xref: /netbsd-src/sys/uvm/uvm_fault.c (revision 3b01aba77a7a698587faaae455bbfe740923c1f5)
1 /*	$NetBSD: uvm_fault.c,v 1.67 2001/06/26 17:55:14 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 #include "opt_uvmhist.h"
38 
39 /*
40  * uvm_fault.c: fault handler
41  */
42 
43 #include <sys/param.h>
44 #include <sys/systm.h>
45 #include <sys/kernel.h>
46 #include <sys/proc.h>
47 #include <sys/malloc.h>
48 #include <sys/mman.h>
49 #include <sys/user.h>
50 
51 #include <uvm/uvm.h>
52 
53 /*
54  *
55  * a word on page faults:
56  *
57  * types of page faults we handle:
58  *
59  * CASE 1: upper layer faults                   CASE 2: lower layer faults
60  *
61  *    CASE 1A         CASE 1B                  CASE 2A        CASE 2B
62  *    read/write1     write>1                  read/write   +-cow_write/zero
63  *         |             |                         |        |
64  *      +--|--+       +--|--+     +-----+       +  |  +     | +-----+
65  * amap |  V  |       |  ----------->new|          |        | |  ^  |
66  *      +-----+       +-----+     +-----+       +  |  +     | +--|--+
67  *                                                 |        |    |
68  *      +-----+       +-----+                   +--|--+     | +--|--+
69  * uobj | d/c |       | d/c |                   |  V  |     +----|  |
70  *      +-----+       +-----+                   +-----+       +-----+
71  *
72  * d/c = don't care
73  *
74  *   case [0]: layerless fault
75  *	no amap or uobj is present.   this is an error.
76  *
77  *   case [1]: upper layer fault [anon active]
78  *     1A: [read] or [write with anon->an_ref == 1]
79  *		I/O takes place in top level anon and uobj is not touched.
80  *     1B: [write with anon->an_ref > 1]
81  *		new anon is alloc'd and data is copied off ["COW"]
82  *
83  *   case [2]: lower layer fault [uobj]
84  *     2A: [read on non-NULL uobj] or [write to non-copy_on_write area]
85  *		I/O takes place directly in object.
86  *     2B: [write to copy_on_write] or [read on NULL uobj]
87  *		data is "promoted" from uobj to a new anon.
88  *		if uobj is null, then we zero fill.
89  *
90  * we follow the standard UVM locking protocol ordering:
91  *
92  * MAPS => AMAP => UOBJ => ANON => PAGE QUEUES (PQ)
93  * we hold a PG_BUSY page if we unlock for I/O
94  *
95  *
96  * the code is structured as follows:
97  *
98  *     - init the "IN" params in the ufi structure
99  *   ReFault:
100  *     - do lookups [locks maps], check protection, handle needs_copy
101  *     - check for case 0 fault (error)
102  *     - establish "range" of fault
103  *     - if we have an amap lock it and extract the anons
104  *     - if sequential advice deactivate pages behind us
105  *     - at the same time check pmap for unmapped areas and anon for pages
106  *	 that we could map in (and do map it if found)
107  *     - check object for resident pages that we could map in
108  *     - if (case 2) goto Case2
109  *     - >>> handle case 1
110  *           - ensure source anon is resident in RAM
111  *           - if case 1B alloc new anon and copy from source
112  *           - map the correct page in
113  *   Case2:
114  *     - >>> handle case 2
115  *           - ensure source page is resident (if uobj)
116  *           - if case 2B alloc new anon and copy from source (could be zero
117  *		fill if uobj == NULL)
118  *           - map the correct page in
119  *     - done!
120  *
121  * note on paging:
122  *   if we have to do I/O we place a PG_BUSY page in the correct object,
123  * unlock everything, and do the I/O.   when I/O is done we must reverify
124  * the state of the world before assuming that our data structures are
125  * valid.   [because mappings could change while the map is unlocked]
126  *
127  *  alternative 1: unbusy the page in question and restart the page fault
128  *    from the top (ReFault).   this is easy but does not take advantage
129  *    of the information that we already have from our previous lookup,
130  *    although it is possible that the "hints" in the vm_map will help here.
131  *
132  * alternative 2: the system already keeps track of a "version" number of
133  *    a map.   [i.e. every time you write-lock a map (e.g. to change a
134  *    mapping) you bump the version number up by one...]   so, we can save
135  *    the version number of the map before we release the lock and start I/O.
136  *    then when I/O is done we can relock and check the version numbers
137  *    to see if anything changed.    this might save us some over 1 because
138  *    we don't have to unbusy the page and may be less compares(?).
139  *
140  * alternative 3: put in backpointers or a way to "hold" part of a map
141  *    in place while I/O is in progress.   this could be complex to
142  *    implement (especially with structures like amap that can be referenced
143  *    by multiple map entries, and figuring out what should wait could be
144  *    complex as well...).
145  *
146  * given that we are not currently multiprocessor or multithreaded we might
147  * as well choose alternative 2 now.   maybe alternative 3 would be useful
148  * in the future.    XXX keep in mind for future consideration//rechecking.
149  */
150 
151 /*
152  * local data structures
153  */
154 
155 struct uvm_advice {
156 	int advice;
157 	int nback;
158 	int nforw;
159 };
160 
161 /*
162  * page range array:
163  * note: index in array must match "advice" value
164  * XXX: borrowed numbers from freebsd.   do they work well for us?
165  */
166 
167 static struct uvm_advice uvmadvice[] = {
168 	{ MADV_NORMAL, 3, 4 },
169 	{ MADV_RANDOM, 0, 0 },
170 	{ MADV_SEQUENTIAL, 8, 7},
171 };
172 
173 #define UVM_MAXRANGE 16	/* must be max() of nback+nforw+1 */
174 
175 /*
176  * private prototypes
177  */
178 
179 static void uvmfault_amapcopy __P((struct uvm_faultinfo *));
180 static __inline void uvmfault_anonflush __P((struct vm_anon **, int));
181 
182 /*
183  * inline functions
184  */
185 
186 /*
187  * uvmfault_anonflush: try and deactivate pages in specified anons
188  *
189  * => does not have to deactivate page if it is busy
190  */
191 
192 static __inline void
193 uvmfault_anonflush(anons, n)
194 	struct vm_anon **anons;
195 	int n;
196 {
197 	int lcv;
198 	struct vm_page *pg;
199 
200 	for (lcv = 0 ; lcv < n ; lcv++) {
201 		if (anons[lcv] == NULL)
202 			continue;
203 		simple_lock(&anons[lcv]->an_lock);
204 		pg = anons[lcv]->u.an_page;
205 		if (pg && (pg->flags & PG_BUSY) == 0 && pg->loan_count == 0) {
206 			uvm_lock_pageq();
207 			if (pg->wire_count == 0) {
208 				pmap_clear_reference(pg);
209 				uvm_pagedeactivate(pg);
210 			}
211 			uvm_unlock_pageq();
212 		}
213 		simple_unlock(&anons[lcv]->an_lock);
214 	}
215 }
216 
217 /*
218  * normal functions
219  */
220 
221 /*
222  * uvmfault_amapcopy: clear "needs_copy" in a map.
223  *
224  * => called with VM data structures unlocked (usually, see below)
225  * => we get a write lock on the maps and clear needs_copy for a VA
226  * => if we are out of RAM we sleep (waiting for more)
227  */
228 
229 static void
230 uvmfault_amapcopy(ufi)
231 	struct uvm_faultinfo *ufi;
232 {
233 
234 	/*
235 	 * while we haven't done the job
236 	 */
237 
238 	while (1) {
239 
240 		/*
241 		 * no mapping?  give up.
242 		 */
243 
244 		if (uvmfault_lookup(ufi, TRUE) == FALSE)
245 			return;
246 
247 		/*
248 		 * copy if needed.
249 		 */
250 
251 		if (UVM_ET_ISNEEDSCOPY(ufi->entry))
252 			amap_copy(ufi->map, ufi->entry, M_NOWAIT, TRUE,
253 				ufi->orig_rvaddr, ufi->orig_rvaddr + 1);
254 
255 		/*
256 		 * didn't work?  must be out of RAM.   unlock and sleep.
257 		 */
258 
259 		if (UVM_ET_ISNEEDSCOPY(ufi->entry)) {
260 			uvmfault_unlockmaps(ufi, TRUE);
261 			uvm_wait("fltamapcopy");
262 			continue;
263 		}
264 
265 		/*
266 		 * got it!   unlock and return.
267 		 */
268 
269 		uvmfault_unlockmaps(ufi, TRUE);
270 		return;
271 	}
272 	/*NOTREACHED*/
273 }
274 
275 /*
276  * uvmfault_anonget: get data in an anon into a non-busy, non-released
277  * page in that anon.
278  *
279  * => maps, amap, and anon locked by caller.
280  * => if we fail (result != 0) we unlock everything.
281  * => if we are successful, we return with everything still locked.
282  * => we don't move the page on the queues [gets moved later]
283  * => if we allocate a new page [we_own], it gets put on the queues.
284  *    either way, the result is that the page is on the queues at return time
285  * => for pages which are on loan from a uvm_object (and thus are not
286  *    owned by the anon): if successful, we return with the owning object
287  *    locked.   the caller must unlock this object when it unlocks everything
288  *    else.
289  */
290 
291 int
292 uvmfault_anonget(ufi, amap, anon)
293 	struct uvm_faultinfo *ufi;
294 	struct vm_amap *amap;
295 	struct vm_anon *anon;
296 {
297 	boolean_t we_own;	/* we own anon's page? */
298 	boolean_t locked;	/* did we relock? */
299 	struct vm_page *pg;
300 	int error;
301 	UVMHIST_FUNC("uvmfault_anonget"); UVMHIST_CALLED(maphist);
302 
303 	LOCK_ASSERT(simple_lock_held(&anon->an_lock));
304 
305 	error = 0;
306 	uvmexp.fltanget++;
307         /* bump rusage counters */
308 	if (anon->u.an_page)
309 		curproc->p_addr->u_stats.p_ru.ru_minflt++;
310 	else
311 		curproc->p_addr->u_stats.p_ru.ru_majflt++;
312 
313 	/*
314 	 * loop until we get it, or fail.
315 	 */
316 
317 	while (1) {
318 
319 		we_own = FALSE;		/* TRUE if we set PG_BUSY on a page */
320 		pg = anon->u.an_page;
321 
322 		/*
323 		 * if there is a resident page and it is loaned, then anon
324 		 * may not own it.   call out to uvm_anon_lockpage() to ensure
325 		 * the real owner of the page has been identified and locked.
326 		 */
327 
328 		if (pg && pg->loan_count)
329 			pg = uvm_anon_lockloanpg(anon);
330 
331 		/*
332 		 * page there?   make sure it is not busy/released.
333 		 */
334 
335 		if (pg) {
336 
337 			/*
338 			 * at this point, if the page has a uobject [meaning
339 			 * we have it on loan], then that uobject is locked
340 			 * by us!   if the page is busy, we drop all the
341 			 * locks (including uobject) and try again.
342 			 */
343 
344 			if ((pg->flags & (PG_BUSY|PG_RELEASED)) == 0) {
345 				UVMHIST_LOG(maphist, "<- OK",0,0,0,0);
346 				return (0);
347 			}
348 			pg->flags |= PG_WANTED;
349 			uvmexp.fltpgwait++;
350 
351 			/*
352 			 * the last unlock must be an atomic unlock+wait on
353 			 * the owner of page
354 			 */
355 			if (pg->uobject) {	/* owner is uobject ? */
356 				uvmfault_unlockall(ufi, amap, NULL, anon);
357 				UVMHIST_LOG(maphist, " unlock+wait on uobj",0,
358 				    0,0,0);
359 				UVM_UNLOCK_AND_WAIT(pg,
360 				    &pg->uobject->vmobjlock,
361 				    FALSE, "anonget1",0);
362 			} else {
363 				/* anon owns page */
364 				uvmfault_unlockall(ufi, amap, NULL, NULL);
365 				UVMHIST_LOG(maphist, " unlock+wait on anon",0,
366 				    0,0,0);
367 				UVM_UNLOCK_AND_WAIT(pg,&anon->an_lock,0,
368 				    "anonget2",0);
369 			}
370 			/* ready to relock and try again */
371 
372 		} else {
373 
374 			/*
375 			 * no page, we must try and bring it in.
376 			 */
377 			pg = uvm_pagealloc(NULL, 0, anon, 0);
378 
379 			if (pg == NULL) {		/* out of RAM.  */
380 
381 				uvmfault_unlockall(ufi, amap, NULL, anon);
382 				uvmexp.fltnoram++;
383 				UVMHIST_LOG(maphist, "  noram -- UVM_WAIT",0,
384 				    0,0,0);
385 				uvm_wait("flt_noram1");
386 				/* ready to relock and try again */
387 
388 			} else {
389 
390 				/* we set the PG_BUSY bit */
391 				we_own = TRUE;
392 				uvmfault_unlockall(ufi, amap, NULL, anon);
393 
394 				/*
395 				 * we are passing a PG_BUSY+PG_FAKE+PG_CLEAN
396 				 * page into the uvm_swap_get function with
397 				 * all data structures unlocked.  note that
398 				 * it is ok to read an_swslot here because
399 				 * we hold PG_BUSY on the page.
400 				 */
401 				uvmexp.pageins++;
402 				error = uvm_swap_get(pg, anon->an_swslot,
403 				    PGO_SYNCIO);
404 
405 				/*
406 				 * we clean up after the i/o below in the
407 				 * "we_own" case
408 				 */
409 				/* ready to relock and try again */
410 			}
411 		}
412 
413 		/*
414 		 * now relock and try again
415 		 */
416 
417 		locked = uvmfault_relock(ufi);
418 		if (locked && amap != NULL) {
419 			amap_lock(amap);
420 		}
421 		if (locked || we_own)
422 			simple_lock(&anon->an_lock);
423 
424 		/*
425 		 * if we own the page (i.e. we set PG_BUSY), then we need
426 		 * to clean up after the I/O. there are three cases to
427 		 * consider:
428 		 *   [1] page released during I/O: free anon and ReFault.
429 		 *   [2] I/O not OK.   free the page and cause the fault
430 		 *       to fail.
431 		 *   [3] I/O OK!   activate the page and sync with the
432 		 *       non-we_own case (i.e. drop anon lock if not locked).
433 		 */
434 
435 		if (we_own) {
436 
437 			if (pg->flags & PG_WANTED) {
438 				/* still holding object lock */
439 				wakeup(pg);
440 			}
441 			/* un-busy! */
442 			pg->flags &= ~(PG_WANTED|PG_BUSY|PG_FAKE);
443 			UVM_PAGE_OWN(pg, NULL);
444 
445 			/*
446 			 * if we were RELEASED during I/O, then our anon is
447 			 * no longer part of an amap.   we need to free the
448 			 * anon and try again.
449 			 */
450 			if (pg->flags & PG_RELEASED) {
451 				pmap_page_protect(pg, VM_PROT_NONE);
452 				simple_unlock(&anon->an_lock);
453 				uvm_anfree(anon);	/* frees page for us */
454 				if (locked)
455 					uvmfault_unlockall(ufi, amap, NULL,
456 							   NULL);
457 				uvmexp.fltpgrele++;
458 				UVMHIST_LOG(maphist, "<- REFAULT", 0,0,0,0);
459 				return (ERESTART);	/* refault! */
460 			}
461 
462 			if (error) {
463 				/* remove page from anon */
464 				anon->u.an_page = NULL;
465 
466 				/*
467 				 * remove the swap slot from the anon
468 				 * and mark the anon as having no real slot.
469 				 * don't free the swap slot, thus preventing
470 				 * it from being used again.
471 				 */
472 				uvm_swap_markbad(anon->an_swslot, 1);
473 				anon->an_swslot = SWSLOT_BAD;
474 
475 				/*
476 				 * note: page was never !PG_BUSY, so it
477 				 * can't be mapped and thus no need to
478 				 * pmap_page_protect it...
479 				 */
480 				uvm_lock_pageq();
481 				uvm_pagefree(pg);
482 				uvm_unlock_pageq();
483 
484 				if (locked)
485 					uvmfault_unlockall(ufi, amap, NULL,
486 					    anon);
487 				else
488 					simple_unlock(&anon->an_lock);
489 				UVMHIST_LOG(maphist, "<- ERROR", 0,0,0,0);
490 				return error;
491 			}
492 
493 			/*
494 			 * must be OK, clear modify (already PG_CLEAN)
495 			 * and activate
496 			 */
497 			pmap_clear_modify(pg);
498 			uvm_lock_pageq();
499 			uvm_pageactivate(pg);
500 			uvm_unlock_pageq();
501 			if (!locked)
502 				simple_unlock(&anon->an_lock);
503 		}
504 
505 		/*
506 		 * we were not able to relock.   restart fault.
507 		 */
508 
509 		if (!locked) {
510 			UVMHIST_LOG(maphist, "<- REFAULT", 0,0,0,0);
511 			return (ERESTART);
512 		}
513 
514 		/*
515 		 * verify no one has touched the amap and moved the anon on us.
516 		 */
517 
518 		if (ufi != NULL &&
519 		    amap_lookup(&ufi->entry->aref,
520 				ufi->orig_rvaddr - ufi->entry->start) != anon) {
521 
522 			uvmfault_unlockall(ufi, amap, NULL, anon);
523 			UVMHIST_LOG(maphist, "<- REFAULT", 0,0,0,0);
524 			return (ERESTART);
525 		}
526 
527 		/*
528 		 * try it again!
529 		 */
530 
531 		uvmexp.fltanretry++;
532 		continue;
533 
534 	} /* while (1) */
535 
536 	/*NOTREACHED*/
537 }
538 
539 /*
540  *   F A U L T   -   m a i n   e n t r y   p o i n t
541  */
542 
543 /*
544  * uvm_fault: page fault handler
545  *
546  * => called from MD code to resolve a page fault
547  * => VM data structures usually should be unlocked.   however, it is
548  *	possible to call here with the main map locked if the caller
549  *	gets a write lock, sets it recusive, and then calls us (c.f.
550  *	uvm_map_pageable).   this should be avoided because it keeps
551  *	the map locked off during I/O.
552  * => MUST NEVER BE CALLED IN INTERRUPT CONTEXT
553  */
554 
555 #define MASK(entry)     (UVM_ET_ISCOPYONWRITE(entry) ? \
556 			 ~VM_PROT_WRITE : VM_PROT_ALL)
557 
558 int
559 uvm_fault(orig_map, vaddr, fault_type, access_type)
560 	struct vm_map *orig_map;
561 	vaddr_t vaddr;
562 	vm_fault_t fault_type;
563 	vm_prot_t access_type;
564 {
565 	struct uvm_faultinfo ufi;
566 	vm_prot_t enter_prot;
567 	boolean_t wired, narrow, promote, locked, shadowed;
568 	int npages, nback, nforw, centeridx, error, lcv, gotpages;
569 	vaddr_t startva, objaddr, currva, offset, uoff;
570 	paddr_t pa;
571 	struct vm_amap *amap;
572 	struct uvm_object *uobj;
573 	struct vm_anon *anons_store[UVM_MAXRANGE], **anons, *anon, *oanon;
574 	struct vm_page *pages[UVM_MAXRANGE], *pg, *uobjpage;
575 	UVMHIST_FUNC("uvm_fault"); UVMHIST_CALLED(maphist);
576 
577 	UVMHIST_LOG(maphist, "(map=0x%x, vaddr=0x%x, ft=%d, at=%d)",
578 	      orig_map, vaddr, fault_type, access_type);
579 
580 	anon = NULL;
581 	pg = NULL;
582 
583 	uvmexp.faults++;	/* XXX: locking? */
584 
585 	/*
586 	 * init the IN parameters in the ufi
587 	 */
588 
589 	ufi.orig_map = orig_map;
590 	ufi.orig_rvaddr = trunc_page(vaddr);
591 	ufi.orig_size = PAGE_SIZE;	/* can't get any smaller than this */
592 	if (fault_type == VM_FAULT_WIRE)
593 		narrow = TRUE;		/* don't look for neighborhood
594 					 * pages on wire */
595 	else
596 		narrow = FALSE;		/* normal fault */
597 
598 	/*
599 	 * "goto ReFault" means restart the page fault from ground zero.
600 	 */
601 ReFault:
602 
603 	/*
604 	 * lookup and lock the maps
605 	 */
606 
607 	if (uvmfault_lookup(&ufi, FALSE) == FALSE) {
608 		UVMHIST_LOG(maphist, "<- no mapping @ 0x%x", vaddr, 0,0,0);
609 		return (EFAULT);
610 	}
611 	/* locked: maps(read) */
612 
613 #ifdef DIAGNOSTIC
614 	if ((ufi.map->flags & VM_MAP_PAGEABLE) == 0) {
615 		printf("Page fault on non-pageable map:\n");
616 		printf("ufi.map = %p\n", ufi.map);
617 		printf("ufi.orig_map = %p\n", ufi.orig_map);
618 		printf("ufi.orig_rvaddr = 0x%lx\n", (u_long) ufi.orig_rvaddr);
619 		panic("uvm_fault: (ufi.map->flags & VM_MAP_PAGEABLE) == 0");
620 	}
621 #endif
622 
623 	/*
624 	 * check protection
625 	 */
626 
627 	if ((ufi.entry->protection & access_type) != access_type) {
628 		UVMHIST_LOG(maphist,
629 		    "<- protection failure (prot=0x%x, access=0x%x)",
630 		    ufi.entry->protection, access_type, 0, 0);
631 		uvmfault_unlockmaps(&ufi, FALSE);
632 		return EACCES;
633 	}
634 
635 	/*
636 	 * "enter_prot" is the protection we want to enter the page in at.
637 	 * for certain pages (e.g. copy-on-write pages) this protection can
638 	 * be more strict than ufi.entry->protection.  "wired" means either
639 	 * the entry is wired or we are fault-wiring the pg.
640 	 */
641 
642 	enter_prot = ufi.entry->protection;
643 	wired = VM_MAPENT_ISWIRED(ufi.entry) || (fault_type == VM_FAULT_WIRE);
644 	if (wired)
645 		access_type = enter_prot; /* full access for wired */
646 
647 	/*
648 	 * handle "needs_copy" case.   if we need to copy the amap we will
649 	 * have to drop our readlock and relock it with a write lock.  (we
650 	 * need a write lock to change anything in a map entry [e.g.
651 	 * needs_copy]).
652 	 */
653 
654 	if (UVM_ET_ISNEEDSCOPY(ufi.entry)) {
655 		if ((access_type & VM_PROT_WRITE) ||
656 		    (ufi.entry->object.uvm_obj == NULL)) {
657 			/* need to clear */
658 			UVMHIST_LOG(maphist,
659 			    "  need to clear needs_copy and refault",0,0,0,0);
660 			uvmfault_unlockmaps(&ufi, FALSE);
661 			uvmfault_amapcopy(&ufi);
662 			uvmexp.fltamcopy++;
663 			goto ReFault;
664 
665 		} else {
666 
667 			/*
668 			 * ensure that we pmap_enter page R/O since
669 			 * needs_copy is still true
670 			 */
671 			enter_prot &= ~VM_PROT_WRITE;
672 
673 		}
674 	}
675 
676 	/*
677 	 * identify the players
678 	 */
679 
680 	amap = ufi.entry->aref.ar_amap;		/* top layer */
681 	uobj = ufi.entry->object.uvm_obj;	/* bottom layer */
682 
683 	/*
684 	 * check for a case 0 fault.  if nothing backing the entry then
685 	 * error now.
686 	 */
687 
688 	if (amap == NULL && uobj == NULL) {
689 		uvmfault_unlockmaps(&ufi, FALSE);
690 		UVMHIST_LOG(maphist,"<- no backing store, no overlay",0,0,0,0);
691 		return (EFAULT);
692 	}
693 
694 	/*
695 	 * establish range of interest based on advice from mapper
696 	 * and then clip to fit map entry.   note that we only want
697 	 * to do this the first time through the fault.   if we
698 	 * ReFault we will disable this by setting "narrow" to true.
699 	 */
700 
701 	if (narrow == FALSE) {
702 
703 		/* wide fault (!narrow) */
704 		KASSERT(uvmadvice[ufi.entry->advice].advice ==
705 			 ufi.entry->advice);
706 		nback = min(uvmadvice[ufi.entry->advice].nback,
707 			    (ufi.orig_rvaddr - ufi.entry->start) >> PAGE_SHIFT);
708 		startva = ufi.orig_rvaddr - (nback << PAGE_SHIFT);
709 		nforw = min(uvmadvice[ufi.entry->advice].nforw,
710 			    ((ufi.entry->end - ufi.orig_rvaddr) >>
711 			     PAGE_SHIFT) - 1);
712 		/*
713 		 * note: "-1" because we don't want to count the
714 		 * faulting page as forw
715 		 */
716 		npages = nback + nforw + 1;
717 		centeridx = nback;
718 
719 		narrow = TRUE;	/* ensure only once per-fault */
720 
721 	} else {
722 
723 		/* narrow fault! */
724 		nback = nforw = 0;
725 		startva = ufi.orig_rvaddr;
726 		npages = 1;
727 		centeridx = 0;
728 
729 	}
730 
731 	/* locked: maps(read) */
732 	UVMHIST_LOG(maphist, "  narrow=%d, back=%d, forw=%d, startva=0x%x",
733 		    narrow, nback, nforw, startva);
734 	UVMHIST_LOG(maphist, "  entry=0x%x, amap=0x%x, obj=0x%x", ufi.entry,
735 		    amap, uobj, 0);
736 
737 	/*
738 	 * if we've got an amap, lock it and extract current anons.
739 	 */
740 
741 	if (amap) {
742 		amap_lock(amap);
743 		anons = anons_store;
744 		amap_lookups(&ufi.entry->aref, startva - ufi.entry->start,
745 		    anons, npages);
746 	} else {
747 		anons = NULL;	/* to be safe */
748 	}
749 
750 	/* locked: maps(read), amap(if there) */
751 
752 	/*
753 	 * for MADV_SEQUENTIAL mappings we want to deactivate the back pages
754 	 * now and then forget about them (for the rest of the fault).
755 	 */
756 
757 	if (ufi.entry->advice == MADV_SEQUENTIAL) {
758 
759 		UVMHIST_LOG(maphist, "  MADV_SEQUENTIAL: flushing backpages",
760 		    0,0,0,0);
761 		/* flush back-page anons? */
762 		if (amap)
763 			uvmfault_anonflush(anons, nback);
764 
765 		/* flush object? */
766 		if (uobj) {
767 			objaddr =
768 			    (startva - ufi.entry->start) + ufi.entry->offset;
769 			simple_lock(&uobj->vmobjlock);
770 			(void) uobj->pgops->pgo_flush(uobj, objaddr, objaddr +
771 				    (nback << PAGE_SHIFT), PGO_DEACTIVATE);
772 			simple_unlock(&uobj->vmobjlock);
773 		}
774 
775 		/* now forget about the backpages */
776 		if (amap)
777 			anons += nback;
778 		startva += (nback << PAGE_SHIFT);
779 		npages -= nback;
780 		nback = centeridx = 0;
781 	}
782 
783 	/* locked: maps(read), amap(if there) */
784 
785 	/*
786 	 * map in the backpages and frontpages we found in the amap in hopes
787 	 * of preventing future faults.    we also init the pages[] array as
788 	 * we go.
789 	 */
790 
791 	currva = startva;
792 	shadowed = FALSE;
793 	for (lcv = 0 ; lcv < npages ; lcv++, currva += PAGE_SIZE) {
794 
795 		/*
796 		 * dont play with VAs that are already mapped
797 		 * except for center)
798 		 */
799 		if (lcv != centeridx &&
800 		    pmap_extract(ufi.orig_map->pmap, currva, &pa)) {
801 			pages[lcv] = PGO_DONTCARE;
802 			continue;
803 		}
804 
805 		/*
806 		 * unmapped or center page.   check if any anon at this level.
807 		 */
808 		if (amap == NULL || anons[lcv] == NULL) {
809 			pages[lcv] = NULL;
810 			continue;
811 		}
812 
813 		/*
814 		 * check for present page and map if possible.   re-activate it.
815 		 */
816 
817 		pages[lcv] = PGO_DONTCARE;
818 		if (lcv == centeridx) {		/* save center for later! */
819 			shadowed = TRUE;
820 			continue;
821 		}
822 		anon = anons[lcv];
823 		simple_lock(&anon->an_lock);
824 		/* ignore loaned pages */
825 		if (anon->u.an_page && anon->u.an_page->loan_count == 0 &&
826 			(anon->u.an_page->flags & (PG_RELEASED|PG_BUSY)) == 0) {
827 			uvm_lock_pageq();
828 			uvm_pageactivate(anon->u.an_page);	/* reactivate */
829 			uvm_unlock_pageq();
830 			UVMHIST_LOG(maphist,
831 			    "  MAPPING: n anon: pm=0x%x, va=0x%x, pg=0x%x",
832 			    ufi.orig_map->pmap, currva, anon->u.an_page, 0);
833 			uvmexp.fltnamap++;
834 
835 			/*
836 			 * Since this isn't the page that's actually faulting,
837 			 * ignore pmap_enter() failures; it's not critical
838 			 * that we enter these right now.
839 			 */
840 
841 			(void) pmap_enter(ufi.orig_map->pmap, currva,
842 			    VM_PAGE_TO_PHYS(anon->u.an_page),
843 			    (anon->an_ref > 1) ? (enter_prot & ~VM_PROT_WRITE) :
844 			    enter_prot,
845 			    PMAP_CANFAIL |
846 			     (VM_MAPENT_ISWIRED(ufi.entry) ? PMAP_WIRED : 0));
847 		}
848 		simple_unlock(&anon->an_lock);
849 		pmap_update();
850 	}
851 
852 	/* locked: maps(read), amap(if there) */
853 	/* (shadowed == TRUE) if there is an anon at the faulting address */
854 	UVMHIST_LOG(maphist, "  shadowed=%d, will_get=%d", shadowed,
855 	    (uobj && shadowed == FALSE),0,0);
856 
857 	/*
858 	 * note that if we are really short of RAM we could sleep in the above
859 	 * call to pmap_enter with everything locked.   bad?
860 	 *
861 	 * XXX Actually, that is bad; pmap_enter() should just fail in that
862 	 * XXX case.  --thorpej
863 	 */
864 
865 	/*
866 	 * if the desired page is not shadowed by the amap and we have a
867 	 * backing object, then we check to see if the backing object would
868 	 * prefer to handle the fault itself (rather than letting us do it
869 	 * with the usual pgo_get hook).  the backing object signals this by
870 	 * providing a pgo_fault routine.
871 	 */
872 
873 	if (uobj && shadowed == FALSE && uobj->pgops->pgo_fault != NULL) {
874 		simple_lock(&uobj->vmobjlock);
875 
876 		/* locked: maps(read), amap (if there), uobj */
877 		error = uobj->pgops->pgo_fault(&ufi, startva, pages, npages,
878 		    centeridx, fault_type, access_type, PGO_LOCKED|PGO_SYNCIO);
879 
880 		/* locked: nothing, pgo_fault has unlocked everything */
881 
882 		if (error == ERESTART)
883 			goto ReFault;		/* try again! */
884 		/*
885 		 * object fault routine responsible for pmap_update().
886 		 */
887 		return error;
888 	}
889 
890 	/*
891 	 * now, if the desired page is not shadowed by the amap and we have
892 	 * a backing object that does not have a special fault routine, then
893 	 * we ask (with pgo_get) the object for resident pages that we care
894 	 * about and attempt to map them in.  we do not let pgo_get block
895 	 * (PGO_LOCKED).
896 	 *
897 	 * ("get" has the option of doing a pmap_enter for us)
898 	 */
899 
900 	if (uobj && shadowed == FALSE) {
901 		simple_lock(&uobj->vmobjlock);
902 
903 		/* locked (!shadowed): maps(read), amap (if there), uobj */
904 		/*
905 		 * the following call to pgo_get does _not_ change locking state
906 		 */
907 
908 		uvmexp.fltlget++;
909 		gotpages = npages;
910 		(void) uobj->pgops->pgo_get(uobj, ufi.entry->offset +
911 				(startva - ufi.entry->start),
912 				pages, &gotpages, centeridx,
913 				access_type & MASK(ufi.entry),
914 				ufi.entry->advice, PGO_LOCKED);
915 
916 		/*
917 		 * check for pages to map, if we got any
918 		 */
919 
920 		uobjpage = NULL;
921 
922 		if (gotpages) {
923 			currva = startva;
924 			for (lcv = 0 ; lcv < npages ;
925 			    lcv++, currva += PAGE_SIZE) {
926 
927 				if (pages[lcv] == NULL ||
928 				    pages[lcv] == PGO_DONTCARE)
929 					continue;
930 
931 				KASSERT((pages[lcv]->flags & PG_RELEASED) == 0);
932 
933 				/*
934 				 * if center page is resident and not
935 				 * PG_BUSY|PG_RELEASED then pgo_get
936 				 * made it PG_BUSY for us and gave
937 				 * us a handle to it.   remember this
938 				 * page as "uobjpage." (for later use).
939 				 */
940 
941 				if (lcv == centeridx) {
942 					uobjpage = pages[lcv];
943 					UVMHIST_LOG(maphist, "  got uobjpage "
944 					    "(0x%x) with locked get",
945 					    uobjpage, 0,0,0);
946 					continue;
947 				}
948 
949 				/*
950 				 * note: calling pgo_get with locked data
951 				 * structures returns us pages which are
952 				 * neither busy nor released, so we don't
953 				 * need to check for this.   we can just
954 				 * directly enter the page (after moving it
955 				 * to the head of the active queue [useful?]).
956 				 */
957 
958 				uvm_lock_pageq();
959 				uvm_pageactivate(pages[lcv]);	/* reactivate */
960 				uvm_unlock_pageq();
961 				UVMHIST_LOG(maphist,
962 				  "  MAPPING: n obj: pm=0x%x, va=0x%x, pg=0x%x",
963 				  ufi.orig_map->pmap, currva, pages[lcv], 0);
964 				uvmexp.fltnomap++;
965 
966 				/*
967 				 * Since this page isn't the page that's
968 				 * actually fauling, ignore pmap_enter()
969 				 * failures; it's not critical that we
970 				 * enter these right now.
971 				 */
972 
973 				(void) pmap_enter(ufi.orig_map->pmap, currva,
974 				    VM_PAGE_TO_PHYS(pages[lcv]),
975 				    pages[lcv]->flags & PG_RDONLY ?
976 				    VM_PROT_READ : enter_prot & MASK(ufi.entry),
977 				    PMAP_CANFAIL |
978 				     (wired ? PMAP_WIRED : 0));
979 
980 				/*
981 				 * NOTE: page can't be PG_WANTED or PG_RELEASED
982 				 * because we've held the lock the whole time
983 				 * we've had the handle.
984 				 */
985 
986 				pages[lcv]->flags &= ~(PG_BUSY); /* un-busy! */
987 				UVM_PAGE_OWN(pages[lcv], NULL);
988 			}	/* for "lcv" loop */
989 			pmap_update();
990 		}   /* "gotpages" != 0 */
991 		/* note: object still _locked_ */
992 	} else {
993 		uobjpage = NULL;
994 	}
995 
996 	/* locked (shadowed): maps(read), amap */
997 	/* locked (!shadowed): maps(read), amap(if there),
998 		 uobj(if !null), uobjpage(if !null) */
999 
1000 	/*
1001 	 * note that at this point we are done with any front or back pages.
1002 	 * we are now going to focus on the center page (i.e. the one we've
1003 	 * faulted on).  if we have faulted on the top (anon) layer
1004 	 * [i.e. case 1], then the anon we want is anons[centeridx] (we have
1005 	 * not touched it yet).  if we have faulted on the bottom (uobj)
1006 	 * layer [i.e. case 2] and the page was both present and available,
1007 	 * then we've got a pointer to it as "uobjpage" and we've already
1008 	 * made it BUSY.
1009 	 */
1010 
1011 	/*
1012 	 * there are four possible cases we must address: 1A, 1B, 2A, and 2B
1013 	 */
1014 
1015 	/*
1016 	 * redirect case 2: if we are not shadowed, go to case 2.
1017 	 */
1018 
1019 	if (shadowed == FALSE)
1020 		goto Case2;
1021 
1022 	/* locked: maps(read), amap */
1023 
1024 	/*
1025 	 * handle case 1: fault on an anon in our amap
1026 	 */
1027 
1028 	anon = anons[centeridx];
1029 	UVMHIST_LOG(maphist, "  case 1 fault: anon=0x%x", anon, 0,0,0);
1030 	simple_lock(&anon->an_lock);
1031 
1032 	/* locked: maps(read), amap, anon */
1033 
1034 	/*
1035 	 * no matter if we have case 1A or case 1B we are going to need to
1036 	 * have the anon's memory resident.   ensure that now.
1037 	 */
1038 
1039 	/*
1040 	 * let uvmfault_anonget do the dirty work.
1041 	 * if it fails (!OK) it will unlock everything for us.
1042 	 * if it succeeds, locks are still valid and locked.
1043 	 * also, if it is OK, then the anon's page is on the queues.
1044 	 * if the page is on loan from a uvm_object, then anonget will
1045 	 * lock that object for us if it does not fail.
1046 	 */
1047 
1048 	error = uvmfault_anonget(&ufi, amap, anon);
1049 	switch (error) {
1050 	case 0:
1051 		break;
1052 
1053 	case ERESTART:
1054 		goto ReFault;
1055 
1056 	case EAGAIN:
1057 		tsleep(&lbolt, PVM, "fltagain1", 0);
1058 		goto ReFault;
1059 
1060 	default:
1061 		return error;
1062 	}
1063 
1064 	/*
1065 	 * uobj is non null if the page is on loan from an object (i.e. uobj)
1066 	 */
1067 
1068 	uobj = anon->u.an_page->uobject;	/* locked by anonget if !NULL */
1069 
1070 	/* locked: maps(read), amap, anon, uobj(if one) */
1071 
1072 	/*
1073 	 * special handling for loaned pages
1074 	 */
1075 
1076 	if (anon->u.an_page->loan_count) {
1077 
1078 		if ((access_type & VM_PROT_WRITE) == 0) {
1079 
1080 			/*
1081 			 * for read faults on loaned pages we just cap the
1082 			 * protection at read-only.
1083 			 */
1084 
1085 			enter_prot = enter_prot & ~VM_PROT_WRITE;
1086 
1087 		} else {
1088 			/*
1089 			 * note that we can't allow writes into a loaned page!
1090 			 *
1091 			 * if we have a write fault on a loaned page in an
1092 			 * anon then we need to look at the anon's ref count.
1093 			 * if it is greater than one then we are going to do
1094 			 * a normal copy-on-write fault into a new anon (this
1095 			 * is not a problem).  however, if the reference count
1096 			 * is one (a case where we would normally allow a
1097 			 * write directly to the page) then we need to kill
1098 			 * the loan before we continue.
1099 			 */
1100 
1101 			/* >1 case is already ok */
1102 			if (anon->an_ref == 1) {
1103 
1104 				/* get new un-owned replacement page */
1105 				pg = uvm_pagealloc(NULL, 0, NULL, 0);
1106 				if (pg == NULL) {
1107 					uvmfault_unlockall(&ufi, amap, uobj,
1108 					    anon);
1109 					uvm_wait("flt_noram2");
1110 					goto ReFault;
1111 				}
1112 
1113 				/*
1114 				 * copy data, kill loan, and drop uobj lock
1115 				 * (if any)
1116 				 */
1117 				/* copy old -> new */
1118 				uvm_pagecopy(anon->u.an_page, pg);
1119 
1120 				/* force reload */
1121 				pmap_page_protect(anon->u.an_page,
1122 						  VM_PROT_NONE);
1123 				uvm_lock_pageq();	  /* KILL loan */
1124 				if (uobj)
1125 					/* if we were loaning */
1126 					anon->u.an_page->loan_count--;
1127 				anon->u.an_page->uanon = NULL;
1128 				/* in case we owned */
1129 				anon->u.an_page->pqflags &= ~PQ_ANON;
1130 				uvm_unlock_pageq();
1131 				if (uobj) {
1132 					simple_unlock(&uobj->vmobjlock);
1133 					uobj = NULL;
1134 				}
1135 
1136 				/* install new page in anon */
1137 				anon->u.an_page = pg;
1138 				pg->uanon = anon;
1139 				pg->pqflags |= PQ_ANON;
1140 				pg->flags &= ~(PG_BUSY|PG_FAKE);
1141 				UVM_PAGE_OWN(pg, NULL);
1142 
1143 				/* done! */
1144 			}     /* ref == 1 */
1145 		}       /* write fault */
1146 	}         /* loan count */
1147 
1148 	/*
1149 	 * if we are case 1B then we will need to allocate a new blank
1150 	 * anon to transfer the data into.   note that we have a lock
1151 	 * on anon, so no one can busy or release the page until we are done.
1152 	 * also note that the ref count can't drop to zero here because
1153 	 * it is > 1 and we are only dropping one ref.
1154 	 *
1155 	 * in the (hopefully very rare) case that we are out of RAM we
1156 	 * will unlock, wait for more RAM, and refault.
1157 	 *
1158 	 * if we are out of anon VM we kill the process (XXX: could wait?).
1159 	 */
1160 
1161 	if ((access_type & VM_PROT_WRITE) != 0 && anon->an_ref > 1) {
1162 
1163 		UVMHIST_LOG(maphist, "  case 1B: COW fault",0,0,0,0);
1164 		uvmexp.flt_acow++;
1165 		oanon = anon;		/* oanon = old, locked anon */
1166 		anon = uvm_analloc();
1167 		if (anon) {
1168 			/* new anon is locked! */
1169 			pg = uvm_pagealloc(NULL, 0, anon, 0);
1170 		}
1171 
1172 		/* check for out of RAM */
1173 		if (anon == NULL || pg == NULL) {
1174 			if (anon) {
1175 				anon->an_ref--;
1176 				simple_unlock(&anon->an_lock);
1177 				uvm_anfree(anon);
1178 			}
1179 			uvmfault_unlockall(&ufi, amap, uobj, oanon);
1180 			KASSERT(uvmexp.swpgonly <= uvmexp.swpages);
1181 			if (anon == NULL || uvmexp.swpgonly == uvmexp.swpages) {
1182 				UVMHIST_LOG(maphist,
1183 				    "<- failed.  out of VM",0,0,0,0);
1184 				uvmexp.fltnoanon++;
1185 				return ENOMEM;
1186 			}
1187 
1188 			uvmexp.fltnoram++;
1189 			uvm_wait("flt_noram3");	/* out of RAM, wait for more */
1190 			goto ReFault;
1191 		}
1192 
1193 		/* got all resources, replace anon with nanon */
1194 
1195 		uvm_pagecopy(oanon->u.an_page, pg);	/* pg now !PG_CLEAN */
1196 		pg->flags &= ~(PG_BUSY|PG_FAKE);	/* un-busy! new page */
1197 		UVM_PAGE_OWN(pg, NULL);
1198 		amap_add(&ufi.entry->aref, ufi.orig_rvaddr - ufi.entry->start,
1199 		    anon, 1);
1200 
1201 		/* deref: can not drop to zero here by defn! */
1202 		oanon->an_ref--;
1203 
1204 		/*
1205 		 * note: oanon is still locked, as is the new anon.  we
1206 		 * need to check for this later when we unlock oanon; if
1207 		 * oanon != anon, we'll have to unlock anon, too.
1208 		 */
1209 
1210 	} else {
1211 
1212 		uvmexp.flt_anon++;
1213 		oanon = anon;		/* old, locked anon is same as anon */
1214 		pg = anon->u.an_page;
1215 		if (anon->an_ref > 1)     /* disallow writes to ref > 1 anons */
1216 			enter_prot = enter_prot & ~VM_PROT_WRITE;
1217 
1218 	}
1219 
1220 	/* locked: maps(read), amap, oanon, anon (if different from oanon) */
1221 
1222 	/*
1223 	 * now map the page in ...
1224 	 * XXX: old fault unlocks object before pmap_enter.  this seems
1225 	 * suspect since some other thread could blast the page out from
1226 	 * under us between the unlock and the pmap_enter.
1227 	 */
1228 
1229 	UVMHIST_LOG(maphist, "  MAPPING: anon: pm=0x%x, va=0x%x, pg=0x%x",
1230 	    ufi.orig_map->pmap, ufi.orig_rvaddr, pg, 0);
1231 	if (pmap_enter(ufi.orig_map->pmap, ufi.orig_rvaddr, VM_PAGE_TO_PHYS(pg),
1232 	    enter_prot, access_type | PMAP_CANFAIL | (wired ? PMAP_WIRED : 0))
1233 	    != 0) {
1234 		/*
1235 		 * No need to undo what we did; we can simply think of
1236 		 * this as the pmap throwing away the mapping information.
1237 		 *
1238 		 * We do, however, have to go through the ReFault path,
1239 		 * as the map may change while we're asleep.
1240 		 */
1241 		if (anon != oanon)
1242 			simple_unlock(&anon->an_lock);
1243 		uvmfault_unlockall(&ufi, amap, uobj, oanon);
1244 		KASSERT(uvmexp.swpgonly <= uvmexp.swpages);
1245 		if (uvmexp.swpgonly == uvmexp.swpages) {
1246 			UVMHIST_LOG(maphist,
1247 			    "<- failed.  out of VM",0,0,0,0);
1248 			/* XXX instrumentation */
1249 			return ENOMEM;
1250 		}
1251 		/* XXX instrumentation */
1252 		uvm_wait("flt_pmfail1");
1253 		goto ReFault;
1254 	}
1255 
1256 	/*
1257 	 * ... update the page queues.
1258 	 */
1259 
1260 	uvm_lock_pageq();
1261 
1262 	if (fault_type == VM_FAULT_WIRE) {
1263 		uvm_pagewire(pg);
1264 
1265 		/*
1266 		 * since the now-wired page cannot be paged out,
1267 		 * release its swap resources for others to use.
1268 		 * since an anon with no swap cannot be PG_CLEAN,
1269 		 * clear its clean flag now.
1270 		 */
1271 
1272 		pg->flags &= ~(PG_CLEAN);
1273 		uvm_anon_dropswap(anon);
1274 	} else {
1275 		/* activate it */
1276 		uvm_pageactivate(pg);
1277 	}
1278 
1279 	uvm_unlock_pageq();
1280 
1281 	/*
1282 	 * done case 1!  finish up by unlocking everything and returning success
1283 	 */
1284 
1285 	if (anon != oanon)
1286 		simple_unlock(&anon->an_lock);
1287 	uvmfault_unlockall(&ufi, amap, uobj, oanon);
1288 	pmap_update();
1289 	return 0;
1290 
1291 
1292 Case2:
1293 	/*
1294 	 * handle case 2: faulting on backing object or zero fill
1295 	 */
1296 
1297 	/*
1298 	 * locked:
1299 	 * maps(read), amap(if there), uobj(if !null), uobjpage(if !null)
1300 	 */
1301 
1302 	/*
1303 	 * note that uobjpage can not be PGO_DONTCARE at this point.  we now
1304 	 * set uobjpage to PGO_DONTCARE if we are doing a zero fill.  if we
1305 	 * have a backing object, check and see if we are going to promote
1306 	 * the data up to an anon during the fault.
1307 	 */
1308 
1309 	if (uobj == NULL) {
1310 		uobjpage = PGO_DONTCARE;
1311 		promote = TRUE;		/* always need anon here */
1312 	} else {
1313 		KASSERT(uobjpage != PGO_DONTCARE);
1314 		promote = (access_type & VM_PROT_WRITE) &&
1315 		     UVM_ET_ISCOPYONWRITE(ufi.entry);
1316 	}
1317 	UVMHIST_LOG(maphist, "  case 2 fault: promote=%d, zfill=%d",
1318 	    promote, (uobj == NULL), 0,0);
1319 
1320 	/*
1321 	 * if uobjpage is not null then we do not need to do I/O to get the
1322 	 * uobjpage.
1323 	 *
1324 	 * if uobjpage is null, then we need to unlock and ask the pager to
1325 	 * get the data for us.   once we have the data, we need to reverify
1326 	 * the state the world.   we are currently not holding any resources.
1327 	 */
1328 
1329 	if (uobjpage) {
1330 		/* update rusage counters */
1331 		curproc->p_addr->u_stats.p_ru.ru_minflt++;
1332 	} else {
1333 		/* update rusage counters */
1334 		curproc->p_addr->u_stats.p_ru.ru_majflt++;
1335 
1336 		/* locked: maps(read), amap(if there), uobj */
1337 		uvmfault_unlockall(&ufi, amap, NULL, NULL);
1338 		/* locked: uobj */
1339 
1340 		uvmexp.fltget++;
1341 		gotpages = 1;
1342 		uoff = (ufi.orig_rvaddr - ufi.entry->start) + ufi.entry->offset;
1343 		error = uobj->pgops->pgo_get(uobj, uoff, &uobjpage, &gotpages,
1344 		    0, access_type & MASK(ufi.entry), ufi.entry->advice,
1345 		    PGO_SYNCIO);
1346 
1347 		/* locked: uobjpage(if no error) */
1348 
1349 		/*
1350 		 * recover from I/O
1351 		 */
1352 
1353 		if (error) {
1354 			if (error == EAGAIN) {
1355 				UVMHIST_LOG(maphist,
1356 				    "  pgo_get says TRY AGAIN!",0,0,0,0);
1357 				tsleep(&lbolt, PVM, "fltagain2", 0);
1358 				goto ReFault;
1359 			}
1360 
1361 			UVMHIST_LOG(maphist, "<- pgo_get failed (code %d)",
1362 			    error, 0,0,0);
1363 			return error;
1364 		}
1365 
1366 		/* locked: uobjpage */
1367 
1368 		/*
1369 		 * re-verify the state of the world by first trying to relock
1370 		 * the maps.  always relock the object.
1371 		 */
1372 
1373 		locked = uvmfault_relock(&ufi);
1374 		if (locked && amap)
1375 			amap_lock(amap);
1376 		simple_lock(&uobj->vmobjlock);
1377 
1378 		/* locked(locked): maps(read), amap(if !null), uobj, uobjpage */
1379 		/* locked(!locked): uobj, uobjpage */
1380 
1381 		/*
1382 		 * verify that the page has not be released and re-verify
1383 		 * that amap slot is still free.   if there is a problem,
1384 		 * we unlock and clean up.
1385 		 */
1386 
1387 		if ((uobjpage->flags & PG_RELEASED) != 0 ||
1388 		    (locked && amap &&
1389 		    amap_lookup(&ufi.entry->aref,
1390 		      ufi.orig_rvaddr - ufi.entry->start))) {
1391 			if (locked)
1392 				uvmfault_unlockall(&ufi, amap, NULL, NULL);
1393 			locked = FALSE;
1394 		}
1395 
1396 		/*
1397 		 * didn't get the lock?   release the page and retry.
1398 		 */
1399 
1400 		if (locked == FALSE) {
1401 
1402 			UVMHIST_LOG(maphist,
1403 			    "  wasn't able to relock after fault: retry",
1404 			    0,0,0,0);
1405 			if (uobjpage->flags & PG_WANTED)
1406 				/* still holding object lock */
1407 				wakeup(uobjpage);
1408 
1409 			if (uobjpage->flags & PG_RELEASED) {
1410 				uvmexp.fltpgrele++;
1411 				KASSERT(uobj->pgops->pgo_releasepg != NULL);
1412 
1413 				/* frees page */
1414 				if (uobj->pgops->pgo_releasepg(uobjpage,NULL))
1415 					/* unlock if still alive */
1416 					simple_unlock(&uobj->vmobjlock);
1417 				goto ReFault;
1418 			}
1419 
1420 			uvm_lock_pageq();
1421 			/* make sure it is in queues */
1422 			uvm_pageactivate(uobjpage);
1423 
1424 			uvm_unlock_pageq();
1425 			uobjpage->flags &= ~(PG_BUSY|PG_WANTED);
1426 			UVM_PAGE_OWN(uobjpage, NULL);
1427 			simple_unlock(&uobj->vmobjlock);
1428 			goto ReFault;
1429 
1430 		}
1431 
1432 		/*
1433 		 * we have the data in uobjpage which is PG_BUSY and
1434 		 * !PG_RELEASED.  we are holding object lock (so the page
1435 		 * can't be released on us).
1436 		 */
1437 
1438 		/* locked: maps(read), amap(if !null), uobj, uobjpage */
1439 	}
1440 
1441 	/*
1442 	 * locked:
1443 	 * maps(read), amap(if !null), uobj(if !null), uobjpage(if uobj)
1444 	 */
1445 
1446 	/*
1447 	 * notes:
1448 	 *  - at this point uobjpage can not be NULL
1449 	 *  - at this point uobjpage can not be PG_RELEASED (since we checked
1450 	 *  for it above)
1451 	 *  - at this point uobjpage could be PG_WANTED (handle later)
1452 	 */
1453 
1454 	if (promote == FALSE) {
1455 
1456 		/*
1457 		 * we are not promoting.   if the mapping is COW ensure that we
1458 		 * don't give more access than we should (e.g. when doing a read
1459 		 * fault on a COPYONWRITE mapping we want to map the COW page in
1460 		 * R/O even though the entry protection could be R/W).
1461 		 *
1462 		 * set "pg" to the page we want to map in (uobjpage, usually)
1463 		 */
1464 
1465 		/* no anon in this case. */
1466 		anon = NULL;
1467 
1468 		uvmexp.flt_obj++;
1469 		if (UVM_ET_ISCOPYONWRITE(ufi.entry))
1470 			enter_prot &= ~VM_PROT_WRITE;
1471 		pg = uobjpage;		/* map in the actual object */
1472 
1473 		/* assert(uobjpage != PGO_DONTCARE) */
1474 
1475 		/*
1476 		 * we are faulting directly on the page.   be careful
1477 		 * about writing to loaned pages...
1478 		 */
1479 		if (uobjpage->loan_count) {
1480 
1481 			if ((access_type & VM_PROT_WRITE) == 0) {
1482 				/* read fault: cap the protection at readonly */
1483 				/* cap! */
1484 				enter_prot = enter_prot & ~VM_PROT_WRITE;
1485 			} else {
1486 				/* write fault: must break the loan here */
1487 
1488 				/* alloc new un-owned page */
1489 				pg = uvm_pagealloc(NULL, 0, NULL, 0);
1490 
1491 				if (pg == NULL) {
1492 					/*
1493 					 * drop ownership of page, it can't
1494 					 * be released
1495 					 */
1496 					if (uobjpage->flags & PG_WANTED)
1497 						wakeup(uobjpage);
1498 					uobjpage->flags &= ~(PG_BUSY|PG_WANTED);
1499 					UVM_PAGE_OWN(uobjpage, NULL);
1500 
1501 					uvm_lock_pageq();
1502 					/* activate: we will need it later */
1503 					uvm_pageactivate(uobjpage);
1504 
1505 					uvm_unlock_pageq();
1506 					uvmfault_unlockall(&ufi, amap, uobj,
1507 					  NULL);
1508 					UVMHIST_LOG(maphist,
1509 					  "  out of RAM breaking loan, waiting",
1510 					  0,0,0,0);
1511 					uvmexp.fltnoram++;
1512 					uvm_wait("flt_noram4");
1513 					goto ReFault;
1514 				}
1515 
1516 				/*
1517 				 * copy the data from the old page to the new
1518 				 * one and clear the fake/clean flags on the
1519 				 * new page (keep it busy).  force a reload
1520 				 * of the old page by clearing it from all
1521 				 * pmaps.  then lock the page queues to
1522 				 * rename the pages.
1523 				 */
1524 				uvm_pagecopy(uobjpage, pg);	/* old -> new */
1525 				pg->flags &= ~(PG_FAKE|PG_CLEAN);
1526 				pmap_page_protect(uobjpage, VM_PROT_NONE);
1527 				if (uobjpage->flags & PG_WANTED)
1528 					wakeup(uobjpage);
1529 				/* uobj still locked */
1530 				uobjpage->flags &= ~(PG_WANTED|PG_BUSY);
1531 				UVM_PAGE_OWN(uobjpage, NULL);
1532 
1533 				uvm_lock_pageq();
1534 				offset = uobjpage->offset;
1535 				/* remove old page */
1536 				uvm_pagerealloc(uobjpage, NULL, 0);
1537 
1538 				/*
1539 				 * at this point we have absolutely no
1540 				 * control over uobjpage
1541 				 */
1542 				/* install new page */
1543 				uvm_pagerealloc(pg, uobj, offset);
1544 				uvm_unlock_pageq();
1545 
1546 				/*
1547 				 * done!  loan is broken and "pg" is
1548 				 * PG_BUSY.   it can now replace uobjpage.
1549 				 */
1550 
1551 				uobjpage = pg;
1552 
1553 			}		/* write fault case */
1554 		}		/* if loan_count */
1555 
1556 	} else {
1557 
1558 		/*
1559 		 * if we are going to promote the data to an anon we
1560 		 * allocate a blank anon here and plug it into our amap.
1561 		 */
1562 #if DIAGNOSTIC
1563 		if (amap == NULL)
1564 			panic("uvm_fault: want to promote data, but no anon");
1565 #endif
1566 
1567 		anon = uvm_analloc();
1568 		if (anon) {
1569 			/*
1570 			 * The new anon is locked.
1571 			 *
1572 			 * In `Fill in data...' below, if
1573 			 * uobjpage == PGO_DONTCARE, we want
1574 			 * a zero'd, dirty page, so have
1575 			 * uvm_pagealloc() do that for us.
1576 			 */
1577 			pg = uvm_pagealloc(NULL, 0, anon,
1578 			    (uobjpage == PGO_DONTCARE) ? UVM_PGA_ZERO : 0);
1579 		}
1580 
1581 		/*
1582 		 * out of memory resources?
1583 		 */
1584 		if (anon == NULL || pg == NULL) {
1585 
1586 			if (anon != NULL) {
1587 				anon->an_ref--;
1588 				simple_unlock(&anon->an_lock);
1589 				uvm_anfree(anon);
1590 			}
1591 
1592 			/*
1593 			 * arg!  must unbusy our page and fail or sleep.
1594 			 */
1595 			if (uobjpage != PGO_DONTCARE) {
1596 				if (uobjpage->flags & PG_WANTED)
1597 					/* still holding object lock */
1598 					wakeup(uobjpage);
1599 
1600 				uvm_lock_pageq();
1601 				uvm_pageactivate(uobjpage);
1602 				uvm_unlock_pageq();
1603 				uobjpage->flags &= ~(PG_BUSY|PG_WANTED);
1604 				UVM_PAGE_OWN(uobjpage, NULL);
1605 			}
1606 
1607 			/* unlock and fail ... */
1608 			uvmfault_unlockall(&ufi, amap, uobj, NULL);
1609 			KASSERT(uvmexp.swpgonly <= uvmexp.swpages);
1610 			if (anon == NULL || uvmexp.swpgonly == uvmexp.swpages) {
1611 				UVMHIST_LOG(maphist, "  promote: out of VM",
1612 				    0,0,0,0);
1613 				uvmexp.fltnoanon++;
1614 				return ENOMEM;
1615 			}
1616 
1617 			UVMHIST_LOG(maphist, "  out of RAM, waiting for more",
1618 			    0,0,0,0);
1619 			uvmexp.fltnoram++;
1620 			uvm_wait("flt_noram5");
1621 			goto ReFault;
1622 		}
1623 
1624 		/*
1625 		 * fill in the data
1626 		 */
1627 
1628 		if (uobjpage != PGO_DONTCARE) {
1629 			uvmexp.flt_prcopy++;
1630 			/* copy page [pg now dirty] */
1631 			uvm_pagecopy(uobjpage, pg);
1632 
1633 			/*
1634 			 * promote to shared amap?  make sure all sharing
1635 			 * procs see it
1636 			 */
1637 			if ((amap_flags(amap) & AMAP_SHARED) != 0) {
1638 				pmap_page_protect(uobjpage, VM_PROT_NONE);
1639 				/*
1640 				 * XXX: PAGE MIGHT BE WIRED!
1641 				 */
1642 			}
1643 
1644 			/*
1645 			 * dispose of uobjpage.  it can't be PG_RELEASED
1646 			 * since we still hold the object lock.
1647 			 * drop handle to uobj as well.
1648 			 */
1649 
1650 			if (uobjpage->flags & PG_WANTED)
1651 				/* still have the obj lock */
1652 				wakeup(uobjpage);
1653 			uobjpage->flags &= ~(PG_BUSY|PG_WANTED);
1654 			UVM_PAGE_OWN(uobjpage, NULL);
1655 			uvm_lock_pageq();
1656 			uvm_pageactivate(uobjpage);
1657 			uvm_unlock_pageq();
1658 			simple_unlock(&uobj->vmobjlock);
1659 			uobj = NULL;
1660 
1661 			UVMHIST_LOG(maphist,
1662 			    "  promote uobjpage 0x%x to anon/page 0x%x/0x%x",
1663 			    uobjpage, anon, pg, 0);
1664 
1665 		} else {
1666 			uvmexp.flt_przero++;
1667 			/*
1668 			 * Page is zero'd and marked dirty by uvm_pagealloc()
1669 			 * above.
1670 			 */
1671 			UVMHIST_LOG(maphist,"  zero fill anon/page 0x%x/0%x",
1672 			    anon, pg, 0, 0);
1673 		}
1674 
1675 		amap_add(&ufi.entry->aref, ufi.orig_rvaddr - ufi.entry->start,
1676 		    anon, 0);
1677 	}
1678 
1679 	/*
1680 	 * locked:
1681 	 * maps(read), amap(if !null), uobj(if !null), uobjpage(if uobj),
1682 	 *   anon(if !null), pg(if anon)
1683 	 *
1684 	 * note: pg is either the uobjpage or the new page in the new anon
1685 	 */
1686 
1687 	/*
1688 	 * all resources are present.   we can now map it in and free our
1689 	 * resources.
1690 	 */
1691 
1692 	UVMHIST_LOG(maphist,
1693 	    "  MAPPING: case2: pm=0x%x, va=0x%x, pg=0x%x, promote=%d",
1694 	    ufi.orig_map->pmap, ufi.orig_rvaddr, pg, promote);
1695 	KASSERT(access_type == VM_PROT_READ || (pg->flags & PG_RDONLY) == 0);
1696 	if (pmap_enter(ufi.orig_map->pmap, ufi.orig_rvaddr, VM_PAGE_TO_PHYS(pg),
1697 	    pg->flags & PG_RDONLY ? VM_PROT_READ : enter_prot,
1698 	    access_type | PMAP_CANFAIL | (wired ? PMAP_WIRED : 0)) != 0) {
1699 
1700 		/*
1701 		 * No need to undo what we did; we can simply think of
1702 		 * this as the pmap throwing away the mapping information.
1703 		 *
1704 		 * We do, however, have to go through the ReFault path,
1705 		 * as the map may change while we're asleep.
1706 		 */
1707 
1708 		if (pg->flags & PG_WANTED)
1709 			wakeup(pg);		/* lock still held */
1710 
1711 		/*
1712 		 * note that pg can't be PG_RELEASED since we did not drop
1713 		 * the object lock since the last time we checked.
1714 		 */
1715 
1716 		pg->flags &= ~(PG_BUSY|PG_FAKE|PG_WANTED);
1717 		UVM_PAGE_OWN(pg, NULL);
1718 		uvmfault_unlockall(&ufi, amap, uobj, anon);
1719 		KASSERT(uvmexp.swpgonly <= uvmexp.swpages);
1720 		if (uvmexp.swpgonly == uvmexp.swpages) {
1721 			UVMHIST_LOG(maphist,
1722 			    "<- failed.  out of VM",0,0,0,0);
1723 			/* XXX instrumentation */
1724 			return ENOMEM;
1725 		}
1726 		/* XXX instrumentation */
1727 		uvm_wait("flt_pmfail2");
1728 		goto ReFault;
1729 	}
1730 
1731 	uvm_lock_pageq();
1732 
1733 	if (fault_type == VM_FAULT_WIRE) {
1734 		uvm_pagewire(pg);
1735 		if (pg->pqflags & PQ_AOBJ) {
1736 
1737 			/*
1738 			 * since the now-wired page cannot be paged out,
1739 			 * release its swap resources for others to use.
1740 			 * since an aobj page with no swap cannot be PG_CLEAN,
1741 			 * clear its clean flag now.
1742 			 */
1743 
1744 			pg->flags &= ~(PG_CLEAN);
1745 			uao_dropswap(uobj, pg->offset >> PAGE_SHIFT);
1746 		}
1747 	} else {
1748 		/* activate it */
1749 		uvm_pageactivate(pg);
1750 	}
1751 	uvm_unlock_pageq();
1752 
1753 	if (pg->flags & PG_WANTED)
1754 		wakeup(pg);		/* lock still held */
1755 
1756 	/*
1757 	 * note that pg can't be PG_RELEASED since we did not drop the object
1758 	 * lock since the last time we checked.
1759 	 */
1760 
1761 	pg->flags &= ~(PG_BUSY|PG_FAKE|PG_WANTED);
1762 	UVM_PAGE_OWN(pg, NULL);
1763 	uvmfault_unlockall(&ufi, amap, uobj, anon);
1764 
1765 	pmap_update();
1766 
1767 	UVMHIST_LOG(maphist, "<- done (SUCCESS!)",0,0,0,0);
1768 	return 0;
1769 }
1770 
1771 
1772 /*
1773  * uvm_fault_wire: wire down a range of virtual addresses in a map.
1774  *
1775  * => map may be read-locked by caller, but MUST NOT be write-locked.
1776  * => if map is read-locked, any operations which may cause map to
1777  *	be write-locked in uvm_fault() must be taken care of by
1778  *	the caller.  See uvm_map_pageable().
1779  */
1780 
1781 int
1782 uvm_fault_wire(map, start, end, access_type)
1783 	struct vm_map *map;
1784 	vaddr_t start, end;
1785 	vm_prot_t access_type;
1786 {
1787 	vaddr_t va;
1788 	int error;
1789 
1790 	/*
1791 	 * now fault it in a page at a time.   if the fault fails then we have
1792 	 * to undo what we have done.   note that in uvm_fault VM_PROT_NONE
1793 	 * is replaced with the max protection if fault_type is VM_FAULT_WIRE.
1794 	 */
1795 
1796 	/*
1797 	 * XXX work around overflowing a vaddr_t.  this prevents us from
1798 	 * wiring the last page in the address space, though.
1799 	 */
1800 	if (start > end) {
1801 		return EFAULT;
1802 	}
1803 
1804 	for (va = start ; va < end ; va += PAGE_SIZE) {
1805 		error = uvm_fault(map, va, VM_FAULT_WIRE, access_type);
1806 		if (error) {
1807 			if (va != start) {
1808 				uvm_fault_unwire(map, start, va);
1809 			}
1810 			return error;
1811 		}
1812 	}
1813 	return 0;
1814 }
1815 
1816 /*
1817  * uvm_fault_unwire(): unwire range of virtual space.
1818  */
1819 
1820 void
1821 uvm_fault_unwire(map, start, end)
1822 	struct vm_map *map;
1823 	vaddr_t start, end;
1824 {
1825 
1826 	vm_map_lock_read(map);
1827 	uvm_fault_unwire_locked(map, start, end);
1828 	vm_map_unlock_read(map);
1829 }
1830 
1831 /*
1832  * uvm_fault_unwire_locked(): the guts of uvm_fault_unwire().
1833  *
1834  * => map must be at least read-locked.
1835  */
1836 
1837 void
1838 uvm_fault_unwire_locked(map, start, end)
1839 	struct vm_map *map;
1840 	vaddr_t start, end;
1841 {
1842 	struct vm_map_entry *entry;
1843 	pmap_t pmap = vm_map_pmap(map);
1844 	vaddr_t va;
1845 	paddr_t pa;
1846 	struct vm_page *pg;
1847 
1848 	KASSERT((map->flags & VM_MAP_INTRSAFE) == 0);
1849 
1850 	/*
1851 	 * we assume that the area we are unwiring has actually been wired
1852 	 * in the first place.   this means that we should be able to extract
1853 	 * the PAs from the pmap.   we also lock out the page daemon so that
1854 	 * we can call uvm_pageunwire.
1855 	 */
1856 
1857 	uvm_lock_pageq();
1858 
1859 	/*
1860 	 * find the beginning map entry for the region.
1861 	 */
1862 	KASSERT(start >= vm_map_min(map) && end <= vm_map_max(map));
1863 	if (uvm_map_lookup_entry(map, start, &entry) == FALSE)
1864 		panic("uvm_fault_unwire_locked: address not in map");
1865 
1866 	for (va = start; va < end ; va += PAGE_SIZE) {
1867 		if (pmap_extract(pmap, va, &pa) == FALSE)
1868 			panic("uvm_fault_unwire_locked: unwiring "
1869 			    "non-wired memory");
1870 
1871 		/*
1872 		 * make sure the current entry is for the address we're
1873 		 * dealing with.  if not, grab the next entry.
1874 		 */
1875 
1876 		KASSERT(va >= entry->start);
1877 		if (va >= entry->end) {
1878 			KASSERT(entry->next != &map->header &&
1879 				entry->next->start <= entry->end);
1880 			entry = entry->next;
1881 		}
1882 
1883 		/*
1884 		 * if the entry is no longer wired, tell the pmap.
1885 		 */
1886 		if (VM_MAPENT_ISWIRED(entry) == 0)
1887 			pmap_unwire(pmap, va);
1888 
1889 		pg = PHYS_TO_VM_PAGE(pa);
1890 		if (pg)
1891 			uvm_pageunwire(pg);
1892 	}
1893 
1894 	uvm_unlock_pageq();
1895 }
1896