xref: /netbsd-src/sys/uvm/uvm_fault.c (revision 4472dbe5e3bd91ef2540bada7a7ca7384627ff9b)
1 /*	$NetBSD: uvm_fault.c,v 1.48 2000/04/10 01:17:41 thorpej Exp $	*/
2 
3 /*
4  *
5  * Copyright (c) 1997 Charles D. Cranor and Washington University.
6  * All rights reserved.
7  *
8  * Redistribution and use in source and binary forms, with or without
9  * modification, are permitted provided that the following conditions
10  * are met:
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in the
15  *    documentation and/or other materials provided with the distribution.
16  * 3. All advertising materials mentioning features or use of this software
17  *    must display the following acknowledgement:
18  *      This product includes software developed by Charles D. Cranor and
19  *      Washington University.
20  * 4. The name of the author may not be used to endorse or promote products
21  *    derived from this software without specific prior written permission.
22  *
23  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
24  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
25  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
26  * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
27  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
28  * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
29  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
30  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
31  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
32  * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
33  *
34  * from: Id: uvm_fault.c,v 1.1.2.23 1998/02/06 05:29:05 chs Exp
35  */
36 
37 #include "opt_uvmhist.h"
38 
39 /*
40  * uvm_fault.c: fault handler
41  */
42 
43 #include <sys/param.h>
44 #include <sys/systm.h>
45 #include <sys/kernel.h>
46 #include <sys/proc.h>
47 #include <sys/malloc.h>
48 #include <sys/mman.h>
49 #include <sys/user.h>
50 
51 #include <vm/vm.h>
52 #include <vm/vm_page.h>
53 #include <vm/vm_kern.h>
54 
55 #include <uvm/uvm.h>
56 
57 /*
58  *
59  * a word on page faults:
60  *
61  * types of page faults we handle:
62  *
63  * CASE 1: upper layer faults                   CASE 2: lower layer faults
64  *
65  *    CASE 1A         CASE 1B                  CASE 2A        CASE 2B
66  *    read/write1     write>1                  read/write   +-cow_write/zero
67  *         |             |                         |        |
68  *      +--|--+       +--|--+     +-----+       +  |  +     | +-----+
69  * amap |  V  |       |  ----------->new|          |        | |  ^  |
70  *      +-----+       +-----+     +-----+       +  |  +     | +--|--+
71  *                                                 |        |    |
72  *      +-----+       +-----+                   +--|--+     | +--|--+
73  * uobj | d/c |       | d/c |                   |  V  |     +----|  |
74  *      +-----+       +-----+                   +-----+       +-----+
75  *
76  * d/c = don't care
77  *
78  *   case [0]: layerless fault
79  *	no amap or uobj is present.   this is an error.
80  *
81  *   case [1]: upper layer fault [anon active]
82  *     1A: [read] or [write with anon->an_ref == 1]
83  *		I/O takes place in top level anon and uobj is not touched.
84  *     1B: [write with anon->an_ref > 1]
85  *		new anon is alloc'd and data is copied off ["COW"]
86  *
87  *   case [2]: lower layer fault [uobj]
88  *     2A: [read on non-NULL uobj] or [write to non-copy_on_write area]
89  *		I/O takes place directly in object.
90  *     2B: [write to copy_on_write] or [read on NULL uobj]
91  *		data is "promoted" from uobj to a new anon.
92  *		if uobj is null, then we zero fill.
93  *
94  * we follow the standard UVM locking protocol ordering:
95  *
96  * MAPS => AMAP => UOBJ => ANON => PAGE QUEUES (PQ)
97  * we hold a PG_BUSY page if we unlock for I/O
98  *
99  *
100  * the code is structured as follows:
101  *
102  *     - init the "IN" params in the ufi structure
103  *   ReFault:
104  *     - do lookups [locks maps], check protection, handle needs_copy
105  *     - check for case 0 fault (error)
106  *     - establish "range" of fault
107  *     - if we have an amap lock it and extract the anons
108  *     - if sequential advice deactivate pages behind us
109  *     - at the same time check pmap for unmapped areas and anon for pages
110  *	 that we could map in (and do map it if found)
111  *     - check object for resident pages that we could map in
112  *     - if (case 2) goto Case2
113  *     - >>> handle case 1
114  *           - ensure source anon is resident in RAM
115  *           - if case 1B alloc new anon and copy from source
116  *           - map the correct page in
117  *   Case2:
118  *     - >>> handle case 2
119  *           - ensure source page is resident (if uobj)
120  *           - if case 2B alloc new anon and copy from source (could be zero
121  *		fill if uobj == NULL)
122  *           - map the correct page in
123  *     - done!
124  *
125  * note on paging:
126  *   if we have to do I/O we place a PG_BUSY page in the correct object,
127  * unlock everything, and do the I/O.   when I/O is done we must reverify
128  * the state of the world before assuming that our data structures are
129  * valid.   [because mappings could change while the map is unlocked]
130  *
131  *  alternative 1: unbusy the page in question and restart the page fault
132  *    from the top (ReFault).   this is easy but does not take advantage
133  *    of the information that we already have from our previous lookup,
134  *    although it is possible that the "hints" in the vm_map will help here.
135  *
136  * alternative 2: the system already keeps track of a "version" number of
137  *    a map.   [i.e. every time you write-lock a map (e.g. to change a
138  *    mapping) you bump the version number up by one...]   so, we can save
139  *    the version number of the map before we release the lock and start I/O.
140  *    then when I/O is done we can relock and check the version numbers
141  *    to see if anything changed.    this might save us some over 1 because
142  *    we don't have to unbusy the page and may be less compares(?).
143  *
144  * alternative 3: put in backpointers or a way to "hold" part of a map
145  *    in place while I/O is in progress.   this could be complex to
146  *    implement (especially with structures like amap that can be referenced
147  *    by multiple map entries, and figuring out what should wait could be
148  *    complex as well...).
149  *
150  * given that we are not currently multiprocessor or multithreaded we might
151  * as well choose alternative 2 now.   maybe alternative 3 would be useful
152  * in the future.    XXX keep in mind for future consideration//rechecking.
153  */
154 
155 /*
156  * local data structures
157  */
158 
159 struct uvm_advice {
160 	int advice;
161 	int nback;
162 	int nforw;
163 };
164 
165 /*
166  * page range array:
167  * note: index in array must match "advice" value
168  * XXX: borrowed numbers from freebsd.   do they work well for us?
169  */
170 
171 static struct uvm_advice uvmadvice[] = {
172 	{ MADV_NORMAL, 3, 4 },
173 	{ MADV_RANDOM, 0, 0 },
174 	{ MADV_SEQUENTIAL, 8, 7},
175 };
176 
177 #define UVM_MAXRANGE 16	/* must be max() of nback+nforw+1 */
178 
179 /*
180  * private prototypes
181  */
182 
183 static void uvmfault_amapcopy __P((struct uvm_faultinfo *));
184 static __inline void uvmfault_anonflush __P((struct vm_anon **, int));
185 
186 /*
187  * inline functions
188  */
189 
190 /*
191  * uvmfault_anonflush: try and deactivate pages in specified anons
192  *
193  * => does not have to deactivate page if it is busy
194  */
195 
196 static __inline void
197 uvmfault_anonflush(anons, n)
198 	struct vm_anon **anons;
199 	int n;
200 {
201 	int lcv;
202 	struct vm_page *pg;
203 
204 	for (lcv = 0 ; lcv < n ; lcv++) {
205 		if (anons[lcv] == NULL)
206 			continue;
207 		simple_lock(&anons[lcv]->an_lock);
208 		pg = anons[lcv]->u.an_page;
209 		if (pg && (pg->flags & PG_BUSY) == 0 && pg->loan_count == 0) {
210 			uvm_lock_pageq();
211 			if (pg->wire_count == 0) {
212 				pmap_page_protect(pg, VM_PROT_NONE);
213 				uvm_pagedeactivate(pg);
214 			}
215 			uvm_unlock_pageq();
216 		}
217 		simple_unlock(&anons[lcv]->an_lock);
218 	}
219 }
220 
221 /*
222  * normal functions
223  */
224 
225 /*
226  * uvmfault_amapcopy: clear "needs_copy" in a map.
227  *
228  * => called with VM data structures unlocked (usually, see below)
229  * => we get a write lock on the maps and clear needs_copy for a VA
230  * => if we are out of RAM we sleep (waiting for more)
231  */
232 
233 static void
234 uvmfault_amapcopy(ufi)
235 	struct uvm_faultinfo *ufi;
236 {
237 
238 	/*
239 	 * while we haven't done the job
240 	 */
241 
242 	while (1) {
243 
244 		/*
245 		 * no mapping?  give up.
246 		 */
247 
248 		if (uvmfault_lookup(ufi, TRUE) == FALSE)
249 			return;
250 
251 		/*
252 		 * copy if needed.
253 		 */
254 
255 		if (UVM_ET_ISNEEDSCOPY(ufi->entry))
256 			amap_copy(ufi->map, ufi->entry, M_NOWAIT, TRUE,
257 				ufi->orig_rvaddr, ufi->orig_rvaddr + 1);
258 
259 		/*
260 		 * didn't work?  must be out of RAM.   unlock and sleep.
261 		 */
262 
263 		if (UVM_ET_ISNEEDSCOPY(ufi->entry)) {
264 			uvmfault_unlockmaps(ufi, TRUE);
265 			uvm_wait("fltamapcopy");
266 			continue;
267 		}
268 
269 		/*
270 		 * got it!   unlock and return.
271 		 */
272 
273 		uvmfault_unlockmaps(ufi, TRUE);
274 		return;
275 	}
276 	/*NOTREACHED*/
277 }
278 
279 /*
280  * uvmfault_anonget: get data in an anon into a non-busy, non-released
281  * page in that anon.
282  *
283  * => maps, amap, and anon locked by caller.
284  * => if we fail (result != VM_PAGER_OK) we unlock everything except anon.
285  * => if we are successful, we return with everything still locked.
286  * => we don't move the page on the queues [gets moved later]
287  * => if we allocate a new page [we_own], it gets put on the queues.
288  *    either way, the result is that the page is on the queues at return time
289  * => for pages which are on loan from a uvm_object (and thus are not
290  *    owned by the anon): if successful, we return with the owning object
291  *    locked.   the caller must unlock this object when it unlocks everything
292  *    else.
293  */
294 
295 int
296 uvmfault_anonget(ufi, amap, anon)
297 	struct uvm_faultinfo *ufi;
298 	struct vm_amap *amap;
299 	struct vm_anon *anon;
300 {
301 	boolean_t we_own;	/* we own anon's page? */
302 	boolean_t locked;	/* did we relock? */
303 	struct vm_page *pg;
304 	int result;
305 	UVMHIST_FUNC("uvmfault_anonget"); UVMHIST_CALLED(maphist);
306 
307 	result = 0;		/* XXX shut up gcc */
308 	uvmexp.fltanget++;
309         /* bump rusage counters */
310 	if (anon->u.an_page)
311 		curproc->p_addr->u_stats.p_ru.ru_minflt++;
312 	else
313 		curproc->p_addr->u_stats.p_ru.ru_majflt++;
314 
315 	/*
316 	 * loop until we get it, or fail.
317 	 */
318 
319 	while (1) {
320 
321 		we_own = FALSE;		/* TRUE if we set PG_BUSY on a page */
322 		pg = anon->u.an_page;
323 
324 		/*
325 		 * if there is a resident page and it is loaned, then anon
326 		 * may not own it.   call out to uvm_anon_lockpage() to ensure
327 		 * the real owner of the page has been identified and locked.
328 		 */
329 
330 		if (pg && pg->loan_count)
331 			pg = uvm_anon_lockloanpg(anon);
332 
333 		/*
334 		 * page there?   make sure it is not busy/released.
335 		 */
336 
337 		if (pg) {
338 
339 			/*
340 			 * at this point, if the page has a uobject [meaning
341 			 * we have it on loan], then that uobject is locked
342 			 * by us!   if the page is busy, we drop all the
343 			 * locks (including uobject) and try again.
344 			 */
345 
346 			if ((pg->flags & (PG_BUSY|PG_RELEASED)) == 0) {
347 				UVMHIST_LOG(maphist, "<- OK",0,0,0,0);
348 				return (VM_PAGER_OK);
349 			}
350 			pg->flags |= PG_WANTED;
351 			uvmexp.fltpgwait++;
352 
353 			/*
354 			 * the last unlock must be an atomic unlock+wait on
355 			 * the owner of page
356 			 */
357 			if (pg->uobject) {	/* owner is uobject ? */
358 				uvmfault_unlockall(ufi, amap, NULL, anon);
359 				UVMHIST_LOG(maphist, " unlock+wait on uobj",0,
360 				    0,0,0);
361 				UVM_UNLOCK_AND_WAIT(pg,
362 				    &pg->uobject->vmobjlock,
363 				    FALSE, "anonget1",0);
364 			} else {
365 				/* anon owns page */
366 				uvmfault_unlockall(ufi, amap, NULL, NULL);
367 				UVMHIST_LOG(maphist, " unlock+wait on anon",0,
368 				    0,0,0);
369 				UVM_UNLOCK_AND_WAIT(pg,&anon->an_lock,0,
370 				    "anonget2",0);
371 			}
372 			/* ready to relock and try again */
373 
374 		} else {
375 
376 			/*
377 			 * no page, we must try and bring it in.
378 			 */
379 			pg = uvm_pagealloc(NULL, 0, anon, 0);
380 
381 			if (pg == NULL) {		/* out of RAM.  */
382 
383 				uvmfault_unlockall(ufi, amap, NULL, anon);
384 				uvmexp.fltnoram++;
385 				UVMHIST_LOG(maphist, "  noram -- UVM_WAIT",0,
386 				    0,0,0);
387 				uvm_wait("flt_noram1");
388 				/* ready to relock and try again */
389 
390 			} else {
391 
392 				/* we set the PG_BUSY bit */
393 				we_own = TRUE;
394 				uvmfault_unlockall(ufi, amap, NULL, anon);
395 
396 				/*
397 				 * we are passing a PG_BUSY+PG_FAKE+PG_CLEAN
398 				 * page into the uvm_swap_get function with
399 				 * all data structures unlocked.  note that
400 				 * it is ok to read an_swslot here because
401 				 * we hold PG_BUSY on the page.
402 				 */
403 				uvmexp.pageins++;
404 				result = uvm_swap_get(pg, anon->an_swslot,
405 				    PGO_SYNCIO);
406 
407 				/*
408 				 * we clean up after the i/o below in the
409 				 * "we_own" case
410 				 */
411 				/* ready to relock and try again */
412 			}
413 		}
414 
415 		/*
416 		 * now relock and try again
417 		 */
418 
419 		locked = uvmfault_relock(ufi);
420 		if (locked && amap != NULL) {
421 			amap_lock(amap);
422 		}
423 		if (locked || we_own)
424 			simple_lock(&anon->an_lock);
425 
426 		/*
427 		 * if we own the page (i.e. we set PG_BUSY), then we need
428 		 * to clean up after the I/O. there are three cases to
429 		 * consider:
430 		 *   [1] page released during I/O: free anon and ReFault.
431 		 *   [2] I/O not OK.   free the page and cause the fault
432 		 *       to fail.
433 		 *   [3] I/O OK!   activate the page and sync with the
434 		 *       non-we_own case (i.e. drop anon lock if not locked).
435 		 */
436 
437 		if (we_own) {
438 
439 			if (pg->flags & PG_WANTED) {
440 				/* still holding object lock */
441 				wakeup(pg);
442 			}
443 			/* un-busy! */
444 			pg->flags &= ~(PG_WANTED|PG_BUSY|PG_FAKE);
445 			UVM_PAGE_OWN(pg, NULL);
446 
447 			/*
448 			 * if we were RELEASED during I/O, then our anon is
449 			 * no longer part of an amap.   we need to free the
450 			 * anon and try again.
451 			 */
452 			if (pg->flags & PG_RELEASED) {
453 				pmap_page_protect(pg, VM_PROT_NONE);
454 				simple_unlock(&anon->an_lock);
455 				uvm_anfree(anon);	/* frees page for us */
456 				if (locked)
457 					uvmfault_unlockall(ufi, amap, NULL,
458 							   NULL);
459 				uvmexp.fltpgrele++;
460 				UVMHIST_LOG(maphist, "<- REFAULT", 0,0,0,0);
461 				return (VM_PAGER_REFAULT);	/* refault! */
462 			}
463 
464 			if (result != VM_PAGER_OK) {
465 #ifdef DIAGNOSTIC
466 				if (result == VM_PAGER_PEND) {
467 					panic("uvmfault_anonget: "
468 					      "got PENDING for non-async I/O");
469 				}
470 #endif
471 				/* remove page from anon */
472 				anon->u.an_page = NULL;
473 
474 				/*
475 				 * remove the swap slot from the anon
476 				 * and mark the anon as having no real slot.
477 				 * don't free the swap slot, thus preventing
478 				 * it from being used again.
479 				 */
480 				uvm_swap_markbad(anon->an_swslot, 1);
481 				anon->an_swslot = SWSLOT_BAD;
482 
483 				/*
484 				 * note: page was never !PG_BUSY, so it
485 				 * can't be mapped and thus no need to
486 				 * pmap_page_protect it...
487 				 */
488 				uvm_lock_pageq();
489 				uvm_pagefree(pg);
490 				uvm_unlock_pageq();
491 
492 				if (locked)
493 					uvmfault_unlockall(ufi, amap, NULL,
494 					    anon);
495 				else
496 					simple_unlock(&anon->an_lock);
497 				UVMHIST_LOG(maphist, "<- ERROR", 0,0,0,0);
498 				return (VM_PAGER_ERROR);
499 			}
500 
501 			/*
502 			 * must be OK, clear modify (already PG_CLEAN)
503 			 * and activate
504 			 */
505 			pmap_clear_modify(pg);
506 			uvm_lock_pageq();
507 			uvm_pageactivate(pg);
508 			uvm_unlock_pageq();
509 			if (!locked)
510 				simple_unlock(&anon->an_lock);
511 		}
512 
513 		/*
514 		 * we were not able to relock.   restart fault.
515 		 */
516 
517 		if (!locked) {
518 			UVMHIST_LOG(maphist, "<- REFAULT", 0,0,0,0);
519 			return (VM_PAGER_REFAULT);
520 		}
521 
522 		/*
523 		 * verify no one has touched the amap and moved the anon on us.
524 		 */
525 
526 		if (ufi != NULL &&
527 		    amap_lookup(&ufi->entry->aref,
528 				ufi->orig_rvaddr - ufi->entry->start) != anon) {
529 
530 			uvmfault_unlockall(ufi, amap, NULL, anon);
531 			UVMHIST_LOG(maphist, "<- REFAULT", 0,0,0,0);
532 			return (VM_PAGER_REFAULT);
533 		}
534 
535 		/*
536 		 * try it again!
537 		 */
538 
539 		uvmexp.fltanretry++;
540 		continue;
541 
542 	} /* while (1) */
543 
544 	/*NOTREACHED*/
545 }
546 
547 /*
548  *   F A U L T   -   m a i n   e n t r y   p o i n t
549  */
550 
551 /*
552  * uvm_fault: page fault handler
553  *
554  * => called from MD code to resolve a page fault
555  * => VM data structures usually should be unlocked.   however, it is
556  *	possible to call here with the main map locked if the caller
557  *	gets a write lock, sets it recusive, and then calls us (c.f.
558  *	uvm_map_pageable).   this should be avoided because it keeps
559  *	the map locked off during I/O.
560  */
561 
562 #define MASK(entry)     (UVM_ET_ISCOPYONWRITE(entry) ? \
563 			 ~VM_PROT_WRITE : VM_PROT_ALL)
564 
565 int
566 uvm_fault(orig_map, vaddr, fault_type, access_type)
567 	vm_map_t orig_map;
568 	vaddr_t vaddr;
569 	vm_fault_t fault_type;
570 	vm_prot_t access_type;
571 {
572 	struct uvm_faultinfo ufi;
573 	vm_prot_t enter_prot;
574 	boolean_t wired, narrow, promote, locked, shadowed;
575 	int npages, nback, nforw, centeridx, result, lcv, gotpages;
576 	vaddr_t startva, objaddr, currva, offset;
577 	paddr_t pa;
578 	struct vm_amap *amap;
579 	struct uvm_object *uobj;
580 	struct vm_anon *anons_store[UVM_MAXRANGE], **anons, *anon, *oanon;
581 	struct vm_page *pages[UVM_MAXRANGE], *pg, *uobjpage;
582 	UVMHIST_FUNC("uvm_fault"); UVMHIST_CALLED(maphist);
583 
584 	UVMHIST_LOG(maphist, "(map=0x%x, vaddr=0x%x, ft=%d, at=%d)",
585 	      orig_map, vaddr, fault_type, access_type);
586 
587 	anon = NULL; /* XXX: shut up gcc */
588 
589 	uvmexp.faults++;	/* XXX: locking? */
590 
591 	/*
592 	 * init the IN parameters in the ufi
593 	 */
594 
595 	ufi.orig_map = orig_map;
596 	ufi.orig_rvaddr = trunc_page(vaddr);
597 	ufi.orig_size = PAGE_SIZE;	/* can't get any smaller than this */
598 	if (fault_type == VM_FAULT_WIRE)
599 		narrow = TRUE;		/* don't look for neighborhood
600 					 * pages on wire */
601 	else
602 		narrow = FALSE;		/* normal fault */
603 
604 	/*
605 	 * before we do anything else, if this is a fault on a kernel
606 	 * address, check to see if the address is managed by an
607 	 * interrupt-safe map.  If it is, we fail immediately.  Intrsafe
608 	 * maps are never pageable, and this approach avoids an evil
609 	 * locking mess.
610 	 */
611 	if (orig_map == kernel_map && uvmfault_check_intrsafe(&ufi)) {
612 		UVMHIST_LOG(maphist, "<- VA 0x%lx in intrsafe map %p",
613 		    ufi.orig_rvaddr, ufi.map, 0, 0);
614 		return (KERN_FAILURE);
615 	}
616 
617 	/*
618 	 * "goto ReFault" means restart the page fault from ground zero.
619 	 */
620 ReFault:
621 
622 	/*
623 	 * lookup and lock the maps
624 	 */
625 
626 	if (uvmfault_lookup(&ufi, FALSE) == FALSE) {
627 		UVMHIST_LOG(maphist, "<- no mapping @ 0x%x", vaddr, 0,0,0);
628 		return (KERN_INVALID_ADDRESS);
629 	}
630 	/* locked: maps(read) */
631 
632 	/*
633 	 * check protection
634 	 */
635 
636 	if ((ufi.entry->protection & access_type) != access_type) {
637 		UVMHIST_LOG(maphist,
638 		    "<- protection failure (prot=0x%x, access=0x%x)",
639 		    ufi.entry->protection, access_type, 0, 0);
640 		uvmfault_unlockmaps(&ufi, FALSE);
641 		return (KERN_PROTECTION_FAILURE);
642 	}
643 
644 	/*
645 	 * if the map is not a pageable map, a page fault always fails.
646 	 */
647 
648 	if ((ufi.map->flags & VM_MAP_PAGEABLE) == 0) {
649 		UVMHIST_LOG(maphist,
650 		    "<- map %p not pageable", ufi.map, 0, 0, 0);
651 		uvmfault_unlockmaps(&ufi, FALSE);
652 		return (KERN_FAILURE);
653 	}
654 
655 	/*
656 	 * "enter_prot" is the protection we want to enter the page in at.
657 	 * for certain pages (e.g. copy-on-write pages) this protection can
658 	 * be more strict than ufi.entry->protection.  "wired" means either
659 	 * the entry is wired or we are fault-wiring the pg.
660 	 */
661 
662 	enter_prot = ufi.entry->protection;
663 	wired = VM_MAPENT_ISWIRED(ufi.entry) || (fault_type == VM_FAULT_WIRE);
664 	if (wired)
665 		access_type = enter_prot; /* full access for wired */
666 
667 	/*
668 	 * handle "needs_copy" case.   if we need to copy the amap we will
669 	 * have to drop our readlock and relock it with a write lock.  (we
670 	 * need a write lock to change anything in a map entry [e.g.
671 	 * needs_copy]).
672 	 */
673 
674 	if (UVM_ET_ISNEEDSCOPY(ufi.entry)) {
675 		if ((access_type & VM_PROT_WRITE) ||
676 		    (ufi.entry->object.uvm_obj == NULL)) {
677 			/* need to clear */
678 			UVMHIST_LOG(maphist,
679 			    "  need to clear needs_copy and refault",0,0,0,0);
680 			uvmfault_unlockmaps(&ufi, FALSE);
681 			uvmfault_amapcopy(&ufi);
682 			uvmexp.fltamcopy++;
683 			goto ReFault;
684 
685 		} else {
686 
687 			/*
688 			 * ensure that we pmap_enter page R/O since
689 			 * needs_copy is still true
690 			 */
691 			enter_prot &= ~VM_PROT_WRITE;
692 
693 		}
694 	}
695 
696 	/*
697 	 * identify the players
698 	 */
699 
700 	amap = ufi.entry->aref.ar_amap;		/* top layer */
701 	uobj = ufi.entry->object.uvm_obj;	/* bottom layer */
702 
703 	/*
704 	 * check for a case 0 fault.  if nothing backing the entry then
705 	 * error now.
706 	 */
707 
708 	if (amap == NULL && uobj == NULL) {
709 		uvmfault_unlockmaps(&ufi, FALSE);
710 		UVMHIST_LOG(maphist,"<- no backing store, no overlay",0,0,0,0);
711 		return (KERN_INVALID_ADDRESS);
712 	}
713 
714 	/*
715 	 * establish range of interest based on advice from mapper
716 	 * and then clip to fit map entry.   note that we only want
717 	 * to do this the first time through the fault.   if we
718 	 * ReFault we will disable this by setting "narrow" to true.
719 	 */
720 
721 	if (narrow == FALSE) {
722 
723 		/* wide fault (!narrow) */
724 #ifdef DIAGNOSTIC
725 		if (uvmadvice[ufi.entry->advice].advice != ufi.entry->advice)
726 			panic("fault: advice mismatch!");
727 #endif
728 		nback = min(uvmadvice[ufi.entry->advice].nback,
729 			    (ufi.orig_rvaddr - ufi.entry->start) >> PAGE_SHIFT);
730 		startva = ufi.orig_rvaddr - (nback << PAGE_SHIFT);
731 		nforw = min(uvmadvice[ufi.entry->advice].nforw,
732 			    ((ufi.entry->end - ufi.orig_rvaddr) >>
733 			     PAGE_SHIFT) - 1);
734 		/*
735 		 * note: "-1" because we don't want to count the
736 		 * faulting page as forw
737 		 */
738 		npages = nback + nforw + 1;
739 		centeridx = nback;
740 
741 		narrow = TRUE;	/* ensure only once per-fault */
742 
743 	} else {
744 
745 		/* narrow fault! */
746 		nback = nforw = 0;
747 		startva = ufi.orig_rvaddr;
748 		npages = 1;
749 		centeridx = 0;
750 
751 	}
752 
753 	/* locked: maps(read) */
754 	UVMHIST_LOG(maphist, "  narrow=%d, back=%d, forw=%d, startva=0x%x",
755 		    narrow, nback, nforw, startva);
756 	UVMHIST_LOG(maphist, "  entry=0x%x, amap=0x%x, obj=0x%x", ufi.entry,
757 		    amap, uobj, 0);
758 
759 	/*
760 	 * if we've got an amap, lock it and extract current anons.
761 	 */
762 
763 	if (amap) {
764 		amap_lock(amap);
765 		anons = anons_store;
766 		amap_lookups(&ufi.entry->aref, startva - ufi.entry->start,
767 		    anons, npages);
768 	} else {
769 		anons = NULL;	/* to be safe */
770 	}
771 
772 	/* locked: maps(read), amap(if there) */
773 
774 	/*
775 	 * for MADV_SEQUENTIAL mappings we want to deactivate the back pages
776 	 * now and then forget about them (for the rest of the fault).
777 	 */
778 
779 	if (ufi.entry->advice == MADV_SEQUENTIAL) {
780 
781 		UVMHIST_LOG(maphist, "  MADV_SEQUENTIAL: flushing backpages",
782 		    0,0,0,0);
783 		/* flush back-page anons? */
784 		if (amap)
785 			uvmfault_anonflush(anons, nback);
786 
787 		/* flush object? */
788 		if (uobj) {
789 			objaddr =
790 			    (startva - ufi.entry->start) + ufi.entry->offset;
791 			simple_lock(&uobj->vmobjlock);
792 			(void) uobj->pgops->pgo_flush(uobj, objaddr, objaddr +
793 				    (nback << PAGE_SHIFT), PGO_DEACTIVATE);
794 			simple_unlock(&uobj->vmobjlock);
795 		}
796 
797 		/* now forget about the backpages */
798 		if (amap)
799 			anons += nback;
800 		startva = startva + (nback << PAGE_SHIFT);
801 		npages -= nback;
802 		nback = centeridx = 0;
803 	}
804 
805 	/* locked: maps(read), amap(if there) */
806 
807 	/*
808 	 * map in the backpages and frontpages we found in the amap in hopes
809 	 * of preventing future faults.    we also init the pages[] array as
810 	 * we go.
811 	 */
812 
813 	currva = startva;
814 	shadowed = FALSE;
815 	for (lcv = 0 ; lcv < npages ; lcv++, currva += PAGE_SIZE) {
816 
817 		/*
818 		 * dont play with VAs that are already mapped
819 		 * except for center)
820 		 */
821 		if (lcv != centeridx) {
822 			if (pmap_extract(ufi.orig_map->pmap, currva, &pa) ==
823 			    TRUE) {
824 				pages[lcv] = PGO_DONTCARE;
825 				continue;
826 			}
827 		}
828 
829 		/*
830 		 * unmapped or center page.   check if any anon at this level.
831 		 */
832 		if (amap == NULL || anons[lcv] == NULL) {
833 			pages[lcv] = NULL;
834 			continue;
835 		}
836 
837 		/*
838 		 * check for present page and map if possible.   re-activate it.
839 		 */
840 
841 		pages[lcv] = PGO_DONTCARE;
842 		if (lcv == centeridx) {		/* save center for later! */
843 			shadowed = TRUE;
844 			continue;
845 		}
846 		anon = anons[lcv];
847 		simple_lock(&anon->an_lock);
848 		/* ignore loaned pages */
849 		if (anon->u.an_page && anon->u.an_page->loan_count == 0 &&
850 			(anon->u.an_page->flags & (PG_RELEASED|PG_BUSY)) == 0) {
851 			uvm_lock_pageq();
852 			uvm_pageactivate(anon->u.an_page);	/* reactivate */
853 			uvm_unlock_pageq();
854 			UVMHIST_LOG(maphist,
855 			    "  MAPPING: n anon: pm=0x%x, va=0x%x, pg=0x%x",
856 			    ufi.orig_map->pmap, currva, anon->u.an_page, 0);
857 			uvmexp.fltnamap++;
858 			/*
859 			 * Since this isn't the page that's actually faulting,
860 			 * ignore pmap_enter() failures; it's not critical
861 			 * that we enter these right now.
862 			 */
863 			(void) pmap_enter(ufi.orig_map->pmap, currva,
864 			    VM_PAGE_TO_PHYS(anon->u.an_page),
865 			    (anon->an_ref > 1) ? (enter_prot & ~VM_PROT_WRITE) :
866 			    enter_prot,
867 			    PMAP_CANFAIL |
868 			     (VM_MAPENT_ISWIRED(ufi.entry) ? PMAP_WIRED : 0));
869 		}
870 		simple_unlock(&anon->an_lock);
871 	}
872 
873 	/* locked: maps(read), amap(if there) */
874 	/* (shadowed == TRUE) if there is an anon at the faulting address */
875 	UVMHIST_LOG(maphist, "  shadowed=%d, will_get=%d", shadowed,
876 	    (uobj && shadowed == FALSE),0,0);
877 
878 	/*
879 	 * note that if we are really short of RAM we could sleep in the above
880 	 * call to pmap_enter with everything locked.   bad?
881 	 *
882 	 * XXX Actually, that is bad; pmap_enter() should just fail in that
883 	 * XXX case.  --thorpej
884 	 */
885 
886 	/*
887 	 * if the desired page is not shadowed by the amap and we have a
888 	 * backing object, then we check to see if the backing object would
889 	 * prefer to handle the fault itself (rather than letting us do it
890 	 * with the usual pgo_get hook).  the backing object signals this by
891 	 * providing a pgo_fault routine.
892 	 */
893 
894 	if (uobj && shadowed == FALSE && uobj->pgops->pgo_fault != NULL) {
895 
896 		simple_lock(&uobj->vmobjlock);
897 
898 		/* locked: maps(read), amap (if there), uobj */
899 		result = uobj->pgops->pgo_fault(&ufi, startva, pages, npages,
900 				    centeridx, fault_type, access_type,
901 				    PGO_LOCKED);
902 		/* locked: nothing, pgo_fault has unlocked everything */
903 
904 		if (result == VM_PAGER_OK)
905 			return (KERN_SUCCESS);	/* pgo_fault did pmap enter */
906 		else if (result == VM_PAGER_REFAULT)
907 			goto ReFault;		/* try again! */
908 		else
909 			return (KERN_PROTECTION_FAILURE);
910 	}
911 
912 	/*
913 	 * now, if the desired page is not shadowed by the amap and we have
914 	 * a backing object that does not have a special fault routine, then
915 	 * we ask (with pgo_get) the object for resident pages that we care
916 	 * about and attempt to map them in.  we do not let pgo_get block
917 	 * (PGO_LOCKED).
918 	 *
919 	 * ("get" has the option of doing a pmap_enter for us)
920 	 */
921 
922 	if (uobj && shadowed == FALSE) {
923 		simple_lock(&uobj->vmobjlock);
924 
925 		/* locked (!shadowed): maps(read), amap (if there), uobj */
926 		/*
927 		 * the following call to pgo_get does _not_ change locking state
928 		 */
929 
930 		uvmexp.fltlget++;
931 		gotpages = npages;
932 		result = uobj->pgops->pgo_get(uobj, ufi.entry->offset +
933 				(startva - ufi.entry->start),
934 				pages, &gotpages, centeridx,
935 				access_type & MASK(ufi.entry),
936 				ufi.entry->advice, PGO_LOCKED);
937 
938 		/*
939 		 * check for pages to map, if we got any
940 		 */
941 
942 		uobjpage = NULL;
943 
944 		if (gotpages) {
945 			currva = startva;
946 			for (lcv = 0 ; lcv < npages ;
947 			    lcv++, currva += PAGE_SIZE) {
948 
949 				if (pages[lcv] == NULL ||
950 				    pages[lcv] == PGO_DONTCARE)
951 					continue;
952 
953 #ifdef DIAGNOSTIC
954 					/*
955 					 * pager sanity check: pgo_get with
956 					 * PGO_LOCKED should never return a
957 					 * released page to us.
958 					 */
959 					if (pages[lcv]->flags & PG_RELEASED)
960 		panic("uvm_fault: pgo_get PGO_LOCKED gave us a RELEASED page");
961 #endif
962 
963 					/*
964 					 * if center page is resident and not
965 					 * PG_BUSY|PG_RELEASED then pgo_get
966 					 * made it PG_BUSY for us and gave
967 					 * us a handle to it.   remember this
968 					 * page as "uobjpage." (for later use).
969 					 */
970 
971 					if (lcv == centeridx) {
972 						uobjpage = pages[lcv];
973 	UVMHIST_LOG(maphist, "  got uobjpage (0x%x) with locked get",
974 					    uobjpage, 0,0,0);
975 						continue;
976 				}
977 
978 				/*
979 				 * note: calling pgo_get with locked data
980 				 * structures returns us pages which are
981 				 * neither busy nor released, so we don't
982 				 * need to check for this.   we can just
983 				 * directly enter the page (after moving it
984 				 * to the head of the active queue [useful?]).
985 				 */
986 
987 				uvm_lock_pageq();
988 				uvm_pageactivate(pages[lcv]);	/* reactivate */
989 				uvm_unlock_pageq();
990 				UVMHIST_LOG(maphist,
991 				  "  MAPPING: n obj: pm=0x%x, va=0x%x, pg=0x%x",
992 				  ufi.orig_map->pmap, currva, pages[lcv], 0);
993 				uvmexp.fltnomap++;
994 				/*
995 				 * Since this page isn't the page that's
996 				 * actually fauling, ignore pmap_enter()
997 				 * failures; it's not critical that we
998 				 * enter these right now.
999 				 */
1000 				(void) pmap_enter(ufi.orig_map->pmap, currva,
1001 				    VM_PAGE_TO_PHYS(pages[lcv]),
1002 				    enter_prot & MASK(ufi.entry),
1003 				    PMAP_CANFAIL |
1004 				     (wired ? PMAP_WIRED : 0));
1005 
1006 				/*
1007 				 * NOTE: page can't be PG_WANTED or PG_RELEASED
1008 				 * because we've held the lock the whole time
1009 				 * we've had the handle.
1010 				 */
1011 				pages[lcv]->flags &= ~(PG_BUSY); /* un-busy! */
1012 				UVM_PAGE_OWN(pages[lcv], NULL);
1013 
1014 				/* done! */
1015 			}	/* for "lcv" loop */
1016 		}   /* "gotpages" != 0 */
1017 
1018 		/* note: object still _locked_ */
1019 	} else {
1020 
1021 		uobjpage = NULL;
1022 
1023 	}
1024 
1025 	/* locked (shadowed): maps(read), amap */
1026 	/* locked (!shadowed): maps(read), amap(if there),
1027 		 uobj(if !null), uobjpage(if !null) */
1028 
1029 	/*
1030 	 * note that at this point we are done with any front or back pages.
1031 	 * we are now going to focus on the center page (i.e. the one we've
1032 	 * faulted on).  if we have faulted on the top (anon) layer
1033 	 * [i.e. case 1], then the anon we want is anons[centeridx] (we have
1034 	 * not touched it yet).  if we have faulted on the bottom (uobj)
1035 	 * layer [i.e. case 2] and the page was both present and available,
1036 	 * then we've got a pointer to it as "uobjpage" and we've already
1037 	 * made it BUSY.
1038 	 */
1039 
1040 	/*
1041 	 * there are four possible cases we must address: 1A, 1B, 2A, and 2B
1042 	 */
1043 
1044 	/*
1045 	 * redirect case 2: if we are not shadowed, go to case 2.
1046 	 */
1047 
1048 	if (shadowed == FALSE)
1049 		goto Case2;
1050 
1051 	/* locked: maps(read), amap */
1052 
1053 	/*
1054 	 * handle case 1: fault on an anon in our amap
1055 	 */
1056 
1057 	anon = anons[centeridx];
1058 	UVMHIST_LOG(maphist, "  case 1 fault: anon=0x%x", anon, 0,0,0);
1059 	simple_lock(&anon->an_lock);
1060 
1061 	/* locked: maps(read), amap, anon */
1062 
1063 	/*
1064 	 * no matter if we have case 1A or case 1B we are going to need to
1065 	 * have the anon's memory resident.   ensure that now.
1066 	 */
1067 
1068 	/*
1069 	 * let uvmfault_anonget do the dirty work.
1070 	 * if it fails (!OK) it will unlock all but the anon for us.
1071 	 * if it succeeds, locks are still valid and locked.
1072 	 * also, if it is OK, then the anon's page is on the queues.
1073 	 * if the page is on loan from a uvm_object, then anonget will
1074 	 * lock that object for us if it does not fail.
1075 	 */
1076 
1077 	result = uvmfault_anonget(&ufi, amap, anon);
1078 	if (result != VM_PAGER_OK) {
1079 		simple_unlock(&anon->an_lock);
1080 	}
1081 
1082 	if (result == VM_PAGER_REFAULT)
1083 		goto ReFault;
1084 
1085 	if (result == VM_PAGER_AGAIN) {
1086 		tsleep((caddr_t)&lbolt, PVM, "fltagain1", 0);
1087 		goto ReFault;
1088 	}
1089 
1090 	if (result != VM_PAGER_OK)
1091 		return (KERN_PROTECTION_FAILURE);		/* XXX??? */
1092 
1093 	/*
1094 	 * uobj is non null if the page is on loan from an object (i.e. uobj)
1095 	 */
1096 
1097 	uobj = anon->u.an_page->uobject;	/* locked by anonget if !NULL */
1098 
1099 	/* locked: maps(read), amap, anon, uobj(if one) */
1100 
1101 	/*
1102 	 * special handling for loaned pages
1103 	 */
1104 	if (anon->u.an_page->loan_count) {
1105 
1106 		if ((access_type & VM_PROT_WRITE) == 0) {
1107 
1108 			/*
1109 			 * for read faults on loaned pages we just cap the
1110 			 * protection at read-only.
1111 			 */
1112 
1113 			enter_prot = enter_prot & ~VM_PROT_WRITE;
1114 
1115 		} else {
1116 			/*
1117 			 * note that we can't allow writes into a loaned page!
1118 			 *
1119 			 * if we have a write fault on a loaned page in an
1120 			 * anon then we need to look at the anon's ref count.
1121 			 * if it is greater than one then we are going to do
1122 			 * a normal copy-on-write fault into a new anon (this
1123 			 * is not a problem).  however, if the reference count
1124 			 * is one (a case where we would normally allow a
1125 			 * write directly to the page) then we need to kill
1126 			 * the loan before we continue.
1127 			 */
1128 
1129 			/* >1 case is already ok */
1130 			if (anon->an_ref == 1) {
1131 
1132 				/* get new un-owned replacement page */
1133 				pg = uvm_pagealloc(NULL, 0, NULL, 0);
1134 				if (pg == NULL) {
1135 					uvmfault_unlockall(&ufi, amap, uobj,
1136 					    anon);
1137 					uvm_wait("flt_noram2");
1138 					goto ReFault;
1139 				}
1140 
1141 				/*
1142 				 * copy data, kill loan, and drop uobj lock
1143 				 * (if any)
1144 				 */
1145 				/* copy old -> new */
1146 				uvm_pagecopy(anon->u.an_page, pg);
1147 
1148 				/* force reload */
1149 				pmap_page_protect(anon->u.an_page,
1150 						  VM_PROT_NONE);
1151 				uvm_lock_pageq();	  /* KILL loan */
1152 				if (uobj)
1153 					/* if we were loaning */
1154 					anon->u.an_page->loan_count--;
1155 				anon->u.an_page->uanon = NULL;
1156 				/* in case we owned */
1157 				anon->u.an_page->pqflags &= ~PQ_ANON;
1158 				uvm_unlock_pageq();
1159 				if (uobj) {
1160 					simple_unlock(&uobj->vmobjlock);
1161 					uobj = NULL;
1162 				}
1163 
1164 				/* install new page in anon */
1165 				anon->u.an_page = pg;
1166 				pg->uanon = anon;
1167 				pg->pqflags |= PQ_ANON;
1168 				pg->flags &= ~(PG_BUSY|PG_FAKE);
1169 				UVM_PAGE_OWN(pg, NULL);
1170 
1171 				/* done! */
1172 			}     /* ref == 1 */
1173 		}       /* write fault */
1174 	}         /* loan count */
1175 
1176 	/*
1177 	 * if we are case 1B then we will need to allocate a new blank
1178 	 * anon to transfer the data into.   note that we have a lock
1179 	 * on anon, so no one can busy or release the page until we are done.
1180 	 * also note that the ref count can't drop to zero here because
1181 	 * it is > 1 and we are only dropping one ref.
1182 	 *
1183 	 * in the (hopefully very rare) case that we are out of RAM we
1184 	 * will unlock, wait for more RAM, and refault.
1185 	 *
1186 	 * if we are out of anon VM we kill the process (XXX: could wait?).
1187 	 */
1188 
1189 	if ((access_type & VM_PROT_WRITE) != 0 && anon->an_ref > 1) {
1190 
1191 		UVMHIST_LOG(maphist, "  case 1B: COW fault",0,0,0,0);
1192 		uvmexp.flt_acow++;
1193 		oanon = anon;		/* oanon = old, locked anon */
1194 		anon = uvm_analloc();
1195 		if (anon)
1196 			pg = uvm_pagealloc(NULL, 0, anon, 0);
1197 #ifdef __GNUC__
1198 		else
1199 			pg = NULL; /* XXX: gcc */
1200 #endif
1201 
1202 		/* check for out of RAM */
1203 		if (anon == NULL || pg == NULL) {
1204 			if (anon)
1205 				uvm_anfree(anon);
1206 			uvmfault_unlockall(&ufi, amap, uobj, oanon);
1207 #ifdef DIAGNOSTIC
1208 			if (uvmexp.swpgonly > uvmexp.swpages) {
1209 				panic("uvmexp.swpgonly botch");
1210 			}
1211 #endif
1212 			if (anon == NULL || uvmexp.swpgonly == uvmexp.swpages) {
1213 				UVMHIST_LOG(maphist,
1214 				    "<- failed.  out of VM",0,0,0,0);
1215 				uvmexp.fltnoanon++;
1216 				return (KERN_RESOURCE_SHORTAGE);
1217 			}
1218 
1219 			uvmexp.fltnoram++;
1220 			uvm_wait("flt_noram3");	/* out of RAM, wait for more */
1221 			goto ReFault;
1222 		}
1223 
1224 		/* got all resources, replace anon with nanon */
1225 
1226 		uvm_pagecopy(oanon->u.an_page, pg);	/* pg now !PG_CLEAN */
1227 		pg->flags &= ~(PG_BUSY|PG_FAKE);	/* un-busy! new page */
1228 		UVM_PAGE_OWN(pg, NULL);
1229 		amap_add(&ufi.entry->aref, ufi.orig_rvaddr - ufi.entry->start,
1230 		    anon, 1);
1231 
1232 		/* deref: can not drop to zero here by defn! */
1233 		oanon->an_ref--;
1234 
1235 		/*
1236 		 * note: oanon still locked.   anon is _not_ locked, but we
1237 		 * have the sole references to in from amap which _is_ locked.
1238 		 * thus, no one can get at it until we are done with it.
1239 		 */
1240 
1241 	} else {
1242 
1243 		uvmexp.flt_anon++;
1244 		oanon = anon;		/* old, locked anon is same as anon */
1245 		pg = anon->u.an_page;
1246 		if (anon->an_ref > 1)     /* disallow writes to ref > 1 anons */
1247 			enter_prot = enter_prot & ~VM_PROT_WRITE;
1248 
1249 	}
1250 
1251 	/* locked: maps(read), amap, anon */
1252 
1253 	/*
1254 	 * now map the page in ...
1255 	 * XXX: old fault unlocks object before pmap_enter.  this seems
1256 	 * suspect since some other thread could blast the page out from
1257 	 * under us between the unlock and the pmap_enter.
1258 	 */
1259 
1260 	UVMHIST_LOG(maphist, "  MAPPING: anon: pm=0x%x, va=0x%x, pg=0x%x",
1261 	    ufi.orig_map->pmap, ufi.orig_rvaddr, pg, 0);
1262 	if (pmap_enter(ufi.orig_map->pmap, ufi.orig_rvaddr, VM_PAGE_TO_PHYS(pg),
1263 	    enter_prot, access_type | PMAP_CANFAIL | (wired ? PMAP_WIRED : 0))
1264 	    != KERN_SUCCESS) {
1265 		/*
1266 		 * No need to undo what we did; we can simply think of
1267 		 * this as the pmap throwing away the mapping information.
1268 		 *
1269 		 * We do, however, have to go through the ReFault path,
1270 		 * as the map may change while we're asleep.
1271 		 */
1272 		uvmfault_unlockall(&ufi, amap, uobj, oanon);
1273 #ifdef DIAGNOSTIC
1274 		if (uvmexp.swpgonly > uvmexp.swpages)
1275 			panic("uvmexp.swpgonly botch");
1276 #endif
1277 		if (uvmexp.swpgonly == uvmexp.swpages) {
1278 			UVMHIST_LOG(maphist,
1279 			    "<- failed.  out of VM",0,0,0,0);
1280 			/* XXX instrumentation */
1281 			return (KERN_RESOURCE_SHORTAGE);
1282 		}
1283 		/* XXX instrumentation */
1284 		uvm_wait("flt_pmfail1");
1285 		goto ReFault;
1286 	}
1287 
1288 	/*
1289 	 * ... update the page queues.
1290 	 */
1291 
1292 	uvm_lock_pageq();
1293 
1294 	if (fault_type == VM_FAULT_WIRE) {
1295 		uvm_pagewire(pg);
1296 
1297 		/*
1298 		 * since the now-wired page cannot be paged out,
1299 		 * release its swap resources for others to use.
1300 		 * since an anon with no swap cannot be PG_CLEAN,
1301 		 * clear its clean flag now.
1302 		 */
1303 
1304 		pg->flags &= ~(PG_CLEAN);
1305 		uvm_anon_dropswap(anon);
1306 	} else {
1307 		/* activate it */
1308 		uvm_pageactivate(pg);
1309 	}
1310 
1311 	uvm_unlock_pageq();
1312 
1313 	/*
1314 	 * done case 1!  finish up by unlocking everything and returning success
1315 	 */
1316 
1317 	uvmfault_unlockall(&ufi, amap, uobj, oanon);
1318 	return (KERN_SUCCESS);
1319 
1320 
1321 Case2:
1322 	/*
1323 	 * handle case 2: faulting on backing object or zero fill
1324 	 */
1325 
1326 	/*
1327 	 * locked:
1328 	 * maps(read), amap(if there), uobj(if !null), uobjpage(if !null)
1329 	 */
1330 
1331 	/*
1332 	 * note that uobjpage can not be PGO_DONTCARE at this point.  we now
1333 	 * set uobjpage to PGO_DONTCARE if we are doing a zero fill.  if we
1334 	 * have a backing object, check and see if we are going to promote
1335 	 * the data up to an anon during the fault.
1336 	 */
1337 
1338 	if (uobj == NULL) {
1339 		uobjpage = PGO_DONTCARE;
1340 		promote = TRUE;		/* always need anon here */
1341 	} else {
1342 		/* assert(uobjpage != PGO_DONTCARE) */
1343 		promote = (access_type & VM_PROT_WRITE) &&
1344 		     UVM_ET_ISCOPYONWRITE(ufi.entry);
1345 	}
1346 	UVMHIST_LOG(maphist, "  case 2 fault: promote=%d, zfill=%d",
1347 	    promote, (uobj == NULL), 0,0);
1348 
1349 	/*
1350 	 * if uobjpage is not null then we do not need to do I/O to get the
1351 	 * uobjpage.
1352 	 *
1353 	 * if uobjpage is null, then we need to unlock and ask the pager to
1354 	 * get the data for us.   once we have the data, we need to reverify
1355 	 * the state the world.   we are currently not holding any resources.
1356 	 */
1357 
1358 	if (uobjpage) {
1359 		/* update rusage counters */
1360 		curproc->p_addr->u_stats.p_ru.ru_minflt++;
1361 	} else {
1362 		/* update rusage counters */
1363 		curproc->p_addr->u_stats.p_ru.ru_majflt++;
1364 
1365 		/* locked: maps(read), amap(if there), uobj */
1366 		uvmfault_unlockall(&ufi, amap, NULL, NULL);
1367 		/* locked: uobj */
1368 
1369 		uvmexp.fltget++;
1370 		gotpages = 1;
1371 		result = uobj->pgops->pgo_get(uobj,
1372 		    (ufi.orig_rvaddr - ufi.entry->start) + ufi.entry->offset,
1373 		    &uobjpage, &gotpages, 0,
1374 			access_type & MASK(ufi.entry),
1375 			ufi.entry->advice, 0);
1376 
1377 		/* locked: uobjpage(if result OK) */
1378 
1379 		/*
1380 		 * recover from I/O
1381 		 */
1382 
1383 		if (result != VM_PAGER_OK) {
1384 #ifdef DIAGNOSTIC
1385 			if (result == VM_PAGER_PEND)
1386 				panic("uvm_fault: pgo_get got PENDing "
1387 				    "on non-async I/O");
1388 #endif
1389 
1390 			if (result == VM_PAGER_AGAIN) {
1391 				UVMHIST_LOG(maphist,
1392 				    "  pgo_get says TRY AGAIN!",0,0,0,0);
1393 				tsleep((caddr_t)&lbolt, PVM, "fltagain2", 0);
1394 				goto ReFault;
1395 			}
1396 
1397 			UVMHIST_LOG(maphist, "<- pgo_get failed (code %d)",
1398 			    result, 0,0,0);
1399 			return (KERN_PROTECTION_FAILURE); /* XXX i/o error */
1400 		}
1401 
1402 		/* locked: uobjpage */
1403 
1404 		/*
1405 		 * re-verify the state of the world by first trying to relock
1406 		 * the maps.  always relock the object.
1407 		 */
1408 
1409 		locked = uvmfault_relock(&ufi);
1410 		if (locked && amap)
1411 			amap_lock(amap);
1412 		simple_lock(&uobj->vmobjlock);
1413 
1414 		/* locked(locked): maps(read), amap(if !null), uobj, uobjpage */
1415 		/* locked(!locked): uobj, uobjpage */
1416 
1417 		/*
1418 		 * verify that the page has not be released and re-verify
1419 		 * that amap slot is still free.   if there is a problem,
1420 		 * we unlock and clean up.
1421 		 */
1422 
1423 		if ((uobjpage->flags & PG_RELEASED) != 0 ||
1424 		    (locked && amap &&
1425 		    amap_lookup(&ufi.entry->aref,
1426 		      ufi.orig_rvaddr - ufi.entry->start))) {
1427 			if (locked)
1428 				uvmfault_unlockall(&ufi, amap, NULL, NULL);
1429 			locked = FALSE;
1430 		}
1431 
1432 		/*
1433 		 * didn't get the lock?   release the page and retry.
1434 		 */
1435 
1436 		if (locked == FALSE) {
1437 
1438 			UVMHIST_LOG(maphist,
1439 			    "  wasn't able to relock after fault: retry",
1440 			    0,0,0,0);
1441 			if (uobjpage->flags & PG_WANTED)
1442 				/* still holding object lock */
1443 				wakeup(uobjpage);
1444 
1445 			if (uobjpage->flags & PG_RELEASED) {
1446 				uvmexp.fltpgrele++;
1447 #ifdef DIAGNOSTIC
1448 				if (uobj->pgops->pgo_releasepg == NULL)
1449 					panic("uvm_fault: object has no "
1450 					    "releasepg function");
1451 #endif
1452 				/* frees page */
1453 				if (uobj->pgops->pgo_releasepg(uobjpage,NULL))
1454 					/* unlock if still alive */
1455 					simple_unlock(&uobj->vmobjlock);
1456 				goto ReFault;
1457 			}
1458 
1459 			uvm_lock_pageq();
1460 			/* make sure it is in queues */
1461 			uvm_pageactivate(uobjpage);
1462 
1463 			uvm_unlock_pageq();
1464 			uobjpage->flags &= ~(PG_BUSY|PG_WANTED);
1465 			UVM_PAGE_OWN(uobjpage, NULL);
1466 			simple_unlock(&uobj->vmobjlock);
1467 			goto ReFault;
1468 
1469 		}
1470 
1471 		/*
1472 		 * we have the data in uobjpage which is PG_BUSY and
1473 		 * !PG_RELEASED.  we are holding object lock (so the page
1474 		 * can't be released on us).
1475 		 */
1476 
1477 		/* locked: maps(read), amap(if !null), uobj, uobjpage */
1478 
1479 	}
1480 
1481 	/*
1482 	 * locked:
1483 	 * maps(read), amap(if !null), uobj(if !null), uobjpage(if uobj)
1484 	 */
1485 
1486 	/*
1487 	 * notes:
1488 	 *  - at this point uobjpage can not be NULL
1489 	 *  - at this point uobjpage can not be PG_RELEASED (since we checked
1490 	 *  for it above)
1491 	 *  - at this point uobjpage could be PG_WANTED (handle later)
1492 	 */
1493 
1494 	if (promote == FALSE) {
1495 
1496 		/*
1497 		 * we are not promoting.   if the mapping is COW ensure that we
1498 		 * don't give more access than we should (e.g. when doing a read
1499 		 * fault on a COPYONWRITE mapping we want to map the COW page in
1500 		 * R/O even though the entry protection could be R/W).
1501 		 *
1502 		 * set "pg" to the page we want to map in (uobjpage, usually)
1503 		 */
1504 
1505 		uvmexp.flt_obj++;
1506 		if (UVM_ET_ISCOPYONWRITE(ufi.entry))
1507 			enter_prot &= ~VM_PROT_WRITE;
1508 		pg = uobjpage;		/* map in the actual object */
1509 
1510 		/* assert(uobjpage != PGO_DONTCARE) */
1511 
1512 		/*
1513 		 * we are faulting directly on the page.   be careful
1514 		 * about writing to loaned pages...
1515 		 */
1516 		if (uobjpage->loan_count) {
1517 
1518 			if ((access_type & VM_PROT_WRITE) == 0) {
1519 				/* read fault: cap the protection at readonly */
1520 				/* cap! */
1521 				enter_prot = enter_prot & ~VM_PROT_WRITE;
1522 			} else {
1523 				/* write fault: must break the loan here */
1524 
1525 				/* alloc new un-owned page */
1526 				pg = uvm_pagealloc(NULL, 0, NULL, 0);
1527 
1528 				if (pg == NULL) {
1529 					/*
1530 					 * drop ownership of page, it can't
1531 					 * be released
1532 					 */
1533 					if (uobjpage->flags & PG_WANTED)
1534 						wakeup(uobjpage);
1535 					uobjpage->flags &= ~(PG_BUSY|PG_WANTED);
1536 					UVM_PAGE_OWN(uobjpage, NULL);
1537 
1538 					uvm_lock_pageq();
1539 					/* activate: we will need it later */
1540 					uvm_pageactivate(uobjpage);
1541 
1542 					uvm_unlock_pageq();
1543 					uvmfault_unlockall(&ufi, amap, uobj,
1544 					  NULL);
1545 					UVMHIST_LOG(maphist,
1546 					  "  out of RAM breaking loan, waiting",
1547 					  0,0,0,0);
1548 					uvmexp.fltnoram++;
1549 					uvm_wait("flt_noram4");
1550 					goto ReFault;
1551 				}
1552 
1553 				/*
1554 				 * copy the data from the old page to the new
1555 				 * one and clear the fake/clean flags on the
1556 				 * new page (keep it busy).  force a reload
1557 				 * of the old page by clearing it from all
1558 				 * pmaps.  then lock the page queues to
1559 				 * rename the pages.
1560 				 */
1561 				uvm_pagecopy(uobjpage, pg);	/* old -> new */
1562 				pg->flags &= ~(PG_FAKE|PG_CLEAN);
1563 				pmap_page_protect(uobjpage, VM_PROT_NONE);
1564 				if (uobjpage->flags & PG_WANTED)
1565 					wakeup(uobjpage);
1566 				/* uobj still locked */
1567 				uobjpage->flags &= ~(PG_WANTED|PG_BUSY);
1568 				UVM_PAGE_OWN(uobjpage, NULL);
1569 
1570 				uvm_lock_pageq();
1571 				offset = uobjpage->offset;
1572 				/* remove old page */
1573 				uvm_pagerealloc(uobjpage, NULL, 0);
1574 
1575 				/*
1576 				 * at this point we have absolutely no
1577 				 * control over uobjpage
1578 				 */
1579 				/* install new page */
1580 				uvm_pagerealloc(pg, uobj, offset);
1581 				uvm_unlock_pageq();
1582 
1583 				/*
1584 				 * done!  loan is broken and "pg" is
1585 				 * PG_BUSY.   it can now replace uobjpage.
1586 				 */
1587 
1588 				uobjpage = pg;
1589 
1590 			}		/* write fault case */
1591 		}		/* if loan_count */
1592 
1593 	} else {
1594 
1595 		/*
1596 		 * if we are going to promote the data to an anon we
1597 		 * allocate a blank anon here and plug it into our amap.
1598 		 */
1599 #if DIAGNOSTIC
1600 		if (amap == NULL)
1601 			panic("uvm_fault: want to promote data, but no anon");
1602 #endif
1603 
1604 		anon = uvm_analloc();
1605 		if (anon) {
1606 			/*
1607 			 * In `Fill in data...' below, if
1608 			 * uobjpage == PGO_DONTCARE, we want
1609 			 * a zero'd, dirty page, so have
1610 			 * uvm_pagealloc() do that for us.
1611 			 */
1612 			pg = uvm_pagealloc(NULL, 0, anon,
1613 			    (uobjpage == PGO_DONTCARE) ? UVM_PGA_ZERO : 0);
1614 		}
1615 #ifdef __GNUC__
1616 		else
1617 			pg = NULL; /* XXX: gcc */
1618 #endif
1619 
1620 		/*
1621 		 * out of memory resources?
1622 		 */
1623 		if (anon == NULL || pg == NULL) {
1624 
1625 			/*
1626 			 * arg!  must unbusy our page and fail or sleep.
1627 			 */
1628 			if (uobjpage != PGO_DONTCARE) {
1629 				if (uobjpage->flags & PG_WANTED)
1630 					/* still holding object lock */
1631 					wakeup(uobjpage);
1632 
1633 				uvm_lock_pageq();
1634 				/* make sure it is in queues */
1635 				uvm_pageactivate(uobjpage);
1636 				uvm_unlock_pageq();
1637 				/* un-busy! (still locked) */
1638 				uobjpage->flags &= ~(PG_BUSY|PG_WANTED);
1639 				UVM_PAGE_OWN(uobjpage, NULL);
1640 			}
1641 
1642 			/* unlock and fail ... */
1643 			uvmfault_unlockall(&ufi, amap, uobj, NULL);
1644 #ifdef DIAGNOSTIC
1645 			if (uvmexp.swpgonly > uvmexp.swpages) {
1646 				panic("uvmexp.swpgonly botch");
1647 			}
1648 #endif
1649 			if (anon == NULL || uvmexp.swpgonly == uvmexp.swpages) {
1650 				UVMHIST_LOG(maphist, "  promote: out of VM",
1651 				    0,0,0,0);
1652 				uvmexp.fltnoanon++;
1653 				return (KERN_RESOURCE_SHORTAGE);
1654 			}
1655 
1656 			UVMHIST_LOG(maphist, "  out of RAM, waiting for more",
1657 			    0,0,0,0);
1658 			uvm_anfree(anon);
1659 			uvmexp.fltnoram++;
1660 			uvm_wait("flt_noram5");
1661 			goto ReFault;
1662 		}
1663 
1664 		/*
1665 		 * fill in the data
1666 		 */
1667 
1668 		if (uobjpage != PGO_DONTCARE) {
1669 			uvmexp.flt_prcopy++;
1670 			/* copy page [pg now dirty] */
1671 			uvm_pagecopy(uobjpage, pg);
1672 
1673 			/*
1674 			 * promote to shared amap?  make sure all sharing
1675 			 * procs see it
1676 			 */
1677 			if ((amap_flags(amap) & AMAP_SHARED) != 0) {
1678 				pmap_page_protect(uobjpage, VM_PROT_NONE);
1679 			}
1680 
1681 			/*
1682 			 * dispose of uobjpage.  it can't be PG_RELEASED
1683 			 * since we still hold the object lock.   drop
1684 			 * handle to uobj as well.
1685 			 */
1686 
1687 			if (uobjpage->flags & PG_WANTED)
1688 				/* still have the obj lock */
1689 				wakeup(uobjpage);
1690 			uobjpage->flags &= ~(PG_BUSY|PG_WANTED);
1691 			UVM_PAGE_OWN(uobjpage, NULL);
1692 			uvm_lock_pageq();
1693 			uvm_pageactivate(uobjpage);	/* put it back */
1694 			uvm_unlock_pageq();
1695 			simple_unlock(&uobj->vmobjlock);
1696 			uobj = NULL;
1697 			UVMHIST_LOG(maphist,
1698 			    "  promote uobjpage 0x%x to anon/page 0x%x/0x%x",
1699 			    uobjpage, anon, pg, 0);
1700 
1701 		} else {
1702 			uvmexp.flt_przero++;
1703 			/*
1704 			 * Page is zero'd and marked dirty by uvm_pagealloc()
1705 			 * above.
1706 			 */
1707 			UVMHIST_LOG(maphist,"  zero fill anon/page 0x%x/0%x",
1708 			    anon, pg, 0, 0);
1709 		}
1710 
1711 		amap_add(&ufi.entry->aref, ufi.orig_rvaddr - ufi.entry->start,
1712 		    anon, 0);
1713 
1714 	}
1715 
1716 	/*
1717 	 * locked:
1718 	 * maps(read), amap(if !null), uobj(if !null), uobjpage(if uobj)
1719 	 *
1720 	 * note: pg is either the uobjpage or the new page in the new anon
1721 	 */
1722 
1723 	/*
1724 	 * all resources are present.   we can now map it in and free our
1725 	 * resources.
1726 	 */
1727 
1728 	UVMHIST_LOG(maphist,
1729 	    "  MAPPING: case2: pm=0x%x, va=0x%x, pg=0x%x, promote=%d",
1730 	    ufi.orig_map->pmap, ufi.orig_rvaddr, pg, promote);
1731 	if (pmap_enter(ufi.orig_map->pmap, ufi.orig_rvaddr, VM_PAGE_TO_PHYS(pg),
1732 	    enter_prot, access_type | PMAP_CANFAIL | (wired ? PMAP_WIRED : 0))
1733 	    != KERN_SUCCESS) {
1734 		/*
1735 		 * No need to undo what we did; we can simply think of
1736 		 * this as the pmap throwing away the mapping information.
1737 		 *
1738 		 * We do, however, have to go through the ReFault path,
1739 		 * as the map may change while we're asleep.
1740 		 */
1741 		if (pg->flags & PG_WANTED)
1742 			wakeup(pg);		/* lock still held */
1743 
1744 		/*
1745 		 * note that pg can't be PG_RELEASED since we did not drop
1746 		 * the object lock since the last time we checked.
1747 		 */
1748 
1749 		pg->flags &= ~(PG_BUSY|PG_FAKE|PG_WANTED);
1750 		UVM_PAGE_OWN(pg, NULL);
1751 		uvmfault_unlockall(&ufi, amap, uobj, NULL);
1752 #ifdef DIAGNOSTIC
1753 		if (uvmexp.swpgonly > uvmexp.swpages)
1754 			panic("uvmexp.swpgonly botch");
1755 #endif
1756 		if (uvmexp.swpgonly == uvmexp.swpages) {
1757 			UVMHIST_LOG(maphist,
1758 			    "<- failed.  out of VM",0,0,0,0);
1759 			/* XXX instrumentation */
1760 			return (KERN_RESOURCE_SHORTAGE);
1761 		}
1762 		/* XXX instrumentation */
1763 		uvm_wait("flt_pmfail2");
1764 		goto ReFault;
1765 	}
1766 
1767 	uvm_lock_pageq();
1768 
1769 	if (fault_type == VM_FAULT_WIRE) {
1770 		uvm_pagewire(pg);
1771 		if (pg->pqflags & PQ_AOBJ) {
1772 
1773 			/*
1774 			 * since the now-wired page cannot be paged out,
1775 			 * release its swap resources for others to use.
1776 			 * since an aobj page with no swap cannot be PG_CLEAN,
1777 			 * clear its clean flag now.
1778 			 */
1779 
1780 			pg->flags &= ~(PG_CLEAN);
1781 			uao_dropswap(uobj, pg->offset >> PAGE_SHIFT);
1782 		}
1783 	} else {
1784 		/* activate it */
1785 		uvm_pageactivate(pg);
1786 	}
1787 
1788 	uvm_unlock_pageq();
1789 
1790 	if (pg->flags & PG_WANTED)
1791 		wakeup(pg);		/* lock still held */
1792 
1793 	/*
1794 	 * note that pg can't be PG_RELEASED since we did not drop the object
1795 	 * lock since the last time we checked.
1796 	 */
1797 
1798 	pg->flags &= ~(PG_BUSY|PG_FAKE|PG_WANTED);
1799 	UVM_PAGE_OWN(pg, NULL);
1800 	uvmfault_unlockall(&ufi, amap, uobj, NULL);
1801 
1802 	UVMHIST_LOG(maphist, "<- done (SUCCESS!)",0,0,0,0);
1803 	return (KERN_SUCCESS);
1804 }
1805 
1806 
1807 /*
1808  * uvm_fault_wire: wire down a range of virtual addresses in a map.
1809  *
1810  * => map may be read-locked by caller, but MUST NOT be write-locked.
1811  * => if map is read-locked, any operations which may cause map to
1812  *	be write-locked in uvm_fault() must be taken care of by
1813  *	the caller.  See uvm_map_pageable().
1814  */
1815 
1816 int
1817 uvm_fault_wire(map, start, end, access_type)
1818 	vm_map_t map;
1819 	vaddr_t start, end;
1820 	vm_prot_t access_type;
1821 {
1822 	vaddr_t va;
1823 	pmap_t  pmap;
1824 	int rv;
1825 
1826 	pmap = vm_map_pmap(map);
1827 
1828 	/*
1829 	 * now fault it in a page at a time.   if the fault fails then we have
1830 	 * to undo what we have done.   note that in uvm_fault VM_PROT_NONE
1831 	 * is replaced with the max protection if fault_type is VM_FAULT_WIRE.
1832 	 */
1833 
1834 	for (va = start ; va < end ; va += PAGE_SIZE) {
1835 		rv = uvm_fault(map, va, VM_FAULT_WIRE, access_type);
1836 		if (rv) {
1837 			if (va != start) {
1838 				uvm_fault_unwire(map, start, va);
1839 			}
1840 			return (rv);
1841 		}
1842 	}
1843 
1844 	return (KERN_SUCCESS);
1845 }
1846 
1847 /*
1848  * uvm_fault_unwire(): unwire range of virtual space.
1849  */
1850 
1851 void
1852 uvm_fault_unwire(map, start, end)
1853 	vm_map_t map;
1854 	vaddr_t start, end;
1855 {
1856 
1857 	vm_map_lock_read(map);
1858 	uvm_fault_unwire_locked(map, start, end);
1859 	vm_map_unlock_read(map);
1860 }
1861 
1862 /*
1863  * uvm_fault_unwire_locked(): the guts of uvm_fault_unwire().
1864  *
1865  * => map must be at least read-locked.
1866  */
1867 
1868 void
1869 uvm_fault_unwire_locked(map, start, end)
1870 	vm_map_t map;
1871 	vaddr_t start, end;
1872 {
1873 	vm_map_entry_t entry;
1874 	pmap_t pmap = vm_map_pmap(map);
1875 	vaddr_t va;
1876 	paddr_t pa;
1877 	struct vm_page *pg;
1878 
1879 #ifdef DIAGNOSTIC
1880 	if (map->flags & VM_MAP_INTRSAFE)
1881 		panic("uvm_fault_unwire_locked: intrsafe map");
1882 #endif
1883 
1884 	/*
1885 	 * we assume that the area we are unwiring has actually been wired
1886 	 * in the first place.   this means that we should be able to extract
1887 	 * the PAs from the pmap.   we also lock out the page daemon so that
1888 	 * we can call uvm_pageunwire.
1889 	 */
1890 
1891 	uvm_lock_pageq();
1892 
1893 	/*
1894 	 * find the beginning map entry for the region.
1895 	 */
1896 #ifdef DIAGNOSTIC
1897 	if (start < vm_map_min(map) || end > vm_map_max(map))
1898 		panic("uvm_fault_unwire_locked: address out of range");
1899 #endif
1900 	if (uvm_map_lookup_entry(map, start, &entry) == FALSE)
1901 		panic("uvm_fault_unwire_locked: address not in map");
1902 
1903 	for (va = start; va < end ; va += PAGE_SIZE) {
1904 		if (pmap_extract(pmap, va, &pa) == FALSE)
1905 			panic("uvm_fault_unwire_locked: unwiring "
1906 			    "non-wired memory");
1907 
1908 		/*
1909 		 * make sure the current entry is for the address we're
1910 		 * dealing with.  if not, grab the next entry.
1911 		 */
1912 #ifdef DIAGNOSTIC
1913 		if (va < entry->start)
1914 			panic("uvm_fault_unwire_locked: hole 1");
1915 #endif
1916 		if (va >= entry->end) {
1917 #ifdef DIAGNOSTIC
1918 			if (entry->next == &map->header ||
1919 			    entry->next->start > entry->end)
1920 				panic("uvm_fault_unwire_locked: hole 2");
1921 #endif
1922 			entry = entry->next;
1923 		}
1924 
1925 		/*
1926 		 * if the entry is no longer wired, tell the pmap.
1927 		 */
1928 		if (VM_MAPENT_ISWIRED(entry) == 0)
1929 			pmap_unwire(pmap, va);
1930 
1931 		pg = PHYS_TO_VM_PAGE(pa);
1932 		if (pg)
1933 			uvm_pageunwire(pg);
1934 	}
1935 
1936 	uvm_unlock_pageq();
1937 }
1938