xref: /openbsd-src/sys/uvm/uvm_fault.c (revision 850e275390052b330d93020bf619a739a3c277ac)
1 /*	$OpenBSD: uvm_fault.c,v 1.50 2008/09/16 18:52:52 chl 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]->an_page;
204 		if (pg && (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->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->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->pg_flags & (PG_BUSY|PG_RELEASED)) == 0) {
346 				UVMHIST_LOG(maphist, "<- OK",0,0,0,0);
347 				return (VM_PAGER_OK);
348 			}
349 			atomic_setbits_int(&pg->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 || we_own)
420 			simple_lock(&anon->an_lock);
421 
422 		/*
423 		 * if we own the page (i.e. we set PG_BUSY), then we need
424 		 * to clean up after the I/O. there are three cases to
425 		 * consider:
426 		 *   [1] page released during I/O: free anon and ReFault.
427 		 *   [2] I/O not OK.   free the page and cause the fault
428 		 *       to fail.
429 		 *   [3] I/O OK!   activate the page and sync with the
430 		 *       non-we_own case (i.e. drop anon lock if not locked).
431 		 */
432 
433 		if (we_own) {
434 
435 			if (pg->pg_flags & PG_WANTED) {
436 				/* still holding object lock */
437 				wakeup(pg);
438 			}
439 			/* un-busy! */
440 			atomic_clearbits_int(&pg->pg_flags,
441 			    PG_WANTED|PG_BUSY|PG_FAKE);
442 			UVM_PAGE_OWN(pg, NULL);
443 
444 			/*
445 			 * if we were RELEASED during I/O, then our anon is
446 			 * no longer part of an amap.   we need to free the
447 			 * anon and try again.
448 			 */
449 			if (pg->pg_flags & PG_RELEASED) {
450 				pmap_page_protect(pg, VM_PROT_NONE);
451 				simple_unlock(&anon->an_lock);
452 				uvm_anfree(anon);	/* frees page for us */
453 				if (locked)
454 					uvmfault_unlockall(ufi, amap, NULL,
455 							   NULL);
456 				uvmexp.fltpgrele++;
457 				UVMHIST_LOG(maphist, "<- REFAULT", 0,0,0,0);
458 				return (VM_PAGER_REFAULT);	/* refault! */
459 			}
460 
461 			if (result != VM_PAGER_OK) {
462 				KASSERT(result != VM_PAGER_PEND);
463 
464 				/* remove page from anon */
465 				anon->an_page = NULL;
466 
467 				/*
468 				 * remove the swap slot from the anon
469 				 * and mark the anon as having no real slot.
470 				 * don't free the swap slot, thus preventing
471 				 * it from being used again.
472 				 */
473 				uvm_swap_markbad(anon->an_swslot, 1);
474 				anon->an_swslot = SWSLOT_BAD;
475 
476 				/*
477 				 * note: page was never !PG_BUSY, so it
478 				 * can't be mapped and thus no need to
479 				 * pmap_page_protect it...
480 				 */
481 				uvm_lock_pageq();
482 				uvm_pagefree(pg);
483 				uvm_unlock_pageq();
484 
485 				if (locked)
486 					uvmfault_unlockall(ufi, amap, NULL,
487 					    anon);
488 				else
489 					simple_unlock(&anon->an_lock);
490 				UVMHIST_LOG(maphist, "<- ERROR", 0,0,0,0);
491 				return (VM_PAGER_ERROR);
492 			}
493 
494 			/*
495 			 * must be OK, clear modify (already PG_CLEAN)
496 			 * and activate
497 			 */
498 			pmap_clear_modify(pg);
499 			uvm_lock_pageq();
500 			uvm_pageactivate(pg);
501 			uvm_unlock_pageq();
502 			if (!locked)
503 				simple_unlock(&anon->an_lock);
504 		}
505 
506 		/*
507 		 * we were not able to relock.   restart fault.
508 		 */
509 
510 		if (!locked) {
511 			UVMHIST_LOG(maphist, "<- REFAULT", 0,0,0,0);
512 			return (VM_PAGER_REFAULT);
513 		}
514 
515 		/*
516 		 * verify no one has touched the amap and moved the anon on us.
517 		 */
518 
519 		if (ufi != NULL &&
520 		    amap_lookup(&ufi->entry->aref,
521 				ufi->orig_rvaddr - ufi->entry->start) != anon) {
522 
523 			uvmfault_unlockall(ufi, amap, NULL, anon);
524 			UVMHIST_LOG(maphist, "<- REFAULT", 0,0,0,0);
525 			return (VM_PAGER_REFAULT);
526 		}
527 
528 		/*
529 		 * try it again!
530 		 */
531 
532 		uvmexp.fltanretry++;
533 		continue;
534 
535 	} /* while (1) */
536 
537 	/*NOTREACHED*/
538 }
539 
540 /*
541  *   F A U L T   -   m a i n   e n t r y   p o i n t
542  */
543 
544 /*
545  * uvm_fault: page fault handler
546  *
547  * => called from MD code to resolve a page fault
548  * => VM data structures usually should be unlocked.   however, it is
549  *	possible to call here with the main map locked if the caller
550  *	gets a write lock, sets it recursive, and then calls us (c.f.
551  *	uvm_map_pageable).   this should be avoided because it keeps
552  *	the map locked off during I/O.
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 	vm_map_t 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, result, lcv, gotpages;
569 	vaddr_t startva, currva;
570 	voff_t uoff;
571 	paddr_t pa;
572 	struct vm_amap *amap;
573 	struct uvm_object *uobj;
574 	struct vm_anon *anons_store[UVM_MAXRANGE], **anons, *anon, *oanon;
575 	struct vm_page *pages[UVM_MAXRANGE], *pg, *uobjpage;
576 	UVMHIST_FUNC("uvm_fault"); UVMHIST_CALLED(maphist);
577 
578 	UVMHIST_LOG(maphist, "(map=%p, vaddr=0x%lx, ft=%ld, at=%ld)",
579 	      orig_map, vaddr, fault_type, access_type);
580 
581 	anon = NULL;
582 	pg = NULL;
583 
584 	uvmexp.faults++;	/* XXX: locking? */
585 
586 	/*
587 	 * init the IN parameters in the ufi
588 	 */
589 
590 	ufi.orig_map = orig_map;
591 	ufi.orig_rvaddr = trunc_page(vaddr);
592 	ufi.orig_size = PAGE_SIZE;	/* can't get any smaller than this */
593 	if (fault_type == VM_FAULT_WIRE)
594 		narrow = TRUE;		/* don't look for neighborhood
595 					 * pages on wire */
596 	else
597 		narrow = FALSE;		/* normal fault */
598 
599 	/*
600 	 * "goto ReFault" means restart the page fault from ground zero.
601 	 */
602 ReFault:
603 
604 	/*
605 	 * lookup and lock the maps
606 	 */
607 
608 	if (uvmfault_lookup(&ufi, FALSE) == FALSE) {
609 		UVMHIST_LOG(maphist, "<- no mapping @ 0x%lx", vaddr, 0,0,0);
610 		return (EFAULT);
611 	}
612 	/* locked: maps(read) */
613 
614 #ifdef DIAGNOSTIC
615 	if ((ufi.map->flags & VM_MAP_PAGEABLE) == 0)
616 		panic("uvm_fault: fault on non-pageable map (%p, 0x%lx)",
617 		    ufi.map, vaddr);
618 #endif
619 
620 	/*
621 	 * check protection
622 	 */
623 
624 	if ((ufi.entry->protection & access_type) != access_type) {
625 		UVMHIST_LOG(maphist,
626 		    "<- protection failure (prot=0x%lx, access=0x%lx)",
627 		    ufi.entry->protection, access_type, 0, 0);
628 		uvmfault_unlockmaps(&ufi, FALSE);
629 		return (EACCES);
630 	}
631 
632 	/*
633 	 * "enter_prot" is the protection we want to enter the page in at.
634 	 * for certain pages (e.g. copy-on-write pages) this protection can
635 	 * be more strict than ufi.entry->protection.  "wired" means either
636 	 * the entry is wired or we are fault-wiring the pg.
637 	 */
638 
639 	enter_prot = ufi.entry->protection;
640 	wired = VM_MAPENT_ISWIRED(ufi.entry) || (fault_type == VM_FAULT_WIRE);
641 	if (wired)
642 		access_type = enter_prot; /* full access for wired */
643 
644 	/*
645 	 * handle "needs_copy" case.   if we need to copy the amap we will
646 	 * have to drop our readlock and relock it with a write lock.  (we
647 	 * need a write lock to change anything in a map entry [e.g.
648 	 * needs_copy]).
649 	 */
650 
651 	if (UVM_ET_ISNEEDSCOPY(ufi.entry)) {
652 		if ((access_type & VM_PROT_WRITE) ||
653 		    (ufi.entry->object.uvm_obj == NULL)) {
654 			/* need to clear */
655 			UVMHIST_LOG(maphist,
656 			    "  need to clear needs_copy and refault",0,0,0,0);
657 			uvmfault_unlockmaps(&ufi, FALSE);
658 			uvmfault_amapcopy(&ufi);
659 			uvmexp.fltamcopy++;
660 			goto ReFault;
661 
662 		} else {
663 
664 			/*
665 			 * ensure that we pmap_enter page R/O since
666 			 * needs_copy is still true
667 			 */
668 			enter_prot &= ~VM_PROT_WRITE;
669 
670 		}
671 	}
672 
673 	/*
674 	 * identify the players
675 	 */
676 
677 	amap = ufi.entry->aref.ar_amap;		/* top layer */
678 	uobj = ufi.entry->object.uvm_obj;	/* bottom layer */
679 
680 	/*
681 	 * check for a case 0 fault.  if nothing backing the entry then
682 	 * error now.
683 	 */
684 
685 	if (amap == NULL && uobj == NULL) {
686 		uvmfault_unlockmaps(&ufi, FALSE);
687 		UVMHIST_LOG(maphist,"<- no backing store, no overlay",0,0,0,0);
688 		return (EFAULT);
689 	}
690 
691 	/*
692 	 * establish range of interest based on advice from mapper
693 	 * and then clip to fit map entry.   note that we only want
694 	 * to do this the first time through the fault.   if we
695 	 * ReFault we will disable this by setting "narrow" to true.
696 	 */
697 
698 	if (narrow == FALSE) {
699 
700 		/* wide fault (!narrow) */
701 		KASSERT(uvmadvice[ufi.entry->advice].advice ==
702 			 ufi.entry->advice);
703 		nback = min(uvmadvice[ufi.entry->advice].nback,
704 			    (ufi.orig_rvaddr - ufi.entry->start) >> PAGE_SHIFT);
705 		startva = ufi.orig_rvaddr - (nback << PAGE_SHIFT);
706 		nforw = min(uvmadvice[ufi.entry->advice].nforw,
707 			    ((ufi.entry->end - ufi.orig_rvaddr) >>
708 			     PAGE_SHIFT) - 1);
709 		/*
710 		 * note: "-1" because we don't want to count the
711 		 * faulting page as forw
712 		 */
713 		npages = nback + nforw + 1;
714 		centeridx = nback;
715 
716 		narrow = TRUE;	/* ensure only once per-fault */
717 
718 	} else {
719 
720 		/* narrow fault! */
721 		nback = nforw = 0;
722 		startva = ufi.orig_rvaddr;
723 		npages = 1;
724 		centeridx = 0;
725 
726 	}
727 
728 	/* locked: maps(read) */
729 	UVMHIST_LOG(maphist, "  narrow=%ld, back=%ld, forw=%ld, startva=0x%lx",
730 		    narrow, nback, nforw, startva);
731 	UVMHIST_LOG(maphist, "  entry=%p, amap=%p, obj=%p", ufi.entry,
732 		    amap, uobj, 0);
733 
734 	/*
735 	 * if we've got an amap, lock it and extract current anons.
736 	 */
737 
738 	if (amap) {
739 		anons = anons_store;
740 		amap_lookups(&ufi.entry->aref, startva - ufi.entry->start,
741 		    anons, npages);
742 	} else {
743 		anons = NULL;	/* to be safe */
744 	}
745 
746 	/* locked: maps(read), amap(if there) */
747 
748 	/*
749 	 * for MADV_SEQUENTIAL mappings we want to deactivate the back pages
750 	 * now and then forget about them (for the rest of the fault).
751 	 */
752 
753 	if (ufi.entry->advice == MADV_SEQUENTIAL && nback != 0) {
754 
755 		UVMHIST_LOG(maphist, "  MADV_SEQUENTIAL: flushing backpages",
756 		    0,0,0,0);
757 		/* flush back-page anons? */
758 		if (amap)
759 			uvmfault_anonflush(anons, nback);
760 
761 		/* flush object? */
762 		if (uobj) {
763 			uoff = (startva - ufi.entry->start) + ufi.entry->offset;
764 			simple_lock(&uobj->vmobjlock);
765 			(void) uobj->pgops->pgo_flush(uobj, uoff, uoff +
766 				    (nback << PAGE_SHIFT), PGO_DEACTIVATE);
767 			simple_unlock(&uobj->vmobjlock);
768 		}
769 
770 		/* now forget about the backpages */
771 		if (amap)
772 			anons += nback;
773 		startva += (nback << PAGE_SHIFT);
774 		npages -= nback;
775 		centeridx = 0;
776 	}
777 
778 	/* locked: maps(read), amap(if there) */
779 
780 	/*
781 	 * map in the backpages and frontpages we found in the amap in hopes
782 	 * of preventing future faults.    we also init the pages[] array as
783 	 * we go.
784 	 */
785 
786 	currva = startva;
787 	shadowed = FALSE;
788 	for (lcv = 0 ; lcv < npages ; lcv++, currva += PAGE_SIZE) {
789 
790 		/*
791 		 * dont play with VAs that are already mapped
792 		 * except for center)
793 		 */
794 		if (lcv != centeridx &&
795 		    pmap_extract(ufi.orig_map->pmap, currva, &pa)) {
796 			pages[lcv] = PGO_DONTCARE;
797 			continue;
798 		}
799 
800 		/*
801 		 * unmapped or center page.   check if any anon at this level.
802 		 */
803 		if (amap == NULL || anons[lcv] == NULL) {
804 			pages[lcv] = NULL;
805 			continue;
806 		}
807 
808 		/*
809 		 * check for present page and map if possible.   re-activate it.
810 		 */
811 
812 		pages[lcv] = PGO_DONTCARE;
813 		if (lcv == centeridx) {		/* save center for later! */
814 			shadowed = TRUE;
815 			continue;
816 		}
817 		anon = anons[lcv];
818 		simple_lock(&anon->an_lock);
819 		/* ignore loaned pages */
820 		if (anon->an_page && anon->an_page->loan_count == 0 &&
821 		    (anon->an_page->pg_flags & (PG_RELEASED|PG_BUSY)) == 0) {
822 			uvm_lock_pageq();
823 			uvm_pageactivate(anon->an_page);	/* reactivate */
824 			uvm_unlock_pageq();
825 			UVMHIST_LOG(maphist,
826 			    "  MAPPING: n anon: pm=%p, va=0x%lx, pg=%p",
827 			    ufi.orig_map->pmap, currva, anon->an_page, 0);
828 			uvmexp.fltnamap++;
829 
830 			/*
831 			 * Since this isn't the page that's actually faulting,
832 			 * ignore pmap_enter() failures; it's not critical
833 			 * that we enter these right now.
834 			 */
835 
836 			(void) pmap_enter(ufi.orig_map->pmap, currva,
837 			    VM_PAGE_TO_PHYS(anon->an_page),
838 			    (anon->an_ref > 1) ? (enter_prot & ~VM_PROT_WRITE) :
839 			    enter_prot,
840 			    PMAP_CANFAIL |
841 			     (VM_MAPENT_ISWIRED(ufi.entry) ? PMAP_WIRED : 0));
842 		}
843 		simple_unlock(&anon->an_lock);
844 		pmap_update(ufi.orig_map->pmap);
845 	}
846 
847 	/* locked: maps(read), amap(if there) */
848 	/* (shadowed == TRUE) if there is an anon at the faulting address */
849 	UVMHIST_LOG(maphist, "  shadowed=%ld, will_get=%ld", shadowed,
850 	    (uobj && shadowed == FALSE),0,0);
851 
852 	/*
853 	 * note that if we are really short of RAM we could sleep in the above
854 	 * call to pmap_enter with everything locked.   bad?
855 	 *
856 	 * XXX Actually, that is bad; pmap_enter() should just fail in that
857 	 * XXX case.  --thorpej
858 	 */
859 
860 	/*
861 	 * if the desired page is not shadowed by the amap and we have a
862 	 * backing object, then we check to see if the backing object would
863 	 * prefer to handle the fault itself (rather than letting us do it
864 	 * with the usual pgo_get hook).  the backing object signals this by
865 	 * providing a pgo_fault routine.
866 	 */
867 
868 	if (uobj && shadowed == FALSE && uobj->pgops->pgo_fault != NULL) {
869 		simple_lock(&uobj->vmobjlock);
870 
871 		/* locked: maps(read), amap (if there), uobj */
872 		result = uobj->pgops->pgo_fault(&ufi, startva, pages, npages,
873 				    centeridx, fault_type, access_type,
874 				    PGO_LOCKED);
875 
876 		/* locked: nothing, pgo_fault has unlocked everything */
877 
878 		if (result == VM_PAGER_OK)
879 			return (0);		/* pgo_fault did pmap enter */
880 		else if (result == VM_PAGER_REFAULT)
881 			goto ReFault;		/* try again! */
882 		else
883 			return (EACCES);
884 	}
885 
886 	/*
887 	 * now, if the desired page is not shadowed by the amap and we have
888 	 * a backing object that does not have a special fault routine, then
889 	 * we ask (with pgo_get) the object for resident pages that we care
890 	 * about and attempt to map them in.  we do not let pgo_get block
891 	 * (PGO_LOCKED).
892 	 *
893 	 * ("get" has the option of doing a pmap_enter for us)
894 	 */
895 
896 	if (uobj && shadowed == FALSE) {
897 		simple_lock(&uobj->vmobjlock);
898 
899 		/* locked (!shadowed): maps(read), amap (if there), uobj */
900 		/*
901 		 * the following call to pgo_get does _not_ change locking state
902 		 */
903 
904 		uvmexp.fltlget++;
905 		gotpages = npages;
906 		(void) uobj->pgops->pgo_get(uobj, ufi.entry->offset +
907 				(startva - ufi.entry->start),
908 				pages, &gotpages, centeridx,
909 				access_type & MASK(ufi.entry),
910 				ufi.entry->advice, PGO_LOCKED);
911 
912 		/*
913 		 * check for pages to map, if we got any
914 		 */
915 
916 		uobjpage = NULL;
917 
918 		if (gotpages) {
919 			currva = startva;
920 			for (lcv = 0 ; lcv < npages ;
921 			    lcv++, currva += PAGE_SIZE) {
922 
923 				if (pages[lcv] == NULL ||
924 				    pages[lcv] == PGO_DONTCARE)
925 					continue;
926 
927 				KASSERT((pages[lcv]->pg_flags & PG_RELEASED) == 0);
928 
929 				/*
930 				 * if center page is resident and not
931 				 * PG_BUSY|PG_RELEASED then pgo_get
932 				 * made it PG_BUSY for us and gave
933 				 * us a handle to it.   remember this
934 				 * page as "uobjpage." (for later use).
935 				 */
936 
937 				if (lcv == centeridx) {
938 					uobjpage = pages[lcv];
939 					UVMHIST_LOG(maphist, "  got uobjpage "
940 					    "(%p) with locked get",
941 					    uobjpage, 0,0,0);
942 					continue;
943 				}
944 
945 				/*
946 				 * note: calling pgo_get with locked data
947 				 * structures returns us pages which are
948 				 * neither busy nor released, so we don't
949 				 * need to check for this.   we can just
950 				 * directly enter the page (after moving it
951 				 * to the head of the active queue [useful?]).
952 				 */
953 
954 				uvm_lock_pageq();
955 				uvm_pageactivate(pages[lcv]);	/* reactivate */
956 				uvm_unlock_pageq();
957 				UVMHIST_LOG(maphist,
958 				  "  MAPPING: n obj: pm=%p, va=0x%lx, pg=%p",
959 				  ufi.orig_map->pmap, currva, pages[lcv], 0);
960 				uvmexp.fltnomap++;
961 
962 				/*
963 				 * Since this page isn't the page that's
964 				 * actually fauling, ignore pmap_enter()
965 				 * failures; it's not critical that we
966 				 * enter these right now.
967 				 */
968 
969 				(void) pmap_enter(ufi.orig_map->pmap, currva,
970 				    VM_PAGE_TO_PHYS(pages[lcv]),
971 				    enter_prot & MASK(ufi.entry),
972 				    PMAP_CANFAIL |
973 				     (wired ? PMAP_WIRED : 0));
974 
975 				/*
976 				 * NOTE: page can't be PG_WANTED or PG_RELEASED
977 				 * because we've held the lock the whole time
978 				 * we've had the handle.
979 				 */
980 
981 				atomic_clearbits_int(&pages[lcv]->pg_flags,
982 				    PG_BUSY);
983 				UVM_PAGE_OWN(pages[lcv], NULL);
984 			}	/* for "lcv" loop */
985 			pmap_update(ufi.orig_map->pmap);
986 		}   /* "gotpages" != 0 */
987 		/* note: object still _locked_ */
988 	} else {
989 		uobjpage = NULL;
990 	}
991 
992 	/* locked (shadowed): maps(read), amap */
993 	/* locked (!shadowed): maps(read), amap(if there),
994 		 uobj(if !null), uobjpage(if !null) */
995 
996 	/*
997 	 * note that at this point we are done with any front or back pages.
998 	 * we are now going to focus on the center page (i.e. the one we've
999 	 * faulted on).  if we have faulted on the top (anon) layer
1000 	 * [i.e. case 1], then the anon we want is anons[centeridx] (we have
1001 	 * not touched it yet).  if we have faulted on the bottom (uobj)
1002 	 * layer [i.e. case 2] and the page was both present and available,
1003 	 * then we've got a pointer to it as "uobjpage" and we've already
1004 	 * made it BUSY.
1005 	 */
1006 
1007 	/*
1008 	 * there are four possible cases we must address: 1A, 1B, 2A, and 2B
1009 	 */
1010 
1011 	/*
1012 	 * redirect case 2: if we are not shadowed, go to case 2.
1013 	 */
1014 
1015 	if (shadowed == FALSE)
1016 		goto Case2;
1017 
1018 	/* locked: maps(read), amap */
1019 
1020 	/*
1021 	 * handle case 1: fault on an anon in our amap
1022 	 */
1023 
1024 	anon = anons[centeridx];
1025 	UVMHIST_LOG(maphist, "  case 1 fault: anon=%p", anon, 0,0,0);
1026 	simple_lock(&anon->an_lock);
1027 
1028 	/* locked: maps(read), amap, anon */
1029 
1030 	/*
1031 	 * no matter if we have case 1A or case 1B we are going to need to
1032 	 * have the anon's memory resident.   ensure that now.
1033 	 */
1034 
1035 	/*
1036 	 * let uvmfault_anonget do the dirty work.
1037 	 * if it fails (!OK) it will unlock everything for us.
1038 	 * if it succeeds, locks are still valid and locked.
1039 	 * also, if it is OK, then the anon's page is on the queues.
1040 	 * if the page is on loan from a uvm_object, then anonget will
1041 	 * lock that object for us if it does not fail.
1042 	 */
1043 
1044 	result = uvmfault_anonget(&ufi, amap, anon);
1045 	switch (result) {
1046 	case VM_PAGER_OK:
1047 		break;
1048 
1049 	case VM_PAGER_REFAULT:
1050 		goto ReFault;
1051 
1052 	case VM_PAGER_ERROR:
1053 		/*
1054 		 * An error occured while trying to bring in the
1055 		 * page -- this is the only error we return right
1056 		 * now.
1057 		 */
1058 		return (EACCES);	/* XXX */
1059 
1060 	default:
1061 #ifdef DIAGNOSTIC
1062 		panic("uvm_fault: uvmfault_anonget -> %d", result);
1063 #else
1064 		return (EACCES);
1065 #endif
1066 	}
1067 
1068 	/*
1069 	 * uobj is non null if the page is on loan from an object (i.e. uobj)
1070 	 */
1071 
1072 	uobj = anon->an_page->uobject;	/* locked by anonget if !NULL */
1073 
1074 	/* locked: maps(read), amap, anon, uobj(if one) */
1075 
1076 	/*
1077 	 * special handling for loaned pages
1078 	 */
1079 
1080 	if (anon->an_page->loan_count) {
1081 
1082 		if ((access_type & VM_PROT_WRITE) == 0) {
1083 
1084 			/*
1085 			 * for read faults on loaned pages we just cap the
1086 			 * protection at read-only.
1087 			 */
1088 
1089 			enter_prot = enter_prot & ~VM_PROT_WRITE;
1090 
1091 		} else {
1092 			/*
1093 			 * note that we can't allow writes into a loaned page!
1094 			 *
1095 			 * if we have a write fault on a loaned page in an
1096 			 * anon then we need to look at the anon's ref count.
1097 			 * if it is greater than one then we are going to do
1098 			 * a normal copy-on-write fault into a new anon (this
1099 			 * is not a problem).  however, if the reference count
1100 			 * is one (a case where we would normally allow a
1101 			 * write directly to the page) then we need to kill
1102 			 * the loan before we continue.
1103 			 */
1104 
1105 			/* >1 case is already ok */
1106 			if (anon->an_ref == 1) {
1107 
1108 				/* get new un-owned replacement page */
1109 				pg = uvm_pagealloc(NULL, 0, NULL, 0);
1110 				if (pg == NULL) {
1111 					uvmfault_unlockall(&ufi, amap, uobj,
1112 					    anon);
1113 					uvm_wait("flt_noram2");
1114 					goto ReFault;
1115 				}
1116 
1117 				/*
1118 				 * copy data, kill loan, and drop uobj lock
1119 				 * (if any)
1120 				 */
1121 				/* copy old -> new */
1122 				uvm_pagecopy(anon->an_page, pg);
1123 
1124 				/* force reload */
1125 				pmap_page_protect(anon->an_page,
1126 						  VM_PROT_NONE);
1127 				uvm_lock_pageq();	  /* KILL loan */
1128 				if (uobj)
1129 					/* if we were loaning */
1130 					anon->an_page->loan_count--;
1131 				anon->an_page->uanon = NULL;
1132 				/* in case we owned */
1133 				atomic_clearbits_int(
1134 				    &anon->an_page->pg_flags, PQ_ANON);
1135 				uvm_pageactivate(pg);
1136 				uvm_unlock_pageq();
1137 				if (uobj) {
1138 					simple_unlock(&uobj->vmobjlock);
1139 					uobj = NULL;
1140 				}
1141 
1142 				/* install new page in anon */
1143 				anon->an_page = pg;
1144 				pg->uanon = anon;
1145 				atomic_setbits_int(&pg->pg_flags, PQ_ANON);
1146 				atomic_clearbits_int(&pg->pg_flags,
1147 				    PG_BUSY|PG_FAKE);
1148 				UVM_PAGE_OWN(pg, NULL);
1149 
1150 				/* done! */
1151 			}     /* ref == 1 */
1152 		}       /* write fault */
1153 	}         /* loan count */
1154 
1155 	/*
1156 	 * if we are case 1B then we will need to allocate a new blank
1157 	 * anon to transfer the data into.   note that we have a lock
1158 	 * on anon, so no one can busy or release the page until we are done.
1159 	 * also note that the ref count can't drop to zero here because
1160 	 * it is > 1 and we are only dropping one ref.
1161 	 *
1162 	 * in the (hopefully very rare) case that we are out of RAM we
1163 	 * will unlock, wait for more RAM, and refault.
1164 	 *
1165 	 * if we are out of anon VM we kill the process (XXX: could wait?).
1166 	 */
1167 
1168 	if ((access_type & VM_PROT_WRITE) != 0 && anon->an_ref > 1) {
1169 
1170 		UVMHIST_LOG(maphist, "  case 1B: COW fault",0,0,0,0);
1171 		uvmexp.flt_acow++;
1172 		oanon = anon;		/* oanon = old, locked anon */
1173 		anon = uvm_analloc();
1174 		if (anon) {
1175 			pg = uvm_pagealloc(NULL, 0, anon, 0);
1176 		}
1177 
1178 		/* check for out of RAM */
1179 		if (anon == NULL || pg == NULL) {
1180 			if (anon)
1181 				uvm_anfree(anon);
1182 			uvmfault_unlockall(&ufi, amap, uobj, oanon);
1183 			KASSERT(uvmexp.swpgonly <= uvmexp.swpages);
1184 			if (anon == NULL || uvmexp.swpgonly == uvmexp.swpages) {
1185 				UVMHIST_LOG(maphist,
1186 				    "<- failed.  out of VM",0,0,0,0);
1187 				uvmexp.fltnoanon++;
1188 				return (ENOMEM);
1189 			}
1190 
1191 			uvmexp.fltnoram++;
1192 			uvm_wait("flt_noram3");	/* out of RAM, wait for more */
1193 			goto ReFault;
1194 		}
1195 
1196 		/* got all resources, replace anon with nanon */
1197 
1198 		uvm_pagecopy(oanon->an_page, pg);	/* pg now !PG_CLEAN */
1199 		/* un-busy! new page */
1200 		atomic_clearbits_int(&pg->pg_flags, PG_BUSY|PG_FAKE);
1201 		UVM_PAGE_OWN(pg, NULL);
1202 		amap_add(&ufi.entry->aref, ufi.orig_rvaddr - ufi.entry->start,
1203 		    anon, 1);
1204 
1205 		/* deref: can not drop to zero here by defn! */
1206 		oanon->an_ref--;
1207 
1208 		/*
1209 		 * note: oanon still locked.   anon is _not_ locked, but we
1210 		 * have the sole references to in from amap which _is_ locked.
1211 		 * thus, no one can get at it until we are done with it.
1212 		 */
1213 
1214 	} else {
1215 
1216 		uvmexp.flt_anon++;
1217 		oanon = anon;		/* old, locked anon is same as anon */
1218 		pg = anon->an_page;
1219 		if (anon->an_ref > 1)     /* disallow writes to ref > 1 anons */
1220 			enter_prot = enter_prot & ~VM_PROT_WRITE;
1221 
1222 	}
1223 
1224 	/* locked: maps(read), amap, oanon */
1225 
1226 	/*
1227 	 * now map the page in ...
1228 	 * XXX: old fault unlocks object before pmap_enter.  this seems
1229 	 * suspect since some other thread could blast the page out from
1230 	 * under us between the unlock and the pmap_enter.
1231 	 */
1232 
1233 	UVMHIST_LOG(maphist, "  MAPPING: anon: pm=%p, va=0x%lx, pg=%p",
1234 	    ufi.orig_map->pmap, ufi.orig_rvaddr, pg, 0);
1235 	if (pmap_enter(ufi.orig_map->pmap, ufi.orig_rvaddr, VM_PAGE_TO_PHYS(pg),
1236 	    enter_prot, access_type | PMAP_CANFAIL | (wired ? PMAP_WIRED : 0))
1237 	    != 0) {
1238 		/*
1239 		 * No need to undo what we did; we can simply think of
1240 		 * this as the pmap throwing away the mapping information.
1241 		 *
1242 		 * We do, however, have to go through the ReFault path,
1243 		 * as the map may change while we're asleep.
1244 		 */
1245 		uvmfault_unlockall(&ufi, amap, uobj, oanon);
1246 		KASSERT(uvmexp.swpgonly <= uvmexp.swpages);
1247 		if (uvmexp.swpgonly == uvmexp.swpages) {
1248 			UVMHIST_LOG(maphist,
1249 			    "<- failed.  out of VM",0,0,0,0);
1250 			/* XXX instrumentation */
1251 			return (ENOMEM);
1252 		}
1253 		/* XXX instrumentation */
1254 		uvm_wait("flt_pmfail1");
1255 		goto ReFault;
1256 	}
1257 
1258 	/*
1259 	 * ... update the page queues.
1260 	 */
1261 
1262 	uvm_lock_pageq();
1263 
1264 	if (fault_type == VM_FAULT_WIRE) {
1265 		uvm_pagewire(pg);
1266 
1267 		/*
1268 		 * since the now-wired page cannot be paged out,
1269 		 * release its swap resources for others to use.
1270 		 * since an anon with no swap cannot be PG_CLEAN,
1271 		 * clear its clean flag now.
1272 		 */
1273 		atomic_clearbits_int(&pg->pg_flags, PG_CLEAN);
1274 		uvm_anon_dropswap(anon);
1275 	} else {
1276 		/* activate it */
1277 		uvm_pageactivate(pg);
1278 	}
1279 
1280 	uvm_unlock_pageq();
1281 
1282 	/*
1283 	 * done case 1!  finish up by unlocking everything and returning success
1284 	 */
1285 
1286 	uvmfault_unlockall(&ufi, amap, uobj, oanon);
1287 	pmap_update(ufi.orig_map->pmap);
1288 	return (0);
1289 
1290 
1291 Case2:
1292 	/*
1293 	 * handle case 2: faulting on backing object or zero fill
1294 	 */
1295 
1296 	/*
1297 	 * locked:
1298 	 * maps(read), amap(if there), uobj(if !null), uobjpage(if !null)
1299 	 */
1300 
1301 	/*
1302 	 * note that uobjpage can not be PGO_DONTCARE at this point.  we now
1303 	 * set uobjpage to PGO_DONTCARE if we are doing a zero fill.  if we
1304 	 * have a backing object, check and see if we are going to promote
1305 	 * the data up to an anon during the fault.
1306 	 */
1307 
1308 	if (uobj == NULL) {
1309 		uobjpage = PGO_DONTCARE;
1310 		promote = TRUE;		/* always need anon here */
1311 	} else {
1312 		KASSERT(uobjpage != PGO_DONTCARE);
1313 		promote = (access_type & VM_PROT_WRITE) &&
1314 		     UVM_ET_ISCOPYONWRITE(ufi.entry);
1315 	}
1316 	UVMHIST_LOG(maphist, "  case 2 fault: promote=%ld, zfill=%ld",
1317 	    promote, (uobj == NULL), 0,0);
1318 
1319 	/*
1320 	 * if uobjpage is not null then we do not need to do I/O to get the
1321 	 * uobjpage.
1322 	 *
1323 	 * if uobjpage is null, then we need to unlock and ask the pager to
1324 	 * get the data for us.   once we have the data, we need to reverify
1325 	 * the state the world.   we are currently not holding any resources.
1326 	 */
1327 
1328 	if (uobjpage) {
1329 		/* update rusage counters */
1330 		curproc->p_addr->u_stats.p_ru.ru_minflt++;
1331 	} else {
1332 		/* update rusage counters */
1333 		curproc->p_addr->u_stats.p_ru.ru_majflt++;
1334 
1335 		/* locked: maps(read), amap(if there), uobj */
1336 		uvmfault_unlockall(&ufi, amap, NULL, NULL);
1337 		/* locked: uobj */
1338 
1339 		uvmexp.fltget++;
1340 		gotpages = 1;
1341 		uoff = (ufi.orig_rvaddr - ufi.entry->start) + ufi.entry->offset;
1342 		result = uobj->pgops->pgo_get(uobj, uoff, &uobjpage, &gotpages,
1343 		    0, access_type & MASK(ufi.entry), ufi.entry->advice,
1344 		    PGO_SYNCIO);
1345 
1346 		/* locked: uobjpage(if result OK) */
1347 
1348 		/*
1349 		 * recover from I/O
1350 		 */
1351 
1352 		if (result != VM_PAGER_OK) {
1353 			KASSERT(result != VM_PAGER_PEND);
1354 
1355 			if (result == VM_PAGER_AGAIN) {
1356 				UVMHIST_LOG(maphist,
1357 				    "  pgo_get says TRY AGAIN!",0,0,0,0);
1358 				tsleep((caddr_t)&lbolt, PVM, "fltagain2", 0);
1359 				goto ReFault;
1360 			}
1361 
1362 			UVMHIST_LOG(maphist, "<- pgo_get failed (code %ld)",
1363 			    result, 0,0,0);
1364 			return (EACCES); /* XXX i/o error */
1365 		}
1366 
1367 		/* locked: uobjpage */
1368 
1369 		/*
1370 		 * re-verify the state of the world by first trying to relock
1371 		 * the maps.  always relock the object.
1372 		 */
1373 
1374 		locked = uvmfault_relock(&ufi);
1375 		simple_lock(&uobj->vmobjlock);
1376 
1377 		/* locked(locked): maps(read), amap(if !null), uobj, uobjpage */
1378 		/* locked(!locked): uobj, uobjpage */
1379 
1380 		/*
1381 		 * verify that the page has not be released and re-verify
1382 		 * that amap slot is still free.   if there is a problem,
1383 		 * we unlock and clean up.
1384 		 */
1385 
1386 		if ((uobjpage->pg_flags & PG_RELEASED) != 0 ||
1387 		    (locked && amap &&
1388 		    amap_lookup(&ufi.entry->aref,
1389 		      ufi.orig_rvaddr - ufi.entry->start))) {
1390 			if (locked)
1391 				uvmfault_unlockall(&ufi, amap, NULL, NULL);
1392 			locked = FALSE;
1393 		}
1394 
1395 		/*
1396 		 * didn't get the lock?   release the page and retry.
1397 		 */
1398 
1399 		if (locked == FALSE) {
1400 
1401 			UVMHIST_LOG(maphist,
1402 			    "  wasn't able to relock after fault: retry",
1403 			    0,0,0,0);
1404 			if (uobjpage->pg_flags & PG_WANTED)
1405 				/* still holding object lock */
1406 				wakeup(uobjpage);
1407 
1408 			if (uobjpage->pg_flags & PG_RELEASED) {
1409 				uvmexp.fltpgrele++;
1410 				KASSERT(uobj->pgops->pgo_releasepg != NULL);
1411 
1412 				/* frees page */
1413 				if (uobj->pgops->pgo_releasepg(uobjpage,NULL))
1414 					/* unlock if still alive */
1415 					simple_unlock(&uobj->vmobjlock);
1416 				goto ReFault;
1417 			}
1418 
1419 			uvm_lock_pageq();
1420 			/* make sure it is in queues */
1421 			uvm_pageactivate(uobjpage);
1422 
1423 			uvm_unlock_pageq();
1424 			atomic_clearbits_int(&uobjpage->pg_flags,
1425 			    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 		uvmexp.flt_obj++;
1466 		if (UVM_ET_ISCOPYONWRITE(ufi.entry))
1467 			enter_prot &= ~VM_PROT_WRITE;
1468 		pg = uobjpage;		/* map in the actual object */
1469 
1470 		/* assert(uobjpage != PGO_DONTCARE) */
1471 
1472 		/*
1473 		 * we are faulting directly on the page.   be careful
1474 		 * about writing to loaned pages...
1475 		 */
1476 		if (uobjpage->loan_count) {
1477 
1478 			if ((access_type & VM_PROT_WRITE) == 0) {
1479 				/* read fault: cap the protection at readonly */
1480 				/* cap! */
1481 				enter_prot = enter_prot & ~VM_PROT_WRITE;
1482 			} else {
1483 				/* write fault: must break the loan here */
1484 
1485 				/* alloc new un-owned page */
1486 				pg = uvm_pagealloc(NULL, 0, NULL, 0);
1487 
1488 				if (pg == NULL) {
1489 					/*
1490 					 * drop ownership of page, it can't
1491 					 * be released
1492 					 */
1493 					if (uobjpage->pg_flags & PG_WANTED)
1494 						wakeup(uobjpage);
1495 					atomic_clearbits_int(
1496 					    &uobjpage->pg_flags,
1497 					    PG_BUSY|PG_WANTED);
1498 					UVM_PAGE_OWN(uobjpage, NULL);
1499 
1500 					uvm_lock_pageq();
1501 					/* activate: we will need it later */
1502 					uvm_pageactivate(uobjpage);
1503 
1504 					uvm_unlock_pageq();
1505 					uvmfault_unlockall(&ufi, amap, uobj,
1506 					  NULL);
1507 					UVMHIST_LOG(maphist,
1508 					  "  out of RAM breaking loan, waiting",
1509 					  0,0,0,0);
1510 					uvmexp.fltnoram++;
1511 					uvm_wait("flt_noram4");
1512 					goto ReFault;
1513 				}
1514 
1515 				/*
1516 				 * copy the data from the old page to the new
1517 				 * one and clear the fake/clean flags on the
1518 				 * new page (keep it busy).  force a reload
1519 				 * of the old page by clearing it from all
1520 				 * pmaps.  then lock the page queues to
1521 				 * rename the pages.
1522 				 */
1523 				uvm_pagecopy(uobjpage, pg);	/* old -> new */
1524 				atomic_clearbits_int(&pg->pg_flags,
1525 				    PG_FAKE|PG_CLEAN);
1526 				pmap_page_protect(uobjpage, VM_PROT_NONE);
1527 				if (uobjpage->pg_flags & PG_WANTED)
1528 					wakeup(uobjpage);
1529 				/* uobj still locked */
1530 				atomic_clearbits_int(&uobjpage->pg_flags,
1531 				    PG_BUSY|PG_WANTED);
1532 				UVM_PAGE_OWN(uobjpage, NULL);
1533 
1534 				uvm_lock_pageq();
1535 				uoff = uobjpage->offset;
1536 				/* remove old page */
1537 				uvm_pagerealloc(uobjpage, NULL, 0);
1538 
1539 				/*
1540 				 * at this point we have absolutely no
1541 				 * control over uobjpage
1542 				 */
1543 				/* install new page */
1544 				uvm_pagerealloc(pg, uobj, uoff);
1545 				uvm_unlock_pageq();
1546 
1547 				/*
1548 				 * done!  loan is broken and "pg" is
1549 				 * PG_BUSY.   it can now replace uobjpage.
1550 				 */
1551 
1552 				uobjpage = pg;
1553 
1554 			}		/* write fault case */
1555 		}		/* if loan_count */
1556 
1557 	} else {
1558 
1559 		/*
1560 		 * if we are going to promote the data to an anon we
1561 		 * allocate a blank anon here and plug it into our amap.
1562 		 */
1563 #ifdef DIAGNOSTIC
1564 		if (amap == NULL)
1565 			panic("uvm_fault: want to promote data, but no anon");
1566 #endif
1567 
1568 		anon = uvm_analloc();
1569 		if (anon) {
1570 			/*
1571 			 * In `Fill in data...' below, if
1572 			 * uobjpage == PGO_DONTCARE, we want
1573 			 * a zero'd, dirty page, so have
1574 			 * uvm_pagealloc() do that for us.
1575 			 */
1576 			pg = uvm_pagealloc(NULL, 0, anon,
1577 			    (uobjpage == PGO_DONTCARE) ? UVM_PGA_ZERO : 0);
1578 		}
1579 
1580 		/*
1581 		 * out of memory resources?
1582 		 */
1583 		if (anon == NULL || pg == NULL) {
1584 
1585 			/*
1586 			 * arg!  must unbusy our page and fail or sleep.
1587 			 */
1588 			if (uobjpage != PGO_DONTCARE) {
1589 				if (uobjpage->pg_flags & PG_WANTED)
1590 					/* still holding object lock */
1591 					wakeup(uobjpage);
1592 
1593 				uvm_lock_pageq();
1594 				uvm_pageactivate(uobjpage);
1595 				uvm_unlock_pageq();
1596 				atomic_clearbits_int(&uobjpage->pg_flags,
1597 				    PG_BUSY|PG_WANTED);
1598 				UVM_PAGE_OWN(uobjpage, NULL);
1599 			}
1600 
1601 			/* unlock and fail ... */
1602 			uvmfault_unlockall(&ufi, amap, uobj, NULL);
1603 			KASSERT(uvmexp.swpgonly <= uvmexp.swpages);
1604 			if (anon == NULL || uvmexp.swpgonly == uvmexp.swpages) {
1605 				UVMHIST_LOG(maphist, "  promote: out of VM",
1606 				    0,0,0,0);
1607 				uvmexp.fltnoanon++;
1608 				return (ENOMEM);
1609 			}
1610 
1611 			UVMHIST_LOG(maphist, "  out of RAM, waiting for more",
1612 			    0,0,0,0);
1613 			uvm_anfree(anon);
1614 			uvmexp.fltnoram++;
1615 			uvm_wait("flt_noram5");
1616 			goto ReFault;
1617 		}
1618 
1619 		/*
1620 		 * fill in the data
1621 		 */
1622 
1623 		if (uobjpage != PGO_DONTCARE) {
1624 			uvmexp.flt_prcopy++;
1625 			/* copy page [pg now dirty] */
1626 			uvm_pagecopy(uobjpage, pg);
1627 
1628 			/*
1629 			 * promote to shared amap?  make sure all sharing
1630 			 * procs see it
1631 			 */
1632 			if ((amap_flags(amap) & AMAP_SHARED) != 0) {
1633 				pmap_page_protect(uobjpage, VM_PROT_NONE);
1634 			}
1635 
1636 			/*
1637 			 * dispose of uobjpage.  it can't be PG_RELEASED
1638 			 * since we still hold the object lock.
1639 			 * drop handle to uobj as well.
1640 			 */
1641 
1642 			if (uobjpage->pg_flags & PG_WANTED)
1643 				/* still have the obj lock */
1644 				wakeup(uobjpage);
1645 			atomic_clearbits_int(&uobjpage->pg_flags,
1646 			    PG_BUSY|PG_WANTED);
1647 			UVM_PAGE_OWN(uobjpage, NULL);
1648 			uvm_lock_pageq();
1649 			uvm_pageactivate(uobjpage);
1650 			uvm_unlock_pageq();
1651 			simple_unlock(&uobj->vmobjlock);
1652 			uobj = NULL;
1653 
1654 			UVMHIST_LOG(maphist,
1655 			    "  promote uobjpage %p to anon/page %p/%p",
1656 			    uobjpage, anon, pg, 0);
1657 
1658 		} else {
1659 			uvmexp.flt_przero++;
1660 			/*
1661 			 * Page is zero'd and marked dirty by uvm_pagealloc()
1662 			 * above.
1663 			 */
1664 			UVMHIST_LOG(maphist,"  zero fill anon/page %p/%p",
1665 			    anon, pg, 0, 0);
1666 		}
1667 
1668 		amap_add(&ufi.entry->aref, ufi.orig_rvaddr - ufi.entry->start,
1669 		    anon, 0);
1670 	}
1671 
1672 	/*
1673 	 * locked:
1674 	 * maps(read), amap(if !null), uobj(if !null), uobjpage(if uobj)
1675 	 *
1676 	 * note: pg is either the uobjpage or the new page in the new anon
1677 	 */
1678 
1679 	/*
1680 	 * all resources are present.   we can now map it in and free our
1681 	 * resources.
1682 	 */
1683 
1684 	UVMHIST_LOG(maphist,
1685 	    "  MAPPING: case2: pm=%p, va=0x%lx, pg=%p, promote=%ld",
1686 	    ufi.orig_map->pmap, ufi.orig_rvaddr, pg, promote);
1687 	if (pmap_enter(ufi.orig_map->pmap, ufi.orig_rvaddr, VM_PAGE_TO_PHYS(pg),
1688 	    enter_prot, access_type | PMAP_CANFAIL | (wired ? PMAP_WIRED : 0))
1689 	    != 0) {
1690 
1691 		/*
1692 		 * No need to undo what we did; we can simply think of
1693 		 * this as the pmap throwing away the mapping information.
1694 		 *
1695 		 * We do, however, have to go through the ReFault path,
1696 		 * as the map may change while we're asleep.
1697 		 */
1698 
1699 		if (pg->pg_flags & PG_WANTED)
1700 			wakeup(pg);		/* lock still held */
1701 
1702 		/*
1703 		 * note that pg can't be PG_RELEASED since we did not drop
1704 		 * the object lock since the last time we checked.
1705 		 */
1706 
1707 		atomic_clearbits_int(&pg->pg_flags, PG_BUSY|PG_FAKE|PG_WANTED);
1708 		UVM_PAGE_OWN(pg, NULL);
1709 		uvmfault_unlockall(&ufi, amap, uobj, NULL);
1710 		KASSERT(uvmexp.swpgonly <= uvmexp.swpages);
1711 		if (uvmexp.swpgonly == uvmexp.swpages) {
1712 			UVMHIST_LOG(maphist,
1713 			    "<- failed.  out of VM",0,0,0,0);
1714 			/* XXX instrumentation */
1715 			return (ENOMEM);
1716 		}
1717 		/* XXX instrumentation */
1718 		uvm_wait("flt_pmfail2");
1719 		goto ReFault;
1720 	}
1721 
1722 	uvm_lock_pageq();
1723 
1724 	if (fault_type == VM_FAULT_WIRE) {
1725 		uvm_pagewire(pg);
1726 		if (pg->pg_flags & PQ_AOBJ) {
1727 
1728 			/*
1729 			 * since the now-wired page cannot be paged out,
1730 			 * release its swap resources for others to use.
1731 			 * since an aobj page with no swap cannot be PG_CLEAN,
1732 			 * clear its clean flag now.
1733 			 */
1734 			atomic_clearbits_int(&pg->pg_flags, PG_CLEAN);
1735 			uao_dropswap(uobj, pg->offset >> PAGE_SHIFT);
1736 		}
1737 	} else {
1738 		/* activate it */
1739 		uvm_pageactivate(pg);
1740 	}
1741 	uvm_unlock_pageq();
1742 
1743 	if (pg->pg_flags & PG_WANTED)
1744 		wakeup(pg);		/* lock still held */
1745 
1746 	/*
1747 	 * note that pg can't be PG_RELEASED since we did not drop the object
1748 	 * lock since the last time we checked.
1749 	 */
1750 
1751 	atomic_clearbits_int(&pg->pg_flags, PG_BUSY|PG_FAKE|PG_WANTED);
1752 	UVM_PAGE_OWN(pg, NULL);
1753 	uvmfault_unlockall(&ufi, amap, uobj, NULL);
1754 	pmap_update(ufi.orig_map->pmap);
1755 
1756 	UVMHIST_LOG(maphist, "<- done (SUCCESS!)",0,0,0,0);
1757 	return (0);
1758 }
1759 
1760 
1761 /*
1762  * uvm_fault_wire: wire down a range of virtual addresses in a map.
1763  *
1764  * => map may be read-locked by caller, but MUST NOT be write-locked.
1765  * => if map is read-locked, any operations which may cause map to
1766  *	be write-locked in uvm_fault() must be taken care of by
1767  *	the caller.  See uvm_map_pageable().
1768  */
1769 
1770 int
1771 uvm_fault_wire(map, start, end, access_type)
1772 	vm_map_t map;
1773 	vaddr_t start, end;
1774 	vm_prot_t access_type;
1775 {
1776 	vaddr_t va;
1777 	pmap_t  pmap;
1778 	int rv;
1779 
1780 	pmap = vm_map_pmap(map);
1781 
1782 	/*
1783 	 * now fault it in a page at a time.   if the fault fails then we have
1784 	 * to undo what we have done.   note that in uvm_fault VM_PROT_NONE
1785 	 * is replaced with the max protection if fault_type is VM_FAULT_WIRE.
1786 	 */
1787 
1788 	for (va = start ; va < end ; va += PAGE_SIZE) {
1789 		rv = uvm_fault(map, va, VM_FAULT_WIRE, access_type);
1790 		if (rv) {
1791 			if (va != start) {
1792 				uvm_fault_unwire(map, start, va);
1793 			}
1794 			return (rv);
1795 		}
1796 	}
1797 
1798 	return (0);
1799 }
1800 
1801 /*
1802  * uvm_fault_unwire(): unwire range of virtual space.
1803  */
1804 
1805 void
1806 uvm_fault_unwire(map, start, end)
1807 	vm_map_t map;
1808 	vaddr_t start, end;
1809 {
1810 
1811 	vm_map_lock_read(map);
1812 	uvm_fault_unwire_locked(map, start, end);
1813 	vm_map_unlock_read(map);
1814 }
1815 
1816 /*
1817  * uvm_fault_unwire_locked(): the guts of uvm_fault_unwire().
1818  *
1819  * => map must be at least read-locked.
1820  */
1821 
1822 void
1823 uvm_fault_unwire_locked(map, start, end)
1824 	vm_map_t map;
1825 	vaddr_t start, end;
1826 {
1827 	vm_map_entry_t entry;
1828 	pmap_t pmap = vm_map_pmap(map);
1829 	vaddr_t va;
1830 	paddr_t pa;
1831 	struct vm_page *pg;
1832 
1833 	KASSERT((map->flags & VM_MAP_INTRSAFE) == 0);
1834 
1835 	/*
1836 	 * we assume that the area we are unwiring has actually been wired
1837 	 * in the first place.   this means that we should be able to extract
1838 	 * the PAs from the pmap.   we also lock out the page daemon so that
1839 	 * we can call uvm_pageunwire.
1840 	 */
1841 
1842 	uvm_lock_pageq();
1843 
1844 	/*
1845 	 * find the beginning map entry for the region.
1846 	 */
1847 	KASSERT(start >= vm_map_min(map) && end <= vm_map_max(map));
1848 	if (uvm_map_lookup_entry(map, start, &entry) == FALSE)
1849 		panic("uvm_fault_unwire_locked: address not in map");
1850 
1851 	for (va = start; va < end ; va += PAGE_SIZE) {
1852 		if (pmap_extract(pmap, va, &pa) == FALSE)
1853 			continue;
1854 
1855 		/*
1856 		 * find the map entry for the current address.
1857 		 */
1858 		KASSERT(va >= entry->start);
1859 		while (va >= entry->end) {
1860 			KASSERT(entry->next != &map->header &&
1861 				entry->next->start <= entry->end);
1862 			entry = entry->next;
1863 		}
1864 
1865 		/*
1866 		 * if the entry is no longer wired, tell the pmap.
1867 		 */
1868 		if (VM_MAPENT_ISWIRED(entry) == 0)
1869 			pmap_unwire(pmap, va);
1870 
1871 		pg = PHYS_TO_VM_PAGE(pa);
1872 		if (pg)
1873 			uvm_pageunwire(pg);
1874 	}
1875 
1876 	uvm_unlock_pageq();
1877 }
1878