xref: /netbsd-src/sys/uvm/uvm_fault.c (revision 23c8222edbfb0f0932d88a8351d3a0cf817dfb9e)
1 /*	$NetBSD: uvm_fault.c,v 1.88 2004/05/05 11:54:32 yamt Exp $	*/
2 
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  * 3. All advertising materials mentioning features or use of this software
17  *    must display the following acknowledgement:
18  *      This product includes software developed by Charles D. Cranor and
19  *      Washington University.
20  * 4. The name of the author may not be used to endorse or promote products
21  *    derived from this software without specific prior written permission.
22  *
23  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
24  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
25  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
26  * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
27  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
28  * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
29  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
30  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
31  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
32  * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
33  *
34  * from: Id: uvm_fault.c,v 1.1.2.23 1998/02/06 05:29:05 chs Exp
35  */
36 
37 /*
38  * uvm_fault.c: fault handler
39  */
40 
41 #include <sys/cdefs.h>
42 __KERNEL_RCSID(0, "$NetBSD: uvm_fault.c,v 1.88 2004/05/05 11:54:32 yamt Exp $");
43 
44 #include "opt_uvmhist.h"
45 
46 #include <sys/param.h>
47 #include <sys/systm.h>
48 #include <sys/kernel.h>
49 #include <sys/proc.h>
50 #include <sys/malloc.h>
51 #include <sys/mman.h>
52 #include <sys/user.h>
53 
54 #include <uvm/uvm.h>
55 
56 /*
57  *
58  * a word on page faults:
59  *
60  * types of page faults we handle:
61  *
62  * CASE 1: upper layer faults                   CASE 2: lower layer faults
63  *
64  *    CASE 1A         CASE 1B                  CASE 2A        CASE 2B
65  *    read/write1     write>1                  read/write   +-cow_write/zero
66  *         |             |                         |        |
67  *      +--|--+       +--|--+     +-----+       +  |  +     | +-----+
68  * amap |  V  |       |  ----------->new|          |        | |  ^  |
69  *      +-----+       +-----+     +-----+       +  |  +     | +--|--+
70  *                                                 |        |    |
71  *      +-----+       +-----+                   +--|--+     | +--|--+
72  * uobj | d/c |       | d/c |                   |  V  |     +----|  |
73  *      +-----+       +-----+                   +-----+       +-----+
74  *
75  * d/c = don't care
76  *
77  *   case [0]: layerless fault
78  *	no amap or uobj is present.   this is an error.
79  *
80  *   case [1]: upper layer fault [anon active]
81  *     1A: [read] or [write with anon->an_ref == 1]
82  *		I/O takes place in top level anon and uobj is not touched.
83  *     1B: [write with anon->an_ref > 1]
84  *		new anon is alloc'd and data is copied off ["COW"]
85  *
86  *   case [2]: lower layer fault [uobj]
87  *     2A: [read on non-NULL uobj] or [write to non-copy_on_write area]
88  *		I/O takes place directly in object.
89  *     2B: [write to copy_on_write] or [read on NULL uobj]
90  *		data is "promoted" from uobj to a new anon.
91  *		if uobj is null, then we zero fill.
92  *
93  * we follow the standard UVM locking protocol ordering:
94  *
95  * MAPS => AMAP => UOBJ => ANON => PAGE QUEUES (PQ)
96  * we hold a PG_BUSY page if we unlock for I/O
97  *
98  *
99  * the code is structured as follows:
100  *
101  *     - init the "IN" params in the ufi structure
102  *   ReFault:
103  *     - do lookups [locks maps], check protection, handle needs_copy
104  *     - check for case 0 fault (error)
105  *     - establish "range" of fault
106  *     - if we have an amap lock it and extract the anons
107  *     - if sequential advice deactivate pages behind us
108  *     - at the same time check pmap for unmapped areas and anon for pages
109  *	 that we could map in (and do map it if found)
110  *     - check object for resident pages that we could map in
111  *     - if (case 2) goto Case2
112  *     - >>> handle case 1
113  *           - ensure source anon is resident in RAM
114  *           - if case 1B alloc new anon and copy from source
115  *           - map the correct page in
116  *   Case2:
117  *     - >>> handle case 2
118  *           - ensure source page is resident (if uobj)
119  *           - if case 2B alloc new anon and copy from source (could be zero
120  *		fill if uobj == NULL)
121  *           - map the correct page in
122  *     - done!
123  *
124  * note on paging:
125  *   if we have to do I/O we place a PG_BUSY page in the correct object,
126  * unlock everything, and do the I/O.   when I/O is done we must reverify
127  * the state of the world before assuming that our data structures are
128  * valid.   [because mappings could change while the map is unlocked]
129  *
130  *  alternative 1: unbusy the page in question and restart the page fault
131  *    from the top (ReFault).   this is easy but does not take advantage
132  *    of the information that we already have from our previous lookup,
133  *    although it is possible that the "hints" in the vm_map will help here.
134  *
135  * alternative 2: the system already keeps track of a "version" number of
136  *    a map.   [i.e. every time you write-lock a map (e.g. to change a
137  *    mapping) you bump the version number up by one...]   so, we can save
138  *    the version number of the map before we release the lock and start I/O.
139  *    then when I/O is done we can relock and check the version numbers
140  *    to see if anything changed.    this might save us some over 1 because
141  *    we don't have to unbusy the page and may be less compares(?).
142  *
143  * alternative 3: put in backpointers or a way to "hold" part of a map
144  *    in place while I/O is in progress.   this could be complex to
145  *    implement (especially with structures like amap that can be referenced
146  *    by multiple map entries, and figuring out what should wait could be
147  *    complex as well...).
148  *
149  * given that we are not currently multiprocessor or multithreaded we might
150  * as well choose alternative 2 now.   maybe alternative 3 would be useful
151  * in the future.    XXX keep in mind for future consideration//rechecking.
152  */
153 
154 /*
155  * local data structures
156  */
157 
158 struct uvm_advice {
159 	int advice;
160 	int nback;
161 	int nforw;
162 };
163 
164 /*
165  * page range array:
166  * note: index in array must match "advice" value
167  * XXX: borrowed numbers from freebsd.   do they work well for us?
168  */
169 
170 static struct uvm_advice uvmadvice[] = {
171 	{ MADV_NORMAL, 3, 4 },
172 	{ MADV_RANDOM, 0, 0 },
173 	{ MADV_SEQUENTIAL, 8, 7},
174 };
175 
176 #define UVM_MAXRANGE 16	/* must be MAX() of nback+nforw+1 */
177 
178 /*
179  * private prototypes
180  */
181 
182 static void uvmfault_amapcopy(struct uvm_faultinfo *);
183 static __inline void uvmfault_anonflush(struct vm_anon **, int);
184 
185 /*
186  * inline functions
187  */
188 
189 /*
190  * uvmfault_anonflush: try and deactivate pages in specified anons
191  *
192  * => does not have to deactivate page if it is busy
193  */
194 
195 static __inline void
196 uvmfault_anonflush(anons, n)
197 	struct vm_anon **anons;
198 	int n;
199 {
200 	int lcv;
201 	struct vm_page *pg;
202 
203 	for (lcv = 0 ; lcv < n ; lcv++) {
204 		if (anons[lcv] == NULL)
205 			continue;
206 		simple_lock(&anons[lcv]->an_lock);
207 		pg = anons[lcv]->u.an_page;
208 		if (pg && (pg->flags & PG_BUSY) == 0 && pg->loan_count == 0) {
209 			uvm_lock_pageq();
210 			if (pg->wire_count == 0) {
211 				pmap_clear_reference(pg);
212 				uvm_pagedeactivate(pg);
213 			}
214 			uvm_unlock_pageq();
215 		}
216 		simple_unlock(&anons[lcv]->an_lock);
217 	}
218 }
219 
220 /*
221  * normal functions
222  */
223 
224 /*
225  * uvmfault_amapcopy: clear "needs_copy" in a map.
226  *
227  * => called with VM data structures unlocked (usually, see below)
228  * => we get a write lock on the maps and clear needs_copy for a VA
229  * => if we are out of RAM we sleep (waiting for more)
230  */
231 
232 static void
233 uvmfault_amapcopy(ufi)
234 	struct uvm_faultinfo *ufi;
235 {
236 	for (;;) {
237 
238 		/*
239 		 * no mapping?  give up.
240 		 */
241 
242 		if (uvmfault_lookup(ufi, TRUE) == FALSE)
243 			return;
244 
245 		/*
246 		 * copy if needed.
247 		 */
248 
249 		if (UVM_ET_ISNEEDSCOPY(ufi->entry))
250 			amap_copy(ufi->map, ufi->entry, M_NOWAIT, TRUE,
251 				ufi->orig_rvaddr, ufi->orig_rvaddr + 1);
252 
253 		/*
254 		 * didn't work?  must be out of RAM.   unlock and sleep.
255 		 */
256 
257 		if (UVM_ET_ISNEEDSCOPY(ufi->entry)) {
258 			uvmfault_unlockmaps(ufi, TRUE);
259 			uvm_wait("fltamapcopy");
260 			continue;
261 		}
262 
263 		/*
264 		 * got it!   unlock and return.
265 		 */
266 
267 		uvmfault_unlockmaps(ufi, TRUE);
268 		return;
269 	}
270 	/*NOTREACHED*/
271 }
272 
273 /*
274  * uvmfault_anonget: get data in an anon into a non-busy, non-released
275  * page in that anon.
276  *
277  * => maps, amap, and anon locked by caller.
278  * => if we fail (result != 0) we unlock everything.
279  * => if we are successful, we return with everything still locked.
280  * => we don't move the page on the queues [gets moved later]
281  * => if we allocate a new page [we_own], it gets put on the queues.
282  *    either way, the result is that the page is on the queues at return time
283  * => for pages which are on loan from a uvm_object (and thus are not
284  *    owned by the anon): if successful, we return with the owning object
285  *    locked.   the caller must unlock this object when it unlocks everything
286  *    else.
287  */
288 
289 int
290 uvmfault_anonget(ufi, amap, anon)
291 	struct uvm_faultinfo *ufi;
292 	struct vm_amap *amap;
293 	struct vm_anon *anon;
294 {
295 	boolean_t we_own;	/* we own anon's page? */
296 	boolean_t locked;	/* did we relock? */
297 	struct vm_page *pg;
298 	int error;
299 	UVMHIST_FUNC("uvmfault_anonget"); UVMHIST_CALLED(maphist);
300 
301 	LOCK_ASSERT(simple_lock_held(&anon->an_lock));
302 
303 	error = 0;
304 	uvmexp.fltanget++;
305         /* bump rusage counters */
306 	if (anon->u.an_page)
307 		curproc->p_stats->p_ru.ru_minflt++;
308 	else
309 		curproc->p_stats->p_ru.ru_majflt++;
310 
311 	/*
312 	 * loop until we get it, or fail.
313 	 */
314 
315 	for (;;) {
316 		we_own = FALSE;		/* TRUE if we set PG_BUSY on a page */
317 		pg = anon->u.an_page;
318 
319 		/*
320 		 * if there is a resident page and it is loaned, then anon
321 		 * may not own it.   call out to uvm_anon_lockpage() to ensure
322 		 * the real owner of the page has been identified and locked.
323 		 */
324 
325 		if (pg && pg->loan_count)
326 			pg = uvm_anon_lockloanpg(anon);
327 
328 		/*
329 		 * page there?   make sure it is not busy/released.
330 		 */
331 
332 		if (pg) {
333 
334 			/*
335 			 * at this point, if the page has a uobject [meaning
336 			 * we have it on loan], then that uobject is locked
337 			 * by us!   if the page is busy, we drop all the
338 			 * locks (including uobject) and try again.
339 			 */
340 
341 			if ((pg->flags & PG_BUSY) == 0) {
342 				UVMHIST_LOG(maphist, "<- OK",0,0,0,0);
343 				return (0);
344 			}
345 			pg->flags |= PG_WANTED;
346 			uvmexp.fltpgwait++;
347 
348 			/*
349 			 * the last unlock must be an atomic unlock+wait on
350 			 * the owner of page
351 			 */
352 
353 			if (pg->uobject) {	/* owner is uobject ? */
354 				uvmfault_unlockall(ufi, amap, NULL, anon);
355 				UVMHIST_LOG(maphist, " unlock+wait on uobj",0,
356 				    0,0,0);
357 				UVM_UNLOCK_AND_WAIT(pg,
358 				    &pg->uobject->vmobjlock,
359 				    FALSE, "anonget1",0);
360 			} else {
361 				/* anon owns page */
362 				uvmfault_unlockall(ufi, amap, NULL, NULL);
363 				UVMHIST_LOG(maphist, " unlock+wait on anon",0,
364 				    0,0,0);
365 				UVM_UNLOCK_AND_WAIT(pg,&anon->an_lock,0,
366 				    "anonget2",0);
367 			}
368 		} else {
369 
370 			/*
371 			 * no page, we must try and bring it in.
372 			 */
373 
374 			pg = uvm_pagealloc(NULL, 0, anon, 0);
375 			if (pg == NULL) {		/* out of RAM.  */
376 				uvmfault_unlockall(ufi, amap, NULL, anon);
377 				uvmexp.fltnoram++;
378 				UVMHIST_LOG(maphist, "  noram -- UVM_WAIT",0,
379 				    0,0,0);
380 				uvm_wait("flt_noram1");
381 			} else {
382 				/* we set the PG_BUSY bit */
383 				we_own = TRUE;
384 				uvmfault_unlockall(ufi, amap, NULL, anon);
385 
386 				/*
387 				 * we are passing a PG_BUSY+PG_FAKE+PG_CLEAN
388 				 * page into the uvm_swap_get function with
389 				 * all data structures unlocked.  note that
390 				 * it is ok to read an_swslot here because
391 				 * we hold PG_BUSY on the page.
392 				 */
393 				uvmexp.pageins++;
394 				error = uvm_swap_get(pg, anon->an_swslot,
395 				    PGO_SYNCIO);
396 
397 				/*
398 				 * we clean up after the i/o below in the
399 				 * "we_own" case
400 				 */
401 			}
402 		}
403 
404 		/*
405 		 * now relock and try again
406 		 */
407 
408 		locked = uvmfault_relock(ufi);
409 		if (locked && amap != NULL) {
410 			amap_lock(amap);
411 		}
412 		if (locked || we_own)
413 			simple_lock(&anon->an_lock);
414 
415 		/*
416 		 * if we own the page (i.e. we set PG_BUSY), then we need
417 		 * to clean up after the I/O. there are three cases to
418 		 * consider:
419 		 *   [1] page released during I/O: free anon and ReFault.
420 		 *   [2] I/O not OK.   free the page and cause the fault
421 		 *       to fail.
422 		 *   [3] I/O OK!   activate the page and sync with the
423 		 *       non-we_own case (i.e. drop anon lock if not locked).
424 		 */
425 
426 		if (we_own) {
427 			if (pg->flags & PG_WANTED) {
428 				wakeup(pg);
429 			}
430 			if (error) {
431 
432 				/*
433 				 * remove the swap slot from the anon
434 				 * and mark the anon as having no real slot.
435 				 * don't free the swap slot, thus preventing
436 				 * it from being used again.
437 				 */
438 
439 				if (anon->an_swslot > 0)
440 					uvm_swap_markbad(anon->an_swslot, 1);
441 				anon->an_swslot = SWSLOT_BAD;
442 
443 				if ((pg->flags & PG_RELEASED) != 0)
444 					goto released;
445 
446 				/*
447 				 * note: page was never !PG_BUSY, so it
448 				 * can't be mapped and thus no need to
449 				 * pmap_page_protect it...
450 				 */
451 
452 				uvm_lock_pageq();
453 				uvm_pagefree(pg);
454 				uvm_unlock_pageq();
455 
456 				if (locked)
457 					uvmfault_unlockall(ufi, amap, NULL,
458 					    anon);
459 				else
460 					simple_unlock(&anon->an_lock);
461 				UVMHIST_LOG(maphist, "<- ERROR", 0,0,0,0);
462 				return error;
463 			}
464 
465 			if ((pg->flags & PG_RELEASED) != 0) {
466 released:
467 				KASSERT(anon->an_ref == 0);
468 
469 				/*
470 				 * released while we unlocked amap.
471 				 */
472 
473 				if (locked)
474 					uvmfault_unlockall(ufi, amap, NULL,
475 					    NULL);
476 
477 				uvm_anon_release(anon);
478 
479 				if (error) {
480 					UVMHIST_LOG(maphist,
481 					    "<- ERROR/RELEASED", 0,0,0,0);
482 					return error;
483 				}
484 
485 				UVMHIST_LOG(maphist, "<- RELEASED", 0,0,0,0);
486 				return ERESTART;
487 			}
488 
489 			/*
490 			 * we've successfully read the page, activate it.
491 			 */
492 
493 			uvm_lock_pageq();
494 			uvm_pageactivate(pg);
495 			uvm_unlock_pageq();
496 			pg->flags &= ~(PG_WANTED|PG_BUSY|PG_FAKE);
497 			UVM_PAGE_OWN(pg, NULL);
498 			if (!locked)
499 				simple_unlock(&anon->an_lock);
500 		}
501 
502 		/*
503 		 * we were not able to relock.   restart fault.
504 		 */
505 
506 		if (!locked) {
507 			UVMHIST_LOG(maphist, "<- REFAULT", 0,0,0,0);
508 			return (ERESTART);
509 		}
510 
511 		/*
512 		 * verify no one has touched the amap and moved the anon on us.
513 		 */
514 
515 		if (ufi != NULL &&
516 		    amap_lookup(&ufi->entry->aref,
517 				ufi->orig_rvaddr - ufi->entry->start) != anon) {
518 
519 			uvmfault_unlockall(ufi, amap, NULL, anon);
520 			UVMHIST_LOG(maphist, "<- REFAULT", 0,0,0,0);
521 			return (ERESTART);
522 		}
523 
524 		/*
525 		 * try it again!
526 		 */
527 
528 		uvmexp.fltanretry++;
529 		continue;
530 	}
531 	/*NOTREACHED*/
532 }
533 
534 /*
535  *   F A U L T   -   m a i n   e n t r y   p o i n t
536  */
537 
538 /*
539  * uvm_fault: page fault handler
540  *
541  * => called from MD code to resolve a page fault
542  * => VM data structures usually should be unlocked.   however, it is
543  *	possible to call here with the main map locked if the caller
544  *	gets a write lock, sets it recusive, and then calls us (c.f.
545  *	uvm_map_pageable).   this should be avoided because it keeps
546  *	the map locked off during I/O.
547  * => MUST NEVER BE CALLED IN INTERRUPT CONTEXT
548  */
549 
550 #define MASK(entry)     (UVM_ET_ISCOPYONWRITE(entry) ? \
551 			 ~VM_PROT_WRITE : VM_PROT_ALL)
552 
553 int
554 uvm_fault(orig_map, vaddr, fault_type, access_type)
555 	struct vm_map *orig_map;
556 	vaddr_t vaddr;
557 	vm_fault_t fault_type;
558 	vm_prot_t access_type;
559 {
560 	struct uvm_faultinfo ufi;
561 	vm_prot_t enter_prot, check_prot;
562 	boolean_t wired, narrow, promote, locked, shadowed, wire_fault, cow_now;
563 	int npages, nback, nforw, centeridx, error, lcv, gotpages;
564 	vaddr_t startva, objaddr, currva;
565 	voff_t uoff;
566 	paddr_t pa;
567 	struct vm_amap *amap;
568 	struct uvm_object *uobj;
569 	struct vm_anon *anons_store[UVM_MAXRANGE], **anons, *anon, *oanon;
570 	struct vm_page *pages[UVM_MAXRANGE], *pg, *uobjpage;
571 	UVMHIST_FUNC("uvm_fault"); UVMHIST_CALLED(maphist);
572 
573 	UVMHIST_LOG(maphist, "(map=0x%x, vaddr=0x%x, ft=%d, at=%d)",
574 	      orig_map, vaddr, fault_type, access_type);
575 
576 	anon = NULL;
577 	pg = NULL;
578 
579 	uvmexp.faults++;	/* XXX: locking? */
580 
581 	/*
582 	 * init the IN parameters in the ufi
583 	 */
584 
585 	ufi.orig_map = orig_map;
586 	ufi.orig_rvaddr = trunc_page(vaddr);
587 	ufi.orig_size = PAGE_SIZE;	/* can't get any smaller than this */
588 	wire_fault = fault_type == VM_FAULT_WIRE ||
589 	    fault_type == VM_FAULT_WIREMAX;
590 	if (wire_fault)
591 		narrow = TRUE;		/* don't look for neighborhood
592 					 * pages on wire */
593 	else
594 		narrow = FALSE;		/* normal fault */
595 
596 	/*
597 	 * "goto ReFault" means restart the page fault from ground zero.
598 	 */
599 ReFault:
600 
601 	/*
602 	 * lookup and lock the maps
603 	 */
604 
605 	if (uvmfault_lookup(&ufi, FALSE) == FALSE) {
606 		UVMHIST_LOG(maphist, "<- no mapping @ 0x%x", vaddr, 0,0,0);
607 		return (EFAULT);
608 	}
609 	/* locked: maps(read) */
610 
611 #ifdef DIAGNOSTIC
612 	if ((ufi.map->flags & VM_MAP_PAGEABLE) == 0) {
613 		printf("Page fault on non-pageable map:\n");
614 		printf("ufi.map = %p\n", ufi.map);
615 		printf("ufi.orig_map = %p\n", ufi.orig_map);
616 		printf("ufi.orig_rvaddr = 0x%lx\n", (u_long) ufi.orig_rvaddr);
617 		panic("uvm_fault: (ufi.map->flags & VM_MAP_PAGEABLE) == 0");
618 	}
619 #endif
620 
621 	/*
622 	 * check protection
623 	 */
624 
625 	check_prot = fault_type == VM_FAULT_WIREMAX ?
626 	    ufi.entry->max_protection : ufi.entry->protection;
627 	if ((check_prot & access_type) != access_type) {
628 		UVMHIST_LOG(maphist,
629 		    "<- protection failure (prot=0x%x, access=0x%x)",
630 		    ufi.entry->protection, access_type, 0, 0);
631 		uvmfault_unlockmaps(&ufi, FALSE);
632 		return EACCES;
633 	}
634 
635 	/*
636 	 * "enter_prot" is the protection we want to enter the page in at.
637 	 * for certain pages (e.g. copy-on-write pages) this protection can
638 	 * be more strict than ufi.entry->protection.  "wired" means either
639 	 * the entry is wired or we are fault-wiring the pg.
640 	 */
641 
642 	enter_prot = ufi.entry->protection;
643 	wired = VM_MAPENT_ISWIRED(ufi.entry) || wire_fault;
644 	if (wired) {
645 		access_type = enter_prot; /* full access for wired */
646 		cow_now = (check_prot & VM_PROT_WRITE) != 0;
647 	} else {
648 		cow_now = (access_type & VM_PROT_WRITE) != 0;
649 	}
650 
651 	/*
652 	 * handle "needs_copy" case.   if we need to copy the amap we will
653 	 * have to drop our readlock and relock it with a write lock.  (we
654 	 * need a write lock to change anything in a map entry [e.g.
655 	 * needs_copy]).
656 	 */
657 
658 	if (UVM_ET_ISNEEDSCOPY(ufi.entry)) {
659 		KASSERT(fault_type != VM_FAULT_WIREMAX);
660 		if (cow_now || (ufi.entry->object.uvm_obj == NULL)) {
661 			/* need to clear */
662 			UVMHIST_LOG(maphist,
663 			    "  need to clear needs_copy and refault",0,0,0,0);
664 			uvmfault_unlockmaps(&ufi, FALSE);
665 			uvmfault_amapcopy(&ufi);
666 			uvmexp.fltamcopy++;
667 			goto ReFault;
668 
669 		} else {
670 
671 			/*
672 			 * ensure that we pmap_enter page R/O since
673 			 * needs_copy is still true
674 			 */
675 
676 			enter_prot &= ~VM_PROT_WRITE;
677 		}
678 	}
679 
680 	/*
681 	 * identify the players
682 	 */
683 
684 	amap = ufi.entry->aref.ar_amap;		/* top layer */
685 	uobj = ufi.entry->object.uvm_obj;	/* bottom layer */
686 
687 	/*
688 	 * check for a case 0 fault.  if nothing backing the entry then
689 	 * error now.
690 	 */
691 
692 	if (amap == NULL && uobj == NULL) {
693 		uvmfault_unlockmaps(&ufi, FALSE);
694 		UVMHIST_LOG(maphist,"<- no backing store, no overlay",0,0,0,0);
695 		return (EFAULT);
696 	}
697 
698 	/*
699 	 * establish range of interest based on advice from mapper
700 	 * and then clip to fit map entry.   note that we only want
701 	 * to do this the first time through the fault.   if we
702 	 * ReFault we will disable this by setting "narrow" to true.
703 	 */
704 
705 	if (narrow == FALSE) {
706 
707 		/* wide fault (!narrow) */
708 		KASSERT(uvmadvice[ufi.entry->advice].advice ==
709 			 ufi.entry->advice);
710 		nback = MIN(uvmadvice[ufi.entry->advice].nback,
711 			    (ufi.orig_rvaddr - ufi.entry->start) >> PAGE_SHIFT);
712 		startva = ufi.orig_rvaddr - (nback << PAGE_SHIFT);
713 		nforw = MIN(uvmadvice[ufi.entry->advice].nforw,
714 			    ((ufi.entry->end - ufi.orig_rvaddr) >>
715 			     PAGE_SHIFT) - 1);
716 		/*
717 		 * note: "-1" because we don't want to count the
718 		 * faulting page as forw
719 		 */
720 		npages = nback + nforw + 1;
721 		centeridx = nback;
722 
723 		narrow = TRUE;	/* ensure only once per-fault */
724 
725 	} else {
726 
727 		/* narrow fault! */
728 		nback = nforw = 0;
729 		startva = ufi.orig_rvaddr;
730 		npages = 1;
731 		centeridx = 0;
732 
733 	}
734 
735 	/* locked: maps(read) */
736 	UVMHIST_LOG(maphist, "  narrow=%d, back=%d, forw=%d, startva=0x%x",
737 		    narrow, nback, nforw, startva);
738 	UVMHIST_LOG(maphist, "  entry=0x%x, amap=0x%x, obj=0x%x", ufi.entry,
739 		    amap, uobj, 0);
740 
741 	/*
742 	 * if we've got an amap, lock it and extract current anons.
743 	 */
744 
745 	if (amap) {
746 		amap_lock(amap);
747 		anons = anons_store;
748 		amap_lookups(&ufi.entry->aref, startva - ufi.entry->start,
749 		    anons, npages);
750 	} else {
751 		anons = NULL;	/* to be safe */
752 	}
753 
754 	/* locked: maps(read), amap(if there) */
755 
756 	/*
757 	 * for MADV_SEQUENTIAL mappings we want to deactivate the back pages
758 	 * now and then forget about them (for the rest of the fault).
759 	 */
760 
761 	if (ufi.entry->advice == MADV_SEQUENTIAL && nback != 0) {
762 
763 		UVMHIST_LOG(maphist, "  MADV_SEQUENTIAL: flushing backpages",
764 		    0,0,0,0);
765 		/* flush back-page anons? */
766 		if (amap)
767 			uvmfault_anonflush(anons, nback);
768 
769 		/* flush object? */
770 		if (uobj) {
771 			objaddr =
772 			    (startva - ufi.entry->start) + ufi.entry->offset;
773 			simple_lock(&uobj->vmobjlock);
774 			(void) (uobj->pgops->pgo_put)(uobj, objaddr, objaddr +
775 				    (nback << PAGE_SHIFT), PGO_DEACTIVATE);
776 		}
777 
778 		/* now forget about the backpages */
779 		if (amap)
780 			anons += nback;
781 		startva += (nback << PAGE_SHIFT);
782 		npages -= nback;
783 		nback = centeridx = 0;
784 	}
785 
786 	/* locked: maps(read), amap(if there) */
787 
788 	/*
789 	 * map in the backpages and frontpages we found in the amap in hopes
790 	 * of preventing future faults.    we also init the pages[] array as
791 	 * we go.
792 	 */
793 
794 	currva = startva;
795 	shadowed = FALSE;
796 	for (lcv = 0 ; lcv < npages ; lcv++, currva += PAGE_SIZE) {
797 
798 		/*
799 		 * dont play with VAs that are already mapped
800 		 * except for center)
801 		 */
802 		if (lcv != centeridx &&
803 		    pmap_extract(ufi.orig_map->pmap, currva, &pa)) {
804 			pages[lcv] = PGO_DONTCARE;
805 			continue;
806 		}
807 
808 		/*
809 		 * unmapped or center page.   check if any anon at this level.
810 		 */
811 		if (amap == NULL || anons[lcv] == NULL) {
812 			pages[lcv] = NULL;
813 			continue;
814 		}
815 
816 		/*
817 		 * check for present page and map if possible.   re-activate it.
818 		 */
819 
820 		pages[lcv] = PGO_DONTCARE;
821 		if (lcv == centeridx) {		/* save center for later! */
822 			shadowed = TRUE;
823 			continue;
824 		}
825 		anon = anons[lcv];
826 		simple_lock(&anon->an_lock);
827 		/* ignore loaned pages */
828 		if (anon->u.an_page && anon->u.an_page->loan_count == 0 &&
829 		    (anon->u.an_page->flags & PG_BUSY) == 0) {
830 			uvm_lock_pageq();
831 			uvm_pageactivate(anon->u.an_page);
832 			uvm_unlock_pageq();
833 			UVMHIST_LOG(maphist,
834 			    "  MAPPING: n anon: pm=0x%x, va=0x%x, pg=0x%x",
835 			    ufi.orig_map->pmap, currva, anon->u.an_page, 0);
836 			uvmexp.fltnamap++;
837 
838 			/*
839 			 * Since this isn't the page that's actually faulting,
840 			 * ignore pmap_enter() failures; it's not critical
841 			 * that we enter these right now.
842 			 */
843 
844 			(void) pmap_enter(ufi.orig_map->pmap, currva,
845 			    VM_PAGE_TO_PHYS(anon->u.an_page),
846 			    (anon->an_ref > 1) ? (enter_prot & ~VM_PROT_WRITE) :
847 			    enter_prot,
848 			    PMAP_CANFAIL |
849 			     (VM_MAPENT_ISWIRED(ufi.entry) ? PMAP_WIRED : 0));
850 		}
851 		simple_unlock(&anon->an_lock);
852 		pmap_update(ufi.orig_map->pmap);
853 	}
854 
855 	/* locked: maps(read), amap(if there) */
856 	/* (shadowed == TRUE) if there is an anon at the faulting address */
857 	UVMHIST_LOG(maphist, "  shadowed=%d, will_get=%d", shadowed,
858 	    (uobj && shadowed == FALSE),0,0);
859 
860 	/*
861 	 * note that if we are really short of RAM we could sleep in the above
862 	 * call to pmap_enter with everything locked.   bad?
863 	 *
864 	 * XXX Actually, that is bad; pmap_enter() should just fail in that
865 	 * XXX case.  --thorpej
866 	 */
867 
868 	/*
869 	 * if the desired page is not shadowed by the amap and we have a
870 	 * backing object, then we check to see if the backing object would
871 	 * prefer to handle the fault itself (rather than letting us do it
872 	 * with the usual pgo_get hook).  the backing object signals this by
873 	 * providing a pgo_fault routine.
874 	 */
875 
876 	if (uobj && shadowed == FALSE && uobj->pgops->pgo_fault != NULL) {
877 		simple_lock(&uobj->vmobjlock);
878 
879 		/* locked: maps(read), amap (if there), uobj */
880 		error = uobj->pgops->pgo_fault(&ufi, startva, pages, npages,
881 		    centeridx, fault_type, access_type, PGO_LOCKED|PGO_SYNCIO);
882 
883 		/* locked: nothing, pgo_fault has unlocked everything */
884 
885 		if (error == ERESTART)
886 			goto ReFault;		/* try again! */
887 		/*
888 		 * object fault routine responsible for pmap_update().
889 		 */
890 		return error;
891 	}
892 
893 	/*
894 	 * now, if the desired page is not shadowed by the amap and we have
895 	 * a backing object that does not have a special fault routine, then
896 	 * we ask (with pgo_get) the object for resident pages that we care
897 	 * about and attempt to map them in.  we do not let pgo_get block
898 	 * (PGO_LOCKED).
899 	 */
900 
901 	if (uobj && shadowed == FALSE) {
902 		simple_lock(&uobj->vmobjlock);
903 
904 		/* locked (!shadowed): maps(read), amap (if there), uobj */
905 		/*
906 		 * the following call to pgo_get does _not_ change locking state
907 		 */
908 
909 		uvmexp.fltlget++;
910 		gotpages = npages;
911 		(void) uobj->pgops->pgo_get(uobj, ufi.entry->offset +
912 				(startva - ufi.entry->start),
913 				pages, &gotpages, centeridx,
914 				access_type & MASK(ufi.entry),
915 				ufi.entry->advice, PGO_LOCKED);
916 
917 		/*
918 		 * check for pages to map, if we got any
919 		 */
920 
921 		uobjpage = NULL;
922 
923 		if (gotpages) {
924 			currva = startva;
925 			for (lcv = 0; lcv < npages;
926 			     lcv++, currva += PAGE_SIZE) {
927 				struct vm_page *curpg;
928 				boolean_t readonly;
929 
930 				curpg = pages[lcv];
931 				if (curpg == NULL || curpg == PGO_DONTCARE) {
932 					continue;
933 				}
934 
935 				/*
936 				 * if center page is resident and not
937 				 * PG_BUSY|PG_RELEASED then pgo_get
938 				 * made it PG_BUSY for us and gave
939 				 * us a handle to it.   remember this
940 				 * page as "uobjpage." (for later use).
941 				 */
942 
943 				if (lcv == centeridx) {
944 					uobjpage = curpg;
945 					UVMHIST_LOG(maphist, "  got uobjpage "
946 					    "(0x%x) with locked get",
947 					    uobjpage, 0,0,0);
948 					continue;
949 				}
950 
951 				/*
952 				 * calling pgo_get with PGO_LOCKED returns us
953 				 * pages which are neither busy nor released,
954 				 * so we don't need to check for this.
955 				 * we can just directly enter the pages.
956 				 */
957 
958 				uvm_lock_pageq();
959 				uvm_pageactivate(curpg);
960 				uvm_unlock_pageq();
961 				UVMHIST_LOG(maphist,
962 				  "  MAPPING: n obj: pm=0x%x, va=0x%x, pg=0x%x",
963 				  ufi.orig_map->pmap, currva, curpg, 0);
964 				uvmexp.fltnomap++;
965 
966 				/*
967 				 * Since this page isn't the page that's
968 				 * actually faulting, ignore pmap_enter()
969 				 * failures; it's not critical that we
970 				 * enter these right now.
971 				 */
972 				KASSERT((curpg->flags & PG_PAGEOUT) == 0);
973 				KASSERT((curpg->flags & PG_RELEASED) == 0);
974 				readonly = (curpg->flags & PG_RDONLY)
975 				    || (curpg->loan_count > 0);
976 
977 				(void) pmap_enter(ufi.orig_map->pmap, currva,
978 				    VM_PAGE_TO_PHYS(curpg),
979 				    readonly ?
980 				    enter_prot & ~VM_PROT_WRITE :
981 				    enter_prot & MASK(ufi.entry),
982 				    PMAP_CANFAIL |
983 				     (wired ? PMAP_WIRED : 0));
984 
985 				/*
986 				 * NOTE: page can't be PG_WANTED or PG_RELEASED
987 				 * because we've held the lock the whole time
988 				 * we've had the handle.
989 				 */
990 
991 				curpg->flags &= ~(PG_BUSY);
992 				UVM_PAGE_OWN(curpg, NULL);
993 			}
994 			pmap_update(ufi.orig_map->pmap);
995 		}
996 	} else {
997 		uobjpage = NULL;
998 	}
999 
1000 	/* locked (shadowed): maps(read), amap */
1001 	/* locked (!shadowed): maps(read), amap(if there),
1002 		 uobj(if !null), uobjpage(if !null) */
1003 
1004 	/*
1005 	 * note that at this point we are done with any front or back pages.
1006 	 * we are now going to focus on the center page (i.e. the one we've
1007 	 * faulted on).  if we have faulted on the top (anon) layer
1008 	 * [i.e. case 1], then the anon we want is anons[centeridx] (we have
1009 	 * not touched it yet).  if we have faulted on the bottom (uobj)
1010 	 * layer [i.e. case 2] and the page was both present and available,
1011 	 * then we've got a pointer to it as "uobjpage" and we've already
1012 	 * made it BUSY.
1013 	 */
1014 
1015 	/*
1016 	 * there are four possible cases we must address: 1A, 1B, 2A, and 2B
1017 	 */
1018 
1019 	/*
1020 	 * redirect case 2: if we are not shadowed, go to case 2.
1021 	 */
1022 
1023 	if (shadowed == FALSE)
1024 		goto Case2;
1025 
1026 	/* locked: maps(read), amap */
1027 
1028 	/*
1029 	 * handle case 1: fault on an anon in our amap
1030 	 */
1031 
1032 	anon = anons[centeridx];
1033 	UVMHIST_LOG(maphist, "  case 1 fault: anon=0x%x", anon, 0,0,0);
1034 	simple_lock(&anon->an_lock);
1035 
1036 	/* locked: maps(read), amap, anon */
1037 
1038 	/*
1039 	 * no matter if we have case 1A or case 1B we are going to need to
1040 	 * have the anon's memory resident.   ensure that now.
1041 	 */
1042 
1043 	/*
1044 	 * let uvmfault_anonget do the dirty work.
1045 	 * if it fails (!OK) it will unlock everything for us.
1046 	 * if it succeeds, locks are still valid and locked.
1047 	 * also, if it is OK, then the anon's page is on the queues.
1048 	 * if the page is on loan from a uvm_object, then anonget will
1049 	 * lock that object for us if it does not fail.
1050 	 */
1051 
1052 	error = uvmfault_anonget(&ufi, amap, anon);
1053 	switch (error) {
1054 	case 0:
1055 		break;
1056 
1057 	case ERESTART:
1058 		goto ReFault;
1059 
1060 	case EAGAIN:
1061 		tsleep(&lbolt, PVM, "fltagain1", 0);
1062 		goto ReFault;
1063 
1064 	default:
1065 		return error;
1066 	}
1067 
1068 	/*
1069 	 * uobj is non null if the page is on loan from an object (i.e. uobj)
1070 	 */
1071 
1072 	uobj = anon->u.an_page->uobject;	/* locked by anonget if !NULL */
1073 
1074 	/* locked: maps(read), amap, anon, uobj(if one) */
1075 
1076 	/*
1077 	 * special handling for loaned pages
1078 	 */
1079 
1080 	if (anon->u.an_page->loan_count) {
1081 
1082 		if (!cow_now) {
1083 
1084 			/*
1085 			 * for read faults on loaned pages we just cap the
1086 			 * protection at read-only.
1087 			 */
1088 
1089 			enter_prot = enter_prot & ~VM_PROT_WRITE;
1090 
1091 		} else {
1092 			/*
1093 			 * note that we can't allow writes into a loaned page!
1094 			 *
1095 			 * if we have a write fault on a loaned page in an
1096 			 * anon then we need to look at the anon's ref count.
1097 			 * if it is greater than one then we are going to do
1098 			 * a normal copy-on-write fault into a new anon (this
1099 			 * is not a problem).  however, if the reference count
1100 			 * is one (a case where we would normally allow a
1101 			 * write directly to the page) then we need to kill
1102 			 * the loan before we continue.
1103 			 */
1104 
1105 			/* >1 case is already ok */
1106 			if (anon->an_ref == 1) {
1107 
1108 				/* get new un-owned replacement page */
1109 				pg = uvm_pagealloc(NULL, 0, NULL, 0);
1110 				if (pg == NULL) {
1111 					uvmfault_unlockall(&ufi, amap, uobj,
1112 					    anon);
1113 					uvm_wait("flt_noram2");
1114 					goto ReFault;
1115 				}
1116 
1117 				/*
1118 				 * copy data, kill loan, and drop uobj lock
1119 				 * (if any)
1120 				 */
1121 				/* copy old -> new */
1122 				uvm_pagecopy(anon->u.an_page, pg);
1123 
1124 				/* force reload */
1125 				pmap_page_protect(anon->u.an_page,
1126 						  VM_PROT_NONE);
1127 				uvm_lock_pageq();	  /* KILL loan */
1128 
1129 				anon->u.an_page->uanon = NULL;
1130 				/* in case we owned */
1131 				anon->u.an_page->pqflags &= ~PQ_ANON;
1132 
1133 				if (uobj) {
1134 					/* if we were receiver of loan */
1135 					anon->u.an_page->loan_count--;
1136 				} else {
1137 					/*
1138 					 * we were the lender (A->K); need
1139 					 * to remove the page from pageq's.
1140 					 */
1141 					uvm_pagedequeue(anon->u.an_page);
1142 				}
1143 
1144 				uvm_pageactivate(pg);
1145 				uvm_unlock_pageq();
1146 				if (uobj) {
1147 					simple_unlock(&uobj->vmobjlock);
1148 					uobj = NULL;
1149 				}
1150 
1151 				/* install new page in anon */
1152 				anon->u.an_page = pg;
1153 				pg->uanon = anon;
1154 				pg->pqflags |= PQ_ANON;
1155 				pg->flags &= ~(PG_BUSY|PG_FAKE);
1156 				UVM_PAGE_OWN(pg, NULL);
1157 
1158 				/* done! */
1159 			}     /* ref == 1 */
1160 		}       /* write fault */
1161 	}         /* loan count */
1162 
1163 	/*
1164 	 * if we are case 1B then we will need to allocate a new blank
1165 	 * anon to transfer the data into.   note that we have a lock
1166 	 * on anon, so no one can busy or release the page until we are done.
1167 	 * also note that the ref count can't drop to zero here because
1168 	 * it is > 1 and we are only dropping one ref.
1169 	 *
1170 	 * in the (hopefully very rare) case that we are out of RAM we
1171 	 * will unlock, wait for more RAM, and refault.
1172 	 *
1173 	 * if we are out of anon VM we kill the process (XXX: could wait?).
1174 	 */
1175 
1176 	if (cow_now && anon->an_ref > 1) {
1177 
1178 		UVMHIST_LOG(maphist, "  case 1B: COW fault",0,0,0,0);
1179 		uvmexp.flt_acow++;
1180 		oanon = anon;		/* oanon = old, locked anon */
1181 		anon = uvm_analloc();
1182 		if (anon) {
1183 			/* new anon is locked! */
1184 			pg = uvm_pagealloc(NULL, 0, anon, 0);
1185 		}
1186 
1187 		/* check for out of RAM */
1188 		if (anon == NULL || pg == NULL) {
1189 			if (anon) {
1190 				anon->an_ref--;
1191 				simple_unlock(&anon->an_lock);
1192 				uvm_anfree(anon);
1193 			}
1194 			uvmfault_unlockall(&ufi, amap, uobj, oanon);
1195 			if (anon == NULL || uvm_swapisfull()) {
1196 				UVMHIST_LOG(maphist,
1197 				    "<- failed.  out of VM",0,0,0,0);
1198 				uvmexp.fltnoanon++;
1199 				return ENOMEM;
1200 			}
1201 
1202 			uvmexp.fltnoram++;
1203 			uvm_wait("flt_noram3");	/* out of RAM, wait for more */
1204 			goto ReFault;
1205 		}
1206 
1207 		/* got all resources, replace anon with nanon */
1208 		uvm_pagecopy(oanon->u.an_page, pg);
1209 		uvm_lock_pageq();
1210 		uvm_pageactivate(pg);
1211 		pg->flags &= ~(PG_BUSY|PG_FAKE);
1212 		uvm_unlock_pageq();
1213 		UVM_PAGE_OWN(pg, NULL);
1214 		amap_add(&ufi.entry->aref, ufi.orig_rvaddr - ufi.entry->start,
1215 		    anon, 1);
1216 
1217 		/* deref: can not drop to zero here by defn! */
1218 		oanon->an_ref--;
1219 
1220 		/*
1221 		 * note: oanon is still locked, as is the new anon.  we
1222 		 * need to check for this later when we unlock oanon; if
1223 		 * oanon != anon, we'll have to unlock anon, too.
1224 		 */
1225 
1226 	} else {
1227 
1228 		uvmexp.flt_anon++;
1229 		oanon = anon;		/* old, locked anon is same as anon */
1230 		pg = anon->u.an_page;
1231 		if (anon->an_ref > 1)     /* disallow writes to ref > 1 anons */
1232 			enter_prot = enter_prot & ~VM_PROT_WRITE;
1233 
1234 	}
1235 
1236 	/* locked: maps(read), amap, oanon, anon (if different from oanon) */
1237 
1238 	/*
1239 	 * now map the page in.
1240 	 */
1241 
1242 	UVMHIST_LOG(maphist, "  MAPPING: anon: pm=0x%x, va=0x%x, pg=0x%x",
1243 	    ufi.orig_map->pmap, ufi.orig_rvaddr, pg, 0);
1244 	if (pmap_enter(ufi.orig_map->pmap, ufi.orig_rvaddr, VM_PAGE_TO_PHYS(pg),
1245 	    enter_prot, access_type | PMAP_CANFAIL | (wired ? PMAP_WIRED : 0))
1246 	    != 0) {
1247 
1248 		/*
1249 		 * No need to undo what we did; we can simply think of
1250 		 * this as the pmap throwing away the mapping information.
1251 		 *
1252 		 * We do, however, have to go through the ReFault path,
1253 		 * as the map may change while we're asleep.
1254 		 */
1255 
1256 		if (anon != oanon)
1257 			simple_unlock(&anon->an_lock);
1258 		uvmfault_unlockall(&ufi, amap, uobj, oanon);
1259 		if (uvm_swapisfull()) {
1260 			UVMHIST_LOG(maphist,
1261 			    "<- failed.  out of VM",0,0,0,0);
1262 			/* XXX instrumentation */
1263 			return ENOMEM;
1264 		}
1265 		/* XXX instrumentation */
1266 		uvm_wait("flt_pmfail1");
1267 		goto ReFault;
1268 	}
1269 
1270 	/*
1271 	 * ... update the page queues.
1272 	 */
1273 
1274 	uvm_lock_pageq();
1275 	if (wire_fault) {
1276 		uvm_pagewire(pg);
1277 
1278 		/*
1279 		 * since the now-wired page cannot be paged out,
1280 		 * release its swap resources for others to use.
1281 		 * since an anon with no swap cannot be PG_CLEAN,
1282 		 * clear its clean flag now.
1283 		 */
1284 
1285 		pg->flags &= ~(PG_CLEAN);
1286 		uvm_anon_dropswap(anon);
1287 	} else {
1288 		uvm_pageactivate(pg);
1289 	}
1290 	uvm_unlock_pageq();
1291 
1292 	/*
1293 	 * done case 1!  finish up by unlocking everything and returning success
1294 	 */
1295 
1296 	if (anon != oanon)
1297 		simple_unlock(&anon->an_lock);
1298 	uvmfault_unlockall(&ufi, amap, uobj, oanon);
1299 	pmap_update(ufi.orig_map->pmap);
1300 	return 0;
1301 
1302 Case2:
1303 	/*
1304 	 * handle case 2: faulting on backing object or zero fill
1305 	 */
1306 
1307 	/*
1308 	 * locked:
1309 	 * maps(read), amap(if there), uobj(if !null), uobjpage(if !null)
1310 	 */
1311 
1312 	/*
1313 	 * note that uobjpage can not be PGO_DONTCARE at this point.  we now
1314 	 * set uobjpage to PGO_DONTCARE if we are doing a zero fill.  if we
1315 	 * have a backing object, check and see if we are going to promote
1316 	 * the data up to an anon during the fault.
1317 	 */
1318 
1319 	if (uobj == NULL) {
1320 		uobjpage = PGO_DONTCARE;
1321 		promote = TRUE;		/* always need anon here */
1322 	} else {
1323 		KASSERT(uobjpage != PGO_DONTCARE);
1324 		promote = cow_now && UVM_ET_ISCOPYONWRITE(ufi.entry);
1325 	}
1326 	UVMHIST_LOG(maphist, "  case 2 fault: promote=%d, zfill=%d",
1327 	    promote, (uobj == NULL), 0,0);
1328 
1329 	/*
1330 	 * if uobjpage is not null then we do not need to do I/O to get the
1331 	 * uobjpage.
1332 	 *
1333 	 * if uobjpage is null, then we need to unlock and ask the pager to
1334 	 * get the data for us.   once we have the data, we need to reverify
1335 	 * the state the world.   we are currently not holding any resources.
1336 	 */
1337 
1338 	if (uobjpage) {
1339 		/* update rusage counters */
1340 		curproc->p_stats->p_ru.ru_minflt++;
1341 	} else {
1342 		/* update rusage counters */
1343 		curproc->p_stats->p_ru.ru_majflt++;
1344 
1345 		/* locked: maps(read), amap(if there), uobj */
1346 		uvmfault_unlockall(&ufi, amap, NULL, NULL);
1347 		/* locked: uobj */
1348 
1349 		uvmexp.fltget++;
1350 		gotpages = 1;
1351 		uoff = (ufi.orig_rvaddr - ufi.entry->start) + ufi.entry->offset;
1352 		error = uobj->pgops->pgo_get(uobj, uoff, &uobjpage, &gotpages,
1353 		    0, access_type & MASK(ufi.entry), ufi.entry->advice,
1354 		    PGO_SYNCIO);
1355 		/* locked: uobjpage(if no error) */
1356 
1357 		/*
1358 		 * recover from I/O
1359 		 */
1360 
1361 		if (error) {
1362 			if (error == EAGAIN) {
1363 				UVMHIST_LOG(maphist,
1364 				    "  pgo_get says TRY AGAIN!",0,0,0,0);
1365 				tsleep(&lbolt, PVM, "fltagain2", 0);
1366 				goto ReFault;
1367 			}
1368 
1369 			UVMHIST_LOG(maphist, "<- pgo_get failed (code %d)",
1370 			    error, 0,0,0);
1371 			return error;
1372 		}
1373 
1374 		/* locked: uobjpage */
1375 
1376 		uvm_lock_pageq();
1377 		uvm_pageactivate(uobjpage);
1378 		uvm_unlock_pageq();
1379 
1380 		/*
1381 		 * re-verify the state of the world by first trying to relock
1382 		 * the maps.  always relock the object.
1383 		 */
1384 
1385 		locked = uvmfault_relock(&ufi);
1386 		if (locked && amap)
1387 			amap_lock(amap);
1388 		simple_lock(&uobj->vmobjlock);
1389 
1390 		/* locked(locked): maps(read), amap(if !null), uobj, uobjpage */
1391 		/* locked(!locked): uobj, uobjpage */
1392 
1393 		/*
1394 		 * verify that the page has not be released and re-verify
1395 		 * that amap slot is still free.   if there is a problem,
1396 		 * we unlock and clean up.
1397 		 */
1398 
1399 		if ((uobjpage->flags & PG_RELEASED) != 0 ||
1400 		    (locked && amap &&
1401 		    amap_lookup(&ufi.entry->aref,
1402 		      ufi.orig_rvaddr - ufi.entry->start))) {
1403 			if (locked)
1404 				uvmfault_unlockall(&ufi, amap, NULL, NULL);
1405 			locked = FALSE;
1406 		}
1407 
1408 		/*
1409 		 * didn't get the lock?   release the page and retry.
1410 		 */
1411 
1412 		if (locked == FALSE) {
1413 			UVMHIST_LOG(maphist,
1414 			    "  wasn't able to relock after fault: retry",
1415 			    0,0,0,0);
1416 			if (uobjpage->flags & PG_WANTED)
1417 				wakeup(uobjpage);
1418 			if (uobjpage->flags & PG_RELEASED) {
1419 				uvmexp.fltpgrele++;
1420 				uvm_pagefree(uobjpage);
1421 				goto ReFault;
1422 			}
1423 			uobjpage->flags &= ~(PG_BUSY|PG_WANTED);
1424 			UVM_PAGE_OWN(uobjpage, NULL);
1425 			simple_unlock(&uobj->vmobjlock);
1426 			goto ReFault;
1427 		}
1428 
1429 		/*
1430 		 * we have the data in uobjpage which is busy and
1431 		 * not released.  we are holding object lock (so the page
1432 		 * can't be released on us).
1433 		 */
1434 
1435 		/* locked: maps(read), amap(if !null), uobj, uobjpage */
1436 	}
1437 
1438 	/*
1439 	 * locked:
1440 	 * maps(read), amap(if !null), uobj(if !null), uobjpage(if uobj)
1441 	 */
1442 
1443 	/*
1444 	 * notes:
1445 	 *  - at this point uobjpage can not be NULL
1446 	 *  - at this point uobjpage can not be PG_RELEASED (since we checked
1447 	 *  for it above)
1448 	 *  - at this point uobjpage could be PG_WANTED (handle later)
1449 	 */
1450 
1451 	if (promote == FALSE) {
1452 
1453 		/*
1454 		 * we are not promoting.   if the mapping is COW ensure that we
1455 		 * don't give more access than we should (e.g. when doing a read
1456 		 * fault on a COPYONWRITE mapping we want to map the COW page in
1457 		 * R/O even though the entry protection could be R/W).
1458 		 *
1459 		 * set "pg" to the page we want to map in (uobjpage, usually)
1460 		 */
1461 
1462 		/* no anon in this case. */
1463 		anon = NULL;
1464 
1465 		uvmexp.flt_obj++;
1466 		if (UVM_ET_ISCOPYONWRITE(ufi.entry))
1467 			enter_prot &= ~VM_PROT_WRITE;
1468 		pg = uobjpage;		/* map in the actual object */
1469 
1470 		/* assert(uobjpage != PGO_DONTCARE) */
1471 
1472 		/*
1473 		 * we are faulting directly on the page.   be careful
1474 		 * about writing to loaned pages...
1475 		 */
1476 
1477 		if (uobjpage->loan_count) {
1478 			if (!cow_now) {
1479 				/* read fault: cap the protection at readonly */
1480 				/* cap! */
1481 				enter_prot = enter_prot & ~VM_PROT_WRITE;
1482 			} else {
1483 				/* write fault: must break the loan here */
1484 
1485 				pg = uvm_loanbreak(uobjpage);
1486 				if (pg == NULL) {
1487 
1488 					/*
1489 					 * drop ownership of page, it can't
1490 					 * be released
1491 					 */
1492 
1493 					if (uobjpage->flags & PG_WANTED)
1494 						wakeup(uobjpage);
1495 					uobjpage->flags &= ~(PG_BUSY|PG_WANTED);
1496 					UVM_PAGE_OWN(uobjpage, NULL);
1497 
1498 					uvmfault_unlockall(&ufi, amap, uobj,
1499 					  NULL);
1500 					UVMHIST_LOG(maphist,
1501 					  "  out of RAM breaking loan, waiting",
1502 					  0,0,0,0);
1503 					uvmexp.fltnoram++;
1504 					uvm_wait("flt_noram4");
1505 					goto ReFault;
1506 				}
1507 				uobjpage = pg;
1508 			}
1509 		}
1510 	} else {
1511 
1512 		/*
1513 		 * if we are going to promote the data to an anon we
1514 		 * allocate a blank anon here and plug it into our amap.
1515 		 */
1516 #if DIAGNOSTIC
1517 		if (amap == NULL)
1518 			panic("uvm_fault: want to promote data, but no anon");
1519 #endif
1520 
1521 		anon = uvm_analloc();
1522 		if (anon) {
1523 
1524 			/*
1525 			 * The new anon is locked.
1526 			 *
1527 			 * In `Fill in data...' below, if
1528 			 * uobjpage == PGO_DONTCARE, we want
1529 			 * a zero'd, dirty page, so have
1530 			 * uvm_pagealloc() do that for us.
1531 			 */
1532 
1533 			pg = uvm_pagealloc(NULL, 0, anon,
1534 			    (uobjpage == PGO_DONTCARE) ? UVM_PGA_ZERO : 0);
1535 		}
1536 
1537 		/*
1538 		 * out of memory resources?
1539 		 */
1540 
1541 		if (anon == NULL || pg == NULL) {
1542 			if (anon != NULL) {
1543 				anon->an_ref--;
1544 				simple_unlock(&anon->an_lock);
1545 				uvm_anfree(anon);
1546 			}
1547 
1548 			/*
1549 			 * arg!  must unbusy our page and fail or sleep.
1550 			 */
1551 
1552 			if (uobjpage != PGO_DONTCARE) {
1553 				if (uobjpage->flags & PG_WANTED)
1554 					/* still holding object lock */
1555 					wakeup(uobjpage);
1556 
1557 				uobjpage->flags &= ~(PG_BUSY|PG_WANTED);
1558 				UVM_PAGE_OWN(uobjpage, NULL);
1559 			}
1560 
1561 			/* unlock and fail ... */
1562 			uvmfault_unlockall(&ufi, amap, uobj, NULL);
1563 			if (anon == NULL || uvm_swapisfull()) {
1564 				UVMHIST_LOG(maphist, "  promote: out of VM",
1565 				    0,0,0,0);
1566 				uvmexp.fltnoanon++;
1567 				return ENOMEM;
1568 			}
1569 
1570 			UVMHIST_LOG(maphist, "  out of RAM, waiting for more",
1571 			    0,0,0,0);
1572 			uvmexp.fltnoram++;
1573 			uvm_wait("flt_noram5");
1574 			goto ReFault;
1575 		}
1576 
1577 		/*
1578 		 * fill in the data
1579 		 */
1580 
1581 		if (uobjpage != PGO_DONTCARE) {
1582 			uvmexp.flt_prcopy++;
1583 			/* copy page [pg now dirty] */
1584 			uvm_pagecopy(uobjpage, pg);
1585 
1586 			/*
1587 			 * promote to shared amap?  make sure all sharing
1588 			 * procs see it
1589 			 */
1590 
1591 			if ((amap_flags(amap) & AMAP_SHARED) != 0) {
1592 				pmap_page_protect(uobjpage, VM_PROT_NONE);
1593 				/*
1594 				 * XXX: PAGE MIGHT BE WIRED!
1595 				 */
1596 			}
1597 
1598 			/*
1599 			 * dispose of uobjpage.  it can't be PG_RELEASED
1600 			 * since we still hold the object lock.
1601 			 * drop handle to uobj as well.
1602 			 */
1603 
1604 			if (uobjpage->flags & PG_WANTED)
1605 				/* still have the obj lock */
1606 				wakeup(uobjpage);
1607 			uobjpage->flags &= ~(PG_BUSY|PG_WANTED);
1608 			UVM_PAGE_OWN(uobjpage, NULL);
1609 			simple_unlock(&uobj->vmobjlock);
1610 			uobj = NULL;
1611 
1612 			UVMHIST_LOG(maphist,
1613 			    "  promote uobjpage 0x%x to anon/page 0x%x/0x%x",
1614 			    uobjpage, anon, pg, 0);
1615 
1616 		} else {
1617 			uvmexp.flt_przero++;
1618 
1619 			/*
1620 			 * Page is zero'd and marked dirty by uvm_pagealloc()
1621 			 * above.
1622 			 */
1623 
1624 			UVMHIST_LOG(maphist,"  zero fill anon/page 0x%x/0%x",
1625 			    anon, pg, 0, 0);
1626 		}
1627 		amap_add(&ufi.entry->aref, ufi.orig_rvaddr - ufi.entry->start,
1628 		    anon, 0);
1629 	}
1630 
1631 	/*
1632 	 * locked:
1633 	 * maps(read), amap(if !null), uobj(if !null), uobjpage(if uobj),
1634 	 *   anon(if !null), pg(if anon)
1635 	 *
1636 	 * note: pg is either the uobjpage or the new page in the new anon
1637 	 */
1638 
1639 	/*
1640 	 * all resources are present.   we can now map it in and free our
1641 	 * resources.
1642 	 */
1643 
1644 	UVMHIST_LOG(maphist,
1645 	    "  MAPPING: case2: pm=0x%x, va=0x%x, pg=0x%x, promote=%d",
1646 	    ufi.orig_map->pmap, ufi.orig_rvaddr, pg, promote);
1647 	KASSERT((access_type & VM_PROT_WRITE) == 0 ||
1648 		(pg->flags & PG_RDONLY) == 0);
1649 	if (pmap_enter(ufi.orig_map->pmap, ufi.orig_rvaddr, VM_PAGE_TO_PHYS(pg),
1650 	    pg->flags & PG_RDONLY ? enter_prot & ~VM_PROT_WRITE : enter_prot,
1651 	    access_type | PMAP_CANFAIL | (wired ? PMAP_WIRED : 0)) != 0) {
1652 
1653 		/*
1654 		 * No need to undo what we did; we can simply think of
1655 		 * this as the pmap throwing away the mapping information.
1656 		 *
1657 		 * We do, however, have to go through the ReFault path,
1658 		 * as the map may change while we're asleep.
1659 		 */
1660 
1661 		if (pg->flags & PG_WANTED)
1662 			wakeup(pg);
1663 
1664 		/*
1665 		 * note that pg can't be PG_RELEASED since we did not drop
1666 		 * the object lock since the last time we checked.
1667 		 */
1668 
1669 		pg->flags &= ~(PG_BUSY|PG_FAKE|PG_WANTED);
1670 		UVM_PAGE_OWN(pg, NULL);
1671 		uvmfault_unlockall(&ufi, amap, uobj, anon);
1672 		if (uvm_swapisfull()) {
1673 			UVMHIST_LOG(maphist,
1674 			    "<- failed.  out of VM",0,0,0,0);
1675 			/* XXX instrumentation */
1676 			return ENOMEM;
1677 		}
1678 		/* XXX instrumentation */
1679 		uvm_wait("flt_pmfail2");
1680 		goto ReFault;
1681 	}
1682 
1683 	uvm_lock_pageq();
1684 	if (wire_fault) {
1685 		uvm_pagewire(pg);
1686 		if (pg->pqflags & PQ_AOBJ) {
1687 
1688 			/*
1689 			 * since the now-wired page cannot be paged out,
1690 			 * release its swap resources for others to use.
1691 			 * since an aobj page with no swap cannot be PG_CLEAN,
1692 			 * clear its clean flag now.
1693 			 */
1694 
1695 			pg->flags &= ~(PG_CLEAN);
1696 			uao_dropswap(uobj, pg->offset >> PAGE_SHIFT);
1697 		}
1698 	} else {
1699 		uvm_pageactivate(pg);
1700 	}
1701 	uvm_unlock_pageq();
1702 	if (pg->flags & PG_WANTED)
1703 		wakeup(pg);
1704 
1705 	/*
1706 	 * note that pg can't be PG_RELEASED since we did not drop the object
1707 	 * lock since the last time we checked.
1708 	 */
1709 
1710 	pg->flags &= ~(PG_BUSY|PG_FAKE|PG_WANTED);
1711 	UVM_PAGE_OWN(pg, NULL);
1712 	uvmfault_unlockall(&ufi, amap, uobj, anon);
1713 	pmap_update(ufi.orig_map->pmap);
1714 	UVMHIST_LOG(maphist, "<- done (SUCCESS!)",0,0,0,0);
1715 	return 0;
1716 }
1717 
1718 /*
1719  * uvm_fault_wire: wire down a range of virtual addresses in a map.
1720  *
1721  * => map may be read-locked by caller, but MUST NOT be write-locked.
1722  * => if map is read-locked, any operations which may cause map to
1723  *	be write-locked in uvm_fault() must be taken care of by
1724  *	the caller.  See uvm_map_pageable().
1725  */
1726 
1727 int
1728 uvm_fault_wire(map, start, end, fault_type, access_type)
1729 	struct vm_map *map;
1730 	vaddr_t start, end;
1731 	vm_fault_t fault_type;
1732 	vm_prot_t access_type;
1733 {
1734 	vaddr_t va;
1735 	int error;
1736 
1737 	/*
1738 	 * now fault it in a page at a time.   if the fault fails then we have
1739 	 * to undo what we have done.   note that in uvm_fault VM_PROT_NONE
1740 	 * is replaced with the max protection if fault_type is VM_FAULT_WIRE.
1741 	 */
1742 
1743 	/*
1744 	 * XXX work around overflowing a vaddr_t.  this prevents us from
1745 	 * wiring the last page in the address space, though.
1746 	 */
1747 	if (start > end) {
1748 		return EFAULT;
1749 	}
1750 
1751 	for (va = start ; va < end ; va += PAGE_SIZE) {
1752 		error = uvm_fault(map, va, fault_type, access_type);
1753 		if (error) {
1754 			if (va != start) {
1755 				uvm_fault_unwire(map, start, va);
1756 			}
1757 			return error;
1758 		}
1759 	}
1760 	return 0;
1761 }
1762 
1763 /*
1764  * uvm_fault_unwire(): unwire range of virtual space.
1765  */
1766 
1767 void
1768 uvm_fault_unwire(map, start, end)
1769 	struct vm_map *map;
1770 	vaddr_t start, end;
1771 {
1772 	vm_map_lock_read(map);
1773 	uvm_fault_unwire_locked(map, start, end);
1774 	vm_map_unlock_read(map);
1775 }
1776 
1777 /*
1778  * uvm_fault_unwire_locked(): the guts of uvm_fault_unwire().
1779  *
1780  * => map must be at least read-locked.
1781  */
1782 
1783 void
1784 uvm_fault_unwire_locked(map, start, end)
1785 	struct vm_map *map;
1786 	vaddr_t start, end;
1787 {
1788 	struct vm_map_entry *entry;
1789 	pmap_t pmap = vm_map_pmap(map);
1790 	vaddr_t va;
1791 	paddr_t pa;
1792 	struct vm_page *pg;
1793 
1794 	KASSERT((map->flags & VM_MAP_INTRSAFE) == 0);
1795 
1796 	/*
1797 	 * we assume that the area we are unwiring has actually been wired
1798 	 * in the first place.   this means that we should be able to extract
1799 	 * the PAs from the pmap.   we also lock out the page daemon so that
1800 	 * we can call uvm_pageunwire.
1801 	 */
1802 
1803 	uvm_lock_pageq();
1804 
1805 	/*
1806 	 * find the beginning map entry for the region.
1807 	 */
1808 
1809 	KASSERT(start >= vm_map_min(map) && end <= vm_map_max(map));
1810 	if (uvm_map_lookup_entry(map, start, &entry) == FALSE)
1811 		panic("uvm_fault_unwire_locked: address not in map");
1812 
1813 	for (va = start; va < end; va += PAGE_SIZE) {
1814 		if (pmap_extract(pmap, va, &pa) == FALSE)
1815 			continue;
1816 
1817 		/*
1818 		 * find the map entry for the current address.
1819 		 */
1820 
1821 		KASSERT(va >= entry->start);
1822 		while (va >= entry->end) {
1823 			KASSERT(entry->next != &map->header &&
1824 				entry->next->start <= entry->end);
1825 			entry = entry->next;
1826 		}
1827 
1828 		/*
1829 		 * if the entry is no longer wired, tell the pmap.
1830 		 */
1831 
1832 		if (VM_MAPENT_ISWIRED(entry) == 0)
1833 			pmap_unwire(pmap, va);
1834 
1835 		pg = PHYS_TO_VM_PAGE(pa);
1836 		if (pg)
1837 			uvm_pageunwire(pg);
1838 	}
1839 
1840 	uvm_unlock_pageq();
1841 }
1842