4 * Copyright (c) 1991, 1993
5 * The Regents of the University of California. All rights reserved.
7 * This code is derived from software contributed to Berkeley by
8 * The Mach Operating System project at Carnegie-Mellon University.
10 * Redistribution and use in source and binary forms, with or without
11 * modification, are permitted provided that the following conditions
13 * 1. Redistributions of source code must retain the above copyright
14 * notice, this list of conditions and the following disclaimer.
15 * 2. Redistributions in binary form must reproduce the above copyright
16 * notice, this list of conditions and the following disclaimer in the
17 * documentation and/or other materials provided with the distribution.
18 * 3. Neither the name of the University nor the names of its contributors
19 * may be used to endorse or promote products derived from this software
20 * without specific prior written permission.
22 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
23 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
24 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
25 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
26 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
27 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
28 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
29 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
30 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
31 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
34 * from: @(#)vm_map.c 8.3 (Berkeley) 1/12/94
37 * Copyright (c) 1987, 1990 Carnegie-Mellon University.
38 * All rights reserved.
40 * Authors: Avadis Tevanian, Jr., Michael Wayne Young
42 * Permission to use, copy, modify and distribute this software and
43 * its documentation is hereby granted, provided that both the copyright
44 * notice and this permission notice appear in all copies of the
45 * software, derivative works or modified versions, and any portions
46 * thereof, and that both notices appear in supporting documentation.
48 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
49 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
50 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
52 * Carnegie Mellon requests users of this software to return to
54 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU
55 * School of Computer Science
56 * Carnegie Mellon University
57 * Pittsburgh PA 15213-3890
59 * any improvements or extensions that they make and grant Carnegie the
60 * rights to redistribute these changes.
62 * $FreeBSD: src/sys/vm/vm_map.c,v 1.187.2.19 2003/05/27 00:47:02 alc Exp $
66 * Virtual memory mapping module.
69 #include <sys/param.h>
70 #include <sys/systm.h>
71 #include <sys/kernel.h>
73 #include <sys/serialize.h>
75 #include <sys/vmmeter.h>
77 #include <sys/vnode.h>
78 #include <sys/resourcevar.h>
81 #include <sys/malloc.h>
82 #include <sys/objcache.h>
85 #include <vm/vm_param.h>
87 #include <vm/vm_map.h>
88 #include <vm/vm_page.h>
89 #include <vm/vm_object.h>
90 #include <vm/vm_pager.h>
91 #include <vm/vm_kern.h>
92 #include <vm/vm_extern.h>
93 #include <vm/swap_pager.h>
94 #include <vm/vm_zone.h>
96 #include <sys/thread2.h>
97 #include <sys/random.h>
98 #include <sys/sysctl.h>
101 * Virtual memory maps provide for the mapping, protection, and sharing
102 * of virtual memory objects. In addition, this module provides for an
103 * efficient virtual copy of memory from one map to another.
105 * Synchronization is required prior to most operations.
107 * Maps consist of an ordered doubly-linked list of simple entries.
108 * A hint and a RB tree is used to speed-up lookups.
110 * Callers looking to modify maps specify start/end addresses which cause
111 * the related map entry to be clipped if necessary, and then later
112 * recombined if the pieces remained compatible.
114 * Virtual copy operations are performed by copying VM object references
115 * from one map to another, and then marking both regions as copy-on-write.
117 static boolean_t
vmspace_ctor(void *obj
, void *privdata
, int ocflags
);
118 static void vmspace_dtor(void *obj
, void *privdata
);
119 static void vmspace_terminate(struct vmspace
*vm
, int final
);
121 MALLOC_DEFINE(M_VMSPACE
, "vmspace", "vmspace objcache backingstore");
122 static struct objcache
*vmspace_cache
;
125 * per-cpu page table cross mappings are initialized in early boot
126 * and might require a considerable number of vm_map_entry structures.
128 #define MAPENTRYBSP_CACHE (MAXCPU+1)
129 #define MAPENTRYAP_CACHE 8
131 static struct vm_zone mapentzone_store
;
132 static vm_zone_t mapentzone
;
133 static struct vm_object mapentobj
;
135 static struct vm_map_entry map_entry_init
[MAX_MAPENT
];
136 static struct vm_map_entry cpu_map_entry_init_bsp
[MAPENTRYBSP_CACHE
];
137 static struct vm_map_entry cpu_map_entry_init_ap
[MAXCPU
][MAPENTRYAP_CACHE
];
139 static int randomize_mmap
;
140 SYSCTL_INT(_vm
, OID_AUTO
, randomize_mmap
, CTLFLAG_RW
, &randomize_mmap
, 0,
141 "Randomize mmap offsets");
142 static int vm_map_relock_enable
= 1;
143 SYSCTL_INT(_vm
, OID_AUTO
, map_relock_enable
, CTLFLAG_RW
,
144 &vm_map_relock_enable
, 0, "Randomize mmap offsets");
146 static void vmspace_drop_notoken(struct vmspace
*vm
);
147 static void vm_map_entry_shadow(vm_map_entry_t entry
, int addref
);
148 static vm_map_entry_t
vm_map_entry_create(vm_map_t map
, int *);
149 static void vm_map_entry_dispose (vm_map_t map
, vm_map_entry_t entry
, int *);
150 static void _vm_map_clip_end (vm_map_t
, vm_map_entry_t
, vm_offset_t
, int *);
151 static void _vm_map_clip_start (vm_map_t
, vm_map_entry_t
, vm_offset_t
, int *);
152 static void vm_map_entry_delete (vm_map_t
, vm_map_entry_t
, int *);
153 static void vm_map_entry_unwire (vm_map_t
, vm_map_entry_t
);
154 static void vm_map_copy_entry (vm_map_t
, vm_map_t
, vm_map_entry_t
,
156 static void vm_map_unclip_range (vm_map_t map
, vm_map_entry_t start_entry
, vm_offset_t start
, vm_offset_t end
, int *count
, int flags
);
159 * Initialize the vm_map module. Must be called before any other vm_map
162 * Map and entry structures are allocated from the general purpose
163 * memory pool with some exceptions:
165 * - The kernel map is allocated statically.
166 * - Initial kernel map entries are allocated out of a static pool.
167 * - We must set ZONE_SPECIAL here or the early boot code can get
168 * stuck if there are >63 cores.
170 * These restrictions are necessary since malloc() uses the
171 * maps and requires map entries.
173 * Called from the low level boot code only.
178 mapentzone
= &mapentzone_store
;
179 zbootinit(mapentzone
, "MAP ENTRY", sizeof (struct vm_map_entry
),
180 map_entry_init
, MAX_MAPENT
);
181 mapentzone_store
.zflags
|= ZONE_SPECIAL
;
185 * Called prior to any vmspace allocations.
187 * Called from the low level boot code only.
192 vmspace_cache
= objcache_create_mbacked(M_VMSPACE
,
193 sizeof(struct vmspace
),
195 vmspace_ctor
, vmspace_dtor
,
197 zinitna(mapentzone
, &mapentobj
, NULL
, 0, 0,
198 ZONE_USE_RESERVE
| ZONE_SPECIAL
);
204 * objcache support. We leave the pmap root cached as long as possible
205 * for performance reasons.
209 vmspace_ctor(void *obj
, void *privdata
, int ocflags
)
211 struct vmspace
*vm
= obj
;
213 bzero(vm
, sizeof(*vm
));
214 vm
->vm_refcnt
= VM_REF_DELETED
;
221 vmspace_dtor(void *obj
, void *privdata
)
223 struct vmspace
*vm
= obj
;
225 KKASSERT(vm
->vm_refcnt
== VM_REF_DELETED
);
226 pmap_puninit(vmspace_pmap(vm
));
230 * Red black tree functions
232 * The caller must hold the related map lock.
234 static int rb_vm_map_compare(vm_map_entry_t a
, vm_map_entry_t b
);
235 RB_GENERATE(vm_map_rb_tree
, vm_map_entry
, rb_entry
, rb_vm_map_compare
);
237 /* a->start is address, and the only field has to be initialized */
239 rb_vm_map_compare(vm_map_entry_t a
, vm_map_entry_t b
)
241 if (a
->start
< b
->start
)
243 else if (a
->start
> b
->start
)
249 * Initialize vmspace ref/hold counts vmspace0. There is a holdcnt for
253 vmspace_initrefs(struct vmspace
*vm
)
260 * Allocate a vmspace structure, including a vm_map and pmap.
261 * Initialize numerous fields. While the initial allocation is zerod,
262 * subsequence reuse from the objcache leaves elements of the structure
263 * intact (particularly the pmap), so portions must be zerod.
265 * Returns a referenced vmspace.
270 vmspace_alloc(vm_offset_t min
, vm_offset_t max
)
274 vm
= objcache_get(vmspace_cache
, M_WAITOK
);
276 bzero(&vm
->vm_startcopy
,
277 (char *)&vm
->vm_endcopy
- (char *)&vm
->vm_startcopy
);
278 vm_map_init(&vm
->vm_map
, min
, max
, NULL
); /* initializes token */
281 * NOTE: hold to acquires token for safety.
283 * On return vmspace is referenced (refs=1, hold=1). That is,
284 * each refcnt also has a holdcnt. There can be additional holds
285 * (holdcnt) above and beyond the refcnt. Finalization is handled in
286 * two stages, one on refs 1->0, and the the second on hold 1->0.
288 KKASSERT(vm
->vm_holdcnt
== 0);
289 KKASSERT(vm
->vm_refcnt
== VM_REF_DELETED
);
290 vmspace_initrefs(vm
);
292 pmap_pinit(vmspace_pmap(vm
)); /* (some fields reused) */
293 vm
->vm_map
.pmap
= vmspace_pmap(vm
); /* XXX */
296 cpu_vmspace_alloc(vm
);
303 * NOTE: Can return 0 if the vmspace is exiting.
306 vmspace_getrefs(struct vmspace
*vm
)
308 return ((int)(vm
->vm_refcnt
& ~VM_REF_DELETED
));
312 vmspace_hold(struct vmspace
*vm
)
314 atomic_add_int(&vm
->vm_holdcnt
, 1);
315 lwkt_gettoken(&vm
->vm_map
.token
);
319 vmspace_drop(struct vmspace
*vm
)
321 lwkt_reltoken(&vm
->vm_map
.token
);
322 vmspace_drop_notoken(vm
);
326 vmspace_drop_notoken(struct vmspace
*vm
)
328 if (atomic_fetchadd_int(&vm
->vm_holdcnt
, -1) == 1) {
329 if (vm
->vm_refcnt
& VM_REF_DELETED
)
330 vmspace_terminate(vm
, 1);
335 * A vmspace object must not be in a terminated state to be able to obtain
336 * additional refs on it.
338 * These are official references to the vmspace, the count is used to check
339 * for vmspace sharing. Foreign accessors should use 'hold' and not 'ref'.
341 * XXX we need to combine hold & ref together into one 64-bit field to allow
342 * holds to prevent stage-1 termination.
345 vmspace_ref(struct vmspace
*vm
)
349 n
= atomic_fetchadd_int(&vm
->vm_refcnt
, 1);
350 KKASSERT((n
& VM_REF_DELETED
) == 0);
354 * Release a ref on the vmspace. On the 1->0 transition we do stage-1
355 * termination of the vmspace. Then, on the final drop of the hold we
356 * will do stage-2 final termination.
359 vmspace_rel(struct vmspace
*vm
)
366 KKASSERT((int)n
> 0); /* at least one ref & not deleted */
370 * We must have a hold first to interlock the
371 * VM_REF_DELETED check that the drop tests.
373 atomic_add_int(&vm
->vm_holdcnt
, 1);
374 if (atomic_cmpset_int(&vm
->vm_refcnt
, n
,
376 vmspace_terminate(vm
, 0);
377 vmspace_drop_notoken(vm
);
380 vmspace_drop_notoken(vm
);
381 } else if (atomic_cmpset_int(&vm
->vm_refcnt
, n
, n
- 1)) {
388 * This is called during exit indicating that the vmspace is no
389 * longer in used by an exiting process, but the process has not yet
392 * We drop refs, allowing for stage-1 termination, but maintain a holdcnt
393 * to prevent stage-2 until the process is reaped. Note hte order of
394 * operation, we must hold first.
399 vmspace_relexit(struct vmspace
*vm
)
401 atomic_add_int(&vm
->vm_holdcnt
, 1);
406 * Called during reap to disconnect the remainder of the vmspace from
407 * the process. On the hold drop the vmspace termination is finalized.
412 vmspace_exitfree(struct proc
*p
)
418 vmspace_drop_notoken(vm
);
422 * Called in two cases:
424 * (1) When the last refcnt is dropped and the vmspace becomes inactive,
425 * called with final == 0. refcnt will be (u_int)-1 at this point,
426 * and holdcnt will still be non-zero.
428 * (2) When holdcnt becomes 0, called with final == 1. There should no
429 * longer be anyone with access to the vmspace.
431 * VMSPACE_EXIT1 flags the primary deactivation
432 * VMSPACE_EXIT2 flags the last reap
435 vmspace_terminate(struct vmspace
*vm
, int final
)
439 lwkt_gettoken(&vm
->vm_map
.token
);
441 KKASSERT((vm
->vm_flags
& VMSPACE_EXIT1
) == 0);
442 vm
->vm_flags
|= VMSPACE_EXIT1
;
445 * Get rid of most of the resources. Leave the kernel pmap
448 * If the pmap does not contain wired pages we can bulk-delete
449 * the pmap as a performance optimization before removing the
452 * If the pmap contains wired pages we cannot do this
453 * pre-optimization because currently vm_fault_unwire()
454 * expects the pmap pages to exist and will not decrement
455 * p->wire_count if they do not.
458 if (vmspace_pmap(vm
)->pm_stats
.wired_count
) {
459 vm_map_remove(&vm
->vm_map
, VM_MIN_USER_ADDRESS
,
460 VM_MAX_USER_ADDRESS
);
461 pmap_remove_pages(vmspace_pmap(vm
), VM_MIN_USER_ADDRESS
,
462 VM_MAX_USER_ADDRESS
);
464 pmap_remove_pages(vmspace_pmap(vm
), VM_MIN_USER_ADDRESS
,
465 VM_MAX_USER_ADDRESS
);
466 vm_map_remove(&vm
->vm_map
, VM_MIN_USER_ADDRESS
,
467 VM_MAX_USER_ADDRESS
);
469 lwkt_reltoken(&vm
->vm_map
.token
);
471 KKASSERT((vm
->vm_flags
& VMSPACE_EXIT1
) != 0);
472 KKASSERT((vm
->vm_flags
& VMSPACE_EXIT2
) == 0);
475 * Get rid of remaining basic resources.
477 vm
->vm_flags
|= VMSPACE_EXIT2
;
480 count
= vm_map_entry_reserve(MAP_RESERVE_COUNT
);
481 vm_map_lock(&vm
->vm_map
);
482 cpu_vmspace_free(vm
);
485 * Lock the map, to wait out all other references to it.
486 * Delete all of the mappings and pages they hold, then call
487 * the pmap module to reclaim anything left.
489 vm_map_delete(&vm
->vm_map
, vm
->vm_map
.min_offset
,
490 vm
->vm_map
.max_offset
, &count
);
491 vm_map_unlock(&vm
->vm_map
);
492 vm_map_entry_release(count
);
494 pmap_release(vmspace_pmap(vm
));
495 lwkt_reltoken(&vm
->vm_map
.token
);
496 objcache_put(vmspace_cache
, vm
);
501 * Swap useage is determined by taking the proportional swap used by
502 * VM objects backing the VM map. To make up for fractional losses,
503 * if the VM object has any swap use at all the associated map entries
504 * count for at least 1 swap page.
509 vmspace_swap_count(struct vmspace
*vm
)
511 vm_map_t map
= &vm
->vm_map
;
514 vm_offset_t count
= 0;
518 for (cur
= map
->header
.next
; cur
!= &map
->header
; cur
= cur
->next
) {
519 switch(cur
->maptype
) {
520 case VM_MAPTYPE_NORMAL
:
521 case VM_MAPTYPE_VPAGETABLE
:
522 if ((object
= cur
->object
.vm_object
) == NULL
)
524 if (object
->swblock_count
) {
525 n
= (cur
->end
- cur
->start
) / PAGE_SIZE
;
526 count
+= object
->swblock_count
*
527 SWAP_META_PAGES
* n
/ object
->size
+ 1;
540 * Calculate the approximate number of anonymous pages in use by
541 * this vmspace. To make up for fractional losses, we count each
542 * VM object as having at least 1 anonymous page.
547 vmspace_anonymous_count(struct vmspace
*vm
)
549 vm_map_t map
= &vm
->vm_map
;
552 vm_offset_t count
= 0;
555 for (cur
= map
->header
.next
; cur
!= &map
->header
; cur
= cur
->next
) {
556 switch(cur
->maptype
) {
557 case VM_MAPTYPE_NORMAL
:
558 case VM_MAPTYPE_VPAGETABLE
:
559 if ((object
= cur
->object
.vm_object
) == NULL
)
561 if (object
->type
!= OBJT_DEFAULT
&&
562 object
->type
!= OBJT_SWAP
) {
565 count
+= object
->resident_page_count
;
577 * Initialize an existing vm_map structure such as that in the vmspace
578 * structure. The pmap is initialized elsewhere.
583 vm_map_init(struct vm_map
*map
, vm_offset_t min
, vm_offset_t max
, pmap_t pmap
)
585 map
->header
.next
= map
->header
.prev
= &map
->header
;
586 RB_INIT(&map
->rb_root
);
590 map
->min_offset
= min
;
591 map
->max_offset
= max
;
593 map
->first_free
= &map
->header
;
594 map
->hint
= &map
->header
;
597 lwkt_token_init(&map
->token
, "vm_map");
598 lockinit(&map
->lock
, "vm_maplk", (hz
+ 9) / 10, 0);
602 * Shadow the vm_map_entry's object. This typically needs to be done when
603 * a write fault is taken on an entry which had previously been cloned by
604 * fork(). The shared object (which might be NULL) must become private so
605 * we add a shadow layer above it.
607 * Object allocation for anonymous mappings is defered as long as possible.
608 * When creating a shadow, however, the underlying object must be instantiated
609 * so it can be shared.
611 * If the map segment is governed by a virtual page table then it is
612 * possible to address offsets beyond the mapped area. Just allocate
613 * a maximally sized object for this case.
615 * If addref is non-zero an additional reference is added to the returned
616 * entry. This mechanic exists because the additional reference might have
617 * to be added atomically and not after return to prevent a premature
620 * The vm_map must be exclusively locked.
621 * No other requirements.
625 vm_map_entry_shadow(vm_map_entry_t entry
, int addref
)
627 if (entry
->maptype
== VM_MAPTYPE_VPAGETABLE
) {
628 vm_object_shadow(&entry
->object
.vm_object
, &entry
->offset
,
629 0x7FFFFFFF, addref
); /* XXX */
631 vm_object_shadow(&entry
->object
.vm_object
, &entry
->offset
,
632 atop(entry
->end
- entry
->start
), addref
);
634 entry
->eflags
&= ~MAP_ENTRY_NEEDS_COPY
;
638 * Allocate an object for a vm_map_entry.
640 * Object allocation for anonymous mappings is defered as long as possible.
641 * This function is called when we can defer no longer, generally when a map
642 * entry might be split or forked or takes a page fault.
644 * If the map segment is governed by a virtual page table then it is
645 * possible to address offsets beyond the mapped area. Just allocate
646 * a maximally sized object for this case.
648 * The vm_map must be exclusively locked.
649 * No other requirements.
652 vm_map_entry_allocate_object(vm_map_entry_t entry
)
656 if (entry
->maptype
== VM_MAPTYPE_VPAGETABLE
) {
657 obj
= vm_object_allocate(OBJT_DEFAULT
, 0x7FFFFFFF); /* XXX */
659 obj
= vm_object_allocate(OBJT_DEFAULT
,
660 atop(entry
->end
- entry
->start
));
662 entry
->object
.vm_object
= obj
;
667 * Set an initial negative count so the first attempt to reserve
668 * space preloads a bunch of vm_map_entry's for this cpu. Also
669 * pre-allocate 2 vm_map_entries which will be needed by zalloc() to
670 * map a new page for vm_map_entry structures. SMP systems are
671 * particularly sensitive.
673 * This routine is called in early boot so we cannot just call
674 * vm_map_entry_reserve().
676 * Called from the low level boot code only (for each cpu)
678 * WARNING! Take care not to have too-big a static/BSS structure here
679 * as MAXCPU can be 256+, otherwise the loader's 64MB heap
680 * can get blown out by the kernel plus the initrd image.
683 vm_map_entry_reserve_cpu_init(globaldata_t gd
)
685 vm_map_entry_t entry
;
689 gd
->gd_vme_avail
-= MAP_RESERVE_COUNT
* 2;
690 if (gd
->gd_cpuid
== 0) {
691 entry
= &cpu_map_entry_init_bsp
[0];
692 count
= MAPENTRYBSP_CACHE
;
694 entry
= &cpu_map_entry_init_ap
[gd
->gd_cpuid
][0];
695 count
= MAPENTRYAP_CACHE
;
697 for (i
= 0; i
< count
; ++i
, ++entry
) {
698 entry
->next
= gd
->gd_vme_base
;
699 gd
->gd_vme_base
= entry
;
704 * Reserves vm_map_entry structures so code later on can manipulate
705 * map_entry structures within a locked map without blocking trying
706 * to allocate a new vm_map_entry.
711 vm_map_entry_reserve(int count
)
713 struct globaldata
*gd
= mycpu
;
714 vm_map_entry_t entry
;
717 * Make sure we have enough structures in gd_vme_base to handle
718 * the reservation request.
720 * The critical section protects access to the per-cpu gd.
723 while (gd
->gd_vme_avail
< count
) {
724 entry
= zalloc(mapentzone
);
725 entry
->next
= gd
->gd_vme_base
;
726 gd
->gd_vme_base
= entry
;
729 gd
->gd_vme_avail
-= count
;
736 * Releases previously reserved vm_map_entry structures that were not
737 * used. If we have too much junk in our per-cpu cache clean some of
743 vm_map_entry_release(int count
)
745 struct globaldata
*gd
= mycpu
;
746 vm_map_entry_t entry
;
749 gd
->gd_vme_avail
+= count
;
750 while (gd
->gd_vme_avail
> MAP_RESERVE_SLOP
) {
751 entry
= gd
->gd_vme_base
;
752 KKASSERT(entry
!= NULL
);
753 gd
->gd_vme_base
= entry
->next
;
756 zfree(mapentzone
, entry
);
763 * Reserve map entry structures for use in kernel_map itself. These
764 * entries have *ALREADY* been reserved on a per-cpu basis when the map
765 * was inited. This function is used by zalloc() to avoid a recursion
766 * when zalloc() itself needs to allocate additional kernel memory.
768 * This function works like the normal reserve but does not load the
769 * vm_map_entry cache (because that would result in an infinite
770 * recursion). Note that gd_vme_avail may go negative. This is expected.
772 * Any caller of this function must be sure to renormalize after
773 * potentially eating entries to ensure that the reserve supply
779 vm_map_entry_kreserve(int count
)
781 struct globaldata
*gd
= mycpu
;
784 gd
->gd_vme_avail
-= count
;
786 KASSERT(gd
->gd_vme_base
!= NULL
,
787 ("no reserved entries left, gd_vme_avail = %d",
793 * Release previously reserved map entries for kernel_map. We do not
794 * attempt to clean up like the normal release function as this would
795 * cause an unnecessary (but probably not fatal) deep procedure call.
800 vm_map_entry_krelease(int count
)
802 struct globaldata
*gd
= mycpu
;
805 gd
->gd_vme_avail
+= count
;
810 * Allocates a VM map entry for insertion. No entry fields are filled in.
812 * The entries should have previously been reserved. The reservation count
813 * is tracked in (*countp).
817 static vm_map_entry_t
818 vm_map_entry_create(vm_map_t map
, int *countp
)
820 struct globaldata
*gd
= mycpu
;
821 vm_map_entry_t entry
;
823 KKASSERT(*countp
> 0);
826 entry
= gd
->gd_vme_base
;
827 KASSERT(entry
!= NULL
, ("gd_vme_base NULL! count %d", *countp
));
828 gd
->gd_vme_base
= entry
->next
;
835 * Dispose of a vm_map_entry that is no longer being referenced.
840 vm_map_entry_dispose(vm_map_t map
, vm_map_entry_t entry
, int *countp
)
842 struct globaldata
*gd
= mycpu
;
844 KKASSERT(map
->hint
!= entry
);
845 KKASSERT(map
->first_free
!= entry
);
849 entry
->next
= gd
->gd_vme_base
;
850 gd
->gd_vme_base
= entry
;
856 * Insert/remove entries from maps.
858 * The related map must be exclusively locked.
859 * The caller must hold map->token
860 * No other requirements.
863 vm_map_entry_link(vm_map_t map
,
864 vm_map_entry_t after_where
,
865 vm_map_entry_t entry
)
867 ASSERT_VM_MAP_LOCKED(map
);
870 entry
->prev
= after_where
;
871 entry
->next
= after_where
->next
;
872 entry
->next
->prev
= entry
;
873 after_where
->next
= entry
;
874 if (vm_map_rb_tree_RB_INSERT(&map
->rb_root
, entry
))
875 panic("vm_map_entry_link: dup addr map %p ent %p", map
, entry
);
879 vm_map_entry_unlink(vm_map_t map
,
880 vm_map_entry_t entry
)
885 ASSERT_VM_MAP_LOCKED(map
);
887 if (entry
->eflags
& MAP_ENTRY_IN_TRANSITION
) {
888 panic("vm_map_entry_unlink: attempt to mess with "
889 "locked entry! %p", entry
);
895 vm_map_rb_tree_RB_REMOVE(&map
->rb_root
, entry
);
900 * Finds the map entry containing (or immediately preceding) the specified
901 * address in the given map. The entry is returned in (*entry).
903 * The boolean result indicates whether the address is actually contained
906 * The related map must be locked.
907 * No other requirements.
910 vm_map_lookup_entry(vm_map_t map
, vm_offset_t address
, vm_map_entry_t
*entry
)
915 ASSERT_VM_MAP_LOCKED(map
);
918 * XXX TEMPORARILY DISABLED. For some reason our attempt to revive
919 * the hint code with the red-black lookup meets with system crashes
920 * and lockups. We do not yet know why.
922 * It is possible that the problem is related to the setting
923 * of the hint during map_entry deletion, in the code specified
924 * at the GGG comment later on in this file.
926 * YYY More likely it's because this function can be called with
927 * a shared lock on the map, resulting in map->hint updates possibly
928 * racing. Fixed now but untested.
931 * Quickly check the cached hint, there's a good chance of a match.
935 if (tmp
!= &map
->header
) {
936 if (address
>= tmp
->start
&& address
< tmp
->end
) {
944 * Locate the record from the top of the tree. 'last' tracks the
945 * closest prior record and is returned if no match is found, which
946 * in binary tree terms means tracking the most recent right-branch
947 * taken. If there is no prior record, &map->header is returned.
950 tmp
= RB_ROOT(&map
->rb_root
);
953 if (address
>= tmp
->start
) {
954 if (address
< tmp
->end
) {
960 tmp
= RB_RIGHT(tmp
, rb_entry
);
962 tmp
= RB_LEFT(tmp
, rb_entry
);
970 * Inserts the given whole VM object into the target map at the specified
971 * address range. The object's size should match that of the address range.
973 * The map must be exclusively locked.
974 * The object must be held.
975 * The caller must have reserved sufficient vm_map_entry structures.
977 * If object is non-NULL, ref count must be bumped by caller prior to
978 * making call to account for the new entry.
981 vm_map_insert(vm_map_t map
, int *countp
, void *map_object
, void *map_aux
,
982 vm_ooffset_t offset
, vm_offset_t start
, vm_offset_t end
,
983 vm_maptype_t maptype
, vm_subsys_t id
,
984 vm_prot_t prot
, vm_prot_t max
, int cow
)
986 vm_map_entry_t new_entry
;
987 vm_map_entry_t prev_entry
;
988 vm_map_entry_t temp_entry
;
989 vm_eflags_t protoeflags
;
993 if (maptype
== VM_MAPTYPE_UKSMAP
)
998 ASSERT_VM_MAP_LOCKED(map
);
1000 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object
));
1003 * Check that the start and end points are not bogus.
1005 if ((start
< map
->min_offset
) || (end
> map
->max_offset
) ||
1007 return (KERN_INVALID_ADDRESS
);
1010 * Find the entry prior to the proposed starting address; if it's part
1011 * of an existing entry, this range is bogus.
1013 if (vm_map_lookup_entry(map
, start
, &temp_entry
))
1014 return (KERN_NO_SPACE
);
1016 prev_entry
= temp_entry
;
1019 * Assert that the next entry doesn't overlap the end point.
1022 if ((prev_entry
->next
!= &map
->header
) &&
1023 (prev_entry
->next
->start
< end
))
1024 return (KERN_NO_SPACE
);
1028 if (cow
& MAP_COPY_ON_WRITE
)
1029 protoeflags
|= MAP_ENTRY_COW
|MAP_ENTRY_NEEDS_COPY
;
1031 if (cow
& MAP_NOFAULT
) {
1032 protoeflags
|= MAP_ENTRY_NOFAULT
;
1034 KASSERT(object
== NULL
,
1035 ("vm_map_insert: paradoxical MAP_NOFAULT request"));
1037 if (cow
& MAP_DISABLE_SYNCER
)
1038 protoeflags
|= MAP_ENTRY_NOSYNC
;
1039 if (cow
& MAP_DISABLE_COREDUMP
)
1040 protoeflags
|= MAP_ENTRY_NOCOREDUMP
;
1041 if (cow
& MAP_IS_STACK
)
1042 protoeflags
|= MAP_ENTRY_STACK
;
1043 if (cow
& MAP_IS_KSTACK
)
1044 protoeflags
|= MAP_ENTRY_KSTACK
;
1046 lwkt_gettoken(&map
->token
);
1050 * When object is non-NULL, it could be shared with another
1051 * process. We have to set or clear OBJ_ONEMAPPING
1054 * NOTE: This flag is only applicable to DEFAULT and SWAP
1055 * objects and will already be clear in other types
1056 * of objects, so a shared object lock is ok for
1059 if ((object
->ref_count
> 1) || (object
->shadow_count
!= 0)) {
1060 vm_object_clear_flag(object
, OBJ_ONEMAPPING
);
1063 else if ((prev_entry
!= &map
->header
) &&
1064 (prev_entry
->eflags
== protoeflags
) &&
1065 (prev_entry
->end
== start
) &&
1066 (prev_entry
->wired_count
== 0) &&
1067 (prev_entry
->id
== id
) &&
1068 prev_entry
->maptype
== maptype
&&
1069 maptype
== VM_MAPTYPE_NORMAL
&&
1070 ((prev_entry
->object
.vm_object
== NULL
) ||
1071 vm_object_coalesce(prev_entry
->object
.vm_object
,
1072 OFF_TO_IDX(prev_entry
->offset
),
1073 (vm_size_t
)(prev_entry
->end
- prev_entry
->start
),
1074 (vm_size_t
)(end
- prev_entry
->end
)))) {
1076 * We were able to extend the object. Determine if we
1077 * can extend the previous map entry to include the
1078 * new range as well.
1080 if ((prev_entry
->inheritance
== VM_INHERIT_DEFAULT
) &&
1081 (prev_entry
->protection
== prot
) &&
1082 (prev_entry
->max_protection
== max
)) {
1083 map
->size
+= (end
- prev_entry
->end
);
1084 prev_entry
->end
= end
;
1085 vm_map_simplify_entry(map
, prev_entry
, countp
);
1086 lwkt_reltoken(&map
->token
);
1087 return (KERN_SUCCESS
);
1091 * If we can extend the object but cannot extend the
1092 * map entry, we have to create a new map entry. We
1093 * must bump the ref count on the extended object to
1094 * account for it. object may be NULL.
1096 * XXX if object is NULL should we set offset to 0 here ?
1098 object
= prev_entry
->object
.vm_object
;
1099 offset
= prev_entry
->offset
+
1100 (prev_entry
->end
- prev_entry
->start
);
1102 vm_object_hold(object
);
1103 vm_object_chain_wait(object
, 0);
1104 vm_object_reference_locked(object
);
1106 map_object
= object
;
1111 * NOTE: if conditionals fail, object can be NULL here. This occurs
1112 * in things like the buffer map where we manage kva but do not manage
1117 * Create a new entry
1120 new_entry
= vm_map_entry_create(map
, countp
);
1121 new_entry
->start
= start
;
1122 new_entry
->end
= end
;
1125 new_entry
->maptype
= maptype
;
1126 new_entry
->eflags
= protoeflags
;
1127 new_entry
->object
.map_object
= map_object
;
1128 new_entry
->aux
.master_pde
= 0; /* in case size is different */
1129 new_entry
->aux
.map_aux
= map_aux
;
1130 new_entry
->offset
= offset
;
1132 new_entry
->inheritance
= VM_INHERIT_DEFAULT
;
1133 new_entry
->protection
= prot
;
1134 new_entry
->max_protection
= max
;
1135 new_entry
->wired_count
= 0;
1138 * Insert the new entry into the list
1141 vm_map_entry_link(map
, prev_entry
, new_entry
);
1142 map
->size
+= new_entry
->end
- new_entry
->start
;
1145 * Update the free space hint. Entries cannot overlap.
1146 * An exact comparison is needed to avoid matching
1147 * against the map->header.
1149 if ((map
->first_free
== prev_entry
) &&
1150 (prev_entry
->end
== new_entry
->start
)) {
1151 map
->first_free
= new_entry
;
1156 * Temporarily removed to avoid MAP_STACK panic, due to
1157 * MAP_STACK being a huge hack. Will be added back in
1158 * when MAP_STACK (and the user stack mapping) is fixed.
1161 * It may be possible to simplify the entry
1163 vm_map_simplify_entry(map
, new_entry
, countp
);
1167 * Try to pre-populate the page table. Mappings governed by virtual
1168 * page tables cannot be prepopulated without a lot of work, so
1171 if ((cow
& (MAP_PREFAULT
|MAP_PREFAULT_PARTIAL
)) &&
1172 maptype
!= VM_MAPTYPE_VPAGETABLE
&&
1173 maptype
!= VM_MAPTYPE_UKSMAP
) {
1175 if (vm_map_relock_enable
&& (cow
& MAP_PREFAULT_RELOCK
)) {
1177 vm_object_lock_swap();
1178 vm_object_drop(object
);
1180 pmap_object_init_pt(map
->pmap
, start
, prot
,
1181 object
, OFF_TO_IDX(offset
), end
- start
,
1182 cow
& MAP_PREFAULT_PARTIAL
);
1184 vm_object_hold(object
);
1185 vm_object_lock_swap();
1189 vm_object_drop(object
);
1191 lwkt_reltoken(&map
->token
);
1192 return (KERN_SUCCESS
);
1196 * Find sufficient space for `length' bytes in the given map, starting at
1197 * `start'. Returns 0 on success, 1 on no space.
1199 * This function will returned an arbitrarily aligned pointer. If no
1200 * particular alignment is required you should pass align as 1. Note that
1201 * the map may return PAGE_SIZE aligned pointers if all the lengths used in
1202 * the map are a multiple of PAGE_SIZE, even if you pass a smaller align
1205 * 'align' should be a power of 2 but is not required to be.
1207 * The map must be exclusively locked.
1208 * No other requirements.
1211 vm_map_findspace(vm_map_t map
, vm_offset_t start
, vm_size_t length
,
1212 vm_size_t align
, int flags
, vm_offset_t
*addr
)
1214 vm_map_entry_t entry
, next
;
1216 vm_offset_t align_mask
;
1218 if (start
< map
->min_offset
)
1219 start
= map
->min_offset
;
1220 if (start
> map
->max_offset
)
1224 * If the alignment is not a power of 2 we will have to use
1225 * a mod/division, set align_mask to a special value.
1227 if ((align
| (align
- 1)) + 1 != (align
<< 1))
1228 align_mask
= (vm_offset_t
)-1;
1230 align_mask
= align
- 1;
1233 * Look for the first possible address; if there's already something
1234 * at this address, we have to start after it.
1236 if (start
== map
->min_offset
) {
1237 if ((entry
= map
->first_free
) != &map
->header
)
1242 if (vm_map_lookup_entry(map
, start
, &tmp
))
1248 * Look through the rest of the map, trying to fit a new region in the
1249 * gap between existing regions, or after the very last region.
1251 for (;; start
= (entry
= next
)->end
) {
1253 * Adjust the proposed start by the requested alignment,
1254 * be sure that we didn't wrap the address.
1256 if (align_mask
== (vm_offset_t
)-1)
1257 end
= roundup(start
, align
);
1259 end
= (start
+ align_mask
) & ~align_mask
;
1264 * Find the end of the proposed new region. Be sure we didn't
1265 * go beyond the end of the map, or wrap around the address.
1266 * Then check to see if this is the last entry or if the
1267 * proposed end fits in the gap between this and the next
1270 end
= start
+ length
;
1271 if (end
> map
->max_offset
|| end
< start
)
1276 * If the next entry's start address is beyond the desired
1277 * end address we may have found a good entry.
1279 * If the next entry is a stack mapping we do not map into
1280 * the stack's reserved space.
1282 * XXX continue to allow mapping into the stack's reserved
1283 * space if doing a MAP_STACK mapping inside a MAP_STACK
1284 * mapping, for backwards compatibility. But the caller
1285 * really should use MAP_STACK | MAP_TRYFIXED if they
1288 if (next
== &map
->header
)
1290 if (next
->start
>= end
) {
1291 if ((next
->eflags
& MAP_ENTRY_STACK
) == 0)
1293 if (flags
& MAP_STACK
)
1295 if (next
->start
- next
->aux
.avail_ssize
>= end
)
1302 * Grow the kernel_map if necessary. pmap_growkernel() will panic
1303 * if it fails. The kernel_map is locked and nothing can steal
1304 * our address space if pmap_growkernel() blocks.
1306 * NOTE: This may be unconditionally called for kldload areas on
1307 * x86_64 because these do not bump kernel_vm_end (which would
1308 * fill 128G worth of page tables!). Therefore we must not
1311 if (map
== &kernel_map
) {
1314 kstop
= round_page(start
+ length
);
1315 if (kstop
> kernel_vm_end
)
1316 pmap_growkernel(start
, kstop
);
1323 * vm_map_find finds an unallocated region in the target address map with
1324 * the given length and allocates it. The search is defined to be first-fit
1325 * from the specified address; the region found is returned in the same
1328 * If object is non-NULL, ref count must be bumped by caller
1329 * prior to making call to account for the new entry.
1331 * No requirements. This function will lock the map temporarily.
1334 vm_map_find(vm_map_t map
, void *map_object
, void *map_aux
,
1335 vm_ooffset_t offset
, vm_offset_t
*addr
,
1336 vm_size_t length
, vm_size_t align
, boolean_t fitit
,
1337 vm_maptype_t maptype
, vm_subsys_t id
,
1338 vm_prot_t prot
, vm_prot_t max
, int cow
)
1345 if (maptype
== VM_MAPTYPE_UKSMAP
)
1348 object
= map_object
;
1352 count
= vm_map_entry_reserve(MAP_RESERVE_COUNT
);
1355 vm_object_hold_shared(object
);
1357 if (vm_map_findspace(map
, start
, length
, align
, 0, addr
)) {
1359 vm_object_drop(object
);
1361 vm_map_entry_release(count
);
1362 return (KERN_NO_SPACE
);
1366 result
= vm_map_insert(map
, &count
, map_object
, map_aux
,
1367 offset
, start
, start
+ length
,
1368 maptype
, id
, prot
, max
, cow
);
1370 vm_object_drop(object
);
1372 vm_map_entry_release(count
);
1378 * Simplify the given map entry by merging with either neighbor. This
1379 * routine also has the ability to merge with both neighbors.
1381 * This routine guarentees that the passed entry remains valid (though
1382 * possibly extended). When merging, this routine may delete one or
1383 * both neighbors. No action is taken on entries which have their
1384 * in-transition flag set.
1386 * The map must be exclusively locked.
1389 vm_map_simplify_entry(vm_map_t map
, vm_map_entry_t entry
, int *countp
)
1391 vm_map_entry_t next
, prev
;
1392 vm_size_t prevsize
, esize
;
1394 if (entry
->eflags
& MAP_ENTRY_IN_TRANSITION
) {
1395 ++mycpu
->gd_cnt
.v_intrans_coll
;
1399 if (entry
->maptype
== VM_MAPTYPE_SUBMAP
)
1401 if (entry
->maptype
== VM_MAPTYPE_UKSMAP
)
1405 if (prev
!= &map
->header
) {
1406 prevsize
= prev
->end
- prev
->start
;
1407 if ( (prev
->end
== entry
->start
) &&
1408 (prev
->maptype
== entry
->maptype
) &&
1409 (prev
->object
.vm_object
== entry
->object
.vm_object
) &&
1410 (!prev
->object
.vm_object
||
1411 (prev
->offset
+ prevsize
== entry
->offset
)) &&
1412 (prev
->eflags
== entry
->eflags
) &&
1413 (prev
->protection
== entry
->protection
) &&
1414 (prev
->max_protection
== entry
->max_protection
) &&
1415 (prev
->inheritance
== entry
->inheritance
) &&
1416 (prev
->id
== entry
->id
) &&
1417 (prev
->wired_count
== entry
->wired_count
)) {
1418 if (map
->first_free
== prev
)
1419 map
->first_free
= entry
;
1420 if (map
->hint
== prev
)
1422 vm_map_entry_unlink(map
, prev
);
1423 entry
->start
= prev
->start
;
1424 entry
->offset
= prev
->offset
;
1425 if (prev
->object
.vm_object
)
1426 vm_object_deallocate(prev
->object
.vm_object
);
1427 vm_map_entry_dispose(map
, prev
, countp
);
1432 if (next
!= &map
->header
) {
1433 esize
= entry
->end
- entry
->start
;
1434 if ((entry
->end
== next
->start
) &&
1435 (next
->maptype
== entry
->maptype
) &&
1436 (next
->object
.vm_object
== entry
->object
.vm_object
) &&
1437 (!entry
->object
.vm_object
||
1438 (entry
->offset
+ esize
== next
->offset
)) &&
1439 (next
->eflags
== entry
->eflags
) &&
1440 (next
->protection
== entry
->protection
) &&
1441 (next
->max_protection
== entry
->max_protection
) &&
1442 (next
->inheritance
== entry
->inheritance
) &&
1443 (next
->id
== entry
->id
) &&
1444 (next
->wired_count
== entry
->wired_count
)) {
1445 if (map
->first_free
== next
)
1446 map
->first_free
= entry
;
1447 if (map
->hint
== next
)
1449 vm_map_entry_unlink(map
, next
);
1450 entry
->end
= next
->end
;
1451 if (next
->object
.vm_object
)
1452 vm_object_deallocate(next
->object
.vm_object
);
1453 vm_map_entry_dispose(map
, next
, countp
);
1459 * Asserts that the given entry begins at or after the specified address.
1460 * If necessary, it splits the entry into two.
1462 #define vm_map_clip_start(map, entry, startaddr, countp) \
1464 if (startaddr > entry->start) \
1465 _vm_map_clip_start(map, entry, startaddr, countp); \
1469 * This routine is called only when it is known that the entry must be split.
1471 * The map must be exclusively locked.
1474 _vm_map_clip_start(vm_map_t map
, vm_map_entry_t entry
, vm_offset_t start
,
1477 vm_map_entry_t new_entry
;
1480 * Split off the front portion -- note that we must insert the new
1481 * entry BEFORE this one, so that this entry has the specified
1485 vm_map_simplify_entry(map
, entry
, countp
);
1488 * If there is no object backing this entry, we might as well create
1489 * one now. If we defer it, an object can get created after the map
1490 * is clipped, and individual objects will be created for the split-up
1491 * map. This is a bit of a hack, but is also about the best place to
1492 * put this improvement.
1494 if (entry
->object
.vm_object
== NULL
&& !map
->system_map
) {
1495 vm_map_entry_allocate_object(entry
);
1498 new_entry
= vm_map_entry_create(map
, countp
);
1499 *new_entry
= *entry
;
1501 new_entry
->end
= start
;
1502 entry
->offset
+= (start
- entry
->start
);
1503 entry
->start
= start
;
1505 vm_map_entry_link(map
, entry
->prev
, new_entry
);
1507 switch(entry
->maptype
) {
1508 case VM_MAPTYPE_NORMAL
:
1509 case VM_MAPTYPE_VPAGETABLE
:
1510 if (new_entry
->object
.vm_object
) {
1511 vm_object_hold(new_entry
->object
.vm_object
);
1512 vm_object_chain_wait(new_entry
->object
.vm_object
, 0);
1513 vm_object_reference_locked(new_entry
->object
.vm_object
);
1514 vm_object_drop(new_entry
->object
.vm_object
);
1523 * Asserts that the given entry ends at or before the specified address.
1524 * If necessary, it splits the entry into two.
1526 * The map must be exclusively locked.
1528 #define vm_map_clip_end(map, entry, endaddr, countp) \
1530 if (endaddr < entry->end) \
1531 _vm_map_clip_end(map, entry, endaddr, countp); \
1535 * This routine is called only when it is known that the entry must be split.
1537 * The map must be exclusively locked.
1540 _vm_map_clip_end(vm_map_t map
, vm_map_entry_t entry
, vm_offset_t end
,
1543 vm_map_entry_t new_entry
;
1546 * If there is no object backing this entry, we might as well create
1547 * one now. If we defer it, an object can get created after the map
1548 * is clipped, and individual objects will be created for the split-up
1549 * map. This is a bit of a hack, but is also about the best place to
1550 * put this improvement.
1553 if (entry
->object
.vm_object
== NULL
&& !map
->system_map
) {
1554 vm_map_entry_allocate_object(entry
);
1558 * Create a new entry and insert it AFTER the specified entry
1561 new_entry
= vm_map_entry_create(map
, countp
);
1562 *new_entry
= *entry
;
1564 new_entry
->start
= entry
->end
= end
;
1565 new_entry
->offset
+= (end
- entry
->start
);
1567 vm_map_entry_link(map
, entry
, new_entry
);
1569 switch(entry
->maptype
) {
1570 case VM_MAPTYPE_NORMAL
:
1571 case VM_MAPTYPE_VPAGETABLE
:
1572 if (new_entry
->object
.vm_object
) {
1573 vm_object_hold(new_entry
->object
.vm_object
);
1574 vm_object_chain_wait(new_entry
->object
.vm_object
, 0);
1575 vm_object_reference_locked(new_entry
->object
.vm_object
);
1576 vm_object_drop(new_entry
->object
.vm_object
);
1585 * Asserts that the starting and ending region addresses fall within the
1586 * valid range for the map.
1588 #define VM_MAP_RANGE_CHECK(map, start, end) \
1590 if (start < vm_map_min(map)) \
1591 start = vm_map_min(map); \
1592 if (end > vm_map_max(map)) \
1593 end = vm_map_max(map); \
1599 * Used to block when an in-transition collison occurs. The map
1600 * is unlocked for the sleep and relocked before the return.
1603 vm_map_transition_wait(vm_map_t map
)
1605 tsleep_interlock(map
, 0);
1607 tsleep(map
, PINTERLOCKED
, "vment", 0);
1612 * When we do blocking operations with the map lock held it is
1613 * possible that a clip might have occured on our in-transit entry,
1614 * requiring an adjustment to the entry in our loop. These macros
1615 * help the pageable and clip_range code deal with the case. The
1616 * conditional costs virtually nothing if no clipping has occured.
1619 #define CLIP_CHECK_BACK(entry, save_start) \
1621 while (entry->start != save_start) { \
1622 entry = entry->prev; \
1623 KASSERT(entry != &map->header, ("bad entry clip")); \
1627 #define CLIP_CHECK_FWD(entry, save_end) \
1629 while (entry->end != save_end) { \
1630 entry = entry->next; \
1631 KASSERT(entry != &map->header, ("bad entry clip")); \
1637 * Clip the specified range and return the base entry. The
1638 * range may cover several entries starting at the returned base
1639 * and the first and last entry in the covering sequence will be
1640 * properly clipped to the requested start and end address.
1642 * If no holes are allowed you should pass the MAP_CLIP_NO_HOLES
1645 * The MAP_ENTRY_IN_TRANSITION flag will be set for the entries
1646 * covered by the requested range.
1648 * The map must be exclusively locked on entry and will remain locked
1649 * on return. If no range exists or the range contains holes and you
1650 * specified that no holes were allowed, NULL will be returned. This
1651 * routine may temporarily unlock the map in order avoid a deadlock when
1656 vm_map_clip_range(vm_map_t map
, vm_offset_t start
, vm_offset_t end
,
1657 int *countp
, int flags
)
1659 vm_map_entry_t start_entry
;
1660 vm_map_entry_t entry
;
1663 * Locate the entry and effect initial clipping. The in-transition
1664 * case does not occur very often so do not try to optimize it.
1667 if (vm_map_lookup_entry(map
, start
, &start_entry
) == FALSE
)
1669 entry
= start_entry
;
1670 if (entry
->eflags
& MAP_ENTRY_IN_TRANSITION
) {
1671 entry
->eflags
|= MAP_ENTRY_NEEDS_WAKEUP
;
1672 ++mycpu
->gd_cnt
.v_intrans_coll
;
1673 ++mycpu
->gd_cnt
.v_intrans_wait
;
1674 vm_map_transition_wait(map
);
1676 * entry and/or start_entry may have been clipped while
1677 * we slept, or may have gone away entirely. We have
1678 * to restart from the lookup.
1684 * Since we hold an exclusive map lock we do not have to restart
1685 * after clipping, even though clipping may block in zalloc.
1687 vm_map_clip_start(map
, entry
, start
, countp
);
1688 vm_map_clip_end(map
, entry
, end
, countp
);
1689 entry
->eflags
|= MAP_ENTRY_IN_TRANSITION
;
1692 * Scan entries covered by the range. When working on the next
1693 * entry a restart need only re-loop on the current entry which
1694 * we have already locked, since 'next' may have changed. Also,
1695 * even though entry is safe, it may have been clipped so we
1696 * have to iterate forwards through the clip after sleeping.
1698 while (entry
->next
!= &map
->header
&& entry
->next
->start
< end
) {
1699 vm_map_entry_t next
= entry
->next
;
1701 if (flags
& MAP_CLIP_NO_HOLES
) {
1702 if (next
->start
> entry
->end
) {
1703 vm_map_unclip_range(map
, start_entry
,
1704 start
, entry
->end
, countp
, flags
);
1709 if (next
->eflags
& MAP_ENTRY_IN_TRANSITION
) {
1710 vm_offset_t save_end
= entry
->end
;
1711 next
->eflags
|= MAP_ENTRY_NEEDS_WAKEUP
;
1712 ++mycpu
->gd_cnt
.v_intrans_coll
;
1713 ++mycpu
->gd_cnt
.v_intrans_wait
;
1714 vm_map_transition_wait(map
);
1717 * clips might have occured while we blocked.
1719 CLIP_CHECK_FWD(entry
, save_end
);
1720 CLIP_CHECK_BACK(start_entry
, start
);
1724 * No restart necessary even though clip_end may block, we
1725 * are holding the map lock.
1727 vm_map_clip_end(map
, next
, end
, countp
);
1728 next
->eflags
|= MAP_ENTRY_IN_TRANSITION
;
1731 if (flags
& MAP_CLIP_NO_HOLES
) {
1732 if (entry
->end
!= end
) {
1733 vm_map_unclip_range(map
, start_entry
,
1734 start
, entry
->end
, countp
, flags
);
1738 return(start_entry
);
1742 * Undo the effect of vm_map_clip_range(). You should pass the same
1743 * flags and the same range that you passed to vm_map_clip_range().
1744 * This code will clear the in-transition flag on the entries and
1745 * wake up anyone waiting. This code will also simplify the sequence
1746 * and attempt to merge it with entries before and after the sequence.
1748 * The map must be locked on entry and will remain locked on return.
1750 * Note that you should also pass the start_entry returned by
1751 * vm_map_clip_range(). However, if you block between the two calls
1752 * with the map unlocked please be aware that the start_entry may
1753 * have been clipped and you may need to scan it backwards to find
1754 * the entry corresponding with the original start address. You are
1755 * responsible for this, vm_map_unclip_range() expects the correct
1756 * start_entry to be passed to it and will KASSERT otherwise.
1760 vm_map_unclip_range(vm_map_t map
, vm_map_entry_t start_entry
,
1761 vm_offset_t start
, vm_offset_t end
,
1762 int *countp
, int flags
)
1764 vm_map_entry_t entry
;
1766 entry
= start_entry
;
1768 KASSERT(entry
->start
== start
, ("unclip_range: illegal base entry"));
1769 while (entry
!= &map
->header
&& entry
->start
< end
) {
1770 KASSERT(entry
->eflags
& MAP_ENTRY_IN_TRANSITION
,
1771 ("in-transition flag not set during unclip on: %p",
1773 KASSERT(entry
->end
<= end
,
1774 ("unclip_range: tail wasn't clipped"));
1775 entry
->eflags
&= ~MAP_ENTRY_IN_TRANSITION
;
1776 if (entry
->eflags
& MAP_ENTRY_NEEDS_WAKEUP
) {
1777 entry
->eflags
&= ~MAP_ENTRY_NEEDS_WAKEUP
;
1780 entry
= entry
->next
;
1784 * Simplification does not block so there is no restart case.
1786 entry
= start_entry
;
1787 while (entry
!= &map
->header
&& entry
->start
< end
) {
1788 vm_map_simplify_entry(map
, entry
, countp
);
1789 entry
= entry
->next
;
1794 * Mark the given range as handled by a subordinate map.
1796 * This range must have been created with vm_map_find(), and no other
1797 * operations may have been performed on this range prior to calling
1800 * Submappings cannot be removed.
1805 vm_map_submap(vm_map_t map
, vm_offset_t start
, vm_offset_t end
, vm_map_t submap
)
1807 vm_map_entry_t entry
;
1808 int result
= KERN_INVALID_ARGUMENT
;
1811 count
= vm_map_entry_reserve(MAP_RESERVE_COUNT
);
1814 VM_MAP_RANGE_CHECK(map
, start
, end
);
1816 if (vm_map_lookup_entry(map
, start
, &entry
)) {
1817 vm_map_clip_start(map
, entry
, start
, &count
);
1819 entry
= entry
->next
;
1822 vm_map_clip_end(map
, entry
, end
, &count
);
1824 if ((entry
->start
== start
) && (entry
->end
== end
) &&
1825 ((entry
->eflags
& MAP_ENTRY_COW
) == 0) &&
1826 (entry
->object
.vm_object
== NULL
)) {
1827 entry
->object
.sub_map
= submap
;
1828 entry
->maptype
= VM_MAPTYPE_SUBMAP
;
1829 result
= KERN_SUCCESS
;
1832 vm_map_entry_release(count
);
1838 * Sets the protection of the specified address region in the target map.
1839 * If "set_max" is specified, the maximum protection is to be set;
1840 * otherwise, only the current protection is affected.
1842 * The protection is not applicable to submaps, but is applicable to normal
1843 * maps and maps governed by virtual page tables. For example, when operating
1844 * on a virtual page table our protection basically controls how COW occurs
1845 * on the backing object, whereas the virtual page table abstraction itself
1846 * is an abstraction for userland.
1851 vm_map_protect(vm_map_t map
, vm_offset_t start
, vm_offset_t end
,
1852 vm_prot_t new_prot
, boolean_t set_max
)
1854 vm_map_entry_t current
;
1855 vm_map_entry_t entry
;
1858 count
= vm_map_entry_reserve(MAP_RESERVE_COUNT
);
1861 VM_MAP_RANGE_CHECK(map
, start
, end
);
1863 if (vm_map_lookup_entry(map
, start
, &entry
)) {
1864 vm_map_clip_start(map
, entry
, start
, &count
);
1866 entry
= entry
->next
;
1870 * Make a first pass to check for protection violations.
1873 while ((current
!= &map
->header
) && (current
->start
< end
)) {
1874 if (current
->maptype
== VM_MAPTYPE_SUBMAP
) {
1876 vm_map_entry_release(count
);
1877 return (KERN_INVALID_ARGUMENT
);
1879 if ((new_prot
& current
->max_protection
) != new_prot
) {
1881 vm_map_entry_release(count
);
1882 return (KERN_PROTECTION_FAILURE
);
1884 current
= current
->next
;
1888 * Go back and fix up protections. [Note that clipping is not
1889 * necessary the second time.]
1893 while ((current
!= &map
->header
) && (current
->start
< end
)) {
1896 vm_map_clip_end(map
, current
, end
, &count
);
1898 old_prot
= current
->protection
;
1900 current
->max_protection
= new_prot
;
1901 current
->protection
= new_prot
& old_prot
;
1903 current
->protection
= new_prot
;
1907 * Update physical map if necessary. Worry about copy-on-write
1908 * here -- CHECK THIS XXX
1911 if (current
->protection
!= old_prot
) {
1912 #define MASK(entry) (((entry)->eflags & MAP_ENTRY_COW) ? ~VM_PROT_WRITE : \
1915 pmap_protect(map
->pmap
, current
->start
,
1917 current
->protection
& MASK(current
));
1921 vm_map_simplify_entry(map
, current
, &count
);
1923 current
= current
->next
;
1927 vm_map_entry_release(count
);
1928 return (KERN_SUCCESS
);
1932 * This routine traverses a processes map handling the madvise
1933 * system call. Advisories are classified as either those effecting
1934 * the vm_map_entry structure, or those effecting the underlying
1937 * The <value> argument is used for extended madvise calls.
1942 vm_map_madvise(vm_map_t map
, vm_offset_t start
, vm_offset_t end
,
1943 int behav
, off_t value
)
1945 vm_map_entry_t current
, entry
;
1951 * Some madvise calls directly modify the vm_map_entry, in which case
1952 * we need to use an exclusive lock on the map and we need to perform
1953 * various clipping operations. Otherwise we only need a read-lock
1956 count
= vm_map_entry_reserve(MAP_RESERVE_COUNT
);
1960 case MADV_SEQUENTIAL
:
1974 vm_map_lock_read(map
);
1977 vm_map_entry_release(count
);
1982 * Locate starting entry and clip if necessary.
1985 VM_MAP_RANGE_CHECK(map
, start
, end
);
1987 if (vm_map_lookup_entry(map
, start
, &entry
)) {
1989 vm_map_clip_start(map
, entry
, start
, &count
);
1991 entry
= entry
->next
;
1996 * madvise behaviors that are implemented in the vm_map_entry.
1998 * We clip the vm_map_entry so that behavioral changes are
1999 * limited to the specified address range.
2001 for (current
= entry
;
2002 (current
!= &map
->header
) && (current
->start
< end
);
2003 current
= current
->next
2005 if (current
->maptype
== VM_MAPTYPE_SUBMAP
)
2008 vm_map_clip_end(map
, current
, end
, &count
);
2012 vm_map_entry_set_behavior(current
, MAP_ENTRY_BEHAV_NORMAL
);
2014 case MADV_SEQUENTIAL
:
2015 vm_map_entry_set_behavior(current
, MAP_ENTRY_BEHAV_SEQUENTIAL
);
2018 vm_map_entry_set_behavior(current
, MAP_ENTRY_BEHAV_RANDOM
);
2021 current
->eflags
|= MAP_ENTRY_NOSYNC
;
2024 current
->eflags
&= ~MAP_ENTRY_NOSYNC
;
2027 current
->eflags
|= MAP_ENTRY_NOCOREDUMP
;
2030 current
->eflags
&= ~MAP_ENTRY_NOCOREDUMP
;
2034 * Set the page directory page for a map
2035 * governed by a virtual page table. Mark
2036 * the entry as being governed by a virtual
2037 * page table if it is not.
2039 * XXX the page directory page is stored
2040 * in the avail_ssize field if the map_entry.
2042 * XXX the map simplification code does not
2043 * compare this field so weird things may
2044 * happen if you do not apply this function
2045 * to the entire mapping governed by the
2046 * virtual page table.
2048 if (current
->maptype
!= VM_MAPTYPE_VPAGETABLE
) {
2052 current
->aux
.master_pde
= value
;
2053 pmap_remove(map
->pmap
,
2054 current
->start
, current
->end
);
2058 * Invalidate the related pmap entries, used
2059 * to flush portions of the real kernel's
2060 * pmap when the caller has removed or
2061 * modified existing mappings in a virtual
2064 * (exclusive locked map version)
2066 pmap_remove(map
->pmap
,
2067 current
->start
, current
->end
);
2073 vm_map_simplify_entry(map
, current
, &count
);
2081 * madvise behaviors that are implemented in the underlying
2084 * Since we don't clip the vm_map_entry, we have to clip
2085 * the vm_object pindex and count.
2087 * NOTE! These functions are only supported on normal maps,
2088 * except MADV_INVAL which is also supported on
2089 * virtual page tables.
2091 for (current
= entry
;
2092 (current
!= &map
->header
) && (current
->start
< end
);
2093 current
= current
->next
2095 vm_offset_t useStart
;
2097 if (current
->maptype
!= VM_MAPTYPE_NORMAL
&&
2098 (current
->maptype
!= VM_MAPTYPE_VPAGETABLE
||
2099 behav
!= MADV_INVAL
)) {
2103 pindex
= OFF_TO_IDX(current
->offset
);
2104 delta
= atop(current
->end
- current
->start
);
2105 useStart
= current
->start
;
2107 if (current
->start
< start
) {
2108 pindex
+= atop(start
- current
->start
);
2109 delta
-= atop(start
- current
->start
);
2112 if (current
->end
> end
)
2113 delta
-= atop(current
->end
- end
);
2115 if ((vm_spindex_t
)delta
<= 0)
2118 if (behav
== MADV_INVAL
) {
2120 * Invalidate the related pmap entries, used
2121 * to flush portions of the real kernel's
2122 * pmap when the caller has removed or
2123 * modified existing mappings in a virtual
2126 * (shared locked map version)
2128 KASSERT(useStart
>= VM_MIN_USER_ADDRESS
&&
2129 useStart
+ ptoa(delta
) <=
2130 VM_MAX_USER_ADDRESS
,
2131 ("Bad range %016jx-%016jx (%016jx)",
2132 useStart
, useStart
+ ptoa(delta
),
2134 pmap_remove(map
->pmap
,
2136 useStart
+ ptoa(delta
));
2138 vm_object_madvise(current
->object
.vm_object
,
2139 pindex
, delta
, behav
);
2143 * Try to populate the page table. Mappings governed
2144 * by virtual page tables cannot be pre-populated
2145 * without a lot of work so don't try.
2147 if (behav
== MADV_WILLNEED
&&
2148 current
->maptype
!= VM_MAPTYPE_VPAGETABLE
) {
2149 pmap_object_init_pt(
2152 current
->protection
,
2153 current
->object
.vm_object
,
2155 (count
<< PAGE_SHIFT
),
2156 MAP_PREFAULT_MADVISE
2160 vm_map_unlock_read(map
);
2162 vm_map_entry_release(count
);
2168 * Sets the inheritance of the specified address range in the target map.
2169 * Inheritance affects how the map will be shared with child maps at the
2170 * time of vm_map_fork.
2173 vm_map_inherit(vm_map_t map
, vm_offset_t start
, vm_offset_t end
,
2174 vm_inherit_t new_inheritance
)
2176 vm_map_entry_t entry
;
2177 vm_map_entry_t temp_entry
;
2180 switch (new_inheritance
) {
2181 case VM_INHERIT_NONE
:
2182 case VM_INHERIT_COPY
:
2183 case VM_INHERIT_SHARE
:
2186 return (KERN_INVALID_ARGUMENT
);
2189 count
= vm_map_entry_reserve(MAP_RESERVE_COUNT
);
2192 VM_MAP_RANGE_CHECK(map
, start
, end
);
2194 if (vm_map_lookup_entry(map
, start
, &temp_entry
)) {
2196 vm_map_clip_start(map
, entry
, start
, &count
);
2198 entry
= temp_entry
->next
;
2200 while ((entry
!= &map
->header
) && (entry
->start
< end
)) {
2201 vm_map_clip_end(map
, entry
, end
, &count
);
2203 entry
->inheritance
= new_inheritance
;
2205 vm_map_simplify_entry(map
, entry
, &count
);
2207 entry
= entry
->next
;
2210 vm_map_entry_release(count
);
2211 return (KERN_SUCCESS
);
2215 * Implement the semantics of mlock
2218 vm_map_unwire(vm_map_t map
, vm_offset_t start
, vm_offset_t real_end
,
2219 boolean_t new_pageable
)
2221 vm_map_entry_t entry
;
2222 vm_map_entry_t start_entry
;
2224 int rv
= KERN_SUCCESS
;
2227 count
= vm_map_entry_reserve(MAP_RESERVE_COUNT
);
2229 VM_MAP_RANGE_CHECK(map
, start
, real_end
);
2232 start_entry
= vm_map_clip_range(map
, start
, end
, &count
,
2234 if (start_entry
== NULL
) {
2236 vm_map_entry_release(count
);
2237 return (KERN_INVALID_ADDRESS
);
2240 if (new_pageable
== 0) {
2241 entry
= start_entry
;
2242 while ((entry
!= &map
->header
) && (entry
->start
< end
)) {
2243 vm_offset_t save_start
;
2244 vm_offset_t save_end
;
2247 * Already user wired or hard wired (trivial cases)
2249 if (entry
->eflags
& MAP_ENTRY_USER_WIRED
) {
2250 entry
= entry
->next
;
2253 if (entry
->wired_count
!= 0) {
2254 entry
->wired_count
++;
2255 entry
->eflags
|= MAP_ENTRY_USER_WIRED
;
2256 entry
= entry
->next
;
2261 * A new wiring requires instantiation of appropriate
2262 * management structures and the faulting in of the
2265 if (entry
->maptype
== VM_MAPTYPE_NORMAL
||
2266 entry
->maptype
== VM_MAPTYPE_VPAGETABLE
) {
2267 int copyflag
= entry
->eflags
&
2268 MAP_ENTRY_NEEDS_COPY
;
2269 if (copyflag
&& ((entry
->protection
&
2270 VM_PROT_WRITE
) != 0)) {
2271 vm_map_entry_shadow(entry
, 0);
2272 } else if (entry
->object
.vm_object
== NULL
&&
2274 vm_map_entry_allocate_object(entry
);
2277 entry
->wired_count
++;
2278 entry
->eflags
|= MAP_ENTRY_USER_WIRED
;
2281 * Now fault in the area. Note that vm_fault_wire()
2282 * may release the map lock temporarily, it will be
2283 * relocked on return. The in-transition
2284 * flag protects the entries.
2286 save_start
= entry
->start
;
2287 save_end
= entry
->end
;
2288 rv
= vm_fault_wire(map
, entry
, TRUE
, 0);
2290 CLIP_CHECK_BACK(entry
, save_start
);
2292 KASSERT(entry
->wired_count
== 1, ("bad wired_count on entry"));
2293 entry
->eflags
&= ~MAP_ENTRY_USER_WIRED
;
2294 entry
->wired_count
= 0;
2295 if (entry
->end
== save_end
)
2297 entry
= entry
->next
;
2298 KASSERT(entry
!= &map
->header
, ("bad entry clip during backout"));
2300 end
= save_start
; /* unwire the rest */
2304 * note that even though the entry might have been
2305 * clipped, the USER_WIRED flag we set prevents
2306 * duplication so we do not have to do a
2309 entry
= entry
->next
;
2313 * If we failed fall through to the unwiring section to
2314 * unwire what we had wired so far. 'end' has already
2321 * start_entry might have been clipped if we unlocked the
2322 * map and blocked. No matter how clipped it has gotten
2323 * there should be a fragment that is on our start boundary.
2325 CLIP_CHECK_BACK(start_entry
, start
);
2329 * Deal with the unwiring case.
2333 * This is the unwiring case. We must first ensure that the
2334 * range to be unwired is really wired down. We know there
2337 entry
= start_entry
;
2338 while ((entry
!= &map
->header
) && (entry
->start
< end
)) {
2339 if ((entry
->eflags
& MAP_ENTRY_USER_WIRED
) == 0) {
2340 rv
= KERN_INVALID_ARGUMENT
;
2343 KASSERT(entry
->wired_count
!= 0, ("wired count was 0 with USER_WIRED set! %p", entry
));
2344 entry
= entry
->next
;
2348 * Now decrement the wiring count for each region. If a region
2349 * becomes completely unwired, unwire its physical pages and
2353 * The map entries are processed in a loop, checking to
2354 * make sure the entry is wired and asserting it has a wired
2355 * count. However, another loop was inserted more-or-less in
2356 * the middle of the unwiring path. This loop picks up the
2357 * "entry" loop variable from the first loop without first
2358 * setting it to start_entry. Naturally, the secound loop
2359 * is never entered and the pages backing the entries are
2360 * never unwired. This can lead to a leak of wired pages.
2362 entry
= start_entry
;
2363 while ((entry
!= &map
->header
) && (entry
->start
< end
)) {
2364 KASSERT(entry
->eflags
& MAP_ENTRY_USER_WIRED
,
2365 ("expected USER_WIRED on entry %p", entry
));
2366 entry
->eflags
&= ~MAP_ENTRY_USER_WIRED
;
2367 entry
->wired_count
--;
2368 if (entry
->wired_count
== 0)
2369 vm_fault_unwire(map
, entry
);
2370 entry
= entry
->next
;
2374 vm_map_unclip_range(map
, start_entry
, start
, real_end
, &count
,
2378 vm_map_entry_release(count
);
2383 * Sets the pageability of the specified address range in the target map.
2384 * Regions specified as not pageable require locked-down physical
2385 * memory and physical page maps.
2387 * The map must not be locked, but a reference must remain to the map
2388 * throughout the call.
2390 * This function may be called via the zalloc path and must properly
2391 * reserve map entries for kernel_map.
2396 vm_map_wire(vm_map_t map
, vm_offset_t start
, vm_offset_t real_end
, int kmflags
)
2398 vm_map_entry_t entry
;
2399 vm_map_entry_t start_entry
;
2401 int rv
= KERN_SUCCESS
;
2404 if (kmflags
& KM_KRESERVE
)
2405 count
= vm_map_entry_kreserve(MAP_RESERVE_COUNT
);
2407 count
= vm_map_entry_reserve(MAP_RESERVE_COUNT
);
2409 VM_MAP_RANGE_CHECK(map
, start
, real_end
);
2412 start_entry
= vm_map_clip_range(map
, start
, end
, &count
,
2414 if (start_entry
== NULL
) {
2416 rv
= KERN_INVALID_ADDRESS
;
2419 if ((kmflags
& KM_PAGEABLE
) == 0) {
2423 * 1. Holding the write lock, we create any shadow or zero-fill
2424 * objects that need to be created. Then we clip each map
2425 * entry to the region to be wired and increment its wiring
2426 * count. We create objects before clipping the map entries
2427 * to avoid object proliferation.
2429 * 2. We downgrade to a read lock, and call vm_fault_wire to
2430 * fault in the pages for any newly wired area (wired_count is
2433 * Downgrading to a read lock for vm_fault_wire avoids a
2434 * possible deadlock with another process that may have faulted
2435 * on one of the pages to be wired (it would mark the page busy,
2436 * blocking us, then in turn block on the map lock that we
2437 * hold). Because of problems in the recursive lock package,
2438 * we cannot upgrade to a write lock in vm_map_lookup. Thus,
2439 * any actions that require the write lock must be done
2440 * beforehand. Because we keep the read lock on the map, the
2441 * copy-on-write status of the entries we modify here cannot
2444 entry
= start_entry
;
2445 while ((entry
!= &map
->header
) && (entry
->start
< end
)) {
2447 * Trivial case if the entry is already wired
2449 if (entry
->wired_count
) {
2450 entry
->wired_count
++;
2451 entry
= entry
->next
;
2456 * The entry is being newly wired, we have to setup
2457 * appropriate management structures. A shadow
2458 * object is required for a copy-on-write region,
2459 * or a normal object for a zero-fill region. We
2460 * do not have to do this for entries that point to sub
2461 * maps because we won't hold the lock on the sub map.
2463 if (entry
->maptype
== VM_MAPTYPE_NORMAL
||
2464 entry
->maptype
== VM_MAPTYPE_VPAGETABLE
) {
2465 int copyflag
= entry
->eflags
&
2466 MAP_ENTRY_NEEDS_COPY
;
2467 if (copyflag
&& ((entry
->protection
&
2468 VM_PROT_WRITE
) != 0)) {
2469 vm_map_entry_shadow(entry
, 0);
2470 } else if (entry
->object
.vm_object
== NULL
&&
2472 vm_map_entry_allocate_object(entry
);
2476 entry
->wired_count
++;
2477 entry
= entry
->next
;
2485 * HACK HACK HACK HACK
2487 * vm_fault_wire() temporarily unlocks the map to avoid
2488 * deadlocks. The in-transition flag from vm_map_clip_range
2489 * call should protect us from changes while the map is
2492 * NOTE: Previously this comment stated that clipping might
2493 * still occur while the entry is unlocked, but from
2494 * what I can tell it actually cannot.
2496 * It is unclear whether the CLIP_CHECK_*() calls
2497 * are still needed but we keep them in anyway.
2499 * HACK HACK HACK HACK
2502 entry
= start_entry
;
2503 while (entry
!= &map
->header
&& entry
->start
< end
) {
2505 * If vm_fault_wire fails for any page we need to undo
2506 * what has been done. We decrement the wiring count
2507 * for those pages which have not yet been wired (now)
2508 * and unwire those that have (later).
2510 vm_offset_t save_start
= entry
->start
;
2511 vm_offset_t save_end
= entry
->end
;
2513 if (entry
->wired_count
== 1)
2514 rv
= vm_fault_wire(map
, entry
, FALSE
, kmflags
);
2516 CLIP_CHECK_BACK(entry
, save_start
);
2518 KASSERT(entry
->wired_count
== 1, ("wired_count changed unexpectedly"));
2519 entry
->wired_count
= 0;
2520 if (entry
->end
== save_end
)
2522 entry
= entry
->next
;
2523 KASSERT(entry
!= &map
->header
, ("bad entry clip during backout"));
2528 CLIP_CHECK_FWD(entry
, save_end
);
2529 entry
= entry
->next
;
2533 * If a failure occured undo everything by falling through
2534 * to the unwiring code. 'end' has already been adjusted
2538 kmflags
|= KM_PAGEABLE
;
2541 * start_entry is still IN_TRANSITION but may have been
2542 * clipped since vm_fault_wire() unlocks and relocks the
2543 * map. No matter how clipped it has gotten there should
2544 * be a fragment that is on our start boundary.
2546 CLIP_CHECK_BACK(start_entry
, start
);
2549 if (kmflags
& KM_PAGEABLE
) {
2551 * This is the unwiring case. We must first ensure that the
2552 * range to be unwired is really wired down. We know there
2555 entry
= start_entry
;
2556 while ((entry
!= &map
->header
) && (entry
->start
< end
)) {
2557 if (entry
->wired_count
== 0) {
2558 rv
= KERN_INVALID_ARGUMENT
;
2561 entry
= entry
->next
;
2565 * Now decrement the wiring count for each region. If a region
2566 * becomes completely unwired, unwire its physical pages and
2569 entry
= start_entry
;
2570 while ((entry
!= &map
->header
) && (entry
->start
< end
)) {
2571 entry
->wired_count
--;
2572 if (entry
->wired_count
== 0)
2573 vm_fault_unwire(map
, entry
);
2574 entry
= entry
->next
;
2578 vm_map_unclip_range(map
, start_entry
, start
, real_end
,
2579 &count
, MAP_CLIP_NO_HOLES
);
2583 if (kmflags
& KM_KRESERVE
)
2584 vm_map_entry_krelease(count
);
2586 vm_map_entry_release(count
);
2591 * Mark a newly allocated address range as wired but do not fault in
2592 * the pages. The caller is expected to load the pages into the object.
2594 * The map must be locked on entry and will remain locked on return.
2595 * No other requirements.
2598 vm_map_set_wired_quick(vm_map_t map
, vm_offset_t addr
, vm_size_t size
,
2601 vm_map_entry_t scan
;
2602 vm_map_entry_t entry
;
2604 entry
= vm_map_clip_range(map
, addr
, addr
+ size
,
2605 countp
, MAP_CLIP_NO_HOLES
);
2607 scan
!= &map
->header
&& scan
->start
< addr
+ size
;
2608 scan
= scan
->next
) {
2609 KKASSERT(scan
->wired_count
== 0);
2610 scan
->wired_count
= 1;
2612 vm_map_unclip_range(map
, entry
, addr
, addr
+ size
,
2613 countp
, MAP_CLIP_NO_HOLES
);
2617 * Push any dirty cached pages in the address range to their pager.
2618 * If syncio is TRUE, dirty pages are written synchronously.
2619 * If invalidate is TRUE, any cached pages are freed as well.
2621 * This routine is called by sys_msync()
2623 * Returns an error if any part of the specified range is not mapped.
2628 vm_map_clean(vm_map_t map
, vm_offset_t start
, vm_offset_t end
,
2629 boolean_t syncio
, boolean_t invalidate
)
2631 vm_map_entry_t current
;
2632 vm_map_entry_t entry
;
2636 vm_ooffset_t offset
;
2638 vm_map_lock_read(map
);
2639 VM_MAP_RANGE_CHECK(map
, start
, end
);
2640 if (!vm_map_lookup_entry(map
, start
, &entry
)) {
2641 vm_map_unlock_read(map
);
2642 return (KERN_INVALID_ADDRESS
);
2644 lwkt_gettoken(&map
->token
);
2647 * Make a first pass to check for holes.
2649 for (current
= entry
; current
->start
< end
; current
= current
->next
) {
2650 if (current
->maptype
== VM_MAPTYPE_SUBMAP
) {
2651 lwkt_reltoken(&map
->token
);
2652 vm_map_unlock_read(map
);
2653 return (KERN_INVALID_ARGUMENT
);
2655 if (end
> current
->end
&&
2656 (current
->next
== &map
->header
||
2657 current
->end
!= current
->next
->start
)) {
2658 lwkt_reltoken(&map
->token
);
2659 vm_map_unlock_read(map
);
2660 return (KERN_INVALID_ADDRESS
);
2665 pmap_remove(vm_map_pmap(map
), start
, end
);
2668 * Make a second pass, cleaning/uncaching pages from the indicated
2671 for (current
= entry
; current
->start
< end
; current
= current
->next
) {
2672 offset
= current
->offset
+ (start
- current
->start
);
2673 size
= (end
<= current
->end
? end
: current
->end
) - start
;
2675 switch(current
->maptype
) {
2676 case VM_MAPTYPE_SUBMAP
:
2679 vm_map_entry_t tentry
;
2682 smap
= current
->object
.sub_map
;
2683 vm_map_lock_read(smap
);
2684 vm_map_lookup_entry(smap
, offset
, &tentry
);
2685 tsize
= tentry
->end
- offset
;
2688 object
= tentry
->object
.vm_object
;
2689 offset
= tentry
->offset
+ (offset
- tentry
->start
);
2690 vm_map_unlock_read(smap
);
2693 case VM_MAPTYPE_NORMAL
:
2694 case VM_MAPTYPE_VPAGETABLE
:
2695 object
= current
->object
.vm_object
;
2703 vm_object_hold(object
);
2706 * Note that there is absolutely no sense in writing out
2707 * anonymous objects, so we track down the vnode object
2709 * We invalidate (remove) all pages from the address space
2710 * anyway, for semantic correctness.
2712 * note: certain anonymous maps, such as MAP_NOSYNC maps,
2713 * may start out with a NULL object.
2715 while (object
&& (tobj
= object
->backing_object
) != NULL
) {
2716 vm_object_hold(tobj
);
2717 if (tobj
== object
->backing_object
) {
2718 vm_object_lock_swap();
2719 offset
+= object
->backing_object_offset
;
2720 vm_object_drop(object
);
2722 if (object
->size
< OFF_TO_IDX(offset
+ size
))
2723 size
= IDX_TO_OFF(object
->size
) -
2727 vm_object_drop(tobj
);
2729 if (object
&& (object
->type
== OBJT_VNODE
) &&
2730 (current
->protection
& VM_PROT_WRITE
) &&
2731 (object
->flags
& OBJ_NOMSYNC
) == 0) {
2733 * Flush pages if writing is allowed, invalidate them
2734 * if invalidation requested. Pages undergoing I/O
2735 * will be ignored by vm_object_page_remove().
2737 * We cannot lock the vnode and then wait for paging
2738 * to complete without deadlocking against vm_fault.
2739 * Instead we simply call vm_object_page_remove() and
2740 * allow it to block internally on a page-by-page
2741 * basis when it encounters pages undergoing async
2746 /* no chain wait needed for vnode objects */
2747 vm_object_reference_locked(object
);
2748 vn_lock(object
->handle
, LK_EXCLUSIVE
| LK_RETRY
);
2749 flags
= (syncio
|| invalidate
) ? OBJPC_SYNC
: 0;
2750 flags
|= invalidate
? OBJPC_INVAL
: 0;
2753 * When operating on a virtual page table just
2754 * flush the whole object. XXX we probably ought
2757 switch(current
->maptype
) {
2758 case VM_MAPTYPE_NORMAL
:
2759 vm_object_page_clean(object
,
2761 OFF_TO_IDX(offset
+ size
+ PAGE_MASK
),
2764 case VM_MAPTYPE_VPAGETABLE
:
2765 vm_object_page_clean(object
, 0, 0, flags
);
2768 vn_unlock(((struct vnode
*)object
->handle
));
2769 vm_object_deallocate_locked(object
);
2771 if (object
&& invalidate
&&
2772 ((object
->type
== OBJT_VNODE
) ||
2773 (object
->type
== OBJT_DEVICE
) ||
2774 (object
->type
== OBJT_MGTDEVICE
))) {
2776 ((object
->type
== OBJT_DEVICE
) ||
2777 (object
->type
== OBJT_MGTDEVICE
)) ? FALSE
: TRUE
;
2778 /* no chain wait needed for vnode/device objects */
2779 vm_object_reference_locked(object
);
2780 switch(current
->maptype
) {
2781 case VM_MAPTYPE_NORMAL
:
2782 vm_object_page_remove(object
,
2784 OFF_TO_IDX(offset
+ size
+ PAGE_MASK
),
2787 case VM_MAPTYPE_VPAGETABLE
:
2788 vm_object_page_remove(object
, 0, 0, clean_only
);
2791 vm_object_deallocate_locked(object
);
2795 vm_object_drop(object
);
2798 lwkt_reltoken(&map
->token
);
2799 vm_map_unlock_read(map
);
2801 return (KERN_SUCCESS
);
2805 * Make the region specified by this entry pageable.
2807 * The vm_map must be exclusively locked.
2810 vm_map_entry_unwire(vm_map_t map
, vm_map_entry_t entry
)
2812 entry
->eflags
&= ~MAP_ENTRY_USER_WIRED
;
2813 entry
->wired_count
= 0;
2814 vm_fault_unwire(map
, entry
);
2818 * Deallocate the given entry from the target map.
2820 * The vm_map must be exclusively locked.
2823 vm_map_entry_delete(vm_map_t map
, vm_map_entry_t entry
, int *countp
)
2825 vm_map_entry_unlink(map
, entry
);
2826 map
->size
-= entry
->end
- entry
->start
;
2828 switch(entry
->maptype
) {
2829 case VM_MAPTYPE_NORMAL
:
2830 case VM_MAPTYPE_VPAGETABLE
:
2831 case VM_MAPTYPE_SUBMAP
:
2832 vm_object_deallocate(entry
->object
.vm_object
);
2834 case VM_MAPTYPE_UKSMAP
:
2841 vm_map_entry_dispose(map
, entry
, countp
);
2845 * Deallocates the given address range from the target map.
2847 * The vm_map must be exclusively locked.
2850 vm_map_delete(vm_map_t map
, vm_offset_t start
, vm_offset_t end
, int *countp
)
2853 vm_map_entry_t entry
;
2854 vm_map_entry_t first_entry
;
2856 ASSERT_VM_MAP_LOCKED(map
);
2857 lwkt_gettoken(&map
->token
);
2860 * Find the start of the region, and clip it. Set entry to point
2861 * at the first record containing the requested address or, if no
2862 * such record exists, the next record with a greater address. The
2863 * loop will run from this point until a record beyond the termination
2864 * address is encountered.
2866 * map->hint must be adjusted to not point to anything we delete,
2867 * so set it to the entry prior to the one being deleted.
2869 * GGG see other GGG comment.
2871 if (vm_map_lookup_entry(map
, start
, &first_entry
)) {
2872 entry
= first_entry
;
2873 vm_map_clip_start(map
, entry
, start
, countp
);
2874 map
->hint
= entry
->prev
; /* possible problem XXX */
2876 map
->hint
= first_entry
; /* possible problem XXX */
2877 entry
= first_entry
->next
;
2881 * If a hole opens up prior to the current first_free then
2882 * adjust first_free. As with map->hint, map->first_free
2883 * cannot be left set to anything we might delete.
2885 if (entry
== &map
->header
) {
2886 map
->first_free
= &map
->header
;
2887 } else if (map
->first_free
->start
>= start
) {
2888 map
->first_free
= entry
->prev
;
2892 * Step through all entries in this region
2894 while ((entry
!= &map
->header
) && (entry
->start
< end
)) {
2895 vm_map_entry_t next
;
2897 vm_pindex_t offidxstart
, offidxend
, count
;
2900 * If we hit an in-transition entry we have to sleep and
2901 * retry. It's easier (and not really slower) to just retry
2902 * since this case occurs so rarely and the hint is already
2903 * pointing at the right place. We have to reset the
2904 * start offset so as not to accidently delete an entry
2905 * another process just created in vacated space.
2907 if (entry
->eflags
& MAP_ENTRY_IN_TRANSITION
) {
2908 entry
->eflags
|= MAP_ENTRY_NEEDS_WAKEUP
;
2909 start
= entry
->start
;
2910 ++mycpu
->gd_cnt
.v_intrans_coll
;
2911 ++mycpu
->gd_cnt
.v_intrans_wait
;
2912 vm_map_transition_wait(map
);
2915 vm_map_clip_end(map
, entry
, end
, countp
);
2921 offidxstart
= OFF_TO_IDX(entry
->offset
);
2922 count
= OFF_TO_IDX(e
- s
);
2924 switch(entry
->maptype
) {
2925 case VM_MAPTYPE_NORMAL
:
2926 case VM_MAPTYPE_VPAGETABLE
:
2927 case VM_MAPTYPE_SUBMAP
:
2928 object
= entry
->object
.vm_object
;
2936 * Unwire before removing addresses from the pmap; otherwise,
2937 * unwiring will put the entries back in the pmap.
2939 if (entry
->wired_count
!= 0)
2940 vm_map_entry_unwire(map
, entry
);
2942 offidxend
= offidxstart
+ count
;
2944 if (object
== &kernel_object
) {
2945 vm_object_hold(object
);
2946 vm_object_page_remove(object
, offidxstart
,
2948 vm_object_drop(object
);
2949 } else if (object
&& object
->type
!= OBJT_DEFAULT
&&
2950 object
->type
!= OBJT_SWAP
) {
2952 * vnode object routines cannot be chain-locked,
2953 * but since we aren't removing pages from the
2954 * object here we can use a shared hold.
2956 vm_object_hold_shared(object
);
2957 pmap_remove(map
->pmap
, s
, e
);
2958 vm_object_drop(object
);
2959 } else if (object
) {
2960 vm_object_hold(object
);
2961 vm_object_chain_acquire(object
, 0);
2962 pmap_remove(map
->pmap
, s
, e
);
2964 if (object
!= NULL
&&
2965 object
->ref_count
!= 1 &&
2966 (object
->flags
& (OBJ_NOSPLIT
|OBJ_ONEMAPPING
)) ==
2968 (object
->type
== OBJT_DEFAULT
||
2969 object
->type
== OBJT_SWAP
)) {
2970 vm_object_collapse(object
, NULL
);
2971 vm_object_page_remove(object
, offidxstart
,
2973 if (object
->type
== OBJT_SWAP
) {
2974 swap_pager_freespace(object
,
2978 if (offidxend
>= object
->size
&&
2979 offidxstart
< object
->size
) {
2980 object
->size
= offidxstart
;
2983 vm_object_chain_release(object
);
2984 vm_object_drop(object
);
2985 } else if (entry
->maptype
== VM_MAPTYPE_UKSMAP
) {
2986 pmap_remove(map
->pmap
, s
, e
);
2990 * Delete the entry (which may delete the object) only after
2991 * removing all pmap entries pointing to its pages.
2992 * (Otherwise, its page frames may be reallocated, and any
2993 * modify bits will be set in the wrong object!)
2995 vm_map_entry_delete(map
, entry
, countp
);
2998 lwkt_reltoken(&map
->token
);
2999 return (KERN_SUCCESS
);
3003 * Remove the given address range from the target map.
3004 * This is the exported form of vm_map_delete.
3009 vm_map_remove(vm_map_t map
, vm_offset_t start
, vm_offset_t end
)
3014 count
= vm_map_entry_reserve(MAP_RESERVE_COUNT
);
3016 VM_MAP_RANGE_CHECK(map
, start
, end
);
3017 result
= vm_map_delete(map
, start
, end
, &count
);
3019 vm_map_entry_release(count
);
3025 * Assert that the target map allows the specified privilege on the
3026 * entire address region given. The entire region must be allocated.
3028 * The caller must specify whether the vm_map is already locked or not.
3031 vm_map_check_protection(vm_map_t map
, vm_offset_t start
, vm_offset_t end
,
3032 vm_prot_t protection
, boolean_t have_lock
)
3034 vm_map_entry_t entry
;
3035 vm_map_entry_t tmp_entry
;
3038 if (have_lock
== FALSE
)
3039 vm_map_lock_read(map
);
3041 if (!vm_map_lookup_entry(map
, start
, &tmp_entry
)) {
3042 if (have_lock
== FALSE
)
3043 vm_map_unlock_read(map
);
3049 while (start
< end
) {
3050 if (entry
== &map
->header
) {
3058 if (start
< entry
->start
) {
3063 * Check protection associated with entry.
3066 if ((entry
->protection
& protection
) != protection
) {
3070 /* go to next entry */
3073 entry
= entry
->next
;
3075 if (have_lock
== FALSE
)
3076 vm_map_unlock_read(map
);
3081 * If appropriate this function shadows the original object with a new object
3082 * and moves the VM pages from the original object to the new object.
3083 * The original object will also be collapsed, if possible.
3085 * We can only do this for normal memory objects with a single mapping, and
3086 * it only makes sense to do it if there are 2 or more refs on the original
3087 * object. i.e. typically a memory object that has been extended into
3088 * multiple vm_map_entry's with non-overlapping ranges.
3090 * This makes it easier to remove unused pages and keeps object inheritance
3091 * from being a negative impact on memory usage.
3093 * On return the (possibly new) entry->object.vm_object will have an
3094 * additional ref on it for the caller to dispose of (usually by cloning
3095 * the vm_map_entry). The additional ref had to be done in this routine
3096 * to avoid racing a collapse. The object's ONEMAPPING flag will also be
3099 * The vm_map must be locked and its token held.
3102 vm_map_split(vm_map_entry_t entry
)
3105 vm_object_t oobject
, nobject
, bobject
;
3108 vm_pindex_t offidxstart
, offidxend
, idx
;
3110 vm_ooffset_t offset
;
3114 * Optimize away object locks for vnode objects. Important exit/exec
3117 * OBJ_ONEMAPPING doesn't apply to vnode objects but clear the flag
3120 oobject
= entry
->object
.vm_object
;
3121 if (oobject
->type
!= OBJT_DEFAULT
&& oobject
->type
!= OBJT_SWAP
) {
3122 vm_object_reference_quick(oobject
);
3123 vm_object_clear_flag(oobject
, OBJ_ONEMAPPING
);
3128 * Setup. Chain lock the original object throughout the entire
3129 * routine to prevent new page faults from occuring.
3131 * XXX can madvise WILLNEED interfere with us too?
3133 vm_object_hold(oobject
);
3134 vm_object_chain_acquire(oobject
, 0);
3137 * Original object cannot be split? Might have also changed state.
3139 if (oobject
->handle
== NULL
|| (oobject
->type
!= OBJT_DEFAULT
&&
3140 oobject
->type
!= OBJT_SWAP
)) {
3141 vm_object_chain_release(oobject
);
3142 vm_object_reference_locked(oobject
);
3143 vm_object_clear_flag(oobject
, OBJ_ONEMAPPING
);
3144 vm_object_drop(oobject
);
3149 * Collapse original object with its backing store as an
3150 * optimization to reduce chain lengths when possible.
3152 * If ref_count <= 1 there aren't other non-overlapping vm_map_entry's
3153 * for oobject, so there's no point collapsing it.
3155 * Then re-check whether the object can be split.
3157 vm_object_collapse(oobject
, NULL
);
3159 if (oobject
->ref_count
<= 1 ||
3160 (oobject
->type
!= OBJT_DEFAULT
&& oobject
->type
!= OBJT_SWAP
) ||
3161 (oobject
->flags
& (OBJ_NOSPLIT
|OBJ_ONEMAPPING
)) != OBJ_ONEMAPPING
) {
3162 vm_object_chain_release(oobject
);
3163 vm_object_reference_locked(oobject
);
3164 vm_object_clear_flag(oobject
, OBJ_ONEMAPPING
);
3165 vm_object_drop(oobject
);
3170 * Acquire the chain lock on the backing object.
3172 * Give bobject an additional ref count for when it will be shadowed
3176 if ((bobject
= oobject
->backing_object
) != NULL
) {
3177 if (bobject
->type
!= OBJT_VNODE
) {
3179 vm_object_hold(bobject
);
3180 vm_object_chain_wait(bobject
, 0);
3181 /* ref for shadowing below */
3182 vm_object_reference_locked(bobject
);
3183 vm_object_chain_acquire(bobject
, 0);
3184 KKASSERT(bobject
->backing_object
== bobject
);
3185 KKASSERT((bobject
->flags
& OBJ_DEAD
) == 0);
3188 * vnodes are not placed on the shadow list but
3189 * they still get another ref for the backing_object
3192 vm_object_reference_quick(bobject
);
3197 * Calculate the object page range and allocate the new object.
3199 offset
= entry
->offset
;
3203 offidxstart
= OFF_TO_IDX(offset
);
3204 offidxend
= offidxstart
+ OFF_TO_IDX(e
- s
);
3205 size
= offidxend
- offidxstart
;
3207 switch(oobject
->type
) {
3209 nobject
= default_pager_alloc(NULL
, IDX_TO_OFF(size
),
3213 nobject
= swap_pager_alloc(NULL
, IDX_TO_OFF(size
),
3222 if (nobject
== NULL
) {
3224 if (useshadowlist
) {
3225 vm_object_chain_release(bobject
);
3226 vm_object_deallocate(bobject
);
3227 vm_object_drop(bobject
);
3229 vm_object_deallocate(bobject
);
3232 vm_object_chain_release(oobject
);
3233 vm_object_reference_locked(oobject
);
3234 vm_object_clear_flag(oobject
, OBJ_ONEMAPPING
);
3235 vm_object_drop(oobject
);
3240 * The new object will replace entry->object.vm_object so it needs
3241 * a second reference (the caller expects an additional ref).
3243 vm_object_hold(nobject
);
3244 vm_object_reference_locked(nobject
);
3245 vm_object_chain_acquire(nobject
, 0);
3248 * nobject shadows bobject (oobject already shadows bobject).
3250 * Adding an object to bobject's shadow list requires refing bobject
3251 * which we did above in the useshadowlist case.
3254 nobject
->backing_object_offset
=
3255 oobject
->backing_object_offset
+ IDX_TO_OFF(offidxstart
);
3256 nobject
->backing_object
= bobject
;
3257 if (useshadowlist
) {
3258 bobject
->shadow_count
++;
3259 atomic_add_int(&bobject
->generation
, 1);
3260 LIST_INSERT_HEAD(&bobject
->shadow_head
,
3261 nobject
, shadow_list
);
3262 vm_object_clear_flag(bobject
, OBJ_ONEMAPPING
); /*XXX*/
3263 vm_object_chain_release(bobject
);
3264 vm_object_drop(bobject
);
3265 vm_object_set_flag(nobject
, OBJ_ONSHADOW
);
3270 * Move the VM pages from oobject to nobject
3272 for (idx
= 0; idx
< size
; idx
++) {
3275 m
= vm_page_lookup_busy_wait(oobject
, offidxstart
+ idx
,
3281 * We must wait for pending I/O to complete before we can
3284 * We do not have to VM_PROT_NONE the page as mappings should
3285 * not be changed by this operation.
3287 * NOTE: The act of renaming a page updates chaingen for both
3290 vm_page_rename(m
, nobject
, idx
);
3291 /* page automatically made dirty by rename and cache handled */
3292 /* page remains busy */
3295 if (oobject
->type
== OBJT_SWAP
) {
3296 vm_object_pip_add(oobject
, 1);
3298 * copy oobject pages into nobject and destroy unneeded
3299 * pages in shadow object.
3301 swap_pager_copy(oobject
, nobject
, offidxstart
, 0);
3302 vm_object_pip_wakeup(oobject
);
3306 * Wakeup the pages we played with. No spl protection is needed
3307 * for a simple wakeup.
3309 for (idx
= 0; idx
< size
; idx
++) {
3310 m
= vm_page_lookup(nobject
, idx
);
3312 KKASSERT(m
->flags
& PG_BUSY
);
3316 entry
->object
.vm_object
= nobject
;
3317 entry
->offset
= 0LL;
3322 * NOTE: There is no need to remove OBJ_ONEMAPPING from oobject, the
3323 * related pages were moved and are no longer applicable to the
3326 * NOTE: Deallocate oobject (due to its entry->object.vm_object being
3327 * replaced by nobject).
3329 vm_object_chain_release(nobject
);
3330 vm_object_drop(nobject
);
3331 if (bobject
&& useshadowlist
) {
3332 vm_object_chain_release(bobject
);
3333 vm_object_drop(bobject
);
3335 vm_object_chain_release(oobject
);
3336 /*vm_object_clear_flag(oobject, OBJ_ONEMAPPING);*/
3337 vm_object_deallocate_locked(oobject
);
3338 vm_object_drop(oobject
);
3342 * Copies the contents of the source entry to the destination
3343 * entry. The entries *must* be aligned properly.
3345 * The vm_maps must be exclusively locked.
3346 * The vm_map's token must be held.
3348 * Because the maps are locked no faults can be in progress during the
3352 vm_map_copy_entry(vm_map_t src_map
, vm_map_t dst_map
,
3353 vm_map_entry_t src_entry
, vm_map_entry_t dst_entry
)
3355 vm_object_t src_object
;
3357 if (dst_entry
->maptype
== VM_MAPTYPE_SUBMAP
||
3358 dst_entry
->maptype
== VM_MAPTYPE_UKSMAP
)
3360 if (src_entry
->maptype
== VM_MAPTYPE_SUBMAP
||
3361 src_entry
->maptype
== VM_MAPTYPE_UKSMAP
)
3364 if (src_entry
->wired_count
== 0) {
3366 * If the source entry is marked needs_copy, it is already
3369 if ((src_entry
->eflags
& MAP_ENTRY_NEEDS_COPY
) == 0) {
3370 pmap_protect(src_map
->pmap
,
3373 src_entry
->protection
& ~VM_PROT_WRITE
);
3377 * Make a copy of the object.
3379 * The object must be locked prior to checking the object type
3380 * and for the call to vm_object_collapse() and vm_map_split().
3381 * We cannot use *_hold() here because the split code will
3382 * probably try to destroy the object. The lock is a pool
3383 * token and doesn't care.
3385 * We must bump src_map->timestamp when setting
3386 * MAP_ENTRY_NEEDS_COPY to force any concurrent fault
3387 * to retry, otherwise the concurrent fault might improperly
3388 * install a RW pte when its supposed to be a RO(COW) pte.
3389 * This race can occur because a vnode-backed fault may have
3390 * to temporarily release the map lock.
3392 if (src_entry
->object
.vm_object
!= NULL
) {
3393 vm_map_split(src_entry
);
3394 src_object
= src_entry
->object
.vm_object
;
3395 dst_entry
->object
.vm_object
= src_object
;
3396 src_entry
->eflags
|= (MAP_ENTRY_COW
|
3397 MAP_ENTRY_NEEDS_COPY
);
3398 dst_entry
->eflags
|= (MAP_ENTRY_COW
|
3399 MAP_ENTRY_NEEDS_COPY
);
3400 dst_entry
->offset
= src_entry
->offset
;
3401 ++src_map
->timestamp
;
3403 dst_entry
->object
.vm_object
= NULL
;
3404 dst_entry
->offset
= 0;
3407 pmap_copy(dst_map
->pmap
, src_map
->pmap
, dst_entry
->start
,
3408 dst_entry
->end
- dst_entry
->start
, src_entry
->start
);
3411 * Of course, wired down pages can't be set copy-on-write.
3412 * Cause wired pages to be copied into the new map by
3413 * simulating faults (the new pages are pageable)
3415 vm_fault_copy_entry(dst_map
, src_map
, dst_entry
, src_entry
);
3421 * Create a new process vmspace structure and vm_map
3422 * based on those of an existing process. The new map
3423 * is based on the old map, according to the inheritance
3424 * values on the regions in that map.
3426 * The source map must not be locked.
3429 static void vmspace_fork_normal_entry(vm_map_t old_map
, vm_map_t new_map
,
3430 vm_map_entry_t old_entry
, int *countp
);
3431 static void vmspace_fork_uksmap_entry(vm_map_t old_map
, vm_map_t new_map
,
3432 vm_map_entry_t old_entry
, int *countp
);
3435 vmspace_fork(struct vmspace
*vm1
)
3437 struct vmspace
*vm2
;
3438 vm_map_t old_map
= &vm1
->vm_map
;
3440 vm_map_entry_t old_entry
;
3443 lwkt_gettoken(&vm1
->vm_map
.token
);
3444 vm_map_lock(old_map
);
3446 vm2
= vmspace_alloc(old_map
->min_offset
, old_map
->max_offset
);
3447 lwkt_gettoken(&vm2
->vm_map
.token
);
3448 bcopy(&vm1
->vm_startcopy
, &vm2
->vm_startcopy
,
3449 (caddr_t
)&vm1
->vm_endcopy
- (caddr_t
)&vm1
->vm_startcopy
);
3450 new_map
= &vm2
->vm_map
; /* XXX */
3451 new_map
->timestamp
= 1;
3453 vm_map_lock(new_map
);
3456 old_entry
= old_map
->header
.next
;
3457 while (old_entry
!= &old_map
->header
) {
3459 old_entry
= old_entry
->next
;
3462 count
= vm_map_entry_reserve(count
+ MAP_RESERVE_COUNT
);
3464 old_entry
= old_map
->header
.next
;
3465 while (old_entry
!= &old_map
->header
) {
3466 switch(old_entry
->maptype
) {
3467 case VM_MAPTYPE_SUBMAP
:
3468 panic("vm_map_fork: encountered a submap");
3470 case VM_MAPTYPE_UKSMAP
:
3471 vmspace_fork_uksmap_entry(old_map
, new_map
,
3474 case VM_MAPTYPE_NORMAL
:
3475 case VM_MAPTYPE_VPAGETABLE
:
3476 vmspace_fork_normal_entry(old_map
, new_map
,
3480 old_entry
= old_entry
->next
;
3483 new_map
->size
= old_map
->size
;
3484 vm_map_unlock(old_map
);
3485 vm_map_unlock(new_map
);
3486 vm_map_entry_release(count
);
3488 lwkt_reltoken(&vm2
->vm_map
.token
);
3489 lwkt_reltoken(&vm1
->vm_map
.token
);
3496 vmspace_fork_normal_entry(vm_map_t old_map
, vm_map_t new_map
,
3497 vm_map_entry_t old_entry
, int *countp
)
3499 vm_map_entry_t new_entry
;
3502 switch (old_entry
->inheritance
) {
3503 case VM_INHERIT_NONE
:
3505 case VM_INHERIT_SHARE
:
3507 * Clone the entry, creating the shared object if
3510 if (old_entry
->object
.vm_object
== NULL
)
3511 vm_map_entry_allocate_object(old_entry
);
3513 if (old_entry
->eflags
& MAP_ENTRY_NEEDS_COPY
) {
3515 * Shadow a map_entry which needs a copy,
3516 * replacing its object with a new object
3517 * that points to the old one. Ask the
3518 * shadow code to automatically add an
3519 * additional ref. We can't do it afterwords
3520 * because we might race a collapse. The call
3521 * to vm_map_entry_shadow() will also clear
3524 vm_map_entry_shadow(old_entry
, 1);
3525 } else if (old_entry
->object
.vm_object
) {
3527 * We will make a shared copy of the object,
3528 * and must clear OBJ_ONEMAPPING.
3530 * Optimize vnode objects. OBJ_ONEMAPPING
3531 * is non-applicable but clear it anyway,
3532 * and its terminal so we don'th ave to deal
3533 * with chains. Reduces SMP conflicts.
3535 * XXX assert that object.vm_object != NULL
3536 * since we allocate it above.
3538 object
= old_entry
->object
.vm_object
;
3539 if (object
->type
== OBJT_VNODE
) {
3540 vm_object_reference_quick(object
);
3541 vm_object_clear_flag(object
,
3544 vm_object_hold(object
);
3545 vm_object_chain_wait(object
, 0);
3546 vm_object_reference_locked(object
);
3547 vm_object_clear_flag(object
,
3549 vm_object_drop(object
);
3554 * Clone the entry. We've already bumped the ref on
3557 new_entry
= vm_map_entry_create(new_map
, countp
);
3558 *new_entry
= *old_entry
;
3559 new_entry
->eflags
&= ~MAP_ENTRY_USER_WIRED
;
3560 new_entry
->wired_count
= 0;
3563 * Insert the entry into the new map -- we know we're
3564 * inserting at the end of the new map.
3567 vm_map_entry_link(new_map
, new_map
->header
.prev
,
3571 * Update the physical map
3573 pmap_copy(new_map
->pmap
, old_map
->pmap
,
3575 (old_entry
->end
- old_entry
->start
),
3578 case VM_INHERIT_COPY
:
3580 * Clone the entry and link into the map.
3582 new_entry
= vm_map_entry_create(new_map
, countp
);
3583 *new_entry
= *old_entry
;
3584 new_entry
->eflags
&= ~MAP_ENTRY_USER_WIRED
;
3585 new_entry
->wired_count
= 0;
3586 new_entry
->object
.vm_object
= NULL
;
3587 vm_map_entry_link(new_map
, new_map
->header
.prev
,
3589 vm_map_copy_entry(old_map
, new_map
, old_entry
,
3596 * When forking user-kernel shared maps, the map might change in the
3597 * child so do not try to copy the underlying pmap entries.
3601 vmspace_fork_uksmap_entry(vm_map_t old_map
, vm_map_t new_map
,
3602 vm_map_entry_t old_entry
, int *countp
)
3604 vm_map_entry_t new_entry
;
3606 new_entry
= vm_map_entry_create(new_map
, countp
);
3607 *new_entry
= *old_entry
;
3608 new_entry
->eflags
&= ~MAP_ENTRY_USER_WIRED
;
3609 new_entry
->wired_count
= 0;
3610 vm_map_entry_link(new_map
, new_map
->header
.prev
,
3615 * Create an auto-grow stack entry
3620 vm_map_stack (vm_map_t map
, vm_offset_t addrbos
, vm_size_t max_ssize
,
3621 int flags
, vm_prot_t prot
, vm_prot_t max
, int cow
)
3623 vm_map_entry_t prev_entry
;
3624 vm_map_entry_t new_stack_entry
;
3625 vm_size_t init_ssize
;
3628 vm_offset_t tmpaddr
;
3630 cow
|= MAP_IS_STACK
;
3632 if (max_ssize
< sgrowsiz
)
3633 init_ssize
= max_ssize
;
3635 init_ssize
= sgrowsiz
;
3637 count
= vm_map_entry_reserve(MAP_RESERVE_COUNT
);
3641 * Find space for the mapping
3643 if ((flags
& (MAP_FIXED
| MAP_TRYFIXED
)) == 0) {
3644 if (vm_map_findspace(map
, addrbos
, max_ssize
, 1,
3647 vm_map_entry_release(count
);
3648 return (KERN_NO_SPACE
);
3653 /* If addr is already mapped, no go */
3654 if (vm_map_lookup_entry(map
, addrbos
, &prev_entry
)) {
3656 vm_map_entry_release(count
);
3657 return (KERN_NO_SPACE
);
3661 /* XXX already handled by kern_mmap() */
3662 /* If we would blow our VMEM resource limit, no go */
3663 if (map
->size
+ init_ssize
>
3664 curproc
->p_rlimit
[RLIMIT_VMEM
].rlim_cur
) {
3666 vm_map_entry_release(count
);
3667 return (KERN_NO_SPACE
);
3672 * If we can't accomodate max_ssize in the current mapping,
3673 * no go. However, we need to be aware that subsequent user
3674 * mappings might map into the space we have reserved for
3675 * stack, and currently this space is not protected.
3677 * Hopefully we will at least detect this condition
3678 * when we try to grow the stack.
3680 if ((prev_entry
->next
!= &map
->header
) &&
3681 (prev_entry
->next
->start
< addrbos
+ max_ssize
)) {
3683 vm_map_entry_release(count
);
3684 return (KERN_NO_SPACE
);
3688 * We initially map a stack of only init_ssize. We will
3689 * grow as needed later. Since this is to be a grow
3690 * down stack, we map at the top of the range.
3692 * Note: we would normally expect prot and max to be
3693 * VM_PROT_ALL, and cow to be 0. Possibly we should
3694 * eliminate these as input parameters, and just
3695 * pass these values here in the insert call.
3697 rv
= vm_map_insert(map
, &count
, NULL
, NULL
,
3698 0, addrbos
+ max_ssize
- init_ssize
,
3699 addrbos
+ max_ssize
,
3701 VM_SUBSYS_STACK
, prot
, max
, cow
);
3703 /* Now set the avail_ssize amount */
3704 if (rv
== KERN_SUCCESS
) {
3705 if (prev_entry
!= &map
->header
)
3706 vm_map_clip_end(map
, prev_entry
, addrbos
+ max_ssize
- init_ssize
, &count
);
3707 new_stack_entry
= prev_entry
->next
;
3708 if (new_stack_entry
->end
!= addrbos
+ max_ssize
||
3709 new_stack_entry
->start
!= addrbos
+ max_ssize
- init_ssize
)
3710 panic ("Bad entry start/end for new stack entry");
3712 new_stack_entry
->aux
.avail_ssize
= max_ssize
- init_ssize
;
3716 vm_map_entry_release(count
);
3721 * Attempts to grow a vm stack entry. Returns KERN_SUCCESS if the
3722 * desired address is already mapped, or if we successfully grow
3723 * the stack. Also returns KERN_SUCCESS if addr is outside the
3724 * stack range (this is strange, but preserves compatibility with
3725 * the grow function in vm_machdep.c).
3730 vm_map_growstack (vm_map_t map
, vm_offset_t addr
)
3732 vm_map_entry_t prev_entry
;
3733 vm_map_entry_t stack_entry
;
3734 vm_map_entry_t new_stack_entry
;
3740 int rv
= KERN_SUCCESS
;
3742 int use_read_lock
= 1;
3748 lp
= curthread
->td_lwp
;
3749 p
= curthread
->td_proc
;
3750 KKASSERT(lp
!= NULL
);
3751 vm
= lp
->lwp_vmspace
;
3752 KKASSERT(map
== &vm
->vm_map
);
3754 count
= vm_map_entry_reserve(MAP_RESERVE_COUNT
);
3757 vm_map_lock_read(map
);
3761 /* If addr is already in the entry range, no need to grow.*/
3762 if (vm_map_lookup_entry(map
, addr
, &prev_entry
))
3765 if ((stack_entry
= prev_entry
->next
) == &map
->header
)
3767 if (prev_entry
== &map
->header
)
3768 end
= stack_entry
->start
- stack_entry
->aux
.avail_ssize
;
3770 end
= prev_entry
->end
;
3773 * This next test mimics the old grow function in vm_machdep.c.
3774 * It really doesn't quite make sense, but we do it anyway
3775 * for compatibility.
3777 * If not growable stack, return success. This signals the
3778 * caller to proceed as he would normally with normal vm.
3780 if (stack_entry
->aux
.avail_ssize
< 1 ||
3781 addr
>= stack_entry
->start
||
3782 addr
< stack_entry
->start
- stack_entry
->aux
.avail_ssize
) {
3786 /* Find the minimum grow amount */
3787 grow_amount
= roundup (stack_entry
->start
- addr
, PAGE_SIZE
);
3788 if (grow_amount
> stack_entry
->aux
.avail_ssize
) {
3794 * If there is no longer enough space between the entries
3795 * nogo, and adjust the available space. Note: this
3796 * should only happen if the user has mapped into the
3797 * stack area after the stack was created, and is
3798 * probably an error.
3800 * This also effectively destroys any guard page the user
3801 * might have intended by limiting the stack size.
3803 if (grow_amount
> stack_entry
->start
- end
) {
3804 if (use_read_lock
&& vm_map_lock_upgrade(map
)) {
3810 stack_entry
->aux
.avail_ssize
= stack_entry
->start
- end
;
3815 is_procstack
= addr
>= (vm_offset_t
)vm
->vm_maxsaddr
;
3817 /* If this is the main process stack, see if we're over the
3820 if (is_procstack
&& (ctob(vm
->vm_ssize
) + grow_amount
>
3821 p
->p_rlimit
[RLIMIT_STACK
].rlim_cur
)) {
3826 /* Round up the grow amount modulo SGROWSIZ */
3827 grow_amount
= roundup (grow_amount
, sgrowsiz
);
3828 if (grow_amount
> stack_entry
->aux
.avail_ssize
) {
3829 grow_amount
= stack_entry
->aux
.avail_ssize
;
3831 if (is_procstack
&& (ctob(vm
->vm_ssize
) + grow_amount
>
3832 p
->p_rlimit
[RLIMIT_STACK
].rlim_cur
)) {
3833 grow_amount
= p
->p_rlimit
[RLIMIT_STACK
].rlim_cur
-
3837 /* If we would blow our VMEM resource limit, no go */
3838 if (map
->size
+ grow_amount
> p
->p_rlimit
[RLIMIT_VMEM
].rlim_cur
) {
3843 if (use_read_lock
&& vm_map_lock_upgrade(map
)) {
3850 /* Get the preliminary new entry start value */
3851 addr
= stack_entry
->start
- grow_amount
;
3853 /* If this puts us into the previous entry, cut back our growth
3854 * to the available space. Also, see the note above.
3857 stack_entry
->aux
.avail_ssize
= stack_entry
->start
- end
;
3861 rv
= vm_map_insert(map
, &count
, NULL
, NULL
,
3862 0, addr
, stack_entry
->start
,
3864 VM_SUBSYS_STACK
, VM_PROT_ALL
, VM_PROT_ALL
, 0);
3866 /* Adjust the available stack space by the amount we grew. */
3867 if (rv
== KERN_SUCCESS
) {
3868 if (prev_entry
!= &map
->header
)
3869 vm_map_clip_end(map
, prev_entry
, addr
, &count
);
3870 new_stack_entry
= prev_entry
->next
;
3871 if (new_stack_entry
->end
!= stack_entry
->start
||
3872 new_stack_entry
->start
!= addr
)
3873 panic ("Bad stack grow start/end in new stack entry");
3875 new_stack_entry
->aux
.avail_ssize
=
3876 stack_entry
->aux
.avail_ssize
-
3877 (new_stack_entry
->end
- new_stack_entry
->start
);
3879 vm
->vm_ssize
+= btoc(new_stack_entry
->end
-
3880 new_stack_entry
->start
);
3883 if (map
->flags
& MAP_WIREFUTURE
)
3884 vm_map_unwire(map
, new_stack_entry
->start
,
3885 new_stack_entry
->end
, FALSE
);
3890 vm_map_unlock_read(map
);
3893 vm_map_entry_release(count
);
3898 * Unshare the specified VM space for exec. If other processes are
3899 * mapped to it, then create a new one. The new vmspace is null.
3904 vmspace_exec(struct proc
*p
, struct vmspace
*vmcopy
)
3906 struct vmspace
*oldvmspace
= p
->p_vmspace
;
3907 struct vmspace
*newvmspace
;
3908 vm_map_t map
= &p
->p_vmspace
->vm_map
;
3911 * If we are execing a resident vmspace we fork it, otherwise
3912 * we create a new vmspace. Note that exitingcnt is not
3913 * copied to the new vmspace.
3915 lwkt_gettoken(&oldvmspace
->vm_map
.token
);
3917 newvmspace
= vmspace_fork(vmcopy
);
3918 lwkt_gettoken(&newvmspace
->vm_map
.token
);
3920 newvmspace
= vmspace_alloc(map
->min_offset
, map
->max_offset
);
3921 lwkt_gettoken(&newvmspace
->vm_map
.token
);
3922 bcopy(&oldvmspace
->vm_startcopy
, &newvmspace
->vm_startcopy
,
3923 (caddr_t
)&oldvmspace
->vm_endcopy
-
3924 (caddr_t
)&oldvmspace
->vm_startcopy
);
3928 * Finish initializing the vmspace before assigning it
3929 * to the process. The vmspace will become the current vmspace
3932 pmap_pinit2(vmspace_pmap(newvmspace
));
3933 pmap_replacevm(p
, newvmspace
, 0);
3934 lwkt_reltoken(&newvmspace
->vm_map
.token
);
3935 lwkt_reltoken(&oldvmspace
->vm_map
.token
);
3936 vmspace_rel(oldvmspace
);
3940 * Unshare the specified VM space for forcing COW. This
3941 * is called by rfork, for the (RFMEM|RFPROC) == 0 case.
3944 vmspace_unshare(struct proc
*p
)
3946 struct vmspace
*oldvmspace
= p
->p_vmspace
;
3947 struct vmspace
*newvmspace
;
3949 lwkt_gettoken(&oldvmspace
->vm_map
.token
);
3950 if (vmspace_getrefs(oldvmspace
) == 1) {
3951 lwkt_reltoken(&oldvmspace
->vm_map
.token
);
3954 newvmspace
= vmspace_fork(oldvmspace
);
3955 lwkt_gettoken(&newvmspace
->vm_map
.token
);
3956 pmap_pinit2(vmspace_pmap(newvmspace
));
3957 pmap_replacevm(p
, newvmspace
, 0);
3958 lwkt_reltoken(&newvmspace
->vm_map
.token
);
3959 lwkt_reltoken(&oldvmspace
->vm_map
.token
);
3960 vmspace_rel(oldvmspace
);
3964 * vm_map_hint: return the beginning of the best area suitable for
3965 * creating a new mapping with "prot" protection.
3970 vm_map_hint(struct proc
*p
, vm_offset_t addr
, vm_prot_t prot
)
3972 struct vmspace
*vms
= p
->p_vmspace
;
3974 if (!randomize_mmap
|| addr
!= 0) {
3976 * Set a reasonable start point for the hint if it was
3977 * not specified or if it falls within the heap space.
3978 * Hinted mmap()s do not allocate out of the heap space.
3981 (addr
>= round_page((vm_offset_t
)vms
->vm_taddr
) &&
3982 addr
< round_page((vm_offset_t
)vms
->vm_daddr
+ maxdsiz
))) {
3983 addr
= round_page((vm_offset_t
)vms
->vm_daddr
+ maxdsiz
);
3988 addr
= (vm_offset_t
)vms
->vm_daddr
+ MAXDSIZ
;
3989 addr
+= karc4random() & (MIN((256 * 1024 * 1024), MAXDSIZ
) - 1);
3991 return (round_page(addr
));
3995 * Finds the VM object, offset, and protection for a given virtual address
3996 * in the specified map, assuming a page fault of the type specified.
3998 * Leaves the map in question locked for read; return values are guaranteed
3999 * until a vm_map_lookup_done call is performed. Note that the map argument
4000 * is in/out; the returned map must be used in the call to vm_map_lookup_done.
4002 * A handle (out_entry) is returned for use in vm_map_lookup_done, to make
4005 * If a lookup is requested with "write protection" specified, the map may
4006 * be changed to perform virtual copying operations, although the data
4007 * referenced will remain the same.
4012 vm_map_lookup(vm_map_t
*var_map
, /* IN/OUT */
4014 vm_prot_t fault_typea
,
4015 vm_map_entry_t
*out_entry
, /* OUT */
4016 vm_object_t
*object
, /* OUT */
4017 vm_pindex_t
*pindex
, /* OUT */
4018 vm_prot_t
*out_prot
, /* OUT */
4019 boolean_t
*wired
) /* OUT */
4021 vm_map_entry_t entry
;
4022 vm_map_t map
= *var_map
;
4024 vm_prot_t fault_type
= fault_typea
;
4025 int use_read_lock
= 1;
4026 int rv
= KERN_SUCCESS
;
4030 vm_map_lock_read(map
);
4035 * If the map has an interesting hint, try it before calling full
4036 * blown lookup routine.
4043 if ((entry
== &map
->header
) ||
4044 (vaddr
< entry
->start
) || (vaddr
>= entry
->end
)) {
4045 vm_map_entry_t tmp_entry
;
4048 * Entry was either not a valid hint, or the vaddr was not
4049 * contained in the entry, so do a full lookup.
4051 if (!vm_map_lookup_entry(map
, vaddr
, &tmp_entry
)) {
4052 rv
= KERN_INVALID_ADDRESS
;
4063 if (entry
->maptype
== VM_MAPTYPE_SUBMAP
) {
4064 vm_map_t old_map
= map
;
4066 *var_map
= map
= entry
->object
.sub_map
;
4068 vm_map_unlock_read(old_map
);
4070 vm_map_unlock(old_map
);
4076 * Check whether this task is allowed to have this page.
4077 * Note the special case for MAP_ENTRY_COW pages with an override.
4078 * This is to implement a forced COW for debuggers.
4080 if (fault_type
& VM_PROT_OVERRIDE_WRITE
)
4081 prot
= entry
->max_protection
;
4083 prot
= entry
->protection
;
4085 fault_type
&= (VM_PROT_READ
|VM_PROT_WRITE
|VM_PROT_EXECUTE
);
4086 if ((fault_type
& prot
) != fault_type
) {
4087 rv
= KERN_PROTECTION_FAILURE
;
4091 if ((entry
->eflags
& MAP_ENTRY_USER_WIRED
) &&
4092 (entry
->eflags
& MAP_ENTRY_COW
) &&
4093 (fault_type
& VM_PROT_WRITE
) &&
4094 (fault_typea
& VM_PROT_OVERRIDE_WRITE
) == 0) {
4095 rv
= KERN_PROTECTION_FAILURE
;
4100 * If this page is not pageable, we have to get it for all possible
4103 *wired
= (entry
->wired_count
!= 0);
4105 prot
= fault_type
= entry
->protection
;
4108 * Virtual page tables may need to update the accessed (A) bit
4109 * in a page table entry. Upgrade the fault to a write fault for
4110 * that case if the map will support it. If the map does not support
4111 * it the page table entry simply will not be updated.
4113 if (entry
->maptype
== VM_MAPTYPE_VPAGETABLE
) {
4114 if (prot
& VM_PROT_WRITE
)
4115 fault_type
|= VM_PROT_WRITE
;
4118 if (curthread
->td_lwp
&& curthread
->td_lwp
->lwp_vmspace
&&
4119 pmap_emulate_ad_bits(&curthread
->td_lwp
->lwp_vmspace
->vm_pmap
)) {
4120 if ((prot
& VM_PROT_WRITE
) == 0)
4121 fault_type
|= VM_PROT_WRITE
;
4125 * Only NORMAL and VPAGETABLE maps are object-based. UKSMAPs are not.
4127 if (entry
->maptype
!= VM_MAPTYPE_NORMAL
&&
4128 entry
->maptype
!= VM_MAPTYPE_VPAGETABLE
) {
4134 * If the entry was copy-on-write, we either ...
4136 if (entry
->eflags
& MAP_ENTRY_NEEDS_COPY
) {
4138 * If we want to write the page, we may as well handle that
4139 * now since we've got the map locked.
4141 * If we don't need to write the page, we just demote the
4142 * permissions allowed.
4145 if (fault_type
& VM_PROT_WRITE
) {
4147 * Not allowed if TDF_NOFAULT is set as the shadowing
4148 * operation can deadlock against the faulting
4149 * function due to the copy-on-write.
4151 if (curthread
->td_flags
& TDF_NOFAULT
) {
4152 rv
= KERN_FAILURE_NOFAULT
;
4157 * Make a new object, and place it in the object
4158 * chain. Note that no new references have appeared
4159 * -- one just moved from the map to the new
4163 if (use_read_lock
&& vm_map_lock_upgrade(map
)) {
4170 vm_map_entry_shadow(entry
, 0);
4173 * We're attempting to read a copy-on-write page --
4174 * don't allow writes.
4177 prot
&= ~VM_PROT_WRITE
;
4182 * Create an object if necessary.
4184 if (entry
->object
.vm_object
== NULL
&& !map
->system_map
) {
4185 if (use_read_lock
&& vm_map_lock_upgrade(map
)) {
4191 vm_map_entry_allocate_object(entry
);
4195 * Return the object/offset from this entry. If the entry was
4196 * copy-on-write or empty, it has been fixed up.
4198 *object
= entry
->object
.vm_object
;
4201 *pindex
= OFF_TO_IDX((vaddr
- entry
->start
) + entry
->offset
);
4204 * Return whether this is the only map sharing this data. On
4205 * success we return with a read lock held on the map. On failure
4206 * we return with the map unlocked.
4210 if (rv
== KERN_SUCCESS
) {
4211 if (use_read_lock
== 0)
4212 vm_map_lock_downgrade(map
);
4213 } else if (use_read_lock
) {
4214 vm_map_unlock_read(map
);
4222 * Releases locks acquired by a vm_map_lookup()
4223 * (according to the handle returned by that lookup).
4225 * No other requirements.
4228 vm_map_lookup_done(vm_map_t map
, vm_map_entry_t entry
, int count
)
4231 * Unlock the main-level map
4233 vm_map_unlock_read(map
);
4235 vm_map_entry_release(count
);
4238 #include "opt_ddb.h"
4240 #include <sys/kernel.h>
4242 #include <ddb/ddb.h>
4247 DB_SHOW_COMMAND(map
, vm_map_print
)
4250 /* XXX convert args. */
4251 vm_map_t map
= (vm_map_t
)addr
;
4252 boolean_t full
= have_addr
;
4254 vm_map_entry_t entry
;
4256 db_iprintf("Task map %p: pmap=%p, nentries=%d, version=%u\n",
4258 (void *)map
->pmap
, map
->nentries
, map
->timestamp
);
4261 if (!full
&& db_indent
)
4265 for (entry
= map
->header
.next
; entry
!= &map
->header
;
4266 entry
= entry
->next
) {
4267 db_iprintf("map entry %p: start=%p, end=%p\n",
4268 (void *)entry
, (void *)entry
->start
, (void *)entry
->end
);
4271 static char *inheritance_name
[4] =
4272 {"share", "copy", "none", "donate_copy"};
4274 db_iprintf(" prot=%x/%x/%s",
4276 entry
->max_protection
,
4277 inheritance_name
[(int)(unsigned char)
4278 entry
->inheritance
]);
4279 if (entry
->wired_count
!= 0)
4280 db_printf(", wired");
4282 switch(entry
->maptype
) {
4283 case VM_MAPTYPE_SUBMAP
:
4284 /* XXX no %qd in kernel. Truncate entry->offset. */
4285 db_printf(", share=%p, offset=0x%lx\n",
4286 (void *)entry
->object
.sub_map
,
4287 (long)entry
->offset
);
4289 if ((entry
->prev
== &map
->header
) ||
4290 (entry
->prev
->object
.sub_map
!=
4291 entry
->object
.sub_map
)) {
4293 vm_map_print((db_expr_t
)(intptr_t)
4294 entry
->object
.sub_map
,
4299 case VM_MAPTYPE_NORMAL
:
4300 case VM_MAPTYPE_VPAGETABLE
:
4301 /* XXX no %qd in kernel. Truncate entry->offset. */
4302 db_printf(", object=%p, offset=0x%lx",
4303 (void *)entry
->object
.vm_object
,
4304 (long)entry
->offset
);
4305 if (entry
->eflags
& MAP_ENTRY_COW
)
4306 db_printf(", copy (%s)",
4307 (entry
->eflags
& MAP_ENTRY_NEEDS_COPY
) ? "needed" : "done");
4311 if ((entry
->prev
== &map
->header
) ||
4312 (entry
->prev
->object
.vm_object
!=
4313 entry
->object
.vm_object
)) {
4315 vm_object_print((db_expr_t
)(intptr_t)
4316 entry
->object
.vm_object
,
4322 case VM_MAPTYPE_UKSMAP
:
4323 db_printf(", uksmap=%p, offset=0x%lx",
4324 (void *)entry
->object
.uksmap
,
4325 (long)entry
->offset
);
4326 if (entry
->eflags
& MAP_ENTRY_COW
)
4327 db_printf(", copy (%s)",
4328 (entry
->eflags
& MAP_ENTRY_NEEDS_COPY
) ? "needed" : "done");
4344 DB_SHOW_COMMAND(procvm
, procvm
)
4349 p
= (struct proc
*) addr
;
4354 db_printf("p = %p, vmspace = %p, map = %p, pmap = %p\n",
4355 (void *)p
, (void *)p
->p_vmspace
, (void *)&p
->p_vmspace
->vm_map
,
4356 (void *)vmspace_pmap(p
->p_vmspace
));
4358 vm_map_print((db_expr_t
)(intptr_t)&p
->p_vmspace
->vm_map
, 1, 0, NULL
);