xref: /netbsd-src/sys/uvm/uvm_fault.c (revision 220b5c059a84c51ea44107ea8951a57ffaecdc8c)
1 /*	$NetBSD: uvm_fault.c,v 1.71 2001/11/10 07:36:59 lukem 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 1.71 2001/11/10 07:36:59 lukem 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 __P((struct uvm_faultinfo *));
183 static __inline void uvmfault_anonflush __P((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_addr->u_stats.p_ru.ru_minflt++;
308 	else
309 		curproc->p_addr->u_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 				/* remove page from anon */
432 				anon->u.an_page = NULL;
433 
434 				/*
435 				 * remove the swap slot from the anon
436 				 * and mark the anon as having no real slot.
437 				 * don't free the swap slot, thus preventing
438 				 * it from being used again.
439 				 */
440 
441 				uvm_swap_markbad(anon->an_swslot, 1);
442 				anon->an_swslot = SWSLOT_BAD;
443 
444 				/*
445 				 * note: page was never !PG_BUSY, so it
446 				 * can't be mapped and thus no need to
447 				 * pmap_page_protect it...
448 				 */
449 
450 				uvm_lock_pageq();
451 				uvm_pagefree(pg);
452 				uvm_unlock_pageq();
453 
454 				if (locked)
455 					uvmfault_unlockall(ufi, amap, NULL,
456 					    anon);
457 				else
458 					simple_unlock(&anon->an_lock);
459 				UVMHIST_LOG(maphist, "<- ERROR", 0,0,0,0);
460 				return error;
461 			}
462 
463 			/*
464 			 * we've successfully read the page, activate it.
465 			 */
466 
467 			uvm_lock_pageq();
468 			uvm_pageactivate(pg);
469 			uvm_unlock_pageq();
470 			pg->flags &= ~(PG_WANTED|PG_BUSY|PG_FAKE);
471 			UVM_PAGE_OWN(pg, NULL);
472 			if (!locked)
473 				simple_unlock(&anon->an_lock);
474 		}
475 
476 		/*
477 		 * we were not able to relock.   restart fault.
478 		 */
479 
480 		if (!locked) {
481 			UVMHIST_LOG(maphist, "<- REFAULT", 0,0,0,0);
482 			return (ERESTART);
483 		}
484 
485 		/*
486 		 * verify no one has touched the amap and moved the anon on us.
487 		 */
488 
489 		if (ufi != NULL &&
490 		    amap_lookup(&ufi->entry->aref,
491 				ufi->orig_rvaddr - ufi->entry->start) != anon) {
492 
493 			uvmfault_unlockall(ufi, amap, NULL, anon);
494 			UVMHIST_LOG(maphist, "<- REFAULT", 0,0,0,0);
495 			return (ERESTART);
496 		}
497 
498 		/*
499 		 * try it again!
500 		 */
501 
502 		uvmexp.fltanretry++;
503 		continue;
504 	}
505 	/*NOTREACHED*/
506 }
507 
508 /*
509  *   F A U L T   -   m a i n   e n t r y   p o i n t
510  */
511 
512 /*
513  * uvm_fault: page fault handler
514  *
515  * => called from MD code to resolve a page fault
516  * => VM data structures usually should be unlocked.   however, it is
517  *	possible to call here with the main map locked if the caller
518  *	gets a write lock, sets it recusive, and then calls us (c.f.
519  *	uvm_map_pageable).   this should be avoided because it keeps
520  *	the map locked off during I/O.
521  * => MUST NEVER BE CALLED IN INTERRUPT CONTEXT
522  */
523 
524 #define MASK(entry)     (UVM_ET_ISCOPYONWRITE(entry) ? \
525 			 ~VM_PROT_WRITE : VM_PROT_ALL)
526 
527 int
528 uvm_fault(orig_map, vaddr, fault_type, access_type)
529 	struct vm_map *orig_map;
530 	vaddr_t vaddr;
531 	vm_fault_t fault_type;
532 	vm_prot_t access_type;
533 {
534 	struct uvm_faultinfo ufi;
535 	vm_prot_t enter_prot;
536 	boolean_t wired, narrow, promote, locked, shadowed;
537 	int npages, nback, nforw, centeridx, error, lcv, gotpages;
538 	vaddr_t startva, objaddr, currva, offset, uoff;
539 	paddr_t pa;
540 	struct vm_amap *amap;
541 	struct uvm_object *uobj;
542 	struct vm_anon *anons_store[UVM_MAXRANGE], **anons, *anon, *oanon;
543 	struct vm_page *pages[UVM_MAXRANGE], *pg, *uobjpage;
544 	UVMHIST_FUNC("uvm_fault"); UVMHIST_CALLED(maphist);
545 
546 	UVMHIST_LOG(maphist, "(map=0x%x, vaddr=0x%x, ft=%d, at=%d)",
547 	      orig_map, vaddr, fault_type, access_type);
548 
549 	anon = NULL;
550 	pg = NULL;
551 
552 	uvmexp.faults++;	/* XXX: locking? */
553 
554 	/*
555 	 * init the IN parameters in the ufi
556 	 */
557 
558 	ufi.orig_map = orig_map;
559 	ufi.orig_rvaddr = trunc_page(vaddr);
560 	ufi.orig_size = PAGE_SIZE;	/* can't get any smaller than this */
561 	if (fault_type == VM_FAULT_WIRE)
562 		narrow = TRUE;		/* don't look for neighborhood
563 					 * pages on wire */
564 	else
565 		narrow = FALSE;		/* normal fault */
566 
567 	/*
568 	 * "goto ReFault" means restart the page fault from ground zero.
569 	 */
570 ReFault:
571 
572 	/*
573 	 * lookup and lock the maps
574 	 */
575 
576 	if (uvmfault_lookup(&ufi, FALSE) == FALSE) {
577 		UVMHIST_LOG(maphist, "<- no mapping @ 0x%x", vaddr, 0,0,0);
578 		return (EFAULT);
579 	}
580 	/* locked: maps(read) */
581 
582 #ifdef DIAGNOSTIC
583 	if ((ufi.map->flags & VM_MAP_PAGEABLE) == 0) {
584 		printf("Page fault on non-pageable map:\n");
585 		printf("ufi.map = %p\n", ufi.map);
586 		printf("ufi.orig_map = %p\n", ufi.orig_map);
587 		printf("ufi.orig_rvaddr = 0x%lx\n", (u_long) ufi.orig_rvaddr);
588 		panic("uvm_fault: (ufi.map->flags & VM_MAP_PAGEABLE) == 0");
589 	}
590 #endif
591 
592 	/*
593 	 * check protection
594 	 */
595 
596 	if ((ufi.entry->protection & access_type) != access_type) {
597 		UVMHIST_LOG(maphist,
598 		    "<- protection failure (prot=0x%x, access=0x%x)",
599 		    ufi.entry->protection, access_type, 0, 0);
600 		uvmfault_unlockmaps(&ufi, FALSE);
601 		return EACCES;
602 	}
603 
604 	/*
605 	 * "enter_prot" is the protection we want to enter the page in at.
606 	 * for certain pages (e.g. copy-on-write pages) this protection can
607 	 * be more strict than ufi.entry->protection.  "wired" means either
608 	 * the entry is wired or we are fault-wiring the pg.
609 	 */
610 
611 	enter_prot = ufi.entry->protection;
612 	wired = VM_MAPENT_ISWIRED(ufi.entry) || (fault_type == VM_FAULT_WIRE);
613 	if (wired)
614 		access_type = enter_prot; /* full access for wired */
615 
616 	/*
617 	 * handle "needs_copy" case.   if we need to copy the amap we will
618 	 * have to drop our readlock and relock it with a write lock.  (we
619 	 * need a write lock to change anything in a map entry [e.g.
620 	 * needs_copy]).
621 	 */
622 
623 	if (UVM_ET_ISNEEDSCOPY(ufi.entry)) {
624 		if ((access_type & VM_PROT_WRITE) ||
625 		    (ufi.entry->object.uvm_obj == NULL)) {
626 			/* need to clear */
627 			UVMHIST_LOG(maphist,
628 			    "  need to clear needs_copy and refault",0,0,0,0);
629 			uvmfault_unlockmaps(&ufi, FALSE);
630 			uvmfault_amapcopy(&ufi);
631 			uvmexp.fltamcopy++;
632 			goto ReFault;
633 
634 		} else {
635 
636 			/*
637 			 * ensure that we pmap_enter page R/O since
638 			 * needs_copy is still true
639 			 */
640 			enter_prot &= ~VM_PROT_WRITE;
641 
642 		}
643 	}
644 
645 	/*
646 	 * identify the players
647 	 */
648 
649 	amap = ufi.entry->aref.ar_amap;		/* top layer */
650 	uobj = ufi.entry->object.uvm_obj;	/* bottom layer */
651 
652 	/*
653 	 * check for a case 0 fault.  if nothing backing the entry then
654 	 * error now.
655 	 */
656 
657 	if (amap == NULL && uobj == NULL) {
658 		uvmfault_unlockmaps(&ufi, FALSE);
659 		UVMHIST_LOG(maphist,"<- no backing store, no overlay",0,0,0,0);
660 		return (EFAULT);
661 	}
662 
663 	/*
664 	 * establish range of interest based on advice from mapper
665 	 * and then clip to fit map entry.   note that we only want
666 	 * to do this the first time through the fault.   if we
667 	 * ReFault we will disable this by setting "narrow" to true.
668 	 */
669 
670 	if (narrow == FALSE) {
671 
672 		/* wide fault (!narrow) */
673 		KASSERT(uvmadvice[ufi.entry->advice].advice ==
674 			 ufi.entry->advice);
675 		nback = MIN(uvmadvice[ufi.entry->advice].nback,
676 			    (ufi.orig_rvaddr - ufi.entry->start) >> PAGE_SHIFT);
677 		startva = ufi.orig_rvaddr - (nback << PAGE_SHIFT);
678 		nforw = MIN(uvmadvice[ufi.entry->advice].nforw,
679 			    ((ufi.entry->end - ufi.orig_rvaddr) >>
680 			     PAGE_SHIFT) - 1);
681 		/*
682 		 * note: "-1" because we don't want to count the
683 		 * faulting page as forw
684 		 */
685 		npages = nback + nforw + 1;
686 		centeridx = nback;
687 
688 		narrow = TRUE;	/* ensure only once per-fault */
689 
690 	} else {
691 
692 		/* narrow fault! */
693 		nback = nforw = 0;
694 		startva = ufi.orig_rvaddr;
695 		npages = 1;
696 		centeridx = 0;
697 
698 	}
699 
700 	/* locked: maps(read) */
701 	UVMHIST_LOG(maphist, "  narrow=%d, back=%d, forw=%d, startva=0x%x",
702 		    narrow, nback, nforw, startva);
703 	UVMHIST_LOG(maphist, "  entry=0x%x, amap=0x%x, obj=0x%x", ufi.entry,
704 		    amap, uobj, 0);
705 
706 	/*
707 	 * if we've got an amap, lock it and extract current anons.
708 	 */
709 
710 	if (amap) {
711 		amap_lock(amap);
712 		anons = anons_store;
713 		amap_lookups(&ufi.entry->aref, startva - ufi.entry->start,
714 		    anons, npages);
715 	} else {
716 		anons = NULL;	/* to be safe */
717 	}
718 
719 	/* locked: maps(read), amap(if there) */
720 
721 	/*
722 	 * for MADV_SEQUENTIAL mappings we want to deactivate the back pages
723 	 * now and then forget about them (for the rest of the fault).
724 	 */
725 
726 	if (ufi.entry->advice == MADV_SEQUENTIAL && nback != 0) {
727 
728 		UVMHIST_LOG(maphist, "  MADV_SEQUENTIAL: flushing backpages",
729 		    0,0,0,0);
730 		/* flush back-page anons? */
731 		if (amap)
732 			uvmfault_anonflush(anons, nback);
733 
734 		/* flush object? */
735 		if (uobj) {
736 			objaddr =
737 			    (startva - ufi.entry->start) + ufi.entry->offset;
738 			simple_lock(&uobj->vmobjlock);
739 			(void) (uobj->pgops->pgo_put)(uobj, objaddr, objaddr +
740 				    (nback << PAGE_SHIFT), PGO_DEACTIVATE);
741 		}
742 
743 		/* now forget about the backpages */
744 		if (amap)
745 			anons += nback;
746 		startva += (nback << PAGE_SHIFT);
747 		npages -= nback;
748 		nback = centeridx = 0;
749 	}
750 
751 	/* locked: maps(read), amap(if there) */
752 
753 	/*
754 	 * map in the backpages and frontpages we found in the amap in hopes
755 	 * of preventing future faults.    we also init the pages[] array as
756 	 * we go.
757 	 */
758 
759 	currva = startva;
760 	shadowed = FALSE;
761 	for (lcv = 0 ; lcv < npages ; lcv++, currva += PAGE_SIZE) {
762 
763 		/*
764 		 * dont play with VAs that are already mapped
765 		 * except for center)
766 		 */
767 		if (lcv != centeridx &&
768 		    pmap_extract(ufi.orig_map->pmap, currva, &pa)) {
769 			pages[lcv] = PGO_DONTCARE;
770 			continue;
771 		}
772 
773 		/*
774 		 * unmapped or center page.   check if any anon at this level.
775 		 */
776 		if (amap == NULL || anons[lcv] == NULL) {
777 			pages[lcv] = NULL;
778 			continue;
779 		}
780 
781 		/*
782 		 * check for present page and map if possible.   re-activate it.
783 		 */
784 
785 		pages[lcv] = PGO_DONTCARE;
786 		if (lcv == centeridx) {		/* save center for later! */
787 			shadowed = TRUE;
788 			continue;
789 		}
790 		anon = anons[lcv];
791 		simple_lock(&anon->an_lock);
792 		/* ignore loaned pages */
793 		if (anon->u.an_page && anon->u.an_page->loan_count == 0 &&
794 		    (anon->u.an_page->flags & PG_BUSY) == 0) {
795 			uvm_lock_pageq();
796 			uvm_pageactivate(anon->u.an_page);
797 			uvm_unlock_pageq();
798 			UVMHIST_LOG(maphist,
799 			    "  MAPPING: n anon: pm=0x%x, va=0x%x, pg=0x%x",
800 			    ufi.orig_map->pmap, currva, anon->u.an_page, 0);
801 			uvmexp.fltnamap++;
802 
803 			/*
804 			 * Since this isn't the page that's actually faulting,
805 			 * ignore pmap_enter() failures; it's not critical
806 			 * that we enter these right now.
807 			 */
808 
809 			(void) pmap_enter(ufi.orig_map->pmap, currva,
810 			    VM_PAGE_TO_PHYS(anon->u.an_page),
811 			    (anon->an_ref > 1) ? (enter_prot & ~VM_PROT_WRITE) :
812 			    enter_prot,
813 			    PMAP_CANFAIL |
814 			     (VM_MAPENT_ISWIRED(ufi.entry) ? PMAP_WIRED : 0));
815 		}
816 		simple_unlock(&anon->an_lock);
817 		pmap_update(ufi.orig_map->pmap);
818 	}
819 
820 	/* locked: maps(read), amap(if there) */
821 	/* (shadowed == TRUE) if there is an anon at the faulting address */
822 	UVMHIST_LOG(maphist, "  shadowed=%d, will_get=%d", shadowed,
823 	    (uobj && shadowed == FALSE),0,0);
824 
825 	/*
826 	 * note that if we are really short of RAM we could sleep in the above
827 	 * call to pmap_enter with everything locked.   bad?
828 	 *
829 	 * XXX Actually, that is bad; pmap_enter() should just fail in that
830 	 * XXX case.  --thorpej
831 	 */
832 
833 	/*
834 	 * if the desired page is not shadowed by the amap and we have a
835 	 * backing object, then we check to see if the backing object would
836 	 * prefer to handle the fault itself (rather than letting us do it
837 	 * with the usual pgo_get hook).  the backing object signals this by
838 	 * providing a pgo_fault routine.
839 	 */
840 
841 	if (uobj && shadowed == FALSE && uobj->pgops->pgo_fault != NULL) {
842 		simple_lock(&uobj->vmobjlock);
843 
844 		/* locked: maps(read), amap (if there), uobj */
845 		error = uobj->pgops->pgo_fault(&ufi, startva, pages, npages,
846 		    centeridx, fault_type, access_type, PGO_LOCKED|PGO_SYNCIO);
847 
848 		/* locked: nothing, pgo_fault has unlocked everything */
849 
850 		if (error == ERESTART)
851 			goto ReFault;		/* try again! */
852 		/*
853 		 * object fault routine responsible for pmap_update().
854 		 */
855 		return error;
856 	}
857 
858 	/*
859 	 * now, if the desired page is not shadowed by the amap and we have
860 	 * a backing object that does not have a special fault routine, then
861 	 * we ask (with pgo_get) the object for resident pages that we care
862 	 * about and attempt to map them in.  we do not let pgo_get block
863 	 * (PGO_LOCKED).
864 	 */
865 
866 	if (uobj && shadowed == FALSE) {
867 		simple_lock(&uobj->vmobjlock);
868 
869 		/* locked (!shadowed): maps(read), amap (if there), uobj */
870 		/*
871 		 * the following call to pgo_get does _not_ change locking state
872 		 */
873 
874 		uvmexp.fltlget++;
875 		gotpages = npages;
876 		(void) uobj->pgops->pgo_get(uobj, ufi.entry->offset +
877 				(startva - ufi.entry->start),
878 				pages, &gotpages, centeridx,
879 				access_type & MASK(ufi.entry),
880 				ufi.entry->advice, PGO_LOCKED);
881 
882 		/*
883 		 * check for pages to map, if we got any
884 		 */
885 
886 		uobjpage = NULL;
887 
888 		if (gotpages) {
889 			currva = startva;
890 			for (lcv = 0; lcv < npages;
891 			     lcv++, currva += PAGE_SIZE) {
892 				if (pages[lcv] == NULL ||
893 				    pages[lcv] == PGO_DONTCARE) {
894 					continue;
895 				}
896 
897 				/*
898 				 * if center page is resident and not
899 				 * PG_BUSY|PG_RELEASED then pgo_get
900 				 * made it PG_BUSY for us and gave
901 				 * us a handle to it.   remember this
902 				 * page as "uobjpage." (for later use).
903 				 */
904 
905 				if (lcv == centeridx) {
906 					uobjpage = pages[lcv];
907 					UVMHIST_LOG(maphist, "  got uobjpage "
908 					    "(0x%x) with locked get",
909 					    uobjpage, 0,0,0);
910 					continue;
911 				}
912 
913 				/*
914 				 * calling pgo_get with PGO_LOCKED returns us
915 				 * pages which are neither busy nor released,
916 				 * so we don't need to check for this.
917 				 * we can just directly enter the pages.
918 				 */
919 
920 				uvm_lock_pageq();
921 				uvm_pageactivate(pages[lcv]);
922 				uvm_unlock_pageq();
923 				UVMHIST_LOG(maphist,
924 				  "  MAPPING: n obj: pm=0x%x, va=0x%x, pg=0x%x",
925 				  ufi.orig_map->pmap, currva, pages[lcv], 0);
926 				uvmexp.fltnomap++;
927 
928 				/*
929 				 * Since this page isn't the page that's
930 				 * actually fauling, ignore pmap_enter()
931 				 * failures; it's not critical that we
932 				 * enter these right now.
933 				 */
934 
935 				(void) pmap_enter(ufi.orig_map->pmap, currva,
936 				    VM_PAGE_TO_PHYS(pages[lcv]),
937 				    pages[lcv]->flags & PG_RDONLY ?
938 				    VM_PROT_READ : enter_prot & MASK(ufi.entry),
939 				    PMAP_CANFAIL |
940 				     (wired ? PMAP_WIRED : 0));
941 
942 				/*
943 				 * NOTE: page can't be PG_WANTED or PG_RELEASED
944 				 * because we've held the lock the whole time
945 				 * we've had the handle.
946 				 */
947 
948 				pages[lcv]->flags &= ~(PG_BUSY);
949 				UVM_PAGE_OWN(pages[lcv], NULL);
950 			}
951 			pmap_update(ufi.orig_map->pmap);
952 		}
953 	} else {
954 		uobjpage = NULL;
955 	}
956 
957 	/* locked (shadowed): maps(read), amap */
958 	/* locked (!shadowed): maps(read), amap(if there),
959 		 uobj(if !null), uobjpage(if !null) */
960 
961 	/*
962 	 * note that at this point we are done with any front or back pages.
963 	 * we are now going to focus on the center page (i.e. the one we've
964 	 * faulted on).  if we have faulted on the top (anon) layer
965 	 * [i.e. case 1], then the anon we want is anons[centeridx] (we have
966 	 * not touched it yet).  if we have faulted on the bottom (uobj)
967 	 * layer [i.e. case 2] and the page was both present and available,
968 	 * then we've got a pointer to it as "uobjpage" and we've already
969 	 * made it BUSY.
970 	 */
971 
972 	/*
973 	 * there are four possible cases we must address: 1A, 1B, 2A, and 2B
974 	 */
975 
976 	/*
977 	 * redirect case 2: if we are not shadowed, go to case 2.
978 	 */
979 
980 	if (shadowed == FALSE)
981 		goto Case2;
982 
983 	/* locked: maps(read), amap */
984 
985 	/*
986 	 * handle case 1: fault on an anon in our amap
987 	 */
988 
989 	anon = anons[centeridx];
990 	UVMHIST_LOG(maphist, "  case 1 fault: anon=0x%x", anon, 0,0,0);
991 	simple_lock(&anon->an_lock);
992 
993 	/* locked: maps(read), amap, anon */
994 
995 	/*
996 	 * no matter if we have case 1A or case 1B we are going to need to
997 	 * have the anon's memory resident.   ensure that now.
998 	 */
999 
1000 	/*
1001 	 * let uvmfault_anonget do the dirty work.
1002 	 * if it fails (!OK) it will unlock everything for us.
1003 	 * if it succeeds, locks are still valid and locked.
1004 	 * also, if it is OK, then the anon's page is on the queues.
1005 	 * if the page is on loan from a uvm_object, then anonget will
1006 	 * lock that object for us if it does not fail.
1007 	 */
1008 
1009 	error = uvmfault_anonget(&ufi, amap, anon);
1010 	switch (error) {
1011 	case 0:
1012 		break;
1013 
1014 	case ERESTART:
1015 		goto ReFault;
1016 
1017 	case EAGAIN:
1018 		tsleep(&lbolt, PVM, "fltagain1", 0);
1019 		goto ReFault;
1020 
1021 	default:
1022 		return error;
1023 	}
1024 
1025 	/*
1026 	 * uobj is non null if the page is on loan from an object (i.e. uobj)
1027 	 */
1028 
1029 	uobj = anon->u.an_page->uobject;	/* locked by anonget if !NULL */
1030 
1031 	/* locked: maps(read), amap, anon, uobj(if one) */
1032 
1033 	/*
1034 	 * special handling for loaned pages
1035 	 */
1036 
1037 	if (anon->u.an_page->loan_count) {
1038 
1039 		if ((access_type & VM_PROT_WRITE) == 0) {
1040 
1041 			/*
1042 			 * for read faults on loaned pages we just cap the
1043 			 * protection at read-only.
1044 			 */
1045 
1046 			enter_prot = enter_prot & ~VM_PROT_WRITE;
1047 
1048 		} else {
1049 			/*
1050 			 * note that we can't allow writes into a loaned page!
1051 			 *
1052 			 * if we have a write fault on a loaned page in an
1053 			 * anon then we need to look at the anon's ref count.
1054 			 * if it is greater than one then we are going to do
1055 			 * a normal copy-on-write fault into a new anon (this
1056 			 * is not a problem).  however, if the reference count
1057 			 * is one (a case where we would normally allow a
1058 			 * write directly to the page) then we need to kill
1059 			 * the loan before we continue.
1060 			 */
1061 
1062 			/* >1 case is already ok */
1063 			if (anon->an_ref == 1) {
1064 
1065 				/* get new un-owned replacement page */
1066 				pg = uvm_pagealloc(NULL, 0, NULL, 0);
1067 				if (pg == NULL) {
1068 					uvmfault_unlockall(&ufi, amap, uobj,
1069 					    anon);
1070 					uvm_wait("flt_noram2");
1071 					goto ReFault;
1072 				}
1073 
1074 				/*
1075 				 * copy data, kill loan, and drop uobj lock
1076 				 * (if any)
1077 				 */
1078 				/* copy old -> new */
1079 				uvm_pagecopy(anon->u.an_page, pg);
1080 
1081 				/* force reload */
1082 				pmap_page_protect(anon->u.an_page,
1083 						  VM_PROT_NONE);
1084 				uvm_lock_pageq();	  /* KILL loan */
1085 				if (uobj)
1086 					/* if we were loaning */
1087 					anon->u.an_page->loan_count--;
1088 				anon->u.an_page->uanon = NULL;
1089 				/* in case we owned */
1090 				anon->u.an_page->pqflags &= ~PQ_ANON;
1091 				uvm_unlock_pageq();
1092 				if (uobj) {
1093 					simple_unlock(&uobj->vmobjlock);
1094 					uobj = NULL;
1095 				}
1096 
1097 				/* install new page in anon */
1098 				anon->u.an_page = pg;
1099 				pg->uanon = anon;
1100 				pg->pqflags |= PQ_ANON;
1101 				pg->flags &= ~(PG_BUSY|PG_FAKE);
1102 				UVM_PAGE_OWN(pg, NULL);
1103 
1104 				/* done! */
1105 			}     /* ref == 1 */
1106 		}       /* write fault */
1107 	}         /* loan count */
1108 
1109 	/*
1110 	 * if we are case 1B then we will need to allocate a new blank
1111 	 * anon to transfer the data into.   note that we have a lock
1112 	 * on anon, so no one can busy or release the page until we are done.
1113 	 * also note that the ref count can't drop to zero here because
1114 	 * it is > 1 and we are only dropping one ref.
1115 	 *
1116 	 * in the (hopefully very rare) case that we are out of RAM we
1117 	 * will unlock, wait for more RAM, and refault.
1118 	 *
1119 	 * if we are out of anon VM we kill the process (XXX: could wait?).
1120 	 */
1121 
1122 	if ((access_type & VM_PROT_WRITE) != 0 && anon->an_ref > 1) {
1123 
1124 		UVMHIST_LOG(maphist, "  case 1B: COW fault",0,0,0,0);
1125 		uvmexp.flt_acow++;
1126 		oanon = anon;		/* oanon = old, locked anon */
1127 		anon = uvm_analloc();
1128 		if (anon) {
1129 			/* new anon is locked! */
1130 			pg = uvm_pagealloc(NULL, 0, anon, 0);
1131 		}
1132 
1133 		/* check for out of RAM */
1134 		if (anon == NULL || pg == NULL) {
1135 			if (anon) {
1136 				anon->an_ref--;
1137 				simple_unlock(&anon->an_lock);
1138 				uvm_anfree(anon);
1139 			}
1140 			uvmfault_unlockall(&ufi, amap, uobj, oanon);
1141 			KASSERT(uvmexp.swpgonly <= uvmexp.swpages);
1142 			if (anon == NULL || uvmexp.swpgonly == uvmexp.swpages) {
1143 				UVMHIST_LOG(maphist,
1144 				    "<- failed.  out of VM",0,0,0,0);
1145 				uvmexp.fltnoanon++;
1146 				return ENOMEM;
1147 			}
1148 
1149 			uvmexp.fltnoram++;
1150 			uvm_wait("flt_noram3");	/* out of RAM, wait for more */
1151 			goto ReFault;
1152 		}
1153 
1154 		/* got all resources, replace anon with nanon */
1155 		uvm_pagecopy(oanon->u.an_page, pg);
1156 		uvm_pageactivate(pg);
1157 		pg->flags &= ~(PG_BUSY|PG_FAKE);
1158 		UVM_PAGE_OWN(pg, NULL);
1159 		amap_add(&ufi.entry->aref, ufi.orig_rvaddr - ufi.entry->start,
1160 		    anon, 1);
1161 
1162 		/* deref: can not drop to zero here by defn! */
1163 		oanon->an_ref--;
1164 
1165 		/*
1166 		 * note: oanon is still locked, as is the new anon.  we
1167 		 * need to check for this later when we unlock oanon; if
1168 		 * oanon != anon, we'll have to unlock anon, too.
1169 		 */
1170 
1171 	} else {
1172 
1173 		uvmexp.flt_anon++;
1174 		oanon = anon;		/* old, locked anon is same as anon */
1175 		pg = anon->u.an_page;
1176 		if (anon->an_ref > 1)     /* disallow writes to ref > 1 anons */
1177 			enter_prot = enter_prot & ~VM_PROT_WRITE;
1178 
1179 	}
1180 
1181 	/* locked: maps(read), amap, oanon, anon (if different from oanon) */
1182 
1183 	/*
1184 	 * now map the page in.
1185 	 */
1186 
1187 	UVMHIST_LOG(maphist, "  MAPPING: anon: pm=0x%x, va=0x%x, pg=0x%x",
1188 	    ufi.orig_map->pmap, ufi.orig_rvaddr, pg, 0);
1189 	if (pmap_enter(ufi.orig_map->pmap, ufi.orig_rvaddr, VM_PAGE_TO_PHYS(pg),
1190 	    enter_prot, access_type | PMAP_CANFAIL | (wired ? PMAP_WIRED : 0))
1191 	    != 0) {
1192 
1193 		/*
1194 		 * No need to undo what we did; we can simply think of
1195 		 * this as the pmap throwing away the mapping information.
1196 		 *
1197 		 * We do, however, have to go through the ReFault path,
1198 		 * as the map may change while we're asleep.
1199 		 */
1200 
1201 		if (anon != oanon)
1202 			simple_unlock(&anon->an_lock);
1203 		uvmfault_unlockall(&ufi, amap, uobj, oanon);
1204 		KASSERT(uvmexp.swpgonly <= uvmexp.swpages);
1205 		if (uvmexp.swpgonly == uvmexp.swpages) {
1206 			UVMHIST_LOG(maphist,
1207 			    "<- failed.  out of VM",0,0,0,0);
1208 			/* XXX instrumentation */
1209 			return ENOMEM;
1210 		}
1211 		/* XXX instrumentation */
1212 		uvm_wait("flt_pmfail1");
1213 		goto ReFault;
1214 	}
1215 
1216 	/*
1217 	 * ... update the page queues.
1218 	 */
1219 
1220 	uvm_lock_pageq();
1221 	if (fault_type == VM_FAULT_WIRE) {
1222 		uvm_pagewire(pg);
1223 
1224 		/*
1225 		 * since the now-wired page cannot be paged out,
1226 		 * release its swap resources for others to use.
1227 		 * since an anon with no swap cannot be PG_CLEAN,
1228 		 * clear its clean flag now.
1229 		 */
1230 
1231 		pg->flags &= ~(PG_CLEAN);
1232 		uvm_anon_dropswap(anon);
1233 	} else {
1234 		uvm_pageactivate(pg);
1235 	}
1236 	uvm_unlock_pageq();
1237 
1238 	/*
1239 	 * done case 1!  finish up by unlocking everything and returning success
1240 	 */
1241 
1242 	if (anon != oanon)
1243 		simple_unlock(&anon->an_lock);
1244 	uvmfault_unlockall(&ufi, amap, uobj, oanon);
1245 	pmap_update(ufi.orig_map->pmap);
1246 	return 0;
1247 
1248 Case2:
1249 	/*
1250 	 * handle case 2: faulting on backing object or zero fill
1251 	 */
1252 
1253 	/*
1254 	 * locked:
1255 	 * maps(read), amap(if there), uobj(if !null), uobjpage(if !null)
1256 	 */
1257 
1258 	/*
1259 	 * note that uobjpage can not be PGO_DONTCARE at this point.  we now
1260 	 * set uobjpage to PGO_DONTCARE if we are doing a zero fill.  if we
1261 	 * have a backing object, check and see if we are going to promote
1262 	 * the data up to an anon during the fault.
1263 	 */
1264 
1265 	if (uobj == NULL) {
1266 		uobjpage = PGO_DONTCARE;
1267 		promote = TRUE;		/* always need anon here */
1268 	} else {
1269 		KASSERT(uobjpage != PGO_DONTCARE);
1270 		promote = (access_type & VM_PROT_WRITE) &&
1271 		     UVM_ET_ISCOPYONWRITE(ufi.entry);
1272 	}
1273 	UVMHIST_LOG(maphist, "  case 2 fault: promote=%d, zfill=%d",
1274 	    promote, (uobj == NULL), 0,0);
1275 
1276 	/*
1277 	 * if uobjpage is not null then we do not need to do I/O to get the
1278 	 * uobjpage.
1279 	 *
1280 	 * if uobjpage is null, then we need to unlock and ask the pager to
1281 	 * get the data for us.   once we have the data, we need to reverify
1282 	 * the state the world.   we are currently not holding any resources.
1283 	 */
1284 
1285 	if (uobjpage) {
1286 		/* update rusage counters */
1287 		curproc->p_addr->u_stats.p_ru.ru_minflt++;
1288 	} else {
1289 		/* update rusage counters */
1290 		curproc->p_addr->u_stats.p_ru.ru_majflt++;
1291 
1292 		/* locked: maps(read), amap(if there), uobj */
1293 		uvmfault_unlockall(&ufi, amap, NULL, NULL);
1294 		/* locked: uobj */
1295 
1296 		uvmexp.fltget++;
1297 		gotpages = 1;
1298 		uoff = (ufi.orig_rvaddr - ufi.entry->start) + ufi.entry->offset;
1299 		error = uobj->pgops->pgo_get(uobj, uoff, &uobjpage, &gotpages,
1300 		    0, access_type & MASK(ufi.entry), ufi.entry->advice,
1301 		    PGO_SYNCIO);
1302 		/* locked: uobjpage(if no error) */
1303 
1304 		/*
1305 		 * recover from I/O
1306 		 */
1307 
1308 		if (error) {
1309 			if (error == EAGAIN) {
1310 				UVMHIST_LOG(maphist,
1311 				    "  pgo_get says TRY AGAIN!",0,0,0,0);
1312 				tsleep(&lbolt, PVM, "fltagain2", 0);
1313 				goto ReFault;
1314 			}
1315 
1316 			UVMHIST_LOG(maphist, "<- pgo_get failed (code %d)",
1317 			    error, 0,0,0);
1318 			return error;
1319 		}
1320 
1321 		/* locked: uobjpage */
1322 
1323 		uvm_lock_pageq();
1324 		uvm_pageactivate(uobjpage);
1325 		uvm_unlock_pageq();
1326 
1327 		/*
1328 		 * re-verify the state of the world by first trying to relock
1329 		 * the maps.  always relock the object.
1330 		 */
1331 
1332 		locked = uvmfault_relock(&ufi);
1333 		if (locked && amap)
1334 			amap_lock(amap);
1335 		simple_lock(&uobj->vmobjlock);
1336 
1337 		/* locked(locked): maps(read), amap(if !null), uobj, uobjpage */
1338 		/* locked(!locked): uobj, uobjpage */
1339 
1340 		/*
1341 		 * verify that the page has not be released and re-verify
1342 		 * that amap slot is still free.   if there is a problem,
1343 		 * we unlock and clean up.
1344 		 */
1345 
1346 		if ((uobjpage->flags & PG_RELEASED) != 0 ||
1347 		    (locked && amap &&
1348 		    amap_lookup(&ufi.entry->aref,
1349 		      ufi.orig_rvaddr - ufi.entry->start))) {
1350 			if (locked)
1351 				uvmfault_unlockall(&ufi, amap, NULL, NULL);
1352 			locked = FALSE;
1353 		}
1354 
1355 		/*
1356 		 * didn't get the lock?   release the page and retry.
1357 		 */
1358 
1359 		if (locked == FALSE) {
1360 			UVMHIST_LOG(maphist,
1361 			    "  wasn't able to relock after fault: retry",
1362 			    0,0,0,0);
1363 			if (uobjpage->flags & PG_WANTED)
1364 				wakeup(uobjpage);
1365 			if (uobjpage->flags & PG_RELEASED) {
1366 				uvmexp.fltpgrele++;
1367 				uvm_pagefree(uobjpage);
1368 				goto ReFault;
1369 			}
1370 			uobjpage->flags &= ~(PG_BUSY|PG_WANTED);
1371 			UVM_PAGE_OWN(uobjpage, NULL);
1372 			simple_unlock(&uobj->vmobjlock);
1373 			goto ReFault;
1374 		}
1375 
1376 		/*
1377 		 * we have the data in uobjpage which is busy and
1378 		 * not released.  we are holding object lock (so the page
1379 		 * can't be released on us).
1380 		 */
1381 
1382 		/* locked: maps(read), amap(if !null), uobj, uobjpage */
1383 	}
1384 
1385 	/*
1386 	 * locked:
1387 	 * maps(read), amap(if !null), uobj(if !null), uobjpage(if uobj)
1388 	 */
1389 
1390 	/*
1391 	 * notes:
1392 	 *  - at this point uobjpage can not be NULL
1393 	 *  - at this point uobjpage can not be PG_RELEASED (since we checked
1394 	 *  for it above)
1395 	 *  - at this point uobjpage could be PG_WANTED (handle later)
1396 	 */
1397 
1398 	if (promote == FALSE) {
1399 
1400 		/*
1401 		 * we are not promoting.   if the mapping is COW ensure that we
1402 		 * don't give more access than we should (e.g. when doing a read
1403 		 * fault on a COPYONWRITE mapping we want to map the COW page in
1404 		 * R/O even though the entry protection could be R/W).
1405 		 *
1406 		 * set "pg" to the page we want to map in (uobjpage, usually)
1407 		 */
1408 
1409 		/* no anon in this case. */
1410 		anon = NULL;
1411 
1412 		uvmexp.flt_obj++;
1413 		if (UVM_ET_ISCOPYONWRITE(ufi.entry))
1414 			enter_prot &= ~VM_PROT_WRITE;
1415 		pg = uobjpage;		/* map in the actual object */
1416 
1417 		/* assert(uobjpage != PGO_DONTCARE) */
1418 
1419 		/*
1420 		 * we are faulting directly on the page.   be careful
1421 		 * about writing to loaned pages...
1422 		 */
1423 
1424 		if (uobjpage->loan_count) {
1425 			if ((access_type & VM_PROT_WRITE) == 0) {
1426 				/* read fault: cap the protection at readonly */
1427 				/* cap! */
1428 				enter_prot = enter_prot & ~VM_PROT_WRITE;
1429 			} else {
1430 				/* write fault: must break the loan here */
1431 
1432 				/* alloc new un-owned page */
1433 				pg = uvm_pagealloc(NULL, 0, NULL, 0);
1434 
1435 				if (pg == NULL) {
1436 
1437 					/*
1438 					 * drop ownership of page, it can't
1439 					 * be released
1440 					 */
1441 
1442 					if (uobjpage->flags & PG_WANTED)
1443 						wakeup(uobjpage);
1444 					uobjpage->flags &= ~(PG_BUSY|PG_WANTED);
1445 					UVM_PAGE_OWN(uobjpage, NULL);
1446 
1447 					uvmfault_unlockall(&ufi, amap, uobj,
1448 					  NULL);
1449 					UVMHIST_LOG(maphist,
1450 					  "  out of RAM breaking loan, waiting",
1451 					  0,0,0,0);
1452 					uvmexp.fltnoram++;
1453 					uvm_wait("flt_noram4");
1454 					goto ReFault;
1455 				}
1456 
1457 				/*
1458 				 * copy the data from the old page to the new
1459 				 * one and clear the fake/clean flags on the
1460 				 * new page (keep it busy).  force a reload
1461 				 * of the old page by clearing it from all
1462 				 * pmaps.  then lock the page queues to
1463 				 * rename the pages.
1464 				 */
1465 
1466 				uvm_pagecopy(uobjpage, pg);	/* old -> new */
1467 				pg->flags &= ~(PG_FAKE|PG_CLEAN);
1468 				pmap_page_protect(uobjpage, VM_PROT_NONE);
1469 				if (uobjpage->flags & PG_WANTED)
1470 					wakeup(uobjpage);
1471 				/* uobj still locked */
1472 				uobjpage->flags &= ~(PG_WANTED|PG_BUSY);
1473 				UVM_PAGE_OWN(uobjpage, NULL);
1474 
1475 				uvm_lock_pageq();
1476 				offset = uobjpage->offset;
1477 				uvm_pagerealloc(uobjpage, NULL, 0);
1478 
1479 				/*
1480 				 * at this point we have absolutely no
1481 				 * control over uobjpage
1482 				 */
1483 
1484 				/* install new page */
1485 				uvm_pageactivate(pg);
1486 				uvm_pagerealloc(pg, uobj, offset);
1487 				uvm_unlock_pageq();
1488 
1489 				/*
1490 				 * done!  loan is broken and "pg" is
1491 				 * PG_BUSY.   it can now replace uobjpage.
1492 				 */
1493 
1494 				uobjpage = pg;
1495 			}
1496 		}
1497 	} else {
1498 
1499 		/*
1500 		 * if we are going to promote the data to an anon we
1501 		 * allocate a blank anon here and plug it into our amap.
1502 		 */
1503 #if DIAGNOSTIC
1504 		if (amap == NULL)
1505 			panic("uvm_fault: want to promote data, but no anon");
1506 #endif
1507 
1508 		anon = uvm_analloc();
1509 		if (anon) {
1510 
1511 			/*
1512 			 * The new anon is locked.
1513 			 *
1514 			 * In `Fill in data...' below, if
1515 			 * uobjpage == PGO_DONTCARE, we want
1516 			 * a zero'd, dirty page, so have
1517 			 * uvm_pagealloc() do that for us.
1518 			 */
1519 
1520 			pg = uvm_pagealloc(NULL, 0, anon,
1521 			    (uobjpage == PGO_DONTCARE) ? UVM_PGA_ZERO : 0);
1522 		}
1523 
1524 		/*
1525 		 * out of memory resources?
1526 		 */
1527 
1528 		if (anon == NULL || pg == NULL) {
1529 			if (anon != NULL) {
1530 				anon->an_ref--;
1531 				simple_unlock(&anon->an_lock);
1532 				uvm_anfree(anon);
1533 			}
1534 
1535 			/*
1536 			 * arg!  must unbusy our page and fail or sleep.
1537 			 */
1538 
1539 			if (uobjpage != PGO_DONTCARE) {
1540 				if (uobjpage->flags & PG_WANTED)
1541 					/* still holding object lock */
1542 					wakeup(uobjpage);
1543 
1544 				uobjpage->flags &= ~(PG_BUSY|PG_WANTED);
1545 				UVM_PAGE_OWN(uobjpage, NULL);
1546 			}
1547 
1548 			/* unlock and fail ... */
1549 			uvmfault_unlockall(&ufi, amap, uobj, NULL);
1550 			KASSERT(uvmexp.swpgonly <= uvmexp.swpages);
1551 			if (anon == NULL || uvmexp.swpgonly == uvmexp.swpages) {
1552 				UVMHIST_LOG(maphist, "  promote: out of VM",
1553 				    0,0,0,0);
1554 				uvmexp.fltnoanon++;
1555 				return ENOMEM;
1556 			}
1557 
1558 			UVMHIST_LOG(maphist, "  out of RAM, waiting for more",
1559 			    0,0,0,0);
1560 			uvmexp.fltnoram++;
1561 			uvm_wait("flt_noram5");
1562 			goto ReFault;
1563 		}
1564 
1565 		/*
1566 		 * fill in the data
1567 		 */
1568 
1569 		if (uobjpage != PGO_DONTCARE) {
1570 			uvmexp.flt_prcopy++;
1571 			/* copy page [pg now dirty] */
1572 			uvm_pagecopy(uobjpage, pg);
1573 
1574 			/*
1575 			 * promote to shared amap?  make sure all sharing
1576 			 * procs see it
1577 			 */
1578 
1579 			if ((amap_flags(amap) & AMAP_SHARED) != 0) {
1580 				pmap_page_protect(uobjpage, VM_PROT_NONE);
1581 				/*
1582 				 * XXX: PAGE MIGHT BE WIRED!
1583 				 */
1584 			}
1585 
1586 			/*
1587 			 * dispose of uobjpage.  it can't be PG_RELEASED
1588 			 * since we still hold the object lock.
1589 			 * drop handle to uobj as well.
1590 			 */
1591 
1592 			if (uobjpage->flags & PG_WANTED)
1593 				/* still have the obj lock */
1594 				wakeup(uobjpage);
1595 			uobjpage->flags &= ~(PG_BUSY|PG_WANTED);
1596 			UVM_PAGE_OWN(uobjpage, NULL);
1597 			simple_unlock(&uobj->vmobjlock);
1598 			uobj = NULL;
1599 
1600 			UVMHIST_LOG(maphist,
1601 			    "  promote uobjpage 0x%x to anon/page 0x%x/0x%x",
1602 			    uobjpage, anon, pg, 0);
1603 
1604 		} else {
1605 			uvmexp.flt_przero++;
1606 
1607 			/*
1608 			 * Page is zero'd and marked dirty by uvm_pagealloc()
1609 			 * above.
1610 			 */
1611 
1612 			UVMHIST_LOG(maphist,"  zero fill anon/page 0x%x/0%x",
1613 			    anon, pg, 0, 0);
1614 		}
1615 		amap_add(&ufi.entry->aref, ufi.orig_rvaddr - ufi.entry->start,
1616 		    anon, 0);
1617 	}
1618 
1619 	/*
1620 	 * locked:
1621 	 * maps(read), amap(if !null), uobj(if !null), uobjpage(if uobj),
1622 	 *   anon(if !null), pg(if anon)
1623 	 *
1624 	 * note: pg is either the uobjpage or the new page in the new anon
1625 	 */
1626 
1627 	/*
1628 	 * all resources are present.   we can now map it in and free our
1629 	 * resources.
1630 	 */
1631 
1632 	UVMHIST_LOG(maphist,
1633 	    "  MAPPING: case2: pm=0x%x, va=0x%x, pg=0x%x, promote=%d",
1634 	    ufi.orig_map->pmap, ufi.orig_rvaddr, pg, promote);
1635 	KASSERT(access_type == VM_PROT_READ || (pg->flags & PG_RDONLY) == 0);
1636 	if (pmap_enter(ufi.orig_map->pmap, ufi.orig_rvaddr, VM_PAGE_TO_PHYS(pg),
1637 	    pg->flags & PG_RDONLY ? VM_PROT_READ : enter_prot,
1638 	    access_type | PMAP_CANFAIL | (wired ? PMAP_WIRED : 0)) != 0) {
1639 
1640 		/*
1641 		 * No need to undo what we did; we can simply think of
1642 		 * this as the pmap throwing away the mapping information.
1643 		 *
1644 		 * We do, however, have to go through the ReFault path,
1645 		 * as the map may change while we're asleep.
1646 		 */
1647 
1648 		if (pg->flags & PG_WANTED)
1649 			wakeup(pg);
1650 
1651 		/*
1652 		 * note that pg can't be PG_RELEASED since we did not drop
1653 		 * the object lock since the last time we checked.
1654 		 */
1655 
1656 		pg->flags &= ~(PG_BUSY|PG_FAKE|PG_WANTED);
1657 		UVM_PAGE_OWN(pg, NULL);
1658 		uvmfault_unlockall(&ufi, amap, uobj, anon);
1659 		KASSERT(uvmexp.swpgonly <= uvmexp.swpages);
1660 		if (uvmexp.swpgonly == uvmexp.swpages) {
1661 			UVMHIST_LOG(maphist,
1662 			    "<- failed.  out of VM",0,0,0,0);
1663 			/* XXX instrumentation */
1664 			return ENOMEM;
1665 		}
1666 		/* XXX instrumentation */
1667 		uvm_wait("flt_pmfail2");
1668 		goto ReFault;
1669 	}
1670 
1671 	uvm_lock_pageq();
1672 	if (fault_type == VM_FAULT_WIRE) {
1673 		uvm_pagewire(pg);
1674 		if (pg->pqflags & PQ_AOBJ) {
1675 
1676 			/*
1677 			 * since the now-wired page cannot be paged out,
1678 			 * release its swap resources for others to use.
1679 			 * since an aobj page with no swap cannot be PG_CLEAN,
1680 			 * clear its clean flag now.
1681 			 */
1682 
1683 			pg->flags &= ~(PG_CLEAN);
1684 			uao_dropswap(uobj, pg->offset >> PAGE_SHIFT);
1685 		}
1686 	} else {
1687 		uvm_pageactivate(pg);
1688 	}
1689 	uvm_unlock_pageq();
1690 	if (pg->flags & PG_WANTED)
1691 		wakeup(pg);
1692 
1693 	/*
1694 	 * note that pg can't be PG_RELEASED since we did not drop the object
1695 	 * lock since the last time we checked.
1696 	 */
1697 
1698 	pg->flags &= ~(PG_BUSY|PG_FAKE|PG_WANTED);
1699 	UVM_PAGE_OWN(pg, NULL);
1700 	uvmfault_unlockall(&ufi, amap, uobj, anon);
1701 	pmap_update(ufi.orig_map->pmap);
1702 	UVMHIST_LOG(maphist, "<- done (SUCCESS!)",0,0,0,0);
1703 	return 0;
1704 }
1705 
1706 /*
1707  * uvm_fault_wire: wire down a range of virtual addresses in a map.
1708  *
1709  * => map may be read-locked by caller, but MUST NOT be write-locked.
1710  * => if map is read-locked, any operations which may cause map to
1711  *	be write-locked in uvm_fault() must be taken care of by
1712  *	the caller.  See uvm_map_pageable().
1713  */
1714 
1715 int
1716 uvm_fault_wire(map, start, end, access_type)
1717 	struct vm_map *map;
1718 	vaddr_t start, end;
1719 	vm_prot_t access_type;
1720 {
1721 	vaddr_t va;
1722 	int error;
1723 
1724 	/*
1725 	 * now fault it in a page at a time.   if the fault fails then we have
1726 	 * to undo what we have done.   note that in uvm_fault VM_PROT_NONE
1727 	 * is replaced with the max protection if fault_type is VM_FAULT_WIRE.
1728 	 */
1729 
1730 	/*
1731 	 * XXX work around overflowing a vaddr_t.  this prevents us from
1732 	 * wiring the last page in the address space, though.
1733 	 */
1734 	if (start > end) {
1735 		return EFAULT;
1736 	}
1737 
1738 	for (va = start ; va < end ; va += PAGE_SIZE) {
1739 		error = uvm_fault(map, va, VM_FAULT_WIRE, access_type);
1740 		if (error) {
1741 			if (va != start) {
1742 				uvm_fault_unwire(map, start, va);
1743 			}
1744 			return error;
1745 		}
1746 	}
1747 	return 0;
1748 }
1749 
1750 /*
1751  * uvm_fault_unwire(): unwire range of virtual space.
1752  */
1753 
1754 void
1755 uvm_fault_unwire(map, start, end)
1756 	struct vm_map *map;
1757 	vaddr_t start, end;
1758 {
1759 	vm_map_lock_read(map);
1760 	uvm_fault_unwire_locked(map, start, end);
1761 	vm_map_unlock_read(map);
1762 }
1763 
1764 /*
1765  * uvm_fault_unwire_locked(): the guts of uvm_fault_unwire().
1766  *
1767  * => map must be at least read-locked.
1768  */
1769 
1770 void
1771 uvm_fault_unwire_locked(map, start, end)
1772 	struct vm_map *map;
1773 	vaddr_t start, end;
1774 {
1775 	struct vm_map_entry *entry;
1776 	pmap_t pmap = vm_map_pmap(map);
1777 	vaddr_t va;
1778 	paddr_t pa;
1779 	struct vm_page *pg;
1780 
1781 	KASSERT((map->flags & VM_MAP_INTRSAFE) == 0);
1782 
1783 	/*
1784 	 * we assume that the area we are unwiring has actually been wired
1785 	 * in the first place.   this means that we should be able to extract
1786 	 * the PAs from the pmap.   we also lock out the page daemon so that
1787 	 * we can call uvm_pageunwire.
1788 	 */
1789 
1790 	uvm_lock_pageq();
1791 
1792 	/*
1793 	 * find the beginning map entry for the region.
1794 	 */
1795 	KASSERT(start >= vm_map_min(map) && end <= vm_map_max(map));
1796 	if (uvm_map_lookup_entry(map, start, &entry) == FALSE)
1797 		panic("uvm_fault_unwire_locked: address not in map");
1798 
1799 	for (va = start; va < end; va += PAGE_SIZE) {
1800 		if (pmap_extract(pmap, va, &pa) == FALSE)
1801 			panic("uvm_fault_unwire_locked: unwiring "
1802 			    "non-wired memory");
1803 
1804 		/*
1805 		 * make sure the current entry is for the address we're
1806 		 * dealing with.  if not, grab the next entry.
1807 		 */
1808 
1809 		KASSERT(va >= entry->start);
1810 		if (va >= entry->end) {
1811 			KASSERT(entry->next != &map->header &&
1812 				entry->next->start <= entry->end);
1813 			entry = entry->next;
1814 		}
1815 
1816 		/*
1817 		 * if the entry is no longer wired, tell the pmap.
1818 		 */
1819 		if (VM_MAPENT_ISWIRED(entry) == 0)
1820 			pmap_unwire(pmap, va);
1821 
1822 		pg = PHYS_TO_VM_PAGE(pa);
1823 		if (pg)
1824 			uvm_pageunwire(pg);
1825 	}
1826 
1827 	uvm_unlock_pageq();
1828 }
1829