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, vm->vm_map.header.start,
515 vm->vm_map.header.end, &count);
516 vm_map_unlock(&vm->vm_map);
517 vm_map_entry_release(count);
519 pmap_release(vmspace_pmap(vm));
520 lwkt_reltoken(&vm->vm_map.token);
521 objcache_put(vmspace_cache, vm);
526 * Swap useage is determined by taking the proportional swap used by
527 * VM objects backing the VM map. To make up for fractional losses,
528 * if the VM object has any swap use at all the associated map entries
529 * count for at least 1 swap page.
534 vmspace_swap_count(struct vmspace *vm)
536 vm_map_t map = &vm->vm_map;
539 vm_offset_t count = 0;
543 for (cur = map->header.next; cur != &map->header; cur = cur->next) {
544 switch(cur->maptype) {
545 case VM_MAPTYPE_NORMAL:
546 case VM_MAPTYPE_VPAGETABLE:
547 if ((object = cur->object.vm_object) == NULL)
549 if (object->swblock_count) {
550 n = (cur->end - cur->start) / PAGE_SIZE;
551 count += object->swblock_count *
552 SWAP_META_PAGES * n / object->size + 1;
565 * Calculate the approximate number of anonymous pages in use by
566 * this vmspace. To make up for fractional losses, we count each
567 * VM object as having at least 1 anonymous page.
572 vmspace_anonymous_count(struct vmspace *vm)
574 vm_map_t map = &vm->vm_map;
577 vm_offset_t count = 0;
580 for (cur = map->header.next; cur != &map->header; cur = cur->next) {
581 switch(cur->maptype) {
582 case VM_MAPTYPE_NORMAL:
583 case VM_MAPTYPE_VPAGETABLE:
584 if ((object = cur->object.vm_object) == NULL)
586 if (object->type != OBJT_DEFAULT &&
587 object->type != OBJT_SWAP) {
590 count += object->resident_page_count;
602 * Initialize an existing vm_map structure such as that in the vmspace
603 * structure. The pmap is initialized elsewhere.
608 vm_map_init(struct vm_map *map, vm_offset_t min, vm_offset_t max, pmap_t pmap)
610 map->header.next = map->header.prev = &map->header;
611 RB_INIT(&map->rb_root);
612 spin_init(&map->ilock_spin, "ilock");
613 map->ilock_base = NULL;
617 map->header.start = min;
618 map->header.end = max;
622 bzero(&map->freehint, sizeof(map->freehint));
623 lwkt_token_init(&map->token, "vm_map");
624 lockinit(&map->lock, "vm_maplk", (hz + 9) / 10, 0);
628 * Find the first possible free address for the specified request length.
629 * Returns 0 if we don't have one cached.
633 vm_map_freehint_find(vm_map_t map, vm_size_t length, vm_size_t align)
635 vm_map_freehint_t *scan;
637 scan = &map->freehint[0];
638 while (scan < &map->freehint[VM_MAP_FFCOUNT]) {
639 if (scan->length == length && scan->align == align)
647 * Unconditionally set the freehint. Called by vm_map_findspace() after
648 * it finds an address. This will help us iterate optimally on the next
653 vm_map_freehint_update(vm_map_t map, vm_offset_t start,
654 vm_size_t length, vm_size_t align)
656 vm_map_freehint_t *scan;
658 scan = &map->freehint[0];
659 while (scan < &map->freehint[VM_MAP_FFCOUNT]) {
660 if (scan->length == length && scan->align == align) {
666 scan = &map->freehint[map->freehint_newindex & VM_MAP_FFMASK];
669 scan->length = length;
670 ++map->freehint_newindex;
674 * Update any existing freehints (for any alignment), for the hole we just
679 vm_map_freehint_hole(vm_map_t map, vm_offset_t start, vm_size_t length)
681 vm_map_freehint_t *scan;
683 scan = &map->freehint[0];
684 while (scan < &map->freehint[VM_MAP_FFCOUNT]) {
685 if (scan->length <= length && scan->start > start)
692 * Shadow the vm_map_entry's object. This typically needs to be done when
693 * a write fault is taken on an entry which had previously been cloned by
694 * fork(). The shared object (which might be NULL) must become private so
695 * we add a shadow layer above it.
697 * Object allocation for anonymous mappings is defered as long as possible.
698 * When creating a shadow, however, the underlying object must be instantiated
699 * so it can be shared.
701 * If the map segment is governed by a virtual page table then it is
702 * possible to address offsets beyond the mapped area. Just allocate
703 * a maximally sized object for this case.
705 * If addref is non-zero an additional reference is added to the returned
706 * entry. This mechanic exists because the additional reference might have
707 * to be added atomically and not after return to prevent a premature
710 * The vm_map must be exclusively locked.
711 * No other requirements.
715 vm_map_entry_shadow(vm_map_entry_t entry, int addref)
717 if (entry->maptype == VM_MAPTYPE_VPAGETABLE) {
718 vm_object_shadow(&entry->object.vm_object, &entry->offset,
719 0x7FFFFFFF, addref); /* XXX */
721 vm_object_shadow(&entry->object.vm_object, &entry->offset,
722 atop(entry->end - entry->start), addref);
724 entry->eflags &= ~MAP_ENTRY_NEEDS_COPY;
728 * Allocate an object for a vm_map_entry.
730 * Object allocation for anonymous mappings is defered as long as possible.
731 * This function is called when we can defer no longer, generally when a map
732 * entry might be split or forked or takes a page fault.
734 * If the map segment is governed by a virtual page table then it is
735 * possible to address offsets beyond the mapped area. Just allocate
736 * a maximally sized object for this case.
738 * The vm_map must be exclusively locked.
739 * No other requirements.
742 vm_map_entry_allocate_object(vm_map_entry_t entry)
746 if (entry->maptype == VM_MAPTYPE_VPAGETABLE) {
747 obj = vm_object_allocate(OBJT_DEFAULT, 0x7FFFFFFF); /* XXX */
749 obj = vm_object_allocate(OBJT_DEFAULT,
750 atop(entry->end - entry->start));
752 entry->object.vm_object = obj;
757 * Set an initial negative count so the first attempt to reserve
758 * space preloads a bunch of vm_map_entry's for this cpu. Also
759 * pre-allocate 2 vm_map_entries which will be needed by zalloc() to
760 * map a new page for vm_map_entry structures. SMP systems are
761 * particularly sensitive.
763 * This routine is called in early boot so we cannot just call
764 * vm_map_entry_reserve().
766 * Called from the low level boot code only (for each cpu)
768 * WARNING! Take care not to have too-big a static/BSS structure here
769 * as MAXCPU can be 256+, otherwise the loader's 64MB heap
770 * can get blown out by the kernel plus the initrd image.
773 vm_map_entry_reserve_cpu_init(globaldata_t gd)
775 vm_map_entry_t entry;
779 atomic_add_int(&gd->gd_vme_avail, -MAP_RESERVE_COUNT * 2);
780 if (gd->gd_cpuid == 0) {
781 entry = &cpu_map_entry_init_bsp[0];
782 count = MAPENTRYBSP_CACHE;
784 entry = &cpu_map_entry_init_ap[gd->gd_cpuid][0];
785 count = MAPENTRYAP_CACHE;
787 for (i = 0; i < count; ++i, ++entry) {
788 entry->next = gd->gd_vme_base;
789 gd->gd_vme_base = entry;
794 * Reserves vm_map_entry structures so code later-on can manipulate
795 * map_entry structures within a locked map without blocking trying
796 * to allocate a new vm_map_entry.
800 * WARNING! We must not decrement gd_vme_avail until after we have
801 * ensured that sufficient entries exist, otherwise we can
802 * get into an endless call recursion in the zalloc code
806 vm_map_entry_reserve(int count)
808 struct globaldata *gd = mycpu;
809 vm_map_entry_t entry;
812 * Make sure we have enough structures in gd_vme_base to handle
813 * the reservation request.
815 * Use a critical section to protect against VM faults. It might
816 * not be needed, but we have to be careful here.
818 if (gd->gd_vme_avail < count) {
820 while (gd->gd_vme_avail < count) {
821 entry = zalloc(mapentzone);
822 entry->next = gd->gd_vme_base;
823 gd->gd_vme_base = entry;
824 atomic_add_int(&gd->gd_vme_avail, 1);
828 atomic_add_int(&gd->gd_vme_avail, -count);
834 * Releases previously reserved vm_map_entry structures that were not
835 * used. If we have too much junk in our per-cpu cache clean some of
841 vm_map_entry_release(int count)
843 struct globaldata *gd = mycpu;
844 vm_map_entry_t entry;
845 vm_map_entry_t efree;
847 count = atomic_fetchadd_int(&gd->gd_vme_avail, count) + count;
848 if (gd->gd_vme_avail > MAP_RESERVE_SLOP) {
851 while (gd->gd_vme_avail > MAP_RESERVE_HYST) {
852 entry = gd->gd_vme_base;
853 KKASSERT(entry != NULL);
854 gd->gd_vme_base = entry->next;
855 atomic_add_int(&gd->gd_vme_avail, -1);
860 while ((entry = efree) != NULL) {
862 zfree(mapentzone, entry);
868 * Reserve map entry structures for use in kernel_map itself. These
869 * entries have *ALREADY* been reserved on a per-cpu basis when the map
870 * was inited. This function is used by zalloc() to avoid a recursion
871 * when zalloc() itself needs to allocate additional kernel memory.
873 * This function works like the normal reserve but does not load the
874 * vm_map_entry cache (because that would result in an infinite
875 * recursion). Note that gd_vme_avail may go negative. This is expected.
877 * Any caller of this function must be sure to renormalize after
878 * potentially eating entries to ensure that the reserve supply
884 vm_map_entry_kreserve(int count)
886 struct globaldata *gd = mycpu;
888 atomic_add_int(&gd->gd_vme_avail, -count);
889 KASSERT(gd->gd_vme_base != NULL,
890 ("no reserved entries left, gd_vme_avail = %d",
896 * Release previously reserved map entries for kernel_map. We do not
897 * attempt to clean up like the normal release function as this would
898 * cause an unnecessary (but probably not fatal) deep procedure call.
903 vm_map_entry_krelease(int count)
905 struct globaldata *gd = mycpu;
907 atomic_add_int(&gd->gd_vme_avail, count);
911 * Allocates a VM map entry for insertion. No entry fields are filled in.
913 * The entries should have previously been reserved. The reservation count
914 * is tracked in (*countp).
918 static vm_map_entry_t
919 vm_map_entry_create(vm_map_t map, int *countp)
921 struct globaldata *gd = mycpu;
922 vm_map_entry_t entry;
924 KKASSERT(*countp > 0);
927 entry = gd->gd_vme_base;
928 KASSERT(entry != NULL, ("gd_vme_base NULL! count %d", *countp));
929 gd->gd_vme_base = entry->next;
936 * Dispose of a vm_map_entry that is no longer being referenced.
941 vm_map_entry_dispose(vm_map_t map, vm_map_entry_t entry, int *countp)
943 struct globaldata *gd = mycpu;
947 entry->next = gd->gd_vme_base;
948 gd->gd_vme_base = entry;
954 * Insert/remove entries from maps.
956 * The related map must be exclusively locked.
957 * The caller must hold map->token
958 * No other requirements.
961 vm_map_entry_link(vm_map_t map,
962 vm_map_entry_t after_where,
963 vm_map_entry_t entry)
965 ASSERT_VM_MAP_LOCKED(map);
968 entry->prev = after_where;
969 entry->next = after_where->next;
970 entry->next->prev = entry;
971 after_where->next = entry;
972 if (vm_map_rb_tree_RB_INSERT(&map->rb_root, entry))
973 panic("vm_map_entry_link: dup addr map %p ent %p", map, entry);
977 vm_map_entry_unlink(vm_map_t map,
978 vm_map_entry_t entry)
983 ASSERT_VM_MAP_LOCKED(map);
985 if (entry->eflags & MAP_ENTRY_IN_TRANSITION) {
986 panic("vm_map_entry_unlink: attempt to mess with "
987 "locked entry! %p", entry);
993 vm_map_rb_tree_RB_REMOVE(&map->rb_root, entry);
998 * Finds the map entry containing (or immediately preceding) the specified
999 * address in the given map. The entry is returned in (*entry).
1001 * The boolean result indicates whether the address is actually contained
1004 * The related map must be locked.
1005 * No other requirements.
1008 vm_map_lookup_entry(vm_map_t map, vm_offset_t address, vm_map_entry_t *entry)
1011 vm_map_entry_t last;
1013 ASSERT_VM_MAP_LOCKED(map);
1016 * Locate the record from the top of the tree. 'last' tracks the
1017 * closest prior record and is returned if no match is found, which
1018 * in binary tree terms means tracking the most recent right-branch
1019 * taken. If there is no prior record, &map->header is returned.
1021 last = &map->header;
1022 tmp = RB_ROOT(&map->rb_root);
1025 if (address >= tmp->start) {
1026 if (address < tmp->end) {
1031 tmp = RB_RIGHT(tmp, rb_entry);
1033 tmp = RB_LEFT(tmp, rb_entry);
1041 * Inserts the given whole VM object into the target map at the specified
1042 * address range. The object's size should match that of the address range.
1044 * The map must be exclusively locked.
1045 * The object must be held.
1046 * The caller must have reserved sufficient vm_map_entry structures.
1048 * If object is non-NULL, ref count must be bumped by caller prior to
1049 * making call to account for the new entry.
1052 vm_map_insert(vm_map_t map, int *countp, void *map_object, void *map_aux,
1053 vm_ooffset_t offset, vm_offset_t start, vm_offset_t end,
1054 vm_maptype_t maptype, vm_subsys_t id,
1055 vm_prot_t prot, vm_prot_t max, int cow)
1057 vm_map_entry_t new_entry;
1058 vm_map_entry_t prev_entry;
1059 vm_map_entry_t temp_entry;
1060 vm_eflags_t protoeflags;
1064 if (maptype == VM_MAPTYPE_UKSMAP)
1067 object = map_object;
1069 ASSERT_VM_MAP_LOCKED(map);
1071 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
1074 * Check that the start and end points are not bogus.
1076 if ((start < map->header.start) || (end > map->header.end) ||
1078 return (KERN_INVALID_ADDRESS);
1081 * Find the entry prior to the proposed starting address; if it's part
1082 * of an existing entry, this range is bogus.
1084 if (vm_map_lookup_entry(map, start, &temp_entry))
1085 return (KERN_NO_SPACE);
1087 prev_entry = temp_entry;
1090 * Assert that the next entry doesn't overlap the end point.
1093 if ((prev_entry->next != &map->header) &&
1094 (prev_entry->next->start < end))
1095 return (KERN_NO_SPACE);
1099 if (cow & MAP_COPY_ON_WRITE)
1100 protoeflags |= MAP_ENTRY_COW|MAP_ENTRY_NEEDS_COPY;
1102 if (cow & MAP_NOFAULT) {
1103 protoeflags |= MAP_ENTRY_NOFAULT;
1105 KASSERT(object == NULL,
1106 ("vm_map_insert: paradoxical MAP_NOFAULT request"));
1108 if (cow & MAP_DISABLE_SYNCER)
1109 protoeflags |= MAP_ENTRY_NOSYNC;
1110 if (cow & MAP_DISABLE_COREDUMP)
1111 protoeflags |= MAP_ENTRY_NOCOREDUMP;
1112 if (cow & MAP_IS_STACK)
1113 protoeflags |= MAP_ENTRY_STACK;
1114 if (cow & MAP_IS_KSTACK)
1115 protoeflags |= MAP_ENTRY_KSTACK;
1117 lwkt_gettoken(&map->token);
1121 * When object is non-NULL, it could be shared with another
1122 * process. We have to set or clear OBJ_ONEMAPPING
1125 * NOTE: This flag is only applicable to DEFAULT and SWAP
1126 * objects and will already be clear in other types
1127 * of objects, so a shared object lock is ok for
1130 if ((object->ref_count > 1) || (object->shadow_count != 0)) {
1131 vm_object_clear_flag(object, OBJ_ONEMAPPING);
1134 else if ((prev_entry != &map->header) &&
1135 (prev_entry->eflags == protoeflags) &&
1136 (prev_entry->end == start) &&
1137 (prev_entry->wired_count == 0) &&
1138 (prev_entry->id == id) &&
1139 prev_entry->maptype == maptype &&
1140 maptype == VM_MAPTYPE_NORMAL &&
1141 ((prev_entry->object.vm_object == NULL) ||
1142 vm_object_coalesce(prev_entry->object.vm_object,
1143 OFF_TO_IDX(prev_entry->offset),
1144 (vm_size_t)(prev_entry->end - prev_entry->start),
1145 (vm_size_t)(end - prev_entry->end)))) {
1147 * We were able to extend the object. Determine if we
1148 * can extend the previous map entry to include the
1149 * new range as well.
1151 if ((prev_entry->inheritance == VM_INHERIT_DEFAULT) &&
1152 (prev_entry->protection == prot) &&
1153 (prev_entry->max_protection == max)) {
1154 map->size += (end - prev_entry->end);
1155 prev_entry->end = end;
1156 vm_map_simplify_entry(map, prev_entry, countp);
1157 lwkt_reltoken(&map->token);
1158 return (KERN_SUCCESS);
1162 * If we can extend the object but cannot extend the
1163 * map entry, we have to create a new map entry. We
1164 * must bump the ref count on the extended object to
1165 * account for it. object may be NULL.
1167 * XXX if object is NULL should we set offset to 0 here ?
1169 object = prev_entry->object.vm_object;
1170 offset = prev_entry->offset +
1171 (prev_entry->end - prev_entry->start);
1173 vm_object_hold(object);
1174 vm_object_chain_wait(object, 0);
1175 vm_object_reference_locked(object);
1177 map_object = object;
1182 * NOTE: if conditionals fail, object can be NULL here. This occurs
1183 * in things like the buffer map where we manage kva but do not manage
1188 * Create a new entry
1191 new_entry = vm_map_entry_create(map, countp);
1192 new_entry->start = start;
1193 new_entry->end = end;
1196 new_entry->maptype = maptype;
1197 new_entry->eflags = protoeflags;
1198 new_entry->object.map_object = map_object;
1199 new_entry->aux.master_pde = 0; /* in case size is different */
1200 new_entry->aux.map_aux = map_aux;
1201 new_entry->offset = offset;
1203 new_entry->inheritance = VM_INHERIT_DEFAULT;
1204 new_entry->protection = prot;
1205 new_entry->max_protection = max;
1206 new_entry->wired_count = 0;
1209 * Insert the new entry into the list
1212 vm_map_entry_link(map, prev_entry, new_entry);
1213 map->size += new_entry->end - new_entry->start;
1216 * Don't worry about updating freehint[] when inserting, allow
1217 * addresses to be lower than the actual first free spot.
1221 * Temporarily removed to avoid MAP_STACK panic, due to
1222 * MAP_STACK being a huge hack. Will be added back in
1223 * when MAP_STACK (and the user stack mapping) is fixed.
1226 * It may be possible to simplify the entry
1228 vm_map_simplify_entry(map, new_entry, countp);
1232 * Try to pre-populate the page table. Mappings governed by virtual
1233 * page tables cannot be prepopulated without a lot of work, so
1236 if ((cow & (MAP_PREFAULT|MAP_PREFAULT_PARTIAL)) &&
1237 maptype != VM_MAPTYPE_VPAGETABLE &&
1238 maptype != VM_MAPTYPE_UKSMAP) {
1240 if (vm_map_relock_enable && (cow & MAP_PREFAULT_RELOCK)) {
1242 vm_object_lock_swap();
1243 vm_object_drop(object);
1245 pmap_object_init_pt(map->pmap, start, prot,
1246 object, OFF_TO_IDX(offset), end - start,
1247 cow & MAP_PREFAULT_PARTIAL);
1249 vm_object_hold(object);
1250 vm_object_lock_swap();
1254 vm_object_drop(object);
1256 lwkt_reltoken(&map->token);
1257 return (KERN_SUCCESS);
1261 * Find sufficient space for `length' bytes in the given map, starting at
1262 * `start'. Returns 0 on success, 1 on no space.
1264 * This function will returned an arbitrarily aligned pointer. If no
1265 * particular alignment is required you should pass align as 1. Note that
1266 * the map may return PAGE_SIZE aligned pointers if all the lengths used in
1267 * the map are a multiple of PAGE_SIZE, even if you pass a smaller align
1270 * 'align' should be a power of 2 but is not required to be.
1272 * The map must be exclusively locked.
1273 * No other requirements.
1276 vm_map_findspace(vm_map_t map, vm_offset_t start, vm_size_t length,
1277 vm_size_t align, int flags, vm_offset_t *addr)
1279 vm_map_entry_t entry, next;
1281 vm_offset_t hole_start;
1283 vm_offset_t align_mask;
1285 if (start < map->header.start)
1286 start = map->header.start;
1287 if (start > map->header.end)
1291 * If the alignment is not a power of 2 we will have to use
1292 * a mod/division, set align_mask to a special value.
1294 if ((align | (align - 1)) + 1 != (align << 1))
1295 align_mask = (vm_offset_t)-1;
1297 align_mask = align - 1;
1300 * Use freehint to adjust the start point, hopefully reducing
1301 * the iteration to O(1).
1303 hole_start = vm_map_freehint_find(map, length, align);
1304 if (start < hole_start)
1306 if (vm_map_lookup_entry(map, start, &tmp))
1311 * Look through the rest of the map, trying to fit a new region in the
1312 * gap between existing regions, or after the very last region.
1314 for (;; start = (entry = next)->end) {
1316 * Adjust the proposed start by the requested alignment,
1317 * be sure that we didn't wrap the address.
1319 if (align_mask == (vm_offset_t)-1)
1320 end = roundup(start, align);
1322 end = (start + align_mask) & ~align_mask;
1328 * Find the end of the proposed new region. Be sure we didn't
1329 * go beyond the end of the map, or wrap around the address.
1330 * Then check to see if this is the last entry or if the
1331 * proposed end fits in the gap between this and the next
1334 end = start + length;
1335 if (end > map->header.end || end < start)
1340 * If the next entry's start address is beyond the desired
1341 * end address we may have found a good entry.
1343 * If the next entry is a stack mapping we do not map into
1344 * the stack's reserved space.
1346 * XXX continue to allow mapping into the stack's reserved
1347 * space if doing a MAP_STACK mapping inside a MAP_STACK
1348 * mapping, for backwards compatibility. But the caller
1349 * really should use MAP_STACK | MAP_TRYFIXED if they
1352 if (next == &map->header)
1354 if (next->start >= end) {
1355 if ((next->eflags & MAP_ENTRY_STACK) == 0)
1357 if (flags & MAP_STACK)
1359 if (next->start - next->aux.avail_ssize >= end)
1365 * Update the freehint
1367 vm_map_freehint_update(map, start, length, align);
1370 * Grow the kernel_map if necessary. pmap_growkernel() will panic
1371 * if it fails. The kernel_map is locked and nothing can steal
1372 * our address space if pmap_growkernel() blocks.
1374 * NOTE: This may be unconditionally called for kldload areas on
1375 * x86_64 because these do not bump kernel_vm_end (which would
1376 * fill 128G worth of page tables!). Therefore we must not
1379 if (map == &kernel_map) {
1382 kstop = round_page(start + length);
1383 if (kstop > kernel_vm_end)
1384 pmap_growkernel(start, kstop);
1391 * vm_map_find finds an unallocated region in the target address map with
1392 * the given length and allocates it. The search is defined to be first-fit
1393 * from the specified address; the region found is returned in the same
1396 * If object is non-NULL, ref count must be bumped by caller
1397 * prior to making call to account for the new entry.
1399 * No requirements. This function will lock the map temporarily.
1402 vm_map_find(vm_map_t map, void *map_object, void *map_aux,
1403 vm_ooffset_t offset, vm_offset_t *addr,
1404 vm_size_t length, vm_size_t align, boolean_t fitit,
1405 vm_maptype_t maptype, vm_subsys_t id,
1406 vm_prot_t prot, vm_prot_t max, int cow)
1413 if (maptype == VM_MAPTYPE_UKSMAP)
1416 object = map_object;
1420 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
1423 vm_object_hold_shared(object);
1425 if (vm_map_findspace(map, start, length, align, 0, addr)) {
1427 vm_object_drop(object);
1429 vm_map_entry_release(count);
1430 return (KERN_NO_SPACE);
1434 result = vm_map_insert(map, &count, map_object, map_aux,
1435 offset, start, start + length,
1436 maptype, id, prot, max, cow);
1438 vm_object_drop(object);
1440 vm_map_entry_release(count);
1446 * Simplify the given map entry by merging with either neighbor. This
1447 * routine also has the ability to merge with both neighbors.
1449 * This routine guarentees that the passed entry remains valid (though
1450 * possibly extended). When merging, this routine may delete one or
1451 * both neighbors. No action is taken on entries which have their
1452 * in-transition flag set.
1454 * The map must be exclusively locked.
1457 vm_map_simplify_entry(vm_map_t map, vm_map_entry_t entry, int *countp)
1459 vm_map_entry_t next, prev;
1460 vm_size_t prevsize, esize;
1462 if (entry->eflags & MAP_ENTRY_IN_TRANSITION) {
1463 ++mycpu->gd_cnt.v_intrans_coll;
1467 if (entry->maptype == VM_MAPTYPE_SUBMAP)
1469 if (entry->maptype == VM_MAPTYPE_UKSMAP)
1473 if (prev != &map->header) {
1474 prevsize = prev->end - prev->start;
1475 if ( (prev->end == entry->start) &&
1476 (prev->maptype == entry->maptype) &&
1477 (prev->object.vm_object == entry->object.vm_object) &&
1478 (!prev->object.vm_object ||
1479 (prev->offset + prevsize == entry->offset)) &&
1480 (prev->eflags == entry->eflags) &&
1481 (prev->protection == entry->protection) &&
1482 (prev->max_protection == entry->max_protection) &&
1483 (prev->inheritance == entry->inheritance) &&
1484 (prev->id == entry->id) &&
1485 (prev->wired_count == entry->wired_count)) {
1486 vm_map_entry_unlink(map, prev);
1487 entry->start = prev->start;
1488 entry->offset = prev->offset;
1489 if (prev->object.vm_object)
1490 vm_object_deallocate(prev->object.vm_object);
1491 vm_map_entry_dispose(map, prev, countp);
1496 if (next != &map->header) {
1497 esize = entry->end - entry->start;
1498 if ((entry->end == next->start) &&
1499 (next->maptype == entry->maptype) &&
1500 (next->object.vm_object == entry->object.vm_object) &&
1501 (!entry->object.vm_object ||
1502 (entry->offset + esize == next->offset)) &&
1503 (next->eflags == entry->eflags) &&
1504 (next->protection == entry->protection) &&
1505 (next->max_protection == entry->max_protection) &&
1506 (next->inheritance == entry->inheritance) &&
1507 (next->id == entry->id) &&
1508 (next->wired_count == entry->wired_count)) {
1509 vm_map_entry_unlink(map, next);
1510 entry->end = next->end;
1511 if (next->object.vm_object)
1512 vm_object_deallocate(next->object.vm_object);
1513 vm_map_entry_dispose(map, next, countp);
1519 * Asserts that the given entry begins at or after the specified address.
1520 * If necessary, it splits the entry into two.
1522 #define vm_map_clip_start(map, entry, startaddr, countp) \
1524 if (startaddr > entry->start) \
1525 _vm_map_clip_start(map, entry, startaddr, countp); \
1529 * This routine is called only when it is known that the entry must be split.
1531 * The map must be exclusively locked.
1534 _vm_map_clip_start(vm_map_t map, vm_map_entry_t entry, vm_offset_t start,
1537 vm_map_entry_t new_entry;
1540 * Split off the front portion -- note that we must insert the new
1541 * entry BEFORE this one, so that this entry has the specified
1545 vm_map_simplify_entry(map, entry, countp);
1548 * If there is no object backing this entry, we might as well create
1549 * one now. If we defer it, an object can get created after the map
1550 * is clipped, and individual objects will be created for the split-up
1551 * map. This is a bit of a hack, but is also about the best place to
1552 * put this improvement.
1554 if (entry->object.vm_object == NULL && !map->system_map &&
1555 VM_MAP_ENTRY_WITHIN_PARTITION(entry)) {
1556 vm_map_entry_allocate_object(entry);
1559 new_entry = vm_map_entry_create(map, countp);
1560 *new_entry = *entry;
1562 new_entry->end = start;
1563 entry->offset += (start - entry->start);
1564 entry->start = start;
1566 vm_map_entry_link(map, entry->prev, new_entry);
1568 switch(entry->maptype) {
1569 case VM_MAPTYPE_NORMAL:
1570 case VM_MAPTYPE_VPAGETABLE:
1571 if (new_entry->object.vm_object) {
1572 vm_object_hold(new_entry->object.vm_object);
1573 vm_object_chain_wait(new_entry->object.vm_object, 0);
1574 vm_object_reference_locked(new_entry->object.vm_object);
1575 vm_object_drop(new_entry->object.vm_object);
1584 * Asserts that the given entry ends at or before the specified address.
1585 * If necessary, it splits the entry into two.
1587 * The map must be exclusively locked.
1589 #define vm_map_clip_end(map, entry, endaddr, countp) \
1591 if (endaddr < entry->end) \
1592 _vm_map_clip_end(map, entry, endaddr, countp); \
1596 * This routine is called only when it is known that the entry must be split.
1598 * The map must be exclusively locked.
1601 _vm_map_clip_end(vm_map_t map, vm_map_entry_t entry, vm_offset_t end,
1604 vm_map_entry_t new_entry;
1607 * If there is no object backing this entry, we might as well create
1608 * one now. If we defer it, an object can get created after the map
1609 * is clipped, and individual objects will be created for the split-up
1610 * map. This is a bit of a hack, but is also about the best place to
1611 * put this improvement.
1614 if (entry->object.vm_object == NULL && !map->system_map &&
1615 VM_MAP_ENTRY_WITHIN_PARTITION(entry)) {
1616 vm_map_entry_allocate_object(entry);
1620 * Create a new entry and insert it AFTER the specified entry
1622 new_entry = vm_map_entry_create(map, countp);
1623 *new_entry = *entry;
1625 new_entry->start = entry->end = end;
1626 new_entry->offset += (end - entry->start);
1628 vm_map_entry_link(map, entry, new_entry);
1630 switch(entry->maptype) {
1631 case VM_MAPTYPE_NORMAL:
1632 case VM_MAPTYPE_VPAGETABLE:
1633 if (new_entry->object.vm_object) {
1634 vm_object_hold(new_entry->object.vm_object);
1635 vm_object_chain_wait(new_entry->object.vm_object, 0);
1636 vm_object_reference_locked(new_entry->object.vm_object);
1637 vm_object_drop(new_entry->object.vm_object);
1646 * Asserts that the starting and ending region addresses fall within the
1647 * valid range for the map.
1649 #define VM_MAP_RANGE_CHECK(map, start, end) \
1651 if (start < vm_map_min(map)) \
1652 start = vm_map_min(map); \
1653 if (end > vm_map_max(map)) \
1654 end = vm_map_max(map); \
1660 * Used to block when an in-transition collison occurs. The map
1661 * is unlocked for the sleep and relocked before the return.
1664 vm_map_transition_wait(vm_map_t map, int relock)
1666 tsleep_interlock(map, 0);
1668 tsleep(map, PINTERLOCKED, "vment", 0);
1674 * When we do blocking operations with the map lock held it is
1675 * possible that a clip might have occured on our in-transit entry,
1676 * requiring an adjustment to the entry in our loop. These macros
1677 * help the pageable and clip_range code deal with the case. The
1678 * conditional costs virtually nothing if no clipping has occured.
1681 #define CLIP_CHECK_BACK(entry, save_start) \
1683 while (entry->start != save_start) { \
1684 entry = entry->prev; \
1685 KASSERT(entry != &map->header, ("bad entry clip")); \
1689 #define CLIP_CHECK_FWD(entry, save_end) \
1691 while (entry->end != save_end) { \
1692 entry = entry->next; \
1693 KASSERT(entry != &map->header, ("bad entry clip")); \
1699 * Clip the specified range and return the base entry. The
1700 * range may cover several entries starting at the returned base
1701 * and the first and last entry in the covering sequence will be
1702 * properly clipped to the requested start and end address.
1704 * If no holes are allowed you should pass the MAP_CLIP_NO_HOLES
1707 * The MAP_ENTRY_IN_TRANSITION flag will be set for the entries
1708 * covered by the requested range.
1710 * The map must be exclusively locked on entry and will remain locked
1711 * on return. If no range exists or the range contains holes and you
1712 * specified that no holes were allowed, NULL will be returned. This
1713 * routine may temporarily unlock the map in order avoid a deadlock when
1718 vm_map_clip_range(vm_map_t map, vm_offset_t start, vm_offset_t end,
1719 int *countp, int flags)
1721 vm_map_entry_t start_entry;
1722 vm_map_entry_t entry;
1725 * Locate the entry and effect initial clipping. The in-transition
1726 * case does not occur very often so do not try to optimize it.
1729 if (vm_map_lookup_entry(map, start, &start_entry) == FALSE)
1731 entry = start_entry;
1732 if (entry->eflags & MAP_ENTRY_IN_TRANSITION) {
1733 entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP;
1734 ++mycpu->gd_cnt.v_intrans_coll;
1735 ++mycpu->gd_cnt.v_intrans_wait;
1736 vm_map_transition_wait(map, 1);
1738 * entry and/or start_entry may have been clipped while
1739 * we slept, or may have gone away entirely. We have
1740 * to restart from the lookup.
1746 * Since we hold an exclusive map lock we do not have to restart
1747 * after clipping, even though clipping may block in zalloc.
1749 vm_map_clip_start(map, entry, start, countp);
1750 vm_map_clip_end(map, entry, end, countp);
1751 entry->eflags |= MAP_ENTRY_IN_TRANSITION;
1754 * Scan entries covered by the range. When working on the next
1755 * entry a restart need only re-loop on the current entry which
1756 * we have already locked, since 'next' may have changed. Also,
1757 * even though entry is safe, it may have been clipped so we
1758 * have to iterate forwards through the clip after sleeping.
1760 while (entry->next != &map->header && entry->next->start < end) {
1761 vm_map_entry_t next = entry->next;
1763 if (flags & MAP_CLIP_NO_HOLES) {
1764 if (next->start > entry->end) {
1765 vm_map_unclip_range(map, start_entry,
1766 start, entry->end, countp, flags);
1771 if (next->eflags & MAP_ENTRY_IN_TRANSITION) {
1772 vm_offset_t save_end = entry->end;
1773 next->eflags |= MAP_ENTRY_NEEDS_WAKEUP;
1774 ++mycpu->gd_cnt.v_intrans_coll;
1775 ++mycpu->gd_cnt.v_intrans_wait;
1776 vm_map_transition_wait(map, 1);
1779 * clips might have occured while we blocked.
1781 CLIP_CHECK_FWD(entry, save_end);
1782 CLIP_CHECK_BACK(start_entry, start);
1787 * No restart necessary even though clip_end may block, we
1788 * are holding the map lock.
1790 vm_map_clip_end(map, next, end, countp);
1791 next->eflags |= MAP_ENTRY_IN_TRANSITION;
1794 if (flags & MAP_CLIP_NO_HOLES) {
1795 if (entry->end != end) {
1796 vm_map_unclip_range(map, start_entry,
1797 start, entry->end, countp, flags);
1801 return(start_entry);
1805 * Undo the effect of vm_map_clip_range(). You should pass the same
1806 * flags and the same range that you passed to vm_map_clip_range().
1807 * This code will clear the in-transition flag on the entries and
1808 * wake up anyone waiting. This code will also simplify the sequence
1809 * and attempt to merge it with entries before and after the sequence.
1811 * The map must be locked on entry and will remain locked on return.
1813 * Note that you should also pass the start_entry returned by
1814 * vm_map_clip_range(). However, if you block between the two calls
1815 * with the map unlocked please be aware that the start_entry may
1816 * have been clipped and you may need to scan it backwards to find
1817 * the entry corresponding with the original start address. You are
1818 * responsible for this, vm_map_unclip_range() expects the correct
1819 * start_entry to be passed to it and will KASSERT otherwise.
1823 vm_map_unclip_range(vm_map_t map, vm_map_entry_t start_entry,
1824 vm_offset_t start, vm_offset_t end,
1825 int *countp, int flags)
1827 vm_map_entry_t entry;
1829 entry = start_entry;
1831 KASSERT(entry->start == start, ("unclip_range: illegal base entry"));
1832 while (entry != &map->header && entry->start < end) {
1833 KASSERT(entry->eflags & MAP_ENTRY_IN_TRANSITION,
1834 ("in-transition flag not set during unclip on: %p",
1836 KASSERT(entry->end <= end,
1837 ("unclip_range: tail wasn't clipped"));
1838 entry->eflags &= ~MAP_ENTRY_IN_TRANSITION;
1839 if (entry->eflags & MAP_ENTRY_NEEDS_WAKEUP) {
1840 entry->eflags &= ~MAP_ENTRY_NEEDS_WAKEUP;
1843 entry = entry->next;
1847 * Simplification does not block so there is no restart case.
1849 entry = start_entry;
1850 while (entry != &map->header && entry->start < end) {
1851 vm_map_simplify_entry(map, entry, countp);
1852 entry = entry->next;
1857 * Mark the given range as handled by a subordinate map.
1859 * This range must have been created with vm_map_find(), and no other
1860 * operations may have been performed on this range prior to calling
1863 * Submappings cannot be removed.
1868 vm_map_submap(vm_map_t map, vm_offset_t start, vm_offset_t end, vm_map_t submap)
1870 vm_map_entry_t entry;
1871 int result = KERN_INVALID_ARGUMENT;
1874 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
1877 VM_MAP_RANGE_CHECK(map, start, end);
1879 if (vm_map_lookup_entry(map, start, &entry)) {
1880 vm_map_clip_start(map, entry, start, &count);
1882 entry = entry->next;
1885 vm_map_clip_end(map, entry, end, &count);
1887 if ((entry->start == start) && (entry->end == end) &&
1888 ((entry->eflags & MAP_ENTRY_COW) == 0) &&
1889 (entry->object.vm_object == NULL)) {
1890 entry->object.sub_map = submap;
1891 entry->maptype = VM_MAPTYPE_SUBMAP;
1892 result = KERN_SUCCESS;
1895 vm_map_entry_release(count);
1901 * Sets the protection of the specified address region in the target map.
1902 * If "set_max" is specified, the maximum protection is to be set;
1903 * otherwise, only the current protection is affected.
1905 * The protection is not applicable to submaps, but is applicable to normal
1906 * maps and maps governed by virtual page tables. For example, when operating
1907 * on a virtual page table our protection basically controls how COW occurs
1908 * on the backing object, whereas the virtual page table abstraction itself
1909 * is an abstraction for userland.
1914 vm_map_protect(vm_map_t map, vm_offset_t start, vm_offset_t end,
1915 vm_prot_t new_prot, boolean_t set_max)
1917 vm_map_entry_t current;
1918 vm_map_entry_t entry;
1921 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
1924 VM_MAP_RANGE_CHECK(map, start, end);
1926 if (vm_map_lookup_entry(map, start, &entry)) {
1927 vm_map_clip_start(map, entry, start, &count);
1929 entry = entry->next;
1933 * Make a first pass to check for protection violations.
1936 while ((current != &map->header) && (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 = current->next;
1971 * Go back and fix up protections. [Note that clipping is not
1972 * necessary the second time.]
1976 while ((current != &map->header) && (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 = current->next;
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 = entry->next;
2077 * madvise behaviors that are implemented in the vm_map_entry.
2079 * We clip the vm_map_entry so that behavioral changes are
2080 * limited to the specified address range.
2082 for (current = entry;
2083 (current != &map->header) && (current->start < end);
2084 current = current->next
2086 if (current->maptype == VM_MAPTYPE_SUBMAP)
2089 vm_map_clip_end(map, current, end, &count);
2093 vm_map_entry_set_behavior(current, MAP_ENTRY_BEHAV_NORMAL);
2095 case MADV_SEQUENTIAL:
2096 vm_map_entry_set_behavior(current, MAP_ENTRY_BEHAV_SEQUENTIAL);
2099 vm_map_entry_set_behavior(current, MAP_ENTRY_BEHAV_RANDOM);
2102 current->eflags |= MAP_ENTRY_NOSYNC;
2105 current->eflags &= ~MAP_ENTRY_NOSYNC;
2108 current->eflags |= MAP_ENTRY_NOCOREDUMP;
2111 current->eflags &= ~MAP_ENTRY_NOCOREDUMP;
2115 * Set the page directory page for a map
2116 * governed by a virtual page table. Mark
2117 * the entry as being governed by a virtual
2118 * page table if it is not.
2120 * XXX the page directory page is stored
2121 * in the avail_ssize field if the map_entry.
2123 * XXX the map simplification code does not
2124 * compare this field so weird things may
2125 * happen if you do not apply this function
2126 * to the entire mapping governed by the
2127 * virtual page table.
2129 if (current->maptype != VM_MAPTYPE_VPAGETABLE) {
2133 current->aux.master_pde = value;
2134 pmap_remove(map->pmap,
2135 current->start, current->end);
2139 * Invalidate the related pmap entries, used
2140 * to flush portions of the real kernel's
2141 * pmap when the caller has removed or
2142 * modified existing mappings in a virtual
2145 * (exclusive locked map version does not
2146 * need the range interlock).
2148 pmap_remove(map->pmap,
2149 current->start, current->end);
2155 vm_map_simplify_entry(map, current, &count);
2163 * madvise behaviors that are implemented in the underlying
2166 * Since we don't clip the vm_map_entry, we have to clip
2167 * the vm_object pindex and count.
2169 * NOTE! These functions are only supported on normal maps,
2170 * except MADV_INVAL which is also supported on
2171 * virtual page tables.
2173 for (current = entry;
2174 (current != &map->header) && (current->start < end);
2175 current = current->next
2177 vm_offset_t useStart;
2179 if (current->maptype != VM_MAPTYPE_NORMAL &&
2180 (current->maptype != VM_MAPTYPE_VPAGETABLE ||
2181 behav != MADV_INVAL)) {
2185 pindex = OFF_TO_IDX(current->offset);
2186 delta = atop(current->end - current->start);
2187 useStart = current->start;
2189 if (current->start < start) {
2190 pindex += atop(start - current->start);
2191 delta -= atop(start - current->start);
2194 if (current->end > end)
2195 delta -= atop(current->end - end);
2197 if ((vm_spindex_t)delta <= 0)
2200 if (behav == MADV_INVAL) {
2202 * Invalidate the related pmap entries, used
2203 * to flush portions of the real kernel's
2204 * pmap when the caller has removed or
2205 * modified existing mappings in a virtual
2208 * (shared locked map version needs the
2209 * interlock, see vm_fault()).
2211 struct vm_map_ilock ilock;
2213 KASSERT(useStart >= VM_MIN_USER_ADDRESS &&
2214 useStart + ptoa(delta) <=
2215 VM_MAX_USER_ADDRESS,
2216 ("Bad range %016jx-%016jx (%016jx)",
2217 useStart, useStart + ptoa(delta),
2219 vm_map_interlock(map, &ilock,
2221 useStart + ptoa(delta));
2222 pmap_remove(map->pmap,
2224 useStart + ptoa(delta));
2225 vm_map_deinterlock(map, &ilock);
2227 vm_object_madvise(current->object.vm_object,
2228 pindex, delta, behav);
2232 * Try to populate the page table. Mappings governed
2233 * by virtual page tables cannot be pre-populated
2234 * without a lot of work so don't try.
2236 if (behav == MADV_WILLNEED &&
2237 current->maptype != VM_MAPTYPE_VPAGETABLE) {
2238 pmap_object_init_pt(
2241 current->protection,
2242 current->object.vm_object,
2244 (count << PAGE_SHIFT),
2245 MAP_PREFAULT_MADVISE
2249 vm_map_unlock_read(map);
2251 vm_map_entry_release(count);
2257 * Sets the inheritance of the specified address range in the target map.
2258 * Inheritance affects how the map will be shared with child maps at the
2259 * time of vm_map_fork.
2262 vm_map_inherit(vm_map_t map, vm_offset_t start, vm_offset_t end,
2263 vm_inherit_t new_inheritance)
2265 vm_map_entry_t entry;
2266 vm_map_entry_t temp_entry;
2269 switch (new_inheritance) {
2270 case VM_INHERIT_NONE:
2271 case VM_INHERIT_COPY:
2272 case VM_INHERIT_SHARE:
2275 return (KERN_INVALID_ARGUMENT);
2278 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
2281 VM_MAP_RANGE_CHECK(map, start, end);
2283 if (vm_map_lookup_entry(map, start, &temp_entry)) {
2285 vm_map_clip_start(map, entry, start, &count);
2287 entry = temp_entry->next;
2289 while ((entry != &map->header) && (entry->start < end)) {
2290 vm_map_clip_end(map, entry, end, &count);
2292 entry->inheritance = new_inheritance;
2294 vm_map_simplify_entry(map, entry, &count);
2296 entry = entry->next;
2299 vm_map_entry_release(count);
2300 return (KERN_SUCCESS);
2304 * Implement the semantics of mlock
2307 vm_map_unwire(vm_map_t map, vm_offset_t start, vm_offset_t real_end,
2308 boolean_t new_pageable)
2310 vm_map_entry_t entry;
2311 vm_map_entry_t start_entry;
2313 int rv = KERN_SUCCESS;
2316 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
2318 VM_MAP_RANGE_CHECK(map, start, real_end);
2321 start_entry = vm_map_clip_range(map, start, end, &count,
2323 if (start_entry == NULL) {
2325 vm_map_entry_release(count);
2326 return (KERN_INVALID_ADDRESS);
2329 if (new_pageable == 0) {
2330 entry = start_entry;
2331 while ((entry != &map->header) && (entry->start < end)) {
2332 vm_offset_t save_start;
2333 vm_offset_t save_end;
2336 * Already user wired or hard wired (trivial cases)
2338 if (entry->eflags & MAP_ENTRY_USER_WIRED) {
2339 entry = entry->next;
2342 if (entry->wired_count != 0) {
2343 entry->wired_count++;
2344 entry->eflags |= MAP_ENTRY_USER_WIRED;
2345 entry = entry->next;
2350 * A new wiring requires instantiation of appropriate
2351 * management structures and the faulting in of the
2354 if (entry->maptype == VM_MAPTYPE_NORMAL ||
2355 entry->maptype == VM_MAPTYPE_VPAGETABLE) {
2356 int copyflag = entry->eflags &
2357 MAP_ENTRY_NEEDS_COPY;
2358 if (copyflag && ((entry->protection &
2359 VM_PROT_WRITE) != 0)) {
2360 vm_map_entry_shadow(entry, 0);
2361 } else if (entry->object.vm_object == NULL &&
2363 vm_map_entry_allocate_object(entry);
2366 entry->wired_count++;
2367 entry->eflags |= MAP_ENTRY_USER_WIRED;
2370 * Now fault in the area. Note that vm_fault_wire()
2371 * may release the map lock temporarily, it will be
2372 * relocked on return. The in-transition
2373 * flag protects the entries.
2375 save_start = entry->start;
2376 save_end = entry->end;
2377 rv = vm_fault_wire(map, entry, TRUE, 0);
2379 CLIP_CHECK_BACK(entry, save_start);
2381 KASSERT(entry->wired_count == 1, ("bad wired_count on entry"));
2382 entry->eflags &= ~MAP_ENTRY_USER_WIRED;
2383 entry->wired_count = 0;
2384 if (entry->end == save_end)
2386 entry = entry->next;
2387 KASSERT(entry != &map->header, ("bad entry clip during backout"));
2389 end = save_start; /* unwire the rest */
2393 * note that even though the entry might have been
2394 * clipped, the USER_WIRED flag we set prevents
2395 * duplication so we do not have to do a
2398 entry = entry->next;
2402 * If we failed fall through to the unwiring section to
2403 * unwire what we had wired so far. 'end' has already
2410 * start_entry might have been clipped if we unlocked the
2411 * map and blocked. No matter how clipped it has gotten
2412 * there should be a fragment that is on our start boundary.
2414 CLIP_CHECK_BACK(start_entry, start);
2418 * Deal with the unwiring case.
2422 * This is the unwiring case. We must first ensure that the
2423 * range to be unwired is really wired down. We know there
2426 entry = start_entry;
2427 while ((entry != &map->header) && (entry->start < end)) {
2428 if ((entry->eflags & MAP_ENTRY_USER_WIRED) == 0) {
2429 rv = KERN_INVALID_ARGUMENT;
2432 KASSERT(entry->wired_count != 0, ("wired count was 0 with USER_WIRED set! %p", entry));
2433 entry = entry->next;
2437 * Now decrement the wiring count for each region. If a region
2438 * becomes completely unwired, unwire its physical pages and
2442 * The map entries are processed in a loop, checking to
2443 * make sure the entry is wired and asserting it has a wired
2444 * count. However, another loop was inserted more-or-less in
2445 * the middle of the unwiring path. This loop picks up the
2446 * "entry" loop variable from the first loop without first
2447 * setting it to start_entry. Naturally, the secound loop
2448 * is never entered and the pages backing the entries are
2449 * never unwired. This can lead to a leak of wired pages.
2451 entry = start_entry;
2452 while ((entry != &map->header) && (entry->start < end)) {
2453 KASSERT(entry->eflags & MAP_ENTRY_USER_WIRED,
2454 ("expected USER_WIRED on entry %p", entry));
2455 entry->eflags &= ~MAP_ENTRY_USER_WIRED;
2456 entry->wired_count--;
2457 if (entry->wired_count == 0)
2458 vm_fault_unwire(map, entry);
2459 entry = entry->next;
2463 vm_map_unclip_range(map, start_entry, start, real_end, &count,
2466 vm_map_entry_release(count);
2472 * Sets the pageability of the specified address range in the target map.
2473 * Regions specified as not pageable require locked-down physical
2474 * memory and physical page maps.
2476 * The map must not be locked, but a reference must remain to the map
2477 * throughout the call.
2479 * This function may be called via the zalloc path and must properly
2480 * reserve map entries for kernel_map.
2485 vm_map_wire(vm_map_t map, vm_offset_t start, vm_offset_t real_end, int kmflags)
2487 vm_map_entry_t entry;
2488 vm_map_entry_t start_entry;
2490 int rv = KERN_SUCCESS;
2493 if (kmflags & KM_KRESERVE)
2494 count = vm_map_entry_kreserve(MAP_RESERVE_COUNT);
2496 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
2498 VM_MAP_RANGE_CHECK(map, start, real_end);
2501 start_entry = vm_map_clip_range(map, start, end, &count,
2503 if (start_entry == NULL) {
2505 rv = KERN_INVALID_ADDRESS;
2508 if ((kmflags & KM_PAGEABLE) == 0) {
2512 * 1. Holding the write lock, we create any shadow or zero-fill
2513 * objects that need to be created. Then we clip each map
2514 * entry to the region to be wired and increment its wiring
2515 * count. We create objects before clipping the map entries
2516 * to avoid object proliferation.
2518 * 2. We downgrade to a read lock, and call vm_fault_wire to
2519 * fault in the pages for any newly wired area (wired_count is
2522 * Downgrading to a read lock for vm_fault_wire avoids a
2523 * possible deadlock with another process that may have faulted
2524 * on one of the pages to be wired (it would mark the page busy,
2525 * blocking us, then in turn block on the map lock that we
2526 * hold). Because of problems in the recursive lock package,
2527 * we cannot upgrade to a write lock in vm_map_lookup. Thus,
2528 * any actions that require the write lock must be done
2529 * beforehand. Because we keep the read lock on the map, the
2530 * copy-on-write status of the entries we modify here cannot
2533 entry = start_entry;
2534 while ((entry != &map->header) && (entry->start < end)) {
2536 * Trivial case if the entry is already wired
2538 if (entry->wired_count) {
2539 entry->wired_count++;
2540 entry = entry->next;
2545 * The entry is being newly wired, we have to setup
2546 * appropriate management structures. A shadow
2547 * object is required for a copy-on-write region,
2548 * or a normal object for a zero-fill region. We
2549 * do not have to do this for entries that point to sub
2550 * maps because we won't hold the lock on the sub map.
2552 if (entry->maptype == VM_MAPTYPE_NORMAL ||
2553 entry->maptype == VM_MAPTYPE_VPAGETABLE) {
2554 int copyflag = entry->eflags &
2555 MAP_ENTRY_NEEDS_COPY;
2556 if (copyflag && ((entry->protection &
2557 VM_PROT_WRITE) != 0)) {
2558 vm_map_entry_shadow(entry, 0);
2559 } else if (entry->object.vm_object == NULL &&
2561 vm_map_entry_allocate_object(entry);
2565 entry->wired_count++;
2566 entry = entry->next;
2574 * HACK HACK HACK HACK
2576 * vm_fault_wire() temporarily unlocks the map to avoid
2577 * deadlocks. The in-transition flag from vm_map_clip_range
2578 * call should protect us from changes while the map is
2581 * NOTE: Previously this comment stated that clipping might
2582 * still occur while the entry is unlocked, but from
2583 * what I can tell it actually cannot.
2585 * It is unclear whether the CLIP_CHECK_*() calls
2586 * are still needed but we keep them in anyway.
2588 * HACK HACK HACK HACK
2591 entry = start_entry;
2592 while (entry != &map->header && entry->start < end) {
2594 * If vm_fault_wire fails for any page we need to undo
2595 * what has been done. We decrement the wiring count
2596 * for those pages which have not yet been wired (now)
2597 * and unwire those that have (later).
2599 vm_offset_t save_start = entry->start;
2600 vm_offset_t save_end = entry->end;
2602 if (entry->wired_count == 1)
2603 rv = vm_fault_wire(map, entry, FALSE, kmflags);
2605 CLIP_CHECK_BACK(entry, save_start);
2607 KASSERT(entry->wired_count == 1, ("wired_count changed unexpectedly"));
2608 entry->wired_count = 0;
2609 if (entry->end == save_end)
2611 entry = entry->next;
2612 KASSERT(entry != &map->header, ("bad entry clip during backout"));
2617 CLIP_CHECK_FWD(entry, save_end);
2618 entry = entry->next;
2622 * If a failure occured undo everything by falling through
2623 * to the unwiring code. 'end' has already been adjusted
2627 kmflags |= KM_PAGEABLE;
2630 * start_entry is still IN_TRANSITION but may have been
2631 * clipped since vm_fault_wire() unlocks and relocks the
2632 * map. No matter how clipped it has gotten there should
2633 * be a fragment that is on our start boundary.
2635 CLIP_CHECK_BACK(start_entry, start);
2638 if (kmflags & KM_PAGEABLE) {
2640 * This is the unwiring case. We must first ensure that the
2641 * range to be unwired is really wired down. We know there
2644 entry = start_entry;
2645 while ((entry != &map->header) && (entry->start < end)) {
2646 if (entry->wired_count == 0) {
2647 rv = KERN_INVALID_ARGUMENT;
2650 entry = entry->next;
2654 * Now decrement the wiring count for each region. If a region
2655 * becomes completely unwired, unwire its physical pages and
2658 entry = start_entry;
2659 while ((entry != &map->header) && (entry->start < end)) {
2660 entry->wired_count--;
2661 if (entry->wired_count == 0)
2662 vm_fault_unwire(map, entry);
2663 entry = entry->next;
2667 vm_map_unclip_range(map, start_entry, start, real_end,
2668 &count, MAP_CLIP_NO_HOLES);
2671 if (kmflags & KM_KRESERVE)
2672 vm_map_entry_krelease(count);
2674 vm_map_entry_release(count);
2679 * Mark a newly allocated address range as wired but do not fault in
2680 * the pages. The caller is expected to load the pages into the object.
2682 * The map must be locked on entry and will remain locked on return.
2683 * No other requirements.
2686 vm_map_set_wired_quick(vm_map_t map, vm_offset_t addr, vm_size_t size,
2689 vm_map_entry_t scan;
2690 vm_map_entry_t entry;
2692 entry = vm_map_clip_range(map, addr, addr + size,
2693 countp, MAP_CLIP_NO_HOLES);
2695 scan != &map->header && scan->start < addr + size;
2696 scan = scan->next) {
2697 KKASSERT(scan->wired_count == 0);
2698 scan->wired_count = 1;
2700 vm_map_unclip_range(map, entry, addr, addr + size,
2701 countp, MAP_CLIP_NO_HOLES);
2705 * Push any dirty cached pages in the address range to their pager.
2706 * If syncio is TRUE, dirty pages are written synchronously.
2707 * If invalidate is TRUE, any cached pages are freed as well.
2709 * This routine is called by sys_msync()
2711 * Returns an error if any part of the specified range is not mapped.
2716 vm_map_clean(vm_map_t map, vm_offset_t start, vm_offset_t end,
2717 boolean_t syncio, boolean_t invalidate)
2719 vm_map_entry_t current;
2720 vm_map_entry_t entry;
2724 vm_ooffset_t offset;
2726 vm_map_lock_read(map);
2727 VM_MAP_RANGE_CHECK(map, start, end);
2728 if (!vm_map_lookup_entry(map, start, &entry)) {
2729 vm_map_unlock_read(map);
2730 return (KERN_INVALID_ADDRESS);
2732 lwkt_gettoken(&map->token);
2735 * Make a first pass to check for holes.
2737 for (current = entry; current->start < end; current = current->next) {
2738 if (current->maptype == VM_MAPTYPE_SUBMAP) {
2739 lwkt_reltoken(&map->token);
2740 vm_map_unlock_read(map);
2741 return (KERN_INVALID_ARGUMENT);
2743 if (end > current->end &&
2744 (current->next == &map->header ||
2745 current->end != current->next->start)) {
2746 lwkt_reltoken(&map->token);
2747 vm_map_unlock_read(map);
2748 return (KERN_INVALID_ADDRESS);
2753 pmap_remove(vm_map_pmap(map), start, end);
2756 * Make a second pass, cleaning/uncaching pages from the indicated
2759 for (current = entry; current->start < end; current = current->next) {
2760 offset = current->offset + (start - current->start);
2761 size = (end <= current->end ? end : current->end) - start;
2763 switch(current->maptype) {
2764 case VM_MAPTYPE_SUBMAP:
2767 vm_map_entry_t tentry;
2770 smap = current->object.sub_map;
2771 vm_map_lock_read(smap);
2772 vm_map_lookup_entry(smap, offset, &tentry);
2773 tsize = tentry->end - offset;
2776 object = tentry->object.vm_object;
2777 offset = tentry->offset + (offset - tentry->start);
2778 vm_map_unlock_read(smap);
2781 case VM_MAPTYPE_NORMAL:
2782 case VM_MAPTYPE_VPAGETABLE:
2783 object = current->object.vm_object;
2791 vm_object_hold(object);
2794 * Note that there is absolutely no sense in writing out
2795 * anonymous objects, so we track down the vnode object
2797 * We invalidate (remove) all pages from the address space
2798 * anyway, for semantic correctness.
2800 * note: certain anonymous maps, such as MAP_NOSYNC maps,
2801 * may start out with a NULL object.
2803 while (object && (tobj = object->backing_object) != NULL) {
2804 vm_object_hold(tobj);
2805 if (tobj == object->backing_object) {
2806 vm_object_lock_swap();
2807 offset += object->backing_object_offset;
2808 vm_object_drop(object);
2810 if (object->size < OFF_TO_IDX(offset + size))
2811 size = IDX_TO_OFF(object->size) -
2815 vm_object_drop(tobj);
2817 if (object && (object->type == OBJT_VNODE) &&
2818 (current->protection & VM_PROT_WRITE) &&
2819 (object->flags & OBJ_NOMSYNC) == 0) {
2821 * Flush pages if writing is allowed, invalidate them
2822 * if invalidation requested. Pages undergoing I/O
2823 * will be ignored by vm_object_page_remove().
2825 * We cannot lock the vnode and then wait for paging
2826 * to complete without deadlocking against vm_fault.
2827 * Instead we simply call vm_object_page_remove() and
2828 * allow it to block internally on a page-by-page
2829 * basis when it encounters pages undergoing async
2834 /* no chain wait needed for vnode objects */
2835 vm_object_reference_locked(object);
2836 vn_lock(object->handle, LK_EXCLUSIVE | LK_RETRY);
2837 flags = (syncio || invalidate) ? OBJPC_SYNC : 0;
2838 flags |= invalidate ? OBJPC_INVAL : 0;
2841 * When operating on a virtual page table just
2842 * flush the whole object. XXX we probably ought
2845 switch(current->maptype) {
2846 case VM_MAPTYPE_NORMAL:
2847 vm_object_page_clean(object,
2849 OFF_TO_IDX(offset + size + PAGE_MASK),
2852 case VM_MAPTYPE_VPAGETABLE:
2853 vm_object_page_clean(object, 0, 0, flags);
2856 vn_unlock(((struct vnode *)object->handle));
2857 vm_object_deallocate_locked(object);
2859 if (object && invalidate &&
2860 ((object->type == OBJT_VNODE) ||
2861 (object->type == OBJT_DEVICE) ||
2862 (object->type == OBJT_MGTDEVICE))) {
2864 ((object->type == OBJT_DEVICE) ||
2865 (object->type == OBJT_MGTDEVICE)) ? FALSE : TRUE;
2866 /* no chain wait needed for vnode/device objects */
2867 vm_object_reference_locked(object);
2868 switch(current->maptype) {
2869 case VM_MAPTYPE_NORMAL:
2870 vm_object_page_remove(object,
2872 OFF_TO_IDX(offset + size + PAGE_MASK),
2875 case VM_MAPTYPE_VPAGETABLE:
2876 vm_object_page_remove(object, 0, 0, clean_only);
2879 vm_object_deallocate_locked(object);
2883 vm_object_drop(object);
2886 lwkt_reltoken(&map->token);
2887 vm_map_unlock_read(map);
2889 return (KERN_SUCCESS);
2893 * Make the region specified by this entry pageable.
2895 * The vm_map must be exclusively locked.
2898 vm_map_entry_unwire(vm_map_t map, vm_map_entry_t entry)
2900 entry->eflags &= ~MAP_ENTRY_USER_WIRED;
2901 entry->wired_count = 0;
2902 vm_fault_unwire(map, entry);
2906 * Deallocate the given entry from the target map.
2908 * The vm_map must be exclusively locked.
2911 vm_map_entry_delete(vm_map_t map, vm_map_entry_t entry, int *countp)
2913 vm_map_entry_unlink(map, entry);
2914 map->size -= entry->end - entry->start;
2916 switch(entry->maptype) {
2917 case VM_MAPTYPE_NORMAL:
2918 case VM_MAPTYPE_VPAGETABLE:
2919 case VM_MAPTYPE_SUBMAP:
2920 vm_object_deallocate(entry->object.vm_object);
2922 case VM_MAPTYPE_UKSMAP:
2929 vm_map_entry_dispose(map, entry, countp);
2933 * Deallocates the given address range from the target map.
2935 * The vm_map must be exclusively locked.
2938 vm_map_delete(vm_map_t map, vm_offset_t start, vm_offset_t end, int *countp)
2941 vm_map_entry_t entry;
2942 vm_map_entry_t first_entry;
2943 vm_offset_t hole_start;
2945 ASSERT_VM_MAP_LOCKED(map);
2946 lwkt_gettoken(&map->token);
2949 * Find the start of the region, and clip it. Set entry to point
2950 * at the first record containing the requested address or, if no
2951 * such record exists, the next record with a greater address. The
2952 * loop will run from this point until a record beyond the termination
2953 * address is encountered.
2955 * Adjust freehint[] for either the clip case or the extension case.
2957 * GGG see other GGG comment.
2959 if (vm_map_lookup_entry(map, start, &first_entry)) {
2960 entry = first_entry;
2961 vm_map_clip_start(map, entry, start, countp);
2964 entry = first_entry->next;
2965 if (entry == &map->header)
2966 hole_start = first_entry->start;
2968 hole_start = first_entry->end;
2972 * Step through all entries in this region
2974 while ((entry != &map->header) && (entry->start < end)) {
2975 vm_map_entry_t next;
2977 vm_pindex_t offidxstart, offidxend, count;
2980 * If we hit an in-transition entry we have to sleep and
2981 * retry. It's easier (and not really slower) to just retry
2982 * since this case occurs so rarely and the hint is already
2983 * pointing at the right place. We have to reset the
2984 * start offset so as not to accidently delete an entry
2985 * another process just created in vacated space.
2987 if (entry->eflags & MAP_ENTRY_IN_TRANSITION) {
2988 entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP;
2989 start = entry->start;
2990 ++mycpu->gd_cnt.v_intrans_coll;
2991 ++mycpu->gd_cnt.v_intrans_wait;
2992 vm_map_transition_wait(map, 1);
2995 vm_map_clip_end(map, entry, end, countp);
3001 offidxstart = OFF_TO_IDX(entry->offset);
3002 count = OFF_TO_IDX(e - s);
3004 switch(entry->maptype) {
3005 case VM_MAPTYPE_NORMAL:
3006 case VM_MAPTYPE_VPAGETABLE:
3007 case VM_MAPTYPE_SUBMAP:
3008 object = entry->object.vm_object;
3016 * Unwire before removing addresses from the pmap; otherwise,
3017 * unwiring will put the entries back in the pmap.
3019 * Generally speaking, doing a bulk pmap_remove() before
3020 * removing the pages from the VM object is better at
3021 * reducing unnecessary IPIs. The pmap code is now optimized
3022 * to not blindly iterate the range when pt and pd pages
3025 if (entry->wired_count != 0)
3026 vm_map_entry_unwire(map, entry);
3028 offidxend = offidxstart + count;
3030 if (object == &kernel_object) {
3031 pmap_remove(map->pmap, s, e);
3032 vm_object_hold(object);
3033 vm_object_page_remove(object, offidxstart,
3035 vm_object_drop(object);
3036 } else if (object && object->type != OBJT_DEFAULT &&
3037 object->type != OBJT_SWAP) {
3039 * vnode object routines cannot be chain-locked,
3040 * but since we aren't removing pages from the
3041 * object here we can use a shared hold.
3043 vm_object_hold_shared(object);
3044 pmap_remove(map->pmap, s, e);
3045 vm_object_drop(object);
3046 } else if (object) {
3047 vm_object_hold(object);
3048 vm_object_chain_acquire(object, 0);
3049 pmap_remove(map->pmap, s, e);
3051 if (object != NULL &&
3052 object->ref_count != 1 &&
3053 (object->flags & (OBJ_NOSPLIT|OBJ_ONEMAPPING)) ==
3055 (object->type == OBJT_DEFAULT ||
3056 object->type == OBJT_SWAP)) {
3058 * When ONEMAPPING is set we can destroy the
3059 * pages underlying the entry's range.
3061 vm_object_collapse(object, NULL);
3062 vm_object_page_remove(object, offidxstart,
3064 if (object->type == OBJT_SWAP) {
3065 swap_pager_freespace(object,
3069 if (offidxend >= object->size &&
3070 offidxstart < object->size) {
3071 object->size = offidxstart;
3074 vm_object_chain_release(object);
3075 vm_object_drop(object);
3076 } else if (entry->maptype == VM_MAPTYPE_UKSMAP) {
3077 pmap_remove(map->pmap, s, e);
3081 * Delete the entry (which may delete the object) only after
3082 * removing all pmap entries pointing to its pages.
3083 * (Otherwise, its page frames may be reallocated, and any
3084 * modify bits will be set in the wrong object!)
3086 vm_map_entry_delete(map, entry, countp);
3089 if (entry == &map->header)
3090 vm_map_freehint_hole(map, hole_start, entry->end - hole_start);
3092 vm_map_freehint_hole(map, hole_start,
3093 entry->start - hole_start);
3095 lwkt_reltoken(&map->token);
3097 return (KERN_SUCCESS);
3101 * Remove the given address range from the target map.
3102 * This is the exported form of vm_map_delete.
3107 vm_map_remove(vm_map_t map, vm_offset_t start, vm_offset_t end)
3112 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
3114 VM_MAP_RANGE_CHECK(map, start, end);
3115 result = vm_map_delete(map, start, end, &count);
3117 vm_map_entry_release(count);
3123 * Assert that the target map allows the specified privilege on the
3124 * entire address region given. The entire region must be allocated.
3126 * The caller must specify whether the vm_map is already locked or not.
3129 vm_map_check_protection(vm_map_t map, vm_offset_t start, vm_offset_t end,
3130 vm_prot_t protection, boolean_t have_lock)
3132 vm_map_entry_t entry;
3133 vm_map_entry_t tmp_entry;
3136 if (have_lock == FALSE)
3137 vm_map_lock_read(map);
3139 if (!vm_map_lookup_entry(map, start, &tmp_entry)) {
3140 if (have_lock == FALSE)
3141 vm_map_unlock_read(map);
3147 while (start < end) {
3148 if (entry == &map->header) {
3156 if (start < entry->start) {
3161 * Check protection associated with entry.
3164 if ((entry->protection & protection) != protection) {
3168 /* go to next entry */
3171 entry = entry->next;
3173 if (have_lock == FALSE)
3174 vm_map_unlock_read(map);
3179 * If appropriate this function shadows the original object with a new object
3180 * and moves the VM pages from the original object to the new object.
3181 * The original object will also be collapsed, if possible.
3183 * Caller must supply entry->object.vm_object held and chain_acquired, and
3184 * should chain_release and drop the object upon return.
3186 * We can only do this for normal memory objects with a single mapping, and
3187 * it only makes sense to do it if there are 2 or more refs on the original
3188 * object. i.e. typically a memory object that has been extended into
3189 * multiple vm_map_entry's with non-overlapping ranges.
3191 * This makes it easier to remove unused pages and keeps object inheritance
3192 * from being a negative impact on memory usage.
3194 * On return the (possibly new) entry->object.vm_object will have an
3195 * additional ref on it for the caller to dispose of (usually by cloning
3196 * the vm_map_entry). The additional ref had to be done in this routine
3197 * to avoid racing a collapse. The object's ONEMAPPING flag will also be
3200 * The vm_map must be locked and its token held.
3203 vm_map_split(vm_map_entry_t entry, vm_object_t oobject)
3206 vm_object_t nobject, bobject;
3209 vm_pindex_t offidxstart, offidxend, idx;
3211 vm_ooffset_t offset;
3215 * Optimize away object locks for vnode objects. Important exit/exec
3218 * OBJ_ONEMAPPING doesn't apply to vnode objects but clear the flag
3221 if (oobject->type != OBJT_DEFAULT && oobject->type != OBJT_SWAP) {
3222 vm_object_reference_quick(oobject);
3223 vm_object_clear_flag(oobject, OBJ_ONEMAPPING);
3229 * Original object cannot be split?
3231 if (oobject->handle == NULL) {
3232 vm_object_reference_locked_chain_held(oobject);
3233 vm_object_clear_flag(oobject, OBJ_ONEMAPPING);
3239 * Collapse original object with its backing store as an
3240 * optimization to reduce chain lengths when possible.
3242 * If ref_count <= 1 there aren't other non-overlapping vm_map_entry's
3243 * for oobject, so there's no point collapsing it.
3245 * Then re-check whether the object can be split.
3247 vm_object_collapse(oobject, NULL);
3249 if (oobject->ref_count <= 1 ||
3250 (oobject->type != OBJT_DEFAULT && oobject->type != OBJT_SWAP) ||
3251 (oobject->flags & (OBJ_NOSPLIT|OBJ_ONEMAPPING)) != OBJ_ONEMAPPING) {
3252 vm_object_reference_locked_chain_held(oobject);
3253 vm_object_clear_flag(oobject, OBJ_ONEMAPPING);
3258 * Acquire the chain lock on the backing object.
3260 * Give bobject an additional ref count for when it will be shadowed
3264 if ((bobject = oobject->backing_object) != NULL) {
3265 if (bobject->type != OBJT_VNODE) {
3267 vm_object_hold(bobject);
3268 vm_object_chain_wait(bobject, 0);
3269 /* ref for shadowing below */
3270 vm_object_reference_locked(bobject);
3271 vm_object_chain_acquire(bobject, 0);
3272 KKASSERT(oobject->backing_object == bobject);
3273 KKASSERT((bobject->flags & OBJ_DEAD) == 0);
3276 * vnodes are not placed on the shadow list but
3277 * they still get another ref for the backing_object
3280 vm_object_reference_quick(bobject);
3285 * Calculate the object page range and allocate the new object.
3287 offset = entry->offset;
3291 offidxstart = OFF_TO_IDX(offset);
3292 offidxend = offidxstart + OFF_TO_IDX(e - s);
3293 size = offidxend - offidxstart;
3295 switch(oobject->type) {
3297 nobject = default_pager_alloc(NULL, IDX_TO_OFF(size),
3301 nobject = swap_pager_alloc(NULL, IDX_TO_OFF(size),
3311 * If we could not allocate nobject just clear ONEMAPPING on
3312 * oobject and return.
3314 if (nobject == NULL) {
3316 if (useshadowlist) {
3317 vm_object_chain_release(bobject);
3318 vm_object_deallocate(bobject);
3319 vm_object_drop(bobject);
3321 vm_object_deallocate(bobject);
3324 vm_object_reference_locked_chain_held(oobject);
3325 vm_object_clear_flag(oobject, OBJ_ONEMAPPING);
3330 * The new object will replace entry->object.vm_object so it needs
3331 * a second reference (the caller expects an additional ref).
3333 vm_object_hold(nobject);
3334 vm_object_reference_locked(nobject);
3335 vm_object_chain_acquire(nobject, 0);
3338 * nobject shadows bobject (oobject already shadows bobject).
3340 * Adding an object to bobject's shadow list requires refing bobject
3341 * which we did above in the useshadowlist case.
3343 * XXX it is unclear if we need to clear ONEMAPPING on bobject here
3347 nobject->backing_object_offset =
3348 oobject->backing_object_offset + IDX_TO_OFF(offidxstart);
3349 nobject->backing_object = bobject;
3350 if (useshadowlist) {
3351 bobject->shadow_count++;
3352 atomic_add_int(&bobject->generation, 1);
3353 LIST_INSERT_HEAD(&bobject->shadow_head,
3354 nobject, shadow_list);
3355 vm_object_clear_flag(bobject, OBJ_ONEMAPPING); /*XXX*/
3356 vm_object_set_flag(nobject, OBJ_ONSHADOW);
3361 * Move the VM pages from oobject to nobject
3363 for (idx = 0; idx < size; idx++) {
3366 m = vm_page_lookup_busy_wait(oobject, offidxstart + idx,
3372 * We must wait for pending I/O to complete before we can
3375 * We do not have to VM_PROT_NONE the page as mappings should
3376 * not be changed by this operation.
3378 * NOTE: The act of renaming a page updates chaingen for both
3381 vm_page_rename(m, nobject, idx);
3382 /* page automatically made dirty by rename and cache handled */
3383 /* page remains busy */
3386 if (oobject->type == OBJT_SWAP) {
3387 vm_object_pip_add(oobject, 1);
3389 * copy oobject pages into nobject and destroy unneeded
3390 * pages in shadow object.
3392 swap_pager_copy(oobject, nobject, offidxstart, 0);
3393 vm_object_pip_wakeup(oobject);
3397 * Wakeup the pages we played with. No spl protection is needed
3398 * for a simple wakeup.
3400 for (idx = 0; idx < size; idx++) {
3401 m = vm_page_lookup(nobject, idx);
3403 KKASSERT(m->busy_count & PBUSY_LOCKED);
3407 entry->object.vm_object = nobject;
3408 entry->offset = 0LL;
3411 * The map is being split and nobject is going to wind up on both
3412 * vm_map_entry's, so make sure OBJ_ONEMAPPING is cleared on
3415 vm_object_clear_flag(nobject, OBJ_ONEMAPPING);
3420 * NOTE: There is no need to remove OBJ_ONEMAPPING from oobject, the
3421 * related pages were moved and are no longer applicable to the
3424 * NOTE: Deallocate oobject (due to its entry->object.vm_object being
3425 * replaced by nobject).
3427 vm_object_chain_release(nobject);
3428 vm_object_drop(nobject);
3429 if (bobject && useshadowlist) {
3430 vm_object_chain_release(bobject);
3431 vm_object_drop(bobject);
3435 if (oobject->resident_page_count) {
3436 kprintf("oobject %p still contains %jd pages!\n",
3437 oobject, (intmax_t)oobject->resident_page_count);
3438 for (idx = 0; idx < size; idx++) {
3441 m = vm_page_lookup_busy_wait(oobject, offidxstart + idx,
3444 kprintf("oobject %p idx %jd\n",
3452 /*vm_object_clear_flag(oobject, OBJ_ONEMAPPING);*/
3453 vm_object_deallocate_locked(oobject);
3457 * Copies the contents of the source entry to the destination
3458 * entry. The entries *must* be aligned properly.
3460 * The vm_maps must be exclusively locked.
3461 * The vm_map's token must be held.
3463 * Because the maps are locked no faults can be in progress during the
3467 vm_map_copy_entry(vm_map_t src_map, vm_map_t dst_map,
3468 vm_map_entry_t src_entry, vm_map_entry_t dst_entry)
3470 vm_object_t src_object;
3471 vm_object_t oobject;
3473 if (dst_entry->maptype == VM_MAPTYPE_SUBMAP ||
3474 dst_entry->maptype == VM_MAPTYPE_UKSMAP)
3476 if (src_entry->maptype == VM_MAPTYPE_SUBMAP ||
3477 src_entry->maptype == VM_MAPTYPE_UKSMAP)
3480 if (src_entry->wired_count == 0) {
3482 * If the source entry is marked needs_copy, it is already
3485 * To avoid interacting with a vm_fault that might have
3486 * released its vm_map, we must acquire the fronting
3489 oobject = src_entry->object.vm_object;
3491 vm_object_hold(oobject);
3492 vm_object_chain_acquire(oobject, 0);
3495 if ((src_entry->eflags & MAP_ENTRY_NEEDS_COPY) == 0) {
3496 pmap_protect(src_map->pmap,
3499 src_entry->protection & ~VM_PROT_WRITE);
3503 * Make a copy of the object.
3505 * The object must be locked prior to checking the object type
3506 * and for the call to vm_object_collapse() and vm_map_split().
3507 * We cannot use *_hold() here because the split code will
3508 * probably try to destroy the object. The lock is a pool
3509 * token and doesn't care.
3511 * We must bump src_map->timestamp when setting
3512 * MAP_ENTRY_NEEDS_COPY to force any concurrent fault
3513 * to retry, otherwise the concurrent fault might improperly
3514 * install a RW pte when its supposed to be a RO(COW) pte.
3515 * This race can occur because a vnode-backed fault may have
3516 * to temporarily release the map lock. This was handled
3517 * when the caller locked the map exclusively.
3520 vm_map_split(src_entry, oobject);
3522 src_object = src_entry->object.vm_object;
3523 dst_entry->object.vm_object = src_object;
3524 src_entry->eflags |= (MAP_ENTRY_COW |
3525 MAP_ENTRY_NEEDS_COPY);
3526 dst_entry->eflags |= (MAP_ENTRY_COW |
3527 MAP_ENTRY_NEEDS_COPY);
3528 dst_entry->offset = src_entry->offset;
3530 dst_entry->object.vm_object = NULL;
3531 dst_entry->offset = 0;
3533 pmap_copy(dst_map->pmap, src_map->pmap, dst_entry->start,
3534 dst_entry->end - dst_entry->start,
3537 vm_object_chain_release(oobject);
3538 vm_object_drop(oobject);
3542 * Of course, wired down pages can't be set copy-on-write.
3543 * Cause wired pages to be copied into the new map by
3544 * simulating faults (the new pages are pageable)
3546 vm_fault_copy_entry(dst_map, src_map, dst_entry, src_entry);
3552 * Create a new process vmspace structure and vm_map
3553 * based on those of an existing process. The new map
3554 * is based on the old map, according to the inheritance
3555 * values on the regions in that map.
3557 * The source map must not be locked.
3560 static void vmspace_fork_normal_entry(vm_map_t old_map, vm_map_t new_map,
3561 vm_map_entry_t old_entry, int *countp);
3562 static void vmspace_fork_uksmap_entry(vm_map_t old_map, vm_map_t new_map,
3563 vm_map_entry_t old_entry, int *countp);
3566 vmspace_fork(struct vmspace *vm1)
3568 struct vmspace *vm2;
3569 vm_map_t old_map = &vm1->vm_map;
3571 vm_map_entry_t old_entry;
3574 lwkt_gettoken(&vm1->vm_map.token);
3575 vm_map_lock(old_map);
3577 vm2 = vmspace_alloc(old_map->header.start, old_map->header.end);
3578 lwkt_gettoken(&vm2->vm_map.token);
3581 * We must bump the timestamp to force any concurrent fault
3584 bcopy(&vm1->vm_startcopy, &vm2->vm_startcopy,
3585 (caddr_t)&vm1->vm_endcopy - (caddr_t)&vm1->vm_startcopy);
3586 new_map = &vm2->vm_map; /* XXX */
3587 new_map->timestamp = 1;
3589 vm_map_lock(new_map);
3592 old_entry = old_map->header.next;
3593 while (old_entry != &old_map->header) {
3595 old_entry = old_entry->next;
3598 count = vm_map_entry_reserve(count + MAP_RESERVE_COUNT);
3600 old_entry = old_map->header.next;
3601 while (old_entry != &old_map->header) {
3602 switch(old_entry->maptype) {
3603 case VM_MAPTYPE_SUBMAP:
3604 panic("vm_map_fork: encountered a submap");
3606 case VM_MAPTYPE_UKSMAP:
3607 vmspace_fork_uksmap_entry(old_map, new_map,
3610 case VM_MAPTYPE_NORMAL:
3611 case VM_MAPTYPE_VPAGETABLE:
3612 vmspace_fork_normal_entry(old_map, new_map,
3616 old_entry = old_entry->next;
3619 new_map->size = old_map->size;
3620 vm_map_unlock(old_map);
3621 vm_map_unlock(new_map);
3622 vm_map_entry_release(count);
3624 lwkt_reltoken(&vm2->vm_map.token);
3625 lwkt_reltoken(&vm1->vm_map.token);
3632 vmspace_fork_normal_entry(vm_map_t old_map, vm_map_t new_map,
3633 vm_map_entry_t old_entry, int *countp)
3635 vm_map_entry_t new_entry;
3638 switch (old_entry->inheritance) {
3639 case VM_INHERIT_NONE:
3641 case VM_INHERIT_SHARE:
3643 * Clone the entry, creating the shared object if
3646 if (old_entry->object.vm_object == NULL)
3647 vm_map_entry_allocate_object(old_entry);
3649 if (old_entry->eflags & MAP_ENTRY_NEEDS_COPY) {
3651 * Shadow a map_entry which needs a copy,
3652 * replacing its object with a new object
3653 * that points to the old one. Ask the
3654 * shadow code to automatically add an
3655 * additional ref. We can't do it afterwords
3656 * because we might race a collapse. The call
3657 * to vm_map_entry_shadow() will also clear
3660 vm_map_entry_shadow(old_entry, 1);
3661 } else if (old_entry->object.vm_object) {
3663 * We will make a shared copy of the object,
3664 * and must clear OBJ_ONEMAPPING.
3666 * Optimize vnode objects. OBJ_ONEMAPPING
3667 * is non-applicable but clear it anyway,
3668 * and its terminal so we don't have to deal
3669 * with chains. Reduces SMP conflicts.
3671 * XXX assert that object.vm_object != NULL
3672 * since we allocate it above.
3674 object = old_entry->object.vm_object;
3675 if (object->type == OBJT_VNODE) {
3676 vm_object_reference_quick(object);
3677 vm_object_clear_flag(object,
3680 vm_object_hold(object);
3681 vm_object_chain_wait(object, 0);
3682 vm_object_reference_locked(object);
3683 vm_object_clear_flag(object,
3685 vm_object_drop(object);
3690 * Clone the entry. We've already bumped the ref on
3693 new_entry = vm_map_entry_create(new_map, countp);
3694 *new_entry = *old_entry;
3695 new_entry->eflags &= ~MAP_ENTRY_USER_WIRED;
3696 new_entry->wired_count = 0;
3699 * Insert the entry into the new map -- we know we're
3700 * inserting at the end of the new map.
3702 vm_map_entry_link(new_map, new_map->header.prev,
3706 * Update the physical map
3708 pmap_copy(new_map->pmap, old_map->pmap,
3710 (old_entry->end - old_entry->start),
3713 case VM_INHERIT_COPY:
3715 * Clone the entry and link into the map.
3717 new_entry = vm_map_entry_create(new_map, countp);
3718 *new_entry = *old_entry;
3719 new_entry->eflags &= ~MAP_ENTRY_USER_WIRED;
3720 new_entry->wired_count = 0;
3721 new_entry->object.vm_object = NULL;
3722 vm_map_entry_link(new_map, new_map->header.prev,
3724 vm_map_copy_entry(old_map, new_map, old_entry,
3731 * When forking user-kernel shared maps, the map might change in the
3732 * child so do not try to copy the underlying pmap entries.
3736 vmspace_fork_uksmap_entry(vm_map_t old_map, vm_map_t new_map,
3737 vm_map_entry_t old_entry, int *countp)
3739 vm_map_entry_t new_entry;
3741 new_entry = vm_map_entry_create(new_map, countp);
3742 *new_entry = *old_entry;
3743 new_entry->eflags &= ~MAP_ENTRY_USER_WIRED;
3744 new_entry->wired_count = 0;
3745 vm_map_entry_link(new_map, new_map->header.prev,
3750 * Create an auto-grow stack entry
3755 vm_map_stack (vm_map_t map, vm_offset_t addrbos, vm_size_t max_ssize,
3756 int flags, vm_prot_t prot, vm_prot_t max, int cow)
3758 vm_map_entry_t prev_entry;
3759 vm_map_entry_t new_stack_entry;
3760 vm_size_t init_ssize;
3763 vm_offset_t tmpaddr;
3765 cow |= MAP_IS_STACK;
3767 if (max_ssize < sgrowsiz)
3768 init_ssize = max_ssize;
3770 init_ssize = sgrowsiz;
3772 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
3776 * Find space for the mapping
3778 if ((flags & (MAP_FIXED | MAP_TRYFIXED)) == 0) {
3779 if (vm_map_findspace(map, addrbos, max_ssize, 1,
3782 vm_map_entry_release(count);
3783 return (KERN_NO_SPACE);
3788 /* If addr is already mapped, no go */
3789 if (vm_map_lookup_entry(map, addrbos, &prev_entry)) {
3791 vm_map_entry_release(count);
3792 return (KERN_NO_SPACE);
3796 /* XXX already handled by kern_mmap() */
3797 /* If we would blow our VMEM resource limit, no go */
3798 if (map->size + init_ssize >
3799 curproc->p_rlimit[RLIMIT_VMEM].rlim_cur) {
3801 vm_map_entry_release(count);
3802 return (KERN_NO_SPACE);
3807 * If we can't accomodate max_ssize in the current mapping,
3808 * no go. However, we need to be aware that subsequent user
3809 * mappings might map into the space we have reserved for
3810 * stack, and currently this space is not protected.
3812 * Hopefully we will at least detect this condition
3813 * when we try to grow the stack.
3815 if ((prev_entry->next != &map->header) &&
3816 (prev_entry->next->start < addrbos + max_ssize)) {
3818 vm_map_entry_release(count);
3819 return (KERN_NO_SPACE);
3823 * We initially map a stack of only init_ssize. We will
3824 * grow as needed later. Since this is to be a grow
3825 * down stack, we map at the top of the range.
3827 * Note: we would normally expect prot and max to be
3828 * VM_PROT_ALL, and cow to be 0. Possibly we should
3829 * eliminate these as input parameters, and just
3830 * pass these values here in the insert call.
3832 rv = vm_map_insert(map, &count, NULL, NULL,
3833 0, addrbos + max_ssize - init_ssize,
3834 addrbos + max_ssize,
3836 VM_SUBSYS_STACK, prot, max, cow);
3838 /* Now set the avail_ssize amount */
3839 if (rv == KERN_SUCCESS) {
3840 if (prev_entry != &map->header)
3841 vm_map_clip_end(map, prev_entry, addrbos + max_ssize - init_ssize, &count);
3842 new_stack_entry = prev_entry->next;
3843 if (new_stack_entry->end != addrbos + max_ssize ||
3844 new_stack_entry->start != addrbos + max_ssize - init_ssize)
3845 panic ("Bad entry start/end for new stack entry");
3847 new_stack_entry->aux.avail_ssize = max_ssize - init_ssize;
3851 vm_map_entry_release(count);
3856 * Attempts to grow a vm stack entry. Returns KERN_SUCCESS if the
3857 * desired address is already mapped, or if we successfully grow
3858 * the stack. Also returns KERN_SUCCESS if addr is outside the
3859 * stack range (this is strange, but preserves compatibility with
3860 * the grow function in vm_machdep.c).
3865 vm_map_growstack (vm_map_t map, vm_offset_t addr)
3867 vm_map_entry_t prev_entry;
3868 vm_map_entry_t stack_entry;
3869 vm_map_entry_t new_stack_entry;
3875 int rv = KERN_SUCCESS;
3877 int use_read_lock = 1;
3883 lp = curthread->td_lwp;
3884 p = curthread->td_proc;
3885 KKASSERT(lp != NULL);
3886 vm = lp->lwp_vmspace;
3889 * Growstack is only allowed on the current process. We disallow
3890 * other use cases, e.g. trying to access memory via procfs that
3891 * the stack hasn't grown into.
3893 if (map != &vm->vm_map) {
3894 return KERN_FAILURE;
3897 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
3900 vm_map_lock_read(map);
3904 /* If addr is already in the entry range, no need to grow.*/
3905 if (vm_map_lookup_entry(map, addr, &prev_entry))
3908 if ((stack_entry = prev_entry->next) == &map->header)
3910 if (prev_entry == &map->header)
3911 end = stack_entry->start - stack_entry->aux.avail_ssize;
3913 end = prev_entry->end;
3916 * This next test mimics the old grow function in vm_machdep.c.
3917 * It really doesn't quite make sense, but we do it anyway
3918 * for compatibility.
3920 * If not growable stack, return success. This signals the
3921 * caller to proceed as he would normally with normal vm.
3923 if (stack_entry->aux.avail_ssize < 1 ||
3924 addr >= stack_entry->start ||
3925 addr < stack_entry->start - stack_entry->aux.avail_ssize) {
3929 /* Find the minimum grow amount */
3930 grow_amount = roundup (stack_entry->start - addr, PAGE_SIZE);
3931 if (grow_amount > stack_entry->aux.avail_ssize) {
3937 * If there is no longer enough space between the entries
3938 * nogo, and adjust the available space. Note: this
3939 * should only happen if the user has mapped into the
3940 * stack area after the stack was created, and is
3941 * probably an error.
3943 * This also effectively destroys any guard page the user
3944 * might have intended by limiting the stack size.
3946 if (grow_amount > stack_entry->start - end) {
3947 if (use_read_lock && vm_map_lock_upgrade(map)) {
3953 stack_entry->aux.avail_ssize = stack_entry->start - end;
3958 is_procstack = addr >= (vm_offset_t)vm->vm_maxsaddr;
3960 /* If this is the main process stack, see if we're over the
3963 if (is_procstack && (vm->vm_ssize + grow_amount >
3964 p->p_rlimit[RLIMIT_STACK].rlim_cur)) {
3969 /* Round up the grow amount modulo SGROWSIZ */
3970 grow_amount = roundup (grow_amount, sgrowsiz);
3971 if (grow_amount > stack_entry->aux.avail_ssize) {
3972 grow_amount = stack_entry->aux.avail_ssize;
3974 if (is_procstack && (vm->vm_ssize + grow_amount >
3975 p->p_rlimit[RLIMIT_STACK].rlim_cur)) {
3976 grow_amount = p->p_rlimit[RLIMIT_STACK].rlim_cur - vm->vm_ssize;
3979 /* If we would blow our VMEM resource limit, no go */
3980 if (map->size + grow_amount > p->p_rlimit[RLIMIT_VMEM].rlim_cur) {
3985 if (use_read_lock && vm_map_lock_upgrade(map)) {
3992 /* Get the preliminary new entry start value */
3993 addr = stack_entry->start - grow_amount;
3995 /* If this puts us into the previous entry, cut back our growth
3996 * to the available space. Also, see the note above.
3999 stack_entry->aux.avail_ssize = stack_entry->start - end;
4003 rv = vm_map_insert(map, &count, NULL, NULL,
4004 0, addr, stack_entry->start,
4006 VM_SUBSYS_STACK, VM_PROT_ALL, VM_PROT_ALL, 0);
4008 /* Adjust the available stack space by the amount we grew. */
4009 if (rv == KERN_SUCCESS) {
4010 if (prev_entry != &map->header)
4011 vm_map_clip_end(map, prev_entry, addr, &count);
4012 new_stack_entry = prev_entry->next;
4013 if (new_stack_entry->end != stack_entry->start ||
4014 new_stack_entry->start != addr)
4015 panic ("Bad stack grow start/end in new stack entry");
4017 new_stack_entry->aux.avail_ssize =
4018 stack_entry->aux.avail_ssize -
4019 (new_stack_entry->end - new_stack_entry->start);
4021 vm->vm_ssize += new_stack_entry->end -
4022 new_stack_entry->start;
4026 if (map->flags & MAP_WIREFUTURE)
4027 vm_map_unwire(map, new_stack_entry->start,
4028 new_stack_entry->end, FALSE);
4033 vm_map_unlock_read(map);
4036 vm_map_entry_release(count);
4041 * Unshare the specified VM space for exec. If other processes are
4042 * mapped to it, then create a new one. The new vmspace is null.
4047 vmspace_exec(struct proc *p, struct vmspace *vmcopy)
4049 struct vmspace *oldvmspace = p->p_vmspace;
4050 struct vmspace *newvmspace;
4051 vm_map_t map = &p->p_vmspace->vm_map;
4054 * If we are execing a resident vmspace we fork it, otherwise
4055 * we create a new vmspace. Note that exitingcnt is not
4056 * copied to the new vmspace.
4058 lwkt_gettoken(&oldvmspace->vm_map.token);
4060 newvmspace = vmspace_fork(vmcopy);
4061 lwkt_gettoken(&newvmspace->vm_map.token);
4063 newvmspace = vmspace_alloc(map->header.start, map->header.end);
4064 lwkt_gettoken(&newvmspace->vm_map.token);
4065 bcopy(&oldvmspace->vm_startcopy, &newvmspace->vm_startcopy,
4066 (caddr_t)&oldvmspace->vm_endcopy -
4067 (caddr_t)&oldvmspace->vm_startcopy);
4071 * Finish initializing the vmspace before assigning it
4072 * to the process. The vmspace will become the current vmspace
4075 pmap_pinit2(vmspace_pmap(newvmspace));
4076 pmap_replacevm(p, newvmspace, 0);
4077 lwkt_reltoken(&newvmspace->vm_map.token);
4078 lwkt_reltoken(&oldvmspace->vm_map.token);
4079 vmspace_rel(oldvmspace);
4083 * Unshare the specified VM space for forcing COW. This
4084 * is called by rfork, for the (RFMEM|RFPROC) == 0 case.
4087 vmspace_unshare(struct proc *p)
4089 struct vmspace *oldvmspace = p->p_vmspace;
4090 struct vmspace *newvmspace;
4092 lwkt_gettoken(&oldvmspace->vm_map.token);
4093 if (vmspace_getrefs(oldvmspace) == 1) {
4094 lwkt_reltoken(&oldvmspace->vm_map.token);
4097 newvmspace = vmspace_fork(oldvmspace);
4098 lwkt_gettoken(&newvmspace->vm_map.token);
4099 pmap_pinit2(vmspace_pmap(newvmspace));
4100 pmap_replacevm(p, newvmspace, 0);
4101 lwkt_reltoken(&newvmspace->vm_map.token);
4102 lwkt_reltoken(&oldvmspace->vm_map.token);
4103 vmspace_rel(oldvmspace);
4107 * vm_map_hint: return the beginning of the best area suitable for
4108 * creating a new mapping with "prot" protection.
4113 vm_map_hint(struct proc *p, vm_offset_t addr, vm_prot_t prot)
4115 struct vmspace *vms = p->p_vmspace;
4116 struct rlimit limit;
4120 * Acquire datasize limit for mmap() operation,
4121 * calculate nearest power of 2.
4123 if (kern_getrlimit(RLIMIT_DATA, &limit))
4124 limit.rlim_cur = maxdsiz;
4125 dsiz = limit.rlim_cur;
4127 if (!randomize_mmap || addr != 0) {
4129 * Set a reasonable start point for the hint if it was
4130 * not specified or if it falls within the heap space.
4131 * Hinted mmap()s do not allocate out of the heap space.
4134 (addr >= round_page((vm_offset_t)vms->vm_taddr) &&
4135 addr < round_page((vm_offset_t)vms->vm_daddr + dsiz))) {
4136 addr = round_page((vm_offset_t)vms->vm_daddr + dsiz);
4143 * randomize_mmap && addr == 0. For now randomize the
4144 * address within a dsiz range beyond the data limit.
4146 addr = (vm_offset_t)vms->vm_daddr + dsiz;
4148 addr += (karc4random64() & 0x7FFFFFFFFFFFFFFFLU) % dsiz;
4149 return (round_page(addr));
4153 * Finds the VM object, offset, and protection for a given virtual address
4154 * in the specified map, assuming a page fault of the type specified.
4156 * Leaves the map in question locked for read; return values are guaranteed
4157 * until a vm_map_lookup_done call is performed. Note that the map argument
4158 * is in/out; the returned map must be used in the call to vm_map_lookup_done.
4160 * A handle (out_entry) is returned for use in vm_map_lookup_done, to make
4163 * If a lookup is requested with "write protection" specified, the map may
4164 * be changed to perform virtual copying operations, although the data
4165 * referenced will remain the same.
4170 vm_map_lookup(vm_map_t *var_map, /* IN/OUT */
4172 vm_prot_t fault_typea,
4173 vm_map_entry_t *out_entry, /* OUT */
4174 vm_object_t *object, /* OUT */
4175 vm_pindex_t *pindex, /* OUT */
4176 vm_prot_t *out_prot, /* OUT */
4177 int *wflags) /* OUT */
4179 vm_map_entry_t entry;
4180 vm_map_t map = *var_map;
4182 vm_prot_t fault_type = fault_typea;
4183 int use_read_lock = 1;
4184 int rv = KERN_SUCCESS;
4186 thread_t td = curthread;
4189 * vm_map_entry_reserve() implements an important mitigation
4190 * against mmap() span running the kernel out of vm_map_entry
4191 * structures, but it can also cause an infinite call recursion.
4192 * Use td_nest_count to prevent an infinite recursion (allows
4193 * the vm_map code to dig into the pcpu vm_map_entry reserve).
4196 if (td->td_nest_count == 0) {
4197 ++td->td_nest_count;
4198 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
4199 --td->td_nest_count;
4203 vm_map_lock_read(map);
4208 * Always do a full lookup. The hint doesn't get us much anymore
4209 * now that the map is RB'd.
4212 *out_entry = &map->header;
4216 vm_map_entry_t tmp_entry;
4218 if (!vm_map_lookup_entry(map, vaddr, &tmp_entry)) {
4219 rv = KERN_INVALID_ADDRESS;
4229 if (entry->maptype == VM_MAPTYPE_SUBMAP) {
4230 vm_map_t old_map = map;
4232 *var_map = map = entry->object.sub_map;
4234 vm_map_unlock_read(old_map);
4236 vm_map_unlock(old_map);
4242 * Check whether this task is allowed to have this page.
4243 * Note the special case for MAP_ENTRY_COW pages with an override.
4244 * This is to implement a forced COW for debuggers.
4246 if (fault_type & VM_PROT_OVERRIDE_WRITE)
4247 prot = entry->max_protection;
4249 prot = entry->protection;
4251 fault_type &= (VM_PROT_READ|VM_PROT_WRITE|VM_PROT_EXECUTE);
4252 if ((fault_type & prot) != fault_type) {
4253 rv = KERN_PROTECTION_FAILURE;
4257 if ((entry->eflags & MAP_ENTRY_USER_WIRED) &&
4258 (entry->eflags & MAP_ENTRY_COW) &&
4259 (fault_type & VM_PROT_WRITE) &&
4260 (fault_typea & VM_PROT_OVERRIDE_WRITE) == 0) {
4261 rv = KERN_PROTECTION_FAILURE;
4266 * If this page is not pageable, we have to get it for all possible
4270 if (entry->wired_count) {
4271 *wflags |= FW_WIRED;
4272 prot = fault_type = entry->protection;
4276 * Virtual page tables may need to update the accessed (A) bit
4277 * in a page table entry. Upgrade the fault to a write fault for
4278 * that case if the map will support it. If the map does not support
4279 * it the page table entry simply will not be updated.
4281 if (entry->maptype == VM_MAPTYPE_VPAGETABLE) {
4282 if (prot & VM_PROT_WRITE)
4283 fault_type |= VM_PROT_WRITE;
4286 if (curthread->td_lwp && curthread->td_lwp->lwp_vmspace &&
4287 pmap_emulate_ad_bits(&curthread->td_lwp->lwp_vmspace->vm_pmap)) {
4288 if ((prot & VM_PROT_WRITE) == 0)
4289 fault_type |= VM_PROT_WRITE;
4293 * Only NORMAL and VPAGETABLE maps are object-based. UKSMAPs are not.
4295 if (entry->maptype != VM_MAPTYPE_NORMAL &&
4296 entry->maptype != VM_MAPTYPE_VPAGETABLE) {
4302 * If the entry was copy-on-write, we either ...
4304 if (entry->eflags & MAP_ENTRY_NEEDS_COPY) {
4306 * If we want to write the page, we may as well handle that
4307 * now since we've got the map locked.
4309 * If we don't need to write the page, we just demote the
4310 * permissions allowed.
4312 if (fault_type & VM_PROT_WRITE) {
4314 * Not allowed if TDF_NOFAULT is set as the shadowing
4315 * operation can deadlock against the faulting
4316 * function due to the copy-on-write.
4318 if (curthread->td_flags & TDF_NOFAULT) {
4319 rv = KERN_FAILURE_NOFAULT;
4324 * Make a new object, and place it in the object
4325 * chain. Note that no new references have appeared
4326 * -- one just moved from the map to the new
4329 if (use_read_lock && vm_map_lock_upgrade(map)) {
4335 vm_map_entry_shadow(entry, 0);
4336 *wflags |= FW_DIDCOW;
4339 * We're attempting to read a copy-on-write page --
4340 * don't allow writes.
4342 prot &= ~VM_PROT_WRITE;
4347 * Create an object if necessary. This code also handles
4348 * partitioning large entries to improve vm_fault performance.
4350 if (entry->object.vm_object == NULL && !map->system_map) {
4351 if (use_read_lock && vm_map_lock_upgrade(map)) {
4359 * Partition large entries, giving each its own VM object,
4360 * to improve concurrent fault performance. This is only
4361 * applicable to userspace.
4363 if (map != &kernel_map &&
4364 entry->maptype == VM_MAPTYPE_NORMAL &&
4365 ((entry->start ^ entry->end) & ~MAP_ENTRY_PARTITION_MASK) &&
4366 vm_map_partition_enable) {
4367 if (entry->eflags & MAP_ENTRY_IN_TRANSITION) {
4368 entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP;
4369 ++mycpu->gd_cnt.v_intrans_coll;
4370 ++mycpu->gd_cnt.v_intrans_wait;
4371 vm_map_transition_wait(map, 0);
4374 vm_map_entry_partition(map, entry, vaddr, &count);
4376 vm_map_entry_allocate_object(entry);
4380 * Return the object/offset from this entry. If the entry was
4381 * copy-on-write or empty, it has been fixed up.
4383 *object = entry->object.vm_object;
4386 *pindex = OFF_TO_IDX((vaddr - entry->start) + entry->offset);
4389 * Return whether this is the only map sharing this data. On
4390 * success we return with a read lock held on the map. On failure
4391 * we return with the map unlocked.
4395 if (rv == KERN_SUCCESS) {
4396 if (use_read_lock == 0)
4397 vm_map_lock_downgrade(map);
4398 } else if (use_read_lock) {
4399 vm_map_unlock_read(map);
4404 vm_map_entry_release(count);
4410 * Releases locks acquired by a vm_map_lookup()
4411 * (according to the handle returned by that lookup).
4413 * No other requirements.
4416 vm_map_lookup_done(vm_map_t map, vm_map_entry_t entry, int count)
4419 * Unlock the main-level map
4421 vm_map_unlock_read(map);
4423 vm_map_entry_release(count);
4427 vm_map_entry_partition(vm_map_t map, vm_map_entry_t entry,
4428 vm_offset_t vaddr, int *countp)
4430 vaddr &= ~MAP_ENTRY_PARTITION_MASK;
4431 vm_map_clip_start(map, entry, vaddr, countp);
4432 vaddr += MAP_ENTRY_PARTITION_SIZE;
4433 vm_map_clip_end(map, entry, vaddr, countp);
4437 * Quick hack, needs some help to make it more SMP friendly.
4440 vm_map_interlock(vm_map_t map, struct vm_map_ilock *ilock,
4441 vm_offset_t ran_beg, vm_offset_t ran_end)
4443 struct vm_map_ilock *scan;
4445 ilock->ran_beg = ran_beg;
4446 ilock->ran_end = ran_end;
4449 spin_lock(&map->ilock_spin);
4451 for (scan = map->ilock_base; scan; scan = scan->next) {
4452 if (ran_end > scan->ran_beg && ran_beg < scan->ran_end) {
4453 scan->flags |= ILOCK_WAITING;
4454 ssleep(scan, &map->ilock_spin, 0, "ilock", 0);
4458 ilock->next = map->ilock_base;
4459 map->ilock_base = ilock;
4460 spin_unlock(&map->ilock_spin);
4464 vm_map_deinterlock(vm_map_t map, struct vm_map_ilock *ilock)
4466 struct vm_map_ilock *scan;
4467 struct vm_map_ilock **scanp;
4469 spin_lock(&map->ilock_spin);
4470 scanp = &map->ilock_base;
4471 while ((scan = *scanp) != NULL) {
4472 if (scan == ilock) {
4473 *scanp = ilock->next;
4474 spin_unlock(&map->ilock_spin);
4475 if (ilock->flags & ILOCK_WAITING)
4479 scanp = &scan->next;
4481 spin_unlock(&map->ilock_spin);
4482 panic("vm_map_deinterlock: missing ilock!");
4485 #include "opt_ddb.h"
4487 #include <ddb/ddb.h>
4492 DB_SHOW_COMMAND(map, vm_map_print)
4495 /* XXX convert args. */
4496 vm_map_t map = (vm_map_t)addr;
4497 boolean_t full = have_addr;
4499 vm_map_entry_t entry;
4501 db_iprintf("Task map %p: pmap=%p, nentries=%d, version=%u\n",
4503 (void *)map->pmap, map->nentries, map->timestamp);
4506 if (!full && db_indent)
4510 for (entry = map->header.next; entry != &map->header;
4511 entry = entry->next) {
4512 db_iprintf("map entry %p: start=%p, end=%p\n",
4513 (void *)entry, (void *)entry->start, (void *)entry->end);
4516 static char *inheritance_name[4] =
4517 {"share", "copy", "none", "donate_copy"};
4519 db_iprintf(" prot=%x/%x/%s",
4521 entry->max_protection,
4522 inheritance_name[(int)(unsigned char)
4523 entry->inheritance]);
4524 if (entry->wired_count != 0)
4525 db_printf(", wired");
4527 switch(entry->maptype) {
4528 case VM_MAPTYPE_SUBMAP:
4529 /* XXX no %qd in kernel. Truncate entry->offset. */
4530 db_printf(", share=%p, offset=0x%lx\n",
4531 (void *)entry->object.sub_map,
4532 (long)entry->offset);
4534 if ((entry->prev == &map->header) ||
4535 (entry->prev->object.sub_map !=
4536 entry->object.sub_map)) {
4538 vm_map_print((db_expr_t)(intptr_t)
4539 entry->object.sub_map,
4544 case VM_MAPTYPE_NORMAL:
4545 case VM_MAPTYPE_VPAGETABLE:
4546 /* XXX no %qd in kernel. Truncate entry->offset. */
4547 db_printf(", object=%p, offset=0x%lx",
4548 (void *)entry->object.vm_object,
4549 (long)entry->offset);
4550 if (entry->eflags & MAP_ENTRY_COW)
4551 db_printf(", copy (%s)",
4552 (entry->eflags & MAP_ENTRY_NEEDS_COPY) ? "needed" : "done");
4556 if ((entry->prev == &map->header) ||
4557 (entry->prev->object.vm_object !=
4558 entry->object.vm_object)) {
4560 vm_object_print((db_expr_t)(intptr_t)
4561 entry->object.vm_object,
4567 case VM_MAPTYPE_UKSMAP:
4568 db_printf(", uksmap=%p, offset=0x%lx",
4569 (void *)entry->object.uksmap,
4570 (long)entry->offset);
4571 if (entry->eflags & MAP_ENTRY_COW)
4572 db_printf(", copy (%s)",
4573 (entry->eflags & MAP_ENTRY_NEEDS_COPY) ? "needed" : "done");
4589 DB_SHOW_COMMAND(procvm, procvm)
4594 p = (struct proc *) addr;
4599 db_printf("p = %p, vmspace = %p, map = %p, pmap = %p\n",
4600 (void *)p, (void *)p->p_vmspace, (void *)&p->p_vmspace->vm_map,
4601 (void *)vmspace_pmap(p->p_vmspace));
4603 vm_map_print((db_expr_t)(intptr_t)&p->p_vmspace->vm_map, 1, 0, NULL);