4 * Copyright (c) 1991, 1993
5 * The Regents of the University of California. All rights reserved.
7 * This code is derived from software contributed to Berkeley by
8 * The Mach Operating System project at Carnegie-Mellon University.
10 * Redistribution and use in source and binary forms, with or without
11 * modification, are permitted provided that the following conditions
13 * 1. Redistributions of source code must retain the above copyright
14 * notice, this list of conditions and the following disclaimer.
15 * 2. Redistributions in binary form must reproduce the above copyright
16 * notice, this list of conditions and the following disclaimer in the
17 * documentation and/or other materials provided with the distribution.
18 * 3. Neither the name of the University nor the names of its contributors
19 * may be used to endorse or promote products derived from this software
20 * without specific prior written permission.
22 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
23 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
24 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
25 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
26 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
27 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
28 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
29 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
30 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
31 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
34 * from: @(#)vm_map.c 8.3 (Berkeley) 1/12/94
37 * Copyright (c) 1987, 1990 Carnegie-Mellon University.
38 * All rights reserved.
40 * Authors: Avadis Tevanian, Jr., Michael Wayne Young
42 * Permission to use, copy, modify and distribute this software and
43 * its documentation is hereby granted, provided that both the copyright
44 * notice and this permission notice appear in all copies of the
45 * software, derivative works or modified versions, and any portions
46 * thereof, and that both notices appear in supporting documentation.
48 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
49 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
50 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
52 * Carnegie Mellon requests users of this software to return to
54 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU
55 * School of Computer Science
56 * Carnegie Mellon University
57 * Pittsburgh PA 15213-3890
59 * any improvements or extensions that they make and grant Carnegie the
60 * rights to redistribute these changes.
62 * $FreeBSD: src/sys/vm/vm_map.c,v 1.187.2.19 2003/05/27 00:47:02 alc Exp $
66 * Virtual memory mapping module.
69 #include <sys/param.h>
70 #include <sys/systm.h>
71 #include <sys/kernel.h>
73 #include <sys/serialize.h>
75 #include <sys/vmmeter.h>
77 #include <sys/vnode.h>
78 #include <sys/resourcevar.h>
81 #include <sys/malloc.h>
82 #include <sys/objcache.h>
85 #include <vm/vm_param.h>
87 #include <vm/vm_map.h>
88 #include <vm/vm_page.h>
89 #include <vm/vm_object.h>
90 #include <vm/vm_pager.h>
91 #include <vm/vm_kern.h>
92 #include <vm/vm_extern.h>
93 #include <vm/swap_pager.h>
94 #include <vm/vm_zone.h>
96 #include <sys/thread2.h>
97 #include <sys/random.h>
98 #include <sys/sysctl.h>
101 * Virtual memory maps provide for the mapping, protection, and sharing
102 * of virtual memory objects. In addition, this module provides for an
103 * efficient virtual copy of memory from one map to another.
105 * Synchronization is required prior to most operations.
107 * Maps consist of an ordered doubly-linked list of simple entries.
108 * A hint and a RB tree is used to speed-up lookups.
110 * Callers looking to modify maps specify start/end addresses which cause
111 * the related map entry to be clipped if necessary, and then later
112 * recombined if the pieces remained compatible.
114 * Virtual copy operations are performed by copying VM object references
115 * from one map to another, and then marking both regions as copy-on-write.
117 static boolean_t
vmspace_ctor(void *obj
, void *privdata
, int ocflags
);
118 static void vmspace_dtor(void *obj
, void *privdata
);
119 static void vmspace_terminate(struct vmspace
*vm
, int final
);
121 MALLOC_DEFINE(M_VMSPACE
, "vmspace", "vmspace objcache backingstore");
122 static struct objcache
*vmspace_cache
;
125 * per-cpu page table cross mappings are initialized in early boot
126 * and might require a considerable number of vm_map_entry structures.
128 #define MAPENTRYBSP_CACHE (MAXCPU+1)
129 #define MAPENTRYAP_CACHE 8
131 static struct vm_zone mapentzone_store
, mapzone_store
;
132 static vm_zone_t mapentzone
, mapzone
;
133 static struct vm_object mapentobj
, mapobj
;
135 static struct vm_map_entry map_entry_init
[MAX_MAPENT
];
136 static struct vm_map_entry cpu_map_entry_init_bsp
[MAPENTRYBSP_CACHE
];
137 static struct vm_map_entry cpu_map_entry_init_ap
[MAXCPU
][MAPENTRYAP_CACHE
];
138 static struct vm_map map_init
[MAX_KMAP
];
140 static int randomize_mmap
;
141 SYSCTL_INT(_vm
, OID_AUTO
, randomize_mmap
, CTLFLAG_RW
, &randomize_mmap
, 0,
142 "Randomize mmap offsets");
143 static int vm_map_relock_enable
= 1;
144 SYSCTL_INT(_vm
, OID_AUTO
, map_relock_enable
, CTLFLAG_RW
,
145 &vm_map_relock_enable
, 0, "Randomize mmap offsets");
147 static void vm_map_entry_shadow(vm_map_entry_t entry
, int addref
);
148 static vm_map_entry_t
vm_map_entry_create(vm_map_t map
, int *);
149 static void vm_map_entry_dispose (vm_map_t map
, vm_map_entry_t entry
, int *);
150 static void _vm_map_clip_end (vm_map_t
, vm_map_entry_t
, vm_offset_t
, int *);
151 static void _vm_map_clip_start (vm_map_t
, vm_map_entry_t
, vm_offset_t
, int *);
152 static void vm_map_entry_delete (vm_map_t
, vm_map_entry_t
, int *);
153 static void vm_map_entry_unwire (vm_map_t
, vm_map_entry_t
);
154 static void vm_map_copy_entry (vm_map_t
, vm_map_t
, vm_map_entry_t
,
156 static void vm_map_unclip_range (vm_map_t map
, vm_map_entry_t start_entry
, vm_offset_t start
, vm_offset_t end
, int *count
, int flags
);
159 * Initialize the vm_map module. Must be called before any other vm_map
162 * Map and entry structures are allocated from the general purpose
163 * memory pool with some exceptions:
165 * - The kernel map is allocated statically.
166 * - Initial kernel map entries are allocated out of a static pool.
167 * - We must set ZONE_SPECIAL here or the early boot code can get
168 * stuck if there are >63 cores.
170 * These restrictions are necessary since malloc() uses the
171 * maps and requires map entries.
173 * Called from the low level boot code only.
178 mapzone
= &mapzone_store
;
179 zbootinit(mapzone
, "MAP", sizeof (struct vm_map
),
181 mapentzone
= &mapentzone_store
;
182 zbootinit(mapentzone
, "MAP ENTRY", sizeof (struct vm_map_entry
),
183 map_entry_init
, MAX_MAPENT
);
184 mapentzone_store
.zflags
|= ZONE_SPECIAL
;
188 * Called prior to any vmspace allocations.
190 * Called from the low level boot code only.
195 vmspace_cache
= objcache_create_mbacked(M_VMSPACE
,
196 sizeof(struct vmspace
),
198 vmspace_ctor
, vmspace_dtor
,
200 zinitna(mapentzone
, &mapentobj
, NULL
, 0, 0,
201 ZONE_USE_RESERVE
| ZONE_SPECIAL
);
202 zinitna(mapzone
, &mapobj
, NULL
, 0, 0, 0);
208 * objcache support. We leave the pmap root cached as long as possible
209 * for performance reasons.
213 vmspace_ctor(void *obj
, void *privdata
, int ocflags
)
215 struct vmspace
*vm
= obj
;
217 bzero(vm
, sizeof(*vm
));
218 vm
->vm_refcnt
= (u_int
)-1;
225 vmspace_dtor(void *obj
, void *privdata
)
227 struct vmspace
*vm
= obj
;
229 KKASSERT(vm
->vm_refcnt
== (u_int
)-1);
230 pmap_puninit(vmspace_pmap(vm
));
234 * Red black tree functions
236 * The caller must hold the related map lock.
238 static int rb_vm_map_compare(vm_map_entry_t a
, vm_map_entry_t b
);
239 RB_GENERATE(vm_map_rb_tree
, vm_map_entry
, rb_entry
, rb_vm_map_compare
);
241 /* a->start is address, and the only field has to be initialized */
243 rb_vm_map_compare(vm_map_entry_t a
, vm_map_entry_t b
)
245 if (a
->start
< b
->start
)
247 else if (a
->start
> b
->start
)
253 * Initialize vmspace ref/hold counts vmspace0. There is a holdcnt for
257 vmspace_initrefs(struct vmspace
*vm
)
264 * Allocate a vmspace structure, including a vm_map and pmap.
265 * Initialize numerous fields. While the initial allocation is zerod,
266 * subsequence reuse from the objcache leaves elements of the structure
267 * intact (particularly the pmap), so portions must be zerod.
269 * Returns a referenced vmspace.
274 vmspace_alloc(vm_offset_t min
, vm_offset_t max
)
278 vm
= objcache_get(vmspace_cache
, M_WAITOK
);
280 bzero(&vm
->vm_startcopy
,
281 (char *)&vm
->vm_endcopy
- (char *)&vm
->vm_startcopy
);
282 vm_map_init(&vm
->vm_map
, min
, max
, NULL
); /* initializes token */
285 * NOTE: hold to acquires token for safety.
287 * On return vmspace is referenced (refs=1, hold=1). That is,
288 * each refcnt also has a holdcnt. There can be additional holds
289 * (holdcnt) above and beyond the refcnt. Finalization is handled in
290 * two stages, one on refs 1->0, and the the second on hold 1->0.
292 KKASSERT(vm
->vm_holdcnt
== 0);
293 KKASSERT(vm
->vm_refcnt
== (u_int
)-1);
294 vmspace_initrefs(vm
);
296 pmap_pinit(vmspace_pmap(vm
)); /* (some fields reused) */
297 vm
->vm_map
.pmap
= vmspace_pmap(vm
); /* XXX */
300 cpu_vmspace_alloc(vm
);
307 * NOTE: Can return -1 if the vmspace is exiting.
310 vmspace_getrefs(struct vmspace
*vm
)
312 return ((int)vm
->vm_refcnt
);
316 * A vmspace object must already have a non-zero hold to be able to gain
317 * further holds on it.
320 vmspace_hold_notoken(struct vmspace
*vm
)
322 KKASSERT(vm
->vm_holdcnt
!= 0);
323 refcount_acquire(&vm
->vm_holdcnt
);
327 vmspace_drop_notoken(struct vmspace
*vm
)
329 if (refcount_release(&vm
->vm_holdcnt
)) {
330 if (vm
->vm_refcnt
== (u_int
)-1) {
331 vmspace_terminate(vm
, 1);
337 vmspace_hold(struct vmspace
*vm
)
339 vmspace_hold_notoken(vm
);
340 lwkt_gettoken(&vm
->vm_map
.token
);
344 vmspace_drop(struct vmspace
*vm
)
346 lwkt_reltoken(&vm
->vm_map
.token
);
347 vmspace_drop_notoken(vm
);
351 * A vmspace object must not be in a terminated state to be able to obtain
352 * additional refs on it.
354 * Ref'ing a vmspace object also increments its hold count.
357 vmspace_ref(struct vmspace
*vm
)
359 KKASSERT((int)vm
->vm_refcnt
>= 0);
360 vmspace_hold_notoken(vm
);
361 refcount_acquire(&vm
->vm_refcnt
);
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
)
372 if (refcount_release(&vm
->vm_refcnt
)) {
373 vm
->vm_refcnt
= (u_int
)-1; /* no other refs possible */
374 vmspace_terminate(vm
, 0);
376 vmspace_drop_notoken(vm
);
380 * This is called during exit indicating that the vmspace is no
381 * longer in used by an exiting process, but the process has not yet
384 * We release the refcnt but not the associated holdcnt.
389 vmspace_relexit(struct vmspace
*vm
)
391 if (refcount_release(&vm
->vm_refcnt
)) {
392 vm
->vm_refcnt
= (u_int
)-1; /* no other refs possible */
393 vmspace_terminate(vm
, 0);
398 * Called during reap to disconnect the remainder of the vmspace from
399 * the process. On the hold drop the vmspace termination is finalized.
404 vmspace_exitfree(struct proc
*p
)
410 vmspace_drop_notoken(vm
);
414 * Called in two cases:
416 * (1) When the last refcnt is dropped and the vmspace becomes inactive,
417 * called with final == 0. refcnt will be (u_int)-1 at this point,
418 * and holdcnt will still be non-zero.
420 * (2) When holdcnt becomes 0, called with final == 1. There should no
421 * longer be anyone with access to the vmspace.
423 * VMSPACE_EXIT1 flags the primary deactivation
424 * VMSPACE_EXIT2 flags the last reap
427 vmspace_terminate(struct vmspace
*vm
, int final
)
431 lwkt_gettoken(&vm
->vm_map
.token
);
433 KKASSERT((vm
->vm_flags
& VMSPACE_EXIT1
) == 0);
436 * Get rid of most of the resources. Leave the kernel pmap
439 * If the pmap does not contain wired pages we can bulk-delete
440 * the pmap as a performance optimization before removing the related mappings.
442 * If the pmap contains wired pages we cannot do this pre-optimization
443 * because currently vm_fault_unwire() expects the pmap pages to exist
444 * and will not decrement p->wire_count if they do not.
446 vm
->vm_flags
|= VMSPACE_EXIT1
;
448 if (vmspace_pmap(vm
)->pm_stats
.wired_count
) {
449 vm_map_remove(&vm
->vm_map
, VM_MIN_USER_ADDRESS
,
450 VM_MAX_USER_ADDRESS
);
451 pmap_remove_pages(vmspace_pmap(vm
), VM_MIN_USER_ADDRESS
,
452 VM_MAX_USER_ADDRESS
);
454 pmap_remove_pages(vmspace_pmap(vm
), VM_MIN_USER_ADDRESS
,
455 VM_MAX_USER_ADDRESS
);
456 vm_map_remove(&vm
->vm_map
, VM_MIN_USER_ADDRESS
,
457 VM_MAX_USER_ADDRESS
);
459 lwkt_reltoken(&vm
->vm_map
.token
);
461 KKASSERT((vm
->vm_flags
& VMSPACE_EXIT1
) != 0);
462 KKASSERT((vm
->vm_flags
& VMSPACE_EXIT2
) == 0);
465 * Get rid of remaining basic resources.
467 vm
->vm_flags
|= VMSPACE_EXIT2
;
468 cpu_vmspace_free(vm
);
472 * Lock the map, to wait out all other references to it.
473 * Delete all of the mappings and pages they hold, then call
474 * the pmap module to reclaim anything left.
476 count
= vm_map_entry_reserve(MAP_RESERVE_COUNT
);
477 vm_map_lock(&vm
->vm_map
);
478 vm_map_delete(&vm
->vm_map
, vm
->vm_map
.min_offset
,
479 vm
->vm_map
.max_offset
, &count
);
480 vm_map_unlock(&vm
->vm_map
);
481 vm_map_entry_release(count
);
483 lwkt_gettoken(&vmspace_pmap(vm
)->pm_token
);
484 pmap_release(vmspace_pmap(vm
));
485 lwkt_reltoken(&vmspace_pmap(vm
)->pm_token
);
486 lwkt_reltoken(&vm
->vm_map
.token
);
487 objcache_put(vmspace_cache
, vm
);
492 * Swap useage is determined by taking the proportional swap used by
493 * VM objects backing the VM map. To make up for fractional losses,
494 * if the VM object has any swap use at all the associated map entries
495 * count for at least 1 swap page.
500 vmspace_swap_count(struct vmspace
*vm
)
502 vm_map_t map
= &vm
->vm_map
;
509 for (cur
= map
->header
.next
; cur
!= &map
->header
; cur
= cur
->next
) {
510 switch(cur
->maptype
) {
511 case VM_MAPTYPE_NORMAL
:
512 case VM_MAPTYPE_VPAGETABLE
:
513 if ((object
= cur
->object
.vm_object
) == NULL
)
515 if (object
->swblock_count
) {
516 n
= (cur
->end
- cur
->start
) / PAGE_SIZE
;
517 count
+= object
->swblock_count
*
518 SWAP_META_PAGES
* n
/ object
->size
+ 1;
531 * Calculate the approximate number of anonymous pages in use by
532 * this vmspace. To make up for fractional losses, we count each
533 * VM object as having at least 1 anonymous page.
538 vmspace_anonymous_count(struct vmspace
*vm
)
540 vm_map_t map
= &vm
->vm_map
;
546 for (cur
= map
->header
.next
; cur
!= &map
->header
; cur
= cur
->next
) {
547 switch(cur
->maptype
) {
548 case VM_MAPTYPE_NORMAL
:
549 case VM_MAPTYPE_VPAGETABLE
:
550 if ((object
= cur
->object
.vm_object
) == NULL
)
552 if (object
->type
!= OBJT_DEFAULT
&&
553 object
->type
!= OBJT_SWAP
) {
556 count
+= object
->resident_page_count
;
568 * Creates and returns a new empty VM map with the given physical map
569 * structure, and having the given lower and upper address bounds.
574 vm_map_create(vm_map_t result
, pmap_t pmap
, vm_offset_t min
, vm_offset_t max
)
577 result
= zalloc(mapzone
);
578 vm_map_init(result
, min
, max
, pmap
);
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
);
596 map
->min_offset
= min
;
597 map
->max_offset
= max
;
599 map
->first_free
= &map
->header
;
600 map
->hint
= &map
->header
;
603 lwkt_token_init(&map
->token
, "vm_map");
604 lockinit(&map
->lock
, "vm_maplk", (hz
+ 9) / 10, 0);
608 * Shadow the vm_map_entry's object. This typically needs to be done when
609 * a write fault is taken on an entry which had previously been cloned by
610 * fork(). The shared object (which might be NULL) must become private so
611 * we add a shadow layer above it.
613 * Object allocation for anonymous mappings is defered as long as possible.
614 * When creating a shadow, however, the underlying object must be instantiated
615 * so it can be shared.
617 * If the map segment is governed by a virtual page table then it is
618 * possible to address offsets beyond the mapped area. Just allocate
619 * a maximally sized object for this case.
621 * If addref is non-zero an additional reference is added to the returned
622 * entry. This mechanic exists because the additional reference might have
623 * to be added atomically and not after return to prevent a premature
626 * The vm_map must be exclusively locked.
627 * No other requirements.
631 vm_map_entry_shadow(vm_map_entry_t entry
, int addref
)
633 if (entry
->maptype
== VM_MAPTYPE_VPAGETABLE
) {
634 vm_object_shadow(&entry
->object
.vm_object
, &entry
->offset
,
635 0x7FFFFFFF, addref
); /* XXX */
637 vm_object_shadow(&entry
->object
.vm_object
, &entry
->offset
,
638 atop(entry
->end
- entry
->start
), addref
);
640 entry
->eflags
&= ~MAP_ENTRY_NEEDS_COPY
;
644 * Allocate an object for a vm_map_entry.
646 * Object allocation for anonymous mappings is defered as long as possible.
647 * This function is called when we can defer no longer, generally when a map
648 * entry might be split or forked or takes a page fault.
650 * If the map segment is governed by a virtual page table then it is
651 * possible to address offsets beyond the mapped area. Just allocate
652 * a maximally sized object for this case.
654 * The vm_map must be exclusively locked.
655 * No other requirements.
658 vm_map_entry_allocate_object(vm_map_entry_t entry
)
662 if (entry
->maptype
== VM_MAPTYPE_VPAGETABLE
) {
663 obj
= vm_object_allocate(OBJT_DEFAULT
, 0x7FFFFFFF); /* XXX */
665 obj
= vm_object_allocate(OBJT_DEFAULT
,
666 atop(entry
->end
- entry
->start
));
668 entry
->object
.vm_object
= obj
;
673 * Set an initial negative count so the first attempt to reserve
674 * space preloads a bunch of vm_map_entry's for this cpu. Also
675 * pre-allocate 2 vm_map_entries which will be needed by zalloc() to
676 * map a new page for vm_map_entry structures. SMP systems are
677 * particularly sensitive.
679 * This routine is called in early boot so we cannot just call
680 * vm_map_entry_reserve().
682 * Called from the low level boot code only (for each cpu)
684 * WARNING! Take care not to have too-big a static/BSS structure here
685 * as MAXCPU can be 256+, otherwise the loader's 64MB heap
686 * can get blown out by the kernel plus the initrd image.
689 vm_map_entry_reserve_cpu_init(globaldata_t gd
)
691 vm_map_entry_t entry
;
695 gd
->gd_vme_avail
-= MAP_RESERVE_COUNT
* 2;
696 if (gd
->gd_cpuid
== 0) {
697 entry
= &cpu_map_entry_init_bsp
[0];
698 count
= MAPENTRYBSP_CACHE
;
700 entry
= &cpu_map_entry_init_ap
[gd
->gd_cpuid
][0];
701 count
= MAPENTRYAP_CACHE
;
703 for (i
= 0; i
< count
; ++i
, ++entry
) {
704 entry
->next
= gd
->gd_vme_base
;
705 gd
->gd_vme_base
= entry
;
710 * Reserves vm_map_entry structures so code later on can manipulate
711 * map_entry structures within a locked map without blocking trying
712 * to allocate a new vm_map_entry.
717 vm_map_entry_reserve(int count
)
719 struct globaldata
*gd
= mycpu
;
720 vm_map_entry_t entry
;
723 * Make sure we have enough structures in gd_vme_base to handle
724 * the reservation request.
726 * The critical section protects access to the per-cpu gd.
729 while (gd
->gd_vme_avail
< count
) {
730 entry
= zalloc(mapentzone
);
731 entry
->next
= gd
->gd_vme_base
;
732 gd
->gd_vme_base
= entry
;
735 gd
->gd_vme_avail
-= count
;
742 * Releases previously reserved vm_map_entry structures that were not
743 * used. If we have too much junk in our per-cpu cache clean some of
749 vm_map_entry_release(int count
)
751 struct globaldata
*gd
= mycpu
;
752 vm_map_entry_t entry
;
755 gd
->gd_vme_avail
+= count
;
756 while (gd
->gd_vme_avail
> MAP_RESERVE_SLOP
) {
757 entry
= gd
->gd_vme_base
;
758 KKASSERT(entry
!= NULL
);
759 gd
->gd_vme_base
= entry
->next
;
762 zfree(mapentzone
, entry
);
769 * Reserve map entry structures for use in kernel_map itself. These
770 * entries have *ALREADY* been reserved on a per-cpu basis when the map
771 * was inited. This function is used by zalloc() to avoid a recursion
772 * when zalloc() itself needs to allocate additional kernel memory.
774 * This function works like the normal reserve but does not load the
775 * vm_map_entry cache (because that would result in an infinite
776 * recursion). Note that gd_vme_avail may go negative. This is expected.
778 * Any caller of this function must be sure to renormalize after
779 * potentially eating entries to ensure that the reserve supply
785 vm_map_entry_kreserve(int count
)
787 struct globaldata
*gd
= mycpu
;
790 gd
->gd_vme_avail
-= count
;
792 KASSERT(gd
->gd_vme_base
!= NULL
,
793 ("no reserved entries left, gd_vme_avail = %d",
799 * Release previously reserved map entries for kernel_map. We do not
800 * attempt to clean up like the normal release function as this would
801 * cause an unnecessary (but probably not fatal) deep procedure call.
806 vm_map_entry_krelease(int count
)
808 struct globaldata
*gd
= mycpu
;
811 gd
->gd_vme_avail
+= count
;
816 * Allocates a VM map entry for insertion. No entry fields are filled in.
818 * The entries should have previously been reserved. The reservation count
819 * is tracked in (*countp).
823 static vm_map_entry_t
824 vm_map_entry_create(vm_map_t map
, int *countp
)
826 struct globaldata
*gd
= mycpu
;
827 vm_map_entry_t entry
;
829 KKASSERT(*countp
> 0);
832 entry
= gd
->gd_vme_base
;
833 KASSERT(entry
!= NULL
, ("gd_vme_base NULL! count %d", *countp
));
834 gd
->gd_vme_base
= entry
->next
;
841 * Dispose of a vm_map_entry that is no longer being referenced.
846 vm_map_entry_dispose(vm_map_t map
, vm_map_entry_t entry
, int *countp
)
848 struct globaldata
*gd
= mycpu
;
850 KKASSERT(map
->hint
!= entry
);
851 KKASSERT(map
->first_free
!= entry
);
855 entry
->next
= gd
->gd_vme_base
;
856 gd
->gd_vme_base
= entry
;
862 * Insert/remove entries from maps.
864 * The related map must be exclusively locked.
865 * The caller must hold map->token
866 * No other requirements.
869 vm_map_entry_link(vm_map_t map
,
870 vm_map_entry_t after_where
,
871 vm_map_entry_t entry
)
873 ASSERT_VM_MAP_LOCKED(map
);
876 entry
->prev
= after_where
;
877 entry
->next
= after_where
->next
;
878 entry
->next
->prev
= entry
;
879 after_where
->next
= entry
;
880 if (vm_map_rb_tree_RB_INSERT(&map
->rb_root
, entry
))
881 panic("vm_map_entry_link: dup addr map %p ent %p", map
, entry
);
885 vm_map_entry_unlink(vm_map_t map
,
886 vm_map_entry_t entry
)
891 ASSERT_VM_MAP_LOCKED(map
);
893 if (entry
->eflags
& MAP_ENTRY_IN_TRANSITION
) {
894 panic("vm_map_entry_unlink: attempt to mess with "
895 "locked entry! %p", entry
);
901 vm_map_rb_tree_RB_REMOVE(&map
->rb_root
, entry
);
906 * Finds the map entry containing (or immediately preceding) the specified
907 * address in the given map. The entry is returned in (*entry).
909 * The boolean result indicates whether the address is actually contained
912 * The related map must be locked.
913 * No other requirements.
916 vm_map_lookup_entry(vm_map_t map
, vm_offset_t address
, vm_map_entry_t
*entry
)
921 ASSERT_VM_MAP_LOCKED(map
);
924 * XXX TEMPORARILY DISABLED. For some reason our attempt to revive
925 * the hint code with the red-black lookup meets with system crashes
926 * and lockups. We do not yet know why.
928 * It is possible that the problem is related to the setting
929 * of the hint during map_entry deletion, in the code specified
930 * at the GGG comment later on in this file.
932 * YYY More likely it's because this function can be called with
933 * a shared lock on the map, resulting in map->hint updates possibly
934 * racing. Fixed now but untested.
937 * Quickly check the cached hint, there's a good chance of a match.
941 if (tmp
!= &map
->header
) {
942 if (address
>= tmp
->start
&& address
< tmp
->end
) {
950 * Locate the record from the top of the tree. 'last' tracks the
951 * closest prior record and is returned if no match is found, which
952 * in binary tree terms means tracking the most recent right-branch
953 * taken. If there is no prior record, &map->header is returned.
956 tmp
= RB_ROOT(&map
->rb_root
);
959 if (address
>= tmp
->start
) {
960 if (address
< tmp
->end
) {
966 tmp
= RB_RIGHT(tmp
, rb_entry
);
968 tmp
= RB_LEFT(tmp
, rb_entry
);
976 * Inserts the given whole VM object into the target map at the specified
977 * address range. The object's size should match that of the address range.
979 * The map must be exclusively locked.
980 * The object must be held.
981 * The caller must have reserved sufficient vm_map_entry structures.
983 * If object is non-NULL, ref count must be bumped by caller prior to
984 * making call to account for the new entry.
987 vm_map_insert(vm_map_t map
, int *countp
, void *map_object
, void *map_aux
,
988 vm_ooffset_t offset
, vm_offset_t start
, vm_offset_t end
,
989 vm_maptype_t maptype
, vm_subsys_t id
,
990 vm_prot_t prot
, vm_prot_t max
, int cow
)
992 vm_map_entry_t new_entry
;
993 vm_map_entry_t prev_entry
;
994 vm_map_entry_t temp_entry
;
995 vm_eflags_t protoeflags
;
999 if (maptype
== VM_MAPTYPE_UKSMAP
)
1002 object
= map_object
;
1004 ASSERT_VM_MAP_LOCKED(map
);
1006 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object
));
1009 * Check that the start and end points are not bogus.
1011 if ((start
< map
->min_offset
) || (end
> map
->max_offset
) ||
1013 return (KERN_INVALID_ADDRESS
);
1016 * Find the entry prior to the proposed starting address; if it's part
1017 * of an existing entry, this range is bogus.
1019 if (vm_map_lookup_entry(map
, start
, &temp_entry
))
1020 return (KERN_NO_SPACE
);
1022 prev_entry
= temp_entry
;
1025 * Assert that the next entry doesn't overlap the end point.
1028 if ((prev_entry
->next
!= &map
->header
) &&
1029 (prev_entry
->next
->start
< end
))
1030 return (KERN_NO_SPACE
);
1034 if (cow
& MAP_COPY_ON_WRITE
)
1035 protoeflags
|= MAP_ENTRY_COW
|MAP_ENTRY_NEEDS_COPY
;
1037 if (cow
& MAP_NOFAULT
) {
1038 protoeflags
|= MAP_ENTRY_NOFAULT
;
1040 KASSERT(object
== NULL
,
1041 ("vm_map_insert: paradoxical MAP_NOFAULT request"));
1043 if (cow
& MAP_DISABLE_SYNCER
)
1044 protoeflags
|= MAP_ENTRY_NOSYNC
;
1045 if (cow
& MAP_DISABLE_COREDUMP
)
1046 protoeflags
|= MAP_ENTRY_NOCOREDUMP
;
1047 if (cow
& MAP_IS_STACK
)
1048 protoeflags
|= MAP_ENTRY_STACK
;
1049 if (cow
& MAP_IS_KSTACK
)
1050 protoeflags
|= MAP_ENTRY_KSTACK
;
1052 lwkt_gettoken(&map
->token
);
1056 * When object is non-NULL, it could be shared with another
1057 * process. We have to set or clear OBJ_ONEMAPPING
1060 * NOTE: This flag is only applicable to DEFAULT and SWAP
1061 * objects and will already be clear in other types
1062 * of objects, so a shared object lock is ok for
1065 if ((object
->ref_count
> 1) || (object
->shadow_count
!= 0)) {
1066 vm_object_clear_flag(object
, OBJ_ONEMAPPING
);
1069 else if ((prev_entry
!= &map
->header
) &&
1070 (prev_entry
->eflags
== protoeflags
) &&
1071 (prev_entry
->end
== start
) &&
1072 (prev_entry
->wired_count
== 0) &&
1073 (prev_entry
->id
== id
) &&
1074 prev_entry
->maptype
== maptype
&&
1075 maptype
== VM_MAPTYPE_NORMAL
&&
1076 ((prev_entry
->object
.vm_object
== NULL
) ||
1077 vm_object_coalesce(prev_entry
->object
.vm_object
,
1078 OFF_TO_IDX(prev_entry
->offset
),
1079 (vm_size_t
)(prev_entry
->end
- prev_entry
->start
),
1080 (vm_size_t
)(end
- prev_entry
->end
)))) {
1082 * We were able to extend the object. Determine if we
1083 * can extend the previous map entry to include the
1084 * new range as well.
1086 if ((prev_entry
->inheritance
== VM_INHERIT_DEFAULT
) &&
1087 (prev_entry
->protection
== prot
) &&
1088 (prev_entry
->max_protection
== max
)) {
1089 map
->size
+= (end
- prev_entry
->end
);
1090 prev_entry
->end
= end
;
1091 vm_map_simplify_entry(map
, prev_entry
, countp
);
1092 lwkt_reltoken(&map
->token
);
1093 return (KERN_SUCCESS
);
1097 * If we can extend the object but cannot extend the
1098 * map entry, we have to create a new map entry. We
1099 * must bump the ref count on the extended object to
1100 * account for it. object may be NULL.
1102 * XXX if object is NULL should we set offset to 0 here ?
1104 object
= prev_entry
->object
.vm_object
;
1105 offset
= prev_entry
->offset
+
1106 (prev_entry
->end
- prev_entry
->start
);
1108 vm_object_hold(object
);
1109 vm_object_chain_wait(object
, 0);
1110 vm_object_reference_locked(object
);
1112 map_object
= object
;
1117 * NOTE: if conditionals fail, object can be NULL here. This occurs
1118 * in things like the buffer map where we manage kva but do not manage
1123 * Create a new entry
1126 new_entry
= vm_map_entry_create(map
, countp
);
1127 new_entry
->start
= start
;
1128 new_entry
->end
= end
;
1131 new_entry
->maptype
= maptype
;
1132 new_entry
->eflags
= protoeflags
;
1133 new_entry
->object
.map_object
= map_object
;
1134 new_entry
->aux
.master_pde
= 0; /* in case size is different */
1135 new_entry
->aux
.map_aux
= map_aux
;
1136 new_entry
->offset
= offset
;
1138 new_entry
->inheritance
= VM_INHERIT_DEFAULT
;
1139 new_entry
->protection
= prot
;
1140 new_entry
->max_protection
= max
;
1141 new_entry
->wired_count
= 0;
1144 * Insert the new entry into the list
1147 vm_map_entry_link(map
, prev_entry
, new_entry
);
1148 map
->size
+= new_entry
->end
- new_entry
->start
;
1151 * Update the free space hint. Entries cannot overlap.
1152 * An exact comparison is needed to avoid matching
1153 * against the map->header.
1155 if ((map
->first_free
== prev_entry
) &&
1156 (prev_entry
->end
== new_entry
->start
)) {
1157 map
->first_free
= new_entry
;
1162 * Temporarily removed to avoid MAP_STACK panic, due to
1163 * MAP_STACK being a huge hack. Will be added back in
1164 * when MAP_STACK (and the user stack mapping) is fixed.
1167 * It may be possible to simplify the entry
1169 vm_map_simplify_entry(map
, new_entry
, countp
);
1173 * Try to pre-populate the page table. Mappings governed by virtual
1174 * page tables cannot be prepopulated without a lot of work, so
1177 if ((cow
& (MAP_PREFAULT
|MAP_PREFAULT_PARTIAL
)) &&
1178 maptype
!= VM_MAPTYPE_VPAGETABLE
&&
1179 maptype
!= VM_MAPTYPE_UKSMAP
) {
1181 if (vm_map_relock_enable
&& (cow
& MAP_PREFAULT_RELOCK
)) {
1183 vm_object_lock_swap();
1184 vm_object_drop(object
);
1186 pmap_object_init_pt(map
->pmap
, start
, prot
,
1187 object
, OFF_TO_IDX(offset
), end
- start
,
1188 cow
& MAP_PREFAULT_PARTIAL
);
1190 vm_object_hold(object
);
1191 vm_object_lock_swap();
1195 vm_object_drop(object
);
1197 lwkt_reltoken(&map
->token
);
1198 return (KERN_SUCCESS
);
1202 * Find sufficient space for `length' bytes in the given map, starting at
1203 * `start'. Returns 0 on success, 1 on no space.
1205 * This function will returned an arbitrarily aligned pointer. If no
1206 * particular alignment is required you should pass align as 1. Note that
1207 * the map may return PAGE_SIZE aligned pointers if all the lengths used in
1208 * the map are a multiple of PAGE_SIZE, even if you pass a smaller align
1211 * 'align' should be a power of 2 but is not required to be.
1213 * The map must be exclusively locked.
1214 * No other requirements.
1217 vm_map_findspace(vm_map_t map
, vm_offset_t start
, vm_size_t length
,
1218 vm_size_t align
, int flags
, vm_offset_t
*addr
)
1220 vm_map_entry_t entry
, next
;
1222 vm_offset_t align_mask
;
1224 if (start
< map
->min_offset
)
1225 start
= map
->min_offset
;
1226 if (start
> map
->max_offset
)
1230 * If the alignment is not a power of 2 we will have to use
1231 * a mod/division, set align_mask to a special value.
1233 if ((align
| (align
- 1)) + 1 != (align
<< 1))
1234 align_mask
= (vm_offset_t
)-1;
1236 align_mask
= align
- 1;
1239 * Look for the first possible address; if there's already something
1240 * at this address, we have to start after it.
1242 if (start
== map
->min_offset
) {
1243 if ((entry
= map
->first_free
) != &map
->header
)
1248 if (vm_map_lookup_entry(map
, start
, &tmp
))
1254 * Look through the rest of the map, trying to fit a new region in the
1255 * gap between existing regions, or after the very last region.
1257 for (;; start
= (entry
= next
)->end
) {
1259 * Adjust the proposed start by the requested alignment,
1260 * be sure that we didn't wrap the address.
1262 if (align_mask
== (vm_offset_t
)-1)
1263 end
= roundup(start
, align
);
1265 end
= (start
+ align_mask
) & ~align_mask
;
1270 * Find the end of the proposed new region. Be sure we didn't
1271 * go beyond the end of the map, or wrap around the address.
1272 * Then check to see if this is the last entry or if the
1273 * proposed end fits in the gap between this and the next
1276 end
= start
+ length
;
1277 if (end
> map
->max_offset
|| end
< start
)
1282 * If the next entry's start address is beyond the desired
1283 * end address we may have found a good entry.
1285 * If the next entry is a stack mapping we do not map into
1286 * the stack's reserved space.
1288 * XXX continue to allow mapping into the stack's reserved
1289 * space if doing a MAP_STACK mapping inside a MAP_STACK
1290 * mapping, for backwards compatibility. But the caller
1291 * really should use MAP_STACK | MAP_TRYFIXED if they
1294 if (next
== &map
->header
)
1296 if (next
->start
>= end
) {
1297 if ((next
->eflags
& MAP_ENTRY_STACK
) == 0)
1299 if (flags
& MAP_STACK
)
1301 if (next
->start
- next
->aux
.avail_ssize
>= end
)
1308 * Grow the kernel_map if necessary. pmap_growkernel() will panic
1309 * if it fails. The kernel_map is locked and nothing can steal
1310 * our address space if pmap_growkernel() blocks.
1312 * NOTE: This may be unconditionally called for kldload areas on
1313 * x86_64 because these do not bump kernel_vm_end (which would
1314 * fill 128G worth of page tables!). Therefore we must not
1317 if (map
== &kernel_map
) {
1320 kstop
= round_page(start
+ length
);
1321 if (kstop
> kernel_vm_end
)
1322 pmap_growkernel(start
, kstop
);
1329 * vm_map_find finds an unallocated region in the target address map with
1330 * the given length and allocates it. The search is defined to be first-fit
1331 * from the specified address; the region found is returned in the same
1334 * If object is non-NULL, ref count must be bumped by caller
1335 * prior to making call to account for the new entry.
1337 * No requirements. This function will lock the map temporarily.
1340 vm_map_find(vm_map_t map
, void *map_object
, void *map_aux
,
1341 vm_ooffset_t offset
, vm_offset_t
*addr
,
1342 vm_size_t length
, vm_size_t align
, boolean_t fitit
,
1343 vm_maptype_t maptype
, vm_subsys_t id
,
1344 vm_prot_t prot
, vm_prot_t max
, int cow
)
1351 if (maptype
== VM_MAPTYPE_UKSMAP
)
1354 object
= map_object
;
1358 count
= vm_map_entry_reserve(MAP_RESERVE_COUNT
);
1361 vm_object_hold_shared(object
);
1363 if (vm_map_findspace(map
, start
, length
, align
, 0, addr
)) {
1365 vm_object_drop(object
);
1367 vm_map_entry_release(count
);
1368 return (KERN_NO_SPACE
);
1372 result
= vm_map_insert(map
, &count
, map_object
, map_aux
,
1373 offset
, start
, start
+ length
,
1374 maptype
, id
, prot
, max
, cow
);
1376 vm_object_drop(object
);
1378 vm_map_entry_release(count
);
1384 * Simplify the given map entry by merging with either neighbor. This
1385 * routine also has the ability to merge with both neighbors.
1387 * This routine guarentees that the passed entry remains valid (though
1388 * possibly extended). When merging, this routine may delete one or
1389 * both neighbors. No action is taken on entries which have their
1390 * in-transition flag set.
1392 * The map must be exclusively locked.
1395 vm_map_simplify_entry(vm_map_t map
, vm_map_entry_t entry
, int *countp
)
1397 vm_map_entry_t next
, prev
;
1398 vm_size_t prevsize
, esize
;
1400 if (entry
->eflags
& MAP_ENTRY_IN_TRANSITION
) {
1401 ++mycpu
->gd_cnt
.v_intrans_coll
;
1405 if (entry
->maptype
== VM_MAPTYPE_SUBMAP
)
1407 if (entry
->maptype
== VM_MAPTYPE_UKSMAP
)
1411 if (prev
!= &map
->header
) {
1412 prevsize
= prev
->end
- prev
->start
;
1413 if ( (prev
->end
== entry
->start
) &&
1414 (prev
->maptype
== entry
->maptype
) &&
1415 (prev
->object
.vm_object
== entry
->object
.vm_object
) &&
1416 (!prev
->object
.vm_object
||
1417 (prev
->offset
+ prevsize
== entry
->offset
)) &&
1418 (prev
->eflags
== entry
->eflags
) &&
1419 (prev
->protection
== entry
->protection
) &&
1420 (prev
->max_protection
== entry
->max_protection
) &&
1421 (prev
->inheritance
== entry
->inheritance
) &&
1422 (prev
->id
== entry
->id
) &&
1423 (prev
->wired_count
== entry
->wired_count
)) {
1424 if (map
->first_free
== prev
)
1425 map
->first_free
= entry
;
1426 if (map
->hint
== prev
)
1428 vm_map_entry_unlink(map
, prev
);
1429 entry
->start
= prev
->start
;
1430 entry
->offset
= prev
->offset
;
1431 if (prev
->object
.vm_object
)
1432 vm_object_deallocate(prev
->object
.vm_object
);
1433 vm_map_entry_dispose(map
, prev
, countp
);
1438 if (next
!= &map
->header
) {
1439 esize
= entry
->end
- entry
->start
;
1440 if ((entry
->end
== next
->start
) &&
1441 (next
->maptype
== entry
->maptype
) &&
1442 (next
->object
.vm_object
== entry
->object
.vm_object
) &&
1443 (!entry
->object
.vm_object
||
1444 (entry
->offset
+ esize
== next
->offset
)) &&
1445 (next
->eflags
== entry
->eflags
) &&
1446 (next
->protection
== entry
->protection
) &&
1447 (next
->max_protection
== entry
->max_protection
) &&
1448 (next
->inheritance
== entry
->inheritance
) &&
1449 (next
->id
== entry
->id
) &&
1450 (next
->wired_count
== entry
->wired_count
)) {
1451 if (map
->first_free
== next
)
1452 map
->first_free
= entry
;
1453 if (map
->hint
== next
)
1455 vm_map_entry_unlink(map
, next
);
1456 entry
->end
= next
->end
;
1457 if (next
->object
.vm_object
)
1458 vm_object_deallocate(next
->object
.vm_object
);
1459 vm_map_entry_dispose(map
, next
, countp
);
1465 * Asserts that the given entry begins at or after the specified address.
1466 * If necessary, it splits the entry into two.
1468 #define vm_map_clip_start(map, entry, startaddr, countp) \
1470 if (startaddr > entry->start) \
1471 _vm_map_clip_start(map, entry, startaddr, countp); \
1475 * This routine is called only when it is known that the entry must be split.
1477 * The map must be exclusively locked.
1480 _vm_map_clip_start(vm_map_t map
, vm_map_entry_t entry
, vm_offset_t start
,
1483 vm_map_entry_t new_entry
;
1486 * Split off the front portion -- note that we must insert the new
1487 * entry BEFORE this one, so that this entry has the specified
1491 vm_map_simplify_entry(map
, entry
, countp
);
1494 * If there is no object backing this entry, we might as well create
1495 * one now. If we defer it, an object can get created after the map
1496 * is clipped, and individual objects will be created for the split-up
1497 * map. This is a bit of a hack, but is also about the best place to
1498 * put this improvement.
1500 if (entry
->object
.vm_object
== NULL
&& !map
->system_map
) {
1501 vm_map_entry_allocate_object(entry
);
1504 new_entry
= vm_map_entry_create(map
, countp
);
1505 *new_entry
= *entry
;
1507 new_entry
->end
= start
;
1508 entry
->offset
+= (start
- entry
->start
);
1509 entry
->start
= start
;
1511 vm_map_entry_link(map
, entry
->prev
, new_entry
);
1513 switch(entry
->maptype
) {
1514 case VM_MAPTYPE_NORMAL
:
1515 case VM_MAPTYPE_VPAGETABLE
:
1516 if (new_entry
->object
.vm_object
) {
1517 vm_object_hold(new_entry
->object
.vm_object
);
1518 vm_object_chain_wait(new_entry
->object
.vm_object
, 0);
1519 vm_object_reference_locked(new_entry
->object
.vm_object
);
1520 vm_object_drop(new_entry
->object
.vm_object
);
1529 * Asserts that the given entry ends at or before the specified address.
1530 * If necessary, it splits the entry into two.
1532 * The map must be exclusively locked.
1534 #define vm_map_clip_end(map, entry, endaddr, countp) \
1536 if (endaddr < entry->end) \
1537 _vm_map_clip_end(map, entry, endaddr, countp); \
1541 * This routine is called only when it is known that the entry must be split.
1543 * The map must be exclusively locked.
1546 _vm_map_clip_end(vm_map_t map
, vm_map_entry_t entry
, vm_offset_t end
,
1549 vm_map_entry_t new_entry
;
1552 * If there is no object backing this entry, we might as well create
1553 * one now. If we defer it, an object can get created after the map
1554 * is clipped, and individual objects will be created for the split-up
1555 * map. This is a bit of a hack, but is also about the best place to
1556 * put this improvement.
1559 if (entry
->object
.vm_object
== NULL
&& !map
->system_map
) {
1560 vm_map_entry_allocate_object(entry
);
1564 * Create a new entry and insert it AFTER the specified entry
1567 new_entry
= vm_map_entry_create(map
, countp
);
1568 *new_entry
= *entry
;
1570 new_entry
->start
= entry
->end
= end
;
1571 new_entry
->offset
+= (end
- entry
->start
);
1573 vm_map_entry_link(map
, entry
, new_entry
);
1575 switch(entry
->maptype
) {
1576 case VM_MAPTYPE_NORMAL
:
1577 case VM_MAPTYPE_VPAGETABLE
:
1578 if (new_entry
->object
.vm_object
) {
1579 vm_object_hold(new_entry
->object
.vm_object
);
1580 vm_object_chain_wait(new_entry
->object
.vm_object
, 0);
1581 vm_object_reference_locked(new_entry
->object
.vm_object
);
1582 vm_object_drop(new_entry
->object
.vm_object
);
1591 * Asserts that the starting and ending region addresses fall within the
1592 * valid range for the map.
1594 #define VM_MAP_RANGE_CHECK(map, start, end) \
1596 if (start < vm_map_min(map)) \
1597 start = vm_map_min(map); \
1598 if (end > vm_map_max(map)) \
1599 end = vm_map_max(map); \
1605 * Used to block when an in-transition collison occurs. The map
1606 * is unlocked for the sleep and relocked before the return.
1609 vm_map_transition_wait(vm_map_t map
)
1611 tsleep_interlock(map
, 0);
1613 tsleep(map
, PINTERLOCKED
, "vment", 0);
1618 * When we do blocking operations with the map lock held it is
1619 * possible that a clip might have occured on our in-transit entry,
1620 * requiring an adjustment to the entry in our loop. These macros
1621 * help the pageable and clip_range code deal with the case. The
1622 * conditional costs virtually nothing if no clipping has occured.
1625 #define CLIP_CHECK_BACK(entry, save_start) \
1627 while (entry->start != save_start) { \
1628 entry = entry->prev; \
1629 KASSERT(entry != &map->header, ("bad entry clip")); \
1633 #define CLIP_CHECK_FWD(entry, save_end) \
1635 while (entry->end != save_end) { \
1636 entry = entry->next; \
1637 KASSERT(entry != &map->header, ("bad entry clip")); \
1643 * Clip the specified range and return the base entry. The
1644 * range may cover several entries starting at the returned base
1645 * and the first and last entry in the covering sequence will be
1646 * properly clipped to the requested start and end address.
1648 * If no holes are allowed you should pass the MAP_CLIP_NO_HOLES
1651 * The MAP_ENTRY_IN_TRANSITION flag will be set for the entries
1652 * covered by the requested range.
1654 * The map must be exclusively locked on entry and will remain locked
1655 * on return. If no range exists or the range contains holes and you
1656 * specified that no holes were allowed, NULL will be returned. This
1657 * routine may temporarily unlock the map in order avoid a deadlock when
1662 vm_map_clip_range(vm_map_t map
, vm_offset_t start
, vm_offset_t end
,
1663 int *countp
, int flags
)
1665 vm_map_entry_t start_entry
;
1666 vm_map_entry_t entry
;
1669 * Locate the entry and effect initial clipping. The in-transition
1670 * case does not occur very often so do not try to optimize it.
1673 if (vm_map_lookup_entry(map
, start
, &start_entry
) == FALSE
)
1675 entry
= start_entry
;
1676 if (entry
->eflags
& MAP_ENTRY_IN_TRANSITION
) {
1677 entry
->eflags
|= MAP_ENTRY_NEEDS_WAKEUP
;
1678 ++mycpu
->gd_cnt
.v_intrans_coll
;
1679 ++mycpu
->gd_cnt
.v_intrans_wait
;
1680 vm_map_transition_wait(map
);
1682 * entry and/or start_entry may have been clipped while
1683 * we slept, or may have gone away entirely. We have
1684 * to restart from the lookup.
1690 * Since we hold an exclusive map lock we do not have to restart
1691 * after clipping, even though clipping may block in zalloc.
1693 vm_map_clip_start(map
, entry
, start
, countp
);
1694 vm_map_clip_end(map
, entry
, end
, countp
);
1695 entry
->eflags
|= MAP_ENTRY_IN_TRANSITION
;
1698 * Scan entries covered by the range. When working on the next
1699 * entry a restart need only re-loop on the current entry which
1700 * we have already locked, since 'next' may have changed. Also,
1701 * even though entry is safe, it may have been clipped so we
1702 * have to iterate forwards through the clip after sleeping.
1704 while (entry
->next
!= &map
->header
&& entry
->next
->start
< end
) {
1705 vm_map_entry_t next
= entry
->next
;
1707 if (flags
& MAP_CLIP_NO_HOLES
) {
1708 if (next
->start
> entry
->end
) {
1709 vm_map_unclip_range(map
, start_entry
,
1710 start
, entry
->end
, countp
, flags
);
1715 if (next
->eflags
& MAP_ENTRY_IN_TRANSITION
) {
1716 vm_offset_t save_end
= entry
->end
;
1717 next
->eflags
|= MAP_ENTRY_NEEDS_WAKEUP
;
1718 ++mycpu
->gd_cnt
.v_intrans_coll
;
1719 ++mycpu
->gd_cnt
.v_intrans_wait
;
1720 vm_map_transition_wait(map
);
1723 * clips might have occured while we blocked.
1725 CLIP_CHECK_FWD(entry
, save_end
);
1726 CLIP_CHECK_BACK(start_entry
, start
);
1730 * No restart necessary even though clip_end may block, we
1731 * are holding the map lock.
1733 vm_map_clip_end(map
, next
, end
, countp
);
1734 next
->eflags
|= MAP_ENTRY_IN_TRANSITION
;
1737 if (flags
& MAP_CLIP_NO_HOLES
) {
1738 if (entry
->end
!= end
) {
1739 vm_map_unclip_range(map
, start_entry
,
1740 start
, entry
->end
, countp
, flags
);
1744 return(start_entry
);
1748 * Undo the effect of vm_map_clip_range(). You should pass the same
1749 * flags and the same range that you passed to vm_map_clip_range().
1750 * This code will clear the in-transition flag on the entries and
1751 * wake up anyone waiting. This code will also simplify the sequence
1752 * and attempt to merge it with entries before and after the sequence.
1754 * The map must be locked on entry and will remain locked on return.
1756 * Note that you should also pass the start_entry returned by
1757 * vm_map_clip_range(). However, if you block between the two calls
1758 * with the map unlocked please be aware that the start_entry may
1759 * have been clipped and you may need to scan it backwards to find
1760 * the entry corresponding with the original start address. You are
1761 * responsible for this, vm_map_unclip_range() expects the correct
1762 * start_entry to be passed to it and will KASSERT otherwise.
1766 vm_map_unclip_range(vm_map_t map
, vm_map_entry_t start_entry
,
1767 vm_offset_t start
, vm_offset_t end
,
1768 int *countp
, int flags
)
1770 vm_map_entry_t entry
;
1772 entry
= start_entry
;
1774 KASSERT(entry
->start
== start
, ("unclip_range: illegal base entry"));
1775 while (entry
!= &map
->header
&& entry
->start
< end
) {
1776 KASSERT(entry
->eflags
& MAP_ENTRY_IN_TRANSITION
,
1777 ("in-transition flag not set during unclip on: %p",
1779 KASSERT(entry
->end
<= end
,
1780 ("unclip_range: tail wasn't clipped"));
1781 entry
->eflags
&= ~MAP_ENTRY_IN_TRANSITION
;
1782 if (entry
->eflags
& MAP_ENTRY_NEEDS_WAKEUP
) {
1783 entry
->eflags
&= ~MAP_ENTRY_NEEDS_WAKEUP
;
1786 entry
= entry
->next
;
1790 * Simplification does not block so there is no restart case.
1792 entry
= start_entry
;
1793 while (entry
!= &map
->header
&& entry
->start
< end
) {
1794 vm_map_simplify_entry(map
, entry
, countp
);
1795 entry
= entry
->next
;
1800 * Mark the given range as handled by a subordinate map.
1802 * This range must have been created with vm_map_find(), and no other
1803 * operations may have been performed on this range prior to calling
1806 * Submappings cannot be removed.
1811 vm_map_submap(vm_map_t map
, vm_offset_t start
, vm_offset_t end
, vm_map_t submap
)
1813 vm_map_entry_t entry
;
1814 int result
= KERN_INVALID_ARGUMENT
;
1817 count
= vm_map_entry_reserve(MAP_RESERVE_COUNT
);
1820 VM_MAP_RANGE_CHECK(map
, start
, end
);
1822 if (vm_map_lookup_entry(map
, start
, &entry
)) {
1823 vm_map_clip_start(map
, entry
, start
, &count
);
1825 entry
= entry
->next
;
1828 vm_map_clip_end(map
, entry
, end
, &count
);
1830 if ((entry
->start
== start
) && (entry
->end
== end
) &&
1831 ((entry
->eflags
& MAP_ENTRY_COW
) == 0) &&
1832 (entry
->object
.vm_object
== NULL
)) {
1833 entry
->object
.sub_map
= submap
;
1834 entry
->maptype
= VM_MAPTYPE_SUBMAP
;
1835 result
= KERN_SUCCESS
;
1838 vm_map_entry_release(count
);
1844 * Sets the protection of the specified address region in the target map.
1845 * If "set_max" is specified, the maximum protection is to be set;
1846 * otherwise, only the current protection is affected.
1848 * The protection is not applicable to submaps, but is applicable to normal
1849 * maps and maps governed by virtual page tables. For example, when operating
1850 * on a virtual page table our protection basically controls how COW occurs
1851 * on the backing object, whereas the virtual page table abstraction itself
1852 * is an abstraction for userland.
1857 vm_map_protect(vm_map_t map
, vm_offset_t start
, vm_offset_t end
,
1858 vm_prot_t new_prot
, boolean_t set_max
)
1860 vm_map_entry_t current
;
1861 vm_map_entry_t entry
;
1864 count
= vm_map_entry_reserve(MAP_RESERVE_COUNT
);
1867 VM_MAP_RANGE_CHECK(map
, start
, end
);
1869 if (vm_map_lookup_entry(map
, start
, &entry
)) {
1870 vm_map_clip_start(map
, entry
, start
, &count
);
1872 entry
= entry
->next
;
1876 * Make a first pass to check for protection violations.
1879 while ((current
!= &map
->header
) && (current
->start
< end
)) {
1880 if (current
->maptype
== VM_MAPTYPE_SUBMAP
) {
1882 vm_map_entry_release(count
);
1883 return (KERN_INVALID_ARGUMENT
);
1885 if ((new_prot
& current
->max_protection
) != new_prot
) {
1887 vm_map_entry_release(count
);
1888 return (KERN_PROTECTION_FAILURE
);
1890 current
= current
->next
;
1894 * Go back and fix up protections. [Note that clipping is not
1895 * necessary the second time.]
1899 while ((current
!= &map
->header
) && (current
->start
< end
)) {
1902 vm_map_clip_end(map
, current
, end
, &count
);
1904 old_prot
= current
->protection
;
1906 current
->protection
=
1907 (current
->max_protection
= new_prot
) &
1910 current
->protection
= new_prot
;
1914 * Update physical map if necessary. Worry about copy-on-write
1915 * here -- CHECK THIS XXX
1918 if (current
->protection
!= old_prot
) {
1919 #define MASK(entry) (((entry)->eflags & MAP_ENTRY_COW) ? ~VM_PROT_WRITE : \
1922 pmap_protect(map
->pmap
, current
->start
,
1924 current
->protection
& MASK(current
));
1928 vm_map_simplify_entry(map
, current
, &count
);
1930 current
= current
->next
;
1934 vm_map_entry_release(count
);
1935 return (KERN_SUCCESS
);
1939 * This routine traverses a processes map handling the madvise
1940 * system call. Advisories are classified as either those effecting
1941 * the vm_map_entry structure, or those effecting the underlying
1944 * The <value> argument is used for extended madvise calls.
1949 vm_map_madvise(vm_map_t map
, vm_offset_t start
, vm_offset_t end
,
1950 int behav
, off_t value
)
1952 vm_map_entry_t current
, entry
;
1958 * Some madvise calls directly modify the vm_map_entry, in which case
1959 * we need to use an exclusive lock on the map and we need to perform
1960 * various clipping operations. Otherwise we only need a read-lock
1964 count
= vm_map_entry_reserve(MAP_RESERVE_COUNT
);
1968 case MADV_SEQUENTIAL
:
1982 vm_map_lock_read(map
);
1985 vm_map_entry_release(count
);
1990 * Locate starting entry and clip if necessary.
1993 VM_MAP_RANGE_CHECK(map
, start
, end
);
1995 if (vm_map_lookup_entry(map
, start
, &entry
)) {
1997 vm_map_clip_start(map
, entry
, start
, &count
);
1999 entry
= entry
->next
;
2004 * madvise behaviors that are implemented in the vm_map_entry.
2006 * We clip the vm_map_entry so that behavioral changes are
2007 * limited to the specified address range.
2009 for (current
= entry
;
2010 (current
!= &map
->header
) && (current
->start
< end
);
2011 current
= current
->next
2013 if (current
->maptype
== VM_MAPTYPE_SUBMAP
)
2016 vm_map_clip_end(map
, current
, end
, &count
);
2020 vm_map_entry_set_behavior(current
, MAP_ENTRY_BEHAV_NORMAL
);
2022 case MADV_SEQUENTIAL
:
2023 vm_map_entry_set_behavior(current
, MAP_ENTRY_BEHAV_SEQUENTIAL
);
2026 vm_map_entry_set_behavior(current
, MAP_ENTRY_BEHAV_RANDOM
);
2029 current
->eflags
|= MAP_ENTRY_NOSYNC
;
2032 current
->eflags
&= ~MAP_ENTRY_NOSYNC
;
2035 current
->eflags
|= MAP_ENTRY_NOCOREDUMP
;
2038 current
->eflags
&= ~MAP_ENTRY_NOCOREDUMP
;
2042 * Invalidate the related pmap entries, used
2043 * to flush portions of the real kernel's
2044 * pmap when the caller has removed or
2045 * modified existing mappings in a virtual
2048 pmap_remove(map
->pmap
,
2049 current
->start
, current
->end
);
2053 * Set the page directory page for a map
2054 * governed by a virtual page table. Mark
2055 * the entry as being governed by a virtual
2056 * page table if it is not.
2058 * XXX the page directory page is stored
2059 * in the avail_ssize field if the map_entry.
2061 * XXX the map simplification code does not
2062 * compare this field so weird things may
2063 * happen if you do not apply this function
2064 * to the entire mapping governed by the
2065 * virtual page table.
2067 if (current
->maptype
!= VM_MAPTYPE_VPAGETABLE
) {
2071 current
->aux
.master_pde
= value
;
2072 pmap_remove(map
->pmap
,
2073 current
->start
, current
->end
);
2079 vm_map_simplify_entry(map
, current
, &count
);
2087 * madvise behaviors that are implemented in the underlying
2090 * Since we don't clip the vm_map_entry, we have to clip
2091 * the vm_object pindex and count.
2093 * NOTE! We currently do not support these functions on
2094 * virtual page tables.
2096 for (current
= entry
;
2097 (current
!= &map
->header
) && (current
->start
< end
);
2098 current
= current
->next
2100 vm_offset_t useStart
;
2102 if (current
->maptype
!= VM_MAPTYPE_NORMAL
)
2105 pindex
= OFF_TO_IDX(current
->offset
);
2106 count
= atop(current
->end
- current
->start
);
2107 useStart
= current
->start
;
2109 if (current
->start
< start
) {
2110 pindex
+= atop(start
- current
->start
);
2111 count
-= atop(start
- current
->start
);
2114 if (current
->end
> end
)
2115 count
-= atop(current
->end
- end
);
2120 vm_object_madvise(current
->object
.vm_object
,
2121 pindex
, count
, behav
);
2124 * Try to populate the page table. Mappings governed
2125 * by virtual page tables cannot be pre-populated
2126 * without a lot of work so don't try.
2128 if (behav
== MADV_WILLNEED
&&
2129 current
->maptype
!= VM_MAPTYPE_VPAGETABLE
) {
2130 pmap_object_init_pt(
2133 current
->protection
,
2134 current
->object
.vm_object
,
2136 (count
<< PAGE_SHIFT
),
2137 MAP_PREFAULT_MADVISE
2141 vm_map_unlock_read(map
);
2143 vm_map_entry_release(count
);
2149 * Sets the inheritance of the specified address range in the target map.
2150 * Inheritance affects how the map will be shared with child maps at the
2151 * time of vm_map_fork.
2154 vm_map_inherit(vm_map_t map
, vm_offset_t start
, vm_offset_t end
,
2155 vm_inherit_t new_inheritance
)
2157 vm_map_entry_t entry
;
2158 vm_map_entry_t temp_entry
;
2161 switch (new_inheritance
) {
2162 case VM_INHERIT_NONE
:
2163 case VM_INHERIT_COPY
:
2164 case VM_INHERIT_SHARE
:
2167 return (KERN_INVALID_ARGUMENT
);
2170 count
= vm_map_entry_reserve(MAP_RESERVE_COUNT
);
2173 VM_MAP_RANGE_CHECK(map
, start
, end
);
2175 if (vm_map_lookup_entry(map
, start
, &temp_entry
)) {
2177 vm_map_clip_start(map
, entry
, start
, &count
);
2179 entry
= temp_entry
->next
;
2181 while ((entry
!= &map
->header
) && (entry
->start
< end
)) {
2182 vm_map_clip_end(map
, entry
, end
, &count
);
2184 entry
->inheritance
= new_inheritance
;
2186 vm_map_simplify_entry(map
, entry
, &count
);
2188 entry
= entry
->next
;
2191 vm_map_entry_release(count
);
2192 return (KERN_SUCCESS
);
2196 * Implement the semantics of mlock
2199 vm_map_unwire(vm_map_t map
, vm_offset_t start
, vm_offset_t real_end
,
2200 boolean_t new_pageable
)
2202 vm_map_entry_t entry
;
2203 vm_map_entry_t start_entry
;
2205 int rv
= KERN_SUCCESS
;
2208 count
= vm_map_entry_reserve(MAP_RESERVE_COUNT
);
2210 VM_MAP_RANGE_CHECK(map
, start
, real_end
);
2213 start_entry
= vm_map_clip_range(map
, start
, end
, &count
,
2215 if (start_entry
== NULL
) {
2217 vm_map_entry_release(count
);
2218 return (KERN_INVALID_ADDRESS
);
2221 if (new_pageable
== 0) {
2222 entry
= start_entry
;
2223 while ((entry
!= &map
->header
) && (entry
->start
< end
)) {
2224 vm_offset_t save_start
;
2225 vm_offset_t save_end
;
2228 * Already user wired or hard wired (trivial cases)
2230 if (entry
->eflags
& MAP_ENTRY_USER_WIRED
) {
2231 entry
= entry
->next
;
2234 if (entry
->wired_count
!= 0) {
2235 entry
->wired_count
++;
2236 entry
->eflags
|= MAP_ENTRY_USER_WIRED
;
2237 entry
= entry
->next
;
2242 * A new wiring requires instantiation of appropriate
2243 * management structures and the faulting in of the
2246 if (entry
->maptype
== VM_MAPTYPE_NORMAL
||
2247 entry
->maptype
== VM_MAPTYPE_VPAGETABLE
) {
2248 int copyflag
= entry
->eflags
&
2249 MAP_ENTRY_NEEDS_COPY
;
2250 if (copyflag
&& ((entry
->protection
&
2251 VM_PROT_WRITE
) != 0)) {
2252 vm_map_entry_shadow(entry
, 0);
2253 } else if (entry
->object
.vm_object
== NULL
&&
2255 vm_map_entry_allocate_object(entry
);
2258 entry
->wired_count
++;
2259 entry
->eflags
|= MAP_ENTRY_USER_WIRED
;
2262 * Now fault in the area. Note that vm_fault_wire()
2263 * may release the map lock temporarily, it will be
2264 * relocked on return. The in-transition
2265 * flag protects the entries.
2267 save_start
= entry
->start
;
2268 save_end
= entry
->end
;
2269 rv
= vm_fault_wire(map
, entry
, TRUE
, 0);
2271 CLIP_CHECK_BACK(entry
, save_start
);
2273 KASSERT(entry
->wired_count
== 1, ("bad wired_count on entry"));
2274 entry
->eflags
&= ~MAP_ENTRY_USER_WIRED
;
2275 entry
->wired_count
= 0;
2276 if (entry
->end
== save_end
)
2278 entry
= entry
->next
;
2279 KASSERT(entry
!= &map
->header
, ("bad entry clip during backout"));
2281 end
= save_start
; /* unwire the rest */
2285 * note that even though the entry might have been
2286 * clipped, the USER_WIRED flag we set prevents
2287 * duplication so we do not have to do a
2290 entry
= entry
->next
;
2294 * If we failed fall through to the unwiring section to
2295 * unwire what we had wired so far. 'end' has already
2302 * start_entry might have been clipped if we unlocked the
2303 * map and blocked. No matter how clipped it has gotten
2304 * there should be a fragment that is on our start boundary.
2306 CLIP_CHECK_BACK(start_entry
, start
);
2310 * Deal with the unwiring case.
2314 * This is the unwiring case. We must first ensure that the
2315 * range to be unwired is really wired down. We know there
2318 entry
= start_entry
;
2319 while ((entry
!= &map
->header
) && (entry
->start
< end
)) {
2320 if ((entry
->eflags
& MAP_ENTRY_USER_WIRED
) == 0) {
2321 rv
= KERN_INVALID_ARGUMENT
;
2324 KASSERT(entry
->wired_count
!= 0, ("wired count was 0 with USER_WIRED set! %p", entry
));
2325 entry
= entry
->next
;
2329 * Now decrement the wiring count for each region. If a region
2330 * becomes completely unwired, unwire its physical pages and
2334 * The map entries are processed in a loop, checking to
2335 * make sure the entry is wired and asserting it has a wired
2336 * count. However, another loop was inserted more-or-less in
2337 * the middle of the unwiring path. This loop picks up the
2338 * "entry" loop variable from the first loop without first
2339 * setting it to start_entry. Naturally, the secound loop
2340 * is never entered and the pages backing the entries are
2341 * never unwired. This can lead to a leak of wired pages.
2343 entry
= start_entry
;
2344 while ((entry
!= &map
->header
) && (entry
->start
< end
)) {
2345 KASSERT(entry
->eflags
& MAP_ENTRY_USER_WIRED
,
2346 ("expected USER_WIRED on entry %p", entry
));
2347 entry
->eflags
&= ~MAP_ENTRY_USER_WIRED
;
2348 entry
->wired_count
--;
2349 if (entry
->wired_count
== 0)
2350 vm_fault_unwire(map
, entry
);
2351 entry
= entry
->next
;
2355 vm_map_unclip_range(map
, start_entry
, start
, real_end
, &count
,
2359 vm_map_entry_release(count
);
2364 * Sets the pageability of the specified address range in the target map.
2365 * Regions specified as not pageable require locked-down physical
2366 * memory and physical page maps.
2368 * The map must not be locked, but a reference must remain to the map
2369 * throughout the call.
2371 * This function may be called via the zalloc path and must properly
2372 * reserve map entries for kernel_map.
2377 vm_map_wire(vm_map_t map
, vm_offset_t start
, vm_offset_t real_end
, int kmflags
)
2379 vm_map_entry_t entry
;
2380 vm_map_entry_t start_entry
;
2382 int rv
= KERN_SUCCESS
;
2385 if (kmflags
& KM_KRESERVE
)
2386 count
= vm_map_entry_kreserve(MAP_RESERVE_COUNT
);
2388 count
= vm_map_entry_reserve(MAP_RESERVE_COUNT
);
2390 VM_MAP_RANGE_CHECK(map
, start
, real_end
);
2393 start_entry
= vm_map_clip_range(map
, start
, end
, &count
,
2395 if (start_entry
== NULL
) {
2397 rv
= KERN_INVALID_ADDRESS
;
2400 if ((kmflags
& KM_PAGEABLE
) == 0) {
2404 * 1. Holding the write lock, we create any shadow or zero-fill
2405 * objects that need to be created. Then we clip each map
2406 * entry to the region to be wired and increment its wiring
2407 * count. We create objects before clipping the map entries
2408 * to avoid object proliferation.
2410 * 2. We downgrade to a read lock, and call vm_fault_wire to
2411 * fault in the pages for any newly wired area (wired_count is
2414 * Downgrading to a read lock for vm_fault_wire avoids a
2415 * possible deadlock with another process that may have faulted
2416 * on one of the pages to be wired (it would mark the page busy,
2417 * blocking us, then in turn block on the map lock that we
2418 * hold). Because of problems in the recursive lock package,
2419 * we cannot upgrade to a write lock in vm_map_lookup. Thus,
2420 * any actions that require the write lock must be done
2421 * beforehand. Because we keep the read lock on the map, the
2422 * copy-on-write status of the entries we modify here cannot
2425 entry
= start_entry
;
2426 while ((entry
!= &map
->header
) && (entry
->start
< end
)) {
2428 * Trivial case if the entry is already wired
2430 if (entry
->wired_count
) {
2431 entry
->wired_count
++;
2432 entry
= entry
->next
;
2437 * The entry is being newly wired, we have to setup
2438 * appropriate management structures. A shadow
2439 * object is required for a copy-on-write region,
2440 * or a normal object for a zero-fill region. We
2441 * do not have to do this for entries that point to sub
2442 * maps because we won't hold the lock on the sub map.
2444 if (entry
->maptype
== VM_MAPTYPE_NORMAL
||
2445 entry
->maptype
== VM_MAPTYPE_VPAGETABLE
) {
2446 int copyflag
= entry
->eflags
&
2447 MAP_ENTRY_NEEDS_COPY
;
2448 if (copyflag
&& ((entry
->protection
&
2449 VM_PROT_WRITE
) != 0)) {
2450 vm_map_entry_shadow(entry
, 0);
2451 } else if (entry
->object
.vm_object
== NULL
&&
2453 vm_map_entry_allocate_object(entry
);
2457 entry
->wired_count
++;
2458 entry
= entry
->next
;
2466 * HACK HACK HACK HACK
2468 * vm_fault_wire() temporarily unlocks the map to avoid
2469 * deadlocks. The in-transition flag from vm_map_clip_range
2470 * call should protect us from changes while the map is
2473 * NOTE: Previously this comment stated that clipping might
2474 * still occur while the entry is unlocked, but from
2475 * what I can tell it actually cannot.
2477 * It is unclear whether the CLIP_CHECK_*() calls
2478 * are still needed but we keep them in anyway.
2480 * HACK HACK HACK HACK
2483 entry
= start_entry
;
2484 while (entry
!= &map
->header
&& entry
->start
< end
) {
2486 * If vm_fault_wire fails for any page we need to undo
2487 * what has been done. We decrement the wiring count
2488 * for those pages which have not yet been wired (now)
2489 * and unwire those that have (later).
2491 vm_offset_t save_start
= entry
->start
;
2492 vm_offset_t save_end
= entry
->end
;
2494 if (entry
->wired_count
== 1)
2495 rv
= vm_fault_wire(map
, entry
, FALSE
, kmflags
);
2497 CLIP_CHECK_BACK(entry
, save_start
);
2499 KASSERT(entry
->wired_count
== 1, ("wired_count changed unexpectedly"));
2500 entry
->wired_count
= 0;
2501 if (entry
->end
== save_end
)
2503 entry
= entry
->next
;
2504 KASSERT(entry
!= &map
->header
, ("bad entry clip during backout"));
2509 CLIP_CHECK_FWD(entry
, save_end
);
2510 entry
= entry
->next
;
2514 * If a failure occured undo everything by falling through
2515 * to the unwiring code. 'end' has already been adjusted
2519 kmflags
|= KM_PAGEABLE
;
2522 * start_entry is still IN_TRANSITION but may have been
2523 * clipped since vm_fault_wire() unlocks and relocks the
2524 * map. No matter how clipped it has gotten there should
2525 * be a fragment that is on our start boundary.
2527 CLIP_CHECK_BACK(start_entry
, start
);
2530 if (kmflags
& KM_PAGEABLE
) {
2532 * This is the unwiring case. We must first ensure that the
2533 * range to be unwired is really wired down. We know there
2536 entry
= start_entry
;
2537 while ((entry
!= &map
->header
) && (entry
->start
< end
)) {
2538 if (entry
->wired_count
== 0) {
2539 rv
= KERN_INVALID_ARGUMENT
;
2542 entry
= entry
->next
;
2546 * Now decrement the wiring count for each region. If a region
2547 * becomes completely unwired, unwire its physical pages and
2550 entry
= start_entry
;
2551 while ((entry
!= &map
->header
) && (entry
->start
< end
)) {
2552 entry
->wired_count
--;
2553 if (entry
->wired_count
== 0)
2554 vm_fault_unwire(map
, entry
);
2555 entry
= entry
->next
;
2559 vm_map_unclip_range(map
, start_entry
, start
, real_end
,
2560 &count
, MAP_CLIP_NO_HOLES
);
2564 if (kmflags
& KM_KRESERVE
)
2565 vm_map_entry_krelease(count
);
2567 vm_map_entry_release(count
);
2572 * Mark a newly allocated address range as wired but do not fault in
2573 * the pages. The caller is expected to load the pages into the object.
2575 * The map must be locked on entry and will remain locked on return.
2576 * No other requirements.
2579 vm_map_set_wired_quick(vm_map_t map
, vm_offset_t addr
, vm_size_t size
,
2582 vm_map_entry_t scan
;
2583 vm_map_entry_t entry
;
2585 entry
= vm_map_clip_range(map
, addr
, addr
+ size
,
2586 countp
, MAP_CLIP_NO_HOLES
);
2588 scan
!= &map
->header
&& scan
->start
< addr
+ size
;
2589 scan
= scan
->next
) {
2590 KKASSERT(scan
->wired_count
== 0);
2591 scan
->wired_count
= 1;
2593 vm_map_unclip_range(map
, entry
, addr
, addr
+ size
,
2594 countp
, MAP_CLIP_NO_HOLES
);
2598 * Push any dirty cached pages in the address range to their pager.
2599 * If syncio is TRUE, dirty pages are written synchronously.
2600 * If invalidate is TRUE, any cached pages are freed as well.
2602 * This routine is called by sys_msync()
2604 * Returns an error if any part of the specified range is not mapped.
2609 vm_map_clean(vm_map_t map
, vm_offset_t start
, vm_offset_t end
,
2610 boolean_t syncio
, boolean_t invalidate
)
2612 vm_map_entry_t current
;
2613 vm_map_entry_t entry
;
2617 vm_ooffset_t offset
;
2619 vm_map_lock_read(map
);
2620 VM_MAP_RANGE_CHECK(map
, start
, end
);
2621 if (!vm_map_lookup_entry(map
, start
, &entry
)) {
2622 vm_map_unlock_read(map
);
2623 return (KERN_INVALID_ADDRESS
);
2625 lwkt_gettoken(&map
->token
);
2628 * Make a first pass to check for holes.
2630 for (current
= entry
; current
->start
< end
; current
= current
->next
) {
2631 if (current
->maptype
== VM_MAPTYPE_SUBMAP
) {
2632 lwkt_reltoken(&map
->token
);
2633 vm_map_unlock_read(map
);
2634 return (KERN_INVALID_ARGUMENT
);
2636 if (end
> current
->end
&&
2637 (current
->next
== &map
->header
||
2638 current
->end
!= current
->next
->start
)) {
2639 lwkt_reltoken(&map
->token
);
2640 vm_map_unlock_read(map
);
2641 return (KERN_INVALID_ADDRESS
);
2646 pmap_remove(vm_map_pmap(map
), start
, end
);
2649 * Make a second pass, cleaning/uncaching pages from the indicated
2652 for (current
= entry
; current
->start
< end
; current
= current
->next
) {
2653 offset
= current
->offset
+ (start
- current
->start
);
2654 size
= (end
<= current
->end
? end
: current
->end
) - start
;
2656 switch(current
->maptype
) {
2657 case VM_MAPTYPE_SUBMAP
:
2660 vm_map_entry_t tentry
;
2663 smap
= current
->object
.sub_map
;
2664 vm_map_lock_read(smap
);
2665 vm_map_lookup_entry(smap
, offset
, &tentry
);
2666 tsize
= tentry
->end
- offset
;
2669 object
= tentry
->object
.vm_object
;
2670 offset
= tentry
->offset
+ (offset
- tentry
->start
);
2671 vm_map_unlock_read(smap
);
2674 case VM_MAPTYPE_NORMAL
:
2675 case VM_MAPTYPE_VPAGETABLE
:
2676 object
= current
->object
.vm_object
;
2684 vm_object_hold(object
);
2687 * Note that there is absolutely no sense in writing out
2688 * anonymous objects, so we track down the vnode object
2690 * We invalidate (remove) all pages from the address space
2691 * anyway, for semantic correctness.
2693 * note: certain anonymous maps, such as MAP_NOSYNC maps,
2694 * may start out with a NULL object.
2696 while (object
&& (tobj
= object
->backing_object
) != NULL
) {
2697 vm_object_hold(tobj
);
2698 if (tobj
== object
->backing_object
) {
2699 vm_object_lock_swap();
2700 offset
+= object
->backing_object_offset
;
2701 vm_object_drop(object
);
2703 if (object
->size
< OFF_TO_IDX(offset
+ size
))
2704 size
= IDX_TO_OFF(object
->size
) -
2708 vm_object_drop(tobj
);
2710 if (object
&& (object
->type
== OBJT_VNODE
) &&
2711 (current
->protection
& VM_PROT_WRITE
) &&
2712 (object
->flags
& OBJ_NOMSYNC
) == 0) {
2714 * Flush pages if writing is allowed, invalidate them
2715 * if invalidation requested. Pages undergoing I/O
2716 * will be ignored by vm_object_page_remove().
2718 * We cannot lock the vnode and then wait for paging
2719 * to complete without deadlocking against vm_fault.
2720 * Instead we simply call vm_object_page_remove() and
2721 * allow it to block internally on a page-by-page
2722 * basis when it encounters pages undergoing async
2727 /* no chain wait needed for vnode objects */
2728 vm_object_reference_locked(object
);
2729 vn_lock(object
->handle
, LK_EXCLUSIVE
| LK_RETRY
);
2730 flags
= (syncio
|| invalidate
) ? OBJPC_SYNC
: 0;
2731 flags
|= invalidate
? OBJPC_INVAL
: 0;
2734 * When operating on a virtual page table just
2735 * flush the whole object. XXX we probably ought
2738 switch(current
->maptype
) {
2739 case VM_MAPTYPE_NORMAL
:
2740 vm_object_page_clean(object
,
2742 OFF_TO_IDX(offset
+ size
+ PAGE_MASK
),
2745 case VM_MAPTYPE_VPAGETABLE
:
2746 vm_object_page_clean(object
, 0, 0, flags
);
2749 vn_unlock(((struct vnode
*)object
->handle
));
2750 vm_object_deallocate_locked(object
);
2752 if (object
&& invalidate
&&
2753 ((object
->type
== OBJT_VNODE
) ||
2754 (object
->type
== OBJT_DEVICE
) ||
2755 (object
->type
== OBJT_MGTDEVICE
))) {
2757 ((object
->type
== OBJT_DEVICE
) ||
2758 (object
->type
== OBJT_MGTDEVICE
)) ? FALSE
: TRUE
;
2759 /* no chain wait needed for vnode/device objects */
2760 vm_object_reference_locked(object
);
2761 switch(current
->maptype
) {
2762 case VM_MAPTYPE_NORMAL
:
2763 vm_object_page_remove(object
,
2765 OFF_TO_IDX(offset
+ size
+ PAGE_MASK
),
2768 case VM_MAPTYPE_VPAGETABLE
:
2769 vm_object_page_remove(object
, 0, 0, clean_only
);
2772 vm_object_deallocate_locked(object
);
2776 vm_object_drop(object
);
2779 lwkt_reltoken(&map
->token
);
2780 vm_map_unlock_read(map
);
2782 return (KERN_SUCCESS
);
2786 * Make the region specified by this entry pageable.
2788 * The vm_map must be exclusively locked.
2791 vm_map_entry_unwire(vm_map_t map
, vm_map_entry_t entry
)
2793 entry
->eflags
&= ~MAP_ENTRY_USER_WIRED
;
2794 entry
->wired_count
= 0;
2795 vm_fault_unwire(map
, entry
);
2799 * Deallocate the given entry from the target map.
2801 * The vm_map must be exclusively locked.
2804 vm_map_entry_delete(vm_map_t map
, vm_map_entry_t entry
, int *countp
)
2806 vm_map_entry_unlink(map
, entry
);
2807 map
->size
-= entry
->end
- entry
->start
;
2809 switch(entry
->maptype
) {
2810 case VM_MAPTYPE_NORMAL
:
2811 case VM_MAPTYPE_VPAGETABLE
:
2812 case VM_MAPTYPE_SUBMAP
:
2813 vm_object_deallocate(entry
->object
.vm_object
);
2815 case VM_MAPTYPE_UKSMAP
:
2822 vm_map_entry_dispose(map
, entry
, countp
);
2826 * Deallocates the given address range from the target map.
2828 * The vm_map must be exclusively locked.
2831 vm_map_delete(vm_map_t map
, vm_offset_t start
, vm_offset_t end
, int *countp
)
2834 vm_map_entry_t entry
;
2835 vm_map_entry_t first_entry
;
2837 ASSERT_VM_MAP_LOCKED(map
);
2838 lwkt_gettoken(&map
->token
);
2841 * Find the start of the region, and clip it. Set entry to point
2842 * at the first record containing the requested address or, if no
2843 * such record exists, the next record with a greater address. The
2844 * loop will run from this point until a record beyond the termination
2845 * address is encountered.
2847 * map->hint must be adjusted to not point to anything we delete,
2848 * so set it to the entry prior to the one being deleted.
2850 * GGG see other GGG comment.
2852 if (vm_map_lookup_entry(map
, start
, &first_entry
)) {
2853 entry
= first_entry
;
2854 vm_map_clip_start(map
, entry
, start
, countp
);
2855 map
->hint
= entry
->prev
; /* possible problem XXX */
2857 map
->hint
= first_entry
; /* possible problem XXX */
2858 entry
= first_entry
->next
;
2862 * If a hole opens up prior to the current first_free then
2863 * adjust first_free. As with map->hint, map->first_free
2864 * cannot be left set to anything we might delete.
2866 if (entry
== &map
->header
) {
2867 map
->first_free
= &map
->header
;
2868 } else if (map
->first_free
->start
>= start
) {
2869 map
->first_free
= entry
->prev
;
2873 * Step through all entries in this region
2875 while ((entry
!= &map
->header
) && (entry
->start
< end
)) {
2876 vm_map_entry_t next
;
2878 vm_pindex_t offidxstart
, offidxend
, count
;
2881 * If we hit an in-transition entry we have to sleep and
2882 * retry. It's easier (and not really slower) to just retry
2883 * since this case occurs so rarely and the hint is already
2884 * pointing at the right place. We have to reset the
2885 * start offset so as not to accidently delete an entry
2886 * another process just created in vacated space.
2888 if (entry
->eflags
& MAP_ENTRY_IN_TRANSITION
) {
2889 entry
->eflags
|= MAP_ENTRY_NEEDS_WAKEUP
;
2890 start
= entry
->start
;
2891 ++mycpu
->gd_cnt
.v_intrans_coll
;
2892 ++mycpu
->gd_cnt
.v_intrans_wait
;
2893 vm_map_transition_wait(map
);
2896 vm_map_clip_end(map
, entry
, end
, countp
);
2902 offidxstart
= OFF_TO_IDX(entry
->offset
);
2903 count
= OFF_TO_IDX(e
- s
);
2905 switch(entry
->maptype
) {
2906 case VM_MAPTYPE_NORMAL
:
2907 case VM_MAPTYPE_VPAGETABLE
:
2908 case VM_MAPTYPE_SUBMAP
:
2909 object
= entry
->object
.vm_object
;
2917 * Unwire before removing addresses from the pmap; otherwise,
2918 * unwiring will put the entries back in the pmap.
2920 if (entry
->wired_count
!= 0)
2921 vm_map_entry_unwire(map
, entry
);
2923 offidxend
= offidxstart
+ count
;
2925 if (object
== &kernel_object
) {
2926 vm_object_hold(object
);
2927 vm_object_page_remove(object
, offidxstart
,
2929 vm_object_drop(object
);
2930 } else if (object
&& object
->type
!= OBJT_DEFAULT
&&
2931 object
->type
!= OBJT_SWAP
) {
2933 * vnode object routines cannot be chain-locked,
2934 * but since we aren't removing pages from the
2935 * object here we can use a shared hold.
2937 vm_object_hold_shared(object
);
2938 pmap_remove(map
->pmap
, s
, e
);
2939 vm_object_drop(object
);
2940 } else if (object
) {
2941 vm_object_hold(object
);
2942 vm_object_chain_acquire(object
, 0);
2943 pmap_remove(map
->pmap
, s
, e
);
2945 if (object
!= NULL
&&
2946 object
->ref_count
!= 1 &&
2947 (object
->flags
& (OBJ_NOSPLIT
|OBJ_ONEMAPPING
)) ==
2949 (object
->type
== OBJT_DEFAULT
||
2950 object
->type
== OBJT_SWAP
)) {
2951 vm_object_collapse(object
, NULL
);
2952 vm_object_page_remove(object
, offidxstart
,
2954 if (object
->type
== OBJT_SWAP
) {
2955 swap_pager_freespace(object
,
2959 if (offidxend
>= object
->size
&&
2960 offidxstart
< object
->size
) {
2961 object
->size
= offidxstart
;
2964 vm_object_chain_release(object
);
2965 vm_object_drop(object
);
2966 } else if (entry
->maptype
== VM_MAPTYPE_UKSMAP
) {
2967 pmap_remove(map
->pmap
, s
, e
);
2971 * Delete the entry (which may delete the object) only after
2972 * removing all pmap entries pointing to its pages.
2973 * (Otherwise, its page frames may be reallocated, and any
2974 * modify bits will be set in the wrong object!)
2976 vm_map_entry_delete(map
, entry
, countp
);
2979 lwkt_reltoken(&map
->token
);
2980 return (KERN_SUCCESS
);
2984 * Remove the given address range from the target map.
2985 * This is the exported form of vm_map_delete.
2990 vm_map_remove(vm_map_t map
, vm_offset_t start
, vm_offset_t end
)
2995 count
= vm_map_entry_reserve(MAP_RESERVE_COUNT
);
2997 VM_MAP_RANGE_CHECK(map
, start
, end
);
2998 result
= vm_map_delete(map
, start
, end
, &count
);
3000 vm_map_entry_release(count
);
3006 * Assert that the target map allows the specified privilege on the
3007 * entire address region given. The entire region must be allocated.
3009 * The caller must specify whether the vm_map is already locked or not.
3012 vm_map_check_protection(vm_map_t map
, vm_offset_t start
, vm_offset_t end
,
3013 vm_prot_t protection
, boolean_t have_lock
)
3015 vm_map_entry_t entry
;
3016 vm_map_entry_t tmp_entry
;
3019 if (have_lock
== FALSE
)
3020 vm_map_lock_read(map
);
3022 if (!vm_map_lookup_entry(map
, start
, &tmp_entry
)) {
3023 if (have_lock
== FALSE
)
3024 vm_map_unlock_read(map
);
3030 while (start
< end
) {
3031 if (entry
== &map
->header
) {
3039 if (start
< entry
->start
) {
3044 * Check protection associated with entry.
3047 if ((entry
->protection
& protection
) != protection
) {
3051 /* go to next entry */
3054 entry
= entry
->next
;
3056 if (have_lock
== FALSE
)
3057 vm_map_unlock_read(map
);
3062 * If appropriate this function shadows the original object with a new object
3063 * and moves the VM pages from the original object to the new object.
3064 * The original object will also be collapsed, if possible.
3066 * We can only do this for normal memory objects with a single mapping, and
3067 * it only makes sense to do it if there are 2 or more refs on the original
3068 * object. i.e. typically a memory object that has been extended into
3069 * multiple vm_map_entry's with non-overlapping ranges.
3071 * This makes it easier to remove unused pages and keeps object inheritance
3072 * from being a negative impact on memory usage.
3074 * On return the (possibly new) entry->object.vm_object will have an
3075 * additional ref on it for the caller to dispose of (usually by cloning
3076 * the vm_map_entry). The additional ref had to be done in this routine
3077 * to avoid racing a collapse. The object's ONEMAPPING flag will also be
3080 * The vm_map must be locked and its token held.
3083 vm_map_split(vm_map_entry_t entry
)
3086 vm_object_t oobject
, nobject
, bobject
;
3089 vm_pindex_t offidxstart
, offidxend
, idx
;
3091 vm_ooffset_t offset
;
3095 * Optimize away object locks for vnode objects. Important exit/exec
3098 * OBJ_ONEMAPPING doesn't apply to vnode objects but clear the flag
3101 oobject
= entry
->object
.vm_object
;
3102 if (oobject
->type
!= OBJT_DEFAULT
&& oobject
->type
!= OBJT_SWAP
) {
3103 vm_object_reference_quick(oobject
);
3104 vm_object_clear_flag(oobject
, OBJ_ONEMAPPING
);
3109 * Setup. Chain lock the original object throughout the entire
3110 * routine to prevent new page faults from occuring.
3112 * XXX can madvise WILLNEED interfere with us too?
3114 vm_object_hold(oobject
);
3115 vm_object_chain_acquire(oobject
, 0);
3118 * Original object cannot be split? Might have also changed state.
3120 if (oobject
->handle
== NULL
|| (oobject
->type
!= OBJT_DEFAULT
&&
3121 oobject
->type
!= OBJT_SWAP
)) {
3122 vm_object_chain_release(oobject
);
3123 vm_object_reference_locked(oobject
);
3124 vm_object_clear_flag(oobject
, OBJ_ONEMAPPING
);
3125 vm_object_drop(oobject
);
3130 * Collapse original object with its backing store as an
3131 * optimization to reduce chain lengths when possible.
3133 * If ref_count <= 1 there aren't other non-overlapping vm_map_entry's
3134 * for oobject, so there's no point collapsing it.
3136 * Then re-check whether the object can be split.
3138 vm_object_collapse(oobject
, NULL
);
3140 if (oobject
->ref_count
<= 1 ||
3141 (oobject
->type
!= OBJT_DEFAULT
&& oobject
->type
!= OBJT_SWAP
) ||
3142 (oobject
->flags
& (OBJ_NOSPLIT
|OBJ_ONEMAPPING
)) != OBJ_ONEMAPPING
) {
3143 vm_object_chain_release(oobject
);
3144 vm_object_reference_locked(oobject
);
3145 vm_object_clear_flag(oobject
, OBJ_ONEMAPPING
);
3146 vm_object_drop(oobject
);
3151 * Acquire the chain lock on the backing object.
3153 * Give bobject an additional ref count for when it will be shadowed
3157 if ((bobject
= oobject
->backing_object
) != NULL
) {
3158 if (bobject
->type
!= OBJT_VNODE
) {
3160 vm_object_hold(bobject
);
3161 vm_object_chain_wait(bobject
, 0);
3162 /* ref for shadowing below */
3163 vm_object_reference_locked(bobject
);
3164 vm_object_chain_acquire(bobject
, 0);
3165 KKASSERT(bobject
->backing_object
== bobject
);
3166 KKASSERT((bobject
->flags
& OBJ_DEAD
) == 0);
3169 * vnodes are not placed on the shadow list but
3170 * they still get another ref for the backing_object
3173 vm_object_reference_quick(bobject
);
3178 * Calculate the object page range and allocate the new object.
3180 offset
= entry
->offset
;
3184 offidxstart
= OFF_TO_IDX(offset
);
3185 offidxend
= offidxstart
+ OFF_TO_IDX(e
- s
);
3186 size
= offidxend
- offidxstart
;
3188 switch(oobject
->type
) {
3190 nobject
= default_pager_alloc(NULL
, IDX_TO_OFF(size
),
3194 nobject
= swap_pager_alloc(NULL
, IDX_TO_OFF(size
),
3203 if (nobject
== NULL
) {
3205 if (useshadowlist
) {
3206 vm_object_chain_release(bobject
);
3207 vm_object_deallocate(bobject
);
3208 vm_object_drop(bobject
);
3210 vm_object_deallocate(bobject
);
3213 vm_object_chain_release(oobject
);
3214 vm_object_reference_locked(oobject
);
3215 vm_object_clear_flag(oobject
, OBJ_ONEMAPPING
);
3216 vm_object_drop(oobject
);
3221 * The new object will replace entry->object.vm_object so it needs
3222 * a second reference (the caller expects an additional ref).
3224 vm_object_hold(nobject
);
3225 vm_object_reference_locked(nobject
);
3226 vm_object_chain_acquire(nobject
, 0);
3229 * nobject shadows bobject (oobject already shadows bobject).
3231 * Adding an object to bobject's shadow list requires refing bobject
3232 * which we did above in the useshadowlist case.
3235 nobject
->backing_object_offset
=
3236 oobject
->backing_object_offset
+ IDX_TO_OFF(offidxstart
);
3237 nobject
->backing_object
= bobject
;
3238 if (useshadowlist
) {
3239 bobject
->shadow_count
++;
3240 bobject
->generation
++;
3241 LIST_INSERT_HEAD(&bobject
->shadow_head
,
3242 nobject
, shadow_list
);
3243 vm_object_clear_flag(bobject
, OBJ_ONEMAPPING
); /*XXX*/
3244 vm_object_chain_release(bobject
);
3245 vm_object_drop(bobject
);
3246 vm_object_set_flag(nobject
, OBJ_ONSHADOW
);
3251 * Move the VM pages from oobject to nobject
3253 for (idx
= 0; idx
< size
; idx
++) {
3256 m
= vm_page_lookup_busy_wait(oobject
, offidxstart
+ idx
,
3262 * We must wait for pending I/O to complete before we can
3265 * We do not have to VM_PROT_NONE the page as mappings should
3266 * not be changed by this operation.
3268 * NOTE: The act of renaming a page updates chaingen for both
3271 vm_page_rename(m
, nobject
, idx
);
3272 /* page automatically made dirty by rename and cache handled */
3273 /* page remains busy */
3276 if (oobject
->type
== OBJT_SWAP
) {
3277 vm_object_pip_add(oobject
, 1);
3279 * copy oobject pages into nobject and destroy unneeded
3280 * pages in shadow object.
3282 swap_pager_copy(oobject
, nobject
, offidxstart
, 0);
3283 vm_object_pip_wakeup(oobject
);
3287 * Wakeup the pages we played with. No spl protection is needed
3288 * for a simple wakeup.
3290 for (idx
= 0; idx
< size
; idx
++) {
3291 m
= vm_page_lookup(nobject
, idx
);
3293 KKASSERT(m
->flags
& PG_BUSY
);
3297 entry
->object
.vm_object
= nobject
;
3298 entry
->offset
= 0LL;
3303 * NOTE: There is no need to remove OBJ_ONEMAPPING from oobject, the
3304 * related pages were moved and are no longer applicable to the
3307 * NOTE: Deallocate oobject (due to its entry->object.vm_object being
3308 * replaced by nobject).
3310 vm_object_chain_release(nobject
);
3311 vm_object_drop(nobject
);
3312 if (bobject
&& useshadowlist
) {
3313 vm_object_chain_release(bobject
);
3314 vm_object_drop(bobject
);
3316 vm_object_chain_release(oobject
);
3317 /*vm_object_clear_flag(oobject, OBJ_ONEMAPPING);*/
3318 vm_object_deallocate_locked(oobject
);
3319 vm_object_drop(oobject
);
3323 * Copies the contents of the source entry to the destination
3324 * entry. The entries *must* be aligned properly.
3326 * The vm_maps must be exclusively locked.
3327 * The vm_map's token must be held.
3329 * Because the maps are locked no faults can be in progress during the
3333 vm_map_copy_entry(vm_map_t src_map
, vm_map_t dst_map
,
3334 vm_map_entry_t src_entry
, vm_map_entry_t dst_entry
)
3336 vm_object_t src_object
;
3338 if (dst_entry
->maptype
== VM_MAPTYPE_SUBMAP
||
3339 dst_entry
->maptype
== VM_MAPTYPE_UKSMAP
)
3341 if (src_entry
->maptype
== VM_MAPTYPE_SUBMAP
||
3342 src_entry
->maptype
== VM_MAPTYPE_UKSMAP
)
3345 if (src_entry
->wired_count
== 0) {
3347 * If the source entry is marked needs_copy, it is already
3350 if ((src_entry
->eflags
& MAP_ENTRY_NEEDS_COPY
) == 0) {
3351 pmap_protect(src_map
->pmap
,
3354 src_entry
->protection
& ~VM_PROT_WRITE
);
3358 * Make a copy of the object.
3360 * The object must be locked prior to checking the object type
3361 * and for the call to vm_object_collapse() and vm_map_split().
3362 * We cannot use *_hold() here because the split code will
3363 * probably try to destroy the object. The lock is a pool
3364 * token and doesn't care.
3366 * We must bump src_map->timestamp when setting
3367 * MAP_ENTRY_NEEDS_COPY to force any concurrent fault
3368 * to retry, otherwise the concurrent fault might improperly
3369 * install a RW pte when its supposed to be a RO(COW) pte.
3370 * This race can occur because a vnode-backed fault may have
3371 * to temporarily release the map lock.
3373 if (src_entry
->object
.vm_object
!= NULL
) {
3374 vm_map_split(src_entry
);
3375 src_object
= src_entry
->object
.vm_object
;
3376 dst_entry
->object
.vm_object
= src_object
;
3377 src_entry
->eflags
|= (MAP_ENTRY_COW
|
3378 MAP_ENTRY_NEEDS_COPY
);
3379 dst_entry
->eflags
|= (MAP_ENTRY_COW
|
3380 MAP_ENTRY_NEEDS_COPY
);
3381 dst_entry
->offset
= src_entry
->offset
;
3382 ++src_map
->timestamp
;
3384 dst_entry
->object
.vm_object
= NULL
;
3385 dst_entry
->offset
= 0;
3388 pmap_copy(dst_map
->pmap
, src_map
->pmap
, dst_entry
->start
,
3389 dst_entry
->end
- dst_entry
->start
, src_entry
->start
);
3392 * Of course, wired down pages can't be set copy-on-write.
3393 * Cause wired pages to be copied into the new map by
3394 * simulating faults (the new pages are pageable)
3396 vm_fault_copy_entry(dst_map
, src_map
, dst_entry
, src_entry
);
3402 * Create a new process vmspace structure and vm_map
3403 * based on those of an existing process. The new map
3404 * is based on the old map, according to the inheritance
3405 * values on the regions in that map.
3407 * The source map must not be locked.
3410 static void vmspace_fork_normal_entry(vm_map_t old_map
, vm_map_t new_map
,
3411 vm_map_entry_t old_entry
, int *countp
);
3412 static void vmspace_fork_uksmap_entry(vm_map_t old_map
, vm_map_t new_map
,
3413 vm_map_entry_t old_entry
, int *countp
);
3416 vmspace_fork(struct vmspace
*vm1
)
3418 struct vmspace
*vm2
;
3419 vm_map_t old_map
= &vm1
->vm_map
;
3421 vm_map_entry_t old_entry
;
3424 lwkt_gettoken(&vm1
->vm_map
.token
);
3425 vm_map_lock(old_map
);
3427 vm2
= vmspace_alloc(old_map
->min_offset
, old_map
->max_offset
);
3428 lwkt_gettoken(&vm2
->vm_map
.token
);
3429 bcopy(&vm1
->vm_startcopy
, &vm2
->vm_startcopy
,
3430 (caddr_t
)&vm1
->vm_endcopy
- (caddr_t
)&vm1
->vm_startcopy
);
3431 new_map
= &vm2
->vm_map
; /* XXX */
3432 new_map
->timestamp
= 1;
3434 vm_map_lock(new_map
);
3437 old_entry
= old_map
->header
.next
;
3438 while (old_entry
!= &old_map
->header
) {
3440 old_entry
= old_entry
->next
;
3443 count
= vm_map_entry_reserve(count
+ MAP_RESERVE_COUNT
);
3445 old_entry
= old_map
->header
.next
;
3446 while (old_entry
!= &old_map
->header
) {
3447 switch(old_entry
->maptype
) {
3448 case VM_MAPTYPE_SUBMAP
:
3449 panic("vm_map_fork: encountered a submap");
3451 case VM_MAPTYPE_UKSMAP
:
3452 vmspace_fork_uksmap_entry(old_map
, new_map
,
3455 case VM_MAPTYPE_NORMAL
:
3456 case VM_MAPTYPE_VPAGETABLE
:
3457 vmspace_fork_normal_entry(old_map
, new_map
,
3461 old_entry
= old_entry
->next
;
3464 new_map
->size
= old_map
->size
;
3465 vm_map_unlock(old_map
);
3466 vm_map_unlock(new_map
);
3467 vm_map_entry_release(count
);
3469 lwkt_reltoken(&vm2
->vm_map
.token
);
3470 lwkt_reltoken(&vm1
->vm_map
.token
);
3477 vmspace_fork_normal_entry(vm_map_t old_map
, vm_map_t new_map
,
3478 vm_map_entry_t old_entry
, int *countp
)
3480 vm_map_entry_t new_entry
;
3483 switch (old_entry
->inheritance
) {
3484 case VM_INHERIT_NONE
:
3486 case VM_INHERIT_SHARE
:
3488 * Clone the entry, creating the shared object if
3491 if (old_entry
->object
.vm_object
== NULL
)
3492 vm_map_entry_allocate_object(old_entry
);
3494 if (old_entry
->eflags
& MAP_ENTRY_NEEDS_COPY
) {
3496 * Shadow a map_entry which needs a copy,
3497 * replacing its object with a new object
3498 * that points to the old one. Ask the
3499 * shadow code to automatically add an
3500 * additional ref. We can't do it afterwords
3501 * because we might race a collapse. The call
3502 * to vm_map_entry_shadow() will also clear
3505 vm_map_entry_shadow(old_entry
, 1);
3506 } else if (old_entry
->object
.vm_object
) {
3508 * We will make a shared copy of the object,
3509 * and must clear OBJ_ONEMAPPING.
3511 * Optimize vnode objects. OBJ_ONEMAPPING
3512 * is non-applicable but clear it anyway,
3513 * and its terminal so we don'th ave to deal
3514 * with chains. Reduces SMP conflicts.
3516 * XXX assert that object.vm_object != NULL
3517 * since we allocate it above.
3519 object
= old_entry
->object
.vm_object
;
3520 if (object
->type
== OBJT_VNODE
) {
3521 vm_object_reference_quick(object
);
3522 vm_object_clear_flag(object
,
3525 vm_object_hold(object
);
3526 vm_object_chain_wait(object
, 0);
3527 vm_object_reference_locked(object
);
3528 vm_object_clear_flag(object
,
3530 vm_object_drop(object
);
3535 * Clone the entry. We've already bumped the ref on
3538 new_entry
= vm_map_entry_create(new_map
, countp
);
3539 *new_entry
= *old_entry
;
3540 new_entry
->eflags
&= ~MAP_ENTRY_USER_WIRED
;
3541 new_entry
->wired_count
= 0;
3544 * Insert the entry into the new map -- we know we're
3545 * inserting at the end of the new map.
3548 vm_map_entry_link(new_map
, new_map
->header
.prev
,
3552 * Update the physical map
3554 pmap_copy(new_map
->pmap
, old_map
->pmap
,
3556 (old_entry
->end
- old_entry
->start
),
3559 case VM_INHERIT_COPY
:
3561 * Clone the entry and link into the map.
3563 new_entry
= vm_map_entry_create(new_map
, countp
);
3564 *new_entry
= *old_entry
;
3565 new_entry
->eflags
&= ~MAP_ENTRY_USER_WIRED
;
3566 new_entry
->wired_count
= 0;
3567 new_entry
->object
.vm_object
= NULL
;
3568 vm_map_entry_link(new_map
, new_map
->header
.prev
,
3570 vm_map_copy_entry(old_map
, new_map
, old_entry
,
3577 * When forking user-kernel shared maps, the map might change in the
3578 * child so do not try to copy the underlying pmap entries.
3582 vmspace_fork_uksmap_entry(vm_map_t old_map
, vm_map_t new_map
,
3583 vm_map_entry_t old_entry
, int *countp
)
3585 vm_map_entry_t new_entry
;
3587 new_entry
= vm_map_entry_create(new_map
, countp
);
3588 *new_entry
= *old_entry
;
3589 new_entry
->eflags
&= ~MAP_ENTRY_USER_WIRED
;
3590 new_entry
->wired_count
= 0;
3591 vm_map_entry_link(new_map
, new_map
->header
.prev
,
3596 * Create an auto-grow stack entry
3601 vm_map_stack (vm_map_t map
, vm_offset_t addrbos
, vm_size_t max_ssize
,
3602 int flags
, vm_prot_t prot
, vm_prot_t max
, int cow
)
3604 vm_map_entry_t prev_entry
;
3605 vm_map_entry_t new_stack_entry
;
3606 vm_size_t init_ssize
;
3609 vm_offset_t tmpaddr
;
3611 cow
|= MAP_IS_STACK
;
3613 if (max_ssize
< sgrowsiz
)
3614 init_ssize
= max_ssize
;
3616 init_ssize
= sgrowsiz
;
3618 count
= vm_map_entry_reserve(MAP_RESERVE_COUNT
);
3622 * Find space for the mapping
3624 if ((flags
& (MAP_FIXED
| MAP_TRYFIXED
)) == 0) {
3625 if (vm_map_findspace(map
, addrbos
, max_ssize
, 1,
3628 vm_map_entry_release(count
);
3629 return (KERN_NO_SPACE
);
3634 /* If addr is already mapped, no go */
3635 if (vm_map_lookup_entry(map
, addrbos
, &prev_entry
)) {
3637 vm_map_entry_release(count
);
3638 return (KERN_NO_SPACE
);
3642 /* XXX already handled by kern_mmap() */
3643 /* If we would blow our VMEM resource limit, no go */
3644 if (map
->size
+ init_ssize
>
3645 curproc
->p_rlimit
[RLIMIT_VMEM
].rlim_cur
) {
3647 vm_map_entry_release(count
);
3648 return (KERN_NO_SPACE
);
3653 * If we can't accomodate max_ssize in the current mapping,
3654 * no go. However, we need to be aware that subsequent user
3655 * mappings might map into the space we have reserved for
3656 * stack, and currently this space is not protected.
3658 * Hopefully we will at least detect this condition
3659 * when we try to grow the stack.
3661 if ((prev_entry
->next
!= &map
->header
) &&
3662 (prev_entry
->next
->start
< addrbos
+ max_ssize
)) {
3664 vm_map_entry_release(count
);
3665 return (KERN_NO_SPACE
);
3669 * We initially map a stack of only init_ssize. We will
3670 * grow as needed later. Since this is to be a grow
3671 * down stack, we map at the top of the range.
3673 * Note: we would normally expect prot and max to be
3674 * VM_PROT_ALL, and cow to be 0. Possibly we should
3675 * eliminate these as input parameters, and just
3676 * pass these values here in the insert call.
3678 rv
= vm_map_insert(map
, &count
, NULL
, NULL
,
3679 0, addrbos
+ max_ssize
- init_ssize
,
3680 addrbos
+ max_ssize
,
3682 VM_SUBSYS_STACK
, prot
, max
, cow
);
3684 /* Now set the avail_ssize amount */
3685 if (rv
== KERN_SUCCESS
) {
3686 if (prev_entry
!= &map
->header
)
3687 vm_map_clip_end(map
, prev_entry
, addrbos
+ max_ssize
- init_ssize
, &count
);
3688 new_stack_entry
= prev_entry
->next
;
3689 if (new_stack_entry
->end
!= addrbos
+ max_ssize
||
3690 new_stack_entry
->start
!= addrbos
+ max_ssize
- init_ssize
)
3691 panic ("Bad entry start/end for new stack entry");
3693 new_stack_entry
->aux
.avail_ssize
= max_ssize
- init_ssize
;
3697 vm_map_entry_release(count
);
3702 * Attempts to grow a vm stack entry. Returns KERN_SUCCESS if the
3703 * desired address is already mapped, or if we successfully grow
3704 * the stack. Also returns KERN_SUCCESS if addr is outside the
3705 * stack range (this is strange, but preserves compatibility with
3706 * the grow function in vm_machdep.c).
3711 vm_map_growstack (struct proc
*p
, vm_offset_t addr
)
3713 vm_map_entry_t prev_entry
;
3714 vm_map_entry_t stack_entry
;
3715 vm_map_entry_t new_stack_entry
;
3716 struct vmspace
*vm
= p
->p_vmspace
;
3717 vm_map_t map
= &vm
->vm_map
;
3720 int rv
= KERN_SUCCESS
;
3722 int use_read_lock
= 1;
3725 count
= vm_map_entry_reserve(MAP_RESERVE_COUNT
);
3728 vm_map_lock_read(map
);
3732 /* If addr is already in the entry range, no need to grow.*/
3733 if (vm_map_lookup_entry(map
, addr
, &prev_entry
))
3736 if ((stack_entry
= prev_entry
->next
) == &map
->header
)
3738 if (prev_entry
== &map
->header
)
3739 end
= stack_entry
->start
- stack_entry
->aux
.avail_ssize
;
3741 end
= prev_entry
->end
;
3744 * This next test mimics the old grow function in vm_machdep.c.
3745 * It really doesn't quite make sense, but we do it anyway
3746 * for compatibility.
3748 * If not growable stack, return success. This signals the
3749 * caller to proceed as he would normally with normal vm.
3751 if (stack_entry
->aux
.avail_ssize
< 1 ||
3752 addr
>= stack_entry
->start
||
3753 addr
< stack_entry
->start
- stack_entry
->aux
.avail_ssize
) {
3757 /* Find the minimum grow amount */
3758 grow_amount
= roundup (stack_entry
->start
- addr
, PAGE_SIZE
);
3759 if (grow_amount
> stack_entry
->aux
.avail_ssize
) {
3765 * If there is no longer enough space between the entries
3766 * nogo, and adjust the available space. Note: this
3767 * should only happen if the user has mapped into the
3768 * stack area after the stack was created, and is
3769 * probably an error.
3771 * This also effectively destroys any guard page the user
3772 * might have intended by limiting the stack size.
3774 if (grow_amount
> stack_entry
->start
- end
) {
3775 if (use_read_lock
&& vm_map_lock_upgrade(map
)) {
3781 stack_entry
->aux
.avail_ssize
= stack_entry
->start
- end
;
3786 is_procstack
= addr
>= (vm_offset_t
)vm
->vm_maxsaddr
;
3788 /* If this is the main process stack, see if we're over the
3791 if (is_procstack
&& (ctob(vm
->vm_ssize
) + grow_amount
>
3792 p
->p_rlimit
[RLIMIT_STACK
].rlim_cur
)) {
3797 /* Round up the grow amount modulo SGROWSIZ */
3798 grow_amount
= roundup (grow_amount
, sgrowsiz
);
3799 if (grow_amount
> stack_entry
->aux
.avail_ssize
) {
3800 grow_amount
= stack_entry
->aux
.avail_ssize
;
3802 if (is_procstack
&& (ctob(vm
->vm_ssize
) + grow_amount
>
3803 p
->p_rlimit
[RLIMIT_STACK
].rlim_cur
)) {
3804 grow_amount
= p
->p_rlimit
[RLIMIT_STACK
].rlim_cur
-
3808 /* If we would blow our VMEM resource limit, no go */
3809 if (map
->size
+ grow_amount
> p
->p_rlimit
[RLIMIT_VMEM
].rlim_cur
) {
3814 if (use_read_lock
&& vm_map_lock_upgrade(map
)) {
3821 /* Get the preliminary new entry start value */
3822 addr
= stack_entry
->start
- grow_amount
;
3824 /* If this puts us into the previous entry, cut back our growth
3825 * to the available space. Also, see the note above.
3828 stack_entry
->aux
.avail_ssize
= stack_entry
->start
- end
;
3832 rv
= vm_map_insert(map
, &count
, NULL
, NULL
,
3833 0, addr
, stack_entry
->start
,
3835 VM_SUBSYS_STACK
, VM_PROT_ALL
, VM_PROT_ALL
, 0);
3837 /* Adjust the available stack space by the amount we grew. */
3838 if (rv
== KERN_SUCCESS
) {
3839 if (prev_entry
!= &map
->header
)
3840 vm_map_clip_end(map
, prev_entry
, addr
, &count
);
3841 new_stack_entry
= prev_entry
->next
;
3842 if (new_stack_entry
->end
!= stack_entry
->start
||
3843 new_stack_entry
->start
!= addr
)
3844 panic ("Bad stack grow start/end in new stack entry");
3846 new_stack_entry
->aux
.avail_ssize
=
3847 stack_entry
->aux
.avail_ssize
-
3848 (new_stack_entry
->end
- new_stack_entry
->start
);
3850 vm
->vm_ssize
+= btoc(new_stack_entry
->end
-
3851 new_stack_entry
->start
);
3854 if (map
->flags
& MAP_WIREFUTURE
)
3855 vm_map_unwire(map
, new_stack_entry
->start
,
3856 new_stack_entry
->end
, FALSE
);
3861 vm_map_unlock_read(map
);
3864 vm_map_entry_release(count
);
3869 * Unshare the specified VM space for exec. If other processes are
3870 * mapped to it, then create a new one. The new vmspace is null.
3875 vmspace_exec(struct proc
*p
, struct vmspace
*vmcopy
)
3877 struct vmspace
*oldvmspace
= p
->p_vmspace
;
3878 struct vmspace
*newvmspace
;
3879 vm_map_t map
= &p
->p_vmspace
->vm_map
;
3882 * If we are execing a resident vmspace we fork it, otherwise
3883 * we create a new vmspace. Note that exitingcnt is not
3884 * copied to the new vmspace.
3886 lwkt_gettoken(&oldvmspace
->vm_map
.token
);
3888 newvmspace
= vmspace_fork(vmcopy
);
3889 lwkt_gettoken(&newvmspace
->vm_map
.token
);
3891 newvmspace
= vmspace_alloc(map
->min_offset
, map
->max_offset
);
3892 lwkt_gettoken(&newvmspace
->vm_map
.token
);
3893 bcopy(&oldvmspace
->vm_startcopy
, &newvmspace
->vm_startcopy
,
3894 (caddr_t
)&oldvmspace
->vm_endcopy
-
3895 (caddr_t
)&oldvmspace
->vm_startcopy
);
3899 * Finish initializing the vmspace before assigning it
3900 * to the process. The vmspace will become the current vmspace
3903 pmap_pinit2(vmspace_pmap(newvmspace
));
3904 pmap_replacevm(p
, newvmspace
, 0);
3905 lwkt_reltoken(&newvmspace
->vm_map
.token
);
3906 lwkt_reltoken(&oldvmspace
->vm_map
.token
);
3907 vmspace_rel(oldvmspace
);
3911 * Unshare the specified VM space for forcing COW. This
3912 * is called by rfork, for the (RFMEM|RFPROC) == 0 case.
3915 vmspace_unshare(struct proc
*p
)
3917 struct vmspace
*oldvmspace
= p
->p_vmspace
;
3918 struct vmspace
*newvmspace
;
3920 lwkt_gettoken(&oldvmspace
->vm_map
.token
);
3921 if (vmspace_getrefs(oldvmspace
) == 1) {
3922 lwkt_reltoken(&oldvmspace
->vm_map
.token
);
3925 newvmspace
= vmspace_fork(oldvmspace
);
3926 lwkt_gettoken(&newvmspace
->vm_map
.token
);
3927 pmap_pinit2(vmspace_pmap(newvmspace
));
3928 pmap_replacevm(p
, newvmspace
, 0);
3929 lwkt_reltoken(&newvmspace
->vm_map
.token
);
3930 lwkt_reltoken(&oldvmspace
->vm_map
.token
);
3931 vmspace_rel(oldvmspace
);
3935 * vm_map_hint: return the beginning of the best area suitable for
3936 * creating a new mapping with "prot" protection.
3941 vm_map_hint(struct proc
*p
, vm_offset_t addr
, vm_prot_t prot
)
3943 struct vmspace
*vms
= p
->p_vmspace
;
3945 if (!randomize_mmap
|| addr
!= 0) {
3947 * Set a reasonable start point for the hint if it was
3948 * not specified or if it falls within the heap space.
3949 * Hinted mmap()s do not allocate out of the heap space.
3952 (addr
>= round_page((vm_offset_t
)vms
->vm_taddr
) &&
3953 addr
< round_page((vm_offset_t
)vms
->vm_daddr
+ maxdsiz
))) {
3954 addr
= round_page((vm_offset_t
)vms
->vm_daddr
+ maxdsiz
);
3959 addr
= (vm_offset_t
)vms
->vm_daddr
+ MAXDSIZ
;
3960 addr
+= karc4random() & (MIN((256 * 1024 * 1024), MAXDSIZ
) - 1);
3962 return (round_page(addr
));
3966 * Finds the VM object, offset, and protection for a given virtual address
3967 * in the specified map, assuming a page fault of the type specified.
3969 * Leaves the map in question locked for read; return values are guaranteed
3970 * until a vm_map_lookup_done call is performed. Note that the map argument
3971 * is in/out; the returned map must be used in the call to vm_map_lookup_done.
3973 * A handle (out_entry) is returned for use in vm_map_lookup_done, to make
3976 * If a lookup is requested with "write protection" specified, the map may
3977 * be changed to perform virtual copying operations, although the data
3978 * referenced will remain the same.
3983 vm_map_lookup(vm_map_t
*var_map
, /* IN/OUT */
3985 vm_prot_t fault_typea
,
3986 vm_map_entry_t
*out_entry
, /* OUT */
3987 vm_object_t
*object
, /* OUT */
3988 vm_pindex_t
*pindex
, /* OUT */
3989 vm_prot_t
*out_prot
, /* OUT */
3990 boolean_t
*wired
) /* OUT */
3992 vm_map_entry_t entry
;
3993 vm_map_t map
= *var_map
;
3995 vm_prot_t fault_type
= fault_typea
;
3996 int use_read_lock
= 1;
3997 int rv
= KERN_SUCCESS
;
4001 vm_map_lock_read(map
);
4006 * If the map has an interesting hint, try it before calling full
4007 * blown lookup routine.
4014 if ((entry
== &map
->header
) ||
4015 (vaddr
< entry
->start
) || (vaddr
>= entry
->end
)) {
4016 vm_map_entry_t tmp_entry
;
4019 * Entry was either not a valid hint, or the vaddr was not
4020 * contained in the entry, so do a full lookup.
4022 if (!vm_map_lookup_entry(map
, vaddr
, &tmp_entry
)) {
4023 rv
= KERN_INVALID_ADDRESS
;
4034 if (entry
->maptype
== VM_MAPTYPE_SUBMAP
) {
4035 vm_map_t old_map
= map
;
4037 *var_map
= map
= entry
->object
.sub_map
;
4039 vm_map_unlock_read(old_map
);
4041 vm_map_unlock(old_map
);
4047 * Check whether this task is allowed to have this page.
4048 * Note the special case for MAP_ENTRY_COW
4049 * pages with an override. This is to implement a forced
4050 * COW for debuggers.
4053 if (fault_type
& VM_PROT_OVERRIDE_WRITE
)
4054 prot
= entry
->max_protection
;
4056 prot
= entry
->protection
;
4058 fault_type
&= (VM_PROT_READ
|VM_PROT_WRITE
|VM_PROT_EXECUTE
);
4059 if ((fault_type
& prot
) != fault_type
) {
4060 rv
= KERN_PROTECTION_FAILURE
;
4064 if ((entry
->eflags
& MAP_ENTRY_USER_WIRED
) &&
4065 (entry
->eflags
& MAP_ENTRY_COW
) &&
4066 (fault_type
& VM_PROT_WRITE
) &&
4067 (fault_typea
& VM_PROT_OVERRIDE_WRITE
) == 0) {
4068 rv
= KERN_PROTECTION_FAILURE
;
4073 * If this page is not pageable, we have to get it for all possible
4076 *wired
= (entry
->wired_count
!= 0);
4078 prot
= fault_type
= entry
->protection
;
4081 * Virtual page tables may need to update the accessed (A) bit
4082 * in a page table entry. Upgrade the fault to a write fault for
4083 * that case if the map will support it. If the map does not support
4084 * it the page table entry simply will not be updated.
4086 if (entry
->maptype
== VM_MAPTYPE_VPAGETABLE
) {
4087 if (prot
& VM_PROT_WRITE
)
4088 fault_type
|= VM_PROT_WRITE
;
4091 if (curthread
->td_lwp
&& curthread
->td_lwp
->lwp_vmspace
&&
4092 pmap_emulate_ad_bits(&curthread
->td_lwp
->lwp_vmspace
->vm_pmap
)) {
4093 if ((prot
& VM_PROT_WRITE
) == 0)
4094 fault_type
|= VM_PROT_WRITE
;
4098 * Only NORMAL and VPAGETABLE maps are object-based. UKSMAPs are not.
4100 if (entry
->maptype
!= VM_MAPTYPE_NORMAL
&&
4101 entry
->maptype
!= VM_MAPTYPE_VPAGETABLE
) {
4107 * If the entry was copy-on-write, we either ...
4109 if (entry
->eflags
& MAP_ENTRY_NEEDS_COPY
) {
4111 * If we want to write the page, we may as well handle that
4112 * now since we've got the map locked.
4114 * If we don't need to write the page, we just demote the
4115 * permissions allowed.
4118 if (fault_type
& VM_PROT_WRITE
) {
4120 * Not allowed if TDF_NOFAULT is set as the shadowing
4121 * operation can deadlock against the faulting
4122 * function due to the copy-on-write.
4124 if (curthread
->td_flags
& TDF_NOFAULT
) {
4125 rv
= KERN_FAILURE_NOFAULT
;
4130 * Make a new object, and place it in the object
4131 * chain. Note that no new references have appeared
4132 * -- one just moved from the map to the new
4136 if (use_read_lock
&& vm_map_lock_upgrade(map
)) {
4143 vm_map_entry_shadow(entry
, 0);
4146 * We're attempting to read a copy-on-write page --
4147 * don't allow writes.
4150 prot
&= ~VM_PROT_WRITE
;
4155 * Create an object if necessary.
4157 if (entry
->object
.vm_object
== NULL
&& !map
->system_map
) {
4158 if (use_read_lock
&& vm_map_lock_upgrade(map
)) {
4164 vm_map_entry_allocate_object(entry
);
4168 * Return the object/offset from this entry. If the entry was
4169 * copy-on-write or empty, it has been fixed up.
4171 *object
= entry
->object
.vm_object
;
4174 *pindex
= OFF_TO_IDX((vaddr
- entry
->start
) + entry
->offset
);
4177 * Return whether this is the only map sharing this data. On
4178 * success we return with a read lock held on the map. On failure
4179 * we return with the map unlocked.
4183 if (rv
== KERN_SUCCESS
) {
4184 if (use_read_lock
== 0)
4185 vm_map_lock_downgrade(map
);
4186 } else if (use_read_lock
) {
4187 vm_map_unlock_read(map
);
4195 * Releases locks acquired by a vm_map_lookup()
4196 * (according to the handle returned by that lookup).
4198 * No other requirements.
4201 vm_map_lookup_done(vm_map_t map
, vm_map_entry_t entry
, int count
)
4204 * Unlock the main-level map
4206 vm_map_unlock_read(map
);
4208 vm_map_entry_release(count
);
4211 #include "opt_ddb.h"
4213 #include <sys/kernel.h>
4215 #include <ddb/ddb.h>
4220 DB_SHOW_COMMAND(map
, vm_map_print
)
4223 /* XXX convert args. */
4224 vm_map_t map
= (vm_map_t
)addr
;
4225 boolean_t full
= have_addr
;
4227 vm_map_entry_t entry
;
4229 db_iprintf("Task map %p: pmap=%p, nentries=%d, version=%u\n",
4231 (void *)map
->pmap
, map
->nentries
, map
->timestamp
);
4234 if (!full
&& db_indent
)
4238 for (entry
= map
->header
.next
; entry
!= &map
->header
;
4239 entry
= entry
->next
) {
4240 db_iprintf("map entry %p: start=%p, end=%p\n",
4241 (void *)entry
, (void *)entry
->start
, (void *)entry
->end
);
4244 static char *inheritance_name
[4] =
4245 {"share", "copy", "none", "donate_copy"};
4247 db_iprintf(" prot=%x/%x/%s",
4249 entry
->max_protection
,
4250 inheritance_name
[(int)(unsigned char)entry
->inheritance
]);
4251 if (entry
->wired_count
!= 0)
4252 db_printf(", wired");
4254 switch(entry
->maptype
) {
4255 case VM_MAPTYPE_SUBMAP
:
4256 /* XXX no %qd in kernel. Truncate entry->offset. */
4257 db_printf(", share=%p, offset=0x%lx\n",
4258 (void *)entry
->object
.sub_map
,
4259 (long)entry
->offset
);
4261 if ((entry
->prev
== &map
->header
) ||
4262 (entry
->prev
->object
.sub_map
!=
4263 entry
->object
.sub_map
)) {
4265 vm_map_print((db_expr_t
)(intptr_t)
4266 entry
->object
.sub_map
,
4271 case VM_MAPTYPE_NORMAL
:
4272 case VM_MAPTYPE_VPAGETABLE
:
4273 /* XXX no %qd in kernel. Truncate entry->offset. */
4274 db_printf(", object=%p, offset=0x%lx",
4275 (void *)entry
->object
.vm_object
,
4276 (long)entry
->offset
);
4277 if (entry
->eflags
& MAP_ENTRY_COW
)
4278 db_printf(", copy (%s)",
4279 (entry
->eflags
& MAP_ENTRY_NEEDS_COPY
) ? "needed" : "done");
4283 if ((entry
->prev
== &map
->header
) ||
4284 (entry
->prev
->object
.vm_object
!=
4285 entry
->object
.vm_object
)) {
4287 vm_object_print((db_expr_t
)(intptr_t)
4288 entry
->object
.vm_object
,
4294 case VM_MAPTYPE_UKSMAP
:
4295 db_printf(", uksmap=%p, offset=0x%lx",
4296 (void *)entry
->object
.uksmap
,
4297 (long)entry
->offset
);
4298 if (entry
->eflags
& MAP_ENTRY_COW
)
4299 db_printf(", copy (%s)",
4300 (entry
->eflags
& MAP_ENTRY_NEEDS_COPY
) ? "needed" : "done");
4316 DB_SHOW_COMMAND(procvm
, procvm
)
4321 p
= (struct proc
*) addr
;
4326 db_printf("p = %p, vmspace = %p, map = %p, pmap = %p\n",
4327 (void *)p
, (void *)p
->p_vmspace
, (void *)&p
->p_vmspace
->vm_map
,
4328 (void *)vmspace_pmap(p
->p_vmspace
));
4330 vm_map_print((db_expr_t
)(intptr_t)&p
->p_vmspace
->vm_map
, 1, 0, NULL
);