e3fd77e578264b99e604ace120f2fa247ca9daf5
[dragonfly.git] / sys / vm / vm_map.c
1 /*
2  * (MPSAFE)
3  *
4  * Copyright (c) 1991, 1993
5  *      The Regents of the University of California.  All rights reserved.
6  *
7  * This code is derived from software contributed to Berkeley by
8  * The Mach Operating System project at Carnegie-Mellon University.
9  *
10  * Redistribution and use in source and binary forms, with or without
11  * modification, are permitted provided that the following conditions
12  * are met:
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. All advertising materials mentioning features or use of this software
19  *    must display the following acknowledgement:
20  *      This product includes software developed by the University of
21  *      California, Berkeley and its contributors.
22  * 4. Neither the name of the University nor the names of its contributors
23  *    may be used to endorse or promote products derived from this software
24  *    without specific prior written permission.
25  *
26  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
27  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
28  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
29  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
30  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
31  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
32  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
33  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
34  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
35  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
36  * SUCH DAMAGE.
37  *
38  *      from: @(#)vm_map.c      8.3 (Berkeley) 1/12/94
39  *
40  *
41  * Copyright (c) 1987, 1990 Carnegie-Mellon University.
42  * All rights reserved.
43  *
44  * Authors: Avadis Tevanian, Jr., Michael Wayne Young
45  *
46  * Permission to use, copy, modify and distribute this software and
47  * its documentation is hereby granted, provided that both the copyright
48  * notice and this permission notice appear in all copies of the
49  * software, derivative works or modified versions, and any portions
50  * thereof, and that both notices appear in supporting documentation.
51  *
52  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
53  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
54  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
55  *
56  * Carnegie Mellon requests users of this software to return to
57  *
58  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
59  *  School of Computer Science
60  *  Carnegie Mellon University
61  *  Pittsburgh PA 15213-3890
62  *
63  * any improvements or extensions that they make and grant Carnegie the
64  * rights to redistribute these changes.
65  *
66  * $FreeBSD: src/sys/vm/vm_map.c,v 1.187.2.19 2003/05/27 00:47:02 alc Exp $
67  * $DragonFly: src/sys/vm/vm_map.c,v 1.56 2007/04/29 18:25:41 dillon Exp $
68  */
69
70 /*
71  *      Virtual memory mapping module.
72  */
73
74 #include <sys/param.h>
75 #include <sys/systm.h>
76 #include <sys/kernel.h>
77 #include <sys/proc.h>
78 #include <sys/serialize.h>
79 #include <sys/lock.h>
80 #include <sys/vmmeter.h>
81 #include <sys/mman.h>
82 #include <sys/vnode.h>
83 #include <sys/resourcevar.h>
84 #include <sys/shm.h>
85 #include <sys/tree.h>
86 #include <sys/malloc.h>
87
88 #include <vm/vm.h>
89 #include <vm/vm_param.h>
90 #include <vm/pmap.h>
91 #include <vm/vm_map.h>
92 #include <vm/vm_page.h>
93 #include <vm/vm_object.h>
94 #include <vm/vm_pager.h>
95 #include <vm/vm_kern.h>
96 #include <vm/vm_extern.h>
97 #include <vm/swap_pager.h>
98 #include <vm/vm_zone.h>
99
100 #include <sys/thread2.h>
101 #include <sys/sysref2.h>
102 #include <sys/random.h>
103 #include <sys/sysctl.h>
104
105 /*
106  * Virtual memory maps provide for the mapping, protection, and sharing
107  * of virtual memory objects.  In addition, this module provides for an
108  * efficient virtual copy of memory from one map to another.
109  *
110  * Synchronization is required prior to most operations.
111  *
112  * Maps consist of an ordered doubly-linked list of simple entries.
113  * A hint and a RB tree is used to speed-up lookups.
114  *
115  * Callers looking to modify maps specify start/end addresses which cause
116  * the related map entry to be clipped if necessary, and then later
117  * recombined if the pieces remained compatible.
118  *
119  * Virtual copy operations are performed by copying VM object references
120  * from one map to another, and then marking both regions as copy-on-write.
121  */
122 static void vmspace_terminate(struct vmspace *vm);
123 static void vmspace_lock(struct vmspace *vm);
124 static void vmspace_unlock(struct vmspace *vm);
125 static void vmspace_dtor(void *obj, void *private);
126
127 MALLOC_DEFINE(M_VMSPACE, "vmspace", "vmspace objcache backingstore");
128
129 struct sysref_class vmspace_sysref_class = {
130         .name =         "vmspace",
131         .mtype =        M_VMSPACE,
132         .proto =        SYSREF_PROTO_VMSPACE,
133         .offset =       offsetof(struct vmspace, vm_sysref),
134         .objsize =      sizeof(struct vmspace),
135         .nom_cache =    32,
136         .flags = SRC_MANAGEDINIT,
137         .dtor = vmspace_dtor,
138         .ops = {
139                 .terminate = (sysref_terminate_func_t)vmspace_terminate,
140                 .lock = (sysref_lock_func_t)vmspace_lock,
141                 .unlock = (sysref_lock_func_t)vmspace_unlock
142         }
143 };
144
145 /*
146  * per-cpu page table cross mappings are initialized in early boot
147  * and might require a considerable number of vm_map_entry structures.
148  */
149 #define VMEPERCPU       (MAXCPU+1)
150
151 static struct vm_zone mapentzone_store, mapzone_store;
152 static vm_zone_t mapentzone, mapzone;
153 static struct vm_object mapentobj, mapobj;
154
155 static struct vm_map_entry map_entry_init[MAX_MAPENT];
156 static struct vm_map_entry cpu_map_entry_init[MAXCPU][VMEPERCPU];
157 static struct vm_map map_init[MAX_KMAP];
158
159 static int randomize_mmap;
160 SYSCTL_INT(_vm, OID_AUTO, randomize_mmap, CTLFLAG_RW, &randomize_mmap, 0,
161     "Randomize mmap offsets");
162
163 static void vm_map_entry_shadow(vm_map_entry_t entry, int addref);
164 static vm_map_entry_t vm_map_entry_create(vm_map_t map, int *);
165 static void vm_map_entry_dispose (vm_map_t map, vm_map_entry_t entry, int *);
166 static void _vm_map_clip_end (vm_map_t, vm_map_entry_t, vm_offset_t, int *);
167 static void _vm_map_clip_start (vm_map_t, vm_map_entry_t, vm_offset_t, int *);
168 static void vm_map_entry_delete (vm_map_t, vm_map_entry_t, int *);
169 static void vm_map_entry_unwire (vm_map_t, vm_map_entry_t);
170 static void vm_map_copy_entry (vm_map_t, vm_map_t, vm_map_entry_t,
171                 vm_map_entry_t);
172 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);
173
174 /*
175  * Initialize the vm_map module.  Must be called before any other vm_map
176  * routines.
177  *
178  * Map and entry structures are allocated from the general purpose
179  * memory pool with some exceptions:
180  *
181  *      - The kernel map is allocated statically.
182  *      - Initial kernel map entries are allocated out of a static pool.
183  *
184  *      These restrictions are necessary since malloc() uses the
185  *      maps and requires map entries.
186  *
187  * Called from the low level boot code only.
188  */
189 void
190 vm_map_startup(void)
191 {
192         mapzone = &mapzone_store;
193         zbootinit(mapzone, "MAP", sizeof (struct vm_map),
194                 map_init, MAX_KMAP);
195         mapentzone = &mapentzone_store;
196         zbootinit(mapentzone, "MAP ENTRY", sizeof (struct vm_map_entry),
197                 map_entry_init, MAX_MAPENT);
198 }
199
200 /*
201  * Called prior to any vmspace allocations.
202  *
203  * Called from the low level boot code only.
204  */
205 void
206 vm_init2(void) 
207 {
208         zinitna(mapentzone, &mapentobj, NULL, 0, 0, 
209                 ZONE_USE_RESERVE | ZONE_SPECIAL, 1);
210         zinitna(mapzone, &mapobj, NULL, 0, 0, 0, 1);
211         pmap_init2();
212         vm_object_init2();
213 }
214
215
216 /*
217  * Red black tree functions
218  *
219  * The caller must hold the related map lock.
220  */
221 static int rb_vm_map_compare(vm_map_entry_t a, vm_map_entry_t b);
222 RB_GENERATE(vm_map_rb_tree, vm_map_entry, rb_entry, rb_vm_map_compare);
223
224 /* a->start is address, and the only field has to be initialized */
225 static int
226 rb_vm_map_compare(vm_map_entry_t a, vm_map_entry_t b)
227 {
228         if (a->start < b->start)
229                 return(-1);
230         else if (a->start > b->start)
231                 return(1);
232         return(0);
233 }
234
235 /*
236  * Allocate a vmspace structure, including a vm_map and pmap.
237  * Initialize numerous fields.  While the initial allocation is zerod,
238  * subsequence reuse from the objcache leaves elements of the structure
239  * intact (particularly the pmap), so portions must be zerod.
240  *
241  * The structure is not considered activated until we call sysref_activate().
242  *
243  * No requirements.
244  */
245 struct vmspace *
246 vmspace_alloc(vm_offset_t min, vm_offset_t max)
247 {
248         struct vmspace *vm;
249
250         vm = sysref_alloc(&vmspace_sysref_class);
251         bzero(&vm->vm_startcopy,
252               (char *)&vm->vm_endcopy - (char *)&vm->vm_startcopy);
253         vm_map_init(&vm->vm_map, min, max, NULL);       /* initializes token */
254
255         /*
256          * Use a hold to prevent any additional racing hold from terminating
257          * the vmspace before we manage to activate it.  This also acquires
258          * the token for safety.
259          */
260         KKASSERT(vm->vm_holdcount == 0);
261         KKASSERT(vm->vm_exitingcnt == 0);
262         vmspace_hold(vm);
263         pmap_pinit(vmspace_pmap(vm));           /* (some fields reused) */
264         vm->vm_map.pmap = vmspace_pmap(vm);             /* XXX */
265         vm->vm_shm = NULL;
266         vm->vm_flags = 0;
267         cpu_vmspace_alloc(vm);
268         sysref_activate(&vm->vm_sysref);
269         vmspace_drop(vm);
270
271         return (vm);
272 }
273
274 /*
275  * Free a primary reference to a vmspace.  This can trigger a
276  * stage-1 termination.
277  */
278 void
279 vmspace_free(struct vmspace *vm)
280 {
281         /*
282          * We want all finalization to occur via vmspace_drop() so we
283          * need to hold the vm around the put.
284          */
285         vmspace_hold(vm);
286         sysref_put(&vm->vm_sysref);
287         vmspace_drop(vm);
288 }
289
290 void
291 vmspace_ref(struct vmspace *vm)
292 {
293         sysref_get(&vm->vm_sysref);
294 }
295
296 void
297 vmspace_hold(struct vmspace *vm)
298 {
299         refcount_acquire(&vm->vm_holdcount);
300         lwkt_gettoken(&vm->vm_map.token);
301 }
302
303 void
304 vmspace_drop(struct vmspace *vm)
305 {
306         lwkt_reltoken(&vm->vm_map.token);
307         if (refcount_release(&vm->vm_holdcount)) {
308                 if (vm->vm_exitingcnt == 0 &&
309                     sysref_isinactive(&vm->vm_sysref)) {
310                         vmspace_terminate(vm);
311                 }
312         }
313 }
314
315 /*
316  * dtor function - Some elements of the pmap are retained in the
317  * free-cached vmspaces to improve performance.  We have to clean them up
318  * here before returning the vmspace to the memory pool.
319  *
320  * No requirements.
321  */
322 static void
323 vmspace_dtor(void *obj, void *private)
324 {
325         struct vmspace *vm = obj;
326
327         pmap_puninit(vmspace_pmap(vm));
328 }
329
330 /*
331  * Called in three cases:
332  *
333  * (1) When the last sysref is dropped and the vmspace becomes inactive.
334  *     (holdcount will not be 0 because the vmspace is held through the op)
335  *
336  * (2) When exitingcount becomes 0 on the last reap
337  *     (holdcount will not be 0 because the vmspace is held through the op)
338  *
339  * (3) When the holdcount becomes 0 in addition to the above two
340  *
341  * sysref will not scrap the object until we call sysref_put() once more
342  * after the last ref has been dropped.
343  *
344  * VMSPACE_EXIT1 flags the primary deactivation
345  * VMSPACE_EXIT2 flags the last reap
346  */
347 static void
348 vmspace_terminate(struct vmspace *vm)
349 {
350         int count;
351
352         /*
353          *
354          */
355         lwkt_gettoken(&vm->vm_map.token);
356         if ((vm->vm_flags & VMSPACE_EXIT1) == 0) {
357                 vm->vm_flags |= VMSPACE_EXIT1;
358                 shmexit(vm);
359                 pmap_remove_pages(vmspace_pmap(vm), VM_MIN_USER_ADDRESS,
360                                   VM_MAX_USER_ADDRESS);
361                 vm_map_remove(&vm->vm_map, VM_MIN_USER_ADDRESS,
362                               VM_MAX_USER_ADDRESS);
363         }
364         if ((vm->vm_flags & VMSPACE_EXIT2) == 0 && vm->vm_exitingcnt == 0) {
365                 vm->vm_flags |= VMSPACE_EXIT2;
366                 cpu_vmspace_free(vm);
367                 shmexit(vm);
368                 KKASSERT(vm->vm_upcalls == NULL);
369
370                 /*
371                  * Lock the map, to wait out all other references to it.
372                  * Delete all of the mappings and pages they hold, then call
373                  * the pmap module to reclaim anything left.
374                  */
375                 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
376                 vm_map_lock(&vm->vm_map);
377                 vm_map_delete(&vm->vm_map, vm->vm_map.min_offset,
378                               vm->vm_map.max_offset, &count);
379                 vm_map_unlock(&vm->vm_map);
380                 vm_map_entry_release(count);
381
382                 lwkt_gettoken(&vmspace_pmap(vm)->pm_token);
383                 pmap_release(vmspace_pmap(vm));
384                 lwkt_reltoken(&vmspace_pmap(vm)->pm_token);
385         }
386
387         lwkt_reltoken(&vm->vm_map.token);
388         if (vm->vm_exitingcnt == 0 && vm->vm_holdcount == 0) {
389                 KKASSERT(vm->vm_flags & VMSPACE_EXIT1);
390                 KKASSERT(vm->vm_flags & VMSPACE_EXIT2);
391                 sysref_put(&vm->vm_sysref);
392         }
393 }
394
395 /*
396  * vmspaces are not currently locked.
397  */
398 static void
399 vmspace_lock(struct vmspace *vm __unused)
400 {
401 }
402
403 static void
404 vmspace_unlock(struct vmspace *vm __unused)
405 {
406 }
407
408 /*
409  * This is called during exit indicating that the vmspace is no
410  * longer in used by an exiting process, but the process has not yet
411  * been reaped.
412  *
413  * No requirements.
414  */
415 void
416 vmspace_exitbump(struct vmspace *vm)
417 {
418         vmspace_hold(vm);
419         ++vm->vm_exitingcnt;
420         vmspace_drop(vm);       /* handles termination sequencing */
421 }
422
423 /*
424  * Decrement the exitingcnt and issue the stage-2 termination if it becomes
425  * zero and the stage1 termination has already occured.
426  *
427  * No requirements.
428  */
429 void
430 vmspace_exitfree(struct proc *p)
431 {
432         struct vmspace *vm;
433
434         vm = p->p_vmspace;
435         p->p_vmspace = NULL;
436         vmspace_hold(vm);
437         KKASSERT(vm->vm_exitingcnt > 0);
438         if (--vm->vm_exitingcnt == 0 && sysref_isinactive(&vm->vm_sysref))
439                 vmspace_terminate(vm);
440         vmspace_drop(vm);       /* handles termination sequencing */
441 }
442
443 /*
444  * Swap useage is determined by taking the proportional swap used by
445  * VM objects backing the VM map.  To make up for fractional losses,
446  * if the VM object has any swap use at all the associated map entries
447  * count for at least 1 swap page.
448  *
449  * No requirements.
450  */
451 int
452 vmspace_swap_count(struct vmspace *vm)
453 {
454         vm_map_t map = &vm->vm_map;
455         vm_map_entry_t cur;
456         vm_object_t object;
457         int count = 0;
458         int n;
459
460         vmspace_hold(vm);
461         for (cur = map->header.next; cur != &map->header; cur = cur->next) {
462                 switch(cur->maptype) {
463                 case VM_MAPTYPE_NORMAL:
464                 case VM_MAPTYPE_VPAGETABLE:
465                         if ((object = cur->object.vm_object) == NULL)
466                                 break;
467                         if (object->swblock_count) {
468                                 n = (cur->end - cur->start) / PAGE_SIZE;
469                                 count += object->swblock_count *
470                                     SWAP_META_PAGES * n / object->size + 1;
471                         }
472                         break;
473                 default:
474                         break;
475                 }
476         }
477         vmspace_drop(vm);
478
479         return(count);
480 }
481
482 /*
483  * Calculate the approximate number of anonymous pages in use by
484  * this vmspace.  To make up for fractional losses, we count each
485  * VM object as having at least 1 anonymous page.
486  *
487  * No requirements.
488  */
489 int
490 vmspace_anonymous_count(struct vmspace *vm)
491 {
492         vm_map_t map = &vm->vm_map;
493         vm_map_entry_t cur;
494         vm_object_t object;
495         int count = 0;
496
497         vmspace_hold(vm);
498         for (cur = map->header.next; cur != &map->header; cur = cur->next) {
499                 switch(cur->maptype) {
500                 case VM_MAPTYPE_NORMAL:
501                 case VM_MAPTYPE_VPAGETABLE:
502                         if ((object = cur->object.vm_object) == NULL)
503                                 break;
504                         if (object->type != OBJT_DEFAULT &&
505                             object->type != OBJT_SWAP) {
506                                 break;
507                         }
508                         count += object->resident_page_count;
509                         break;
510                 default:
511                         break;
512                 }
513         }
514         vmspace_drop(vm);
515
516         return(count);
517 }
518
519 /*
520  * Creates and returns a new empty VM map with the given physical map
521  * structure, and having the given lower and upper address bounds.
522  *
523  * No requirements.
524  */
525 vm_map_t
526 vm_map_create(vm_map_t result, pmap_t pmap, vm_offset_t min, vm_offset_t max)
527 {
528         if (result == NULL)
529                 result = zalloc(mapzone);
530         vm_map_init(result, min, max, pmap);
531         return (result);
532 }
533
534 /*
535  * Initialize an existing vm_map structure such as that in the vmspace
536  * structure.  The pmap is initialized elsewhere.
537  *
538  * No requirements.
539  */
540 void
541 vm_map_init(struct vm_map *map, vm_offset_t min, vm_offset_t max, pmap_t pmap)
542 {
543         map->header.next = map->header.prev = &map->header;
544         RB_INIT(&map->rb_root);
545         map->nentries = 0;
546         map->size = 0;
547         map->system_map = 0;
548         map->min_offset = min;
549         map->max_offset = max;
550         map->pmap = pmap;
551         map->first_free = &map->header;
552         map->hint = &map->header;
553         map->timestamp = 0;
554         map->flags = 0;
555         lwkt_token_init(&map->token, "vm_map");
556         lockinit(&map->lock, "thrd_sleep", (hz + 9) / 10, 0);
557         TUNABLE_INT("vm.cache_vmspaces", &vmspace_sysref_class.nom_cache);
558 }
559
560 /*
561  * Shadow the vm_map_entry's object.  This typically needs to be done when
562  * a write fault is taken on an entry which had previously been cloned by
563  * fork().  The shared object (which might be NULL) must become private so
564  * we add a shadow layer above it.
565  *
566  * Object allocation for anonymous mappings is defered as long as possible.
567  * When creating a shadow, however, the underlying object must be instantiated
568  * so it can be shared.
569  *
570  * If the map segment is governed by a virtual page table then it is
571  * possible to address offsets beyond the mapped area.  Just allocate
572  * a maximally sized object for this case.
573  *
574  * The vm_map must be exclusively locked.
575  * No other requirements.
576  */
577 static
578 void
579 vm_map_entry_shadow(vm_map_entry_t entry, int addref)
580 {
581         if (entry->maptype == VM_MAPTYPE_VPAGETABLE) {
582                 vm_object_shadow(&entry->object.vm_object, &entry->offset,
583                                  0x7FFFFFFF, addref);   /* XXX */
584         } else {
585                 vm_object_shadow(&entry->object.vm_object, &entry->offset,
586                                  atop(entry->end - entry->start), addref);
587         }
588         entry->eflags &= ~MAP_ENTRY_NEEDS_COPY;
589 }
590
591 /*
592  * Allocate an object for a vm_map_entry.
593  *
594  * Object allocation for anonymous mappings is defered as long as possible.
595  * This function is called when we can defer no longer, generally when a map
596  * entry might be split or forked or takes a page fault.
597  *
598  * If the map segment is governed by a virtual page table then it is
599  * possible to address offsets beyond the mapped area.  Just allocate
600  * a maximally sized object for this case.
601  *
602  * The vm_map must be exclusively locked.
603  * No other requirements.
604  */
605 void 
606 vm_map_entry_allocate_object(vm_map_entry_t entry)
607 {
608         vm_object_t obj;
609
610         if (entry->maptype == VM_MAPTYPE_VPAGETABLE) {
611                 obj = vm_object_allocate(OBJT_DEFAULT, 0x7FFFFFFF); /* XXX */
612         } else {
613                 obj = vm_object_allocate(OBJT_DEFAULT,
614                                          atop(entry->end - entry->start));
615         }
616         entry->object.vm_object = obj;
617         entry->offset = 0;
618 }
619
620 /*
621  * Set an initial negative count so the first attempt to reserve
622  * space preloads a bunch of vm_map_entry's for this cpu.  Also
623  * pre-allocate 2 vm_map_entries which will be needed by zalloc() to
624  * map a new page for vm_map_entry structures.  SMP systems are
625  * particularly sensitive.
626  *
627  * This routine is called in early boot so we cannot just call
628  * vm_map_entry_reserve().
629  *
630  * Called from the low level boot code only (for each cpu)
631  */
632 void
633 vm_map_entry_reserve_cpu_init(globaldata_t gd)
634 {
635         vm_map_entry_t entry;
636         int i;
637
638         gd->gd_vme_avail -= MAP_RESERVE_COUNT * 2;
639         entry = &cpu_map_entry_init[gd->gd_cpuid][0];
640         for (i = 0; i < VMEPERCPU; ++i, ++entry) {
641                 entry->next = gd->gd_vme_base;
642                 gd->gd_vme_base = entry;
643         }
644 }
645
646 /*
647  * Reserves vm_map_entry structures so code later on can manipulate
648  * map_entry structures within a locked map without blocking trying
649  * to allocate a new vm_map_entry.
650  *
651  * No requirements.
652  */
653 int
654 vm_map_entry_reserve(int count)
655 {
656         struct globaldata *gd = mycpu;
657         vm_map_entry_t entry;
658
659         /*
660          * Make sure we have enough structures in gd_vme_base to handle
661          * the reservation request.
662          *
663          * The critical section protects access to the per-cpu gd.
664          */
665         crit_enter();
666         while (gd->gd_vme_avail < count) {
667                 entry = zalloc(mapentzone);
668                 entry->next = gd->gd_vme_base;
669                 gd->gd_vme_base = entry;
670                 ++gd->gd_vme_avail;
671         }
672         gd->gd_vme_avail -= count;
673         crit_exit();
674
675         return(count);
676 }
677
678 /*
679  * Releases previously reserved vm_map_entry structures that were not
680  * used.  If we have too much junk in our per-cpu cache clean some of
681  * it out.
682  *
683  * No requirements.
684  */
685 void
686 vm_map_entry_release(int count)
687 {
688         struct globaldata *gd = mycpu;
689         vm_map_entry_t entry;
690
691         crit_enter();
692         gd->gd_vme_avail += count;
693         while (gd->gd_vme_avail > MAP_RESERVE_SLOP) {
694                 entry = gd->gd_vme_base;
695                 KKASSERT(entry != NULL);
696                 gd->gd_vme_base = entry->next;
697                 --gd->gd_vme_avail;
698                 crit_exit();
699                 zfree(mapentzone, entry);
700                 crit_enter();
701         }
702         crit_exit();
703 }
704
705 /*
706  * Reserve map entry structures for use in kernel_map itself.  These
707  * entries have *ALREADY* been reserved on a per-cpu basis when the map
708  * was inited.  This function is used by zalloc() to avoid a recursion
709  * when zalloc() itself needs to allocate additional kernel memory.
710  *
711  * This function works like the normal reserve but does not load the
712  * vm_map_entry cache (because that would result in an infinite
713  * recursion).  Note that gd_vme_avail may go negative.  This is expected.
714  *
715  * Any caller of this function must be sure to renormalize after
716  * potentially eating entries to ensure that the reserve supply
717  * remains intact.
718  *
719  * No requirements.
720  */
721 int
722 vm_map_entry_kreserve(int count)
723 {
724         struct globaldata *gd = mycpu;
725
726         crit_enter();
727         gd->gd_vme_avail -= count;
728         crit_exit();
729         KASSERT(gd->gd_vme_base != NULL,
730                 ("no reserved entries left, gd_vme_avail = %d\n",
731                 gd->gd_vme_avail));
732         return(count);
733 }
734
735 /*
736  * Release previously reserved map entries for kernel_map.  We do not
737  * attempt to clean up like the normal release function as this would
738  * cause an unnecessary (but probably not fatal) deep procedure call.
739  *
740  * No requirements.
741  */
742 void
743 vm_map_entry_krelease(int count)
744 {
745         struct globaldata *gd = mycpu;
746
747         crit_enter();
748         gd->gd_vme_avail += count;
749         crit_exit();
750 }
751
752 /*
753  * Allocates a VM map entry for insertion.  No entry fields are filled in.
754  *
755  * The entries should have previously been reserved.  The reservation count
756  * is tracked in (*countp).
757  *
758  * No requirements.
759  */
760 static vm_map_entry_t
761 vm_map_entry_create(vm_map_t map, int *countp)
762 {
763         struct globaldata *gd = mycpu;
764         vm_map_entry_t entry;
765
766         KKASSERT(*countp > 0);
767         --*countp;
768         crit_enter();
769         entry = gd->gd_vme_base;
770         KASSERT(entry != NULL, ("gd_vme_base NULL! count %d", *countp));
771         gd->gd_vme_base = entry->next;
772         crit_exit();
773
774         return(entry);
775 }
776
777 /*
778  * Dispose of a vm_map_entry that is no longer being referenced.
779  *
780  * No requirements.
781  */
782 static void
783 vm_map_entry_dispose(vm_map_t map, vm_map_entry_t entry, int *countp)
784 {
785         struct globaldata *gd = mycpu;
786
787         KKASSERT(map->hint != entry);
788         KKASSERT(map->first_free != entry);
789
790         ++*countp;
791         crit_enter();
792         entry->next = gd->gd_vme_base;
793         gd->gd_vme_base = entry;
794         crit_exit();
795 }
796
797
798 /*
799  * Insert/remove entries from maps.
800  *
801  * The related map must be exclusively locked.
802  * The caller must hold map->token
803  * No other requirements.
804  */
805 static __inline void
806 vm_map_entry_link(vm_map_t map,
807                   vm_map_entry_t after_where,
808                   vm_map_entry_t entry)
809 {
810         ASSERT_VM_MAP_LOCKED(map);
811
812         map->nentries++;
813         entry->prev = after_where;
814         entry->next = after_where->next;
815         entry->next->prev = entry;
816         after_where->next = entry;
817         if (vm_map_rb_tree_RB_INSERT(&map->rb_root, entry))
818                 panic("vm_map_entry_link: dup addr map %p ent %p", map, entry);
819 }
820
821 static __inline void
822 vm_map_entry_unlink(vm_map_t map,
823                     vm_map_entry_t entry)
824 {
825         vm_map_entry_t prev;
826         vm_map_entry_t next;
827
828         ASSERT_VM_MAP_LOCKED(map);
829
830         if (entry->eflags & MAP_ENTRY_IN_TRANSITION) {
831                 panic("vm_map_entry_unlink: attempt to mess with "
832                       "locked entry! %p", entry);
833         }
834         prev = entry->prev;
835         next = entry->next;
836         next->prev = prev;
837         prev->next = next;
838         vm_map_rb_tree_RB_REMOVE(&map->rb_root, entry);
839         map->nentries--;
840 }
841
842 /*
843  * Finds the map entry containing (or immediately preceding) the specified
844  * address in the given map.  The entry is returned in (*entry).
845  *
846  * The boolean result indicates whether the address is actually contained
847  * in the map.
848  *
849  * The related map must be locked.
850  * No other requirements.
851  */
852 boolean_t
853 vm_map_lookup_entry(vm_map_t map, vm_offset_t address, vm_map_entry_t *entry)
854 {
855         vm_map_entry_t tmp;
856         vm_map_entry_t last;
857
858         ASSERT_VM_MAP_LOCKED(map);
859 #if 0
860         /*
861          * XXX TEMPORARILY DISABLED.  For some reason our attempt to revive
862          * the hint code with the red-black lookup meets with system crashes
863          * and lockups.  We do not yet know why.
864          *
865          * It is possible that the problem is related to the setting
866          * of the hint during map_entry deletion, in the code specified
867          * at the GGG comment later on in this file.
868          *
869          * YYY More likely it's because this function can be called with
870          * a shared lock on the map, resulting in map->hint updates possibly
871          * racing.  Fixed now but untested.
872          */
873         /*
874          * Quickly check the cached hint, there's a good chance of a match.
875          */
876         tmp = map->hint;
877         cpu_ccfence();
878         if (tmp != &map->header) {
879                 if (address >= tmp->start && address < tmp->end) {
880                         *entry = tmp;
881                         return(TRUE);
882                 }
883         }
884 #endif
885
886         /*
887          * Locate the record from the top of the tree.  'last' tracks the
888          * closest prior record and is returned if no match is found, which
889          * in binary tree terms means tracking the most recent right-branch
890          * taken.  If there is no prior record, &map->header is returned.
891          */
892         last = &map->header;
893         tmp = RB_ROOT(&map->rb_root);
894
895         while (tmp) {
896                 if (address >= tmp->start) {
897                         if (address < tmp->end) {
898                                 *entry = tmp;
899                                 map->hint = tmp;
900                                 return(TRUE);
901                         }
902                         last = tmp;
903                         tmp = RB_RIGHT(tmp, rb_entry);
904                 } else {
905                         tmp = RB_LEFT(tmp, rb_entry);
906                 }
907         }
908         *entry = last;
909         return (FALSE);
910 }
911
912 /*
913  * Inserts the given whole VM object into the target map at the specified
914  * address range.  The object's size should match that of the address range.
915  *
916  * The map must be exclusively locked.
917  * The object must be held.
918  * The caller must have reserved sufficient vm_map_entry structures.
919  *
920  * If object is non-NULL, ref count must be bumped by caller prior to
921  * making call to account for the new entry.
922  */
923 int
924 vm_map_insert(vm_map_t map, int *countp,
925               vm_object_t object, vm_ooffset_t offset,
926               vm_offset_t start, vm_offset_t end,
927               vm_maptype_t maptype,
928               vm_prot_t prot, vm_prot_t max,
929               int cow)
930 {
931         vm_map_entry_t new_entry;
932         vm_map_entry_t prev_entry;
933         vm_map_entry_t temp_entry;
934         vm_eflags_t protoeflags;
935         int must_drop = 0;
936
937         ASSERT_VM_MAP_LOCKED(map);
938         if (object)
939                 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
940
941         /*
942          * Check that the start and end points are not bogus.
943          */
944         if ((start < map->min_offset) || (end > map->max_offset) ||
945             (start >= end))
946                 return (KERN_INVALID_ADDRESS);
947
948         /*
949          * Find the entry prior to the proposed starting address; if it's part
950          * of an existing entry, this range is bogus.
951          */
952         if (vm_map_lookup_entry(map, start, &temp_entry))
953                 return (KERN_NO_SPACE);
954
955         prev_entry = temp_entry;
956
957         /*
958          * Assert that the next entry doesn't overlap the end point.
959          */
960
961         if ((prev_entry->next != &map->header) &&
962             (prev_entry->next->start < end))
963                 return (KERN_NO_SPACE);
964
965         protoeflags = 0;
966
967         if (cow & MAP_COPY_ON_WRITE)
968                 protoeflags |= MAP_ENTRY_COW|MAP_ENTRY_NEEDS_COPY;
969
970         if (cow & MAP_NOFAULT) {
971                 protoeflags |= MAP_ENTRY_NOFAULT;
972
973                 KASSERT(object == NULL,
974                         ("vm_map_insert: paradoxical MAP_NOFAULT request"));
975         }
976         if (cow & MAP_DISABLE_SYNCER)
977                 protoeflags |= MAP_ENTRY_NOSYNC;
978         if (cow & MAP_DISABLE_COREDUMP)
979                 protoeflags |= MAP_ENTRY_NOCOREDUMP;
980         if (cow & MAP_IS_STACK)
981                 protoeflags |= MAP_ENTRY_STACK;
982         if (cow & MAP_IS_KSTACK)
983                 protoeflags |= MAP_ENTRY_KSTACK;
984
985         lwkt_gettoken(&map->token);
986
987         if (object) {
988                 /*
989                  * When object is non-NULL, it could be shared with another
990                  * process.  We have to set or clear OBJ_ONEMAPPING 
991                  * appropriately.
992                  */
993                 if ((object->ref_count > 1) || (object->shadow_count != 0)) {
994                         vm_object_clear_flag(object, OBJ_ONEMAPPING);
995                 }
996         }
997         else if ((prev_entry != &map->header) &&
998                  (prev_entry->eflags == protoeflags) &&
999                  (prev_entry->end == start) &&
1000                  (prev_entry->wired_count == 0) &&
1001                  prev_entry->maptype == maptype &&
1002                  ((prev_entry->object.vm_object == NULL) ||
1003                   vm_object_coalesce(prev_entry->object.vm_object,
1004                                      OFF_TO_IDX(prev_entry->offset),
1005                                      (vm_size_t)(prev_entry->end - prev_entry->start),
1006                                      (vm_size_t)(end - prev_entry->end)))) {
1007                 /*
1008                  * We were able to extend the object.  Determine if we
1009                  * can extend the previous map entry to include the 
1010                  * new range as well.
1011                  */
1012                 if ((prev_entry->inheritance == VM_INHERIT_DEFAULT) &&
1013                     (prev_entry->protection == prot) &&
1014                     (prev_entry->max_protection == max)) {
1015                         map->size += (end - prev_entry->end);
1016                         prev_entry->end = end;
1017                         vm_map_simplify_entry(map, prev_entry, countp);
1018                         lwkt_reltoken(&map->token);
1019                         return (KERN_SUCCESS);
1020                 }
1021
1022                 /*
1023                  * If we can extend the object but cannot extend the
1024                  * map entry, we have to create a new map entry.  We
1025                  * must bump the ref count on the extended object to
1026                  * account for it.  object may be NULL.
1027                  */
1028                 object = prev_entry->object.vm_object;
1029                 offset = prev_entry->offset +
1030                         (prev_entry->end - prev_entry->start);
1031                 if (object) {
1032                         vm_object_hold(object);
1033                         vm_object_chain_wait(object);
1034                         vm_object_reference_locked(object);
1035                         must_drop = 1;
1036                 }
1037         }
1038
1039         /*
1040          * NOTE: if conditionals fail, object can be NULL here.  This occurs
1041          * in things like the buffer map where we manage kva but do not manage
1042          * backing objects.
1043          */
1044
1045         /*
1046          * Create a new entry
1047          */
1048
1049         new_entry = vm_map_entry_create(map, countp);
1050         new_entry->start = start;
1051         new_entry->end = end;
1052
1053         new_entry->maptype = maptype;
1054         new_entry->eflags = protoeflags;
1055         new_entry->object.vm_object = object;
1056         new_entry->offset = offset;
1057         new_entry->aux.master_pde = 0;
1058
1059         new_entry->inheritance = VM_INHERIT_DEFAULT;
1060         new_entry->protection = prot;
1061         new_entry->max_protection = max;
1062         new_entry->wired_count = 0;
1063
1064         /*
1065          * Insert the new entry into the list
1066          */
1067
1068         vm_map_entry_link(map, prev_entry, new_entry);
1069         map->size += new_entry->end - new_entry->start;
1070
1071         /*
1072          * Update the free space hint.  Entries cannot overlap.
1073          * An exact comparison is needed to avoid matching
1074          * against the map->header.
1075          */
1076         if ((map->first_free == prev_entry) &&
1077             (prev_entry->end == new_entry->start)) {
1078                 map->first_free = new_entry;
1079         }
1080
1081 #if 0
1082         /*
1083          * Temporarily removed to avoid MAP_STACK panic, due to
1084          * MAP_STACK being a huge hack.  Will be added back in
1085          * when MAP_STACK (and the user stack mapping) is fixed.
1086          */
1087         /*
1088          * It may be possible to simplify the entry
1089          */
1090         vm_map_simplify_entry(map, new_entry, countp);
1091 #endif
1092
1093         /*
1094          * Try to pre-populate the page table.  Mappings governed by virtual
1095          * page tables cannot be prepopulated without a lot of work, so
1096          * don't try.
1097          */
1098         if ((cow & (MAP_PREFAULT|MAP_PREFAULT_PARTIAL)) &&
1099             maptype != VM_MAPTYPE_VPAGETABLE) {
1100                 pmap_object_init_pt(map->pmap, start, prot,
1101                                     object, OFF_TO_IDX(offset), end - start,
1102                                     cow & MAP_PREFAULT_PARTIAL);
1103         }
1104         if (must_drop)
1105                 vm_object_drop(object);
1106
1107         lwkt_reltoken(&map->token);
1108         return (KERN_SUCCESS);
1109 }
1110
1111 /*
1112  * Find sufficient space for `length' bytes in the given map, starting at
1113  * `start'.  Returns 0 on success, 1 on no space.
1114  *
1115  * This function will returned an arbitrarily aligned pointer.  If no
1116  * particular alignment is required you should pass align as 1.  Note that
1117  * the map may return PAGE_SIZE aligned pointers if all the lengths used in
1118  * the map are a multiple of PAGE_SIZE, even if you pass a smaller align
1119  * argument.
1120  *
1121  * 'align' should be a power of 2 but is not required to be.
1122  *
1123  * The map must be exclusively locked.
1124  * No other requirements.
1125  */
1126 int
1127 vm_map_findspace(vm_map_t map, vm_offset_t start, vm_size_t length,
1128                  vm_size_t align, int flags, vm_offset_t *addr)
1129 {
1130         vm_map_entry_t entry, next;
1131         vm_offset_t end;
1132         vm_offset_t align_mask;
1133
1134         if (start < map->min_offset)
1135                 start = map->min_offset;
1136         if (start > map->max_offset)
1137                 return (1);
1138
1139         /*
1140          * If the alignment is not a power of 2 we will have to use
1141          * a mod/division, set align_mask to a special value.
1142          */
1143         if ((align | (align - 1)) + 1 != (align << 1))
1144                 align_mask = (vm_offset_t)-1;
1145         else
1146                 align_mask = align - 1;
1147
1148         /*
1149          * Look for the first possible address; if there's already something
1150          * at this address, we have to start after it.
1151          */
1152         if (start == map->min_offset) {
1153                 if ((entry = map->first_free) != &map->header)
1154                         start = entry->end;
1155         } else {
1156                 vm_map_entry_t tmp;
1157
1158                 if (vm_map_lookup_entry(map, start, &tmp))
1159                         start = tmp->end;
1160                 entry = tmp;
1161         }
1162
1163         /*
1164          * Look through the rest of the map, trying to fit a new region in the
1165          * gap between existing regions, or after the very last region.
1166          */
1167         for (;; start = (entry = next)->end) {
1168                 /*
1169                  * Adjust the proposed start by the requested alignment,
1170                  * be sure that we didn't wrap the address.
1171                  */
1172                 if (align_mask == (vm_offset_t)-1)
1173                         end = ((start + align - 1) / align) * align;
1174                 else
1175                         end = (start + align_mask) & ~align_mask;
1176                 if (end < start)
1177                         return (1);
1178                 start = end;
1179                 /*
1180                  * Find the end of the proposed new region.  Be sure we didn't
1181                  * go beyond the end of the map, or wrap around the address.
1182                  * Then check to see if this is the last entry or if the 
1183                  * proposed end fits in the gap between this and the next
1184                  * entry.
1185                  */
1186                 end = start + length;
1187                 if (end > map->max_offset || end < start)
1188                         return (1);
1189                 next = entry->next;
1190
1191                 /*
1192                  * If the next entry's start address is beyond the desired
1193                  * end address we may have found a good entry.
1194                  *
1195                  * If the next entry is a stack mapping we do not map into
1196                  * the stack's reserved space.
1197                  *
1198                  * XXX continue to allow mapping into the stack's reserved
1199                  * space if doing a MAP_STACK mapping inside a MAP_STACK
1200                  * mapping, for backwards compatibility.  But the caller
1201                  * really should use MAP_STACK | MAP_TRYFIXED if they
1202                  * want to do that.
1203                  */
1204                 if (next == &map->header)
1205                         break;
1206                 if (next->start >= end) {
1207                         if ((next->eflags & MAP_ENTRY_STACK) == 0)
1208                                 break;
1209                         if (flags & MAP_STACK)
1210                                 break;
1211                         if (next->start - next->aux.avail_ssize >= end)
1212                                 break;
1213                 }
1214         }
1215         map->hint = entry;
1216
1217         /*
1218          * Grow the kernel_map if necessary.  pmap_growkernel() will panic
1219          * if it fails.  The kernel_map is locked and nothing can steal
1220          * our address space if pmap_growkernel() blocks.
1221          *
1222          * NOTE: This may be unconditionally called for kldload areas on
1223          *       x86_64 because these do not bump kernel_vm_end (which would
1224          *       fill 128G worth of page tables!).  Therefore we must not
1225          *       retry.
1226          */
1227         if (map == &kernel_map) {
1228                 vm_offset_t kstop;
1229
1230                 kstop = round_page(start + length);
1231                 if (kstop > kernel_vm_end)
1232                         pmap_growkernel(start, kstop);
1233         }
1234         *addr = start;
1235         return (0);
1236 }
1237
1238 /*
1239  * vm_map_find finds an unallocated region in the target address map with
1240  * the given length and allocates it.  The search is defined to be first-fit
1241  * from the specified address; the region found is returned in the same
1242  * parameter.
1243  *
1244  * If object is non-NULL, ref count must be bumped by caller
1245  * prior to making call to account for the new entry.
1246  *
1247  * No requirements.  This function will lock the map temporarily.
1248  */
1249 int
1250 vm_map_find(vm_map_t map, vm_object_t object, vm_ooffset_t offset,
1251             vm_offset_t *addr,  vm_size_t length, vm_size_t align,
1252             boolean_t fitit,
1253             vm_maptype_t maptype,
1254             vm_prot_t prot, vm_prot_t max,
1255             int cow)
1256 {
1257         vm_offset_t start;
1258         int result;
1259         int count;
1260
1261         start = *addr;
1262
1263         count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
1264         vm_map_lock(map);
1265         if (object)
1266                 vm_object_hold(object);
1267         if (fitit) {
1268                 if (vm_map_findspace(map, start, length, align, 0, addr)) {
1269                         vm_map_unlock(map);
1270                         vm_map_entry_release(count);
1271                         return (KERN_NO_SPACE);
1272                 }
1273                 start = *addr;
1274         }
1275         result = vm_map_insert(map, &count, object, offset,
1276                                start, start + length,
1277                                maptype,
1278                                prot, max,
1279                                cow);
1280         if (object)
1281                 vm_object_drop(object);
1282         vm_map_unlock(map);
1283         vm_map_entry_release(count);
1284
1285         return (result);
1286 }
1287
1288 /*
1289  * Simplify the given map entry by merging with either neighbor.  This
1290  * routine also has the ability to merge with both neighbors.
1291  *
1292  * This routine guarentees that the passed entry remains valid (though
1293  * possibly extended).  When merging, this routine may delete one or
1294  * both neighbors.  No action is taken on entries which have their
1295  * in-transition flag set.
1296  *
1297  * The map must be exclusively locked.
1298  */
1299 void
1300 vm_map_simplify_entry(vm_map_t map, vm_map_entry_t entry, int *countp)
1301 {
1302         vm_map_entry_t next, prev;
1303         vm_size_t prevsize, esize;
1304
1305         if (entry->eflags & MAP_ENTRY_IN_TRANSITION) {
1306                 ++mycpu->gd_cnt.v_intrans_coll;
1307                 return;
1308         }
1309
1310         if (entry->maptype == VM_MAPTYPE_SUBMAP)
1311                 return;
1312
1313         prev = entry->prev;
1314         if (prev != &map->header) {
1315                 prevsize = prev->end - prev->start;
1316                 if ( (prev->end == entry->start) &&
1317                      (prev->maptype == entry->maptype) &&
1318                      (prev->object.vm_object == entry->object.vm_object) &&
1319                      (!prev->object.vm_object ||
1320                         (prev->offset + prevsize == entry->offset)) &&
1321                      (prev->eflags == entry->eflags) &&
1322                      (prev->protection == entry->protection) &&
1323                      (prev->max_protection == entry->max_protection) &&
1324                      (prev->inheritance == entry->inheritance) &&
1325                      (prev->wired_count == entry->wired_count)) {
1326                         if (map->first_free == prev)
1327                                 map->first_free = entry;
1328                         if (map->hint == prev)
1329                                 map->hint = entry;
1330                         vm_map_entry_unlink(map, prev);
1331                         entry->start = prev->start;
1332                         entry->offset = prev->offset;
1333                         if (prev->object.vm_object)
1334                                 vm_object_deallocate(prev->object.vm_object);
1335                         vm_map_entry_dispose(map, prev, countp);
1336                 }
1337         }
1338
1339         next = entry->next;
1340         if (next != &map->header) {
1341                 esize = entry->end - entry->start;
1342                 if ((entry->end == next->start) &&
1343                     (next->maptype == entry->maptype) &&
1344                     (next->object.vm_object == entry->object.vm_object) &&
1345                      (!entry->object.vm_object ||
1346                         (entry->offset + esize == next->offset)) &&
1347                     (next->eflags == entry->eflags) &&
1348                     (next->protection == entry->protection) &&
1349                     (next->max_protection == entry->max_protection) &&
1350                     (next->inheritance == entry->inheritance) &&
1351                     (next->wired_count == entry->wired_count)) {
1352                         if (map->first_free == next)
1353                                 map->first_free = entry;
1354                         if (map->hint == next)
1355                                 map->hint = entry;
1356                         vm_map_entry_unlink(map, next);
1357                         entry->end = next->end;
1358                         if (next->object.vm_object)
1359                                 vm_object_deallocate(next->object.vm_object);
1360                         vm_map_entry_dispose(map, next, countp);
1361                 }
1362         }
1363 }
1364
1365 /*
1366  * Asserts that the given entry begins at or after the specified address.
1367  * If necessary, it splits the entry into two.
1368  */
1369 #define vm_map_clip_start(map, entry, startaddr, countp)                \
1370 {                                                                       \
1371         if (startaddr > entry->start)                                   \
1372                 _vm_map_clip_start(map, entry, startaddr, countp);      \
1373 }
1374
1375 /*
1376  * This routine is called only when it is known that the entry must be split.
1377  *
1378  * The map must be exclusively locked.
1379  */
1380 static void
1381 _vm_map_clip_start(vm_map_t map, vm_map_entry_t entry, vm_offset_t start,
1382                    int *countp)
1383 {
1384         vm_map_entry_t new_entry;
1385
1386         /*
1387          * Split off the front portion -- note that we must insert the new
1388          * entry BEFORE this one, so that this entry has the specified
1389          * starting address.
1390          */
1391
1392         vm_map_simplify_entry(map, entry, countp);
1393
1394         /*
1395          * If there is no object backing this entry, we might as well create
1396          * one now.  If we defer it, an object can get created after the map
1397          * is clipped, and individual objects will be created for the split-up
1398          * map.  This is a bit of a hack, but is also about the best place to
1399          * put this improvement.
1400          */
1401         if (entry->object.vm_object == NULL && !map->system_map) {
1402                 vm_map_entry_allocate_object(entry);
1403         }
1404
1405         new_entry = vm_map_entry_create(map, countp);
1406         *new_entry = *entry;
1407
1408         new_entry->end = start;
1409         entry->offset += (start - entry->start);
1410         entry->start = start;
1411
1412         vm_map_entry_link(map, entry->prev, new_entry);
1413
1414         switch(entry->maptype) {
1415         case VM_MAPTYPE_NORMAL:
1416         case VM_MAPTYPE_VPAGETABLE:
1417                 if (new_entry->object.vm_object) {
1418                         vm_object_hold(new_entry->object.vm_object);
1419                         vm_object_chain_wait(new_entry->object.vm_object);
1420                         vm_object_reference_locked(new_entry->object.vm_object);
1421                         vm_object_drop(new_entry->object.vm_object);
1422                 }
1423                 break;
1424         default:
1425                 break;
1426         }
1427 }
1428
1429 /*
1430  * Asserts that the given entry ends at or before the specified address.
1431  * If necessary, it splits the entry into two.
1432  *
1433  * The map must be exclusively locked.
1434  */
1435 #define vm_map_clip_end(map, entry, endaddr, countp)            \
1436 {                                                               \
1437         if (endaddr < entry->end)                               \
1438                 _vm_map_clip_end(map, entry, endaddr, countp);  \
1439 }
1440
1441 /*
1442  * This routine is called only when it is known that the entry must be split.
1443  *
1444  * The map must be exclusively locked.
1445  */
1446 static void
1447 _vm_map_clip_end(vm_map_t map, vm_map_entry_t entry, vm_offset_t end,
1448                  int *countp)
1449 {
1450         vm_map_entry_t new_entry;
1451
1452         /*
1453          * If there is no object backing this entry, we might as well create
1454          * one now.  If we defer it, an object can get created after the map
1455          * is clipped, and individual objects will be created for the split-up
1456          * map.  This is a bit of a hack, but is also about the best place to
1457          * put this improvement.
1458          */
1459
1460         if (entry->object.vm_object == NULL && !map->system_map) {
1461                 vm_map_entry_allocate_object(entry);
1462         }
1463
1464         /*
1465          * Create a new entry and insert it AFTER the specified entry
1466          */
1467
1468         new_entry = vm_map_entry_create(map, countp);
1469         *new_entry = *entry;
1470
1471         new_entry->start = entry->end = end;
1472         new_entry->offset += (end - entry->start);
1473
1474         vm_map_entry_link(map, entry, new_entry);
1475
1476         switch(entry->maptype) {
1477         case VM_MAPTYPE_NORMAL:
1478         case VM_MAPTYPE_VPAGETABLE:
1479                 if (new_entry->object.vm_object) {
1480                         vm_object_hold(new_entry->object.vm_object);
1481                         vm_object_chain_wait(new_entry->object.vm_object);
1482                         vm_object_reference_locked(new_entry->object.vm_object);
1483                         vm_object_drop(new_entry->object.vm_object);
1484                 }
1485                 break;
1486         default:
1487                 break;
1488         }
1489 }
1490
1491 /*
1492  * Asserts that the starting and ending region addresses fall within the
1493  * valid range for the map.
1494  */
1495 #define VM_MAP_RANGE_CHECK(map, start, end)     \
1496 {                                               \
1497         if (start < vm_map_min(map))            \
1498                 start = vm_map_min(map);        \
1499         if (end > vm_map_max(map))              \
1500                 end = vm_map_max(map);          \
1501         if (start > end)                        \
1502                 start = end;                    \
1503 }
1504
1505 /*
1506  * Used to block when an in-transition collison occurs.  The map
1507  * is unlocked for the sleep and relocked before the return.
1508  */
1509 void
1510 vm_map_transition_wait(vm_map_t map)
1511 {
1512         tsleep_interlock(map, 0);
1513         vm_map_unlock(map);
1514         tsleep(map, PINTERLOCKED, "vment", 0);
1515         vm_map_lock(map);
1516 }
1517
1518 /*
1519  * When we do blocking operations with the map lock held it is
1520  * possible that a clip might have occured on our in-transit entry,
1521  * requiring an adjustment to the entry in our loop.  These macros
1522  * help the pageable and clip_range code deal with the case.  The
1523  * conditional costs virtually nothing if no clipping has occured.
1524  */
1525
1526 #define CLIP_CHECK_BACK(entry, save_start)              \
1527     do {                                                \
1528             while (entry->start != save_start) {        \
1529                     entry = entry->prev;                \
1530                     KASSERT(entry != &map->header, ("bad entry clip")); \
1531             }                                           \
1532     } while(0)
1533
1534 #define CLIP_CHECK_FWD(entry, save_end)                 \
1535     do {                                                \
1536             while (entry->end != save_end) {            \
1537                     entry = entry->next;                \
1538                     KASSERT(entry != &map->header, ("bad entry clip")); \
1539             }                                           \
1540     } while(0)
1541
1542
1543 /*
1544  * Clip the specified range and return the base entry.  The
1545  * range may cover several entries starting at the returned base
1546  * and the first and last entry in the covering sequence will be
1547  * properly clipped to the requested start and end address.
1548  *
1549  * If no holes are allowed you should pass the MAP_CLIP_NO_HOLES
1550  * flag.
1551  *
1552  * The MAP_ENTRY_IN_TRANSITION flag will be set for the entries
1553  * covered by the requested range.
1554  *
1555  * The map must be exclusively locked on entry and will remain locked
1556  * on return. If no range exists or the range contains holes and you
1557  * specified that no holes were allowed, NULL will be returned.  This
1558  * routine may temporarily unlock the map in order avoid a deadlock when
1559  * sleeping.
1560  */
1561 static
1562 vm_map_entry_t
1563 vm_map_clip_range(vm_map_t map, vm_offset_t start, vm_offset_t end, 
1564                   int *countp, int flags)
1565 {
1566         vm_map_entry_t start_entry;
1567         vm_map_entry_t entry;
1568
1569         /*
1570          * Locate the entry and effect initial clipping.  The in-transition
1571          * case does not occur very often so do not try to optimize it.
1572          */
1573 again:
1574         if (vm_map_lookup_entry(map, start, &start_entry) == FALSE)
1575                 return (NULL);
1576         entry = start_entry;
1577         if (entry->eflags & MAP_ENTRY_IN_TRANSITION) {
1578                 entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP;
1579                 ++mycpu->gd_cnt.v_intrans_coll;
1580                 ++mycpu->gd_cnt.v_intrans_wait;
1581                 vm_map_transition_wait(map);
1582                 /*
1583                  * entry and/or start_entry may have been clipped while
1584                  * we slept, or may have gone away entirely.  We have
1585                  * to restart from the lookup.
1586                  */
1587                 goto again;
1588         }
1589
1590         /*
1591          * Since we hold an exclusive map lock we do not have to restart
1592          * after clipping, even though clipping may block in zalloc.
1593          */
1594         vm_map_clip_start(map, entry, start, countp);
1595         vm_map_clip_end(map, entry, end, countp);
1596         entry->eflags |= MAP_ENTRY_IN_TRANSITION;
1597
1598         /*
1599          * Scan entries covered by the range.  When working on the next
1600          * entry a restart need only re-loop on the current entry which
1601          * we have already locked, since 'next' may have changed.  Also,
1602          * even though entry is safe, it may have been clipped so we
1603          * have to iterate forwards through the clip after sleeping.
1604          */
1605         while (entry->next != &map->header && entry->next->start < end) {
1606                 vm_map_entry_t next = entry->next;
1607
1608                 if (flags & MAP_CLIP_NO_HOLES) {
1609                         if (next->start > entry->end) {
1610                                 vm_map_unclip_range(map, start_entry,
1611                                         start, entry->end, countp, flags);
1612                                 return(NULL);
1613                         }
1614                 }
1615
1616                 if (next->eflags & MAP_ENTRY_IN_TRANSITION) {
1617                         vm_offset_t save_end = entry->end;
1618                         next->eflags |= MAP_ENTRY_NEEDS_WAKEUP;
1619                         ++mycpu->gd_cnt.v_intrans_coll;
1620                         ++mycpu->gd_cnt.v_intrans_wait;
1621                         vm_map_transition_wait(map);
1622
1623                         /*
1624                          * clips might have occured while we blocked.
1625                          */
1626                         CLIP_CHECK_FWD(entry, save_end);
1627                         CLIP_CHECK_BACK(start_entry, start);
1628                         continue;
1629                 }
1630                 /*
1631                  * No restart necessary even though clip_end may block, we
1632                  * are holding the map lock.
1633                  */
1634                 vm_map_clip_end(map, next, end, countp);
1635                 next->eflags |= MAP_ENTRY_IN_TRANSITION;
1636                 entry = next;
1637         }
1638         if (flags & MAP_CLIP_NO_HOLES) {
1639                 if (entry->end != end) {
1640                         vm_map_unclip_range(map, start_entry,
1641                                 start, entry->end, countp, flags);
1642                         return(NULL);
1643                 }
1644         }
1645         return(start_entry);
1646 }
1647
1648 /*
1649  * Undo the effect of vm_map_clip_range().  You should pass the same
1650  * flags and the same range that you passed to vm_map_clip_range().
1651  * This code will clear the in-transition flag on the entries and
1652  * wake up anyone waiting.  This code will also simplify the sequence
1653  * and attempt to merge it with entries before and after the sequence.
1654  *
1655  * The map must be locked on entry and will remain locked on return.
1656  *
1657  * Note that you should also pass the start_entry returned by
1658  * vm_map_clip_range().  However, if you block between the two calls
1659  * with the map unlocked please be aware that the start_entry may
1660  * have been clipped and you may need to scan it backwards to find
1661  * the entry corresponding with the original start address.  You are
1662  * responsible for this, vm_map_unclip_range() expects the correct
1663  * start_entry to be passed to it and will KASSERT otherwise.
1664  */
1665 static
1666 void
1667 vm_map_unclip_range(vm_map_t map, vm_map_entry_t start_entry,
1668                     vm_offset_t start, vm_offset_t end,
1669                     int *countp, int flags)
1670 {
1671         vm_map_entry_t entry;
1672
1673         entry = start_entry;
1674
1675         KASSERT(entry->start == start, ("unclip_range: illegal base entry"));
1676         while (entry != &map->header && entry->start < end) {
1677                 KASSERT(entry->eflags & MAP_ENTRY_IN_TRANSITION,
1678                         ("in-transition flag not set during unclip on: %p",
1679                         entry));
1680                 KASSERT(entry->end <= end,
1681                         ("unclip_range: tail wasn't clipped"));
1682                 entry->eflags &= ~MAP_ENTRY_IN_TRANSITION;
1683                 if (entry->eflags & MAP_ENTRY_NEEDS_WAKEUP) {
1684                         entry->eflags &= ~MAP_ENTRY_NEEDS_WAKEUP;
1685                         wakeup(map);
1686                 }
1687                 entry = entry->next;
1688         }
1689
1690         /*
1691          * Simplification does not block so there is no restart case.
1692          */
1693         entry = start_entry;
1694         while (entry != &map->header && entry->start < end) {
1695                 vm_map_simplify_entry(map, entry, countp);
1696                 entry = entry->next;
1697         }
1698 }
1699
1700 /*
1701  * Mark the given range as handled by a subordinate map.
1702  *
1703  * This range must have been created with vm_map_find(), and no other
1704  * operations may have been performed on this range prior to calling
1705  * vm_map_submap().
1706  *
1707  * Submappings cannot be removed.
1708  *
1709  * No requirements.
1710  */
1711 int
1712 vm_map_submap(vm_map_t map, vm_offset_t start, vm_offset_t end, vm_map_t submap)
1713 {
1714         vm_map_entry_t entry;
1715         int result = KERN_INVALID_ARGUMENT;
1716         int count;
1717
1718         count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
1719         vm_map_lock(map);
1720
1721         VM_MAP_RANGE_CHECK(map, start, end);
1722
1723         if (vm_map_lookup_entry(map, start, &entry)) {
1724                 vm_map_clip_start(map, entry, start, &count);
1725         } else {
1726                 entry = entry->next;
1727         }
1728
1729         vm_map_clip_end(map, entry, end, &count);
1730
1731         if ((entry->start == start) && (entry->end == end) &&
1732             ((entry->eflags & MAP_ENTRY_COW) == 0) &&
1733             (entry->object.vm_object == NULL)) {
1734                 entry->object.sub_map = submap;
1735                 entry->maptype = VM_MAPTYPE_SUBMAP;
1736                 result = KERN_SUCCESS;
1737         }
1738         vm_map_unlock(map);
1739         vm_map_entry_release(count);
1740
1741         return (result);
1742 }
1743
1744 /*
1745  * Sets the protection of the specified address region in the target map. 
1746  * If "set_max" is specified, the maximum protection is to be set;
1747  * otherwise, only the current protection is affected.
1748  *
1749  * The protection is not applicable to submaps, but is applicable to normal
1750  * maps and maps governed by virtual page tables.  For example, when operating
1751  * on a virtual page table our protection basically controls how COW occurs
1752  * on the backing object, whereas the virtual page table abstraction itself
1753  * is an abstraction for userland.
1754  *
1755  * No requirements.
1756  */
1757 int
1758 vm_map_protect(vm_map_t map, vm_offset_t start, vm_offset_t end,
1759                vm_prot_t new_prot, boolean_t set_max)
1760 {
1761         vm_map_entry_t current;
1762         vm_map_entry_t entry;
1763         int count;
1764
1765         count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
1766         vm_map_lock(map);
1767
1768         VM_MAP_RANGE_CHECK(map, start, end);
1769
1770         if (vm_map_lookup_entry(map, start, &entry)) {
1771                 vm_map_clip_start(map, entry, start, &count);
1772         } else {
1773                 entry = entry->next;
1774         }
1775
1776         /*
1777          * Make a first pass to check for protection violations.
1778          */
1779         current = entry;
1780         while ((current != &map->header) && (current->start < end)) {
1781                 if (current->maptype == VM_MAPTYPE_SUBMAP) {
1782                         vm_map_unlock(map);
1783                         vm_map_entry_release(count);
1784                         return (KERN_INVALID_ARGUMENT);
1785                 }
1786                 if ((new_prot & current->max_protection) != new_prot) {
1787                         vm_map_unlock(map);
1788                         vm_map_entry_release(count);
1789                         return (KERN_PROTECTION_FAILURE);
1790                 }
1791                 current = current->next;
1792         }
1793
1794         /*
1795          * Go back and fix up protections. [Note that clipping is not
1796          * necessary the second time.]
1797          */
1798         current = entry;
1799
1800         while ((current != &map->header) && (current->start < end)) {
1801                 vm_prot_t old_prot;
1802
1803                 vm_map_clip_end(map, current, end, &count);
1804
1805                 old_prot = current->protection;
1806                 if (set_max) {
1807                         current->protection =
1808                             (current->max_protection = new_prot) &
1809                             old_prot;
1810                 } else {
1811                         current->protection = new_prot;
1812                 }
1813
1814                 /*
1815                  * Update physical map if necessary. Worry about copy-on-write
1816                  * here -- CHECK THIS XXX
1817                  */
1818
1819                 if (current->protection != old_prot) {
1820 #define MASK(entry)     (((entry)->eflags & MAP_ENTRY_COW) ? ~VM_PROT_WRITE : \
1821                                                         VM_PROT_ALL)
1822
1823                         pmap_protect(map->pmap, current->start,
1824                             current->end,
1825                             current->protection & MASK(current));
1826 #undef  MASK
1827                 }
1828
1829                 vm_map_simplify_entry(map, current, &count);
1830
1831                 current = current->next;
1832         }
1833
1834         vm_map_unlock(map);
1835         vm_map_entry_release(count);
1836         return (KERN_SUCCESS);
1837 }
1838
1839 /*
1840  * This routine traverses a processes map handling the madvise
1841  * system call.  Advisories are classified as either those effecting
1842  * the vm_map_entry structure, or those effecting the underlying
1843  * objects.
1844  *
1845  * The <value> argument is used for extended madvise calls.
1846  *
1847  * No requirements.
1848  */
1849 int
1850 vm_map_madvise(vm_map_t map, vm_offset_t start, vm_offset_t end,
1851                int behav, off_t value)
1852 {
1853         vm_map_entry_t current, entry;
1854         int modify_map = 0;
1855         int error = 0;
1856         int count;
1857
1858         /*
1859          * Some madvise calls directly modify the vm_map_entry, in which case
1860          * we need to use an exclusive lock on the map and we need to perform 
1861          * various clipping operations.  Otherwise we only need a read-lock
1862          * on the map.
1863          */
1864
1865         count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
1866
1867         switch(behav) {
1868         case MADV_NORMAL:
1869         case MADV_SEQUENTIAL:
1870         case MADV_RANDOM:
1871         case MADV_NOSYNC:
1872         case MADV_AUTOSYNC:
1873         case MADV_NOCORE:
1874         case MADV_CORE:
1875         case MADV_SETMAP:
1876         case MADV_INVAL:
1877                 modify_map = 1;
1878                 vm_map_lock(map);
1879                 break;
1880         case MADV_WILLNEED:
1881         case MADV_DONTNEED:
1882         case MADV_FREE:
1883                 vm_map_lock_read(map);
1884                 break;
1885         default:
1886                 vm_map_entry_release(count);
1887                 return (EINVAL);
1888         }
1889
1890         /*
1891          * Locate starting entry and clip if necessary.
1892          */
1893
1894         VM_MAP_RANGE_CHECK(map, start, end);
1895
1896         if (vm_map_lookup_entry(map, start, &entry)) {
1897                 if (modify_map)
1898                         vm_map_clip_start(map, entry, start, &count);
1899         } else {
1900                 entry = entry->next;
1901         }
1902
1903         if (modify_map) {
1904                 /*
1905                  * madvise behaviors that are implemented in the vm_map_entry.
1906                  *
1907                  * We clip the vm_map_entry so that behavioral changes are
1908                  * limited to the specified address range.
1909                  */
1910                 for (current = entry;
1911                      (current != &map->header) && (current->start < end);
1912                      current = current->next
1913                 ) {
1914                         if (current->maptype == VM_MAPTYPE_SUBMAP)
1915                                 continue;
1916
1917                         vm_map_clip_end(map, current, end, &count);
1918
1919                         switch (behav) {
1920                         case MADV_NORMAL:
1921                                 vm_map_entry_set_behavior(current, MAP_ENTRY_BEHAV_NORMAL);
1922                                 break;
1923                         case MADV_SEQUENTIAL:
1924                                 vm_map_entry_set_behavior(current, MAP_ENTRY_BEHAV_SEQUENTIAL);
1925                                 break;
1926                         case MADV_RANDOM:
1927                                 vm_map_entry_set_behavior(current, MAP_ENTRY_BEHAV_RANDOM);
1928                                 break;
1929                         case MADV_NOSYNC:
1930                                 current->eflags |= MAP_ENTRY_NOSYNC;
1931                                 break;
1932                         case MADV_AUTOSYNC:
1933                                 current->eflags &= ~MAP_ENTRY_NOSYNC;
1934                                 break;
1935                         case MADV_NOCORE:
1936                                 current->eflags |= MAP_ENTRY_NOCOREDUMP;
1937                                 break;
1938                         case MADV_CORE:
1939                                 current->eflags &= ~MAP_ENTRY_NOCOREDUMP;
1940                                 break;
1941                         case MADV_INVAL:
1942                                 /*
1943                                  * Invalidate the related pmap entries, used
1944                                  * to flush portions of the real kernel's
1945                                  * pmap when the caller has removed or
1946                                  * modified existing mappings in a virtual
1947                                  * page table.
1948                                  */
1949                                 pmap_remove(map->pmap,
1950                                             current->start, current->end);
1951                                 break;
1952                         case MADV_SETMAP:
1953                                 /*
1954                                  * Set the page directory page for a map
1955                                  * governed by a virtual page table.  Mark
1956                                  * the entry as being governed by a virtual
1957                                  * page table if it is not.
1958                                  *
1959                                  * XXX the page directory page is stored
1960                                  * in the avail_ssize field if the map_entry.
1961                                  *
1962                                  * XXX the map simplification code does not
1963                                  * compare this field so weird things may
1964                                  * happen if you do not apply this function
1965                                  * to the entire mapping governed by the
1966                                  * virtual page table.
1967                                  */
1968                                 if (current->maptype != VM_MAPTYPE_VPAGETABLE) {
1969                                         error = EINVAL;
1970                                         break;
1971                                 }
1972                                 current->aux.master_pde = value;
1973                                 pmap_remove(map->pmap,
1974                                             current->start, current->end);
1975                                 break;
1976                         default:
1977                                 error = EINVAL;
1978                                 break;
1979                         }
1980                         vm_map_simplify_entry(map, current, &count);
1981                 }
1982                 vm_map_unlock(map);
1983         } else {
1984                 vm_pindex_t pindex;
1985                 int count;
1986
1987                 /*
1988                  * madvise behaviors that are implemented in the underlying
1989                  * vm_object.
1990                  *
1991                  * Since we don't clip the vm_map_entry, we have to clip
1992                  * the vm_object pindex and count.
1993                  *
1994                  * NOTE!  We currently do not support these functions on
1995                  * virtual page tables.
1996                  */
1997                 for (current = entry;
1998                      (current != &map->header) && (current->start < end);
1999                      current = current->next
2000                 ) {
2001                         vm_offset_t useStart;
2002
2003                         if (current->maptype != VM_MAPTYPE_NORMAL)
2004                                 continue;
2005
2006                         pindex = OFF_TO_IDX(current->offset);
2007                         count = atop(current->end - current->start);
2008                         useStart = current->start;
2009
2010                         if (current->start < start) {
2011                                 pindex += atop(start - current->start);
2012                                 count -= atop(start - current->start);
2013                                 useStart = start;
2014                         }
2015                         if (current->end > end)
2016                                 count -= atop(current->end - end);
2017
2018                         if (count <= 0)
2019                                 continue;
2020
2021                         vm_object_madvise(current->object.vm_object,
2022                                           pindex, count, behav);
2023
2024                         /*
2025                          * Try to populate the page table.  Mappings governed
2026                          * by virtual page tables cannot be pre-populated
2027                          * without a lot of work so don't try.
2028                          */
2029                         if (behav == MADV_WILLNEED &&
2030                             current->maptype != VM_MAPTYPE_VPAGETABLE) {
2031                                 pmap_object_init_pt(
2032                                     map->pmap, 
2033                                     useStart,
2034                                     current->protection,
2035                                     current->object.vm_object,
2036                                     pindex, 
2037                                     (count << PAGE_SHIFT),
2038                                     MAP_PREFAULT_MADVISE
2039                                 );
2040                         }
2041                 }
2042                 vm_map_unlock_read(map);
2043         }
2044         vm_map_entry_release(count);
2045         return(error);
2046 }       
2047
2048
2049 /*
2050  * Sets the inheritance of the specified address range in the target map.
2051  * Inheritance affects how the map will be shared with child maps at the
2052  * time of vm_map_fork.
2053  */
2054 int
2055 vm_map_inherit(vm_map_t map, vm_offset_t start, vm_offset_t end,
2056                vm_inherit_t new_inheritance)
2057 {
2058         vm_map_entry_t entry;
2059         vm_map_entry_t temp_entry;
2060         int count;
2061
2062         switch (new_inheritance) {
2063         case VM_INHERIT_NONE:
2064         case VM_INHERIT_COPY:
2065         case VM_INHERIT_SHARE:
2066                 break;
2067         default:
2068                 return (KERN_INVALID_ARGUMENT);
2069         }
2070
2071         count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
2072         vm_map_lock(map);
2073
2074         VM_MAP_RANGE_CHECK(map, start, end);
2075
2076         if (vm_map_lookup_entry(map, start, &temp_entry)) {
2077                 entry = temp_entry;
2078                 vm_map_clip_start(map, entry, start, &count);
2079         } else
2080                 entry = temp_entry->next;
2081
2082         while ((entry != &map->header) && (entry->start < end)) {
2083                 vm_map_clip_end(map, entry, end, &count);
2084
2085                 entry->inheritance = new_inheritance;
2086
2087                 vm_map_simplify_entry(map, entry, &count);
2088
2089                 entry = entry->next;
2090         }
2091         vm_map_unlock(map);
2092         vm_map_entry_release(count);
2093         return (KERN_SUCCESS);
2094 }
2095
2096 /*
2097  * Implement the semantics of mlock
2098  */
2099 int
2100 vm_map_unwire(vm_map_t map, vm_offset_t start, vm_offset_t real_end,
2101               boolean_t new_pageable)
2102 {
2103         vm_map_entry_t entry;
2104         vm_map_entry_t start_entry;
2105         vm_offset_t end;
2106         int rv = KERN_SUCCESS;
2107         int count;
2108
2109         count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
2110         vm_map_lock(map);
2111         VM_MAP_RANGE_CHECK(map, start, real_end);
2112         end = real_end;
2113
2114         start_entry = vm_map_clip_range(map, start, end, &count,
2115                                         MAP_CLIP_NO_HOLES);
2116         if (start_entry == NULL) {
2117                 vm_map_unlock(map);
2118                 vm_map_entry_release(count);
2119                 return (KERN_INVALID_ADDRESS);
2120         }
2121
2122         if (new_pageable == 0) {
2123                 entry = start_entry;
2124                 while ((entry != &map->header) && (entry->start < end)) {
2125                         vm_offset_t save_start;
2126                         vm_offset_t save_end;
2127
2128                         /*
2129                          * Already user wired or hard wired (trivial cases)
2130                          */
2131                         if (entry->eflags & MAP_ENTRY_USER_WIRED) {
2132                                 entry = entry->next;
2133                                 continue;
2134                         }
2135                         if (entry->wired_count != 0) {
2136                                 entry->wired_count++;
2137                                 entry->eflags |= MAP_ENTRY_USER_WIRED;
2138                                 entry = entry->next;
2139                                 continue;
2140                         }
2141
2142                         /*
2143                          * A new wiring requires instantiation of appropriate
2144                          * management structures and the faulting in of the
2145                          * page.
2146                          */
2147                         if (entry->maptype != VM_MAPTYPE_SUBMAP) {
2148                                 int copyflag = entry->eflags &
2149                                                MAP_ENTRY_NEEDS_COPY;
2150                                 if (copyflag && ((entry->protection &
2151                                                   VM_PROT_WRITE) != 0)) {
2152                                         vm_map_entry_shadow(entry, 0);
2153                                 } else if (entry->object.vm_object == NULL &&
2154                                            !map->system_map) {
2155                                         vm_map_entry_allocate_object(entry);
2156                                 }
2157                         }
2158                         entry->wired_count++;
2159                         entry->eflags |= MAP_ENTRY_USER_WIRED;
2160
2161                         /*
2162                          * Now fault in the area.  Note that vm_fault_wire()
2163                          * may release the map lock temporarily, it will be
2164                          * relocked on return.  The in-transition
2165                          * flag protects the entries. 
2166                          */
2167                         save_start = entry->start;
2168                         save_end = entry->end;
2169                         rv = vm_fault_wire(map, entry, TRUE);
2170                         if (rv) {
2171                                 CLIP_CHECK_BACK(entry, save_start);
2172                                 for (;;) {
2173                                         KASSERT(entry->wired_count == 1, ("bad wired_count on entry"));
2174                                         entry->eflags &= ~MAP_ENTRY_USER_WIRED;
2175                                         entry->wired_count = 0;
2176                                         if (entry->end == save_end)
2177                                                 break;
2178                                         entry = entry->next;
2179                                         KASSERT(entry != &map->header, ("bad entry clip during backout"));
2180                                 }
2181                                 end = save_start;       /* unwire the rest */
2182                                 break;
2183                         }
2184                         /*
2185                          * note that even though the entry might have been
2186                          * clipped, the USER_WIRED flag we set prevents
2187                          * duplication so we do not have to do a 
2188                          * clip check.
2189                          */
2190                         entry = entry->next;
2191                 }
2192
2193                 /*
2194                  * If we failed fall through to the unwiring section to
2195                  * unwire what we had wired so far.  'end' has already
2196                  * been adjusted.
2197                  */
2198                 if (rv)
2199                         new_pageable = 1;
2200
2201                 /*
2202                  * start_entry might have been clipped if we unlocked the
2203                  * map and blocked.  No matter how clipped it has gotten
2204                  * there should be a fragment that is on our start boundary.
2205                  */
2206                 CLIP_CHECK_BACK(start_entry, start);
2207         }
2208
2209         /*
2210          * Deal with the unwiring case.
2211          */
2212         if (new_pageable) {
2213                 /*
2214                  * This is the unwiring case.  We must first ensure that the
2215                  * range to be unwired is really wired down.  We know there
2216                  * are no holes.
2217                  */
2218                 entry = start_entry;
2219                 while ((entry != &map->header) && (entry->start < end)) {
2220                         if ((entry->eflags & MAP_ENTRY_USER_WIRED) == 0) {
2221                                 rv = KERN_INVALID_ARGUMENT;
2222                                 goto done;
2223                         }
2224                         KASSERT(entry->wired_count != 0, ("wired count was 0 with USER_WIRED set! %p", entry));
2225                         entry = entry->next;
2226                 }
2227
2228                 /*
2229                  * Now decrement the wiring count for each region. If a region
2230                  * becomes completely unwired, unwire its physical pages and
2231                  * mappings.
2232                  */
2233                 /*
2234                  * The map entries are processed in a loop, checking to
2235                  * make sure the entry is wired and asserting it has a wired
2236                  * count. However, another loop was inserted more-or-less in
2237                  * the middle of the unwiring path. This loop picks up the
2238                  * "entry" loop variable from the first loop without first
2239                  * setting it to start_entry. Naturally, the secound loop
2240                  * is never entered and the pages backing the entries are
2241                  * never unwired. This can lead to a leak of wired pages.
2242                  */
2243                 entry = start_entry;
2244                 while ((entry != &map->header) && (entry->start < end)) {
2245                         KASSERT(entry->eflags & MAP_ENTRY_USER_WIRED,
2246                                 ("expected USER_WIRED on entry %p", entry));
2247                         entry->eflags &= ~MAP_ENTRY_USER_WIRED;
2248                         entry->wired_count--;
2249                         if (entry->wired_count == 0)
2250                                 vm_fault_unwire(map, entry);
2251                         entry = entry->next;
2252                 }
2253         }
2254 done:
2255         vm_map_unclip_range(map, start_entry, start, real_end, &count,
2256                 MAP_CLIP_NO_HOLES);
2257         map->timestamp++;
2258         vm_map_unlock(map);
2259         vm_map_entry_release(count);
2260         return (rv);
2261 }
2262
2263 /*
2264  * Sets the pageability of the specified address range in the target map.
2265  * Regions specified as not pageable require locked-down physical
2266  * memory and physical page maps.
2267  *
2268  * The map must not be locked, but a reference must remain to the map
2269  * throughout the call.
2270  *
2271  * This function may be called via the zalloc path and must properly
2272  * reserve map entries for kernel_map.
2273  *
2274  * No requirements.
2275  */
2276 int
2277 vm_map_wire(vm_map_t map, vm_offset_t start, vm_offset_t real_end, int kmflags)
2278 {
2279         vm_map_entry_t entry;
2280         vm_map_entry_t start_entry;
2281         vm_offset_t end;
2282         int rv = KERN_SUCCESS;
2283         int count;
2284
2285         if (kmflags & KM_KRESERVE)
2286                 count = vm_map_entry_kreserve(MAP_RESERVE_COUNT);
2287         else
2288                 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
2289         vm_map_lock(map);
2290         VM_MAP_RANGE_CHECK(map, start, real_end);
2291         end = real_end;
2292
2293         start_entry = vm_map_clip_range(map, start, end, &count,
2294                                         MAP_CLIP_NO_HOLES);
2295         if (start_entry == NULL) {
2296                 vm_map_unlock(map);
2297                 rv = KERN_INVALID_ADDRESS;
2298                 goto failure;
2299         }
2300         if ((kmflags & KM_PAGEABLE) == 0) {
2301                 /*
2302                  * Wiring.  
2303                  *
2304                  * 1.  Holding the write lock, we create any shadow or zero-fill
2305                  * objects that need to be created. Then we clip each map
2306                  * entry to the region to be wired and increment its wiring
2307                  * count.  We create objects before clipping the map entries
2308                  * to avoid object proliferation.
2309                  *
2310                  * 2.  We downgrade to a read lock, and call vm_fault_wire to
2311                  * fault in the pages for any newly wired area (wired_count is
2312                  * 1).
2313                  *
2314                  * Downgrading to a read lock for vm_fault_wire avoids a 
2315                  * possible deadlock with another process that may have faulted
2316                  * on one of the pages to be wired (it would mark the page busy,
2317                  * blocking us, then in turn block on the map lock that we
2318                  * hold).  Because of problems in the recursive lock package,
2319                  * we cannot upgrade to a write lock in vm_map_lookup.  Thus,
2320                  * any actions that require the write lock must be done
2321                  * beforehand.  Because we keep the read lock on the map, the
2322                  * copy-on-write status of the entries we modify here cannot
2323                  * change.
2324                  */
2325                 entry = start_entry;
2326                 while ((entry != &map->header) && (entry->start < end)) {
2327                         /*
2328                          * Trivial case if the entry is already wired
2329                          */
2330                         if (entry->wired_count) {
2331                                 entry->wired_count++;
2332                                 entry = entry->next;
2333                                 continue;
2334                         }
2335
2336                         /*
2337                          * The entry is being newly wired, we have to setup
2338                          * appropriate management structures.  A shadow 
2339                          * object is required for a copy-on-write region,
2340                          * or a normal object for a zero-fill region.  We
2341                          * do not have to do this for entries that point to sub
2342                          * maps because we won't hold the lock on the sub map.
2343                          */
2344                         if (entry->maptype != VM_MAPTYPE_SUBMAP) {
2345                                 int copyflag = entry->eflags &
2346                                                MAP_ENTRY_NEEDS_COPY;
2347                                 if (copyflag && ((entry->protection &
2348                                                   VM_PROT_WRITE) != 0)) {
2349                                         vm_map_entry_shadow(entry, 0);
2350                                 } else if (entry->object.vm_object == NULL &&
2351                                            !map->system_map) {
2352                                         vm_map_entry_allocate_object(entry);
2353                                 }
2354                         }
2355
2356                         entry->wired_count++;
2357                         entry = entry->next;
2358                 }
2359
2360                 /*
2361                  * Pass 2.
2362                  */
2363
2364                 /*
2365                  * HACK HACK HACK HACK
2366                  *
2367                  * vm_fault_wire() temporarily unlocks the map to avoid
2368                  * deadlocks.  The in-transition flag from vm_map_clip_range
2369                  * call should protect us from changes while the map is
2370                  * unlocked.  T
2371                  *
2372                  * NOTE: Previously this comment stated that clipping might
2373                  *       still occur while the entry is unlocked, but from
2374                  *       what I can tell it actually cannot.
2375                  *
2376                  *       It is unclear whether the CLIP_CHECK_*() calls
2377                  *       are still needed but we keep them in anyway.
2378                  *
2379                  * HACK HACK HACK HACK
2380                  */
2381
2382                 entry = start_entry;
2383                 while (entry != &map->header && entry->start < end) {
2384                         /*
2385                          * If vm_fault_wire fails for any page we need to undo
2386                          * what has been done.  We decrement the wiring count
2387                          * for those pages which have not yet been wired (now)
2388                          * and unwire those that have (later).
2389                          */
2390                         vm_offset_t save_start = entry->start;
2391                         vm_offset_t save_end = entry->end;
2392
2393                         if (entry->wired_count == 1)
2394                                 rv = vm_fault_wire(map, entry, FALSE);
2395                         if (rv) {
2396                                 CLIP_CHECK_BACK(entry, save_start);
2397                                 for (;;) {
2398                                         KASSERT(entry->wired_count == 1, ("wired_count changed unexpectedly"));
2399                                         entry->wired_count = 0;
2400                                         if (entry->end == save_end)
2401                                                 break;
2402                                         entry = entry->next;
2403                                         KASSERT(entry != &map->header, ("bad entry clip during backout"));
2404                                 }
2405                                 end = save_start;
2406                                 break;
2407                         }
2408                         CLIP_CHECK_FWD(entry, save_end);
2409                         entry = entry->next;
2410                 }
2411
2412                 /*
2413                  * If a failure occured undo everything by falling through
2414                  * to the unwiring code.  'end' has already been adjusted
2415                  * appropriately.
2416                  */
2417                 if (rv)
2418                         kmflags |= KM_PAGEABLE;
2419
2420                 /*
2421                  * start_entry is still IN_TRANSITION but may have been 
2422                  * clipped since vm_fault_wire() unlocks and relocks the
2423                  * map.  No matter how clipped it has gotten there should
2424                  * be a fragment that is on our start boundary.
2425                  */
2426                 CLIP_CHECK_BACK(start_entry, start);
2427         }
2428
2429         if (kmflags & KM_PAGEABLE) {
2430                 /*
2431                  * This is the unwiring case.  We must first ensure that the
2432                  * range to be unwired is really wired down.  We know there
2433                  * are no holes.
2434                  */
2435                 entry = start_entry;
2436                 while ((entry != &map->header) && (entry->start < end)) {
2437                         if (entry->wired_count == 0) {
2438                                 rv = KERN_INVALID_ARGUMENT;
2439                                 goto done;
2440                         }
2441                         entry = entry->next;
2442                 }
2443
2444                 /*
2445                  * Now decrement the wiring count for each region. If a region
2446                  * becomes completely unwired, unwire its physical pages and
2447                  * mappings.
2448                  */
2449                 entry = start_entry;
2450                 while ((entry != &map->header) && (entry->start < end)) {
2451                         entry->wired_count--;
2452                         if (entry->wired_count == 0)
2453                                 vm_fault_unwire(map, entry);
2454                         entry = entry->next;
2455                 }
2456         }
2457 done:
2458         vm_map_unclip_range(map, start_entry, start, real_end,
2459                             &count, MAP_CLIP_NO_HOLES);
2460         map->timestamp++;
2461         vm_map_unlock(map);
2462 failure:
2463         if (kmflags & KM_KRESERVE)
2464                 vm_map_entry_krelease(count);
2465         else
2466                 vm_map_entry_release(count);
2467         return (rv);
2468 }
2469
2470 /*
2471  * Mark a newly allocated address range as wired but do not fault in
2472  * the pages.  The caller is expected to load the pages into the object.
2473  *
2474  * The map must be locked on entry and will remain locked on return.
2475  * No other requirements.
2476  */
2477 void
2478 vm_map_set_wired_quick(vm_map_t map, vm_offset_t addr, vm_size_t size,
2479                        int *countp)
2480 {
2481         vm_map_entry_t scan;
2482         vm_map_entry_t entry;
2483
2484         entry = vm_map_clip_range(map, addr, addr + size,
2485                                   countp, MAP_CLIP_NO_HOLES);
2486         for (scan = entry;
2487              scan != &map->header && scan->start < addr + size;
2488              scan = scan->next) {
2489             KKASSERT(entry->wired_count == 0);
2490             entry->wired_count = 1;                                              
2491         }
2492         vm_map_unclip_range(map, entry, addr, addr + size,
2493                             countp, MAP_CLIP_NO_HOLES);
2494 }
2495
2496 /*
2497  * Push any dirty cached pages in the address range to their pager.
2498  * If syncio is TRUE, dirty pages are written synchronously.
2499  * If invalidate is TRUE, any cached pages are freed as well.
2500  *
2501  * This routine is called by sys_msync()
2502  *
2503  * Returns an error if any part of the specified range is not mapped.
2504  *
2505  * No requirements.
2506  */
2507 int
2508 vm_map_clean(vm_map_t map, vm_offset_t start, vm_offset_t end,
2509              boolean_t syncio, boolean_t invalidate)
2510 {
2511         vm_map_entry_t current;
2512         vm_map_entry_t entry;
2513         vm_size_t size;
2514         vm_object_t object;
2515         vm_object_t tobj;
2516         vm_ooffset_t offset;
2517
2518         vm_map_lock_read(map);
2519         VM_MAP_RANGE_CHECK(map, start, end);
2520         if (!vm_map_lookup_entry(map, start, &entry)) {
2521                 vm_map_unlock_read(map);
2522                 return (KERN_INVALID_ADDRESS);
2523         }
2524         lwkt_gettoken(&map->token);
2525
2526         /*
2527          * Make a first pass to check for holes.
2528          */
2529         for (current = entry; current->start < end; current = current->next) {
2530                 if (current->maptype == VM_MAPTYPE_SUBMAP) {
2531                         lwkt_reltoken(&map->token);
2532                         vm_map_unlock_read(map);
2533                         return (KERN_INVALID_ARGUMENT);
2534                 }
2535                 if (end > current->end &&
2536                     (current->next == &map->header ||
2537                         current->end != current->next->start)) {
2538                         lwkt_reltoken(&map->token);
2539                         vm_map_unlock_read(map);
2540                         return (KERN_INVALID_ADDRESS);
2541                 }
2542         }
2543
2544         if (invalidate)
2545                 pmap_remove(vm_map_pmap(map), start, end);
2546
2547         /*
2548          * Make a second pass, cleaning/uncaching pages from the indicated
2549          * objects as we go.
2550          */
2551         for (current = entry; current->start < end; current = current->next) {
2552                 offset = current->offset + (start - current->start);
2553                 size = (end <= current->end ? end : current->end) - start;
2554                 if (current->maptype == VM_MAPTYPE_SUBMAP) {
2555                         vm_map_t smap;
2556                         vm_map_entry_t tentry;
2557                         vm_size_t tsize;
2558
2559                         smap = current->object.sub_map;
2560                         vm_map_lock_read(smap);
2561                         vm_map_lookup_entry(smap, offset, &tentry);
2562                         tsize = tentry->end - offset;
2563                         if (tsize < size)
2564                                 size = tsize;
2565                         object = tentry->object.vm_object;
2566                         offset = tentry->offset + (offset - tentry->start);
2567                         vm_map_unlock_read(smap);
2568                 } else {
2569                         object = current->object.vm_object;
2570                 }
2571
2572                 if (object)
2573                         vm_object_hold(object);
2574
2575                 /*
2576                  * Note that there is absolutely no sense in writing out
2577                  * anonymous objects, so we track down the vnode object
2578                  * to write out.
2579                  * We invalidate (remove) all pages from the address space
2580                  * anyway, for semantic correctness.
2581                  *
2582                  * note: certain anonymous maps, such as MAP_NOSYNC maps,
2583                  * may start out with a NULL object.
2584                  */
2585                 while (object && (tobj = object->backing_object) != NULL) {
2586                         vm_object_hold(tobj);
2587                         if (tobj == object->backing_object) {
2588                                 vm_object_lock_swap();
2589                                 offset += object->backing_object_offset;
2590                                 vm_object_drop(object);
2591                                 object = tobj;
2592                                 if (object->size < OFF_TO_IDX(offset + size))
2593                                         size = IDX_TO_OFF(object->size) -
2594                                                offset;
2595                                 break;
2596                         }
2597                         vm_object_drop(tobj);
2598                 }
2599                 if (object && (object->type == OBJT_VNODE) && 
2600                     (current->protection & VM_PROT_WRITE) &&
2601                     (object->flags & OBJ_NOMSYNC) == 0) {
2602                         /*
2603                          * Flush pages if writing is allowed, invalidate them
2604                          * if invalidation requested.  Pages undergoing I/O
2605                          * will be ignored by vm_object_page_remove().
2606                          *
2607                          * We cannot lock the vnode and then wait for paging
2608                          * to complete without deadlocking against vm_fault.
2609                          * Instead we simply call vm_object_page_remove() and
2610                          * allow it to block internally on a page-by-page 
2611                          * basis when it encounters pages undergoing async 
2612                          * I/O.
2613                          */
2614                         int flags;
2615
2616                         /* no chain wait needed for vnode objects */
2617                         vm_object_reference_locked(object);
2618                         vn_lock(object->handle, LK_EXCLUSIVE | LK_RETRY);
2619                         flags = (syncio || invalidate) ? OBJPC_SYNC : 0;
2620                         flags |= invalidate ? OBJPC_INVAL : 0;
2621
2622                         /*
2623                          * When operating on a virtual page table just
2624                          * flush the whole object.  XXX we probably ought
2625                          * to 
2626                          */
2627                         switch(current->maptype) {
2628                         case VM_MAPTYPE_NORMAL:
2629                                 vm_object_page_clean(object,
2630                                     OFF_TO_IDX(offset),
2631                                     OFF_TO_IDX(offset + size + PAGE_MASK),
2632                                     flags);
2633                                 break;
2634                         case VM_MAPTYPE_VPAGETABLE:
2635                                 vm_object_page_clean(object, 0, 0, flags);
2636                                 break;
2637                         }
2638                         vn_unlock(((struct vnode *)object->handle));
2639                         vm_object_deallocate_locked(object);
2640                 }
2641                 if (object && invalidate &&
2642                    ((object->type == OBJT_VNODE) ||
2643                     (object->type == OBJT_DEVICE))) {
2644                         int clean_only = 
2645                                 (object->type == OBJT_DEVICE) ? FALSE : TRUE;
2646                         /* no chain wait needed for vnode/device objects */
2647                         vm_object_reference_locked(object);
2648                         switch(current->maptype) {
2649                         case VM_MAPTYPE_NORMAL:
2650                                 vm_object_page_remove(object,
2651                                     OFF_TO_IDX(offset),
2652                                     OFF_TO_IDX(offset + size + PAGE_MASK),
2653                                     clean_only);
2654                                 break;
2655                         case VM_MAPTYPE_VPAGETABLE:
2656                                 vm_object_page_remove(object, 0, 0, clean_only);
2657                                 break;
2658                         }
2659                         vm_object_deallocate_locked(object);
2660                 }
2661                 start += size;
2662                 if (object)
2663                         vm_object_drop(object);
2664         }
2665
2666         lwkt_reltoken(&map->token);
2667         vm_map_unlock_read(map);
2668
2669         return (KERN_SUCCESS);
2670 }
2671
2672 /*
2673  * Make the region specified by this entry pageable.
2674  *
2675  * The vm_map must be exclusively locked.
2676  */
2677 static void 
2678 vm_map_entry_unwire(vm_map_t map, vm_map_entry_t entry)
2679 {
2680         entry->eflags &= ~MAP_ENTRY_USER_WIRED;
2681         entry->wired_count = 0;
2682         vm_fault_unwire(map, entry);
2683 }
2684
2685 /*
2686  * Deallocate the given entry from the target map.
2687  *
2688  * The vm_map must be exclusively locked.
2689  */
2690 static void
2691 vm_map_entry_delete(vm_map_t map, vm_map_entry_t entry, int *countp)
2692 {
2693         vm_map_entry_unlink(map, entry);
2694         map->size -= entry->end - entry->start;
2695
2696         switch(entry->maptype) {
2697         case VM_MAPTYPE_NORMAL:
2698         case VM_MAPTYPE_VPAGETABLE:
2699                 vm_object_deallocate(entry->object.vm_object);
2700                 break;
2701         default:
2702                 break;
2703         }
2704
2705         vm_map_entry_dispose(map, entry, countp);
2706 }
2707
2708 /*
2709  * Deallocates the given address range from the target map.
2710  *
2711  * The vm_map must be exclusively locked.
2712  */
2713 int
2714 vm_map_delete(vm_map_t map, vm_offset_t start, vm_offset_t end, int *countp)
2715 {
2716         vm_object_t object;
2717         vm_map_entry_t entry;
2718         vm_map_entry_t first_entry;
2719
2720         ASSERT_VM_MAP_LOCKED(map);
2721         lwkt_gettoken(&map->token);
2722 again:
2723         /*
2724          * Find the start of the region, and clip it.  Set entry to point
2725          * at the first record containing the requested address or, if no
2726          * such record exists, the next record with a greater address.  The
2727          * loop will run from this point until a record beyond the termination
2728          * address is encountered.
2729          *
2730          * map->hint must be adjusted to not point to anything we delete,
2731          * so set it to the entry prior to the one being deleted.
2732          *
2733          * GGG see other GGG comment.
2734          */
2735         if (vm_map_lookup_entry(map, start, &first_entry)) {
2736                 entry = first_entry;
2737                 vm_map_clip_start(map, entry, start, countp);
2738                 map->hint = entry->prev;        /* possible problem XXX */
2739         } else {
2740                 map->hint = first_entry;        /* possible problem XXX */
2741                 entry = first_entry->next;
2742         }
2743
2744         /*
2745          * If a hole opens up prior to the current first_free then
2746          * adjust first_free.  As with map->hint, map->first_free
2747          * cannot be left set to anything we might delete.
2748          */
2749         if (entry == &map->header) {
2750                 map->first_free = &map->header;
2751         } else if (map->first_free->start >= start) {
2752                 map->first_free = entry->prev;
2753         }
2754
2755         /*
2756          * Step through all entries in this region
2757          */
2758         while ((entry != &map->header) && (entry->start < end)) {
2759                 vm_map_entry_t next;
2760                 vm_offset_t s, e;
2761                 vm_pindex_t offidxstart, offidxend, count;
2762
2763                 /*
2764                  * If we hit an in-transition entry we have to sleep and
2765                  * retry.  It's easier (and not really slower) to just retry
2766                  * since this case occurs so rarely and the hint is already
2767                  * pointing at the right place.  We have to reset the
2768                  * start offset so as not to accidently delete an entry
2769                  * another process just created in vacated space.
2770                  */
2771                 if (entry->eflags & MAP_ENTRY_IN_TRANSITION) {
2772                         entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP;
2773                         start = entry->start;
2774                         ++mycpu->gd_cnt.v_intrans_coll;
2775                         ++mycpu->gd_cnt.v_intrans_wait;
2776                         vm_map_transition_wait(map);
2777                         goto again;
2778                 }
2779                 vm_map_clip_end(map, entry, end, countp);
2780
2781                 s = entry->start;
2782                 e = entry->end;
2783                 next = entry->next;
2784
2785                 offidxstart = OFF_TO_IDX(entry->offset);
2786                 count = OFF_TO_IDX(e - s);
2787                 object = entry->object.vm_object;
2788
2789                 /*
2790                  * Unwire before removing addresses from the pmap; otherwise,
2791                  * unwiring will put the entries back in the pmap.
2792                  */
2793                 if (entry->wired_count != 0)
2794                         vm_map_entry_unwire(map, entry);
2795
2796                 offidxend = offidxstart + count;
2797
2798                 if (object == &kernel_object) {
2799                         vm_object_hold(object);
2800                         vm_object_page_remove(object, offidxstart,
2801                                               offidxend, FALSE);
2802                         vm_object_drop(object);
2803                 } else if (object && object->type != OBJT_DEFAULT &&
2804                            object->type != OBJT_SWAP) {
2805                         /*
2806                          * vnode object routines cannot be chain-locked
2807                          */
2808                         vm_object_hold(object);
2809                         pmap_remove(map->pmap, s, e);
2810                         vm_object_drop(object);
2811                 } else if (object) {
2812                         vm_object_hold(object);
2813                         vm_object_chain_acquire(object);
2814                         pmap_remove(map->pmap, s, e);
2815
2816                         if (object != NULL &&
2817                             object->ref_count != 1 &&
2818                             (object->flags & (OBJ_NOSPLIT|OBJ_ONEMAPPING)) ==
2819                              OBJ_ONEMAPPING &&
2820                             (object->type == OBJT_DEFAULT ||
2821                              object->type == OBJT_SWAP)) {
2822                                 vm_object_collapse(object, NULL);
2823                                 vm_object_page_remove(object, offidxstart,
2824                                                       offidxend, FALSE);
2825                                 if (object->type == OBJT_SWAP) {
2826                                         swap_pager_freespace(object,
2827                                                              offidxstart,
2828                                                              count);
2829                                 }
2830                                 if (offidxend >= object->size &&
2831                                     offidxstart < object->size) {
2832                                         object->size = offidxstart;
2833                                 }
2834                         }
2835                         vm_object_chain_release(object);
2836                         vm_object_drop(object);
2837                 }
2838
2839                 /*
2840                  * Delete the entry (which may delete the object) only after
2841                  * removing all pmap entries pointing to its pages.
2842                  * (Otherwise, its page frames may be reallocated, and any
2843                  * modify bits will be set in the wrong object!)
2844                  */
2845                 vm_map_entry_delete(map, entry, countp);
2846                 entry = next;
2847         }
2848         lwkt_reltoken(&map->token);
2849         return (KERN_SUCCESS);
2850 }
2851
2852 /*
2853  * Remove the given address range from the target map.
2854  * This is the exported form of vm_map_delete.
2855  *
2856  * No requirements.
2857  */
2858 int
2859 vm_map_remove(vm_map_t map, vm_offset_t start, vm_offset_t end)
2860 {
2861         int result;
2862         int count;
2863
2864         count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
2865         vm_map_lock(map);
2866         VM_MAP_RANGE_CHECK(map, start, end);
2867         result = vm_map_delete(map, start, end, &count);
2868         vm_map_unlock(map);
2869         vm_map_entry_release(count);
2870
2871         return (result);
2872 }
2873
2874 /*
2875  * Assert that the target map allows the specified privilege on the
2876  * entire address region given.  The entire region must be allocated.
2877  *
2878  * The caller must specify whether the vm_map is already locked or not.
2879  */
2880 boolean_t
2881 vm_map_check_protection(vm_map_t map, vm_offset_t start, vm_offset_t end,
2882                         vm_prot_t protection, boolean_t have_lock)
2883 {
2884         vm_map_entry_t entry;
2885         vm_map_entry_t tmp_entry;
2886         boolean_t result;
2887
2888         if (have_lock == FALSE)
2889                 vm_map_lock_read(map);
2890
2891         if (!vm_map_lookup_entry(map, start, &tmp_entry)) {
2892                 if (have_lock == FALSE)
2893                         vm_map_unlock_read(map);
2894                 return (FALSE);
2895         }
2896         entry = tmp_entry;
2897
2898         result = TRUE;
2899         while (start < end) {
2900                 if (entry == &map->header) {
2901                         result = FALSE;
2902                         break;
2903                 }
2904                 /*
2905                  * No holes allowed!
2906                  */
2907
2908                 if (start < entry->start) {
2909                         result = FALSE;
2910                         break;
2911                 }
2912                 /*
2913                  * Check protection associated with entry.
2914                  */
2915
2916                 if ((entry->protection & protection) != protection) {
2917                         result = FALSE;
2918                         break;
2919                 }
2920                 /* go to next entry */
2921
2922                 start = entry->end;
2923                 entry = entry->next;
2924         }
2925         if (have_lock == FALSE)
2926                 vm_map_unlock_read(map);
2927         return (result);
2928 }
2929
2930 /*
2931  * If appropriate this function shadows the original object with a new object
2932  * and moves the VM pages from the original object to the new object.
2933  * The original object will also be collapsed, if possible.
2934  *
2935  * We can only do this for normal memory objects with a single mapping, and
2936  * it only makes sense to do it if there are 2 or more refs on the original
2937  * object.  i.e. typically a memory object that has been extended into
2938  * multiple vm_map_entry's with non-overlapping ranges.
2939  *
2940  * This makes it easier to remove unused pages and keeps object inheritance
2941  * from being a negative impact on memory usage.
2942  *
2943  * On return the (possibly new) entry->object.vm_object will have an
2944  * additional ref on it for the caller to dispose of (usually by cloning
2945  * the vm_map_entry).  The additional ref had to be done in this routine
2946  * to avoid racing a collapse.  The object's ONEMAPPING flag will also be
2947  * cleared.
2948  *
2949  * The vm_map must be locked and its token held.
2950  */
2951 static void
2952 vm_map_split(vm_map_entry_t entry)
2953 {
2954 #if 0
2955         /* UNOPTIMIZED */
2956         vm_object_t oobject;
2957
2958         oobject = entry->object.vm_object;
2959         vm_object_hold(oobject);
2960         vm_object_chain_wait(oobject);
2961         vm_object_reference_locked(oobject);
2962         vm_object_clear_flag(oobject, OBJ_ONEMAPPING);
2963         vm_object_drop(oobject);
2964 #else
2965         /* OPTIMIZED */
2966         vm_object_t oobject, nobject, bobject;
2967         vm_offset_t s, e;
2968         vm_page_t m;
2969         vm_pindex_t offidxstart, offidxend, idx;
2970         vm_size_t size;
2971         vm_ooffset_t offset;
2972
2973         /*
2974          * Setup.  Chain lock the original object throughout the entire
2975          * routine to prevent new page faults from occuring.
2976          *
2977          * XXX can madvise WILLNEED interfere with us too?
2978          */
2979         oobject = entry->object.vm_object;
2980         vm_object_hold(oobject);
2981         vm_object_chain_acquire(oobject);
2982
2983         /*
2984          * Original object cannot be split?
2985          */
2986         if (oobject->handle == NULL || (oobject->type != OBJT_DEFAULT &&
2987                                         oobject->type != OBJT_SWAP)) {
2988                 vm_object_chain_release(oobject);
2989                 vm_object_reference_locked(oobject);
2990                 vm_object_clear_flag(oobject, OBJ_ONEMAPPING);
2991                 vm_object_drop(oobject);
2992                 return;
2993         }
2994
2995         /*
2996          * Collapse original object with its backing store as an
2997          * optimization to reduce chain lengths when possible.
2998          *
2999          * If ref_count <= 1 there aren't other non-overlapping vm_map_entry's
3000          * for oobject, so there's no point collapsing it.
3001          *
3002          * Then re-check whether the object can be split.
3003          */
3004         vm_object_collapse(oobject, NULL);
3005
3006         if (oobject->ref_count <= 1 ||
3007             (oobject->type != OBJT_DEFAULT && oobject->type != OBJT_SWAP) ||
3008             (oobject->flags & (OBJ_NOSPLIT|OBJ_ONEMAPPING)) != OBJ_ONEMAPPING) {
3009                 vm_object_chain_release(oobject);
3010                 vm_object_reference_locked(oobject);
3011                 vm_object_clear_flag(oobject, OBJ_ONEMAPPING);
3012                 vm_object_drop(oobject);
3013                 return;
3014         }
3015
3016         /*
3017          * Acquire the chain lock on the backing object.
3018          *
3019          * Give bobject an additional ref count for when it will be shadowed
3020          * by nobject.
3021          */
3022         if ((bobject = oobject->backing_object) != NULL) {
3023                 vm_object_hold(bobject);
3024                 vm_object_chain_wait(bobject);
3025                 vm_object_reference_locked(bobject);
3026                 vm_object_chain_acquire(bobject);
3027                 KKASSERT(bobject->backing_object == bobject);
3028                 KKASSERT((bobject->flags & OBJ_DEAD) == 0);
3029         }
3030
3031         /*
3032          * Calculate the object page range and allocate the new object.
3033          */
3034         offset = entry->offset;
3035         s = entry->start;
3036         e = entry->end;
3037
3038         offidxstart = OFF_TO_IDX(offset);
3039         offidxend = offidxstart + OFF_TO_IDX(e - s);
3040         size = offidxend - offidxstart;
3041
3042         switch(oobject->type) {
3043         case OBJT_DEFAULT:
3044                 nobject = default_pager_alloc(NULL, IDX_TO_OFF(size),
3045                                               VM_PROT_ALL, 0);
3046                 break;
3047         case OBJT_SWAP:
3048                 nobject = swap_pager_alloc(NULL, IDX_TO_OFF(size),
3049                                            VM_PROT_ALL, 0);
3050                 break;
3051         default:
3052                 /* not reached */
3053                 nobject = NULL;
3054                 KKASSERT(0);
3055         }
3056
3057         if (nobject == NULL) {
3058                 if (bobject) {
3059                         vm_object_chain_release(bobject);
3060                         vm_object_deallocate(bobject);
3061                         vm_object_drop(bobject);
3062                 }
3063                 vm_object_chain_release(oobject);
3064                 vm_object_reference_locked(oobject);
3065                 vm_object_clear_flag(oobject, OBJ_ONEMAPPING);
3066                 vm_object_drop(oobject);
3067                 return;
3068         }
3069
3070         /*
3071          * The new object will replace entry->object.vm_object so it needs
3072          * a second reference (the caller expects an additional ref).
3073          */
3074         vm_object_hold(nobject);
3075         vm_object_reference_locked(nobject);
3076         vm_object_chain_acquire(nobject);
3077
3078         /*
3079          * nobject shadows bobject (oobject already shadows bobject).
3080          */
3081         if (bobject) {
3082                 nobject->backing_object_offset =
3083                     oobject->backing_object_offset + IDX_TO_OFF(offidxstart);
3084                 nobject->backing_object = bobject;
3085                 bobject->shadow_count++;
3086                 bobject->generation++;
3087                 LIST_INSERT_HEAD(&bobject->shadow_head, nobject, shadow_list);
3088                 vm_object_clear_flag(bobject, OBJ_ONEMAPPING); /* XXX? */
3089                 vm_object_chain_release(bobject);
3090                 vm_object_drop(bobject);
3091         }
3092
3093         /*
3094          * Move the VM pages from oobject to nobject
3095          */
3096         for (idx = 0; idx < size; idx++) {
3097                 vm_page_t m;
3098
3099                 m = vm_page_lookup_busy_wait(oobject, offidxstart + idx,
3100                                              TRUE, "vmpg");
3101                 if (m == NULL)
3102                         continue;
3103
3104                 /*
3105                  * We must wait for pending I/O to complete before we can
3106                  * rename the page.
3107                  *
3108                  * We do not have to VM_PROT_NONE the page as mappings should
3109                  * not be changed by this operation.
3110                  *
3111                  * NOTE: The act of renaming a page updates chaingen for both
3112                  *       objects.
3113                  */
3114                 vm_page_rename(m, nobject, idx);
3115                 /* page automatically made dirty by rename and cache handled */
3116                 /* page remains busy */
3117         }
3118
3119         if (oobject->type == OBJT_SWAP) {
3120                 vm_object_pip_add(oobject, 1);
3121                 /*
3122                  * copy oobject pages into nobject and destroy unneeded
3123                  * pages in shadow object.
3124                  */
3125                 swap_pager_copy(oobject, nobject, offidxstart, 0);
3126                 vm_object_pip_wakeup(oobject);
3127         }
3128
3129         /*
3130          * Wakeup the pages we played with.  No spl protection is needed
3131          * for a simple wakeup.
3132          */
3133         for (idx = 0; idx < size; idx++) {
3134                 m = vm_page_lookup(nobject, idx);
3135                 if (m) {
3136                         KKASSERT(m->flags & PG_BUSY);
3137                         vm_page_wakeup(m);
3138                 }
3139         }
3140         entry->object.vm_object = nobject;
3141         entry->offset = 0LL;
3142
3143         /*
3144          * Cleanup
3145          *
3146          * NOTE: There is no need to remove OBJ_ONEMAPPING from oobject, the
3147          *       related pages were moved and are no longer applicable to the
3148          *       original object.
3149          *
3150          * NOTE: Deallocate oobject (due to its entry->object.vm_object being
3151          *       replaced by nobject).
3152          */
3153         vm_object_chain_release(nobject);
3154         vm_object_drop(nobject);
3155         if (bobject) {
3156                 vm_object_chain_release(bobject);
3157                 vm_object_drop(bobject);
3158         }
3159         vm_object_chain_release(oobject);
3160         /*vm_object_clear_flag(oobject, OBJ_ONEMAPPING);*/
3161         vm_object_deallocate_locked(oobject);
3162         vm_object_drop(oobject);
3163 #endif
3164 }
3165
3166 /*
3167  * Copies the contents of the source entry to the destination
3168  * entry.  The entries *must* be aligned properly.
3169  *
3170  * The vm_maps must be exclusively locked.
3171  * The vm_map's token must be held.
3172  *
3173  * Because the maps are locked no faults can be in progress during the
3174  * operation.
3175  */
3176 static void
3177 vm_map_copy_entry(vm_map_t src_map, vm_map_t dst_map,
3178                   vm_map_entry_t src_entry, vm_map_entry_t dst_entry)
3179 {
3180         vm_object_t src_object;
3181
3182         if (dst_entry->maptype == VM_MAPTYPE_SUBMAP)
3183                 return;
3184         if (src_entry->maptype == VM_MAPTYPE_SUBMAP)
3185                 return;
3186
3187         if (src_entry->wired_count == 0) {
3188                 /*
3189                  * If the source entry is marked needs_copy, it is already
3190                  * write-protected.
3191                  */
3192                 if ((src_entry->eflags & MAP_ENTRY_NEEDS_COPY) == 0) {
3193                         pmap_protect(src_map->pmap,
3194                             src_entry->start,
3195                             src_entry->end,
3196                             src_entry->protection & ~VM_PROT_WRITE);
3197                 }
3198
3199                 /*
3200                  * Make a copy of the object.
3201                  *
3202                  * The object must be locked prior to checking the object type
3203                  * and for the call to vm_object_collapse() and vm_map_split().
3204                  * We cannot use *_hold() here because the split code will
3205                  * probably try to destroy the object.  The lock is a pool
3206                  * token and doesn't care.
3207                  *
3208                  * We must bump src_map->timestamp when setting
3209                  * MAP_ENTRY_NEEDS_COPY to force any concurrent fault
3210                  * to retry, otherwise the concurrent fault might improperly
3211                  * install a RW pte when its supposed to be a RO(COW) pte.
3212                  * This race can occur because a vnode-backed fault may have
3213                  * to temporarily release the map lock.
3214                  */
3215                 if (src_entry->object.vm_object != NULL) {
3216                         vm_map_split(src_entry);
3217                         src_object = src_entry->object.vm_object;
3218                         dst_entry->object.vm_object = src_object;
3219                         src_entry->eflags |= (MAP_ENTRY_COW |
3220                                               MAP_ENTRY_NEEDS_COPY);
3221                         dst_entry->eflags |= (MAP_ENTRY_COW |
3222                                               MAP_ENTRY_NEEDS_COPY);
3223                         dst_entry->offset = src_entry->offset;
3224                         ++src_map->timestamp;
3225                 } else {
3226                         dst_entry->object.vm_object = NULL;
3227                         dst_entry->offset = 0;
3228                 }
3229
3230                 pmap_copy(dst_map->pmap, src_map->pmap, dst_entry->start,
3231                     dst_entry->end - dst_entry->start, src_entry->start);
3232         } else {
3233                 /*
3234                  * Of course, wired down pages can't be set copy-on-write.
3235                  * Cause wired pages to be copied into the new map by
3236                  * simulating faults (the new pages are pageable)
3237                  */
3238                 vm_fault_copy_entry(dst_map, src_map, dst_entry, src_entry);
3239         }
3240 }
3241
3242 /*
3243  * vmspace_fork:
3244  * Create a new process vmspace structure and vm_map
3245  * based on those of an existing process.  The new map
3246  * is based on the old map, according to the inheritance
3247  * values on the regions in that map.
3248  *
3249  * The source map must not be locked.
3250  * No requirements.
3251  */
3252 struct vmspace *
3253 vmspace_fork(struct vmspace *vm1)
3254 {
3255         struct vmspace *vm2;
3256         vm_map_t old_map = &vm1->vm_map;
3257         vm_map_t new_map;
3258         vm_map_entry_t old_entry;
3259         vm_map_entry_t new_entry;
3260         vm_object_t object;
3261         int count;
3262
3263         lwkt_gettoken(&vm1->vm_map.token);
3264         vm_map_lock(old_map);
3265
3266         /*
3267          * XXX Note: upcalls are not copied.
3268          */
3269         vm2 = vmspace_alloc(old_map->min_offset, old_map->max_offset);
3270         lwkt_gettoken(&vm2->vm_map.token);
3271         bcopy(&vm1->vm_startcopy, &vm2->vm_startcopy,
3272             (caddr_t)&vm1->vm_endcopy - (caddr_t)&vm1->vm_startcopy);
3273         new_map = &vm2->vm_map; /* XXX */
3274         new_map->timestamp = 1;
3275
3276         vm_map_lock(new_map);
3277
3278         count = 0;
3279         old_entry = old_map->header.next;
3280         while (old_entry != &old_map->header) {
3281                 ++count;
3282                 old_entry = old_entry->next;
3283         }
3284
3285         count = vm_map_entry_reserve(count + MAP_RESERVE_COUNT);
3286
3287         old_entry = old_map->header.next;
3288         while (old_entry != &old_map->header) {
3289                 if (old_entry->maptype == VM_MAPTYPE_SUBMAP)
3290                         panic("vm_map_fork: encountered a submap");
3291
3292                 switch (old_entry->inheritance) {
3293                 case VM_INHERIT_NONE:
3294                         break;
3295                 case VM_INHERIT_SHARE:
3296                         /*
3297                          * Clone the entry, creating the shared object if
3298                          * necessary.
3299                          */
3300                         if (old_entry->object.vm_object == NULL)
3301                                 vm_map_entry_allocate_object(old_entry);
3302
3303                         if (old_entry->eflags & MAP_ENTRY_NEEDS_COPY) {
3304                                 /*
3305                                  * Shadow a map_entry which needs a copy,
3306                                  * replacing its object with a new object
3307                                  * that points to the old one.  Ask the
3308                                  * shadow code to automatically add an
3309                                  * additional ref.  We can't do it afterwords
3310                                  * because we might race a collapse.  The call
3311                                  * to vm_map_entry_shadow() will also clear
3312                                  * OBJ_ONEMAPPING.
3313                                  */
3314                                 vm_map_entry_shadow(old_entry, 1);
3315                         } else {
3316                                 /*
3317                                  * We will make a shared copy of the object,
3318                                  * and must clear OBJ_ONEMAPPING.
3319                                  *
3320                                  * XXX assert that object.vm_object != NULL
3321                                  *     since we allocate it above.
3322                                  */
3323                                 if (old_entry->object.vm_object) {
3324                                         object = old_entry->object.vm_object;
3325                                         vm_object_hold(object);
3326                                         vm_object_chain_wait(object);
3327                                         vm_object_reference_locked(object);
3328                                         vm_object_clear_flag(object,
3329                                                              OBJ_ONEMAPPING);
3330                                         vm_object_drop(object);
3331                                 }
3332                         }
3333
3334                         /*
3335                          * Clone the entry.  We've already bumped the ref on
3336                          * any vm_object.
3337                          */
3338                         new_entry = vm_map_entry_create(new_map, &count);
3339                         *new_entry = *old_entry;
3340                         new_entry->eflags &= ~MAP_ENTRY_USER_WIRED;
3341                         new_entry->wired_count = 0;
3342
3343                         /*
3344                          * Insert the entry into the new map -- we know we're
3345                          * inserting at the end of the new map.
3346                          */
3347
3348                         vm_map_entry_link(new_map, new_map->header.prev,
3349                                           new_entry);
3350
3351                         /*
3352                          * Update the physical map
3353                          */
3354                         pmap_copy(new_map->pmap, old_map->pmap,
3355                             new_entry->start,
3356                             (old_entry->end - old_entry->start),
3357                             old_entry->start);
3358                         break;
3359                 case VM_INHERIT_COPY:
3360                         /*
3361                          * Clone the entry and link into the map.
3362                          */
3363                         new_entry = vm_map_entry_create(new_map, &count);
3364                         *new_entry = *old_entry;
3365                         new_entry->eflags &= ~MAP_ENTRY_USER_WIRED;
3366                         new_entry->wired_count = 0;
3367                         new_entry->object.vm_object = NULL;
3368                         vm_map_entry_link(new_map, new_map->header.prev,
3369                                           new_entry);
3370                         vm_map_copy_entry(old_map, new_map, old_entry,
3371                                           new_entry);
3372                         break;
3373                 }
3374                 old_entry = old_entry->next;
3375         }
3376
3377         new_map->size = old_map->size;
3378         vm_map_unlock(old_map);
3379         vm_map_unlock(new_map);
3380         vm_map_entry_release(count);
3381
3382         lwkt_reltoken(&vm2->vm_map.token);
3383         lwkt_reltoken(&vm1->vm_map.token);
3384
3385         return (vm2);
3386 }
3387
3388 /*
3389  * Create an auto-grow stack entry
3390  *
3391  * No requirements.
3392  */
3393 int
3394 vm_map_stack (vm_map_t map, vm_offset_t addrbos, vm_size_t max_ssize,
3395               int flags, vm_prot_t prot, vm_prot_t max, int cow)
3396 {
3397         vm_map_entry_t  prev_entry;
3398         vm_map_entry_t  new_stack_entry;
3399         vm_size_t       init_ssize;
3400         int             rv;
3401         int             count;
3402         vm_offset_t     tmpaddr;
3403
3404         cow |= MAP_IS_STACK;
3405
3406         if (max_ssize < sgrowsiz)
3407                 init_ssize = max_ssize;
3408         else
3409                 init_ssize = sgrowsiz;
3410
3411         count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
3412         vm_map_lock(map);
3413
3414         /*
3415          * Find space for the mapping
3416          */
3417         if ((flags & (MAP_FIXED | MAP_TRYFIXED)) == 0) {
3418                 if (vm_map_findspace(map, addrbos, max_ssize, 1,
3419                                      flags, &tmpaddr)) {
3420                         vm_map_unlock(map);
3421                         vm_map_entry_release(count);
3422                         return (KERN_NO_SPACE);
3423                 }
3424                 addrbos = tmpaddr;
3425         }
3426
3427         /* If addr is already mapped, no go */
3428         if (vm_map_lookup_entry(map, addrbos, &prev_entry)) {
3429                 vm_map_unlock(map);
3430                 vm_map_entry_release(count);
3431                 return (KERN_NO_SPACE);
3432         }
3433
3434 #if 0
3435         /* XXX already handled by kern_mmap() */
3436         /* If we would blow our VMEM resource limit, no go */
3437         if (map->size + init_ssize >
3438             curproc->p_rlimit[RLIMIT_VMEM].rlim_cur) {
3439                 vm_map_unlock(map);
3440                 vm_map_entry_release(count);
3441                 return (KERN_NO_SPACE);
3442         }
3443 #endif
3444
3445         /*
3446          * If we can't accomodate max_ssize in the current mapping,
3447          * no go.  However, we need to be aware that subsequent user
3448          * mappings might map into the space we have reserved for
3449          * stack, and currently this space is not protected.  
3450          * 
3451          * Hopefully we will at least detect this condition 
3452          * when we try to grow the stack.
3453          */
3454         if ((prev_entry->next != &map->header) &&
3455             (prev_entry->next->start < addrbos + max_ssize)) {
3456                 vm_map_unlock(map);
3457                 vm_map_entry_release(count);
3458                 return (KERN_NO_SPACE);
3459         }
3460
3461         /*
3462          * We initially map a stack of only init_ssize.  We will
3463          * grow as needed later.  Since this is to be a grow 
3464          * down stack, we map at the top of the range.
3465          *
3466          * Note: we would normally expect prot and max to be
3467          * VM_PROT_ALL, and cow to be 0.  Possibly we should
3468          * eliminate these as input parameters, and just
3469          * pass these values here in the insert call.
3470          */
3471         rv = vm_map_insert(map, &count,
3472                            NULL, 0, addrbos + max_ssize - init_ssize,
3473                            addrbos + max_ssize,
3474                            VM_MAPTYPE_NORMAL,
3475                            prot, max,
3476                            cow);
3477
3478         /* Now set the avail_ssize amount */
3479         if (rv == KERN_SUCCESS) {
3480                 if (prev_entry != &map->header)
3481                         vm_map_clip_end(map, prev_entry, addrbos + max_ssize - init_ssize, &count);
3482                 new_stack_entry = prev_entry->next;
3483                 if (new_stack_entry->end   != addrbos + max_ssize ||
3484                     new_stack_entry->start != addrbos + max_ssize - init_ssize)
3485                         panic ("Bad entry start/end for new stack entry");
3486                 else 
3487                         new_stack_entry->aux.avail_ssize = max_ssize - init_ssize;
3488         }
3489
3490         vm_map_unlock(map);
3491         vm_map_entry_release(count);
3492         return (rv);
3493 }
3494
3495 /*
3496  * Attempts to grow a vm stack entry.  Returns KERN_SUCCESS if the
3497  * desired address is already mapped, or if we successfully grow
3498  * the stack.  Also returns KERN_SUCCESS if addr is outside the
3499  * stack range (this is strange, but preserves compatibility with
3500  * the grow function in vm_machdep.c).
3501  *
3502  * No requirements.
3503  */
3504 int
3505 vm_map_growstack (struct proc *p, vm_offset_t addr)
3506 {
3507         vm_map_entry_t prev_entry;
3508         vm_map_entry_t stack_entry;
3509         vm_map_entry_t new_stack_entry;
3510         struct vmspace *vm = p->p_vmspace;
3511         vm_map_t map = &vm->vm_map;
3512         vm_offset_t    end;
3513         int grow_amount;
3514         int rv = KERN_SUCCESS;
3515         int is_procstack;
3516         int use_read_lock = 1;
3517         int count;
3518
3519         count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
3520 Retry:
3521         if (use_read_lock)
3522                 vm_map_lock_read(map);
3523         else
3524                 vm_map_lock(map);
3525
3526         /* If addr is already in the entry range, no need to grow.*/
3527         if (vm_map_lookup_entry(map, addr, &prev_entry))
3528                 goto done;
3529
3530         if ((stack_entry = prev_entry->next) == &map->header)
3531                 goto done;
3532         if (prev_entry == &map->header) 
3533                 end = stack_entry->start - stack_entry->aux.avail_ssize;
3534         else
3535                 end = prev_entry->end;
3536
3537         /*
3538          * This next test mimics the old grow function in vm_machdep.c.
3539          * It really doesn't quite make sense, but we do it anyway
3540          * for compatibility.
3541          *
3542          * If not growable stack, return success.  This signals the
3543          * caller to proceed as he would normally with normal vm.
3544          */
3545         if (stack_entry->aux.avail_ssize < 1 ||
3546             addr >= stack_entry->start ||
3547             addr <  stack_entry->start - stack_entry->aux.avail_ssize) {
3548                 goto done;
3549         } 
3550         
3551         /* Find the minimum grow amount */
3552         grow_amount = roundup (stack_entry->start - addr, PAGE_SIZE);
3553         if (grow_amount > stack_entry->aux.avail_ssize) {
3554                 rv = KERN_NO_SPACE;
3555                 goto done;
3556         }
3557
3558         /*
3559          * If there is no longer enough space between the entries
3560          * nogo, and adjust the available space.  Note: this 
3561          * should only happen if the user has mapped into the
3562          * stack area after the stack was created, and is
3563          * probably an error.
3564          *
3565          * This also effectively destroys any guard page the user
3566          * might have intended by limiting the stack size.
3567          */
3568         if (grow_amount > stack_entry->start - end) {
3569                 if (use_read_lock && vm_map_lock_upgrade(map)) {
3570                         /* lost lock */
3571                         use_read_lock = 0;
3572                         goto Retry;
3573                 }
3574                 use_read_lock = 0;
3575                 stack_entry->aux.avail_ssize = stack_entry->start - end;
3576                 rv = KERN_NO_SPACE;
3577                 goto done;
3578         }
3579
3580         is_procstack = addr >= (vm_offset_t)vm->vm_maxsaddr;
3581
3582         /* If this is the main process stack, see if we're over the 
3583          * stack limit.
3584          */
3585         if (is_procstack && (ctob(vm->vm_ssize) + grow_amount >
3586                              p->p_rlimit[RLIMIT_STACK].rlim_cur)) {
3587                 rv = KERN_NO_SPACE;
3588                 goto done;
3589         }
3590
3591         /* Round up the grow amount modulo SGROWSIZ */
3592         grow_amount = roundup (grow_amount, sgrowsiz);
3593         if (grow_amount > stack_entry->aux.avail_ssize) {
3594                 grow_amount = stack_entry->aux.avail_ssize;
3595         }
3596         if (is_procstack && (ctob(vm->vm_ssize) + grow_amount >
3597                              p->p_rlimit[RLIMIT_STACK].rlim_cur)) {
3598                 grow_amount = p->p_rlimit[RLIMIT_STACK].rlim_cur -
3599                               ctob(vm->vm_ssize);
3600         }
3601
3602         /* If we would blow our VMEM resource limit, no go */
3603         if (map->size + grow_amount > p->p_rlimit[RLIMIT_VMEM].rlim_cur) {
3604                 rv = KERN_NO_SPACE;
3605                 goto done;
3606         }
3607
3608         if (use_read_lock && vm_map_lock_upgrade(map)) {
3609                 /* lost lock */
3610                 use_read_lock = 0;
3611                 goto Retry;
3612         }
3613         use_read_lock = 0;
3614
3615         /* Get the preliminary new entry start value */
3616         addr = stack_entry->start - grow_amount;
3617
3618         /* If this puts us into the previous entry, cut back our growth
3619          * to the available space.  Also, see the note above.
3620          */
3621         if (addr < end) {
3622                 stack_entry->aux.avail_ssize = stack_entry->start - end;
3623                 addr = end;
3624         }
3625
3626         rv = vm_map_insert(map, &count,
3627                            NULL, 0, addr, stack_entry->start,
3628                            VM_MAPTYPE_NORMAL,
3629                            VM_PROT_ALL, VM_PROT_ALL,
3630                            0);
3631
3632         /* Adjust the available stack space by the amount we grew. */
3633         if (rv == KERN_SUCCESS) {
3634                 if (prev_entry != &map->header)
3635                         vm_map_clip_end(map, prev_entry, addr, &count);
3636                 new_stack_entry = prev_entry->next;
3637                 if (new_stack_entry->end   != stack_entry->start  ||
3638                     new_stack_entry->start != addr)
3639                         panic ("Bad stack grow start/end in new stack entry");
3640                 else {
3641                         new_stack_entry->aux.avail_ssize =
3642                                 stack_entry->aux.avail_ssize -
3643                                 (new_stack_entry->end - new_stack_entry->start);
3644                         if (is_procstack)
3645                                 vm->vm_ssize += btoc(new_stack_entry->end -
3646                                                      new_stack_entry->start);
3647                 }
3648
3649                 if (map->flags & MAP_WIREFUTURE)
3650                         vm_map_unwire(map, new_stack_entry->start,
3651                                       new_stack_entry->end, FALSE);
3652         }
3653
3654 done:
3655         if (use_read_lock)
3656                 vm_map_unlock_read(map);
3657         else
3658                 vm_map_unlock(map);
3659         vm_map_entry_release(count);
3660         return (rv);
3661 }
3662
3663 /*
3664  * Unshare the specified VM space for exec.  If other processes are
3665  * mapped to it, then create a new one.  The new vmspace is null.
3666  *
3667  * No requirements.
3668  */
3669 void
3670 vmspace_exec(struct proc *p, struct vmspace *vmcopy) 
3671 {
3672         struct vmspace *oldvmspace = p->p_vmspace;
3673         struct vmspace *newvmspace;
3674         vm_map_t map = &p->p_vmspace->vm_map;
3675
3676         /*
3677          * If we are execing a resident vmspace we fork it, otherwise
3678          * we create a new vmspace.  Note that exitingcnt and upcalls
3679          * are not copied to the new vmspace.
3680          */
3681         lwkt_gettoken(&oldvmspace->vm_map.token);
3682         if (vmcopy)  {
3683                 newvmspace = vmspace_fork(vmcopy);
3684                 lwkt_gettoken(&newvmspace->vm_map.token);
3685         } else {
3686                 newvmspace = vmspace_alloc(map->min_offset, map->max_offset);
3687                 lwkt_gettoken(&newvmspace->vm_map.token);
3688                 bcopy(&oldvmspace->vm_startcopy, &newvmspace->vm_startcopy,
3689                       (caddr_t)&oldvmspace->vm_endcopy -
3690                        (caddr_t)&oldvmspace->vm_startcopy);
3691         }
3692
3693         /*
3694          * Finish initializing the vmspace before assigning it
3695          * to the process.  The vmspace will become the current vmspace
3696          * if p == curproc.
3697          */
3698         pmap_pinit2(vmspace_pmap(newvmspace));
3699         pmap_replacevm(p, newvmspace, 0);
3700         lwkt_reltoken(&newvmspace->vm_map.token);
3701         lwkt_reltoken(&oldvmspace->vm_map.token);
3702         vmspace_free(oldvmspace);
3703 }
3704
3705 /*
3706  * Unshare the specified VM space for forcing COW.  This
3707  * is called by rfork, for the (RFMEM|RFPROC) == 0 case.
3708  */
3709 void
3710 vmspace_unshare(struct proc *p) 
3711 {
3712         struct vmspace *oldvmspace = p->p_vmspace;
3713         struct vmspace *newvmspace;
3714
3715         lwkt_gettoken(&oldvmspace->vm_map.token);
3716         if (oldvmspace->vm_sysref.refcnt == 1) {
3717                 lwkt_reltoken(&oldvmspace->vm_map.token);
3718                 return;
3719         }
3720         newvmspace = vmspace_fork(oldvmspace);
3721         lwkt_gettoken(&newvmspace->vm_map.token);
3722         pmap_pinit2(vmspace_pmap(newvmspace));
3723         pmap_replacevm(p, newvmspace, 0);
3724         lwkt_reltoken(&newvmspace->vm_map.token);
3725         lwkt_reltoken(&oldvmspace->vm_map.token);
3726         vmspace_free(oldvmspace);
3727 }
3728
3729 /*
3730  * vm_map_hint: return the beginning of the best area suitable for
3731  * creating a new mapping with "prot" protection.
3732  *
3733  * No requirements.
3734  */
3735 vm_offset_t
3736 vm_map_hint(struct proc *p, vm_offset_t addr, vm_prot_t prot)
3737 {
3738         struct vmspace *vms = p->p_vmspace;
3739
3740         if (!randomize_mmap) {
3741                 /*
3742                  * Set a reasonable start point for the hint if it was
3743                  * not specified or if it falls within the heap space.
3744                  * Hinted mmap()s do not allocate out of the heap space.
3745                  */
3746                 if (addr == 0 ||
3747                     (addr >= round_page((vm_offset_t)vms->vm_taddr) &&
3748                      addr < round_page((vm_offset_t)vms->vm_daddr + maxdsiz))) {
3749                         addr = round_page((vm_offset_t)vms->vm_daddr + maxdsiz);
3750                 }
3751
3752                 return addr;
3753         }
3754
3755         if (addr != 0 && addr >= (vm_offset_t)vms->vm_daddr)
3756                 return addr;
3757
3758 #ifdef notyet
3759 #ifdef __i386__
3760         /*
3761          * If executable skip first two pages, otherwise start
3762          * after data + heap region.
3763          */
3764         if ((prot & VM_PROT_EXECUTE) &&
3765             ((vm_offset_t)vms->vm_daddr >= I386_MAX_EXE_ADDR)) {
3766                 addr = (PAGE_SIZE * 2) +
3767                     (karc4random() & (I386_MAX_EXE_ADDR / 2 - 1));
3768                 return (round_page(addr));
3769         }
3770 #endif /* __i386__ */
3771 #endif /* notyet */
3772
3773         addr = (vm_offset_t)vms->vm_daddr + MAXDSIZ;
3774         addr += karc4random() & (MIN((256 * 1024 * 1024), MAXDSIZ) - 1);
3775
3776         return (round_page(addr));
3777 }
3778
3779 /*
3780  * Finds the VM object, offset, and protection for a given virtual address
3781  * in the specified map, assuming a page fault of the type specified.
3782  *
3783  * Leaves the map in question locked for read; return values are guaranteed
3784  * until a vm_map_lookup_done call is performed.  Note that the map argument
3785  * is in/out; the returned map must be used in the call to vm_map_lookup_done.
3786  *
3787  * A handle (out_entry) is returned for use in vm_map_lookup_done, to make
3788  * that fast.
3789  *
3790  * If a lookup is requested with "write protection" specified, the map may
3791  * be changed to perform virtual copying operations, although the data
3792  * referenced will remain the same.
3793  *
3794  * No requirements.
3795  */
3796 int
3797 vm_map_lookup(vm_map_t *var_map,                /* IN/OUT */
3798               vm_offset_t vaddr,
3799               vm_prot_t fault_typea,
3800               vm_map_entry_t *out_entry,        /* OUT */
3801               vm_object_t *object,              /* OUT */
3802               vm_pindex_t *pindex,              /* OUT */
3803               vm_prot_t *out_prot,              /* OUT */
3804               boolean_t *wired)                 /* OUT */
3805 {
3806         vm_map_entry_t entry;
3807         vm_map_t map = *var_map;
3808         vm_prot_t prot;
3809         vm_prot_t fault_type = fault_typea;
3810         int use_read_lock = 1;
3811         int rv = KERN_SUCCESS;
3812
3813 RetryLookup:
3814         if (use_read_lock)
3815                 vm_map_lock_read(map);
3816         else
3817                 vm_map_lock(map);
3818
3819         /*
3820          * If the map has an interesting hint, try it before calling full
3821          * blown lookup routine.
3822          */
3823         entry = map->hint;
3824         cpu_ccfence();
3825         *out_entry = entry;
3826         *object = NULL;
3827
3828         if ((entry == &map->header) ||
3829             (vaddr < entry->start) || (vaddr >= entry->end)) {
3830                 vm_map_entry_t tmp_entry;
3831
3832                 /*
3833                  * Entry was either not a valid hint, or the vaddr was not
3834                  * contained in the entry, so do a full lookup.
3835                  */
3836                 if (!vm_map_lookup_entry(map, vaddr, &tmp_entry)) {
3837                         rv = KERN_INVALID_ADDRESS;
3838                         goto done;
3839                 }
3840
3841                 entry = tmp_entry;
3842                 *out_entry = entry;
3843         }
3844         
3845         /*
3846          * Handle submaps.
3847          */
3848         if (entry->maptype == VM_MAPTYPE_SUBMAP) {
3849                 vm_map_t old_map = map;
3850
3851                 *var_map = map = entry->object.sub_map;
3852                 if (use_read_lock)
3853                         vm_map_unlock_read(old_map);
3854                 else
3855                         vm_map_unlock(old_map);
3856                 use_read_lock = 1;
3857                 goto RetryLookup;
3858         }
3859
3860         /*
3861          * Check whether this task is allowed to have this page.
3862          * Note the special case for MAP_ENTRY_COW
3863          * pages with an override.  This is to implement a forced
3864          * COW for debuggers.
3865          */
3866
3867         if (fault_type & VM_PROT_OVERRIDE_WRITE)
3868                 prot = entry->max_protection;
3869         else
3870                 prot = entry->protection;
3871
3872         fault_type &= (VM_PROT_READ|VM_PROT_WRITE|VM_PROT_EXECUTE);
3873         if ((fault_type & prot) != fault_type) {
3874                 rv = KERN_PROTECTION_FAILURE;
3875                 goto done;
3876         }
3877
3878         if ((entry->eflags & MAP_ENTRY_USER_WIRED) &&
3879             (entry->eflags & MAP_ENTRY_COW) &&
3880             (fault_type & VM_PROT_WRITE) &&
3881             (fault_typea & VM_PROT_OVERRIDE_WRITE) == 0) {
3882                 rv = KERN_PROTECTION_FAILURE;
3883                 goto done;
3884         }
3885
3886         /*
3887          * If this page is not pageable, we have to get it for all possible
3888          * accesses.
3889          */
3890         *wired = (entry->wired_count != 0);
3891         if (*wired)
3892                 prot = fault_type = entry->protection;
3893
3894         /*
3895          * Virtual page tables may need to update the accessed (A) bit
3896          * in a page table entry.  Upgrade the fault to a write fault for
3897          * that case if the map will support it.  If the map does not support
3898          * it the page table entry simply will not be updated.
3899          */
3900         if (entry->maptype == VM_MAPTYPE_VPAGETABLE) {
3901                 if (prot & VM_PROT_WRITE)
3902                         fault_type |= VM_PROT_WRITE;
3903         }
3904
3905         /*
3906          * If the entry was copy-on-write, we either ...
3907          */
3908         if (entry->eflags & MAP_ENTRY_NEEDS_COPY) {
3909                 /*
3910                  * If we want to write the page, we may as well handle that
3911                  * now since we've got the map locked.
3912                  *
3913                  * If we don't need to write the page, we just demote the
3914                  * permissions allowed.
3915                  */
3916
3917                 if (fault_type & VM_PROT_WRITE) {
3918                         /*
3919                          * Make a new object, and place it in the object
3920                          * chain.  Note that no new references have appeared
3921                          * -- one just moved from the map to the new
3922                          * object.
3923                          */
3924
3925                         if (use_read_lock && vm_map_lock_upgrade(map)) {
3926                                 /* lost lock */
3927                                 use_read_lock = 0;
3928                                 goto RetryLookup;
3929                         }
3930                         use_read_lock = 0;
3931
3932                         vm_map_entry_shadow(entry, 0);
3933                 } else {
3934                         /*
3935                          * We're attempting to read a copy-on-write page --
3936                          * don't allow writes.
3937                          */
3938
3939                         prot &= ~VM_PROT_WRITE;
3940                 }
3941         }
3942
3943         /*
3944          * Create an object if necessary.
3945          */
3946         if (entry->object.vm_object == NULL && !map->system_map) {
3947                 if (use_read_lock && vm_map_lock_upgrade(map))  {
3948                         /* lost lock */
3949                         use_read_lock = 0;
3950                         goto RetryLookup;
3951                 }
3952                 use_read_lock = 0;
3953                 vm_map_entry_allocate_object(entry);
3954         }
3955
3956         /*
3957          * Return the object/offset from this entry.  If the entry was
3958          * copy-on-write or empty, it has been fixed up.
3959          */
3960
3961         *pindex = OFF_TO_IDX((vaddr - entry->start) + entry->offset);
3962         *object = entry->object.vm_object;
3963
3964         /*
3965          * Return whether this is the only map sharing this data.  On
3966          * success we return with a read lock held on the map.  On failure
3967          * we return with the map unlocked.
3968          */
3969         *out_prot = prot;
3970 done:
3971         if (rv == KERN_SUCCESS) {
3972                 if (use_read_lock == 0)
3973                         vm_map_lock_downgrade(map);
3974         } else if (use_read_lock) {
3975                 vm_map_unlock_read(map);
3976         } else {
3977                 vm_map_unlock(map);
3978         }
3979         return (rv);
3980 }
3981
3982 /*
3983  * Releases locks acquired by a vm_map_lookup()
3984  * (according to the handle returned by that lookup).
3985  *
3986  * No other requirements.
3987  */
3988 void
3989 vm_map_lookup_done(vm_map_t map, vm_map_entry_t entry, int count)
3990 {
3991         /*
3992          * Unlock the main-level map
3993          */
3994         vm_map_unlock_read(map);
3995         if (count)
3996                 vm_map_entry_release(count);
3997 }
3998
3999 #include "opt_ddb.h"
4000 #ifdef DDB
4001 #include <sys/kernel.h>
4002
4003 #include <ddb/ddb.h>
4004
4005 /*
4006  * Debugging only
4007  */
4008 DB_SHOW_COMMAND(map, vm_map_print)
4009 {
4010         static int nlines;
4011         /* XXX convert args. */
4012         vm_map_t map = (vm_map_t)addr;
4013         boolean_t full = have_addr;
4014
4015         vm_map_entry_t entry;
4016
4017         db_iprintf("Task map %p: pmap=%p, nentries=%d, version=%u\n",
4018             (void *)map,
4019             (void *)map->pmap, map->nentries, map->timestamp);
4020         nlines++;
4021
4022         if (!full && db_indent)
4023                 return;
4024
4025         db_indent += 2;
4026         for (entry = map->header.next; entry != &map->header;
4027             entry = entry->next) {
4028                 db_iprintf("map entry %p: start=%p, end=%p\n",
4029                     (void *)entry, (void *)entry->start, (void *)entry->end);
4030                 nlines++;
4031                 {
4032                         static char *inheritance_name[4] =
4033                         {"share", "copy", "none", "donate_copy"};
4034
4035                         db_iprintf(" prot=%x/%x/%s",
4036                             entry->protection,
4037                             entry->max_protection,
4038                             inheritance_name[(int)(unsigned char)entry->inheritance]);
4039                         if (entry->wired_count != 0)