xref: /openbsd-src/sys/uvm/uvm_fault.c (revision daf88648c0e349d5c02e1504293082072c981640)
1 /*	$OpenBSD: uvm_fault.c,v 1.41 2006/07/31 11:51:29 mickey Exp $	*/
2 /*	$NetBSD: uvm_fault.c,v 1.51 2000/08/06 00:22:53 thorpej Exp $	*/
3 
4 /*
5  *
6  * Copyright (c) 1997 Charles D. Cranor and Washington University.
7  * All rights reserved.
8  *
9  * Redistribution and use in source and binary forms, with or without
10  * modification, are permitted provided that the following conditions
11  * are met:
12  * 1. Redistributions of source code must retain the above copyright
13  *    notice, this list of conditions and the following disclaimer.
14  * 2. Redistributions in binary form must reproduce the above copyright
15  *    notice, this list of conditions and the following disclaimer in the
16  *    documentation and/or other materials provided with the distribution.
17  * 3. All advertising materials mentioning features or use of this software
18  *    must display the following acknowledgement:
19  *      This product includes software developed by Charles D. Cranor and
20  *      Washington University.
21  * 4. The name of the author may not be used to endorse or promote products
22  *    derived from this software without specific prior written permission.
23  *
24  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
25  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
26  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
27  * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
28  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
29  * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
30  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
31  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
32  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
33  * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
34  *
35  * from: Id: uvm_fault.c,v 1.1.2.23 1998/02/06 05:29:05 chs Exp
36  */
37 
38 /*
39  * uvm_fault.c: fault handler
40  */
41 
42 #include <sys/param.h>
43 #include <sys/systm.h>
44 #include <sys/kernel.h>
45 #include <sys/proc.h>
46 #include <sys/malloc.h>
47 #include <sys/mman.h>
48 #include <sys/user.h>
49 
50 #include <uvm/uvm.h>
51 
52 /*
53  *
54  * a word on page faults:
55  *
56  * types of page faults we handle:
57  *
58  * CASE 1: upper layer faults                   CASE 2: lower layer faults
59  *
60  *    CASE 1A         CASE 1B                  CASE 2A        CASE 2B
61  *    read/write1     write>1                  read/write   +-cow_write/zero
62  *         |             |                         |        |
63  *      +--|--+       +--|--+     +-----+       +  |  +     | +-----+
64  * amap |  V  |       |  ----------->new|          |        | |  ^  |
65  *      +-----+       +-----+     +-----+       +  |  +     | +--|--+
66  *                                                 |        |    |
67  *      +-----+       +-----+                   +--|--+     | +--|--+
68  * uobj | d/c |       | d/c |                   |  V  |     +----|  |
69  *      +-----+       +-----+                   +-----+       +-----+
70  *
71  * d/c = don't care
72  *
73  *   case [0]: layerless fault
74  *	no amap or uobj is present.   this is an error.
75  *
76  *   case [1]: upper layer fault [anon active]
77  *     1A: [read] or [write with anon->an_ref == 1]
78  *		I/O takes place in top level anon and uobj is not touched.
79  *     1B: [write with anon->an_ref > 1]
80  *		new anon is alloc'd and data is copied off ["COW"]
81  *
82  *   case [2]: lower layer fault [uobj]
83  *     2A: [read on non-NULL uobj] or [write to non-copy_on_write area]
84  *		I/O takes place directly in object.
85  *     2B: [write to copy_on_write] or [read on NULL uobj]
86  *		data is "promoted" from uobj to a new anon.
87  *		if uobj is null, then we zero fill.
88  *
89  * we follow the standard UVM locking protocol ordering:
90  *
91  * MAPS => AMAP => UOBJ => ANON => PAGE QUEUES (PQ)
92  * we hold a PG_BUSY page if we unlock for I/O
93  *
94  *
95  * the code is structured as follows:
96  *
97  *     - init the "IN" params in the ufi structure
98  *   ReFault:
99  *     - do lookups [locks maps], check protection, handle needs_copy
100  *     - check for case 0 fault (error)
101  *     - establish "range" of fault
102  *     - if we have an amap lock it and extract the anons
103  *     - if sequential advice deactivate pages behind us
104  *     - at the same time check pmap for unmapped areas and anon for pages
105  *	 that we could map in (and do map it if found)
106  *     - check object for resident pages that we could map in
107  *     - if (case 2) goto Case2
108  *     - >>> handle case 1
109  *           - ensure source anon is resident in RAM
110  *           - if case 1B alloc new anon and copy from source
111  *           - map the correct page in
112  *   Case2:
113  *     - >>> handle case 2
114  *           - ensure source page is resident (if uobj)
115  *           - if case 2B alloc new anon and copy from source (could be zero
116  *		fill if uobj == NULL)
117  *           - map the correct page in
118  *     - done!
119  *
120  * note on paging:
121  *   if we have to do I/O we place a PG_BUSY page in the correct object,
122  * unlock everything, and do the I/O.   when I/O is done we must reverify
123  * the state of the world before assuming that our data structures are
124  * valid.   [because mappings could change while the map is unlocked]
125  *
126  *  alternative 1: unbusy the page in question and restart the page fault
127  *    from the top (ReFault).   this is easy but does not take advantage
128  *    of the information that we already have from our previous lookup,
129  *    although it is possible that the "hints" in the vm_map will help here.
130  *
131  * alternative 2: the system already keeps track of a "version" number of
132  *    a map.   [i.e. every time you write-lock a map (e.g. to change a
133  *    mapping) you bump the version number up by one...]   so, we can save
134  *    the version number of the map before we release the lock and start I/O.
135  *    then when I/O is done we can relock and check the version numbers
136  *    to see if anything changed.    this might save us some over 1 because
137  *    we don't have to unbusy the page and may be less compares(?).
138  *
139  * alternative 3: put in backpointers or a way to "hold" part of a map
140  *    in place while I/O is in progress.   this could be complex to
141  *    implement (especially with structures like amap that can be referenced
142  *    by multiple map entries, and figuring out what should wait could be
143  *    complex as well...).
144  *
145  * given that we are not currently multiprocessor or multithreaded we might
146  * as well choose alternative 2 now.   maybe alternative 3 would be useful
147  * in the future.    XXX keep in mind for future consideration//rechecking.
148  */
149 
150 /*
151  * local data structures
152  */
153 
154 struct uvm_advice {
155 	int advice;
156 	int nback;
157 	int nforw;
158 };
159 
160 /*
161  * page range array:
162  * note: index in array must match "advice" value
163  * XXX: borrowed numbers from freebsd.   do they work well for us?
164  */
165 
166 static struct uvm_advice uvmadvice[] = {
167 	{ MADV_NORMAL, 3, 4 },
168 	{ MADV_RANDOM, 0, 0 },
169 	{ MADV_SEQUENTIAL, 8, 7},
170 };
171 
172 #define UVM_MAXRANGE 16	/* must be max() of nback+nforw+1 */
173 
174 /*
175  * private prototypes
176  */
177 
178 static void uvmfault_amapcopy(struct uvm_faultinfo *);
179 static __inline void uvmfault_anonflush(struct vm_anon **, int);
180 
181 /*
182  * inline functions
183  */
184 
185 /*
186  * uvmfault_anonflush: try and deactivate pages in specified anons
187  *
188  * => does not have to deactivate page if it is busy
189  */
190 
191 static __inline void
192 uvmfault_anonflush(anons, n)
193 	struct vm_anon **anons;
194 	int n;
195 {
196 	int lcv;
197 	struct vm_page *pg;
198 
199 	for (lcv = 0 ; lcv < n ; lcv++) {
200 		if (anons[lcv] == NULL)
201 			continue;
202 		simple_lock(&anons[lcv]->an_lock);
203 		pg = anons[lcv]->u.an_page;
204 		if (pg && (pg->flags & PG_BUSY) == 0 && pg->loan_count == 0) {
205 			uvm_lock_pageq();
206 			if (pg->wire_count == 0) {
207 #ifdef UBC
208 				pmap_clear_reference(pg);
209 #else
210 				pmap_page_protect(pg, VM_PROT_NONE);
211 #endif
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 
237 	/*
238 	 * while we haven't done the job
239 	 */
240 
241 	while (1) {
242 
243 		/*
244 		 * no mapping?  give up.
245 		 */
246 
247 		if (uvmfault_lookup(ufi, TRUE) == FALSE)
248 			return;
249 
250 		/*
251 		 * copy if needed.
252 		 */
253 
254 		if (UVM_ET_ISNEEDSCOPY(ufi->entry))
255 			amap_copy(ufi->map, ufi->entry, M_NOWAIT, TRUE,
256 				ufi->orig_rvaddr, ufi->orig_rvaddr + 1);
257 
258 		/*
259 		 * didn't work?  must be out of RAM.   unlock and sleep.
260 		 */
261 
262 		if (UVM_ET_ISNEEDSCOPY(ufi->entry)) {
263 			uvmfault_unlockmaps(ufi, TRUE);
264 			uvm_wait("fltamapcopy");
265 			continue;
266 		}
267 
268 		/*
269 		 * got it!   unlock and return.
270 		 */
271 
272 		uvmfault_unlockmaps(ufi, TRUE);
273 		return;
274 	}
275 	/*NOTREACHED*/
276 }
277 
278 /*
279  * uvmfault_anonget: get data in an anon into a non-busy, non-released
280  * page in that anon.
281  *
282  * => maps, amap, and anon locked by caller.
283  * => if we fail (result != VM_PAGER_OK) we unlock everything.
284  * => if we are successful, we return with everything still locked.
285  * => we don't move the page on the queues [gets moved later]
286  * => if we allocate a new page [we_own], it gets put on the queues.
287  *    either way, the result is that the page is on the queues at return time
288  * => for pages which are on loan from a uvm_object (and thus are not
289  *    owned by the anon): if successful, we return with the owning object
290  *    locked.   the caller must unlock this object when it unlocks everything
291  *    else.
292  */
293 
294 int
295 uvmfault_anonget(ufi, amap, anon)
296 	struct uvm_faultinfo *ufi;
297 	struct vm_amap *amap;
298 	struct vm_anon *anon;
299 {
300 	boolean_t we_own;	/* we own anon's page? */
301 	boolean_t locked;	/* did we relock? */
302 	struct vm_page *pg;
303 	int result;
304 	UVMHIST_FUNC("uvmfault_anonget"); UVMHIST_CALLED(maphist);
305 
306 	result = 0;		/* XXX shut up gcc */
307 	uvmexp.fltanget++;
308         /* bump rusage counters */
309 	if (anon->u.an_page)
310 		curproc->p_addr->u_stats.p_ru.ru_minflt++;
311 	else
312 		curproc->p_addr->u_stats.p_ru.ru_majflt++;
313 
314 	/*
315 	 * loop until we get it, or fail.
316 	 */
317 
318 	while (1) {
319 
320 		we_own = FALSE;		/* TRUE if we set PG_BUSY on a page */
321 		pg = anon->u.an_page;
322 
323 		/*
324 		 * if there is a resident page and it is loaned, then anon
325 		 * may not own it.   call out to uvm_anon_lockpage() to ensure
326 		 * the real owner of the page has been identified and locked.
327 		 */
328 
329 		if (pg && pg->loan_count)
330 			pg = uvm_anon_lockloanpg(anon);
331 
332 		/*
333 		 * page there?   make sure it is not busy/released.
334 		 */
335 
336 		if (pg) {
337 
338 			/*
339 			 * at this point, if the page has a uobject [meaning
340 			 * we have it on loan], then that uobject is locked
341 			 * by us!   if the page is busy, we drop all the
342 			 * locks (including uobject) and try again.
343 			 */
344 
345 			if ((pg->flags & (PG_BUSY|PG_RELEASED)) == 0) {
346 				UVMHIST_LOG(maphist, "<- OK",0,0,0,0);
347 				return (VM_PAGER_OK);
348 			}
349 			pg->flags |= PG_WANTED;
350 			uvmexp.fltpgwait++;
351 
352 			/*
353 			 * the last unlock must be an atomic unlock+wait on
354 			 * the owner of page
355 			 */
356 			if (pg->uobject) {	/* owner is uobject ? */
357 				uvmfault_unlockall(ufi, amap, NULL, anon);
358 				UVMHIST_LOG(maphist, " unlock+wait on uobj",0,
359 				    0,0,0);
360 				UVM_UNLOCK_AND_WAIT(pg,
361 				    &pg->uobject->vmobjlock,
362 				    FALSE, "anonget1",0);
363 			} else {
364 				/* anon owns page */
365 				uvmfault_unlockall(ufi, amap, NULL, NULL);
366 				UVMHIST_LOG(maphist, " unlock+wait on anon",0,
367 				    0,0,0);
368 				UVM_UNLOCK_AND_WAIT(pg,&anon->an_lock,0,
369 				    "anonget2",0);
370 			}
371 			/* ready to relock and try again */
372 
373 		} else {
374 
375 			/*
376 			 * no page, we must try and bring it in.
377 			 */
378 			pg = uvm_pagealloc(NULL, 0, anon, 0);
379 
380 			if (pg == NULL) {		/* out of RAM.  */
381 
382 				uvmfault_unlockall(ufi, amap, NULL, anon);
383 				uvmexp.fltnoram++;
384 				UVMHIST_LOG(maphist, "  noram -- UVM_WAIT",0,
385 				    0,0,0);
386 				uvm_wait("flt_noram1");
387 				/* ready to relock and try again */
388 
389 			} else {
390 
391 				/* we set the PG_BUSY bit */
392 				we_own = TRUE;
393 				uvmfault_unlockall(ufi, amap, NULL, anon);
394 
395 				/*
396 				 * we are passing a PG_BUSY+PG_FAKE+PG_CLEAN
397 				 * page into the uvm_swap_get function with
398 				 * all data structures unlocked.  note that
399 				 * it is ok to read an_swslot here because
400 				 * we hold PG_BUSY on the page.
401 				 */
402 				uvmexp.pageins++;
403 				result = uvm_swap_get(pg, anon->an_swslot,
404 				    PGO_SYNCIO);
405 
406 				/*
407 				 * we clean up after the i/o below in the
408 				 * "we_own" case
409 				 */
410 				/* ready to relock and try again */
411 			}
412 		}
413 
414 		/*
415 		 * now relock and try again
416 		 */
417 
418 		locked = uvmfault_relock(ufi);
419 		if (locked && amap != NULL) {
420 			amap_lock(amap);
421 		}
422 		if (locked || we_own)
423 			simple_lock(&anon->an_lock);
424 
425 		/*
426 		 * if we own the page (i.e. we set PG_BUSY), then we need
427 		 * to clean up after the I/O. there are three cases to
428 		 * consider:
429 		 *   [1] page released during I/O: free anon and ReFault.
430 		 *   [2] I/O not OK.   free the page and cause the fault
431 		 *       to fail.
432 		 *   [3] I/O OK!   activate the page and sync with the
433 		 *       non-we_own case (i.e. drop anon lock if not locked).
434 		 */
435 
436 		if (we_own) {
437 
438 			if (pg->flags & PG_WANTED) {
439 				/* still holding object lock */
440 				wakeup(pg);
441 			}
442 			/* un-busy! */
443 			pg->flags &= ~(PG_WANTED|PG_BUSY|PG_FAKE);
444 			UVM_PAGE_OWN(pg, NULL);
445 
446 			/*
447 			 * if we were RELEASED during I/O, then our anon is
448 			 * no longer part of an amap.   we need to free the
449 			 * anon and try again.
450 			 */
451 			if (pg->flags & PG_RELEASED) {
452 				pmap_page_protect(pg, VM_PROT_NONE);
453 				simple_unlock(&anon->an_lock);
454 				uvm_anfree(anon);	/* frees page for us */
455 				if (locked)
456 					uvmfault_unlockall(ufi, amap, NULL,
457 							   NULL);
458 				uvmexp.fltpgrele++;
459 				UVMHIST_LOG(maphist, "<- REFAULT", 0,0,0,0);
460 				return (VM_PAGER_REFAULT);	/* refault! */
461 			}
462 
463 			if (result != VM_PAGER_OK) {
464 				KASSERT(result != VM_PAGER_PEND);
465 
466 				/* remove page from anon */
467 				anon->u.an_page = NULL;
468 
469 				/*
470 				 * remove the swap slot from the anon
471 				 * and mark the anon as having no real slot.
472 				 * don't free the swap slot, thus preventing
473 				 * it from being used again.
474 				 */
475 				uvm_swap_markbad(anon->an_swslot, 1);
476 				anon->an_swslot = SWSLOT_BAD;
477 
478 				/*
479 				 * note: page was never !PG_BUSY, so it
480 				 * can't be mapped and thus no need to
481 				 * pmap_page_protect it...
482 				 */
483 				uvm_lock_pageq();
484 				uvm_pagefree(pg);
485 				uvm_unlock_pageq();
486 
487 				if (locked)
488 					uvmfault_unlockall(ufi, amap, NULL,
489 					    anon);
490 				else
491 					simple_unlock(&anon->an_lock);
492 				UVMHIST_LOG(maphist, "<- ERROR", 0,0,0,0);
493 				return (VM_PAGER_ERROR);
494 			}
495 
496 			/*
497 			 * must be OK, clear modify (already PG_CLEAN)
498 			 * and activate
499 			 */
500 			pmap_clear_modify(pg);
501 			uvm_lock_pageq();
502 			uvm_pageactivate(pg);
503 			uvm_unlock_pageq();
504 			if (!locked)
505 				simple_unlock(&anon->an_lock);
506 		}
507 
508 		/*
509 		 * we were not able to relock.   restart fault.
510 		 */
511 
512 		if (!locked) {
513 			UVMHIST_LOG(maphist, "<- REFAULT", 0,0,0,0);
514 			return (VM_PAGER_REFAULT);
515 		}
516 
517 		/*
518 		 * verify no one has touched the amap and moved the anon on us.
519 		 */
520 
521 		if (ufi != NULL &&
522 		    amap_lookup(&ufi->entry->aref,
523 				ufi->orig_rvaddr - ufi->entry->start) != anon) {
524 
525 			uvmfault_unlockall(ufi, amap, NULL, anon);
526 			UVMHIST_LOG(maphist, "<- REFAULT", 0,0,0,0);
527 			return (VM_PAGER_REFAULT);
528 		}
529 
530 		/*
531 		 * try it again!
532 		 */
533 
534 		uvmexp.fltanretry++;
535 		continue;
536 
537 	} /* while (1) */
538 
539 	/*NOTREACHED*/
540 }
541 
542 /*
543  *   F A U L T   -   m a i n   e n t r y   p o i n t
544  */
545 
546 /*
547  * uvm_fault: page fault handler
548  *
549  * => called from MD code to resolve a page fault
550  * => VM data structures usually should be unlocked.   however, it is
551  *	possible to call here with the main map locked if the caller
552  *	gets a write lock, sets it recursive, and then calls us (c.f.
553  *	uvm_map_pageable).   this should be avoided because it keeps
554  *	the map locked off during I/O.
555  */
556 
557 #define MASK(entry)     (UVM_ET_ISCOPYONWRITE(entry) ? \
558 			 ~VM_PROT_WRITE : VM_PROT_ALL)
559 
560 int
561 uvm_fault(orig_map, vaddr, fault_type, access_type)
562 	vm_map_t orig_map;
563 	vaddr_t vaddr;
564 	vm_fault_t fault_type;
565 	vm_prot_t access_type;
566 {
567 	struct uvm_faultinfo ufi;
568 	vm_prot_t enter_prot;
569 	boolean_t wired, narrow, promote, locked, shadowed;
570 	int npages, nback, nforw, centeridx, result, lcv, gotpages;
571 	vaddr_t startva, currva;
572 	voff_t uoff;
573 	paddr_t pa;
574 	struct vm_amap *amap;
575 	struct uvm_object *uobj;
576 	struct vm_anon *anons_store[UVM_MAXRANGE], **anons, *anon, *oanon;
577 	struct vm_page *pages[UVM_MAXRANGE], *pg, *uobjpage;
578 	UVMHIST_FUNC("uvm_fault"); UVMHIST_CALLED(maphist);
579 
580 	UVMHIST_LOG(maphist, "(map=%p, vaddr=0x%lx, ft=%ld, at=%ld)",
581 	      orig_map, vaddr, fault_type, access_type);
582 
583 	anon = NULL;
584 	pg = NULL;
585 
586 	uvmexp.faults++;	/* XXX: locking? */
587 
588 	/*
589 	 * init the IN parameters in the ufi
590 	 */
591 
592 	ufi.orig_map = orig_map;
593 	ufi.orig_rvaddr = trunc_page(vaddr);
594 	ufi.orig_size = PAGE_SIZE;	/* can't get any smaller than this */
595 	if (fault_type == VM_FAULT_WIRE)
596 		narrow = TRUE;		/* don't look for neighborhood
597 					 * pages on wire */
598 	else
599 		narrow = FALSE;		/* normal fault */
600 
601 	/*
602 	 * before we do anything else, if this is a fault on a kernel
603 	 * address, check to see if the address is managed by an
604 	 * interrupt-safe map.  If it is, we fail immediately.  Intrsafe
605 	 * maps are never pageable, and this approach avoids an evil
606 	 * locking mess.
607 	 */
608 	if (orig_map == kernel_map && uvmfault_check_intrsafe(&ufi)) {
609 		UVMHIST_LOG(maphist, "<- VA 0x%lx in intrsafe map %p",
610 		    ufi.orig_rvaddr, ufi.map, 0, 0);
611 		return (KERN_FAILURE);
612 	}
613 
614 	/*
615 	 * "goto ReFault" means restart the page fault from ground zero.
616 	 */
617 ReFault:
618 
619 	/*
620 	 * lookup and lock the maps
621 	 */
622 
623 	if (uvmfault_lookup(&ufi, FALSE) == FALSE) {
624 		UVMHIST_LOG(maphist, "<- no mapping @ 0x%lx", vaddr, 0,0,0);
625 		return (KERN_INVALID_ADDRESS);
626 	}
627 	/* locked: maps(read) */
628 
629 	/*
630 	 * check protection
631 	 */
632 
633 	if ((ufi.entry->protection & access_type) != access_type) {
634 		UVMHIST_LOG(maphist,
635 		    "<- protection failure (prot=0x%lx, access=0x%lx)",
636 		    ufi.entry->protection, access_type, 0, 0);
637 		uvmfault_unlockmaps(&ufi, FALSE);
638 		return (KERN_PROTECTION_FAILURE);
639 	}
640 
641 	/*
642 	 * if the map is not a pageable map, a page fault always fails.
643 	 */
644 
645 	if ((ufi.map->flags & VM_MAP_PAGEABLE) == 0) {
646 		UVMHIST_LOG(maphist,
647 		    "<- map %p not pageable", ufi.map, 0, 0, 0);
648 		uvmfault_unlockmaps(&ufi, FALSE);
649 		return (KERN_FAILURE);
650 	}
651 
652 	/*
653 	 * "enter_prot" is the protection we want to enter the page in at.
654 	 * for certain pages (e.g. copy-on-write pages) this protection can
655 	 * be more strict than ufi.entry->protection.  "wired" means either
656 	 * the entry is wired or we are fault-wiring the pg.
657 	 */
658 
659 	enter_prot = ufi.entry->protection;
660 	wired = VM_MAPENT_ISWIRED(ufi.entry) || (fault_type == VM_FAULT_WIRE);
661 	if (wired)
662 		access_type = enter_prot; /* full access for wired */
663 
664 	/*
665 	 * handle "needs_copy" case.   if we need to copy the amap we will
666 	 * have to drop our readlock and relock it with a write lock.  (we
667 	 * need a write lock to change anything in a map entry [e.g.
668 	 * needs_copy]).
669 	 */
670 
671 	if (UVM_ET_ISNEEDSCOPY(ufi.entry)) {
672 		if ((access_type & VM_PROT_WRITE) ||
673 		    (ufi.entry->object.uvm_obj == NULL)) {
674 			/* need to clear */
675 			UVMHIST_LOG(maphist,
676 			    "  need to clear needs_copy and refault",0,0,0,0);
677 			uvmfault_unlockmaps(&ufi, FALSE);
678 			uvmfault_amapcopy(&ufi);
679 			uvmexp.fltamcopy++;
680 			goto ReFault;
681 
682 		} else {
683 
684 			/*
685 			 * ensure that we pmap_enter page R/O since
686 			 * needs_copy is still true
687 			 */
688 			enter_prot &= ~VM_PROT_WRITE;
689 
690 		}
691 	}
692 
693 	/*
694 	 * identify the players
695 	 */
696 
697 	amap = ufi.entry->aref.ar_amap;		/* top layer */
698 	uobj = ufi.entry->object.uvm_obj;	/* bottom layer */
699 
700 	/*
701 	 * check for a case 0 fault.  if nothing backing the entry then
702 	 * error now.
703 	 */
704 
705 	if (amap == NULL && uobj == NULL) {
706 		uvmfault_unlockmaps(&ufi, FALSE);
707 		UVMHIST_LOG(maphist,"<- no backing store, no overlay",0,0,0,0);
708 		return (KERN_INVALID_ADDRESS);
709 	}
710 
711 	/*
712 	 * establish range of interest based on advice from mapper
713 	 * and then clip to fit map entry.   note that we only want
714 	 * to do this the first time through the fault.   if we
715 	 * ReFault we will disable this by setting "narrow" to true.
716 	 */
717 
718 	if (narrow == FALSE) {
719 
720 		/* wide fault (!narrow) */
721 		KASSERT(uvmadvice[ufi.entry->advice].advice ==
722 			 ufi.entry->advice);
723 		nback = min(uvmadvice[ufi.entry->advice].nback,
724 			    (ufi.orig_rvaddr - ufi.entry->start) >> PAGE_SHIFT);
725 		startva = ufi.orig_rvaddr - (nback << PAGE_SHIFT);
726 		nforw = min(uvmadvice[ufi.entry->advice].nforw,
727 			    ((ufi.entry->end - ufi.orig_rvaddr) >>
728 			     PAGE_SHIFT) - 1);
729 		/*
730 		 * note: "-1" because we don't want to count the
731 		 * faulting page as forw
732 		 */
733 		npages = nback + nforw + 1;
734 		centeridx = nback;
735 
736 		narrow = TRUE;	/* ensure only once per-fault */
737 
738 	} else {
739 
740 		/* narrow fault! */
741 		nback = nforw = 0;
742 		startva = ufi.orig_rvaddr;
743 		npages = 1;
744 		centeridx = 0;
745 
746 	}
747 
748 	/* locked: maps(read) */
749 	UVMHIST_LOG(maphist, "  narrow=%ld, back=%ld, forw=%ld, startva=0x%lx",
750 		    narrow, nback, nforw, startva);
751 	UVMHIST_LOG(maphist, "  entry=%p, amap=%p, obj=%p", ufi.entry,
752 		    amap, uobj, 0);
753 
754 	/*
755 	 * if we've got an amap, lock it and extract current anons.
756 	 */
757 
758 	if (amap) {
759 		amap_lock(amap);
760 		anons = anons_store;
761 		amap_lookups(&ufi.entry->aref, startva - ufi.entry->start,
762 		    anons, npages);
763 	} else {
764 		anons = NULL;	/* to be safe */
765 	}
766 
767 	/* locked: maps(read), amap(if there) */
768 
769 	/*
770 	 * for MADV_SEQUENTIAL mappings we want to deactivate the back pages
771 	 * now and then forget about them (for the rest of the fault).
772 	 */
773 
774 	if (ufi.entry->advice == MADV_SEQUENTIAL && nback != 0) {
775 
776 		UVMHIST_LOG(maphist, "  MADV_SEQUENTIAL: flushing backpages",
777 		    0,0,0,0);
778 		/* flush back-page anons? */
779 		if (amap)
780 			uvmfault_anonflush(anons, nback);
781 
782 		/* flush object? */
783 		if (uobj) {
784 			uoff = (startva - ufi.entry->start) + ufi.entry->offset;
785 			simple_lock(&uobj->vmobjlock);
786 			(void) uobj->pgops->pgo_flush(uobj, uoff, uoff +
787 				    (nback << PAGE_SHIFT), PGO_DEACTIVATE);
788 			simple_unlock(&uobj->vmobjlock);
789 		}
790 
791 		/* now forget about the backpages */
792 		if (amap)
793 			anons += nback;
794 		startva += (nback << PAGE_SHIFT);
795 		npages -= nback;
796 		nback = centeridx = 0;
797 	}
798 
799 	/* locked: maps(read), amap(if there) */
800 
801 	/*
802 	 * map in the backpages and frontpages we found in the amap in hopes
803 	 * of preventing future faults.    we also init the pages[] array as
804 	 * we go.
805 	 */
806 
807 	currva = startva;
808 	shadowed = FALSE;
809 	for (lcv = 0 ; lcv < npages ; lcv++, currva += PAGE_SIZE) {
810 
811 		/*
812 		 * dont play with VAs that are already mapped
813 		 * except for center)
814 		 */
815 		if (lcv != centeridx &&
816 		    pmap_extract(ufi.orig_map->pmap, currva, &pa)) {
817 			pages[lcv] = PGO_DONTCARE;
818 			continue;
819 		}
820 
821 		/*
822 		 * unmapped or center page.   check if any anon at this level.
823 		 */
824 		if (amap == NULL || anons[lcv] == NULL) {
825 			pages[lcv] = NULL;
826 			continue;
827 		}
828 
829 		/*
830 		 * check for present page and map if possible.   re-activate it.
831 		 */
832 
833 		pages[lcv] = PGO_DONTCARE;
834 		if (lcv == centeridx) {		/* save center for later! */
835 			shadowed = TRUE;
836 			continue;
837 		}
838 		anon = anons[lcv];
839 		simple_lock(&anon->an_lock);
840 		/* ignore loaned pages */
841 		if (anon->u.an_page && anon->u.an_page->loan_count == 0 &&
842 			(anon->u.an_page->flags & (PG_RELEASED|PG_BUSY)) == 0) {
843 			uvm_lock_pageq();
844 			uvm_pageactivate(anon->u.an_page);	/* reactivate */
845 			uvm_unlock_pageq();
846 			UVMHIST_LOG(maphist,
847 			    "  MAPPING: n anon: pm=%p, va=0x%lx, pg=%p",
848 			    ufi.orig_map->pmap, currva, anon->u.an_page, 0);
849 			uvmexp.fltnamap++;
850 
851 			/*
852 			 * Since this isn't the page that's actually faulting,
853 			 * ignore pmap_enter() failures; it's not critical
854 			 * that we enter these right now.
855 			 */
856 
857 			(void) pmap_enter(ufi.orig_map->pmap, currva,
858 			    VM_PAGE_TO_PHYS(anon->u.an_page),
859 			    (anon->an_ref > 1) ? (enter_prot & ~VM_PROT_WRITE) :
860 			    enter_prot,
861 			    PMAP_CANFAIL |
862 			     (VM_MAPENT_ISWIRED(ufi.entry) ? PMAP_WIRED : 0));
863 		}
864 		simple_unlock(&anon->an_lock);
865 		pmap_update(ufi.orig_map->pmap);
866 	}
867 
868 	/* locked: maps(read), amap(if there) */
869 	/* (shadowed == TRUE) if there is an anon at the faulting address */
870 	UVMHIST_LOG(maphist, "  shadowed=%ld, will_get=%ld", shadowed,
871 	    (uobj && shadowed == FALSE),0,0);
872 
873 	/*
874 	 * note that if we are really short of RAM we could sleep in the above
875 	 * call to pmap_enter with everything locked.   bad?
876 	 *
877 	 * XXX Actually, that is bad; pmap_enter() should just fail in that
878 	 * XXX case.  --thorpej
879 	 */
880 
881 	/*
882 	 * if the desired page is not shadowed by the amap and we have a
883 	 * backing object, then we check to see if the backing object would
884 	 * prefer to handle the fault itself (rather than letting us do it
885 	 * with the usual pgo_get hook).  the backing object signals this by
886 	 * providing a pgo_fault routine.
887 	 */
888 
889 	if (uobj && shadowed == FALSE && uobj->pgops->pgo_fault != NULL) {
890 		simple_lock(&uobj->vmobjlock);
891 
892 		/* locked: maps(read), amap (if there), uobj */
893 		result = uobj->pgops->pgo_fault(&ufi, startva, pages, npages,
894 				    centeridx, fault_type, access_type,
895 				    PGO_LOCKED);
896 
897 		/* locked: nothing, pgo_fault has unlocked everything */
898 
899 		if (result == VM_PAGER_OK)
900 			return (KERN_SUCCESS);	/* pgo_fault did pmap enter */
901 		else if (result == VM_PAGER_REFAULT)
902 			goto ReFault;		/* try again! */
903 		else
904 			return (KERN_PROTECTION_FAILURE);
905 	}
906 
907 	/*
908 	 * now, if the desired page is not shadowed by the amap and we have
909 	 * a backing object that does not have a special fault routine, then
910 	 * we ask (with pgo_get) the object for resident pages that we care
911 	 * about and attempt to map them in.  we do not let pgo_get block
912 	 * (PGO_LOCKED).
913 	 *
914 	 * ("get" has the option of doing a pmap_enter for us)
915 	 */
916 
917 	if (uobj && shadowed == FALSE) {
918 		simple_lock(&uobj->vmobjlock);
919 
920 		/* locked (!shadowed): maps(read), amap (if there), uobj */
921 		/*
922 		 * the following call to pgo_get does _not_ change locking state
923 		 */
924 
925 		uvmexp.fltlget++;
926 		gotpages = npages;
927 		(void) uobj->pgops->pgo_get(uobj, ufi.entry->offset +
928 				(startva - ufi.entry->start),
929 				pages, &gotpages, centeridx,
930 				access_type & MASK(ufi.entry),
931 				ufi.entry->advice, PGO_LOCKED);
932 
933 		/*
934 		 * check for pages to map, if we got any
935 		 */
936 
937 		uobjpage = NULL;
938 
939 		if (gotpages) {
940 			currva = startva;
941 			for (lcv = 0 ; lcv < npages ;
942 			    lcv++, currva += PAGE_SIZE) {
943 
944 				if (pages[lcv] == NULL ||
945 				    pages[lcv] == PGO_DONTCARE)
946 					continue;
947 
948 				KASSERT((pages[lcv]->flags & PG_RELEASED) == 0);
949 
950 				/*
951 				 * if center page is resident and not
952 				 * PG_BUSY|PG_RELEASED then pgo_get
953 				 * made it PG_BUSY for us and gave
954 				 * us a handle to it.   remember this
955 				 * page as "uobjpage." (for later use).
956 				 */
957 
958 				if (lcv == centeridx) {
959 					uobjpage = pages[lcv];
960 					UVMHIST_LOG(maphist, "  got uobjpage "
961 					    "(%p) with locked get",
962 					    uobjpage, 0,0,0);
963 					continue;
964 				}
965 
966 				/*
967 				 * note: calling pgo_get with locked data
968 				 * structures returns us pages which are
969 				 * neither busy nor released, so we don't
970 				 * need to check for this.   we can just
971 				 * directly enter the page (after moving it
972 				 * to the head of the active queue [useful?]).
973 				 */
974 
975 				uvm_lock_pageq();
976 				uvm_pageactivate(pages[lcv]);	/* reactivate */
977 				uvm_unlock_pageq();
978 				UVMHIST_LOG(maphist,
979 				  "  MAPPING: n obj: pm=%p, va=0x%lx, pg=%p",
980 				  ufi.orig_map->pmap, currva, pages[lcv], 0);
981 				uvmexp.fltnomap++;
982 
983 				/*
984 				 * Since this page isn't the page that's
985 				 * actually fauling, ignore pmap_enter()
986 				 * failures; it's not critical that we
987 				 * enter these right now.
988 				 */
989 
990 				(void) pmap_enter(ufi.orig_map->pmap, currva,
991 				    VM_PAGE_TO_PHYS(pages[lcv]),
992 				    enter_prot & MASK(ufi.entry),
993 				    PMAP_CANFAIL |
994 				     (wired ? PMAP_WIRED : 0));
995 
996 				/*
997 				 * NOTE: page can't be PG_WANTED or PG_RELEASED
998 				 * because we've held the lock the whole time
999 				 * we've had the handle.
1000 				 */
1001 
1002 				pages[lcv]->flags &= ~(PG_BUSY); /* un-busy! */
1003 				UVM_PAGE_OWN(pages[lcv], NULL);
1004 			}	/* for "lcv" loop */
1005 				pmap_update(ufi.orig_map->pmap);
1006 		}   /* "gotpages" != 0 */
1007 		/* note: object still _locked_ */
1008 	} else {
1009 		uobjpage = NULL;
1010 	}
1011 
1012 	/* locked (shadowed): maps(read), amap */
1013 	/* locked (!shadowed): maps(read), amap(if there),
1014 		 uobj(if !null), uobjpage(if !null) */
1015 
1016 	/*
1017 	 * note that at this point we are done with any front or back pages.
1018 	 * we are now going to focus on the center page (i.e. the one we've
1019 	 * faulted on).  if we have faulted on the top (anon) layer
1020 	 * [i.e. case 1], then the anon we want is anons[centeridx] (we have
1021 	 * not touched it yet).  if we have faulted on the bottom (uobj)
1022 	 * layer [i.e. case 2] and the page was both present and available,
1023 	 * then we've got a pointer to it as "uobjpage" and we've already
1024 	 * made it BUSY.
1025 	 */
1026 
1027 	/*
1028 	 * there are four possible cases we must address: 1A, 1B, 2A, and 2B
1029 	 */
1030 
1031 	/*
1032 	 * redirect case 2: if we are not shadowed, go to case 2.
1033 	 */
1034 
1035 	if (shadowed == FALSE)
1036 		goto Case2;
1037 
1038 	/* locked: maps(read), amap */
1039 
1040 	/*
1041 	 * handle case 1: fault on an anon in our amap
1042 	 */
1043 
1044 	anon = anons[centeridx];
1045 	UVMHIST_LOG(maphist, "  case 1 fault: anon=%p", anon, 0,0,0);
1046 	simple_lock(&anon->an_lock);
1047 
1048 	/* locked: maps(read), amap, anon */
1049 
1050 	/*
1051 	 * no matter if we have case 1A or case 1B we are going to need to
1052 	 * have the anon's memory resident.   ensure that now.
1053 	 */
1054 
1055 	/*
1056 	 * let uvmfault_anonget do the dirty work.
1057 	 * if it fails (!OK) it will unlock everything for us.
1058 	 * if it succeeds, locks are still valid and locked.
1059 	 * also, if it is OK, then the anon's page is on the queues.
1060 	 * if the page is on loan from a uvm_object, then anonget will
1061 	 * lock that object for us if it does not fail.
1062 	 */
1063 
1064 	result = uvmfault_anonget(&ufi, amap, anon);
1065 	switch (result) {
1066 	case VM_PAGER_OK:
1067 		break;
1068 
1069 	case VM_PAGER_REFAULT:
1070 		goto ReFault;
1071 
1072 	case VM_PAGER_ERROR:
1073 		/*
1074 		 * An error occured while trying to bring in the
1075 		 * page -- this is the only error we return right
1076 		 * now.
1077 		 */
1078 		return (KERN_PROTECTION_FAILURE);	/* XXX */
1079 
1080 	default:
1081 #ifdef DIAGNOSTIC
1082 		panic("uvm_fault: uvmfault_anonget -> %d", result);
1083 #else
1084 		return (KERN_PROTECTION_FAILURE);
1085 #endif
1086 	}
1087 
1088 	/*
1089 	 * uobj is non null if the page is on loan from an object (i.e. uobj)
1090 	 */
1091 
1092 	uobj = anon->u.an_page->uobject;	/* locked by anonget if !NULL */
1093 
1094 	/* locked: maps(read), amap, anon, uobj(if one) */
1095 
1096 	/*
1097 	 * special handling for loaned pages
1098 	 */
1099 
1100 	if (anon->u.an_page->loan_count) {
1101 
1102 		if ((access_type & VM_PROT_WRITE) == 0) {
1103 
1104 			/*
1105 			 * for read faults on loaned pages we just cap the
1106 			 * protection at read-only.
1107 			 */
1108 
1109 			enter_prot = enter_prot & ~VM_PROT_WRITE;
1110 
1111 		} else {
1112 			/*
1113 			 * note that we can't allow writes into a loaned page!
1114 			 *
1115 			 * if we have a write fault on a loaned page in an
1116 			 * anon then we need to look at the anon's ref count.
1117 			 * if it is greater than one then we are going to do
1118 			 * a normal copy-on-write fault into a new anon (this
1119 			 * is not a problem).  however, if the reference count
1120 			 * is one (a case where we would normally allow a
1121 			 * write directly to the page) then we need to kill
1122 			 * the loan before we continue.
1123 			 */
1124 
1125 			/* >1 case is already ok */
1126 			if (anon->an_ref == 1) {
1127 
1128 				/* get new un-owned replacement page */
1129 				pg = uvm_pagealloc(NULL, 0, NULL, 0);
1130 				if (pg == NULL) {
1131 					uvmfault_unlockall(&ufi, amap, uobj,
1132 					    anon);
1133 					uvm_wait("flt_noram2");
1134 					goto ReFault;
1135 				}
1136 
1137 				/*
1138 				 * copy data, kill loan, and drop uobj lock
1139 				 * (if any)
1140 				 */
1141 				/* copy old -> new */
1142 				uvm_pagecopy(anon->u.an_page, pg);
1143 
1144 				/* force reload */
1145 				pmap_page_protect(anon->u.an_page,
1146 						  VM_PROT_NONE);
1147 				uvm_lock_pageq();	  /* KILL loan */
1148 				if (uobj)
1149 					/* if we were loaning */
1150 					anon->u.an_page->loan_count--;
1151 				anon->u.an_page->uanon = NULL;
1152 				/* in case we owned */
1153 				anon->u.an_page->pqflags &= ~PQ_ANON;
1154 				uvm_pageactivate(pg);
1155 				uvm_unlock_pageq();
1156 				if (uobj) {
1157 					simple_unlock(&uobj->vmobjlock);
1158 					uobj = NULL;
1159 				}
1160 
1161 				/* install new page in anon */
1162 				anon->u.an_page = pg;
1163 				pg->uanon = anon;
1164 				pg->pqflags |= PQ_ANON;
1165 				pg->flags &= ~(PG_BUSY|PG_FAKE);
1166 				UVM_PAGE_OWN(pg, NULL);
1167 
1168 				/* done! */
1169 			}     /* ref == 1 */
1170 		}       /* write fault */
1171 	}         /* loan count */
1172 
1173 	/*
1174 	 * if we are case 1B then we will need to allocate a new blank
1175 	 * anon to transfer the data into.   note that we have a lock
1176 	 * on anon, so no one can busy or release the page until we are done.
1177 	 * also note that the ref count can't drop to zero here because
1178 	 * it is > 1 and we are only dropping one ref.
1179 	 *
1180 	 * in the (hopefully very rare) case that we are out of RAM we
1181 	 * will unlock, wait for more RAM, and refault.
1182 	 *
1183 	 * if we are out of anon VM we kill the process (XXX: could wait?).
1184 	 */
1185 
1186 	if ((access_type & VM_PROT_WRITE) != 0 && anon->an_ref > 1) {
1187 
1188 		UVMHIST_LOG(maphist, "  case 1B: COW fault",0,0,0,0);
1189 		uvmexp.flt_acow++;
1190 		oanon = anon;		/* oanon = old, locked anon */
1191 		anon = uvm_analloc();
1192 		if (anon) {
1193 			pg = uvm_pagealloc(NULL, 0, anon, 0);
1194 		}
1195 
1196 		/* check for out of RAM */
1197 		if (anon == NULL || pg == NULL) {
1198 			if (anon)
1199 				uvm_anfree(anon);
1200 			uvmfault_unlockall(&ufi, amap, uobj, oanon);
1201 			KASSERT(uvmexp.swpgonly <= uvmexp.swpages);
1202 			if (anon == NULL || uvmexp.swpgonly == uvmexp.swpages) {
1203 				UVMHIST_LOG(maphist,
1204 				    "<- failed.  out of VM",0,0,0,0);
1205 				uvmexp.fltnoanon++;
1206 				return (KERN_RESOURCE_SHORTAGE);
1207 			}
1208 
1209 			uvmexp.fltnoram++;
1210 			uvm_wait("flt_noram3");	/* out of RAM, wait for more */
1211 			goto ReFault;
1212 		}
1213 
1214 		/* got all resources, replace anon with nanon */
1215 
1216 		uvm_pagecopy(oanon->u.an_page, pg);	/* pg now !PG_CLEAN */
1217 		pg->flags &= ~(PG_BUSY|PG_FAKE);	/* un-busy! new page */
1218 		UVM_PAGE_OWN(pg, NULL);
1219 		amap_add(&ufi.entry->aref, ufi.orig_rvaddr - ufi.entry->start,
1220 		    anon, 1);
1221 
1222 		/* deref: can not drop to zero here by defn! */
1223 		oanon->an_ref--;
1224 
1225 		/*
1226 		 * note: oanon still locked.   anon is _not_ locked, but we
1227 		 * have the sole references to in from amap which _is_ locked.
1228 		 * thus, no one can get at it until we are done with it.
1229 		 */
1230 
1231 	} else {
1232 
1233 		uvmexp.flt_anon++;
1234 		oanon = anon;		/* old, locked anon is same as anon */
1235 		pg = anon->u.an_page;
1236 		if (anon->an_ref > 1)     /* disallow writes to ref > 1 anons */
1237 			enter_prot = enter_prot & ~VM_PROT_WRITE;
1238 
1239 	}
1240 
1241 	/* locked: maps(read), amap, oanon */
1242 
1243 	/*
1244 	 * now map the page in ...
1245 	 * XXX: old fault unlocks object before pmap_enter.  this seems
1246 	 * suspect since some other thread could blast the page out from
1247 	 * under us between the unlock and the pmap_enter.
1248 	 */
1249 
1250 	UVMHIST_LOG(maphist, "  MAPPING: anon: pm=%p, va=0x%lx, pg=%p",
1251 	    ufi.orig_map->pmap, ufi.orig_rvaddr, pg, 0);
1252 	if (pmap_enter(ufi.orig_map->pmap, ufi.orig_rvaddr, VM_PAGE_TO_PHYS(pg),
1253 	    enter_prot, access_type | PMAP_CANFAIL | (wired ? PMAP_WIRED : 0))
1254 	    != KERN_SUCCESS) {
1255 		/*
1256 		 * No need to undo what we did; we can simply think of
1257 		 * this as the pmap throwing away the mapping information.
1258 		 *
1259 		 * We do, however, have to go through the ReFault path,
1260 		 * as the map may change while we're asleep.
1261 		 */
1262 		uvmfault_unlockall(&ufi, amap, uobj, oanon);
1263 		KASSERT(uvmexp.swpgonly <= uvmexp.swpages);
1264 		if (uvmexp.swpgonly == uvmexp.swpages) {
1265 			UVMHIST_LOG(maphist,
1266 			    "<- failed.  out of VM",0,0,0,0);
1267 			/* XXX instrumentation */
1268 			return (KERN_RESOURCE_SHORTAGE);
1269 		}
1270 		/* XXX instrumentation */
1271 		uvm_wait("flt_pmfail1");
1272 		goto ReFault;
1273 	}
1274 
1275 	/*
1276 	 * ... update the page queues.
1277 	 */
1278 
1279 	uvm_lock_pageq();
1280 
1281 	if (fault_type == VM_FAULT_WIRE) {
1282 		uvm_pagewire(pg);
1283 
1284 		/*
1285 		 * since the now-wired page cannot be paged out,
1286 		 * release its swap resources for others to use.
1287 		 * since an anon with no swap cannot be PG_CLEAN,
1288 		 * clear its clean flag now.
1289 		 */
1290 
1291 		pg->flags &= ~(PG_CLEAN);
1292 		uvm_anon_dropswap(anon);
1293 	} else {
1294 		/* activate it */
1295 		uvm_pageactivate(pg);
1296 	}
1297 
1298 	uvm_unlock_pageq();
1299 
1300 	/*
1301 	 * done case 1!  finish up by unlocking everything and returning success
1302 	 */
1303 
1304 	uvmfault_unlockall(&ufi, amap, uobj, oanon);
1305 	pmap_update(ufi.orig_map->pmap);
1306 	return (KERN_SUCCESS);
1307 
1308 
1309 Case2:
1310 	/*
1311 	 * handle case 2: faulting on backing object or zero fill
1312 	 */
1313 
1314 	/*
1315 	 * locked:
1316 	 * maps(read), amap(if there), uobj(if !null), uobjpage(if !null)
1317 	 */
1318 
1319 	/*
1320 	 * note that uobjpage can not be PGO_DONTCARE at this point.  we now
1321 	 * set uobjpage to PGO_DONTCARE if we are doing a zero fill.  if we
1322 	 * have a backing object, check and see if we are going to promote
1323 	 * the data up to an anon during the fault.
1324 	 */
1325 
1326 	if (uobj == NULL) {
1327 		uobjpage = PGO_DONTCARE;
1328 		promote = TRUE;		/* always need anon here */
1329 	} else {
1330 		KASSERT(uobjpage != PGO_DONTCARE);
1331 		promote = (access_type & VM_PROT_WRITE) &&
1332 		     UVM_ET_ISCOPYONWRITE(ufi.entry);
1333 	}
1334 	UVMHIST_LOG(maphist, "  case 2 fault: promote=%ld, zfill=%ld",
1335 	    promote, (uobj == NULL), 0,0);
1336 
1337 	/*
1338 	 * if uobjpage is not null then we do not need to do I/O to get the
1339 	 * uobjpage.
1340 	 *
1341 	 * if uobjpage is null, then we need to unlock and ask the pager to
1342 	 * get the data for us.   once we have the data, we need to reverify
1343 	 * the state the world.   we are currently not holding any resources.
1344 	 */
1345 
1346 	if (uobjpage) {
1347 		/* update rusage counters */
1348 		curproc->p_addr->u_stats.p_ru.ru_minflt++;
1349 	} else {
1350 		/* update rusage counters */
1351 		curproc->p_addr->u_stats.p_ru.ru_majflt++;
1352 
1353 		/* locked: maps(read), amap(if there), uobj */
1354 		uvmfault_unlockall(&ufi, amap, NULL, NULL);
1355 		/* locked: uobj */
1356 
1357 		uvmexp.fltget++;
1358 		gotpages = 1;
1359 		uoff = (ufi.orig_rvaddr - ufi.entry->start) + ufi.entry->offset;
1360 		result = uobj->pgops->pgo_get(uobj, uoff, &uobjpage, &gotpages,
1361 		    0, access_type & MASK(ufi.entry), ufi.entry->advice,
1362 		    PGO_SYNCIO);
1363 
1364 		/* locked: uobjpage(if result OK) */
1365 
1366 		/*
1367 		 * recover from I/O
1368 		 */
1369 
1370 		if (result != VM_PAGER_OK) {
1371 			KASSERT(result != VM_PAGER_PEND);
1372 
1373 			if (result == VM_PAGER_AGAIN) {
1374 				UVMHIST_LOG(maphist,
1375 				    "  pgo_get says TRY AGAIN!",0,0,0,0);
1376 				tsleep((caddr_t)&lbolt, PVM, "fltagain2", 0);
1377 				goto ReFault;
1378 			}
1379 
1380 			UVMHIST_LOG(maphist, "<- pgo_get failed (code %ld)",
1381 			    result, 0,0,0);
1382 			return (KERN_PROTECTION_FAILURE); /* XXX i/o error */
1383 		}
1384 
1385 		/* locked: uobjpage */
1386 
1387 		/*
1388 		 * re-verify the state of the world by first trying to relock
1389 		 * the maps.  always relock the object.
1390 		 */
1391 
1392 		locked = uvmfault_relock(&ufi);
1393 		if (locked && amap)
1394 			amap_lock(amap);
1395 		simple_lock(&uobj->vmobjlock);
1396 
1397 		/* locked(locked): maps(read), amap(if !null), uobj, uobjpage */
1398 		/* locked(!locked): uobj, uobjpage */
1399 
1400 		/*
1401 		 * verify that the page has not be released and re-verify
1402 		 * that amap slot is still free.   if there is a problem,
1403 		 * we unlock and clean up.
1404 		 */
1405 
1406 		if ((uobjpage->flags & PG_RELEASED) != 0 ||
1407 		    (locked && amap &&
1408 		    amap_lookup(&ufi.entry->aref,
1409 		      ufi.orig_rvaddr - ufi.entry->start))) {
1410 			if (locked)
1411 				uvmfault_unlockall(&ufi, amap, NULL, NULL);
1412 			locked = FALSE;
1413 		}
1414 
1415 		/*
1416 		 * didn't get the lock?   release the page and retry.
1417 		 */
1418 
1419 		if (locked == FALSE) {
1420 
1421 			UVMHIST_LOG(maphist,
1422 			    "  wasn't able to relock after fault: retry",
1423 			    0,0,0,0);
1424 			if (uobjpage->flags & PG_WANTED)
1425 				/* still holding object lock */
1426 				wakeup(uobjpage);
1427 
1428 			if (uobjpage->flags & PG_RELEASED) {
1429 				uvmexp.fltpgrele++;
1430 				KASSERT(uobj->pgops->pgo_releasepg != NULL);
1431 
1432 				/* frees page */
1433 				if (uobj->pgops->pgo_releasepg(uobjpage,NULL))
1434 					/* unlock if still alive */
1435 					simple_unlock(&uobj->vmobjlock);
1436 				goto ReFault;
1437 			}
1438 
1439 			uvm_lock_pageq();
1440 			/* make sure it is in queues */
1441 			uvm_pageactivate(uobjpage);
1442 
1443 			uvm_unlock_pageq();
1444 			uobjpage->flags &= ~(PG_BUSY|PG_WANTED);
1445 			UVM_PAGE_OWN(uobjpage, NULL);
1446 			simple_unlock(&uobj->vmobjlock);
1447 			goto ReFault;
1448 
1449 		}
1450 
1451 		/*
1452 		 * we have the data in uobjpage which is PG_BUSY and
1453 		 * !PG_RELEASED.  we are holding object lock (so the page
1454 		 * can't be released on us).
1455 		 */
1456 
1457 		/* locked: maps(read), amap(if !null), uobj, uobjpage */
1458 	}
1459 
1460 	/*
1461 	 * locked:
1462 	 * maps(read), amap(if !null), uobj(if !null), uobjpage(if uobj)
1463 	 */
1464 
1465 	/*
1466 	 * notes:
1467 	 *  - at this point uobjpage can not be NULL
1468 	 *  - at this point uobjpage can not be PG_RELEASED (since we checked
1469 	 *  for it above)
1470 	 *  - at this point uobjpage could be PG_WANTED (handle later)
1471 	 */
1472 
1473 	if (promote == FALSE) {
1474 
1475 		/*
1476 		 * we are not promoting.   if the mapping is COW ensure that we
1477 		 * don't give more access than we should (e.g. when doing a read
1478 		 * fault on a COPYONWRITE mapping we want to map the COW page in
1479 		 * R/O even though the entry protection could be R/W).
1480 		 *
1481 		 * set "pg" to the page we want to map in (uobjpage, usually)
1482 		 */
1483 
1484 		uvmexp.flt_obj++;
1485 		if (UVM_ET_ISCOPYONWRITE(ufi.entry))
1486 			enter_prot &= ~VM_PROT_WRITE;
1487 		pg = uobjpage;		/* map in the actual object */
1488 
1489 		/* assert(uobjpage != PGO_DONTCARE) */
1490 
1491 		/*
1492 		 * we are faulting directly on the page.   be careful
1493 		 * about writing to loaned pages...
1494 		 */
1495 		if (uobjpage->loan_count) {
1496 
1497 			if ((access_type & VM_PROT_WRITE) == 0) {
1498 				/* read fault: cap the protection at readonly */
1499 				/* cap! */
1500 				enter_prot = enter_prot & ~VM_PROT_WRITE;
1501 			} else {
1502 				/* write fault: must break the loan here */
1503 
1504 				/* alloc new un-owned page */
1505 				pg = uvm_pagealloc(NULL, 0, NULL, 0);
1506 
1507 				if (pg == NULL) {
1508 					/*
1509 					 * drop ownership of page, it can't
1510 					 * be released
1511 					 */
1512 					if (uobjpage->flags & PG_WANTED)
1513 						wakeup(uobjpage);
1514 					uobjpage->flags &= ~(PG_BUSY|PG_WANTED);
1515 					UVM_PAGE_OWN(uobjpage, NULL);
1516 
1517 					uvm_lock_pageq();
1518 					/* activate: we will need it later */
1519 					uvm_pageactivate(uobjpage);
1520 
1521 					uvm_unlock_pageq();
1522 					uvmfault_unlockall(&ufi, amap, uobj,
1523 					  NULL);
1524 					UVMHIST_LOG(maphist,
1525 					  "  out of RAM breaking loan, waiting",
1526 					  0,0,0,0);
1527 					uvmexp.fltnoram++;
1528 					uvm_wait("flt_noram4");
1529 					goto ReFault;
1530 				}
1531 
1532 				/*
1533 				 * copy the data from the old page to the new
1534 				 * one and clear the fake/clean flags on the
1535 				 * new page (keep it busy).  force a reload
1536 				 * of the old page by clearing it from all
1537 				 * pmaps.  then lock the page queues to
1538 				 * rename the pages.
1539 				 */
1540 				uvm_pagecopy(uobjpage, pg);	/* old -> new */
1541 				pg->flags &= ~(PG_FAKE|PG_CLEAN);
1542 				pmap_page_protect(uobjpage, VM_PROT_NONE);
1543 				if (uobjpage->flags & PG_WANTED)
1544 					wakeup(uobjpage);
1545 				/* uobj still locked */
1546 				uobjpage->flags &= ~(PG_WANTED|PG_BUSY);
1547 				UVM_PAGE_OWN(uobjpage, NULL);
1548 
1549 				uvm_lock_pageq();
1550 				uoff = uobjpage->offset;
1551 				/* remove old page */
1552 				uvm_pagerealloc(uobjpage, NULL, 0);
1553 
1554 				/*
1555 				 * at this point we have absolutely no
1556 				 * control over uobjpage
1557 				 */
1558 				/* install new page */
1559 				uvm_pagerealloc(pg, uobj, uoff);
1560 				uvm_unlock_pageq();
1561 
1562 				/*
1563 				 * done!  loan is broken and "pg" is
1564 				 * PG_BUSY.   it can now replace uobjpage.
1565 				 */
1566 
1567 				uobjpage = pg;
1568 
1569 			}		/* write fault case */
1570 		}		/* if loan_count */
1571 
1572 	} else {
1573 
1574 		/*
1575 		 * if we are going to promote the data to an anon we
1576 		 * allocate a blank anon here and plug it into our amap.
1577 		 */
1578 #ifdef DIAGNOSTIC
1579 		if (amap == NULL)
1580 			panic("uvm_fault: want to promote data, but no anon");
1581 #endif
1582 
1583 		anon = uvm_analloc();
1584 		if (anon) {
1585 			/*
1586 			 * In `Fill in data...' below, if
1587 			 * uobjpage == PGO_DONTCARE, we want
1588 			 * a zero'd, dirty page, so have
1589 			 * uvm_pagealloc() do that for us.
1590 			 */
1591 			pg = uvm_pagealloc(NULL, 0, anon,
1592 			    (uobjpage == PGO_DONTCARE) ? UVM_PGA_ZERO : 0);
1593 		}
1594 
1595 		/*
1596 		 * out of memory resources?
1597 		 */
1598 		if (anon == NULL || pg == NULL) {
1599 
1600 			/*
1601 			 * arg!  must unbusy our page and fail or sleep.
1602 			 */
1603 			if (uobjpage != PGO_DONTCARE) {
1604 				if (uobjpage->flags & PG_WANTED)
1605 					/* still holding object lock */
1606 					wakeup(uobjpage);
1607 
1608 				uvm_lock_pageq();
1609 				uvm_pageactivate(uobjpage);
1610 				uvm_unlock_pageq();
1611 				uobjpage->flags &= ~(PG_BUSY|PG_WANTED);
1612 				UVM_PAGE_OWN(uobjpage, NULL);
1613 			}
1614 
1615 			/* unlock and fail ... */
1616 			uvmfault_unlockall(&ufi, amap, uobj, NULL);
1617 			KASSERT(uvmexp.swpgonly <= uvmexp.swpages);
1618 			if (anon == NULL || uvmexp.swpgonly == uvmexp.swpages) {
1619 				UVMHIST_LOG(maphist, "  promote: out of VM",
1620 				    0,0,0,0);
1621 				uvmexp.fltnoanon++;
1622 				return (KERN_RESOURCE_SHORTAGE);
1623 			}
1624 
1625 			UVMHIST_LOG(maphist, "  out of RAM, waiting for more",
1626 			    0,0,0,0);
1627 			uvm_anfree(anon);
1628 			uvmexp.fltnoram++;
1629 			uvm_wait("flt_noram5");
1630 			goto ReFault;
1631 		}
1632 
1633 		/*
1634 		 * fill in the data
1635 		 */
1636 
1637 		if (uobjpage != PGO_DONTCARE) {
1638 			uvmexp.flt_prcopy++;
1639 			/* copy page [pg now dirty] */
1640 			uvm_pagecopy(uobjpage, pg);
1641 
1642 			/*
1643 			 * promote to shared amap?  make sure all sharing
1644 			 * procs see it
1645 			 */
1646 			if ((amap_flags(amap) & AMAP_SHARED) != 0) {
1647 				pmap_page_protect(uobjpage, VM_PROT_NONE);
1648 			}
1649 
1650 			/*
1651 			 * dispose of uobjpage.  it can't be PG_RELEASED
1652 			 * since we still hold the object lock.
1653 			 * drop handle to uobj as well.
1654 			 */
1655 
1656 			if (uobjpage->flags & PG_WANTED)
1657 				/* still have the obj lock */
1658 				wakeup(uobjpage);
1659 			uobjpage->flags &= ~(PG_BUSY|PG_WANTED);
1660 			UVM_PAGE_OWN(uobjpage, NULL);
1661 			uvm_lock_pageq();
1662 			uvm_pageactivate(uobjpage);
1663 			uvm_unlock_pageq();
1664 			simple_unlock(&uobj->vmobjlock);
1665 			uobj = NULL;
1666 
1667 			UVMHIST_LOG(maphist,
1668 			    "  promote uobjpage %p to anon/page %p/%p",
1669 			    uobjpage, anon, pg, 0);
1670 
1671 		} else {
1672 			uvmexp.flt_przero++;
1673 			/*
1674 			 * Page is zero'd and marked dirty by uvm_pagealloc()
1675 			 * above.
1676 			 */
1677 			UVMHIST_LOG(maphist,"  zero fill anon/page %p/%p",
1678 			    anon, pg, 0, 0);
1679 		}
1680 
1681 		amap_add(&ufi.entry->aref, ufi.orig_rvaddr - ufi.entry->start,
1682 		    anon, 0);
1683 	}
1684 
1685 	/*
1686 	 * locked:
1687 	 * maps(read), amap(if !null), uobj(if !null), uobjpage(if uobj)
1688 	 *
1689 	 * note: pg is either the uobjpage or the new page in the new anon
1690 	 */
1691 
1692 	/*
1693 	 * all resources are present.   we can now map it in and free our
1694 	 * resources.
1695 	 */
1696 
1697 	UVMHIST_LOG(maphist,
1698 	    "  MAPPING: case2: pm=%p, va=0x%lx, pg=%p, promote=%ld",
1699 	    ufi.orig_map->pmap, ufi.orig_rvaddr, pg, promote);
1700 	if (pmap_enter(ufi.orig_map->pmap, ufi.orig_rvaddr, VM_PAGE_TO_PHYS(pg),
1701 	    enter_prot, access_type | PMAP_CANFAIL | (wired ? PMAP_WIRED : 0))
1702 	    != KERN_SUCCESS) {
1703 
1704 		/*
1705 		 * No need to undo what we did; we can simply think of
1706 		 * this as the pmap throwing away the mapping information.
1707 		 *
1708 		 * We do, however, have to go through the ReFault path,
1709 		 * as the map may change while we're asleep.
1710 		 */
1711 
1712 		if (pg->flags & PG_WANTED)
1713 			wakeup(pg);		/* lock still held */
1714 
1715 		/*
1716 		 * note that pg can't be PG_RELEASED since we did not drop
1717 		 * the object lock since the last time we checked.
1718 		 */
1719 
1720 		pg->flags &= ~(PG_BUSY|PG_FAKE|PG_WANTED);
1721 		UVM_PAGE_OWN(pg, NULL);
1722 		uvmfault_unlockall(&ufi, amap, uobj, NULL);
1723 		KASSERT(uvmexp.swpgonly <= uvmexp.swpages);
1724 		if (uvmexp.swpgonly == uvmexp.swpages) {
1725 			UVMHIST_LOG(maphist,
1726 			    "<- failed.  out of VM",0,0,0,0);
1727 			/* XXX instrumentation */
1728 			return (KERN_RESOURCE_SHORTAGE);
1729 		}
1730 		/* XXX instrumentation */
1731 		uvm_wait("flt_pmfail2");
1732 		goto ReFault;
1733 	}
1734 
1735 	uvm_lock_pageq();
1736 
1737 	if (fault_type == VM_FAULT_WIRE) {
1738 		uvm_pagewire(pg);
1739 		if (pg->pqflags & PQ_AOBJ) {
1740 
1741 			/*
1742 			 * since the now-wired page cannot be paged out,
1743 			 * release its swap resources for others to use.
1744 			 * since an aobj page with no swap cannot be PG_CLEAN,
1745 			 * clear its clean flag now.
1746 			 */
1747 
1748 			pg->flags &= ~(PG_CLEAN);
1749 			uao_dropswap(uobj, pg->offset >> PAGE_SHIFT);
1750 		}
1751 	} else {
1752 		/* activate it */
1753 		uvm_pageactivate(pg);
1754 	}
1755 	uvm_unlock_pageq();
1756 
1757 	if (pg->flags & PG_WANTED)
1758 		wakeup(pg);		/* lock still held */
1759 
1760 	/*
1761 	 * note that pg can't be PG_RELEASED since we did not drop the object
1762 	 * lock since the last time we checked.
1763 	 */
1764 
1765 	pg->flags &= ~(PG_BUSY|PG_FAKE|PG_WANTED);
1766 	UVM_PAGE_OWN(pg, NULL);
1767 	uvmfault_unlockall(&ufi, amap, uobj, NULL);
1768 	pmap_update(ufi.orig_map->pmap);
1769 
1770 	UVMHIST_LOG(maphist, "<- done (SUCCESS!)",0,0,0,0);
1771 	return (KERN_SUCCESS);
1772 }
1773 
1774 
1775 /*
1776  * uvm_fault_wire: wire down a range of virtual addresses in a map.
1777  *
1778  * => map may be read-locked by caller, but MUST NOT be write-locked.
1779  * => if map is read-locked, any operations which may cause map to
1780  *	be write-locked in uvm_fault() must be taken care of by
1781  *	the caller.  See uvm_map_pageable().
1782  */
1783 
1784 int
1785 uvm_fault_wire(map, start, end, access_type)
1786 	vm_map_t map;
1787 	vaddr_t start, end;
1788 	vm_prot_t access_type;
1789 {
1790 	vaddr_t va;
1791 	pmap_t  pmap;
1792 	int rv;
1793 
1794 	pmap = vm_map_pmap(map);
1795 
1796 	/*
1797 	 * now fault it in a page at a time.   if the fault fails then we have
1798 	 * to undo what we have done.   note that in uvm_fault VM_PROT_NONE
1799 	 * is replaced with the max protection if fault_type is VM_FAULT_WIRE.
1800 	 */
1801 
1802 	for (va = start ; va < end ; va += PAGE_SIZE) {
1803 		rv = uvm_fault(map, va, VM_FAULT_WIRE, access_type);
1804 		if (rv) {
1805 			if (va != start) {
1806 				uvm_fault_unwire(map, start, va);
1807 			}
1808 			return (rv);
1809 		}
1810 	}
1811 
1812 	return (KERN_SUCCESS);
1813 }
1814 
1815 /*
1816  * uvm_fault_unwire(): unwire range of virtual space.
1817  */
1818 
1819 void
1820 uvm_fault_unwire(map, start, end)
1821 	vm_map_t map;
1822 	vaddr_t start, end;
1823 {
1824 
1825 	vm_map_lock_read(map);
1826 	uvm_fault_unwire_locked(map, start, end);
1827 	vm_map_unlock_read(map);
1828 }
1829 
1830 /*
1831  * uvm_fault_unwire_locked(): the guts of uvm_fault_unwire().
1832  *
1833  * => map must be at least read-locked.
1834  */
1835 
1836 void
1837 uvm_fault_unwire_locked(map, start, end)
1838 	vm_map_t map;
1839 	vaddr_t start, end;
1840 {
1841 	vm_map_entry_t entry;
1842 	pmap_t pmap = vm_map_pmap(map);
1843 	vaddr_t va;
1844 	paddr_t pa;
1845 	struct vm_page *pg;
1846 
1847 	KASSERT((map->flags & VM_MAP_INTRSAFE) == 0);
1848 
1849 	/*
1850 	 * we assume that the area we are unwiring has actually been wired
1851 	 * in the first place.   this means that we should be able to extract
1852 	 * the PAs from the pmap.   we also lock out the page daemon so that
1853 	 * we can call uvm_pageunwire.
1854 	 */
1855 
1856 	uvm_lock_pageq();
1857 
1858 	/*
1859 	 * find the beginning map entry for the region.
1860 	 */
1861 	KASSERT(start >= vm_map_min(map) && end <= vm_map_max(map));
1862 	if (uvm_map_lookup_entry(map, start, &entry) == FALSE)
1863 		panic("uvm_fault_unwire_locked: address not in map");
1864 
1865 	for (va = start; va < end ; va += PAGE_SIZE) {
1866 		if (pmap_extract(pmap, va, &pa) == FALSE)
1867 			continue;
1868 
1869 		/*
1870 		 * find the map entry for the current address.
1871 		 */
1872 		KASSERT(va >= entry->start);
1873 		while (va >= entry->end) {
1874 			KASSERT(entry->next != &map->header &&
1875 				entry->next->start <= entry->end);
1876 			entry = entry->next;
1877 		}
1878 
1879 		/*
1880 		 * if the entry is no longer wired, tell the pmap.
1881 		 */
1882 		if (VM_MAPENT_ISWIRED(entry) == 0)
1883 			pmap_unwire(pmap, va);
1884 
1885 		pg = PHYS_TO_VM_PAGE(pa);
1886 		if (pg)
1887 			uvm_pageunwire(pg);
1888 	}
1889 
1890 	uvm_unlock_pageq();
1891 }
1892