4 * Copyright (c) 1991, 1993
5 * The Regents of the University of California. All rights reserved.
7 * This code is derived from software contributed to Berkeley by
8 * The Mach Operating System project at Carnegie-Mellon University.
10 * Redistribution and use in source and binary forms, with or without
11 * modification, are permitted provided that the following conditions
13 * 1. Redistributions of source code must retain the above copyright
14 * notice, this list of conditions and the following disclaimer.
15 * 2. Redistributions in binary form must reproduce the above copyright
16 * notice, this list of conditions and the following disclaimer in the
17 * documentation and/or other materials provided with the distribution.
18 * 3. Neither the name of the University nor the names of its contributors
19 * may be used to endorse or promote products derived from this software
20 * without specific prior written permission.
22 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
23 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
24 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
25 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
26 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
27 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
28 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
29 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
30 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
31 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
34 * from: @(#)vm_map.c 8.3 (Berkeley) 1/12/94
37 * Copyright (c) 1987, 1990 Carnegie-Mellon University.
38 * All rights reserved.
40 * Authors: Avadis Tevanian, Jr., Michael Wayne Young
42 * Permission to use, copy, modify and distribute this software and
43 * its documentation is hereby granted, provided that both the copyright
44 * notice and this permission notice appear in all copies of the
45 * software, derivative works or modified versions, and any portions
46 * thereof, and that both notices appear in supporting documentation.
48 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
49 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
50 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
52 * Carnegie Mellon requests users of this software to return to
54 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU
55 * School of Computer Science
56 * Carnegie Mellon University
57 * Pittsburgh PA 15213-3890
59 * any improvements or extensions that they make and grant Carnegie the
60 * rights to redistribute these changes.
62 * $FreeBSD: src/sys/vm/vm_map.c,v 1.187.2.19 2003/05/27 00:47:02 alc Exp $
66 * Virtual memory mapping module.
69 #include <sys/param.h>
70 #include <sys/systm.h>
71 #include <sys/kernel.h>
73 #include <sys/serialize.h>
75 #include <sys/vmmeter.h>
77 #include <sys/vnode.h>
78 #include <sys/resourcevar.h>
81 #include <sys/malloc.h>
82 #include <sys/objcache.h>
85 #include <vm/vm_param.h>
87 #include <vm/vm_map.h>
88 #include <vm/vm_page.h>
89 #include <vm/vm_object.h>
90 #include <vm/vm_pager.h>
91 #include <vm/vm_kern.h>
92 #include <vm/vm_extern.h>
93 #include <vm/swap_pager.h>
94 #include <vm/vm_zone.h>
96 #include <sys/random.h>
97 #include <sys/sysctl.h>
98 #include <sys/spinlock.h>
100 #include <sys/thread2.h>
101 #include <sys/spinlock2.h>
104 * Virtual memory maps provide for the mapping, protection, and sharing
105 * of virtual memory objects. In addition, this module provides for an
106 * efficient virtual copy of memory from one map to another.
108 * Synchronization is required prior to most operations.
110 * Maps consist of an ordered doubly-linked list of simple entries.
111 * A hint and a RB tree is used to speed-up lookups.
113 * Callers looking to modify maps specify start/end addresses which cause
114 * the related map entry to be clipped if necessary, and then later
115 * recombined if the pieces remained compatible.
117 * Virtual copy operations are performed by copying VM object references
118 * from one map to another, and then marking both regions as copy-on-write.
120 static boolean_t
vmspace_ctor(void *obj
, void *privdata
, int ocflags
);
121 static void vmspace_dtor(void *obj
, void *privdata
);
122 static void vmspace_terminate(struct vmspace
*vm
, int final
);
124 MALLOC_DEFINE(M_VMSPACE
, "vmspace", "vmspace objcache backingstore");
125 static struct objcache
*vmspace_cache
;
128 * per-cpu page table cross mappings are initialized in early boot
129 * and might require a considerable number of vm_map_entry structures.
131 #define MAPENTRYBSP_CACHE (MAXCPU+1)
132 #define MAPENTRYAP_CACHE 8
134 static struct vm_zone mapentzone_store
;
135 static vm_zone_t mapentzone
;
136 static struct vm_object mapentobj
;
138 static struct vm_map_entry map_entry_init
[MAX_MAPENT
];
139 static struct vm_map_entry cpu_map_entry_init_bsp
[MAPENTRYBSP_CACHE
];
140 static struct vm_map_entry cpu_map_entry_init_ap
[MAXCPU
][MAPENTRYAP_CACHE
];
142 static int randomize_mmap
;
143 SYSCTL_INT(_vm
, OID_AUTO
, randomize_mmap
, CTLFLAG_RW
, &randomize_mmap
, 0,
144 "Randomize mmap offsets");
145 static int vm_map_relock_enable
= 1;
146 SYSCTL_INT(_vm
, OID_AUTO
, map_relock_enable
, CTLFLAG_RW
,
147 &vm_map_relock_enable
, 0, "Randomize mmap offsets");
149 static void vmspace_drop_notoken(struct vmspace
*vm
);
150 static void vm_map_entry_shadow(vm_map_entry_t entry
, int addref
);
151 static vm_map_entry_t
vm_map_entry_create(vm_map_t map
, int *);
152 static void vm_map_entry_dispose (vm_map_t map
, vm_map_entry_t entry
, int *);
153 static void _vm_map_clip_end (vm_map_t
, vm_map_entry_t
, vm_offset_t
, int *);
154 static void _vm_map_clip_start (vm_map_t
, vm_map_entry_t
, vm_offset_t
, int *);
155 static void vm_map_entry_delete (vm_map_t
, vm_map_entry_t
, int *);
156 static void vm_map_entry_unwire (vm_map_t
, vm_map_entry_t
);
157 static void vm_map_copy_entry (vm_map_t
, vm_map_t
, vm_map_entry_t
,
159 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
);
162 * Initialize the vm_map module. Must be called before any other vm_map
165 * Map and entry structures are allocated from the general purpose
166 * memory pool with some exceptions:
168 * - The kernel map is allocated statically.
169 * - Initial kernel map entries are allocated out of a static pool.
170 * - We must set ZONE_SPECIAL here or the early boot code can get
171 * stuck if there are >63 cores.
173 * These restrictions are necessary since malloc() uses the
174 * maps and requires map entries.
176 * Called from the low level boot code only.
181 mapentzone
= &mapentzone_store
;
182 zbootinit(mapentzone
, "MAP ENTRY", sizeof (struct vm_map_entry
),
183 map_entry_init
, MAX_MAPENT
);
184 mapentzone_store
.zflags
|= ZONE_SPECIAL
;
188 * Called prior to any vmspace allocations.
190 * Called from the low level boot code only.
195 vmspace_cache
= objcache_create_mbacked(M_VMSPACE
,
196 sizeof(struct vmspace
),
198 vmspace_ctor
, vmspace_dtor
,
200 zinitna(mapentzone
, &mapentobj
, NULL
, 0, 0,
201 ZONE_USE_RESERVE
| ZONE_SPECIAL
);
207 * objcache support. We leave the pmap root cached as long as possible
208 * for performance reasons.
212 vmspace_ctor(void *obj
, void *privdata
, int ocflags
)
214 struct vmspace
*vm
= obj
;
216 bzero(vm
, sizeof(*vm
));
217 vm
->vm_refcnt
= VM_REF_DELETED
;
224 vmspace_dtor(void *obj
, void *privdata
)
226 struct vmspace
*vm
= obj
;
228 KKASSERT(vm
->vm_refcnt
== VM_REF_DELETED
);
229 pmap_puninit(vmspace_pmap(vm
));
233 * Red black tree functions
235 * The caller must hold the related map lock.
237 static int rb_vm_map_compare(vm_map_entry_t a
, vm_map_entry_t b
);
238 RB_GENERATE(vm_map_rb_tree
, vm_map_entry
, rb_entry
, rb_vm_map_compare
);
240 /* a->start is address, and the only field has to be initialized */
242 rb_vm_map_compare(vm_map_entry_t a
, vm_map_entry_t b
)
244 if (a
->start
< b
->start
)
246 else if (a
->start
> b
->start
)
252 * Initialize vmspace ref/hold counts vmspace0. There is a holdcnt for
256 vmspace_initrefs(struct vmspace
*vm
)
263 * Allocate a vmspace structure, including a vm_map and pmap.
264 * Initialize numerous fields. While the initial allocation is zerod,
265 * subsequence reuse from the objcache leaves elements of the structure
266 * intact (particularly the pmap), so portions must be zerod.
268 * Returns a referenced vmspace.
273 vmspace_alloc(vm_offset_t min
, vm_offset_t max
)
277 vm
= objcache_get(vmspace_cache
, M_WAITOK
);
279 bzero(&vm
->vm_startcopy
,
280 (char *)&vm
->vm_endcopy
- (char *)&vm
->vm_startcopy
);
281 vm_map_init(&vm
->vm_map
, min
, max
, NULL
); /* initializes token */
284 * NOTE: hold to acquires token for safety.
286 * On return vmspace is referenced (refs=1, hold=1). That is,
287 * each refcnt also has a holdcnt. There can be additional holds
288 * (holdcnt) above and beyond the refcnt. Finalization is handled in
289 * two stages, one on refs 1->0, and the the second on hold 1->0.
291 KKASSERT(vm
->vm_holdcnt
== 0);
292 KKASSERT(vm
->vm_refcnt
== VM_REF_DELETED
);
293 vmspace_initrefs(vm
);
295 pmap_pinit(vmspace_pmap(vm
)); /* (some fields reused) */
296 vm
->vm_map
.pmap
= vmspace_pmap(vm
); /* XXX */
299 cpu_vmspace_alloc(vm
);
306 * NOTE: Can return 0 if the vmspace is exiting.
309 vmspace_getrefs(struct vmspace
*vm
)
311 return ((int)(vm
->vm_refcnt
& ~VM_REF_DELETED
));
315 vmspace_hold(struct vmspace
*vm
)
317 atomic_add_int(&vm
->vm_holdcnt
, 1);
318 lwkt_gettoken(&vm
->vm_map
.token
);
322 vmspace_drop(struct vmspace
*vm
)
324 lwkt_reltoken(&vm
->vm_map
.token
);
325 vmspace_drop_notoken(vm
);
329 vmspace_drop_notoken(struct vmspace
*vm
)
331 if (atomic_fetchadd_int(&vm
->vm_holdcnt
, -1) == 1) {
332 if (vm
->vm_refcnt
& VM_REF_DELETED
)
333 vmspace_terminate(vm
, 1);
338 * A vmspace object must not be in a terminated state to be able to obtain
339 * additional refs on it.
341 * These are official references to the vmspace, the count is used to check
342 * for vmspace sharing. Foreign accessors should use 'hold' and not 'ref'.
344 * XXX we need to combine hold & ref together into one 64-bit field to allow
345 * holds to prevent stage-1 termination.
348 vmspace_ref(struct vmspace
*vm
)
352 n
= atomic_fetchadd_int(&vm
->vm_refcnt
, 1);
353 KKASSERT((n
& VM_REF_DELETED
) == 0);
357 * Release a ref on the vmspace. On the 1->0 transition we do stage-1
358 * termination of the vmspace. Then, on the final drop of the hold we
359 * will do stage-2 final termination.
362 vmspace_rel(struct vmspace
*vm
)
369 KKASSERT((int)n
> 0); /* at least one ref & not deleted */
373 * We must have a hold first to interlock the
374 * VM_REF_DELETED check that the drop tests.
376 atomic_add_int(&vm
->vm_holdcnt
, 1);
377 if (atomic_cmpset_int(&vm
->vm_refcnt
, n
,
379 vmspace_terminate(vm
, 0);
380 vmspace_drop_notoken(vm
);
383 vmspace_drop_notoken(vm
);
384 } else if (atomic_cmpset_int(&vm
->vm_refcnt
, n
, n
- 1)) {
391 * This is called during exit indicating that the vmspace is no
392 * longer in used by an exiting process, but the process has not yet
395 * We drop refs, allowing for stage-1 termination, but maintain a holdcnt
396 * to prevent stage-2 until the process is reaped. Note hte order of
397 * operation, we must hold first.
402 vmspace_relexit(struct vmspace
*vm
)
404 atomic_add_int(&vm
->vm_holdcnt
, 1);
409 * Called during reap to disconnect the remainder of the vmspace from
410 * the process. On the hold drop the vmspace termination is finalized.
415 vmspace_exitfree(struct proc
*p
)
421 vmspace_drop_notoken(vm
);
425 * Called in two cases:
427 * (1) When the last refcnt is dropped and the vmspace becomes inactive,
428 * called with final == 0. refcnt will be (u_int)-1 at this point,
429 * and holdcnt will still be non-zero.
431 * (2) When holdcnt becomes 0, called with final == 1. There should no
432 * longer be anyone with access to the vmspace.
434 * VMSPACE_EXIT1 flags the primary deactivation
435 * VMSPACE_EXIT2 flags the last reap
438 vmspace_terminate(struct vmspace
*vm
, int final
)
442 lwkt_gettoken(&vm
->vm_map
.token
);
444 KKASSERT((vm
->vm_flags
& VMSPACE_EXIT1
) == 0);
445 vm
->vm_flags
|= VMSPACE_EXIT1
;
448 * Get rid of most of the resources. Leave the kernel pmap
451 * If the pmap does not contain wired pages we can bulk-delete
452 * the pmap as a performance optimization before removing the
455 * If the pmap contains wired pages we cannot do this
456 * pre-optimization because currently vm_fault_unwire()
457 * expects the pmap pages to exist and will not decrement
458 * p->wire_count if they do not.
461 if (vmspace_pmap(vm
)->pm_stats
.wired_count
) {
462 vm_map_remove(&vm
->vm_map
, VM_MIN_USER_ADDRESS
,
463 VM_MAX_USER_ADDRESS
);
464 pmap_remove_pages(vmspace_pmap(vm
), VM_MIN_USER_ADDRESS
,
465 VM_MAX_USER_ADDRESS
);
467 pmap_remove_pages(vmspace_pmap(vm
), VM_MIN_USER_ADDRESS
,
468 VM_MAX_USER_ADDRESS
);
469 vm_map_remove(&vm
->vm_map
, VM_MIN_USER_ADDRESS
,
470 VM_MAX_USER_ADDRESS
);
472 lwkt_reltoken(&vm
->vm_map
.token
);
474 KKASSERT((vm
->vm_flags
& VMSPACE_EXIT1
) != 0);
475 KKASSERT((vm
->vm_flags
& VMSPACE_EXIT2
) == 0);
478 * Get rid of remaining basic resources.
480 vm
->vm_flags
|= VMSPACE_EXIT2
;
483 count
= vm_map_entry_reserve(MAP_RESERVE_COUNT
);
484 vm_map_lock(&vm
->vm_map
);
485 cpu_vmspace_free(vm
);
488 * Lock the map, to wait out all other references to it.
489 * Delete all of the mappings and pages they hold, then call
490 * the pmap module to reclaim anything left.
492 vm_map_delete(&vm
->vm_map
, vm
->vm_map
.min_offset
,
493 vm
->vm_map
.max_offset
, &count
);
494 vm_map_unlock(&vm
->vm_map
);
495 vm_map_entry_release(count
);
497 pmap_release(vmspace_pmap(vm
));
498 lwkt_reltoken(&vm
->vm_map
.token
);
499 objcache_put(vmspace_cache
, vm
);
504 * Swap useage is determined by taking the proportional swap used by
505 * VM objects backing the VM map. To make up for fractional losses,
506 * if the VM object has any swap use at all the associated map entries
507 * count for at least 1 swap page.
512 vmspace_swap_count(struct vmspace
*vm
)
514 vm_map_t map
= &vm
->vm_map
;
517 vm_offset_t count
= 0;
521 for (cur
= map
->header
.next
; cur
!= &map
->header
; cur
= cur
->next
) {
522 switch(cur
->maptype
) {
523 case VM_MAPTYPE_NORMAL
:
524 case VM_MAPTYPE_VPAGETABLE
:
525 if ((object
= cur
->object
.vm_object
) == NULL
)
527 if (object
->swblock_count
) {
528 n
= (cur
->end
- cur
->start
) / PAGE_SIZE
;
529 count
+= object
->swblock_count
*
530 SWAP_META_PAGES
* n
/ object
->size
+ 1;
543 * Calculate the approximate number of anonymous pages in use by
544 * this vmspace. To make up for fractional losses, we count each
545 * VM object as having at least 1 anonymous page.
550 vmspace_anonymous_count(struct vmspace
*vm
)
552 vm_map_t map
= &vm
->vm_map
;
555 vm_offset_t count
= 0;
558 for (cur
= map
->header
.next
; cur
!= &map
->header
; cur
= cur
->next
) {
559 switch(cur
->maptype
) {
560 case VM_MAPTYPE_NORMAL
:
561 case VM_MAPTYPE_VPAGETABLE
:
562 if ((object
= cur
->object
.vm_object
) == NULL
)
564 if (object
->type
!= OBJT_DEFAULT
&&
565 object
->type
!= OBJT_SWAP
) {
568 count
+= object
->resident_page_count
;
580 * Initialize an existing vm_map structure such as that in the vmspace
581 * structure. The pmap is initialized elsewhere.
586 vm_map_init(struct vm_map
*map
, vm_offset_t min
, vm_offset_t max
, pmap_t pmap
)
588 map
->header
.next
= map
->header
.prev
= &map
->header
;
589 RB_INIT(&map
->rb_root
);
590 spin_init(&map
->ilock_spin
, "ilock");
591 map
->ilock_base
= NULL
;
595 map
->min_offset
= min
;
596 map
->max_offset
= max
;
598 map
->first_free
= &map
->header
;
599 map
->hint
= &map
->header
;
602 lwkt_token_init(&map
->token
, "vm_map");
603 lockinit(&map
->lock
, "vm_maplk", (hz
+ 9) / 10, 0);
607 * Shadow the vm_map_entry's object. This typically needs to be done when
608 * a write fault is taken on an entry which had previously been cloned by
609 * fork(). The shared object (which might be NULL) must become private so
610 * we add a shadow layer above it.
612 * Object allocation for anonymous mappings is defered as long as possible.
613 * When creating a shadow, however, the underlying object must be instantiated
614 * so it can be shared.
616 * If the map segment is governed by a virtual page table then it is
617 * possible to address offsets beyond the mapped area. Just allocate
618 * a maximally sized object for this case.
620 * If addref is non-zero an additional reference is added to the returned
621 * entry. This mechanic exists because the additional reference might have
622 * to be added atomically and not after return to prevent a premature
625 * The vm_map must be exclusively locked.
626 * No other requirements.
630 vm_map_entry_shadow(vm_map_entry_t entry
, int addref
)
632 if (entry
->maptype
== VM_MAPTYPE_VPAGETABLE
) {
633 vm_object_shadow(&entry
->object
.vm_object
, &entry
->offset
,
634 0x7FFFFFFF, addref
); /* XXX */
636 vm_object_shadow(&entry
->object
.vm_object
, &entry
->offset
,
637 atop(entry
->end
- entry
->start
), addref
);
639 entry
->eflags
&= ~MAP_ENTRY_NEEDS_COPY
;
643 * Allocate an object for a vm_map_entry.
645 * Object allocation for anonymous mappings is defered as long as possible.
646 * This function is called when we can defer no longer, generally when a map
647 * entry might be split or forked or takes a page fault.
649 * If the map segment is governed by a virtual page table then it is
650 * possible to address offsets beyond the mapped area. Just allocate
651 * a maximally sized object for this case.
653 * The vm_map must be exclusively locked.
654 * No other requirements.
657 vm_map_entry_allocate_object(vm_map_entry_t entry
)
661 if (entry
->maptype
== VM_MAPTYPE_VPAGETABLE
) {
662 obj
= vm_object_allocate(OBJT_DEFAULT
, 0x7FFFFFFF); /* XXX */
664 obj
= vm_object_allocate(OBJT_DEFAULT
,
665 atop(entry
->end
- entry
->start
));
667 entry
->object
.vm_object
= obj
;
672 * Set an initial negative count so the first attempt to reserve
673 * space preloads a bunch of vm_map_entry's for this cpu. Also
674 * pre-allocate 2 vm_map_entries which will be needed by zalloc() to
675 * map a new page for vm_map_entry structures. SMP systems are
676 * particularly sensitive.
678 * This routine is called in early boot so we cannot just call
679 * vm_map_entry_reserve().
681 * Called from the low level boot code only (for each cpu)
683 * WARNING! Take care not to have too-big a static/BSS structure here
684 * as MAXCPU can be 256+, otherwise the loader's 64MB heap
685 * can get blown out by the kernel plus the initrd image.
688 vm_map_entry_reserve_cpu_init(globaldata_t gd
)
690 vm_map_entry_t entry
;
694 gd
->gd_vme_avail
-= MAP_RESERVE_COUNT
* 2;
695 if (gd
->gd_cpuid
== 0) {
696 entry
= &cpu_map_entry_init_bsp
[0];
697 count
= MAPENTRYBSP_CACHE
;
699 entry
= &cpu_map_entry_init_ap
[gd
->gd_cpuid
][0];
700 count
= MAPENTRYAP_CACHE
;
702 for (i
= 0; i
< count
; ++i
, ++entry
) {
703 entry
->next
= gd
->gd_vme_base
;
704 gd
->gd_vme_base
= entry
;
709 * Reserves vm_map_entry structures so code later on can manipulate
710 * map_entry structures within a locked map without blocking trying
711 * to allocate a new vm_map_entry.
716 vm_map_entry_reserve(int count
)
718 struct globaldata
*gd
= mycpu
;
719 vm_map_entry_t entry
;
722 * Make sure we have enough structures in gd_vme_base to handle
723 * the reservation request.
725 * The critical section protects access to the per-cpu gd.
728 while (gd
->gd_vme_avail
< count
) {
729 entry
= zalloc(mapentzone
);
730 entry
->next
= gd
->gd_vme_base
;
731 gd
->gd_vme_base
= entry
;
734 gd
->gd_vme_avail
-= count
;
741 * Releases previously reserved vm_map_entry structures that were not
742 * used. If we have too much junk in our per-cpu cache clean some of
748 vm_map_entry_release(int count
)
750 struct globaldata
*gd
= mycpu
;
751 vm_map_entry_t entry
;
754 gd
->gd_vme_avail
+= count
;
755 while (gd
->gd_vme_avail
> MAP_RESERVE_SLOP
) {
756 entry
= gd
->gd_vme_base
;
757 KKASSERT(entry
!= NULL
);
758 gd
->gd_vme_base
= entry
->next
;
761 zfree(mapentzone
, entry
);
768 * Reserve map entry structures for use in kernel_map itself. These
769 * entries have *ALREADY* been reserved on a per-cpu basis when the map
770 * was inited. This function is used by zalloc() to avoid a recursion
771 * when zalloc() itself needs to allocate additional kernel memory.
773 * This function works like the normal reserve but does not load the
774 * vm_map_entry cache (because that would result in an infinite
775 * recursion). Note that gd_vme_avail may go negative. This is expected.
777 * Any caller of this function must be sure to renormalize after
778 * potentially eating entries to ensure that the reserve supply
784 vm_map_entry_kreserve(int count
)
786 struct globaldata
*gd
= mycpu
;
789 gd
->gd_vme_avail
-= count
;
791 KASSERT(gd
->gd_vme_base
!= NULL
,
792 ("no reserved entries left, gd_vme_avail = %d",
798 * Release previously reserved map entries for kernel_map. We do not
799 * attempt to clean up like the normal release function as this would
800 * cause an unnecessary (but probably not fatal) deep procedure call.
805 vm_map_entry_krelease(int count
)
807 struct globaldata
*gd
= mycpu
;
810 gd
->gd_vme_avail
+= count
;
815 * Allocates a VM map entry for insertion. No entry fields are filled in.
817 * The entries should have previously been reserved. The reservation count
818 * is tracked in (*countp).
822 static vm_map_entry_t
823 vm_map_entry_create(vm_map_t map
, int *countp
)
825 struct globaldata
*gd
= mycpu
;
826 vm_map_entry_t entry
;
828 KKASSERT(*countp
> 0);
831 entry
= gd
->gd_vme_base
;
832 KASSERT(entry
!= NULL
, ("gd_vme_base NULL! count %d", *countp
));
833 gd
->gd_vme_base
= entry
->next
;
840 * Dispose of a vm_map_entry that is no longer being referenced.
845 vm_map_entry_dispose(vm_map_t map
, vm_map_entry_t entry
, int *countp
)
847 struct globaldata
*gd
= mycpu
;
849 KKASSERT(map
->hint
!= entry
);
850 KKASSERT(map
->first_free
!= entry
);
854 entry
->next
= gd
->gd_vme_base
;
855 gd
->gd_vme_base
= entry
;
861 * Insert/remove entries from maps.
863 * The related map must be exclusively locked.
864 * The caller must hold map->token
865 * No other requirements.
868 vm_map_entry_link(vm_map_t map
,
869 vm_map_entry_t after_where
,
870 vm_map_entry_t entry
)
872 ASSERT_VM_MAP_LOCKED(map
);
875 entry
->prev
= after_where
;
876 entry
->next
= after_where
->next
;
877 entry
->next
->prev
= entry
;
878 after_where
->next
= entry
;
879 if (vm_map_rb_tree_RB_INSERT(&map
->rb_root
, entry
))
880 panic("vm_map_entry_link: dup addr map %p ent %p", map
, entry
);
884 vm_map_entry_unlink(vm_map_t map
,
885 vm_map_entry_t entry
)
890 ASSERT_VM_MAP_LOCKED(map
);
892 if (entry
->eflags
& MAP_ENTRY_IN_TRANSITION
) {
893 panic("vm_map_entry_unlink: attempt to mess with "
894 "locked entry! %p", entry
);
900 vm_map_rb_tree_RB_REMOVE(&map
->rb_root
, entry
);
905 * Finds the map entry containing (or immediately preceding) the specified
906 * address in the given map. The entry is returned in (*entry).
908 * The boolean result indicates whether the address is actually contained
911 * The related map must be locked.
912 * No other requirements.
915 vm_map_lookup_entry(vm_map_t map
, vm_offset_t address
, vm_map_entry_t
*entry
)
920 ASSERT_VM_MAP_LOCKED(map
);
923 * XXX TEMPORARILY DISABLED. For some reason our attempt to revive
924 * the hint code with the red-black lookup meets with system crashes
925 * and lockups. We do not yet know why.
927 * It is possible that the problem is related to the setting
928 * of the hint during map_entry deletion, in the code specified
929 * at the GGG comment later on in this file.
931 * YYY More likely it's because this function can be called with
932 * a shared lock on the map, resulting in map->hint updates possibly
933 * racing. Fixed now but untested.
936 * Quickly check the cached hint, there's a good chance of a match.
940 if (tmp
!= &map
->header
) {
941 if (address
>= tmp
->start
&& address
< tmp
->end
) {
949 * Locate the record from the top of the tree. 'last' tracks the
950 * closest prior record and is returned if no match is found, which
951 * in binary tree terms means tracking the most recent right-branch
952 * taken. If there is no prior record, &map->header is returned.
955 tmp
= RB_ROOT(&map
->rb_root
);
958 if (address
>= tmp
->start
) {
959 if (address
< tmp
->end
) {
965 tmp
= RB_RIGHT(tmp
, rb_entry
);
967 tmp
= RB_LEFT(tmp
, rb_entry
);
975 * Inserts the given whole VM object into the target map at the specified
976 * address range. The object's size should match that of the address range.
978 * The map must be exclusively locked.
979 * The object must be held.
980 * The caller must have reserved sufficient vm_map_entry structures.
982 * If object is non-NULL, ref count must be bumped by caller prior to
983 * making call to account for the new entry.
986 vm_map_insert(vm_map_t map
, int *countp
, void *map_object
, void *map_aux
,
987 vm_ooffset_t offset
, vm_offset_t start
, vm_offset_t end
,
988 vm_maptype_t maptype
, vm_subsys_t id
,
989 vm_prot_t prot
, vm_prot_t max
, int cow
)
991 vm_map_entry_t new_entry
;
992 vm_map_entry_t prev_entry
;
993 vm_map_entry_t temp_entry
;
994 vm_eflags_t protoeflags
;
998 if (maptype
== VM_MAPTYPE_UKSMAP
)
1001 object
= map_object
;
1003 ASSERT_VM_MAP_LOCKED(map
);
1005 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object
));
1008 * Check that the start and end points are not bogus.
1010 if ((start
< map
->min_offset
) || (end
> map
->max_offset
) ||
1012 return (KERN_INVALID_ADDRESS
);
1015 * Find the entry prior to the proposed starting address; if it's part
1016 * of an existing entry, this range is bogus.
1018 if (vm_map_lookup_entry(map
, start
, &temp_entry
))
1019 return (KERN_NO_SPACE
);
1021 prev_entry
= temp_entry
;
1024 * Assert that the next entry doesn't overlap the end point.
1027 if ((prev_entry
->next
!= &map
->header
) &&
1028 (prev_entry
->next
->start
< end
))
1029 return (KERN_NO_SPACE
);
1033 if (cow
& MAP_COPY_ON_WRITE
)
1034 protoeflags
|= MAP_ENTRY_COW
|MAP_ENTRY_NEEDS_COPY
;
1036 if (cow
& MAP_NOFAULT
) {
1037 protoeflags
|= MAP_ENTRY_NOFAULT
;
1039 KASSERT(object
== NULL
,
1040 ("vm_map_insert: paradoxical MAP_NOFAULT request"));
1042 if (cow
& MAP_DISABLE_SYNCER
)
1043 protoeflags
|= MAP_ENTRY_NOSYNC
;
1044 if (cow
& MAP_DISABLE_COREDUMP
)
1045 protoeflags
|= MAP_ENTRY_NOCOREDUMP
;
1046 if (cow
& MAP_IS_STACK
)
1047 protoeflags
|= MAP_ENTRY_STACK
;
1048 if (cow
& MAP_IS_KSTACK
)
1049 protoeflags
|= MAP_ENTRY_KSTACK
;
1051 lwkt_gettoken(&map
->token
);
1055 * When object is non-NULL, it could be shared with another
1056 * process. We have to set or clear OBJ_ONEMAPPING
1059 * NOTE: This flag is only applicable to DEFAULT and SWAP
1060 * objects and will already be clear in other types
1061 * of objects, so a shared object lock is ok for
1064 if ((object
->ref_count
> 1) || (object
->shadow_count
!= 0)) {
1065 vm_object_clear_flag(object
, OBJ_ONEMAPPING
);
1068 else if ((prev_entry
!= &map
->header
) &&
1069 (prev_entry
->eflags
== protoeflags
) &&
1070 (prev_entry
->end
== start
) &&
1071 (prev_entry
->wired_count
== 0) &&
1072 (prev_entry
->id
== id
) &&
1073 prev_entry
->maptype
== maptype
&&
1074 maptype
== VM_MAPTYPE_NORMAL
&&
1075 ((prev_entry
->object
.vm_object
== NULL
) ||
1076 vm_object_coalesce(prev_entry
->object
.vm_object
,
1077 OFF_TO_IDX(prev_entry
->offset
),
1078 (vm_size_t
)(prev_entry
->end
- prev_entry
->start
),
1079 (vm_size_t
)(end
- prev_entry
->end
)))) {
1081 * We were able to extend the object. Determine if we
1082 * can extend the previous map entry to include the
1083 * new range as well.
1085 if ((prev_entry
->inheritance
== VM_INHERIT_DEFAULT
) &&
1086 (prev_entry
->protection
== prot
) &&
1087 (prev_entry
->max_protection
== max
)) {
1088 map
->size
+= (end
- prev_entry
->end
);
1089 prev_entry
->end
= end
;
1090 vm_map_simplify_entry(map
, prev_entry
, countp
);
1091 lwkt_reltoken(&map
->token
);
1092 return (KERN_SUCCESS
);
1096 * If we can extend the object but cannot extend the
1097 * map entry, we have to create a new map entry. We
1098 * must bump the ref count on the extended object to
1099 * account for it. object may be NULL.
1101 * XXX if object is NULL should we set offset to 0 here ?
1103 object
= prev_entry
->object
.vm_object
;
1104 offset
= prev_entry
->offset
+
1105 (prev_entry
->end
- prev_entry
->start
);
1107 vm_object_hold(object
);
1108 vm_object_chain_wait(object
, 0);
1109 vm_object_reference_locked(object
);
1111 map_object
= object
;
1116 * NOTE: if conditionals fail, object can be NULL here. This occurs
1117 * in things like the buffer map where we manage kva but do not manage
1122 * Create a new entry
1125 new_entry
= vm_map_entry_create(map
, countp
);
1126 new_entry
->start
= start
;
1127 new_entry
->end
= end
;
1130 new_entry
->maptype
= maptype
;
1131 new_entry
->eflags
= protoeflags
;
1132 new_entry
->object
.map_object
= map_object
;
1133 new_entry
->aux
.master_pde
= 0; /* in case size is different */
1134 new_entry
->aux
.map_aux
= map_aux
;
1135 new_entry
->offset
= offset
;
1137 new_entry
->inheritance
= VM_INHERIT_DEFAULT
;
1138 new_entry
->protection
= prot
;
1139 new_entry
->max_protection
= max
;
1140 new_entry
->wired_count
= 0;
1143 * Insert the new entry into the list
1146 vm_map_entry_link(map
, prev_entry
, new_entry
);
1147 map
->size
+= new_entry
->end
- new_entry
->start
;
1150 * Update the free space hint. Entries cannot overlap.
1151 * An exact comparison is needed to avoid matching
1152 * against the map->header.
1154 if ((map
->first_free
== prev_entry
) &&
1155 (prev_entry
->end
== new_entry
->start
)) {
1156 map
->first_free
= new_entry
;
1161 * Temporarily removed to avoid MAP_STACK panic, due to
1162 * MAP_STACK being a huge hack. Will be added back in
1163 * when MAP_STACK (and the user stack mapping) is fixed.
1166 * It may be possible to simplify the entry
1168 vm_map_simplify_entry(map
, new_entry
, countp
);
1172 * Try to pre-populate the page table. Mappings governed by virtual
1173 * page tables cannot be prepopulated without a lot of work, so
1176 if ((cow
& (MAP_PREFAULT
|MAP_PREFAULT_PARTIAL
)) &&
1177 maptype
!= VM_MAPTYPE_VPAGETABLE
&&
1178 maptype
!= VM_MAPTYPE_UKSMAP
) {
1180 if (vm_map_relock_enable
&& (cow
& MAP_PREFAULT_RELOCK
)) {
1182 vm_object_lock_swap();
1183 vm_object_drop(object
);
1185 pmap_object_init_pt(map
->pmap
, start
, prot
,
1186 object
, OFF_TO_IDX(offset
), end
- start
,
1187 cow
& MAP_PREFAULT_PARTIAL
);
1189 vm_object_hold(object
);
1190 vm_object_lock_swap();
1194 vm_object_drop(object
);
1196 lwkt_reltoken(&map
->token
);
1197 return (KERN_SUCCESS
);
1201 * Find sufficient space for `length' bytes in the given map, starting at
1202 * `start'. Returns 0 on success, 1 on no space.
1204 * This function will returned an arbitrarily aligned pointer. If no
1205 * particular alignment is required you should pass align as 1. Note that
1206 * the map may return PAGE_SIZE aligned pointers if all the lengths used in
1207 * the map are a multiple of PAGE_SIZE, even if you pass a smaller align
1210 * 'align' should be a power of 2 but is not required to be.
1212 * The map must be exclusively locked.
1213 * No other requirements.
1216 vm_map_findspace(vm_map_t map
, vm_offset_t start
, vm_size_t length
,
1217 vm_size_t align
, int flags
, vm_offset_t
*addr
)
1219 vm_map_entry_t entry
, next
;
1221 vm_offset_t align_mask
;
1223 if (start
< map
->min_offset
)
1224 start
= map
->min_offset
;
1225 if (start
> map
->max_offset
)
1229 * If the alignment is not a power of 2 we will have to use
1230 * a mod/division, set align_mask to a special value.
1232 if ((align
| (align
- 1)) + 1 != (align
<< 1))
1233 align_mask
= (vm_offset_t
)-1;
1235 align_mask
= align
- 1;
1238 * Look for the first possible address; if there's already something
1239 * at this address, we have to start after it.
1241 if (start
== map
->min_offset
) {
1242 if ((entry
= map
->first_free
) != &map
->header
)
1247 if (vm_map_lookup_entry(map
, start
, &tmp
))
1253 * Look through the rest of the map, trying to fit a new region in the
1254 * gap between existing regions, or after the very last region.
1256 for (;; start
= (entry
= next
)->end
) {
1258 * Adjust the proposed start by the requested alignment,
1259 * be sure that we didn't wrap the address.
1261 if (align_mask
== (vm_offset_t
)-1)
1262 end
= roundup(start
, align
);
1264 end
= (start
+ align_mask
) & ~align_mask
;
1269 * Find the end of the proposed new region. Be sure we didn't
1270 * go beyond the end of the map, or wrap around the address.
1271 * Then check to see if this is the last entry or if the
1272 * proposed end fits in the gap between this and the next
1275 end
= start
+ length
;
1276 if (end
> map
->max_offset
|| end
< start
)
1281 * If the next entry's start address is beyond the desired
1282 * end address we may have found a good entry.
1284 * If the next entry is a stack mapping we do not map into
1285 * the stack's reserved space.
1287 * XXX continue to allow mapping into the stack's reserved
1288 * space if doing a MAP_STACK mapping inside a MAP_STACK
1289 * mapping, for backwards compatibility. But the caller
1290 * really should use MAP_STACK | MAP_TRYFIXED if they
1293 if (next
== &map
->header
)
1295 if (next
->start
>= end
) {
1296 if ((next
->eflags
& MAP_ENTRY_STACK
) == 0)
1298 if (flags
& MAP_STACK
)
1300 if (next
->start
- next
->aux
.avail_ssize
>= end
)
1307 * Grow the kernel_map if necessary. pmap_growkernel() will panic
1308 * if it fails. The kernel_map is locked and nothing can steal
1309 * our address space if pmap_growkernel() blocks.
1311 * NOTE: This may be unconditionally called for kldload areas on
1312 * x86_64 because these do not bump kernel_vm_end (which would
1313 * fill 128G worth of page tables!). Therefore we must not
1316 if (map
== &kernel_map
) {
1319 kstop
= round_page(start
+ length
);
1320 if (kstop
> kernel_vm_end
)
1321 pmap_growkernel(start
, kstop
);
1328 * vm_map_find finds an unallocated region in the target address map with
1329 * the given length and allocates it. The search is defined to be first-fit
1330 * from the specified address; the region found is returned in the same
1333 * If object is non-NULL, ref count must be bumped by caller
1334 * prior to making call to account for the new entry.
1336 * No requirements. This function will lock the map temporarily.
1339 vm_map_find(vm_map_t map
, void *map_object
, void *map_aux
,
1340 vm_ooffset_t offset
, vm_offset_t
*addr
,
1341 vm_size_t length
, vm_size_t align
, boolean_t fitit
,
1342 vm_maptype_t maptype
, vm_subsys_t id
,
1343 vm_prot_t prot
, vm_prot_t max
, int cow
)
1350 if (maptype
== VM_MAPTYPE_UKSMAP
)
1353 object
= map_object
;
1357 count
= vm_map_entry_reserve(MAP_RESERVE_COUNT
);
1360 vm_object_hold_shared(object
);
1362 if (vm_map_findspace(map
, start
, length
, align
, 0, addr
)) {
1364 vm_object_drop(object
);
1366 vm_map_entry_release(count
);
1367 return (KERN_NO_SPACE
);
1371 result
= vm_map_insert(map
, &count
, map_object
, map_aux
,
1372 offset
, start
, start
+ length
,
1373 maptype
, id
, prot
, max
, cow
);
1375 vm_object_drop(object
);
1377 vm_map_entry_release(count
);
1383 * Simplify the given map entry by merging with either neighbor. This
1384 * routine also has the ability to merge with both neighbors.
1386 * This routine guarentees that the passed entry remains valid (though
1387 * possibly extended). When merging, this routine may delete one or
1388 * both neighbors. No action is taken on entries which have their
1389 * in-transition flag set.
1391 * The map must be exclusively locked.
1394 vm_map_simplify_entry(vm_map_t map
, vm_map_entry_t entry
, int *countp
)
1396 vm_map_entry_t next
, prev
;
1397 vm_size_t prevsize
, esize
;
1399 if (entry
->eflags
& MAP_ENTRY_IN_TRANSITION
) {
1400 ++mycpu
->gd_cnt
.v_intrans_coll
;
1404 if (entry
->maptype
== VM_MAPTYPE_SUBMAP
)
1406 if (entry
->maptype
== VM_MAPTYPE_UKSMAP
)
1410 if (prev
!= &map
->header
) {
1411 prevsize
= prev
->end
- prev
->start
;
1412 if ( (prev
->end
== entry
->start
) &&
1413 (prev
->maptype
== entry
->maptype
) &&
1414 (prev
->object
.vm_object
== entry
->object
.vm_object
) &&
1415 (!prev
->object
.vm_object
||
1416 (prev
->offset
+ prevsize
== entry
->offset
)) &&
1417 (prev
->eflags
== entry
->eflags
) &&
1418 (prev
->protection
== entry
->protection
) &&
1419 (prev
->max_protection
== entry
->max_protection
) &&
1420 (prev
->inheritance
== entry
->inheritance
) &&
1421 (prev
->id
== entry
->id
) &&
1422 (prev
->wired_count
== entry
->wired_count
)) {
1423 if (map
->first_free
== prev
)
1424 map
->first_free
= entry
;
1425 if (map
->hint
== prev
)
1427 vm_map_entry_unlink(map
, prev
);
1428 entry
->start
= prev
->start
;
1429 entry
->offset
= prev
->offset
;
1430 if (prev
->object
.vm_object
)
1431 vm_object_deallocate(prev
->object
.vm_object
);
1432 vm_map_entry_dispose(map
, prev
, countp
);
1437 if (next
!= &map
->header
) {
1438 esize
= entry
->end
- entry
->start
;
1439 if ((entry
->end
== next
->start
) &&
1440 (next
->maptype
== entry
->maptype
) &&
1441 (next
->object
.vm_object
== entry
->object
.vm_object
) &&
1442 (!entry
->object
.vm_object
||
1443 (entry
->offset
+ esize
== next
->offset
)) &&
1444 (next
->eflags
== entry
->eflags
) &&
1445 (next
->protection
== entry
->protection
) &&
1446 (next
->max_protection
== entry
->max_protection
) &&
1447 (next
->inheritance
== entry
->inheritance
) &&
1448 (next
->id
== entry
->id
) &&
1449 (next
->wired_count
== entry
->wired_count
)) {
1450 if (map
->first_free
== next
)
1451 map
->first_free
= entry
;
1452 if (map
->hint
== next
)
1454 vm_map_entry_unlink(map
, next
);
1455 entry
->end
= next
->end
;
1456 if (next
->object
.vm_object
)
1457 vm_object_deallocate(next
->object
.vm_object
);
1458 vm_map_entry_dispose(map
, next
, countp
);
1464 * Asserts that the given entry begins at or after the specified address.
1465 * If necessary, it splits the entry into two.
1467 #define vm_map_clip_start(map, entry, startaddr, countp) \
1469 if (startaddr > entry->start) \
1470 _vm_map_clip_start(map, entry, startaddr, countp); \
1474 * This routine is called only when it is known that the entry must be split.
1476 * The map must be exclusively locked.
1479 _vm_map_clip_start(vm_map_t map
, vm_map_entry_t entry
, vm_offset_t start
,
1482 vm_map_entry_t new_entry
;
1485 * Split off the front portion -- note that we must insert the new
1486 * entry BEFORE this one, so that this entry has the specified
1490 vm_map_simplify_entry(map
, entry
, countp
);
1493 * If there is no object backing this entry, we might as well create
1494 * one now. If we defer it, an object can get created after the map
1495 * is clipped, and individual objects will be created for the split-up
1496 * map. This is a bit of a hack, but is also about the best place to
1497 * put this improvement.
1499 if (entry
->object
.vm_object
== NULL
&& !map
->system_map
) {
1500 vm_map_entry_allocate_object(entry
);
1503 new_entry
= vm_map_entry_create(map
, countp
);
1504 *new_entry
= *entry
;
1506 new_entry
->end
= start
;
1507 entry
->offset
+= (start
- entry
->start
);
1508 entry
->start
= start
;
1510 vm_map_entry_link(map
, entry
->prev
, new_entry
);
1512 switch(entry
->maptype
) {
1513 case VM_MAPTYPE_NORMAL
:
1514 case VM_MAPTYPE_VPAGETABLE
:
1515 if (new_entry
->object
.vm_object
) {
1516 vm_object_hold(new_entry
->object
.vm_object
);
1517 vm_object_chain_wait(new_entry
->object
.vm_object
, 0);
1518 vm_object_reference_locked(new_entry
->object
.vm_object
);
1519 vm_object_drop(new_entry
->object
.vm_object
);
1528 * Asserts that the given entry ends at or before the specified address.
1529 * If necessary, it splits the entry into two.
1531 * The map must be exclusively locked.
1533 #define vm_map_clip_end(map, entry, endaddr, countp) \
1535 if (endaddr < entry->end) \
1536 _vm_map_clip_end(map, entry, endaddr, countp); \
1540 * This routine is called only when it is known that the entry must be split.
1542 * The map must be exclusively locked.
1545 _vm_map_clip_end(vm_map_t map
, vm_map_entry_t entry
, vm_offset_t end
,
1548 vm_map_entry_t new_entry
;
1551 * If there is no object backing this entry, we might as well create
1552 * one now. If we defer it, an object can get created after the map
1553 * is clipped, and individual objects will be created for the split-up
1554 * map. This is a bit of a hack, but is also about the best place to
1555 * put this improvement.
1558 if (entry
->object
.vm_object
== NULL
&& !map
->system_map
) {
1559 vm_map_entry_allocate_object(entry
);
1563 * Create a new entry and insert it AFTER the specified entry
1566 new_entry
= vm_map_entry_create(map
, countp
);
1567 *new_entry
= *entry
;
1569 new_entry
->start
= entry
->end
= end
;
1570 new_entry
->offset
+= (end
- entry
->start
);
1572 vm_map_entry_link(map
, entry
, new_entry
);
1574 switch(entry
->maptype
) {
1575 case VM_MAPTYPE_NORMAL
:
1576 case VM_MAPTYPE_VPAGETABLE
:
1577 if (new_entry
->object
.vm_object
) {
1578 vm_object_hold(new_entry
->object
.vm_object
);
1579 vm_object_chain_wait(new_entry
->object
.vm_object
, 0);
1580 vm_object_reference_locked(new_entry
->object
.vm_object
);
1581 vm_object_drop(new_entry
->object
.vm_object
);
1590 * Asserts that the starting and ending region addresses fall within the
1591 * valid range for the map.
1593 #define VM_MAP_RANGE_CHECK(map, start, end) \
1595 if (start < vm_map_min(map)) \
1596 start = vm_map_min(map); \
1597 if (end > vm_map_max(map)) \
1598 end = vm_map_max(map); \
1604 * Used to block when an in-transition collison occurs. The map
1605 * is unlocked for the sleep and relocked before the return.
1608 vm_map_transition_wait(vm_map_t map
)
1610 tsleep_interlock(map
, 0);
1612 tsleep(map
, PINTERLOCKED
, "vment", 0);
1617 * When we do blocking operations with the map lock held it is
1618 * possible that a clip might have occured on our in-transit entry,
1619 * requiring an adjustment to the entry in our loop. These macros
1620 * help the pageable and clip_range code deal with the case. The
1621 * conditional costs virtually nothing if no clipping has occured.
1624 #define CLIP_CHECK_BACK(entry, save_start) \
1626 while (entry->start != save_start) { \
1627 entry = entry->prev; \
1628 KASSERT(entry != &map->header, ("bad entry clip")); \
1632 #define CLIP_CHECK_FWD(entry, save_end) \
1634 while (entry->end != save_end) { \
1635 entry = entry->next; \
1636 KASSERT(entry != &map->header, ("bad entry clip")); \
1642 * Clip the specified range and return the base entry. The
1643 * range may cover several entries starting at the returned base
1644 * and the first and last entry in the covering sequence will be
1645 * properly clipped to the requested start and end address.
1647 * If no holes are allowed you should pass the MAP_CLIP_NO_HOLES
1650 * The MAP_ENTRY_IN_TRANSITION flag will be set for the entries
1651 * covered by the requested range.
1653 * The map must be exclusively locked on entry and will remain locked
1654 * on return. If no range exists or the range contains holes and you
1655 * specified that no holes were allowed, NULL will be returned. This
1656 * routine may temporarily unlock the map in order avoid a deadlock when
1661 vm_map_clip_range(vm_map_t map
, vm_offset_t start
, vm_offset_t end
,
1662 int *countp
, int flags
)
1664 vm_map_entry_t start_entry
;
1665 vm_map_entry_t entry
;
1668 * Locate the entry and effect initial clipping. The in-transition
1669 * case does not occur very often so do not try to optimize it.
1672 if (vm_map_lookup_entry(map
, start
, &start_entry
) == FALSE
)
1674 entry
= start_entry
;
1675 if (entry
->eflags
& MAP_ENTRY_IN_TRANSITION
) {
1676 entry
->eflags
|= MAP_ENTRY_NEEDS_WAKEUP
;
1677 ++mycpu
->gd_cnt
.v_intrans_coll
;
1678 ++mycpu
->gd_cnt
.v_intrans_wait
;
1679 vm_map_transition_wait(map
);
1681 * entry and/or start_entry may have been clipped while
1682 * we slept, or may have gone away entirely. We have
1683 * to restart from the lookup.
1689 * Since we hold an exclusive map lock we do not have to restart
1690 * after clipping, even though clipping may block in zalloc.
1692 vm_map_clip_start(map
, entry
, start
, countp
);
1693 vm_map_clip_end(map
, entry
, end
, countp
);
1694 entry
->eflags
|= MAP_ENTRY_IN_TRANSITION
;
1697 * Scan entries covered by the range. When working on the next
1698 * entry a restart need only re-loop on the current entry which
1699 * we have already locked, since 'next' may have changed. Also,
1700 * even though entry is safe, it may have been clipped so we
1701 * have to iterate forwards through the clip after sleeping.
1703 while (entry
->next
!= &map
->header
&& entry
->next
->start
< end
) {
1704 vm_map_entry_t next
= entry
->next
;
1706 if (flags
& MAP_CLIP_NO_HOLES
) {
1707 if (next
->start
> entry
->end
) {
1708 vm_map_unclip_range(map
, start_entry
,
1709 start
, entry
->end
, countp
, flags
);
1714 if (next
->eflags
& MAP_ENTRY_IN_TRANSITION
) {
1715 vm_offset_t save_end
= entry
->end
;
1716 next
->eflags
|= MAP_ENTRY_NEEDS_WAKEUP
;
1717 ++mycpu
->gd_cnt
.v_intrans_coll
;
1718 ++mycpu
->gd_cnt
.v_intrans_wait
;
1719 vm_map_transition_wait(map
);
1722 * clips might have occured while we blocked.
1724 CLIP_CHECK_FWD(entry
, save_end
);
1725 CLIP_CHECK_BACK(start_entry
, start
);
1729 * No restart necessary even though clip_end may block, we
1730 * are holding the map lock.
1732 vm_map_clip_end(map
, next
, end
, countp
);
1733 next
->eflags
|= MAP_ENTRY_IN_TRANSITION
;
1736 if (flags
& MAP_CLIP_NO_HOLES
) {
1737 if (entry
->end
!= end
) {
1738 vm_map_unclip_range(map
, start_entry
,
1739 start
, entry
->end
, countp
, flags
);
1743 return(start_entry
);
1747 * Undo the effect of vm_map_clip_range(). You should pass the same
1748 * flags and the same range that you passed to vm_map_clip_range().
1749 * This code will clear the in-transition flag on the entries and
1750 * wake up anyone waiting. This code will also simplify the sequence
1751 * and attempt to merge it with entries before and after the sequence.
1753 * The map must be locked on entry and will remain locked on return.
1755 * Note that you should also pass the start_entry returned by
1756 * vm_map_clip_range(). However, if you block between the two calls
1757 * with the map unlocked please be aware that the start_entry may
1758 * have been clipped and you may need to scan it backwards to find
1759 * the entry corresponding with the original start address. You are
1760 * responsible for this, vm_map_unclip_range() expects the correct
1761 * start_entry to be passed to it and will KASSERT otherwise.
1765 vm_map_unclip_range(vm_map_t map
, vm_map_entry_t start_entry
,
1766 vm_offset_t start
, vm_offset_t end
,
1767 int *countp
, int flags
)
1769 vm_map_entry_t entry
;
1771 entry
= start_entry
;
1773 KASSERT(entry
->start
== start
, ("unclip_range: illegal base entry"));
1774 while (entry
!= &map
->header
&& entry
->start
< end
) {
1775 KASSERT(entry
->eflags
& MAP_ENTRY_IN_TRANSITION
,
1776 ("in-transition flag not set during unclip on: %p",
1778 KASSERT(entry
->end
<= end
,
1779 ("unclip_range: tail wasn't clipped"));
1780 entry
->eflags
&= ~MAP_ENTRY_IN_TRANSITION
;
1781 if (entry
->eflags
& MAP_ENTRY_NEEDS_WAKEUP
) {
1782 entry
->eflags
&= ~MAP_ENTRY_NEEDS_WAKEUP
;
1785 entry
= entry
->next
;
1789 * Simplification does not block so there is no restart case.
1791 entry
= start_entry
;
1792 while (entry
!= &map
->header
&& entry
->start
< end
) {
1793 vm_map_simplify_entry(map
, entry
, countp
);
1794 entry
= entry
->next
;
1799 * Mark the given range as handled by a subordinate map.
1801 * This range must have been created with vm_map_find(), and no other
1802 * operations may have been performed on this range prior to calling
1805 * Submappings cannot be removed.
1810 vm_map_submap(vm_map_t map
, vm_offset_t start
, vm_offset_t end
, vm_map_t submap
)
1812 vm_map_entry_t entry
;
1813 int result
= KERN_INVALID_ARGUMENT
;
1816 count
= vm_map_entry_reserve(MAP_RESERVE_COUNT
);
1819 VM_MAP_RANGE_CHECK(map
, start
, end
);
1821 if (vm_map_lookup_entry(map
, start
, &entry
)) {
1822 vm_map_clip_start(map
, entry
, start
, &count
);
1824 entry
= entry
->next
;
1827 vm_map_clip_end(map
, entry
, end
, &count
);
1829 if ((entry
->start
== start
) && (entry
->end
== end
) &&
1830 ((entry
->eflags
& MAP_ENTRY_COW
) == 0) &&
1831 (entry
->object
.vm_object
== NULL
)) {
1832 entry
->object
.sub_map
= submap
;
1833 entry
->maptype
= VM_MAPTYPE_SUBMAP
;
1834 result
= KERN_SUCCESS
;
1837 vm_map_entry_release(count
);
1843 * Sets the protection of the specified address region in the target map.
1844 * If "set_max" is specified, the maximum protection is to be set;
1845 * otherwise, only the current protection is affected.
1847 * The protection is not applicable to submaps, but is applicable to normal
1848 * maps and maps governed by virtual page tables. For example, when operating
1849 * on a virtual page table our protection basically controls how COW occurs
1850 * on the backing object, whereas the virtual page table abstraction itself
1851 * is an abstraction for userland.
1856 vm_map_protect(vm_map_t map
, vm_offset_t start
, vm_offset_t end
,
1857 vm_prot_t new_prot
, boolean_t set_max
)
1859 vm_map_entry_t current
;
1860 vm_map_entry_t entry
;
1863 count
= vm_map_entry_reserve(MAP_RESERVE_COUNT
);
1866 VM_MAP_RANGE_CHECK(map
, start
, end
);
1868 if (vm_map_lookup_entry(map
, start
, &entry
)) {
1869 vm_map_clip_start(map
, entry
, start
, &count
);
1871 entry
= entry
->next
;
1875 * Make a first pass to check for protection violations.
1878 while ((current
!= &map
->header
) && (current
->start
< end
)) {
1879 if (current
->maptype
== VM_MAPTYPE_SUBMAP
) {
1881 vm_map_entry_release(count
);
1882 return (KERN_INVALID_ARGUMENT
);
1884 if ((new_prot
& current
->max_protection
) != new_prot
) {
1886 vm_map_entry_release(count
);
1887 return (KERN_PROTECTION_FAILURE
);
1889 current
= current
->next
;
1893 * Go back and fix up protections. [Note that clipping is not
1894 * necessary the second time.]
1898 while ((current
!= &map
->header
) && (current
->start
< end
)) {
1901 vm_map_clip_end(map
, current
, end
, &count
);
1903 old_prot
= current
->protection
;
1905 current
->max_protection
= new_prot
;
1906 current
->protection
= new_prot
& old_prot
;
1908 current
->protection
= new_prot
;
1912 * Update physical map if necessary. Worry about copy-on-write
1913 * here -- CHECK THIS XXX
1916 if (current
->protection
!= old_prot
) {
1917 #define MASK(entry) (((entry)->eflags & MAP_ENTRY_COW) ? ~VM_PROT_WRITE : \
1920 pmap_protect(map
->pmap
, current
->start
,
1922 current
->protection
& MASK(current
));
1926 vm_map_simplify_entry(map
, current
, &count
);
1928 current
= current
->next
;
1932 vm_map_entry_release(count
);
1933 return (KERN_SUCCESS
);
1937 * This routine traverses a processes map handling the madvise
1938 * system call. Advisories are classified as either those effecting
1939 * the vm_map_entry structure, or those effecting the underlying
1942 * The <value> argument is used for extended madvise calls.
1947 vm_map_madvise(vm_map_t map
, vm_offset_t start
, vm_offset_t end
,
1948 int behav
, off_t value
)
1950 vm_map_entry_t current
, entry
;
1956 * Some madvise calls directly modify the vm_map_entry, in which case
1957 * we need to use an exclusive lock on the map and we need to perform
1958 * various clipping operations. Otherwise we only need a read-lock
1961 count
= vm_map_entry_reserve(MAP_RESERVE_COUNT
);
1965 case MADV_SEQUENTIAL
:
1979 vm_map_lock_read(map
);
1982 vm_map_entry_release(count
);
1987 * Locate starting entry and clip if necessary.
1990 VM_MAP_RANGE_CHECK(map
, start
, end
);
1992 if (vm_map_lookup_entry(map
, start
, &entry
)) {
1994 vm_map_clip_start(map
, entry
, start
, &count
);
1996 entry
= entry
->next
;
2001 * madvise behaviors that are implemented in the vm_map_entry.
2003 * We clip the vm_map_entry so that behavioral changes are
2004 * limited to the specified address range.
2006 for (current
= entry
;
2007 (current
!= &map
->header
) && (current
->start
< end
);
2008 current
= current
->next
2010 if (current
->maptype
== VM_MAPTYPE_SUBMAP
)
2013 vm_map_clip_end(map
, current
, end
, &count
);
2017 vm_map_entry_set_behavior(current
, MAP_ENTRY_BEHAV_NORMAL
);
2019 case MADV_SEQUENTIAL
:
2020 vm_map_entry_set_behavior(current
, MAP_ENTRY_BEHAV_SEQUENTIAL
);
2023 vm_map_entry_set_behavior(current
, MAP_ENTRY_BEHAV_RANDOM
);
2026 current
->eflags
|= MAP_ENTRY_NOSYNC
;
2029 current
->eflags
&= ~MAP_ENTRY_NOSYNC
;
2032 current
->eflags
|= MAP_ENTRY_NOCOREDUMP
;
2035 current
->eflags
&= ~MAP_ENTRY_NOCOREDUMP
;
2039 * Set the page directory page for a map
2040 * governed by a virtual page table. Mark
2041 * the entry as being governed by a virtual
2042 * page table if it is not.
2044 * XXX the page directory page is stored
2045 * in the avail_ssize field if the map_entry.
2047 * XXX the map simplification code does not
2048 * compare this field so weird things may
2049 * happen if you do not apply this function
2050 * to the entire mapping governed by the
2051 * virtual page table.
2053 if (current
->maptype
!= VM_MAPTYPE_VPAGETABLE
) {
2057 current
->aux
.master_pde
= value
;
2058 pmap_remove(map
->pmap
,
2059 current
->start
, current
->end
);
2063 * Invalidate the related pmap entries, used
2064 * to flush portions of the real kernel's
2065 * pmap when the caller has removed or
2066 * modified existing mappings in a virtual
2069 * (exclusive locked map version does not
2070 * need the range interlock).
2072 pmap_remove(map
->pmap
,
2073 current
->start
, current
->end
);
2079 vm_map_simplify_entry(map
, current
, &count
);
2087 * madvise behaviors that are implemented in the underlying
2090 * Since we don't clip the vm_map_entry, we have to clip
2091 * the vm_object pindex and count.
2093 * NOTE! These functions are only supported on normal maps,
2094 * except MADV_INVAL which is also supported on
2095 * virtual page tables.
2097 for (current
= entry
;
2098 (current
!= &map
->header
) && (current
->start
< end
);
2099 current
= current
->next
2101 vm_offset_t useStart
;
2103 if (current
->maptype
!= VM_MAPTYPE_NORMAL
&&
2104 (current
->maptype
!= VM_MAPTYPE_VPAGETABLE
||
2105 behav
!= MADV_INVAL
)) {
2109 pindex
= OFF_TO_IDX(current
->offset
);
2110 delta
= atop(current
->end
- current
->start
);
2111 useStart
= current
->start
;
2113 if (current
->start
< start
) {
2114 pindex
+= atop(start
- current
->start
);
2115 delta
-= atop(start
- current
->start
);
2118 if (current
->end
> end
)
2119 delta
-= atop(current
->end
- end
);
2121 if ((vm_spindex_t
)delta
<= 0)
2124 if (behav
== MADV_INVAL
) {
2126 * Invalidate the related pmap entries, used
2127 * to flush portions of the real kernel's
2128 * pmap when the caller has removed or
2129 * modified existing mappings in a virtual
2132 * (shared locked map version needs the
2133 * interlock, see vm_fault()).
2135 struct vm_map_ilock ilock
;
2137 KASSERT(useStart
>= VM_MIN_USER_ADDRESS
&&
2138 useStart
+ ptoa(delta
) <=
2139 VM_MAX_USER_ADDRESS
,
2140 ("Bad range %016jx-%016jx (%016jx)",
2141 useStart
, useStart
+ ptoa(delta
),
2143 vm_map_interlock(map
, &ilock
,
2145 useStart
+ ptoa(delta
));
2146 pmap_remove(map
->pmap
,
2148 useStart
+ ptoa(delta
));
2149 vm_map_deinterlock(map
, &ilock
);
2151 vm_object_madvise(current
->object
.vm_object
,
2152 pindex
, delta
, behav
);
2156 * Try to populate the page table. Mappings governed
2157 * by virtual page tables cannot be pre-populated
2158 * without a lot of work so don't try.
2160 if (behav
== MADV_WILLNEED
&&
2161 current
->maptype
!= VM_MAPTYPE_VPAGETABLE
) {
2162 pmap_object_init_pt(
2165 current
->protection
,
2166 current
->object
.vm_object
,
2168 (count
<< PAGE_SHIFT
),
2169 MAP_PREFAULT_MADVISE
2173 vm_map_unlock_read(map
);
2175 vm_map_entry_release(count
);
2181 * Sets the inheritance of the specified address range in the target map.
2182 * Inheritance affects how the map will be shared with child maps at the
2183 * time of vm_map_fork.
2186 vm_map_inherit(vm_map_t map
, vm_offset_t start
, vm_offset_t end
,
2187 vm_inherit_t new_inheritance
)
2189 vm_map_entry_t entry
;
2190 vm_map_entry_t temp_entry
;
2193 switch (new_inheritance
) {
2194 case VM_INHERIT_NONE
:
2195 case VM_INHERIT_COPY
:
2196 case VM_INHERIT_SHARE
:
2199 return (KERN_INVALID_ARGUMENT
);
2202 count
= vm_map_entry_reserve(MAP_RESERVE_COUNT
);
2205 VM_MAP_RANGE_CHECK(map
, start
, end
);
2207 if (vm_map_lookup_entry(map
, start
, &temp_entry
)) {
2209 vm_map_clip_start(map
, entry
, start
, &count
);
2211 entry
= temp_entry
->next
;
2213 while ((entry
!= &map
->header
) && (entry
->start
< end
)) {
2214 vm_map_clip_end(map
, entry
, end
, &count
);
2216 entry
->inheritance
= new_inheritance
;
2218 vm_map_simplify_entry(map
, entry
, &count
);
2220 entry
= entry
->next
;
2223 vm_map_entry_release(count
);
2224 return (KERN_SUCCESS
);
2228 * Implement the semantics of mlock
2231 vm_map_unwire(vm_map_t map
, vm_offset_t start
, vm_offset_t real_end
,
2232 boolean_t new_pageable
)
2234 vm_map_entry_t entry
;
2235 vm_map_entry_t start_entry
;
2237 int rv
= KERN_SUCCESS
;
2240 count
= vm_map_entry_reserve(MAP_RESERVE_COUNT
);
2242 VM_MAP_RANGE_CHECK(map
, start
, real_end
);
2245 start_entry
= vm_map_clip_range(map
, start
, end
, &count
,
2247 if (start_entry
== NULL
) {
2249 vm_map_entry_release(count
);
2250 return (KERN_INVALID_ADDRESS
);
2253 if (new_pageable
== 0) {
2254 entry
= start_entry
;
2255 while ((entry
!= &map
->header
) && (entry
->start
< end
)) {
2256 vm_offset_t save_start
;
2257 vm_offset_t save_end
;
2260 * Already user wired or hard wired (trivial cases)
2262 if (entry
->eflags
& MAP_ENTRY_USER_WIRED
) {
2263 entry
= entry
->next
;
2266 if (entry
->wired_count
!= 0) {
2267 entry
->wired_count
++;
2268 entry
->eflags
|= MAP_ENTRY_USER_WIRED
;
2269 entry
= entry
->next
;
2274 * A new wiring requires instantiation of appropriate
2275 * management structures and the faulting in of the
2278 if (entry
->maptype
== VM_MAPTYPE_NORMAL
||
2279 entry
->maptype
== VM_MAPTYPE_VPAGETABLE
) {
2280 int copyflag
= entry
->eflags
&
2281 MAP_ENTRY_NEEDS_COPY
;
2282 if (copyflag
&& ((entry
->protection
&
2283 VM_PROT_WRITE
) != 0)) {
2284 vm_map_entry_shadow(entry
, 0);
2285 } else if (entry
->object
.vm_object
== NULL
&&
2287 vm_map_entry_allocate_object(entry
);
2290 entry
->wired_count
++;
2291 entry
->eflags
|= MAP_ENTRY_USER_WIRED
;
2294 * Now fault in the area. Note that vm_fault_wire()
2295 * may release the map lock temporarily, it will be
2296 * relocked on return. The in-transition
2297 * flag protects the entries.
2299 save_start
= entry
->start
;
2300 save_end
= entry
->end
;
2301 rv
= vm_fault_wire(map
, entry
, TRUE
, 0);
2303 CLIP_CHECK_BACK(entry
, save_start
);
2305 KASSERT(entry
->wired_count
== 1, ("bad wired_count on entry"));
2306 entry
->eflags
&= ~MAP_ENTRY_USER_WIRED
;
2307 entry
->wired_count
= 0;
2308 if (entry
->end
== save_end
)
2310 entry
= entry
->next
;
2311 KASSERT(entry
!= &map
->header
, ("bad entry clip during backout"));
2313 end
= save_start
; /* unwire the rest */
2317 * note that even though the entry might have been
2318 * clipped, the USER_WIRED flag we set prevents
2319 * duplication so we do not have to do a
2322 entry
= entry
->next
;
2326 * If we failed fall through to the unwiring section to
2327 * unwire what we had wired so far. 'end' has already
2334 * start_entry might have been clipped if we unlocked the
2335 * map and blocked. No matter how clipped it has gotten
2336 * there should be a fragment that is on our start boundary.
2338 CLIP_CHECK_BACK(start_entry
, start
);
2342 * Deal with the unwiring case.
2346 * This is the unwiring case. We must first ensure that the
2347 * range to be unwired is really wired down. We know there
2350 entry
= start_entry
;
2351 while ((entry
!= &map
->header
) && (entry
->start
< end
)) {
2352 if ((entry
->eflags
& MAP_ENTRY_USER_WIRED
) == 0) {
2353 rv
= KERN_INVALID_ARGUMENT
;
2356 KASSERT(entry
->wired_count
!= 0, ("wired count was 0 with USER_WIRED set! %p", entry
));
2357 entry
= entry
->next
;
2361 * Now decrement the wiring count for each region. If a region
2362 * becomes completely unwired, unwire its physical pages and
2366 * The map entries are processed in a loop, checking to
2367 * make sure the entry is wired and asserting it has a wired
2368 * count. However, another loop was inserted more-or-less in
2369 * the middle of the unwiring path. This loop picks up the
2370 * "entry" loop variable from the first loop without first
2371 * setting it to start_entry. Naturally, the secound loop
2372 * is never entered and the pages backing the entries are
2373 * never unwired. This can lead to a leak of wired pages.
2375 entry
= start_entry
;
2376 while ((entry
!= &map
->header
) && (entry
->start
< end
)) {
2377 KASSERT(entry
->eflags
& MAP_ENTRY_USER_WIRED
,
2378 ("expected USER_WIRED on entry %p", entry
));
2379 entry
->eflags
&= ~MAP_ENTRY_USER_WIRED
;
2380 entry
->wired_count
--;
2381 if (entry
->wired_count
== 0)
2382 vm_fault_unwire(map
, entry
);
2383 entry
= entry
->next
;
2387 vm_map_unclip_range(map
, start_entry
, start
, real_end
, &count
,
2391 vm_map_entry_release(count
);
2396 * Sets the pageability of the specified address range in the target map.
2397 * Regions specified as not pageable require locked-down physical
2398 * memory and physical page maps.
2400 * The map must not be locked, but a reference must remain to the map
2401 * throughout the call.
2403 * This function may be called via the zalloc path and must properly
2404 * reserve map entries for kernel_map.
2409 vm_map_wire(vm_map_t map
, vm_offset_t start
, vm_offset_t real_end
, int kmflags
)
2411 vm_map_entry_t entry
;
2412 vm_map_entry_t start_entry
;
2414 int rv
= KERN_SUCCESS
;
2417 if (kmflags
& KM_KRESERVE
)
2418 count
= vm_map_entry_kreserve(MAP_RESERVE_COUNT
);
2420 count
= vm_map_entry_reserve(MAP_RESERVE_COUNT
);
2422 VM_MAP_RANGE_CHECK(map
, start
, real_end
);
2425 start_entry
= vm_map_clip_range(map
, start
, end
, &count
,
2427 if (start_entry
== NULL
) {
2429 rv
= KERN_INVALID_ADDRESS
;
2432 if ((kmflags
& KM_PAGEABLE
) == 0) {
2436 * 1. Holding the write lock, we create any shadow or zero-fill
2437 * objects that need to be created. Then we clip each map
2438 * entry to the region to be wired and increment its wiring
2439 * count. We create objects before clipping the map entries
2440 * to avoid object proliferation.
2442 * 2. We downgrade to a read lock, and call vm_fault_wire to
2443 * fault in the pages for any newly wired area (wired_count is
2446 * Downgrading to a read lock for vm_fault_wire avoids a
2447 * possible deadlock with another process that may have faulted
2448 * on one of the pages to be wired (it would mark the page busy,
2449 * blocking us, then in turn block on the map lock that we
2450 * hold). Because of problems in the recursive lock package,
2451 * we cannot upgrade to a write lock in vm_map_lookup. Thus,
2452 * any actions that require the write lock must be done
2453 * beforehand. Because we keep the read lock on the map, the
2454 * copy-on-write status of the entries we modify here cannot
2457 entry
= start_entry
;
2458 while ((entry
!= &map
->header
) && (entry
->start
< end
)) {
2460 * Trivial case if the entry is already wired
2462 if (entry
->wired_count
) {
2463 entry
->wired_count
++;
2464 entry
= entry
->next
;
2469 * The entry is being newly wired, we have to setup
2470 * appropriate management structures. A shadow
2471 * object is required for a copy-on-write region,
2472 * or a normal object for a zero-fill region. We
2473 * do not have to do this for entries that point to sub
2474 * maps because we won't hold the lock on the sub map.
2476 if (entry
->maptype
== VM_MAPTYPE_NORMAL
||
2477 entry
->maptype
== VM_MAPTYPE_VPAGETABLE
) {
2478 int copyflag
= entry
->eflags
&
2479 MAP_ENTRY_NEEDS_COPY
;
2480 if (copyflag
&& ((entry
->protection
&
2481 VM_PROT_WRITE
) != 0)) {
2482 vm_map_entry_shadow(entry
, 0);
2483 } else if (entry
->object
.vm_object
== NULL
&&
2485 vm_map_entry_allocate_object(entry
);
2489 entry
->wired_count
++;
2490 entry
= entry
->next
;
2498 * HACK HACK HACK HACK
2500 * vm_fault_wire() temporarily unlocks the map to avoid
2501 * deadlocks. The in-transition flag from vm_map_clip_range
2502 * call should protect us from changes while the map is
2505 * NOTE: Previously this comment stated that clipping might
2506 * still occur while the entry is unlocked, but from
2507 * what I can tell it actually cannot.
2509 * It is unclear whether the CLIP_CHECK_*() calls
2510 * are still needed but we keep them in anyway.
2512 * HACK HACK HACK HACK
2515 entry
= start_entry
;
2516 while (entry
!= &map
->header
&& entry
->start
< end
) {
2518 * If vm_fault_wire fails for any page we need to undo
2519 * what has been done. We decrement the wiring count
2520 * for those pages which have not yet been wired (now)
2521 * and unwire those that have (later).
2523 vm_offset_t save_start
= entry
->start
;
2524 vm_offset_t save_end
= entry
->end
;
2526 if (entry
->wired_count
== 1)
2527 rv
= vm_fault_wire(map
, entry
, FALSE
, kmflags
);
2529 CLIP_CHECK_BACK(entry
, save_start
);
2531 KASSERT(entry
->wired_count
== 1, ("wired_count changed unexpectedly"));
2532 entry
->wired_count
= 0;
2533 if (entry
->end
== save_end
)
2535 entry
= entry
->next
;
2536 KASSERT(entry
!= &map
->header
, ("bad entry clip during backout"));
2541 CLIP_CHECK_FWD(entry
, save_end
);
2542 entry
= entry
->next
;
2546 * If a failure occured undo everything by falling through
2547 * to the unwiring code. 'end' has already been adjusted
2551 kmflags
|= KM_PAGEABLE
;
2554 * start_entry is still IN_TRANSITION but may have been
2555 * clipped since vm_fault_wire() unlocks and relocks the
2556 * map. No matter how clipped it has gotten there should
2557 * be a fragment that is on our start boundary.
2559 CLIP_CHECK_BACK(start_entry
, start
);
2562 if (kmflags
& KM_PAGEABLE
) {
2564 * This is the unwiring case. We must first ensure that the
2565 * range to be unwired is really wired down. We know there
2568 entry
= start_entry
;
2569 while ((entry
!= &map
->header
) && (entry
->start
< end
)) {
2570 if (entry
->wired_count
== 0) {
2571 rv
= KERN_INVALID_ARGUMENT
;
2574 entry
= entry
->next
;
2578 * Now decrement the wiring count for each region. If a region
2579 * becomes completely unwired, unwire its physical pages and
2582 entry
= start_entry
;
2583 while ((entry
!= &map
->header
) && (entry
->start
< end
)) {
2584 entry
->wired_count
--;
2585 if (entry
->wired_count
== 0)
2586 vm_fault_unwire(map
, entry
);
2587 entry
= entry
->next
;
2591 vm_map_unclip_range(map
, start_entry
, start
, real_end
,
2592 &count
, MAP_CLIP_NO_HOLES
);
2596 if (kmflags
& KM_KRESERVE
)
2597 vm_map_entry_krelease(count
);
2599 vm_map_entry_release(count
);
2604 * Mark a newly allocated address range as wired but do not fault in
2605 * the pages. The caller is expected to load the pages into the object.
2607 * The map must be locked on entry and will remain locked on return.
2608 * No other requirements.
2611 vm_map_set_wired_quick(vm_map_t map
, vm_offset_t addr
, vm_size_t size
,
2614 vm_map_entry_t scan
;
2615 vm_map_entry_t entry
;
2617 entry
= vm_map_clip_range(map
, addr
, addr
+ size
,
2618 countp
, MAP_CLIP_NO_HOLES
);
2620 scan
!= &map
->header
&& scan
->start
< addr
+ size
;
2621 scan
= scan
->next
) {
2622 KKASSERT(scan
->wired_count
== 0);
2623 scan
->wired_count
= 1;
2625 vm_map_unclip_range(map
, entry
, addr
, addr
+ size
,
2626 countp
, MAP_CLIP_NO_HOLES
);
2630 * Push any dirty cached pages in the address range to their pager.
2631 * If syncio is TRUE, dirty pages are written synchronously.
2632 * If invalidate is TRUE, any cached pages are freed as well.
2634 * This routine is called by sys_msync()
2636 * Returns an error if any part of the specified range is not mapped.
2641 vm_map_clean(vm_map_t map
, vm_offset_t start
, vm_offset_t end
,
2642 boolean_t syncio
, boolean_t invalidate
)
2644 vm_map_entry_t current
;
2645 vm_map_entry_t entry
;
2649 vm_ooffset_t offset
;
2651 vm_map_lock_read(map
);
2652 VM_MAP_RANGE_CHECK(map
, start
, end
);
2653 if (!vm_map_lookup_entry(map
, start
, &entry
)) {
2654 vm_map_unlock_read(map
);
2655 return (KERN_INVALID_ADDRESS
);
2657 lwkt_gettoken(&map
->token
);
2660 * Make a first pass to check for holes.
2662 for (current
= entry
; current
->start
< end
; current
= current
->next
) {
2663 if (current
->maptype
== VM_MAPTYPE_SUBMAP
) {
2664 lwkt_reltoken(&map
->token
);
2665 vm_map_unlock_read(map
);
2666 return (KERN_INVALID_ARGUMENT
);
2668 if (end
> current
->end
&&
2669 (current
->next
== &map
->header
||
2670 current
->end
!= current
->next
->start
)) {
2671 lwkt_reltoken(&map
->token
);
2672 vm_map_unlock_read(map
);
2673 return (KERN_INVALID_ADDRESS
);
2678 pmap_remove(vm_map_pmap(map
), start
, end
);
2681 * Make a second pass, cleaning/uncaching pages from the indicated
2684 for (current
= entry
; current
->start
< end
; current
= current
->next
) {
2685 offset
= current
->offset
+ (start
- current
->start
);
2686 size
= (end
<= current
->end
? end
: current
->end
) - start
;
2688 switch(current
->maptype
) {
2689 case VM_MAPTYPE_SUBMAP
:
2692 vm_map_entry_t tentry
;
2695 smap
= current
->object
.sub_map
;
2696 vm_map_lock_read(smap
);
2697 vm_map_lookup_entry(smap
, offset
, &tentry
);
2698 tsize
= tentry
->end
- offset
;
2701 object
= tentry
->object
.vm_object
;
2702 offset
= tentry
->offset
+ (offset
- tentry
->start
);
2703 vm_map_unlock_read(smap
);
2706 case VM_MAPTYPE_NORMAL
:
2707 case VM_MAPTYPE_VPAGETABLE
:
2708 object
= current
->object
.vm_object
;
2716 vm_object_hold(object
);
2719 * Note that there is absolutely no sense in writing out
2720 * anonymous objects, so we track down the vnode object
2722 * We invalidate (remove) all pages from the address space
2723 * anyway, for semantic correctness.
2725 * note: certain anonymous maps, such as MAP_NOSYNC maps,
2726 * may start out with a NULL object.
2728 while (object
&& (tobj
= object
->backing_object
) != NULL
) {
2729 vm_object_hold(tobj
);
2730 if (tobj
== object
->backing_object
) {
2731 vm_object_lock_swap();
2732 offset
+= object
->backing_object_offset
;
2733 vm_object_drop(object
);
2735 if (object
->size
< OFF_TO_IDX(offset
+ size
))
2736 size
= IDX_TO_OFF(object
->size
) -
2740 vm_object_drop(tobj
);
2742 if (object
&& (object
->type
== OBJT_VNODE
) &&
2743 (current
->protection
& VM_PROT_WRITE
) &&
2744 (object
->flags
& OBJ_NOMSYNC
) == 0) {
2746 * Flush pages if writing is allowed, invalidate them
2747 * if invalidation requested. Pages undergoing I/O
2748 * will be ignored by vm_object_page_remove().
2750 * We cannot lock the vnode and then wait for paging
2751 * to complete without deadlocking against vm_fault.
2752 * Instead we simply call vm_object_page_remove() and
2753 * allow it to block internally on a page-by-page
2754 * basis when it encounters pages undergoing async
2759 /* no chain wait needed for vnode objects */
2760 vm_object_reference_locked(object
);
2761 vn_lock(object
->handle
, LK_EXCLUSIVE
| LK_RETRY
);
2762 flags
= (syncio
|| invalidate
) ? OBJPC_SYNC
: 0;
2763 flags
|= invalidate
? OBJPC_INVAL
: 0;
2766 * When operating on a virtual page table just
2767 * flush the whole object. XXX we probably ought
2770 switch(current
->maptype
) {
2771 case VM_MAPTYPE_NORMAL
:
2772 vm_object_page_clean(object
,
2774 OFF_TO_IDX(offset
+ size
+ PAGE_MASK
),
2777 case VM_MAPTYPE_VPAGETABLE
:
2778 vm_object_page_clean(object
, 0, 0, flags
);
2781 vn_unlock(((struct vnode
*)object
->handle
));
2782 vm_object_deallocate_locked(object
);
2784 if (object
&& invalidate
&&
2785 ((object
->type
== OBJT_VNODE
) ||
2786 (object
->type
== OBJT_DEVICE
) ||
2787 (object
->type
== OBJT_MGTDEVICE
))) {
2789 ((object
->type
== OBJT_DEVICE
) ||
2790 (object
->type
== OBJT_MGTDEVICE
)) ? FALSE
: TRUE
;
2791 /* no chain wait needed for vnode/device objects */
2792 vm_object_reference_locked(object
);
2793 switch(current
->maptype
) {
2794 case VM_MAPTYPE_NORMAL
:
2795 vm_object_page_remove(object
,
2797 OFF_TO_IDX(offset
+ size
+ PAGE_MASK
),
2800 case VM_MAPTYPE_VPAGETABLE
:
2801 vm_object_page_remove(object
, 0, 0, clean_only
);
2804 vm_object_deallocate_locked(object
);
2808 vm_object_drop(object
);
2811 lwkt_reltoken(&map
->token
);
2812 vm_map_unlock_read(map
);
2814 return (KERN_SUCCESS
);
2818 * Make the region specified by this entry pageable.
2820 * The vm_map must be exclusively locked.
2823 vm_map_entry_unwire(vm_map_t map
, vm_map_entry_t entry
)
2825 entry
->eflags
&= ~MAP_ENTRY_USER_WIRED
;
2826 entry
->wired_count
= 0;
2827 vm_fault_unwire(map
, entry
);
2831 * Deallocate the given entry from the target map.
2833 * The vm_map must be exclusively locked.
2836 vm_map_entry_delete(vm_map_t map
, vm_map_entry_t entry
, int *countp
)
2838 vm_map_entry_unlink(map
, entry
);
2839 map
->size
-= entry
->end
- entry
->start
;
2841 switch(entry
->maptype
) {
2842 case VM_MAPTYPE_NORMAL
:
2843 case VM_MAPTYPE_VPAGETABLE
:
2844 case VM_MAPTYPE_SUBMAP
:
2845 vm_object_deallocate(entry
->object
.vm_object
);
2847 case VM_MAPTYPE_UKSMAP
:
2854 vm_map_entry_dispose(map
, entry
, countp
);
2858 * Deallocates the given address range from the target map.
2860 * The vm_map must be exclusively locked.
2863 vm_map_delete(vm_map_t map
, vm_offset_t start
, vm_offset_t end
, int *countp
)
2866 vm_map_entry_t entry
;
2867 vm_map_entry_t first_entry
;
2869 ASSERT_VM_MAP_LOCKED(map
);
2870 lwkt_gettoken(&map
->token
);
2873 * Find the start of the region, and clip it. Set entry to point
2874 * at the first record containing the requested address or, if no
2875 * such record exists, the next record with a greater address. The
2876 * loop will run from this point until a record beyond the termination
2877 * address is encountered.
2879 * map->hint must be adjusted to not point to anything we delete,
2880 * so set it to the entry prior to the one being deleted.
2882 * GGG see other GGG comment.
2884 if (vm_map_lookup_entry(map
, start
, &first_entry
)) {
2885 entry
= first_entry
;
2886 vm_map_clip_start(map
, entry
, start
, countp
);
2887 map
->hint
= entry
->prev
; /* possible problem XXX */
2889 map
->hint
= first_entry
; /* possible problem XXX */
2890 entry
= first_entry
->next
;
2894 * If a hole opens up prior to the current first_free then
2895 * adjust first_free. As with map->hint, map->first_free
2896 * cannot be left set to anything we might delete.
2898 if (entry
== &map
->header
) {
2899 map
->first_free
= &map
->header
;
2900 } else if (map
->first_free
->start
>= start
) {
2901 map
->first_free
= entry
->prev
;
2905 * Step through all entries in this region
2907 while ((entry
!= &map
->header
) && (entry
->start
< end
)) {
2908 vm_map_entry_t next
;
2910 vm_pindex_t offidxstart
, offidxend
, count
;
2913 * If we hit an in-transition entry we have to sleep and
2914 * retry. It's easier (and not really slower) to just retry
2915 * since this case occurs so rarely and the hint is already
2916 * pointing at the right place. We have to reset the
2917 * start offset so as not to accidently delete an entry
2918 * another process just created in vacated space.
2920 if (entry
->eflags
& MAP_ENTRY_IN_TRANSITION
) {
2921 entry
->eflags
|= MAP_ENTRY_NEEDS_WAKEUP
;
2922 start
= entry
->start
;
2923 ++mycpu
->gd_cnt
.v_intrans_coll
;
2924 ++mycpu
->gd_cnt
.v_intrans_wait
;
2925 vm_map_transition_wait(map
);
2928 vm_map_clip_end(map
, entry
, end
, countp
);
2934 offidxstart
= OFF_TO_IDX(entry
->offset
);
2935 count
= OFF_TO_IDX(e
- s
);
2937 switch(entry
->maptype
) {
2938 case VM_MAPTYPE_NORMAL
:
2939 case VM_MAPTYPE_VPAGETABLE
:
2940 case VM_MAPTYPE_SUBMAP
:
2941 object
= entry
->object
.vm_object
;
2949 * Unwire before removing addresses from the pmap; otherwise,
2950 * unwiring will put the entries back in the pmap.
2952 if (entry
->wired_count
!= 0)
2953 vm_map_entry_unwire(map
, entry
);
2955 offidxend
= offidxstart
+ count
;
2957 if (object
== &kernel_object
) {
2958 vm_object_hold(object
);
2959 vm_object_page_remove(object
, offidxstart
,
2961 vm_object_drop(object
);
2962 } else if (object
&& object
->type
!= OBJT_DEFAULT
&&
2963 object
->type
!= OBJT_SWAP
) {
2965 * vnode object routines cannot be chain-locked,
2966 * but since we aren't removing pages from the
2967 * object here we can use a shared hold.
2969 vm_object_hold_shared(object
);
2970 pmap_remove(map
->pmap
, s
, e
);
2971 vm_object_drop(object
);
2972 } else if (object
) {
2973 vm_object_hold(object
);
2974 vm_object_chain_acquire(object
, 0);
2975 pmap_remove(map
->pmap
, s
, e
);
2977 if (object
!= NULL
&&
2978 object
->ref_count
!= 1 &&
2979 (object
->flags
& (OBJ_NOSPLIT
|OBJ_ONEMAPPING
)) ==
2981 (object
->type
== OBJT_DEFAULT
||
2982 object
->type
== OBJT_SWAP
)) {
2983 vm_object_collapse(object
, NULL
);
2984 vm_object_page_remove(object
, offidxstart
,
2986 if (object
->type
== OBJT_SWAP
) {
2987 swap_pager_freespace(object
,
2991 if (offidxend
>= object
->size
&&
2992 offidxstart
< object
->size
) {
2993 object
->size
= offidxstart
;
2996 vm_object_chain_release(object
);
2997 vm_object_drop(object
);
2998 } else if (entry
->maptype
== VM_MAPTYPE_UKSMAP
) {
2999 pmap_remove(map
->pmap
, s
, e
);
3003 * Delete the entry (which may delete the object) only after
3004 * removing all pmap entries pointing to its pages.
3005 * (Otherwise, its page frames may be reallocated, and any
3006 * modify bits will be set in the wrong object!)
3008 vm_map_entry_delete(map
, entry
, countp
);
3011 lwkt_reltoken(&map
->token
);
3012 return (KERN_SUCCESS
);
3016 * Remove the given address range from the target map.
3017 * This is the exported form of vm_map_delete.
3022 vm_map_remove(vm_map_t map
, vm_offset_t start
, vm_offset_t end
)
3027 count
= vm_map_entry_reserve(MAP_RESERVE_COUNT
);
3029 VM_MAP_RANGE_CHECK(map
, start
, end
);
3030 result
= vm_map_delete(map
, start
, end
, &count
);
3032 vm_map_entry_release(count
);
3038 * Assert that the target map allows the specified privilege on the
3039 * entire address region given. The entire region must be allocated.
3041 * The caller must specify whether the vm_map is already locked or not.
3044 vm_map_check_protection(vm_map_t map
, vm_offset_t start
, vm_offset_t end
,
3045 vm_prot_t protection
, boolean_t have_lock
)
3047 vm_map_entry_t entry
;
3048 vm_map_entry_t tmp_entry
;
3051 if (have_lock
== FALSE
)
3052 vm_map_lock_read(map
);
3054 if (!vm_map_lookup_entry(map
, start
, &tmp_entry
)) {
3055 if (have_lock
== FALSE
)
3056 vm_map_unlock_read(map
);
3062 while (start
< end
) {
3063 if (entry
== &map
->header
) {
3071 if (start
< entry
->start
) {
3076 * Check protection associated with entry.
3079 if ((entry
->protection
& protection
) != protection
) {
3083 /* go to next entry */
3086 entry
= entry
->next
;
3088 if (have_lock
== FALSE
)
3089 vm_map_unlock_read(map
);
3094 * If appropriate this function shadows the original object with a new object
3095 * and moves the VM pages from the original object to the new object.
3096 * The original object will also be collapsed, if possible.
3098 * We can only do this for normal memory objects with a single mapping, and
3099 * it only makes sense to do it if there are 2 or more refs on the original
3100 * object. i.e. typically a memory object that has been extended into
3101 * multiple vm_map_entry's with non-overlapping ranges.
3103 * This makes it easier to remove unused pages and keeps object inheritance
3104 * from being a negative impact on memory usage.
3106 * On return the (possibly new) entry->object.vm_object will have an
3107 * additional ref on it for the caller to dispose of (usually by cloning
3108 * the vm_map_entry). The additional ref had to be done in this routine
3109 * to avoid racing a collapse. The object's ONEMAPPING flag will also be
3112 * The vm_map must be locked and its token held.
3115 vm_map_split(vm_map_entry_t entry
)
3118 vm_object_t oobject
, nobject
, bobject
;
3121 vm_pindex_t offidxstart
, offidxend
, idx
;
3123 vm_ooffset_t offset
;
3127 * Optimize away object locks for vnode objects. Important exit/exec
3130 * OBJ_ONEMAPPING doesn't apply to vnode objects but clear the flag
3133 oobject
= entry
->object
.vm_object
;
3134 if (oobject
->type
!= OBJT_DEFAULT
&& oobject
->type
!= OBJT_SWAP
) {
3135 vm_object_reference_quick(oobject
);
3136 vm_object_clear_flag(oobject
, OBJ_ONEMAPPING
);
3141 * Setup. Chain lock the original object throughout the entire
3142 * routine to prevent new page faults from occuring.
3144 * XXX can madvise WILLNEED interfere with us too?
3146 vm_object_hold(oobject
);
3147 vm_object_chain_acquire(oobject
, 0);
3150 * Original object cannot be split? Might have also changed state.
3152 if (oobject
->handle
== NULL
|| (oobject
->type
!= OBJT_DEFAULT
&&
3153 oobject
->type
!= OBJT_SWAP
)) {
3154 vm_object_chain_release(oobject
);
3155 vm_object_reference_locked(oobject
);
3156 vm_object_clear_flag(oobject
, OBJ_ONEMAPPING
);
3157 vm_object_drop(oobject
);
3162 * Collapse original object with its backing store as an
3163 * optimization to reduce chain lengths when possible.
3165 * If ref_count <= 1 there aren't other non-overlapping vm_map_entry's
3166 * for oobject, so there's no point collapsing it.
3168 * Then re-check whether the object can be split.
3170 vm_object_collapse(oobject
, NULL
);
3172 if (oobject
->ref_count
<= 1 ||
3173 (oobject
->type
!= OBJT_DEFAULT
&& oobject
->type
!= OBJT_SWAP
) ||
3174 (oobject
->flags
& (OBJ_NOSPLIT
|OBJ_ONEMAPPING
)) != OBJ_ONEMAPPING
) {
3175 vm_object_chain_release(oobject
);
3176 vm_object_reference_locked(oobject
);
3177 vm_object_clear_flag(oobject
, OBJ_ONEMAPPING
);
3178 vm_object_drop(oobject
);
3183 * Acquire the chain lock on the backing object.
3185 * Give bobject an additional ref count for when it will be shadowed
3189 if ((bobject
= oobject
->backing_object
) != NULL
) {
3190 if (bobject
->type
!= OBJT_VNODE
) {
3192 vm_object_hold(bobject
);
3193 vm_object_chain_wait(bobject
, 0);
3194 /* ref for shadowing below */
3195 vm_object_reference_locked(bobject
);
3196 vm_object_chain_acquire(bobject
, 0);
3197 KKASSERT(bobject
->backing_object
== bobject
);
3198 KKASSERT((bobject
->flags
& OBJ_DEAD
) == 0);
3201 * vnodes are not placed on the shadow list but
3202 * they still get another ref for the backing_object
3205 vm_object_reference_quick(bobject
);
3210 * Calculate the object page range and allocate the new object.
3212 offset
= entry
->offset
;
3216 offidxstart
= OFF_TO_IDX(offset
);
3217 offidxend
= offidxstart
+ OFF_TO_IDX(e
- s
);
3218 size
= offidxend
- offidxstart
;
3220 switch(oobject
->type
) {
3222 nobject
= default_pager_alloc(NULL
, IDX_TO_OFF(size
),
3226 nobject
= swap_pager_alloc(NULL
, IDX_TO_OFF(size
),
3235 if (nobject
== NULL
) {
3237 if (useshadowlist
) {
3238 vm_object_chain_release(bobject
);
3239 vm_object_deallocate(bobject
);
3240 vm_object_drop(bobject
);
3242 vm_object_deallocate(bobject
);
3245 vm_object_chain_release(oobject
);
3246 vm_object_reference_locked(oobject
);
3247 vm_object_clear_flag(oobject
, OBJ_ONEMAPPING
);
3248 vm_object_drop(oobject
);
3253 * The new object will replace entry->object.vm_object so it needs
3254 * a second reference (the caller expects an additional ref).
3256 vm_object_hold(nobject
);
3257 vm_object_reference_locked(nobject
);
3258 vm_object_chain_acquire(nobject
, 0);
3261 * nobject shadows bobject (oobject already shadows bobject).
3263 * Adding an object to bobject's shadow list requires refing bobject
3264 * which we did above in the useshadowlist case.
3267 nobject
->backing_object_offset
=
3268 oobject
->backing_object_offset
+ IDX_TO_OFF(offidxstart
);
3269 nobject
->backing_object
= bobject
;
3270 if (useshadowlist
) {
3271 bobject
->shadow_count
++;
3272 atomic_add_int(&bobject
->generation
, 1);
3273 LIST_INSERT_HEAD(&bobject
->shadow_head
,
3274 nobject
, shadow_list
);
3275 vm_object_clear_flag(bobject
, OBJ_ONEMAPPING
); /*XXX*/
3276 vm_object_chain_release(bobject
);
3277 vm_object_drop(bobject
);
3278 vm_object_set_flag(nobject
, OBJ_ONSHADOW
);
3283 * Move the VM pages from oobject to nobject
3285 for (idx
= 0; idx
< size
; idx
++) {
3288 m
= vm_page_lookup_busy_wait(oobject
, offidxstart
+ idx
,
3294 * We must wait for pending I/O to complete before we can
3297 * We do not have to VM_PROT_NONE the page as mappings should
3298 * not be changed by this operation.
3300 * NOTE: The act of renaming a page updates chaingen for both
3303 vm_page_rename(m
, nobject
, idx
);
3304 /* page automatically made dirty by rename and cache handled */
3305 /* page remains busy */
3308 if (oobject
->type
== OBJT_SWAP
) {
3309 vm_object_pip_add(oobject
, 1);
3311 * copy oobject pages into nobject and destroy unneeded
3312 * pages in shadow object.
3314 swap_pager_copy(oobject
, nobject
, offidxstart
, 0);
3315 vm_object_pip_wakeup(oobject
);
3319 * Wakeup the pages we played with. No spl protection is needed
3320 * for a simple wakeup.
3322 for (idx
= 0; idx
< size
; idx
++) {
3323 m
= vm_page_lookup(nobject
, idx
);
3325 KKASSERT(m
->flags
& PG_BUSY
);
3329 entry
->object
.vm_object
= nobject
;
3330 entry
->offset
= 0LL;
3335 * NOTE: There is no need to remove OBJ_ONEMAPPING from oobject, the
3336 * related pages were moved and are no longer applicable to the
3339 * NOTE: Deallocate oobject (due to its entry->object.vm_object being
3340 * replaced by nobject).
3342 vm_object_chain_release(nobject
);
3343 vm_object_drop(nobject
);
3344 if (bobject
&& useshadowlist
) {
3345 vm_object_chain_release(bobject
);
3346 vm_object_drop(bobject
);
3348 vm_object_chain_release(oobject
);
3349 /*vm_object_clear_flag(oobject, OBJ_ONEMAPPING);*/
3350 vm_object_deallocate_locked(oobject
);
3351 vm_object_drop(oobject
);
3355 * Copies the contents of the source entry to the destination
3356 * entry. The entries *must* be aligned properly.
3358 * The vm_maps must be exclusively locked.
3359 * The vm_map's token must be held.
3361 * Because the maps are locked no faults can be in progress during the
3365 vm_map_copy_entry(vm_map_t src_map
, vm_map_t dst_map
,
3366 vm_map_entry_t src_entry
, vm_map_entry_t dst_entry
)
3368 vm_object_t src_object
;
3370 if (dst_entry
->maptype
== VM_MAPTYPE_SUBMAP
||
3371 dst_entry
->maptype
== VM_MAPTYPE_UKSMAP
)
3373 if (src_entry
->maptype
== VM_MAPTYPE_SUBMAP
||
3374 src_entry
->maptype
== VM_MAPTYPE_UKSMAP
)
3377 if (src_entry
->wired_count
== 0) {
3379 * If the source entry is marked needs_copy, it is already
3382 if ((src_entry
->eflags
& MAP_ENTRY_NEEDS_COPY
) == 0) {
3383 pmap_protect(src_map
->pmap
,
3386 src_entry
->protection
& ~VM_PROT_WRITE
);
3390 * Make a copy of the object.
3392 * The object must be locked prior to checking the object type
3393 * and for the call to vm_object_collapse() and vm_map_split().
3394 * We cannot use *_hold() here because the split code will
3395 * probably try to destroy the object. The lock is a pool
3396 * token and doesn't care.
3398 * We must bump src_map->timestamp when setting
3399 * MAP_ENTRY_NEEDS_COPY to force any concurrent fault
3400 * to retry, otherwise the concurrent fault might improperly
3401 * install a RW pte when its supposed to be a RO(COW) pte.
3402 * This race can occur because a vnode-backed fault may have
3403 * to temporarily release the map lock.
3405 if (src_entry
->object
.vm_object
!= NULL
) {
3406 vm_map_split(src_entry
);
3407 src_object
= src_entry
->object
.vm_object
;
3408 dst_entry
->object
.vm_object
= src_object
;
3409 src_entry
->eflags
|= (MAP_ENTRY_COW
|
3410 MAP_ENTRY_NEEDS_COPY
);
3411 dst_entry
->eflags
|= (MAP_ENTRY_COW
|
3412 MAP_ENTRY_NEEDS_COPY
);
3413 dst_entry
->offset
= src_entry
->offset
;
3414 ++src_map
->timestamp
;
3416 dst_entry
->object
.vm_object
= NULL
;
3417 dst_entry
->offset
= 0;
3420 pmap_copy(dst_map
->pmap
, src_map
->pmap
, dst_entry
->start
,
3421 dst_entry
->end
- dst_entry
->start
, src_entry
->start
);
3424 * Of course, wired down pages can't be set copy-on-write.
3425 * Cause wired pages to be copied into the new map by
3426 * simulating faults (the new pages are pageable)
3428 vm_fault_copy_entry(dst_map
, src_map
, dst_entry
, src_entry
);
3434 * Create a new process vmspace structure and vm_map
3435 * based on those of an existing process. The new map
3436 * is based on the old map, according to the inheritance
3437 * values on the regions in that map.
3439 * The source map must not be locked.
3442 static void vmspace_fork_normal_entry(vm_map_t old_map
, vm_map_t new_map
,
3443 vm_map_entry_t old_entry
, int *countp
);
3444 static void vmspace_fork_uksmap_entry(vm_map_t old_map
, vm_map_t new_map
,
3445 vm_map_entry_t old_entry
, int *countp
);
3448 vmspace_fork(struct vmspace
*vm1
)
3450 struct vmspace
*vm2
;
3451 vm_map_t old_map
= &vm1
->vm_map
;
3453 vm_map_entry_t old_entry
;
3456 lwkt_gettoken(&vm1
->vm_map
.token
);
3457 vm_map_lock(old_map
);
3459 vm2
= vmspace_alloc(old_map
->min_offset
, old_map
->max_offset
);
3460 lwkt_gettoken(&vm2
->vm_map
.token
);
3461 bcopy(&vm1
->vm_startcopy
, &vm2
->vm_startcopy
,
3462 (caddr_t
)&vm1
->vm_endcopy
- (caddr_t
)&vm1
->vm_startcopy
);
3463 new_map
= &vm2
->vm_map
; /* XXX */
3464 new_map
->timestamp
= 1;
3466 vm_map_lock(new_map
);
3469 old_entry
= old_map
->header
.next
;
3470 while (old_entry
!= &old_map
->header
) {
3472 old_entry
= old_entry
->next
;
3475 count
= vm_map_entry_reserve(count
+ MAP_RESERVE_COUNT
);
3477 old_entry
= old_map
->header
.next
;
3478 while (old_entry
!= &old_map
->header
) {
3479 switch(old_entry
->maptype
) {
3480 case VM_MAPTYPE_SUBMAP
:
3481 panic("vm_map_fork: encountered a submap");
3483 case VM_MAPTYPE_UKSMAP
:
3484 vmspace_fork_uksmap_entry(old_map
, new_map
,
3487 case VM_MAPTYPE_NORMAL
:
3488 case VM_MAPTYPE_VPAGETABLE
:
3489 vmspace_fork_normal_entry(old_map
, new_map
,
3493 old_entry
= old_entry
->next
;
3496 new_map
->size
= old_map
->size
;
3497 vm_map_unlock(old_map
);
3498 vm_map_unlock(new_map
);
3499 vm_map_entry_release(count
);
3501 lwkt_reltoken(&vm2
->vm_map
.token
);
3502 lwkt_reltoken(&vm1
->vm_map
.token
);
3509 vmspace_fork_normal_entry(vm_map_t old_map
, vm_map_t new_map
,
3510 vm_map_entry_t old_entry
, int *countp
)
3512 vm_map_entry_t new_entry
;
3515 switch (old_entry
->inheritance
) {
3516 case VM_INHERIT_NONE
:
3518 case VM_INHERIT_SHARE
:
3520 * Clone the entry, creating the shared object if
3523 if (old_entry
->object
.vm_object
== NULL
)
3524 vm_map_entry_allocate_object(old_entry
);
3526 if (old_entry
->eflags
& MAP_ENTRY_NEEDS_COPY
) {
3528 * Shadow a map_entry which needs a copy,
3529 * replacing its object with a new object
3530 * that points to the old one. Ask the
3531 * shadow code to automatically add an
3532 * additional ref. We can't do it afterwords
3533 * because we might race a collapse. The call
3534 * to vm_map_entry_shadow() will also clear
3537 vm_map_entry_shadow(old_entry
, 1);
3538 } else if (old_entry
->object
.vm_object
) {
3540 * We will make a shared copy of the object,
3541 * and must clear OBJ_ONEMAPPING.
3543 * Optimize vnode objects. OBJ_ONEMAPPING
3544 * is non-applicable but clear it anyway,
3545 * and its terminal so we don'th ave to deal
3546 * with chains. Reduces SMP conflicts.
3548 * XXX assert that object.vm_object != NULL
3549 * since we allocate it above.
3551 object
= old_entry
->object
.vm_object
;
3552 if (object
->type
== OBJT_VNODE
) {
3553 vm_object_reference_quick(object
);
3554 vm_object_clear_flag(object
,
3557 vm_object_hold(object
);
3558 vm_object_chain_wait(object
, 0);
3559 vm_object_reference_locked(object
);
3560 vm_object_clear_flag(object
,
3562 vm_object_drop(object
);
3567 * Clone the entry. We've already bumped the ref on
3570 new_entry
= vm_map_entry_create(new_map
, countp
);
3571 *new_entry
= *old_entry
;
3572 new_entry
->eflags
&= ~MAP_ENTRY_USER_WIRED
;
3573 new_entry
->wired_count
= 0;
3576 * Insert the entry into the new map -- we know we're
3577 * inserting at the end of the new map.
3580 vm_map_entry_link(new_map
, new_map
->header
.prev
,
3584 * Update the physical map
3586 pmap_copy(new_map
->pmap
, old_map
->pmap
,
3588 (old_entry
->end
- old_entry
->start
),
3591 case VM_INHERIT_COPY
:
3593 * Clone the entry and link into the map.
3595 new_entry
= vm_map_entry_create(new_map
, countp
);
3596 *new_entry
= *old_entry
;
3597 new_entry
->eflags
&= ~MAP_ENTRY_USER_WIRED
;
3598 new_entry
->wired_count
= 0;
3599 new_entry
->object
.vm_object
= NULL
;
3600 vm_map_entry_link(new_map
, new_map
->header
.prev
,
3602 vm_map_copy_entry(old_map
, new_map
, old_entry
,
3609 * When forking user-kernel shared maps, the map might change in the
3610 * child so do not try to copy the underlying pmap entries.
3614 vmspace_fork_uksmap_entry(vm_map_t old_map
, vm_map_t new_map
,
3615 vm_map_entry_t old_entry
, int *countp
)
3617 vm_map_entry_t new_entry
;
3619 new_entry
= vm_map_entry_create(new_map
, countp
);
3620 *new_entry
= *old_entry
;
3621 new_entry
->eflags
&= ~MAP_ENTRY_USER_WIRED
;
3622 new_entry
->wired_count
= 0;
3623 vm_map_entry_link(new_map
, new_map
->header
.prev
,
3628 * Create an auto-grow stack entry
3633 vm_map_stack (vm_map_t map
, vm_offset_t addrbos
, vm_size_t max_ssize
,
3634 int flags
, vm_prot_t prot
, vm_prot_t max
, int cow
)
3636 vm_map_entry_t prev_entry
;
3637 vm_map_entry_t new_stack_entry
;
3638 vm_size_t init_ssize
;
3641 vm_offset_t tmpaddr
;
3643 cow
|= MAP_IS_STACK
;
3645 if (max_ssize
< sgrowsiz
)
3646 init_ssize
= max_ssize
;
3648 init_ssize
= sgrowsiz
;
3650 count
= vm_map_entry_reserve(MAP_RESERVE_COUNT
);
3654 * Find space for the mapping
3656 if ((flags
& (MAP_FIXED
| MAP_TRYFIXED
)) == 0) {
3657 if (vm_map_findspace(map
, addrbos
, max_ssize
, 1,
3660 vm_map_entry_release(count
);
3661 return (KERN_NO_SPACE
);
3666 /* If addr is already mapped, no go */
3667 if (vm_map_lookup_entry(map
, addrbos
, &prev_entry
)) {
3669 vm_map_entry_release(count
);
3670 return (KERN_NO_SPACE
);
3674 /* XXX already handled by kern_mmap() */
3675 /* If we would blow our VMEM resource limit, no go */
3676 if (map
->size
+ init_ssize
>
3677 curproc
->p_rlimit
[RLIMIT_VMEM
].rlim_cur
) {
3679 vm_map_entry_release(count
);
3680 return (KERN_NO_SPACE
);
3685 * If we can't accomodate max_ssize in the current mapping,
3686 * no go. However, we need to be aware that subsequent user
3687 * mappings might map into the space we have reserved for
3688 * stack, and currently this space is not protected.
3690 * Hopefully we will at least detect this condition
3691 * when we try to grow the stack.
3693 if ((prev_entry
->next
!= &map
->header
) &&
3694 (prev_entry
->next
->start
< addrbos
+ max_ssize
)) {
3696 vm_map_entry_release(count
);
3697 return (KERN_NO_SPACE
);
3701 * We initially map a stack of only init_ssize. We will
3702 * grow as needed later. Since this is to be a grow
3703 * down stack, we map at the top of the range.
3705 * Note: we would normally expect prot and max to be
3706 * VM_PROT_ALL, and cow to be 0. Possibly we should
3707 * eliminate these as input parameters, and just
3708 * pass these values here in the insert call.
3710 rv
= vm_map_insert(map
, &count
, NULL
, NULL
,
3711 0, addrbos
+ max_ssize
- init_ssize
,
3712 addrbos
+ max_ssize
,
3714 VM_SUBSYS_STACK
, prot
, max
, cow
);
3716 /* Now set the avail_ssize amount */
3717 if (rv
== KERN_SUCCESS
) {
3718 if (prev_entry
!= &map
->header
)
3719 vm_map_clip_end(map
, prev_entry
, addrbos
+ max_ssize
- init_ssize
, &count
);
3720 new_stack_entry
= prev_entry
->next
;
3721 if (new_stack_entry
->end
!= addrbos
+ max_ssize
||
3722 new_stack_entry
->start
!= addrbos
+ max_ssize
- init_ssize
)
3723 panic ("Bad entry start/end for new stack entry");
3725 new_stack_entry
->aux
.avail_ssize
= max_ssize
- init_ssize
;
3729 vm_map_entry_release(count
);
3734 * Attempts to grow a vm stack entry. Returns KERN_SUCCESS if the
3735 * desired address is already mapped, or if we successfully grow
3736 * the stack. Also returns KERN_SUCCESS if addr is outside the
3737 * stack range (this is strange, but preserves compatibility with
3738 * the grow function in vm_machdep.c).
3743 vm_map_growstack (vm_map_t map
, vm_offset_t addr
)
3745 vm_map_entry_t prev_entry
;
3746 vm_map_entry_t stack_entry
;
3747 vm_map_entry_t new_stack_entry
;
3753 int rv
= KERN_SUCCESS
;
3755 int use_read_lock
= 1;
3761 lp
= curthread
->td_lwp
;
3762 p
= curthread
->td_proc
;
3763 KKASSERT(lp
!= NULL
);
3764 vm
= lp
->lwp_vmspace
;
3765 KKASSERT(map
== &vm
->vm_map
);
3767 count
= vm_map_entry_reserve(MAP_RESERVE_COUNT
);
3770 vm_map_lock_read(map
);
3774 /* If addr is already in the entry range, no need to grow.*/
3775 if (vm_map_lookup_entry(map
, addr
, &prev_entry
))
3778 if ((stack_entry
= prev_entry
->next
) == &map
->header
)
3780 if (prev_entry
== &map
->header
)
3781 end
= stack_entry
->start
- stack_entry
->aux
.avail_ssize
;
3783 end
= prev_entry
->end
;
3786 * This next test mimics the old grow function in vm_machdep.c.
3787 * It really doesn't quite make sense, but we do it anyway
3788 * for compatibility.
3790 * If not growable stack, return success. This signals the
3791 * caller to proceed as he would normally with normal vm.
3793 if (stack_entry
->aux
.avail_ssize
< 1 ||
3794 addr
>= stack_entry
->start
||
3795 addr
< stack_entry
->start
- stack_entry
->aux
.avail_ssize
) {
3799 /* Find the minimum grow amount */
3800 grow_amount
= roundup (stack_entry
->start
- addr
, PAGE_SIZE
);
3801 if (grow_amount
> stack_entry
->aux
.avail_ssize
) {
3807 * If there is no longer enough space between the entries
3808 * nogo, and adjust the available space. Note: this
3809 * should only happen if the user has mapped into the
3810 * stack area after the stack was created, and is
3811 * probably an error.
3813 * This also effectively destroys any guard page the user
3814 * might have intended by limiting the stack size.
3816 if (grow_amount
> stack_entry
->start
- end
) {
3817 if (use_read_lock
&& vm_map_lock_upgrade(map
)) {
3823 stack_entry
->aux
.avail_ssize
= stack_entry
->start
- end
;
3828 is_procstack
= addr
>= (vm_offset_t
)vm
->vm_maxsaddr
;
3830 /* If this is the main process stack, see if we're over the
3833 if (is_procstack
&& (ctob(vm
->vm_ssize
) + grow_amount
>
3834 p
->p_rlimit
[RLIMIT_STACK
].rlim_cur
)) {
3839 /* Round up the grow amount modulo SGROWSIZ */
3840 grow_amount
= roundup (grow_amount
, sgrowsiz
);
3841 if (grow_amount
> stack_entry
->aux
.avail_ssize
) {
3842 grow_amount
= stack_entry
->aux
.avail_ssize
;
3844 if (is_procstack
&& (ctob(vm
->vm_ssize
) + grow_amount
>
3845 p
->p_rlimit
[RLIMIT_STACK
].rlim_cur
)) {
3846 grow_amount
= p
->p_rlimit
[RLIMIT_STACK
].rlim_cur
-
3850 /* If we would blow our VMEM resource limit, no go */
3851 if (map
->size
+ grow_amount
> p
->p_rlimit
[RLIMIT_VMEM
].rlim_cur
) {
3856 if (use_read_lock
&& vm_map_lock_upgrade(map
)) {
3863 /* Get the preliminary new entry start value */
3864 addr
= stack_entry
->start
- grow_amount
;
3866 /* If this puts us into the previous entry, cut back our growth
3867 * to the available space. Also, see the note above.
3870 stack_entry
->aux
.avail_ssize
= stack_entry
->start
- end
;
3874 rv
= vm_map_insert(map
, &count
, NULL
, NULL
,
3875 0, addr
, stack_entry
->start
,
3877 VM_SUBSYS_STACK
, VM_PROT_ALL
, VM_PROT_ALL
, 0);
3879 /* Adjust the available stack space by the amount we grew. */
3880 if (rv
== KERN_SUCCESS
) {
3881 if (prev_entry
!= &map
->header
)
3882 vm_map_clip_end(map
, prev_entry
, addr
, &count
);
3883 new_stack_entry
= prev_entry
->next
;
3884 if (new_stack_entry
->end
!= stack_entry
->start
||
3885 new_stack_entry
->start
!= addr
)
3886 panic ("Bad stack grow start/end in new stack entry");
3888 new_stack_entry
->aux
.avail_ssize
=
3889 stack_entry
->aux
.avail_ssize
-
3890 (new_stack_entry
->end
- new_stack_entry
->start
);
3892 vm
->vm_ssize
+= btoc(new_stack_entry
->end
-
3893 new_stack_entry
->start
);
3896 if (map
->flags
& MAP_WIREFUTURE
)
3897 vm_map_unwire(map
, new_stack_entry
->start
,
3898 new_stack_entry
->end
, FALSE
);
3903 vm_map_unlock_read(map
);
3906 vm_map_entry_release(count
);
3911 * Unshare the specified VM space for exec. If other processes are
3912 * mapped to it, then create a new one. The new vmspace is null.
3917 vmspace_exec(struct proc
*p
, struct vmspace
*vmcopy
)
3919 struct vmspace
*oldvmspace
= p
->p_vmspace
;
3920 struct vmspace
*newvmspace
;
3921 vm_map_t map
= &p
->p_vmspace
->vm_map
;
3924 * If we are execing a resident vmspace we fork it, otherwise
3925 * we create a new vmspace. Note that exitingcnt is not
3926 * copied to the new vmspace.
3928 lwkt_gettoken(&oldvmspace
->vm_map
.token
);
3930 newvmspace
= vmspace_fork(vmcopy
);
3931 lwkt_gettoken(&newvmspace
->vm_map
.token
);
3933 newvmspace
= vmspace_alloc(map
->min_offset
, map
->max_offset
);
3934 lwkt_gettoken(&newvmspace
->vm_map
.token
);
3935 bcopy(&oldvmspace
->vm_startcopy
, &newvmspace
->vm_startcopy
,
3936 (caddr_t
)&oldvmspace
->vm_endcopy
-
3937 (caddr_t
)&oldvmspace
->vm_startcopy
);
3941 * Finish initializing the vmspace before assigning it
3942 * to the process. The vmspace will become the current vmspace
3945 pmap_pinit2(vmspace_pmap(newvmspace
));
3946 pmap_replacevm(p
, newvmspace
, 0);
3947 lwkt_reltoken(&newvmspace
->vm_map
.token
);
3948 lwkt_reltoken(&oldvmspace
->vm_map
.token
);
3949 vmspace_rel(oldvmspace
);
3953 * Unshare the specified VM space for forcing COW. This
3954 * is called by rfork, for the (RFMEM|RFPROC) == 0 case.
3957 vmspace_unshare(struct proc
*p
)
3959 struct vmspace
*oldvmspace
= p
->p_vmspace
;
3960 struct vmspace
*newvmspace
;
3962 lwkt_gettoken(&oldvmspace
->vm_map
.token
);
3963 if (vmspace_getrefs(oldvmspace
) == 1) {
3964 lwkt_reltoken(&oldvmspace
->vm_map
.token
);
3967 newvmspace
= vmspace_fork(oldvmspace
);
3968 lwkt_gettoken(&newvmspace
->vm_map
.token
);
3969 pmap_pinit2(vmspace_pmap(newvmspace
));
3970 pmap_replacevm(p
, newvmspace
, 0);
3971 lwkt_reltoken(&newvmspace
->vm_map
.token
);
3972 lwkt_reltoken(&oldvmspace
->vm_map
.token
);
3973 vmspace_rel(oldvmspace
);
3977 * vm_map_hint: return the beginning of the best area suitable for
3978 * creating a new mapping with "prot" protection.
3983 vm_map_hint(struct proc
*p
, vm_offset_t addr
, vm_prot_t prot
)
3985 struct vmspace
*vms
= p
->p_vmspace
;
3987 if (!randomize_mmap
|| addr
!= 0) {
3989 * Set a reasonable start point for the hint if it was
3990 * not specified or if it falls within the heap space.
3991 * Hinted mmap()s do not allocate out of the heap space.
3994 (addr
>= round_page((vm_offset_t
)vms
->vm_taddr
) &&
3995 addr
< round_page((vm_offset_t
)vms
->vm_daddr
+ maxdsiz
))) {
3996 addr
= round_page((vm_offset_t
)vms
->vm_daddr
+ maxdsiz
);
4001 addr
= (vm_offset_t
)vms
->vm_daddr
+ MAXDSIZ
;
4002 addr
+= karc4random() & (MIN((256 * 1024 * 1024), MAXDSIZ
) - 1);
4004 return (round_page(addr
));
4008 * Finds the VM object, offset, and protection for a given virtual address
4009 * in the specified map, assuming a page fault of the type specified.
4011 * Leaves the map in question locked for read; return values are guaranteed
4012 * until a vm_map_lookup_done call is performed. Note that the map argument
4013 * is in/out; the returned map must be used in the call to vm_map_lookup_done.
4015 * A handle (out_entry) is returned for use in vm_map_lookup_done, to make
4018 * If a lookup is requested with "write protection" specified, the map may
4019 * be changed to perform virtual copying operations, although the data
4020 * referenced will remain the same.
4025 vm_map_lookup(vm_map_t
*var_map
, /* IN/OUT */
4027 vm_prot_t fault_typea
,
4028 vm_map_entry_t
*out_entry
, /* OUT */
4029 vm_object_t
*object
, /* OUT */
4030 vm_pindex_t
*pindex
, /* OUT */
4031 vm_prot_t
*out_prot
, /* OUT */
4032 boolean_t
*wired
) /* OUT */
4034 vm_map_entry_t entry
;
4035 vm_map_t map
= *var_map
;
4037 vm_prot_t fault_type
= fault_typea
;
4038 int use_read_lock
= 1;
4039 int rv
= KERN_SUCCESS
;
4043 vm_map_lock_read(map
);
4048 * If the map has an interesting hint, try it before calling full
4049 * blown lookup routine.
4056 if ((entry
== &map
->header
) ||
4057 (vaddr
< entry
->start
) || (vaddr
>= entry
->end
)) {
4058 vm_map_entry_t tmp_entry
;
4061 * Entry was either not a valid hint, or the vaddr was not
4062 * contained in the entry, so do a full lookup.
4064 if (!vm_map_lookup_entry(map
, vaddr
, &tmp_entry
)) {
4065 rv
= KERN_INVALID_ADDRESS
;
4076 if (entry
->maptype
== VM_MAPTYPE_SUBMAP
) {
4077 vm_map_t old_map
= map
;
4079 *var_map
= map
= entry
->object
.sub_map
;
4081 vm_map_unlock_read(old_map
);
4083 vm_map_unlock(old_map
);
4089 * Check whether this task is allowed to have this page.
4090 * Note the special case for MAP_ENTRY_COW pages with an override.
4091 * This is to implement a forced COW for debuggers.
4093 if (fault_type
& VM_PROT_OVERRIDE_WRITE
)
4094 prot
= entry
->max_protection
;
4096 prot
= entry
->protection
;
4098 fault_type
&= (VM_PROT_READ
|VM_PROT_WRITE
|VM_PROT_EXECUTE
);
4099 if ((fault_type
& prot
) != fault_type
) {
4100 rv
= KERN_PROTECTION_FAILURE
;
4104 if ((entry
->eflags
& MAP_ENTRY_USER_WIRED
) &&
4105 (entry
->eflags
& MAP_ENTRY_COW
) &&
4106 (fault_type
& VM_PROT_WRITE
) &&
4107 (fault_typea
& VM_PROT_OVERRIDE_WRITE
) == 0) {
4108 rv
= KERN_PROTECTION_FAILURE
;
4113 * If this page is not pageable, we have to get it for all possible
4116 *wired
= (entry
->wired_count
!= 0);
4118 prot
= fault_type
= entry
->protection
;
4121 * Virtual page tables may need to update the accessed (A) bit
4122 * in a page table entry. Upgrade the fault to a write fault for
4123 * that case if the map will support it. If the map does not support
4124 * it the page table entry simply will not be updated.
4126 if (entry
->maptype
== VM_MAPTYPE_VPAGETABLE
) {
4127 if (prot
& VM_PROT_WRITE
)
4128 fault_type
|= VM_PROT_WRITE
;
4131 if (curthread
->td_lwp
&& curthread
->td_lwp
->lwp_vmspace
&&
4132 pmap_emulate_ad_bits(&curthread
->td_lwp
->lwp_vmspace
->vm_pmap
)) {
4133 if ((prot
& VM_PROT_WRITE
) == 0)
4134 fault_type
|= VM_PROT_WRITE
;
4138 * Only NORMAL and VPAGETABLE maps are object-based. UKSMAPs are not.
4140 if (entry
->maptype
!= VM_MAPTYPE_NORMAL
&&
4141 entry
->maptype
!= VM_MAPTYPE_VPAGETABLE
) {
4147 * If the entry was copy-on-write, we either ...
4149 if (entry
->eflags
& MAP_ENTRY_NEEDS_COPY
) {
4151 * If we want to write the page, we may as well handle that
4152 * now since we've got the map locked.
4154 * If we don't need to write the page, we just demote the
4155 * permissions allowed.
4158 if (fault_type
& VM_PROT_WRITE
) {
4160 * Not allowed if TDF_NOFAULT is set as the shadowing
4161 * operation can deadlock against the faulting
4162 * function due to the copy-on-write.
4164 if (curthread
->td_flags
& TDF_NOFAULT
) {
4165 rv
= KERN_FAILURE_NOFAULT
;
4170 * Make a new object, and place it in the object
4171 * chain. Note that no new references have appeared
4172 * -- one just moved from the map to the new
4176 if (use_read_lock
&& vm_map_lock_upgrade(map
)) {
4183 vm_map_entry_shadow(entry
, 0);
4186 * We're attempting to read a copy-on-write page --
4187 * don't allow writes.
4190 prot
&= ~VM_PROT_WRITE
;
4195 * Create an object if necessary.
4197 if (entry
->object
.vm_object
== NULL
&& !map
->system_map
) {
4198 if (use_read_lock
&& vm_map_lock_upgrade(map
)) {
4204 vm_map_entry_allocate_object(entry
);
4208 * Return the object/offset from this entry. If the entry was
4209 * copy-on-write or empty, it has been fixed up.
4211 *object
= entry
->object
.vm_object
;
4214 *pindex
= OFF_TO_IDX((vaddr
- entry
->start
) + entry
->offset
);
4217 * Return whether this is the only map sharing this data. On
4218 * success we return with a read lock held on the map. On failure
4219 * we return with the map unlocked.
4223 if (rv
== KERN_SUCCESS
) {
4224 if (use_read_lock
== 0)
4225 vm_map_lock_downgrade(map
);
4226 } else if (use_read_lock
) {
4227 vm_map_unlock_read(map
);
4235 * Releases locks acquired by a vm_map_lookup()
4236 * (according to the handle returned by that lookup).
4238 * No other requirements.
4241 vm_map_lookup_done(vm_map_t map
, vm_map_entry_t entry
, int count
)
4244 * Unlock the main-level map
4246 vm_map_unlock_read(map
);
4248 vm_map_entry_release(count
);
4252 * Quick hack, needs some help to make it more SMP friendly.
4255 vm_map_interlock(vm_map_t map
, struct vm_map_ilock
*ilock
,
4256 vm_offset_t ran_beg
, vm_offset_t ran_end
)
4258 struct vm_map_ilock
*scan
;
4260 ilock
->ran_beg
= ran_beg
;
4261 ilock
->ran_end
= ran_end
;
4264 spin_lock(&map
->ilock_spin
);
4266 for (scan
= map
->ilock_base
; scan
; scan
= scan
->next
) {
4267 if (ran_end
> scan
->ran_beg
&& ran_beg
< scan
->ran_end
) {
4268 scan
->flags
|= ILOCK_WAITING
;
4269 ssleep(scan
, &map
->ilock_spin
, 0, "ilock", 0);
4273 ilock
->next
= map
->ilock_base
;
4274 map
->ilock_base
= ilock
;
4275 spin_unlock(&map
->ilock_spin
);
4279 vm_map_deinterlock(vm_map_t map
, struct vm_map_ilock
*ilock
)
4281 struct vm_map_ilock
*scan
;
4282 struct vm_map_ilock
**scanp
;
4284 spin_lock(&map
->ilock_spin
);
4285 scanp
= &map
->ilock_base
;
4286 while ((scan
= *scanp
) != NULL
) {
4287 if (scan
== ilock
) {
4288 *scanp
= ilock
->next
;
4289 spin_unlock(&map
->ilock_spin
);
4290 if (ilock
->flags
& ILOCK_WAITING
)
4294 scanp
= &scan
->next
;
4296 spin_unlock(&map
->ilock_spin
);
4297 panic("vm_map_deinterlock: missing ilock!");
4300 #include "opt_ddb.h"
4302 #include <sys/kernel.h>
4304 #include <ddb/ddb.h>
4309 DB_SHOW_COMMAND(map
, vm_map_print
)
4312 /* XXX convert args. */
4313 vm_map_t map
= (vm_map_t
)addr
;
4314 boolean_t full
= have_addr
;
4316 vm_map_entry_t entry
;
4318 db_iprintf("Task map %p: pmap=%p, nentries=%d, version=%u\n",
4320 (void *)map
->pmap
, map
->nentries
, map
->timestamp
);
4323 if (!full
&& db_indent
)
4327 for (entry
= map
->header
.next
; entry
!= &map
->header
;
4328 entry
= entry
->next
) {
4329 db_iprintf("map entry %p: start=%p, end=%p\n",
4330 (void *)entry
, (void *)entry
->start
, (void *)entry
->end
);
4333 static char *inheritance_name
[4] =
4334 {"share", "copy", "none", "donate_copy"};
4336 db_iprintf(" prot=%x/%x/%s",
4338 entry
->max_protection
,
4339 inheritance_name
[(int)(unsigned char)
4340 entry
->inheritance
]);
4341 if (entry
->wired_count
!= 0)
4342 db_printf(", wired");
4344 switch(entry
->maptype
) {
4345 case VM_MAPTYPE_SUBMAP
:
4346 /* XXX no %qd in kernel. Truncate entry->offset. */
4347 db_printf(", share=%p, offset=0x%lx\n",
4348 (void *)entry
->object
.sub_map
,
4349 (long)entry
->offset
);
4351 if ((entry
->prev
== &map
->header
) ||
4352 (entry
->prev
->object
.sub_map
!=
4353 entry
->object
.sub_map
)) {
4355 vm_map_print((db_expr_t
)(intptr_t)
4356 entry
->object
.sub_map
,
4361 case VM_MAPTYPE_NORMAL
:
4362 case VM_MAPTYPE_VPAGETABLE
:
4363 /* XXX no %qd in kernel. Truncate entry->offset. */
4364 db_printf(", object=%p, offset=0x%lx",
4365 (void *)entry
->object
.vm_object
,
4366 (long)entry
->offset
);
4367 if (entry
->eflags
& MAP_ENTRY_COW
)
4368 db_printf(", copy (%s)",
4369 (entry
->eflags
& MAP_ENTRY_NEEDS_COPY
) ? "needed" : "done");
4373 if ((entry
->prev
== &map
->header
) ||
4374 (entry
->prev
->object
.vm_object
!=
4375 entry
->object
.vm_object
)) {
4377 vm_object_print((db_expr_t
)(intptr_t)
4378 entry
->object
.vm_object
,
4384 case VM_MAPTYPE_UKSMAP
:
4385 db_printf(", uksmap=%p, offset=0x%lx",
4386 (void *)entry
->object
.uksmap
,
4387 (long)entry
->offset
);
4388 if (entry
->eflags
& MAP_ENTRY_COW
)
4389 db_printf(", copy (%s)",
4390 (entry
->eflags
& MAP_ENTRY_NEEDS_COPY
) ? "needed" : "done");
4406 DB_SHOW_COMMAND(procvm
, procvm
)
4411 p
= (struct proc
*) addr
;
4416 db_printf("p = %p, vmspace = %p, map = %p, pmap = %p\n",
4417 (void *)p
, (void *)p
->p_vmspace
, (void *)&p
->p_vmspace
->vm_map
,
4418 (void *)vmspace_pmap(p
->p_vmspace
));
4420 vm_map_print((db_expr_t
)(intptr_t)&p
->p_vmspace
->vm_map
, 1, 0, NULL
);