kernel - Fix long-standing vm_map token panic
[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                         if (object)
1270                                 vm_object_drop(object);
1271                         vm_map_unlock(map);
1272                         vm_map_entry_release(count);
1273                         return (KERN_NO_SPACE);
1274                 }
1275                 start = *addr;
1276         }
1277         result = vm_map_insert(map, &count, object, offset,
1278                                start, start + length,
1279                                maptype,
1280                                prot, max,
1281                                cow);
1282         if (object)
1283                 vm_object_drop(object);
1284         vm_map_unlock(map);
1285         vm_map_entry_release(count);
1286
1287         return (result);
1288 }
1289
1290 /*
1291  * Simplify the given map entry by merging with either neighbor.  This
1292  * routine also has the ability to merge with both neighbors.
1293  *
1294  * This routine guarentees that the passed entry remains valid (though
1295  * possibly extended).  When merging, this routine may delete one or
1296  * both neighbors.  No action is taken on entries which have their
1297  * in-transition flag set.
1298  *
1299  * The map must be exclusively locked.
1300  */
1301 void
1302 vm_map_simplify_entry(vm_map_t map, vm_map_entry_t entry, int *countp)
1303 {
1304         vm_map_entry_t next, prev;
1305         vm_size_t prevsize, esize;
1306
1307         if (entry->eflags & MAP_ENTRY_IN_TRANSITION) {
1308                 ++mycpu->gd_cnt.v_intrans_coll;
1309                 return;
1310         }
1311
1312         if (entry->maptype == VM_MAPTYPE_SUBMAP)
1313                 return;
1314
1315         prev = entry->prev;
1316         if (prev != &map->header) {
1317                 prevsize = prev->end - prev->start;
1318                 if ( (prev->end == entry->start) &&
1319                      (prev->maptype == entry->maptype) &&
1320                      (prev->object.vm_object == entry->object.vm_object) &&
1321                      (!prev->object.vm_object ||
1322                         (prev->offset + prevsize == entry->offset)) &&
1323                      (prev->eflags == entry->eflags) &&
1324                      (prev->protection == entry->protection) &&
1325                      (prev->max_protection == entry->max_protection) &&
1326                      (prev->inheritance == entry->inheritance) &&
1327                      (prev->wired_count == entry->wired_count)) {
1328                         if (map->first_free == prev)
1329                                 map->first_free = entry;
1330                         if (map->hint == prev)
1331                                 map->hint = entry;
1332                         vm_map_entry_unlink(map, prev);
1333                         entry->start = prev->start;
1334                         entry->offset = prev->offset;
1335                         if (prev->object.vm_object)
1336                                 vm_object_deallocate(prev->object.vm_object);
1337                         vm_map_entry_dispose(map, prev, countp);
1338                 }
1339         }
1340
1341         next = entry->next;
1342         if (next != &map->header) {
1343                 esize = entry->end - entry->start;
1344                 if ((entry->end == next->start) &&
1345                     (next->maptype == entry->maptype) &&
1346                     (next->object.vm_object == entry->object.vm_object) &&
1347                      (!entry->object.vm_object ||
1348                         (entry->offset + esize == next->offset)) &&
1349                     (next->eflags == entry->eflags) &&
1350                     (next->protection == entry->protection) &&
1351                     (next->max_protection == entry->max_protection) &&
1352                     (next->inheritance == entry->inheritance) &&
1353                     (next->wired_count == entry->wired_count)) {
1354                         if (map->first_free == next)
1355                                 map->first_free = entry;
1356                         if (map->hint == next)
1357                                 map->hint = entry;
1358                         vm_map_entry_unlink(map, next);
1359                         entry->end = next->end;
1360                         if (next->object.vm_object)
1361                                 vm_object_deallocate(next->object.vm_object);
1362                         vm_map_entry_dispose(map, next, countp);
1363                 }
1364         }
1365 }
1366
1367 /*
1368  * Asserts that the given entry begins at or after the specified address.
1369  * If necessary, it splits the entry into two.
1370  */
1371 #define vm_map_clip_start(map, entry, startaddr, countp)                \
1372 {                                                                       \
1373         if (startaddr > entry->start)                                   \
1374                 _vm_map_clip_start(map, entry, startaddr, countp);      \
1375 }
1376
1377 /*
1378  * This routine is called only when it is known that the entry must be split.
1379  *
1380  * The map must be exclusively locked.
1381  */
1382 static void
1383 _vm_map_clip_start(vm_map_t map, vm_map_entry_t entry, vm_offset_t start,
1384                    int *countp)
1385 {
1386         vm_map_entry_t new_entry;
1387
1388         /*
1389          * Split off the front portion -- note that we must insert the new
1390          * entry BEFORE this one, so that this entry has the specified
1391          * starting address.
1392          */
1393
1394         vm_map_simplify_entry(map, entry, countp);
1395
1396         /*
1397          * If there is no object backing this entry, we might as well create
1398          * one now.  If we defer it, an object can get created after the map
1399          * is clipped, and individual objects will be created for the split-up
1400          * map.  This is a bit of a hack, but is also about the best place to
1401          * put this improvement.
1402          */
1403         if (entry->object.vm_object == NULL && !map->system_map) {
1404                 vm_map_entry_allocate_object(entry);
1405         }
1406
1407         new_entry = vm_map_entry_create(map, countp);
1408         *new_entry = *entry;
1409
1410         new_entry->end = start;
1411         entry->offset += (start - entry->start);
1412         entry->start = start;
1413
1414         vm_map_entry_link(map, entry->prev, new_entry);
1415
1416         switch(entry->maptype) {
1417         case VM_MAPTYPE_NORMAL:
1418         case VM_MAPTYPE_VPAGETABLE:
1419                 if (new_entry->object.vm_object) {
1420                         vm_object_hold(new_entry->object.vm_object);
1421                         vm_object_chain_wait(new_entry->object.vm_object);
1422                         vm_object_reference_locked(new_entry->object.vm_object);
1423                         vm_object_drop(new_entry->object.vm_object);
1424                 }
1425                 break;
1426         default:
1427                 break;
1428         }
1429 }
1430
1431 /*
1432  * Asserts that the given entry ends at or before the specified address.
1433  * If necessary, it splits the entry into two.
1434  *
1435  * The map must be exclusively locked.
1436  */
1437 #define vm_map_clip_end(map, entry, endaddr, countp)            \
1438 {                                                               \
1439         if (endaddr < entry->end)                               \
1440                 _vm_map_clip_end(map, entry, endaddr, countp);  \
1441 }
1442
1443 /*
1444  * This routine is called only when it is known that the entry must be split.
1445  *
1446  * The map must be exclusively locked.
1447  */
1448 static void
1449 _vm_map_clip_end(vm_map_t map, vm_map_entry_t entry, vm_offset_t end,
1450                  int *countp)
1451 {
1452         vm_map_entry_t new_entry;
1453
1454         /*
1455          * If there is no object backing this entry, we might as well create
1456          * one now.  If we defer it, an object can get created after the map
1457          * is clipped, and individual objects will be created for the split-up
1458          * map.  This is a bit of a hack, but is also about the best place to
1459          * put this improvement.
1460          */
1461
1462         if (entry->object.vm_object == NULL && !map->system_map) {
1463                 vm_map_entry_allocate_object(entry);
1464         }
1465
1466         /*
1467          * Create a new entry and insert it AFTER the specified entry
1468          */
1469
1470         new_entry = vm_map_entry_create(map, countp);
1471         *new_entry = *entry;
1472
1473         new_entry->start = entry->end = end;
1474         new_entry->offset += (end - entry->start);
1475
1476         vm_map_entry_link(map, entry, new_entry);
1477
1478         switch(entry->maptype) {
1479         case VM_MAPTYPE_NORMAL:
1480         case VM_MAPTYPE_VPAGETABLE:
1481                 if (new_entry->object.vm_object) {
1482                         vm_object_hold(new_entry->object.vm_object);
1483                         vm_object_chain_wait(new_entry->object.vm_object);
1484                         vm_object_reference_locked(new_entry->object.vm_object);
1485                         vm_object_drop(new_entry->object.vm_object);
1486                 }
1487                 break;
1488         default:
1489                 break;
1490         }
1491 }
1492
1493 /*
1494  * Asserts that the starting and ending region addresses fall within the
1495  * valid range for the map.
1496  */
1497 #define VM_MAP_RANGE_CHECK(map, start, end)     \
1498 {                                               \
1499         if (start < vm_map_min(map))            \
1500                 start = vm_map_min(map);        \
1501         if (end > vm_map_max(map))              \
1502                 end = vm_map_max(map);          \
1503         if (start > end)                        \
1504                 start = end;                    \
1505 }
1506
1507 /*
1508  * Used to block when an in-transition collison occurs.  The map
1509  * is unlocked for the sleep and relocked before the return.
1510  */
1511 void
1512 vm_map_transition_wait(vm_map_t map)
1513 {
1514         tsleep_interlock(map, 0);
1515         vm_map_unlock(map);
1516         tsleep(map, PINTERLOCKED, "vment", 0);
1517         vm_map_lock(map);
1518 }
1519
1520 /*
1521  * When we do blocking operations with the map lock held it is
1522  * possible that a clip might have occured on our in-transit entry,
1523  * requiring an adjustment to the entry in our loop.  These macros
1524  * help the pageable and clip_range code deal with the case.  The
1525  * conditional costs virtually nothing if no clipping has occured.
1526  */
1527
1528 #define CLIP_CHECK_BACK(entry, save_start)              \
1529     do {                                                \
1530             while (entry->start != save_start) {        \
1531                     entry = entry->prev;                \
1532                     KASSERT(entry != &map->header, ("bad entry clip")); \
1533             }                                           \
1534     } while(0)
1535
1536 #define CLIP_CHECK_FWD(entry, save_end)                 \
1537     do {                                                \
1538             while (entry->end != save_end) {            \
1539                     entry = entry->next;                \
1540                     KASSERT(entry != &map->header, ("bad entry clip")); \
1541             }                                           \
1542     } while(0)
1543
1544
1545 /*
1546  * Clip the specified range and return the base entry.  The
1547  * range may cover several entries starting at the returned base
1548  * and the first and last entry in the covering sequence will be
1549  * properly clipped to the requested start and end address.
1550  *
1551  * If no holes are allowed you should pass the MAP_CLIP_NO_HOLES
1552  * flag.
1553  *
1554  * The MAP_ENTRY_IN_TRANSITION flag will be set for the entries
1555  * covered by the requested range.
1556  *
1557  * The map must be exclusively locked on entry and will remain locked
1558  * on return. If no range exists or the range contains holes and you
1559  * specified that no holes were allowed, NULL will be returned.  This
1560  * routine may temporarily unlock the map in order avoid a deadlock when
1561  * sleeping.
1562  */
1563 static
1564 vm_map_entry_t
1565 vm_map_clip_range(vm_map_t map, vm_offset_t start, vm_offset_t end, 
1566                   int *countp, int flags)
1567 {
1568         vm_map_entry_t start_entry;
1569         vm_map_entry_t entry;
1570
1571         /*
1572          * Locate the entry and effect initial clipping.  The in-transition
1573          * case does not occur very often so do not try to optimize it.
1574          */
1575 again:
1576         if (vm_map_lookup_entry(map, start, &start_entry) == FALSE)
1577                 return (NULL);
1578         entry = start_entry;
1579         if (entry->eflags & MAP_ENTRY_IN_TRANSITION) {
1580                 entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP;
1581                 ++mycpu->gd_cnt.v_intrans_coll;
1582                 ++mycpu->gd_cnt.v_intrans_wait;
1583                 vm_map_transition_wait(map);
1584                 /*
1585                  * entry and/or start_entry may have been clipped while
1586                  * we slept, or may have gone away entirely.  We have
1587                  * to restart from the lookup.
1588                  */
1589                 goto again;
1590         }
1591
1592         /*
1593          * Since we hold an exclusive map lock we do not have to restart
1594          * after clipping, even though clipping may block in zalloc.
1595          */
1596         vm_map_clip_start(map, entry, start, countp);
1597         vm_map_clip_end(map, entry, end, countp);
1598         entry->eflags |= MAP_ENTRY_IN_TRANSITION;
1599
1600         /*
1601          * Scan entries covered by the range.  When working on the next
1602          * entry a restart need only re-loop on the current entry which
1603          * we have already locked, since 'next' may have changed.  Also,
1604          * even though entry is safe, it may have been clipped so we
1605          * have to iterate forwards through the clip after sleeping.
1606          */
1607         while (entry->next != &map->header && entry->next->start < end) {
1608                 vm_map_entry_t next = entry->next;
1609
1610                 if (flags & MAP_CLIP_NO_HOLES) {
1611                         if (next->start > entry->end) {
1612                                 vm_map_unclip_range(map, start_entry,
1613                                         start, entry->end, countp, flags);
1614                                 return(NULL);
1615                         }
1616                 }
1617
1618                 if (next->eflags & MAP_ENTRY_IN_TRANSITION) {
1619                         vm_offset_t save_end = entry->end;
1620                         next->eflags |= MAP_ENTRY_NEEDS_WAKEUP;
1621                         ++mycpu->gd_cnt.v_intrans_coll;
1622                         ++mycpu->gd_cnt.v_intrans_wait;
1623                         vm_map_transition_wait(map);
1624
1625                         /*
1626                          * clips might have occured while we blocked.
1627                          */
1628                         CLIP_CHECK_FWD(entry, save_end);
1629                         CLIP_CHECK_BACK(start_entry, start);
1630                         continue;
1631                 }
1632                 /*
1633                  * No restart necessary even though clip_end may block, we
1634                  * are holding the map lock.
1635                  */
1636                 vm_map_clip_end(map, next, end, countp);
1637                 next->eflags |= MAP_ENTRY_IN_TRANSITION;
1638                 entry = next;
1639         }
1640         if (flags & MAP_CLIP_NO_HOLES) {
1641                 if (entry->end != end) {
1642                         vm_map_unclip_range(map, start_entry,
1643                                 start, entry->end, countp, flags);
1644                         return(NULL);
1645                 }
1646         }
1647         return(start_entry);
1648 }
1649
1650 /*
1651  * Undo the effect of vm_map_clip_range().  You should pass the same
1652  * flags and the same range that you passed to vm_map_clip_range().
1653  * This code will clear the in-transition flag on the entries and
1654  * wake up anyone waiting.  This code will also simplify the sequence
1655  * and attempt to merge it with entries before and after the sequence.
1656  *
1657  * The map must be locked on entry and will remain locked on return.
1658  *
1659  * Note that you should also pass the start_entry returned by
1660  * vm_map_clip_range().  However, if you block between the two calls
1661  * with the map unlocked please be aware that the start_entry may
1662  * have been clipped and you may need to scan it backwards to find
1663  * the entry corresponding with the original start address.  You are
1664  * responsible for this, vm_map_unclip_range() expects the correct
1665  * start_entry to be passed to it and will KASSERT otherwise.
1666  */
1667 static
1668 void
1669 vm_map_unclip_range(vm_map_t map, vm_map_entry_t start_entry,
1670                     vm_offset_t start, vm_offset_t end,
1671                     int *countp, int flags)
1672 {
1673         vm_map_entry_t entry;
1674
1675         entry = start_entry;
1676
1677         KASSERT(entry->start == start, ("unclip_range: illegal base entry"));
1678         while (entry != &map->header && entry->start < end) {
1679                 KASSERT(entry->eflags & MAP_ENTRY_IN_TRANSITION,
1680                         ("in-transition flag not set during unclip on: %p",
1681                         entry));
1682                 KASSERT(entry->end <= end,
1683                         ("unclip_range: tail wasn't clipped"));
1684                 entry->eflags &= ~MAP_ENTRY_IN_TRANSITION;
1685                 if (entry->eflags & MAP_ENTRY_NEEDS_WAKEUP) {
1686                         entry->eflags &= ~MAP_ENTRY_NEEDS_WAKEUP;
1687                         wakeup(map);
1688                 }
1689                 entry = entry->next;
1690         }
1691
1692         /*
1693          * Simplification does not block so there is no restart case.
1694          */
1695         entry = start_entry;
1696         while (entry != &map->header && entry->start < end) {
1697                 vm_map_simplify_entry(map, entry, countp);
1698                 entry = entry->next;
1699         }
1700 }
1701
1702 /*
1703  * Mark the given range as handled by a subordinate map.
1704  *
1705  * This range must have been created with vm_map_find(), and no other
1706  * operations may have been performed on this range prior to calling
1707  * vm_map_submap().
1708  *
1709  * Submappings cannot be removed.
1710  *
1711  * No requirements.
1712  */
1713 int
1714 vm_map_submap(vm_map_t map, vm_offset_t start, vm_offset_t end, vm_map_t submap)
1715 {
1716         vm_map_entry_t entry;
1717         int result = KERN_INVALID_ARGUMENT;
1718         int count;
1719
1720         count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
1721         vm_map_lock(map);
1722
1723         VM_MAP_RANGE_CHECK(map, start, end);
1724
1725         if (vm_map_lookup_entry(map, start, &entry)) {
1726                 vm_map_clip_start(map, entry, start, &count);
1727         } else {
1728                 entry = entry->next;
1729         }
1730
1731         vm_map_clip_end(map, entry, end, &count);
1732
1733         if ((entry->start == start) && (entry->end == end) &&
1734             ((entry->eflags & MAP_ENTRY_COW) == 0) &&
1735             (entry->object.vm_object == NULL)) {
1736                 entry->object.sub_map = submap;
1737                 entry->maptype = VM_MAPTYPE_SUBMAP;
1738                 result = KERN_SUCCESS;
1739         }
1740         vm_map_unlock(map);
1741         vm_map_entry_release(count);
1742
1743         return (result);
1744 }
1745
1746 /*
1747  * Sets the protection of the specified address region in the target map. 
1748  * If "set_max" is specified, the maximum protection is to be set;
1749  * otherwise, only the current protection is affected.
1750  *
1751  * The protection is not applicable to submaps, but is applicable to normal
1752  * maps and maps governed by virtual page tables.  For example, when operating
1753  * on a virtual page table our protection basically controls how COW occurs
1754  * on the backing object, whereas the virtual page table abstraction itself
1755  * is an abstraction for userland.
1756  *
1757  * No requirements.
1758  */
1759 int
1760 vm_map_protect(vm_map_t map, vm_offset_t start, vm_offset_t end,
1761                vm_prot_t new_prot, boolean_t set_max)
1762 {
1763         vm_map_entry_t current;
1764         vm_map_entry_t entry;
1765         int count;
1766
1767         count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
1768         vm_map_lock(map);
1769
1770         VM_MAP_RANGE_CHECK(map, start, end);
1771
1772         if (vm_map_lookup_entry(map, start, &entry)) {
1773                 vm_map_clip_start(map, entry, start, &count);
1774         } else {
1775                 entry = entry->next;
1776         }
1777
1778         /*
1779          * Make a first pass to check for protection violations.
1780          */
1781         current = entry;
1782         while ((current != &map->header) && (current->start < end)) {
1783                 if (current->maptype == VM_MAPTYPE_SUBMAP) {
1784                         vm_map_unlock(map);
1785                         vm_map_entry_release(count);
1786                         return (KERN_INVALID_ARGUMENT);
1787                 }
1788                 if ((new_prot & current->max_protection) != new_prot) {
1789                         vm_map_unlock(map);
1790                         vm_map_entry_release(count);
1791                         return (KERN_PROTECTION_FAILURE);
1792                 }
1793                 current = current->next;
1794         }
1795
1796         /*
1797          * Go back and fix up protections. [Note that clipping is not
1798          * necessary the second time.]
1799          */
1800         current = entry;
1801
1802         while ((current != &map->header) && (current->start < end)) {
1803                 vm_prot_t old_prot;
1804
1805                 vm_map_clip_end(map, current, end, &count);
1806
1807                 old_prot = current->protection;
1808                 if (set_max) {
1809                         current->protection =
1810                             (current->max_protection = new_prot) &
1811                             old_prot;
1812                 } else {
1813                         current->protection = new_prot;
1814                 }
1815
1816                 /*
1817                  * Update physical map if necessary. Worry about copy-on-write
1818                  * here -- CHECK THIS XXX
1819                  */
1820
1821                 if (current->protection != old_prot) {
1822 #define MASK(entry)     (((entry)->eflags & MAP_ENTRY_COW) ? ~VM_PROT_WRITE : \
1823                                                         VM_PROT_ALL)
1824
1825                         pmap_protect(map->pmap, current->start,
1826                             current->end,
1827                             current->protection & MASK(current));
1828 #undef  MASK
1829                 }
1830
1831                 vm_map_simplify_entry(map, current, &count);
1832
1833                 current = current->next;
1834         }
1835
1836         vm_map_unlock(map);
1837         vm_map_entry_release(count);
1838         return (KERN_SUCCESS);
1839 }
1840
1841 /*
1842  * This routine traverses a processes map handling the madvise
1843  * system call.  Advisories are classified as either those effecting
1844  * the vm_map_entry structure, or those effecting the underlying
1845  * objects.
1846  *
1847  * The <value> argument is used for extended madvise calls.
1848  *
1849  * No requirements.
1850  */
1851 int
1852 vm_map_madvise(vm_map_t map, vm_offset_t start, vm_offset_t end,
1853                int behav, off_t value)
1854 {
1855         vm_map_entry_t current, entry;
1856         int modify_map = 0;
1857         int error = 0;
1858         int count;
1859
1860         /*
1861          * Some madvise calls directly modify the vm_map_entry, in which case
1862          * we need to use an exclusive lock on the map and we need to perform 
1863          * various clipping operations.  Otherwise we only need a read-lock
1864          * on the map.
1865          */
1866
1867         count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
1868
1869         switch(behav) {
1870         case MADV_NORMAL:
1871         case MADV_SEQUENTIAL:
1872         case MADV_RANDOM:
1873         case MADV_NOSYNC:
1874         case MADV_AUTOSYNC:
1875         case MADV_NOCORE:
1876         case MADV_CORE:
1877         case MADV_SETMAP:
1878         case MADV_INVAL:
1879                 modify_map = 1;
1880                 vm_map_lock(map);
1881                 break;
1882         case MADV_WILLNEED:
1883         case MADV_DONTNEED:
1884         case MADV_FREE:
1885                 vm_map_lock_read(map);
1886                 break;
1887         default:
1888                 vm_map_entry_release(count);
1889                 return (EINVAL);
1890         }
1891
1892         /*
1893          * Locate starting entry and clip if necessary.
1894          */
1895
1896         VM_MAP_RANGE_CHECK(map, start, end);
1897
1898         if (vm_map_lookup_entry(map, start, &entry)) {
1899                 if (modify_map)
1900                         vm_map_clip_start(map, entry, start, &count);
1901         } else {
1902                 entry = entry->next;
1903         }
1904
1905         if (modify_map) {
1906                 /*
1907                  * madvise behaviors that are implemented in the vm_map_entry.
1908                  *
1909                  * We clip the vm_map_entry so that behavioral changes are
1910                  * limited to the specified address range.
1911                  */
1912                 for (current = entry;
1913                      (current != &map->header) && (current->start < end);
1914                      current = current->next
1915                 ) {
1916                         if (current->maptype == VM_MAPTYPE_SUBMAP)
1917                                 continue;
1918
1919                         vm_map_clip_end(map, current, end, &count);
1920
1921                         switch (behav) {
1922                         case MADV_NORMAL:
1923                                 vm_map_entry_set_behavior(current, MAP_ENTRY_BEHAV_NORMAL);
1924                                 break;
1925                         case MADV_SEQUENTIAL:
1926                                 vm_map_entry_set_behavior(current, MAP_ENTRY_BEHAV_SEQUENTIAL);
1927                                 break;
1928                         case MADV_RANDOM:
1929                                 vm_map_entry_set_behavior(current, MAP_ENTRY_BEHAV_RANDOM);
1930                                 break;
1931                         case MADV_NOSYNC:
1932                                 current->eflags |= MAP_ENTRY_NOSYNC;
1933                                 break;
1934                         case MADV_AUTOSYNC:
1935                                 current->eflags &= ~MAP_ENTRY_NOSYNC;
1936                                 break;
1937                         case MADV_NOCORE:
1938                                 current->eflags |= MAP_ENTRY_NOCOREDUMP;
1939                                 break;
1940                         case MADV_CORE:
1941                                 current->eflags &= ~MAP_ENTRY_NOCOREDUMP;
1942                                 break;
1943                         case MADV_INVAL:
1944                                 /*
1945                                  * Invalidate the related pmap entries, used
1946                                  * to flush portions of the real kernel's
1947                                  * pmap when the caller has removed or
1948                                  * modified existing mappings in a virtual
1949                                  * page table.
1950                                  */
1951                                 pmap_remove(map->pmap,
1952                                             current->start, current->end);
1953                                 break;
1954                         case MADV_SETMAP:
1955                                 /*
1956                                  * Set the page directory page for a map
1957                                  * governed by a virtual page table.  Mark
1958                                  * the entry as being governed by a virtual
1959                                  * page table if it is not.
1960                                  *
1961                                  * XXX the page directory page is stored
1962                                  * in the avail_ssize field if the map_entry.
1963                                  *
1964                                  * XXX the map simplification code does not
1965                                  * compare this field so weird things may
1966                                  * happen if you do not apply this function
1967                                  * to the entire mapping governed by the
1968                                  * virtual page table.
1969                                  */
1970                                 if (current->maptype != VM_MAPTYPE_VPAGETABLE) {
1971                                         error = EINVAL;
1972                                         break;
1973                                 }
1974                                 current->aux.master_pde = value;
1975                                 pmap_remove(map->pmap,
1976                                             current->start, current->end);
1977                                 break;
1978                         default:
1979                                 error = EINVAL;
1980                                 break;
1981                         }
1982                         vm_map_simplify_entry(map, current, &count);
1983                 }
1984                 vm_map_unlock(map);
1985         } else {
1986                 vm_pindex_t pindex;
1987                 int count;
1988
1989                 /*
1990                  * madvise behaviors that are implemented in the underlying
1991                  * vm_object.
1992                  *
1993                  * Since we don't clip the vm_map_entry, we have to clip
1994                  * the vm_object pindex and count.
1995                  *
1996                  * NOTE!  We currently do not support these functions on
1997                  * virtual page tables.
1998                  */
1999                 for (current = entry;
2000                      (current != &map->header) && (current->start < end);
2001                      current = current->next
2002                 ) {
2003                         vm_offset_t useStart;
2004
2005                         if (current->maptype != VM_MAPTYPE_NORMAL)
2006                                 continue;
2007
2008                         pindex = OFF_TO_IDX(current->offset);
2009                         count = atop(current->end - current->start);
2010                         useStart = current->start;
2011
2012                         if (current->start < start) {
2013                                 pindex += atop(start - current->start);
2014                                 count -= atop(start - current->start);
2015                                 useStart = start;
2016                         }
2017                         if (current->end > end)
2018                                 count -= atop(current->end - end);
2019
2020                         if (count <= 0)
2021                                 continue;
2022
2023                         vm_object_madvise(current->object.vm_object,
2024                                           pindex, count, behav);
2025
2026                         /*
2027                          * Try to populate the page table.  Mappings governed
2028                          * by virtual page tables cannot be pre-populated
2029                          * without a lot of work so don't try.
2030                          */
2031                         if (behav == MADV_WILLNEED &&
2032                             current->maptype != VM_MAPTYPE_VPAGETABLE) {
2033                                 pmap_object_init_pt(
2034                                     map->pmap, 
2035                                     useStart,
2036                                     current->protection,
2037                                     current->object.vm_object,
2038                                     pindex, 
2039                                     (count << PAGE_SHIFT),
2040                                     MAP_PREFAULT_MADVISE
2041                                 );
2042                         }
2043                 }
2044                 vm_map_unlock_read(map);
2045         }
2046         vm_map_entry_release(count);
2047         return(error);
2048 }       
2049
2050
2051 /*
2052  * Sets the inheritance of the specified address range in the target map.
2053  * Inheritance affects how the map will be shared with child maps at the
2054  * time of vm_map_fork.
2055  */
2056 int
2057 vm_map_inherit(vm_map_t map, vm_offset_t start, vm_offset_t end,
2058                vm_inherit_t new_inheritance)
2059 {
2060         vm_map_entry_t entry;
2061         vm_map_entry_t temp_entry;
2062         int count;
2063
2064         switch (new_inheritance) {
2065         case VM_INHERIT_NONE:
2066         case VM_INHERIT_COPY:
2067         case VM_INHERIT_SHARE:
2068                 break;
2069         default:
2070                 return (KERN_INVALID_ARGUMENT);
2071         }
2072
2073         count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
2074         vm_map_lock(map);
2075
2076         VM_MAP_RANGE_CHECK(map, start, end);
2077
2078         if (vm_map_lookup_entry(map, start, &temp_entry)) {
2079                 entry = temp_entry;
2080                 vm_map_clip_start(map, entry, start, &count);
2081         } else
2082                 entry = temp_entry->next;
2083
2084         while ((entry != &map->header) && (entry->start < end)) {
2085                 vm_map_clip_end(map, entry, end, &count);
2086
2087                 entry->inheritance = new_inheritance;
2088
2089                 vm_map_simplify_entry(map, entry, &count);
2090
2091                 entry = entry->next;
2092         }
2093         vm_map_unlock(map);
2094         vm_map_entry_release(count);
2095         return (KERN_SUCCESS);
2096 }
2097
2098 /*
2099  * Implement the semantics of mlock
2100  */
2101 int
2102 vm_map_unwire(vm_map_t map, vm_offset_t start, vm_offset_t real_end,
2103               boolean_t new_pageable)
2104 {
2105         vm_map_entry_t entry;
2106         vm_map_entry_t start_entry;
2107         vm_offset_t end;
2108         int rv = KERN_SUCCESS;
2109         int count;
2110
2111         count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
2112         vm_map_lock(map);
2113         VM_MAP_RANGE_CHECK(map, start, real_end);
2114         end = real_end;
2115
2116         start_entry = vm_map_clip_range(map, start, end, &count,
2117                                         MAP_CLIP_NO_HOLES);
2118         if (start_entry == NULL) {
2119                 vm_map_unlock(map);
2120                 vm_map_entry_release(count);
2121                 return (KERN_INVALID_ADDRESS);
2122         }
2123
2124         if (new_pageable == 0) {
2125                 entry = start_entry;
2126                 while ((entry != &map->header) && (entry->start < end)) {
2127                         vm_offset_t save_start;
2128                         vm_offset_t save_end;
2129
2130                         /*
2131                          * Already user wired or hard wired (trivial cases)
2132                          */
2133                         if (entry->eflags & MAP_ENTRY_USER_WIRED) {
2134                                 entry = entry->next;
2135                                 continue;
2136                         }
2137                         if (entry->wired_count != 0) {
2138                                 entry->wired_count++;
2139                                 entry->eflags |= MAP_ENTRY_USER_WIRED;
2140                                 entry = entry->next;
2141                                 continue;
2142                         }
2143
2144                         /*
2145                          * A new wiring requires instantiation of appropriate
2146                          * management structures and the faulting in of the
2147                          * page.
2148                          */
2149                         if (entry->maptype != VM_MAPTYPE_SUBMAP) {
2150                                 int copyflag = entry->eflags &
2151                                                MAP_ENTRY_NEEDS_COPY;
2152                                 if (copyflag && ((entry->protection &
2153                                                   VM_PROT_WRITE) != 0)) {
2154                                         vm_map_entry_shadow(entry, 0);
2155                                 } else if (entry->object.vm_object == NULL &&
2156                                            !map->system_map) {
2157                                         vm_map_entry_allocate_object(entry);
2158                                 }
2159                         }
2160                         entry->wired_count++;
2161                         entry->eflags |= MAP_ENTRY_USER_WIRED;
2162
2163                         /*
2164                          * Now fault in the area.  Note that vm_fault_wire()
2165                          * may release the map lock temporarily, it will be
2166                          * relocked on return.  The in-transition
2167                          * flag protects the entries. 
2168                          */
2169                         save_start = entry->start;
2170                         save_end = entry->end;
2171                         rv = vm_fault_wire(map, entry, TRUE);
2172                         if (rv) {
2173                                 CLIP_CHECK_BACK(entry, save_start);
2174                                 for (;;) {
2175                                         KASSERT(entry->wired_count == 1, ("bad wired_count on entry"));
2176                                         entry->eflags &= ~MAP_ENTRY_USER_WIRED;
2177                                         entry->wired_count = 0;
2178                                         if (entry->end == save_end)
2179                                                 break;
2180                                         entry = entry->next;
2181                                         KASSERT(entry != &map->header, ("bad entry clip during backout"));
2182                                 }
2183                                 end = save_start;       /* unwire the rest */
2184                                 break;
2185                         }
2186                         /*
2187                          * note that even though the entry might have been
2188                          * clipped, the USER_WIRED flag we set prevents
2189                          * duplication so we do not have to do a 
2190                          * clip check.
2191                          */
2192                         entry = entry->next;
2193                 }
2194
2195                 /*
2196                  * If we failed fall through to the unwiring section to
2197                  * unwire what we had wired so far.  'end' has already
2198                  * been adjusted.
2199                  */
2200                 if (rv)
2201                         new_pageable = 1;
2202
2203                 /*
2204                  * start_entry might have been clipped if we unlocked the
2205                  * map and blocked.  No matter how clipped it has gotten
2206                  * there should be a fragment that is on our start boundary.
2207                  */
2208                 CLIP_CHECK_BACK(start_entry, start);
2209         }
2210
2211         /*
2212          * Deal with the unwiring case.
2213          */
2214         if (new_pageable) {
2215                 /*
2216                  * This is the unwiring case.  We must first ensure that the
2217                  * range to be unwired is really wired down.  We know there
2218                  * are no holes.
2219                  */
2220                 entry = start_entry;
2221                 while ((entry != &map->header) && (entry->start < end)) {
2222                         if ((entry->eflags & MAP_ENTRY_USER_WIRED) == 0) {
2223                                 rv = KERN_INVALID_ARGUMENT;
2224                                 goto done;
2225                         }
2226                         KASSERT(entry->wired_count != 0, ("wired count was 0 with USER_WIRED set! %p", entry));
2227                         entry = entry->next;
2228                 }
2229
2230                 /*
2231                  * Now decrement the wiring count for each region. If a region
2232                  * becomes completely unwired, unwire its physical pages and
2233                  * mappings.
2234                  */
2235                 /*
2236                  * The map entries are processed in a loop, checking to
2237                  * make sure the entry is wired and asserting it has a wired
2238                  * count. However, another loop was inserted more-or-less in
2239                  * the middle of the unwiring path. This loop picks up the
2240                  * "entry" loop variable from the first loop without first
2241                  * setting it to start_entry. Naturally, the secound loop
2242                  * is never entered and the pages backing the entries are
2243                  * never unwired. This can lead to a leak of wired pages.
2244                  */
2245                 entry = start_entry;
2246                 while ((entry != &map->header) && (entry->start < end)) {
2247                         KASSERT(entry->eflags & MAP_ENTRY_USER_WIRED,
2248                                 ("expected USER_WIRED on entry %p", entry));
2249                         entry->eflags &= ~MAP_ENTRY_USER_WIRED;
2250                         entry->wired_count--;
2251                         if (entry->wired_count == 0)
2252                                 vm_fault_unwire(map, entry);
2253                         entry = entry->next;
2254                 }
2255         }
2256 done:
2257         vm_map_unclip_range(map, start_entry, start, real_end, &count,
2258                 MAP_CLIP_NO_HOLES);
2259         map->timestamp++;
2260         vm_map_unlock(map);
2261         vm_map_entry_release(count);
2262         return (rv);
2263 }
2264
2265 /*
2266  * Sets the pageability of the specified address range in the target map.
2267  * Regions specified as not pageable require locked-down physical
2268  * memory and physical page maps.
2269  *
2270  * The map must not be locked, but a reference must remain to the map
2271  * throughout the call.
2272  *
2273  * This function may be called via the zalloc path and must properly
2274  * reserve map entries for kernel_map.
2275  *
2276  * No requirements.
2277  */
2278 int
2279 vm_map_wire(vm_map_t map, vm_offset_t start, vm_offset_t real_end, int kmflags)
2280 {
2281         vm_map_entry_t entry;
2282         vm_map_entry_t start_entry;
2283         vm_offset_t end;
2284         int rv = KERN_SUCCESS;
2285         int count;
2286
2287         if (kmflags & KM_KRESERVE)
2288                 count = vm_map_entry_kreserve(MAP_RESERVE_COUNT);
2289         else
2290                 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
2291         vm_map_lock(map);
2292         VM_MAP_RANGE_CHECK(map, start, real_end);
2293         end = real_end;
2294
2295         start_entry = vm_map_clip_range(map, start, end, &count,
2296                                         MAP_CLIP_NO_HOLES);
2297         if (start_entry == NULL) {
2298                 vm_map_unlock(map);
2299                 rv = KERN_INVALID_ADDRESS;
2300                 goto failure;
2301         }
2302         if ((kmflags & KM_PAGEABLE) == 0) {
2303                 /*
2304                  * Wiring.  
2305                  *
2306                  * 1.  Holding the write lock, we create any shadow or zero-fill
2307                  * objects that need to be created. Then we clip each map
2308                  * entry to the region to be wired and increment its wiring
2309                  * count.  We create objects before clipping the map entries
2310                  * to avoid object proliferation.
2311                  *
2312                  * 2.  We downgrade to a read lock, and call vm_fault_wire to
2313                  * fault in the pages for any newly wired area (wired_count is
2314                  * 1).
2315                  *
2316                  * Downgrading to a read lock for vm_fault_wire avoids a 
2317                  * possible deadlock with another process that may have faulted
2318                  * on one of the pages to be wired (it would mark the page busy,
2319                  * blocking us, then in turn block on the map lock that we
2320                  * hold).  Because of problems in the recursive lock package,
2321                  * we cannot upgrade to a write lock in vm_map_lookup.  Thus,
2322                  * any actions that require the write lock must be done
2323                  * beforehand.  Because we keep the read lock on the map, the
2324                  * copy-on-write status of the entries we modify here cannot
2325                  * change.
2326                  */
2327                 entry = start_entry;
2328                 while ((entry != &map->header) && (entry->start < end)) {
2329                         /*
2330                          * Trivial case if the entry is already wired
2331                          */
2332                         if (entry->wired_count) {
2333                                 entry->wired_count++;
2334                                 entry = entry->next;
2335                                 continue;
2336                         }
2337
2338                         /*
2339                          * The entry is being newly wired, we have to setup
2340                          * appropriate management structures.  A shadow 
2341                          * object is required for a copy-on-write region,
2342                          * or a normal object for a zero-fill region.  We
2343                          * do not have to do this for entries that point to sub
2344                          * maps because we won't hold the lock on the sub map.
2345                          */
2346                         if (entry->maptype != VM_MAPTYPE_SUBMAP) {
2347                                 int copyflag = entry->eflags &
2348                                                MAP_ENTRY_NEEDS_COPY;
2349                                 if (copyflag && ((entry->protection &
2350                                                   VM_PROT_WRITE) != 0)) {
2351                                         vm_map_entry_shadow(entry, 0);
2352                                 } else if (entry->object.vm_object == NULL &&
2353                                            !map->system_map) {
2354                                         vm_map_entry_allocate_object(entry);
2355                                 }
2356                         }
2357
2358                         entry->wired_count++;
2359                         entry = entry->next;
2360                 }
2361
2362                 /*
2363                  * Pass 2.
2364                  */
2365
2366                 /*
2367                  * HACK HACK HACK HACK
2368                  *
2369                  * vm_fault_wire() temporarily unlocks the map to avoid
2370                  * deadlocks.  The in-transition flag from vm_map_clip_range
2371                  * call should protect us from changes while the map is
2372                  * unlocked.  T
2373                  *
2374                  * NOTE: Previously this comment stated that clipping might
2375                  *       still occur while the entry is unlocked, but from
2376                  *       what I can tell it actually cannot.
2377                  *
2378                  *       It is unclear whether the CLIP_CHECK_*() calls
2379                  *       are still needed but we keep them in anyway.
2380                  *
2381                  * HACK HACK HACK HACK
2382                  */
2383
2384                 entry = start_entry;
2385                 while (entry != &map->header && entry->start < end) {
2386                         /*
2387                          * If vm_fault_wire fails for any page we need to undo
2388                          * what has been done.  We decrement the wiring count
2389                          * for those pages which have not yet been wired (now)
2390                          * and unwire those that have (later).
2391                          */
2392                         vm_offset_t save_start = entry->start;
2393                         vm_offset_t save_end = entry->end;
2394
2395                         if (entry->wired_count == 1)
2396                                 rv = vm_fault_wire(map, entry, FALSE);
2397                         if (rv) {
2398                                 CLIP_CHECK_BACK(entry, save_start);
2399                                 for (;;) {
2400                                         KASSERT(entry->wired_count == 1, ("wired_count changed unexpectedly"));
2401                                         entry->wired_count = 0;
2402                                         if (entry->end == save_end)
2403                                                 break;
2404                                         entry = entry->next;
2405                                         KASSERT(entry != &map->header, ("bad entry clip during backout"));
2406                                 }
2407                                 end = save_start;
2408                                 break;
2409                         }
2410                         CLIP_CHECK_FWD(entry, save_end);
2411                         entry = entry->next;
2412                 }
2413
2414                 /*
2415                  * If a failure occured undo everything by falling through
2416                  * to the unwiring code.  'end' has already been adjusted
2417                  * appropriately.
2418                  */
2419                 if (rv)
2420                         kmflags |= KM_PAGEABLE;
2421
2422                 /*
2423                  * start_entry is still IN_TRANSITION but may have been 
2424                  * clipped since vm_fault_wire() unlocks and relocks the
2425                  * map.  No matter how clipped it has gotten there should
2426                  * be a fragment that is on our start boundary.
2427                  */
2428                 CLIP_CHECK_BACK(start_entry, start);
2429         }
2430
2431         if (kmflags & KM_PAGEABLE) {
2432                 /*
2433                  * This is the unwiring case.  We must first ensure that the
2434                  * range to be unwired is really wired down.  We know there
2435                  * are no holes.
2436                  */
2437                 entry = start_entry;
2438                 while ((entry != &map->header) && (entry->start < end)) {
2439                         if (entry->wired_count == 0) {
2440                                 rv = KERN_INVALID_ARGUMENT;
2441                                 goto done;
2442                         }
2443                         entry = entry->next;
2444                 }
2445
2446                 /*
2447                  * Now decrement the wiring count for each region. If a region
2448                  * becomes completely unwired, unwire its physical pages and
2449                  * mappings.
2450                  */
2451                 entry = start_entry;
2452                 while ((entry != &map->header) && (entry->start < end)) {
2453                         entry->wired_count--;
2454                         if (entry->wired_count == 0)
2455                                 vm_fault_unwire(map, entry);
2456                         entry = entry->next;
2457                 }
2458         }
2459 done:
2460         vm_map_unclip_range(map, start_entry, start, real_end,
2461                             &count, MAP_CLIP_NO_HOLES);
2462         map->timestamp++;
2463         vm_map_unlock(map);
2464 failure:
2465         if (kmflags & KM_KRESERVE)
2466                 vm_map_entry_krelease(count);
2467         else
2468                 vm_map_entry_release(count);
2469         return (rv);
2470 }
2471
2472 /*
2473  * Mark a newly allocated address range as wired but do not fault in
2474  * the pages.  The caller is expected to load the pages into the object.
2475  *
2476  * The map must be locked on entry and will remain locked on return.
2477  * No other requirements.
2478  */
2479 void
2480 vm_map_set_wired_quick(vm_map_t map, vm_offset_t addr, vm_size_t size,
2481                        int *countp)
2482 {
2483         vm_map_entry_t scan;
2484         vm_map_entry_t entry;
2485
2486         entry = vm_map_clip_range(map, addr, addr + size,
2487                                   countp, MAP_CLIP_NO_HOLES);
2488         for (scan = entry;
2489              scan != &map->header && scan->start < addr + size;
2490              scan = scan->next) {
2491             KKASSERT(entry->wired_count == 0);
2492             entry->wired_count = 1;                                              
2493         }
2494         vm_map_unclip_range(map, entry, addr, addr + size,
2495                             countp, MAP_CLIP_NO_HOLES);
2496 }
2497
2498 /*
2499  * Push any dirty cached pages in the address range to their pager.
2500  * If syncio is TRUE, dirty pages are written synchronously.
2501  * If invalidate is TRUE, any cached pages are freed as well.
2502  *
2503  * This routine is called by sys_msync()
2504  *
2505  * Returns an error if any part of the specified range is not mapped.
2506  *
2507  * No requirements.
2508  */
2509 int
2510 vm_map_clean(vm_map_t map, vm_offset_t start, vm_offset_t end,
2511              boolean_t syncio, boolean_t invalidate)
2512 {
2513         vm_map_entry_t current;
2514         vm_map_entry_t entry;
2515         vm_size_t size;
2516         vm_object_t object;
2517         vm_object_t tobj;
2518         vm_ooffset_t offset;
2519
2520         vm_map_lock_read(map);
2521         VM_MAP_RANGE_CHECK(map, start, end);
2522         if (!vm_map_lookup_entry(map, start, &entry)) {
2523                 vm_map_unlock_read(map);
2524                 return (KERN_INVALID_ADDRESS);
2525         }
2526         lwkt_gettoken(&map->token);
2527
2528         /*
2529          * Make a first pass to check for holes.
2530          */
2531         for (current = entry; current->start < end; current = current->next) {
2532                 if (current->maptype == VM_MAPTYPE_SUBMAP) {
2533                         lwkt_reltoken(&map->token);
2534                         vm_map_unlock_read(map);
2535                         return (KERN_INVALID_ARGUMENT);
2536                 }
2537                 if (end > current->end &&
2538                     (current->next == &map->header ||
2539                         current->end != current->next->start)) {
2540                         lwkt_reltoken(&map->token);
2541                         vm_map_unlock_read(map);
2542                         return (KERN_INVALID_ADDRESS);
2543                 }
2544         }
2545
2546         if (invalidate)
2547                 pmap_remove(vm_map_pmap(map), start, end);
2548
2549         /*
2550          * Make a second pass, cleaning/uncaching pages from the indicated
2551          * objects as we go.
2552          */
2553         for (current = entry; current->start < end; current = current->next) {
2554                 offset = current->offset + (start - current->start);
2555                 size = (end <= current->end ? end : current->end) - start;
2556                 if (current->maptype == VM_MAPTYPE_SUBMAP) {
2557                         vm_map_t smap;
2558                         vm_map_entry_t tentry;
2559                         vm_size_t tsize;
2560
2561                         smap = current->object.sub_map;
2562                         vm_map_lock_read(smap);
2563                         vm_map_lookup_entry(smap, offset, &tentry);
2564                         tsize = tentry->end - offset;
2565                         if (tsize < size)
2566                                 size = tsize;
2567                         object = tentry->object.vm_object;
2568                         offset = tentry->offset + (offset - tentry->start);
2569                         vm_map_unlock_read(smap);
2570                 } else {
2571                         object = current->object.vm_object;
2572                 }
2573
2574                 if (object)
2575                         vm_object_hold(object);
2576
2577                 /*
2578                  * Note that there is absolutely no sense in writing out
2579                  * anonymous objects, so we track down the vnode object
2580                  * to write out.
2581                  * We invalidate (remove) all pages from the address space
2582                  * anyway, for semantic correctness.
2583                  *
2584                  * note: certain anonymous maps, such as MAP_NOSYNC maps,
2585                  * may start out with a NULL object.
2586                  */
2587                 while (object && (tobj = object->backing_object) != NULL) {
2588                         vm_object_hold(tobj);
2589                         if (tobj == object->backing_object) {
2590                                 vm_object_lock_swap();
2591                                 offset += object->backing_object_offset;
2592                                 vm_object_drop(object);
2593                                 object = tobj;
2594                                 if (object->size < OFF_TO_IDX(offset + size))
2595                                         size = IDX_TO_OFF(object->size) -
2596                                                offset;
2597                                 break;
2598                         }
2599                         vm_object_drop(tobj);
2600                 }
2601                 if (object && (object->type == OBJT_VNODE) && 
2602                     (current->protection & VM_PROT_WRITE) &&
2603                     (object->flags & OBJ_NOMSYNC) == 0) {
2604                         /*
2605                          * Flush pages if writing is allowed, invalidate them
2606                          * if invalidation requested.  Pages undergoing I/O
2607                          * will be ignored by vm_object_page_remove().
2608                          *
2609                          * We cannot lock the vnode and then wait for paging
2610                          * to complete without deadlocking against vm_fault.
2611                          * Instead we simply call vm_object_page_remove() and
2612                          * allow it to block internally on a page-by-page 
2613                          * basis when it encounters pages undergoing async 
2614                          * I/O.
2615                          */
2616                         int flags;
2617
2618                         /* no chain wait needed for vnode objects */
2619                         vm_object_reference_locked(object);
2620                         vn_lock(object->handle, LK_EXCLUSIVE | LK_RETRY);
2621                         flags = (syncio || invalidate) ? OBJPC_SYNC : 0;
2622                         flags |= invalidate ? OBJPC_INVAL : 0;
2623
2624                         /*
2625                          * When operating on a virtual page table just
2626                          * flush the whole object.  XXX we probably ought
2627                          * to 
2628                          */
2629                         switch(current->maptype) {
2630                         case VM_MAPTYPE_NORMAL:
2631                                 vm_object_page_clean(object,
2632                                     OFF_TO_IDX(offset),
2633                                     OFF_TO_IDX(offset + size + PAGE_MASK),
2634                                     flags);
2635                                 break;
2636                         case VM_MAPTYPE_VPAGETABLE:
2637                                 vm_object_page_clean(object, 0, 0, flags);
2638                                 break;
2639                         }
2640                         vn_unlock(((struct vnode *)object->handle));
2641                         vm_object_deallocate_locked(object);
2642                 }
2643                 if (object && invalidate &&
2644                    ((object->type == OBJT_VNODE) ||
2645                     (object->type == OBJT_DEVICE))) {
2646                         int clean_only = 
2647                                 (object->type == OBJT_DEVICE) ? FALSE : TRUE;
2648                         /* no chain wait needed for vnode/device objects */
2649                         vm_object_reference_locked(object);
2650                         switch(current->maptype) {
2651                         case VM_MAPTYPE_NORMAL:
2652                                 vm_object_page_remove(object,
2653                                     OFF_TO_IDX(offset),
2654                                     OFF_TO_IDX(offset + size + PAGE_MASK),
2655                                     clean_only);
2656                                 break;
2657                         case VM_MAPTYPE_VPAGETABLE:
2658                                 vm_object_page_remove(object, 0, 0, clean_only);
2659                                 break;
2660                         }
2661                         vm_object_deallocate_locked(object);
2662                 }
2663                 start += size;
2664                 if (object)
2665                         vm_object_drop(object);
2666         }
2667
2668         lwkt_reltoken(&map->token);
2669         vm_map_unlock_read(map);
2670
2671         return (KERN_SUCCESS);
2672 }
2673
2674 /*
2675  * Make the region specified by this entry pageable.
2676  *
2677  * The vm_map must be exclusively locked.
2678  */
2679 static void 
2680 vm_map_entry_unwire(vm_map_t map, vm_map_entry_t entry)
2681 {
2682         entry->eflags &= ~MAP_ENTRY_USER_WIRED;
2683         entry->wired_count = 0;
2684         vm_fault_unwire(map, entry);
2685 }
2686
2687 /*
2688  * Deallocate the given entry from the target map.
2689  *
2690  * The vm_map must be exclusively locked.
2691  */
2692 static void
2693 vm_map_entry_delete(vm_map_t map, vm_map_entry_t entry, int *countp)
2694 {
2695         vm_map_entry_unlink(map, entry);
2696         map->size -= entry->end - entry->start;
2697
2698         switch(entry->maptype) {
2699         case VM_MAPTYPE_NORMAL:
2700         case VM_MAPTYPE_VPAGETABLE:
2701                 vm_object_deallocate(entry->object.vm_object);
2702                 break;
2703         default:
2704                 break;
2705         }
2706
2707         vm_map_entry_dispose(map, entry, countp);
2708 }
2709
2710 /*
2711  * Deallocates the given address range from the target map.
2712  *
2713  * The vm_map must be exclusively locked.
2714  */
2715 int
2716 vm_map_delete(vm_map_t map, vm_offset_t start, vm_offset_t end, int *countp)
2717 {
2718         vm_object_t object;
2719         vm_map_entry_t entry;
2720         vm_map_entry_t first_entry;
2721
2722         ASSERT_VM_MAP_LOCKED(map);
2723         lwkt_gettoken(&map->token);
2724 again:
2725         /*
2726          * Find the start of the region, and clip it.  Set entry to point
2727          * at the first record containing the requested address or, if no
2728          * such record exists, the next record with a greater address.  The
2729          * loop will run from this point until a record beyond the termination
2730          * address is encountered.
2731          *
2732          * map->hint must be adjusted to not point to anything we delete,
2733          * so set it to the entry prior to the one being deleted.
2734          *
2735          * GGG see other GGG comment.
2736          */
2737         if (vm_map_lookup_entry(map, start, &first_entry)) {
2738                 entry = first_entry;
2739                 vm_map_clip_start(map, entry, start, countp);
2740                 map->hint = entry->prev;        /* possible problem XXX */
2741         } else {
2742                 map->hint = first_entry;        /* possible problem XXX */
2743                 entry = first_entry->next;
2744         }
2745
2746         /*
2747          * If a hole opens up prior to the current first_free then
2748          * adjust first_free.  As with map->hint, map->first_free
2749          * cannot be left set to anything we might delete.
2750          */
2751         if (entry == &map->header) {
2752                 map->first_free = &map->header;
2753         } else if (map->first_free->start >= start) {
2754                 map->first_free = entry->prev;
2755         }
2756
2757         /*
2758          * Step through all entries in this region
2759          */
2760         while ((entry != &map->header) && (entry->start < end)) {
2761                 vm_map_entry_t next;
2762                 vm_offset_t s, e;
2763                 vm_pindex_t offidxstart, offidxend, count;
2764
2765                 /*
2766                  * If we hit an in-transition entry we have to sleep and
2767                  * retry.  It's easier (and not really slower) to just retry
2768                  * since this case occurs so rarely and the hint is already
2769                  * pointing at the right place.  We have to reset the
2770                  * start offset so as not to accidently delete an entry
2771                  * another process just created in vacated space.
2772                  */
2773                 if (entry->eflags & MAP_ENTRY_IN_TRANSITION) {
2774                         entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP;
2775                         start = entry->start;
2776                         ++mycpu->gd_cnt.v_intrans_coll;
2777                         ++mycpu->gd_cnt.v_intrans_wait;
2778                         vm_map_transition_wait(map);
2779                         goto again;
2780                 }
2781                 vm_map_clip_end(map, entry, end, countp);
2782
2783                 s = entry->start;
2784                 e = entry->end;
2785                 next = entry->next;
2786
2787                 offidxstart = OFF_TO_IDX(entry->offset);
2788                 count = OFF_TO_IDX(e - s);
2789                 object = entry->object.vm_object;
2790
2791                 /*
2792                  * Unwire before removing addresses from the pmap; otherwise,
2793                  * unwiring will put the entries back in the pmap.
2794                  */
2795                 if (entry->wired_count != 0)
2796                         vm_map_entry_unwire(map, entry);
2797
2798                 offidxend = offidxstart + count;
2799
2800                 if (object == &kernel_object) {
2801                         vm_object_hold(object);
2802                         vm_object_page_remove(object, offidxstart,
2803                                               offidxend, FALSE);
2804                         vm_object_drop(object);
2805                 } else if (object && object->type != OBJT_DEFAULT &&
2806                            object->type != OBJT_SWAP) {
2807                         /*
2808                          * vnode object routines cannot be chain-locked
2809                          */
2810                         vm_object_hold(object);
2811                         pmap_remove(map->pmap, s, e);
2812                         vm_object_drop(object);
2813                 } else if (object) {
2814                         vm_object_hold(object);
2815                         vm_object_chain_acquire(object);
2816                         pmap_remove(map->pmap, s, e);
2817
2818                         if (object != NULL &&
2819                             object->ref_count != 1 &&
2820                             (object->flags & (OBJ_NOSPLIT|OBJ_ONEMAPPING)) ==
2821                              OBJ_ONEMAPPING &&
2822                             (object->type == OBJT_DEFAULT ||
2823                              object->type == OBJT_SWAP)) {
2824                                 vm_object_collapse(object, NULL);
2825                                 vm_object_page_remove(object, offidxstart,
2826                                                       offidxend, FALSE);
2827                                 if (object->type == OBJT_SWAP) {
2828                                         swap_pager_freespace(object,
2829                                                              offidxstart,
2830                                                              count);
2831                                 }
2832                                 if (offidxend >= object->size &&
2833                                     offidxstart < object->size) {
2834                                         object->size = offidxstart;
2835                                 }
2836                         }
2837                         vm_object_chain_release(object);
2838                         vm_object_drop(object);
2839                 }
2840
2841                 /*
2842                  * Delete the entry (which may delete the object) only after
2843                  * removing all pmap entries pointing to its pages.
2844                  * (Otherwise, its page frames may be reallocated, and any
2845                  * modify bits will be set in the wrong object!)
2846                  */
2847                 vm_map_entry_delete(map, entry, countp);
2848                 entry = next;
2849         }
2850         lwkt_reltoken(&map->token);
2851         return (KERN_SUCCESS);
2852 }
2853
2854 /*
2855  * Remove the given address range from the target map.
2856  * This is the exported form of vm_map_delete.
2857  *
2858  * No requirements.
2859  */
2860 int
2861 vm_map_remove(vm_map_t map, vm_offset_t start, vm_offset_t end)
2862 {
2863         int result;
2864         int count;
2865
2866         count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
2867         vm_map_lock(map);
2868         VM_MAP_RANGE_CHECK(map, start, end);
2869         result = vm_map_delete(map, start, end, &count);
2870         vm_map_unlock(map);
2871         vm_map_entry_release(count);
2872
2873         return (result);
2874 }
2875
2876 /*
2877  * Assert that the target map allows the specified privilege on the
2878  * entire address region given.  The entire region must be allocated.
2879  *
2880  * The caller must specify whether the vm_map is already locked or not.
2881  */
2882 boolean_t
2883 vm_map_check_protection(vm_map_t map, vm_offset_t start, vm_offset_t end,
2884                         vm_prot_t protection, boolean_t have_lock)
2885 {
2886         vm_map_entry_t entry;
2887         vm_map_entry_t tmp_entry;
2888         boolean_t result;
2889
2890         if (have_lock == FALSE)
2891                 vm_map_lock_read(map);
2892
2893         if (!vm_map_lookup_entry(map, start, &tmp_entry)) {
2894                 if (have_lock == FALSE)
2895                         vm_map_unlock_read(map);
2896                 return (FALSE);
2897         }
2898         entry = tmp_entry;
2899
2900         result = TRUE;
2901         while (start < end) {
2902                 if (entry == &map->header) {
2903                         result = FALSE;
2904                         break;
2905                 }
2906                 /*
2907                  * No holes allowed!
2908                  */
2909
2910                 if (start < entry->start) {
2911                         result = FALSE;
2912                         break;
2913                 }
2914                 /*
2915                  * Check protection associated with entry.
2916                  */
2917
2918                 if ((entry->protection & protection) != protection) {
2919                         result = FALSE;
2920                         break;
2921                 }
2922                 /* go to next entry */
2923
2924                 start = entry->end;
2925                 entry = entry->next;
2926         }
2927         if (have_lock == FALSE)
2928                 vm_map_unlock_read(map);
2929         return (result);
2930 }
2931
2932 /*
2933  * If appropriate this function shadows the original object with a new object
2934  * and moves the VM pages from the original object to the new object.
2935  * The original object will also be collapsed, if possible.
2936  *
2937  * We can only do this for normal memory objects with a single mapping, and
2938  * it only makes sense to do it if there are 2 or more refs on the original
2939  * object.  i.e. typically a memory object that has been extended into
2940  * multiple vm_map_entry's with non-overlapping ranges.
2941  *
2942  * This makes it easier to remove unused pages and keeps object inheritance
2943  * from being a negative impact on memory usage.
2944  *
2945  * On return the (possibly new) entry->object.vm_object will have an
2946  * additional ref on it for the caller to dispose of (usually by cloning
2947  * the vm_map_entry).  The additional ref had to be done in this routine
2948  * to avoid racing a collapse.  The object's ONEMAPPING flag will also be
2949  * cleared.
2950  *
2951  * The vm_map must be locked and its token held.
2952  */
2953 static void
2954 vm_map_split(vm_map_entry_t entry)
2955 {
2956 #if 0
2957         /* UNOPTIMIZED */
2958         vm_object_t oobject;
2959
2960         oobject = entry->object.vm_object;
2961         vm_object_hold(oobject);
2962         vm_object_chain_wait(oobject);
2963         vm_object_reference_locked(oobject);
2964         vm_object_clear_flag(oobject, OBJ_ONEMAPPING);
2965         vm_object_drop(oobject);
2966 #else
2967         /* OPTIMIZED */
2968         vm_object_t oobject, nobject, bobject;
2969         vm_offset_t s, e;
2970         vm_page_t m;
2971         vm_pindex_t offidxstart, offidxend, idx;
2972         vm_size_t size;
2973         vm_ooffset_t offset;
2974
2975         /*
2976          * Setup.  Chain lock the original object throughout the entire
2977          * routine to prevent new page faults from occuring.
2978          *
2979          * XXX can madvise WILLNEED interfere with us too?
2980          */
2981         oobject = entry->object.vm_object;
2982         vm_object_hold(oobject);
2983         vm_object_chain_acquire(oobject);
2984
2985         /*
2986          * Original object cannot be split?
2987          */
2988         if (oobject->handle == NULL || (oobject->type != OBJT_DEFAULT &&
2989                                         oobject->type != OBJT_SWAP)) {
2990                 vm_object_chain_release(oobject);
2991                 vm_object_reference_locked(oobject);
2992                 vm_object_clear_flag(oobject, OBJ_ONEMAPPING);
2993                 vm_object_drop(oobject);
2994                 return;
2995         }
2996
2997         /*
2998          * Collapse original object with its backing store as an
2999          * optimization to reduce chain lengths when possible.
3000          *
3001          * If ref_count <= 1 there aren't other non-overlapping vm_map_entry's
3002          * for oobject, so there's no point collapsing it.
3003          *
3004          * Then re-check whether the object can be split.
3005          */
3006         vm_object_collapse(oobject, NULL);
3007
3008         if (oobject->ref_count <= 1 ||
3009             (oobject->type != OBJT_DEFAULT && oobject->type != OBJT_SWAP) ||
3010             (oobject->flags & (OBJ_NOSPLIT|OBJ_ONEMAPPING)) != OBJ_ONEMAPPING) {
3011                 vm_object_chain_release(oobject);
3012                 vm_object_reference_locked(oobject);
3013                 vm_object_clear_flag(oobject, OBJ_ONEMAPPING);
3014                 vm_object_drop(oobject);
3015                 return;
3016         }
3017
3018         /*
3019          * Acquire the chain lock on the backing object.
3020          *
3021          * Give bobject an additional ref count for when it will be shadowed
3022          * by nobject.
3023          */
3024         if ((bobject = oobject->backing_object) != NULL) {
3025                 vm_object_hold(bobject);
3026                 vm_object_chain_wait(bobject);
3027                 vm_object_reference_locked(bobject);
3028                 vm_object_chain_acquire(bobject);
3029                 KKASSERT(bobject->backing_object == bobject);
3030                 KKASSERT((bobject->flags & OBJ_DEAD) == 0);
3031         }
3032
3033         /*
3034          * Calculate the object page range and allocate the new object.
3035          */
3036         offset = entry->offset;
3037         s = entry->start;
3038         e = entry->end;
3039
3040         offidxstart = OFF_TO_IDX(offset);
3041         offidxend = offidxstart + OFF_TO_IDX(e - s);
3042         size = offidxend - offidxstart;
3043
3044         switch(oobject->type) {
3045         case OBJT_DEFAULT:
3046                 nobject = default_pager_alloc(NULL, IDX_TO_OFF(size),
3047                                               VM_PROT_ALL, 0);
3048                 break;
3049         case OBJT_SWAP:
3050                 nobject = swap_pager_alloc(NULL, IDX_TO_OFF(size),
3051                                            VM_PROT_ALL, 0);
3052                 break;
3053         default:
3054                 /* not reached */
3055                 nobject = NULL;
3056                 KKASSERT(0);
3057         }
3058
3059         if (nobject == NULL) {
3060                 if (bobject) {
3061                         vm_object_chain_release(bobject);
3062                         vm_object_deallocate(bobject);
3063                         vm_object_drop(bobject);
3064                 }
3065                 vm_object_chain_release(oobject);
3066                 vm_object_reference_locked(oobject);
3067                 vm_object_clear_flag(oobject, OBJ_ONEMAPPING);
3068                 vm_object_drop(oobject);
3069                 return;
3070         }
3071
3072         /*
3073          * The new object will replace entry->object.vm_object so it needs
3074          * a second reference (the caller expects an additional ref).
3075          */
3076         vm_object_hold(nobject);
3077         vm_object_reference_locked(nobject);
3078         vm_object_chain_acquire(nobject);
3079
3080         /*
3081          * nobject shadows bobject (oobject already shadows bobject).
3082          */
3083         if (bobject) {
3084                 nobject->backing_object_offset =
3085                     oobject->backing_object_offset + IDX_TO_OFF(offidxstart);
3086                 nobject->backing_object = bobject;
3087                 bobject->shadow_count++;
3088                 bobject->generation++;
3089                 LIST_INSERT_HEAD(&bobject->shadow_head, nobject, shadow_list);
3090                 vm_object_clear_flag(bobject, OBJ_ONEMAPPING); /* XXX? */
3091                 vm_object_chain_release(bobject);
3092                 vm_object_drop(bobject);
3093         }
3094
3095         /*
3096          * Move the VM pages from oobject to nobject
3097          */
3098         for (idx = 0; idx < size; idx++) {
3099                 vm_page_t m;
3100
3101                 m = vm_page_lookup_busy_wait(oobject, offidxstart + idx,
3102                                              TRUE, "vmpg");
3103                 if (m == NULL)
3104                         continue;
3105
3106                 /*
3107                  * We must wait for pending I/O to complete before we can
3108                  * rename the page.
3109                  *
3110                  * We do not have to VM_PROT_NONE the page as mappings should
3111                  * not be changed by this operation.
3112                  *
3113                  * NOTE: The act of renaming a page updates chaingen for both
3114                  *       objects.
3115                  */
3116                 vm_page_rename(m, nobject, idx);
3117                 /* page automatically made dirty by rename and cache handled */
3118                 /* page remains busy */
3119         }
3120
3121         if (oobject->type == OBJT_SWAP) {
3122                 vm_object_pip_add(oobject, 1);
3123                 /*
3124                  * copy oobject pages into nobject and destroy unneeded
3125                  * pages in shadow object.
3126                  */
3127                 swap_pager_copy(oobject, nobject, offidxstart, 0);
3128                 vm_object_pip_wakeup(oobject);
3129         }
3130
3131         /*
3132          * Wakeup the pages we played with.  No spl protection is needed
3133          * for a simple wakeup.
3134          */
3135         for (idx = 0; idx < size; idx++) {
3136                 m = vm_page_lookup(nobject, idx);
3137                 if (m) {
3138                         KKASSERT(m->flags & PG_BUSY);
3139                         vm_page_wakeup(m);
3140                 }
3141         }
3142         entry->object.vm_object = nobject;
3143         entry->offset = 0LL;
3144
3145         /*
3146          * Cleanup
3147          *
3148          * NOTE: There is no need to remove OBJ_ONEMAPPING from oobject, the
3149          *       related pages were moved and are no longer applicable to the
3150          *       original object.
3151          *
3152          * NOTE: Deallocate oobject (due to its entry->object.vm_object being
3153          *       replaced by nobject).
3154          */
3155         vm_object_chain_release(nobject);
3156         vm_object_drop(nobject);
3157         if (bobject) {
3158                 vm_object_chain_release(bobject);
3159                 vm_object_drop(bobject);
3160         }
3161         vm_object_chain_release(oobject);
3162         /*vm_object_clear_flag(oobject, OBJ_ONEMAPPING);*/
3163         vm_object_deallocate_locked(oobject);
3164         vm_object_drop(oobject);
3165 #endif
3166 }
3167
3168 /*
3169  * Copies the contents of the source entry to the destination
3170  * entry.  The entries *must* be aligned properly.
3171  *
3172  * The vm_maps must be exclusively locked.
3173  * The vm_map's token must be held.
3174  *
3175  * Because the maps are locked no faults can be in progress during the
3176  * operation.
3177  */
3178 static void
3179 vm_map_copy_entry(vm_map_t src_map, vm_map_t dst_map,
3180                   vm_map_entry_t src_entry, vm_map_entry_t dst_entry)
3181 {
3182         vm_object_t src_object;
3183
3184         if (dst_entry->maptype == VM_MAPTYPE_SUBMAP)
3185                 return;
3186         if (src_entry->maptype == VM_MAPTYPE_SUBMAP)
3187                 return;
3188
3189         if (src_entry->wired_count == 0) {
3190                 /*
3191                  * If the source entry is marked needs_copy, it is already
3192                  * write-protected.
3193                  */
3194                 if ((src_entry->eflags & MAP_ENTRY_NEEDS_COPY) == 0) {
3195                         pmap_protect(src_map->pmap,
3196                             src_entry->start,
3197                             src_entry->end,
3198                             src_entry->protection & ~VM_PROT_WRITE);
3199                 }
3200
3201                 /*
3202                  * Make a copy of the object.
3203                  *
3204                  * The object must be locked prior to checking the object type
3205                  * and for the call to vm_object_collapse() and vm_map_split().
3206                  * We cannot use *_hold() here because the split code will
3207                  * probably try to destroy the object.  The lock is a pool
3208                  * token and doesn't care.
3209                  *
3210                  * We must bump src_map->timestamp when setting
3211                  * MAP_ENTRY_NEEDS_COPY to force any concurrent fault
3212                  * to retry, otherwise the concurrent fault might improperly
3213                  * install a RW pte when its supposed to be a RO(COW) pte.
3214                  * This race can occur because a vnode-backed fault may have
3215                  * to temporarily release the map lock.
3216                  */
3217                 if (src_entry->object.vm_object != NULL) {
3218                         vm_map_split(src_entry);
3219                         src_object = src_entry->object.vm_object;
3220                         dst_entry->object.vm_object = src_object;
3221                         src_entry->eflags |= (MAP_ENTRY_COW |
3222                                               MAP_ENTRY_NEEDS_COPY);
3223                         dst_entry->eflags |= (MAP_ENTRY_COW |
3224                                               MAP_ENTRY_NEEDS_COPY);
3225                         dst_entry->offset = src_entry->offset;
3226                         ++src_map->timestamp;
3227                 } else {
3228                         dst_entry->object.vm_object = NULL;
3229                         dst_entry->offset = 0;
3230                 }
3231
3232                 pmap_copy(dst_map->pmap, src_map->pmap, dst_entry->start,
3233                     dst_entry->end - dst_entry->start, src_entry->start);
3234         } else {
3235                 /*
3236                  * Of course, wired down pages can't be set copy-on-write.
3237                  * Cause wired pages to be copied into the new map by
3238                  * simulating faults (the new pages are pageable)
3239                  */
3240                 vm_fault_copy_entry(dst_map, src_map, dst_entry, src_entry);
3241         }
3242 }
3243
3244 /*
3245  * vmspace_fork:
3246  * Create a new process vmspace structure and vm_map
3247  * based on those of an existing process.  The new map
3248  * is based on the old map, according to the inheritance
3249  * values on the regions in that map.
3250  *
3251  * The source map must not be locked.
3252  * No requirements.
3253  */
3254 struct vmspace *
3255 vmspace_fork(struct vmspace *vm1)
3256 {
3257         struct vmspace *vm2;
3258         vm_map_t old_map = &vm1->vm_map;
3259         vm_map_t new_map;
3260         vm_map_entry_t old_entry;
3261         vm_map_entry_t new_entry;
3262         vm_object_t object;
3263         int count;
3264
3265         lwkt_gettoken(&vm1->vm_map.token);
3266         vm_map_lock(old_map);
3267
3268         /*
3269          * XXX Note: upcalls are not copied.
3270          */
3271         vm2 = vmspace_alloc(old_map->min_offset, old_map->max_offset);
3272         lwkt_gettoken(&vm2->vm_map.token);
3273         bcopy(&vm1->vm_startcopy, &vm2->vm_startcopy,
3274             (caddr_t)&vm1->vm_endcopy - (caddr_t)&vm1->vm_startcopy);
3275         new_map = &vm2->vm_map; /* XXX */
3276         new_map->timestamp = 1;
3277
3278         vm_map_lock(new_map);
3279
3280         count = 0;
3281         old_entry = old_map->header.next;
3282         while (old_entry != &old_map->header) {
3283                 ++count;
3284                 old_entry = old_entry->next;
3285         }
3286
3287         count = vm_map_entry_reserve(count + MAP_RESERVE_COUNT);
3288
3289         old_entry = old_map->header.next;
3290         while (old_entry != &old_map->header) {
3291                 if (old_entry->maptype == VM_MAPTYPE_SUBMAP)
3292                         panic("vm_map_fork: encountered a submap");
3293
3294                 switch (old_entry->inheritance) {
3295                 case VM_INHERIT_NONE:
3296                         break;
3297                 case VM_INHERIT_SHARE:
3298                         /*
3299                          * Clone the entry, creating the shared object if
3300                          * necessary.
3301                          */
3302                         if (old_entry->object.vm_object == NULL)
3303                                 vm_map_entry_allocate_object(old_entry);
3304
3305                         if (old_entry->eflags & MAP_ENTRY_NEEDS_COPY) {
3306                                 /*
3307                                  * Shadow a map_entry which needs a copy,
3308                                  * replacing its object with a new object
3309                                  * that points to the old one.  Ask the
3310                                  * shadow code to automatically add an
3311                                  * additional ref.  We can't do it afterwords
3312                                  * because we might race a collapse.  The call
3313                                  * to vm_map_entry_shadow() will also clear
3314                                  * OBJ_ONEMAPPING.
3315                                  */
3316                                 vm_map_entry_shadow(old_entry, 1);
3317                         } else {
3318                                 /*
3319                                  * We will make a shared copy of the object,
3320                                  * and must clear OBJ_ONEMAPPING.
3321                                  *
3322                                  * XXX assert that object.vm_object != NULL
3323                                  *     since we allocate it above.
3324                                  */
3325                                 if (old_entry->object.vm_object) {
3326                                         object = old_entry->object.vm_object;
3327                                         vm_object_hold(object);
3328                                         vm_object_chain_wait(object);
3329                                         vm_object_reference_locked(object);
3330                                         vm_object_clear_flag(object,
3331                                                              OBJ_ONEMAPPING);
3332                                         vm_object_drop(object);
3333                                 }
3334                         }
3335
3336                         /*
3337                          * Clone the entry.  We've already bumped the ref on
3338                          * any vm_object.
3339                          */
3340                         new_entry = vm_map_entry_create(new_map, &count);
3341                         *new_entry = *old_entry;
3342                         new_entry->eflags &= ~MAP_ENTRY_USER_WIRED;
3343                         new_entry->wired_count = 0;
3344
3345                         /*
3346                          * Insert the entry into the new map -- we know we're
3347                          * inserting at the end of the new map.
3348                          */
3349
3350                         vm_map_entry_link(new_map, new_map->header.prev,
3351                                           new_entry);
3352
3353                         /*
3354                          * Update the physical map
3355                          */
3356                         pmap_copy(new_map->pmap, old_map->pmap,
3357                             new_entry->start,
3358                             (old_entry->end - old_entry->start),
3359                             old_entry->start);
3360                         break;
3361                 case VM_INHERIT_COPY:
3362                         /*
3363                          * Clone the entry and link into the map.
3364                          */
3365                         new_entry = vm_map_entry_create(new_map, &count);
3366                         *new_entry = *old_entry;
3367                         new_entry->eflags &= ~MAP_ENTRY_USER_WIRED;
3368                         new_entry->wired_count = 0;
3369                         new_entry->object.vm_object = NULL;
3370                         vm_map_entry_link(new_map, new_map->header.prev,
3371                                           new_entry);
3372                         vm_map_copy_entry(old_map, new_map, old_entry,
3373                                           new_entry);
3374                         break;
3375                 }
3376                 old_entry = old_entry->next;
3377         }
3378
3379         new_map->size = old_map->size;
3380         vm_map_unlock(old_map);
3381         vm_map_unlock(new_map);
3382         vm_map_entry_release(count);
3383
3384         lwkt_reltoken(&vm2->vm_map.token);
3385         lwkt_reltoken(&vm1->vm_map.token);
3386
3387         return (vm2);
3388 }
3389
3390 /*
3391  * Create an auto-grow stack entry
3392  *
3393  * No requirements.
3394  */
3395 int
3396 vm_map_stack (vm_map_t map, vm_offset_t addrbos, vm_size_t max_ssize,
3397               int flags, vm_prot_t prot, vm_prot_t max, int cow)
3398 {
3399         vm_map_entry_t  prev_entry;
3400         vm_map_entry_t  new_stack_entry;
3401         vm_size_t       init_ssize;
3402         int             rv;
3403         int             count;
3404         vm_offset_t     tmpaddr;
3405
3406         cow |= MAP_IS_STACK;
3407
3408         if (max_ssize < sgrowsiz)
3409                 init_ssize = max_ssize;
3410         else
3411                 init_ssize = sgrowsiz;
3412
3413         count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
3414         vm_map_lock(map);
3415
3416         /*
3417          * Find space for the mapping
3418          */
3419         if ((flags & (MAP_FIXED | MAP_TRYFIXED)) == 0) {
3420                 if (vm_map_findspace(map, addrbos, max_ssize, 1,
3421                                      flags, &tmpaddr)) {
3422                         vm_map_unlock(map);
3423                         vm_map_entry_release(count);
3424                         return (KERN_NO_SPACE);
3425                 }
3426                 addrbos = tmpaddr;
3427         }
3428
3429         /* If addr is already mapped, no go */
3430         if (vm_map_lookup_entry(map, addrbos, &prev_entry)) {
3431                 vm_map_unlock(map);
3432                 vm_map_entry_release(count);
3433                 return (KERN_NO_SPACE);
3434         }
3435
3436 #if 0
3437         /* XXX already handled by kern_mmap() */
3438         /* If we would blow our VMEM resource limit, no go */
3439         if (map->size + init_ssize >
3440             curproc->p_rlimit[RLIMIT_VMEM].rlim_cur) {
3441                 vm_map_unlock(map);
3442                 vm_map_entry_release(count);
3443                 return (KERN_NO_SPACE);
3444         }
3445 #endif
3446
3447         /*
3448          * If we can't accomodate max_ssize in the current mapping,
3449          * no go.  However, we need to be aware that subsequent user
3450          * mappings might map into the space we have reserved for
3451          * stack, and currently this space is not protected.  
3452          * 
3453          * Hopefully we will at least detect this condition 
3454          * when we try to grow the stack.
3455          */
3456         if ((prev_entry->next != &map->header) &&
3457             (prev_entry->next->start < addrbos + max_ssize)) {
3458                 vm_map_unlock(map);
3459                 vm_map_entry_release(count);
3460                 return (KERN_NO_SPACE);
3461         }
3462
3463         /*
3464          * We initially map a stack of only init_ssize.  We will
3465          * grow as needed later.  Since this is to be a grow 
3466          * down stack, we map at the top of the range.
3467          *
3468          * Note: we would normally expect prot and max to be
3469          * VM_PROT_ALL, and cow to be 0.  Possibly we should
3470          * eliminate these as input parameters, and just
3471          * pass these values here in the insert call.
3472          */
3473         rv = vm_map_insert(map, &count,
3474                            NULL, 0, addrbos + max_ssize - init_ssize,
3475                            addrbos + max_ssize,
3476                            VM_MAPTYPE_NORMAL,
3477                            prot, max,
3478                            cow);
3479
3480         /* Now set the avail_ssize amount */
3481         if (rv == KERN_SUCCESS) {
3482                 if (prev_entry != &map->header)
3483                         vm_map_clip_end(map, prev_entry, addrbos + max_ssize - init_ssize, &count);
3484                 new_stack_entry = prev_entry->next;
3485                 if (new_stack_entry->end   != addrbos + max_ssize ||
3486                     new_stack_entry->start != addrbos + max_ssize - init_ssize)
3487                         panic ("Bad entry start/end for new stack entry");
3488                 else 
3489                         new_stack_entry->aux.avail_ssize = max_ssize - init_ssize;
3490         }
3491
3492         vm_map_unlock(map);
3493         vm_map_entry_release(count);
3494         return (rv);
3495 }
3496
3497 /*
3498  * Attempts to grow a vm stack entry.  Returns KERN_SUCCESS if the
3499  * desired address is already mapped, or if we successfully grow
3500  * the stack.  Also returns KERN_SUCCESS if addr is outside the
3501  * stack range (this is strange, but preserves compatibility with
3502  * the grow function in vm_machdep.c).
3503  *
3504  * No requirements.
3505  */
3506 int
3507 vm_map_growstack (struct proc *p, vm_offset_t addr)
3508 {
3509         vm_map_entry_t prev_entry;
3510         vm_map_entry_t stack_entry;
3511         vm_map_entry_t new_stack_entry;
3512         struct vmspace *vm = p->p_vmspace;
3513         vm_map_t map = &vm->vm_map;
3514         vm_offset_t    end;
3515         int grow_amount;
3516         int rv = KERN_SUCCESS;
3517         int is_procstack;
3518         int use_read_lock = 1;
3519         int count;
3520
3521         count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
3522 Retry:
3523         if (use_read_lock)
3524                 vm_map_lock_read(map);
3525         else
3526                 vm_map_lock(map);
3527
3528         /* If addr is already in the entry range, no need to grow.*/
3529         if (vm_map_lookup_entry(map, addr, &prev_entry))
3530                 goto done;
3531
3532         if ((stack_entry = prev_entry->next) == &map->header)
3533                 goto done;
3534         if (prev_entry == &map->header) 
3535                 end = stack_entry->start - stack_entry->aux.avail_ssize;
3536         else
3537                 end = prev_entry->end;
3538
3539         /*
3540          * This next test mimics the old grow function in vm_machdep.c.
3541          * It really doesn't quite make sense, but we do it anyway
3542          * for compatibility.
3543          *
3544          * If not growable stack, return success.  This signals the
3545          * caller to proceed as he would normally with normal vm.
3546          */
3547         if (stack_entry->aux.avail_ssize < 1 ||
3548             addr >= stack_entry->start ||
3549             addr <  stack_entry->start - stack_entry->aux.avail_ssize) {
3550                 goto done;
3551         } 
3552         
3553         /* Find the minimum grow amount */
3554         grow_amount = roundup (stack_entry->start - addr, PAGE_SIZE);
3555         if (grow_amount > stack_entry->aux.avail_ssize) {
3556                 rv = KERN_NO_SPACE;
3557                 goto done;
3558         }
3559
3560         /*
3561          * If there is no longer enough space between the entries
3562          * nogo, and adjust the available space.  Note: this 
3563          * should only happen if the user has mapped into the
3564          * stack area after the stack was created, and is
3565          * probably an error.
3566          *
3567          * This also effectively destroys any guard page the user
3568          * might have intended by limiting the stack size.
3569          */
3570         if (grow_amount > stack_entry->start - end) {
3571                 if (use_read_lock && vm_map_lock_upgrade(map)) {
3572                         /* lost lock */
3573                         use_read_lock = 0;
3574                         goto Retry;
3575                 }
3576                 use_read_lock = 0;
3577                 stack_entry->aux.avail_ssize = stack_entry->start - end;
3578                 rv = KERN_NO_SPACE;
3579                 goto done;
3580         }
3581
3582         is_procstack = addr >= (vm_offset_t)vm->vm_maxsaddr;
3583
3584         /* If this is the main process stack, see if we're over the 
3585          * stack limit.
3586          */
3587         if (is_procstack && (ctob(vm->vm_ssize) + grow_amount >
3588                              p->p_rlimit[RLIMIT_STACK].rlim_cur)) {
3589                 rv = KERN_NO_SPACE;
3590                 goto done;
3591         }
3592
3593         /* Round up the grow amount modulo SGROWSIZ */
3594         grow_amount = roundup (grow_amount, sgrowsiz);
3595         if (grow_amount > stack_entry->aux.avail_ssize) {
3596                 grow_amount = stack_entry->aux.avail_ssize;
3597         }
3598         if (is_procstack && (ctob(vm->vm_ssize) + grow_amount >
3599                              p->p_rlimit[RLIMIT_STACK].rlim_cur)) {
3600                 grow_amount = p->p_rlimit[RLIMIT_STACK].rlim_cur -
3601                               ctob(vm->vm_ssize);
3602         }
3603
3604         /* If we would blow our VMEM resource limit, no go */
3605         if (map->size + grow_amount > p->p_rlimit[RLIMIT_VMEM].rlim_cur) {
3606                 rv = KERN_NO_SPACE;
3607                 goto done;
3608         }
3609
3610         if (use_read_lock && vm_map_lock_upgrade(map)) {
3611                 /* lost lock */
3612                 use_read_lock = 0;
3613                 goto Retry;
3614         }
3615         use_read_lock = 0;
3616
3617         /* Get the preliminary new entry start value */
3618         addr = stack_entry->start - grow_amount;
3619
3620         /* If this puts us into the previous entry, cut back our growth
3621          * to the available space.  Also, see the note above.
3622          */
3623         if (addr < end) {
3624                 stack_entry->aux.avail_ssize = stack_entry->start - end;
3625                 addr = end;
3626         }
3627
3628         rv = vm_map_insert(map, &count,
3629                            NULL, 0, addr, stack_entry->start,
3630                            VM_MAPTYPE_NORMAL,
3631                            VM_PROT_ALL, VM_PROT_ALL,
3632                            0);
3633
3634         /* Adjust the available stack space by the amount we grew. */
3635         if (rv == KERN_SUCCESS) {
3636                 if (prev_entry != &map->header)
3637                         vm_map_clip_end(map, prev_entry, addr, &count);
3638                 new_stack_entry = prev_entry->next;
3639                 if (new_stack_entry->end   != stack_entry->start  ||
3640                     new_stack_entry->start != addr)
3641                         panic ("Bad stack grow start/end in new stack entry");
3642                 else {
3643                         new_stack_entry->aux.avail_ssize =
3644                                 stack_entry->aux.avail_ssize -
3645                                 (new_stack_entry->end - new_stack_entry->start);
3646                         if (is_procstack)
3647                                 vm->vm_ssize += btoc(new_stack_entry->end -
3648                                                      new_stack_entry->start);
3649                 }
3650
3651                 if (map->flags & MAP_WIREFUTURE)
3652                         vm_map_unwire(map, new_stack_entry->start,
3653                                       new_stack_entry->end, FALSE);
3654         }
3655
3656 done:
3657         if (use_read_lock)
3658                 vm_map_unlock_read(map);
3659         else
3660                 vm_map_unlock(map);
3661         vm_map_entry_release(count);
3662         return (rv);
3663 }
3664
3665 /*
3666  * Unshare the specified VM space for exec.  If other processes are
3667  * mapped to it, then create a new one.  The new vmspace is null.
3668  *
3669  * No requirements.
3670  */
3671 void
3672 vmspace_exec(struct proc *p, struct vmspace *vmcopy) 
3673 {
3674         struct vmspace *oldvmspace = p->p_vmspace;
3675         struct vmspace *newvmspace;
3676         vm_map_t map = &p->p_vmspace->vm_map;
3677
3678         /*
3679          * If we are execing a resident vmspace we fork it, otherwise
3680          * we create a new vmspace.  Note that exitingcnt and upcalls
3681          * are not copied to the new vmspace.
3682          */
3683         lwkt_gettoken(&oldvmspace->vm_map.token);
3684         if (vmcopy)  {
3685                 newvmspace = vmspace_fork(vmcopy);
3686                 lwkt_gettoken(&newvmspace->vm_map.token);
3687         } else {
3688                 newvmspace = vmspace_alloc(map->min_offset, map->max_offset);
3689                 lwkt_gettoken(&newvmspace->vm_map.token);
3690                 bcopy(&oldvmspace->vm_startcopy, &newvmspace->vm_startcopy,
3691                       (caddr_t)&oldvmspace->vm_endcopy -
3692                        (caddr_t)&oldvmspace->vm_startcopy);
3693         }
3694
3695         /*
3696          * Finish initializing the vmspace before assigning it
3697          * to the process.  The vmspace will become the current vmspace
3698          * if p == curproc.
3699          */
3700         pmap_pinit2(vmspace_pmap(newvmspace));
3701         pmap_replacevm(p, newvmspace, 0);
3702         lwkt_reltoken(&newvmspace->vm_map.token);
3703         lwkt_reltoken(&oldvmspace->vm_map.token);
3704         vmspace_free(oldvmspace);
3705 }
3706
3707 /*
3708  * Unshare the specified VM space for forcing COW.  This
3709  * is called by rfork, for the (RFMEM|RFPROC) == 0 case.
3710  */
3711 void
3712 vmspace_unshare(struct proc *p) 
3713 {
3714         struct vmspace *oldvmspace = p->p_vmspace;
3715         struct vmspace *newvmspace;
3716
3717         lwkt_gettoken(&oldvmspace->vm_map.token);
3718         if (oldvmspace->vm_sysref.refcnt == 1) {
3719                 lwkt_reltoken(&oldvmspace->vm_map.token);
3720                 return;
3721         }
3722         newvmspace = vmspace_fork(oldvmspace);
3723         lwkt_gettoken(&newvmspace->vm_map.token);
3724         pmap_pinit2(vmspace_pmap(newvmspace));
3725         pmap_replacevm(p, newvmspace, 0);
3726         lwkt_reltoken(&newvmspace->vm_map.token);
3727         lwkt_reltoken(&oldvmspace->vm_map.token);
3728         vmspace_free(oldvmspace);
3729 }
3730
3731 /*
3732  * vm_map_hint: return the beginning of the best area suitable for
3733  * creating a new mapping with "prot" protection.
3734  *
3735  * No requirements.
3736  */
3737 vm_offset_t
3738 vm_map_hint(struct proc *p, vm_offset_t addr, vm_prot_t prot)
3739 {
3740         struct vmspace *vms = p->p_vmspace;
3741
3742         if (!randomize_mmap) {
3743                 /*
3744                  * Set a reasonable start point for the hint if it was
3745                  * not specified or if it falls within the heap space.
3746                  * Hinted mmap()s do not allocate out of the heap space.
3747                  */
3748                 if (addr == 0 ||
3749                     (addr >= round_page((vm_offset_t)vms->vm_taddr) &&
3750                      addr < round_page((vm_offset_t)vms->vm_daddr + maxdsiz))) {
3751                         addr = round_page((vm_offset_t)vms->vm_daddr + maxdsiz);
3752                 }
3753
3754                 return addr;
3755         }
3756
3757         if (addr != 0 && addr >= (vm_offset_t)vms->vm_daddr)
3758                 return addr;
3759
3760 #ifdef notyet
3761 #ifdef __i386__
3762         /*
3763          * If executable skip first two pages, otherwise start
3764          * after data + heap region.
3765          */
3766         if ((prot & VM_PROT_EXECUTE) &&
3767             ((vm_offset_t)vms->vm_daddr >= I386_MAX_EXE_ADDR)) {
3768                 addr = (PAGE_SIZE * 2) +
3769                     (karc4random() & (I386_MAX_EXE_ADDR / 2 - 1));
3770                 return (round_page(addr));
3771         }
3772 #endif /* __i386__ */
3773 #endif /* notyet */
3774
3775         addr = (vm_offset_t)vms->vm_daddr + MAXDSIZ;
3776         addr += karc4random() & (MIN((256 * 1024 * 1024), MAXDSIZ) - 1);
3777
3778         return (round_page(addr));
3779 }
3780
3781 /*
3782  * Finds the VM object, offset, and protection for a given virtual address
3783  * in the specified map, assuming a page fault of the type specified.
3784  *
3785  * Leaves the map in question locked for read; return values are guaranteed
3786  * until a vm_map_lookup_done call is performed.  Note that the map argument
3787  * is in/out; the returned map must be used in the call to vm_map_lookup_done.
3788  *
3789  * A handle (out_entry) is returned for use in vm_map_lookup_done, to make
3790  * that fast.
3791  *
3792  * If a lookup is requested with "write protection" specified, the map may
3793  * be changed to perform virtual copying operations, although the data
3794  * referenced will remain the same.
3795  *
3796  * No requirements.
3797  */
3798 int
3799 vm_map_lookup(vm_map_t *var_map,                /* IN/OUT */
3800               vm_offset_t vaddr,
3801               vm_prot_t fault_typea,
3802               vm_map_entry_t *out_entry,        /* OUT */
3803               vm_object_t *object,              /* OUT */
3804               vm_pindex_t *pindex,              /* OUT */
3805               vm_prot_t *out_prot,              /* OUT */
3806               boolean_t *wired)                 /* OUT */
3807 {
3808         vm_map_entry_t entry;
3809         vm_map_t map = *var_map;
3810         vm_prot_t prot;
3811         vm_prot_t fault_type = fault_typea;
3812         int use_read_lock = 1;
3813         int rv = KERN_SUCCESS;
3814
3815 RetryLookup:
3816         if (use_read_lock)
3817                 vm_map_lock_read(map);
3818         else
3819                 vm_map_lock(map);
3820
3821         /*
3822          * If the map has an interesting hint, try it before calling full
3823          * blown lookup routine.
3824          */
3825         entry = map->hint;
3826         cpu_ccfence();
3827         *out_entry = entry;
3828         *object = NULL;
3829
3830         if ((entry == &map->header) ||
3831             (vaddr < entry->start) || (vaddr >= entry->end)) {
3832                 vm_map_entry_t tmp_entry;
3833
3834                 /*
3835                  * Entry was either not a valid hint, or the vaddr was not
3836                  * contained in the entry, so do a full lookup.
3837                  */
3838                 if (!vm_map_lookup_entry(map, vaddr, &tmp_entry)) {
3839                         rv = KERN_INVALID_ADDRESS;
3840                         goto done;
3841                 }
3842
3843                 entry = tmp_entry;
3844                 *out_entry = entry;
3845         }
3846         
3847         /*
3848          * Handle submaps.
3849          */
3850         if (entry->maptype == VM_MAPTYPE_SUBMAP) {
3851                 vm_map_t old_map = map;
3852
3853                 *var_map = map = entry->object.sub_map;
3854                 if (use_read_lock)
3855                         vm_map_unlock_read(old_map);
3856                 else
3857                         vm_map_unlock(old_map);
3858                 use_read_lock = 1;
3859                 goto RetryLookup;
3860         }
3861
3862         /*
3863          * Check whether this task is allowed to have this page.
3864          * Note the special case for MAP_ENTRY_COW
3865          * pages with an override.  This is to implement a forced
3866          * COW for debuggers.
3867          */
3868
3869         if (fault_type & VM_PROT_OVERRIDE_WRITE)
3870                 prot = entry->max_protection;
3871         else
3872                 prot = entry->protection;
3873
3874         fault_type &= (VM_PROT_READ|VM_PROT_WRITE|VM_PROT_EXECUTE);
3875         if ((fault_type & prot) != fault_type) {
3876                 rv = KERN_PROTECTION_FAILURE;
3877                 goto done;
3878         }
3879
3880         if ((entry->eflags & MAP_ENTRY_USER_WIRED) &&
3881             (entry->eflags & MAP_ENTRY_COW) &&
3882             (fault_type & VM_PROT_WRITE) &&
3883             (fault_typea & VM_PROT_OVERRIDE_WRITE) == 0) {
3884                 rv = KERN_PROTECTION_FAILURE;
3885                 goto done;
3886         }
3887
3888         /*
3889          * If this page is not pageable, we have to get it for all possible
3890          * accesses.
3891          */
3892         *wired = (entry->wired_count != 0);
3893         if (*wired)
3894                 prot = fault_type = entry->protection;
3895
3896         /*
3897          * Virtual page tables may need to update the accessed (A) bit
3898          * in a page table entry.  Upgrade the fault to a write fault for
3899          * that case if the map will support it.  If the map does not support
3900          * it the page table entry simply will not be updated.
3901          */
3902         if (entry->maptype == VM_MAPTYPE_VPAGETABLE) {
3903                 if (prot & VM_PROT_WRITE)
3904                         fault_type |= VM_PROT_WRITE;
3905         }
3906
3907         /*
3908          * If the entry was copy-on-write, we either ...
3909          */
3910         if (entry->eflags & MAP_ENTRY_NEEDS_COPY) {
3911                 /*
3912                  * If we want to write the page, we may as well handle that
3913                  * now since we've got the map locked.
3914                  *
3915                  * If we don't need to write the page, we just demote the
3916                  * permissions allowed.
3917                  */
3918
3919                 if (fault_type & VM_PROT_WRITE) {
3920                         /*
3921                          * Make a new object, and place it in the object
3922                          * chain.  Note that no new references have appeared
3923                          * -- one just moved from the map to the new
3924                          * object.
3925                          */
3926
3927                         if (use_read_lock && vm_map_lock_upgrade(map)) {
3928                                 /* lost lock */
3929                                 use_read_lock = 0;
3930                                 goto RetryLookup;
3931                         }
3932                         use_read_lock = 0;
3933
3934                         vm_map_entry_shadow(entry, 0);
3935                 } else {
3936                         /*
3937                          * We're attempting to read a copy-on-write page --
3938                          * don't allow writes.
3939                          */
3940
3941                         prot &= ~VM_PROT_WRITE;
3942                 }
3943         }
3944
3945         /*
3946          * Create an object if necessary.
3947          */
3948         if (entry->object.vm_object == NULL && !map->system_map) {
3949                 if (use_read_lock && vm_map_lock_upgrade(map))  {
3950                         /* lost lock */
3951                         use_read_lock = 0;
3952                         goto RetryLookup;
3953                 }
3954                 use_read_lock = 0;
3955                 vm_map_entry_allocate_object(entry);
3956         }
3957
3958         /*
3959          * Return the object/offset from this entry.  If the entry was
3960          * copy-on-write or empty, it has been fixed up.
3961          */
3962
3963         *pindex = OFF_TO_IDX((vaddr - entry->start) + entry->offset);
3964         *object = entry->object.vm_object;
3965
3966         /*
3967          * Return whether this is the only map sharing this data.  On
3968          * success we return with a read lock held on the map.  On failure
3969          * we return with the map unlocked.
3970          */
3971         *out_prot = prot;
3972 done:
3973         if (rv == KERN_SUCCESS) {
3974                 if (use_read_lock == 0)
3975                         vm_map_lock_downgrade(map);
3976         } else if (use_read_lock) {
3977                 vm_map_unlock_read(map);
3978         } else {
3979                 vm_map_unlock(map);
3980         }
3981         return (rv);
3982 }
3983
3984 /*
3985  * Releases locks acquired by a vm_map_lookup()
3986  * (according to the handle returned by that lookup).
3987  *
3988  * No other requirements.
3989  */
3990 void
3991 vm_map_lookup_done(vm_map_t map, vm_map_entry_t entry, int count)
3992 {
3993         /*
3994          * Unlock the main-level map
3995          */
3996         vm_map_unlock_read(map);
3997         if (count)
3998                 vm_map_entry_release(count);
3999 }
4000
4001 #include "opt_ddb.h"
4002 #ifdef DDB
4003 #include <sys/kernel.h>
4004
4005 #include <ddb/ddb.h>
4006
4007 /*
4008  * Debugging only
4009  */
4010 DB_SHOW_COMMAND(map, vm_map_print)
4011 {
4012         static int nlines;
4013         /* XXX convert args. */
4014         vm_map_t map = (vm_map_t)addr;
4015         boolean_t full = have_addr;
4016
4017         vm_map_entry_t entry;
4018
4019         db_iprintf("Task map %p: pmap=%p, nentries=%d, version=%u\n",
4020             (void *)map,
4021             (void *)map->pmap, map->nentries, map->timestamp);
4022         nlines++;
4023
4024         if (!full && db_indent)
4025                 return;
4026
4027         db_indent += 2;
4028         for (entry = map->header.next; entry != &map->header;
4029             entry = entry->next) {
4030                 db_iprintf("map entry %p: start=%p, end=%p\n",
4031                     (void *)entry, (void *)entry->start, (void *)entry->end);
4032                 nlines++;
4033                 {
4034                         static char *inheritance_name[4] =
4035                         {"share", "copy", "none", "donate_copy"};
4036
4037                         db_iprintf(" prot=%x/%x/%s",
4038                             entry->protection,
4039                             entry->max_protection,
4040                             inheritance_name[(int)(unsigned char)entry->inheritance]);
4041                         if (entry->wired_count != 0)
4042                                 db_printf(", wired");
4043                 }
4044                 if (entry->maptype == VM_MAPTYPE_SUBMAP) {
4045                         /* XXX no %qd in kernel.  Truncate entry->offset. */
4046                         db_printf(", share=%p, offset=0x%lx\n",
4047                             (void *)entry->object.sub_map,
4048                             (long)entry->offset);