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  */
68
69 /*
70  *      Virtual memory mapping module.
71  */
72
73 #include <sys/param.h>
74 #include <sys/systm.h>
75 #include <sys/kernel.h>
76 #include <sys/proc.h>
77 #include <sys/serialize.h>
78 #include <sys/lock.h>
79 #include <sys/vmmeter.h>
80 #include <sys/mman.h>
81 #include <sys/vnode.h>
82 #include <sys/resourcevar.h>
83 #include <sys/shm.h>
84 #include <sys/tree.h>
85 #include <sys/malloc.h>
86
87 #include <vm/vm.h>
88 #include <vm/vm_param.h>
89 #include <vm/pmap.h>
90 #include <vm/vm_map.h>
91 #include <vm/vm_page.h>
92 #include <vm/vm_object.h>
93 #include <vm/vm_pager.h>
94 #include <vm/vm_kern.h>
95 #include <vm/vm_extern.h>
96 #include <vm/swap_pager.h>
97 #include <vm/vm_zone.h>
98
99 #include <sys/thread2.h>
100 #include <sys/sysref2.h>
101 #include <sys/random.h>
102 #include <sys/sysctl.h>
103
104 /*
105  * Virtual memory maps provide for the mapping, protection, and sharing
106  * of virtual memory objects.  In addition, this module provides for an
107  * efficient virtual copy of memory from one map to another.
108  *
109  * Synchronization is required prior to most operations.
110  *
111  * Maps consist of an ordered doubly-linked list of simple entries.
112  * A hint and a RB tree is used to speed-up lookups.
113  *
114  * Callers looking to modify maps specify start/end addresses which cause
115  * the related map entry to be clipped if necessary, and then later
116  * recombined if the pieces remained compatible.
117  *
118  * Virtual copy operations are performed by copying VM object references
119  * from one map to another, and then marking both regions as copy-on-write.
120  */
121 static void vmspace_terminate(struct vmspace *vm);
122 static void vmspace_lock(struct vmspace *vm);
123 static void vmspace_unlock(struct vmspace *vm);
124 static void vmspace_dtor(void *obj, void *private);
125
126 MALLOC_DEFINE(M_VMSPACE, "vmspace", "vmspace objcache backingstore");
127
128 struct sysref_class vmspace_sysref_class = {
129         .name =         "vmspace",
130         .mtype =        M_VMSPACE,
131         .proto =        SYSREF_PROTO_VMSPACE,
132         .offset =       offsetof(struct vmspace, vm_sysref),
133         .objsize =      sizeof(struct vmspace),
134         .nom_cache =    32,
135         .flags = SRC_MANAGEDINIT,
136         .dtor = vmspace_dtor,
137         .ops = {
138                 .terminate = (sysref_terminate_func_t)vmspace_terminate,
139                 .lock = (sysref_lock_func_t)vmspace_lock,
140                 .unlock = (sysref_lock_func_t)vmspace_unlock
141         }
142 };
143
144 /*
145  * per-cpu page table cross mappings are initialized in early boot
146  * and might require a considerable number of vm_map_entry structures.
147  */
148 #define VMEPERCPU       (MAXCPU+1)
149
150 static struct vm_zone mapentzone_store, mapzone_store;
151 static vm_zone_t mapentzone, mapzone;
152 static struct vm_object mapentobj, mapobj;
153
154 static struct vm_map_entry map_entry_init[MAX_MAPENT];
155 static struct vm_map_entry cpu_map_entry_init[MAXCPU][VMEPERCPU];
156 static struct vm_map map_init[MAX_KMAP];
157
158 static int randomize_mmap;
159 SYSCTL_INT(_vm, OID_AUTO, randomize_mmap, CTLFLAG_RW, &randomize_mmap, 0,
160     "Randomize mmap offsets");
161
162 static void vm_map_entry_shadow(vm_map_entry_t entry, int addref);
163 static vm_map_entry_t vm_map_entry_create(vm_map_t map, int *);
164 static void vm_map_entry_dispose (vm_map_t map, vm_map_entry_t entry, int *);
165 static void _vm_map_clip_end (vm_map_t, vm_map_entry_t, vm_offset_t, int *);
166 static void _vm_map_clip_start (vm_map_t, vm_map_entry_t, vm_offset_t, int *);
167 static void vm_map_entry_delete (vm_map_t, vm_map_entry_t, int *);
168 static void vm_map_entry_unwire (vm_map_t, vm_map_entry_t);
169 static void vm_map_copy_entry (vm_map_t, vm_map_t, vm_map_entry_t,
170                 vm_map_entry_t);
171 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);
172
173 /*
174  * Initialize the vm_map module.  Must be called before any other vm_map
175  * routines.
176  *
177  * Map and entry structures are allocated from the general purpose
178  * memory pool with some exceptions:
179  *
180  *      - The kernel map is allocated statically.
181  *      - Initial kernel map entries are allocated out of a static pool.
182  *
183  *      These restrictions are necessary since malloc() uses the
184  *      maps and requires map entries.
185  *
186  * Called from the low level boot code only.
187  */
188 void
189 vm_map_startup(void)
190 {
191         mapzone = &mapzone_store;
192         zbootinit(mapzone, "MAP", sizeof (struct vm_map),
193                 map_init, MAX_KMAP);
194         mapentzone = &mapentzone_store;
195         zbootinit(mapentzone, "MAP ENTRY", sizeof (struct vm_map_entry),
196                 map_entry_init, MAX_MAPENT);
197 }
198
199 /*
200  * Called prior to any vmspace allocations.
201  *
202  * Called from the low level boot code only.
203  */
204 void
205 vm_init2(void) 
206 {
207         zinitna(mapentzone, &mapentobj, NULL, 0, 0, 
208                 ZONE_USE_RESERVE | ZONE_SPECIAL, 1);
209         zinitna(mapzone, &mapobj, NULL, 0, 0, 0, 1);
210         pmap_init2();
211         vm_object_init2();
212 }
213
214
215 /*
216  * Red black tree functions
217  *
218  * The caller must hold the related map lock.
219  */
220 static int rb_vm_map_compare(vm_map_entry_t a, vm_map_entry_t b);
221 RB_GENERATE(vm_map_rb_tree, vm_map_entry, rb_entry, rb_vm_map_compare);
222
223 /* a->start is address, and the only field has to be initialized */
224 static int
225 rb_vm_map_compare(vm_map_entry_t a, vm_map_entry_t b)
226 {
227         if (a->start < b->start)
228                 return(-1);
229         else if (a->start > b->start)
230                 return(1);
231         return(0);
232 }
233
234 /*
235  * Allocate a vmspace structure, including a vm_map and pmap.
236  * Initialize numerous fields.  While the initial allocation is zerod,
237  * subsequence reuse from the objcache leaves elements of the structure
238  * intact (particularly the pmap), so portions must be zerod.
239  *
240  * The structure is not considered activated until we call sysref_activate().
241  *
242  * No requirements.
243  */
244 struct vmspace *
245 vmspace_alloc(vm_offset_t min, vm_offset_t max)
246 {
247         struct vmspace *vm;
248
249         vm = sysref_alloc(&vmspace_sysref_class);
250         bzero(&vm->vm_startcopy,
251               (char *)&vm->vm_endcopy - (char *)&vm->vm_startcopy);
252         vm_map_init(&vm->vm_map, min, max, NULL);       /* initializes token */
253
254         /*
255          * Use a hold to prevent any additional racing hold from terminating
256          * the vmspace before we manage to activate it.  This also acquires
257          * the token for safety.
258          */
259         KKASSERT(vm->vm_holdcount == 0);
260         KKASSERT(vm->vm_exitingcnt == 0);
261         vmspace_hold(vm);
262         pmap_pinit(vmspace_pmap(vm));           /* (some fields reused) */
263         vm->vm_map.pmap = vmspace_pmap(vm);             /* XXX */
264         vm->vm_shm = NULL;
265         vm->vm_flags = 0;
266         cpu_vmspace_alloc(vm);
267         sysref_activate(&vm->vm_sysref);
268         vmspace_drop(vm);
269
270         return (vm);
271 }
272
273 /*
274  * Free a primary reference to a vmspace.  This can trigger a
275  * stage-1 termination.
276  */
277 void
278 vmspace_free(struct vmspace *vm)
279 {
280         /*
281          * We want all finalization to occur via vmspace_drop() so we
282          * need to hold the vm around the put.
283          */
284         vmspace_hold(vm);
285         sysref_put(&vm->vm_sysref);
286         vmspace_drop(vm);
287 }
288
289 void
290 vmspace_ref(struct vmspace *vm)
291 {
292         sysref_get(&vm->vm_sysref);
293 }
294
295 void
296 vmspace_hold(struct vmspace *vm)
297 {
298         refcount_acquire(&vm->vm_holdcount);
299         lwkt_gettoken(&vm->vm_map.token);
300 }
301
302 void
303 vmspace_drop(struct vmspace *vm)
304 {
305         lwkt_reltoken(&vm->vm_map.token);
306         if (refcount_release(&vm->vm_holdcount)) {
307                 if (vm->vm_exitingcnt == 0 &&
308                     sysref_isinactive(&vm->vm_sysref)) {
309                         vmspace_terminate(vm);
310                 }
311         }
312 }
313
314 /*
315  * dtor function - Some elements of the pmap are retained in the
316  * free-cached vmspaces to improve performance.  We have to clean them up
317  * here before returning the vmspace to the memory pool.
318  *
319  * No requirements.
320  */
321 static void
322 vmspace_dtor(void *obj, void *private)
323 {
324         struct vmspace *vm = obj;
325
326         pmap_puninit(vmspace_pmap(vm));
327 }
328
329 /*
330  * Called in three cases:
331  *
332  * (1) When the last sysref is dropped and the vmspace becomes inactive.
333  *     (holdcount will not be 0 because the vmspace is held through the op)
334  *
335  * (2) When exitingcount becomes 0 on the last reap
336  *     (holdcount will not be 0 because the vmspace is held through the op)
337  *
338  * (3) When the holdcount becomes 0 in addition to the above two
339  *
340  * sysref will not scrap the object until we call sysref_put() once more
341  * after the last ref has been dropped.
342  *
343  * VMSPACE_EXIT1 flags the primary deactivation
344  * VMSPACE_EXIT2 flags the last reap
345  */
346 static void
347 vmspace_terminate(struct vmspace *vm)
348 {
349         int count;
350
351         /*
352          *
353          */
354         lwkt_gettoken(&vm->vm_map.token);
355         if ((vm->vm_flags & VMSPACE_EXIT1) == 0) {
356                 vm->vm_flags |= VMSPACE_EXIT1;
357                 shmexit(vm);
358                 pmap_remove_pages(vmspace_pmap(vm), VM_MIN_USER_ADDRESS,
359                                   VM_MAX_USER_ADDRESS);
360                 vm_map_remove(&vm->vm_map, VM_MIN_USER_ADDRESS,
361                               VM_MAX_USER_ADDRESS);
362         }
363         if ((vm->vm_flags & VMSPACE_EXIT2) == 0 && vm->vm_exitingcnt == 0) {
364                 vm->vm_flags |= VMSPACE_EXIT2;
365                 cpu_vmspace_free(vm);
366                 shmexit(vm);
367                 KKASSERT(vm->vm_upcalls == NULL);
368
369                 /*
370                  * Lock the map, to wait out all other references to it.
371                  * Delete all of the mappings and pages they hold, then call
372                  * the pmap module to reclaim anything left.
373                  */
374                 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
375                 vm_map_lock(&vm->vm_map);
376                 vm_map_delete(&vm->vm_map, vm->vm_map.min_offset,
377                               vm->vm_map.max_offset, &count);
378                 vm_map_unlock(&vm->vm_map);
379                 vm_map_entry_release(count);
380
381                 lwkt_gettoken(&vmspace_pmap(vm)->pm_token);
382                 pmap_release(vmspace_pmap(vm));
383                 lwkt_reltoken(&vmspace_pmap(vm)->pm_token);
384         }
385
386         lwkt_reltoken(&vm->vm_map.token);
387         if (vm->vm_exitingcnt == 0 && vm->vm_holdcount == 0) {
388                 KKASSERT(vm->vm_flags & VMSPACE_EXIT1);
389                 KKASSERT(vm->vm_flags & VMSPACE_EXIT2);
390                 sysref_put(&vm->vm_sysref);
391         }
392 }
393
394 /*
395  * vmspaces are not currently locked.
396  */
397 static void
398 vmspace_lock(struct vmspace *vm __unused)
399 {
400 }
401
402 static void
403 vmspace_unlock(struct vmspace *vm __unused)
404 {
405 }
406
407 /*
408  * This is called during exit indicating that the vmspace is no
409  * longer in used by an exiting process, but the process has not yet
410  * been reaped.
411  *
412  * No requirements.
413  */
414 void
415 vmspace_exitbump(struct vmspace *vm)
416 {
417         vmspace_hold(vm);
418         ++vm->vm_exitingcnt;
419         vmspace_drop(vm);       /* handles termination sequencing */
420 }
421
422 /*
423  * Decrement the exitingcnt and issue the stage-2 termination if it becomes
424  * zero and the stage1 termination has already occured.
425  *
426  * No requirements.
427  */
428 void
429 vmspace_exitfree(struct proc *p)
430 {
431         struct vmspace *vm;
432
433         vm = p->p_vmspace;
434         p->p_vmspace = NULL;
435         vmspace_hold(vm);
436         KKASSERT(vm->vm_exitingcnt > 0);
437         if (--vm->vm_exitingcnt == 0 && sysref_isinactive(&vm->vm_sysref))
438                 vmspace_terminate(vm);
439         vmspace_drop(vm);       /* handles termination sequencing */
440 }
441
442 /*
443  * Swap useage is determined by taking the proportional swap used by
444  * VM objects backing the VM map.  To make up for fractional losses,
445  * if the VM object has any swap use at all the associated map entries
446  * count for at least 1 swap page.
447  *
448  * No requirements.
449  */
450 int
451 vmspace_swap_count(struct vmspace *vm)
452 {
453         vm_map_t map = &vm->vm_map;
454         vm_map_entry_t cur;
455         vm_object_t object;
456         int count = 0;
457         int n;
458
459         vmspace_hold(vm);
460         for (cur = map->header.next; cur != &map->header; cur = cur->next) {
461                 switch(cur->maptype) {
462                 case VM_MAPTYPE_NORMAL:
463                 case VM_MAPTYPE_VPAGETABLE:
464                         if ((object = cur->object.vm_object) == NULL)
465                                 break;
466                         if (object->swblock_count) {
467                                 n = (cur->end - cur->start) / PAGE_SIZE;
468                                 count += object->swblock_count *
469                                     SWAP_META_PAGES * n / object->size + 1;
470                         }
471                         break;
472                 default:
473                         break;
474                 }
475         }
476         vmspace_drop(vm);
477
478         return(count);
479 }
480
481 /*
482  * Calculate the approximate number of anonymous pages in use by
483  * this vmspace.  To make up for fractional losses, we count each
484  * VM object as having at least 1 anonymous page.
485  *
486  * No requirements.
487  */
488 int
489 vmspace_anonymous_count(struct vmspace *vm)
490 {
491         vm_map_t map = &vm->vm_map;
492         vm_map_entry_t cur;
493         vm_object_t object;
494         int count = 0;
495
496         vmspace_hold(vm);
497         for (cur = map->header.next; cur != &map->header; cur = cur->next) {
498                 switch(cur->maptype) {
499                 case VM_MAPTYPE_NORMAL:
500                 case VM_MAPTYPE_VPAGETABLE:
501                         if ((object = cur->object.vm_object) == NULL)
502                                 break;
503                         if (object->type != OBJT_DEFAULT &&
504                             object->type != OBJT_SWAP) {
505                                 break;
506                         }
507                         count += object->resident_page_count;
508                         break;
509                 default:
510                         break;
511                 }
512         }
513         vmspace_drop(vm);
514
515         return(count);
516 }
517
518 /*
519  * Creates and returns a new empty VM map with the given physical map
520  * structure, and having the given lower and upper address bounds.
521  *
522  * No requirements.
523  */
524 vm_map_t
525 vm_map_create(vm_map_t result, pmap_t pmap, vm_offset_t min, vm_offset_t max)
526 {
527         if (result == NULL)
528                 result = zalloc(mapzone);
529         vm_map_init(result, min, max, pmap);
530         return (result);
531 }
532
533 /*
534  * Initialize an existing vm_map structure such as that in the vmspace
535  * structure.  The pmap is initialized elsewhere.
536  *
537  * No requirements.
538  */
539 void
540 vm_map_init(struct vm_map *map, vm_offset_t min, vm_offset_t max, pmap_t pmap)
541 {
542         map->header.next = map->header.prev = &map->header;
543         RB_INIT(&map->rb_root);
544         map->nentries = 0;
545         map->size = 0;
546         map->system_map = 0;
547         map->min_offset = min;
548         map->max_offset = max;
549         map->pmap = pmap;
550         map->first_free = &map->header;
551         map->hint = &map->header;
552         map->timestamp = 0;
553         map->flags = 0;
554         lwkt_token_init(&map->token, "vm_map");
555         lockinit(&map->lock, "thrd_sleep", (hz + 9) / 10, 0);
556         TUNABLE_INT("vm.cache_vmspaces", &vmspace_sysref_class.nom_cache);
557 }
558
559 /*
560  * Shadow the vm_map_entry's object.  This typically needs to be done when
561  * a write fault is taken on an entry which had previously been cloned by
562  * fork().  The shared object (which might be NULL) must become private so
563  * we add a shadow layer above it.
564  *
565  * Object allocation for anonymous mappings is defered as long as possible.
566  * When creating a shadow, however, the underlying object must be instantiated
567  * so it can be shared.
568  *
569  * If the map segment is governed by a virtual page table then it is
570  * possible to address offsets beyond the mapped area.  Just allocate
571  * a maximally sized object for this case.
572  *
573  * The vm_map must be exclusively locked.
574  * No other requirements.
575  */
576 static
577 void
578 vm_map_entry_shadow(vm_map_entry_t entry, int addref)
579 {
580         if (entry->maptype == VM_MAPTYPE_VPAGETABLE) {
581                 vm_object_shadow(&entry->object.vm_object, &entry->offset,
582                                  0x7FFFFFFF, addref);   /* XXX */
583         } else {
584                 vm_object_shadow(&entry->object.vm_object, &entry->offset,
585                                  atop(entry->end - entry->start), addref);
586         }
587         entry->eflags &= ~MAP_ENTRY_NEEDS_COPY;
588 }
589
590 /*
591  * Allocate an object for a vm_map_entry.
592  *
593  * Object allocation for anonymous mappings is defered as long as possible.
594  * This function is called when we can defer no longer, generally when a map
595  * entry might be split or forked or takes a page fault.
596  *
597  * If the map segment is governed by a virtual page table then it is
598  * possible to address offsets beyond the mapped area.  Just allocate
599  * a maximally sized object for this case.
600  *
601  * The vm_map must be exclusively locked.
602  * No other requirements.
603  */
604 void 
605 vm_map_entry_allocate_object(vm_map_entry_t entry)
606 {
607         vm_object_t obj;
608
609         if (entry->maptype == VM_MAPTYPE_VPAGETABLE) {
610                 obj = vm_object_allocate(OBJT_DEFAULT, 0x7FFFFFFF); /* XXX */
611         } else {
612                 obj = vm_object_allocate(OBJT_DEFAULT,
613                                          atop(entry->end - entry->start));
614         }
615         entry->object.vm_object = obj;
616         entry->offset = 0;
617 }
618
619 /*
620  * Set an initial negative count so the first attempt to reserve
621  * space preloads a bunch of vm_map_entry's for this cpu.  Also
622  * pre-allocate 2 vm_map_entries which will be needed by zalloc() to
623  * map a new page for vm_map_entry structures.  SMP systems are
624  * particularly sensitive.
625  *
626  * This routine is called in early boot so we cannot just call
627  * vm_map_entry_reserve().
628  *
629  * Called from the low level boot code only (for each cpu)
630  */
631 void
632 vm_map_entry_reserve_cpu_init(globaldata_t gd)
633 {
634         vm_map_entry_t entry;
635         int i;
636
637         gd->gd_vme_avail -= MAP_RESERVE_COUNT * 2;
638         entry = &cpu_map_entry_init[gd->gd_cpuid][0];
639         for (i = 0; i < VMEPERCPU; ++i, ++entry) {
640                 entry->next = gd->gd_vme_base;
641                 gd->gd_vme_base = entry;
642         }
643 }
644
645 /*
646  * Reserves vm_map_entry structures so code later on can manipulate
647  * map_entry structures within a locked map without blocking trying
648  * to allocate a new vm_map_entry.
649  *
650  * No requirements.
651  */
652 int
653 vm_map_entry_reserve(int count)
654 {
655         struct globaldata *gd = mycpu;
656         vm_map_entry_t entry;
657
658         /*
659          * Make sure we have enough structures in gd_vme_base to handle
660          * the reservation request.
661          *
662          * The critical section protects access to the per-cpu gd.
663          */
664         crit_enter();
665         while (gd->gd_vme_avail < count) {
666                 entry = zalloc(mapentzone);
667                 entry->next = gd->gd_vme_base;
668                 gd->gd_vme_base = entry;
669                 ++gd->gd_vme_avail;
670         }
671         gd->gd_vme_avail -= count;
672         crit_exit();
673
674         return(count);
675 }
676
677 /*
678  * Releases previously reserved vm_map_entry structures that were not
679  * used.  If we have too much junk in our per-cpu cache clean some of
680  * it out.
681  *
682  * No requirements.
683  */
684 void
685 vm_map_entry_release(int count)
686 {
687         struct globaldata *gd = mycpu;
688         vm_map_entry_t entry;
689
690         crit_enter();
691         gd->gd_vme_avail += count;
692         while (gd->gd_vme_avail > MAP_RESERVE_SLOP) {
693                 entry = gd->gd_vme_base;
694                 KKASSERT(entry != NULL);
695                 gd->gd_vme_base = entry->next;
696                 --gd->gd_vme_avail;
697                 crit_exit();
698                 zfree(mapentzone, entry);
699                 crit_enter();
700         }
701         crit_exit();
702 }
703
704 /*
705  * Reserve map entry structures for use in kernel_map itself.  These
706  * entries have *ALREADY* been reserved on a per-cpu basis when the map
707  * was inited.  This function is used by zalloc() to avoid a recursion
708  * when zalloc() itself needs to allocate additional kernel memory.
709  *
710  * This function works like the normal reserve but does not load the
711  * vm_map_entry cache (because that would result in an infinite
712  * recursion).  Note that gd_vme_avail may go negative.  This is expected.
713  *
714  * Any caller of this function must be sure to renormalize after
715  * potentially eating entries to ensure that the reserve supply
716  * remains intact.
717  *
718  * No requirements.
719  */
720 int
721 vm_map_entry_kreserve(int count)
722 {
723         struct globaldata *gd = mycpu;
724
725         crit_enter();
726         gd->gd_vme_avail -= count;
727         crit_exit();
728         KASSERT(gd->gd_vme_base != NULL,
729                 ("no reserved entries left, gd_vme_avail = %d",
730                 gd->gd_vme_avail));
731         return(count);
732 }
733
734 /*
735  * Release previously reserved map entries for kernel_map.  We do not
736  * attempt to clean up like the normal release function as this would
737  * cause an unnecessary (but probably not fatal) deep procedure call.
738  *
739  * No requirements.
740  */
741 void
742 vm_map_entry_krelease(int count)
743 {
744         struct globaldata *gd = mycpu;
745
746         crit_enter();
747         gd->gd_vme_avail += count;
748         crit_exit();
749 }
750
751 /*
752  * Allocates a VM map entry for insertion.  No entry fields are filled in.
753  *
754  * The entries should have previously been reserved.  The reservation count
755  * is tracked in (*countp).
756  *
757  * No requirements.
758  */
759 static vm_map_entry_t
760 vm_map_entry_create(vm_map_t map, int *countp)
761 {
762         struct globaldata *gd = mycpu;
763         vm_map_entry_t entry;
764
765         KKASSERT(*countp > 0);
766         --*countp;
767         crit_enter();
768         entry = gd->gd_vme_base;
769         KASSERT(entry != NULL, ("gd_vme_base NULL! count %d", *countp));
770         gd->gd_vme_base = entry->next;
771         crit_exit();
772
773         return(entry);
774 }
775
776 /*
777  * Dispose of a vm_map_entry that is no longer being referenced.
778  *
779  * No requirements.
780  */
781 static void
782 vm_map_entry_dispose(vm_map_t map, vm_map_entry_t entry, int *countp)
783 {
784         struct globaldata *gd = mycpu;
785
786         KKASSERT(map->hint != entry);
787         KKASSERT(map->first_free != entry);
788
789         ++*countp;
790         crit_enter();
791         entry->next = gd->gd_vme_base;
792         gd->gd_vme_base = entry;
793         crit_exit();
794 }
795
796
797 /*
798  * Insert/remove entries from maps.
799  *
800  * The related map must be exclusively locked.
801  * The caller must hold map->token
802  * No other requirements.
803  */
804 static __inline void
805 vm_map_entry_link(vm_map_t map,
806                   vm_map_entry_t after_where,
807                   vm_map_entry_t entry)
808 {
809         ASSERT_VM_MAP_LOCKED(map);
810
811         map->nentries++;
812         entry->prev = after_where;
813         entry->next = after_where->next;
814         entry->next->prev = entry;
815         after_where->next = entry;
816         if (vm_map_rb_tree_RB_INSERT(&map->rb_root, entry))
817                 panic("vm_map_entry_link: dup addr map %p ent %p", map, entry);
818 }
819
820 static __inline void
821 vm_map_entry_unlink(vm_map_t map,
822                     vm_map_entry_t entry)
823 {
824         vm_map_entry_t prev;
825         vm_map_entry_t next;
826
827         ASSERT_VM_MAP_LOCKED(map);
828
829         if (entry->eflags & MAP_ENTRY_IN_TRANSITION) {
830                 panic("vm_map_entry_unlink: attempt to mess with "
831                       "locked entry! %p", entry);
832         }
833         prev = entry->prev;
834         next = entry->next;
835         next->prev = prev;
836         prev->next = next;
837         vm_map_rb_tree_RB_REMOVE(&map->rb_root, entry);
838         map->nentries--;
839 }
840
841 /*
842  * Finds the map entry containing (or immediately preceding) the specified
843  * address in the given map.  The entry is returned in (*entry).
844  *
845  * The boolean result indicates whether the address is actually contained
846  * in the map.
847  *
848  * The related map must be locked.
849  * No other requirements.
850  */
851 boolean_t
852 vm_map_lookup_entry(vm_map_t map, vm_offset_t address, vm_map_entry_t *entry)
853 {
854         vm_map_entry_t tmp;
855         vm_map_entry_t last;
856
857         ASSERT_VM_MAP_LOCKED(map);
858 #if 0
859         /*
860          * XXX TEMPORARILY DISABLED.  For some reason our attempt to revive
861          * the hint code with the red-black lookup meets with system crashes
862          * and lockups.  We do not yet know why.
863          *
864          * It is possible that the problem is related to the setting
865          * of the hint during map_entry deletion, in the code specified
866          * at the GGG comment later on in this file.
867          *
868          * YYY More likely it's because this function can be called with
869          * a shared lock on the map, resulting in map->hint updates possibly
870          * racing.  Fixed now but untested.
871          */
872         /*
873          * Quickly check the cached hint, there's a good chance of a match.
874          */
875         tmp = map->hint;
876         cpu_ccfence();
877         if (tmp != &map->header) {
878                 if (address >= tmp->start && address < tmp->end) {
879                         *entry = tmp;
880                         return(TRUE);
881                 }
882         }
883 #endif
884
885         /*
886          * Locate the record from the top of the tree.  'last' tracks the
887          * closest prior record and is returned if no match is found, which
888          * in binary tree terms means tracking the most recent right-branch
889          * taken.  If there is no prior record, &map->header is returned.
890          */
891         last = &map->header;
892         tmp = RB_ROOT(&map->rb_root);
893
894         while (tmp) {
895                 if (address >= tmp->start) {
896                         if (address < tmp->end) {
897                                 *entry = tmp;
898                                 map->hint = tmp;
899                                 return(TRUE);
900                         }
901                         last = tmp;
902                         tmp = RB_RIGHT(tmp, rb_entry);
903                 } else {
904                         tmp = RB_LEFT(tmp, rb_entry);
905                 }
906         }
907         *entry = last;
908         return (FALSE);
909 }
910
911 /*
912  * Inserts the given whole VM object into the target map at the specified
913  * address range.  The object's size should match that of the address range.
914  *
915  * The map must be exclusively locked.
916  * The object must be held.
917  * The caller must have reserved sufficient vm_map_entry structures.
918  *
919  * If object is non-NULL, ref count must be bumped by caller prior to
920  * making call to account for the new entry.
921  */
922 int
923 vm_map_insert(vm_map_t map, int *countp,
924               vm_object_t object, vm_ooffset_t offset,
925               vm_offset_t start, vm_offset_t end,
926               vm_maptype_t maptype,
927               vm_prot_t prot, vm_prot_t max,
928               int cow)
929 {
930         vm_map_entry_t new_entry;
931         vm_map_entry_t prev_entry;
932         vm_map_entry_t temp_entry;
933         vm_eflags_t protoeflags;
934         int must_drop = 0;
935
936         ASSERT_VM_MAP_LOCKED(map);
937         if (object)
938                 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
939
940         /*
941          * Check that the start and end points are not bogus.
942          */
943         if ((start < map->min_offset) || (end > map->max_offset) ||
944             (start >= end))
945                 return (KERN_INVALID_ADDRESS);
946
947         /*
948          * Find the entry prior to the proposed starting address; if it's part
949          * of an existing entry, this range is bogus.
950          */
951         if (vm_map_lookup_entry(map, start, &temp_entry))
952                 return (KERN_NO_SPACE);
953
954         prev_entry = temp_entry;
955
956         /*
957          * Assert that the next entry doesn't overlap the end point.
958          */
959
960         if ((prev_entry->next != &map->header) &&
961             (prev_entry->next->start < end))
962                 return (KERN_NO_SPACE);
963
964         protoeflags = 0;
965
966         if (cow & MAP_COPY_ON_WRITE)
967                 protoeflags |= MAP_ENTRY_COW|MAP_ENTRY_NEEDS_COPY;
968
969         if (cow & MAP_NOFAULT) {
970                 protoeflags |= MAP_ENTRY_NOFAULT;
971
972                 KASSERT(object == NULL,
973                         ("vm_map_insert: paradoxical MAP_NOFAULT request"));
974         }
975         if (cow & MAP_DISABLE_SYNCER)
976                 protoeflags |= MAP_ENTRY_NOSYNC;
977         if (cow & MAP_DISABLE_COREDUMP)
978                 protoeflags |= MAP_ENTRY_NOCOREDUMP;
979         if (cow & MAP_IS_STACK)
980                 protoeflags |= MAP_ENTRY_STACK;
981         if (cow & MAP_IS_KSTACK)
982                 protoeflags |= MAP_ENTRY_KSTACK;
983
984         lwkt_gettoken(&map->token);
985
986         if (object) {
987                 /*
988                  * When object is non-NULL, it could be shared with another
989                  * process.  We have to set or clear OBJ_ONEMAPPING 
990                  * appropriately.
991                  */
992                 if ((object->ref_count > 1) || (object->shadow_count != 0)) {
993                         vm_object_clear_flag(object, OBJ_ONEMAPPING);
994                 }
995         }
996         else if ((prev_entry != &map->header) &&
997                  (prev_entry->eflags == protoeflags) &&
998                  (prev_entry->end == start) &&
999                  (prev_entry->wired_count == 0) &&
1000                  prev_entry->maptype == maptype &&
1001                  ((prev_entry->object.vm_object == NULL) ||
1002                   vm_object_coalesce(prev_entry->object.vm_object,
1003                                      OFF_TO_IDX(prev_entry->offset),
1004                                      (vm_size_t)(prev_entry->end - prev_entry->start),
1005                                      (vm_size_t)(end - prev_entry->end)))) {
1006                 /*
1007                  * We were able to extend the object.  Determine if we
1008                  * can extend the previous map entry to include the 
1009                  * new range as well.
1010                  */
1011                 if ((prev_entry->inheritance == VM_INHERIT_DEFAULT) &&
1012                     (prev_entry->protection == prot) &&
1013                     (prev_entry->max_protection == max)) {
1014                         map->size += (end - prev_entry->end);
1015                         prev_entry->end = end;
1016                         vm_map_simplify_entry(map, prev_entry, countp);
1017                         lwkt_reltoken(&map->token);
1018                         return (KERN_SUCCESS);
1019                 }
1020
1021                 /*
1022                  * If we can extend the object but cannot extend the
1023                  * map entry, we have to create a new map entry.  We
1024                  * must bump the ref count on the extended object to
1025                  * account for it.  object may be NULL.
1026                  */
1027                 object = prev_entry->object.vm_object;
1028                 offset = prev_entry->offset +
1029                         (prev_entry->end - prev_entry->start);
1030                 if (object) {
1031                         vm_object_hold(object);
1032                         vm_object_chain_wait(object);
1033                         vm_object_reference_locked(object);
1034                         must_drop = 1;
1035                 }
1036         }
1037
1038         /*
1039          * NOTE: if conditionals fail, object can be NULL here.  This occurs
1040          * in things like the buffer map where we manage kva but do not manage
1041          * backing objects.
1042          */
1043
1044         /*
1045          * Create a new entry
1046          */
1047
1048         new_entry = vm_map_entry_create(map, countp);
1049         new_entry->start = start;
1050         new_entry->end = end;
1051
1052         new_entry->maptype = maptype;
1053         new_entry->eflags = protoeflags;
1054         new_entry->object.vm_object = object;
1055         new_entry->offset = offset;
1056         new_entry->aux.master_pde = 0;
1057
1058         new_entry->inheritance = VM_INHERIT_DEFAULT;
1059         new_entry->protection = prot;
1060         new_entry->max_protection = max;
1061         new_entry->wired_count = 0;
1062
1063         /*
1064          * Insert the new entry into the list
1065          */
1066
1067         vm_map_entry_link(map, prev_entry, new_entry);
1068         map->size += new_entry->end - new_entry->start;
1069
1070         /*
1071          * Update the free space hint.  Entries cannot overlap.
1072          * An exact comparison is needed to avoid matching
1073          * against the map->header.
1074          */
1075         if ((map->first_free == prev_entry) &&
1076             (prev_entry->end == new_entry->start)) {
1077                 map->first_free = new_entry;
1078         }
1079
1080 #if 0
1081         /*
1082          * Temporarily removed to avoid MAP_STACK panic, due to
1083          * MAP_STACK being a huge hack.  Will be added back in
1084          * when MAP_STACK (and the user stack mapping) is fixed.
1085          */
1086         /*
1087          * It may be possible to simplify the entry
1088          */
1089         vm_map_simplify_entry(map, new_entry, countp);
1090 #endif
1091
1092         /*
1093          * Try to pre-populate the page table.  Mappings governed by virtual
1094          * page tables cannot be prepopulated without a lot of work, so
1095          * don't try.
1096          */
1097         if ((cow & (MAP_PREFAULT|MAP_PREFAULT_PARTIAL)) &&
1098             maptype != VM_MAPTYPE_VPAGETABLE) {
1099                 pmap_object_init_pt(map->pmap, start, prot,
1100                                     object, OFF_TO_IDX(offset), end - start,
1101                                     cow & MAP_PREFAULT_PARTIAL);
1102         }
1103         if (must_drop)
1104                 vm_object_drop(object);
1105
1106         lwkt_reltoken(&map->token);
1107         return (KERN_SUCCESS);
1108 }
1109
1110 /*
1111  * Find sufficient space for `length' bytes in the given map, starting at
1112  * `start'.  Returns 0 on success, 1 on no space.
1113  *
1114  * This function will returned an arbitrarily aligned pointer.  If no
1115  * particular alignment is required you should pass align as 1.  Note that
1116  * the map may return PAGE_SIZE aligned pointers if all the lengths used in
1117  * the map are a multiple of PAGE_SIZE, even if you pass a smaller align
1118  * argument.
1119  *
1120  * 'align' should be a power of 2 but is not required to be.
1121  *
1122  * The map must be exclusively locked.
1123  * No other requirements.
1124  */
1125 int
1126 vm_map_findspace(vm_map_t map, vm_offset_t start, vm_size_t length,
1127                  vm_size_t align, int flags, vm_offset_t *addr)
1128 {
1129         vm_map_entry_t entry, next;
1130         vm_offset_t end;
1131         vm_offset_t align_mask;
1132
1133         if (start < map->min_offset)
1134                 start = map->min_offset;
1135         if (start > map->max_offset)
1136                 return (1);
1137
1138         /*
1139          * If the alignment is not a power of 2 we will have to use
1140          * a mod/division, set align_mask to a special value.
1141          */
1142         if ((align | (align - 1)) + 1 != (align << 1))
1143                 align_mask = (vm_offset_t)-1;
1144         else
1145                 align_mask = align - 1;
1146
1147         /*
1148          * Look for the first possible address; if there's already something
1149          * at this address, we have to start after it.
1150          */
1151         if (start == map->min_offset) {
1152                 if ((entry = map->first_free) != &map->header)
1153                         start = entry->end;
1154         } else {
1155                 vm_map_entry_t tmp;
1156
1157                 if (vm_map_lookup_entry(map, start, &tmp))
1158                         start = tmp->end;
1159                 entry = tmp;
1160         }
1161
1162         /*
1163          * Look through the rest of the map, trying to fit a new region in the
1164          * gap between existing regions, or after the very last region.
1165          */
1166         for (;; start = (entry = next)->end) {
1167                 /*
1168                  * Adjust the proposed start by the requested alignment,
1169                  * be sure that we didn't wrap the address.
1170                  */
1171                 if (align_mask == (vm_offset_t)-1)
1172                         end = ((start + align - 1) / align) * align;
1173                 else
1174                         end = (start + align_mask) & ~align_mask;
1175                 if (end < start)
1176                         return (1);
1177                 start = end;
1178                 /*
1179                  * Find the end of the proposed new region.  Be sure we didn't
1180                  * go beyond the end of the map, or wrap around the address.
1181                  * Then check to see if this is the last entry or if the 
1182                  * proposed end fits in the gap between this and the next
1183                  * entry.
1184                  */
1185                 end = start + length;
1186                 if (end > map->max_offset || end < start)
1187                         return (1);
1188                 next = entry->next;
1189
1190                 /*
1191                  * If the next entry's start address is beyond the desired
1192                  * end address we may have found a good entry.
1193                  *
1194                  * If the next entry is a stack mapping we do not map into
1195                  * the stack's reserved space.
1196                  *
1197                  * XXX continue to allow mapping into the stack's reserved
1198                  * space if doing a MAP_STACK mapping inside a MAP_STACK
1199                  * mapping, for backwards compatibility.  But the caller
1200                  * really should use MAP_STACK | MAP_TRYFIXED if they
1201                  * want to do that.
1202                  */
1203                 if (next == &map->header)
1204                         break;
1205                 if (next->start >= end) {
1206                         if ((next->eflags & MAP_ENTRY_STACK) == 0)
1207                                 break;
1208                         if (flags & MAP_STACK)
1209                                 break;
1210                         if (next->start - next->aux.avail_ssize >= end)
1211                                 break;
1212                 }
1213         }
1214         map->hint = entry;
1215
1216         /*
1217          * Grow the kernel_map if necessary.  pmap_growkernel() will panic
1218          * if it fails.  The kernel_map is locked and nothing can steal
1219          * our address space if pmap_growkernel() blocks.
1220          *
1221          * NOTE: This may be unconditionally called for kldload areas on
1222          *       x86_64 because these do not bump kernel_vm_end (which would
1223          *       fill 128G worth of page tables!).  Therefore we must not
1224          *       retry.
1225          */
1226         if (map == &kernel_map) {
1227                 vm_offset_t kstop;
1228
1229                 kstop = round_page(start + length);
1230                 if (kstop > kernel_vm_end)
1231                         pmap_growkernel(start, kstop);
1232         }
1233         *addr = start;
1234         return (0);
1235 }
1236
1237 /*
1238  * vm_map_find finds an unallocated region in the target address map with
1239  * the given length and allocates it.  The search is defined to be first-fit
1240  * from the specified address; the region found is returned in the same
1241  * parameter.
1242  *
1243  * If object is non-NULL, ref count must be bumped by caller
1244  * prior to making call to account for the new entry.
1245  *
1246  * No requirements.  This function will lock the map temporarily.
1247  */
1248 int
1249 vm_map_find(vm_map_t map, vm_object_t object, vm_ooffset_t offset,
1250             vm_offset_t *addr,  vm_size_t length, vm_size_t align,
1251             boolean_t fitit,
1252             vm_maptype_t maptype,
1253             vm_prot_t prot, vm_prot_t max,
1254             int cow)
1255 {
1256         vm_offset_t start;
1257         int result;
1258         int count;
1259
1260         start = *addr;
1261
1262         count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
1263         vm_map_lock(map);
1264         if (object)
1265                 vm_object_hold(object);
1266         if (fitit) {
1267                 if (vm_map_findspace(map, start, length, align, 0, addr)) {
1268                         if (object)
1269                                 vm_object_drop(object);
1270                         vm_map_unlock(map);
1271                         vm_map_entry_release(count);
1272                         return (KERN_NO_SPACE);
1273                 }
1274                 start = *addr;
1275         }
1276         result = vm_map_insert(map, &count, object, offset,
1277                                start, start + length,
1278                                maptype,
1279                                prot, max,
1280                                cow);
1281         if (object)
1282                 vm_object_drop(object);
1283         vm_map_unlock(map);
1284         vm_map_entry_release(count);
1285
1286         return (result);
1287 }
1288
1289 /*
1290  * Simplify the given map entry by merging with either neighbor.  This
1291  * routine also has the ability to merge with both neighbors.
1292  *
1293  * This routine guarentees that the passed entry remains valid (though
1294  * possibly extended).  When merging, this routine may delete one or
1295  * both neighbors.  No action is taken on entries which have their
1296  * in-transition flag set.
1297  *
1298  * The map must be exclusively locked.
1299  */
1300 void
1301 vm_map_simplify_entry(vm_map_t map, vm_map_entry_t entry, int *countp)
1302 {
1303         vm_map_entry_t next, prev;
1304         vm_size_t prevsize, esize;
1305
1306         if (entry->eflags & MAP_ENTRY_IN_TRANSITION) {
1307                 ++mycpu->gd_cnt.v_intrans_coll;
1308                 return;
1309         }
1310
1311         if (entry->maptype == VM_MAPTYPE_SUBMAP)
1312                 return;
1313
1314         prev = entry->prev;
1315         if (prev != &map->header) {
1316                 prevsize = prev->end - prev->start;
1317                 if ( (prev->end == entry->start) &&
1318                      (prev->maptype == entry->maptype) &&
1319                      (prev->object.vm_object == entry->object.vm_object) &&
1320                      (!prev->object.vm_object ||
1321                         (prev->offset + prevsize == entry->offset)) &&
1322                      (prev->eflags == entry->eflags) &&
1323                      (prev->protection == entry->protection) &&
1324                      (prev->max_protection == entry->max_protection) &&
1325                      (prev->inheritance == entry->inheritance) &&
1326                      (prev->wired_count == entry->wired_count)) {
1327                         if (map->first_free == prev)
1328                                 map->first_free = entry;
1329                         if (map->hint == prev)
1330                                 map->hint = entry;
1331                         vm_map_entry_unlink(map, prev);
1332                         entry->start = prev->start;
1333                         entry->offset = prev->offset;
1334                         if (prev->object.vm_object)
1335                                 vm_object_deallocate(prev->object.vm_object);
1336                         vm_map_entry_dispose(map, prev, countp);
1337                 }
1338         }
1339
1340         next = entry->next;
1341         if (next != &map->header) {
1342                 esize = entry->end - entry->start;
1343                 if ((entry->end == next->start) &&
1344                     (next->maptype == entry->maptype) &&
1345                     (next->object.vm_object == entry->object.vm_object) &&
1346                      (!entry->object.vm_object ||
1347                         (entry->offset + esize == next->offset)) &&
1348                     (next->eflags == entry->eflags) &&
1349                     (next->protection == entry->protection) &&
1350                     (next->max_protection == entry->max_protection) &&
1351                     (next->inheritance == entry->inheritance) &&
1352                     (next->wired_count == entry->wired_count)) {
1353                         if (map->first_free == next)
1354                                 map->first_free = entry;
1355                         if (map->hint == next)
1356                                 map->hint = entry;
1357                         vm_map_entry_unlink(map, next);
1358                         entry->end = next->end;
1359                         if (next->object.vm_object)
1360                                 vm_object_deallocate(next->object.vm_object);
1361                         vm_map_entry_dispose(map, next, countp);
1362                 }
1363         }
1364 }
1365
1366 /*
1367  * Asserts that the given entry begins at or after the specified address.
1368  * If necessary, it splits the entry into two.
1369  */
1370 #define vm_map_clip_start(map, entry, startaddr, countp)                \
1371 {                                                                       \
1372         if (startaddr > entry->start)                                   \
1373                 _vm_map_clip_start(map, entry, startaddr, countp);      \
1374 }
1375
1376 /*
1377  * This routine is called only when it is known that the entry must be split.
1378  *
1379  * The map must be exclusively locked.
1380  */
1381 static void
1382 _vm_map_clip_start(vm_map_t map, vm_map_entry_t entry, vm_offset_t start,
1383                    int *countp)
1384 {
1385         vm_map_entry_t new_entry;
1386
1387         /*
1388          * Split off the front portion -- note that we must insert the new
1389          * entry BEFORE this one, so that this entry has the specified
1390          * starting address.
1391          */
1392
1393         vm_map_simplify_entry(map, entry, countp);
1394
1395         /*
1396          * If there is no object backing this entry, we might as well create
1397          * one now.  If we defer it, an object can get created after the map
1398          * is clipped, and individual objects will be created for the split-up
1399          * map.  This is a bit of a hack, but is also about the best place to
1400          * put this improvement.
1401          */
1402         if (entry->object.vm_object == NULL && !map->system_map) {
1403                 vm_map_entry_allocate_object(entry);
1404         }
1405
1406         new_entry = vm_map_entry_create(map, countp);
1407         *new_entry = *entry;
1408
1409         new_entry->end = start;
1410         entry->offset += (start - entry->start);
1411         entry->start = start;
1412
1413         vm_map_entry_link(map, entry->prev, new_entry);
1414
1415         switch(entry->maptype) {
1416         case VM_MAPTYPE_NORMAL:
1417         case VM_MAPTYPE_VPAGETABLE:
1418                 if (new_entry->object.vm_object) {
1419                         vm_object_hold(new_entry->object.vm_object);
1420                         vm_object_chain_wait(new_entry->object.vm_object);
1421                         vm_object_reference_locked(new_entry->object.vm_object);
1422                         vm_object_drop(new_entry->object.vm_object);
1423                 }
1424                 break;
1425         default:
1426                 break;
1427         }
1428 }
1429
1430 /*
1431  * Asserts that the given entry ends at or before the specified address.
1432  * If necessary, it splits the entry into two.
1433  *
1434  * The map must be exclusively locked.
1435  */
1436 #define vm_map_clip_end(map, entry, endaddr, countp)            \
1437 {                                                               \
1438         if (endaddr < entry->end)                               \
1439                 _vm_map_clip_end(map, entry, endaddr, countp);  \
1440 }
1441
1442 /*
1443  * This routine is called only when it is known that the entry must be split.
1444  *
1445  * The map must be exclusively locked.
1446  */
1447 static void
1448 _vm_map_clip_end(vm_map_t map, vm_map_entry_t entry, vm_offset_t end,
1449                  int *countp)
1450 {
1451         vm_map_entry_t new_entry;
1452
1453         /*
1454          * If there is no object backing this entry, we might as well create
1455          * one now.  If we defer it, an object can get created after the map
1456          * is clipped, and individual objects will be created for the split-up
1457          * map.  This is a bit of a hack, but is also about the best place to
1458          * put this improvement.
1459          */
1460
1461         if (entry->object.vm_object == NULL && !map->system_map) {
1462                 vm_map_entry_allocate_object(entry);
1463         }
1464
1465         /*
1466          * Create a new entry and insert it AFTER the specified entry
1467          */
1468
1469         new_entry = vm_map_entry_create(map, countp);
1470         *new_entry = *entry;
1471
1472         new_entry->start = entry->end = end;
1473         new_entry->offset += (end - entry->start);
1474
1475         vm_map_entry_link(map, entry, new_entry);
1476
1477         switch(entry->maptype) {
1478         case VM_MAPTYPE_NORMAL:
1479         case VM_MAPTYPE_VPAGETABLE:
1480                 if (new_entry->object.vm_object) {
1481                         vm_object_hold(new_entry->object.vm_object);
1482                         vm_object_chain_wait(new_entry->object.vm_object);
1483                         vm_object_reference_locked(new_entry->object.vm_object);
1484                         vm_object_drop(new_entry->object.vm_object);
1485                 }
1486                 break;
1487         default:
1488                 break;
1489         }
1490 }
1491
1492 /*
1493  * Asserts that the starting and ending region addresses fall within the
1494  * valid range for the map.
1495  */
1496 #define VM_MAP_RANGE_CHECK(map, start, end)     \
1497 {                                               \
1498         if (start < vm_map_min(map))            \
1499                 start = vm_map_min(map);        \
1500         if (end > vm_map_max(map))              \
1501                 end = vm_map_max(map);          \
1502         if (start > end)                        \
1503                 start = end;                    \
1504 }
1505
1506 /*
1507  * Used to block when an in-transition collison occurs.  The map
1508  * is unlocked for the sleep and relocked before the return.
1509  */
1510 void
1511 vm_map_transition_wait(vm_map_t map)
1512 {
1513         tsleep_interlock(map, 0);
1514         vm_map_unlock(map);
1515         tsleep(map, PINTERLOCKED, "vment", 0);
1516         vm_map_lock(map);
1517 }
1518
1519 /*
1520  * When we do blocking operations with the map lock held it is
1521  * possible that a clip might have occured on our in-transit entry,
1522  * requiring an adjustment to the entry in our loop.  These macros
1523  * help the pageable and clip_range code deal with the case.  The
1524  * conditional costs virtually nothing if no clipping has occured.
1525  */
1526
1527 #define CLIP_CHECK_BACK(entry, save_start)              \
1528     do {                                                \
1529             while (entry->start != save_start) {        \
1530                     entry = entry->prev;                \
1531                     KASSERT(entry != &map->header, ("bad entry clip")); \
1532             }                                           \
1533     } while(0)
1534
1535 #define CLIP_CHECK_FWD(entry, save_end)                 \
1536     do {                                                \
1537             while (entry->end != save_end) {            \
1538                     entry = entry->next;                \
1539                     KASSERT(entry != &map->header, ("bad entry clip")); \
1540             }                                           \
1541     } while(0)
1542
1543
1544 /*
1545  * Clip the specified range and return the base entry.  The
1546  * range may cover several entries starting at the returned base
1547  * and the first and last entry in the covering sequence will be
1548  * properly clipped to the requested start and end address.
1549  *
1550  * If no holes are allowed you should pass the MAP_CLIP_NO_HOLES
1551  * flag.
1552  *
1553  * The MAP_ENTRY_IN_TRANSITION flag will be set for the entries
1554  * covered by the requested range.
1555  *
1556  * The map must be exclusively locked on entry and will remain locked
1557  * on return. If no range exists or the range contains holes and you
1558  * specified that no holes were allowed, NULL will be returned.  This
1559  * routine may temporarily unlock the map in order avoid a deadlock when
1560  * sleeping.
1561  */
1562 static
1563 vm_map_entry_t
1564 vm_map_clip_range(vm_map_t map, vm_offset_t start, vm_offset_t end, 
1565                   int *countp, int flags)
1566 {
1567         vm_map_entry_t start_entry;
1568         vm_map_entry_t entry;
1569
1570         /*
1571          * Locate the entry and effect initial clipping.  The in-transition
1572          * case does not occur very often so do not try to optimize it.
1573          */
1574 again:
1575         if (vm_map_lookup_entry(map, start, &start_entry) == FALSE)
1576                 return (NULL);
1577         entry = start_entry;
1578         if (entry->eflags & MAP_ENTRY_IN_TRANSITION) {
1579                 entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP;
1580                 ++mycpu->gd_cnt.v_intrans_coll;
1581                 ++mycpu->gd_cnt.v_intrans_wait;
1582                 vm_map_transition_wait(map);
1583                 /*
1584                  * entry and/or start_entry may have been clipped while
1585                  * we slept, or may have gone away entirely.  We have
1586                  * to restart from the lookup.
1587                  */
1588                 goto again;
1589         }
1590
1591         /*
1592          * Since we hold an exclusive map lock we do not have to restart
1593          * after clipping, even though clipping may block in zalloc.
1594          */
1595         vm_map_clip_start(map, entry, start, countp);
1596         vm_map_clip_end(map, entry, end, countp);
1597         entry->eflags |= MAP_ENTRY_IN_TRANSITION;
1598
1599         /*
1600          * Scan entries covered by the range.  When working on the next
1601          * entry a restart need only re-loop on the current entry which
1602          * we have already locked, since 'next' may have changed.  Also,
1603          * even though entry is safe, it may have been clipped so we
1604          * have to iterate forwards through the clip after sleeping.
1605          */
1606         while (entry->next != &map->header && entry->next->start < end) {
1607                 vm_map_entry_t next = entry->next;
1608
1609                 if (flags & MAP_CLIP_NO_HOLES) {
1610                         if (next->start > entry->end) {
1611                                 vm_map_unclip_range(map, start_entry,
1612                                         start, entry->end, countp, flags);
1613                                 return(NULL);
1614                         }
1615                 }
1616
1617                 if (next->eflags & MAP_ENTRY_IN_TRANSITION) {
1618                         vm_offset_t save_end = entry->end;
1619                         next->eflags |= MAP_ENTRY_NEEDS_WAKEUP;
1620                         ++mycpu->gd_cnt.v_intrans_coll;
1621                         ++mycpu->gd_cnt.v_intrans_wait;
1622                         vm_map_transition_wait(map);
1623
1624                         /*
1625                          * clips might have occured while we blocked.
1626                          */
1627                         CLIP_CHECK_FWD(entry, save_end);
1628                         CLIP_CHECK_BACK(start_entry, start);
1629                         continue;
1630                 }
1631                 /*
1632                  * No restart necessary even though clip_end may block, we
1633                  * are holding the map lock.
1634                  */
1635                 vm_map_clip_end(map, next, end, countp);
1636                 next->eflags |= MAP_ENTRY_IN_TRANSITION;
1637                 entry = next;
1638         }
1639         if (flags & MAP_CLIP_NO_HOLES) {
1640                 if (entry->end != end) {
1641                         vm_map_unclip_range(map, start_entry,
1642                                 start, entry->end, countp, flags);
1643                         return(NULL);
1644                 }
1645         }
1646         return(start_entry);
1647 }
1648
1649 /*
1650  * Undo the effect of vm_map_clip_range().  You should pass the same
1651  * flags and the same range that you passed to vm_map_clip_range().
1652  * This code will clear the in-transition flag on the entries and
1653  * wake up anyone waiting.  This code will also simplify the sequence
1654  * and attempt to merge it with entries before and after the sequence.
1655  *
1656  * The map must be locked on entry and will remain locked on return.
1657  *
1658  * Note that you should also pass the start_entry returned by
1659  * vm_map_clip_range().  However, if you block between the two calls
1660  * with the map unlocked please be aware that the start_entry may
1661  * have been clipped and you may need to scan it backwards to find
1662  * the entry corresponding with the original start address.  You are
1663  * responsible for this, vm_map_unclip_range() expects the correct
1664  * start_entry to be passed to it and will KASSERT otherwise.
1665  */
1666 static
1667 void
1668 vm_map_unclip_range(vm_map_t map, vm_map_entry_t start_entry,
1669                     vm_offset_t start, vm_offset_t end,
1670                     int *countp, int flags)
1671 {
1672         vm_map_entry_t entry;
1673
1674         entry = start_entry;
1675
1676         KASSERT(entry->start == start, ("unclip_range: illegal base entry"));
1677         while (entry != &map->header && entry->start < end) {
1678                 KASSERT(entry->eflags & MAP_ENTRY_IN_TRANSITION,
1679                         ("in-transition flag not set during unclip on: %p",
1680                         entry));
1681                 KASSERT(entry->end <= end,
1682                         ("unclip_range: tail wasn't clipped"));
1683                 entry->eflags &= ~MAP_ENTRY_IN_TRANSITION;
1684                 if (entry->eflags & MAP_ENTRY_NEEDS_WAKEUP) {
1685                         entry->eflags &= ~MAP_ENTRY_NEEDS_WAKEUP;
1686                         wakeup(map);
1687                 }
1688                 entry = entry->next;
1689         }
1690
1691         /*
1692          * Simplification does not block so there is no restart case.
1693          */
1694         entry = start_entry;
1695         while (entry != &map->header && entry->start < end) {
1696                 vm_map_simplify_entry(map, entry, countp);
1697                 entry = entry->next;
1698         }
1699 }
1700
1701 /*
1702  * Mark the given range as handled by a subordinate map.
1703  *
1704  * This range must have been created with vm_map_find(), and no other
1705  * operations may have been performed on this range prior to calling
1706  * vm_map_submap().
1707  *
1708  * Submappings cannot be removed.
1709  *
1710  * No requirements.
1711  */
1712 int
1713 vm_map_submap(vm_map_t map, vm_offset_t start, vm_offset_t end, vm_map_t submap)
1714 {
1715         vm_map_entry_t entry;
1716         int result = KERN_INVALID_ARGUMENT;
1717         int count;
1718
1719         count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
1720         vm_map_lock(map);
1721
1722         VM_MAP_RANGE_CHECK(map, start, end);
1723
1724         if (vm_map_lookup_entry(map, start, &entry)) {
1725                 vm_map_clip_start(map, entry, start, &count);
1726         } else {
1727                 entry = entry->next;
1728         }
1729
1730         vm_map_clip_end(map, entry, end, &count);
1731
1732         if ((entry->start == start) && (entry->end == end) &&
1733             ((entry->eflags & MAP_ENTRY_COW) == 0) &&
1734             (entry->object.vm_object == NULL)) {
1735                 entry->object.sub_map = submap;
1736                 entry->maptype = VM_MAPTYPE_SUBMAP;
1737                 result = KERN_SUCCESS;
1738         }
1739         vm_map_unlock(map);
1740         vm_map_entry_release(count);
1741
1742         return (result);
1743 }
1744
1745 /*
1746  * Sets the protection of the specified address region in the target map. 
1747  * If "set_max" is specified, the maximum protection is to be set;
1748  * otherwise, only the current protection is affected.
1749  *
1750  * The protection is not applicable to submaps, but is applicable to normal
1751  * maps and maps governed by virtual page tables.  For example, when operating
1752  * on a virtual page table our protection basically controls how COW occurs
1753  * on the backing object, whereas the virtual page table abstraction itself
1754  * is an abstraction for userland.
1755  *
1756  * No requirements.
1757  */
1758 int
1759 vm_map_protect(vm_map_t map, vm_offset_t start, vm_offset_t end,
1760                vm_prot_t new_prot, boolean_t set_max)
1761 {
1762         vm_map_entry_t current;
1763         vm_map_entry_t entry;
1764         int count;
1765
1766         count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
1767         vm_map_lock(map);
1768
1769         VM_MAP_RANGE_CHECK(map, start, end);
1770
1771         if (vm_map_lookup_entry(map, start, &entry)) {
1772                 vm_map_clip_start(map, entry, start, &count);
1773         } else {
1774                 entry = entry->next;
1775         }
1776
1777         /*
1778          * Make a first pass to check for protection violations.
1779          */
1780         current = entry;
1781         while ((current != &map->header) && (current->start < end)) {
1782                 if (current->maptype == VM_MAPTYPE_SUBMAP) {
1783                         vm_map_unlock(map);
1784                         vm_map_entry_release(count);
1785                         return (KERN_INVALID_ARGUMENT);
1786                 }
1787                 if ((new_prot & current->max_protection) != new_prot) {
1788                         vm_map_unlock(map);
1789                         vm_map_entry_release(count);
1790                         return (KERN_PROTECTION_FAILURE);
1791                 }
1792                 current = current->next;
1793         }
1794
1795         /*
1796          * Go back and fix up protections. [Note that clipping is not
1797          * necessary the second time.]
1798          */
1799         current = entry;
1800
1801         while ((current != &map->header) && (current->start < end)) {
1802                 vm_prot_t old_prot;
1803
1804                 vm_map_clip_end(map, current, end, &count);
1805
1806                 old_prot = current->protection;
1807                 if (set_max) {
1808                         current->protection =
1809                             (current->max_protection = new_prot) &
1810                             old_prot;
1811                 } else {
1812                         current->protection = new_prot;
1813                 }
1814
1815                 /*
1816                  * Update physical map if necessary. Worry about copy-on-write
1817                  * here -- CHECK THIS XXX
1818                  */
1819
1820                 if (current->protection != old_prot) {
1821 #define MASK(entry)     (((entry)->eflags & MAP_ENTRY_COW) ? ~VM_PROT_WRITE : \
1822                                                         VM_PROT_ALL)
1823
1824                         pmap_protect(map->pmap, current->start,
1825                             current->end,
1826                             current->protection & MASK(current));
1827 #undef  MASK
1828                 }
1829
1830                 vm_map_simplify_entry(map, current, &count);
1831
1832                 current = current->next;
1833         }
1834
1835         vm_map_unlock(map);
1836         vm_map_entry_release(count);
1837         return (KERN_SUCCESS);
1838 }
1839
1840 /*
1841  * This routine traverses a processes map handling the madvise
1842  * system call.  Advisories are classified as either those effecting
1843  * the vm_map_entry structure, or those effecting the underlying
1844  * objects.
1845  *
1846  * The <value> argument is used for extended madvise calls.
1847  *
1848  * No requirements.
1849  */
1850 int
1851 vm_map_madvise(vm_map_t map, vm_offset_t start, vm_offset_t end,
1852                int behav, off_t value)
1853 {
1854         vm_map_entry_t current, entry;
1855         int modify_map = 0;
1856         int error = 0;
1857         int count;
1858
1859         /*
1860          * Some madvise calls directly modify the vm_map_entry, in which case
1861          * we need to use an exclusive lock on the map and we need to perform 
1862          * various clipping operations.  Otherwise we only need a read-lock
1863          * on the map.
1864          */
1865
1866         count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
1867
1868         switch(behav) {
1869         case MADV_NORMAL:
1870         case MADV_SEQUENTIAL:
1871         case MADV_RANDOM:
1872         case MADV_NOSYNC:
1873         case MADV_AUTOSYNC:
1874         case MADV_NOCORE:
1875         case MADV_CORE:
1876         case MADV_SETMAP:
1877         case MADV_INVAL:
1878                 modify_map = 1;
1879                 vm_map_lock(map);
1880                 break;
1881         case MADV_WILLNEED:
1882         case MADV_DONTNEED:
1883         case MADV_FREE:
1884                 vm_map_lock_read(map);
1885                 break;
1886         default:
1887                 vm_map_entry_release(count);
1888                 return (EINVAL);
1889         }
1890
1891         /*
1892          * Locate starting entry and clip if necessary.
1893          */
1894
1895         VM_MAP_RANGE_CHECK(map, start, end);
1896
1897         if (vm_map_lookup_entry(map, start, &entry)) {
1898                 if (modify_map)
1899                         vm_map_clip_start(map, entry, start, &count);
1900         } else {
1901                 entry = entry->next;
1902         }
1903
1904         if (modify_map) {
1905                 /*
1906                  * madvise behaviors that are implemented in the vm_map_entry.
1907                  *
1908                  * We clip the vm_map_entry so that behavioral changes are
1909                  * limited to the specified address range.
1910                  */
1911                 for (current = entry;
1912                      (current != &map->header) && (current->start < end);
1913                      current = current->next
1914                 ) {
1915                         if (current->maptype == VM_MAPTYPE_SUBMAP)
1916                                 continue;
1917
1918                         vm_map_clip_end(map, current, end, &count);
1919
1920                         switch (behav) {
1921                         case MADV_NORMAL:
1922                                 vm_map_entry_set_behavior(current, MAP_ENTRY_BEHAV_NORMAL);
1923                                 break;
1924                         case MADV_SEQUENTIAL:
1925                                 vm_map_entry_set_behavior(current, MAP_ENTRY_BEHAV_SEQUENTIAL);
1926                                 break;
1927                         case MADV_RANDOM:
1928                                 vm_map_entry_set_behavior(current, MAP_ENTRY_BEHAV_RANDOM);
1929                                 break;
1930                         case MADV_NOSYNC:
1931                                 current->eflags |= MAP_ENTRY_NOSYNC;
1932                                 break;
1933                         case MADV_AUTOSYNC:
1934                                 current->eflags &= ~MAP_ENTRY_NOSYNC;
1935                                 break;
1936                         case MADV_NOCORE:
1937                                 current->eflags |= MAP_ENTRY_NOCOREDUMP;
1938                                 break;
1939                         case MADV_CORE:
1940                                 current->eflags &= ~MAP_ENTRY_NOCOREDUMP;
1941                                 break;
1942                         case MADV_INVAL:
1943                                 /*
1944                                  * Invalidate the related pmap entries, used
1945                                  * to flush portions of the real kernel's
1946                                  * pmap when the caller has removed or
1947                                  * modified existing mappings in a virtual
1948                                  * page table.
1949                                  */
1950                                 pmap_remove(map->pmap,
1951                                             current->start, current->end);
1952                                 break;
1953                         case MADV_SETMAP:
1954                                 /*
1955                                  * Set the page directory page for a map
1956                                  * governed by a virtual page table.  Mark
1957                                  * the entry as being governed by a virtual
1958                                  * page table if it is not.
1959                                  *
1960                                  * XXX the page directory page is stored
1961                                  * in the avail_ssize field if the map_entry.
1962                                  *
1963                                  * XXX the map simplification code does not
1964                                  * compare this field so weird things may
1965                                  * happen if you do not apply this function
1966                                  * to the entire mapping governed by the
1967                                  * virtual page table.
1968                                  */
1969                                 if (current->maptype != VM_MAPTYPE_VPAGETABLE) {
1970                                         error = EINVAL;
1971                                         break;
1972                                 }
1973                                 current->aux.master_pde = value;
1974                                 pmap_remove(map->pmap,
1975                                             current->start, current->end);
1976                                 break;
1977                         default:
1978                                 error = EINVAL;
1979                                 break;
1980                         }
1981                         vm_map_simplify_entry(map, current, &count);
1982                 }
1983                 vm_map_unlock(map);
1984         } else {
1985                 vm_pindex_t pindex;
1986                 int count;
1987
1988                 /*
1989                  * madvise behaviors that are implemented in the underlying
1990                  * vm_object.
1991                  *
1992                  * Since we don't clip the vm_map_entry, we have to clip
1993                  * the vm_object pindex and count.
1994                  *
1995                  * NOTE!  We currently do not support these functions on
1996                  * virtual page tables.
1997                  */
1998                 for (current = entry;
1999                      (current != &map->header) && (current->start < end);
2000                      current = current->next
2001                 ) {
2002                         vm_offset_t useStart;
2003
2004                         if (current->maptype != VM_MAPTYPE_NORMAL)
2005                                 continue;
2006
2007                         pindex = OFF_TO_IDX(current->offset);
2008                         count = atop(current->end - current->start);
2009                         useStart = current->start;
2010
2011                         if (current->start < start) {
2012                                 pindex += atop(start - current->start);
2013                                 count -= atop(start - current->start);
2014                                 useStart = start;
2015                         }
2016                         if (current->end > end)
2017                                 count -= atop(current->end - end);
2018
2019                         if (count <= 0)
2020                                 continue;
2021
2022                         vm_object_madvise(current->object.vm_object,
2023                                           pindex, count, behav);
2024
2025                         /*
2026                          * Try to populate the page table.  Mappings governed
2027                          * by virtual page tables cannot be pre-populated
2028                          * without a lot of work so don't try.
2029                          */
2030                         if (behav == MADV_WILLNEED &&
2031                             current->maptype != VM_MAPTYPE_VPAGETABLE) {
2032                                 pmap_object_init_pt(
2033                                     map->pmap, 
2034                                     useStart,
2035                                     current->protection,
2036                                     current->object.vm_object,
2037                                     pindex, 
2038                                     (count << PAGE_SHIFT),
2039                                     MAP_PREFAULT_MADVISE
2040                                 );
2041                         }
2042                 }
2043                 vm_map_unlock_read(map);
2044         }
2045         vm_map_entry_release(count);
2046         return(error);
2047 }       
2048
2049
2050 /*
2051  * Sets the inheritance of the specified address range in the target map.
2052  * Inheritance affects how the map will be shared with child maps at the
2053  * time of vm_map_fork.
2054  */
2055 int
2056 vm_map_inherit(vm_map_t map, vm_offset_t start, vm_offset_t end,
2057                vm_inherit_t new_inheritance)
2058 {
2059         vm_map_entry_t entry;
2060         vm_map_entry_t temp_entry;
2061         int count;
2062
2063         switch (new_inheritance) {
2064         case VM_INHERIT_NONE:
2065         case VM_INHERIT_COPY:
2066         case VM_INHERIT_SHARE:
2067                 break;
2068         default:
2069                 return (KERN_INVALID_ARGUMENT);
2070         }
2071
2072         count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
2073         vm_map_lock(map);
2074
2075         VM_MAP_RANGE_CHECK(map, start, end);
2076
2077         if (vm_map_lookup_entry(map, start, &temp_entry)) {
2078                 entry = temp_entry;
2079                 vm_map_clip_start(map, entry, start, &count);
2080         } else
2081                 entry = temp_entry->next;
2082
2083         while ((entry != &map->header) && (entry->start < end)) {
2084                 vm_map_clip_end(map, entry, end, &count);
2085
2086                 entry->inheritance = new_inheritance;
2087
2088                 vm_map_simplify_entry(map, entry, &count);
2089
2090                 entry = entry->next;
2091         }
2092         vm_map_unlock(map);
2093         vm_map_entry_release(count);
2094         return (KERN_SUCCESS);
2095 }
2096
2097 /*
2098  * Implement the semantics of mlock
2099  */
2100 int
2101 vm_map_unwire(vm_map_t map, vm_offset_t start, vm_offset_t real_end,
2102               boolean_t new_pageable)
2103 {
2104         vm_map_entry_t entry;
2105         vm_map_entry_t start_entry;
2106         vm_offset_t end;
2107         int rv = KERN_SUCCESS;
2108         int count;
2109
2110         count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
2111         vm_map_lock(map);
2112         VM_MAP_RANGE_CHECK(map, start, real_end);
2113         end = real_end;
2114
2115         start_entry = vm_map_clip_range(map, start, end, &count,
2116                                         MAP_CLIP_NO_HOLES);
2117         if (start_entry == NULL) {
2118                 vm_map_unlock(map);
2119                 vm_map_entry_release(count);
2120                 return (KERN_INVALID_ADDRESS);
2121         }
2122
2123         if (new_pageable == 0) {
2124                 entry = start_entry;
2125                 while ((entry != &map->header) && (entry->start < end)) {
2126                         vm_offset_t save_start;
2127                         vm_offset_t save_end;
2128
2129                         /*
2130                          * Already user wired or hard wired (trivial cases)
2131                          */
2132                         if (entry->eflags & MAP_ENTRY_USER_WIRED) {
2133                                 entry = entry->next;
2134                                 continue;
2135                         }
2136                         if (entry->wired_count != 0) {
2137                                 entry->wired_count++;
2138                                 entry->eflags |= MAP_ENTRY_USER_WIRED;
2139                                 entry = entry->next;
2140                                 continue;
2141                         }
2142
2143                         /*
2144                          * A new wiring requires instantiation of appropriate
2145                          * management structures and the faulting in of the
2146                          * page.
2147                          */
2148                         if (entry->maptype != VM_MAPTYPE_SUBMAP) {
2149                                 int copyflag = entry->eflags &
2150                                                MAP_ENTRY_NEEDS_COPY;
2151                                 if (copyflag && ((entry->protection &
2152                                                   VM_PROT_WRITE) != 0)) {
2153                                         vm_map_entry_shadow(entry, 0);
2154                                 } else if (entry->object.vm_object == NULL &&
2155                                            !map->system_map) {
2156                                         vm_map_entry_allocate_object(entry);
2157                                 }
2158                         }
2159                         entry->wired_count++;
2160                         entry->eflags |= MAP_ENTRY_USER_WIRED;
2161
2162                         /*
2163                          * Now fault in the area.  Note that vm_fault_wire()
2164                          * may release the map lock temporarily, it will be
2165                          * relocked on return.  The in-transition
2166                          * flag protects the entries. 
2167                          */
2168                         save_start = entry->start;
2169                         save_end = entry->end;
2170                         rv = vm_fault_wire(map, entry, TRUE);
2171                         if (rv) {
2172                                 CLIP_CHECK_BACK(entry, save_start);
2173                                 for (;;) {
2174                                         KASSERT(entry->wired_count == 1, ("bad wired_count on entry"));
2175                                         entry->eflags &= ~MAP_ENTRY_USER_WIRED;
2176                                         entry->wired_count = 0;
2177                                         if (entry->end == save_end)
2178                                                 break;
2179                                         entry = entry->next;
2180                                         KASSERT(entry != &map->header, ("bad entry clip during backout"));
2181                                 }
2182                                 end = save_start;       /* unwire the rest */
2183                                 break;
2184                         }
2185                         /*
2186                          * note that even though the entry might have been
2187                          * clipped, the USER_WIRED flag we set prevents
2188                          * duplication so we do not have to do a 
2189                          * clip check.
2190                          */
2191                         entry = entry->next;
2192                 }
2193
2194                 /*
2195                  * If we failed fall through to the unwiring section to
2196                  * unwire what we had wired so far.  'end' has already
2197                  * been adjusted.
2198                  */
2199                 if (rv)
2200                         new_pageable = 1;
2201
2202                 /*
2203                  * start_entry might have been clipped if we unlocked the
2204                  * map and blocked.  No matter how clipped it has gotten
2205                  * there should be a fragment that is on our start boundary.
2206                  */
2207                 CLIP_CHECK_BACK(start_entry, start);
2208         }
2209
2210         /*
2211          * Deal with the unwiring case.
2212          */
2213         if (new_pageable) {
2214                 /*
2215                  * This is the unwiring case.  We must first ensure that the
2216                  * range to be unwired is really wired down.  We know there
2217                  * are no holes.
2218                  */
2219                 entry = start_entry;
2220                 while ((entry != &map->header) && (entry->start < end)) {
2221                         if ((entry->eflags & MAP_ENTRY_USER_WIRED) == 0) {
2222                                 rv = KERN_INVALID_ARGUMENT;
2223                                 goto done;
2224                         }
2225                         KASSERT(entry->wired_count != 0, ("wired count was 0 with USER_WIRED set! %p", entry));
2226                         entry = entry->next;
2227                 }
2228
2229                 /*
2230                  * Now decrement the wiring count for each region. If a region
2231                  * becomes completely unwired, unwire its physical pages and
2232                  * mappings.
2233                  */
2234                 /*
2235                  * The map entries are processed in a loop, checking to
2236                  * make sure the entry is wired and asserting it has a wired
2237                  * count. However, another loop was inserted more-or-less in
2238                  * the middle of the unwiring path. This loop picks up the
2239                  * "entry" loop variable from the first loop without first
2240                  * setting it to start_entry. Naturally, the secound loop
2241                  * is never entered and the pages backing the entries are
2242                  * never unwired. This can lead to a leak of wired pages.
2243                  */
2244                 entry = start_entry;
2245                 while ((entry != &map->header) && (entry->start < end)) {
2246                         KASSERT(entry->eflags & MAP_ENTRY_USER_WIRED,
2247                                 ("expected USER_WIRED on entry %p", entry));
2248                         entry->eflags &= ~MAP_ENTRY_USER_WIRED;
2249                         entry->wired_count--;
2250                         if (entry->wired_count == 0)
2251                                 vm_fault_unwire(map, entry);
2252                         entry = entry->next;
2253                 }
2254         }
2255 done:
2256         vm_map_unclip_range(map, start_entry, start, real_end, &count,
2257                 MAP_CLIP_NO_HOLES);
2258         map->timestamp++;
2259         vm_map_unlock(map);
2260         vm_map_entry_release(count);
2261         return (rv);
2262 }
2263
2264 /*
2265  * Sets the pageability of the specified address range in the target map.
2266  * Regions specified as not pageable require locked-down physical
2267  * memory and physical page maps.
2268  *
2269  * The map must not be locked, but a reference must remain to the map
2270  * throughout the call.
2271  *
2272  * This function may be called via the zalloc path and must properly
2273  * reserve map entries for kernel_map.
2274  *
2275  * No requirements.
2276  */
2277 int
2278 vm_map_wire(vm_map_t map, vm_offset_t start, vm_offset_t real_end, int kmflags)
2279 {
2280         vm_map_entry_t entry;
2281         vm_map_entry_t start_entry;
2282         vm_offset_t end;
2283         int rv = KERN_SUCCESS;
2284         int count;
2285
2286         if (kmflags & KM_KRESERVE)
2287                 count = vm_map_entry_kreserve(MAP_RESERVE_COUNT);
2288         else
2289                 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
2290         vm_map_lock(map);
2291         VM_MAP_RANGE_CHECK(map, start, real_end);
2292         end = real_end;
2293
2294         start_entry = vm_map_clip_range(map, start, end, &count,
2295                                         MAP_CLIP_NO_HOLES);
2296         if (start_entry == NULL) {
2297                 vm_map_unlock(map);
2298                 rv = KERN_INVALID_ADDRESS;
2299                 goto failure;
2300         }
2301         if ((kmflags & KM_PAGEABLE) == 0) {
2302                 /*
2303                  * Wiring.  
2304                  *
2305                  * 1.  Holding the write lock, we create any shadow or zero-fill
2306                  * objects that need to be created. Then we clip each map
2307                  * entry to the region to be wired and increment its wiring
2308                  * count.  We create objects before clipping the map entries
2309                  * to avoid object proliferation.
2310                  *
2311                  * 2.  We downgrade to a read lock, and call vm_fault_wire to
2312                  * fault in the pages for any newly wired area (wired_count is
2313                  * 1).
2314                  *
2315                  * Downgrading to a read lock for vm_fault_wire avoids a 
2316                  * possible deadlock with another process that may have faulted
2317                  * on one of the pages to be wired (it would mark the page busy,
2318                  * blocking us, then in turn block on the map lock that we
2319                  * hold).  Because of problems in the recursive lock package,
2320                  * we cannot upgrade to a write lock in vm_map_lookup.  Thus,
2321                  * any actions that require the write lock must be done
2322                  * beforehand.  Because we keep the read lock on the map, the
2323                  * copy-on-write status of the entries we modify here cannot
2324                  * change.
2325                  */
2326                 entry = start_entry;
2327                 while ((entry != &map->header) && (entry->start < end)) {
2328                         /*
2329                          * Trivial case if the entry is already wired
2330                          */
2331                         if (entry->wired_count) {
2332                                 entry->wired_count++;
2333                                 entry = entry->next;
2334                                 continue;
2335                         }
2336
2337                         /*
2338                          * The entry is being newly wired, we have to setup
2339                          * appropriate management structures.  A shadow 
2340                          * object is required for a copy-on-write region,
2341                          * or a normal object for a zero-fill region.  We
2342                          * do not have to do this for entries that point to sub
2343                          * maps because we won't hold the lock on the sub map.
2344                          */
2345                         if (entry->maptype != VM_MAPTYPE_SUBMAP) {
2346                                 int copyflag = entry->eflags &
2347                                                MAP_ENTRY_NEEDS_COPY;
2348                                 if (copyflag && ((entry->protection &
2349                                                   VM_PROT_WRITE) != 0)) {
2350                                         vm_map_entry_shadow(entry, 0);
2351                                 } else if (entry->object.vm_object == NULL &&
2352                                            !map->system_map) {
2353                                         vm_map_entry_allocate_object(entry);
2354                                 }
2355                         }
2356
2357                         entry->wired_count++;
2358                         entry = entry->next;
2359                 }
2360
2361                 /*
2362                  * Pass 2.
2363                  */
2364
2365                 /*
2366                  * HACK HACK HACK HACK
2367                  *
2368                  * vm_fault_wire() temporarily unlocks the map to avoid
2369                  * deadlocks.  The in-transition flag from vm_map_clip_range
2370                  * call should protect us from changes while the map is
2371                  * unlocked.  T
2372                  *
2373                  * NOTE: Previously this comment stated that clipping might
2374                  *       still occur while the entry is unlocked, but from
2375                  *       what I can tell it actually cannot.
2376                  *
2377                  *       It is unclear whether the CLIP_CHECK_*() calls
2378                  *       are still needed but we keep them in anyway.
2379                  *
2380                  * HACK HACK HACK HACK
2381                  */
2382
2383                 entry = start_entry;
2384                 while (entry != &map->header && entry->start < end) {
2385                         /*
2386                          * If vm_fault_wire fails for any page we need to undo
2387                          * what has been done.  We decrement the wiring count
2388                          * for those pages which have not yet been wired (now)
2389                          * and unwire those that have (later).
2390                          */
2391                         vm_offset_t save_start = entry->start;
2392                         vm_offset_t save_end = entry->end;
2393
2394                         if (entry->wired_count == 1)
2395                                 rv = vm_fault_wire(map, entry, FALSE);
2396                         if (rv) {
2397                                 CLIP_CHECK_BACK(entry, save_start);
2398                                 for (;;) {
2399                                         KASSERT(entry->wired_count == 1, ("wired_count changed unexpectedly"));
2400                                         entry->wired_count = 0;
2401                                         if (entry->end == save_end)
2402                                                 break;
2403                                         entry = entry->next;
2404                                         KASSERT(entry != &map->header, ("bad entry clip during backout"));
2405                                 }
2406                                 end = save_start;
2407                                 break;
2408                         }
2409                         CLIP_CHECK_FWD(entry, save_end);
2410                         entry = entry->next;
2411                 }
2412
2413                 /*
2414                  * If a failure occured undo everything by falling through
2415                  * to the unwiring code.  'end' has already been adjusted
2416                  * appropriately.
2417                  */
2418                 if (rv)
2419                         kmflags |= KM_PAGEABLE;
2420
2421                 /*
2422                  * start_entry is still IN_TRANSITION but may have been 
2423                  * clipped since vm_fault_wire() unlocks and relocks the
2424                  * map.  No matter how clipped it has gotten there should
2425                  * be a fragment that is on our start boundary.
2426                  */
2427                 CLIP_CHECK_BACK(start_entry, start);
2428         }
2429
2430         if (kmflags & KM_PAGEABLE) {
2431                 /*
2432                  * This is the unwiring case.  We must first ensure that the
2433                  * range to be unwired is really wired down.  We know there
2434                  * are no holes.
2435                  */
2436                 entry = start_entry;
2437                 while ((entry != &map->header) && (entry->start < end)) {
2438                         if (entry->wired_count == 0) {
2439                                 rv = KERN_INVALID_ARGUMENT;
2440                                 goto done;
2441                         }
2442                         entry = entry->next;
2443                 }
2444
2445                 /*
2446                  * Now decrement the wiring count for each region. If a region
2447                  * becomes completely unwired, unwire its physical pages and
2448                  * mappings.
2449                  */
2450                 entry = start_entry;
2451                 while ((entry != &map->header) && (entry->start < end)) {
2452                         entry->wired_count--;
2453                         if (entry->wired_count == 0)
2454                                 vm_fault_unwire(map, entry);
2455                         entry = entry->next;
2456                 }
2457         }
2458 done:
2459         vm_map_unclip_range(map, start_entry, start, real_end,
2460                             &count, MAP_CLIP_NO_HOLES);
2461         map->timestamp++;
2462         vm_map_unlock(map);
2463 failure:
2464         if (kmflags & KM_KRESERVE)
2465                 vm_map_entry_krelease(count);
2466         else
2467                 vm_map_entry_release(count);
2468         return (rv);
2469 }
2470
2471 /*
2472  * Mark a newly allocated address range as wired but do not fault in
2473  * the pages.  The caller is expected to load the pages into the object.
2474  *
2475  * The map must be locked on entry and will remain locked on return.
2476  * No other requirements.
2477  */
2478 void
2479 vm_map_set_wired_quick(vm_map_t map, vm_offset_t addr, vm_size_t size,
2480                        int *countp)
2481 {
2482         vm_map_entry_t scan;
2483         vm_map_entry_t entry;
2484
2485         entry = vm_map_clip_range(map, addr, addr + size,
2486                                   countp, MAP_CLIP_NO_HOLES);
2487         for (scan = entry;
2488              scan != &map->header && scan->start < addr + size;
2489              scan = scan->next) {
2490             KKASSERT(entry->wired_count == 0);
2491             entry->wired_count = 1;                                              
2492         }
2493         vm_map_unclip_range(map, entry, addr, addr + size,
2494                             countp, MAP_CLIP_NO_HOLES);
2495 }
2496
2497 /*
2498  * Push any dirty cached pages in the address range to their pager.
2499  * If syncio is TRUE, dirty pages are written synchronously.
2500  * If invalidate is TRUE, any cached pages are freed as well.
2501  *
2502  * This routine is called by sys_msync()
2503  *
2504  * Returns an error if any part of the specified range is not mapped.
2505  *
2506  * No requirements.
2507  */
2508 int
2509 vm_map_clean(vm_map_t map, vm_offset_t start, vm_offset_t end,
2510              boolean_t syncio, boolean_t invalidate)
2511 {
2512         vm_map_entry_t current;
2513         vm_map_entry_t entry;
2514         vm_size_t size;
2515         vm_object_t object;
2516         vm_object_t tobj;
2517         vm_ooffset_t offset;
2518
2519         vm_map_lock_read(map);
2520         VM_MAP_RANGE_CHECK(map, start, end);
2521         if (!vm_map_lookup_entry(map, start, &entry)) {
2522                 vm_map_unlock_read(map);
2523                 return (KERN_INVALID_ADDRESS);
2524         }
2525         lwkt_gettoken(&map->token);
2526
2527         /*
2528          * Make a first pass to check for holes.
2529          */
2530         for (current = entry; current->start < end; current = current->next) {
2531                 if (current->maptype == VM_MAPTYPE_SUBMAP) {
2532                         lwkt_reltoken(&map->token);
2533                         vm_map_unlock_read(map);
2534                         return (KERN_INVALID_ARGUMENT);
2535                 }
2536                 if (end > current->end &&
2537                     (current->next == &map->header ||
2538                         current->end != current->next->start)) {
2539                         lwkt_reltoken(&map->token);
2540                         vm_map_unlock_read(map);
2541                         return (KERN_INVALID_ADDRESS);
2542                 }
2543         }
2544
2545         if (invalidate)
2546                 pmap_remove(vm_map_pmap(map), start, end);
2547
2548         /*
2549          * Make a second pass, cleaning/uncaching pages from the indicated
2550          * objects as we go.
2551          */
2552         for (current = entry; current->start < end; current = current->next) {
2553                 offset = current->offset + (start - current->start);
2554                 size = (end <= current->end ? end : current->end) - start;
2555                 if (current->maptype == VM_MAPTYPE_SUBMAP) {
2556                         vm_map_t smap;
2557                         vm_map_entry_t tentry;
2558                         vm_size_t tsize;
2559
2560                         smap = current->object.sub_map;
2561                         vm_map_lock_read(smap);
2562                         vm_map_lookup_entry(smap, offset, &tentry);
2563                         tsize = tentry->end - offset;
2564                         if (tsize < size)
2565                                 size = tsize;
2566                         object = tentry->object.vm_object;
2567                         offset = tentry->offset + (offset - tentry->start);
2568                         vm_map_unlock_read(smap);
2569                 } else {
2570                         object = current->object.vm_object;
2571                 }
2572
2573                 if (object)
2574                         vm_object_hold(object);
2575
2576                 /*
2577                  * Note that there is absolutely no sense in writing out
2578                  * anonymous objects, so we track down the vnode object
2579                  * to write out.
2580                  * We invalidate (remove) all pages from the address space
2581                  * anyway, for semantic correctness.
2582                  *
2583                  * note: certain anonymous maps, such as MAP_NOSYNC maps,
2584                  * may start out with a NULL object.
2585                  */
2586                 while (object && (tobj = object->backing_object) != NULL) {
2587                         vm_object_hold(tobj);
2588                         if (tobj == object->backing_object) {
2589                                 vm_object_lock_swap();
2590                                 offset += object->backing_object_offset;
2591                                 vm_object_drop(object);
2592                                 object = tobj;
2593                                 if (object->size < OFF_TO_IDX(offset + size))
2594                                         size = IDX_TO_OFF(object->size) -
2595                                                offset;
2596                                 break;
2597                         }
2598                         vm_object_drop(tobj);
2599                 }
2600                 if (object && (object->type == OBJT_VNODE) && 
2601                     (current->protection & VM_PROT_WRITE) &&
2602                     (object->flags & OBJ_NOMSYNC) == 0) {
2603                         /*
2604                          * Flush pages if writing is allowed, invalidate them
2605                          * if invalidation requested.  Pages undergoing I/O
2606                          * will be ignored by vm_object_page_remove().
2607                          *
2608                          * We cannot lock the vnode and then wait for paging
2609                          * to complete without deadlocking against vm_fault.
2610                          * Instead we simply call vm_object_page_remove() and
2611                          * allow it to block internally on a page-by-page 
2612                          * basis when it encounters pages undergoing async 
2613                          * I/O.
2614                          */
2615                         int flags;
2616
2617                         /* no chain wait needed for vnode objects */
2618                         vm_object_reference_locked(object);
2619                         vn_lock(object->handle, LK_EXCLUSIVE | LK_RETRY);
2620                         flags = (syncio || invalidate) ? OBJPC_SYNC : 0;
2621                         flags |= invalidate ? OBJPC_INVAL : 0;
2622
2623                         /*
2624                          * When operating on a virtual page table just
2625                          * flush the whole object.  XXX we probably ought
2626                          * to 
2627                          */
2628                         switch(current->maptype) {
2629                         case VM_MAPTYPE_NORMAL:
2630                                 vm_object_page_clean(object,
2631                                     OFF_TO_IDX(offset),
2632                                     OFF_TO_IDX(offset + size + PAGE_MASK),
2633                                     flags);
2634                                 break;
2635                         case VM_MAPTYPE_VPAGETABLE:
2636                                 vm_object_page_clean(object, 0, 0, flags);
2637                                 break;
2638                         }
2639                         vn_unlock(((struct vnode *)object->handle));
2640                         vm_object_deallocate_locked(object);
2641                 }
2642                 if (object && invalidate &&
2643                    ((object->type == OBJT_VNODE) ||
2644                     (object->type == OBJT_DEVICE))) {
2645                         int clean_only = 
2646                                 (object->type == OBJT_DEVICE) ? FALSE : TRUE;
2647                         /* no chain wait needed for vnode/device objects */
2648                         vm_object_reference_locked(object);
2649                         switch(current->maptype) {
2650                         case VM_MAPTYPE_NORMAL:
2651                                 vm_object_page_remove(object,
2652                                     OFF_TO_IDX(offset),
2653                                     OFF_TO_IDX(offset + size + PAGE_MASK),
2654                                     clean_only);
2655                                 break;
2656                         case VM_MAPTYPE_VPAGETABLE:
2657                                 vm_object_page_remove(object, 0, 0, clean_only);
2658                                 break;
2659                         }
2660                         vm_object_deallocate_locked(object);
2661                 }
2662                 start += size;
2663                 if (object)
2664                         vm_object_drop(object);
2665         }
2666
2667         lwkt_reltoken(&map->token);
2668         vm_map_unlock_read(map);
2669
2670         return (KERN_SUCCESS);
2671 }
2672
2673 /*
2674  * Make the region specified by this entry pageable.
2675  *
2676  * The vm_map must be exclusively locked.
2677  */
2678 static void 
2679 vm_map_entry_unwire(vm_map_t map, vm_map_entry_t entry)
2680 {
2681         entry->eflags &= ~MAP_ENTRY_USER_WIRED;
2682         entry->wired_count = 0;
2683         vm_fault_unwire(map, entry);
2684 }
2685
2686 /*
2687  * Deallocate the given entry from the target map.
2688  *
2689  * The vm_map must be exclusively locked.
2690  */
2691 static void
2692 vm_map_entry_delete(vm_map_t map, vm_map_entry_t entry, int *countp)
2693 {
2694         vm_map_entry_unlink(map, entry);
2695         map->size -= entry->end - entry->start;
2696
2697         switch(entry->maptype) {
2698         case VM_MAPTYPE_NORMAL:
2699         case VM_MAPTYPE_VPAGETABLE:
2700                 vm_object_deallocate(entry->object.vm_object);
2701                 break;
2702         default:
2703                 break;
2704         }
2705
2706         vm_map_entry_dispose(map, entry, countp);
2707 }
2708
2709 /*
2710  * Deallocates the given address range from the target map.
2711  *
2712  * The vm_map must be exclusively locked.
2713  */
2714 int
2715 vm_map_delete(vm_map_t map, vm_offset_t start, vm_offset_t end, int *countp)
2716 {
2717         vm_object_t object;
2718         vm_map_entry_t entry;
2719         vm_map_entry_t first_entry;
2720
2721         ASSERT_VM_MAP_LOCKED(map);
2722         lwkt_gettoken(&map->token);
2723 again:
2724         /*
2725          * Find the start of the region, and clip it.  Set entry to point
2726          * at the first record containing the requested address or, if no
2727          * such record exists, the next record with a greater address.  The
2728          * loop will run from this point until a record beyond the termination
2729          * address is encountered.
2730          *
2731          * map->hint must be adjusted to not point to anything we delete,
2732          * so set it to the entry prior to the one being deleted.
2733          *
2734          * GGG see other GGG comment.
2735          */
2736         if (vm_map_lookup_entry(map, start, &first_entry)) {
2737                 entry = first_entry;
2738                 vm_map_clip_start(map, entry, start, countp);
2739                 map->hint = entry->prev;        /* possible problem XXX */
2740         } else {
2741                 map->hint = first_entry;        /* possible problem XXX */
2742                 entry = first_entry->next;
2743         }
2744
2745         /*
2746          * If a hole opens up prior to the current first_free then
2747          * adjust first_free.  As with map->hint, map->first_free
2748          * cannot be left set to anything we might delete.
2749          */
2750         if (entry == &map->header) {
2751                 map->first_free = &map->header;
2752         } else if (map->first_free->start >= start) {
2753                 map->first_free = entry->prev;
2754         }
2755
2756         /*
2757          * Step through all entries in this region
2758          */
2759         while ((entry != &map->header) && (entry->start < end)) {
2760                 vm_map_entry_t next;
2761                 vm_offset_t s, e;
2762                 vm_pindex_t offidxstart, offidxend, count;
2763
2764                 /*
2765                  * If we hit an in-transition entry we have to sleep and
2766                  * retry.  It's easier (and not really slower) to just retry
2767                  * since this case occurs so rarely and the hint is already
2768                  * pointing at the right place.  We have to reset the
2769                  * start offset so as not to accidently delete an entry
2770                  * another process just created in vacated space.
2771                  */
2772                 if (entry->eflags & MAP_ENTRY_IN_TRANSITION) {
2773                         entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP;
2774                         start = entry->start;
2775                         ++mycpu->gd_cnt.v_intrans_coll;
2776                         ++mycpu->gd_cnt.v_intrans_wait;
2777                         vm_map_transition_wait(map);
2778                         goto again;
2779                 }
2780                 vm_map_clip_end(map, entry, end, countp);
2781
2782                 s = entry->start;
2783                 e = entry->end;
2784                 next = entry->next;
2785
2786                 offidxstart = OFF_TO_IDX(entry->offset);
2787                 count = OFF_TO_IDX(e - s);
2788                 object = entry->object.vm_object;
2789
2790                 /*
2791                  * Unwire before removing addresses from the pmap; otherwise,
2792                  * unwiring will put the entries back in the pmap.
2793                  */
2794                 if (entry->wired_count != 0)
2795                         vm_map_entry_unwire(map, entry);
2796
2797                 offidxend = offidxstart + count;
2798
2799                 if (object == &kernel_object) {
2800                         vm_object_hold(object);
2801                         vm_object_page_remove(object, offidxstart,
2802                                               offidxend, FALSE);
2803                         vm_object_drop(object);
2804                 } else if (object && object->type != OBJT_DEFAULT &&
2805                            object->type != OBJT_SWAP) {
2806                         /*
2807                          * vnode object routines cannot be chain-locked
2808                          */
2809                         vm_object_hold(object);
2810                         pmap_remove(map->pmap, s, e);
2811                         vm_object_drop(object);
2812                 } else if (object) {
2813                         vm_object_hold(object);
2814                         vm_object_chain_acquire(object);
2815                         pmap_remove(map->pmap, s, e);
2816
2817                         if (object != NULL &&
2818                             object->ref_count != 1 &&
2819                             (object->flags & (OBJ_NOSPLIT|OBJ_ONEMAPPING)) ==
2820                              OBJ_ONEMAPPING &&
2821                             (object->type == OBJT_DEFAULT ||
2822                              object->type == OBJT_SWAP)) {
2823                                 vm_object_collapse(object, NULL);
2824                                 vm_object_page_remove(object, offidxstart,
2825                                                       offidxend, FALSE);
2826                                 if (object->type == OBJT_SWAP) {
2827                                         swap_pager_freespace(object,
2828                                                              offidxstart,
2829                                                              count);
2830                                 }
2831                                 if (offidxend >= object->size &&
2832                                     offidxstart < object->size) {
2833                                         object->size = offidxstart;
2834                                 }
2835                         }
2836                         vm_object_chain_release(object);
2837                         vm_object_drop(object);
2838                 }
2839
2840                 /*
2841                  * Delete the entry (which may delete the object) only after
2842                  * removing all pmap entries pointing to its pages.
2843                  * (Otherwise, its page frames may be reallocated, and any
2844                  * modify bits will be set in the wrong object!)
2845                  */
2846                 vm_map_entry_delete(map, entry, countp);
2847                 entry = next;
2848         }
2849         lwkt_reltoken(&map->token);
2850         return (KERN_SUCCESS);
2851 }
2852
2853 /*
2854  * Remove the given address range from the target map.
2855  * This is the exported form of vm_map_delete.
2856  *
2857  * No requirements.
2858  */
2859 int
2860 vm_map_remove(vm_map_t map, vm_offset_t start, vm_offset_t end)
2861 {
2862         int result;
2863         int count;
2864
2865         count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
2866         vm_map_lock(map);
2867         VM_MAP_RANGE_CHECK(map, start, end);
2868         result = vm_map_delete(map, start, end, &count);
2869         vm_map_unlock(map);
2870         vm_map_entry_release(count);
2871
2872         return (result);
2873 }
2874
2875 /*
2876  * Assert that the target map allows the specified privilege on the
2877  * entire address region given.  The entire region must be allocated.
2878  *
2879  * The caller must specify whether the vm_map is already locked or not.
2880  */
2881 boolean_t
2882 vm_map_check_protection(vm_map_t map, vm_offset_t start, vm_offset_t end,
2883                         vm_prot_t protection, boolean_t have_lock)
2884 {
2885         vm_map_entry_t entry;
2886         vm_map_entry_t tmp_entry;
2887         boolean_t result;
2888
2889         if (have_lock == FALSE)
2890                 vm_map_lock_read(map);
2891
2892         if (!vm_map_lookup_entry(map, start, &tmp_entry)) {
2893                 if (have_lock == FALSE)
2894                         vm_map_unlock_read(map);
2895                 return (FALSE);
2896         }
2897         entry = tmp_entry;
2898
2899         result = TRUE;
2900         while (start < end) {
2901                 if (entry == &map->header) {
2902                         result = FALSE;
2903                         break;
2904                 }
2905                 /*
2906                  * No holes allowed!
2907                  */
2908
2909                 if (start < entry->start) {
2910                         result = FALSE;
2911                         break;
2912                 }
2913                 /*
2914                  * Check protection associated with entry.
2915                  */
2916
2917                 if ((entry->protection & protection) != protection) {
2918                         result = FALSE;
2919                         break;
2920                 }
2921                 /* go to next entry */
2922
2923                 start = entry->end;
2924                 entry = entry->next;
2925         }
2926         if (have_lock == FALSE)
2927                 vm_map_unlock_read(map);
2928         return (result);
2929 }
2930
2931 /*
2932  * If appropriate this function shadows the original object with a new object
2933  * and moves the VM pages from the original object to the new object.
2934  * The original object will also be collapsed, if possible.
2935  *
2936  * We can only do this for normal memory objects with a single mapping, and
2937  * it only makes sense to do it if there are 2 or more refs on the original
2938  * object.  i.e. typically a memory object that has been extended into
2939  * multiple vm_map_entry's with non-overlapping ranges.
2940  *
2941  * This makes it easier to remove unused pages and keeps object inheritance
2942  * from being a negative impact on memory usage.
2943  *
2944  * On return the (possibly new) entry->object.vm_object will have an
2945  * additional ref on it for the caller to dispose of (usually by cloning
2946  * the vm_map_entry).  The additional ref had to be done in this routine
2947  * to avoid racing a collapse.  The object's ONEMAPPING flag will also be
2948  * cleared.
2949  *
2950  * The vm_map must be locked and its token held.
2951  */
2952 static void
2953 vm_map_split(vm_map_entry_t entry)
2954 {
2955 #if 0
2956         /* UNOPTIMIZED */
2957         vm_object_t oobject;
2958
2959         oobject = entry->object.vm_object;
2960         vm_object_hold(oobject);
2961         vm_object_chain_wait(oobject);
2962         vm_object_reference_locked(oobject);
2963         vm_object_clear_flag(oobject, OBJ_ONEMAPPING);
2964         vm_object_drop(oobject);
2965 #else
2966         /* OPTIMIZED */
2967         vm_object_t oobject, nobject, bobject;
2968         vm_offset_t s, e;
2969         vm_page_t m;
2970         vm_pindex_t offidxstart, offidxend, idx;
2971         vm_size_t size;
2972         vm_ooffset_t offset;
2973
2974         /*
2975          * Setup.  Chain lock the original object throughout the entire
2976          * routine to prevent new page faults from occuring.
2977          *
2978          * XXX can madvise WILLNEED interfere with us too?
2979          */
2980         oobject = entry->object.vm_object;
2981         vm_object_hold(oobject);
2982         vm_object_chain_acquire(oobject);
2983
2984         /*
2985          * Original object cannot be split?
2986          */
2987         if (oobject->handle == NULL || (oobject->type != OBJT_DEFAULT &&
2988                                         oobject->type != OBJT_SWAP)) {
2989                 vm_object_chain_release(oobject);
2990                 vm_object_reference_locked(oobject);
2991                 vm_object_clear_flag(oobject, OBJ_ONEMAPPING);
2992                 vm_object_drop(oobject);
2993                 return;
2994         }
2995
2996         /*
2997          * Collapse original object with its backing store as an
2998          * optimization to reduce chain lengths when possible.
2999          *
3000          * If ref_count <= 1 there aren't other non-overlapping vm_map_entry's
3001          * for oobject, so there's no point collapsing it.
3002          *
3003          * Then re-check whether the object can be split.
3004          */
3005         vm_object_collapse(oobject, NULL);
3006
3007         if (oobject->ref_count <= 1 ||
3008             (oobject->type != OBJT_DEFAULT && oobject->type != OBJT_SWAP) ||
3009             (oobject->flags & (OBJ_NOSPLIT|OBJ_ONEMAPPING)) != OBJ_ONEMAPPING) {
3010                 vm_object_chain_release(oobject);
3011                 vm_object_reference_locked(oobject);
3012                 vm_object_clear_flag(oobject, OBJ_ONEMAPPING);
3013                 vm_object_drop(oobject);
3014                 return;
3015         }
3016
3017         /*
3018          * Acquire the chain lock on the backing object.
3019          *
3020          * Give bobject an additional ref count for when it will be shadowed
3021          * by nobject.
3022          */
3023         if ((bobject = oobject->backing_object) != NULL) {
3024                 vm_object_hold(bobject);
3025                 vm_object_chain_wait(bobject);
3026                 vm_object_reference_locked(bobject);
3027                 vm_object_chain_acquire(bobject);
3028                 KKASSERT(bobject->backing_object == bobject);
3029                 KKASSERT((bobject->flags & OBJ_DEAD) == 0);
3030         }
3031
3032         /*
3033          * Calculate the object page range and allocate the new object.
3034          */
3035         offset = entry->offset;
3036         s = entry->start;
3037         e = entry->end;
3038
3039         offidxstart = OFF_TO_IDX(offset);
3040         offidxend = offidxstart + OFF_TO_IDX(e - s);
3041         size = offidxend - offidxstart;
3042
3043         switch(oobject->type) {
3044         case OBJT_DEFAULT:
3045                 nobject = default_pager_alloc(NULL, IDX_TO_OFF(size),
3046                                               VM_PROT_ALL, 0);
3047                 break;
3048         case OBJT_SWAP:
3049                 nobject = swap_pager_alloc(NULL, IDX_TO_OFF(size),
3050                                            VM_PROT_ALL, 0);
3051                 break;
3052         default:
3053                 /* not reached */
3054                 nobject = NULL;
3055                 KKASSERT(0);
3056         }
3057
3058         if (nobject == NULL) {
3059                 if (bobject) {
3060                         vm_object_chain_release(bobject);
3061                         vm_object_deallocate(bobject);
3062                         vm_object_drop(bobject);
3063                 }
3064                 vm_object_chain_release(oobject);
3065                 vm_object_reference_locked(oobject);
3066                 vm_object_clear_flag(oobject, OBJ_ONEMAPPING);
3067                 vm_object_drop(oobject);
3068                 return;
3069         }
3070
3071         /*
3072          * The new object will replace entry->object.vm_object so it needs
3073          * a second reference (the caller expects an additional ref).
3074          */
3075         vm_object_hold(nobject);
3076         vm_object_reference_locked(nobject);
3077         vm_object_chain_acquire(nobject);
3078
3079         /*
3080          * nobject shadows bobject (oobject already shadows bobject).
3081          */
3082         if (bobject) {
3083                 nobject->backing_object_offset =
3084                     oobject->backing_object_offset + IDX_TO_OFF(offidxstart);
3085                 nobject->backing_object = bobject;
3086                 bobject->shadow_count++;
3087                 bobject->generation++;
3088                 LIST_INSERT_HEAD(&bobject->shadow_head, nobject, shadow_list);
3089                 vm_object_clear_flag(bobject, OBJ_ONEMAPPING); /* XXX? */
3090                 vm_object_chain_release(bobject);
3091                 vm_object_drop(bobject);
3092         }
3093
3094         /*
3095          * Move the VM pages from oobject to nobject
3096          */
3097         for (idx = 0; idx < size; idx++) {
3098                 vm_page_t m;
3099
3100                 m = vm_page_lookup_busy_wait(oobject, offidxstart + idx,
3101                                              TRUE, "vmpg");
3102                 if (m == NULL)
3103                         continue;
3104
3105                 /*
3106                  * We must wait for pending I/O to complete before we can
3107                  * rename the page.
3108                  *
3109                  * We do not have to VM_PROT_NONE the page as mappings should
3110                  * not be changed by this operation.
3111                  *
3112                  * NOTE: The act of renaming a page updates chaingen for both
3113                  *       objects.
3114                  */
3115                 vm_page_rename(m, nobject, idx);
3116                 /* page automatically made dirty by rename and cache handled */
3117                 /* page remains busy */
3118         }
3119
3120         if (oobject->type == OBJT_SWAP) {
3121                 vm_object_pip_add(oobject, 1);
3122                 /*
3123                  * copy oobject pages into nobject and destroy unneeded
3124                  * pages in shadow object.
3125                  */
3126                 swap_pager_copy(oobject, nobject, offidxstart, 0);
3127                 vm_object_pip_wakeup(oobject);
3128         }
3129
3130         /*
3131          * Wakeup the pages we played with.  No spl protection is needed
3132          * for a simple wakeup.
3133          */
3134         for (idx = 0; idx < size; idx++) {
3135                 m = vm_page_lookup(nobject, idx);
3136                 if (m) {
3137                         KKASSERT(m->flags & PG_BUSY);
3138                         vm_page_wakeup(m);
3139                 }
3140         }
3141         entry->object.vm_object = nobject;
3142         entry->offset = 0LL;
3143
3144         /*
3145          * Cleanup
3146          *
3147          * NOTE: There is no need to remove OBJ_ONEMAPPING from oobject, the
3148          *       related pages were moved and are no longer applicable to the
3149          *       original object.
3150          *
3151          * NOTE: Deallocate oobject (due to its entry->object.vm_object being
3152          *       replaced by nobject).
3153          */
3154         vm_object_chain_release(nobject);
3155         vm_object_drop(nobject);
3156         if (bobject) {
3157                 vm_object_chain_release(bobject);
3158                 vm_object_drop(bobject);
3159         }
3160         vm_object_chain_release(oobject);
3161         /*vm_object_clear_flag(oobject, OBJ_ONEMAPPING);*/
3162         vm_object_deallocate_locked(oobject);
3163         vm_object_drop(oobject);
3164 #endif
3165 }
3166
3167 /*
3168  * Copies the contents of the source entry to the destination
3169  * entry.  The entries *must* be aligned properly.
3170  *
3171  * The vm_maps must be exclusively locked.
3172  * The vm_map's token must be held.
3173  *
3174  * Because the maps are locked no faults can be in progress during the
3175  * operation.
3176  */
3177 static void
3178 vm_map_copy_entry(vm_map_t src_map, vm_map_t dst_map,
3179                   vm_map_entry_t src_entry, vm_map_entry_t dst_entry)
3180 {
3181         vm_object_t src_object;
3182
3183         if (dst_entry->maptype == VM_MAPTYPE_SUBMAP)
3184                 return;
3185         if (src_entry->maptype == VM_MAPTYPE_SUBMAP)
3186                 return;
3187
3188         if (src_entry->wired_count == 0) {
3189                 /*
3190                  * If the source entry is marked needs_copy, it is already
3191                  * write-protected.
3192                  */
3193                 if ((src_entry->eflags & MAP_ENTRY_NEEDS_COPY) == 0) {
3194                         pmap_protect(src_map->pmap,
3195                             src_entry->start,
3196                             src_entry->end,
3197                             src_entry->protection & ~VM_PROT_WRITE);
3198                 }
3199
3200                 /*
3201                  * Make a copy of the object.
3202                  *
3203                  * The object must be locked prior to checking the object type
3204                  * and for the call to vm_object_collapse() and vm_map_split().
3205                  * We cannot use *_hold() here because the split code will
3206                  * probably try to destroy the object.  The lock is a pool
3207                  * token and doesn't care.
3208                  *
3209                  * We must bump src_map->timestamp when setting
3210                  * MAP_ENTRY_NEEDS_COPY to force any concurrent fault
3211                  * to retry, otherwise the concurrent fault might improperly
3212                  * install a RW pte when its supposed to be a RO(COW) pte.
3213                  * This race can occur because a vnode-backed fault may have
3214                  * to temporarily release the map lock.
3215                  */
3216                 if (src_entry->object.vm_object != NULL) {
3217                         vm_map_split(src_entry);
3218                         src_object = src_entry->object.vm_object;
3219                         dst_entry->object.vm_object = src_object;
3220                         src_entry->eflags |= (MAP_ENTRY_COW |
3221                                               MAP_ENTRY_NEEDS_COPY);
3222                         dst_entry->eflags |= (MAP_ENTRY_COW |
3223                                               MAP_ENTRY_NEEDS_COPY);
3224                         dst_entry->offset = src_entry->offset;
3225                         ++src_map->timestamp;
3226                 } else {
3227                         dst_entry->object.vm_object = NULL;
3228                         dst_entry->offset = 0;
3229                 }
3230
3231                 pmap_copy(dst_map->pmap, src_map->pmap, dst_entry->start,
3232                     dst_entry->end - dst_entry->start, src_entry->start);
3233         } else {
3234                 /*
3235                  * Of course, wired down pages can't be set copy-on-write.
3236                  * Cause wired pages to be copied into the new map by
3237                  * simulating faults (the new pages are pageable)
3238                  */
3239                 vm_fault_copy_entry(dst_map, src_map, dst_entry, src_entry);
3240         }
3241 }
3242
3243 /*
3244  * vmspace_fork:
3245  * Create a new process vmspace structure and vm_map
3246  * based on those of an existing process.  The new map
3247  * is based on the old map, according to the inheritance
3248  * values on the regions in that map.
3249  *
3250  * The source map must not be locked.
3251  * No requirements.
3252  */
3253 struct vmspace *
3254 vmspace_fork(struct vmspace *vm1)
3255 {
3256         struct vmspace *vm2;
3257         vm_map_t old_map = &vm1->vm_map;
3258         vm_map_t new_map;
3259         vm_map_entry_t old_entry;
3260         vm_map_entry_t new_entry;
3261         vm_object_t object;
3262         int count;
3263
3264         lwkt_gettoken(&vm1->vm_map.token);
3265         vm_map_lock(old_map);
3266
3267         /*
3268          * XXX Note: upcalls are not copied.
3269          */
3270         vm2 = vmspace_alloc(old_map->min_offset, old_map->max_offset);
3271         lwkt_gettoken(&vm2->vm_map.token);
3272         bcopy(&vm1->vm_startcopy, &vm2->vm_startcopy,
3273             (caddr_t)&vm1->vm_endcopy - (caddr_t)&vm1->vm_startcopy);
3274         new_map = &vm2->vm_map; /* XXX */
3275         new_map->timestamp = 1;
3276
3277         vm_map_lock(new_map);
3278
3279         count = 0;
3280         old_entry = old_map->header.next;
3281         while (old_entry != &old_map->header) {
3282                 ++count;
3283                 old_entry = old_entry->next;
3284         }
3285
3286         count = vm_map_entry_reserve(count + MAP_RESERVE_COUNT);
3287
3288         old_entry = old_map->header.next;
3289         while (old_entry != &old_map->header) {
3290                 if (old_entry->maptype == VM_MAPTYPE_SUBMAP)
3291                         panic("vm_map_fork: encountered a submap");
3292
3293                 switch (old_entry->inheritance) {
3294                 case VM_INHERIT_NONE:
3295                         break;
3296                 case VM_INHERIT_SHARE:
3297                         /*
3298                          * Clone the entry, creating the shared object if
3299                          * necessary.
3300                          */
3301                         if (old_entry->object.vm_object == NULL)
3302                                 vm_map_entry_allocate_object(old_entry);
3303
3304                         if (old_entry->eflags & MAP_ENTRY_NEEDS_COPY) {
3305                                 /*
3306                                  * Shadow a map_entry which needs a copy,
3307                                  * replacing its object with a new object
3308                                  * that points to the old one.  Ask the
3309                                  * shadow code to automatically add an
3310                                  * additional ref.  We can't do it afterwords
3311                                  * because we might race a collapse.  The call
3312                                  * to vm_map_entry_shadow() will also clear
3313                                  * OBJ_ONEMAPPING.
3314                                  */
3315                                 vm_map_entry_shadow(old_entry, 1);
3316                         } else {
3317                                 /*
3318                                  * We will make a shared copy of the object,
3319                                  * and must clear OBJ_ONEMAPPING.
3320                                  *
3321                                  * XXX assert that object.vm_object != NULL
3322                                  *     since we allocate it above.
3323                                  */
3324                                 if (old_entry->object.vm_object) {
3325                                         object = old_entry->object.vm_object;
3326                                         vm_object_hold(object);
3327                                         vm_object_chain_wait(object);
3328                                         vm_object_reference_locked(object);
3329                                         vm_object_clear_flag(object,
3330                                                              OBJ_ONEMAPPING);
3331                                         vm_object_drop(object);
3332                                 }
3333                         }
3334
3335                         /*
3336                          * Clone the entry.  We've already bumped the ref on
3337                          * any vm_object.
3338                          */
3339                         new_entry = vm_map_entry_create(new_map, &count);
3340                         *new_entry = *old_entry;
3341                         new_entry->eflags &= ~MAP_ENTRY_USER_WIRED;
3342                         new_entry->wired_count = 0;
3343
3344                         /*
3345                          * Insert the entry into the new map -- we know we're
3346                          * inserting at the end of the new map.
3347                          */
3348
3349                         vm_map_entry_link(new_map, new_map->header.prev,
3350                                           new_entry);
3351
3352                         /*
3353                          * Update the physical map
3354                          */
3355                         pmap_copy(new_map->pmap, old_map->pmap,
3356                             new_entry->start,
3357                             (old_entry->end - old_entry->start),
3358                             old_entry->start);
3359                         break;
3360                 case VM_INHERIT_COPY:
3361                         /*
3362                          * Clone the entry and link into the map.
3363                          */
3364                         new_entry = vm_map_entry_create(new_map, &count);
3365                         *new_entry = *old_entry;
3366                         new_entry->eflags &= ~MAP_ENTRY_USER_WIRED;
3367                         new_entry->wired_count = 0;
3368                         new_entry->object.vm_object = NULL;
3369                         vm_map_entry_link(new_map, new_map->header.prev,
3370                                           new_entry);
3371                         vm_map_copy_entry(old_map, new_map, old_entry,
3372                                           new_entry);
3373                         break;
3374                 }
3375                 old_entry = old_entry->next;
3376         }
3377
3378         new_map->size = old_map->size;
3379         vm_map_unlock(old_map);
3380         vm_map_unlock(new_map);
3381         vm_map_entry_release(count);
3382
3383         lwkt_reltoken(&vm2->vm_map.token);
3384         lwkt_reltoken(&vm1->vm_map.token);
3385
3386         return (vm2);
3387 }
3388
3389 /*
3390  * Create an auto-grow stack entry
3391  *
3392  * No requirements.
3393  */
3394 int
3395 vm_map_stack (vm_map_t map, vm_offset_t addrbos, vm_size_t max_ssize,
3396               int flags, vm_prot_t prot, vm_prot_t max, int cow)
3397 {
3398         vm_map_entry_t  prev_entry;
3399         vm_map_entry_t  new_stack_entry;
3400         vm_size_t       init_ssize;
3401         int             rv;
3402         int             count;
3403         vm_offset_t     tmpaddr;
3404
3405         cow |= MAP_IS_STACK;
3406
3407         if (max_ssize < sgrowsiz)
3408                 init_ssize = max_ssize;
3409         else
3410                 init_ssize = sgrowsiz;
3411
3412         count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
3413         vm_map_lock(map);
3414
3415         /*
3416          * Find space for the mapping
3417          */
3418         if ((flags & (MAP_FIXED | MAP_TRYFIXED)) == 0) {
3419                 if (vm_map_findspace(map, addrbos, max_ssize, 1,
3420                                      flags, &tmpaddr)) {
3421                         vm_map_unlock(map);
3422                         vm_map_entry_release(count);
3423                         return (KERN_NO_SPACE);
3424                 }
3425                 addrbos = tmpaddr;
3426         }
3427
3428         /* If addr is already mapped, no go */
3429         if (vm_map_lookup_entry(map, addrbos, &prev_entry)) {
3430                 vm_map_unlock(map);
3431                 vm_map_entry_release(count);
3432                 return (KERN_NO_SPACE);
3433         }
3434
3435 #if 0
3436         /* XXX already handled by kern_mmap() */
3437         /* If we would blow our VMEM resource limit, no go */
3438         if (map->size + init_ssize >
3439             curproc->p_rlimit[RLIMIT_VMEM].rlim_cur) {
3440                 vm_map_unlock(map);
3441                 vm_map_entry_release(count);
3442                 return (KERN_NO_SPACE);
3443         }
3444 #endif
3445
3446         /*
3447          * If we can't accomodate max_ssize in the current mapping,
3448          * no go.  However, we need to be aware that subsequent user
3449          * mappings might map into the space we have reserved for
3450          * stack, and currently this space is not protected.  
3451          * 
3452          * Hopefully we will at least detect this condition 
3453          * when we try to grow the stack.
3454          */
3455         if ((prev_entry->next != &map->header) &&
3456             (prev_entry->next->start < addrbos + max_ssize)) {
3457                 vm_map_unlock(map);
3458                 vm_map_entry_release(count);
3459                 return (KERN_NO_SPACE);
3460         }
3461
3462         /*
3463          * We initially map a stack of only init_ssize.  We will
3464          * grow as needed later.  Since this is to be a grow 
3465          * down stack, we map at the top of the range.
3466          *
3467          * Note: we would normally expect prot and max to be
3468          * VM_PROT_ALL, and cow to be 0.  Possibly we should
3469          * eliminate these as input parameters, and just
3470          * pass these values here in the insert call.
3471          */
3472         rv = vm_map_insert(map, &count,
3473                            NULL, 0, addrbos + max_ssize - init_ssize,
3474                            addrbos + max_ssize,
3475                            VM_MAPTYPE_NORMAL,
3476                            prot, max,
3477                            cow);
3478
3479         /* Now set the avail_ssize amount */
3480         if (rv == KERN_SUCCESS) {
3481                 if (prev_entry != &map->header)
3482                         vm_map_clip_end(map, prev_entry, addrbos + max_ssize - init_ssize, &count);
3483                 new_stack_entry = prev_entry->next;
3484                 if (new_stack_entry->end   != addrbos + max_ssize ||
3485                     new_stack_entry->start != addrbos + max_ssize - init_ssize)
3486                         panic ("Bad entry start/end for new stack entry");
3487                 else 
3488                         new_stack_entry->aux.avail_ssize = max_ssize - init_ssize;
3489         }
3490
3491         vm_map_unlock(map);
3492         vm_map_entry_release(count);
3493         return (rv);
3494 }
3495
3496 /*
3497  * Attempts to grow a vm stack entry.  Returns KERN_SUCCESS if the
3498  * desired address is already mapped, or if we successfully grow
3499  * the stack.  Also returns KERN_SUCCESS if addr is outside the
3500  * stack range (this is strange, but preserves compatibility with
3501  * the grow function in vm_machdep.c).
3502  *
3503  * No requirements.
3504  */
3505 int
3506 vm_map_growstack (struct proc *p, vm_offset_t addr)
3507 {
3508         vm_map_entry_t prev_entry;
3509         vm_map_entry_t stack_entry;
3510         vm_map_entry_t new_stack_entry;
3511         struct vmspace *vm = p->p_vmspace;
3512         vm_map_t map = &vm->vm_map;
3513         vm_offset_t    end;
3514         int grow_amount;
3515         int rv = KERN_SUCCESS;
3516         int is_procstack;
3517         int use_read_lock = 1;
3518         int count;
3519
3520         count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
3521 Retry:
3522         if (use_read_lock)
3523                 vm_map_lock_read(map);
3524         else
3525                 vm_map_lock(map);
3526
3527         /* If addr is already in the entry range, no need to grow.*/
3528         if (vm_map_lookup_entry(map, addr, &prev_entry))
3529                 goto done;
3530
3531         if ((stack_entry = prev_entry->next) == &map->header)
3532                 goto done;
3533         if (prev_entry == &map->header) 
3534                 end = stack_entry->start - stack_entry->aux.avail_ssize;
3535         else
3536                 end = prev_entry->end;
3537
3538         /*
3539          * This next test mimics the old grow function in vm_machdep.c.
3540          * It really doesn't quite make sense, but we do it anyway
3541          * for compatibility.
3542          *
3543          * If not growable stack, return success.  This signals the
3544          * caller to proceed as he would normally with normal vm.
3545          */
3546         if (stack_entry->aux.avail_ssize < 1 ||
3547             addr >= stack_entry->start ||
3548             addr <  stack_entry->start - stack_entry->aux.avail_ssize) {
3549                 goto done;
3550         } 
3551         
3552         /* Find the minimum grow amount */
3553         grow_amount = roundup (stack_entry->start - addr, PAGE_SIZE);
3554         if (grow_amount > stack_entry->aux.avail_ssize) {
3555                 rv = KERN_NO_SPACE;
3556                 goto done;
3557         }
3558
3559         /*
3560          * If there is no longer enough space between the entries
3561          * nogo, and adjust the available space.  Note: this 
3562          * should only happen if the user has mapped into the
3563          * stack area after the stack was created, and is
3564          * probably an error.
3565          *
3566          * This also effectively destroys any guard page the user
3567          * might have intended by limiting the stack size.
3568          */
3569         if (grow_amount > stack_entry->start - end) {
3570                 if (use_read_lock && vm_map_lock_upgrade(map)) {
3571                         /* lost lock */
3572                         use_read_lock = 0;
3573                         goto Retry;
3574                 }
3575                 use_read_lock = 0;
3576                 stack_entry->aux.avail_ssize = stack_entry->start - end;
3577                 rv = KERN_NO_SPACE;
3578                 goto done;
3579         }
3580
3581         is_procstack = addr >= (vm_offset_t)vm->vm_maxsaddr;
3582
3583         /* If this is the main process stack, see if we're over the 
3584          * stack limit.
3585          */
3586         if (is_procstack && (ctob(vm->vm_ssize) + grow_amount >
3587                              p->p_rlimit[RLIMIT_STACK].rlim_cur)) {
3588                 rv = KERN_NO_SPACE;
3589                 goto done;
3590         }
3591
3592         /* Round up the grow amount modulo SGROWSIZ */
3593         grow_amount = roundup (grow_amount, sgrowsiz);
3594         if (grow_amount > stack_entry->aux.avail_ssize) {
3595                 grow_amount = stack_entry->aux.avail_ssize;
3596         }
3597         if (is_procstack && (ctob(vm->vm_ssize) + grow_amount >
3598                              p->p_rlimit[RLIMIT_STACK].rlim_cur)) {
3599                 grow_amount = p->p_rlimit[RLIMIT_STACK].rlim_cur -
3600                               ctob(vm->vm_ssize);
3601         }
3602
3603         /* If we would blow our VMEM resource limit, no go */
3604         if (map->size + grow_amount > p->p_rlimit[RLIMIT_VMEM].rlim_cur) {
3605                 rv = KERN_NO_SPACE;
3606                 goto done;
3607         }
3608
3609         if (use_read_lock && vm_map_lock_upgrade(map)) {
3610                 /* lost lock */
3611                 use_read_lock = 0;
3612                 goto Retry;
3613         }
3614         use_read_lock = 0;
3615
3616         /* Get the preliminary new entry start value */
3617         addr = stack_entry->start - grow_amount;
3618
3619         /* If this puts us into the previous entry, cut back our growth
3620          * to the available space.  Also, see the note above.
3621          */
3622         if (addr < end) {
3623                 stack_entry->aux.avail_ssize = stack_entry->start - end;
3624                 addr = end;
3625         }
3626
3627         rv = vm_map_insert(map, &count,
3628                            NULL, 0, addr, stack_entry->start,
3629                            VM_MAPTYPE_NORMAL,
3630                            VM_PROT_ALL, VM_PROT_ALL,
3631                            0);
3632
3633         /* Adjust the available stack space by the amount we grew. */
3634         if (rv == KERN_SUCCESS) {
3635                 if (prev_entry != &map->header)
3636                         vm_map_clip_end(map, prev_entry, addr, &count);
3637                 new_stack_entry = prev_entry->next;
3638                 if (new_stack_entry->end   != stack_entry->start  ||
3639                     new_stack_entry->start != addr)
3640                         panic ("Bad stack grow start/end in new stack entry");
3641                 else {
3642                         new_stack_entry->aux.avail_ssize =
3643                                 stack_entry->aux.avail_ssize -
3644                                 (new_stack_entry->end - new_stack_entry->start);
3645                         if (is_procstack)
3646                                 vm->vm_ssize += btoc(new_stack_entry->end -
3647                                                      new_stack_entry->start);
3648                 }
3649
3650                 if (map->flags & MAP_WIREFUTURE)
3651                         vm_map_unwire(map, new_stack_entry->start,
3652                                       new_stack_entry->end, FALSE);
3653         }
3654
3655 done:
3656         if (use_read_lock)
3657                 vm_map_unlock_read(map);
3658         else
3659                 vm_map_unlock(map);
3660         vm_map_entry_release(count);
3661         return (rv);
3662 }
3663
3664 /*
3665  * Unshare the specified VM space for exec.  If other processes are
3666  * mapped to it, then create a new one.  The new vmspace is null.
3667  *
3668  * No requirements.
3669  */
3670 void
3671 vmspace_exec(struct proc *p, struct vmspace *vmcopy) 
3672 {
3673         struct vmspace *oldvmspace = p->p_vmspace;
3674         struct vmspace *newvmspace;
3675         vm_map_t map = &p->p_vmspace->vm_map;
3676
3677         /*
3678          * If we are execing a resident vmspace we fork it, otherwise
3679          * we create a new vmspace.  Note that exitingcnt and upcalls
3680          * are not copied to the new vmspace.
3681          */
3682         lwkt_gettoken(&oldvmspace->vm_map.token);
3683         if (vmcopy)  {
3684                 newvmspace = vmspace_fork(vmcopy);
3685                 lwkt_gettoken(&newvmspace->vm_map.token);
3686         } else {
3687                 newvmspace = vmspace_alloc(map->min_offset, map->max_offset);
3688                 lwkt_gettoken(&newvmspace->vm_map.token);
3689                 bcopy(&oldvmspace->vm_startcopy, &newvmspace->vm_startcopy,
3690                       (caddr_t)&oldvmspace->vm_endcopy -
3691                        (caddr_t)&oldvmspace->vm_startcopy);
3692         }
3693
3694         /*
3695          * Finish initializing the vmspace before assigning it
3696          * to the process.  The vmspace will become the current vmspace
3697          * if p == curproc.
3698          */
3699         pmap_pinit2(vmspace_pmap(newvmspace));
3700         pmap_replacevm(p, newvmspace, 0);
3701         lwkt_reltoken(&newvmspace->vm_map.token);
3702         lwkt_reltoken(&oldvmspace->vm_map.token);
3703         vmspace_free(oldvmspace);
3704 }
3705
3706 /*
3707  * Unshare the specified VM space for forcing COW.  This
3708  * is called by rfork, for the (RFMEM|RFPROC) == 0 case.
3709  */
3710 void
3711 vmspace_unshare(struct proc *p) 
3712 {
3713         struct vmspace *oldvmspace = p->p_vmspace;
3714         struct vmspace *newvmspace;
3715
3716         lwkt_gettoken(&oldvmspace->vm_map.token);
3717         if (oldvmspace->vm_sysref.refcnt == 1) {
3718                 lwkt_reltoken(&oldvmspace->vm_map.token);
3719                 return;
3720         }
3721         newvmspace = vmspace_fork(oldvmspace);
3722         lwkt_gettoken(&newvmspace->vm_map.token);
3723         pmap_pinit2(vmspace_pmap(newvmspace));
3724         pmap_replacevm(p, newvmspace, 0);
3725         lwkt_reltoken(&newvmspace->vm_map.token);
3726         lwkt_reltoken(&oldvmspace->vm_map.token);
3727         vmspace_free(oldvmspace);
3728 }
3729
3730 /*
3731  * vm_map_hint: return the beginning of the best area suitable for
3732  * creating a new mapping with "prot" protection.
3733  *
3734  * No requirements.
3735  */
3736 vm_offset_t
3737 vm_map_hint(struct proc *p, vm_offset_t addr, vm_prot_t prot)
3738 {
3739         struct vmspace *vms = p->p_vmspace;
3740
3741         if (!randomize_mmap) {
3742                 /*
3743                  * Set a reasonable start point for the hint if it was
3744                  * not specified or if it falls within the heap space.
3745                  * Hinted mmap()s do not allocate out of the heap space.
3746                  */
3747                 if (addr == 0 ||
3748                     (addr >= round_page((vm_offset_t)vms->vm_taddr) &&
3749                      addr < round_page((vm_offset_t)vms->vm_daddr + maxdsiz))) {
3750                         addr = round_page((vm_offset_t)vms->vm_daddr + maxdsiz);
3751                 }
3752
3753                 return addr;
3754         }
3755
3756         if (addr != 0 && addr >= (vm_offset_t)vms->vm_daddr)
3757                 return addr;
3758
3759 #ifdef notyet
3760 #ifdef __i386__
3761         /*
3762          * If executable skip first two pages, otherwise start
3763          * after data + heap region.
3764          */
3765         if ((prot & VM_PROT_EXECUTE) &&
3766             ((vm_offset_t)vms->vm_daddr >= I386_MAX_EXE_ADDR)) {
3767                 addr = (PAGE_SIZE * 2) +
3768                     (karc4random() & (I386_MAX_EXE_ADDR / 2 - 1));
3769                 return (round_page(addr));
3770         }
3771 #endif /* __i386__ */
3772 #endif /* notyet */
3773
3774         addr = (vm_offset_t)vms->vm_daddr + MAXDSIZ;
3775         addr += karc4random() & (MIN((256 * 1024 * 1024), MAXDSIZ) - 1);
3776
3777         return (round_page(addr));
3778 }
3779
3780 /*
3781  * Finds the VM object, offset, and protection for a given virtual address
3782  * in the specified map, assuming a page fault of the type specified.
3783  *
3784  * Leaves the map in question locked for read; return values are guaranteed
3785  * until a vm_map_lookup_done call is performed.  Note that the map argument
3786  * is in/out; the returned map must be used in the call to vm_map_lookup_done.
3787  *
3788  * A handle (out_entry) is returned for use in vm_map_lookup_done, to make
3789  * that fast.
3790  *
3791  * If a lookup is requested with "write protection" specified, the map may
3792  * be changed to perform virtual copying operations, although the data
3793  * referenced will remain the same.
3794  *
3795  * No requirements.
3796  */
3797 int
3798 vm_map_lookup(vm_map_t *var_map,                /* IN/OUT */
3799               vm_offset_t vaddr,
3800               vm_prot_t fault_typea,
3801               vm_map_entry_t *out_entry,        /* OUT */
3802               vm_object_t *object,              /* OUT */
3803               vm_pindex_t *pindex,              /* OUT */
3804               vm_prot_t *out_prot,              /* OUT */
3805               boolean_t *wired)                 /* OUT */
3806 {
3807         vm_map_entry_t entry;
3808         vm_map_t map = *var_map;
3809         vm_prot_t prot;
3810         vm_prot_t fault_type = fault_typea;
3811         int use_read_lock = 1;
3812         int rv = KERN_SUCCESS;
3813
3814 RetryLookup:
3815         if (use_read_lock)
3816                 vm_map_lock_read(map);
3817         else
3818                 vm_map_lock(map);
3819
3820         /*
3821          * If the map has an interesting hint, try it before calling full
3822          * blown lookup routine.
3823          */
3824         entry = map->hint;
3825         cpu_ccfence();
3826         *out_entry = entry;
3827         *object = NULL;
3828
3829         if ((entry == &map->header) ||
3830             (vaddr < entry->start) || (vaddr >= entry->end)) {
3831                 vm_map_entry_t tmp_entry;
3832
3833                 /*
3834                  * Entry was either not a valid hint, or the vaddr was not
3835                  * contained in the entry, so do a full lookup.
3836                  */
3837                 if (!vm_map_lookup_entry(map, vaddr, &tmp_entry)) {
3838                         rv = KERN_INVALID_ADDRESS;
3839                         goto done;
3840                 }
3841
3842                 entry = tmp_entry;
3843                 *out_entry = entry;
3844         }
3845         
3846         /*
3847          * Handle submaps.
3848          */
3849         if (entry->maptype == VM_MAPTYPE_SUBMAP) {
3850                 vm_map_t old_map = map;
3851
3852                 *var_map = map = entry->object.sub_map;
3853                 if (use_read_lock)
3854                         vm_map_unlock_read(old_map);
3855                 else
3856                         vm_map_unlock(old_map);
3857                 use_read_lock = 1;
3858                 goto RetryLookup;
3859         }
3860
3861         /*
3862          * Check whether this task is allowed to have this page.
3863          * Note the special case for MAP_ENTRY_COW
3864          * pages with an override.  This is to implement a forced
3865          * COW for debuggers.
3866          */
3867
3868         if (fault_type & VM_PROT_OVERRIDE_WRITE)
3869                 prot = entry->max_protection;
3870         else
3871                 prot = entry->protection;
3872
3873         fault_type &= (VM_PROT_READ|VM_PROT_WRITE|VM_PROT_EXECUTE);
3874         if ((fault_type & prot) != fault_type) {
3875                 rv = KERN_PROTECTION_FAILURE;
3876                 goto done;
3877         }
3878
3879         if ((entry->eflags & MAP_ENTRY_USER_WIRED) &&
3880             (entry->eflags & MAP_ENTRY_COW) &&
3881             (fault_type & VM_PROT_WRITE) &&
3882             (fault_typea & VM_PROT_OVERRIDE_WRITE) == 0) {
3883                 rv = KERN_PROTECTION_FAILURE;
3884                 goto done;
3885         }
3886
3887         /*
3888          * If this page is not pageable, we have to get it for all possible
3889          * accesses.
3890          */
3891         *wired = (entry->wired_count != 0);
3892         if (*wired)
3893                 prot = fault_type = entry->protection;
3894
3895         /*
3896          * Virtual page tables may need to update the accessed (A) bit
3897          * in a page table entry.  Upgrade the fault to a write fault for
3898          * that case if the map will support it.  If the map does not support
3899          * it the page table entry simply will not be updated.
3900          */
3901         if (entry->maptype == VM_MAPTYPE_VPAGETABLE) {
3902                 if (prot & VM_PROT_WRITE)
3903                         fault_type |= VM_PROT_WRITE;
3904         }
3905
3906         /*
3907          * If the entry was copy-on-write, we either ...
3908          */
3909         if (entry->eflags & MAP_ENTRY_NEEDS_COPY) {
3910                 /*
3911                  * If we want to write the page, we may as well handle that
3912                  * now since we've got the map locked.
3913                  *
3914                  * If we don't need to write the page, we just demote the
3915                  * permissions allowed.
3916                  */
3917
3918                 if (fault_type & VM_PROT_WRITE) {
3919                         /*
3920                          * Make a new object, and place it in the object
3921                          * chain.  Note that no new references have appeared
3922                          * -- one just moved from the map to the new
3923                          * object.
3924                          */
3925
3926                         if (use_read_lock && vm_map_lock_upgrade(map)) {
3927                                 /* lost lock */
3928                                 use_read_lock = 0;
3929                                 goto RetryLookup;
3930                         }
3931                         use_read_lock = 0;
3932
3933                         vm_map_entry_shadow(entry, 0);
3934                 } else {
3935                         /*
3936                          * We're attempting to read a copy-on-write page --
3937                          * don't allow writes.
3938                          */
3939
3940                         prot &= ~VM_PROT_WRITE;
3941                 }
3942         }
3943
3944         /*
3945          * Create an object if necessary.
3946          */
3947         if (entry->object.vm_object == NULL && !map->system_map) {
3948                 if (use_read_lock && vm_map_lock_upgrade(map))  {
3949                         /* lost lock */
3950                         use_read_lock = 0;
3951                         goto RetryLookup;
3952                 }
3953                 use_read_lock = 0;
3954                 vm_map_entry_allocate_object(entry);
3955         }
3956
3957         /*
3958          * Return the object/offset from this entry.  If the entry was
3959          * copy-on-write or empty, it has been fixed up.
3960          */
3961
3962         *pindex = OFF_TO_IDX((vaddr - entry->start) + entry->offset);
3963         *object = entry->object.vm_object;
3964
3965         /*
3966          * Return whether this is the only map sharing this data.  On
3967          * success we return with a read lock held on the map.  On failure
3968          * we return with the map unlocked.
3969          */
3970         *out_prot = prot;
3971 done:
3972         if (rv == KERN_SUCCESS) {
3973                 if (use_read_lock == 0)
3974                         vm_map_lock_downgrade(map);
3975         } else if (use_read_lock) {
3976                 vm_map_unlock_read(map);
3977         } else {
3978                 vm_map_unlock(map);
3979         }
3980         return (rv);
3981 }
3982
3983 /*
3984  * Releases locks acquired by a vm_map_lookup()
3985  * (according to the handle returned by that lookup).
3986  *
3987  * No other requirements.
3988  */
3989 void
3990 vm_map_lookup_done(vm_map_t map, vm_map_entry_t entry, int count)
3991 {
3992         /*
3993          * Unlock the main-level map
3994          */
3995         vm_map_unlock_read(map);
3996         if (count)
3997                 vm_map_entry_release(count);
3998 }
3999
4000 #include "opt_ddb.h"
4001 #ifdef DDB
4002 #include <sys/kernel.h>
4003
4004 #include <ddb/ddb.h>
4005
4006 /*
4007  * Debugging only
4008  */
4009 DB_SHOW_COMMAND(map, vm_map_print)
4010 {
4011         static int nlines;
4012         /* XXX convert args. */
4013         vm_map_t map = (vm_map_t)addr;
4014         boolean_t full = have_addr;
4015
4016         vm_map_entry_t entry;
4017
4018         db_iprintf("Task map %p: pmap=%p, nentries=%d, version=%u\n",
4019             (void *)map,
4020             (void *)map->pmap, map->nentries, map->timestamp);
4021         nlines++;
4022
4023         if (!full && db_indent)
4024                 return;
4025
4026         db_indent += 2;
4027         for (entry = map->header.next; entry != &map->header;
4028             entry = entry->next) {
4029                 db_iprintf("map entry %p: start=%p, end=%p\n",
4030                     (void *)entry, (void *)entry->start, (void *)entry->end);
4031                 nlines++;
4032                 {
4033                         static char *inheritance_name[4] =
4034                         {"share", "copy", "none", "donate_copy"};
4035
4036                         db_iprintf(" prot=%x/%x/%s",
4037                             entry->protection,
4038                             entry->max_protection,
4039                             inheritance_name[(int)(unsigned char)entry->inheritance]);
4040                         if (entry->wired_count != 0)
4041                                 db_printf(", wired");
4042                 }
4043                 if (entry->maptype == VM_MAPTYPE_SUBMAP) {
4044                         /* XXX no %qd in kernel.  Truncate entry->offset. */
4045                         db_printf(", share=%p, offset=0x%lx\n",
4046                             (void *)entry->object.sub_map,
4047                             (long)entry->offset);
4048                         nlines++;