * Insert the entry into the new map -- we know we're
3731 * inserting at the end of the new map.
3733 vm_map_entry_link(new_map, new_entry);
3736 * Update the physical map
3738 pmap_copy(new_map->pmap, old_map->pmap,
3740 (old_entry->end - old_entry->start),
3743 case VM_INHERIT_COPY:
3745 * Clone the entry and link into the map.
3747 new_entry = vm_map_entry_create(new_map, countp);
3748 *new_entry = *old_entry;
3749 new_entry->eflags &= ~MAP_ENTRY_USER_WIRED;
3750 new_entry->wired_count = 0;
3751 new_entry->object.vm_object = NULL;
3752 vm_map_entry_link(new_map, new_entry);
3753 vm_map_copy_entry(old_map, new_map, old_entry,
3760 * When forking user-kernel shared maps, the map might change in the
3761 * child so do not try to copy the underlying pmap entries.
3765 vmspace_fork_uksmap_entry(vm_map_t old_map, vm_map_t new_map,
3766 vm_map_entry_t old_entry, int *countp)
3768 vm_map_entry_t new_entry;
3770 new_entry = vm_map_entry_create(new_map, countp);
3771 *new_entry = *old_entry;
3772 new_entry->eflags &= ~MAP_ENTRY_USER_WIRED;
3773 new_entry->wired_count = 0;
3774 vm_map_entry_link(new_map, new_entry);
3778 * Create an auto-grow stack entry
3783 vm_map_stack (vm_map_t map, vm_offset_t addrbos, vm_size_t max_ssize,
3784 int flags, vm_prot_t prot, vm_prot_t max, int cow)
3786 vm_map_entry_t prev_entry;
3787 vm_map_entry_t next;
3788 vm_size_t init_ssize;
3791 vm_offset_t tmpaddr;
3793 cow |= MAP_IS_STACK;
3795 if (max_ssize < sgrowsiz)
3796 init_ssize = max_ssize;
3798 init_ssize = sgrowsiz;
3800 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
3804 * Find space for the mapping
3806 if ((flags & (MAP_FIXED | MAP_TRYFIXED)) == 0) {
3807 if (vm_map_findspace(map, addrbos, max_ssize, 1,
3810 vm_map_entry_release(count);
3811 return (KERN_NO_SPACE);
3816 /* If addr is already mapped, no go */
3817 if (vm_map_lookup_entry(map, addrbos, &prev_entry)) {
3819 vm_map_entry_release(count);
3820 return (KERN_NO_SPACE);
3824 /* XXX already handled by kern_mmap() */
3825 /* If we would blow our VMEM resource limit, no go */
3826 if (map->size + init_ssize >
3827 curproc->p_rlimit[RLIMIT_VMEM].rlim_cur) {
3829 vm_map_entry_release(count);
3830 return (KERN_NO_SPACE);
3835 * If we can't accomodate max_ssize in the current mapping,
3836 * no go. However, we need to be aware that subsequent user
3837 * mappings might map into the space we have reserved for
3838 * stack, and currently this space is not protected.
3840 * Hopefully we will at least detect this condition
3841 * when we try to grow the stack.
3844 next = vm_map_rb_tree_RB_NEXT(prev_entry);
3846 next = RB_MIN(vm_map_rb_tree, &map->rb_root);
3848 if (next && next->start < addrbos + max_ssize) {
3850 vm_map_entry_release(count);
3851 return (KERN_NO_SPACE);
3855 * We initially map a stack of only init_ssize. We will
3856 * grow as needed later. Since this is to be a grow
3857 * down stack, we map at the top of the range.
3859 * Note: we would normally expect prot and max to be
3860 * VM_PROT_ALL, and cow to be 0. Possibly we should
3861 * eliminate these as input parameters, and just
3862 * pass these values here in the insert call.
3864 rv = vm_map_insert(map, &count, NULL, NULL,
3865 0, addrbos + max_ssize - init_ssize,
3866 addrbos + max_ssize,
3868 VM_SUBSYS_STACK, prot, max, cow);
3870 /* Now set the avail_ssize amount */
3871 if (rv == KERN_SUCCESS) {
3873 next = vm_map_rb_tree_RB_NEXT(prev_entry);
3875 next = RB_MIN(vm_map_rb_tree, &map->rb_root);
3876 if (prev_entry != NULL) {
3877 vm_map_clip_end(map,
3879 addrbos + max_ssize - init_ssize,
3882 if (next->end != addrbos + max_ssize ||
3883 next->start != addrbos + max_ssize - init_ssize){
3884 panic ("Bad entry start/end for new stack entry");
3886 next->aux.avail_ssize = max_ssize - init_ssize;
3891 vm_map_entry_release(count);
3896 * Attempts to grow a vm stack entry. Returns KERN_SUCCESS if the
3897 * desired address is already mapped, or if we successfully grow
3898 * the stack. Also returns KERN_SUCCESS if addr is outside the
3899 * stack range (this is strange, but preserves compatibility with
3900 * the grow function in vm_machdep.c).
3905 vm_map_growstack (vm_map_t map, vm_offset_t addr)
3907 vm_map_entry_t prev_entry;
3908 vm_map_entry_t stack_entry;
3909 vm_map_entry_t next;
3915 int rv = KERN_SUCCESS;
3917 int use_read_lock = 1;
3923 lp = curthread->td_lwp;
3924 p = curthread->td_proc;
3925 KKASSERT(lp != NULL);
3926 vm = lp->lwp_vmspace;
3929 * Growstack is only allowed on the current process. We disallow
3930 * other use cases, e.g. trying to access memory via procfs that
3931 * the stack hasn't grown into.
3933 if (map != &vm->vm_map) {
3934 return KERN_FAILURE;
3937 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
3940 vm_map_lock_read(map);
3945 * If addr is already in the entry range, no need to grow.
3946 * prev_entry returns NULL if addr is at the head.
3948 if (vm_map_lookup_entry(map, addr, &prev_entry))
3951 stack_entry = vm_map_rb_tree_RB_NEXT(prev_entry);
3953 stack_entry = RB_MIN(vm_map_rb_tree, &map->rb_root);
3955 if (stack_entry == NULL)
3957 if (prev_entry == NULL)
3958 end = stack_entry->start - stack_entry->aux.avail_ssize;
3960 end = prev_entry->end;
3963 * This next test mimics the old grow function in vm_machdep.c.
3964 * It really doesn't quite make sense, but we do it anyway
3965 * for compatibility.
3967 * If not growable stack, return success. This signals the
3968 * caller to proceed as he would normally with normal vm.
3970 if (stack_entry->aux.avail_ssize < 1 ||
3971 addr >= stack_entry->start ||
3972 addr < stack_entry->start - stack_entry->aux.avail_ssize) {
3976 /* Find the minimum grow amount */
3977 grow_amount = roundup (stack_entry->start - addr, PAGE_SIZE);
3978 if (grow_amount > stack_entry->aux.avail_ssize) {
3984 * If there is no longer enough space between the entries
3985 * nogo, and adjust the available space. Note: this
3986 * should only happen if the user has mapped into the
3987 * stack area after the stack was created, and is
3988 * probably an error.
3990 * This also effectively destroys any guard page the user
3991 * might have intended by limiting the stack size.
3993 if (grow_amount > stack_entry->start - end) {
3994 if (use_read_lock && vm_map_lock_upgrade(map)) {
4000 stack_entry->aux.avail_ssize = stack_entry->start - end;
4005 is_procstack = addr >= (vm_offset_t)vm->vm_maxsaddr;
4007 /* If this is the main process stack, see if we're over the
4010 if (is_procstack && (vm->vm_ssize + grow_amount >
4011 p->p_rlimit[RLIMIT_STACK].rlim_cur)) {
4016 /* Round up the grow amount modulo SGROWSIZ */
4017 grow_amount = roundup (grow_amount, sgrowsiz);
4018 if (grow_amount > stack_entry->aux.avail_ssize) {
4019 grow_amount = stack_entry->aux.avail_ssize;
4021 if (is_procstack && (vm->vm_ssize + grow_amount >
4022 p->p_rlimit[RLIMIT_STACK].rlim_cur)) {
4023 grow_amount = p->p_rlimit[RLIMIT_STACK].rlim_cur - vm->vm_ssize;
4026 /* If we would blow our VMEM resource limit, no go */
4027 if (map->size + grow_amount > p->p_rlimit[RLIMIT_VMEM].rlim_cur) {
4032 if (use_read_lock && vm_map_lock_upgrade(map)) {
4039 /* Get the preliminary new entry start value */
4040 addr = stack_entry->start - grow_amount;
4042 /* If this puts us into the previous entry, cut back our growth
4043 * to the available space. Also, see the note above.
4046 stack_entry->aux.avail_ssize = stack_entry->start - end;
4050 rv = vm_map_insert(map, &count, NULL, NULL,
4051 0, addr, stack_entry->start,
4053 VM_SUBSYS_STACK, VM_PROT_ALL, VM_PROT_ALL, 0);
4055 /* Adjust the available stack space by the amount we grew. */
4056 if (rv == KERN_SUCCESS) {
4058 vm_map_clip_end(map, prev_entry, addr, &count);
4059 next = vm_map_rb_tree_RB_NEXT(prev_entry);
4061 next = RB_MIN(vm_map_rb_tree, &map->rb_root);
4063 if (next->end != stack_entry->start ||
4064 next->start != addr) {
4065 panic ("Bad stack grow start/end in new stack entry");
4067 next->aux.avail_ssize =
4068 stack_entry->aux.avail_ssize -
4069 (next->end - next->start);
4071 vm->vm_ssize += next->end -
4076 if (map->flags & MAP_WIREFUTURE)
4077 vm_map_unwire(map, next->start, next->end, FALSE);
4082 vm_map_unlock_read(map);
4085 vm_map_entry_release(count);
4090 * Unshare the specified VM space for exec. If other processes are
4091 * mapped to it, then create a new one. The new vmspace is null.
4096 vmspace_exec(struct proc *p, struct vmspace *vmcopy)
4098 struct vmspace *oldvmspace = p->p_vmspace;
4099 struct vmspace *newvmspace;
4100 vm_map_t map = &p->p_vmspace->vm_map;
4103 * If we are execing a resident vmspace we fork it, otherwise
4104 * we create a new vmspace. Note that exitingcnt is not
4105 * copied to the new vmspace.
4107 lwkt_gettoken(&oldvmspace->vm_map.token);
4109 newvmspace = vmspace_fork(vmcopy);
4110 lwkt_gettoken(&newvmspace->vm_map.token);
4112 newvmspace = vmspace_alloc(vm_map_min(map), vm_map_max(map));
4113 lwkt_gettoken(&newvmspace->vm_map.token);
4114 bcopy(&oldvmspace->vm_startcopy, &newvmspace->vm_startcopy,
4115 (caddr_t)&oldvmspace->vm_endcopy -
4116 (caddr_t)&oldvmspace->vm_startcopy);
4120 * Finish initializing the vmspace before assigning it
4121 * to the process. The vmspace will become the current vmspace
4124 pmap_pinit2(vmspace_pmap(newvmspace));
4125 pmap_replacevm(p, newvmspace, 0);
4126 lwkt_reltoken(&newvmspace->vm_map.token);
4127 lwkt_reltoken(&oldvmspace->vm_map.token);
4128 vmspace_rel(oldvmspace);
4132 * Unshare the specified VM space for forcing COW. This
4133 * is called by rfork, for the (RFMEM|RFPROC) == 0 case.
4136 vmspace_unshare(struct proc *p)
4138 struct vmspace *oldvmspace = p->p_vmspace;
4139 struct vmspace *newvmspace;
4141 lwkt_gettoken(&oldvmspace->vm_map.token);
4142 if (vmspace_getrefs(oldvmspace) == 1) {
4143 lwkt_reltoken(&oldvmspace->vm_map.token);
4146 newvmspace = vmspace_fork(oldvmspace);
4147 lwkt_gettoken(&newvmspace->vm_map.token);
4148 pmap_pinit2(vmspace_pmap(newvmspace));
4149 pmap_replacevm(p, newvmspace, 0);
4150 lwkt_reltoken(&newvmspace->vm_map.token);
4151 lwkt_reltoken(&oldvmspace->vm_map.token);
4152 vmspace_rel(oldvmspace);
4156 * vm_map_hint: return the beginning of the best area suitable for
4157 * creating a new mapping with "prot" protection.
4162 vm_map_hint(struct proc *p, vm_offset_t addr, vm_prot_t prot)
4164 struct vmspace *vms = p->p_vmspace;
4165 struct rlimit limit;
4169 * Acquire datasize limit for mmap() operation,
4170 * calculate nearest power of 2.
4172 if (kern_getrlimit(RLIMIT_DATA, &limit))
4173 limit.rlim_cur = maxdsiz;
4174 dsiz = limit.rlim_cur;
4176 if (!randomize_mmap || addr != 0) {
4178 * Set a reasonable start point for the hint if it was
4179 * not specified or if it falls within the heap space.
4180 * Hinted mmap()s do not allocate out of the heap space.
4183 (addr >= round_page((vm_offset_t)vms->vm_taddr) &&
4184 addr < round_page((vm_offset_t)vms->vm_daddr + dsiz))) {
4185 addr = round_page((vm_offset_t)vms->vm_daddr + dsiz);
4192 * randomize_mmap && addr == 0. For now randomize the
4193 * address within a dsiz range beyond the data limit.
4195 addr = (vm_offset_t)vms->vm_daddr + dsiz;
4197 addr += (karc4random64() & 0x7FFFFFFFFFFFFFFFLU) % dsiz;
4198 return (round_page(addr));
4202 * Finds the VM object, offset, and protection for a given virtual address
4203 * in the specified map, assuming a page fault of the type specified.
4205 * Leaves the map in question locked for read; return values are guaranteed
4206 * until a vm_map_lookup_done call is performed. Note that the map argument
4207 * is in/out; the returned map must be used in the call to vm_map_lookup_done.
4209 * A handle (out_entry) is returned for use in vm_map_lookup_done, to make
4212 * If a lookup is requested with "write protection" specified, the map may
4213 * be changed to perform virtual copying operations, although the data
4214 * referenced will remain the same.
4219 vm_map_lookup(vm_map_t *var_map, /* IN/OUT */
4221 vm_prot_t fault_typea,
4222 vm_map_entry_t *out_entry, /* OUT */
4223 vm_object_t *object, /* OUT */
4224 vm_pindex_t *pindex, /* OUT */
4225 vm_prot_t *out_prot, /* OUT */
4226 int *wflags) /* OUT */
4228 vm_map_entry_t entry;
4229 vm_map_t map = *var_map;
4231 vm_prot_t fault_type = fault_typea;
4232 int use_read_lock = 1;
4233 int rv = KERN_SUCCESS;
4235 thread_t td = curthread;
4238 * vm_map_entry_reserve() implements an important mitigation
4239 * against mmap() span running the kernel out of vm_map_entry
4240 * structures, but it can also cause an infinite call recursion.
4241 * Use td_nest_count to prevent an infinite recursion (allows
4242 * the vm_map code to dig into the pcpu vm_map_entry reserve).
4245 if (td->td_nest_count == 0) {
4246 ++td->td_nest_count;
4247 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
4248 --td->td_nest_count;
4252 vm_map_lock_read(map);
4257 * Always do a full lookup. The hint doesn't get us much anymore
4258 * now that the map is RB'd.
4265 vm_map_entry_t tmp_entry;
4267 if (!vm_map_lookup_entry(map, vaddr, &tmp_entry)) {
4268 rv = KERN_INVALID_ADDRESS;
4278 if (entry->maptype == VM_MAPTYPE_SUBMAP) {
4279 vm_map_t old_map = map;
4281 *var_map = map = entry->object.sub_map;
4283 vm_map_unlock_read(old_map);
4285 vm_map_unlock(old_map);
4291 * Check whether this task is allowed to have this page.
4292 * Note the special case for MAP_ENTRY_COW pages with an override.
4293 * This is to implement a forced COW for debuggers.
4295 if (fault_type & VM_PROT_OVERRIDE_WRITE)
4296 prot = entry->max_protection;
4298 prot = entry->protection;
4300 fault_type &= (VM_PROT_READ|VM_PROT_WRITE|VM_PROT_EXECUTE);
4301 if ((fault_type & prot) != fault_type) {
4302 rv = KERN_PROTECTION_FAILURE;
4306 if ((entry->eflags & MAP_ENTRY_USER_WIRED) &&
4307 (entry->eflags & MAP_ENTRY_COW) &&
4308 (fault_type & VM_PROT_WRITE) &&
4309 (fault_typea & VM_PROT_OVERRIDE_WRITE) == 0) {
4310 rv = KERN_PROTECTION_FAILURE;
4315 * If this page is not pageable, we have to get it for all possible
4319 if (entry->wired_count) {
4320 *wflags |= FW_WIRED;
4321 prot = fault_type = entry->protection;
4325 * Virtual page tables may need to update the accessed (A) bit
4326 * in a page table entry. Upgrade the fault to a write fault for
4327 * that case if the map will support it. If the map does not support
4328 * it the page table entry simply will not be updated.
4330 if (entry->maptype == VM_MAPTYPE_VPAGETABLE) {
4331 if (prot & VM_PROT_WRITE)
4332 fault_type |= VM_PROT_WRITE;
4335 if (curthread->td_lwp && curthread->td_lwp->lwp_vmspace &&
4336 pmap_emulate_ad_bits(&curthread->td_lwp->lwp_vmspace->vm_pmap)) {
4337 if ((prot & VM_PROT_WRITE) == 0)
4338 fault_type |= VM_PROT_WRITE;
4342 * Only NORMAL and VPAGETABLE maps are object-based. UKSMAPs are not.
4344 if (entry->maptype != VM_MAPTYPE_NORMAL &&
4345 entry->maptype != VM_MAPTYPE_VPAGETABLE) {
4351 * If the entry was copy-on-write, we either ...
4353 if (entry->eflags & MAP_ENTRY_NEEDS_COPY) {
4355 * If we want to write the page, we may as well handle that
4356 * now since we've got the map locked.
4358 * If we don't need to write the page, we just demote the
4359 * permissions allowed.
4361 if (fault_type & VM_PROT_WRITE) {
4363 * Not allowed if TDF_NOFAULT is set as the shadowing
4364 * operation can deadlock against the faulting
4365 * function due to the copy-on-write.
4367 if (curthread->td_flags & TDF_NOFAULT) {
4368 rv = KERN_FAILURE_NOFAULT;
4373 * Make a new object, and place it in the object
4374 * chain. Note that no new references have appeared
4375 * -- one just moved from the map to the new
4378 if (use_read_lock && vm_map_lock_upgrade(map)) {
4384 vm_map_entry_shadow(entry, 0);
4385 *wflags |= FW_DIDCOW;
4388 * We're attempting to read a copy-on-write page --
4389 * don't allow writes.
4391 prot &= ~VM_PROT_WRITE;
4396 * Create an object if necessary. This code also handles
4397 * partitioning large entries to improve vm_fault performance.
4399 if (entry->object.vm_object == NULL && !map->system_map) {
4400 if (use_read_lock && vm_map_lock_upgrade(map)) {
4408 * Partition large entries, giving each its own VM object,
4409 * to improve concurrent fault performance. This is only
4410 * applicable to userspace.
4412 if (map != &kernel_map &&
4413 entry->maptype == VM_MAPTYPE_NORMAL &&
4414 ((entry->start ^ entry->end) & ~MAP_ENTRY_PARTITION_MASK) &&
4415 vm_map_partition_enable) {
4416 if (entry->eflags & MAP_ENTRY_IN_TRANSITION) {
4417 entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP;
4418 ++mycpu->gd_cnt.v_intrans_coll;
4419 ++mycpu->gd_cnt.v_intrans_wait;
4420 vm_map_transition_wait(map, 0);
4423 vm_map_entry_partition(map, entry, vaddr, &count);
4425 vm_map_entry_allocate_object(entry);
4429 * Return the object/offset from this entry. If the entry was
4430 * copy-on-write or empty, it has been fixed up.
4432 *object = entry->object.vm_object;
4435 *pindex = OFF_TO_IDX((vaddr - entry->start) + entry->offset);
4438 * Return whether this is the only map sharing this data. On
4439 * success we return with a read lock held on the map. On failure
4440 * we return with the map unlocked.
4444 if (rv == KERN_SUCCESS) {
4445 if (use_read_lock == 0)
4446 vm_map_lock_downgrade(map);
4447 } else if (use_read_lock) {
4448 vm_map_unlock_read(map);
4453 vm_map_entry_release(count);
4459 * Releases locks acquired by a vm_map_lookup()
4460 * (according to the handle returned by that lookup).
4462 * No other requirements.
4465 vm_map_lookup_done(vm_map_t map, vm_map_entry_t entry, int count)
4468 * Unlock the main-level map
4470 vm_map_unlock_read(map);
4472 vm_map_entry_release(count);
4476 vm_map_entry_partition(vm_map_t map, vm_map_entry_t entry,
4477 vm_offset_t vaddr, int *countp)
4479 vaddr &= ~MAP_ENTRY_PARTITION_MASK;
4480 vm_map_clip_start(map, entry, vaddr, countp);
4481 vaddr += MAP_ENTRY_PARTITION_SIZE;
4482 vm_map_clip_end(map, entry, vaddr, countp);
4486 * Quick hack, needs some help to make it more SMP friendly.
4489 vm_map_interlock(vm_map_t map, struct vm_map_ilock *ilock,
4490 vm_offset_t ran_beg, vm_offset_t ran_end)
4492 struct vm_map_ilock *scan;
4494 ilock->ran_beg = ran_beg;
4495 ilock->ran_end = ran_end;
4498 spin_lock(&map->ilock_spin);
4500 for (scan = map->ilock_base; scan; scan = scan->next) {
4501 if (ran_end > scan->ran_beg && ran_beg < scan->ran_end) {
4502 scan->flags |= ILOCK_WAITING;
4503 ssleep(scan, &map->ilock_spin, 0, "ilock", 0);
4507 ilock->next = map->ilock_base;
4508 map->ilock_base = ilock;
4509 spin_unlock(&map->ilock_spin);
4513 vm_map_deinterlock(vm_map_t map, struct vm_map_ilock *ilock)
4515 struct vm_map_ilock *scan;
4516 struct vm_map_ilock **scanp;
4518 spin_lock(&map->ilock_spin);
4519 scanp = &map->ilock_base;
4520 while ((scan = *scanp) != NULL) {
4521 if (scan == ilock) {
4522 *scanp = ilock->next;
4523 spin_unlock(&map->ilock_spin);
4524 if (ilock->flags & ILOCK_WAITING)
4528 scanp = &scan->next;
4530 spin_unlock(&map->ilock_spin);
4531 panic("vm_map_deinterlock: missing ilock!");
4534 #include "opt_ddb.h"
4536 #include <ddb/ddb.h>
4541 DB_SHOW_COMMAND(map, vm_map_print)
4544 /* XXX convert args. */
4545 vm_map_t map = (vm_map_t)addr;
4546 boolean_t full = have_addr;
4548 vm_map_entry_t entry;
4550 db_iprintf("Task map %p: pmap=%p, nentries=%d, version=%u\n",
4552 (void *)map->pmap, map->nentries, map->timestamp);
4555 if (!full && db_indent)
4559 RB_FOREACH(entry, vm_map_rb_tree, &map->rb_root) {
4560 db_iprintf("map entry %p: start=%p, end=%p\n",
4561 (void *)entry, (void *)entry->start, (void *)entry->end);
4564 static char *inheritance_name[4] =
4565 {"share", "copy", "none", "donate_copy"};
4567 db_iprintf(" prot=%x/%x/%s",
4569 entry->max_protection,
4570 inheritance_name[(int)(unsigned char)
4571 entry->inheritance]);
4572 if (entry->wired_count != 0)
4573 db_printf(", wired");
4575 switch(entry->maptype) {
4576 case VM_MAPTYPE_SUBMAP:
4577 /* XXX no %qd in kernel. Truncate entry->offset. */
4578 db_printf(", share=%p, offset=0x%lx\n",
4579 (void *)entry->object.sub_map,
4580 (long)entry->offset);
4584 vm_map_print((db_expr_t)(intptr_t)
4585 entry->object.sub_map,
4589 case VM_MAPTYPE_NORMAL:
4590 case VM_MAPTYPE_VPAGETABLE:
4591 /* XXX no %qd in kernel. Truncate entry->offset. */
4592 db_printf(", object=%p, offset=0x%lx",
4593 (void *)entry->object.vm_object,
4594 (long)entry->offset);
4595 if (entry->eflags & MAP_ENTRY_COW)
4596 db_printf(", copy (%s)",
4597 (entry->eflags & MAP_ENTRY_NEEDS_COPY) ? "needed" : "done");
4601 if (entry->object.vm_object) {
4603 vm_object_print((db_expr_t)(intptr_t)
4604 entry->object.vm_object,
4610 case VM_MAPTYPE_UKSMAP:
4611 db_printf(", uksmap=%p, offset=0x%lx",
4612 (void *)entry->object.uksmap,
4613 (long)entry->offset);
4614 if (entry->eflags & MAP_ENTRY_COW)
4615 db_printf(", copy (%s)",
4616 (entry->eflags & MAP_ENTRY_NEEDS_COPY) ? "needed" : "done");
4632 DB_SHOW_COMMAND(procvm, procvm)
4637 p = (struct proc *) addr;
4642 db_printf("p = %p, vmspace = %p, map = %p, pmap = %p\n",
4643 (void *)p, (void *)p->p_vmspace, (void *)&p->p_vmspace->vm_map,
4644 (void *)vmspace_pmap(p->p_vmspace));
4646 vm_map_print((db_expr_t)(intptr_t)&p->p_vmspace->vm_map, 1, 0, NULL);