kernel - Fix VM deadlock in optimized file read path
[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. Neither the name of the University nor the names of its contributors
19  *    may be used to endorse or promote products derived from this software
20  *    without specific prior written permission.
21  *
22  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
23  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
24  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
25  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
26  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
27  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
28  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
29  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
30  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
31  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32  * SUCH DAMAGE.
33  *
34  *      from: @(#)vm_map.c      8.3 (Berkeley) 1/12/94
35  *
36  *
37  * Copyright (c) 1987, 1990 Carnegie-Mellon University.
38  * All rights reserved.
39  *
40  * Authors: Avadis Tevanian, Jr., Michael Wayne Young
41  *
42  * Permission to use, copy, modify and distribute this software and
43  * its documentation is hereby granted, provided that both the copyright
44  * notice and this permission notice appear in all copies of the
45  * software, derivative works or modified versions, and any portions
46  * thereof, and that both notices appear in supporting documentation.
47  *
48  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
49  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
50  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
51  *
52  * Carnegie Mellon requests users of this software to return to
53  *
54  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
55  *  School of Computer Science
56  *  Carnegie Mellon University
57  *  Pittsburgh PA 15213-3890
58  *
59  * any improvements or extensions that they make and grant Carnegie the
60  * rights to redistribute these changes.
61  *
62  * $FreeBSD: src/sys/vm/vm_map.c,v 1.187.2.19 2003/05/27 00:47:02 alc Exp $
63  */
64
65 /*
66  *      Virtual memory mapping module.
67  */
68
69 #include <sys/param.h>
70 #include <sys/systm.h>
71 #include <sys/kernel.h>
72 #include <sys/proc.h>
73 #include <sys/serialize.h>
74 #include <sys/lock.h>
75 #include <sys/vmmeter.h>
76 #include <sys/mman.h>
77 #include <sys/vnode.h>
78 #include <sys/resourcevar.h>
79 #include <sys/shm.h>
80 #include <sys/tree.h>
81 #include <sys/malloc.h>
82 #include <sys/objcache.h>
83
84 #include <vm/vm.h>
85 #include <vm/vm_param.h>
86 #include <vm/pmap.h>
87 #include <vm/vm_map.h>
88 #include <vm/vm_page.h>
89 #include <vm/vm_object.h>
90 #include <vm/vm_pager.h>
91 #include <vm/vm_kern.h>
92 #include <vm/vm_extern.h>
93 #include <vm/swap_pager.h>
94 #include <vm/vm_zone.h>
95
96 #include <sys/thread2.h>
97 #include <sys/random.h>
98 #include <sys/sysctl.h>
99
100 /*
101  * Virtual memory maps provide for the mapping, protection, and sharing
102  * of virtual memory objects.  In addition, this module provides for an
103  * efficient virtual copy of memory from one map to another.
104  *
105  * Synchronization is required prior to most operations.
106  *
107  * Maps consist of an ordered doubly-linked list of simple entries.
108  * A hint and a RB tree is used to speed-up lookups.
109  *
110  * Callers looking to modify maps specify start/end addresses which cause
111  * the related map entry to be clipped if necessary, and then later
112  * recombined if the pieces remained compatible.
113  *
114  * Virtual copy operations are performed by copying VM object references
115  * from one map to another, and then marking both regions as copy-on-write.
116  */
117 static __boolean_t vmspace_ctor(void *obj, void *privdata, int ocflags);
118 static void vmspace_dtor(void *obj, void *privdata);
119 static void vmspace_terminate(struct vmspace *vm, int final);
120
121 MALLOC_DEFINE(M_VMSPACE, "vmspace", "vmspace objcache backingstore");
122 static struct objcache *vmspace_cache;
123
124 /*
125  * per-cpu page table cross mappings are initialized in early boot
126  * and might require a considerable number of vm_map_entry structures.
127  */
128 #define MAPENTRYBSP_CACHE       (MAXCPU+1)
129 #define MAPENTRYAP_CACHE        8
130
131 static struct vm_zone mapentzone_store, mapzone_store;
132 static vm_zone_t mapentzone, mapzone;
133 static struct vm_object mapentobj, mapobj;
134
135 static struct vm_map_entry map_entry_init[MAX_MAPENT];
136 static struct vm_map_entry cpu_map_entry_init_bsp[MAPENTRYBSP_CACHE];
137 static struct vm_map_entry cpu_map_entry_init_ap[MAXCPU][MAPENTRYAP_CACHE];
138 static struct vm_map map_init[MAX_KMAP];
139
140 static int randomize_mmap;
141 SYSCTL_INT(_vm, OID_AUTO, randomize_mmap, CTLFLAG_RW, &randomize_mmap, 0,
142     "Randomize mmap offsets");
143 static int vm_map_relock_enable = 1;
144 SYSCTL_INT(_vm, OID_AUTO, map_relock_enable, CTLFLAG_RW,
145            &vm_map_relock_enable, 0, "Randomize mmap offsets");
146
147 static void vm_map_entry_shadow(vm_map_entry_t entry, int addref);
148 static vm_map_entry_t vm_map_entry_create(vm_map_t map, int *);
149 static void vm_map_entry_dispose (vm_map_t map, vm_map_entry_t entry, int *);
150 static void _vm_map_clip_end (vm_map_t, vm_map_entry_t, vm_offset_t, int *);
151 static void _vm_map_clip_start (vm_map_t, vm_map_entry_t, vm_offset_t, int *);
152 static void vm_map_entry_delete (vm_map_t, vm_map_entry_t, int *);
153 static void vm_map_entry_unwire (vm_map_t, vm_map_entry_t);
154 static void vm_map_copy_entry (vm_map_t, vm_map_t, vm_map_entry_t,
155                 vm_map_entry_t);
156 static void vm_map_unclip_range (vm_map_t map, vm_map_entry_t start_entry, vm_offset_t start, vm_offset_t end, int *count, int flags);
157
158 /*
159  * Initialize the vm_map module.  Must be called before any other vm_map
160  * routines.
161  *
162  * Map and entry structures are allocated from the general purpose
163  * memory pool with some exceptions:
164  *
165  *      - The kernel map is allocated statically.
166  *      - Initial kernel map entries are allocated out of a static pool.
167  *      - We must set ZONE_SPECIAL here or the early boot code can get
168  *        stuck if there are >63 cores.
169  *
170  *      These restrictions are necessary since malloc() uses the
171  *      maps and requires map entries.
172  *
173  * Called from the low level boot code only.
174  */
175 void
176 vm_map_startup(void)
177 {
178         mapzone = &mapzone_store;
179         zbootinit(mapzone, "MAP", sizeof (struct vm_map),
180                 map_init, MAX_KMAP);
181         mapentzone = &mapentzone_store;
182         zbootinit(mapentzone, "MAP ENTRY", sizeof (struct vm_map_entry),
183                   map_entry_init, MAX_MAPENT);
184         mapentzone_store.zflags |= ZONE_SPECIAL;
185 }
186
187 /*
188  * Called prior to any vmspace allocations.
189  *
190  * Called from the low level boot code only.
191  */
192 void
193 vm_init2(void) 
194 {
195         vmspace_cache = objcache_create_mbacked(M_VMSPACE,
196                                                 sizeof(struct vmspace),
197                                                 0, ncpus * 4,
198                                                 vmspace_ctor, vmspace_dtor,
199                                                 NULL);
200         zinitna(mapentzone, &mapentobj, NULL, 0, 0, 
201                 ZONE_USE_RESERVE | ZONE_SPECIAL, 1);
202         zinitna(mapzone, &mapobj, NULL, 0, 0, 0, 1);
203         pmap_init2();
204         vm_object_init2();
205 }
206
207 /*
208  * objcache support.  We leave the pmap root cached as long as possible
209  * for performance reasons.
210  */
211 static
212 __boolean_t
213 vmspace_ctor(void *obj, void *privdata, int ocflags)
214 {
215         struct vmspace *vm = obj;
216
217         bzero(vm, sizeof(*vm));
218         vm->vm_refcnt = (u_int)-1;
219
220         return 1;
221 }
222
223 static
224 void
225 vmspace_dtor(void *obj, void *privdata)
226 {
227         struct vmspace *vm = obj;
228
229         KKASSERT(vm->vm_refcnt == (u_int)-1);
230         pmap_puninit(vmspace_pmap(vm));
231 }
232
233 /*
234  * Red black tree functions
235  *
236  * The caller must hold the related map lock.
237  */
238 static int rb_vm_map_compare(vm_map_entry_t a, vm_map_entry_t b);
239 RB_GENERATE(vm_map_rb_tree, vm_map_entry, rb_entry, rb_vm_map_compare);
240
241 /* a->start is address, and the only field has to be initialized */
242 static int
243 rb_vm_map_compare(vm_map_entry_t a, vm_map_entry_t b)
244 {
245         if (a->start < b->start)
246                 return(-1);
247         else if (a->start > b->start)
248                 return(1);
249         return(0);
250 }
251
252 /*
253  * Initialize vmspace ref/hold counts vmspace0.  There is a holdcnt for
254  * every refcnt.
255  */
256 void
257 vmspace_initrefs(struct vmspace *vm)
258 {
259         vm->vm_refcnt = 1;
260         vm->vm_holdcnt = 1;
261 }
262
263 /*
264  * Allocate a vmspace structure, including a vm_map and pmap.
265  * Initialize numerous fields.  While the initial allocation is zerod,
266  * subsequence reuse from the objcache leaves elements of the structure
267  * intact (particularly the pmap), so portions must be zerod.
268  *
269  * Returns a referenced vmspace.
270  *
271  * No requirements.
272  */
273 struct vmspace *
274 vmspace_alloc(vm_offset_t min, vm_offset_t max)
275 {
276         struct vmspace *vm;
277
278         vm = objcache_get(vmspace_cache, M_WAITOK);
279
280         bzero(&vm->vm_startcopy,
281               (char *)&vm->vm_endcopy - (char *)&vm->vm_startcopy);
282         vm_map_init(&vm->vm_map, min, max, NULL);       /* initializes token */
283
284         /*
285          * NOTE: hold to acquires token for safety.
286          *
287          * On return vmspace is referenced (refs=1, hold=1).  That is,
288          * each refcnt also has a holdcnt.  There can be additional holds
289          * (holdcnt) above and beyond the refcnt.  Finalization is handled in
290          * two stages, one on refs 1->0, and the the second on hold 1->0.
291          */
292         KKASSERT(vm->vm_holdcnt == 0);
293         KKASSERT(vm->vm_refcnt == (u_int)-1);
294         vmspace_initrefs(vm);
295         vmspace_hold(vm);
296         pmap_pinit(vmspace_pmap(vm));           /* (some fields reused) */
297         vm->vm_map.pmap = vmspace_pmap(vm);     /* XXX */
298         vm->vm_shm = NULL;
299         vm->vm_flags = 0;
300         cpu_vmspace_alloc(vm);
301         vmspace_drop(vm);
302
303         return (vm);
304 }
305
306 /*
307  * NOTE: Can return -1 if the vmspace is exiting.
308  */
309 int
310 vmspace_getrefs(struct vmspace *vm)
311 {
312         return ((int)vm->vm_refcnt);
313 }
314
315 /*
316  * A vmspace object must already have a non-zero hold to be able to gain
317  * further holds on it.
318  */
319 static void
320 vmspace_hold_notoken(struct vmspace *vm)
321 {
322         KKASSERT(vm->vm_holdcnt != 0);
323         refcount_acquire(&vm->vm_holdcnt);
324 }
325
326 static void
327 vmspace_drop_notoken(struct vmspace *vm)
328 {
329         if (refcount_release(&vm->vm_holdcnt)) {
330                 if (vm->vm_refcnt == (u_int)-1) {
331                         vmspace_terminate(vm, 1);
332                 }
333         }
334 }
335
336 void
337 vmspace_hold(struct vmspace *vm)
338 {
339         vmspace_hold_notoken(vm);
340         lwkt_gettoken(&vm->vm_map.token);
341 }
342
343 void
344 vmspace_drop(struct vmspace *vm)
345 {
346         lwkt_reltoken(&vm->vm_map.token);
347         vmspace_drop_notoken(vm);
348 }
349
350 /*
351  * A vmspace object must not be in a terminated state to be able to obtain
352  * additional refs on it.
353  *
354  * Ref'ing a vmspace object also increments its hold count.
355  */
356 void
357 vmspace_ref(struct vmspace *vm)
358 {
359         KKASSERT((int)vm->vm_refcnt >= 0);
360         vmspace_hold_notoken(vm);
361         refcount_acquire(&vm->vm_refcnt);
362 }
363
364 /*
365  * Release a ref on the vmspace.  On the 1->0 transition we do stage-1
366  * termination of the vmspace.  Then, on the final drop of the hold we
367  * will do stage-2 final termination.
368  */
369 void
370 vmspace_rel(struct vmspace *vm)
371 {
372         if (refcount_release(&vm->vm_refcnt)) {
373                 vm->vm_refcnt = (u_int)-1;      /* no other refs possible */
374                 vmspace_terminate(vm, 0);
375         }
376         vmspace_drop_notoken(vm);
377 }
378
379 /*
380  * This is called during exit indicating that the vmspace is no
381  * longer in used by an exiting process, but the process has not yet
382  * been reaped.
383  *
384  * We release the refcnt but not the associated holdcnt.
385  *
386  * No requirements.
387  */
388 void
389 vmspace_relexit(struct vmspace *vm)
390 {
391         if (refcount_release(&vm->vm_refcnt)) {
392                 vm->vm_refcnt = (u_int)-1;      /* no other refs possible */
393                 vmspace_terminate(vm, 0);
394         }
395 }
396
397 /*
398  * Called during reap to disconnect the remainder of the vmspace from
399  * the process.  On the hold drop the vmspace termination is finalized.
400  *
401  * No requirements.
402  */
403 void
404 vmspace_exitfree(struct proc *p)
405 {
406         struct vmspace *vm;
407
408         vm = p->p_vmspace;
409         p->p_vmspace = NULL;
410         vmspace_drop_notoken(vm);
411 }
412
413 /*
414  * Called in two cases:
415  *
416  * (1) When the last refcnt is dropped and the vmspace becomes inactive,
417  *     called with final == 0.  refcnt will be (u_int)-1 at this point,
418  *     and holdcnt will still be non-zero.
419  *
420  * (2) When holdcnt becomes 0, called with final == 1.  There should no
421  *     longer be anyone with access to the vmspace.
422  *
423  * VMSPACE_EXIT1 flags the primary deactivation
424  * VMSPACE_EXIT2 flags the last reap
425  */
426 static void
427 vmspace_terminate(struct vmspace *vm, int final)
428 {
429         int count;
430
431         lwkt_gettoken(&vm->vm_map.token);
432         if (final == 0) {
433                 KKASSERT((vm->vm_flags & VMSPACE_EXIT1) == 0);
434
435                 /*
436                  * Get rid of most of the resources.  Leave the kernel pmap
437                  * intact.
438                  */
439                 vm->vm_flags |= VMSPACE_EXIT1;
440                 shmexit(vm);
441                 pmap_remove_pages(vmspace_pmap(vm), VM_MIN_USER_ADDRESS,
442                                   VM_MAX_USER_ADDRESS);
443                 vm_map_remove(&vm->vm_map, VM_MIN_USER_ADDRESS,
444                               VM_MAX_USER_ADDRESS);
445                 lwkt_reltoken(&vm->vm_map.token);
446         } else {
447                 KKASSERT((vm->vm_flags & VMSPACE_EXIT1) != 0);
448                 KKASSERT((vm->vm_flags & VMSPACE_EXIT2) == 0);
449
450                 /*
451                  * Get rid of remaining basic resources.
452                  */
453                 vm->vm_flags |= VMSPACE_EXIT2;
454                 cpu_vmspace_free(vm);
455                 shmexit(vm);
456
457                 /*
458                  * Lock the map, to wait out all other references to it.
459                  * Delete all of the mappings and pages they hold, then call
460                  * the pmap module to reclaim anything left.
461                  */
462                 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
463                 vm_map_lock(&vm->vm_map);
464                 vm_map_delete(&vm->vm_map, vm->vm_map.min_offset,
465                               vm->vm_map.max_offset, &count);
466                 vm_map_unlock(&vm->vm_map);
467                 vm_map_entry_release(count);
468
469                 lwkt_gettoken(&vmspace_pmap(vm)->pm_token);
470                 pmap_release(vmspace_pmap(vm));
471                 lwkt_reltoken(&vmspace_pmap(vm)->pm_token);
472                 lwkt_reltoken(&vm->vm_map.token);
473                 objcache_put(vmspace_cache, vm);
474         }
475 }
476
477 /*
478  * Swap useage is determined by taking the proportional swap used by
479  * VM objects backing the VM map.  To make up for fractional losses,
480  * if the VM object has any swap use at all the associated map entries
481  * count for at least 1 swap page.
482  *
483  * No requirements.
484  */
485 int
486 vmspace_swap_count(struct vmspace *vm)
487 {
488         vm_map_t map = &vm->vm_map;
489         vm_map_entry_t cur;
490         vm_object_t object;
491         int count = 0;
492         int n;
493
494         vmspace_hold(vm);
495         for (cur = map->header.next; cur != &map->header; cur = cur->next) {
496                 switch(cur->maptype) {
497                 case VM_MAPTYPE_NORMAL:
498                 case VM_MAPTYPE_VPAGETABLE:
499                         if ((object = cur->object.vm_object) == NULL)
500                                 break;
501                         if (object->swblock_count) {
502                                 n = (cur->end - cur->start) / PAGE_SIZE;
503                                 count += object->swblock_count *
504                                     SWAP_META_PAGES * n / object->size + 1;
505                         }
506                         break;
507                 default:
508                         break;
509                 }
510         }
511         vmspace_drop(vm);
512
513         return(count);
514 }
515
516 /*
517  * Calculate the approximate number of anonymous pages in use by
518  * this vmspace.  To make up for fractional losses, we count each
519  * VM object as having at least 1 anonymous page.
520  *
521  * No requirements.
522  */
523 int
524 vmspace_anonymous_count(struct vmspace *vm)
525 {
526         vm_map_t map = &vm->vm_map;
527         vm_map_entry_t cur;
528         vm_object_t object;
529         int count = 0;
530
531         vmspace_hold(vm);
532         for (cur = map->header.next; cur != &map->header; cur = cur->next) {
533                 switch(cur->maptype) {
534                 case VM_MAPTYPE_NORMAL:
535                 case VM_MAPTYPE_VPAGETABLE:
536                         if ((object = cur->object.vm_object) == NULL)
537                                 break;
538                         if (object->type != OBJT_DEFAULT &&
539                             object->type != OBJT_SWAP) {
540                                 break;
541                         }
542                         count += object->resident_page_count;
543                         break;
544                 default:
545                         break;
546                 }
547         }
548         vmspace_drop(vm);
549
550         return(count);
551 }
552
553 /*
554  * Creates and returns a new empty VM map with the given physical map
555  * structure, and having the given lower and upper address bounds.
556  *
557  * No requirements.
558  */
559 vm_map_t
560 vm_map_create(vm_map_t result, pmap_t pmap, vm_offset_t min, vm_offset_t max)
561 {
562         if (result == NULL)
563                 result = zalloc(mapzone);
564         vm_map_init(result, min, max, pmap);
565         return (result);
566 }
567
568 /*
569  * Initialize an existing vm_map structure such as that in the vmspace
570  * structure.  The pmap is initialized elsewhere.
571  *
572  * No requirements.
573  */
574 void
575 vm_map_init(struct vm_map *map, vm_offset_t min, vm_offset_t max, pmap_t pmap)
576 {
577         map->header.next = map->header.prev = &map->header;
578         RB_INIT(&map->rb_root);
579         map->nentries = 0;
580         map->size = 0;
581         map->system_map = 0;
582         map->min_offset = min;
583         map->max_offset = max;
584         map->pmap = pmap;
585         map->first_free = &map->header;
586         map->hint = &map->header;
587         map->timestamp = 0;
588         map->flags = 0;
589         lwkt_token_init(&map->token, "vm_map");
590         lockinit(&map->lock, "vm_maplk", (hz + 9) / 10, 0);
591 }
592
593 /*
594  * Shadow the vm_map_entry's object.  This typically needs to be done when
595  * a write fault is taken on an entry which had previously been cloned by
596  * fork().  The shared object (which might be NULL) must become private so
597  * we add a shadow layer above it.
598  *
599  * Object allocation for anonymous mappings is defered as long as possible.
600  * When creating a shadow, however, the underlying object must be instantiated
601  * so it can be shared.
602  *
603  * If the map segment is governed by a virtual page table then it is
604  * possible to address offsets beyond the mapped area.  Just allocate
605  * a maximally sized object for this case.
606  *
607  * If addref is non-zero an additional reference is added to the returned
608  * entry.  This mechanic exists because the additional reference might have
609  * to be added atomically and not after return to prevent a premature
610  * collapse.
611  *
612  * The vm_map must be exclusively locked.
613  * No other requirements.
614  */
615 static
616 void
617 vm_map_entry_shadow(vm_map_entry_t entry, int addref)
618 {
619         if (entry->maptype == VM_MAPTYPE_VPAGETABLE) {
620                 vm_object_shadow(&entry->object.vm_object, &entry->offset,
621                                  0x7FFFFFFF, addref);   /* XXX */
622         } else {
623                 vm_object_shadow(&entry->object.vm_object, &entry->offset,
624                                  atop(entry->end - entry->start), addref);
625         }
626         entry->eflags &= ~MAP_ENTRY_NEEDS_COPY;
627 }
628
629 /*
630  * Allocate an object for a vm_map_entry.
631  *
632  * Object allocation for anonymous mappings is defered as long as possible.
633  * This function is called when we can defer no longer, generally when a map
634  * entry might be split or forked or takes a page fault.
635  *
636  * If the map segment is governed by a virtual page table then it is
637  * possible to address offsets beyond the mapped area.  Just allocate
638  * a maximally sized object for this case.
639  *
640  * The vm_map must be exclusively locked.
641  * No other requirements.
642  */
643 void 
644 vm_map_entry_allocate_object(vm_map_entry_t entry)
645 {
646         vm_object_t obj;
647
648         if (entry->maptype == VM_MAPTYPE_VPAGETABLE) {
649                 obj = vm_object_allocate(OBJT_DEFAULT, 0x7FFFFFFF); /* XXX */
650         } else {
651                 obj = vm_object_allocate(OBJT_DEFAULT,
652                                          atop(entry->end - entry->start));
653         }
654         entry->object.vm_object = obj;
655         entry->offset = 0;
656 }
657
658 /*
659  * Set an initial negative count so the first attempt to reserve
660  * space preloads a bunch of vm_map_entry's for this cpu.  Also
661  * pre-allocate 2 vm_map_entries which will be needed by zalloc() to
662  * map a new page for vm_map_entry structures.  SMP systems are
663  * particularly sensitive.
664  *
665  * This routine is called in early boot so we cannot just call
666  * vm_map_entry_reserve().
667  *
668  * Called from the low level boot code only (for each cpu)
669  *
670  * WARNING! Take care not to have too-big a static/BSS structure here
671  *          as MAXCPU can be 256+, otherwise the loader's 64MB heap
672  *          can get blown out by the kernel plus the initrd image.
673  */
674 void
675 vm_map_entry_reserve_cpu_init(globaldata_t gd)
676 {
677         vm_map_entry_t entry;
678         int count;
679         int i;
680
681         gd->gd_vme_avail -= MAP_RESERVE_COUNT * 2;
682         if (gd->gd_cpuid == 0) {
683                 entry = &cpu_map_entry_init_bsp[0];
684                 count = MAPENTRYBSP_CACHE;
685         } else {
686                 entry = &cpu_map_entry_init_ap[gd->gd_cpuid][0];
687                 count = MAPENTRYAP_CACHE;
688         }
689         for (i = 0; i < count; ++i, ++entry) {
690                 entry->next = gd->gd_vme_base;
691                 gd->gd_vme_base = entry;
692         }
693 }
694
695 /*
696  * Reserves vm_map_entry structures so code later on can manipulate
697  * map_entry structures within a locked map without blocking trying
698  * to allocate a new vm_map_entry.
699  *
700  * No requirements.
701  */
702 int
703 vm_map_entry_reserve(int count)
704 {
705         struct globaldata *gd = mycpu;
706         vm_map_entry_t entry;
707
708         /*
709          * Make sure we have enough structures in gd_vme_base to handle
710          * the reservation request.
711          *
712          * The critical section protects access to the per-cpu gd.
713          */
714         crit_enter();
715         while (gd->gd_vme_avail < count) {
716                 entry = zalloc(mapentzone);
717                 entry->next = gd->gd_vme_base;
718                 gd->gd_vme_base = entry;
719                 ++gd->gd_vme_avail;
720         }
721         gd->gd_vme_avail -= count;
722         crit_exit();
723
724         return(count);
725 }
726
727 /*
728  * Releases previously reserved vm_map_entry structures that were not
729  * used.  If we have too much junk in our per-cpu cache clean some of
730  * it out.
731  *
732  * No requirements.
733  */
734 void
735 vm_map_entry_release(int count)
736 {
737         struct globaldata *gd = mycpu;
738         vm_map_entry_t entry;
739
740         crit_enter();
741         gd->gd_vme_avail += count;
742         while (gd->gd_vme_avail > MAP_RESERVE_SLOP) {
743                 entry = gd->gd_vme_base;
744                 KKASSERT(entry != NULL);
745                 gd->gd_vme_base = entry->next;
746                 --gd->gd_vme_avail;
747                 crit_exit();
748                 zfree(mapentzone, entry);
749                 crit_enter();
750         }
751         crit_exit();
752 }
753
754 /*
755  * Reserve map entry structures for use in kernel_map itself.  These
756  * entries have *ALREADY* been reserved on a per-cpu basis when the map
757  * was inited.  This function is used by zalloc() to avoid a recursion
758  * when zalloc() itself needs to allocate additional kernel memory.
759  *
760  * This function works like the normal reserve but does not load the
761  * vm_map_entry cache (because that would result in an infinite
762  * recursion).  Note that gd_vme_avail may go negative.  This is expected.
763  *
764  * Any caller of this function must be sure to renormalize after
765  * potentially eating entries to ensure that the reserve supply
766  * remains intact.
767  *
768  * No requirements.
769  */
770 int
771 vm_map_entry_kreserve(int count)
772 {
773         struct globaldata *gd = mycpu;
774
775         crit_enter();
776         gd->gd_vme_avail -= count;
777         crit_exit();
778         KASSERT(gd->gd_vme_base != NULL,
779                 ("no reserved entries left, gd_vme_avail = %d",
780                 gd->gd_vme_avail));
781         return(count);
782 }
783
784 /*
785  * Release previously reserved map entries for kernel_map.  We do not
786  * attempt to clean up like the normal release function as this would
787  * cause an unnecessary (but probably not fatal) deep procedure call.
788  *
789  * No requirements.
790  */
791 void
792 vm_map_entry_krelease(int count)
793 {
794         struct globaldata *gd = mycpu;
795
796         crit_enter();
797         gd->gd_vme_avail += count;
798         crit_exit();
799 }
800
801 /*
802  * Allocates a VM map entry for insertion.  No entry fields are filled in.
803  *
804  * The entries should have previously been reserved.  The reservation count
805  * is tracked in (*countp).
806  *
807  * No requirements.
808  */
809 static vm_map_entry_t
810 vm_map_entry_create(vm_map_t map, int *countp)
811 {
812         struct globaldata *gd = mycpu;
813         vm_map_entry_t entry;
814
815         KKASSERT(*countp > 0);
816         --*countp;
817         crit_enter();
818         entry = gd->gd_vme_base;
819         KASSERT(entry != NULL, ("gd_vme_base NULL! count %d", *countp));
820         gd->gd_vme_base = entry->next;
821         crit_exit();
822
823         return(entry);
824 }
825
826 /*
827  * Dispose of a vm_map_entry that is no longer being referenced.
828  *
829  * No requirements.
830  */
831 static void
832 vm_map_entry_dispose(vm_map_t map, vm_map_entry_t entry, int *countp)
833 {
834         struct globaldata *gd = mycpu;
835
836         KKASSERT(map->hint != entry);
837         KKASSERT(map->first_free != entry);
838
839         ++*countp;
840         crit_enter();
841         entry->next = gd->gd_vme_base;
842         gd->gd_vme_base = entry;
843         crit_exit();
844 }
845
846
847 /*
848  * Insert/remove entries from maps.
849  *
850  * The related map must be exclusively locked.
851  * The caller must hold map->token
852  * No other requirements.
853  */
854 static __inline void
855 vm_map_entry_link(vm_map_t map,
856                   vm_map_entry_t after_where,
857                   vm_map_entry_t entry)
858 {
859         ASSERT_VM_MAP_LOCKED(map);
860
861         map->nentries++;
862         entry->prev = after_where;
863         entry->next = after_where->next;
864         entry->next->prev = entry;
865         after_where->next = entry;
866         if (vm_map_rb_tree_RB_INSERT(&map->rb_root, entry))
867                 panic("vm_map_entry_link: dup addr map %p ent %p", map, entry);
868 }
869
870 static __inline void
871 vm_map_entry_unlink(vm_map_t map,
872                     vm_map_entry_t entry)
873 {
874         vm_map_entry_t prev;
875         vm_map_entry_t next;
876
877         ASSERT_VM_MAP_LOCKED(map);
878
879         if (entry->eflags & MAP_ENTRY_IN_TRANSITION) {
880                 panic("vm_map_entry_unlink: attempt to mess with "
881                       "locked entry! %p", entry);
882         }
883         prev = entry->prev;
884         next = entry->next;
885         next->prev = prev;
886         prev->next = next;
887         vm_map_rb_tree_RB_REMOVE(&map->rb_root, entry);
888         map->nentries--;
889 }
890
891 /*
892  * Finds the map entry containing (or immediately preceding) the specified
893  * address in the given map.  The entry is returned in (*entry).
894  *
895  * The boolean result indicates whether the address is actually contained
896  * in the map.
897  *
898  * The related map must be locked.
899  * No other requirements.
900  */
901 boolean_t
902 vm_map_lookup_entry(vm_map_t map, vm_offset_t address, vm_map_entry_t *entry)
903 {
904         vm_map_entry_t tmp;
905         vm_map_entry_t last;
906
907         ASSERT_VM_MAP_LOCKED(map);
908 #if 0
909         /*
910          * XXX TEMPORARILY DISABLED.  For some reason our attempt to revive
911          * the hint code with the red-black lookup meets with system crashes
912          * and lockups.  We do not yet know why.
913          *
914          * It is possible that the problem is related to the setting
915          * of the hint during map_entry deletion, in the code specified
916          * at the GGG comment later on in this file.
917          *
918          * YYY More likely it's because this function can be called with
919          * a shared lock on the map, resulting in map->hint updates possibly
920          * racing.  Fixed now but untested.
921          */
922         /*
923          * Quickly check the cached hint, there's a good chance of a match.
924          */
925         tmp = map->hint;
926         cpu_ccfence();
927         if (tmp != &map->header) {
928                 if (address >= tmp->start && address < tmp->end) {
929                         *entry = tmp;
930                         return(TRUE);
931                 }
932         }
933 #endif
934
935         /*
936          * Locate the record from the top of the tree.  'last' tracks the
937          * closest prior record and is returned if no match is found, which
938          * in binary tree terms means tracking the most recent right-branch
939          * taken.  If there is no prior record, &map->header is returned.
940          */
941         last = &map->header;
942         tmp = RB_ROOT(&map->rb_root);
943
944         while (tmp) {
945                 if (address >= tmp->start) {
946                         if (address < tmp->end) {
947                                 *entry = tmp;
948                                 map->hint = tmp;
949                                 return(TRUE);
950                         }
951                         last = tmp;
952                         tmp = RB_RIGHT(tmp, rb_entry);
953                 } else {
954                         tmp = RB_LEFT(tmp, rb_entry);
955                 }
956         }
957         *entry = last;
958         return (FALSE);
959 }
960
961 /*
962  * Inserts the given whole VM object into the target map at the specified
963  * address range.  The object's size should match that of the address range.
964  *
965  * The map must be exclusively locked.
966  * The object must be held.
967  * The caller must have reserved sufficient vm_map_entry structures.
968  *
969  * If object is non-NULL, ref count must be bumped by caller prior to
970  * making call to account for the new entry.
971  */
972 int
973 vm_map_insert(vm_map_t map, int *countp, void *map_object, void *map_aux,
974               vm_ooffset_t offset, vm_offset_t start, vm_offset_t end,
975               vm_maptype_t maptype,
976               vm_prot_t prot, vm_prot_t max, int cow)
977 {
978         vm_map_entry_t new_entry;
979         vm_map_entry_t prev_entry;
980         vm_map_entry_t temp_entry;
981         vm_eflags_t protoeflags;
982         int must_drop = 0;
983         vm_object_t object;
984
985         if (maptype == VM_MAPTYPE_UKSMAP)
986                 object = NULL;
987         else
988                 object = map_object;
989
990         ASSERT_VM_MAP_LOCKED(map);
991         if (object)
992                 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
993
994         /*
995          * Check that the start and end points are not bogus.
996          */
997         if ((start < map->min_offset) || (end > map->max_offset) ||
998             (start >= end))
999                 return (KERN_INVALID_ADDRESS);
1000
1001         /*
1002          * Find the entry prior to the proposed starting address; if it's part
1003          * of an existing entry, this range is bogus.
1004          */
1005         if (vm_map_lookup_entry(map, start, &temp_entry))
1006                 return (KERN_NO_SPACE);
1007
1008         prev_entry = temp_entry;
1009
1010         /*
1011          * Assert that the next entry doesn't overlap the end point.
1012          */
1013
1014         if ((prev_entry->next != &map->header) &&
1015             (prev_entry->next->start < end))
1016                 return (KERN_NO_SPACE);
1017
1018         protoeflags = 0;
1019
1020         if (cow & MAP_COPY_ON_WRITE)
1021                 protoeflags |= MAP_ENTRY_COW|MAP_ENTRY_NEEDS_COPY;
1022
1023         if (cow & MAP_NOFAULT) {
1024                 protoeflags |= MAP_ENTRY_NOFAULT;
1025
1026                 KASSERT(object == NULL,
1027                         ("vm_map_insert: paradoxical MAP_NOFAULT request"));
1028         }
1029         if (cow & MAP_DISABLE_SYNCER)
1030                 protoeflags |= MAP_ENTRY_NOSYNC;
1031         if (cow & MAP_DISABLE_COREDUMP)
1032                 protoeflags |= MAP_ENTRY_NOCOREDUMP;
1033         if (cow & MAP_IS_STACK)
1034                 protoeflags |= MAP_ENTRY_STACK;
1035         if (cow & MAP_IS_KSTACK)
1036                 protoeflags |= MAP_ENTRY_KSTACK;
1037
1038         lwkt_gettoken(&map->token);
1039
1040         if (object) {
1041                 /*
1042                  * When object is non-NULL, it could be shared with another
1043                  * process.  We have to set or clear OBJ_ONEMAPPING 
1044                  * appropriately.
1045                  *
1046                  * NOTE: This flag is only applicable to DEFAULT and SWAP
1047                  *       objects and will already be clear in other types
1048                  *       of objects, so a shared object lock is ok for
1049                  *       VNODE objects.
1050                  */
1051                 if ((object->ref_count > 1) || (object->shadow_count != 0)) {
1052                         vm_object_clear_flag(object, OBJ_ONEMAPPING);
1053                 }
1054         }
1055         else if ((prev_entry != &map->header) &&
1056                  (prev_entry->eflags == protoeflags) &&
1057                  (prev_entry->end == start) &&
1058                  (prev_entry->wired_count == 0) &&
1059                  prev_entry->maptype == maptype &&
1060                  maptype == VM_MAPTYPE_NORMAL &&
1061                  ((prev_entry->object.vm_object == NULL) ||
1062                   vm_object_coalesce(prev_entry->object.vm_object,
1063                                      OFF_TO_IDX(prev_entry->offset),
1064                                      (vm_size_t)(prev_entry->end - prev_entry->start),
1065                                      (vm_size_t)(end - prev_entry->end)))) {
1066                 /*
1067                  * We were able to extend the object.  Determine if we
1068                  * can extend the previous map entry to include the 
1069                  * new range as well.
1070                  */
1071                 if ((prev_entry->inheritance == VM_INHERIT_DEFAULT) &&
1072                     (prev_entry->protection == prot) &&
1073                     (prev_entry->max_protection == max)) {
1074                         map->size += (end - prev_entry->end);
1075                         prev_entry->end = end;
1076                         vm_map_simplify_entry(map, prev_entry, countp);
1077                         lwkt_reltoken(&map->token);
1078                         return (KERN_SUCCESS);
1079                 }
1080
1081                 /*
1082                  * If we can extend the object but cannot extend the
1083                  * map entry, we have to create a new map entry.  We
1084                  * must bump the ref count on the extended object to
1085                  * account for it.  object may be NULL.
1086                  *
1087                  * XXX if object is NULL should we set offset to 0 here ?
1088                  */
1089                 object = prev_entry->object.vm_object;
1090                 offset = prev_entry->offset +
1091                         (prev_entry->end - prev_entry->start);
1092                 if (object) {
1093                         vm_object_hold(object);
1094                         vm_object_chain_wait(object, 0);
1095                         vm_object_reference_locked(object);
1096                         must_drop = 1;
1097                         map_object = object;
1098                 }
1099         }
1100
1101         /*
1102          * NOTE: if conditionals fail, object can be NULL here.  This occurs
1103          * in things like the buffer map where we manage kva but do not manage
1104          * backing objects.
1105          */
1106
1107         /*
1108          * Create a new entry
1109          */
1110
1111         new_entry = vm_map_entry_create(map, countp);
1112         new_entry->start = start;
1113         new_entry->end = end;
1114
1115         new_entry->maptype = maptype;
1116         new_entry->eflags = protoeflags;
1117         new_entry->object.map_object = map_object;
1118         new_entry->aux.master_pde = 0;          /* in case size is different */
1119         new_entry->aux.map_aux = map_aux;
1120         new_entry->offset = offset;
1121
1122         new_entry->inheritance = VM_INHERIT_DEFAULT;
1123         new_entry->protection = prot;
1124         new_entry->max_protection = max;
1125         new_entry->wired_count = 0;
1126
1127         /*
1128          * Insert the new entry into the list
1129          */
1130
1131         vm_map_entry_link(map, prev_entry, new_entry);
1132         map->size += new_entry->end - new_entry->start;
1133
1134         /*
1135          * Update the free space hint.  Entries cannot overlap.
1136          * An exact comparison is needed to avoid matching
1137          * against the map->header.
1138          */
1139         if ((map->first_free == prev_entry) &&
1140             (prev_entry->end == new_entry->start)) {
1141                 map->first_free = new_entry;
1142         }
1143
1144 #if 0
1145         /*
1146          * Temporarily removed to avoid MAP_STACK panic, due to
1147          * MAP_STACK being a huge hack.  Will be added back in
1148          * when MAP_STACK (and the user stack mapping) is fixed.
1149          */
1150         /*
1151          * It may be possible to simplify the entry
1152          */
1153         vm_map_simplify_entry(map, new_entry, countp);
1154 #endif
1155
1156         /*
1157          * Try to pre-populate the page table.  Mappings governed by virtual
1158          * page tables cannot be prepopulated without a lot of work, so
1159          * don't try.
1160          */
1161         if ((cow & (MAP_PREFAULT|MAP_PREFAULT_PARTIAL)) &&
1162             maptype != VM_MAPTYPE_VPAGETABLE &&
1163             maptype != VM_MAPTYPE_UKSMAP) {
1164                 int dorelock = 0;
1165                 if (vm_map_relock_enable && (cow & MAP_PREFAULT_RELOCK)) {
1166                         dorelock = 1;
1167                         vm_object_lock_swap();
1168                         vm_object_drop(object);
1169                 }
1170                 pmap_object_init_pt(map->pmap, start, prot,
1171                                     object, OFF_TO_IDX(offset), end - start,
1172                                     cow & MAP_PREFAULT_PARTIAL);
1173                 if (dorelock) {
1174                         vm_object_hold(object);
1175                         vm_object_lock_swap();
1176                 }
1177         }
1178         if (must_drop)
1179                 vm_object_drop(object);
1180
1181         lwkt_reltoken(&map->token);
1182         return (KERN_SUCCESS);
1183 }
1184
1185 /*
1186  * Find sufficient space for `length' bytes in the given map, starting at
1187  * `start'.  Returns 0 on success, 1 on no space.
1188  *
1189  * This function will returned an arbitrarily aligned pointer.  If no
1190  * particular alignment is required you should pass align as 1.  Note that
1191  * the map may return PAGE_SIZE aligned pointers if all the lengths used in
1192  * the map are a multiple of PAGE_SIZE, even if you pass a smaller align
1193  * argument.
1194  *
1195  * 'align' should be a power of 2 but is not required to be.
1196  *
1197  * The map must be exclusively locked.
1198  * No other requirements.
1199  */
1200 int
1201 vm_map_findspace(vm_map_t map, vm_offset_t start, vm_size_t length,
1202                  vm_size_t align, int flags, vm_offset_t *addr)
1203 {
1204         vm_map_entry_t entry, next;
1205         vm_offset_t end;
1206         vm_offset_t align_mask;
1207
1208         if (start < map->min_offset)
1209                 start = map->min_offset;
1210         if (start > map->max_offset)
1211                 return (1);
1212
1213         /*
1214          * If the alignment is not a power of 2 we will have to use
1215          * a mod/division, set align_mask to a special value.
1216          */
1217         if ((align | (align - 1)) + 1 != (align << 1))
1218                 align_mask = (vm_offset_t)-1;
1219         else
1220                 align_mask = align - 1;
1221
1222         /*
1223          * Look for the first possible address; if there's already something
1224          * at this address, we have to start after it.
1225          */
1226         if (start == map->min_offset) {
1227                 if ((entry = map->first_free) != &map->header)
1228                         start = entry->end;
1229         } else {
1230                 vm_map_entry_t tmp;
1231
1232                 if (vm_map_lookup_entry(map, start, &tmp))
1233                         start = tmp->end;
1234                 entry = tmp;
1235         }
1236
1237         /*
1238          * Look through the rest of the map, trying to fit a new region in the
1239          * gap between existing regions, or after the very last region.
1240          */
1241         for (;; start = (entry = next)->end) {
1242                 /*
1243                  * Adjust the proposed start by the requested alignment,
1244                  * be sure that we didn't wrap the address.
1245                  */
1246                 if (align_mask == (vm_offset_t)-1)
1247                         end = ((start + align - 1) / align) * align;
1248                 else
1249                         end = (start + align_mask) & ~align_mask;
1250                 if (end < start)
1251                         return (1);
1252                 start = end;
1253                 /*
1254                  * Find the end of the proposed new region.  Be sure we didn't
1255                  * go beyond the end of the map, or wrap around the address.
1256                  * Then check to see if this is the last entry or if the 
1257                  * proposed end fits in the gap between this and the next
1258                  * entry.
1259                  */
1260                 end = start + length;
1261                 if (end > map->max_offset || end < start)
1262                         return (1);
1263                 next = entry->next;
1264
1265                 /*
1266                  * If the next entry's start address is beyond the desired
1267                  * end address we may have found a good entry.
1268                  *
1269                  * If the next entry is a stack mapping we do not map into
1270                  * the stack's reserved space.
1271                  *
1272                  * XXX continue to allow mapping into the stack's reserved
1273                  * space if doing a MAP_STACK mapping inside a MAP_STACK
1274                  * mapping, for backwards compatibility.  But the caller
1275                  * really should use MAP_STACK | MAP_TRYFIXED if they
1276                  * want to do that.
1277                  */
1278                 if (next == &map->header)
1279                         break;
1280                 if (next->start >= end) {
1281                         if ((next->eflags & MAP_ENTRY_STACK) == 0)
1282                                 break;
1283                         if (flags & MAP_STACK)
1284                                 break;
1285                         if (next->start - next->aux.avail_ssize >= end)
1286                                 break;
1287                 }
1288         }
1289         map->hint = entry;
1290
1291         /*
1292          * Grow the kernel_map if necessary.  pmap_growkernel() will panic
1293          * if it fails.  The kernel_map is locked and nothing can steal
1294          * our address space if pmap_growkernel() blocks.
1295          *
1296          * NOTE: This may be unconditionally called for kldload areas on
1297          *       x86_64 because these do not bump kernel_vm_end (which would
1298          *       fill 128G worth of page tables!).  Therefore we must not
1299          *       retry.
1300          */
1301         if (map == &kernel_map) {
1302                 vm_offset_t kstop;
1303
1304                 kstop = round_page(start + length);
1305                 if (kstop > kernel_vm_end)
1306                         pmap_growkernel(start, kstop);
1307         }
1308         *addr = start;
1309         return (0);
1310 }
1311
1312 /*
1313  * vm_map_find finds an unallocated region in the target address map with
1314  * the given length and allocates it.  The search is defined to be first-fit
1315  * from the specified address; the region found is returned in the same
1316  * parameter.
1317  *
1318  * If object is non-NULL, ref count must be bumped by caller
1319  * prior to making call to account for the new entry.
1320  *
1321  * No requirements.  This function will lock the map temporarily.
1322  */
1323 int
1324 vm_map_find(vm_map_t map, void *map_object, void *map_aux,
1325             vm_ooffset_t offset, vm_offset_t *addr,
1326             vm_size_t length, vm_size_t align,
1327             boolean_t fitit,
1328             vm_maptype_t maptype,
1329             vm_prot_t prot, vm_prot_t max,
1330             int cow)
1331 {
1332         vm_offset_t start;
1333         vm_object_t object;
1334         int result;
1335         int count;
1336
1337         if (maptype == VM_MAPTYPE_UKSMAP)
1338                 object = NULL;
1339         else
1340                 object = map_object;
1341
1342         start = *addr;
1343
1344         count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
1345         vm_map_lock(map);
1346         if (object)
1347                 vm_object_hold_shared(object);
1348         if (fitit) {
1349                 if (vm_map_findspace(map, start, length, align, 0, addr)) {
1350                         if (object)
1351                                 vm_object_drop(object);
1352                         vm_map_unlock(map);
1353                         vm_map_entry_release(count);
1354                         return (KERN_NO_SPACE);
1355                 }
1356                 start = *addr;
1357         }
1358         result = vm_map_insert(map, &count, map_object, map_aux,
1359                                offset, start, start + length,
1360                                maptype, prot, max, cow);
1361         if (object)
1362                 vm_object_drop(object);
1363         vm_map_unlock(map);
1364         vm_map_entry_release(count);
1365
1366         return (result);
1367 }
1368
1369 /*
1370  * Simplify the given map entry by merging with either neighbor.  This
1371  * routine also has the ability to merge with both neighbors.
1372  *
1373  * This routine guarentees that the passed entry remains valid (though
1374  * possibly extended).  When merging, this routine may delete one or
1375  * both neighbors.  No action is taken on entries which have their
1376  * in-transition flag set.
1377  *
1378  * The map must be exclusively locked.
1379  */
1380 void
1381 vm_map_simplify_entry(vm_map_t map, vm_map_entry_t entry, int *countp)
1382 {
1383         vm_map_entry_t next, prev;
1384         vm_size_t prevsize, esize;
1385
1386         if (entry->eflags & MAP_ENTRY_IN_TRANSITION) {
1387                 ++mycpu->gd_cnt.v_intrans_coll;
1388                 return;
1389         }
1390
1391         if (entry->maptype == VM_MAPTYPE_SUBMAP)
1392                 return;
1393         if (entry->maptype == VM_MAPTYPE_UKSMAP)
1394                 return;
1395
1396         prev = entry->prev;
1397         if (prev != &map->header) {
1398                 prevsize = prev->end - prev->start;
1399                 if ( (prev->end == entry->start) &&
1400                      (prev->maptype == entry->maptype) &&
1401                      (prev->object.vm_object == entry->object.vm_object) &&
1402                      (!prev->object.vm_object ||
1403                         (prev->offset + prevsize == entry->offset)) &&
1404                      (prev->eflags == entry->eflags) &&
1405                      (prev->protection == entry->protection) &&
1406                      (prev->max_protection == entry->max_protection) &&
1407                      (prev->inheritance == entry->inheritance) &&
1408                      (prev->wired_count == entry->wired_count)) {
1409                         if (map->first_free == prev)
1410                                 map->first_free = entry;
1411                         if (map->hint == prev)
1412                                 map->hint = entry;
1413                         vm_map_entry_unlink(map, prev);
1414                         entry->start = prev->start;
1415                         entry->offset = prev->offset;
1416                         if (prev->object.vm_object)
1417                                 vm_object_deallocate(prev->object.vm_object);
1418                         vm_map_entry_dispose(map, prev, countp);
1419                 }
1420         }
1421
1422         next = entry->next;
1423         if (next != &map->header) {
1424                 esize = entry->end - entry->start;
1425                 if ((entry->end == next->start) &&
1426                     (next->maptype == entry->maptype) &&
1427                     (next->object.vm_object == entry->object.vm_object) &&
1428                      (!entry->object.vm_object ||
1429                         (entry->offset + esize == next->offset)) &&
1430                     (next->eflags == entry->eflags) &&
1431                     (next->protection == entry->protection) &&
1432                     (next->max_protection == entry->max_protection) &&
1433                     (next->inheritance == entry->inheritance) &&
1434                     (next->wired_count == entry->wired_count)) {
1435                         if (map->first_free == next)
1436                                 map->first_free = entry;
1437                         if (map->hint == next)
1438                                 map->hint = entry;
1439                         vm_map_entry_unlink(map, next);
1440                         entry->end = next->end;
1441                         if (next->object.vm_object)
1442                                 vm_object_deallocate(next->object.vm_object);
1443                         vm_map_entry_dispose(map, next, countp);
1444                 }
1445         }
1446 }
1447
1448 /*
1449  * Asserts that the given entry begins at or after the specified address.
1450  * If necessary, it splits the entry into two.
1451  */
1452 #define vm_map_clip_start(map, entry, startaddr, countp)                \
1453 {                                                                       \
1454         if (startaddr > entry->start)                                   \
1455                 _vm_map_clip_start(map, entry, startaddr, countp);      \
1456 }
1457
1458 /*
1459  * This routine is called only when it is known that the entry must be split.
1460  *
1461  * The map must be exclusively locked.
1462  */
1463 static void
1464 _vm_map_clip_start(vm_map_t map, vm_map_entry_t entry, vm_offset_t start,
1465                    int *countp)
1466 {
1467         vm_map_entry_t new_entry;
1468
1469         /*
1470          * Split off the front portion -- note that we must insert the new
1471          * entry BEFORE this one, so that this entry has the specified
1472          * starting address.
1473          */
1474
1475         vm_map_simplify_entry(map, entry, countp);
1476
1477         /*
1478          * If there is no object backing this entry, we might as well create
1479          * one now.  If we defer it, an object can get created after the map
1480          * is clipped, and individual objects will be created for the split-up
1481          * map.  This is a bit of a hack, but is also about the best place to
1482          * put this improvement.
1483          */
1484         if (entry->object.vm_object == NULL && !map->system_map) {
1485                 vm_map_entry_allocate_object(entry);
1486         }
1487
1488         new_entry = vm_map_entry_create(map, countp);
1489         *new_entry = *entry;
1490
1491         new_entry->end = start;
1492         entry->offset += (start - entry->start);
1493         entry->start = start;
1494
1495         vm_map_entry_link(map, entry->prev, new_entry);
1496
1497         switch(entry->maptype) {
1498         case VM_MAPTYPE_NORMAL:
1499         case VM_MAPTYPE_VPAGETABLE:
1500                 if (new_entry->object.vm_object) {
1501                         vm_object_hold(new_entry->object.vm_object);
1502                         vm_object_chain_wait(new_entry->object.vm_object, 0);
1503                         vm_object_reference_locked(new_entry->object.vm_object);
1504                         vm_object_drop(new_entry->object.vm_object);
1505                 }
1506                 break;
1507         default:
1508                 break;
1509         }
1510 }
1511
1512 /*
1513  * Asserts that the given entry ends at or before the specified address.
1514  * If necessary, it splits the entry into two.
1515  *
1516  * The map must be exclusively locked.
1517  */
1518 #define vm_map_clip_end(map, entry, endaddr, countp)            \
1519 {                                                               \
1520         if (endaddr < entry->end)                               \
1521                 _vm_map_clip_end(map, entry, endaddr, countp);  \
1522 }
1523
1524 /*
1525  * This routine is called only when it is known that the entry must be split.
1526  *
1527  * The map must be exclusively locked.
1528  */
1529 static void
1530 _vm_map_clip_end(vm_map_t map, vm_map_entry_t entry, vm_offset_t end,
1531                  int *countp)
1532 {
1533         vm_map_entry_t new_entry;
1534
1535         /*
1536          * If there is no object backing this entry, we might as well create
1537          * one now.  If we defer it, an object can get created after the map
1538          * is clipped, and individual objects will be created for the split-up
1539          * map.  This is a bit of a hack, but is also about the best place to
1540          * put this improvement.
1541          */
1542
1543         if (entry->object.vm_object == NULL && !map->system_map) {
1544                 vm_map_entry_allocate_object(entry);
1545         }
1546
1547         /*
1548          * Create a new entry and insert it AFTER the specified entry
1549          */
1550
1551         new_entry = vm_map_entry_create(map, countp);
1552         *new_entry = *entry;
1553
1554         new_entry->start = entry->end = end;
1555         new_entry->offset += (end - entry->start);
1556
1557         vm_map_entry_link(map, entry, new_entry);
1558
1559         switch(entry->maptype) {
1560         case VM_MAPTYPE_NORMAL:
1561         case VM_MAPTYPE_VPAGETABLE:
1562                 if (new_entry->object.vm_object) {
1563                         vm_object_hold(new_entry->object.vm_object);
1564                         vm_object_chain_wait(new_entry->object.vm_object, 0);
1565                         vm_object_reference_locked(new_entry->object.vm_object);
1566                         vm_object_drop(new_entry->object.vm_object);
1567                 }
1568                 break;
1569         default:
1570                 break;
1571         }
1572 }
1573
1574 /*
1575  * Asserts that the starting and ending region addresses fall within the
1576  * valid range for the map.
1577  */
1578 #define VM_MAP_RANGE_CHECK(map, start, end)     \
1579 {                                               \
1580         if (start < vm_map_min(map))            \
1581                 start = vm_map_min(map);        \
1582         if (end > vm_map_max(map))              \
1583                 end = vm_map_max(map);          \
1584         if (start > end)                        \
1585                 start = end;                    \
1586 }
1587
1588 /*
1589  * Used to block when an in-transition collison occurs.  The map
1590  * is unlocked for the sleep and relocked before the return.
1591  */
1592 void
1593 vm_map_transition_wait(vm_map_t map)
1594 {
1595         tsleep_interlock(map, 0);
1596         vm_map_unlock(map);
1597         tsleep(map, PINTERLOCKED, "vment", 0);
1598         vm_map_lock(map);
1599 }
1600
1601 /*
1602  * When we do blocking operations with the map lock held it is
1603  * possible that a clip might have occured on our in-transit entry,
1604  * requiring an adjustment to the entry in our loop.  These macros
1605  * help the pageable and clip_range code deal with the case.  The
1606  * conditional costs virtually nothing if no clipping has occured.
1607  */
1608
1609 #define CLIP_CHECK_BACK(entry, save_start)              \
1610     do {                                                \
1611             while (entry->start != save_start) {        \
1612                     entry = entry->prev;                \
1613                     KASSERT(entry != &map->header, ("bad entry clip")); \
1614             }                                           \
1615     } while(0)
1616
1617 #define CLIP_CHECK_FWD(entry, save_end)                 \
1618     do {                                                \
1619             while (entry->end != save_end) {            \
1620                     entry = entry->next;                \
1621                     KASSERT(entry != &map->header, ("bad entry clip")); \
1622             }                                           \
1623     } while(0)
1624
1625
1626 /*
1627  * Clip the specified range and return the base entry.  The
1628  * range may cover several entries starting at the returned base
1629  * and the first and last entry in the covering sequence will be
1630  * properly clipped to the requested start and end address.
1631  *
1632  * If no holes are allowed you should pass the MAP_CLIP_NO_HOLES
1633  * flag.
1634  *
1635  * The MAP_ENTRY_IN_TRANSITION flag will be set for the entries
1636  * covered by the requested range.
1637  *
1638  * The map must be exclusively locked on entry and will remain locked
1639  * on return. If no range exists or the range contains holes and you
1640  * specified that no holes were allowed, NULL will be returned.  This
1641  * routine may temporarily unlock the map in order avoid a deadlock when
1642  * sleeping.
1643  */
1644 static
1645 vm_map_entry_t
1646 vm_map_clip_range(vm_map_t map, vm_offset_t start, vm_offset_t end, 
1647                   int *countp, int flags)
1648 {
1649         vm_map_entry_t start_entry;
1650         vm_map_entry_t entry;
1651
1652         /*
1653          * Locate the entry and effect initial clipping.  The in-transition
1654          * case does not occur very often so do not try to optimize it.
1655          */
1656 again:
1657         if (vm_map_lookup_entry(map, start, &start_entry) == FALSE)
1658                 return (NULL);
1659         entry = start_entry;
1660         if (entry->eflags & MAP_ENTRY_IN_TRANSITION) {
1661                 entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP;
1662                 ++mycpu->gd_cnt.v_intrans_coll;
1663                 ++mycpu->gd_cnt.v_intrans_wait;
1664                 vm_map_transition_wait(map);
1665                 /*
1666                  * entry and/or start_entry may have been clipped while
1667                  * we slept, or may have gone away entirely.  We have
1668                  * to restart from the lookup.
1669                  */
1670                 goto again;
1671         }
1672
1673         /*
1674          * Since we hold an exclusive map lock we do not have to restart
1675          * after clipping, even though clipping may block in zalloc.
1676          */
1677         vm_map_clip_start(map, entry, start, countp);
1678         vm_map_clip_end(map, entry, end, countp);
1679         entry->eflags |= MAP_ENTRY_IN_TRANSITION;
1680
1681         /*
1682          * Scan entries covered by the range.  When working on the next
1683          * entry a restart need only re-loop on the current entry which
1684          * we have already locked, since 'next' may have changed.  Also,
1685          * even though entry is safe, it may have been clipped so we
1686          * have to iterate forwards through the clip after sleeping.
1687          */
1688         while (entry->next != &map->header && entry->next->start < end) {
1689                 vm_map_entry_t next = entry->next;
1690
1691                 if (flags & MAP_CLIP_NO_HOLES) {
1692                         if (next->start > entry->end) {
1693                                 vm_map_unclip_range(map, start_entry,
1694                                         start, entry->end, countp, flags);
1695                                 return(NULL);
1696                         }
1697                 }
1698
1699                 if (next->eflags & MAP_ENTRY_IN_TRANSITION) {
1700                         vm_offset_t save_end = entry->end;
1701                         next->eflags |= MAP_ENTRY_NEEDS_WAKEUP;
1702                         ++mycpu->gd_cnt.v_intrans_coll;
1703                         ++mycpu->gd_cnt.v_intrans_wait;
1704                         vm_map_transition_wait(map);
1705
1706                         /*
1707                          * clips might have occured while we blocked.
1708                          */
1709                         CLIP_CHECK_FWD(entry, save_end);
1710                         CLIP_CHECK_BACK(start_entry, start);
1711                         continue;
1712                 }
1713                 /*
1714                  * No restart necessary even though clip_end may block, we
1715                  * are holding the map lock.
1716                  */
1717                 vm_map_clip_end(map, next, end, countp);
1718                 next->eflags |= MAP_ENTRY_IN_TRANSITION;
1719                 entry = next;
1720         }
1721         if (flags & MAP_CLIP_NO_HOLES) {
1722                 if (entry->end != end) {
1723                         vm_map_unclip_range(map, start_entry,
1724                                 start, entry->end, countp, flags);
1725                         return(NULL);
1726                 }
1727         }
1728         return(start_entry);
1729 }
1730
1731 /*
1732  * Undo the effect of vm_map_clip_range().  You should pass the same
1733  * flags and the same range that you passed to vm_map_clip_range().
1734  * This code will clear the in-transition flag on the entries and
1735  * wake up anyone waiting.  This code will also simplify the sequence
1736  * and attempt to merge it with entries before and after the sequence.
1737  *
1738  * The map must be locked on entry and will remain locked on return.
1739  *
1740  * Note that you should also pass the start_entry returned by
1741  * vm_map_clip_range().  However, if you block between the two calls
1742  * with the map unlocked please be aware that the start_entry may
1743  * have been clipped and you may need to scan it backwards to find
1744  * the entry corresponding with the original start address.  You are
1745  * responsible for this, vm_map_unclip_range() expects the correct
1746  * start_entry to be passed to it and will KASSERT otherwise.
1747  */
1748 static
1749 void
1750 vm_map_unclip_range(vm_map_t map, vm_map_entry_t start_entry,
1751                     vm_offset_t start, vm_offset_t end,
1752                     int *countp, int flags)
1753 {
1754         vm_map_entry_t entry;
1755
1756         entry = start_entry;
1757
1758         KASSERT(entry->start == start, ("unclip_range: illegal base entry"));
1759         while (entry != &map->header && entry->start < end) {
1760                 KASSERT(entry->eflags & MAP_ENTRY_IN_TRANSITION,
1761                         ("in-transition flag not set during unclip on: %p",
1762                         entry));
1763                 KASSERT(entry->end <= end,
1764                         ("unclip_range: tail wasn't clipped"));
1765                 entry->eflags &= ~MAP_ENTRY_IN_TRANSITION;
1766                 if (entry->eflags & MAP_ENTRY_NEEDS_WAKEUP) {
1767                         entry->eflags &= ~MAP_ENTRY_NEEDS_WAKEUP;
1768                         wakeup(map);
1769                 }
1770                 entry = entry->next;
1771         }
1772
1773         /*
1774          * Simplification does not block so there is no restart case.
1775          */
1776         entry = start_entry;
1777         while (entry != &map->header && entry->start < end) {
1778                 vm_map_simplify_entry(map, entry, countp);
1779                 entry = entry->next;
1780         }
1781 }
1782
1783 /*
1784  * Mark the given range as handled by a subordinate map.
1785  *
1786  * This range must have been created with vm_map_find(), and no other
1787  * operations may have been performed on this range prior to calling
1788  * vm_map_submap().
1789  *
1790  * Submappings cannot be removed.
1791  *
1792  * No requirements.
1793  */
1794 int
1795 vm_map_submap(vm_map_t map, vm_offset_t start, vm_offset_t end, vm_map_t submap)
1796 {
1797         vm_map_entry_t entry;
1798         int result = KERN_INVALID_ARGUMENT;
1799         int count;
1800
1801         count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
1802         vm_map_lock(map);
1803
1804         VM_MAP_RANGE_CHECK(map, start, end);
1805
1806         if (vm_map_lookup_entry(map, start, &entry)) {
1807                 vm_map_clip_start(map, entry, start, &count);
1808         } else {
1809                 entry = entry->next;
1810         }
1811
1812         vm_map_clip_end(map, entry, end, &count);
1813
1814         if ((entry->start == start) && (entry->end == end) &&
1815             ((entry->eflags & MAP_ENTRY_COW) == 0) &&
1816             (entry->object.vm_object == NULL)) {
1817                 entry->object.sub_map = submap;
1818                 entry->maptype = VM_MAPTYPE_SUBMAP;
1819                 result = KERN_SUCCESS;
1820         }
1821         vm_map_unlock(map);
1822         vm_map_entry_release(count);
1823
1824         return (result);
1825 }
1826
1827 /*
1828  * Sets the protection of the specified address region in the target map. 
1829  * If "set_max" is specified, the maximum protection is to be set;
1830  * otherwise, only the current protection is affected.
1831  *
1832  * The protection is not applicable to submaps, but is applicable to normal
1833  * maps and maps governed by virtual page tables.  For example, when operating
1834  * on a virtual page table our protection basically controls how COW occurs
1835  * on the backing object, whereas the virtual page table abstraction itself
1836  * is an abstraction for userland.
1837  *
1838  * No requirements.
1839  */
1840 int
1841 vm_map_protect(vm_map_t map, vm_offset_t start, vm_offset_t end,
1842                vm_prot_t new_prot, boolean_t set_max)
1843 {
1844         vm_map_entry_t current;
1845         vm_map_entry_t entry;
1846         int count;
1847
1848         count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
1849         vm_map_lock(map);
1850
1851         VM_MAP_RANGE_CHECK(map, start, end);
1852
1853         if (vm_map_lookup_entry(map, start, &entry)) {
1854                 vm_map_clip_start(map, entry, start, &count);
1855         } else {
1856                 entry = entry->next;
1857         }
1858
1859         /*
1860          * Make a first pass to check for protection violations.
1861          */
1862         current = entry;
1863         while ((current != &map->header) && (current->start < end)) {
1864                 if (current->maptype == VM_MAPTYPE_SUBMAP) {
1865                         vm_map_unlock(map);
1866                         vm_map_entry_release(count);
1867                         return (KERN_INVALID_ARGUMENT);
1868                 }
1869                 if ((new_prot & current->max_protection) != new_prot) {
1870                         vm_map_unlock(map);
1871                         vm_map_entry_release(count);
1872                         return (KERN_PROTECTION_FAILURE);
1873                 }
1874                 current = current->next;
1875         }
1876
1877         /*
1878          * Go back and fix up protections. [Note that clipping is not
1879          * necessary the second time.]
1880          */
1881         current = entry;
1882
1883         while ((current != &map->header) && (current->start < end)) {
1884                 vm_prot_t old_prot;
1885
1886                 vm_map_clip_end(map, current, end, &count);
1887
1888                 old_prot = current->protection;
1889                 if (set_max) {
1890                         current->protection =
1891                             (current->max_protection = new_prot) &
1892                             old_prot;
1893                 } else {
1894                         current->protection = new_prot;
1895                 }
1896
1897                 /*
1898                  * Update physical map if necessary. Worry about copy-on-write
1899                  * here -- CHECK THIS XXX
1900                  */
1901
1902                 if (current->protection != old_prot) {
1903 #define MASK(entry)     (((entry)->eflags & MAP_ENTRY_COW) ? ~VM_PROT_WRITE : \
1904                                                         VM_PROT_ALL)
1905
1906                         pmap_protect(map->pmap, current->start,
1907                             current->end,
1908                             current->protection & MASK(current));
1909 #undef  MASK
1910                 }
1911
1912                 vm_map_simplify_entry(map, current, &count);
1913
1914                 current = current->next;
1915         }
1916
1917         vm_map_unlock(map);
1918         vm_map_entry_release(count);
1919         return (KERN_SUCCESS);
1920 }
1921
1922 /*
1923  * This routine traverses a processes map handling the madvise
1924  * system call.  Advisories are classified as either those effecting
1925  * the vm_map_entry structure, or those effecting the underlying
1926  * objects.
1927  *
1928  * The <value> argument is used for extended madvise calls.
1929  *
1930  * No requirements.
1931  */
1932 int
1933 vm_map_madvise(vm_map_t map, vm_offset_t start, vm_offset_t end,
1934                int behav, off_t value)
1935 {
1936         vm_map_entry_t current, entry;
1937         int modify_map = 0;
1938         int error = 0;
1939         int count;
1940
1941         /*
1942          * Some madvise calls directly modify the vm_map_entry, in which case
1943          * we need to use an exclusive lock on the map and we need to perform 
1944          * various clipping operations.  Otherwise we only need a read-lock
1945          * on the map.
1946          */
1947
1948         count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
1949
1950         switch(behav) {
1951         case MADV_NORMAL:
1952         case MADV_SEQUENTIAL:
1953         case MADV_RANDOM:
1954         case MADV_NOSYNC:
1955         case MADV_AUTOSYNC:
1956         case MADV_NOCORE:
1957         case MADV_CORE:
1958         case MADV_SETMAP:
1959         case MADV_INVAL:
1960                 modify_map = 1;
1961                 vm_map_lock(map);
1962                 break;
1963         case MADV_WILLNEED:
1964         case MADV_DONTNEED:
1965         case MADV_FREE:
1966                 vm_map_lock_read(map);
1967                 break;
1968         default:
1969                 vm_map_entry_release(count);
1970                 return (EINVAL);
1971         }
1972
1973         /*
1974          * Locate starting entry and clip if necessary.
1975          */
1976
1977         VM_MAP_RANGE_CHECK(map, start, end);
1978
1979         if (vm_map_lookup_entry(map, start, &entry)) {
1980                 if (modify_map)
1981                         vm_map_clip_start(map, entry, start, &count);
1982         } else {
1983                 entry = entry->next;
1984         }
1985
1986         if (modify_map) {
1987                 /*
1988                  * madvise behaviors that are implemented in the vm_map_entry.
1989                  *
1990                  * We clip the vm_map_entry so that behavioral changes are
1991                  * limited to the specified address range.
1992                  */
1993                 for (current = entry;
1994                      (current != &map->header) && (current->start < end);
1995                      current = current->next
1996                 ) {
1997                         if (current->maptype == VM_MAPTYPE_SUBMAP)
1998                                 continue;
1999
2000                         vm_map_clip_end(map, current, end, &count);
2001
2002                         switch (behav) {
2003                         case MADV_NORMAL:
2004                                 vm_map_entry_set_behavior(current, MAP_ENTRY_BEHAV_NORMAL);
2005                                 break;
2006                         case MADV_SEQUENTIAL:
2007                                 vm_map_entry_set_behavior(current, MAP_ENTRY_BEHAV_SEQUENTIAL);
2008                                 break;
2009                         case MADV_RANDOM:
2010                                 vm_map_entry_set_behavior(current, MAP_ENTRY_BEHAV_RANDOM);
2011                                 break;
2012                         case MADV_NOSYNC:
2013                                 current->eflags |= MAP_ENTRY_NOSYNC;
2014                                 break;
2015                         case MADV_AUTOSYNC:
2016                                 current->eflags &= ~MAP_ENTRY_NOSYNC;
2017                                 break;
2018                         case MADV_NOCORE:
2019                                 current->eflags |= MAP_ENTRY_NOCOREDUMP;
2020                                 break;
2021                         case MADV_CORE:
2022                                 current->eflags &= ~MAP_ENTRY_NOCOREDUMP;
2023                                 break;
2024                         case MADV_INVAL:
2025                                 /*
2026                                  * Invalidate the related pmap entries, used
2027                                  * to flush portions of the real kernel's
2028                                  * pmap when the caller has removed or
2029                                  * modified existing mappings in a virtual
2030                                  * page table.
2031                                  */
2032                                 pmap_remove(map->pmap,
2033                                             current->start, current->end);
2034                                 break;
2035                         case MADV_SETMAP:
2036                                 /*
2037                                  * Set the page directory page for a map
2038                                  * governed by a virtual page table.  Mark
2039                                  * the entry as being governed by a virtual
2040                                  * page table if it is not.
2041                                  *
2042                                  * XXX the page directory page is stored
2043                                  * in the avail_ssize field if the map_entry.
2044                                  *
2045                                  * XXX the map simplification code does not
2046                                  * compare this field so weird things may
2047                                  * happen if you do not apply this function
2048                                  * to the entire mapping governed by the
2049                                  * virtual page table.
2050                                  */
2051                                 if (current->maptype != VM_MAPTYPE_VPAGETABLE) {
2052                                         error = EINVAL;
2053                                         break;
2054                                 }
2055                                 current->aux.master_pde = value;
2056                                 pmap_remove(map->pmap,
2057                                             current->start, current->end);
2058                                 break;
2059                         default:
2060                                 error = EINVAL;
2061                                 break;
2062                         }
2063                         vm_map_simplify_entry(map, current, &count);
2064                 }
2065                 vm_map_unlock(map);
2066         } else {
2067                 vm_pindex_t pindex;
2068                 int count;
2069
2070                 /*
2071                  * madvise behaviors that are implemented in the underlying
2072                  * vm_object.
2073                  *
2074                  * Since we don't clip the vm_map_entry, we have to clip
2075                  * the vm_object pindex and count.
2076                  *
2077                  * NOTE!  We currently do not support these functions on
2078                  * virtual page tables.
2079                  */
2080                 for (current = entry;
2081                      (current != &map->header) && (current->start < end);
2082                      current = current->next
2083                 ) {
2084                         vm_offset_t useStart;
2085
2086                         if (current->maptype != VM_MAPTYPE_NORMAL)
2087                                 continue;
2088
2089                         pindex = OFF_TO_IDX(current->offset);
2090                         count = atop(current->end - current->start);
2091                         useStart = current->start;
2092
2093                         if (current->start < start) {
2094                                 pindex += atop(start - current->start);
2095                                 count -= atop(start - current->start);
2096                                 useStart = start;
2097                         }
2098                         if (current->end > end)
2099                                 count -= atop(current->end - end);
2100
2101                         if (count <= 0)
2102                                 continue;
2103
2104                         vm_object_madvise(current->object.vm_object,
2105                                           pindex, count, behav);
2106
2107                         /*
2108                          * Try to populate the page table.  Mappings governed
2109                          * by virtual page tables cannot be pre-populated
2110                          * without a lot of work so don't try.
2111                          */
2112                         if (behav == MADV_WILLNEED &&
2113                             current->maptype != VM_MAPTYPE_VPAGETABLE) {
2114                                 pmap_object_init_pt(
2115                                     map->pmap, 
2116                                     useStart,
2117                                     current->protection,
2118                                     current->object.vm_object,
2119                                     pindex, 
2120                                     (count << PAGE_SHIFT),
2121                                     MAP_PREFAULT_MADVISE
2122                                 );
2123                         }
2124                 }
2125                 vm_map_unlock_read(map);
2126         }
2127         vm_map_entry_release(count);
2128         return(error);
2129 }       
2130
2131
2132 /*
2133  * Sets the inheritance of the specified address range in the target map.
2134  * Inheritance affects how the map will be shared with child maps at the
2135  * time of vm_map_fork.
2136  */
2137 int
2138 vm_map_inherit(vm_map_t map, vm_offset_t start, vm_offset_t end,
2139                vm_inherit_t new_inheritance)
2140 {
2141         vm_map_entry_t entry;
2142         vm_map_entry_t temp_entry;
2143         int count;
2144
2145         switch (new_inheritance) {
2146         case VM_INHERIT_NONE:
2147         case VM_INHERIT_COPY:
2148         case VM_INHERIT_SHARE:
2149                 break;
2150         default:
2151                 return (KERN_INVALID_ARGUMENT);
2152         }
2153
2154         count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
2155         vm_map_lock(map);
2156
2157         VM_MAP_RANGE_CHECK(map, start, end);
2158
2159         if (vm_map_lookup_entry(map, start, &temp_entry)) {
2160                 entry = temp_entry;
2161                 vm_map_clip_start(map, entry, start, &count);
2162         } else
2163                 entry = temp_entry->next;
2164
2165         while ((entry != &map->header) && (entry->start < end)) {
2166                 vm_map_clip_end(map, entry, end, &count);
2167
2168                 entry->inheritance = new_inheritance;
2169
2170                 vm_map_simplify_entry(map, entry, &count);
2171
2172                 entry = entry->next;
2173         }
2174         vm_map_unlock(map);
2175         vm_map_entry_release(count);
2176         return (KERN_SUCCESS);
2177 }
2178
2179 /*
2180  * Implement the semantics of mlock
2181  */
2182 int
2183 vm_map_unwire(vm_map_t map, vm_offset_t start, vm_offset_t real_end,
2184               boolean_t new_pageable)
2185 {
2186         vm_map_entry_t entry;
2187         vm_map_entry_t start_entry;
2188         vm_offset_t end;
2189         int rv = KERN_SUCCESS;
2190         int count;
2191
2192         count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
2193         vm_map_lock(map);
2194         VM_MAP_RANGE_CHECK(map, start, real_end);
2195         end = real_end;
2196
2197         start_entry = vm_map_clip_range(map, start, end, &count,
2198                                         MAP_CLIP_NO_HOLES);
2199         if (start_entry == NULL) {
2200                 vm_map_unlock(map);
2201                 vm_map_entry_release(count);
2202                 return (KERN_INVALID_ADDRESS);
2203         }
2204
2205         if (new_pageable == 0) {
2206                 entry = start_entry;
2207                 while ((entry != &map->header) && (entry->start < end)) {
2208                         vm_offset_t save_start;
2209                         vm_offset_t save_end;
2210
2211                         /*
2212                          * Already user wired or hard wired (trivial cases)
2213                          */
2214                         if (entry->eflags & MAP_ENTRY_USER_WIRED) {
2215                                 entry = entry->next;
2216                                 continue;
2217                         }
2218                         if (entry->wired_count != 0) {
2219                                 entry->wired_count++;
2220                                 entry->eflags |= MAP_ENTRY_USER_WIRED;
2221                                 entry = entry->next;
2222                                 continue;
2223                         }
2224
2225                         /*
2226                          * A new wiring requires instantiation of appropriate
2227                          * management structures and the faulting in of the
2228                          * page.
2229                          */
2230                         if (entry->maptype == VM_MAPTYPE_NORMAL ||
2231                             entry->maptype == VM_MAPTYPE_VPAGETABLE) {
2232                                 int copyflag = entry->eflags &
2233                                                MAP_ENTRY_NEEDS_COPY;
2234                                 if (copyflag && ((entry->protection &
2235                                                   VM_PROT_WRITE) != 0)) {
2236                                         vm_map_entry_shadow(entry, 0);
2237                                 } else if (entry->object.vm_object == NULL &&
2238                                            !map->system_map) {
2239                                         vm_map_entry_allocate_object(entry);
2240                                 }
2241                         }
2242                         entry->wired_count++;
2243                         entry->eflags |= MAP_ENTRY_USER_WIRED;
2244
2245                         /*
2246                          * Now fault in the area.  Note that vm_fault_wire()
2247                          * may release the map lock temporarily, it will be
2248                          * relocked on return.  The in-transition
2249                          * flag protects the entries. 
2250                          */
2251                         save_start = entry->start;
2252                         save_end = entry->end;
2253                         rv = vm_fault_wire(map, entry, TRUE, 0);
2254                         if (rv) {
2255                                 CLIP_CHECK_BACK(entry, save_start);
2256                                 for (;;) {
2257                                         KASSERT(entry->wired_count == 1, ("bad wired_count on entry"));
2258                                         entry->eflags &= ~MAP_ENTRY_USER_WIRED;
2259                                         entry->wired_count = 0;
2260                                         if (entry->end == save_end)
2261                                                 break;
2262                                         entry = entry->next;
2263                                         KASSERT(entry != &map->header, ("bad entry clip during backout"));
2264                                 }
2265                                 end = save_start;       /* unwire the rest */
2266                                 break;
2267                         }
2268                         /*
2269                          * note that even though the entry might have been
2270                          * clipped, the USER_WIRED flag we set prevents
2271                          * duplication so we do not have to do a 
2272                          * clip check.
2273                          */
2274                         entry = entry->next;
2275                 }
2276
2277                 /*
2278                  * If we failed fall through to the unwiring section to
2279                  * unwire what we had wired so far.  'end' has already
2280                  * been adjusted.
2281                  */
2282                 if (rv)
2283                         new_pageable = 1;
2284
2285                 /*
2286                  * start_entry might have been clipped if we unlocked the
2287                  * map and blocked.  No matter how clipped it has gotten
2288                  * there should be a fragment that is on our start boundary.
2289                  */
2290                 CLIP_CHECK_BACK(start_entry, start);
2291         }
2292
2293         /*
2294          * Deal with the unwiring case.
2295          */
2296         if (new_pageable) {
2297                 /*
2298                  * This is the unwiring case.  We must first ensure that the
2299                  * range to be unwired is really wired down.  We know there
2300                  * are no holes.
2301                  */
2302                 entry = start_entry;
2303                 while ((entry != &map->header) && (entry->start < end)) {
2304                         if ((entry->eflags & MAP_ENTRY_USER_WIRED) == 0) {
2305                                 rv = KERN_INVALID_ARGUMENT;
2306                                 goto done;
2307                         }
2308                         KASSERT(entry->wired_count != 0, ("wired count was 0 with USER_WIRED set! %p", entry));
2309                         entry = entry->next;
2310                 }
2311
2312                 /*
2313                  * Now decrement the wiring count for each region. If a region
2314                  * becomes completely unwired, unwire its physical pages and
2315                  * mappings.
2316                  */
2317                 /*
2318                  * The map entries are processed in a loop, checking to
2319                  * make sure the entry is wired and asserting it has a wired
2320                  * count. However, another loop was inserted more-or-less in
2321                  * the middle of the unwiring path. This loop picks up the
2322                  * "entry" loop variable from the first loop without first
2323                  * setting it to start_entry. Naturally, the secound loop
2324                  * is never entered and the pages backing the entries are
2325                  * never unwired. This can lead to a leak of wired pages.
2326                  */
2327                 entry = start_entry;
2328                 while ((entry != &map->header) && (entry->start < end)) {
2329                         KASSERT(entry->eflags & MAP_ENTRY_USER_WIRED,
2330                                 ("expected USER_WIRED on entry %p", entry));
2331                         entry->eflags &= ~MAP_ENTRY_USER_WIRED;
2332                         entry->wired_count--;
2333                         if (entry->wired_count == 0)
2334                                 vm_fault_unwire(map, entry);
2335                         entry = entry->next;
2336                 }
2337         }
2338 done:
2339         vm_map_unclip_range(map, start_entry, start, real_end, &count,
2340                 MAP_CLIP_NO_HOLES);
2341         map->timestamp++;
2342         vm_map_unlock(map);
2343         vm_map_entry_release(count);
2344         return (rv);
2345 }
2346
2347 /*
2348  * Sets the pageability of the specified address range in the target map.
2349  * Regions specified as not pageable require locked-down physical
2350  * memory and physical page maps.
2351  *
2352  * The map must not be locked, but a reference must remain to the map
2353  * throughout the call.
2354  *
2355  * This function may be called via the zalloc path and must properly
2356  * reserve map entries for kernel_map.
2357  *
2358  * No requirements.
2359  */
2360 int
2361 vm_map_wire(vm_map_t map, vm_offset_t start, vm_offset_t real_end, int kmflags)
2362 {
2363         vm_map_entry_t entry;
2364         vm_map_entry_t start_entry;
2365         vm_offset_t end;
2366         int rv = KERN_SUCCESS;
2367         int count;
2368
2369         if (kmflags & KM_KRESERVE)
2370                 count = vm_map_entry_kreserve(MAP_RESERVE_COUNT);
2371         else
2372                 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
2373         vm_map_lock(map);
2374         VM_MAP_RANGE_CHECK(map, start, real_end);
2375         end = real_end;
2376
2377         start_entry = vm_map_clip_range(map, start, end, &count,
2378                                         MAP_CLIP_NO_HOLES);
2379         if (start_entry == NULL) {
2380                 vm_map_unlock(map);
2381                 rv = KERN_INVALID_ADDRESS;
2382                 goto failure;
2383         }
2384         if ((kmflags & KM_PAGEABLE) == 0) {
2385                 /*
2386                  * Wiring.  
2387                  *
2388                  * 1.  Holding the write lock, we create any shadow or zero-fill
2389                  * objects that need to be created. Then we clip each map
2390                  * entry to the region to be wired and increment its wiring
2391                  * count.  We create objects before clipping the map entries
2392                  * to avoid object proliferation.
2393                  *
2394                  * 2.  We downgrade to a read lock, and call vm_fault_wire to
2395                  * fault in the pages for any newly wired area (wired_count is
2396                  * 1).
2397                  *
2398                  * Downgrading to a read lock for vm_fault_wire avoids a 
2399                  * possible deadlock with another process that may have faulted
2400                  * on one of the pages to be wired (it would mark the page busy,
2401                  * blocking us, then in turn block on the map lock that we
2402                  * hold).  Because of problems in the recursive lock package,
2403                  * we cannot upgrade to a write lock in vm_map_lookup.  Thus,
2404                  * any actions that require the write lock must be done
2405                  * beforehand.  Because we keep the read lock on the map, the
2406                  * copy-on-write status of the entries we modify here cannot
2407                  * change.
2408                  */
2409                 entry = start_entry;
2410                 while ((entry != &map->header) && (entry->start < end)) {
2411                         /*
2412                          * Trivial case if the entry is already wired
2413                          */
2414                         if (entry->wired_count) {
2415                                 entry->wired_count++;
2416                                 entry = entry->next;
2417                                 continue;
2418                         }
2419
2420                         /*
2421                          * The entry is being newly wired, we have to setup
2422                          * appropriate management structures.  A shadow 
2423                          * object is required for a copy-on-write region,
2424                          * or a normal object for a zero-fill region.  We
2425                          * do not have to do this for entries that point to sub
2426                          * maps because we won't hold the lock on the sub map.
2427                          */
2428                         if (entry->maptype == VM_MAPTYPE_NORMAL ||
2429                             entry->maptype == VM_MAPTYPE_VPAGETABLE) {
2430                                 int copyflag = entry->eflags &
2431                                                MAP_ENTRY_NEEDS_COPY;
2432                                 if (copyflag && ((entry->protection &
2433                                                   VM_PROT_WRITE) != 0)) {
2434                                         vm_map_entry_shadow(entry, 0);
2435                                 } else if (entry->object.vm_object == NULL &&
2436                                            !map->system_map) {
2437                                         vm_map_entry_allocate_object(entry);
2438                                 }
2439                         }
2440
2441                         entry->wired_count++;
2442                         entry = entry->next;
2443                 }
2444
2445                 /*
2446                  * Pass 2.
2447                  */
2448
2449                 /*
2450                  * HACK HACK HACK HACK
2451                  *
2452                  * vm_fault_wire() temporarily unlocks the map to avoid
2453                  * deadlocks.  The in-transition flag from vm_map_clip_range
2454                  * call should protect us from changes while the map is
2455                  * unlocked.  T
2456                  *
2457                  * NOTE: Previously this comment stated that clipping might
2458                  *       still occur while the entry is unlocked, but from
2459                  *       what I can tell it actually cannot.
2460                  *
2461                  *       It is unclear whether the CLIP_CHECK_*() calls
2462                  *       are still needed but we keep them in anyway.
2463                  *
2464                  * HACK HACK HACK HACK
2465                  */
2466
2467                 entry = start_entry;
2468                 while (entry != &map->header && entry->start < end) {
2469                         /*
2470                          * If vm_fault_wire fails for any page we need to undo
2471                          * what has been done.  We decrement the wiring count
2472                          * for those pages which have not yet been wired (now)
2473                          * and unwire those that have (later).
2474                          */
2475                         vm_offset_t save_start = entry->start;
2476                         vm_offset_t save_end = entry->end;
2477
2478                         if (entry->wired_count == 1)
2479                                 rv = vm_fault_wire(map, entry, FALSE, kmflags);
2480                         if (rv) {
2481                                 CLIP_CHECK_BACK(entry, save_start);
2482                                 for (;;) {
2483                                         KASSERT(entry->wired_count == 1, ("wired_count changed unexpectedly"));
2484                                         entry->wired_count = 0;
2485                                         if (entry->end == save_end)
2486                                                 break;
2487                                         entry = entry->next;
2488                                         KASSERT(entry != &map->header, ("bad entry clip during backout"));
2489                                 }
2490                                 end = save_start;
2491                                 break;
2492                         }
2493                         CLIP_CHECK_FWD(entry, save_end);
2494                         entry = entry->next;
2495                 }
2496
2497                 /*
2498                  * If a failure occured undo everything by falling through
2499                  * to the unwiring code.  'end' has already been adjusted
2500                  * appropriately.
2501                  */
2502                 if (rv)
2503                         kmflags |= KM_PAGEABLE;
2504
2505                 /*
2506                  * start_entry is still IN_TRANSITION but may have been 
2507                  * clipped since vm_fault_wire() unlocks and relocks the
2508                  * map.  No matter how clipped it has gotten there should
2509                  * be a fragment that is on our start boundary.
2510                  */
2511                 CLIP_CHECK_BACK(start_entry, start);
2512         }
2513
2514         if (kmflags & KM_PAGEABLE) {
2515                 /*
2516                  * This is the unwiring case.  We must first ensure that the
2517                  * range to be unwired is really wired down.  We know there
2518                  * are no holes.
2519                  */
2520                 entry = start_entry;
2521                 while ((entry != &map->header) && (entry->start < end)) {
2522                         if (entry->wired_count == 0) {
2523                                 rv = KERN_INVALID_ARGUMENT;
2524                                 goto done;
2525                         }
2526                         entry = entry->next;
2527                 }
2528
2529                 /*
2530                  * Now decrement the wiring count for each region. If a region
2531                  * becomes completely unwired, unwire its physical pages and
2532                  * mappings.
2533                  */
2534                 entry = start_entry;
2535                 while ((entry != &map->header) && (entry->start < end)) {
2536                         entry->wired_count--;
2537                         if (entry->wired_count == 0)
2538                                 vm_fault_unwire(map, entry);
2539                         entry = entry->next;
2540                 }
2541         }
2542 done:
2543         vm_map_unclip_range(map, start_entry, start, real_end,
2544                             &count, MAP_CLIP_NO_HOLES);
2545         map->timestamp++;
2546         vm_map_unlock(map);
2547 failure:
2548         if (kmflags & KM_KRESERVE)
2549                 vm_map_entry_krelease(count);
2550         else
2551                 vm_map_entry_release(count);
2552         return (rv);
2553 }
2554
2555 /*
2556  * Mark a newly allocated address range as wired but do not fault in
2557  * the pages.  The caller is expected to load the pages into the object.
2558  *
2559  * The map must be locked on entry and will remain locked on return.
2560  * No other requirements.
2561  */
2562 void
2563 vm_map_set_wired_quick(vm_map_t map, vm_offset_t addr, vm_size_t size,
2564                        int *countp)
2565 {
2566         vm_map_entry_t scan;
2567         vm_map_entry_t entry;
2568
2569         entry = vm_map_clip_range(map, addr, addr + size,
2570                                   countp, MAP_CLIP_NO_HOLES);
2571         for (scan = entry;
2572              scan != &map->header && scan->start < addr + size;
2573              scan = scan->next) {
2574             KKASSERT(scan->wired_count == 0);
2575             scan->wired_count = 1;
2576         }
2577         vm_map_unclip_range(map, entry, addr, addr + size,
2578                             countp, MAP_CLIP_NO_HOLES);
2579 }
2580
2581 /*
2582  * Push any dirty cached pages in the address range to their pager.
2583  * If syncio is TRUE, dirty pages are written synchronously.
2584  * If invalidate is TRUE, any cached pages are freed as well.
2585  *
2586  * This routine is called by sys_msync()
2587  *
2588  * Returns an error if any part of the specified range is not mapped.
2589  *
2590  * No requirements.
2591  */
2592 int
2593 vm_map_clean(vm_map_t map, vm_offset_t start, vm_offset_t end,
2594              boolean_t syncio, boolean_t invalidate)
2595 {
2596         vm_map_entry_t current;
2597         vm_map_entry_t entry;
2598         vm_size_t size;
2599         vm_object_t object;
2600         vm_object_t tobj;
2601         vm_ooffset_t offset;
2602
2603         vm_map_lock_read(map);
2604         VM_MAP_RANGE_CHECK(map, start, end);
2605         if (!vm_map_lookup_entry(map, start, &entry)) {
2606                 vm_map_unlock_read(map);
2607                 return (KERN_INVALID_ADDRESS);
2608         }
2609         lwkt_gettoken(&map->token);
2610
2611         /*
2612          * Make a first pass to check for holes.
2613          */
2614         for (current = entry; current->start < end; current = current->next) {
2615                 if (current->maptype == VM_MAPTYPE_SUBMAP) {
2616                         lwkt_reltoken(&map->token);
2617                         vm_map_unlock_read(map);
2618                         return (KERN_INVALID_ARGUMENT);
2619                 }
2620                 if (end > current->end &&
2621                     (current->next == &map->header ||
2622                         current->end != current->next->start)) {
2623                         lwkt_reltoken(&map->token);
2624                         vm_map_unlock_read(map);
2625                         return (KERN_INVALID_ADDRESS);
2626                 }
2627         }
2628
2629         if (invalidate)
2630                 pmap_remove(vm_map_pmap(map), start, end);
2631
2632         /*
2633          * Make a second pass, cleaning/uncaching pages from the indicated
2634          * objects as we go.
2635          */
2636         for (current = entry; current->start < end; current = current->next) {
2637                 offset = current->offset + (start - current->start);
2638                 size = (end <= current->end ? end : current->end) - start;
2639
2640                 switch(current->maptype) {
2641                 case VM_MAPTYPE_SUBMAP:
2642                 {
2643                         vm_map_t smap;
2644                         vm_map_entry_t tentry;
2645                         vm_size_t tsize;
2646
2647                         smap = current->object.sub_map;
2648                         vm_map_lock_read(smap);
2649                         vm_map_lookup_entry(smap, offset, &tentry);
2650                         tsize = tentry->end - offset;
2651                         if (tsize < size)
2652                                 size = tsize;
2653                         object = tentry->object.vm_object;
2654                         offset = tentry->offset + (offset - tentry->start);
2655                         vm_map_unlock_read(smap);
2656                         break;
2657                 }
2658                 case VM_MAPTYPE_NORMAL:
2659                 case VM_MAPTYPE_VPAGETABLE:
2660                         object = current->object.vm_object;
2661                         break;
2662                 default:
2663                         object = NULL;
2664                         break;
2665                 }
2666
2667                 if (object)
2668                         vm_object_hold(object);
2669
2670                 /*
2671                  * Note that there is absolutely no sense in writing out
2672                  * anonymous objects, so we track down the vnode object
2673                  * to write out.
2674                  * We invalidate (remove) all pages from the address space
2675                  * anyway, for semantic correctness.
2676                  *
2677                  * note: certain anonymous maps, such as MAP_NOSYNC maps,
2678                  * may start out with a NULL object.
2679                  */
2680                 while (object && (tobj = object->backing_object) != NULL) {
2681                         vm_object_hold(tobj);
2682                         if (tobj == object->backing_object) {
2683                                 vm_object_lock_swap();
2684                                 offset += object->backing_object_offset;
2685                                 vm_object_drop(object);
2686                                 object = tobj;
2687                                 if (object->size < OFF_TO_IDX(offset + size))
2688                                         size = IDX_TO_OFF(object->size) -
2689                                                offset;
2690                                 break;
2691                         }
2692                         vm_object_drop(tobj);
2693                 }
2694                 if (object && (object->type == OBJT_VNODE) && 
2695                     (current->protection & VM_PROT_WRITE) &&
2696                     (object->flags & OBJ_NOMSYNC) == 0) {
2697                         /*
2698                          * Flush pages if writing is allowed, invalidate them
2699                          * if invalidation requested.  Pages undergoing I/O
2700                          * will be ignored by vm_object_page_remove().
2701                          *
2702                          * We cannot lock the vnode and then wait for paging
2703                          * to complete without deadlocking against vm_fault.
2704                          * Instead we simply call vm_object_page_remove() and
2705                          * allow it to block internally on a page-by-page 
2706                          * basis when it encounters pages undergoing async 
2707                          * I/O.
2708                          */
2709                         int flags;
2710
2711                         /* no chain wait needed for vnode objects */
2712                         vm_object_reference_locked(object);
2713                         vn_lock(object->handle, LK_EXCLUSIVE | LK_RETRY);
2714                         flags = (syncio || invalidate) ? OBJPC_SYNC : 0;
2715                         flags |= invalidate ? OBJPC_INVAL : 0;
2716
2717                         /*
2718                          * When operating on a virtual page table just
2719                          * flush the whole object.  XXX we probably ought
2720                          * to 
2721                          */
2722                         switch(current->maptype) {
2723                         case VM_MAPTYPE_NORMAL:
2724                                 vm_object_page_clean(object,
2725                                     OFF_TO_IDX(offset),
2726                                     OFF_TO_IDX(offset + size + PAGE_MASK),
2727                                     flags);
2728                                 break;
2729                         case VM_MAPTYPE_VPAGETABLE:
2730                                 vm_object_page_clean(object, 0, 0, flags);
2731                                 break;
2732                         }
2733                         vn_unlock(((struct vnode *)object->handle));
2734                         vm_object_deallocate_locked(object);
2735                 }
2736                 if (object && invalidate &&
2737                    ((object->type == OBJT_VNODE) ||
2738                     (object->type == OBJT_DEVICE) ||
2739                     (object->type == OBJT_MGTDEVICE))) {
2740                         int clean_only = 
2741                                 ((object->type == OBJT_DEVICE) ||
2742                                 (object->type == OBJT_MGTDEVICE)) ? FALSE : TRUE;
2743                         /* no chain wait needed for vnode/device objects */
2744                         vm_object_reference_locked(object);
2745                         switch(current->maptype) {
2746                         case VM_MAPTYPE_NORMAL:
2747                                 vm_object_page_remove(object,
2748                                     OFF_TO_IDX(offset),
2749                                     OFF_TO_IDX(offset + size + PAGE_MASK),
2750                                     clean_only);
2751                                 break;
2752                         case VM_MAPTYPE_VPAGETABLE:
2753                                 vm_object_page_remove(object, 0, 0, clean_only);
2754                                 break;
2755                         }
2756                         vm_object_deallocate_locked(object);
2757                 }
2758                 start += size;
2759                 if (object)
2760                         vm_object_drop(object);
2761         }
2762
2763         lwkt_reltoken(&map->token);
2764         vm_map_unlock_read(map);
2765
2766         return (KERN_SUCCESS);
2767 }
2768
2769 /*
2770  * Make the region specified by this entry pageable.
2771  *
2772  * The vm_map must be exclusively locked.
2773  */
2774 static void 
2775 vm_map_entry_unwire(vm_map_t map, vm_map_entry_t entry)
2776 {
2777         entry->eflags &= ~MAP_ENTRY_USER_WIRED;
2778         entry->wired_count = 0;
2779         vm_fault_unwire(map, entry);
2780 }
2781
2782 /*
2783  * Deallocate the given entry from the target map.
2784  *
2785  * The vm_map must be exclusively locked.
2786  */
2787 static void
2788 vm_map_entry_delete(vm_map_t map, vm_map_entry_t entry, int *countp)
2789 {
2790         vm_map_entry_unlink(map, entry);
2791         map->size -= entry->end - entry->start;
2792
2793         switch(entry->maptype) {
2794         case VM_MAPTYPE_NORMAL:
2795         case VM_MAPTYPE_VPAGETABLE:
2796         case VM_MAPTYPE_SUBMAP:
2797                 vm_object_deallocate(entry->object.vm_object);
2798                 break;
2799         case VM_MAPTYPE_UKSMAP:
2800                 /* XXX TODO */
2801                 break;
2802         default:
2803                 break;
2804         }
2805
2806         vm_map_entry_dispose(map, entry, countp);
2807 }
2808
2809 /*
2810  * Deallocates the given address range from the target map.
2811  *
2812  * The vm_map must be exclusively locked.
2813  */
2814 int
2815 vm_map_delete(vm_map_t map, vm_offset_t start, vm_offset_t end, int *countp)
2816 {
2817         vm_object_t object;
2818         vm_map_entry_t entry;
2819         vm_map_entry_t first_entry;
2820
2821         ASSERT_VM_MAP_LOCKED(map);
2822         lwkt_gettoken(&map->token);
2823 again:
2824         /*
2825          * Find the start of the region, and clip it.  Set entry to point
2826          * at the first record containing the requested address or, if no
2827          * such record exists, the next record with a greater address.  The
2828          * loop will run from this point until a record beyond the termination
2829          * address is encountered.
2830          *
2831          * map->hint must be adjusted to not point to anything we delete,
2832          * so set it to the entry prior to the one being deleted.
2833          *
2834          * GGG see other GGG comment.
2835          */
2836         if (vm_map_lookup_entry(map, start, &first_entry)) {
2837                 entry = first_entry;
2838                 vm_map_clip_start(map, entry, start, countp);
2839                 map->hint = entry->prev;        /* possible problem XXX */
2840         } else {
2841                 map->hint = first_entry;        /* possible problem XXX */
2842                 entry = first_entry->next;
2843         }
2844
2845         /*
2846          * If a hole opens up prior to the current first_free then
2847          * adjust first_free.  As with map->hint, map->first_free
2848          * cannot be left set to anything we might delete.
2849          */
2850         if (entry == &map->header) {
2851                 map->first_free = &map->header;
2852         } else if (map->first_free->start >= start) {
2853                 map->first_free = entry->prev;
2854         }
2855
2856         /*
2857          * Step through all entries in this region
2858          */
2859         while ((entry != &map->header) && (entry->start < end)) {
2860                 vm_map_entry_t next;
2861                 vm_offset_t s, e;
2862                 vm_pindex_t offidxstart, offidxend, count;
2863
2864                 /*
2865                  * If we hit an in-transition entry we have to sleep and
2866                  * retry.  It's easier (and not really slower) to just retry
2867                  * since this case occurs so rarely and the hint is already
2868                  * pointing at the right place.  We have to reset the
2869                  * start offset so as not to accidently delete an entry
2870                  * another process just created in vacated space.
2871                  */
2872                 if (entry->eflags & MAP_ENTRY_IN_TRANSITION) {
2873                         entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP;
2874                         start = entry->start;
2875                         ++mycpu->gd_cnt.v_intrans_coll;
2876                         ++mycpu->gd_cnt.v_intrans_wait;
2877                         vm_map_transition_wait(map);
2878                         goto again;
2879                 }
2880                 vm_map_clip_end(map, entry, end, countp);
2881
2882                 s = entry->start;
2883                 e = entry->end;
2884                 next = entry->next;
2885
2886                 offidxstart = OFF_TO_IDX(entry->offset);
2887                 count = OFF_TO_IDX(e - s);
2888
2889                 switch(entry->maptype) {
2890                 case VM_MAPTYPE_NORMAL:
2891                 case VM_MAPTYPE_VPAGETABLE:
2892                 case VM_MAPTYPE_SUBMAP:
2893                         object = entry->object.vm_object;
2894                         break;
2895                 default:
2896                         object = NULL;
2897                         break;
2898                 }
2899
2900                 /*
2901                  * Unwire before removing addresses from the pmap; otherwise,
2902                  * unwiring will put the entries back in the pmap.
2903                  */
2904                 if (entry->wired_count != 0)
2905                         vm_map_entry_unwire(map, entry);
2906
2907                 offidxend = offidxstart + count;
2908
2909                 if (object == &kernel_object) {
2910                         vm_object_hold(object);
2911                         vm_object_page_remove(object, offidxstart,
2912                                               offidxend, FALSE);
2913                         vm_object_drop(object);
2914                 } else if (object && object->type != OBJT_DEFAULT &&
2915                            object->type != OBJT_SWAP) {
2916                         /*
2917                          * vnode object routines cannot be chain-locked,
2918                          * but since we aren't removing pages from the
2919                          * object here we can use a shared hold.
2920                          */
2921                         vm_object_hold_shared(object);
2922                         pmap_remove(map->pmap, s, e);
2923                         vm_object_drop(object);
2924                 } else if (object) {
2925                         vm_object_hold(object);
2926                         vm_object_chain_acquire(object, 0);
2927                         pmap_remove(map->pmap, s, e);
2928
2929                         if (object != NULL &&
2930                             object->ref_count != 1 &&
2931                             (object->flags & (OBJ_NOSPLIT|OBJ_ONEMAPPING)) ==
2932                              OBJ_ONEMAPPING &&
2933                             (object->type == OBJT_DEFAULT ||
2934                              object->type == OBJT_SWAP)) {
2935                                 vm_object_collapse(object, NULL);
2936                                 vm_object_page_remove(object, offidxstart,
2937                                                       offidxend, FALSE);
2938                                 if (object->type == OBJT_SWAP) {
2939                                         swap_pager_freespace(object,
2940                                                              offidxstart,
2941                                                              count);
2942                                 }
2943                                 if (offidxend >= object->size &&
2944                                     offidxstart < object->size) {
2945                                         object->size = offidxstart;
2946                                 }
2947                         }
2948                         vm_object_chain_release(object);
2949                         vm_object_drop(object);
2950                 } else if (entry->maptype == VM_MAPTYPE_UKSMAP) {
2951                         pmap_remove(map->pmap, s, e);
2952                 }
2953
2954                 /*
2955                  * Delete the entry (which may delete the object) only after
2956                  * removing all pmap entries pointing to its pages.
2957                  * (Otherwise, its page frames may be reallocated, and any
2958                  * modify bits will be set in the wrong object!)
2959                  */
2960                 vm_map_entry_delete(map, entry, countp);
2961                 entry = next;
2962         }
2963         lwkt_reltoken(&map->token);
2964         return (KERN_SUCCESS);
2965 }
2966
2967 /*
2968  * Remove the given address range from the target map.
2969  * This is the exported form of vm_map_delete.
2970  *
2971  * No requirements.
2972  */
2973 int
2974 vm_map_remove(vm_map_t map, vm_offset_t start, vm_offset_t end)
2975 {
2976         int result;
2977         int count;
2978
2979         count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
2980         vm_map_lock(map);
2981         VM_MAP_RANGE_CHECK(map, start, end);
2982         result = vm_map_delete(map, start, end, &count);
2983         vm_map_unlock(map);
2984         vm_map_entry_release(count);
2985
2986         return (result);
2987 }
2988
2989 /*
2990  * Assert that the target map allows the specified privilege on the
2991  * entire address region given.  The entire region must be allocated.
2992  *
2993  * The caller must specify whether the vm_map is already locked or not.
2994  */
2995 boolean_t
2996 vm_map_check_protection(vm_map_t map, vm_offset_t start, vm_offset_t end,
2997                         vm_prot_t protection, boolean_t have_lock)
2998 {
2999         vm_map_entry_t entry;
3000         vm_map_entry_t tmp_entry;
3001         boolean_t result;
3002
3003         if (have_lock == FALSE)
3004                 vm_map_lock_read(map);
3005
3006         if (!vm_map_lookup_entry(map, start, &tmp_entry)) {
3007                 if (have_lock == FALSE)
3008                         vm_map_unlock_read(map);
3009                 return (FALSE);
3010         }
3011         entry = tmp_entry;
3012
3013         result = TRUE;
3014         while (start < end) {
3015                 if (entry == &map->header) {
3016                         result = FALSE;
3017                         break;
3018                 }
3019                 /*
3020                  * No holes allowed!
3021                  */
3022
3023                 if (start < entry->start) {
3024                         result = FALSE;
3025                         break;
3026                 }
3027                 /*
3028                  * Check protection associated with entry.
3029                  */
3030
3031                 if ((entry->protection & protection) != protection) {
3032                         result = FALSE;
3033                         break;
3034                 }
3035                 /* go to next entry */
3036
3037                 start = entry->end;
3038                 entry = entry->next;
3039         }
3040         if (have_lock == FALSE)
3041                 vm_map_unlock_read(map);
3042         return (result);
3043 }
3044
3045 /*
3046  * If appropriate this function shadows the original object with a new object
3047  * and moves the VM pages from the original object to the new object.
3048  * The original object will also be collapsed, if possible.
3049  *
3050  * We can only do this for normal memory objects with a single mapping, and
3051  * it only makes sense to do it if there are 2 or more refs on the original
3052  * object.  i.e. typically a memory object that has been extended into
3053  * multiple vm_map_entry's with non-overlapping ranges.
3054  *
3055  * This makes it easier to remove unused pages and keeps object inheritance
3056  * from being a negative impact on memory usage.
3057  *
3058  * On return the (possibly new) entry->object.vm_object will have an
3059  * additional ref on it for the caller to dispose of (usually by cloning
3060  * the vm_map_entry).  The additional ref had to be done in this routine
3061  * to avoid racing a collapse.  The object's ONEMAPPING flag will also be
3062  * cleared.
3063  *
3064  * The vm_map must be locked and its token held.
3065  */
3066 static void
3067 vm_map_split(vm_map_entry_t entry)
3068 {
3069         /* OPTIMIZED */
3070         vm_object_t oobject, nobject, bobject;
3071         vm_offset_t s, e;
3072         vm_page_t m;
3073         vm_pindex_t offidxstart, offidxend, idx;
3074         vm_size_t size;
3075         vm_ooffset_t offset;
3076         int useshadowlist;
3077
3078         /*
3079          * Optimize away object locks for vnode objects.  Important exit/exec
3080          * critical path.
3081          *
3082          * OBJ_ONEMAPPING doesn't apply to vnode objects but clear the flag
3083          * anyway.
3084          */
3085         oobject = entry->object.vm_object;
3086         if (oobject->type != OBJT_DEFAULT && oobject->type != OBJT_SWAP) {
3087                 vm_object_reference_quick(oobject);
3088                 vm_object_clear_flag(oobject, OBJ_ONEMAPPING);
3089                 return;
3090         }
3091
3092         /*
3093          * Setup.  Chain lock the original object throughout the entire
3094          * routine to prevent new page faults from occuring.
3095          *
3096          * XXX can madvise WILLNEED interfere with us too?
3097          */
3098         vm_object_hold(oobject);
3099         vm_object_chain_acquire(oobject, 0);
3100
3101         /*
3102          * Original object cannot be split?  Might have also changed state.
3103          */
3104         if (oobject->handle == NULL || (oobject->type != OBJT_DEFAULT &&
3105                                         oobject->type != OBJT_SWAP)) {
3106                 vm_object_chain_release(oobject);
3107                 vm_object_reference_locked(oobject);
3108                 vm_object_clear_flag(oobject, OBJ_ONEMAPPING);
3109                 vm_object_drop(oobject);
3110                 return;
3111         }
3112
3113         /*
3114          * Collapse original object with its backing store as an
3115          * optimization to reduce chain lengths when possible.
3116          *
3117          * If ref_count <= 1 there aren't other non-overlapping vm_map_entry's
3118          * for oobject, so there's no point collapsing it.
3119          *
3120          * Then re-check whether the object can be split.
3121          */
3122         vm_object_collapse(oobject, NULL);
3123
3124         if (oobject->ref_count <= 1 ||
3125             (oobject->type != OBJT_DEFAULT && oobject->type != OBJT_SWAP) ||
3126             (oobject->flags & (OBJ_NOSPLIT|OBJ_ONEMAPPING)) != OBJ_ONEMAPPING) {
3127                 vm_object_chain_release(oobject);
3128                 vm_object_reference_locked(oobject);
3129                 vm_object_clear_flag(oobject, OBJ_ONEMAPPING);
3130                 vm_object_drop(oobject);
3131                 return;
3132         }
3133
3134         /*
3135          * Acquire the chain lock on the backing object.
3136          *
3137          * Give bobject an additional ref count for when it will be shadowed
3138          * by nobject.
3139          */
3140         useshadowlist = 0;
3141         if ((bobject = oobject->backing_object) != NULL) {
3142                 if (bobject->type != OBJT_VNODE) {
3143                         useshadowlist = 1;
3144                         vm_object_hold(bobject);
3145                         vm_object_chain_wait(bobject, 0);
3146                         /* ref for shadowing below */
3147                         vm_object_reference_locked(bobject);
3148                         vm_object_chain_acquire(bobject, 0);
3149                         KKASSERT(bobject->backing_object == bobject);
3150                         KKASSERT((bobject->flags & OBJ_DEAD) == 0);
3151                 } else {
3152                         /*
3153                          * vnodes are not placed on the shadow list but
3154                          * they still get another ref for the backing_object
3155                          * reference.
3156                          */
3157                         vm_object_reference_quick(bobject);
3158                 }
3159         }
3160
3161         /*
3162          * Calculate the object page range and allocate the new object.
3163          */
3164         offset = entry->offset;
3165         s = entry->start;
3166         e = entry->end;
3167
3168         offidxstart = OFF_TO_IDX(offset);
3169         offidxend = offidxstart + OFF_TO_IDX(e - s);
3170         size = offidxend - offidxstart;
3171
3172         switch(oobject->type) {
3173         case OBJT_DEFAULT:
3174                 nobject = default_pager_alloc(NULL, IDX_TO_OFF(size),
3175                                               VM_PROT_ALL, 0);
3176                 break;
3177         case OBJT_SWAP:
3178                 nobject = swap_pager_alloc(NULL, IDX_TO_OFF(size),
3179                                            VM_PROT_ALL, 0);
3180                 break;
3181         default:
3182                 /* not reached */
3183                 nobject = NULL;
3184                 KKASSERT(0);
3185         }
3186
3187         if (nobject == NULL) {
3188                 if (bobject) {
3189                         if (useshadowlist) {
3190                                 vm_object_chain_release(bobject);
3191                                 vm_object_deallocate(bobject);
3192                                 vm_object_drop(bobject);
3193                         } else {
3194                                 vm_object_deallocate(bobject);
3195                         }
3196                 }
3197                 vm_object_chain_release(oobject);
3198                 vm_object_reference_locked(oobject);
3199                 vm_object_clear_flag(oobject, OBJ_ONEMAPPING);
3200                 vm_object_drop(oobject);
3201                 return;
3202         }
3203
3204         /*
3205          * The new object will replace entry->object.vm_object so it needs
3206          * a second reference (the caller expects an additional ref).
3207          */
3208         vm_object_hold(nobject);
3209         vm_object_reference_locked(nobject);
3210         vm_object_chain_acquire(nobject, 0);
3211
3212         /*
3213          * nobject shadows bobject (oobject already shadows bobject).
3214          *
3215          * Adding an object to bobject's shadow list requires refing bobject
3216          * which we did above in the useshadowlist case.
3217          */
3218         if (bobject) {
3219                 nobject->backing_object_offset =
3220                     oobject->backing_object_offset + IDX_TO_OFF(offidxstart);
3221                 nobject->backing_object = bobject;
3222                 if (useshadowlist) {
3223                         bobject->shadow_count++;
3224                         bobject->generation++;
3225                         LIST_INSERT_HEAD(&bobject->shadow_head,
3226                                          nobject, shadow_list);
3227                         vm_object_clear_flag(bobject, OBJ_ONEMAPPING); /*XXX*/
3228                         vm_object_chain_release(bobject);
3229                         vm_object_drop(bobject);
3230                         vm_object_set_flag(nobject, OBJ_ONSHADOW);
3231                 }
3232         }
3233
3234         /*
3235          * Move the VM pages from oobject to nobject
3236          */
3237         for (idx = 0; idx < size; idx++) {
3238                 vm_page_t m;
3239
3240                 m = vm_page_lookup_busy_wait(oobject, offidxstart + idx,
3241                                              TRUE, "vmpg");
3242                 if (m == NULL)
3243                         continue;
3244
3245                 /*
3246                  * We must wait for pending I/O to complete before we can
3247                  * rename the page.
3248                  *
3249                  * We do not have to VM_PROT_NONE the page as mappings should
3250                  * not be changed by this operation.
3251                  *
3252                  * NOTE: The act of renaming a page updates chaingen for both
3253                  *       objects.
3254                  */
3255                 vm_page_rename(m, nobject, idx);
3256                 /* page automatically made dirty by rename and cache handled */
3257                 /* page remains busy */
3258         }
3259
3260         if (oobject->type == OBJT_SWAP) {
3261                 vm_object_pip_add(oobject, 1);
3262                 /*
3263                  * copy oobject pages into nobject and destroy unneeded
3264                  * pages in shadow object.
3265                  */
3266                 swap_pager_copy(oobject, nobject, offidxstart, 0);
3267                 vm_object_pip_wakeup(oobject);
3268         }
3269
3270         /*
3271          * Wakeup the pages we played with.  No spl protection is needed
3272          * for a simple wakeup.
3273          */
3274         for (idx = 0; idx < size; idx++) {
3275                 m = vm_page_lookup(nobject, idx);
3276                 if (m) {
3277                         KKASSERT(m->flags & PG_BUSY);
3278                         vm_page_wakeup(m);
3279                 }
3280         }
3281         entry->object.vm_object = nobject;
3282         entry->offset = 0LL;
3283
3284         /*
3285          * Cleanup
3286          *
3287          * NOTE: There is no need to remove OBJ_ONEMAPPING from oobject, the
3288          *       related pages were moved and are no longer applicable to the
3289          *       original object.
3290          *
3291          * NOTE: Deallocate oobject (due to its entry->object.vm_object being
3292          *       replaced by nobject).
3293          */
3294         vm_object_chain_release(nobject);
3295         vm_object_drop(nobject);
3296         if (bobject && useshadowlist) {
3297                 vm_object_chain_release(bobject);
3298                 vm_object_drop(bobject);
3299         }
3300         vm_object_chain_release(oobject);
3301         /*vm_object_clear_flag(oobject, OBJ_ONEMAPPING);*/
3302         vm_object_deallocate_locked(oobject);
3303         vm_object_drop(oobject);
3304 }
3305
3306 /*
3307  * Copies the contents of the source entry to the destination
3308  * entry.  The entries *must* be aligned properly.
3309  *
3310  * The vm_maps must be exclusively locked.
3311  * The vm_map's token must be held.
3312  *
3313  * Because the maps are locked no faults can be in progress during the
3314  * operation.
3315  */
3316 static void
3317 vm_map_copy_entry(vm_map_t src_map, vm_map_t dst_map,
3318                   vm_map_entry_t src_entry, vm_map_entry_t dst_entry)
3319 {
3320         vm_object_t src_object;
3321
3322         if (dst_entry->maptype == VM_MAPTYPE_SUBMAP ||
3323             dst_entry->maptype == VM_MAPTYPE_UKSMAP)
3324                 return;
3325         if (src_entry->maptype == VM_MAPTYPE_SUBMAP ||
3326             src_entry->maptype == VM_MAPTYPE_UKSMAP)
3327                 return;
3328
3329         if (src_entry->wired_count == 0) {
3330                 /*
3331                  * If the source entry is marked needs_copy, it is already
3332                  * write-protected.
3333                  */
3334                 if ((src_entry->eflags & MAP_ENTRY_NEEDS_COPY) == 0) {
3335                         pmap_protect(src_map->pmap,
3336                             src_entry->start,
3337                             src_entry->end,
3338                             src_entry->protection & ~VM_PROT_WRITE);
3339                 }
3340
3341                 /*
3342                  * Make a copy of the object.
3343                  *
3344                  * The object must be locked prior to checking the object type
3345                  * and for the call to vm_object_collapse() and vm_map_split().
3346                  * We cannot use *_hold() here because the split code will
3347                  * probably try to destroy the object.  The lock is a pool
3348                  * token and doesn't care.
3349                  *
3350                  * We must bump src_map->timestamp when setting
3351                  * MAP_ENTRY_NEEDS_COPY to force any concurrent fault
3352                  * to retry, otherwise the concurrent fault might improperly
3353                  * install a RW pte when its supposed to be a RO(COW) pte.
3354                  * This race can occur because a vnode-backed fault may have
3355                  * to temporarily release the map lock.
3356                  */
3357                 if (src_entry->object.vm_object != NULL) {
3358                         vm_map_split(src_entry);
3359                         src_object = src_entry->object.vm_object;
3360                         dst_entry->object.vm_object = src_object;
3361                         src_entry->eflags |= (MAP_ENTRY_COW |
3362                                               MAP_ENTRY_NEEDS_COPY);
3363                         dst_entry->eflags |= (MAP_ENTRY_COW |
3364                                               MAP_ENTRY_NEEDS_COPY);
3365                         dst_entry->offset = src_entry->offset;
3366                         ++src_map->timestamp;
3367                 } else {
3368                         dst_entry->object.vm_object = NULL;
3369                         dst_entry->offset = 0;
3370                 }
3371
3372                 pmap_copy(dst_map->pmap, src_map->pmap, dst_entry->start,
3373                     dst_entry->end - dst_entry->start, src_entry->start);
3374         } else {
3375                 /*
3376                  * Of course, wired down pages can't be set copy-on-write.
3377                  * Cause wired pages to be copied into the new map by
3378                  * simulating faults (the new pages are pageable)
3379                  */
3380                 vm_fault_copy_entry(dst_map, src_map, dst_entry, src_entry);
3381         }
3382 }
3383
3384 /*
3385  * vmspace_fork:
3386  * Create a new process vmspace structure and vm_map
3387  * based on those of an existing process.  The new map
3388  * is based on the old map, according to the inheritance
3389  * values on the regions in that map.
3390  *
3391  * The source map must not be locked.
3392  * No requirements.
3393  */
3394 static void vmspace_fork_normal_entry(vm_map_t old_map, vm_map_t new_map,
3395                           vm_map_entry_t old_entry, int *countp);
3396 static void vmspace_fork_uksmap_entry(vm_map_t old_map, vm_map_t new_map,
3397                           vm_map_entry_t old_entry, int *countp);
3398
3399 struct vmspace *
3400 vmspace_fork(struct vmspace *vm1)
3401 {
3402         struct vmspace *vm2;
3403         vm_map_t old_map = &vm1->vm_map;
3404         vm_map_t new_map;
3405         vm_map_entry_t old_entry;
3406         int count;
3407
3408         lwkt_gettoken(&vm1->vm_map.token);
3409         vm_map_lock(old_map);
3410
3411         vm2 = vmspace_alloc(old_map->min_offset, old_map->max_offset);
3412         lwkt_gettoken(&vm2->vm_map.token);
3413         bcopy(&vm1->vm_startcopy, &vm2->vm_startcopy,
3414             (caddr_t)&vm1->vm_endcopy - (caddr_t)&vm1->vm_startcopy);
3415         new_map = &vm2->vm_map; /* XXX */
3416         new_map->timestamp = 1;
3417
3418         vm_map_lock(new_map);
3419
3420         count = 0;
3421         old_entry = old_map->header.next;
3422         while (old_entry != &old_map->header) {
3423                 ++count;
3424                 old_entry = old_entry->next;
3425         }
3426
3427         count = vm_map_entry_reserve(count + MAP_RESERVE_COUNT);
3428
3429         old_entry = old_map->header.next;
3430         while (old_entry != &old_map->header) {
3431                 switch(old_entry->maptype) {
3432                 case VM_MAPTYPE_SUBMAP:
3433                         panic("vm_map_fork: encountered a submap");
3434                         break;
3435                 case VM_MAPTYPE_UKSMAP:
3436                         vmspace_fork_uksmap_entry(old_map, new_map,
3437                                                   old_entry, &count);
3438                         break;
3439                 case VM_MAPTYPE_NORMAL:
3440                 case VM_MAPTYPE_VPAGETABLE:
3441                         vmspace_fork_normal_entry(old_map, new_map,
3442                                                   old_entry, &count);
3443                         break;
3444                 }
3445                 old_entry = old_entry->next;
3446         }
3447
3448         new_map->size = old_map->size;
3449         vm_map_unlock(old_map);
3450         vm_map_unlock(new_map);
3451         vm_map_entry_release(count);
3452
3453         lwkt_reltoken(&vm2->vm_map.token);
3454         lwkt_reltoken(&vm1->vm_map.token);
3455
3456         return (vm2);
3457 }
3458
3459 static
3460 void
3461 vmspace_fork_normal_entry(vm_map_t old_map, vm_map_t new_map,
3462                           vm_map_entry_t old_entry, int *countp)
3463 {
3464         vm_map_entry_t new_entry;
3465         vm_object_t object;
3466
3467         switch (old_entry->inheritance) {
3468         case VM_INHERIT_NONE:
3469                 break;
3470         case VM_INHERIT_SHARE:
3471                 /*
3472                  * Clone the entry, creating the shared object if
3473                  * necessary.
3474                  */
3475                 if (old_entry->object.vm_object == NULL)
3476                         vm_map_entry_allocate_object(old_entry);
3477
3478                 if (old_entry->eflags & MAP_ENTRY_NEEDS_COPY) {
3479                         /*
3480                          * Shadow a map_entry which needs a copy,
3481                          * replacing its object with a new object
3482                          * that points to the old one.  Ask the
3483                          * shadow code to automatically add an
3484                          * additional ref.  We can't do it afterwords
3485                          * because we might race a collapse.  The call
3486                          * to vm_map_entry_shadow() will also clear
3487                          * OBJ_ONEMAPPING.
3488                          */
3489                         vm_map_entry_shadow(old_entry, 1);
3490                 } else if (old_entry->object.vm_object) {
3491                         /*
3492                          * We will make a shared copy of the object,
3493                          * and must clear OBJ_ONEMAPPING.
3494                          *
3495                          * Optimize vnode objects.  OBJ_ONEMAPPING
3496                          * is non-applicable but clear it anyway,
3497                          * and its terminal so we don'th ave to deal
3498                          * with chains.  Reduces SMP conflicts.
3499                          *
3500                          * XXX assert that object.vm_object != NULL
3501                          *     since we allocate it above.
3502                          */
3503                         object = old_entry->object.vm_object;
3504                         if (object->type == OBJT_VNODE) {
3505                                 vm_object_reference_quick(object);
3506                                 vm_object_clear_flag(object,
3507                                                      OBJ_ONEMAPPING);
3508                         } else {
3509                                 vm_object_hold(object);
3510                                 vm_object_chain_wait(object, 0);
3511                                 vm_object_reference_locked(object);
3512                                 vm_object_clear_flag(object,
3513                                                      OBJ_ONEMAPPING);
3514                                 vm_object_drop(object);
3515                         }
3516                 }
3517
3518                 /*
3519                  * Clone the entry.  We've already bumped the ref on
3520                  * any vm_object.
3521                  */
3522                 new_entry = vm_map_entry_create(new_map, countp);
3523                 *new_entry = *old_entry;
3524                 new_entry->eflags &= ~MAP_ENTRY_USER_WIRED;
3525                 new_entry->wired_count = 0;
3526
3527                 /*
3528                  * Insert the entry into the new map -- we know we're
3529                  * inserting at the end of the new map.
3530                  */
3531
3532                 vm_map_entry_link(new_map, new_map->header.prev,
3533                                   new_entry);
3534
3535                 /*
3536                  * Update the physical map
3537                  */
3538                 pmap_copy(new_map->pmap, old_map->pmap,
3539                           new_entry->start,
3540                           (old_entry->end - old_entry->start),
3541                           old_entry->start);
3542                 break;
3543         case VM_INHERIT_COPY:
3544                 /*
3545                  * Clone the entry and link into the map.
3546                  */
3547                 new_entry = vm_map_entry_create(new_map, countp);
3548                 *new_entry = *old_entry;
3549                 new_entry->eflags &= ~MAP_ENTRY_USER_WIRED;
3550                 new_entry->wired_count = 0;
3551                 new_entry->object.vm_object = NULL;
3552                 vm_map_entry_link(new_map, new_map->header.prev,
3553                                   new_entry);
3554                 vm_map_copy_entry(old_map, new_map, old_entry,
3555                                   new_entry);
3556                 break;
3557         }
3558 }
3559
3560 /*
3561  * When forking user-kernel shared maps, the map might change in the
3562  * child so do not try to copy the underlying pmap entries.
3563  */
3564 static
3565 void
3566 vmspace_fork_uksmap_entry(vm_map_t old_map, vm_map_t new_map,
3567                           vm_map_entry_t old_entry, int *countp)
3568 {
3569         vm_map_entry_t new_entry;
3570
3571         new_entry = vm_map_entry_create(new_map, countp);
3572         *new_entry = *old_entry;
3573         new_entry->eflags &= ~MAP_ENTRY_USER_WIRED;
3574         new_entry->wired_count = 0;
3575         vm_map_entry_link(new_map, new_map->header.prev,
3576                           new_entry);
3577 }
3578
3579 /*
3580  * Create an auto-grow stack entry
3581  *
3582  * No requirements.
3583  */
3584 int
3585 vm_map_stack (vm_map_t map, vm_offset_t addrbos, vm_size_t max_ssize,
3586               int flags, vm_prot_t prot, vm_prot_t max, int cow)
3587 {
3588         vm_map_entry_t  prev_entry;
3589         vm_map_entry_t  new_stack_entry;
3590         vm_size_t       init_ssize;
3591         int             rv;
3592         int             count;
3593         vm_offset_t     tmpaddr;
3594
3595         cow |= MAP_IS_STACK;
3596
3597         if (max_ssize < sgrowsiz)
3598                 init_ssize = max_ssize;
3599         else
3600                 init_ssize = sgrowsiz;
3601
3602         count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
3603         vm_map_lock(map);
3604
3605         /*
3606          * Find space for the mapping
3607          */
3608         if ((flags & (MAP_FIXED | MAP_TRYFIXED)) == 0) {
3609                 if (vm_map_findspace(map, addrbos, max_ssize, 1,
3610                                      flags, &tmpaddr)) {
3611                         vm_map_unlock(map);
3612                         vm_map_entry_release(count);
3613                         return (KERN_NO_SPACE);
3614                 }
3615                 addrbos = tmpaddr;
3616         }
3617
3618         /* If addr is already mapped, no go */
3619         if (vm_map_lookup_entry(map, addrbos, &prev_entry)) {
3620                 vm_map_unlock(map);
3621                 vm_map_entry_release(count);
3622                 return (KERN_NO_SPACE);
3623         }
3624
3625 #if 0
3626         /* XXX already handled by kern_mmap() */
3627         /* If we would blow our VMEM resource limit, no go */
3628         if (map->size + init_ssize >
3629             curproc->p_rlimit[RLIMIT_VMEM].rlim_cur) {
3630                 vm_map_unlock(map);
3631                 vm_map_entry_release(count);
3632                 return (KERN_NO_SPACE);
3633         }
3634 #endif
3635
3636         /*
3637          * If we can't accomodate max_ssize in the current mapping,
3638          * no go.  However, we need to be aware that subsequent user
3639          * mappings might map into the space we have reserved for
3640          * stack, and currently this space is not protected.  
3641          * 
3642          * Hopefully we will at least detect this condition 
3643          * when we try to grow the stack.
3644          */
3645         if ((prev_entry->next != &map->header) &&
3646             (prev_entry->next->start < addrbos + max_ssize)) {
3647                 vm_map_unlock(map);
3648                 vm_map_entry_release(count);
3649                 return (KERN_NO_SPACE);
3650         }
3651
3652         /*
3653          * We initially map a stack of only init_ssize.  We will
3654          * grow as needed later.  Since this is to be a grow 
3655          * down stack, we map at the top of the range.
3656          *
3657          * Note: we would normally expect prot and max to be
3658          * VM_PROT_ALL, and cow to be 0.  Possibly we should
3659          * eliminate these as input parameters, and just
3660          * pass these values here in the insert call.
3661          */
3662         rv = vm_map_insert(map, &count, NULL, NULL,
3663                            0, addrbos + max_ssize - init_ssize,
3664                            addrbos + max_ssize,
3665                            VM_MAPTYPE_NORMAL,
3666                            prot, max, cow);
3667
3668         /* Now set the avail_ssize amount */
3669         if (rv == KERN_SUCCESS) {
3670                 if (prev_entry != &map->header)
3671                         vm_map_clip_end(map, prev_entry, addrbos + max_ssize - init_ssize, &count);
3672                 new_stack_entry = prev_entry->next;
3673                 if (new_stack_entry->end   != addrbos + max_ssize ||
3674                     new_stack_entry->start != addrbos + max_ssize - init_ssize)
3675                         panic ("Bad entry start/end for new stack entry");
3676                 else 
3677                         new_stack_entry->aux.avail_ssize = max_ssize - init_ssize;
3678         }
3679
3680         vm_map_unlock(map);
3681         vm_map_entry_release(count);
3682         return (rv);
3683 }
3684
3685 /*
3686  * Attempts to grow a vm stack entry.  Returns KERN_SUCCESS if the
3687  * desired address is already mapped, or if we successfully grow
3688  * the stack.  Also returns KERN_SUCCESS if addr is outside the
3689  * stack range (this is strange, but preserves compatibility with
3690  * the grow function in vm_machdep.c).
3691  *
3692  * No requirements.
3693  */
3694 int
3695 vm_map_growstack (struct proc *p, vm_offset_t addr)
3696 {
3697         vm_map_entry_t prev_entry;
3698         vm_map_entry_t stack_entry;
3699         vm_map_entry_t new_stack_entry;
3700         struct vmspace *vm = p->p_vmspace;
3701         vm_map_t map = &vm->vm_map;
3702         vm_offset_t    end;
3703         int grow_amount;
3704         int rv = KERN_SUCCESS;
3705         int is_procstack;
3706         int use_read_lock = 1;
3707         int count;
3708
3709         count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
3710 Retry:
3711         if (use_read_lock)
3712                 vm_map_lock_read(map);
3713         else
3714                 vm_map_lock(map);
3715
3716         /* If addr is already in the entry range, no need to grow.*/
3717         if (vm_map_lookup_entry(map, addr, &prev_entry))
3718                 goto done;
3719
3720         if ((stack_entry = prev_entry->next) == &map->header)
3721                 goto done;
3722         if (prev_entry == &map->header) 
3723                 end = stack_entry->start - stack_entry->aux.avail_ssize;
3724         else
3725                 end = prev_entry->end;
3726
3727         /*
3728          * This next test mimics the old grow function in vm_machdep.c.
3729          * It really doesn't quite make sense, but we do it anyway
3730          * for compatibility.
3731          *
3732          * If not growable stack, return success.  This signals the
3733          * caller to proceed as he would normally with normal vm.
3734          */
3735         if (stack_entry->aux.avail_ssize < 1 ||
3736             addr >= stack_entry->start ||
3737             addr <  stack_entry->start - stack_entry->aux.avail_ssize) {
3738                 goto done;
3739         } 
3740         
3741         /* Find the minimum grow amount */
3742         grow_amount = roundup (stack_entry->start - addr, PAGE_SIZE);
3743         if (grow_amount > stack_entry->aux.avail_ssize) {
3744                 rv = KERN_NO_SPACE;
3745                 goto done;
3746         }
3747
3748         /*
3749          * If there is no longer enough space between the entries
3750          * nogo, and adjust the available space.  Note: this 
3751          * should only happen if the user has mapped into the
3752          * stack area after the stack was created, and is
3753          * probably an error.
3754          *
3755          * This also effectively destroys any guard page the user
3756          * might have intended by limiting the stack size.
3757          */
3758         if (grow_amount > stack_entry->start - end) {
3759                 if (use_read_lock && vm_map_lock_upgrade(map)) {
3760                         /* lost lock */
3761                         use_read_lock = 0;
3762                         goto Retry;
3763                 }
3764                 use_read_lock = 0;
3765                 stack_entry->aux.avail_ssize = stack_entry->start - end;
3766                 rv = KERN_NO_SPACE;
3767                 goto done;
3768         }
3769
3770         is_procstack = addr >= (vm_offset_t)vm->vm_maxsaddr;
3771
3772         /* If this is the main process stack, see if we're over the 
3773          * stack limit.
3774          */
3775         if (is_procstack && (ctob(vm->vm_ssize) + grow_amount >
3776                              p->p_rlimit[RLIMIT_STACK].rlim_cur)) {
3777                 rv = KERN_NO_SPACE;
3778                 goto done;
3779         }
3780
3781         /* Round up the grow amount modulo SGROWSIZ */
3782         grow_amount = roundup (grow_amount, sgrowsiz);
3783         if (grow_amount > stack_entry->aux.avail_ssize) {
3784                 grow_amount = stack_entry->aux.avail_ssize;
3785         }
3786         if (is_procstack && (ctob(vm->vm_ssize) + grow_amount >
3787                              p->p_rlimit[RLIMIT_STACK].rlim_cur)) {
3788                 grow_amount = p->p_rlimit[RLIMIT_STACK].rlim_cur -
3789                               ctob(vm->vm_ssize);
3790         }
3791
3792         /* If we would blow our VMEM resource limit, no go */
3793         if (map->size + grow_amount > p->p_rlimit[RLIMIT_VMEM].rlim_cur) {
3794                 rv = KERN_NO_SPACE;
3795                 goto done;
3796         }
3797
3798         if (use_read_lock && vm_map_lock_upgrade(map)) {
3799                 /* lost lock */
3800                 use_read_lock = 0;
3801                 goto Retry;
3802         }
3803         use_read_lock = 0;
3804
3805         /* Get the preliminary new entry start value */
3806         addr = stack_entry->start - grow_amount;
3807
3808         /* If this puts us into the previous entry, cut back our growth
3809          * to the available space.  Also, see the note above.
3810          */
3811         if (addr < end) {
3812                 stack_entry->aux.avail_ssize = stack_entry->start - end;
3813                 addr = end;
3814         }
3815
3816         rv = vm_map_insert(map, &count, NULL, NULL,
3817                            0, addr, stack_entry->start,
3818                            VM_MAPTYPE_NORMAL,
3819                            VM_PROT_ALL, VM_PROT_ALL, 0);
3820
3821         /* Adjust the available stack space by the amount we grew. */
3822         if (rv == KERN_SUCCESS) {
3823                 if (prev_entry != &map->header)
3824                         vm_map_clip_end(map, prev_entry, addr, &count);
3825                 new_stack_entry = prev_entry->next;
3826                 if (new_stack_entry->end   != stack_entry->start  ||
3827                     new_stack_entry->start != addr)
3828                         panic ("Bad stack grow start/end in new stack entry");
3829                 else {
3830                         new_stack_entry->aux.avail_ssize =
3831                                 stack_entry->aux.avail_ssize -
3832                                 (new_stack_entry->end - new_stack_entry->start);
3833                         if (is_procstack)
3834                                 vm->vm_ssize += btoc(new_stack_entry->end -
3835                                                      new_stack_entry->start);
3836                 }
3837
3838                 if (map->flags & MAP_WIREFUTURE)
3839                         vm_map_unwire(map, new_stack_entry->start,
3840                                       new_stack_entry->end, FALSE);
3841         }
3842
3843 done:
3844         if (use_read_lock)
3845                 vm_map_unlock_read(map);
3846         else
3847                 vm_map_unlock(map);
3848         vm_map_entry_release(count);
3849         return (rv);
3850 }
3851
3852 /*
3853  * Unshare the specified VM space for exec.  If other processes are
3854  * mapped to it, then create a new one.  The new vmspace is null.
3855  *
3856  * No requirements.
3857  */
3858 void
3859 vmspace_exec(struct proc *p, struct vmspace *vmcopy) 
3860 {
3861         struct vmspace *oldvmspace = p->p_vmspace;
3862         struct vmspace *newvmspace;
3863         vm_map_t map = &p->p_vmspace->vm_map;
3864
3865         /*
3866          * If we are execing a resident vmspace we fork it, otherwise
3867          * we create a new vmspace.  Note that exitingcnt is not
3868          * copied to the new vmspace.
3869          */
3870         lwkt_gettoken(&oldvmspace->vm_map.token);
3871         if (vmcopy)  {
3872                 newvmspace = vmspace_fork(vmcopy);
3873                 lwkt_gettoken(&newvmspace->vm_map.token);
3874         } else {
3875                 newvmspace = vmspace_alloc(map->min_offset, map->max_offset);
3876                 lwkt_gettoken(&newvmspace->vm_map.token);
3877                 bcopy(&oldvmspace->vm_startcopy, &newvmspace->vm_startcopy,
3878                       (caddr_t)&oldvmspace->vm_endcopy -
3879                        (caddr_t)&oldvmspace->vm_startcopy);
3880         }
3881
3882         /*
3883          * Finish initializing the vmspace before assigning it
3884          * to the process.  The vmspace will become the current vmspace
3885          * if p == curproc.
3886          */
3887         pmap_pinit2(vmspace_pmap(newvmspace));
3888         pmap_replacevm(p, newvmspace, 0);
3889         lwkt_reltoken(&newvmspace->vm_map.token);
3890         lwkt_reltoken(&oldvmspace->vm_map.token);
3891         vmspace_rel(oldvmspace);
3892 }
3893
3894 /*
3895  * Unshare the specified VM space for forcing COW.  This
3896  * is called by rfork, for the (RFMEM|RFPROC) == 0 case.
3897  */
3898 void
3899 vmspace_unshare(struct proc *p) 
3900 {
3901         struct vmspace *oldvmspace = p->p_vmspace;
3902         struct vmspace *newvmspace;
3903
3904         lwkt_gettoken(&oldvmspace->vm_map.token);
3905         if (vmspace_getrefs(oldvmspace) == 1) {
3906                 lwkt_reltoken(&oldvmspace->vm_map.token);
3907                 return;
3908         }
3909         newvmspace = vmspace_fork(oldvmspace);
3910         lwkt_gettoken(&newvmspace->vm_map.token);
3911         pmap_pinit2(vmspace_pmap(newvmspace));
3912         pmap_replacevm(p, newvmspace, 0);
3913         lwkt_reltoken(&newvmspace->vm_map.token);
3914         lwkt_reltoken(&oldvmspace->vm_map.token);
3915         vmspace_rel(oldvmspace);
3916 }
3917
3918 /*
3919  * vm_map_hint: return the beginning of the best area suitable for
3920  * creating a new mapping with "prot" protection.
3921  *
3922  * No requirements.
3923  */
3924 vm_offset_t
3925 vm_map_hint(struct proc *p, vm_offset_t addr, vm_prot_t prot)
3926 {
3927         struct vmspace *vms = p->p_vmspace;
3928
3929         if (!randomize_mmap || addr != 0) {
3930                 /*
3931                  * Set a reasonable start point for the hint if it was
3932                  * not specified or if it falls within the heap space.
3933                  * Hinted mmap()s do not allocate out of the heap space.
3934                  */
3935                 if (addr == 0 ||
3936                     (addr >= round_page((vm_offset_t)vms->vm_taddr) &&
3937                      addr < round_page((vm_offset_t)vms->vm_daddr + maxdsiz))) {
3938                         addr = round_page((vm_offset_t)vms->vm_daddr + maxdsiz);
3939                 }
3940
3941                 return addr;
3942         }
3943
3944 #ifdef notyet
3945 #ifdef __i386__
3946         /*
3947          * If executable skip first two pages, otherwise start
3948          * after data + heap region.
3949          */
3950         if ((prot & VM_PROT_EXECUTE) &&
3951             ((vm_offset_t)vms->vm_daddr >= I386_MAX_EXE_ADDR)) {
3952                 addr = (PAGE_SIZE * 2) +
3953                     (karc4random() & (I386_MAX_EXE_ADDR / 2 - 1));
3954                 return (round_page(addr));
3955         }
3956 #endif /* __i386__ */
3957 #endif /* notyet */
3958
3959         addr = (vm_offset_t)vms->vm_daddr + MAXDSIZ;
3960         addr += karc4random() & (MIN((256 * 1024 * 1024), MAXDSIZ) - 1);
3961
3962         return (round_page(addr));
3963 }
3964
3965 /*
3966  * Finds the VM object, offset, and protection for a given virtual address
3967  * in the specified map, assuming a page fault of the type specified.
3968  *
3969  * Leaves the map in question locked for read; return values are guaranteed
3970  * until a vm_map_lookup_done call is performed.  Note that the map argument
3971  * is in/out; the returned map must be used in the call to vm_map_lookup_done.
3972  *
3973  * A handle (out_entry) is returned for use in vm_map_lookup_done, to make
3974  * that fast.
3975  *
3976  * If a lookup is requested with "write protection" specified, the map may
3977  * be changed to perform virtual copying operations, although the data
3978  * referenced will remain the same.
3979  *
3980  * No requirements.
3981  */
3982 int
3983 vm_map_lookup(vm_map_t *var_map,                /* IN/OUT */
3984               vm_offset_t vaddr,
3985               vm_prot_t fault_typea,
3986               vm_map_entry_t *out_entry,        /* OUT */
3987               vm_object_t *object,              /* OUT */
3988               vm_pindex_t *pindex,              /* OUT */
3989               vm_prot_t *out_prot,              /* OUT */
3990               boolean_t *wired)                 /* OUT */
3991 {
3992         vm_map_entry_t entry;
3993         vm_map_t map = *var_map;
3994         vm_prot_t prot;
3995         vm_prot_t fault_type = fault_typea;
3996         int use_read_lock = 1;
3997         int rv = KERN_SUCCESS;
3998
3999 RetryLookup:
4000         if (use_read_lock)
4001                 vm_map_lock_read(map);
4002         else
4003                 vm_map_lock(map);
4004
4005         /*
4006          * If the map has an interesting hint, try it before calling full
4007          * blown lookup routine.
4008          */
4009         entry = map->hint;
4010         cpu_ccfence();
4011         *out_entry = entry;
4012         *object = NULL;
4013
4014         if ((entry == &map->header) ||
4015             (vaddr < entry->start) || (vaddr >= entry->end)) {
4016                 vm_map_entry_t tmp_entry;
4017
4018                 /*
4019                  * Entry was either not a valid hint, or the vaddr was not
4020                  * contained in the entry, so do a full lookup.
4021                  */
4022                 if (!vm_map_lookup_entry(map, vaddr, &tmp_entry)) {
4023                         rv = KERN_INVALID_ADDRESS;
4024                         goto done;
4025                 }
4026
4027                 entry = tmp_entry;
4028                 *out_entry = entry;
4029         }
4030         
4031         /*
4032          * Handle submaps.
4033          */
4034         if (entry->maptype == VM_MAPTYPE_SUBMAP) {
4035                 vm_map_t old_map = map;
4036
4037                 *var_map = map = entry->object.sub_map;
4038                 if (use_read_lock)
4039                         vm_map_unlock_read(old_map);
4040                 else
4041                         vm_map_unlock(old_map);
4042                 use_read_lock = 1;
4043                 goto RetryLookup;
4044         }