xref: /netbsd-src/sys/uvm/uvm_fault.c (revision 08c81a9c2dc8c7300e893321eb65c0925d60871c)
1 /*	$NetBSD: uvm_fault.c,v 1.78 2002/09/02 21:09:50 thorpej 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 1.78 2002/09/02 21:09:50 thorpej 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, check_prot;
536 	boolean_t wired, narrow, promote, locked, shadowed, wire_fault, cow_now;
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 	wire_fault = fault_type == VM_FAULT_WIRE ||
562 	    fault_type == VM_FAULT_WIREMAX;
563 	if (wire_fault)
564 		narrow = TRUE;		/* don't look for neighborhood
565 					 * pages on wire */
566 	else
567 		narrow = FALSE;		/* normal fault */
568 
569 	/*
570 	 * "goto ReFault" means restart the page fault from ground zero.
571 	 */
572 ReFault:
573 
574 	/*
575 	 * lookup and lock the maps
576 	 */
577 
578 	if (uvmfault_lookup(&ufi, FALSE) == FALSE) {
579 		UVMHIST_LOG(maphist, "<- no mapping @ 0x%x", vaddr, 0,0,0);
580 		return (EFAULT);
581 	}
582 	/* locked: maps(read) */
583 
584 #ifdef DIAGNOSTIC
585 	if ((ufi.map->flags & VM_MAP_PAGEABLE) == 0) {
586 		printf("Page fault on non-pageable map:\n");
587 		printf("ufi.map = %p\n", ufi.map);
588 		printf("ufi.orig_map = %p\n", ufi.orig_map);
589 		printf("ufi.orig_rvaddr = 0x%lx\n", (u_long) ufi.orig_rvaddr);
590 		panic("uvm_fault: (ufi.map->flags & VM_MAP_PAGEABLE) == 0");
591 	}
592 #endif
593 
594 	/*
595 	 * check protection
596 	 */
597 
598 	check_prot = fault_type == VM_FAULT_WIREMAX ?
599 	    ufi.entry->max_protection : ufi.entry->protection;
600 	if ((check_prot & access_type) != access_type) {
601 		UVMHIST_LOG(maphist,
602 		    "<- protection failure (prot=0x%x, access=0x%x)",
603 		    ufi.entry->protection, access_type, 0, 0);
604 		uvmfault_unlockmaps(&ufi, FALSE);
605 		return EACCES;
606 	}
607 
608 	/*
609 	 * "enter_prot" is the protection we want to enter the page in at.
610 	 * for certain pages (e.g. copy-on-write pages) this protection can
611 	 * be more strict than ufi.entry->protection.  "wired" means either
612 	 * the entry is wired or we are fault-wiring the pg.
613 	 */
614 
615 	enter_prot = ufi.entry->protection;
616 	wired = VM_MAPENT_ISWIRED(ufi.entry) || wire_fault;
617 	if (wired) {
618 		access_type = enter_prot; /* full access for wired */
619 		cow_now = (check_prot & VM_PROT_WRITE) != 0;
620 	} else {
621 		cow_now = (access_type & VM_PROT_WRITE) != 0;
622 	}
623 
624 	/*
625 	 * handle "needs_copy" case.   if we need to copy the amap we will
626 	 * have to drop our readlock and relock it with a write lock.  (we
627 	 * need a write lock to change anything in a map entry [e.g.
628 	 * needs_copy]).
629 	 */
630 
631 	if (UVM_ET_ISNEEDSCOPY(ufi.entry)) {
632 		KASSERT(fault_type != VM_FAULT_WIREMAX);
633 		if (cow_now || (ufi.entry->object.uvm_obj == NULL)) {
634 			/* need to clear */
635 			UVMHIST_LOG(maphist,
636 			    "  need to clear needs_copy and refault",0,0,0,0);
637 			uvmfault_unlockmaps(&ufi, FALSE);
638 			uvmfault_amapcopy(&ufi);
639 			uvmexp.fltamcopy++;
640 			goto ReFault;
641 
642 		} else {
643 
644 			/*
645 			 * ensure that we pmap_enter page R/O since
646 			 * needs_copy is still true
647 			 */
648 
649 			enter_prot &= ~VM_PROT_WRITE;
650 		}
651 	}
652 
653 	/*
654 	 * identify the players
655 	 */
656 
657 	amap = ufi.entry->aref.ar_amap;		/* top layer */
658 	uobj = ufi.entry->object.uvm_obj;	/* bottom layer */
659 
660 	/*
661 	 * check for a case 0 fault.  if nothing backing the entry then
662 	 * error now.
663 	 */
664 
665 	if (amap == NULL && uobj == NULL) {
666 		uvmfault_unlockmaps(&ufi, FALSE);
667 		UVMHIST_LOG(maphist,"<- no backing store, no overlay",0,0,0,0);
668 		return (EFAULT);
669 	}
670 
671 	/*
672 	 * establish range of interest based on advice from mapper
673 	 * and then clip to fit map entry.   note that we only want
674 	 * to do this the first time through the fault.   if we
675 	 * ReFault we will disable this by setting "narrow" to true.
676 	 */
677 
678 	if (narrow == FALSE) {
679 
680 		/* wide fault (!narrow) */
681 		KASSERT(uvmadvice[ufi.entry->advice].advice ==
682 			 ufi.entry->advice);
683 		nback = MIN(uvmadvice[ufi.entry->advice].nback,
684 			    (ufi.orig_rvaddr - ufi.entry->start) >> PAGE_SHIFT);
685 		startva = ufi.orig_rvaddr - (nback << PAGE_SHIFT);
686 		nforw = MIN(uvmadvice[ufi.entry->advice].nforw,
687 			    ((ufi.entry->end - ufi.orig_rvaddr) >>
688 			     PAGE_SHIFT) - 1);
689 		/*
690 		 * note: "-1" because we don't want to count the
691 		 * faulting page as forw
692 		 */
693 		npages = nback + nforw + 1;
694 		centeridx = nback;
695 
696 		narrow = TRUE;	/* ensure only once per-fault */
697 
698 	} else {
699 
700 		/* narrow fault! */
701 		nback = nforw = 0;
702 		startva = ufi.orig_rvaddr;
703 		npages = 1;
704 		centeridx = 0;
705 
706 	}
707 
708 	/* locked: maps(read) */
709 	UVMHIST_LOG(maphist, "  narrow=%d, back=%d, forw=%d, startva=0x%x",
710 		    narrow, nback, nforw, startva);
711 	UVMHIST_LOG(maphist, "  entry=0x%x, amap=0x%x, obj=0x%x", ufi.entry,
712 		    amap, uobj, 0);
713 
714 	/*
715 	 * if we've got an amap, lock it and extract current anons.
716 	 */
717 
718 	if (amap) {
719 		amap_lock(amap);
720 		anons = anons_store;
721 		amap_lookups(&ufi.entry->aref, startva - ufi.entry->start,
722 		    anons, npages);
723 	} else {
724 		anons = NULL;	/* to be safe */
725 	}
726 
727 	/* locked: maps(read), amap(if there) */
728 
729 	/*
730 	 * for MADV_SEQUENTIAL mappings we want to deactivate the back pages
731 	 * now and then forget about them (for the rest of the fault).
732 	 */
733 
734 	if (ufi.entry->advice == MADV_SEQUENTIAL && nback != 0) {
735 
736 		UVMHIST_LOG(maphist, "  MADV_SEQUENTIAL: flushing backpages",
737 		    0,0,0,0);
738 		/* flush back-page anons? */
739 		if (amap)
740 			uvmfault_anonflush(anons, nback);
741 
742 		/* flush object? */
743 		if (uobj) {
744 			objaddr =
745 			    (startva - ufi.entry->start) + ufi.entry->offset;
746 			simple_lock(&uobj->vmobjlock);
747 			(void) (uobj->pgops->pgo_put)(uobj, objaddr, objaddr +
748 				    (nback << PAGE_SHIFT), PGO_DEACTIVATE);
749 		}
750 
751 		/* now forget about the backpages */
752 		if (amap)
753 			anons += nback;
754 		startva += (nback << PAGE_SHIFT);
755 		npages -= nback;
756 		nback = centeridx = 0;
757 	}
758 
759 	/* locked: maps(read), amap(if there) */
760 
761 	/*
762 	 * map in the backpages and frontpages we found in the amap in hopes
763 	 * of preventing future faults.    we also init the pages[] array as
764 	 * we go.
765 	 */
766 
767 	currva = startva;
768 	shadowed = FALSE;
769 	for (lcv = 0 ; lcv < npages ; lcv++, currva += PAGE_SIZE) {
770 
771 		/*
772 		 * dont play with VAs that are already mapped
773 		 * except for center)
774 		 */
775 		if (lcv != centeridx &&
776 		    pmap_extract(ufi.orig_map->pmap, currva, &pa)) {
777 			pages[lcv] = PGO_DONTCARE;
778 			continue;
779 		}
780 
781 		/*
782 		 * unmapped or center page.   check if any anon at this level.
783 		 */
784 		if (amap == NULL || anons[lcv] == NULL) {
785 			pages[lcv] = NULL;
786 			continue;
787 		}
788 
789 		/*
790 		 * check for present page and map if possible.   re-activate it.
791 		 */
792 
793 		pages[lcv] = PGO_DONTCARE;
794 		if (lcv == centeridx) {		/* save center for later! */
795 			shadowed = TRUE;
796 			continue;
797 		}
798 		anon = anons[lcv];
799 		simple_lock(&anon->an_lock);
800 		/* ignore loaned pages */
801 		if (anon->u.an_page && anon->u.an_page->loan_count == 0 &&
802 		    (anon->u.an_page->flags & PG_BUSY) == 0) {
803 			uvm_lock_pageq();
804 			uvm_pageactivate(anon->u.an_page);
805 			uvm_unlock_pageq();
806 			UVMHIST_LOG(maphist,
807 			    "  MAPPING: n anon: pm=0x%x, va=0x%x, pg=0x%x",
808 			    ufi.orig_map->pmap, currva, anon->u.an_page, 0);
809 			uvmexp.fltnamap++;
810 
811 			/*
812 			 * Since this isn't the page that's actually faulting,
813 			 * ignore pmap_enter() failures; it's not critical
814 			 * that we enter these right now.
815 			 */
816 
817 			(void) pmap_enter(ufi.orig_map->pmap, currva,
818 			    VM_PAGE_TO_PHYS(anon->u.an_page),
819 			    (anon->an_ref > 1) ? (enter_prot & ~VM_PROT_WRITE) :
820 			    enter_prot,
821 			    PMAP_CANFAIL |
822 			     (VM_MAPENT_ISWIRED(ufi.entry) ? PMAP_WIRED : 0));
823 		}
824 		simple_unlock(&anon->an_lock);
825 		pmap_update(ufi.orig_map->pmap);
826 	}
827 
828 	/* locked: maps(read), amap(if there) */
829 	/* (shadowed == TRUE) if there is an anon at the faulting address */
830 	UVMHIST_LOG(maphist, "  shadowed=%d, will_get=%d", shadowed,
831 	    (uobj && shadowed == FALSE),0,0);
832 
833 	/*
834 	 * note that if we are really short of RAM we could sleep in the above
835 	 * call to pmap_enter with everything locked.   bad?
836 	 *
837 	 * XXX Actually, that is bad; pmap_enter() should just fail in that
838 	 * XXX case.  --thorpej
839 	 */
840 
841 	/*
842 	 * if the desired page is not shadowed by the amap and we have a
843 	 * backing object, then we check to see if the backing object would
844 	 * prefer to handle the fault itself (rather than letting us do it
845 	 * with the usual pgo_get hook).  the backing object signals this by
846 	 * providing a pgo_fault routine.
847 	 */
848 
849 	if (uobj && shadowed == FALSE && uobj->pgops->pgo_fault != NULL) {
850 		simple_lock(&uobj->vmobjlock);
851 
852 		/* locked: maps(read), amap (if there), uobj */
853 		error = uobj->pgops->pgo_fault(&ufi, startva, pages, npages,
854 		    centeridx, fault_type, access_type, PGO_LOCKED|PGO_SYNCIO);
855 
856 		/* locked: nothing, pgo_fault has unlocked everything */
857 
858 		if (error == ERESTART)
859 			goto ReFault;		/* try again! */
860 		/*
861 		 * object fault routine responsible for pmap_update().
862 		 */
863 		return error;
864 	}
865 
866 	/*
867 	 * now, if the desired page is not shadowed by the amap and we have
868 	 * a backing object that does not have a special fault routine, then
869 	 * we ask (with pgo_get) the object for resident pages that we care
870 	 * about and attempt to map them in.  we do not let pgo_get block
871 	 * (PGO_LOCKED).
872 	 */
873 
874 	if (uobj && shadowed == FALSE) {
875 		simple_lock(&uobj->vmobjlock);
876 
877 		/* locked (!shadowed): maps(read), amap (if there), uobj */
878 		/*
879 		 * the following call to pgo_get does _not_ change locking state
880 		 */
881 
882 		uvmexp.fltlget++;
883 		gotpages = npages;
884 		(void) uobj->pgops->pgo_get(uobj, ufi.entry->offset +
885 				(startva - ufi.entry->start),
886 				pages, &gotpages, centeridx,
887 				access_type & MASK(ufi.entry),
888 				ufi.entry->advice, PGO_LOCKED);
889 
890 		/*
891 		 * check for pages to map, if we got any
892 		 */
893 
894 		uobjpage = NULL;
895 
896 		if (gotpages) {
897 			currva = startva;
898 			for (lcv = 0; lcv < npages;
899 			     lcv++, currva += PAGE_SIZE) {
900 				if (pages[lcv] == NULL ||
901 				    pages[lcv] == PGO_DONTCARE) {
902 					continue;
903 				}
904 
905 				/*
906 				 * if center page is resident and not
907 				 * PG_BUSY|PG_RELEASED then pgo_get
908 				 * made it PG_BUSY for us and gave
909 				 * us a handle to it.   remember this
910 				 * page as "uobjpage." (for later use).
911 				 */
912 
913 				if (lcv == centeridx) {
914 					uobjpage = pages[lcv];
915 					UVMHIST_LOG(maphist, "  got uobjpage "
916 					    "(0x%x) with locked get",
917 					    uobjpage, 0,0,0);
918 					continue;
919 				}
920 
921 				/*
922 				 * calling pgo_get with PGO_LOCKED returns us
923 				 * pages which are neither busy nor released,
924 				 * so we don't need to check for this.
925 				 * we can just directly enter the pages.
926 				 */
927 
928 				uvm_lock_pageq();
929 				uvm_pageactivate(pages[lcv]);
930 				uvm_unlock_pageq();
931 				UVMHIST_LOG(maphist,
932 				  "  MAPPING: n obj: pm=0x%x, va=0x%x, pg=0x%x",
933 				  ufi.orig_map->pmap, currva, pages[lcv], 0);
934 				uvmexp.fltnomap++;
935 
936 				/*
937 				 * Since this page isn't the page that's
938 				 * actually fauling, ignore pmap_enter()
939 				 * failures; it's not critical that we
940 				 * enter these right now.
941 				 */
942 
943 				(void) pmap_enter(ufi.orig_map->pmap, currva,
944 				    VM_PAGE_TO_PHYS(pages[lcv]),
945 				    pages[lcv]->flags & PG_RDONLY ?
946 				    enter_prot & ~VM_PROT_WRITE :
947 				    enter_prot & MASK(ufi.entry),
948 				    PMAP_CANFAIL |
949 				     (wired ? PMAP_WIRED : 0));
950 
951 				/*
952 				 * NOTE: page can't be PG_WANTED or PG_RELEASED
953 				 * because we've held the lock the whole time
954 				 * we've had the handle.
955 				 */
956 
957 				pages[lcv]->flags &= ~(PG_BUSY);
958 				UVM_PAGE_OWN(pages[lcv], NULL);
959 			}
960 			pmap_update(ufi.orig_map->pmap);
961 		}
962 	} else {
963 		uobjpage = NULL;
964 	}
965 
966 	/* locked (shadowed): maps(read), amap */
967 	/* locked (!shadowed): maps(read), amap(if there),
968 		 uobj(if !null), uobjpage(if !null) */
969 
970 	/*
971 	 * note that at this point we are done with any front or back pages.
972 	 * we are now going to focus on the center page (i.e. the one we've
973 	 * faulted on).  if we have faulted on the top (anon) layer
974 	 * [i.e. case 1], then the anon we want is anons[centeridx] (we have
975 	 * not touched it yet).  if we have faulted on the bottom (uobj)
976 	 * layer [i.e. case 2] and the page was both present and available,
977 	 * then we've got a pointer to it as "uobjpage" and we've already
978 	 * made it BUSY.
979 	 */
980 
981 	/*
982 	 * there are four possible cases we must address: 1A, 1B, 2A, and 2B
983 	 */
984 
985 	/*
986 	 * redirect case 2: if we are not shadowed, go to case 2.
987 	 */
988 
989 	if (shadowed == FALSE)
990 		goto Case2;
991 
992 	/* locked: maps(read), amap */
993 
994 	/*
995 	 * handle case 1: fault on an anon in our amap
996 	 */
997 
998 	anon = anons[centeridx];
999 	UVMHIST_LOG(maphist, "  case 1 fault: anon=0x%x", anon, 0,0,0);
1000 	simple_lock(&anon->an_lock);
1001 
1002 	/* locked: maps(read), amap, anon */
1003 
1004 	/*
1005 	 * no matter if we have case 1A or case 1B we are going to need to
1006 	 * have the anon's memory resident.   ensure that now.
1007 	 */
1008 
1009 	/*
1010 	 * let uvmfault_anonget do the dirty work.
1011 	 * if it fails (!OK) it will unlock everything for us.
1012 	 * if it succeeds, locks are still valid and locked.
1013 	 * also, if it is OK, then the anon's page is on the queues.
1014 	 * if the page is on loan from a uvm_object, then anonget will
1015 	 * lock that object for us if it does not fail.
1016 	 */
1017 
1018 	error = uvmfault_anonget(&ufi, amap, anon);
1019 	switch (error) {
1020 	case 0:
1021 		break;
1022 
1023 	case ERESTART:
1024 		goto ReFault;
1025 
1026 	case EAGAIN:
1027 		tsleep(&lbolt, PVM, "fltagain1", 0);
1028 		goto ReFault;
1029 
1030 	default:
1031 		return error;
1032 	}
1033 
1034 	/*
1035 	 * uobj is non null if the page is on loan from an object (i.e. uobj)
1036 	 */
1037 
1038 	uobj = anon->u.an_page->uobject;	/* locked by anonget if !NULL */
1039 
1040 	/* locked: maps(read), amap, anon, uobj(if one) */
1041 
1042 	/*
1043 	 * special handling for loaned pages
1044 	 */
1045 
1046 	if (anon->u.an_page->loan_count) {
1047 
1048 		if (!cow_now) {
1049 
1050 			/*
1051 			 * for read faults on loaned pages we just cap the
1052 			 * protection at read-only.
1053 			 */
1054 
1055 			enter_prot = enter_prot & ~VM_PROT_WRITE;
1056 
1057 		} else {
1058 			/*
1059 			 * note that we can't allow writes into a loaned page!
1060 			 *
1061 			 * if we have a write fault on a loaned page in an
1062 			 * anon then we need to look at the anon's ref count.
1063 			 * if it is greater than one then we are going to do
1064 			 * a normal copy-on-write fault into a new anon (this
1065 			 * is not a problem).  however, if the reference count
1066 			 * is one (a case where we would normally allow a
1067 			 * write directly to the page) then we need to kill
1068 			 * the loan before we continue.
1069 			 */
1070 
1071 			/* >1 case is already ok */
1072 			if (anon->an_ref == 1) {
1073 
1074 				/* get new un-owned replacement page */
1075 				pg = uvm_pagealloc(NULL, 0, NULL, 0);
1076 				if (pg == NULL) {
1077 					uvmfault_unlockall(&ufi, amap, uobj,
1078 					    anon);
1079 					uvm_wait("flt_noram2");
1080 					goto ReFault;
1081 				}
1082 
1083 				/*
1084 				 * copy data, kill loan, and drop uobj lock
1085 				 * (if any)
1086 				 */
1087 				/* copy old -> new */
1088 				uvm_pagecopy(anon->u.an_page, pg);
1089 
1090 				/* force reload */
1091 				pmap_page_protect(anon->u.an_page,
1092 						  VM_PROT_NONE);
1093 				uvm_lock_pageq();	  /* KILL loan */
1094 
1095 				anon->u.an_page->uanon = NULL;
1096 				/* in case we owned */
1097 				anon->u.an_page->pqflags &= ~PQ_ANON;
1098 
1099 				if (uobj) {
1100 					/* if we were receiver of loan */
1101 					anon->u.an_page->loan_count--;
1102 				} else {
1103 					/*
1104 					 * we were the lender (A->K); need
1105 					 * to remove the page from pageq's.
1106 					 */
1107 					uvm_pagedequeue(anon->u.an_page);
1108 				}
1109 
1110 				uvm_pageactivate(pg);
1111 				uvm_unlock_pageq();
1112 				if (uobj) {
1113 					simple_unlock(&uobj->vmobjlock);
1114 					uobj = NULL;
1115 				}
1116 
1117 				/* install new page in anon */
1118 				anon->u.an_page = pg;
1119 				pg->uanon = anon;
1120 				pg->pqflags |= PQ_ANON;
1121 				pg->flags &= ~(PG_BUSY|PG_FAKE);
1122 				UVM_PAGE_OWN(pg, NULL);
1123 
1124 				/* done! */
1125 			}     /* ref == 1 */
1126 		}       /* write fault */
1127 	}         /* loan count */
1128 
1129 	/*
1130 	 * if we are case 1B then we will need to allocate a new blank
1131 	 * anon to transfer the data into.   note that we have a lock
1132 	 * on anon, so no one can busy or release the page until we are done.
1133 	 * also note that the ref count can't drop to zero here because
1134 	 * it is > 1 and we are only dropping one ref.
1135 	 *
1136 	 * in the (hopefully very rare) case that we are out of RAM we
1137 	 * will unlock, wait for more RAM, and refault.
1138 	 *
1139 	 * if we are out of anon VM we kill the process (XXX: could wait?).
1140 	 */
1141 
1142 	if (cow_now && anon->an_ref > 1) {
1143 
1144 		UVMHIST_LOG(maphist, "  case 1B: COW fault",0,0,0,0);
1145 		uvmexp.flt_acow++;
1146 		oanon = anon;		/* oanon = old, locked anon */
1147 		anon = uvm_analloc();
1148 		if (anon) {
1149 			/* new anon is locked! */
1150 			pg = uvm_pagealloc(NULL, 0, anon, 0);
1151 		}
1152 
1153 		/* check for out of RAM */
1154 		if (anon == NULL || pg == NULL) {
1155 			if (anon) {
1156 				anon->an_ref--;
1157 				simple_unlock(&anon->an_lock);
1158 				uvm_anfree(anon);
1159 			}
1160 			uvmfault_unlockall(&ufi, amap, uobj, oanon);
1161 			KASSERT(uvmexp.swpgonly <= uvmexp.swpages);
1162 			if (anon == NULL || uvmexp.swpgonly == uvmexp.swpages) {
1163 				UVMHIST_LOG(maphist,
1164 				    "<- failed.  out of VM",0,0,0,0);
1165 				uvmexp.fltnoanon++;
1166 				return ENOMEM;
1167 			}
1168 
1169 			uvmexp.fltnoram++;
1170 			uvm_wait("flt_noram3");	/* out of RAM, wait for more */
1171 			goto ReFault;
1172 		}
1173 
1174 		/* got all resources, replace anon with nanon */
1175 		uvm_pagecopy(oanon->u.an_page, pg);
1176 		uvm_pageactivate(pg);
1177 		pg->flags &= ~(PG_BUSY|PG_FAKE);
1178 		UVM_PAGE_OWN(pg, NULL);
1179 		amap_add(&ufi.entry->aref, ufi.orig_rvaddr - ufi.entry->start,
1180 		    anon, 1);
1181 
1182 		/* deref: can not drop to zero here by defn! */
1183 		oanon->an_ref--;
1184 
1185 		/*
1186 		 * note: oanon is still locked, as is the new anon.  we
1187 		 * need to check for this later when we unlock oanon; if
1188 		 * oanon != anon, we'll have to unlock anon, too.
1189 		 */
1190 
1191 	} else {
1192 
1193 		uvmexp.flt_anon++;
1194 		oanon = anon;		/* old, locked anon is same as anon */
1195 		pg = anon->u.an_page;
1196 		if (anon->an_ref > 1)     /* disallow writes to ref > 1 anons */
1197 			enter_prot = enter_prot & ~VM_PROT_WRITE;
1198 
1199 	}
1200 
1201 	/* locked: maps(read), amap, oanon, anon (if different from oanon) */
1202 
1203 	/*
1204 	 * now map the page in.
1205 	 */
1206 
1207 	UVMHIST_LOG(maphist, "  MAPPING: anon: pm=0x%x, va=0x%x, pg=0x%x",
1208 	    ufi.orig_map->pmap, ufi.orig_rvaddr, pg, 0);
1209 	if (pmap_enter(ufi.orig_map->pmap, ufi.orig_rvaddr, VM_PAGE_TO_PHYS(pg),
1210 	    enter_prot, access_type | PMAP_CANFAIL | (wired ? PMAP_WIRED : 0))
1211 	    != 0) {
1212 
1213 		/*
1214 		 * No need to undo what we did; we can simply think of
1215 		 * this as the pmap throwing away the mapping information.
1216 		 *
1217 		 * We do, however, have to go through the ReFault path,
1218 		 * as the map may change while we're asleep.
1219 		 */
1220 
1221 		if (anon != oanon)
1222 			simple_unlock(&anon->an_lock);
1223 		uvmfault_unlockall(&ufi, amap, uobj, oanon);
1224 		KASSERT(uvmexp.swpgonly <= uvmexp.swpages);
1225 		if (uvmexp.swpgonly == uvmexp.swpages) {
1226 			UVMHIST_LOG(maphist,
1227 			    "<- failed.  out of VM",0,0,0,0);
1228 			/* XXX instrumentation */
1229 			return ENOMEM;
1230 		}
1231 		/* XXX instrumentation */
1232 		uvm_wait("flt_pmfail1");
1233 		goto ReFault;
1234 	}
1235 
1236 	/*
1237 	 * ... update the page queues.
1238 	 */
1239 
1240 	uvm_lock_pageq();
1241 	if (wire_fault) {
1242 		uvm_pagewire(pg);
1243 
1244 		/*
1245 		 * since the now-wired page cannot be paged out,
1246 		 * release its swap resources for others to use.
1247 		 * since an anon with no swap cannot be PG_CLEAN,
1248 		 * clear its clean flag now.
1249 		 */
1250 
1251 		pg->flags &= ~(PG_CLEAN);
1252 		uvm_anon_dropswap(anon);
1253 	} else {
1254 		uvm_pageactivate(pg);
1255 	}
1256 	uvm_unlock_pageq();
1257 
1258 	/*
1259 	 * done case 1!  finish up by unlocking everything and returning success
1260 	 */
1261 
1262 	if (anon != oanon)
1263 		simple_unlock(&anon->an_lock);
1264 	uvmfault_unlockall(&ufi, amap, uobj, oanon);
1265 	pmap_update(ufi.orig_map->pmap);
1266 	return 0;
1267 
1268 Case2:
1269 	/*
1270 	 * handle case 2: faulting on backing object or zero fill
1271 	 */
1272 
1273 	/*
1274 	 * locked:
1275 	 * maps(read), amap(if there), uobj(if !null), uobjpage(if !null)
1276 	 */
1277 
1278 	/*
1279 	 * note that uobjpage can not be PGO_DONTCARE at this point.  we now
1280 	 * set uobjpage to PGO_DONTCARE if we are doing a zero fill.  if we
1281 	 * have a backing object, check and see if we are going to promote
1282 	 * the data up to an anon during the fault.
1283 	 */
1284 
1285 	if (uobj == NULL) {
1286 		uobjpage = PGO_DONTCARE;
1287 		promote = TRUE;		/* always need anon here */
1288 	} else {
1289 		KASSERT(uobjpage != PGO_DONTCARE);
1290 		promote = cow_now && UVM_ET_ISCOPYONWRITE(ufi.entry);
1291 	}
1292 	UVMHIST_LOG(maphist, "  case 2 fault: promote=%d, zfill=%d",
1293 	    promote, (uobj == NULL), 0,0);
1294 
1295 	/*
1296 	 * if uobjpage is not null then we do not need to do I/O to get the
1297 	 * uobjpage.
1298 	 *
1299 	 * if uobjpage is null, then we need to unlock and ask the pager to
1300 	 * get the data for us.   once we have the data, we need to reverify
1301 	 * the state the world.   we are currently not holding any resources.
1302 	 */
1303 
1304 	if (uobjpage) {
1305 		/* update rusage counters */
1306 		curproc->p_addr->u_stats.p_ru.ru_minflt++;
1307 	} else {
1308 		/* update rusage counters */
1309 		curproc->p_addr->u_stats.p_ru.ru_majflt++;
1310 
1311 		/* locked: maps(read), amap(if there), uobj */
1312 		uvmfault_unlockall(&ufi, amap, NULL, NULL);
1313 		/* locked: uobj */
1314 
1315 		uvmexp.fltget++;
1316 		gotpages = 1;
1317 		uoff = (ufi.orig_rvaddr - ufi.entry->start) + ufi.entry->offset;
1318 		error = uobj->pgops->pgo_get(uobj, uoff, &uobjpage, &gotpages,
1319 		    0, access_type & MASK(ufi.entry), ufi.entry->advice,
1320 		    PGO_SYNCIO);
1321 		/* locked: uobjpage(if no error) */
1322 
1323 		/*
1324 		 * recover from I/O
1325 		 */
1326 
1327 		if (error) {
1328 			if (error == EAGAIN) {
1329 				UVMHIST_LOG(maphist,
1330 				    "  pgo_get says TRY AGAIN!",0,0,0,0);
1331 				tsleep(&lbolt, PVM, "fltagain2", 0);
1332 				goto ReFault;
1333 			}
1334 
1335 			UVMHIST_LOG(maphist, "<- pgo_get failed (code %d)",
1336 			    error, 0,0,0);
1337 			return error;
1338 		}
1339 
1340 		/* locked: uobjpage */
1341 
1342 		uvm_lock_pageq();
1343 		uvm_pageactivate(uobjpage);
1344 		uvm_unlock_pageq();
1345 
1346 		/*
1347 		 * re-verify the state of the world by first trying to relock
1348 		 * the maps.  always relock the object.
1349 		 */
1350 
1351 		locked = uvmfault_relock(&ufi);
1352 		if (locked && amap)
1353 			amap_lock(amap);
1354 		simple_lock(&uobj->vmobjlock);
1355 
1356 		/* locked(locked): maps(read), amap(if !null), uobj, uobjpage */
1357 		/* locked(!locked): uobj, uobjpage */
1358 
1359 		/*
1360 		 * verify that the page has not be released and re-verify
1361 		 * that amap slot is still free.   if there is a problem,
1362 		 * we unlock and clean up.
1363 		 */
1364 
1365 		if ((uobjpage->flags & PG_RELEASED) != 0 ||
1366 		    (locked && amap &&
1367 		    amap_lookup(&ufi.entry->aref,
1368 		      ufi.orig_rvaddr - ufi.entry->start))) {
1369 			if (locked)
1370 				uvmfault_unlockall(&ufi, amap, NULL, NULL);
1371 			locked = FALSE;
1372 		}
1373 
1374 		/*
1375 		 * didn't get the lock?   release the page and retry.
1376 		 */
1377 
1378 		if (locked == FALSE) {
1379 			UVMHIST_LOG(maphist,
1380 			    "  wasn't able to relock after fault: retry",
1381 			    0,0,0,0);
1382 			if (uobjpage->flags & PG_WANTED)
1383 				wakeup(uobjpage);
1384 			if (uobjpage->flags & PG_RELEASED) {
1385 				uvmexp.fltpgrele++;
1386 				uvm_pagefree(uobjpage);
1387 				goto ReFault;
1388 			}
1389 			uobjpage->flags &= ~(PG_BUSY|PG_WANTED);
1390 			UVM_PAGE_OWN(uobjpage, NULL);
1391 			simple_unlock(&uobj->vmobjlock);
1392 			goto ReFault;
1393 		}
1394 
1395 		/*
1396 		 * we have the data in uobjpage which is busy and
1397 		 * not released.  we are holding object lock (so the page
1398 		 * can't be released on us).
1399 		 */
1400 
1401 		/* locked: maps(read), amap(if !null), uobj, uobjpage */
1402 	}
1403 
1404 	/*
1405 	 * locked:
1406 	 * maps(read), amap(if !null), uobj(if !null), uobjpage(if uobj)
1407 	 */
1408 
1409 	/*
1410 	 * notes:
1411 	 *  - at this point uobjpage can not be NULL
1412 	 *  - at this point uobjpage can not be PG_RELEASED (since we checked
1413 	 *  for it above)
1414 	 *  - at this point uobjpage could be PG_WANTED (handle later)
1415 	 */
1416 
1417 	if (promote == FALSE) {
1418 
1419 		/*
1420 		 * we are not promoting.   if the mapping is COW ensure that we
1421 		 * don't give more access than we should (e.g. when doing a read
1422 		 * fault on a COPYONWRITE mapping we want to map the COW page in
1423 		 * R/O even though the entry protection could be R/W).
1424 		 *
1425 		 * set "pg" to the page we want to map in (uobjpage, usually)
1426 		 */
1427 
1428 		/* no anon in this case. */
1429 		anon = NULL;
1430 
1431 		uvmexp.flt_obj++;
1432 		if (UVM_ET_ISCOPYONWRITE(ufi.entry))
1433 			enter_prot &= ~VM_PROT_WRITE;
1434 		pg = uobjpage;		/* map in the actual object */
1435 
1436 		/* assert(uobjpage != PGO_DONTCARE) */
1437 
1438 		/*
1439 		 * we are faulting directly on the page.   be careful
1440 		 * about writing to loaned pages...
1441 		 */
1442 
1443 		if (uobjpage->loan_count) {
1444 			if (!cow_now) {
1445 				/* read fault: cap the protection at readonly */
1446 				/* cap! */
1447 				enter_prot = enter_prot & ~VM_PROT_WRITE;
1448 			} else {
1449 				/* write fault: must break the loan here */
1450 
1451 				/* alloc new un-owned page */
1452 				pg = uvm_pagealloc(NULL, 0, NULL, 0);
1453 
1454 				if (pg == NULL) {
1455 
1456 					/*
1457 					 * drop ownership of page, it can't
1458 					 * be released
1459 					 */
1460 
1461 					if (uobjpage->flags & PG_WANTED)
1462 						wakeup(uobjpage);
1463 					uobjpage->flags &= ~(PG_BUSY|PG_WANTED);
1464 					UVM_PAGE_OWN(uobjpage, NULL);
1465 
1466 					uvmfault_unlockall(&ufi, amap, uobj,
1467 					  NULL);
1468 					UVMHIST_LOG(maphist,
1469 					  "  out of RAM breaking loan, waiting",
1470 					  0,0,0,0);
1471 					uvmexp.fltnoram++;
1472 					uvm_wait("flt_noram4");
1473 					goto ReFault;
1474 				}
1475 
1476 				/*
1477 				 * copy the data from the old page to the new
1478 				 * one and clear the fake/clean flags on the
1479 				 * new page (keep it busy).  force a reload
1480 				 * of the old page by clearing it from all
1481 				 * pmaps.  then lock the page queues to
1482 				 * rename the pages.
1483 				 */
1484 
1485 				uvm_pagecopy(uobjpage, pg);	/* old -> new */
1486 				pg->flags &= ~(PG_FAKE|PG_CLEAN);
1487 				pmap_page_protect(uobjpage, VM_PROT_NONE);
1488 				if (uobjpage->flags & PG_WANTED)
1489 					wakeup(uobjpage);
1490 				/* uobj still locked */
1491 				uobjpage->flags &= ~(PG_WANTED|PG_BUSY);
1492 				UVM_PAGE_OWN(uobjpage, NULL);
1493 
1494 				uvm_lock_pageq();
1495 				offset = uobjpage->offset;
1496 				uvm_pagerealloc(uobjpage, NULL, 0);
1497 
1498 				/*
1499 				 * if the page is no longer referenced by
1500 				 * an anon (i.e. we are breaking an O->K
1501 				 * loan), then remove it from any pageq's.
1502 				 */
1503 				if (uobjpage->uanon == NULL)
1504 					uvm_pagedequeue(uobjpage);
1505 
1506 				/*
1507 				 * at this point we have absolutely no
1508 				 * control over uobjpage
1509 				 */
1510 
1511 				/* install new page */
1512 				uvm_pageactivate(pg);
1513 				uvm_pagerealloc(pg, uobj, offset);
1514 				uvm_unlock_pageq();
1515 
1516 				/*
1517 				 * done!  loan is broken and "pg" is
1518 				 * PG_BUSY.   it can now replace uobjpage.
1519 				 */
1520 
1521 				uobjpage = pg;
1522 			}
1523 		}
1524 	} else {
1525 
1526 		/*
1527 		 * if we are going to promote the data to an anon we
1528 		 * allocate a blank anon here and plug it into our amap.
1529 		 */
1530 #if DIAGNOSTIC
1531 		if (amap == NULL)
1532 			panic("uvm_fault: want to promote data, but no anon");
1533 #endif
1534 
1535 		anon = uvm_analloc();
1536 		if (anon) {
1537 
1538 			/*
1539 			 * The new anon is locked.
1540 			 *
1541 			 * In `Fill in data...' below, if
1542 			 * uobjpage == PGO_DONTCARE, we want
1543 			 * a zero'd, dirty page, so have
1544 			 * uvm_pagealloc() do that for us.
1545 			 */
1546 
1547 			pg = uvm_pagealloc(NULL, 0, anon,
1548 			    (uobjpage == PGO_DONTCARE) ? UVM_PGA_ZERO : 0);
1549 		}
1550 
1551 		/*
1552 		 * out of memory resources?
1553 		 */
1554 
1555 		if (anon == NULL || pg == NULL) {
1556 			if (anon != NULL) {
1557 				anon->an_ref--;
1558 				simple_unlock(&anon->an_lock);
1559 				uvm_anfree(anon);
1560 			}
1561 
1562 			/*
1563 			 * arg!  must unbusy our page and fail or sleep.
1564 			 */
1565 
1566 			if (uobjpage != PGO_DONTCARE) {
1567 				if (uobjpage->flags & PG_WANTED)
1568 					/* still holding object lock */
1569 					wakeup(uobjpage);
1570 
1571 				uobjpage->flags &= ~(PG_BUSY|PG_WANTED);
1572 				UVM_PAGE_OWN(uobjpage, NULL);
1573 			}
1574 
1575 			/* unlock and fail ... */
1576 			uvmfault_unlockall(&ufi, amap, uobj, NULL);
1577 			KASSERT(uvmexp.swpgonly <= uvmexp.swpages);
1578 			if (anon == NULL || uvmexp.swpgonly == uvmexp.swpages) {
1579 				UVMHIST_LOG(maphist, "  promote: out of VM",
1580 				    0,0,0,0);
1581 				uvmexp.fltnoanon++;
1582 				return ENOMEM;
1583 			}
1584 
1585 			UVMHIST_LOG(maphist, "  out of RAM, waiting for more",
1586 			    0,0,0,0);
1587 			uvmexp.fltnoram++;
1588 			uvm_wait("flt_noram5");
1589 			goto ReFault;
1590 		}
1591 
1592 		/*
1593 		 * fill in the data
1594 		 */
1595 
1596 		if (uobjpage != PGO_DONTCARE) {
1597 			uvmexp.flt_prcopy++;
1598 			/* copy page [pg now dirty] */
1599 			uvm_pagecopy(uobjpage, pg);
1600 
1601 			/*
1602 			 * promote to shared amap?  make sure all sharing
1603 			 * procs see it
1604 			 */
1605 
1606 			if ((amap_flags(amap) & AMAP_SHARED) != 0) {
1607 				pmap_page_protect(uobjpage, VM_PROT_NONE);
1608 				/*
1609 				 * XXX: PAGE MIGHT BE WIRED!
1610 				 */
1611 			}
1612 
1613 			/*
1614 			 * dispose of uobjpage.  it can't be PG_RELEASED
1615 			 * since we still hold the object lock.
1616 			 * drop handle to uobj as well.
1617 			 */
1618 
1619 			if (uobjpage->flags & PG_WANTED)
1620 				/* still have the obj lock */
1621 				wakeup(uobjpage);
1622 			uobjpage->flags &= ~(PG_BUSY|PG_WANTED);
1623 			UVM_PAGE_OWN(uobjpage, NULL);
1624 			simple_unlock(&uobj->vmobjlock);
1625 			uobj = NULL;
1626 
1627 			UVMHIST_LOG(maphist,
1628 			    "  promote uobjpage 0x%x to anon/page 0x%x/0x%x",
1629 			    uobjpage, anon, pg, 0);
1630 
1631 		} else {
1632 			uvmexp.flt_przero++;
1633 
1634 			/*
1635 			 * Page is zero'd and marked dirty by uvm_pagealloc()
1636 			 * above.
1637 			 */
1638 
1639 			UVMHIST_LOG(maphist,"  zero fill anon/page 0x%x/0%x",
1640 			    anon, pg, 0, 0);
1641 		}
1642 		amap_add(&ufi.entry->aref, ufi.orig_rvaddr - ufi.entry->start,
1643 		    anon, 0);
1644 	}
1645 
1646 	/*
1647 	 * locked:
1648 	 * maps(read), amap(if !null), uobj(if !null), uobjpage(if uobj),
1649 	 *   anon(if !null), pg(if anon)
1650 	 *
1651 	 * note: pg is either the uobjpage or the new page in the new anon
1652 	 */
1653 
1654 	/*
1655 	 * all resources are present.   we can now map it in and free our
1656 	 * resources.
1657 	 */
1658 
1659 	UVMHIST_LOG(maphist,
1660 	    "  MAPPING: case2: pm=0x%x, va=0x%x, pg=0x%x, promote=%d",
1661 	    ufi.orig_map->pmap, ufi.orig_rvaddr, pg, promote);
1662 	KASSERT((access_type & VM_PROT_WRITE) == 0 ||
1663 		(pg->flags & PG_RDONLY) == 0);
1664 	if (pmap_enter(ufi.orig_map->pmap, ufi.orig_rvaddr, VM_PAGE_TO_PHYS(pg),
1665 	    pg->flags & PG_RDONLY ? enter_prot & ~VM_PROT_WRITE : enter_prot,
1666 	    access_type | PMAP_CANFAIL | (wired ? PMAP_WIRED : 0)) != 0) {
1667 
1668 		/*
1669 		 * No need to undo what we did; we can simply think of
1670 		 * this as the pmap throwing away the mapping information.
1671 		 *
1672 		 * We do, however, have to go through the ReFault path,
1673 		 * as the map may change while we're asleep.
1674 		 */
1675 
1676 		if (pg->flags & PG_WANTED)
1677 			wakeup(pg);
1678 
1679 		/*
1680 		 * note that pg can't be PG_RELEASED since we did not drop
1681 		 * the object lock since the last time we checked.
1682 		 */
1683 
1684 		pg->flags &= ~(PG_BUSY|PG_FAKE|PG_WANTED);
1685 		UVM_PAGE_OWN(pg, NULL);
1686 		uvmfault_unlockall(&ufi, amap, uobj, anon);
1687 		KASSERT(uvmexp.swpgonly <= uvmexp.swpages);
1688 		if (uvmexp.swpgonly == uvmexp.swpages) {
1689 			UVMHIST_LOG(maphist,
1690 			    "<- failed.  out of VM",0,0,0,0);
1691 			/* XXX instrumentation */
1692 			return ENOMEM;
1693 		}
1694 		/* XXX instrumentation */
1695 		uvm_wait("flt_pmfail2");
1696 		goto ReFault;
1697 	}
1698 
1699 	uvm_lock_pageq();
1700 	if (wire_fault) {
1701 		uvm_pagewire(pg);
1702 		if (pg->pqflags & PQ_AOBJ) {
1703 
1704 			/*
1705 			 * since the now-wired page cannot be paged out,
1706 			 * release its swap resources for others to use.
1707 			 * since an aobj page with no swap cannot be PG_CLEAN,
1708 			 * clear its clean flag now.
1709 			 */
1710 
1711 			pg->flags &= ~(PG_CLEAN);
1712 			uao_dropswap(uobj, pg->offset >> PAGE_SHIFT);
1713 		}
1714 	} else {
1715 		uvm_pageactivate(pg);
1716 	}
1717 	uvm_unlock_pageq();
1718 	if (pg->flags & PG_WANTED)
1719 		wakeup(pg);
1720 
1721 	/*
1722 	 * note that pg can't be PG_RELEASED since we did not drop the object
1723 	 * lock since the last time we checked.
1724 	 */
1725 
1726 	pg->flags &= ~(PG_BUSY|PG_FAKE|PG_WANTED);
1727 	UVM_PAGE_OWN(pg, NULL);
1728 	uvmfault_unlockall(&ufi, amap, uobj, anon);
1729 	pmap_update(ufi.orig_map->pmap);
1730 	UVMHIST_LOG(maphist, "<- done (SUCCESS!)",0,0,0,0);
1731 	return 0;
1732 }
1733 
1734 /*
1735  * uvm_fault_wire: wire down a range of virtual addresses in a map.
1736  *
1737  * => map may be read-locked by caller, but MUST NOT be write-locked.
1738  * => if map is read-locked, any operations which may cause map to
1739  *	be write-locked in uvm_fault() must be taken care of by
1740  *	the caller.  See uvm_map_pageable().
1741  */
1742 
1743 int
1744 uvm_fault_wire(map, start, end, fault_type, access_type)
1745 	struct vm_map *map;
1746 	vaddr_t start, end;
1747 	vm_fault_t fault_type;
1748 	vm_prot_t access_type;
1749 {
1750 	vaddr_t va;
1751 	int error;
1752 
1753 	/*
1754 	 * now fault it in a page at a time.   if the fault fails then we have
1755 	 * to undo what we have done.   note that in uvm_fault VM_PROT_NONE
1756 	 * is replaced with the max protection if fault_type is VM_FAULT_WIRE.
1757 	 */
1758 
1759 	/*
1760 	 * XXX work around overflowing a vaddr_t.  this prevents us from
1761 	 * wiring the last page in the address space, though.
1762 	 */
1763 	if (start > end) {
1764 		return EFAULT;
1765 	}
1766 
1767 	for (va = start ; va < end ; va += PAGE_SIZE) {
1768 		error = uvm_fault(map, va, fault_type, access_type);
1769 		if (error) {
1770 			if (va != start) {
1771 				uvm_fault_unwire(map, start, va);
1772 			}
1773 			return error;
1774 		}
1775 	}
1776 	return 0;
1777 }
1778 
1779 /*
1780  * uvm_fault_unwire(): unwire range of virtual space.
1781  */
1782 
1783 void
1784 uvm_fault_unwire(map, start, end)
1785 	struct vm_map *map;
1786 	vaddr_t start, end;
1787 {
1788 	vm_map_lock_read(map);
1789 	uvm_fault_unwire_locked(map, start, end);
1790 	vm_map_unlock_read(map);
1791 }
1792 
1793 /*
1794  * uvm_fault_unwire_locked(): the guts of uvm_fault_unwire().
1795  *
1796  * => map must be at least read-locked.
1797  */
1798 
1799 void
1800 uvm_fault_unwire_locked(map, start, end)
1801 	struct vm_map *map;
1802 	vaddr_t start, end;
1803 {
1804 	struct vm_map_entry *entry;
1805 	pmap_t pmap = vm_map_pmap(map);
1806 	vaddr_t va;
1807 	paddr_t pa;
1808 	struct vm_page *pg;
1809 
1810 	KASSERT((map->flags & VM_MAP_INTRSAFE) == 0);
1811 
1812 	/*
1813 	 * we assume that the area we are unwiring has actually been wired
1814 	 * in the first place.   this means that we should be able to extract
1815 	 * the PAs from the pmap.   we also lock out the page daemon so that
1816 	 * we can call uvm_pageunwire.
1817 	 */
1818 
1819 	uvm_lock_pageq();
1820 
1821 	/*
1822 	 * find the beginning map entry for the region.
1823 	 */
1824 
1825 	KASSERT(start >= vm_map_min(map) && end <= vm_map_max(map));
1826 	if (uvm_map_lookup_entry(map, start, &entry) == FALSE)
1827 		panic("uvm_fault_unwire_locked: address not in map");
1828 
1829 	for (va = start; va < end; va += PAGE_SIZE) {
1830 		if (pmap_extract(pmap, va, &pa) == FALSE)
1831 			continue;
1832 
1833 		/*
1834 		 * find the map entry for the current address.
1835 		 */
1836 
1837 		KASSERT(va >= entry->start);
1838 		while (va >= entry->end) {
1839 			KASSERT(entry->next != &map->header &&
1840 				entry->next->start <= entry->end);
1841 			entry = entry->next;
1842 		}
1843 
1844 		/*
1845 		 * if the entry is no longer wired, tell the pmap.
1846 		 */
1847 
1848 		if (VM_MAPENT_ISWIRED(entry) == 0)
1849 			pmap_unwire(pmap, va);
1850 
1851 		pg = PHYS_TO_VM_PAGE(pa);
1852 		if (pg)
1853 			uvm_pageunwire(pg);
1854 	}
1855 
1856 	uvm_unlock_pageq();
1857 }
1858