xref: /openbsd-src/sys/uvm/uvm_fault.c (revision 9b9d2a55a62c8e82206c25f94fcc7f4e2765250e)
1 /*	$OpenBSD: uvm_fault.c,v 1.85 2015/08/21 16:04:35 visa Exp $	*/
2 /*	$NetBSD: uvm_fault.c,v 1.51 2000/08/06 00:22:53 thorpej Exp $	*/
3 
4 /*
5  * Copyright (c) 1997 Charles D. Cranor and Washington University.
6  * All rights reserved.
7  *
8  * Redistribution and use in source and binary forms, with or without
9  * modification, are permitted provided that the following conditions
10  * are met:
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in the
15  *    documentation and/or other materials provided with the distribution.
16  *
17  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
18  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
19  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
20  * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
21  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
22  * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
23  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
24  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
25  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
26  * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
27  *
28  * from: Id: uvm_fault.c,v 1.1.2.23 1998/02/06 05:29:05 chs Exp
29  */
30 
31 /*
32  * uvm_fault.c: fault handler
33  */
34 
35 #include <sys/param.h>
36 #include <sys/systm.h>
37 #include <sys/kernel.h>
38 #include <sys/proc.h>
39 #include <sys/malloc.h>
40 #include <sys/mman.h>
41 
42 #include <uvm/uvm.h>
43 
44 /*
45  *
46  * a word on page faults:
47  *
48  * types of page faults we handle:
49  *
50  * CASE 1: upper layer faults                   CASE 2: lower layer faults
51  *
52  *    CASE 1A         CASE 1B                  CASE 2A        CASE 2B
53  *    read/write1     write>1                  read/write   +-cow_write/zero
54  *         |             |                         |        |
55  *      +--|--+       +--|--+     +-----+       +  |  +     | +-----+
56  * amap |  V  |       |  ----------->new|          |        | |  ^  |
57  *      +-----+       +-----+     +-----+       +  |  +     | +--|--+
58  *                                                 |        |    |
59  *      +-----+       +-----+                   +--|--+     | +--|--+
60  * uobj | d/c |       | d/c |                   |  V  |     +----|  |
61  *      +-----+       +-----+                   +-----+       +-----+
62  *
63  * d/c = don't care
64  *
65  *   case [0]: layerless fault
66  *	no amap or uobj is present.   this is an error.
67  *
68  *   case [1]: upper layer fault [anon active]
69  *     1A: [read] or [write with anon->an_ref == 1]
70  *		I/O takes place in top level anon and uobj is not touched.
71  *     1B: [write with anon->an_ref > 1]
72  *		new anon is alloc'd and data is copied off ["COW"]
73  *
74  *   case [2]: lower layer fault [uobj]
75  *     2A: [read on non-NULL uobj] or [write to non-copy_on_write area]
76  *		I/O takes place directly in object.
77  *     2B: [write to copy_on_write] or [read on NULL uobj]
78  *		data is "promoted" from uobj to a new anon.
79  *		if uobj is null, then we zero fill.
80  *
81  * we follow the standard UVM locking protocol ordering:
82  *
83  * MAPS => AMAP => UOBJ => ANON => PAGE QUEUES (PQ)
84  * we hold a PG_BUSY page if we unlock for I/O
85  *
86  *
87  * the code is structured as follows:
88  *
89  *     - init the "IN" params in the ufi structure
90  *   ReFault:
91  *     - do lookups [locks maps], check protection, handle needs_copy
92  *     - check for case 0 fault (error)
93  *     - establish "range" of fault
94  *     - if we have an amap lock it and extract the anons
95  *     - if sequential advice deactivate pages behind us
96  *     - at the same time check pmap for unmapped areas and anon for pages
97  *	 that we could map in (and do map it if found)
98  *     - check object for resident pages that we could map in
99  *     - if (case 2) goto Case2
100  *     - >>> handle case 1
101  *           - ensure source anon is resident in RAM
102  *           - if case 1B alloc new anon and copy from source
103  *           - map the correct page in
104  *   Case2:
105  *     - >>> handle case 2
106  *           - ensure source page is resident (if uobj)
107  *           - if case 2B alloc new anon and copy from source (could be zero
108  *		fill if uobj == NULL)
109  *           - map the correct page in
110  *     - done!
111  *
112  * note on paging:
113  *   if we have to do I/O we place a PG_BUSY page in the correct object,
114  * unlock everything, and do the I/O.   when I/O is done we must reverify
115  * the state of the world before assuming that our data structures are
116  * valid.   [because mappings could change while the map is unlocked]
117  *
118  *  alternative 1: unbusy the page in question and restart the page fault
119  *    from the top (ReFault).   this is easy but does not take advantage
120  *    of the information that we already have from our previous lookup,
121  *    although it is possible that the "hints" in the vm_map will help here.
122  *
123  * alternative 2: the system already keeps track of a "version" number of
124  *    a map.   [i.e. every time you write-lock a map (e.g. to change a
125  *    mapping) you bump the version number up by one...]   so, we can save
126  *    the version number of the map before we release the lock and start I/O.
127  *    then when I/O is done we can relock and check the version numbers
128  *    to see if anything changed.    this might save us some over 1 because
129  *    we don't have to unbusy the page and may be less compares(?).
130  *
131  * alternative 3: put in backpointers or a way to "hold" part of a map
132  *    in place while I/O is in progress.   this could be complex to
133  *    implement (especially with structures like amap that can be referenced
134  *    by multiple map entries, and figuring out what should wait could be
135  *    complex as well...).
136  *
137  * given that we are not currently multiprocessor or multithreaded we might
138  * as well choose alternative 2 now.   maybe alternative 3 would be useful
139  * in the future.    XXX keep in mind for future consideration//rechecking.
140  */
141 
142 /*
143  * local data structures
144  */
145 struct uvm_advice {
146 	int nback;
147 	int nforw;
148 };
149 
150 /*
151  * page range array: set up in uvmfault_init().
152  */
153 static struct uvm_advice uvmadvice[MADV_MASK + 1];
154 
155 #define UVM_MAXRANGE 16	/* must be max() of nback+nforw+1 */
156 
157 /*
158  * private prototypes
159  */
160 static void uvmfault_amapcopy(struct uvm_faultinfo *);
161 static __inline void uvmfault_anonflush(struct vm_anon **, int);
162 void	uvmfault_unlockmaps(struct uvm_faultinfo *, boolean_t);
163 void	uvmfault_update_stats(struct uvm_faultinfo *);
164 
165 /*
166  * inline functions
167  */
168 /*
169  * uvmfault_anonflush: try and deactivate pages in specified anons
170  *
171  * => does not have to deactivate page if it is busy
172  */
173 static __inline void
174 uvmfault_anonflush(struct vm_anon **anons, int n)
175 {
176 	int lcv;
177 	struct vm_page *pg;
178 
179 	for (lcv = 0 ; lcv < n ; lcv++) {
180 		if (anons[lcv] == NULL)
181 			continue;
182 		pg = anons[lcv]->an_page;
183 		if (pg && (pg->pg_flags & PG_BUSY) == 0) {
184 			uvm_lock_pageq();
185 			if (pg->wire_count == 0) {
186 				pmap_page_protect(pg, PROT_NONE);
187 				uvm_pagedeactivate(pg);
188 			}
189 			uvm_unlock_pageq();
190 		}
191 	}
192 }
193 
194 /*
195  * normal functions
196  */
197 /*
198  * uvmfault_init: compute proper values for the uvmadvice[] array.
199  */
200 void
201 uvmfault_init()
202 {
203 	int npages;
204 
205 	npages = atop(16384);
206 	if (npages > 0) {
207 		KASSERT(npages <= UVM_MAXRANGE / 2);
208 		uvmadvice[MADV_NORMAL].nforw = npages;
209 		uvmadvice[MADV_NORMAL].nback = npages - 1;
210 	}
211 
212 	npages = atop(32768);
213 	if (npages > 0) {
214 		KASSERT(npages <= UVM_MAXRANGE / 2);
215 		uvmadvice[MADV_SEQUENTIAL].nforw = npages - 1;
216 		uvmadvice[MADV_SEQUENTIAL].nback = npages;
217 	}
218 }
219 
220 /*
221  * uvmfault_amapcopy: clear "needs_copy" in a map.
222  *
223  * => if we are out of RAM we sleep (waiting for more)
224  */
225 static void
226 uvmfault_amapcopy(struct uvm_faultinfo *ufi)
227 {
228 
229 	/* while we haven't done the job */
230 	while (1) {
231 		/* no mapping?  give up. */
232 		if (uvmfault_lookup(ufi, TRUE) == FALSE)
233 			return;
234 
235 		/* copy if needed. */
236 		if (UVM_ET_ISNEEDSCOPY(ufi->entry))
237 			amap_copy(ufi->map, ufi->entry, M_NOWAIT, TRUE,
238 				ufi->orig_rvaddr, ufi->orig_rvaddr + 1);
239 
240 		/* didn't work?  must be out of RAM.  sleep. */
241 		if (UVM_ET_ISNEEDSCOPY(ufi->entry)) {
242 			uvmfault_unlockmaps(ufi, TRUE);
243 			uvm_wait("fltamapcopy");
244 			continue;
245 		}
246 
247 		/* got it! */
248 		uvmfault_unlockmaps(ufi, TRUE);
249 		return;
250 	}
251 	/*NOTREACHED*/
252 }
253 
254 /*
255  * uvmfault_anonget: get data in an anon into a non-busy, non-released
256  * page in that anon.
257  *
258  * => we don't move the page on the queues [gets moved later]
259  * => if we allocate a new page [we_own], it gets put on the queues.
260  *    either way, the result is that the page is on the queues at return time
261  */
262 int
263 uvmfault_anonget(struct uvm_faultinfo *ufi, struct vm_amap *amap,
264     struct vm_anon *anon)
265 {
266 	boolean_t we_own;	/* we own anon's page? */
267 	boolean_t locked;	/* did we relock? */
268 	struct vm_page *pg;
269 	int result;
270 
271 	result = 0;		/* XXX shut up gcc */
272 	uvmexp.fltanget++;
273         /* bump rusage counters */
274 	if (anon->an_page)
275 		curproc->p_ru.ru_minflt++;
276 	else
277 		curproc->p_ru.ru_majflt++;
278 
279 	/* loop until we get it, or fail. */
280 	while (1) {
281 		we_own = FALSE;		/* TRUE if we set PG_BUSY on a page */
282 		pg = anon->an_page;
283 
284 		/* page there?   make sure it is not busy/released. */
285 		if (pg) {
286 			KASSERT(pg->pg_flags & PQ_ANON);
287 			KASSERT(pg->uanon == anon);
288 
289 			/*
290 			 * if the page is busy, we drop all the locks and
291 			 * try again.
292 			 */
293 			if ((pg->pg_flags & (PG_BUSY|PG_RELEASED)) == 0)
294 				return (VM_PAGER_OK);
295 			atomic_setbits_int(&pg->pg_flags, PG_WANTED);
296 			uvmexp.fltpgwait++;
297 
298 			/*
299 			 * the last unlock must be an atomic unlock+wait on
300 			 * the owner of page
301 			 */
302 			uvmfault_unlockall(ufi, amap, NULL, NULL);
303 			UVM_WAIT(pg, 0, "anonget2", 0);
304 			/* ready to relock and try again */
305 		} else {
306 			/* no page, we must try and bring it in. */
307 			pg = uvm_pagealloc(NULL, 0, anon, 0);
308 
309 			if (pg == NULL) {		/* out of RAM.  */
310 				uvmfault_unlockall(ufi, amap, NULL, anon);
311 				uvmexp.fltnoram++;
312 				uvm_wait("flt_noram1");
313 				/* ready to relock and try again */
314 			} else {
315 				/* we set the PG_BUSY bit */
316 				we_own = TRUE;
317 				uvmfault_unlockall(ufi, amap, NULL, anon);
318 
319 				/*
320 				 * we are passing a PG_BUSY+PG_FAKE+PG_CLEAN
321 				 * page into the uvm_swap_get function with
322 				 * all data structures unlocked.  note that
323 				 * it is ok to read an_swslot here because
324 				 * we hold PG_BUSY on the page.
325 				 */
326 				uvmexp.pageins++;
327 				result = uvm_swap_get(pg, anon->an_swslot,
328 				    PGO_SYNCIO);
329 
330 				/*
331 				 * we clean up after the i/o below in the
332 				 * "we_own" case
333 				 */
334 				/* ready to relock and try again */
335 			}
336 		}
337 
338 		/* now relock and try again */
339 		locked = uvmfault_relock(ufi);
340 
341 		/*
342 		 * if we own the page (i.e. we set PG_BUSY), then we need
343 		 * to clean up after the I/O. there are three cases to
344 		 * consider:
345 		 *   [1] page released during I/O: free anon and ReFault.
346 		 *   [2] I/O not OK.   free the page and cause the fault
347 		 *       to fail.
348 		 *   [3] I/O OK!   activate the page and sync with the
349 		 *       non-we_own case (i.e. drop anon lock if not locked).
350 		 */
351 		if (we_own) {
352 			if (pg->pg_flags & PG_WANTED) {
353 				wakeup(pg);
354 			}
355 			/* un-busy! */
356 			atomic_clearbits_int(&pg->pg_flags,
357 			    PG_WANTED|PG_BUSY|PG_FAKE);
358 			UVM_PAGE_OWN(pg, NULL);
359 
360 			/*
361 			 * if we were RELEASED during I/O, then our anon is
362 			 * no longer part of an amap.   we need to free the
363 			 * anon and try again.
364 			 */
365 			if (pg->pg_flags & PG_RELEASED) {
366 				pmap_page_protect(pg, PROT_NONE);
367 				uvm_anfree(anon);	/* frees page for us */
368 				if (locked)
369 					uvmfault_unlockall(ufi, amap, NULL,
370 							   NULL);
371 				uvmexp.fltpgrele++;
372 				return (VM_PAGER_REFAULT);	/* refault! */
373 			}
374 
375 			if (result != VM_PAGER_OK) {
376 				KASSERT(result != VM_PAGER_PEND);
377 
378 				/* remove page from anon */
379 				anon->an_page = NULL;
380 
381 				/*
382 				 * remove the swap slot from the anon
383 				 * and mark the anon as having no real slot.
384 				 * don't free the swap slot, thus preventing
385 				 * it from being used again.
386 				 */
387 				uvm_swap_markbad(anon->an_swslot, 1);
388 				anon->an_swslot = SWSLOT_BAD;
389 
390 				/*
391 				 * note: page was never !PG_BUSY, so it
392 				 * can't be mapped and thus no need to
393 				 * pmap_page_protect it...
394 				 */
395 				uvm_lock_pageq();
396 				uvm_pagefree(pg);
397 				uvm_unlock_pageq();
398 
399 				if (locked)
400 					uvmfault_unlockall(ufi, amap, NULL,
401 					    anon);
402 				return (VM_PAGER_ERROR);
403 			}
404 
405 			/*
406 			 * must be OK, clear modify (already PG_CLEAN)
407 			 * and activate
408 			 */
409 			pmap_clear_modify(pg);
410 			uvm_lock_pageq();
411 			uvm_pageactivate(pg);
412 			uvm_unlock_pageq();
413 		}
414 
415 		/* we were not able to relock.   restart fault. */
416 		if (!locked)
417 			return (VM_PAGER_REFAULT);
418 
419 		/* verify no one touched the amap and moved the anon on us. */
420 		if (ufi != NULL &&
421 		    amap_lookup(&ufi->entry->aref,
422 				ufi->orig_rvaddr - ufi->entry->start) != anon) {
423 
424 			uvmfault_unlockall(ufi, amap, NULL, anon);
425 			return (VM_PAGER_REFAULT);
426 		}
427 
428 		/* try it again! */
429 		uvmexp.fltanretry++;
430 		continue;
431 
432 	} /* while (1) */
433 	/*NOTREACHED*/
434 }
435 
436 /*
437  * Update statistics after fault resolution.
438  * - maxrss
439  */
440 void
441 uvmfault_update_stats(struct uvm_faultinfo *ufi)
442 {
443 	struct vm_map		*map;
444 	struct proc		*p;
445 	vsize_t			 res;
446 #ifndef pmap_resident_count
447 	struct vm_space		*vm;
448 #endif
449 
450 	map = ufi->orig_map;
451 
452 	/* Update the maxrss for the process. */
453 	if (map->flags & VM_MAP_ISVMSPACE) {
454 		p = curproc;
455 		KASSERT(p != NULL && &p->p_vmspace->vm_map == map);
456 
457 #ifdef pmap_resident_count
458 		res = pmap_resident_count(map->pmap);
459 #else
460 		/*
461 		 * Rather inaccurate, but this is the current anon size
462 		 * of the vmspace.  It's basically the resident size
463 		 * minus the mmapped in files/text.
464 		 */
465 		vm = (struct vmspace*)map;
466 		res = vm->dsize;
467 #endif
468 
469 		/* Convert res from pages to kilobytes. */
470 		res <<= (PAGE_SHIFT - 10);
471 
472 		if (p->p_ru.ru_maxrss < res)
473 			p->p_ru.ru_maxrss = res;
474 	}
475 }
476 
477 /*
478  *   F A U L T   -   m a i n   e n t r y   p o i n t
479  */
480 
481 /*
482  * uvm_fault: page fault handler
483  *
484  * => called from MD code to resolve a page fault
485  * => VM data structures usually should be unlocked.   however, it is
486  *	possible to call here with the main map locked if the caller
487  *	gets a write lock, sets it recursive, and then calls us (c.f.
488  *	uvm_map_pageable).   this should be avoided because it keeps
489  *	the map locked off during I/O.
490  */
491 #define MASK(entry)     (UVM_ET_ISCOPYONWRITE(entry) ? \
492 			 ~PROT_WRITE : PROT_MASK)
493 int
494 uvm_fault(vm_map_t orig_map, vaddr_t vaddr, vm_fault_t fault_type,
495     vm_prot_t access_type)
496 {
497 	struct uvm_faultinfo ufi;
498 	vm_prot_t enter_prot;
499 	boolean_t wired, narrow, promote, locked, shadowed;
500 	int npages, nback, nforw, centeridx, result, lcv, gotpages;
501 	vaddr_t startva, currva;
502 	voff_t uoff;
503 	paddr_t pa;
504 	struct vm_amap *amap;
505 	struct uvm_object *uobj;
506 	struct vm_anon *anons_store[UVM_MAXRANGE], **anons, *anon, *oanon;
507 	struct vm_page *pages[UVM_MAXRANGE], *pg, *uobjpage;
508 
509 	anon = NULL;
510 	pg = NULL;
511 
512 	uvmexp.faults++;	/* XXX: locking? */
513 
514 	/* init the IN parameters in the ufi */
515 	ufi.orig_map = orig_map;
516 	ufi.orig_rvaddr = trunc_page(vaddr);
517 	ufi.orig_size = PAGE_SIZE;	/* can't get any smaller than this */
518 	if (fault_type == VM_FAULT_WIRE)
519 		narrow = TRUE;		/* don't look for neighborhood
520 					 * pages on wire */
521 	else
522 		narrow = FALSE;		/* normal fault */
523 
524 	/* "goto ReFault" means restart the page fault from ground zero. */
525 ReFault:
526 	/* lookup and lock the maps */
527 	if (uvmfault_lookup(&ufi, FALSE) == FALSE) {
528 		return (EFAULT);
529 	}
530 
531 #ifdef DIAGNOSTIC
532 	if ((ufi.map->flags & VM_MAP_PAGEABLE) == 0)
533 		panic("uvm_fault: fault on non-pageable map (%p, 0x%lx)",
534 		    ufi.map, vaddr);
535 #endif
536 
537 	/* check protection */
538 	if ((ufi.entry->protection & access_type) != access_type) {
539 		uvmfault_unlockmaps(&ufi, FALSE);
540 		return (EACCES);
541 	}
542 
543 	/*
544 	 * "enter_prot" is the protection we want to enter the page in at.
545 	 * for certain pages (e.g. copy-on-write pages) this protection can
546 	 * be more strict than ufi.entry->protection.  "wired" means either
547 	 * the entry is wired or we are fault-wiring the pg.
548 	 */
549 
550 	enter_prot = ufi.entry->protection;
551 	wired = VM_MAPENT_ISWIRED(ufi.entry) || (fault_type == VM_FAULT_WIRE);
552 	if (wired)
553 		access_type = enter_prot; /* full access for wired */
554 
555 	/* handle "needs_copy" case. */
556 	if (UVM_ET_ISNEEDSCOPY(ufi.entry)) {
557 		if ((access_type & PROT_WRITE) ||
558 		    (ufi.entry->object.uvm_obj == NULL)) {
559 			/* need to clear */
560 			uvmfault_unlockmaps(&ufi, FALSE);
561 			uvmfault_amapcopy(&ufi);
562 			uvmexp.fltamcopy++;
563 			goto ReFault;
564 		} else {
565 			/*
566 			 * ensure that we pmap_enter page R/O since
567 			 * needs_copy is still true
568 			 */
569 			enter_prot &= ~PROT_WRITE;
570 		}
571 	}
572 
573 	/* identify the players */
574 	amap = ufi.entry->aref.ar_amap;		/* top layer */
575 	uobj = ufi.entry->object.uvm_obj;	/* bottom layer */
576 
577 	/*
578 	 * check for a case 0 fault.  if nothing backing the entry then
579 	 * error now.
580 	 */
581 	if (amap == NULL && uobj == NULL) {
582 		uvmfault_unlockmaps(&ufi, FALSE);
583 		return (EFAULT);
584 	}
585 
586 	/*
587 	 * establish range of interest based on advice from mapper
588 	 * and then clip to fit map entry.   note that we only want
589 	 * to do this the first time through the fault.   if we
590 	 * ReFault we will disable this by setting "narrow" to true.
591 	 */
592 	if (narrow == FALSE) {
593 
594 		/* wide fault (!narrow) */
595 		nback = min(uvmadvice[ufi.entry->advice].nback,
596 			    (ufi.orig_rvaddr - ufi.entry->start) >> PAGE_SHIFT);
597 		startva = ufi.orig_rvaddr - ((vsize_t)nback << PAGE_SHIFT);
598 		nforw = min(uvmadvice[ufi.entry->advice].nforw,
599 			    ((ufi.entry->end - ufi.orig_rvaddr) >>
600 			     PAGE_SHIFT) - 1);
601 		/*
602 		 * note: "-1" because we don't want to count the
603 		 * faulting page as forw
604 		 */
605 		npages = nback + nforw + 1;
606 		centeridx = nback;
607 
608 		narrow = TRUE;	/* ensure only once per-fault */
609 	} else {
610 		/* narrow fault! */
611 		nback = nforw = 0;
612 		startva = ufi.orig_rvaddr;
613 		npages = 1;
614 		centeridx = 0;
615 	}
616 
617 	/* if we've got an amap, extract current anons. */
618 	if (amap) {
619 		anons = anons_store;
620 		amap_lookups(&ufi.entry->aref, startva - ufi.entry->start,
621 		    anons, npages);
622 	} else {
623 		anons = NULL;	/* to be safe */
624 	}
625 
626 	/*
627 	 * for MADV_SEQUENTIAL mappings we want to deactivate the back pages
628 	 * now and then forget about them (for the rest of the fault).
629 	 */
630 	if (ufi.entry->advice == MADV_SEQUENTIAL && nback != 0) {
631 		/* flush back-page anons? */
632 		if (amap)
633 			uvmfault_anonflush(anons, nback);
634 
635 		/* flush object? */
636 		if (uobj) {
637 			uoff = (startva - ufi.entry->start) + ufi.entry->offset;
638 			(void) uobj->pgops->pgo_flush(uobj, uoff, uoff +
639 			    ((vsize_t)nback << PAGE_SHIFT), PGO_DEACTIVATE);
640 		}
641 
642 		/* now forget about the backpages */
643 		if (amap)
644 			anons += nback;
645 		startva += ((vsize_t)nback << PAGE_SHIFT);
646 		npages -= nback;
647 		centeridx = 0;
648 	}
649 
650 	/*
651 	 * map in the backpages and frontpages we found in the amap in hopes
652 	 * of preventing future faults.    we also init the pages[] array as
653 	 * we go.
654 	 */
655 	currva = startva;
656 	shadowed = FALSE;
657 	for (lcv = 0 ; lcv < npages ; lcv++, currva += PAGE_SIZE) {
658 		/*
659 		 * dont play with VAs that are already mapped
660 		 * except for center)
661 		 */
662 		if (lcv != centeridx &&
663 		    pmap_extract(ufi.orig_map->pmap, currva, &pa)) {
664 			pages[lcv] = PGO_DONTCARE;
665 			continue;
666 		}
667 
668 		/* unmapped or center page.   check if any anon at this level. */
669 		if (amap == NULL || anons[lcv] == NULL) {
670 			pages[lcv] = NULL;
671 			continue;
672 		}
673 
674 		/* check for present page and map if possible.   re-activate it. */
675 		pages[lcv] = PGO_DONTCARE;
676 		if (lcv == centeridx) {		/* save center for later! */
677 			shadowed = TRUE;
678 			continue;
679 		}
680 		anon = anons[lcv];
681 		if (anon->an_page &&
682 		    (anon->an_page->pg_flags & (PG_RELEASED|PG_BUSY)) == 0) {
683 			uvm_lock_pageq();
684 			uvm_pageactivate(anon->an_page);	/* reactivate */
685 			uvm_unlock_pageq();
686 			uvmexp.fltnamap++;
687 
688 			/*
689 			 * Since this isn't the page that's actually faulting,
690 			 * ignore pmap_enter() failures; it's not critical
691 			 * that we enter these right now.
692 			 */
693 			(void) pmap_enter(ufi.orig_map->pmap, currva,
694 			    VM_PAGE_TO_PHYS(anon->an_page),
695 			    (anon->an_ref > 1) ? (enter_prot & ~PROT_WRITE) :
696 			    enter_prot,
697 			    PMAP_CANFAIL |
698 			     (VM_MAPENT_ISWIRED(ufi.entry) ? PMAP_WIRED : 0));
699 		}
700 	}
701 	if (npages > 1)
702 		pmap_update(ufi.orig_map->pmap);
703 
704 	/* (shadowed == TRUE) if there is an anon at the faulting address */
705 	/*
706 	 * note that if we are really short of RAM we could sleep in the above
707 	 * call to pmap_enter.   bad?
708 	 *
709 	 * XXX Actually, that is bad; pmap_enter() should just fail in that
710 	 * XXX case.  --thorpej
711 	 */
712 	/*
713 	 * if the desired page is not shadowed by the amap and we have a
714 	 * backing object, then we check to see if the backing object would
715 	 * prefer to handle the fault itself (rather than letting us do it
716 	 * with the usual pgo_get hook).  the backing object signals this by
717 	 * providing a pgo_fault routine.
718 	 */
719 	if (uobj && shadowed == FALSE && uobj->pgops->pgo_fault != NULL) {
720 		result = uobj->pgops->pgo_fault(&ufi, startva, pages, npages,
721 				    centeridx, fault_type, access_type,
722 				    PGO_LOCKED);
723 
724 		if (result == VM_PAGER_OK)
725 			return (0);		/* pgo_fault did pmap enter */
726 		else if (result == VM_PAGER_REFAULT)
727 			goto ReFault;		/* try again! */
728 		else
729 			return (EACCES);
730 	}
731 
732 	/*
733 	 * now, if the desired page is not shadowed by the amap and we have
734 	 * a backing object that does not have a special fault routine, then
735 	 * we ask (with pgo_get) the object for resident pages that we care
736 	 * about and attempt to map them in.  we do not let pgo_get block
737 	 * (PGO_LOCKED).
738 	 *
739 	 * ("get" has the option of doing a pmap_enter for us)
740 	 */
741 	if (uobj && shadowed == FALSE) {
742 		uvmexp.fltlget++;
743 		gotpages = npages;
744 		(void) uobj->pgops->pgo_get(uobj, ufi.entry->offset +
745 				(startva - ufi.entry->start),
746 				pages, &gotpages, centeridx,
747 				access_type & MASK(ufi.entry),
748 				ufi.entry->advice, PGO_LOCKED);
749 
750 		/* check for pages to map, if we got any */
751 		uobjpage = NULL;
752 		if (gotpages) {
753 			currva = startva;
754 			for (lcv = 0 ; lcv < npages ;
755 			    lcv++, currva += PAGE_SIZE) {
756 				if (pages[lcv] == NULL ||
757 				    pages[lcv] == PGO_DONTCARE)
758 					continue;
759 
760 				KASSERT((pages[lcv]->pg_flags & PG_RELEASED) == 0);
761 
762 				/*
763 				 * if center page is resident and not
764 				 * PG_BUSY, then pgo_get made it PG_BUSY
765 				 * for us and gave us a handle to it.
766 				 * remember this page as "uobjpage."
767 				 * (for later use).
768 				 */
769 				if (lcv == centeridx) {
770 					uobjpage = pages[lcv];
771 					continue;
772 				}
773 
774 				/*
775 				 * note: calling pgo_get with locked data
776 				 * structures returns us pages which are
777 				 * neither busy nor released, so we don't
778 				 * need to check for this.   we can just
779 				 * directly enter the page (after moving it
780 				 * to the head of the active queue [useful?]).
781 				 */
782 
783 				uvm_lock_pageq();
784 				uvm_pageactivate(pages[lcv]);	/* reactivate */
785 				uvm_unlock_pageq();
786 				uvmexp.fltnomap++;
787 
788 				/*
789 				 * Since this page isn't the page that's
790 				 * actually faulting, ignore pmap_enter()
791 				 * failures; it's not critical that we
792 				 * enter these right now.
793 				 */
794 				(void) pmap_enter(ufi.orig_map->pmap, currva,
795 				    VM_PAGE_TO_PHYS(pages[lcv]),
796 				    enter_prot & MASK(ufi.entry),
797 				    PMAP_CANFAIL |
798 				     (wired ? PMAP_WIRED : 0));
799 
800 				/*
801 				 * NOTE: page can't be PG_WANTED because
802 				 * we've held the lock the whole time
803 				 * we've had the handle.
804 				 */
805 				atomic_clearbits_int(&pages[lcv]->pg_flags,
806 				    PG_BUSY);
807 				UVM_PAGE_OWN(pages[lcv], NULL);
808 			}	/* for "lcv" loop */
809 			pmap_update(ufi.orig_map->pmap);
810 		}   /* "gotpages" != 0 */
811 		/* note: object still _locked_ */
812 	} else {
813 		uobjpage = NULL;
814 	}
815 
816 	/*
817 	 * note that at this point we are done with any front or back pages.
818 	 * we are now going to focus on the center page (i.e. the one we've
819 	 * faulted on).  if we have faulted on the top (anon) layer
820 	 * [i.e. case 1], then the anon we want is anons[centeridx] (we have
821 	 * not touched it yet).  if we have faulted on the bottom (uobj)
822 	 * layer [i.e. case 2] and the page was both present and available,
823 	 * then we've got a pointer to it as "uobjpage" and we've already
824 	 * made it BUSY.
825 	 */
826 	/*
827 	 * there are four possible cases we must address: 1A, 1B, 2A, and 2B
828 	 */
829 	/* redirect case 2: if we are not shadowed, go to case 2. */
830 	if (shadowed == FALSE)
831 		goto Case2;
832 
833 	/* handle case 1: fault on an anon in our amap */
834 	anon = anons[centeridx];
835 
836 	/*
837 	 * no matter if we have case 1A or case 1B we are going to need to
838 	 * have the anon's memory resident.   ensure that now.
839 	 */
840 	/*
841 	 * let uvmfault_anonget do the dirty work.
842 	 * also, if it is OK, then the anon's page is on the queues.
843 	 */
844 	result = uvmfault_anonget(&ufi, amap, anon);
845 	switch (result) {
846 	case VM_PAGER_OK:
847 		break;
848 
849 	case VM_PAGER_REFAULT:
850 		goto ReFault;
851 
852 	case VM_PAGER_ERROR:
853 		/*
854 		 * An error occured while trying to bring in the
855 		 * page -- this is the only error we return right
856 		 * now.
857 		 */
858 		return (EACCES);	/* XXX */
859 	default:
860 #ifdef DIAGNOSTIC
861 		panic("uvm_fault: uvmfault_anonget -> %d", result);
862 #else
863 		return (EACCES);
864 #endif
865 	}
866 
867 	/*
868 	 * if we are case 1B then we will need to allocate a new blank
869 	 * anon to transfer the data into.   note that we have a lock
870 	 * on anon, so no one can busy or release the page until we are done.
871 	 * also note that the ref count can't drop to zero here because
872 	 * it is > 1 and we are only dropping one ref.
873 	 *
874 	 * in the (hopefully very rare) case that we are out of RAM we
875 	 * will wait for more RAM, and refault.
876 	 *
877 	 * if we are out of anon VM we kill the process (XXX: could wait?).
878 	 */
879 
880 	if ((access_type & PROT_WRITE) != 0 && anon->an_ref > 1) {
881 		uvmexp.flt_acow++;
882 		oanon = anon;		/* oanon = old */
883 		anon = uvm_analloc();
884 		if (anon) {
885 			pg = uvm_pagealloc(NULL, 0, anon, 0);
886 		}
887 
888 		/* check for out of RAM */
889 		if (anon == NULL || pg == NULL) {
890 			if (anon)
891 				uvm_anfree(anon);
892 			uvmfault_unlockall(&ufi, amap, NULL, oanon);
893 			KASSERT(uvmexp.swpgonly <= uvmexp.swpages);
894 			if (anon == NULL || uvmexp.swpgonly == uvmexp.swpages) {
895 				uvmexp.fltnoanon++;
896 				return (ENOMEM);
897 			}
898 
899 			uvmexp.fltnoram++;
900 			uvm_wait("flt_noram3");	/* out of RAM, wait for more */
901 			goto ReFault;
902 		}
903 
904 		/* got all resources, replace anon with nanon */
905 		uvm_pagecopy(oanon->an_page, pg);	/* pg now !PG_CLEAN */
906 		/* un-busy! new page */
907 		atomic_clearbits_int(&pg->pg_flags, PG_BUSY|PG_FAKE);
908 		UVM_PAGE_OWN(pg, NULL);
909 		amap_add(&ufi.entry->aref, ufi.orig_rvaddr - ufi.entry->start,
910 		    anon, 1);
911 
912 		/* deref: can not drop to zero here by defn! */
913 		oanon->an_ref--;
914 
915 		/*
916 		 * note: anon is _not_ locked, but we have the sole references
917 		 * to in from amap.
918 		 * thus, no one can get at it until we are done with it.
919 		 */
920 	} else {
921 		uvmexp.flt_anon++;
922 		oanon = anon;
923 		pg = anon->an_page;
924 		if (anon->an_ref > 1)     /* disallow writes to ref > 1 anons */
925 			enter_prot = enter_prot & ~PROT_WRITE;
926 	}
927 
928 	/*
929 	 * now map the page in ...
930 	 * XXX: old fault unlocks object before pmap_enter.  this seems
931 	 * suspect since some other thread could blast the page out from
932 	 * under us between the unlock and the pmap_enter.
933 	 */
934 	if (pmap_enter(ufi.orig_map->pmap, ufi.orig_rvaddr, VM_PAGE_TO_PHYS(pg),
935 	    enter_prot, access_type | PMAP_CANFAIL | (wired ? PMAP_WIRED : 0))
936 	    != 0) {
937 		/*
938 		 * No need to undo what we did; we can simply think of
939 		 * this as the pmap throwing away the mapping information.
940 		 *
941 		 * We do, however, have to go through the ReFault path,
942 		 * as the map may change while we're asleep.
943 		 */
944 		uvmfault_unlockall(&ufi, amap, NULL, oanon);
945 		KASSERT(uvmexp.swpgonly <= uvmexp.swpages);
946 		if (uvmexp.swpgonly == uvmexp.swpages) {
947 			/* XXX instrumentation */
948 			return (ENOMEM);
949 		}
950 		/* XXX instrumentation */
951 		uvm_wait("flt_pmfail1");
952 		goto ReFault;
953 	}
954 
955 	/* ... update the page queues. */
956 	uvm_lock_pageq();
957 
958 	if (fault_type == VM_FAULT_WIRE) {
959 		uvm_pagewire(pg);
960 		/*
961 		 * since the now-wired page cannot be paged out,
962 		 * release its swap resources for others to use.
963 		 * since an anon with no swap cannot be PG_CLEAN,
964 		 * clear its clean flag now.
965 		 */
966 		atomic_clearbits_int(&pg->pg_flags, PG_CLEAN);
967 		uvm_anon_dropswap(anon);
968 	} else {
969 		/* activate it */
970 		uvm_pageactivate(pg);
971 	}
972 
973 	uvm_unlock_pageq();
974 
975 	/* done case 1!  finish up by unlocking everything and returning success */
976 	uvmfault_unlockall(&ufi, amap, NULL, oanon);
977 	pmap_update(ufi.orig_map->pmap);
978 	return (0);
979 
980 
981 Case2:
982 	/* handle case 2: faulting on backing object or zero fill */
983 	/*
984 	 * note that uobjpage can not be PGO_DONTCARE at this point.  we now
985 	 * set uobjpage to PGO_DONTCARE if we are doing a zero fill.  if we
986 	 * have a backing object, check and see if we are going to promote
987 	 * the data up to an anon during the fault.
988 	 */
989 	if (uobj == NULL) {
990 		uobjpage = PGO_DONTCARE;
991 		promote = TRUE;		/* always need anon here */
992 	} else {
993 		KASSERT(uobjpage != PGO_DONTCARE);
994 		promote = (access_type & PROT_WRITE) &&
995 		     UVM_ET_ISCOPYONWRITE(ufi.entry);
996 	}
997 
998 	/*
999 	 * if uobjpage is not null then we do not need to do I/O to get the
1000 	 * uobjpage.
1001 	 *
1002 	 * if uobjpage is null, then we need to ask the pager to
1003 	 * get the data for us.   once we have the data, we need to reverify
1004 	 * the state the world.   we are currently not holding any resources.
1005 	 */
1006 	if (uobjpage) {
1007 		/* update rusage counters */
1008 		curproc->p_ru.ru_minflt++;
1009 	} else {
1010 		/* update rusage counters */
1011 		curproc->p_ru.ru_majflt++;
1012 
1013 		uvmfault_unlockall(&ufi, amap, NULL, NULL);
1014 
1015 		uvmexp.fltget++;
1016 		gotpages = 1;
1017 		uoff = (ufi.orig_rvaddr - ufi.entry->start) + ufi.entry->offset;
1018 		result = uobj->pgops->pgo_get(uobj, uoff, &uobjpage, &gotpages,
1019 		    0, access_type & MASK(ufi.entry), ufi.entry->advice,
1020 		    PGO_SYNCIO);
1021 
1022 		/* recover from I/O */
1023 		if (result != VM_PAGER_OK) {
1024 			KASSERT(result != VM_PAGER_PEND);
1025 
1026 			if (result == VM_PAGER_AGAIN) {
1027 				tsleep(&lbolt, PVM, "fltagain2", 0);
1028 				goto ReFault;
1029 			}
1030 
1031 			if (!UVM_ET_ISNOFAULT(ufi.entry))
1032 				return (EACCES); /* XXX i/o error */
1033 
1034 			uobjpage = PGO_DONTCARE;
1035 			promote = TRUE;
1036 		}
1037 
1038 		/* re-verify the state of the world.  */
1039 		locked = uvmfault_relock(&ufi);
1040 
1041 		/*
1042 		 * Re-verify that amap slot is still free. if there is
1043 		 * a problem, we clean up.
1044 		 */
1045 		if (locked && amap && amap_lookup(&ufi.entry->aref,
1046 		      ufi.orig_rvaddr - ufi.entry->start)) {
1047 			if (locked)
1048 				uvmfault_unlockall(&ufi, amap, NULL, NULL);
1049 			locked = FALSE;
1050 		}
1051 
1052 		/* didn't get the lock?   release the page and retry. */
1053 		if (locked == FALSE && uobjpage != PGO_DONTCARE) {
1054 			uvm_lock_pageq();
1055 			/* make sure it is in queues */
1056 			uvm_pageactivate(uobjpage);
1057 			uvm_unlock_pageq();
1058 
1059 			if (uobjpage->pg_flags & PG_WANTED)
1060 				/* still holding object lock */
1061 				wakeup(uobjpage);
1062 			atomic_clearbits_int(&uobjpage->pg_flags,
1063 			    PG_BUSY|PG_WANTED);
1064 			UVM_PAGE_OWN(uobjpage, NULL);
1065 			goto ReFault;
1066 		}
1067 
1068 		/*
1069 		 * we have the data in uobjpage which is PG_BUSY
1070 		 */
1071 	}
1072 
1073 	/*
1074 	 * notes:
1075 	 *  - at this point uobjpage can not be NULL
1076 	 *  - at this point uobjpage could be PG_WANTED (handle later)
1077 	 */
1078 	if (promote == FALSE) {
1079 		/*
1080 		 * we are not promoting.   if the mapping is COW ensure that we
1081 		 * don't give more access than we should (e.g. when doing a read
1082 		 * fault on a COPYONWRITE mapping we want to map the COW page in
1083 		 * R/O even though the entry protection could be R/W).
1084 		 *
1085 		 * set "pg" to the page we want to map in (uobjpage, usually)
1086 		 */
1087 		uvmexp.flt_obj++;
1088 		if (UVM_ET_ISCOPYONWRITE(ufi.entry))
1089 			enter_prot &= ~PROT_WRITE;
1090 		pg = uobjpage;		/* map in the actual object */
1091 
1092 		/* assert(uobjpage != PGO_DONTCARE) */
1093 
1094 		/*
1095 		 * we are faulting directly on the page.
1096 		 */
1097 	} else {
1098 		/*
1099 		 * if we are going to promote the data to an anon we
1100 		 * allocate a blank anon here and plug it into our amap.
1101 		 */
1102 #ifdef DIAGNOSTIC
1103 		if (amap == NULL)
1104 			panic("uvm_fault: want to promote data, but no anon");
1105 #endif
1106 
1107 		anon = uvm_analloc();
1108 		if (anon) {
1109 			/*
1110 			 * In `Fill in data...' below, if
1111 			 * uobjpage == PGO_DONTCARE, we want
1112 			 * a zero'd, dirty page, so have
1113 			 * uvm_pagealloc() do that for us.
1114 			 */
1115 			pg = uvm_pagealloc(NULL, 0, anon,
1116 			    (uobjpage == PGO_DONTCARE) ? UVM_PGA_ZERO : 0);
1117 		}
1118 
1119 		/*
1120 		 * out of memory resources?
1121 		 */
1122 		if (anon == NULL || pg == NULL) {
1123 			/* arg!  must unbusy our page and fail or sleep. */
1124 			if (uobjpage != PGO_DONTCARE) {
1125 				uvm_lock_pageq();
1126 				uvm_pageactivate(uobjpage);
1127 				uvm_unlock_pageq();
1128 
1129 				if (uobjpage->pg_flags & PG_WANTED)
1130 					wakeup(uobjpage);
1131 				atomic_clearbits_int(&uobjpage->pg_flags,
1132 				    PG_BUSY|PG_WANTED);
1133 				UVM_PAGE_OWN(uobjpage, NULL);
1134 			}
1135 
1136 			/* unlock and fail ... */
1137 			uvmfault_unlockall(&ufi, amap, uobj, NULL);
1138 			KASSERT(uvmexp.swpgonly <= uvmexp.swpages);
1139 			if (anon == NULL || uvmexp.swpgonly == uvmexp.swpages) {
1140 				uvmexp.fltnoanon++;
1141 				return (ENOMEM);
1142 			}
1143 
1144 			uvm_anfree(anon);
1145 			uvmexp.fltnoram++;
1146 			uvm_wait("flt_noram5");
1147 			goto ReFault;
1148 		}
1149 
1150 		/* fill in the data */
1151 		if (uobjpage != PGO_DONTCARE) {
1152 			uvmexp.flt_prcopy++;
1153 			/* copy page [pg now dirty] */
1154 			uvm_pagecopy(uobjpage, pg);
1155 
1156 			/*
1157 			 * promote to shared amap?  make sure all sharing
1158 			 * procs see it
1159 			 */
1160 			if ((amap_flags(amap) & AMAP_SHARED) != 0) {
1161 				pmap_page_protect(uobjpage, PROT_NONE);
1162 			}
1163 
1164 			/* dispose of uobjpage. drop handle to uobj as well. */
1165 			if (uobjpage->pg_flags & PG_WANTED)
1166 				wakeup(uobjpage);
1167 			atomic_clearbits_int(&uobjpage->pg_flags,
1168 			    PG_BUSY|PG_WANTED);
1169 			UVM_PAGE_OWN(uobjpage, NULL);
1170 			uvm_lock_pageq();
1171 			uvm_pageactivate(uobjpage);
1172 			uvm_unlock_pageq();
1173 			uobj = NULL;
1174 		} else {
1175 			uvmexp.flt_przero++;
1176 			/*
1177 			 * Page is zero'd and marked dirty by uvm_pagealloc()
1178 			 * above.
1179 			 */
1180 		}
1181 
1182 		amap_add(&ufi.entry->aref, ufi.orig_rvaddr - ufi.entry->start,
1183 		    anon, 0);
1184 	}
1185 
1186 	/* note: pg is either the uobjpage or the new page in the new anon */
1187 	/*
1188 	 * all resources are present.   we can now map it in and free our
1189 	 * resources.
1190 	 */
1191 	if (pmap_enter(ufi.orig_map->pmap, ufi.orig_rvaddr, VM_PAGE_TO_PHYS(pg),
1192 	    enter_prot, access_type | PMAP_CANFAIL | (wired ? PMAP_WIRED : 0))
1193 	    != 0) {
1194 		/*
1195 		 * No need to undo what we did; we can simply think of
1196 		 * this as the pmap throwing away the mapping information.
1197 		 *
1198 		 * We do, however, have to go through the ReFault path,
1199 		 * as the map may change while we're asleep.
1200 		 */
1201 		if (pg->pg_flags & PG_WANTED)
1202 			wakeup(pg);
1203 
1204 		atomic_clearbits_int(&pg->pg_flags, PG_BUSY|PG_FAKE|PG_WANTED);
1205 		UVM_PAGE_OWN(pg, NULL);
1206 		uvmfault_unlockall(&ufi, amap, uobj, NULL);
1207 		KASSERT(uvmexp.swpgonly <= uvmexp.swpages);
1208 		if (uvmexp.swpgonly == uvmexp.swpages) {
1209 			/* XXX instrumentation */
1210 			return (ENOMEM);
1211 		}
1212 		/* XXX instrumentation */
1213 		uvm_wait("flt_pmfail2");
1214 		goto ReFault;
1215 	}
1216 
1217 	uvm_lock_pageq();
1218 
1219 	if (fault_type == VM_FAULT_WIRE) {
1220 		uvm_pagewire(pg);
1221 		if (pg->pg_flags & PQ_AOBJ) {
1222 			/*
1223 			 * since the now-wired page cannot be paged out,
1224 			 * release its swap resources for others to use.
1225 			 * since an aobj page with no swap cannot be PG_CLEAN,
1226 			 * clear its clean flag now.
1227 			 */
1228 			atomic_clearbits_int(&pg->pg_flags, PG_CLEAN);
1229 			uao_dropswap(uobj, pg->offset >> PAGE_SHIFT);
1230 		}
1231 	} else {
1232 		/* activate it */
1233 		uvm_pageactivate(pg);
1234 	}
1235 	uvm_unlock_pageq();
1236 
1237 	if (pg->pg_flags & PG_WANTED)
1238 		wakeup(pg);
1239 
1240 	atomic_clearbits_int(&pg->pg_flags, PG_BUSY|PG_FAKE|PG_WANTED);
1241 	UVM_PAGE_OWN(pg, NULL);
1242 	uvmfault_unlockall(&ufi, amap, uobj, NULL);
1243 	pmap_update(ufi.orig_map->pmap);
1244 
1245 	return (0);
1246 }
1247 
1248 
1249 /*
1250  * uvm_fault_wire: wire down a range of virtual addresses in a map.
1251  *
1252  * => map may be read-locked by caller, but MUST NOT be write-locked.
1253  * => if map is read-locked, any operations which may cause map to
1254  *	be write-locked in uvm_fault() must be taken care of by
1255  *	the caller.  See uvm_map_pageable().
1256  */
1257 int
1258 uvm_fault_wire(vm_map_t map, vaddr_t start, vaddr_t end, vm_prot_t access_type)
1259 {
1260 	vaddr_t va;
1261 	pmap_t  pmap;
1262 	int rv;
1263 
1264 	pmap = vm_map_pmap(map);
1265 
1266 	/*
1267 	 * now fault it in a page at a time.   if the fault fails then we have
1268 	 * to undo what we have done.   note that in uvm_fault PROT_NONE
1269 	 * is replaced with the max protection if fault_type is VM_FAULT_WIRE.
1270 	 */
1271 	for (va = start ; va < end ; va += PAGE_SIZE) {
1272 		rv = uvm_fault(map, va, VM_FAULT_WIRE, access_type);
1273 		if (rv) {
1274 			if (va != start) {
1275 				uvm_fault_unwire(map, start, va);
1276 			}
1277 			return (rv);
1278 		}
1279 	}
1280 
1281 	return (0);
1282 }
1283 
1284 /*
1285  * uvm_fault_unwire(): unwire range of virtual space.
1286  */
1287 void
1288 uvm_fault_unwire(vm_map_t map, vaddr_t start, vaddr_t end)
1289 {
1290 
1291 	vm_map_lock_read(map);
1292 	uvm_fault_unwire_locked(map, start, end);
1293 	vm_map_unlock_read(map);
1294 }
1295 
1296 /*
1297  * uvm_fault_unwire_locked(): the guts of uvm_fault_unwire().
1298  *
1299  * => map must be at least read-locked.
1300  */
1301 void
1302 uvm_fault_unwire_locked(vm_map_t map, vaddr_t start, vaddr_t end)
1303 {
1304 	vm_map_entry_t entry, next;
1305 	pmap_t pmap = vm_map_pmap(map);
1306 	vaddr_t va;
1307 	paddr_t pa;
1308 	struct vm_page *pg;
1309 
1310 	KASSERT((map->flags & VM_MAP_INTRSAFE) == 0);
1311 
1312 	/*
1313 	 * we assume that the area we are unwiring has actually been wired
1314 	 * in the first place.   this means that we should be able to extract
1315 	 * the PAs from the pmap.   we also lock out the page daemon so that
1316 	 * we can call uvm_pageunwire.
1317 	 */
1318 	uvm_lock_pageq();
1319 
1320 	/* find the beginning map entry for the region. */
1321 	KASSERT(start >= vm_map_min(map) && end <= vm_map_max(map));
1322 	if (uvm_map_lookup_entry(map, start, &entry) == FALSE)
1323 		panic("uvm_fault_unwire_locked: address not in map");
1324 
1325 	for (va = start; va < end ; va += PAGE_SIZE) {
1326 		if (pmap_extract(pmap, va, &pa) == FALSE)
1327 			continue;
1328 
1329 		/* find the map entry for the current address. */
1330 		KASSERT(va >= entry->start);
1331 		while (va >= entry->end) {
1332 			next = RB_NEXT(uvm_map_addr, &map->addr, entry);
1333 			KASSERT(next != NULL && next->start <= entry->end);
1334 			entry = next;
1335 		}
1336 
1337 		/* if the entry is no longer wired, tell the pmap. */
1338 		if (VM_MAPENT_ISWIRED(entry) == 0)
1339 			pmap_unwire(pmap, va);
1340 
1341 		pg = PHYS_TO_VM_PAGE(pa);
1342 		if (pg)
1343 			uvm_pageunwire(pg);
1344 	}
1345 
1346 	uvm_unlock_pageq();
1347 }
1348 
1349 /*
1350  * uvmfault_unlockmaps: unlock the maps
1351  */
1352 void
1353 uvmfault_unlockmaps(struct uvm_faultinfo *ufi, boolean_t write_locked)
1354 {
1355 	/*
1356 	 * ufi can be NULL when this isn't really a fault,
1357 	 * but merely paging in anon data.
1358 	 */
1359 	if (ufi == NULL) {
1360 		return;
1361 	}
1362 
1363 	uvmfault_update_stats(ufi);
1364 	if (write_locked) {
1365 		vm_map_unlock(ufi->map);
1366 	} else {
1367 		vm_map_unlock_read(ufi->map);
1368 	}
1369 }
1370 
1371 /*
1372  * uvmfault_unlockall: unlock everything passed in.
1373  *
1374  * => maps must be read-locked (not write-locked).
1375  */
1376 void
1377 uvmfault_unlockall(struct uvm_faultinfo *ufi, struct vm_amap *amap,
1378     struct uvm_object *uobj, struct vm_anon *anon)
1379 {
1380 
1381 	uvmfault_unlockmaps(ufi, FALSE);
1382 }
1383 
1384 /*
1385  * uvmfault_lookup: lookup a virtual address in a map
1386  *
1387  * => caller must provide a uvm_faultinfo structure with the IN
1388  *	params properly filled in
1389  * => we will lookup the map entry (handling submaps) as we go
1390  * => if the lookup is a success we will return with the maps locked
1391  * => if "write_lock" is TRUE, we write_lock the map, otherwise we only
1392  *	get a read lock.
1393  * => note that submaps can only appear in the kernel and they are
1394  *	required to use the same virtual addresses as the map they
1395  *	are referenced by (thus address translation between the main
1396  *	map and the submap is unnecessary).
1397  */
1398 
1399 boolean_t
1400 uvmfault_lookup(struct uvm_faultinfo *ufi, boolean_t write_lock)
1401 {
1402 	vm_map_t tmpmap;
1403 
1404 	/* init ufi values for lookup. */
1405 	ufi->map = ufi->orig_map;
1406 	ufi->size = ufi->orig_size;
1407 
1408 	/*
1409 	 * keep going down levels until we are done.   note that there can
1410 	 * only be two levels so we won't loop very long.
1411 	 */
1412 	while (1) {
1413 		if (ufi->orig_rvaddr < ufi->map->min_offset ||
1414 		    ufi->orig_rvaddr >= ufi->map->max_offset)
1415 			return(FALSE);
1416 
1417 		/* lock map */
1418 		if (write_lock) {
1419 			vm_map_lock(ufi->map);
1420 		} else {
1421 			vm_map_lock_read(ufi->map);
1422 		}
1423 
1424 		/* lookup */
1425 		if (!uvm_map_lookup_entry(ufi->map, ufi->orig_rvaddr,
1426 		    &ufi->entry)) {
1427 			uvmfault_unlockmaps(ufi, write_lock);
1428 			return(FALSE);
1429 		}
1430 
1431 		/* reduce size if necessary */
1432 		if (ufi->entry->end - ufi->orig_rvaddr < ufi->size)
1433 			ufi->size = ufi->entry->end - ufi->orig_rvaddr;
1434 
1435 		/*
1436 		 * submap?    replace map with the submap and lookup again.
1437 		 * note: VAs in submaps must match VAs in main map.
1438 		 */
1439 		if (UVM_ET_ISSUBMAP(ufi->entry)) {
1440 			tmpmap = ufi->entry->object.sub_map;
1441 			uvmfault_unlockmaps(ufi, write_lock);
1442 			ufi->map = tmpmap;
1443 			continue;
1444 		}
1445 
1446 		/* got it! */
1447 		ufi->mapv = ufi->map->timestamp;
1448 		return(TRUE);
1449 
1450 	}
1451 	/*NOTREACHED*/
1452 }
1453 
1454 /*
1455  * uvmfault_relock: attempt to relock the same version of the map
1456  *
1457  * => fault data structures should be unlocked before calling.
1458  * => if a success (TRUE) maps will be locked after call.
1459  */
1460 boolean_t
1461 uvmfault_relock(struct uvm_faultinfo *ufi)
1462 {
1463 	/*
1464 	 * ufi can be NULL when this isn't really a fault,
1465 	 * but merely paging in anon data.
1466 	 */
1467 	if (ufi == NULL) {
1468 		return TRUE;
1469 	}
1470 
1471 	uvmexp.fltrelck++;
1472 
1473 	/*
1474 	 * relock map.   fail if version mismatch (in which case nothing
1475 	 * gets locked).
1476 	 */
1477 	vm_map_lock_read(ufi->map);
1478 	if (ufi->mapv != ufi->map->timestamp) {
1479 		vm_map_unlock_read(ufi->map);
1480 		return(FALSE);
1481 	}
1482 
1483 	uvmexp.fltrelckok++;
1484 	return(TRUE);		/* got it! */
1485 }
1486