4 * Copyright (c) 1991, 1993
5 * The Regents of the University of California. All rights reserved.
7 * This code is derived from software contributed to Berkeley by
8 * The Mach Operating System project at Carnegie-Mellon University.
10 * Redistribution and use in source and binary forms, with or without
11 * modification, are permitted provided that the following conditions
13 * 1. Redistributions of source code must retain the above copyright
14 * notice, this list of conditions and the following disclaimer.
15 * 2. Redistributions in binary form must reproduce the above copyright
16 * notice, this list of conditions and the following disclaimer in the
17 * documentation and/or other materials provided with the distribution.
18 * 3. Neither the name of the University nor the names of its contributors
19 * may be used to endorse or promote products derived from this software
20 * without specific prior written permission.
22 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
23 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
24 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
25 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
26 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
27 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
28 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
29 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
30 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
31 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
34 * from: @(#)vm_map.c 8.3 (Berkeley) 1/12/94
37 * Copyright (c) 1987, 1990 Carnegie-Mellon University.
38 * All rights reserved.
40 * Authors: Avadis Tevanian, Jr., Michael Wayne Young
42 * Permission to use, copy, modify and distribute this software and
43 * its documentation is hereby granted, provided that both the copyright
44 * notice and this permission notice appear in all copies of the
45 * software, derivative works or modified versions, and any portions
46 * thereof, and that both notices appear in supporting documentation.
48 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
49 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
50 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
52 * Carnegie Mellon requests users of this software to return to
54 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU
55 * School of Computer Science
56 * Carnegie Mellon University
57 * Pittsburgh PA 15213-3890
59 * any improvements or extensions that they make and grant Carnegie the
60 * rights to redistribute these changes.
62 * $FreeBSD: src/sys/vm/vm_map.c,v 1.187.2.19 2003/05/27 00:47:02 alc Exp $
66 * Virtual memory mapping module.
69 #include <sys/param.h>
70 #include <sys/systm.h>
71 #include <sys/kernel.h>
73 #include <sys/serialize.h>
75 #include <sys/vmmeter.h>
77 #include <sys/vnode.h>
78 #include <sys/resourcevar.h>
81 #include <sys/malloc.h>
82 #include <sys/objcache.h>
85 #include <vm/vm_param.h>
87 #include <vm/vm_map.h>
88 #include <vm/vm_page.h>
89 #include <vm/vm_object.h>
90 #include <vm/vm_pager.h>
91 #include <vm/vm_kern.h>
92 #include <vm/vm_extern.h>
93 #include <vm/swap_pager.h>
94 #include <vm/vm_zone.h>
96 #include <sys/random.h>
97 #include <sys/sysctl.h>
98 #include <sys/spinlock.h>
100 #include <sys/thread2.h>
101 #include <sys/spinlock2.h>
104 * Virtual memory maps provide for the mapping, protection, and sharing
105 * of virtual memory objects. In addition, this module provides for an
106 * efficient virtual copy of memory from one map to another.
108 * Synchronization is required prior to most operations.
110 * Maps consist of an ordered doubly-linked list of simple entries.
111 * A hint and a RB tree is used to speed-up lookups.
113 * Callers looking to modify maps specify start/end addresses which cause
114 * the related map entry to be clipped if necessary, and then later
115 * recombined if the pieces remained compatible.
117 * Virtual copy operations are performed by copying VM object references
118 * from one map to another, and then marking both regions as copy-on-write.
120 static boolean_t
vmspace_ctor(void *obj
, void *privdata
, int ocflags
);
121 static void vmspace_dtor(void *obj
, void *privdata
);
122 static void vmspace_terminate(struct vmspace
*vm
, int final
);
124 MALLOC_DEFINE(M_VMSPACE
, "vmspace", "vmspace objcache backingstore");
125 static struct objcache
*vmspace_cache
;
128 * per-cpu page table cross mappings are initialized in early boot
129 * and might require a considerable number of vm_map_entry structures.
131 #define MAPENTRYBSP_CACHE (MAXCPU+1)
132 #define MAPENTRYAP_CACHE 8
134 static struct vm_zone mapentzone_store
;
135 static vm_zone_t mapentzone
;
137 static struct vm_map_entry map_entry_init
[MAX_MAPENT
];
138 static struct vm_map_entry cpu_map_entry_init_bsp
[MAPENTRYBSP_CACHE
];
139 static struct vm_map_entry cpu_map_entry_init_ap
[MAXCPU
][MAPENTRYAP_CACHE
];
141 static int randomize_mmap
;
142 SYSCTL_INT(_vm
, OID_AUTO
, randomize_mmap
, CTLFLAG_RW
, &randomize_mmap
, 0,
143 "Randomize mmap offsets");
144 static int vm_map_relock_enable
= 1;
145 SYSCTL_INT(_vm
, OID_AUTO
, map_relock_enable
, CTLFLAG_RW
,
146 &vm_map_relock_enable
, 0, "Randomize mmap offsets");
148 static void vmspace_drop_notoken(struct vmspace
*vm
);
149 static void vm_map_entry_shadow(vm_map_entry_t entry
, int addref
);
150 static vm_map_entry_t
vm_map_entry_create(vm_map_t map
, int *);
151 static void vm_map_entry_dispose (vm_map_t map
, vm_map_entry_t entry
, int *);
152 static void _vm_map_clip_end (vm_map_t
, vm_map_entry_t
, vm_offset_t
, int *);
153 static void _vm_map_clip_start (vm_map_t
, vm_map_entry_t
, vm_offset_t
, int *);
154 static void vm_map_entry_delete (vm_map_t
, vm_map_entry_t
, int *);
155 static void vm_map_entry_unwire (vm_map_t
, vm_map_entry_t
);
156 static void vm_map_copy_entry (vm_map_t
, vm_map_t
, vm_map_entry_t
,
158 static void vm_map_unclip_range (vm_map_t map
, vm_map_entry_t start_entry
, vm_offset_t start
, vm_offset_t end
, int *count
, int flags
);
161 * Initialize the vm_map module. Must be called before any other vm_map
164 * Map and entry structures are allocated from the general purpose
165 * memory pool with some exceptions:
167 * - The kernel map is allocated statically.
168 * - Initial kernel map entries are allocated out of a static pool.
169 * - We must set ZONE_SPECIAL here or the early boot code can get
170 * stuck if there are >63 cores.
172 * These restrictions are necessary since malloc() uses the
173 * maps and requires map entries.
175 * Called from the low level boot code only.
180 mapentzone
= &mapentzone_store
;
181 zbootinit(mapentzone
, "MAP ENTRY", sizeof (struct vm_map_entry
),
182 map_entry_init
, MAX_MAPENT
);
183 mapentzone_store
.zflags
|= ZONE_SPECIAL
;
187 * Called prior to any vmspace allocations.
189 * Called from the low level boot code only.
194 vmspace_cache
= objcache_create_mbacked(M_VMSPACE
,
195 sizeof(struct vmspace
),
197 vmspace_ctor
, vmspace_dtor
,
199 zinitna(mapentzone
, NULL
, 0, 0, ZONE_USE_RESERVE
| ZONE_SPECIAL
);
205 * objcache support. We leave the pmap root cached as long as possible
206 * for performance reasons.
210 vmspace_ctor(void *obj
, void *privdata
, int ocflags
)
212 struct vmspace
*vm
= obj
;
214 bzero(vm
, sizeof(*vm
));
215 vm
->vm_refcnt
= VM_REF_DELETED
;
222 vmspace_dtor(void *obj
, void *privdata
)
224 struct vmspace
*vm
= obj
;
226 KKASSERT(vm
->vm_refcnt
== VM_REF_DELETED
);
227 pmap_puninit(vmspace_pmap(vm
));
231 * Red black tree functions
233 * The caller must hold the related map lock.
235 static int rb_vm_map_compare(vm_map_entry_t a
, vm_map_entry_t b
);
236 RB_GENERATE(vm_map_rb_tree
, vm_map_entry
, rb_entry
, rb_vm_map_compare
);
238 /* a->start is address, and the only field has to be initialized */
240 rb_vm_map_compare(vm_map_entry_t a
, vm_map_entry_t b
)
242 if (a
->start
< b
->start
)
244 else if (a
->start
> b
->start
)
250 * Initialize vmspace ref/hold counts vmspace0. There is a holdcnt for
254 vmspace_initrefs(struct vmspace
*vm
)
261 * Allocate a vmspace structure, including a vm_map and pmap.
262 * Initialize numerous fields. While the initial allocation is zerod,
263 * subsequence reuse from the objcache leaves elements of the structure
264 * intact (particularly the pmap), so portions must be zerod.
266 * Returns a referenced vmspace.
271 vmspace_alloc(vm_offset_t min
, vm_offset_t max
)
275 vm
= objcache_get(vmspace_cache
, M_WAITOK
);
277 bzero(&vm
->vm_startcopy
,
278 (char *)&vm
->vm_endcopy
- (char *)&vm
->vm_startcopy
);
279 vm_map_init(&vm
->vm_map
, min
, max
, NULL
); /* initializes token */
282 * NOTE: hold to acquires token for safety.
284 * On return vmspace is referenced (refs=1, hold=1). That is,
285 * each refcnt also has a holdcnt. There can be additional holds
286 * (holdcnt) above and beyond the refcnt. Finalization is handled in
287 * two stages, one on refs 1->0, and the the second on hold 1->0.
289 KKASSERT(vm
->vm_holdcnt
== 0);
290 KKASSERT(vm
->vm_refcnt
== VM_REF_DELETED
);
291 vmspace_initrefs(vm
);
293 pmap_pinit(vmspace_pmap(vm
)); /* (some fields reused) */
294 vm
->vm_map
.pmap
= vmspace_pmap(vm
); /* XXX */
297 cpu_vmspace_alloc(vm
);
304 * NOTE: Can return 0 if the vmspace is exiting.
307 vmspace_getrefs(struct vmspace
*vm
)
313 if (n
& VM_REF_DELETED
)
319 vmspace_hold(struct vmspace
*vm
)
321 atomic_add_int(&vm
->vm_holdcnt
, 1);
322 lwkt_gettoken(&vm
->vm_map
.token
);
326 * Drop with final termination interlock.
329 vmspace_drop(struct vmspace
*vm
)
331 lwkt_reltoken(&vm
->vm_map
.token
);
332 vmspace_drop_notoken(vm
);
336 vmspace_drop_notoken(struct vmspace
*vm
)
338 if (atomic_fetchadd_int(&vm
->vm_holdcnt
, -1) == 1) {
339 if (vm
->vm_refcnt
& VM_REF_DELETED
)
340 vmspace_terminate(vm
, 1);
345 * A vmspace object must not be in a terminated state to be able to obtain
346 * additional refs on it.
348 * These are official references to the vmspace, the count is used to check
349 * for vmspace sharing. Foreign accessors should use 'hold' and not 'ref'.
351 * XXX we need to combine hold & ref together into one 64-bit field to allow
352 * holds to prevent stage-1 termination.
355 vmspace_ref(struct vmspace
*vm
)
359 atomic_add_int(&vm
->vm_holdcnt
, 1);
360 n
= atomic_fetchadd_int(&vm
->vm_refcnt
, 1);
361 KKASSERT((n
& VM_REF_DELETED
) == 0);
365 * Release a ref on the vmspace. On the 1->0 transition we do stage-1
366 * termination of the vmspace. Then, on the final drop of the hold we
367 * will do stage-2 final termination.
370 vmspace_rel(struct vmspace
*vm
)
375 * Drop refs. Each ref also has a hold which is also dropped.
377 * When refs hits 0 compete to get the VM_REF_DELETED flag (hold
378 * prevent finalization) to start termination processing.
379 * Finalization occurs when the last hold count drops to 0.
381 n
= atomic_fetchadd_int(&vm
->vm_refcnt
, -1) - 1;
383 if (atomic_cmpset_int(&vm
->vm_refcnt
, 0, VM_REF_DELETED
)) {
384 vmspace_terminate(vm
, 0);
390 vmspace_drop_notoken(vm
);
394 * This is called during exit indicating that the vmspace is no
395 * longer in used by an exiting process, but the process has not yet
398 * We drop refs, allowing for stage-1 termination, but maintain a holdcnt
399 * to prevent stage-2 until the process is reaped. Note hte order of
400 * operation, we must hold first.
405 vmspace_relexit(struct vmspace
*vm
)
407 atomic_add_int(&vm
->vm_holdcnt
, 1);
412 * Called during reap to disconnect the remainder of the vmspace from
413 * the process. On the hold drop the vmspace termination is finalized.
418 vmspace_exitfree(struct proc
*p
)
424 vmspace_drop_notoken(vm
);
428 * Called in two cases:
430 * (1) When the last refcnt is dropped and the vmspace becomes inactive,
431 * called with final == 0. refcnt will be (u_int)-1 at this point,
432 * and holdcnt will still be non-zero.
434 * (2) When holdcnt becomes 0, called with final == 1. There should no
435 * longer be anyone with access to the vmspace.
437 * VMSPACE_EXIT1 flags the primary deactivation
438 * VMSPACE_EXIT2 flags the last reap
441 vmspace_terminate(struct vmspace
*vm
, int final
)
445 lwkt_gettoken(&vm
->vm_map
.token
);
447 KKASSERT((vm
->vm_flags
& VMSPACE_EXIT1
) == 0);
448 vm
->vm_flags
|= VMSPACE_EXIT1
;
451 * Get rid of most of the resources. Leave the kernel pmap
454 * If the pmap does not contain wired pages we can bulk-delete
455 * the pmap as a performance optimization before removing the
458 * If the pmap contains wired pages we cannot do this
459 * pre-optimization because currently vm_fault_unwire()
460 * expects the pmap pages to exist and will not decrement
461 * p->wire_count if they do not.
464 if (vmspace_pmap(vm
)->pm_stats
.wired_count
) {
465 vm_map_remove(&vm
->vm_map
, VM_MIN_USER_ADDRESS
,
466 VM_MAX_USER_ADDRESS
);
467 pmap_remove_pages(vmspace_pmap(vm
), VM_MIN_USER_ADDRESS
,
468 VM_MAX_USER_ADDRESS
);
470 pmap_remove_pages(vmspace_pmap(vm
), VM_MIN_USER_ADDRESS
,
471 VM_MAX_USER_ADDRESS
);
472 vm_map_remove(&vm
->vm_map
, VM_MIN_USER_ADDRESS
,
473 VM_MAX_USER_ADDRESS
);
475 lwkt_reltoken(&vm
->vm_map
.token
);
477 KKASSERT((vm
->vm_flags
& VMSPACE_EXIT1
) != 0);
478 KKASSERT((vm
->vm_flags
& VMSPACE_EXIT2
) == 0);
481 * Get rid of remaining basic resources.
483 vm
->vm_flags
|= VMSPACE_EXIT2
;
486 count
= vm_map_entry_reserve(MAP_RESERVE_COUNT
);
487 vm_map_lock(&vm
->vm_map
);
488 cpu_vmspace_free(vm
);
491 * Lock the map, to wait out all other references to it.
492 * Delete all of the mappings and pages they hold, then call
493 * the pmap module to reclaim anything left.
495 vm_map_delete(&vm
->vm_map
, vm
->vm_map
.min_offset
,
496 vm
->vm_map
.max_offset
, &count
);
497 vm_map_unlock(&vm
->vm_map
);
498 vm_map_entry_release(count
);
500 pmap_release(vmspace_pmap(vm
));
501 lwkt_reltoken(&vm
->vm_map
.token
);
502 objcache_put(vmspace_cache
, vm
);
507 * Swap useage is determined by taking the proportional swap used by
508 * VM objects backing the VM map. To make up for fractional losses,
509 * if the VM object has any swap use at all the associated map entries
510 * count for at least 1 swap page.
515 vmspace_swap_count(struct vmspace
*vm
)
517 vm_map_t map
= &vm
->vm_map
;
520 vm_offset_t count
= 0;
524 for (cur
= map
->header
.next
; cur
!= &map
->header
; cur
= cur
->next
) {
525 switch(cur
->maptype
) {
526 case VM_MAPTYPE_NORMAL
:
527 case VM_MAPTYPE_VPAGETABLE
:
528 if ((object
= cur
->object
.vm_object
) == NULL
)
530 if (object
->swblock_count
) {
531 n
= (cur
->end
- cur
->start
) / PAGE_SIZE
;
532 count
+= object
->swblock_count
*
533 SWAP_META_PAGES
* n
/ object
->size
+ 1;
546 * Calculate the approximate number of anonymous pages in use by
547 * this vmspace. To make up for fractional losses, we count each
548 * VM object as having at least 1 anonymous page.
553 vmspace_anonymous_count(struct vmspace
*vm
)
555 vm_map_t map
= &vm
->vm_map
;
558 vm_offset_t count
= 0;
561 for (cur
= map
->header
.next
; cur
!= &map
->header
; cur
= cur
->next
) {
562 switch(cur
->maptype
) {
563 case VM_MAPTYPE_NORMAL
:
564 case VM_MAPTYPE_VPAGETABLE
:
565 if ((object
= cur
->object
.vm_object
) == NULL
)
567 if (object
->type
!= OBJT_DEFAULT
&&
568 object
->type
!= OBJT_SWAP
) {
571 count
+= object
->resident_page_count
;
583 * Initialize an existing vm_map structure such as that in the vmspace
584 * structure. The pmap is initialized elsewhere.
589 vm_map_init(struct vm_map
*map
, vm_offset_t min
, vm_offset_t max
, pmap_t pmap
)
591 map
->header
.next
= map
->header
.prev
= &map
->header
;
592 RB_INIT(&map
->rb_root
);
593 spin_init(&map
->ilock_spin
, "ilock");
594 map
->ilock_base
= NULL
;
598 map
->min_offset
= min
;
599 map
->max_offset
= max
;
603 bzero(&map
->freehint
, sizeof(map
->freehint
));
604 lwkt_token_init(&map
->token
, "vm_map");
605 lockinit(&map
->lock
, "vm_maplk", (hz
+ 9) / 10, 0);
609 * Find the first possible free address for the specified request length.
610 * Returns 0 if we don't have one cached.
614 vm_map_freehint_find(vm_map_t map
, vm_size_t length
, vm_size_t align
)
616 vm_map_freehint_t
*scan
;
618 scan
= &map
->freehint
[0];
619 while (scan
< &map
->freehint
[VM_MAP_FFCOUNT
]) {
620 if (scan
->length
== length
&& scan
->align
== align
)
628 * Unconditionally set the freehint. Called by vm_map_findspace() after
629 * it finds an address. This will help us iterate optimally on the next
634 vm_map_freehint_update(vm_map_t map
, vm_offset_t start
,
635 vm_size_t length
, vm_size_t align
)
637 vm_map_freehint_t
*scan
;
639 scan
= &map
->freehint
[0];
640 while (scan
< &map
->freehint
[VM_MAP_FFCOUNT
]) {
641 if (scan
->length
== length
&& scan
->align
== align
) {
647 scan
= &map
->freehint
[map
->freehint_newindex
& VM_MAP_FFMASK
];
650 scan
->length
= length
;
651 ++map
->freehint_newindex
;
655 * Update any existing freehints (for any alignment), for the hole we just
660 vm_map_freehint_hole(vm_map_t map
, vm_offset_t start
, vm_size_t length
)
662 vm_map_freehint_t
*scan
;
664 scan
= &map
->freehint
[0];
665 while (scan
< &map
->freehint
[VM_MAP_FFCOUNT
]) {
666 if (scan
->length
<= length
&& scan
->start
> start
)
673 * Shadow the vm_map_entry's object. This typically needs to be done when
674 * a write fault is taken on an entry which had previously been cloned by
675 * fork(). The shared object (which might be NULL) must become private so
676 * we add a shadow layer above it.
678 * Object allocation for anonymous mappings is defered as long as possible.
679 * When creating a shadow, however, the underlying object must be instantiated
680 * so it can be shared.
682 * If the map segment is governed by a virtual page table then it is
683 * possible to address offsets beyond the mapped area. Just allocate
684 * a maximally sized object for this case.
686 * If addref is non-zero an additional reference is added to the returned
687 * entry. This mechanic exists because the additional reference might have
688 * to be added atomically and not after return to prevent a premature
691 * The vm_map must be exclusively locked.
692 * No other requirements.
696 vm_map_entry_shadow(vm_map_entry_t entry
, int addref
)
698 if (entry
->maptype
== VM_MAPTYPE_VPAGETABLE
) {
699 vm_object_shadow(&entry
->object
.vm_object
, &entry
->offset
,
700 0x7FFFFFFF, addref
); /* XXX */
702 vm_object_shadow(&entry
->object
.vm_object
, &entry
->offset
,
703 atop(entry
->end
- entry
->start
), addref
);
705 entry
->eflags
&= ~MAP_ENTRY_NEEDS_COPY
;
709 * Allocate an object for a vm_map_entry.
711 * Object allocation for anonymous mappings is defered as long as possible.
712 * This function is called when we can defer no longer, generally when a map
713 * entry might be split or forked or takes a page fault.
715 * If the map segment is governed by a virtual page table then it is
716 * possible to address offsets beyond the mapped area. Just allocate
717 * a maximally sized object for this case.
719 * The vm_map must be exclusively locked.
720 * No other requirements.
723 vm_map_entry_allocate_object(vm_map_entry_t entry
)
727 if (entry
->maptype
== VM_MAPTYPE_VPAGETABLE
) {
728 obj
= vm_object_allocate(OBJT_DEFAULT
, 0x7FFFFFFF); /* XXX */
730 obj
= vm_object_allocate(OBJT_DEFAULT
,
731 atop(entry
->end
- entry
->start
));
733 entry
->object
.vm_object
= obj
;
738 * Set an initial negative count so the first attempt to reserve
739 * space preloads a bunch of vm_map_entry's for this cpu. Also
740 * pre-allocate 2 vm_map_entries which will be needed by zalloc() to
741 * map a new page for vm_map_entry structures. SMP systems are
742 * particularly sensitive.
744 * This routine is called in early boot so we cannot just call
745 * vm_map_entry_reserve().
747 * Called from the low level boot code only (for each cpu)
749 * WARNING! Take care not to have too-big a static/BSS structure here
750 * as MAXCPU can be 256+, otherwise the loader's 64MB heap
751 * can get blown out by the kernel plus the initrd image.
754 vm_map_entry_reserve_cpu_init(globaldata_t gd
)
756 vm_map_entry_t entry
;
760 gd
->gd_vme_avail
-= MAP_RESERVE_COUNT
* 2;
761 if (gd
->gd_cpuid
== 0) {
762 entry
= &cpu_map_entry_init_bsp
[0];
763 count
= MAPENTRYBSP_CACHE
;
765 entry
= &cpu_map_entry_init_ap
[gd
->gd_cpuid
][0];
766 count
= MAPENTRYAP_CACHE
;
768 for (i
= 0; i
< count
; ++i
, ++entry
) {
769 entry
->next
= gd
->gd_vme_base
;
770 gd
->gd_vme_base
= entry
;
775 * Reserves vm_map_entry structures so code later on can manipulate
776 * map_entry structures within a locked map without blocking trying
777 * to allocate a new vm_map_entry.
782 vm_map_entry_reserve(int count
)
784 struct globaldata
*gd
= mycpu
;
785 vm_map_entry_t entry
;
788 * Make sure we have enough structures in gd_vme_base to handle
789 * the reservation request.
791 * The critical section protects access to the per-cpu gd.
794 while (gd
->gd_vme_avail
< count
) {
795 entry
= zalloc(mapentzone
);
796 entry
->next
= gd
->gd_vme_base
;
797 gd
->gd_vme_base
= entry
;
800 gd
->gd_vme_avail
-= count
;
807 * Releases previously reserved vm_map_entry structures that were not
808 * used. If we have too much junk in our per-cpu cache clean some of
814 vm_map_entry_release(int count
)
816 struct globaldata
*gd
= mycpu
;
817 vm_map_entry_t entry
;
820 gd
->gd_vme_avail
+= count
;
821 while (gd
->gd_vme_avail
> MAP_RESERVE_SLOP
) {
822 entry
= gd
->gd_vme_base
;
823 KKASSERT(entry
!= NULL
);
824 gd
->gd_vme_base
= entry
->next
;
827 zfree(mapentzone
, entry
);
834 * Reserve map entry structures for use in kernel_map itself. These
835 * entries have *ALREADY* been reserved on a per-cpu basis when the map
836 * was inited. This function is used by zalloc() to avoid a recursion
837 * when zalloc() itself needs to allocate additional kernel memory.
839 * This function works like the normal reserve but does not load the
840 * vm_map_entry cache (because that would result in an infinite
841 * recursion). Note that gd_vme_avail may go negative. This is expected.
843 * Any caller of this function must be sure to renormalize after
844 * potentially eating entries to ensure that the reserve supply
850 vm_map_entry_kreserve(int count
)
852 struct globaldata
*gd
= mycpu
;
855 gd
->gd_vme_avail
-= count
;
857 KASSERT(gd
->gd_vme_base
!= NULL
,
858 ("no reserved entries left, gd_vme_avail = %d",
864 * Release previously reserved map entries for kernel_map. We do not
865 * attempt to clean up like the normal release function as this would
866 * cause an unnecessary (but probably not fatal) deep procedure call.
871 vm_map_entry_krelease(int count
)
873 struct globaldata
*gd
= mycpu
;
876 gd
->gd_vme_avail
+= count
;
881 * Allocates a VM map entry for insertion. No entry fields are filled in.
883 * The entries should have previously been reserved. The reservation count
884 * is tracked in (*countp).
888 static vm_map_entry_t
889 vm_map_entry_create(vm_map_t map
, int *countp
)
891 struct globaldata
*gd
= mycpu
;
892 vm_map_entry_t entry
;
894 KKASSERT(*countp
> 0);
897 entry
= gd
->gd_vme_base
;
898 KASSERT(entry
!= NULL
, ("gd_vme_base NULL! count %d", *countp
));
899 gd
->gd_vme_base
= entry
->next
;
906 * Dispose of a vm_map_entry that is no longer being referenced.
911 vm_map_entry_dispose(vm_map_t map
, vm_map_entry_t entry
, int *countp
)
913 struct globaldata
*gd
= mycpu
;
917 entry
->next
= gd
->gd_vme_base
;
918 gd
->gd_vme_base
= entry
;
924 * Insert/remove entries from maps.
926 * The related map must be exclusively locked.
927 * The caller must hold map->token
928 * No other requirements.
931 vm_map_entry_link(vm_map_t map
,
932 vm_map_entry_t after_where
,
933 vm_map_entry_t entry
)
935 ASSERT_VM_MAP_LOCKED(map
);
938 entry
->prev
= after_where
;
939 entry
->next
= after_where
->next
;
940 entry
->next
->prev
= entry
;
941 after_where
->next
= entry
;
942 if (vm_map_rb_tree_RB_INSERT(&map
->rb_root
, entry
))
943 panic("vm_map_entry_link: dup addr map %p ent %p", map
, entry
);
947 vm_map_entry_unlink(vm_map_t map
,
948 vm_map_entry_t entry
)
953 ASSERT_VM_MAP_LOCKED(map
);
955 if (entry
->eflags
& MAP_ENTRY_IN_TRANSITION
) {
956 panic("vm_map_entry_unlink: attempt to mess with "
957 "locked entry! %p", entry
);
963 vm_map_rb_tree_RB_REMOVE(&map
->rb_root
, entry
);
968 * Finds the map entry containing (or immediately preceding) the specified
969 * address in the given map. The entry is returned in (*entry).
971 * The boolean result indicates whether the address is actually contained
974 * The related map must be locked.
975 * No other requirements.
978 vm_map_lookup_entry(vm_map_t map
, vm_offset_t address
, vm_map_entry_t
*entry
)
983 ASSERT_VM_MAP_LOCKED(map
);
986 * Locate the record from the top of the tree. 'last' tracks the
987 * closest prior record and is returned if no match is found, which
988 * in binary tree terms means tracking the most recent right-branch
989 * taken. If there is no prior record, &map->header is returned.
992 tmp
= RB_ROOT(&map
->rb_root
);
995 if (address
>= tmp
->start
) {
996 if (address
< tmp
->end
) {
1001 tmp
= RB_RIGHT(tmp
, rb_entry
);
1003 tmp
= RB_LEFT(tmp
, rb_entry
);
1011 * Inserts the given whole VM object into the target map at the specified
1012 * address range. The object's size should match that of the address range.
1014 * The map must be exclusively locked.
1015 * The object must be held.
1016 * The caller must have reserved sufficient vm_map_entry structures.
1018 * If object is non-NULL, ref count must be bumped by caller prior to
1019 * making call to account for the new entry.
1022 vm_map_insert(vm_map_t map
, int *countp
, void *map_object
, void *map_aux
,
1023 vm_ooffset_t offset
, vm_offset_t start
, vm_offset_t end
,
1024 vm_maptype_t maptype
, vm_subsys_t id
,
1025 vm_prot_t prot
, vm_prot_t max
, int cow
)
1027 vm_map_entry_t new_entry
;
1028 vm_map_entry_t prev_entry
;
1029 vm_map_entry_t temp_entry
;
1030 vm_eflags_t protoeflags
;
1034 if (maptype
== VM_MAPTYPE_UKSMAP
)
1037 object
= map_object
;
1039 ASSERT_VM_MAP_LOCKED(map
);
1041 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object
));
1044 * Check that the start and end points are not bogus.
1046 if ((start
< map
->min_offset
) || (end
> map
->max_offset
) ||
1048 return (KERN_INVALID_ADDRESS
);
1051 * Find the entry prior to the proposed starting address; if it's part
1052 * of an existing entry, this range is bogus.
1054 if (vm_map_lookup_entry(map
, start
, &temp_entry
))
1055 return (KERN_NO_SPACE
);
1057 prev_entry
= temp_entry
;
1060 * Assert that the next entry doesn't overlap the end point.
1063 if ((prev_entry
->next
!= &map
->header
) &&
1064 (prev_entry
->next
->start
< end
))
1065 return (KERN_NO_SPACE
);
1069 if (cow
& MAP_COPY_ON_WRITE
)
1070 protoeflags
|= MAP_ENTRY_COW
|MAP_ENTRY_NEEDS_COPY
;
1072 if (cow
& MAP_NOFAULT
) {
1073 protoeflags
|= MAP_ENTRY_NOFAULT
;
1075 KASSERT(object
== NULL
,
1076 ("vm_map_insert: paradoxical MAP_NOFAULT request"));
1078 if (cow
& MAP_DISABLE_SYNCER
)
1079 protoeflags
|= MAP_ENTRY_NOSYNC
;
1080 if (cow
& MAP_DISABLE_COREDUMP
)
1081 protoeflags
|= MAP_ENTRY_NOCOREDUMP
;
1082 if (cow
& MAP_IS_STACK
)
1083 protoeflags
|= MAP_ENTRY_STACK
;
1084 if (cow
& MAP_IS_KSTACK
)
1085 protoeflags
|= MAP_ENTRY_KSTACK
;
1087 lwkt_gettoken(&map
->token
);
1091 * When object is non-NULL, it could be shared with another
1092 * process. We have to set or clear OBJ_ONEMAPPING
1095 * NOTE: This flag is only applicable to DEFAULT and SWAP
1096 * objects and will already be clear in other types
1097 * of objects, so a shared object lock is ok for
1100 if ((object
->ref_count
> 1) || (object
->shadow_count
!= 0)) {
1101 vm_object_clear_flag(object
, OBJ_ONEMAPPING
);
1104 else if ((prev_entry
!= &map
->header
) &&
1105 (prev_entry
->eflags
== protoeflags
) &&
1106 (prev_entry
->end
== start
) &&
1107 (prev_entry
->wired_count
== 0) &&
1108 (prev_entry
->id
== id
) &&
1109 prev_entry
->maptype
== maptype
&&
1110 maptype
== VM_MAPTYPE_NORMAL
&&
1111 ((prev_entry
->object
.vm_object
== NULL
) ||
1112 vm_object_coalesce(prev_entry
->object
.vm_object
,
1113 OFF_TO_IDX(prev_entry
->offset
),
1114 (vm_size_t
)(prev_entry
->end
- prev_entry
->start
),
1115 (vm_size_t
)(end
- prev_entry
->end
)))) {
1117 * We were able to extend the object. Determine if we
1118 * can extend the previous map entry to include the
1119 * new range as well.
1121 if ((prev_entry
->inheritance
== VM_INHERIT_DEFAULT
) &&
1122 (prev_entry
->protection
== prot
) &&
1123 (prev_entry
->max_protection
== max
)) {
1124 map
->size
+= (end
- prev_entry
->end
);
1125 prev_entry
->end
= end
;
1126 vm_map_simplify_entry(map
, prev_entry
, countp
);
1127 lwkt_reltoken(&map
->token
);
1128 return (KERN_SUCCESS
);
1132 * If we can extend the object but cannot extend the
1133 * map entry, we have to create a new map entry. We
1134 * must bump the ref count on the extended object to
1135 * account for it. object may be NULL.
1137 * XXX if object is NULL should we set offset to 0 here ?
1139 object
= prev_entry
->object
.vm_object
;
1140 offset
= prev_entry
->offset
+
1141 (prev_entry
->end
- prev_entry
->start
);
1143 vm_object_hold(object
);
1144 vm_object_chain_wait(object
, 0);
1145 vm_object_reference_locked(object
);
1147 map_object
= object
;
1152 * NOTE: if conditionals fail, object can be NULL here. This occurs
1153 * in things like the buffer map where we manage kva but do not manage
1158 * Create a new entry
1161 new_entry
= vm_map_entry_create(map
, countp
);
1162 new_entry
->start
= start
;
1163 new_entry
->end
= end
;
1166 new_entry
->maptype
= maptype
;
1167 new_entry
->eflags
= protoeflags
;
1168 new_entry
->object
.map_object
= map_object
;
1169 new_entry
->aux
.master_pde
= 0; /* in case size is different */
1170 new_entry
->aux
.map_aux
= map_aux
;
1171 new_entry
->offset
= offset
;
1173 new_entry
->inheritance
= VM_INHERIT_DEFAULT
;
1174 new_entry
->protection
= prot
;
1175 new_entry
->max_protection
= max
;
1176 new_entry
->wired_count
= 0;
1179 * Insert the new entry into the list
1182 vm_map_entry_link(map
, prev_entry
, new_entry
);
1183 map
->size
+= new_entry
->end
- new_entry
->start
;
1186 * Don't worry about updating freehint[] when inserting, allow
1187 * addresses to be lower than the actual first free spot.
1191 * Temporarily removed to avoid MAP_STACK panic, due to
1192 * MAP_STACK being a huge hack. Will be added back in
1193 * when MAP_STACK (and the user stack mapping) is fixed.
1196 * It may be possible to simplify the entry
1198 vm_map_simplify_entry(map
, new_entry
, countp
);
1202 * Try to pre-populate the page table. Mappings governed by virtual
1203 * page tables cannot be prepopulated without a lot of work, so
1206 if ((cow
& (MAP_PREFAULT
|MAP_PREFAULT_PARTIAL
)) &&
1207 maptype
!= VM_MAPTYPE_VPAGETABLE
&&
1208 maptype
!= VM_MAPTYPE_UKSMAP
) {
1210 if (vm_map_relock_enable
&& (cow
& MAP_PREFAULT_RELOCK
)) {
1212 vm_object_lock_swap();
1213 vm_object_drop(object
);
1215 pmap_object_init_pt(map
->pmap
, start
, prot
,
1216 object
, OFF_TO_IDX(offset
), end
- start
,
1217 cow
& MAP_PREFAULT_PARTIAL
);
1219 vm_object_hold(object
);
1220 vm_object_lock_swap();
1224 vm_object_drop(object
);
1226 lwkt_reltoken(&map
->token
);
1227 return (KERN_SUCCESS
);
1231 * Find sufficient space for `length' bytes in the given map, starting at
1232 * `start'. Returns 0 on success, 1 on no space.
1234 * This function will returned an arbitrarily aligned pointer. If no
1235 * particular alignment is required you should pass align as 1. Note that
1236 * the map may return PAGE_SIZE aligned pointers if all the lengths used in
1237 * the map are a multiple of PAGE_SIZE, even if you pass a smaller align
1240 * 'align' should be a power of 2 but is not required to be.
1242 * The map must be exclusively locked.
1243 * No other requirements.
1246 vm_map_findspace(vm_map_t map
, vm_offset_t start
, vm_size_t length
,
1247 vm_size_t align
, int flags
, vm_offset_t
*addr
)
1249 vm_map_entry_t entry
, next
;
1251 vm_offset_t hole_start
;
1253 vm_offset_t align_mask
;
1255 if (start
< map
->min_offset
)
1256 start
= map
->min_offset
;
1257 if (start
> map
->max_offset
)
1261 * If the alignment is not a power of 2 we will have to use
1262 * a mod/division, set align_mask to a special value.
1264 if ((align
| (align
- 1)) + 1 != (align
<< 1))
1265 align_mask
= (vm_offset_t
)-1;
1267 align_mask
= align
- 1;
1270 * Use freehint to adjust the start point, hopefully reducing
1271 * the iteration to O(1).
1273 hole_start
= vm_map_freehint_find(map
, length
, align
);
1274 if (start
< hole_start
)
1276 if (vm_map_lookup_entry(map
, start
, &tmp
))
1281 * Look through the rest of the map, trying to fit a new region in the
1282 * gap between existing regions, or after the very last region.
1284 for (;; start
= (entry
= next
)->end
) {
1286 * Adjust the proposed start by the requested alignment,
1287 * be sure that we didn't wrap the address.
1289 if (align_mask
== (vm_offset_t
)-1)
1290 end
= roundup(start
, align
);
1292 end
= (start
+ align_mask
) & ~align_mask
;
1298 * Find the end of the proposed new region. Be sure we didn't
1299 * go beyond the end of the map, or wrap around the address.
1300 * Then check to see if this is the last entry or if the
1301 * proposed end fits in the gap between this and the next
1304 end
= start
+ length
;
1305 if (end
> map
->max_offset
|| end
< start
)
1310 * If the next entry's start address is beyond the desired
1311 * end address we may have found a good entry.
1313 * If the next entry is a stack mapping we do not map into
1314 * the stack's reserved space.
1316 * XXX continue to allow mapping into the stack's reserved
1317 * space if doing a MAP_STACK mapping inside a MAP_STACK
1318 * mapping, for backwards compatibility. But the caller
1319 * really should use MAP_STACK | MAP_TRYFIXED if they
1322 if (next
== &map
->header
)
1324 if (next
->start
>= end
) {
1325 if ((next
->eflags
& MAP_ENTRY_STACK
) == 0)
1327 if (flags
& MAP_STACK
)
1329 if (next
->start
- next
->aux
.avail_ssize
>= end
)
1335 * Update the freehint
1337 vm_map_freehint_update(map
, start
, length
, align
);
1340 * Grow the kernel_map if necessary. pmap_growkernel() will panic
1341 * if it fails. The kernel_map is locked and nothing can steal
1342 * our address space if pmap_growkernel() blocks.
1344 * NOTE: This may be unconditionally called for kldload areas on
1345 * x86_64 because these do not bump kernel_vm_end (which would
1346 * fill 128G worth of page tables!). Therefore we must not
1349 if (map
== &kernel_map
) {
1352 kstop
= round_page(start
+ length
);
1353 if (kstop
> kernel_vm_end
)
1354 pmap_growkernel(start
, kstop
);
1361 * vm_map_find finds an unallocated region in the target address map with
1362 * the given length and allocates it. The search is defined to be first-fit
1363 * from the specified address; the region found is returned in the same
1366 * If object is non-NULL, ref count must be bumped by caller
1367 * prior to making call to account for the new entry.
1369 * No requirements. This function will lock the map temporarily.
1372 vm_map_find(vm_map_t map
, void *map_object
, void *map_aux
,
1373 vm_ooffset_t offset
, vm_offset_t
*addr
,
1374 vm_size_t length
, vm_size_t align
, boolean_t fitit
,
1375 vm_maptype_t maptype
, vm_subsys_t id
,
1376 vm_prot_t prot
, vm_prot_t max
, int cow
)
1383 if (maptype
== VM_MAPTYPE_UKSMAP
)
1386 object
= map_object
;
1390 count
= vm_map_entry_reserve(MAP_RESERVE_COUNT
);
1393 vm_object_hold_shared(object
);
1395 if (vm_map_findspace(map
, start
, length
, align
, 0, addr
)) {
1397 vm_object_drop(object
);
1399 vm_map_entry_release(count
);
1400 return (KERN_NO_SPACE
);
1404 result
= vm_map_insert(map
, &count
, map_object
, map_aux
,
1405 offset
, start
, start
+ length
,
1406 maptype
, id
, prot
, max
, cow
);
1408 vm_object_drop(object
);
1410 vm_map_entry_release(count
);
1416 * Simplify the given map entry by merging with either neighbor. This
1417 * routine also has the ability to merge with both neighbors.
1419 * This routine guarentees that the passed entry remains valid (though
1420 * possibly extended). When merging, this routine may delete one or
1421 * both neighbors. No action is taken on entries which have their
1422 * in-transition flag set.
1424 * The map must be exclusively locked.
1427 vm_map_simplify_entry(vm_map_t map
, vm_map_entry_t entry
, int *countp
)
1429 vm_map_entry_t next
, prev
;
1430 vm_size_t prevsize
, esize
;
1432 if (entry
->eflags
& MAP_ENTRY_IN_TRANSITION
) {
1433 ++mycpu
->gd_cnt
.v_intrans_coll
;
1437 if (entry
->maptype
== VM_MAPTYPE_SUBMAP
)
1439 if (entry
->maptype
== VM_MAPTYPE_UKSMAP
)
1443 if (prev
!= &map
->header
) {
1444 prevsize
= prev
->end
- prev
->start
;
1445 if ( (prev
->end
== entry
->start
) &&
1446 (prev
->maptype
== entry
->maptype
) &&
1447 (prev
->object
.vm_object
== entry
->object
.vm_object
) &&
1448 (!prev
->object
.vm_object
||
1449 (prev
->offset
+ prevsize
== entry
->offset
)) &&
1450 (prev
->eflags
== entry
->eflags
) &&
1451 (prev
->protection
== entry
->protection
) &&
1452 (prev
->max_protection
== entry
->max_protection
) &&
1453 (prev
->inheritance
== entry
->inheritance
) &&
1454 (prev
->id
== entry
->id
) &&
1455 (prev
->wired_count
== entry
->wired_count
)) {
1456 vm_map_entry_unlink(map
, prev
);
1457 entry
->start
= prev
->start
;
1458 entry
->offset
= prev
->offset
;
1459 if (prev
->object
.vm_object
)
1460 vm_object_deallocate(prev
->object
.vm_object
);
1461 vm_map_entry_dispose(map
, prev
, countp
);
1466 if (next
!= &map
->header
) {
1467 esize
= entry
->end
- entry
->start
;
1468 if ((entry
->end
== next
->start
) &&
1469 (next
->maptype
== entry
->maptype
) &&
1470 (next
->object
.vm_object
== entry
->object
.vm_object
) &&
1471 (!entry
->object
.vm_object
||
1472 (entry
->offset
+ esize
== next
->offset
)) &&
1473 (next
->eflags
== entry
->eflags
) &&
1474 (next
->protection
== entry
->protection
) &&
1475 (next
->max_protection
== entry
->max_protection
) &&
1476 (next
->inheritance
== entry
->inheritance
) &&
1477 (next
->id
== entry
->id
) &&
1478 (next
->wired_count
== entry
->wired_count
)) {
1479 vm_map_entry_unlink(map
, next
);
1480 entry
->end
= next
->end
;
1481 if (next
->object
.vm_object
)
1482 vm_object_deallocate(next
->object
.vm_object
);
1483 vm_map_entry_dispose(map
, next
, countp
);
1489 * Asserts that the given entry begins at or after the specified address.
1490 * If necessary, it splits the entry into two.
1492 #define vm_map_clip_start(map, entry, startaddr, countp) \
1494 if (startaddr > entry->start) \
1495 _vm_map_clip_start(map, entry, startaddr, countp); \
1499 * This routine is called only when it is known that the entry must be split.
1501 * The map must be exclusively locked.
1504 _vm_map_clip_start(vm_map_t map
, vm_map_entry_t entry
, vm_offset_t start
,
1507 vm_map_entry_t new_entry
;
1510 * Split off the front portion -- note that we must insert the new
1511 * entry BEFORE this one, so that this entry has the specified
1515 vm_map_simplify_entry(map
, entry
, countp
);
1518 * If there is no object backing this entry, we might as well create
1519 * one now. If we defer it, an object can get created after the map
1520 * is clipped, and individual objects will be created for the split-up
1521 * map. This is a bit of a hack, but is also about the best place to
1522 * put this improvement.
1524 if (entry
->object
.vm_object
== NULL
&& !map
->system_map
) {
1525 vm_map_entry_allocate_object(entry
);
1528 new_entry
= vm_map_entry_create(map
, countp
);
1529 *new_entry
= *entry
;
1531 new_entry
->end
= start
;
1532 entry
->offset
+= (start
- entry
->start
);
1533 entry
->start
= start
;
1535 vm_map_entry_link(map
, entry
->prev
, new_entry
);
1537 switch(entry
->maptype
) {
1538 case VM_MAPTYPE_NORMAL
:
1539 case VM_MAPTYPE_VPAGETABLE
:
1540 if (new_entry
->object
.vm_object
) {
1541 vm_object_hold(new_entry
->object
.vm_object
);
1542 vm_object_chain_wait(new_entry
->object
.vm_object
, 0);
1543 vm_object_reference_locked(new_entry
->object
.vm_object
);
1544 vm_object_drop(new_entry
->object
.vm_object
);
1553 * Asserts that the given entry ends at or before the specified address.
1554 * If necessary, it splits the entry into two.
1556 * The map must be exclusively locked.
1558 #define vm_map_clip_end(map, entry, endaddr, countp) \
1560 if (endaddr < entry->end) \
1561 _vm_map_clip_end(map, entry, endaddr, countp); \
1565 * This routine is called only when it is known that the entry must be split.
1567 * The map must be exclusively locked.
1570 _vm_map_clip_end(vm_map_t map
, vm_map_entry_t entry
, vm_offset_t end
,
1573 vm_map_entry_t new_entry
;
1576 * If there is no object backing this entry, we might as well create
1577 * one now. If we defer it, an object can get created after the map
1578 * is clipped, and individual objects will be created for the split-up
1579 * map. This is a bit of a hack, but is also about the best place to
1580 * put this improvement.
1583 if (entry
->object
.vm_object
== NULL
&& !map
->system_map
) {
1584 vm_map_entry_allocate_object(entry
);
1588 * Create a new entry and insert it AFTER the specified entry
1591 new_entry
= vm_map_entry_create(map
, countp
);
1592 *new_entry
= *entry
;
1594 new_entry
->start
= entry
->end
= end
;
1595 new_entry
->offset
+= (end
- entry
->start
);
1597 vm_map_entry_link(map
, entry
, new_entry
);
1599 switch(entry
->maptype
) {
1600 case VM_MAPTYPE_NORMAL
:
1601 case VM_MAPTYPE_VPAGETABLE
:
1602 if (new_entry
->object
.vm_object
) {
1603 vm_object_hold(new_entry
->object
.vm_object
);
1604 vm_object_chain_wait(new_entry
->object
.vm_object
, 0);
1605 vm_object_reference_locked(new_entry
->object
.vm_object
);
1606 vm_object_drop(new_entry
->object
.vm_object
);
1615 * Asserts that the starting and ending region addresses fall within the
1616 * valid range for the map.
1618 #define VM_MAP_RANGE_CHECK(map, start, end) \
1620 if (start < vm_map_min(map)) \
1621 start = vm_map_min(map); \
1622 if (end > vm_map_max(map)) \
1623 end = vm_map_max(map); \
1629 * Used to block when an in-transition collison occurs. The map
1630 * is unlocked for the sleep and relocked before the return.
1633 vm_map_transition_wait(vm_map_t map
)
1635 tsleep_interlock(map
, 0);
1637 tsleep(map
, PINTERLOCKED
, "vment", 0);
1642 * When we do blocking operations with the map lock held it is
1643 * possible that a clip might have occured on our in-transit entry,
1644 * requiring an adjustment to the entry in our loop. These macros
1645 * help the pageable and clip_range code deal with the case. The
1646 * conditional costs virtually nothing if no clipping has occured.
1649 #define CLIP_CHECK_BACK(entry, save_start) \
1651 while (entry->start != save_start) { \
1652 entry = entry->prev; \
1653 KASSERT(entry != &map->header, ("bad entry clip")); \
1657 #define CLIP_CHECK_FWD(entry, save_end) \
1659 while (entry->end != save_end) { \
1660 entry = entry->next; \
1661 KASSERT(entry != &map->header, ("bad entry clip")); \
1667 * Clip the specified range and return the base entry. The
1668 * range may cover several entries starting at the returned base
1669 * and the first and last entry in the covering sequence will be
1670 * properly clipped to the requested start and end address.
1672 * If no holes are allowed you should pass the MAP_CLIP_NO_HOLES
1675 * The MAP_ENTRY_IN_TRANSITION flag will be set for the entries
1676 * covered by the requested range.
1678 * The map must be exclusively locked on entry and will remain locked
1679 * on return. If no range exists or the range contains holes and you
1680 * specified that no holes were allowed, NULL will be returned. This
1681 * routine may temporarily unlock the map in order avoid a deadlock when
1686 vm_map_clip_range(vm_map_t map
, vm_offset_t start
, vm_offset_t end
,
1687 int *countp
, int flags
)
1689 vm_map_entry_t start_entry
;
1690 vm_map_entry_t entry
;
1693 * Locate the entry and effect initial clipping. The in-transition
1694 * case does not occur very often so do not try to optimize it.
1697 if (vm_map_lookup_entry(map
, start
, &start_entry
) == FALSE
)
1699 entry
= start_entry
;
1700 if (entry
->eflags
& MAP_ENTRY_IN_TRANSITION
) {
1701 entry
->eflags
|= MAP_ENTRY_NEEDS_WAKEUP
;
1702 ++mycpu
->gd_cnt
.v_intrans_coll
;
1703 ++mycpu
->gd_cnt
.v_intrans_wait
;
1704 vm_map_transition_wait(map
);
1706 * entry and/or start_entry may have been clipped while
1707 * we slept, or may have gone away entirely. We have
1708 * to restart from the lookup.
1714 * Since we hold an exclusive map lock we do not have to restart
1715 * after clipping, even though clipping may block in zalloc.
1717 vm_map_clip_start(map
, entry
, start
, countp
);
1718 vm_map_clip_end(map
, entry
, end
, countp
);
1719 entry
->eflags
|= MAP_ENTRY_IN_TRANSITION
;
1722 * Scan entries covered by the range. When working on the next
1723 * entry a restart need only re-loop on the current entry which
1724 * we have already locked, since 'next' may have changed. Also,
1725 * even though entry is safe, it may have been clipped so we
1726 * have to iterate forwards through the clip after sleeping.
1728 while (entry
->next
!= &map
->header
&& entry
->next
->start
< end
) {
1729 vm_map_entry_t next
= entry
->next
;
1731 if (flags
& MAP_CLIP_NO_HOLES
) {
1732 if (next
->start
> entry
->end
) {
1733 vm_map_unclip_range(map
, start_entry
,
1734 start
, entry
->end
, countp
, flags
);
1739 if (next
->eflags
& MAP_ENTRY_IN_TRANSITION
) {
1740 vm_offset_t save_end
= entry
->end
;
1741 next
->eflags
|= MAP_ENTRY_NEEDS_WAKEUP
;
1742 ++mycpu
->gd_cnt
.v_intrans_coll
;
1743 ++mycpu
->gd_cnt
.v_intrans_wait
;
1744 vm_map_transition_wait(map
);
1747 * clips might have occured while we blocked.
1749 CLIP_CHECK_FWD(entry
, save_end
);
1750 CLIP_CHECK_BACK(start_entry
, start
);
1754 * No restart necessary even though clip_end may block, we
1755 * are holding the map lock.
1757 vm_map_clip_end(map
, next
, end
, countp
);
1758 next
->eflags
|= MAP_ENTRY_IN_TRANSITION
;
1761 if (flags
& MAP_CLIP_NO_HOLES
) {
1762 if (entry
->end
!= end
) {
1763 vm_map_unclip_range(map
, start_entry
,
1764 start
, entry
->end
, countp
, flags
);
1768 return(start_entry
);
1772 * Undo the effect of vm_map_clip_range(). You should pass the same
1773 * flags and the same range that you passed to vm_map_clip_range().
1774 * This code will clear the in-transition flag on the entries and
1775 * wake up anyone waiting. This code will also simplify the sequence
1776 * and attempt to merge it with entries before and after the sequence.
1778 * The map must be locked on entry and will remain locked on return.
1780 * Note that you should also pass the start_entry returned by
1781 * vm_map_clip_range(). However, if you block between the two calls
1782 * with the map unlocked please be aware that the start_entry may
1783 * have been clipped and you may need to scan it backwards to find
1784 * the entry corresponding with the original start address. You are
1785 * responsible for this, vm_map_unclip_range() expects the correct
1786 * start_entry to be passed to it and will KASSERT otherwise.
1790 vm_map_unclip_range(vm_map_t map
, vm_map_entry_t start_entry
,
1791 vm_offset_t start
, vm_offset_t end
,
1792 int *countp
, int flags
)
1794 vm_map_entry_t entry
;
1796 entry
= start_entry
;
1798 KASSERT(entry
->start
== start
, ("unclip_range: illegal base entry"));
1799 while (entry
!= &map
->header
&& entry
->start
< end
) {
1800 KASSERT(entry
->eflags
& MAP_ENTRY_IN_TRANSITION
,
1801 ("in-transition flag not set during unclip on: %p",
1803 KASSERT(entry
->end
<= end
,
1804 ("unclip_range: tail wasn't clipped"));
1805 entry
->eflags
&= ~MAP_ENTRY_IN_TRANSITION
;
1806 if (entry
->eflags
& MAP_ENTRY_NEEDS_WAKEUP
) {
1807 entry
->eflags
&= ~MAP_ENTRY_NEEDS_WAKEUP
;
1810 entry
= entry
->next
;
1814 * Simplification does not block so there is no restart case.
1816 entry
= start_entry
;
1817 while (entry
!= &map
->header
&& entry
->start
< end
) {
1818 vm_map_simplify_entry(map
, entry
, countp
);
1819 entry
= entry
->next
;
1824 * Mark the given range as handled by a subordinate map.
1826 * This range must have been created with vm_map_find(), and no other
1827 * operations may have been performed on this range prior to calling
1830 * Submappings cannot be removed.
1835 vm_map_submap(vm_map_t map
, vm_offset_t start
, vm_offset_t end
, vm_map_t submap
)
1837 vm_map_entry_t entry
;
1838 int result
= KERN_INVALID_ARGUMENT
;
1841 count
= vm_map_entry_reserve(MAP_RESERVE_COUNT
);
1844 VM_MAP_RANGE_CHECK(map
, start
, end
);
1846 if (vm_map_lookup_entry(map
, start
, &entry
)) {
1847 vm_map_clip_start(map
, entry
, start
, &count
);
1849 entry
= entry
->next
;
1852 vm_map_clip_end(map
, entry
, end
, &count
);
1854 if ((entry
->start
== start
) && (entry
->end
== end
) &&
1855 ((entry
->eflags
& MAP_ENTRY_COW
) == 0) &&
1856 (entry
->object
.vm_object
== NULL
)) {
1857 entry
->object
.sub_map
= submap
;
1858 entry
->maptype
= VM_MAPTYPE_SUBMAP
;
1859 result
= KERN_SUCCESS
;
1862 vm_map_entry_release(count
);
1868 * Sets the protection of the specified address region in the target map.
1869 * If "set_max" is specified, the maximum protection is to be set;
1870 * otherwise, only the current protection is affected.
1872 * The protection is not applicable to submaps, but is applicable to normal
1873 * maps and maps governed by virtual page tables. For example, when operating
1874 * on a virtual page table our protection basically controls how COW occurs
1875 * on the backing object, whereas the virtual page table abstraction itself
1876 * is an abstraction for userland.
1881 vm_map_protect(vm_map_t map
, vm_offset_t start
, vm_offset_t end
,
1882 vm_prot_t new_prot
, boolean_t set_max
)
1884 vm_map_entry_t current
;
1885 vm_map_entry_t entry
;
1888 count
= vm_map_entry_reserve(MAP_RESERVE_COUNT
);
1891 VM_MAP_RANGE_CHECK(map
, start
, end
);
1893 if (vm_map_lookup_entry(map
, start
, &entry
)) {
1894 vm_map_clip_start(map
, entry
, start
, &count
);
1896 entry
= entry
->next
;
1900 * Make a first pass to check for protection violations.
1903 while ((current
!= &map
->header
) && (current
->start
< end
)) {
1904 if (current
->maptype
== VM_MAPTYPE_SUBMAP
) {
1906 vm_map_entry_release(count
);
1907 return (KERN_INVALID_ARGUMENT
);
1909 if ((new_prot
& current
->max_protection
) != new_prot
) {
1911 vm_map_entry_release(count
);
1912 return (KERN_PROTECTION_FAILURE
);
1914 current
= current
->next
;
1918 * Go back and fix up protections. [Note that clipping is not
1919 * necessary the second time.]
1923 while ((current
!= &map
->header
) && (current
->start
< end
)) {
1926 vm_map_clip_end(map
, current
, end
, &count
);
1928 old_prot
= current
->protection
;
1930 current
->max_protection
= new_prot
;
1931 current
->protection
= new_prot
& old_prot
;
1933 current
->protection
= new_prot
;
1937 * Update physical map if necessary. Worry about copy-on-write
1938 * here -- CHECK THIS XXX
1941 if (current
->protection
!= old_prot
) {
1942 #define MASK(entry) (((entry)->eflags & MAP_ENTRY_COW) ? ~VM_PROT_WRITE : \
1945 pmap_protect(map
->pmap
, current
->start
,
1947 current
->protection
& MASK(current
));
1951 vm_map_simplify_entry(map
, current
, &count
);
1953 current
= current
->next
;
1957 vm_map_entry_release(count
);
1958 return (KERN_SUCCESS
);
1962 * This routine traverses a processes map handling the madvise
1963 * system call. Advisories are classified as either those effecting
1964 * the vm_map_entry structure, or those effecting the underlying
1967 * The <value> argument is used for extended madvise calls.
1972 vm_map_madvise(vm_map_t map
, vm_offset_t start
, vm_offset_t end
,
1973 int behav
, off_t value
)
1975 vm_map_entry_t current
, entry
;
1981 * Some madvise calls directly modify the vm_map_entry, in which case
1982 * we need to use an exclusive lock on the map and we need to perform
1983 * various clipping operations. Otherwise we only need a read-lock
1986 count
= vm_map_entry_reserve(MAP_RESERVE_COUNT
);
1990 case MADV_SEQUENTIAL
:
2004 vm_map_lock_read(map
);
2007 vm_map_entry_release(count
);
2012 * Locate starting entry and clip if necessary.
2015 VM_MAP_RANGE_CHECK(map
, start
, end
);
2017 if (vm_map_lookup_entry(map
, start
, &entry
)) {
2019 vm_map_clip_start(map
, entry
, start
, &count
);
2021 entry
= entry
->next
;
2026 * madvise behaviors that are implemented in the vm_map_entry.
2028 * We clip the vm_map_entry so that behavioral changes are
2029 * limited to the specified address range.
2031 for (current
= entry
;
2032 (current
!= &map
->header
) && (current
->start
< end
);
2033 current
= current
->next
2035 if (current
->maptype
== VM_MAPTYPE_SUBMAP
)
2038 vm_map_clip_end(map
, current
, end
, &count
);
2042 vm_map_entry_set_behavior(current
, MAP_ENTRY_BEHAV_NORMAL
);
2044 case MADV_SEQUENTIAL
:
2045 vm_map_entry_set_behavior(current
, MAP_ENTRY_BEHAV_SEQUENTIAL
);
2048 vm_map_entry_set_behavior(current
, MAP_ENTRY_BEHAV_RANDOM
);
2051 current
->eflags
|= MAP_ENTRY_NOSYNC
;
2054 current
->eflags
&= ~MAP_ENTRY_NOSYNC
;
2057 current
->eflags
|= MAP_ENTRY_NOCOREDUMP
;
2060 current
->eflags
&= ~MAP_ENTRY_NOCOREDUMP
;
2064 * Set the page directory page for a map
2065 * governed by a virtual page table. Mark
2066 * the entry as being governed by a virtual
2067 * page table if it is not.
2069 * XXX the page directory page is stored
2070 * in the avail_ssize field if the map_entry.
2072 * XXX the map simplification code does not
2073 * compare this field so weird things may
2074 * happen if you do not apply this function
2075 * to the entire mapping governed by the
2076 * virtual page table.
2078 if (current
->maptype
!= VM_MAPTYPE_VPAGETABLE
) {
2082 current
->aux
.master_pde
= value
;
2083 pmap_remove(map
->pmap
,
2084 current
->start
, current
->end
);
2088 * Invalidate the related pmap entries, used
2089 * to flush portions of the real kernel's
2090 * pmap when the caller has removed or
2091 * modified existing mappings in a virtual
2094 * (exclusive locked map version does not
2095 * need the range interlock).
2097 pmap_remove(map
->pmap
,
2098 current
->start
, current
->end
);
2104 vm_map_simplify_entry(map
, current
, &count
);
2112 * madvise behaviors that are implemented in the underlying
2115 * Since we don't clip the vm_map_entry, we have to clip
2116 * the vm_object pindex and count.
2118 * NOTE! These functions are only supported on normal maps,
2119 * except MADV_INVAL which is also supported on
2120 * virtual page tables.
2122 for (current
= entry
;
2123 (current
!= &map
->header
) && (current
->start
< end
);
2124 current
= current
->next
2126 vm_offset_t useStart
;
2128 if (current
->maptype
!= VM_MAPTYPE_NORMAL
&&
2129 (current
->maptype
!= VM_MAPTYPE_VPAGETABLE
||
2130 behav
!= MADV_INVAL
)) {
2134 pindex
= OFF_TO_IDX(current
->offset
);
2135 delta
= atop(current
->end
- current
->start
);
2136 useStart
= current
->start
;
2138 if (current
->start
< start
) {
2139 pindex
+= atop(start
- current
->start
);
2140 delta
-= atop(start
- current
->start
);
2143 if (current
->end
> end
)
2144 delta
-= atop(current
->end
- end
);
2146 if ((vm_spindex_t
)delta
<= 0)
2149 if (behav
== MADV_INVAL
) {
2151 * Invalidate the related pmap entries, used
2152 * to flush portions of the real kernel's
2153 * pmap when the caller has removed or
2154 * modified existing mappings in a virtual
2157 * (shared locked map version needs the
2158 * interlock, see vm_fault()).
2160 struct vm_map_ilock ilock
;
2162 KASSERT(useStart
>= VM_MIN_USER_ADDRESS
&&
2163 useStart
+ ptoa(delta
) <=
2164 VM_MAX_USER_ADDRESS
,
2165 ("Bad range %016jx-%016jx (%016jx)",
2166 useStart
, useStart
+ ptoa(delta
),
2168 vm_map_interlock(map
, &ilock
,
2170 useStart
+ ptoa(delta
));
2171 pmap_remove(map
->pmap
,
2173 useStart
+ ptoa(delta
));
2174 vm_map_deinterlock(map
, &ilock
);
2176 vm_object_madvise(current
->object
.vm_object
,
2177 pindex
, delta
, behav
);
2181 * Try to populate the page table. Mappings governed
2182 * by virtual page tables cannot be pre-populated
2183 * without a lot of work so don't try.
2185 if (behav
== MADV_WILLNEED
&&
2186 current
->maptype
!= VM_MAPTYPE_VPAGETABLE
) {
2187 pmap_object_init_pt(
2190 current
->protection
,
2191 current
->object
.vm_object
,
2193 (count
<< PAGE_SHIFT
),
2194 MAP_PREFAULT_MADVISE
2198 vm_map_unlock_read(map
);
2200 vm_map_entry_release(count
);
2206 * Sets the inheritance of the specified address range in the target map.
2207 * Inheritance affects how the map will be shared with child maps at the
2208 * time of vm_map_fork.
2211 vm_map_inherit(vm_map_t map
, vm_offset_t start
, vm_offset_t end
,
2212 vm_inherit_t new_inheritance
)
2214 vm_map_entry_t entry
;
2215 vm_map_entry_t temp_entry
;
2218 switch (new_inheritance
) {
2219 case VM_INHERIT_NONE
:
2220 case VM_INHERIT_COPY
:
2221 case VM_INHERIT_SHARE
:
2224 return (KERN_INVALID_ARGUMENT
);
2227 count
= vm_map_entry_reserve(MAP_RESERVE_COUNT
);
2230 VM_MAP_RANGE_CHECK(map
, start
, end
);
2232 if (vm_map_lookup_entry(map
, start
, &temp_entry
)) {
2234 vm_map_clip_start(map
, entry
, start
, &count
);
2236 entry
= temp_entry
->next
;
2238 while ((entry
!= &map
->header
) && (entry
->start
< end
)) {
2239 vm_map_clip_end(map
, entry
, end
, &count
);
2241 entry
->inheritance
= new_inheritance
;
2243 vm_map_simplify_entry(map
, entry
, &count
);
2245 entry
= entry
->next
;
2248 vm_map_entry_release(count
);
2249 return (KERN_SUCCESS
);
2253 * Implement the semantics of mlock
2256 vm_map_unwire(vm_map_t map
, vm_offset_t start
, vm_offset_t real_end
,
2257 boolean_t new_pageable
)
2259 vm_map_entry_t entry
;
2260 vm_map_entry_t start_entry
;
2262 int rv
= KERN_SUCCESS
;
2265 count
= vm_map_entry_reserve(MAP_RESERVE_COUNT
);
2267 VM_MAP_RANGE_CHECK(map
, start
, real_end
);
2270 start_entry
= vm_map_clip_range(map
, start
, end
, &count
,
2272 if (start_entry
== NULL
) {
2274 vm_map_entry_release(count
);
2275 return (KERN_INVALID_ADDRESS
);
2278 if (new_pageable
== 0) {
2279 entry
= start_entry
;
2280 while ((entry
!= &map
->header
) && (entry
->start
< end
)) {
2281 vm_offset_t save_start
;
2282 vm_offset_t save_end
;
2285 * Already user wired or hard wired (trivial cases)
2287 if (entry
->eflags
& MAP_ENTRY_USER_WIRED
) {
2288 entry
= entry
->next
;
2291 if (entry
->wired_count
!= 0) {
2292 entry
->wired_count
++;
2293 entry
->eflags
|= MAP_ENTRY_USER_WIRED
;
2294 entry
= entry
->next
;
2299 * A new wiring requires instantiation of appropriate
2300 * management structures and the faulting in of the
2303 if (entry
->maptype
== VM_MAPTYPE_NORMAL
||
2304 entry
->maptype
== VM_MAPTYPE_VPAGETABLE
) {
2305 int copyflag
= entry
->eflags
&
2306 MAP_ENTRY_NEEDS_COPY
;
2307 if (copyflag
&& ((entry
->protection
&
2308 VM_PROT_WRITE
) != 0)) {
2309 vm_map_entry_shadow(entry
, 0);
2310 } else if (entry
->object
.vm_object
== NULL
&&
2312 vm_map_entry_allocate_object(entry
);
2315 entry
->wired_count
++;
2316 entry
->eflags
|= MAP_ENTRY_USER_WIRED
;
2319 * Now fault in the area. Note that vm_fault_wire()
2320 * may release the map lock temporarily, it will be
2321 * relocked on return. The in-transition
2322 * flag protects the entries.
2324 save_start
= entry
->start
;
2325 save_end
= entry
->end
;
2326 rv
= vm_fault_wire(map
, entry
, TRUE
, 0);
2328 CLIP_CHECK_BACK(entry
, save_start
);
2330 KASSERT(entry
->wired_count
== 1, ("bad wired_count on entry"));
2331 entry
->eflags
&= ~MAP_ENTRY_USER_WIRED
;
2332 entry
->wired_count
= 0;
2333 if (entry
->end
== save_end
)
2335 entry
= entry
->next
;
2336 KASSERT(entry
!= &map
->header
, ("bad entry clip during backout"));
2338 end
= save_start
; /* unwire the rest */
2342 * note that even though the entry might have been
2343 * clipped, the USER_WIRED flag we set prevents
2344 * duplication so we do not have to do a
2347 entry
= entry
->next
;
2351 * If we failed fall through to the unwiring section to
2352 * unwire what we had wired so far. 'end' has already
2359 * start_entry might have been clipped if we unlocked the
2360 * map and blocked. No matter how clipped it has gotten
2361 * there should be a fragment that is on our start boundary.
2363 CLIP_CHECK_BACK(start_entry
, start
);
2367 * Deal with the unwiring case.
2371 * This is the unwiring case. We must first ensure that the
2372 * range to be unwired is really wired down. We know there
2375 entry
= start_entry
;
2376 while ((entry
!= &map
->header
) && (entry
->start
< end
)) {
2377 if ((entry
->eflags
& MAP_ENTRY_USER_WIRED
) == 0) {
2378 rv
= KERN_INVALID_ARGUMENT
;
2381 KASSERT(entry
->wired_count
!= 0, ("wired count was 0 with USER_WIRED set! %p", entry
));
2382 entry
= entry
->next
;
2386 * Now decrement the wiring count for each region. If a region
2387 * becomes completely unwired, unwire its physical pages and
2391 * The map entries are processed in a loop, checking to
2392 * make sure the entry is wired and asserting it has a wired
2393 * count. However, another loop was inserted more-or-less in
2394 * the middle of the unwiring path. This loop picks up the
2395 * "entry" loop variable from the first loop without first
2396 * setting it to start_entry. Naturally, the secound loop
2397 * is never entered and the pages backing the entries are
2398 * never unwired. This can lead to a leak of wired pages.
2400 entry
= start_entry
;
2401 while ((entry
!= &map
->header
) && (entry
->start
< end
)) {
2402 KASSERT(entry
->eflags
& MAP_ENTRY_USER_WIRED
,
2403 ("expected USER_WIRED on entry %p", entry
));
2404 entry
->eflags
&= ~MAP_ENTRY_USER_WIRED
;
2405 entry
->wired_count
--;
2406 if (entry
->wired_count
== 0)
2407 vm_fault_unwire(map
, entry
);
2408 entry
= entry
->next
;
2412 vm_map_unclip_range(map
, start_entry
, start
, real_end
, &count
,
2416 vm_map_entry_release(count
);
2421 * Sets the pageability of the specified address range in the target map.
2422 * Regions specified as not pageable require locked-down physical
2423 * memory and physical page maps.
2425 * The map must not be locked, but a reference must remain to the map
2426 * throughout the call.
2428 * This function may be called via the zalloc path and must properly
2429 * reserve map entries for kernel_map.
2434 vm_map_wire(vm_map_t map
, vm_offset_t start
, vm_offset_t real_end
, int kmflags
)
2436 vm_map_entry_t entry
;
2437 vm_map_entry_t start_entry
;
2439 int rv
= KERN_SUCCESS
;
2442 if (kmflags
& KM_KRESERVE
)
2443 count
= vm_map_entry_kreserve(MAP_RESERVE_COUNT
);
2445 count
= vm_map_entry_reserve(MAP_RESERVE_COUNT
);
2447 VM_MAP_RANGE_CHECK(map
, start
, real_end
);
2450 start_entry
= vm_map_clip_range(map
, start
, end
, &count
,
2452 if (start_entry
== NULL
) {
2454 rv
= KERN_INVALID_ADDRESS
;
2457 if ((kmflags
& KM_PAGEABLE
) == 0) {
2461 * 1. Holding the write lock, we create any shadow or zero-fill
2462 * objects that need to be created. Then we clip each map
2463 * entry to the region to be wired and increment its wiring
2464 * count. We create objects before clipping the map entries
2465 * to avoid object proliferation.
2467 * 2. We downgrade to a read lock, and call vm_fault_wire to
2468 * fault in the pages for any newly wired area (wired_count is
2471 * Downgrading to a read lock for vm_fault_wire avoids a
2472 * possible deadlock with another process that may have faulted
2473 * on one of the pages to be wired (it would mark the page busy,
2474 * blocking us, then in turn block on the map lock that we
2475 * hold). Because of problems in the recursive lock package,
2476 * we cannot upgrade to a write lock in vm_map_lookup. Thus,
2477 * any actions that require the write lock must be done
2478 * beforehand. Because we keep the read lock on the map, the
2479 * copy-on-write status of the entries we modify here cannot
2482 entry
= start_entry
;
2483 while ((entry
!= &map
->header
) && (entry
->start
< end
)) {
2485 * Trivial case if the entry is already wired
2487 if (entry
->wired_count
) {
2488 entry
->wired_count
++;
2489 entry
= entry
->next
;
2494 * The entry is being newly wired, we have to setup
2495 * appropriate management structures. A shadow
2496 * object is required for a copy-on-write region,
2497 * or a normal object for a zero-fill region. We
2498 * do not have to do this for entries that point to sub
2499 * maps because we won't hold the lock on the sub map.
2501 if (entry
->maptype
== VM_MAPTYPE_NORMAL
||
2502 entry
->maptype
== VM_MAPTYPE_VPAGETABLE
) {
2503 int copyflag
= entry
->eflags
&
2504 MAP_ENTRY_NEEDS_COPY
;
2505 if (copyflag
&& ((entry
->protection
&
2506 VM_PROT_WRITE
) != 0)) {
2507 vm_map_entry_shadow(entry
, 0);
2508 } else if (entry
->object
.vm_object
== NULL
&&
2510 vm_map_entry_allocate_object(entry
);
2514 entry
->wired_count
++;
2515 entry
= entry
->next
;
2523 * HACK HACK HACK HACK
2525 * vm_fault_wire() temporarily unlocks the map to avoid
2526 * deadlocks. The in-transition flag from vm_map_clip_range
2527 * call should protect us from changes while the map is
2530 * NOTE: Previously this comment stated that clipping might
2531 * still occur while the entry is unlocked, but from
2532 * what I can tell it actually cannot.
2534 * It is unclear whether the CLIP_CHECK_*() calls
2535 * are still needed but we keep them in anyway.
2537 * HACK HACK HACK HACK
2540 entry
= start_entry
;
2541 while (entry
!= &map
->header
&& entry
->start
< end
) {
2543 * If vm_fault_wire fails for any page we need to undo
2544 * what has been done. We decrement the wiring count
2545 * for those pages which have not yet been wired (now)
2546 * and unwire those that have (later).
2548 vm_offset_t save_start
= entry
->start
;
2549 vm_offset_t save_end
= entry
->end
;
2551 if (entry
->wired_count
== 1)
2552 rv
= vm_fault_wire(map
, entry
, FALSE
, kmflags
);
2554 CLIP_CHECK_BACK(entry
, save_start
);
2556 KASSERT(entry
->wired_count
== 1, ("wired_count changed unexpectedly"));
2557 entry
->wired_count
= 0;
2558 if (entry
->end
== save_end
)
2560 entry
= entry
->next
;
2561 KASSERT(entry
!= &map
->header
, ("bad entry clip during backout"));
2566 CLIP_CHECK_FWD(entry
, save_end
);
2567 entry
= entry
->next
;
2571 * If a failure occured undo everything by falling through
2572 * to the unwiring code. 'end' has already been adjusted
2576 kmflags
|= KM_PAGEABLE
;
2579 * start_entry is still IN_TRANSITION but may have been
2580 * clipped since vm_fault_wire() unlocks and relocks the
2581 * map. No matter how clipped it has gotten there should
2582 * be a fragment that is on our start boundary.
2584 CLIP_CHECK_BACK(start_entry
, start
);
2587 if (kmflags
& KM_PAGEABLE
) {
2589 * This is the unwiring case. We must first ensure that the
2590 * range to be unwired is really wired down. We know there
2593 entry
= start_entry
;
2594 while ((entry
!= &map
->header
) && (entry
->start
< end
)) {
2595 if (entry
->wired_count
== 0) {
2596 rv
= KERN_INVALID_ARGUMENT
;
2599 entry
= entry
->next
;
2603 * Now decrement the wiring count for each region. If a region
2604 * becomes completely unwired, unwire its physical pages and
2607 entry
= start_entry
;
2608 while ((entry
!= &map
->header
) && (entry
->start
< end
)) {
2609 entry
->wired_count
--;
2610 if (entry
->wired_count
== 0)
2611 vm_fault_unwire(map
, entry
);
2612 entry
= entry
->next
;
2616 vm_map_unclip_range(map
, start_entry
, start
, real_end
,
2617 &count
, MAP_CLIP_NO_HOLES
);
2621 if (kmflags
& KM_KRESERVE
)
2622 vm_map_entry_krelease(count
);
2624 vm_map_entry_release(count
);
2629 * Mark a newly allocated address range as wired but do not fault in
2630 * the pages. The caller is expected to load the pages into the object.
2632 * The map must be locked on entry and will remain locked on return.
2633 * No other requirements.
2636 vm_map_set_wired_quick(vm_map_t map
, vm_offset_t addr
, vm_size_t size
,
2639 vm_map_entry_t scan
;
2640 vm_map_entry_t entry
;
2642 entry
= vm_map_clip_range(map
, addr
, addr
+ size
,
2643 countp
, MAP_CLIP_NO_HOLES
);
2645 scan
!= &map
->header
&& scan
->start
< addr
+ size
;
2646 scan
= scan
->next
) {
2647 KKASSERT(scan
->wired_count
== 0);
2648 scan
->wired_count
= 1;
2650 vm_map_unclip_range(map
, entry
, addr
, addr
+ size
,
2651 countp
, MAP_CLIP_NO_HOLES
);
2655 * Push any dirty cached pages in the address range to their pager.
2656 * If syncio is TRUE, dirty pages are written synchronously.
2657 * If invalidate is TRUE, any cached pages are freed as well.
2659 * This routine is called by sys_msync()
2661 * Returns an error if any part of the specified range is not mapped.
2666 vm_map_clean(vm_map_t map
, vm_offset_t start
, vm_offset_t end
,
2667 boolean_t syncio
, boolean_t invalidate
)
2669 vm_map_entry_t current
;
2670 vm_map_entry_t entry
;
2674 vm_ooffset_t offset
;
2676 vm_map_lock_read(map
);
2677 VM_MAP_RANGE_CHECK(map
, start
, end
);
2678 if (!vm_map_lookup_entry(map
, start
, &entry
)) {
2679 vm_map_unlock_read(map
);
2680 return (KERN_INVALID_ADDRESS
);
2682 lwkt_gettoken(&map
->token
);
2685 * Make a first pass to check for holes.
2687 for (current
= entry
; current
->start
< end
; current
= current
->next
) {
2688 if (current
->maptype
== VM_MAPTYPE_SUBMAP
) {
2689 lwkt_reltoken(&map
->token
);
2690 vm_map_unlock_read(map
);
2691 return (KERN_INVALID_ARGUMENT
);
2693 if (end
> current
->end
&&
2694 (current
->next
== &map
->header
||
2695 current
->end
!= current
->next
->start
)) {
2696 lwkt_reltoken(&map
->token
);
2697 vm_map_unlock_read(map
);
2698 return (KERN_INVALID_ADDRESS
);
2703 pmap_remove(vm_map_pmap(map
), start
, end
);
2706 * Make a second pass, cleaning/uncaching pages from the indicated
2709 for (current
= entry
; current
->start
< end
; current
= current
->next
) {
2710 offset
= current
->offset
+ (start
- current
->start
);
2711 size
= (end
<= current
->end
? end
: current
->end
) - start
;
2713 switch(current
->maptype
) {
2714 case VM_MAPTYPE_SUBMAP
:
2717 vm_map_entry_t tentry
;
2720 smap
= current
->object
.sub_map
;
2721 vm_map_lock_read(smap
);
2722 vm_map_lookup_entry(smap
, offset
, &tentry
);
2723 tsize
= tentry
->end
- offset
;
2726 object
= tentry
->object
.vm_object
;
2727 offset
= tentry
->offset
+ (offset
- tentry
->start
);
2728 vm_map_unlock_read(smap
);
2731 case VM_MAPTYPE_NORMAL
:
2732 case VM_MAPTYPE_VPAGETABLE
:
2733 object
= current
->object
.vm_object
;
2741 vm_object_hold(object
);
2744 * Note that there is absolutely no sense in writing out
2745 * anonymous objects, so we track down the vnode object
2747 * We invalidate (remove) all pages from the address space
2748 * anyway, for semantic correctness.
2750 * note: certain anonymous maps, such as MAP_NOSYNC maps,
2751 * may start out with a NULL object.
2753 while (object
&& (tobj
= object
->backing_object
) != NULL
) {
2754 vm_object_hold(tobj
);
2755 if (tobj
== object
->backing_object
) {
2756 vm_object_lock_swap();
2757 offset
+= object
->backing_object_offset
;
2758 vm_object_drop(object
);
2760 if (object
->size
< OFF_TO_IDX(offset
+ size
))
2761 size
= IDX_TO_OFF(object
->size
) -
2765 vm_object_drop(tobj
);
2767 if (object
&& (object
->type
== OBJT_VNODE
) &&
2768 (current
->protection
& VM_PROT_WRITE
) &&
2769 (object
->flags
& OBJ_NOMSYNC
) == 0) {
2771 * Flush pages if writing is allowed, invalidate them
2772 * if invalidation requested. Pages undergoing I/O
2773 * will be ignored by vm_object_page_remove().
2775 * We cannot lock the vnode and then wait for paging
2776 * to complete without deadlocking against vm_fault.
2777 * Instead we simply call vm_object_page_remove() and
2778 * allow it to block internally on a page-by-page
2779 * basis when it encounters pages undergoing async
2784 /* no chain wait needed for vnode objects */
2785 vm_object_reference_locked(object
);
2786 vn_lock(object
->handle
, LK_EXCLUSIVE
| LK_RETRY
);
2787 flags
= (syncio
|| invalidate
) ? OBJPC_SYNC
: 0;
2788 flags
|= invalidate
? OBJPC_INVAL
: 0;
2791 * When operating on a virtual page table just
2792 * flush the whole object. XXX we probably ought
2795 switch(current
->maptype
) {
2796 case VM_MAPTYPE_NORMAL
:
2797 vm_object_page_clean(object
,
2799 OFF_TO_IDX(offset
+ size
+ PAGE_MASK
),
2802 case VM_MAPTYPE_VPAGETABLE
:
2803 vm_object_page_clean(object
, 0, 0, flags
);
2806 vn_unlock(((struct vnode
*)object
->handle
));
2807 vm_object_deallocate_locked(object
);
2809 if (object
&& invalidate
&&
2810 ((object
->type
== OBJT_VNODE
) ||
2811 (object
->type
== OBJT_DEVICE
) ||
2812 (object
->type
== OBJT_MGTDEVICE
))) {
2814 ((object
->type
== OBJT_DEVICE
) ||
2815 (object
->type
== OBJT_MGTDEVICE
)) ? FALSE
: TRUE
;
2816 /* no chain wait needed for vnode/device objects */
2817 vm_object_reference_locked(object
);
2818 switch(current
->maptype
) {
2819 case VM_MAPTYPE_NORMAL
:
2820 vm_object_page_remove(object
,
2822 OFF_TO_IDX(offset
+ size
+ PAGE_MASK
),
2825 case VM_MAPTYPE_VPAGETABLE
:
2826 vm_object_page_remove(object
, 0, 0, clean_only
);
2829 vm_object_deallocate_locked(object
);
2833 vm_object_drop(object
);
2836 lwkt_reltoken(&map
->token
);
2837 vm_map_unlock_read(map
);
2839 return (KERN_SUCCESS
);
2843 * Make the region specified by this entry pageable.
2845 * The vm_map must be exclusively locked.
2848 vm_map_entry_unwire(vm_map_t map
, vm_map_entry_t entry
)
2850 entry
->eflags
&= ~MAP_ENTRY_USER_WIRED
;
2851 entry
->wired_count
= 0;
2852 vm_fault_unwire(map
, entry
);
2856 * Deallocate the given entry from the target map.
2858 * The vm_map must be exclusively locked.
2861 vm_map_entry_delete(vm_map_t map
, vm_map_entry_t entry
, int *countp
)
2863 vm_map_entry_unlink(map
, entry
);
2864 map
->size
-= entry
->end
- entry
->start
;
2866 switch(entry
->maptype
) {
2867 case VM_MAPTYPE_NORMAL
:
2868 case VM_MAPTYPE_VPAGETABLE
:
2869 case VM_MAPTYPE_SUBMAP
:
2870 vm_object_deallocate(entry
->object
.vm_object
);
2872 case VM_MAPTYPE_UKSMAP
:
2879 vm_map_entry_dispose(map
, entry
, countp
);
2883 * Deallocates the given address range from the target map.
2885 * The vm_map must be exclusively locked.
2888 vm_map_delete(vm_map_t map
, vm_offset_t start
, vm_offset_t end
, int *countp
)
2891 vm_map_entry_t entry
;
2892 vm_map_entry_t first_entry
;
2893 vm_offset_t hole_start
;
2895 ASSERT_VM_MAP_LOCKED(map
);
2896 lwkt_gettoken(&map
->token
);
2899 * Find the start of the region, and clip it. Set entry to point
2900 * at the first record containing the requested address or, if no
2901 * such record exists, the next record with a greater address. The
2902 * loop will run from this point until a record beyond the termination
2903 * address is encountered.
2905 * Adjust freehint[] for either the clip case or the extension case.
2907 * GGG see other GGG comment.
2909 if (vm_map_lookup_entry(map
, start
, &first_entry
)) {
2910 entry
= first_entry
;
2911 vm_map_clip_start(map
, entry
, start
, countp
);
2914 entry
= first_entry
->next
;
2915 if (entry
== &map
->header
)
2916 hole_start
= first_entry
->start
;
2918 hole_start
= first_entry
->end
;
2922 * Step through all entries in this region
2924 while ((entry
!= &map
->header
) && (entry
->start
< end
)) {
2925 vm_map_entry_t next
;
2927 vm_pindex_t offidxstart
, offidxend
, count
;
2930 * If we hit an in-transition entry we have to sleep and
2931 * retry. It's easier (and not really slower) to just retry
2932 * since this case occurs so rarely and the hint is already
2933 * pointing at the right place. We have to reset the
2934 * start offset so as not to accidently delete an entry
2935 * another process just created in vacated space.
2937 if (entry
->eflags
& MAP_ENTRY_IN_TRANSITION
) {
2938 entry
->eflags
|= MAP_ENTRY_NEEDS_WAKEUP
;
2939 start
= entry
->start
;
2940 ++mycpu
->gd_cnt
.v_intrans_coll
;
2941 ++mycpu
->gd_cnt
.v_intrans_wait
;
2942 vm_map_transition_wait(map
);
2945 vm_map_clip_end(map
, entry
, end
, countp
);
2951 offidxstart
= OFF_TO_IDX(entry
->offset
);
2952 count
= OFF_TO_IDX(e
- s
);
2954 switch(entry
->maptype
) {
2955 case VM_MAPTYPE_NORMAL
:
2956 case VM_MAPTYPE_VPAGETABLE
:
2957 case VM_MAPTYPE_SUBMAP
:
2958 object
= entry
->object
.vm_object
;
2966 * Unwire before removing addresses from the pmap; otherwise,
2967 * unwiring will put the entries back in the pmap.
2969 * Generally speaking, doing a bulk pmap_remove() before
2970 * removing the pages from the VM object is better at
2971 * reducing unnecessary IPIs. The pmap code is now optimized
2972 * to not blindly iterate the range when pt and pd pages
2975 if (entry
->wired_count
!= 0)
2976 vm_map_entry_unwire(map
, entry
);
2978 offidxend
= offidxstart
+ count
;
2980 if (object
== &kernel_object
) {
2981 pmap_remove(map
->pmap
, s
, e
);
2982 vm_object_hold(object
);
2983 vm_object_page_remove(object
, offidxstart
,
2985 vm_object_drop(object
);
2986 } else if (object
&& object
->type
!= OBJT_DEFAULT
&&
2987 object
->type
!= OBJT_SWAP
) {
2989 * vnode object routines cannot be chain-locked,
2990 * but since we aren't removing pages from the
2991 * object here we can use a shared hold.
2993 vm_object_hold_shared(object
);
2994 pmap_remove(map
->pmap
, s
, e
);
2995 vm_object_drop(object
);
2996 } else if (object
) {
2997 vm_object_hold(object
);
2998 vm_object_chain_acquire(object
, 0);
2999 pmap_remove(map
->pmap
, s
, e
);
3001 if (object
!= NULL
&&
3002 object
->ref_count
!= 1 &&
3003 (object
->flags
& (OBJ_NOSPLIT
|OBJ_ONEMAPPING
)) ==
3005 (object
->type
== OBJT_DEFAULT
||
3006 object
->type
== OBJT_SWAP
)) {
3007 vm_object_collapse(object
, NULL
);
3008 vm_object_page_remove(object
, offidxstart
,
3010 if (object
->type
== OBJT_SWAP
) {
3011 swap_pager_freespace(object
,
3015 if (offidxend
>= object
->size
&&
3016 offidxstart
< object
->size
) {
3017 object
->size
= offidxstart
;
3020 vm_object_chain_release(object
);
3021 vm_object_drop(object
);
3022 } else if (entry
->maptype
== VM_MAPTYPE_UKSMAP
) {
3023 pmap_remove(map
->pmap
, s
, e
);
3027 * Delete the entry (which may delete the object) only after
3028 * removing all pmap entries pointing to its pages.
3029 * (Otherwise, its page frames may be reallocated, and any
3030 * modify bits will be set in the wrong object!)
3032 vm_map_entry_delete(map
, entry
, countp
);
3035 if (entry
== &map
->header
)
3036 vm_map_freehint_hole(map
, hole_start
, entry
->end
- hole_start
);
3038 vm_map_freehint_hole(map
, hole_start
,
3039 entry
->start
- hole_start
);
3041 lwkt_reltoken(&map
->token
);
3043 return (KERN_SUCCESS
);
3047 * Remove the given address range from the target map.
3048 * This is the exported form of vm_map_delete.
3053 vm_map_remove(vm_map_t map
, vm_offset_t start
, vm_offset_t end
)
3058 count
= vm_map_entry_reserve(MAP_RESERVE_COUNT
);
3060 VM_MAP_RANGE_CHECK(map
, start
, end
);
3061 result
= vm_map_delete(map
, start
, end
, &count
);
3063 vm_map_entry_release(count
);
3069 * Assert that the target map allows the specified privilege on the
3070 * entire address region given. The entire region must be allocated.
3072 * The caller must specify whether the vm_map is already locked or not.
3075 vm_map_check_protection(vm_map_t map
, vm_offset_t start
, vm_offset_t end
,
3076 vm_prot_t protection
, boolean_t have_lock
)
3078 vm_map_entry_t entry
;
3079 vm_map_entry_t tmp_entry
;
3082 if (have_lock
== FALSE
)
3083 vm_map_lock_read(map
);
3085 if (!vm_map_lookup_entry(map
, start
, &tmp_entry
)) {
3086 if (have_lock
== FALSE
)
3087 vm_map_unlock_read(map
);
3093 while (start
< end
) {
3094 if (entry
== &map
->header
) {
3102 if (start
< entry
->start
) {
3107 * Check protection associated with entry.
3110 if ((entry
->protection
& protection
) != protection
) {
3114 /* go to next entry */
3117 entry
= entry
->next
;
3119 if (have_lock
== FALSE
)
3120 vm_map_unlock_read(map
);
3125 * If appropriate this function shadows the original object with a new object
3126 * and moves the VM pages from the original object to the new object.
3127 * The original object will also be collapsed, if possible.
3129 * We can only do this for normal memory objects with a single mapping, and
3130 * it only makes sense to do it if there are 2 or more refs on the original
3131 * object. i.e. typically a memory object that has been extended into
3132 * multiple vm_map_entry's with non-overlapping ranges.
3134 * This makes it easier to remove unused pages and keeps object inheritance
3135 * from being a negative impact on memory usage.
3137 * On return the (possibly new) entry->object.vm_object will have an
3138 * additional ref on it for the caller to dispose of (usually by cloning
3139 * the vm_map_entry). The additional ref had to be done in this routine
3140 * to avoid racing a collapse. The object's ONEMAPPING flag will also be
3143 * The vm_map must be locked and its token held.
3146 vm_map_split(vm_map_entry_t entry
)
3149 vm_object_t oobject
, nobject
, bobject
;
3152 vm_pindex_t offidxstart
, offidxend
, idx
;
3154 vm_ooffset_t offset
;
3158 * Optimize away object locks for vnode objects. Important exit/exec
3161 * OBJ_ONEMAPPING doesn't apply to vnode objects but clear the flag
3164 oobject
= entry
->object
.vm_object
;
3165 if (oobject
->type
!= OBJT_DEFAULT
&& oobject
->type
!= OBJT_SWAP
) {
3166 vm_object_reference_quick(oobject
);
3167 vm_object_clear_flag(oobject
, OBJ_ONEMAPPING
);
3172 * Setup. Chain lock the original object throughout the entire
3173 * routine to prevent new page faults from occuring.
3175 * XXX can madvise WILLNEED interfere with us too?
3177 vm_object_hold(oobject
);
3178 vm_object_chain_acquire(oobject
, 0);
3181 * Original object cannot be split? Might have also changed state.
3183 if (oobject
->handle
== NULL
|| (oobject
->type
!= OBJT_DEFAULT
&&
3184 oobject
->type
!= OBJT_SWAP
)) {
3185 vm_object_chain_release(oobject
);
3186 vm_object_reference_locked(oobject
);
3187 vm_object_clear_flag(oobject
, OBJ_ONEMAPPING
);
3188 vm_object_drop(oobject
);
3193 * Collapse original object with its backing store as an
3194 * optimization to reduce chain lengths when possible.
3196 * If ref_count <= 1 there aren't other non-overlapping vm_map_entry's
3197 * for oobject, so there's no point collapsing it.
3199 * Then re-check whether the object can be split.
3201 vm_object_collapse(oobject
, NULL
);
3203 if (oobject
->ref_count
<= 1 ||
3204 (oobject
->type
!= OBJT_DEFAULT
&& oobject
->type
!= OBJT_SWAP
) ||
3205 (oobject
->flags
& (OBJ_NOSPLIT
|OBJ_ONEMAPPING
)) != OBJ_ONEMAPPING
) {
3206 vm_object_chain_release(oobject
);
3207 vm_object_reference_locked(oobject
);
3208 vm_object_clear_flag(oobject
, OBJ_ONEMAPPING
);
3209 vm_object_drop(oobject
);
3214 * Acquire the chain lock on the backing object.
3216 * Give bobject an additional ref count for when it will be shadowed
3220 if ((bobject
= oobject
->backing_object
) != NULL
) {
3221 if (bobject
->type
!= OBJT_VNODE
) {
3223 vm_object_hold(bobject
);
3224 vm_object_chain_wait(bobject
, 0);
3225 /* ref for shadowing below */
3226 vm_object_reference_locked(bobject
);
3227 vm_object_chain_acquire(bobject
, 0);
3228 KKASSERT(bobject
->backing_object
== bobject
);
3229 KKASSERT((bobject
->flags
& OBJ_DEAD
) == 0);
3232 * vnodes are not placed on the shadow list but
3233 * they still get another ref for the backing_object
3236 vm_object_reference_quick(bobject
);
3241 * Calculate the object page range and allocate the new object.
3243 offset
= entry
->offset
;
3247 offidxstart
= OFF_TO_IDX(offset
);
3248 offidxend
= offidxstart
+ OFF_TO_IDX(e
- s
);
3249 size
= offidxend
- offidxstart
;
3251 switch(oobject
->type
) {
3253 nobject
= default_pager_alloc(NULL
, IDX_TO_OFF(size
),
3257 nobject
= swap_pager_alloc(NULL
, IDX_TO_OFF(size
),
3266 if (nobject
== NULL
) {
3268 if (useshadowlist
) {
3269 vm_object_chain_release(bobject
);
3270 vm_object_deallocate(bobject
);
3271 vm_object_drop(bobject
);
3273 vm_object_deallocate(bobject
);
3276 vm_object_chain_release(oobject
);
3277 vm_object_reference_locked(oobject
);
3278 vm_object_clear_flag(oobject
, OBJ_ONEMAPPING
);
3279 vm_object_drop(oobject
);
3284 * The new object will replace entry->object.vm_object so it needs
3285 * a second reference (the caller expects an additional ref).
3287 vm_object_hold(nobject
);
3288 vm_object_reference_locked(nobject
);
3289 vm_object_chain_acquire(nobject
, 0);
3292 * nobject shadows bobject (oobject already shadows bobject).
3294 * Adding an object to bobject's shadow list requires refing bobject
3295 * which we did above in the useshadowlist case.
3298 nobject
->backing_object_offset
=
3299 oobject
->backing_object_offset
+ IDX_TO_OFF(offidxstart
);
3300 nobject
->backing_object
= bobject
;
3301 if (useshadowlist
) {
3302 bobject
->shadow_count
++;
3303 atomic_add_int(&bobject
->generation
, 1);
3304 LIST_INSERT_HEAD(&bobject
->shadow_head
,
3305 nobject
, shadow_list
);
3306 vm_object_clear_flag(bobject
, OBJ_ONEMAPPING
); /*XXX*/
3307 vm_object_chain_release(bobject
);
3308 vm_object_drop(bobject
);
3309 vm_object_set_flag(nobject
, OBJ_ONSHADOW
);
3314 * Move the VM pages from oobject to nobject
3316 for (idx
= 0; idx
< size
; idx
++) {
3319 m
= vm_page_lookup_busy_wait(oobject
, offidxstart
+ idx
,
3325 * We must wait for pending I/O to complete before we can
3328 * We do not have to VM_PROT_NONE the page as mappings should
3329 * not be changed by this operation.
3331 * NOTE: The act of renaming a page updates chaingen for both
3334 vm_page_rename(m
, nobject
, idx
);
3335 /* page automatically made dirty by rename and cache handled */
3336 /* page remains busy */
3339 if (oobject
->type
== OBJT_SWAP
) {
3340 vm_object_pip_add(oobject
, 1);
3342 * copy oobject pages into nobject and destroy unneeded
3343 * pages in shadow object.
3345 swap_pager_copy(oobject
, nobject
, offidxstart
, 0);
3346 vm_object_pip_wakeup(oobject
);
3350 * Wakeup the pages we played with. No spl protection is needed
3351 * for a simple wakeup.
3353 for (idx
= 0; idx
< size
; idx
++) {
3354 m
= vm_page_lookup(nobject
, idx
);
3356 KKASSERT(m
->flags
& PG_BUSY
);
3360 entry
->object
.vm_object
= nobject
;
3361 entry
->offset
= 0LL;
3366 * NOTE: There is no need to remove OBJ_ONEMAPPING from oobject, the
3367 * related pages were moved and are no longer applicable to the
3370 * NOTE: Deallocate oobject (due to its entry->object.vm_object being
3371 * replaced by nobject).
3373 vm_object_chain_release(nobject
);
3374 vm_object_drop(nobject
);
3375 if (bobject
&& useshadowlist
) {
3376 vm_object_chain_release(bobject
);
3377 vm_object_drop(bobject
);
3379 vm_object_chain_release(oobject
);
3380 /*vm_object_clear_flag(oobject, OBJ_ONEMAPPING);*/
3381 vm_object_deallocate_locked(oobject
);
3382 vm_object_drop(oobject
);
3386 * Copies the contents of the source entry to the destination
3387 * entry. The entries *must* be aligned properly.
3389 * The vm_maps must be exclusively locked.
3390 * The vm_map's token must be held.
3392 * Because the maps are locked no faults can be in progress during the
3396 vm_map_copy_entry(vm_map_t src_map
, vm_map_t dst_map
,
3397 vm_map_entry_t src_entry
, vm_map_entry_t dst_entry
)
3399 vm_object_t src_object
;
3401 if (dst_entry
->maptype
== VM_MAPTYPE_SUBMAP
||
3402 dst_entry
->maptype
== VM_MAPTYPE_UKSMAP
)
3404 if (src_entry
->maptype
== VM_MAPTYPE_SUBMAP
||
3405 src_entry
->maptype
== VM_MAPTYPE_UKSMAP
)
3408 if (src_entry
->wired_count
== 0) {
3410 * If the source entry is marked needs_copy, it is already
3413 if ((src_entry
->eflags
& MAP_ENTRY_NEEDS_COPY
) == 0) {
3414 pmap_protect(src_map
->pmap
,
3417 src_entry
->protection
& ~VM_PROT_WRITE
);
3421 * Make a copy of the object.
3423 * The object must be locked prior to checking the object type
3424 * and for the call to vm_object_collapse() and vm_map_split().
3425 * We cannot use *_hold() here because the split code will
3426 * probably try to destroy the object. The lock is a pool
3427 * token and doesn't care.
3429 * We must bump src_map->timestamp when setting
3430 * MAP_ENTRY_NEEDS_COPY to force any concurrent fault
3431 * to retry, otherwise the concurrent fault might improperly
3432 * install a RW pte when its supposed to be a RO(COW) pte.
3433 * This race can occur because a vnode-backed fault may have
3434 * to temporarily release the map lock.
3436 if (src_entry
->object
.vm_object
!= NULL
) {
3437 vm_map_split(src_entry
);
3438 src_object
= src_entry
->object
.vm_object
;
3439 dst_entry
->object
.vm_object
= src_object
;
3440 src_entry
->eflags
|= (MAP_ENTRY_COW
|
3441 MAP_ENTRY_NEEDS_COPY
);
3442 dst_entry
->eflags
|= (MAP_ENTRY_COW
|
3443 MAP_ENTRY_NEEDS_COPY
);
3444 dst_entry
->offset
= src_entry
->offset
;
3445 ++src_map
->timestamp
;
3447 dst_entry
->object
.vm_object
= NULL
;
3448 dst_entry
->offset
= 0;
3451 pmap_copy(dst_map
->pmap
, src_map
->pmap
, dst_entry
->start
,
3452 dst_entry
->end
- dst_entry
->start
, src_entry
->start
);
3455 * Of course, wired down pages can't be set copy-on-write.
3456 * Cause wired pages to be copied into the new map by
3457 * simulating faults (the new pages are pageable)
3459 vm_fault_copy_entry(dst_map
, src_map
, dst_entry
, src_entry
);
3465 * Create a new process vmspace structure and vm_map
3466 * based on those of an existing process. The new map
3467 * is based on the old map, according to the inheritance
3468 * values on the regions in that map.
3470 * The source map must not be locked.
3473 static void vmspace_fork_normal_entry(vm_map_t old_map
, vm_map_t new_map
,
3474 vm_map_entry_t old_entry
, int *countp
);
3475 static void vmspace_fork_uksmap_entry(vm_map_t old_map
, vm_map_t new_map
,
3476 vm_map_entry_t old_entry
, int *countp
);
3479 vmspace_fork(struct vmspace
*vm1
)
3481 struct vmspace
*vm2
;
3482 vm_map_t old_map
= &vm1
->vm_map
;
3484 vm_map_entry_t old_entry
;
3487 lwkt_gettoken(&vm1
->vm_map
.token
);
3488 vm_map_lock(old_map
);
3490 vm2
= vmspace_alloc(old_map
->min_offset
, old_map
->max_offset
);
3491 lwkt_gettoken(&vm2
->vm_map
.token
);
3492 bcopy(&vm1
->vm_startcopy
, &vm2
->vm_startcopy
,
3493 (caddr_t
)&vm1
->vm_endcopy
- (caddr_t
)&vm1
->vm_startcopy
);
3494 new_map
= &vm2
->vm_map
; /* XXX */
3495 new_map
->timestamp
= 1;
3497 vm_map_lock(new_map
);
3500 old_entry
= old_map
->header
.next
;
3501 while (old_entry
!= &old_map
->header
) {
3503 old_entry
= old_entry
->next
;
3506 count
= vm_map_entry_reserve(count
+ MAP_RESERVE_COUNT
);
3508 old_entry
= old_map
->header
.next
;
3509 while (old_entry
!= &old_map
->header
) {
3510 switch(old_entry
->maptype
) {
3511 case VM_MAPTYPE_SUBMAP
:
3512 panic("vm_map_fork: encountered a submap");
3514 case VM_MAPTYPE_UKSMAP
:
3515 vmspace_fork_uksmap_entry(old_map
, new_map
,
3518 case VM_MAPTYPE_NORMAL
:
3519 case VM_MAPTYPE_VPAGETABLE
:
3520 vmspace_fork_normal_entry(old_map
, new_map
,
3524 old_entry
= old_entry
->next
;
3527 new_map
->size
= old_map
->size
;
3528 vm_map_unlock(old_map
);
3529 vm_map_unlock(new_map
);
3530 vm_map_entry_release(count
);
3532 lwkt_reltoken(&vm2
->vm_map
.token
);
3533 lwkt_reltoken(&vm1
->vm_map
.token
);
3540 vmspace_fork_normal_entry(vm_map_t old_map
, vm_map_t new_map
,
3541 vm_map_entry_t old_entry
, int *countp
)
3543 vm_map_entry_t new_entry
;
3546 switch (old_entry
->inheritance
) {
3547 case VM_INHERIT_NONE
:
3549 case VM_INHERIT_SHARE
:
3551 * Clone the entry, creating the shared object if
3554 if (old_entry
->object
.vm_object
== NULL
)
3555 vm_map_entry_allocate_object(old_entry
);
3557 if (old_entry
->eflags
& MAP_ENTRY_NEEDS_COPY
) {
3559 * Shadow a map_entry which needs a copy,
3560 * replacing its object with a new object
3561 * that points to the old one. Ask the
3562 * shadow code to automatically add an
3563 * additional ref. We can't do it afterwords
3564 * because we might race a collapse. The call
3565 * to vm_map_entry_shadow() will also clear
3568 vm_map_entry_shadow(old_entry
, 1);
3569 } else if (old_entry
->object
.vm_object
) {
3571 * We will make a shared copy of the object,
3572 * and must clear OBJ_ONEMAPPING.
3574 * Optimize vnode objects. OBJ_ONEMAPPING
3575 * is non-applicable but clear it anyway,
3576 * and its terminal so we don'th ave to deal
3577 * with chains. Reduces SMP conflicts.
3579 * XXX assert that object.vm_object != NULL
3580 * since we allocate it above.
3582 object
= old_entry
->object
.vm_object
;
3583 if (object
->type
== OBJT_VNODE
) {
3584 vm_object_reference_quick(object
);
3585 vm_object_clear_flag(object
,
3588 vm_object_hold(object
);
3589 vm_object_chain_wait(object
, 0);
3590 vm_object_reference_locked(object
);
3591 vm_object_clear_flag(object
,
3593 vm_object_drop(object
);
3598 * Clone the entry. We've already bumped the ref on
3601 new_entry
= vm_map_entry_create(new_map
, countp
);
3602 *new_entry
= *old_entry
;
3603 new_entry
->eflags
&= ~MAP_ENTRY_USER_WIRED
;
3604 new_entry
->wired_count
= 0;
3607 * Insert the entry into the new map -- we know we're
3608 * inserting at the end of the new map.
3611 vm_map_entry_link(new_map
, new_map
->header
.prev
,
3615 * Update the physical map
3617 pmap_copy(new_map
->pmap
, old_map
->pmap
,
3619 (old_entry
->end
- old_entry
->start
),
3622 case VM_INHERIT_COPY
:
3624 * Clone the entry and link into the map.
3626 new_entry
= vm_map_entry_create(new_map
, countp
);
3627 *new_entry
= *old_entry
;
3628 new_entry
->eflags
&= ~MAP_ENTRY_USER_WIRED
;
3629 new_entry
->wired_count
= 0;
3630 new_entry
->object
.vm_object
= NULL
;
3631 vm_map_entry_link(new_map
, new_map
->header
.prev
,
3633 vm_map_copy_entry(old_map
, new_map
, old_entry
,
3640 * When forking user-kernel shared maps, the map might change in the
3641 * child so do not try to copy the underlying pmap entries.
3645 vmspace_fork_uksmap_entry(vm_map_t old_map
, vm_map_t new_map
,
3646 vm_map_entry_t old_entry
, int *countp
)
3648 vm_map_entry_t new_entry
;
3650 new_entry
= vm_map_entry_create(new_map
, countp
);
3651 *new_entry
= *old_entry
;
3652 new_entry
->eflags
&= ~MAP_ENTRY_USER_WIRED
;
3653 new_entry
->wired_count
= 0;
3654 vm_map_entry_link(new_map
, new_map
->header
.prev
,
3659 * Create an auto-grow stack entry
3664 vm_map_stack (vm_map_t map
, vm_offset_t addrbos
, vm_size_t max_ssize
,
3665 int flags
, vm_prot_t prot
, vm_prot_t max
, int cow
)
3667 vm_map_entry_t prev_entry
;
3668 vm_map_entry_t new_stack_entry
;
3669 vm_size_t init_ssize
;
3672 vm_offset_t tmpaddr
;
3674 cow
|= MAP_IS_STACK
;
3676 if (max_ssize
< sgrowsiz
)
3677 init_ssize
= max_ssize
;
3679 init_ssize
= sgrowsiz
;
3681 count
= vm_map_entry_reserve(MAP_RESERVE_COUNT
);
3685 * Find space for the mapping
3687 if ((flags
& (MAP_FIXED
| MAP_TRYFIXED
)) == 0) {
3688 if (vm_map_findspace(map
, addrbos
, max_ssize
, 1,
3691 vm_map_entry_release(count
);
3692 return (KERN_NO_SPACE
);
3697 /* If addr is already mapped, no go */
3698 if (vm_map_lookup_entry(map
, addrbos
, &prev_entry
)) {
3700 vm_map_entry_release(count
);
3701 return (KERN_NO_SPACE
);
3705 /* XXX already handled by kern_mmap() */
3706 /* If we would blow our VMEM resource limit, no go */
3707 if (map
->size
+ init_ssize
>
3708 curproc
->p_rlimit
[RLIMIT_VMEM
].rlim_cur
) {
3710 vm_map_entry_release(count
);
3711 return (KERN_NO_SPACE
);
3716 * If we can't accomodate max_ssize in the current mapping,
3717 * no go. However, we need to be aware that subsequent user
3718 * mappings might map into the space we have reserved for
3719 * stack, and currently this space is not protected.
3721 * Hopefully we will at least detect this condition
3722 * when we try to grow the stack.
3724 if ((prev_entry
->next
!= &map
->header
) &&
3725 (prev_entry
->next
->start
< addrbos
+ max_ssize
)) {
3727 vm_map_entry_release(count
);
3728 return (KERN_NO_SPACE
);
3732 * We initially map a stack of only init_ssize. We will
3733 * grow as needed later. Since this is to be a grow
3734 * down stack, we map at the top of the range.
3736 * Note: we would normally expect prot and max to be
3737 * VM_PROT_ALL, and cow to be 0. Possibly we should
3738 * eliminate these as input parameters, and just
3739 * pass these values here in the insert call.
3741 rv
= vm_map_insert(map
, &count
, NULL
, NULL
,
3742 0, addrbos
+ max_ssize
- init_ssize
,
3743 addrbos
+ max_ssize
,
3745 VM_SUBSYS_STACK
, prot
, max
, cow
);
3747 /* Now set the avail_ssize amount */
3748 if (rv
== KERN_SUCCESS
) {
3749 if (prev_entry
!= &map
->header
)
3750 vm_map_clip_end(map
, prev_entry
, addrbos
+ max_ssize
- init_ssize
, &count
);
3751 new_stack_entry
= prev_entry
->next
;
3752 if (new_stack_entry
->end
!= addrbos
+ max_ssize
||
3753 new_stack_entry
->start
!= addrbos
+ max_ssize
- init_ssize
)
3754 panic ("Bad entry start/end for new stack entry");
3756 new_stack_entry
->aux
.avail_ssize
= max_ssize
- init_ssize
;
3760 vm_map_entry_release(count
);
3765 * Attempts to grow a vm stack entry. Returns KERN_SUCCESS if the
3766 * desired address is already mapped, or if we successfully grow
3767 * the stack. Also returns KERN_SUCCESS if addr is outside the
3768 * stack range (this is strange, but preserves compatibility with
3769 * the grow function in vm_machdep.c).
3774 vm_map_growstack (vm_map_t map
, vm_offset_t addr
)
3776 vm_map_entry_t prev_entry
;
3777 vm_map_entry_t stack_entry
;
3778 vm_map_entry_t new_stack_entry
;
3784 int rv
= KERN_SUCCESS
;
3786 int use_read_lock
= 1;
3792 lp
= curthread
->td_lwp
;
3793 p
= curthread
->td_proc
;
3794 KKASSERT(lp
!= NULL
);
3795 vm
= lp
->lwp_vmspace
;
3798 * Growstack is only allowed on the current process. We disallow
3799 * other use cases, e.g. trying to access memory via procfs that
3800 * the stack hasn't grown into.
3802 if (map
!= &vm
->vm_map
) {
3803 return KERN_FAILURE
;
3806 count
= vm_map_entry_reserve(MAP_RESERVE_COUNT
);
3809 vm_map_lock_read(map
);
3813 /* If addr is already in the entry range, no need to grow.*/
3814 if (vm_map_lookup_entry(map
, addr
, &prev_entry
))
3817 if ((stack_entry
= prev_entry
->next
) == &map
->header
)
3819 if (prev_entry
== &map
->header
)
3820 end
= stack_entry
->start
- stack_entry
->aux
.avail_ssize
;
3822 end
= prev_entry
->end
;
3825 * This next test mimics the old grow function in vm_machdep.c.
3826 * It really doesn't quite make sense, but we do it anyway
3827 * for compatibility.
3829 * If not growable stack, return success. This signals the
3830 * caller to proceed as he would normally with normal vm.
3832 if (stack_entry
->aux
.avail_ssize
< 1 ||
3833 addr
>= stack_entry
->start
||
3834 addr
< stack_entry
->start
- stack_entry
->aux
.avail_ssize
) {
3838 /* Find the minimum grow amount */
3839 grow_amount
= roundup (stack_entry
->start
- addr
, PAGE_SIZE
);
3840 if (grow_amount
> stack_entry
->aux
.avail_ssize
) {
3846 * If there is no longer enough space between the entries
3847 * nogo, and adjust the available space. Note: this
3848 * should only happen if the user has mapped into the
3849 * stack area after the stack was created, and is
3850 * probably an error.
3852 * This also effectively destroys any guard page the user
3853 * might have intended by limiting the stack size.
3855 if (grow_amount
> stack_entry
->start
- end
) {
3856 if (use_read_lock
&& vm_map_lock_upgrade(map
)) {
3862 stack_entry
->aux
.avail_ssize
= stack_entry
->start
- end
;
3867 is_procstack
= addr
>= (vm_offset_t
)vm
->vm_maxsaddr
;
3869 /* If this is the main process stack, see if we're over the
3872 if (is_procstack
&& (ctob(vm
->vm_ssize
) + grow_amount
>
3873 p
->p_rlimit
[RLIMIT_STACK
].rlim_cur
)) {
3878 /* Round up the grow amount modulo SGROWSIZ */
3879 grow_amount
= roundup (grow_amount
, sgrowsiz
);
3880 if (grow_amount
> stack_entry
->aux
.avail_ssize
) {
3881 grow_amount
= stack_entry
->aux
.avail_ssize
;
3883 if (is_procstack
&& (ctob(vm
->vm_ssize
) + grow_amount
>
3884 p
->p_rlimit
[RLIMIT_STACK
].rlim_cur
)) {
3885 grow_amount
= p
->p_rlimit
[RLIMIT_STACK
].rlim_cur
-
3889 /* If we would blow our VMEM resource limit, no go */
3890 if (map
->size
+ grow_amount
> p
->p_rlimit
[RLIMIT_VMEM
].rlim_cur
) {
3895 if (use_read_lock
&& vm_map_lock_upgrade(map
)) {
3902 /* Get the preliminary new entry start value */
3903 addr
= stack_entry
->start
- grow_amount
;
3905 /* If this puts us into the previous entry, cut back our growth
3906 * to the available space. Also, see the note above.
3909 stack_entry
->aux
.avail_ssize
= stack_entry
->start
- end
;
3913 rv
= vm_map_insert(map
, &count
, NULL
, NULL
,
3914 0, addr
, stack_entry
->start
,
3916 VM_SUBSYS_STACK
, VM_PROT_ALL
, VM_PROT_ALL
, 0);
3918 /* Adjust the available stack space by the amount we grew. */
3919 if (rv
== KERN_SUCCESS
) {
3920 if (prev_entry
!= &map
->header
)
3921 vm_map_clip_end(map
, prev_entry
, addr
, &count
);
3922 new_stack_entry
= prev_entry
->next
;
3923 if (new_stack_entry
->end
!= stack_entry
->start
||
3924 new_stack_entry
->start
!= addr
)
3925 panic ("Bad stack grow start/end in new stack entry");
3927 new_stack_entry
->aux
.avail_ssize
=
3928 stack_entry
->aux
.avail_ssize
-
3929 (new_stack_entry
->end
- new_stack_entry
->start
);
3931 vm
->vm_ssize
+= btoc(new_stack_entry
->end
-
3932 new_stack_entry
->start
);
3935 if (map
->flags
& MAP_WIREFUTURE
)
3936 vm_map_unwire(map
, new_stack_entry
->start
,
3937 new_stack_entry
->end
, FALSE
);
3942 vm_map_unlock_read(map
);
3945 vm_map_entry_release(count
);
3950 * Unshare the specified VM space for exec. If other processes are
3951 * mapped to it, then create a new one. The new vmspace is null.
3956 vmspace_exec(struct proc
*p
, struct vmspace
*vmcopy
)
3958 struct vmspace
*oldvmspace
= p
->p_vmspace
;
3959 struct vmspace
*newvmspace
;
3960 vm_map_t map
= &p
->p_vmspace
->vm_map
;
3963 * If we are execing a resident vmspace we fork it, otherwise
3964 * we create a new vmspace. Note that exitingcnt is not
3965 * copied to the new vmspace.
3967 lwkt_gettoken(&oldvmspace
->vm_map
.token
);
3969 newvmspace
= vmspace_fork(vmcopy
);
3970 lwkt_gettoken(&newvmspace
->vm_map
.token
);
3972 newvmspace
= vmspace_alloc(map
->min_offset
, map
->max_offset
);
3973 lwkt_gettoken(&newvmspace
->vm_map
.token
);
3974 bcopy(&oldvmspace
->vm_startcopy
, &newvmspace
->vm_startcopy
,
3975 (caddr_t
)&oldvmspace
->vm_endcopy
-
3976 (caddr_t
)&oldvmspace
->vm_startcopy
);
3980 * Finish initializing the vmspace before assigning it
3981 * to the process. The vmspace will become the current vmspace
3984 pmap_pinit2(vmspace_pmap(newvmspace
));
3985 pmap_replacevm(p
, newvmspace
, 0);
3986 lwkt_reltoken(&newvmspace
->vm_map
.token
);
3987 lwkt_reltoken(&oldvmspace
->vm_map
.token
);
3988 vmspace_rel(oldvmspace
);
3992 * Unshare the specified VM space for forcing COW. This
3993 * is called by rfork, for the (RFMEM|RFPROC) == 0 case.
3996 vmspace_unshare(struct proc
*p
)
3998 struct vmspace
*oldvmspace
= p
->p_vmspace
;
3999 struct vmspace
*newvmspace
;
4001 lwkt_gettoken(&oldvmspace
->vm_map
.token
);
4002 if (vmspace_getrefs(oldvmspace
) == 1) {
4003 lwkt_reltoken(&oldvmspace
->vm_map
.token
);
4006 newvmspace
= vmspace_fork(oldvmspace
);
4007 lwkt_gettoken(&newvmspace
->vm_map
.token
);
4008 pmap_pinit2(vmspace_pmap(newvmspace
));
4009 pmap_replacevm(p
, newvmspace
, 0);
4010 lwkt_reltoken(&newvmspace
->vm_map
.token
);
4011 lwkt_reltoken(&oldvmspace
->vm_map
.token
);
4012 vmspace_rel(oldvmspace
);
4016 * vm_map_hint: return the beginning of the best area suitable for
4017 * creating a new mapping with "prot" protection.
4022 vm_map_hint(struct proc
*p
, vm_offset_t addr
, vm_prot_t prot
)
4024 struct vmspace
*vms
= p
->p_vmspace
;
4026 if (!randomize_mmap
|| addr
!= 0) {
4028 * Set a reasonable start point for the hint if it was
4029 * not specified or if it falls within the heap space.
4030 * Hinted mmap()s do not allocate out of the heap space.
4033 (addr
>= round_page((vm_offset_t
)vms
->vm_taddr
) &&
4034 addr
< round_page((vm_offset_t
)vms
->vm_daddr
+ maxdsiz
))) {
4035 addr
= round_page((vm_offset_t
)vms
->vm_daddr
+ maxdsiz
);
4040 addr
= (vm_offset_t
)vms
->vm_daddr
+ MAXDSIZ
;
4041 addr
+= karc4random() & (MIN((256 * 1024 * 1024), MAXDSIZ
) - 1);
4043 return (round_page(addr
));
4047 * Finds the VM object, offset, and protection for a given virtual address
4048 * in the specified map, assuming a page fault of the type specified.
4050 * Leaves the map in question locked for read; return values are guaranteed
4051 * until a vm_map_lookup_done call is performed. Note that the map argument
4052 * is in/out; the returned map must be used in the call to vm_map_lookup_done.
4054 * A handle (out_entry) is returned for use in vm_map_lookup_done, to make
4057 * If a lookup is requested with "write protection" specified, the map may
4058 * be changed to perform virtual copying operations, although the data
4059 * referenced will remain the same.
4064 vm_map_lookup(vm_map_t
*var_map
, /* IN/OUT */
4066 vm_prot_t fault_typea
,
4067 vm_map_entry_t
*out_entry
, /* OUT */
4068 vm_object_t
*object
, /* OUT */
4069 vm_pindex_t
*pindex
, /* OUT */
4070 vm_prot_t
*out_prot
, /* OUT */
4071 boolean_t
*wired
) /* OUT */
4073 vm_map_entry_t entry
;
4074 vm_map_t map
= *var_map
;
4076 vm_prot_t fault_type
= fault_typea
;
4077 int use_read_lock
= 1;
4078 int rv
= KERN_SUCCESS
;
4082 vm_map_lock_read(map
);
4087 * Always do a full lookup. The hint doesn't get us much anymore
4088 * now that the map is RB'd.
4091 *out_entry
= &map
->header
;
4095 vm_map_entry_t tmp_entry
;
4097 if (!vm_map_lookup_entry(map
, vaddr
, &tmp_entry
)) {
4098 rv
= KERN_INVALID_ADDRESS
;
4108 if (entry
->maptype
== VM_MAPTYPE_SUBMAP
) {
4109 vm_map_t old_map
= map
;
4111 *var_map
= map
= entry
->object
.sub_map
;
4113 vm_map_unlock_read(old_map
);
4115 vm_map_unlock(old_map
);
4121 * Check whether this task is allowed to have this page.
4122 * Note the special case for MAP_ENTRY_COW pages with an override.
4123 * This is to implement a forced COW for debuggers.
4125 if (fault_type
& VM_PROT_OVERRIDE_WRITE
)
4126 prot
= entry
->max_protection
;
4128 prot
= entry
->protection
;
4130 fault_type
&= (VM_PROT_READ
|VM_PROT_WRITE
|VM_PROT_EXECUTE
);
4131 if ((fault_type
& prot
) != fault_type
) {
4132 rv
= KERN_PROTECTION_FAILURE
;
4136 if ((entry
->eflags
& MAP_ENTRY_USER_WIRED
) &&
4137 (entry
->eflags
& MAP_ENTRY_COW
) &&
4138 (fault_type
& VM_PROT_WRITE
) &&
4139 (fault_typea
& VM_PROT_OVERRIDE_WRITE
) == 0) {
4140 rv
= KERN_PROTECTION_FAILURE
;
4145 * If this page is not pageable, we have to get it for all possible
4148 *wired
= (entry
->wired_count
!= 0);
4150 prot
= fault_type
= entry
->protection
;
4153 * Virtual page tables may need to update the accessed (A) bit
4154 * in a page table entry. Upgrade the fault to a write fault for
4155 * that case if the map will support it. If the map does not support
4156 * it the page table entry simply will not be updated.
4158 if (entry
->maptype
== VM_MAPTYPE_VPAGETABLE
) {
4159 if (prot
& VM_PROT_WRITE
)
4160 fault_type
|= VM_PROT_WRITE
;
4163 if (curthread
->td_lwp
&& curthread
->td_lwp
->lwp_vmspace
&&
4164 pmap_emulate_ad_bits(&curthread
->td_lwp
->lwp_vmspace
->vm_pmap
)) {
4165 if ((prot
& VM_PROT_WRITE
) == 0)
4166 fault_type
|= VM_PROT_WRITE
;
4170 * Only NORMAL and VPAGETABLE maps are object-based. UKSMAPs are not.
4172 if (entry
->maptype
!= VM_MAPTYPE_NORMAL
&&
4173 entry
->maptype
!= VM_MAPTYPE_VPAGETABLE
) {
4179 * If the entry was copy-on-write, we either ...
4181 if (entry
->eflags
& MAP_ENTRY_NEEDS_COPY
) {
4183 * If we want to write the page, we may as well handle that
4184 * now since we've got the map locked.
4186 * If we don't need to write the page, we just demote the
4187 * permissions allowed.
4190 if (fault_type
& VM_PROT_WRITE
) {
4192 * Not allowed if TDF_NOFAULT is set as the shadowing
4193 * operation can deadlock against the faulting
4194 * function due to the copy-on-write.
4196 if (curthread
->td_flags
& TDF_NOFAULT
) {
4197 rv
= KERN_FAILURE_NOFAULT
;
4202 * Make a new object, and place it in the object
4203 * chain. Note that no new references have appeared
4204 * -- one just moved from the map to the new
4208 if (use_read_lock
&& vm_map_lock_upgrade(map
)) {
4215 vm_map_entry_shadow(entry
, 0);
4218 * We're attempting to read a copy-on-write page --
4219 * don't allow writes.
4222 prot
&= ~VM_PROT_WRITE
;
4227 * Create an object if necessary.
4229 if (entry
->object
.vm_object
== NULL
&& !map
->system_map
) {
4230 if (use_read_lock
&& vm_map_lock_upgrade(map
)) {
4236 vm_map_entry_allocate_object(entry
);
4240 * Return the object/offset from this entry. If the entry was
4241 * copy-on-write or empty, it has been fixed up.
4243 *object
= entry
->object
.vm_object
;
4246 *pindex
= OFF_TO_IDX((vaddr
- entry
->start
) + entry
->offset
);
4249 * Return whether this is the only map sharing this data. On
4250 * success we return with a read lock held on the map. On failure
4251 * we return with the map unlocked.
4255 if (rv
== KERN_SUCCESS
) {
4256 if (use_read_lock
== 0)
4257 vm_map_lock_downgrade(map
);
4258 } else if (use_read_lock
) {
4259 vm_map_unlock_read(map
);
4267 * Releases locks acquired by a vm_map_lookup()
4268 * (according to the handle returned by that lookup).
4270 * No other requirements.
4273 vm_map_lookup_done(vm_map_t map
, vm_map_entry_t entry
, int count
)
4276 * Unlock the main-level map
4278 vm_map_unlock_read(map
);
4280 vm_map_entry_release(count
);
4284 * Quick hack, needs some help to make it more SMP friendly.
4287 vm_map_interlock(vm_map_t map
, struct vm_map_ilock
*ilock
,
4288 vm_offset_t ran_beg
, vm_offset_t ran_end
)
4290 struct vm_map_ilock
*scan
;
4292 ilock
->ran_beg
= ran_beg
;
4293 ilock
->ran_end
= ran_end
;
4296 spin_lock(&map
->ilock_spin
);
4298 for (scan
= map
->ilock_base
; scan
; scan
= scan
->next
) {
4299 if (ran_end
> scan
->ran_beg
&& ran_beg
< scan
->ran_end
) {
4300 scan
->flags
|= ILOCK_WAITING
;
4301 ssleep(scan
, &map
->ilock_spin
, 0, "ilock", 0);
4305 ilock
->next
= map
->ilock_base
;
4306 map
->ilock_base
= ilock
;
4307 spin_unlock(&map
->ilock_spin
);
4311 vm_map_deinterlock(vm_map_t map
, struct vm_map_ilock
*ilock
)
4313 struct vm_map_ilock
*scan
;
4314 struct vm_map_ilock
**scanp
;
4316 spin_lock(&map
->ilock_spin
);
4317 scanp
= &map
->ilock_base
;
4318 while ((scan
= *scanp
) != NULL
) {
4319 if (scan
== ilock
) {
4320 *scanp
= ilock
->next
;
4321 spin_unlock(&map
->ilock_spin
);
4322 if (ilock
->flags
& ILOCK_WAITING
)
4326 scanp
= &scan
->next
;
4328 spin_unlock(&map
->ilock_spin
);
4329 panic("vm_map_deinterlock: missing ilock!");
4332 #include "opt_ddb.h"
4334 #include <ddb/ddb.h>
4339 DB_SHOW_COMMAND(map
, vm_map_print
)
4342 /* XXX convert args. */
4343 vm_map_t map
= (vm_map_t
)addr
;
4344 boolean_t full
= have_addr
;
4346 vm_map_entry_t entry
;
4348 db_iprintf("Task map %p: pmap=%p, nentries=%d, version=%u\n",
4350 (void *)map
->pmap
, map
->nentries
, map
->timestamp
);
4353 if (!full
&& db_indent
)
4357 for (entry
= map
->header
.next
; entry
!= &map
->header
;
4358 entry
= entry
->next
) {
4359 db_iprintf("map entry %p: start=%p, end=%p\n",
4360 (void *)entry
, (void *)entry
->start
, (void *)entry
->end
);
4363 static char *inheritance_name
[4] =
4364 {"share", "copy", "none", "donate_copy"};
4366 db_iprintf(" prot=%x/%x/%s",
4368 entry
->max_protection
,
4369 inheritance_name
[(int)(unsigned char)
4370 entry
->inheritance
]);
4371 if (entry
->wired_count
!= 0)
4372 db_printf(", wired");
4374 switch(entry
->maptype
) {
4375 case VM_MAPTYPE_SUBMAP
:
4376 /* XXX no %qd in kernel. Truncate entry->offset. */
4377 db_printf(", share=%p, offset=0x%lx\n",
4378 (void *)entry
->object
.sub_map
,
4379 (long)entry
->offset
);
4381 if ((entry
->prev
== &map
->header
) ||
4382 (entry
->prev
->object
.sub_map
!=
4383 entry
->object
.sub_map
)) {
4385 vm_map_print((db_expr_t
)(intptr_t)
4386 entry
->object
.sub_map
,
4391 case VM_MAPTYPE_NORMAL
:
4392 case VM_MAPTYPE_VPAGETABLE
:
4393 /* XXX no %qd in kernel. Truncate entry->offset. */
4394 db_printf(", object=%p, offset=0x%lx",
4395 (void *)entry
->object
.vm_object
,
4396 (long)entry
->offset
);
4397 if (entry
->eflags
& MAP_ENTRY_COW
)
4398 db_printf(", copy (%s)",
4399 (entry
->eflags
& MAP_ENTRY_NEEDS_COPY
) ? "needed" : "done");
4403 if ((entry
->prev
== &map
->header
) ||
4404 (entry
->prev
->object
.vm_object
!=
4405 entry
->object
.vm_object
)) {
4407 vm_object_print((db_expr_t
)(intptr_t)
4408 entry
->object
.vm_object
,
4414 case VM_MAPTYPE_UKSMAP
:
4415 db_printf(", uksmap=%p, offset=0x%lx",
4416 (void *)entry
->object
.uksmap
,
4417 (long)entry
->offset
);
4418 if (entry
->eflags
& MAP_ENTRY_COW
)
4419 db_printf(", copy (%s)",
4420 (entry
->eflags
& MAP_ENTRY_NEEDS_COPY
) ? "needed" : "done");
4436 DB_SHOW_COMMAND(procvm
, procvm
)
4441 p
= (struct proc
*) addr
;
4446 db_printf("p = %p, vmspace = %p, map = %p, pmap = %p\n",
4447 (void *)p
, (void *)p
->p_vmspace
, (void *)&p
->p_vmspace
->vm_map
,
4448 (void *)vmspace_pmap(p
->p_vmspace
));
4450 vm_map_print((db_expr_t
)(intptr_t)&p
->p_vmspace
->vm_map
, 1, 0, NULL
);