xref: /netbsd-src/sys/uvm/uvm_fault.c (revision aaf4ece63a859a04e37cf3a7229b5fab0157cc06)
1 /*	$NetBSD: uvm_fault.c,v 1.103 2005/12/24 20:45:10 perry Exp $	*/
2 
3 /*
4  *
5  * Copyright (c) 1997 Charles D. Cranor and Washington University.
6  * All rights reserved.
7  *
8  * Redistribution and use in source and binary forms, with or without
9  * modification, are permitted provided that the following conditions
10  * are met:
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in the
15  *    documentation and/or other materials provided with the distribution.
16  * 3. All advertising materials mentioning features or use of this software
17  *    must display the following acknowledgement:
18  *      This product includes software developed by Charles D. Cranor and
19  *      Washington University.
20  * 4. The name of the author may not be used to endorse or promote products
21  *    derived from this software without specific prior written permission.
22  *
23  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
24  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
25  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
26  * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
27  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
28  * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
29  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
30  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
31  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
32  * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
33  *
34  * from: Id: uvm_fault.c,v 1.1.2.23 1998/02/06 05:29:05 chs Exp
35  */
36 
37 /*
38  * uvm_fault.c: fault handler
39  */
40 
41 #include <sys/cdefs.h>
42 __KERNEL_RCSID(0, "$NetBSD: uvm_fault.c,v 1.103 2005/12/24 20:45:10 perry Exp $");
43 
44 #include "opt_uvmhist.h"
45 
46 #include <sys/param.h>
47 #include <sys/systm.h>
48 #include <sys/kernel.h>
49 #include <sys/proc.h>
50 #include <sys/malloc.h>
51 #include <sys/mman.h>
52 #include <sys/user.h>
53 #include <sys/vnode.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 const 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 /*
184  * inline functions
185  */
186 
187 /*
188  * uvmfault_anonflush: try and deactivate pages in specified anons
189  *
190  * => does not have to deactivate page if it is busy
191  */
192 
193 static inline void
194 uvmfault_anonflush(struct vm_anon **anons, 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->flags & PG_BUSY) == 0 && pg->loan_count == 0) {
205 			uvm_lock_pageq();
206 			if (pg->wire_count == 0) {
207 				pmap_clear_reference(pg);
208 				uvm_pagedeactivate(pg);
209 			}
210 			uvm_unlock_pageq();
211 		}
212 		simple_unlock(&anons[lcv]->an_lock);
213 	}
214 }
215 
216 /*
217  * normal functions
218  */
219 
220 /*
221  * uvmfault_amapcopy: clear "needs_copy" in a map.
222  *
223  * => called with VM data structures unlocked (usually, see below)
224  * => we get a write lock on the maps and clear needs_copy for a VA
225  * => if we are out of RAM we sleep (waiting for more)
226  */
227 
228 static void
229 uvmfault_amapcopy(struct uvm_faultinfo *ufi)
230 {
231 	for (;;) {
232 
233 		/*
234 		 * no mapping?  give up.
235 		 */
236 
237 		if (uvmfault_lookup(ufi, TRUE) == FALSE)
238 			return;
239 
240 		/*
241 		 * copy if needed.
242 		 */
243 
244 		if (UVM_ET_ISNEEDSCOPY(ufi->entry))
245 			amap_copy(ufi->map, ufi->entry, M_NOWAIT, TRUE,
246 				ufi->orig_rvaddr, ufi->orig_rvaddr + 1);
247 
248 		/*
249 		 * didn't work?  must be out of RAM.   unlock and sleep.
250 		 */
251 
252 		if (UVM_ET_ISNEEDSCOPY(ufi->entry)) {
253 			uvmfault_unlockmaps(ufi, TRUE);
254 			uvm_wait("fltamapcopy");
255 			continue;
256 		}
257 
258 		/*
259 		 * got it!   unlock and return.
260 		 */
261 
262 		uvmfault_unlockmaps(ufi, TRUE);
263 		return;
264 	}
265 	/*NOTREACHED*/
266 }
267 
268 /*
269  * uvmfault_anonget: get data in an anon into a non-busy, non-released
270  * page in that anon.
271  *
272  * => maps, amap, and anon locked by caller.
273  * => if we fail (result != 0) we unlock everything.
274  * => if we are successful, we return with everything still locked.
275  * => we don't move the page on the queues [gets moved later]
276  * => if we allocate a new page [we_own], it gets put on the queues.
277  *    either way, the result is that the page is on the queues at return time
278  * => for pages which are on loan from a uvm_object (and thus are not
279  *    owned by the anon): if successful, we return with the owning object
280  *    locked.   the caller must unlock this object when it unlocks everything
281  *    else.
282  */
283 
284 int
285 uvmfault_anonget(struct uvm_faultinfo *ufi, struct vm_amap *amap,
286     struct vm_anon *anon)
287 {
288 	boolean_t we_own;	/* we own anon's page? */
289 	boolean_t locked;	/* did we relock? */
290 	struct vm_page *pg;
291 	int error;
292 	UVMHIST_FUNC("uvmfault_anonget"); UVMHIST_CALLED(maphist);
293 
294 	LOCK_ASSERT(simple_lock_held(&anon->an_lock));
295 
296 	error = 0;
297 	uvmexp.fltanget++;
298         /* bump rusage counters */
299 	if (anon->an_page)
300 		curproc->p_stats->p_ru.ru_minflt++;
301 	else
302 		curproc->p_stats->p_ru.ru_majflt++;
303 
304 	/*
305 	 * loop until we get it, or fail.
306 	 */
307 
308 	for (;;) {
309 		we_own = FALSE;		/* TRUE if we set PG_BUSY on a page */
310 		pg = anon->an_page;
311 
312 		/*
313 		 * if there is a resident page and it is loaned, then anon
314 		 * may not own it.   call out to uvm_anon_lockpage() to ensure
315 		 * the real owner of the page has been identified and locked.
316 		 */
317 
318 		if (pg && pg->loan_count)
319 			pg = uvm_anon_lockloanpg(anon);
320 
321 		/*
322 		 * page there?   make sure it is not busy/released.
323 		 */
324 
325 		if (pg) {
326 
327 			/*
328 			 * at this point, if the page has a uobject [meaning
329 			 * we have it on loan], then that uobject is locked
330 			 * by us!   if the page is busy, we drop all the
331 			 * locks (including uobject) and try again.
332 			 */
333 
334 			if ((pg->flags & PG_BUSY) == 0) {
335 				UVMHIST_LOG(maphist, "<- OK",0,0,0,0);
336 				return (0);
337 			}
338 			pg->flags |= PG_WANTED;
339 			uvmexp.fltpgwait++;
340 
341 			/*
342 			 * the last unlock must be an atomic unlock+wait on
343 			 * the owner of page
344 			 */
345 
346 			if (pg->uobject) {	/* owner is uobject ? */
347 				uvmfault_unlockall(ufi, amap, NULL, anon);
348 				UVMHIST_LOG(maphist, " unlock+wait on uobj",0,
349 				    0,0,0);
350 				UVM_UNLOCK_AND_WAIT(pg,
351 				    &pg->uobject->vmobjlock,
352 				    FALSE, "anonget1",0);
353 			} else {
354 				/* anon owns page */
355 				uvmfault_unlockall(ufi, amap, NULL, NULL);
356 				UVMHIST_LOG(maphist, " unlock+wait on anon",0,
357 				    0,0,0);
358 				UVM_UNLOCK_AND_WAIT(pg,&anon->an_lock,0,
359 				    "anonget2",0);
360 			}
361 		} else {
362 #if defined(VMSWAP)
363 
364 			/*
365 			 * no page, we must try and bring it in.
366 			 */
367 
368 			pg = uvm_pagealloc(NULL, 0, anon, 0);
369 			if (pg == NULL) {		/* out of RAM.  */
370 				uvmfault_unlockall(ufi, amap, NULL, anon);
371 				uvmexp.fltnoram++;
372 				UVMHIST_LOG(maphist, "  noram -- UVM_WAIT",0,
373 				    0,0,0);
374 				if (!uvm_reclaimable()) {
375 					return ENOMEM;
376 				}
377 				uvm_wait("flt_noram1");
378 			} else {
379 				/* we set the PG_BUSY bit */
380 				we_own = TRUE;
381 				uvmfault_unlockall(ufi, amap, NULL, anon);
382 
383 				/*
384 				 * we are passing a PG_BUSY+PG_FAKE+PG_CLEAN
385 				 * page into the uvm_swap_get function with
386 				 * all data structures unlocked.  note that
387 				 * it is ok to read an_swslot here because
388 				 * we hold PG_BUSY on the page.
389 				 */
390 				uvmexp.pageins++;
391 				error = uvm_swap_get(pg, anon->an_swslot,
392 				    PGO_SYNCIO);
393 
394 				/*
395 				 * we clean up after the i/o below in the
396 				 * "we_own" case
397 				 */
398 			}
399 #else /* defined(VMSWAP) */
400 			panic("%s: no page", __func__);
401 #endif /* defined(VMSWAP) */
402 		}
403 
404 		/*
405 		 * now relock and try again
406 		 */
407 
408 		locked = uvmfault_relock(ufi);
409 		if (locked && amap != NULL) {
410 			amap_lock(amap);
411 		}
412 		if (locked || we_own)
413 			simple_lock(&anon->an_lock);
414 
415 		/*
416 		 * if we own the page (i.e. we set PG_BUSY), then we need
417 		 * to clean up after the I/O. there are three cases to
418 		 * consider:
419 		 *   [1] page released during I/O: free anon and ReFault.
420 		 *   [2] I/O not OK.   free the page and cause the fault
421 		 *       to fail.
422 		 *   [3] I/O OK!   activate the page and sync with the
423 		 *       non-we_own case (i.e. drop anon lock if not locked).
424 		 */
425 
426 		if (we_own) {
427 #if defined(VMSWAP)
428 			if (pg->flags & PG_WANTED) {
429 				wakeup(pg);
430 			}
431 			if (error) {
432 
433 				/*
434 				 * remove the swap slot from the anon
435 				 * and mark the anon as having no real slot.
436 				 * don't free the swap slot, thus preventing
437 				 * it from being used again.
438 				 */
439 
440 				if (anon->an_swslot > 0)
441 					uvm_swap_markbad(anon->an_swslot, 1);
442 				anon->an_swslot = SWSLOT_BAD;
443 
444 				if ((pg->flags & PG_RELEASED) != 0)
445 					goto released;
446 
447 				/*
448 				 * note: page was never !PG_BUSY, so it
449 				 * can't be mapped and thus no need to
450 				 * pmap_page_protect it...
451 				 */
452 
453 				uvm_lock_pageq();
454 				uvm_pagefree(pg);
455 				uvm_unlock_pageq();
456 
457 				if (locked)
458 					uvmfault_unlockall(ufi, amap, NULL,
459 					    anon);
460 				else
461 					simple_unlock(&anon->an_lock);
462 				UVMHIST_LOG(maphist, "<- ERROR", 0,0,0,0);
463 				return error;
464 			}
465 
466 			if ((pg->flags & PG_RELEASED) != 0) {
467 released:
468 				KASSERT(anon->an_ref == 0);
469 
470 				/*
471 				 * released while we unlocked amap.
472 				 */
473 
474 				if (locked)
475 					uvmfault_unlockall(ufi, amap, NULL,
476 					    NULL);
477 
478 				uvm_anon_release(anon);
479 
480 				if (error) {
481 					UVMHIST_LOG(maphist,
482 					    "<- ERROR/RELEASED", 0,0,0,0);
483 					return error;
484 				}
485 
486 				UVMHIST_LOG(maphist, "<- RELEASED", 0,0,0,0);
487 				return ERESTART;
488 			}
489 
490 			/*
491 			 * we've successfully read the page, activate it.
492 			 */
493 
494 			uvm_lock_pageq();
495 			uvm_pageactivate(pg);
496 			uvm_unlock_pageq();
497 			pg->flags &= ~(PG_WANTED|PG_BUSY|PG_FAKE);
498 			UVM_PAGE_OWN(pg, NULL);
499 			if (!locked)
500 				simple_unlock(&anon->an_lock);
501 #else /* defined(VMSWAP) */
502 			panic("%s: we_own", __func__);
503 #endif /* defined(VMSWAP) */
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 (ERESTART);
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 (ERESTART);
526 		}
527 
528 		/*
529 		 * try it again!
530 		 */
531 
532 		uvmexp.fltanretry++;
533 		continue;
534 	}
535 	/*NOTREACHED*/
536 }
537 
538 /*
539  *   F A U L T   -   m a i n   e n t r y   p o i n t
540  */
541 
542 /*
543  * uvm_fault: page fault handler
544  *
545  * => called from MD code to resolve a page fault
546  * => VM data structures usually should be unlocked.   however, it is
547  *	possible to call here with the main map locked if the caller
548  *	gets a write lock, sets it recusive, and then calls us (c.f.
549  *	uvm_map_pageable).   this should be avoided because it keeps
550  *	the map locked off during I/O.
551  * => MUST NEVER BE CALLED IN INTERRUPT CONTEXT
552  */
553 
554 #define MASK(entry)     (UVM_ET_ISCOPYONWRITE(entry) ? \
555 			 ~VM_PROT_WRITE : VM_PROT_ALL)
556 
557 int
558 uvm_fault(struct vm_map *orig_map, vaddr_t vaddr, vm_fault_t fault_type,
559     vm_prot_t access_type)
560 {
561 	struct uvm_faultinfo ufi;
562 	vm_prot_t enter_prot, check_prot;
563 	boolean_t wired, narrow, promote, locked, shadowed, wire_fault, cow_now;
564 	int npages, nback, nforw, centeridx, error, lcv, gotpages;
565 	vaddr_t startva, currva;
566 	voff_t uoff;
567 	struct vm_amap *amap;
568 	struct uvm_object *uobj;
569 	struct vm_anon *anons_store[UVM_MAXRANGE], **anons, *anon, *oanon;
570 	struct vm_page *pages[UVM_MAXRANGE], *pg, *uobjpage;
571 	UVMHIST_FUNC("uvm_fault"); UVMHIST_CALLED(maphist);
572 
573 	UVMHIST_LOG(maphist, "(map=0x%x, vaddr=0x%x, ft=%d, at=%d)",
574 	      orig_map, vaddr, fault_type, access_type);
575 
576 	anon = NULL;
577 	pg = NULL;
578 
579 	uvmexp.faults++;	/* XXX: locking? */
580 
581 	/*
582 	 * init the IN parameters in the ufi
583 	 */
584 
585 	ufi.orig_map = orig_map;
586 	ufi.orig_rvaddr = trunc_page(vaddr);
587 	ufi.orig_size = PAGE_SIZE;	/* can't get any smaller than this */
588 	wire_fault = fault_type == VM_FAULT_WIRE ||
589 	    fault_type == VM_FAULT_WIREMAX;
590 	if (wire_fault)
591 		narrow = TRUE;		/* don't look for neighborhood
592 					 * pages on wire */
593 	else
594 		narrow = FALSE;		/* normal fault */
595 
596 	/*
597 	 * "goto ReFault" means restart the page fault from ground zero.
598 	 */
599 ReFault:
600 
601 	/*
602 	 * lookup and lock the maps
603 	 */
604 
605 	if (uvmfault_lookup(&ufi, FALSE) == FALSE) {
606 		UVMHIST_LOG(maphist, "<- no mapping @ 0x%x", vaddr, 0,0,0);
607 		return (EFAULT);
608 	}
609 	/* locked: maps(read) */
610 
611 #ifdef DIAGNOSTIC
612 	if ((ufi.map->flags & VM_MAP_PAGEABLE) == 0) {
613 		printf("Page fault on non-pageable map:\n");
614 		printf("ufi.map = %p\n", ufi.map);
615 		printf("ufi.orig_map = %p\n", ufi.orig_map);
616 		printf("ufi.orig_rvaddr = 0x%lx\n", (u_long) ufi.orig_rvaddr);
617 		panic("uvm_fault: (ufi.map->flags & VM_MAP_PAGEABLE) == 0");
618 	}
619 #endif
620 
621 	/*
622 	 * check protection
623 	 */
624 
625 	check_prot = fault_type == VM_FAULT_WIREMAX ?
626 	    ufi.entry->max_protection : ufi.entry->protection;
627 	if ((check_prot & access_type) != access_type) {
628 		UVMHIST_LOG(maphist,
629 		    "<- protection failure (prot=0x%x, access=0x%x)",
630 		    ufi.entry->protection, access_type, 0, 0);
631 		uvmfault_unlockmaps(&ufi, FALSE);
632 		return EACCES;
633 	}
634 
635 	/*
636 	 * "enter_prot" is the protection we want to enter the page in at.
637 	 * for certain pages (e.g. copy-on-write pages) this protection can
638 	 * be more strict than ufi.entry->protection.  "wired" means either
639 	 * the entry is wired or we are fault-wiring the pg.
640 	 */
641 
642 	enter_prot = ufi.entry->protection;
643 	wired = VM_MAPENT_ISWIRED(ufi.entry) || wire_fault;
644 	if (wired) {
645 		access_type = enter_prot; /* full access for wired */
646 		cow_now = (check_prot & VM_PROT_WRITE) != 0;
647 	} else {
648 		cow_now = (access_type & VM_PROT_WRITE) != 0;
649 	}
650 
651 	/*
652 	 * handle "needs_copy" case.   if we need to copy the amap we will
653 	 * have to drop our readlock and relock it with a write lock.  (we
654 	 * need a write lock to change anything in a map entry [e.g.
655 	 * needs_copy]).
656 	 */
657 
658 	if (UVM_ET_ISNEEDSCOPY(ufi.entry)) {
659 		KASSERT(fault_type != VM_FAULT_WIREMAX);
660 		if (cow_now || (ufi.entry->object.uvm_obj == NULL)) {
661 			/* need to clear */
662 			UVMHIST_LOG(maphist,
663 			    "  need to clear needs_copy and refault",0,0,0,0);
664 			uvmfault_unlockmaps(&ufi, FALSE);
665 			uvmfault_amapcopy(&ufi);
666 			uvmexp.fltamcopy++;
667 			goto ReFault;
668 
669 		} else {
670 
671 			/*
672 			 * ensure that we pmap_enter page R/O since
673 			 * needs_copy is still true
674 			 */
675 
676 			enter_prot &= ~VM_PROT_WRITE;
677 		}
678 	}
679 
680 	/*
681 	 * identify the players
682 	 */
683 
684 	amap = ufi.entry->aref.ar_amap;		/* top layer */
685 	uobj = ufi.entry->object.uvm_obj;	/* bottom layer */
686 
687 	/*
688 	 * check for a case 0 fault.  if nothing backing the entry then
689 	 * error now.
690 	 */
691 
692 	if (amap == NULL && uobj == NULL) {
693 		uvmfault_unlockmaps(&ufi, FALSE);
694 		UVMHIST_LOG(maphist,"<- no backing store, no overlay",0,0,0,0);
695 		return (EFAULT);
696 	}
697 
698 	/*
699 	 * establish range of interest based on advice from mapper
700 	 * and then clip to fit map entry.   note that we only want
701 	 * to do this the first time through the fault.   if we
702 	 * ReFault we will disable this by setting "narrow" to true.
703 	 */
704 
705 	if (narrow == FALSE) {
706 
707 		/* wide fault (!narrow) */
708 		KASSERT(uvmadvice[ufi.entry->advice].advice ==
709 			 ufi.entry->advice);
710 		nback = MIN(uvmadvice[ufi.entry->advice].nback,
711 			    (ufi.orig_rvaddr - ufi.entry->start) >> PAGE_SHIFT);
712 		startva = ufi.orig_rvaddr - (nback << PAGE_SHIFT);
713 		nforw = MIN(uvmadvice[ufi.entry->advice].nforw,
714 			    ((ufi.entry->end - ufi.orig_rvaddr) >>
715 			     PAGE_SHIFT) - 1);
716 		/*
717 		 * note: "-1" because we don't want to count the
718 		 * faulting page as forw
719 		 */
720 		npages = nback + nforw + 1;
721 		centeridx = nback;
722 
723 		narrow = TRUE;	/* ensure only once per-fault */
724 
725 	} else {
726 
727 		/* narrow fault! */
728 		nback = nforw = 0;
729 		startva = ufi.orig_rvaddr;
730 		npages = 1;
731 		centeridx = 0;
732 
733 	}
734 
735 	/* locked: maps(read) */
736 	UVMHIST_LOG(maphist, "  narrow=%d, back=%d, forw=%d, startva=0x%x",
737 		    narrow, nback, nforw, startva);
738 	UVMHIST_LOG(maphist, "  entry=0x%x, amap=0x%x, obj=0x%x", ufi.entry,
739 		    amap, uobj, 0);
740 
741 	/*
742 	 * if we've got an amap, lock it and extract current anons.
743 	 */
744 
745 	if (amap) {
746 		amap_lock(amap);
747 		anons = anons_store;
748 		amap_lookups(&ufi.entry->aref, startva - ufi.entry->start,
749 		    anons, npages);
750 	} else {
751 		anons = NULL;	/* to be safe */
752 	}
753 
754 	/* locked: maps(read), amap(if there) */
755 
756 	/*
757 	 * for MADV_SEQUENTIAL mappings we want to deactivate the back pages
758 	 * now and then forget about them (for the rest of the fault).
759 	 */
760 
761 	if (ufi.entry->advice == MADV_SEQUENTIAL && nback != 0) {
762 
763 		UVMHIST_LOG(maphist, "  MADV_SEQUENTIAL: flushing backpages",
764 		    0,0,0,0);
765 		/* flush back-page anons? */
766 		if (amap)
767 			uvmfault_anonflush(anons, nback);
768 
769 		/* flush object? */
770 		if (uobj) {
771 			uoff = (startva - ufi.entry->start) + ufi.entry->offset;
772 			simple_lock(&uobj->vmobjlock);
773 			(void) (uobj->pgops->pgo_put)(uobj, uoff, uoff +
774 				    (nback << PAGE_SHIFT), PGO_DEACTIVATE);
775 		}
776 
777 		/* now forget about the backpages */
778 		if (amap)
779 			anons += nback;
780 		startva += (nback << PAGE_SHIFT);
781 		npages -= nback;
782 		nback = centeridx = 0;
783 	}
784 
785 	/* locked: maps(read), amap(if there) */
786 
787 	/*
788 	 * map in the backpages and frontpages we found in the amap in hopes
789 	 * of preventing future faults.    we also init the pages[] array as
790 	 * we go.
791 	 */
792 
793 	currva = startva;
794 	shadowed = FALSE;
795 	for (lcv = 0 ; lcv < npages ; lcv++, currva += PAGE_SIZE) {
796 
797 		/*
798 		 * dont play with VAs that are already mapped
799 		 * except for center)
800 		 */
801 		if (lcv != centeridx &&
802 		    pmap_extract(ufi.orig_map->pmap, currva, NULL)) {
803 			pages[lcv] = PGO_DONTCARE;
804 			continue;
805 		}
806 
807 		/*
808 		 * unmapped or center page.   check if any anon at this level.
809 		 */
810 		if (amap == NULL || anons[lcv] == NULL) {
811 			pages[lcv] = NULL;
812 			continue;
813 		}
814 
815 		/*
816 		 * check for present page and map if possible.   re-activate it.
817 		 */
818 
819 		pages[lcv] = PGO_DONTCARE;
820 		if (lcv == centeridx) {		/* save center for later! */
821 			shadowed = TRUE;
822 			continue;
823 		}
824 		anon = anons[lcv];
825 		simple_lock(&anon->an_lock);
826 		/* ignore loaned pages */
827 		if (anon->an_page && anon->an_page->loan_count == 0 &&
828 		    (anon->an_page->flags & PG_BUSY) == 0) {
829 			uvm_lock_pageq();
830 			uvm_pageactivate(anon->an_page);
831 			uvm_unlock_pageq();
832 			UVMHIST_LOG(maphist,
833 			    "  MAPPING: n anon: pm=0x%x, va=0x%x, pg=0x%x",
834 			    ufi.orig_map->pmap, currva, anon->an_page, 0);
835 			uvmexp.fltnamap++;
836 
837 			/*
838 			 * Since this isn't the page that's actually faulting,
839 			 * ignore pmap_enter() failures; it's not critical
840 			 * that we enter these right now.
841 			 */
842 
843 			(void) pmap_enter(ufi.orig_map->pmap, currva,
844 			    VM_PAGE_TO_PHYS(anon->an_page),
845 			    (anon->an_ref > 1) ? (enter_prot & ~VM_PROT_WRITE) :
846 			    enter_prot,
847 			    PMAP_CANFAIL |
848 			     (VM_MAPENT_ISWIRED(ufi.entry) ? PMAP_WIRED : 0));
849 		}
850 		simple_unlock(&anon->an_lock);
851 		pmap_update(ufi.orig_map->pmap);
852 	}
853 
854 	/* locked: maps(read), amap(if there) */
855 	/* (shadowed == TRUE) if there is an anon at the faulting address */
856 	UVMHIST_LOG(maphist, "  shadowed=%d, will_get=%d", shadowed,
857 	    (uobj && shadowed == FALSE),0,0);
858 
859 	/*
860 	 * note that if we are really short of RAM we could sleep in the above
861 	 * call to pmap_enter with everything locked.   bad?
862 	 *
863 	 * XXX Actually, that is bad; pmap_enter() should just fail in that
864 	 * XXX case.  --thorpej
865 	 */
866 
867 	/*
868 	 * if the desired page is not shadowed by the amap and we have a
869 	 * backing object, then we check to see if the backing object would
870 	 * prefer to handle the fault itself (rather than letting us do it
871 	 * with the usual pgo_get hook).  the backing object signals this by
872 	 * providing a pgo_fault routine.
873 	 */
874 
875 	if (uobj && shadowed == FALSE && uobj->pgops->pgo_fault != NULL) {
876 		simple_lock(&uobj->vmobjlock);
877 
878 		/* locked: maps(read), amap (if there), uobj */
879 		error = uobj->pgops->pgo_fault(&ufi, startva, pages, npages,
880 		    centeridx, fault_type, access_type, PGO_LOCKED|PGO_SYNCIO);
881 
882 		/* locked: nothing, pgo_fault has unlocked everything */
883 
884 		if (error == ERESTART)
885 			goto ReFault;		/* try again! */
886 		/*
887 		 * object fault routine responsible for pmap_update().
888 		 */
889 		return error;
890 	}
891 
892 	/*
893 	 * now, if the desired page is not shadowed by the amap and we have
894 	 * a backing object that does not have a special fault routine, then
895 	 * we ask (with pgo_get) the object for resident pages that we care
896 	 * about and attempt to map them in.  we do not let pgo_get block
897 	 * (PGO_LOCKED).
898 	 */
899 
900 	if (uobj && shadowed == FALSE) {
901 		simple_lock(&uobj->vmobjlock);
902 
903 		/* locked (!shadowed): maps(read), amap (if there), uobj */
904 		/*
905 		 * the following call to pgo_get does _not_ change locking state
906 		 */
907 
908 		uvmexp.fltlget++;
909 		gotpages = npages;
910 		(void) uobj->pgops->pgo_get(uobj, ufi.entry->offset +
911 				(startva - ufi.entry->start),
912 				pages, &gotpages, centeridx,
913 				access_type & MASK(ufi.entry),
914 				ufi.entry->advice, PGO_LOCKED);
915 
916 		/*
917 		 * check for pages to map, if we got any
918 		 */
919 
920 		uobjpage = NULL;
921 
922 		if (gotpages) {
923 			currva = startva;
924 			for (lcv = 0; lcv < npages;
925 			     lcv++, currva += PAGE_SIZE) {
926 				struct vm_page *curpg;
927 				boolean_t readonly;
928 
929 				curpg = pages[lcv];
930 				if (curpg == NULL || curpg == PGO_DONTCARE) {
931 					continue;
932 				}
933 
934 				/*
935 				 * if center page is resident and not
936 				 * PG_BUSY|PG_RELEASED then pgo_get
937 				 * made it PG_BUSY for us and gave
938 				 * us a handle to it.   remember this
939 				 * page as "uobjpage." (for later use).
940 				 */
941 
942 				if (lcv == centeridx) {
943 					uobjpage = curpg;
944 					UVMHIST_LOG(maphist, "  got uobjpage "
945 					    "(0x%x) with locked get",
946 					    uobjpage, 0,0,0);
947 					continue;
948 				}
949 
950 				/*
951 				 * calling pgo_get with PGO_LOCKED returns us
952 				 * pages which are neither busy nor released,
953 				 * so we don't need to check for this.
954 				 * we can just directly enter the pages.
955 				 */
956 
957 				uvm_lock_pageq();
958 				uvm_pageactivate(curpg);
959 				uvm_unlock_pageq();
960 				UVMHIST_LOG(maphist,
961 				  "  MAPPING: n obj: pm=0x%x, va=0x%x, pg=0x%x",
962 				  ufi.orig_map->pmap, currva, curpg, 0);
963 				uvmexp.fltnomap++;
964 
965 				/*
966 				 * Since this page isn't the page that's
967 				 * actually faulting, ignore pmap_enter()
968 				 * failures; it's not critical that we
969 				 * enter these right now.
970 				 */
971 				KASSERT((curpg->flags & PG_PAGEOUT) == 0);
972 				KASSERT((curpg->flags & PG_RELEASED) == 0);
973 				KASSERT(!UVM_OBJ_IS_CLEAN(curpg->uobject) ||
974 				    (curpg->flags & PG_CLEAN) != 0);
975 				readonly = (curpg->flags & PG_RDONLY)
976 				    || (curpg->loan_count > 0)
977 				    || UVM_OBJ_NEEDS_WRITEFAULT(curpg->uobject);
978 
979 				(void) pmap_enter(ufi.orig_map->pmap, currva,
980 				    VM_PAGE_TO_PHYS(curpg),
981 				    readonly ?
982 				    enter_prot & ~VM_PROT_WRITE :
983 				    enter_prot & MASK(ufi.entry),
984 				    PMAP_CANFAIL |
985 				     (wired ? PMAP_WIRED : 0));
986 
987 				/*
988 				 * NOTE: page can't be PG_WANTED or PG_RELEASED
989 				 * because we've held the lock the whole time
990 				 * we've had the handle.
991 				 */
992 
993 				curpg->flags &= ~(PG_BUSY);
994 				UVM_PAGE_OWN(curpg, NULL);
995 			}
996 			pmap_update(ufi.orig_map->pmap);
997 		}
998 	} else {
999 		uobjpage = NULL;
1000 	}
1001 
1002 	/* locked (shadowed): maps(read), amap */
1003 	/* locked (!shadowed): maps(read), amap(if there),
1004 		 uobj(if !null), uobjpage(if !null) */
1005 
1006 	/*
1007 	 * note that at this point we are done with any front or back pages.
1008 	 * we are now going to focus on the center page (i.e. the one we've
1009 	 * faulted on).  if we have faulted on the top (anon) layer
1010 	 * [i.e. case 1], then the anon we want is anons[centeridx] (we have
1011 	 * not touched it yet).  if we have faulted on the bottom (uobj)
1012 	 * layer [i.e. case 2] and the page was both present and available,
1013 	 * then we've got a pointer to it as "uobjpage" and we've already
1014 	 * made it BUSY.
1015 	 */
1016 
1017 	/*
1018 	 * there are four possible cases we must address: 1A, 1B, 2A, and 2B
1019 	 */
1020 
1021 	/*
1022 	 * redirect case 2: if we are not shadowed, go to case 2.
1023 	 */
1024 
1025 	if (shadowed == FALSE)
1026 		goto Case2;
1027 
1028 	/* locked: maps(read), amap */
1029 
1030 	/*
1031 	 * handle case 1: fault on an anon in our amap
1032 	 */
1033 
1034 	anon = anons[centeridx];
1035 	UVMHIST_LOG(maphist, "  case 1 fault: anon=0x%x", anon, 0,0,0);
1036 	simple_lock(&anon->an_lock);
1037 
1038 	/* locked: maps(read), amap, anon */
1039 
1040 	/*
1041 	 * no matter if we have case 1A or case 1B we are going to need to
1042 	 * have the anon's memory resident.   ensure that now.
1043 	 */
1044 
1045 	/*
1046 	 * let uvmfault_anonget do the dirty work.
1047 	 * if it fails (!OK) it will unlock everything for us.
1048 	 * if it succeeds, locks are still valid and locked.
1049 	 * also, if it is OK, then the anon's page is on the queues.
1050 	 * if the page is on loan from a uvm_object, then anonget will
1051 	 * lock that object for us if it does not fail.
1052 	 */
1053 
1054 	error = uvmfault_anonget(&ufi, amap, anon);
1055 	switch (error) {
1056 	case 0:
1057 		break;
1058 
1059 	case ERESTART:
1060 		goto ReFault;
1061 
1062 	case EAGAIN:
1063 		tsleep(&lbolt, PVM, "fltagain1", 0);
1064 		goto ReFault;
1065 
1066 	default:
1067 		return error;
1068 	}
1069 
1070 	/*
1071 	 * uobj is non null if the page is on loan from an object (i.e. uobj)
1072 	 */
1073 
1074 	uobj = anon->an_page->uobject;	/* locked by anonget if !NULL */
1075 
1076 	/* locked: maps(read), amap, anon, uobj(if one) */
1077 
1078 	/*
1079 	 * special handling for loaned pages
1080 	 */
1081 
1082 	if (anon->an_page->loan_count) {
1083 
1084 		if (!cow_now) {
1085 
1086 			/*
1087 			 * for read faults on loaned pages we just cap the
1088 			 * protection at read-only.
1089 			 */
1090 
1091 			enter_prot = enter_prot & ~VM_PROT_WRITE;
1092 
1093 		} else {
1094 			/*
1095 			 * note that we can't allow writes into a loaned page!
1096 			 *
1097 			 * if we have a write fault on a loaned page in an
1098 			 * anon then we need to look at the anon's ref count.
1099 			 * if it is greater than one then we are going to do
1100 			 * a normal copy-on-write fault into a new anon (this
1101 			 * is not a problem).  however, if the reference count
1102 			 * is one (a case where we would normally allow a
1103 			 * write directly to the page) then we need to kill
1104 			 * the loan before we continue.
1105 			 */
1106 
1107 			/* >1 case is already ok */
1108 			if (anon->an_ref == 1) {
1109 
1110 				/* get new un-owned replacement page */
1111 				pg = uvm_pagealloc(NULL, 0, NULL, 0);
1112 				if (pg == NULL) {
1113 					uvmfault_unlockall(&ufi, amap, uobj,
1114 					    anon);
1115 					uvm_wait("flt_noram2");
1116 					goto ReFault;
1117 				}
1118 
1119 				/*
1120 				 * copy data, kill loan, and drop uobj lock
1121 				 * (if any)
1122 				 */
1123 				/* copy old -> new */
1124 				uvm_pagecopy(anon->an_page, pg);
1125 
1126 				/* force reload */
1127 				pmap_page_protect(anon->an_page,
1128 						  VM_PROT_NONE);
1129 				uvm_lock_pageq();	  /* KILL loan */
1130 
1131 				anon->an_page->uanon = NULL;
1132 				/* in case we owned */
1133 				anon->an_page->pqflags &= ~PQ_ANON;
1134 
1135 				if (uobj) {
1136 					/* if we were receiver of loan */
1137 					anon->an_page->loan_count--;
1138 				} else {
1139 					/*
1140 					 * we were the lender (A->K); need
1141 					 * to remove the page from pageq's.
1142 					 */
1143 					uvm_pagedequeue(anon->an_page);
1144 				}
1145 
1146 				uvm_pageactivate(pg);
1147 				uvm_unlock_pageq();
1148 				if (uobj) {
1149 					simple_unlock(&uobj->vmobjlock);
1150 					uobj = NULL;
1151 				}
1152 
1153 				/* install new page in anon */
1154 				anon->an_page = pg;
1155 				pg->uanon = anon;
1156 				pg->pqflags |= PQ_ANON;
1157 				pg->flags &= ~(PG_BUSY|PG_FAKE);
1158 				UVM_PAGE_OWN(pg, NULL);
1159 
1160 				/* done! */
1161 			}     /* ref == 1 */
1162 		}       /* write fault */
1163 	}         /* loan count */
1164 
1165 	/*
1166 	 * if we are case 1B then we will need to allocate a new blank
1167 	 * anon to transfer the data into.   note that we have a lock
1168 	 * on anon, so no one can busy or release the page until we are done.
1169 	 * also note that the ref count can't drop to zero here because
1170 	 * it is > 1 and we are only dropping one ref.
1171 	 *
1172 	 * in the (hopefully very rare) case that we are out of RAM we
1173 	 * will unlock, wait for more RAM, and refault.
1174 	 *
1175 	 * if we are out of anon VM we kill the process (XXX: could wait?).
1176 	 */
1177 
1178 	if (cow_now && anon->an_ref > 1) {
1179 
1180 		UVMHIST_LOG(maphist, "  case 1B: COW fault",0,0,0,0);
1181 		uvmexp.flt_acow++;
1182 		oanon = anon;		/* oanon = old, locked anon */
1183 		anon = uvm_analloc();
1184 		if (anon) {
1185 			/* new anon is locked! */
1186 			pg = uvm_pagealloc(NULL, 0, anon, 0);
1187 		}
1188 
1189 		/* check for out of RAM */
1190 		if (anon == NULL || pg == NULL) {
1191 			if (anon) {
1192 				anon->an_ref--;
1193 				simple_unlock(&anon->an_lock);
1194 				uvm_anfree(anon);
1195 			}
1196 			uvmfault_unlockall(&ufi, amap, uobj, oanon);
1197 			if (!uvm_reclaimable()) {
1198 				UVMHIST_LOG(maphist,
1199 				    "<- failed.  out of VM",0,0,0,0);
1200 				uvmexp.fltnoanon++;
1201 				return ENOMEM;
1202 			}
1203 
1204 			uvmexp.fltnoram++;
1205 			uvm_wait("flt_noram3");	/* out of RAM, wait for more */
1206 			goto ReFault;
1207 		}
1208 
1209 		/* got all resources, replace anon with nanon */
1210 		uvm_pagecopy(oanon->an_page, pg);
1211 		uvm_lock_pageq();
1212 		uvm_pageactivate(pg);
1213 		pg->flags &= ~(PG_BUSY|PG_FAKE);
1214 		uvm_unlock_pageq();
1215 		UVM_PAGE_OWN(pg, NULL);
1216 		amap_add(&ufi.entry->aref, ufi.orig_rvaddr - ufi.entry->start,
1217 		    anon, TRUE);
1218 
1219 		/* deref: can not drop to zero here by defn! */
1220 		oanon->an_ref--;
1221 
1222 		/*
1223 		 * note: oanon is still locked, as is the new anon.  we
1224 		 * need to check for this later when we unlock oanon; if
1225 		 * oanon != anon, we'll have to unlock anon, too.
1226 		 */
1227 
1228 	} else {
1229 
1230 		uvmexp.flt_anon++;
1231 		oanon = anon;		/* old, locked anon is same as anon */
1232 		pg = anon->an_page;
1233 		if (anon->an_ref > 1)     /* disallow writes to ref > 1 anons */
1234 			enter_prot = enter_prot & ~VM_PROT_WRITE;
1235 
1236 	}
1237 
1238 	/* locked: maps(read), amap, oanon, anon (if different from oanon) */
1239 
1240 	/*
1241 	 * now map the page in.
1242 	 */
1243 
1244 	UVMHIST_LOG(maphist, "  MAPPING: anon: pm=0x%x, va=0x%x, pg=0x%x",
1245 	    ufi.orig_map->pmap, ufi.orig_rvaddr, pg, 0);
1246 	if (pmap_enter(ufi.orig_map->pmap, ufi.orig_rvaddr, VM_PAGE_TO_PHYS(pg),
1247 	    enter_prot, access_type | PMAP_CANFAIL | (wired ? PMAP_WIRED : 0))
1248 	    != 0) {
1249 
1250 		/*
1251 		 * No need to undo what we did; we can simply think of
1252 		 * this as the pmap throwing away the mapping information.
1253 		 *
1254 		 * We do, however, have to go through the ReFault path,
1255 		 * as the map may change while we're asleep.
1256 		 */
1257 
1258 		if (anon != oanon)
1259 			simple_unlock(&anon->an_lock);
1260 		uvmfault_unlockall(&ufi, amap, uobj, oanon);
1261 		if (!uvm_reclaimable()) {
1262 			UVMHIST_LOG(maphist,
1263 			    "<- failed.  out of VM",0,0,0,0);
1264 			/* XXX instrumentation */
1265 			return ENOMEM;
1266 		}
1267 		/* XXX instrumentation */
1268 		uvm_wait("flt_pmfail1");
1269 		goto ReFault;
1270 	}
1271 
1272 	/*
1273 	 * ... update the page queues.
1274 	 */
1275 
1276 	uvm_lock_pageq();
1277 	if (wire_fault) {
1278 		uvm_pagewire(pg);
1279 
1280 		/*
1281 		 * since the now-wired page cannot be paged out,
1282 		 * release its swap resources for others to use.
1283 		 * since an anon with no swap cannot be PG_CLEAN,
1284 		 * clear its clean flag now.
1285 		 */
1286 
1287 		pg->flags &= ~(PG_CLEAN);
1288 		uvm_anon_dropswap(anon);
1289 	} else {
1290 		uvm_pageactivate(pg);
1291 	}
1292 	uvm_unlock_pageq();
1293 
1294 	/*
1295 	 * done case 1!  finish up by unlocking everything and returning success
1296 	 */
1297 
1298 	if (anon != oanon)
1299 		simple_unlock(&anon->an_lock);
1300 	uvmfault_unlockall(&ufi, amap, uobj, oanon);
1301 	pmap_update(ufi.orig_map->pmap);
1302 	return 0;
1303 
1304 Case2:
1305 	/*
1306 	 * handle case 2: faulting on backing object or zero fill
1307 	 */
1308 
1309 	/*
1310 	 * locked:
1311 	 * maps(read), amap(if there), uobj(if !null), uobjpage(if !null)
1312 	 */
1313 
1314 	/*
1315 	 * note that uobjpage can not be PGO_DONTCARE at this point.  we now
1316 	 * set uobjpage to PGO_DONTCARE if we are doing a zero fill.  if we
1317 	 * have a backing object, check and see if we are going to promote
1318 	 * the data up to an anon during the fault.
1319 	 */
1320 
1321 	if (uobj == NULL) {
1322 		uobjpage = PGO_DONTCARE;
1323 		promote = TRUE;		/* always need anon here */
1324 	} else {
1325 		KASSERT(uobjpage != PGO_DONTCARE);
1326 		promote = cow_now && UVM_ET_ISCOPYONWRITE(ufi.entry);
1327 	}
1328 	UVMHIST_LOG(maphist, "  case 2 fault: promote=%d, zfill=%d",
1329 	    promote, (uobj == NULL), 0,0);
1330 
1331 	/*
1332 	 * if uobjpage is not null then we do not need to do I/O to get the
1333 	 * uobjpage.
1334 	 *
1335 	 * if uobjpage is null, then we need to unlock and ask the pager to
1336 	 * get the data for us.   once we have the data, we need to reverify
1337 	 * the state the world.   we are currently not holding any resources.
1338 	 */
1339 
1340 	if (uobjpage) {
1341 		/* update rusage counters */
1342 		curproc->p_stats->p_ru.ru_minflt++;
1343 	} else {
1344 		/* update rusage counters */
1345 		curproc->p_stats->p_ru.ru_majflt++;
1346 
1347 		/* locked: maps(read), amap(if there), uobj */
1348 		uvmfault_unlockall(&ufi, amap, NULL, NULL);
1349 		/* locked: uobj */
1350 
1351 		uvmexp.fltget++;
1352 		gotpages = 1;
1353 		uoff = (ufi.orig_rvaddr - ufi.entry->start) + ufi.entry->offset;
1354 		error = uobj->pgops->pgo_get(uobj, uoff, &uobjpage, &gotpages,
1355 		    0, access_type & MASK(ufi.entry), ufi.entry->advice,
1356 		    PGO_SYNCIO);
1357 		/* locked: uobjpage(if no error) */
1358 
1359 		/*
1360 		 * recover from I/O
1361 		 */
1362 
1363 		if (error) {
1364 			if (error == EAGAIN) {
1365 				UVMHIST_LOG(maphist,
1366 				    "  pgo_get says TRY AGAIN!",0,0,0,0);
1367 				tsleep(&lbolt, PVM, "fltagain2", 0);
1368 				goto ReFault;
1369 			}
1370 
1371 			UVMHIST_LOG(maphist, "<- pgo_get failed (code %d)",
1372 			    error, 0,0,0);
1373 			return error;
1374 		}
1375 
1376 		/* locked: uobjpage */
1377 
1378 		uvm_lock_pageq();
1379 		uvm_pageactivate(uobjpage);
1380 		uvm_unlock_pageq();
1381 
1382 		/*
1383 		 * re-verify the state of the world by first trying to relock
1384 		 * the maps.  always relock the object.
1385 		 */
1386 
1387 		locked = uvmfault_relock(&ufi);
1388 		if (locked && amap)
1389 			amap_lock(amap);
1390 		simple_lock(&uobj->vmobjlock);
1391 
1392 		/* locked(locked): maps(read), amap(if !null), uobj, uobjpage */
1393 		/* locked(!locked): uobj, uobjpage */
1394 
1395 		/*
1396 		 * verify that the page has not be released and re-verify
1397 		 * that amap slot is still free.   if there is a problem,
1398 		 * we unlock and clean up.
1399 		 */
1400 
1401 		if ((uobjpage->flags & PG_RELEASED) != 0 ||
1402 		    (locked && amap &&
1403 		    amap_lookup(&ufi.entry->aref,
1404 		      ufi.orig_rvaddr - ufi.entry->start))) {
1405 			if (locked)
1406 				uvmfault_unlockall(&ufi, amap, NULL, NULL);
1407 			locked = FALSE;
1408 		}
1409 
1410 		/*
1411 		 * didn't get the lock?   release the page and retry.
1412 		 */
1413 
1414 		if (locked == FALSE) {
1415 			UVMHIST_LOG(maphist,
1416 			    "  wasn't able to relock after fault: retry",
1417 			    0,0,0,0);
1418 			if (uobjpage->flags & PG_WANTED)
1419 				wakeup(uobjpage);
1420 			if (uobjpage->flags & PG_RELEASED) {
1421 				uvmexp.fltpgrele++;
1422 				uvm_pagefree(uobjpage);
1423 				goto ReFault;
1424 			}
1425 			uobjpage->flags &= ~(PG_BUSY|PG_WANTED);
1426 			UVM_PAGE_OWN(uobjpage, NULL);
1427 			simple_unlock(&uobj->vmobjlock);
1428 			goto ReFault;
1429 		}
1430 
1431 		/*
1432 		 * we have the data in uobjpage which is busy and
1433 		 * not released.  we are holding object lock (so the page
1434 		 * can't be released on us).
1435 		 */
1436 
1437 		/* locked: maps(read), amap(if !null), uobj, uobjpage */
1438 	}
1439 
1440 	/*
1441 	 * locked:
1442 	 * maps(read), amap(if !null), uobj(if !null), uobjpage(if uobj)
1443 	 */
1444 
1445 	/*
1446 	 * notes:
1447 	 *  - at this point uobjpage can not be NULL
1448 	 *  - at this point uobjpage can not be PG_RELEASED (since we checked
1449 	 *  for it above)
1450 	 *  - at this point uobjpage could be PG_WANTED (handle later)
1451 	 */
1452 
1453 	KASSERT(uobj == NULL || !UVM_OBJ_IS_CLEAN(uobjpage->uobject) ||
1454 	    (uobjpage->flags & PG_CLEAN) != 0);
1455 	if (promote == FALSE) {
1456 
1457 		/*
1458 		 * we are not promoting.   if the mapping is COW ensure that we
1459 		 * don't give more access than we should (e.g. when doing a read
1460 		 * fault on a COPYONWRITE mapping we want to map the COW page in
1461 		 * R/O even though the entry protection could be R/W).
1462 		 *
1463 		 * set "pg" to the page we want to map in (uobjpage, usually)
1464 		 */
1465 
1466 		/* no anon in this case. */
1467 		anon = NULL;
1468 
1469 		uvmexp.flt_obj++;
1470 		if (UVM_ET_ISCOPYONWRITE(ufi.entry) ||
1471 		    UVM_OBJ_NEEDS_WRITEFAULT(uobjpage->uobject))
1472 			enter_prot &= ~VM_PROT_WRITE;
1473 		pg = uobjpage;		/* map in the actual object */
1474 
1475 		/* assert(uobjpage != PGO_DONTCARE) */
1476 
1477 		/*
1478 		 * we are faulting directly on the page.   be careful
1479 		 * about writing to loaned pages...
1480 		 */
1481 
1482 		if (uobjpage->loan_count) {
1483 			if (!cow_now) {
1484 				/* read fault: cap the protection at readonly */
1485 				/* cap! */
1486 				enter_prot = enter_prot & ~VM_PROT_WRITE;
1487 			} else {
1488 				/* write fault: must break the loan here */
1489 
1490 				pg = uvm_loanbreak(uobjpage);
1491 				if (pg == NULL) {
1492 
1493 					/*
1494 					 * drop ownership of page, it can't
1495 					 * be released
1496 					 */
1497 
1498 					if (uobjpage->flags & PG_WANTED)
1499 						wakeup(uobjpage);
1500 					uobjpage->flags &= ~(PG_BUSY|PG_WANTED);
1501 					UVM_PAGE_OWN(uobjpage, NULL);
1502 
1503 					uvmfault_unlockall(&ufi, amap, uobj,
1504 					  NULL);
1505 					UVMHIST_LOG(maphist,
1506 					  "  out of RAM breaking loan, waiting",
1507 					  0,0,0,0);
1508 					uvmexp.fltnoram++;
1509 					uvm_wait("flt_noram4");
1510 					goto ReFault;
1511 				}
1512 				uobjpage = pg;
1513 			}
1514 		}
1515 	} else {
1516 
1517 		/*
1518 		 * if we are going to promote the data to an anon we
1519 		 * allocate a blank anon here and plug it into our amap.
1520 		 */
1521 #if DIAGNOSTIC
1522 		if (amap == NULL)
1523 			panic("uvm_fault: want to promote data, but no anon");
1524 #endif
1525 
1526 		anon = uvm_analloc();
1527 		if (anon) {
1528 
1529 			/*
1530 			 * The new anon is locked.
1531 			 *
1532 			 * In `Fill in data...' below, if
1533 			 * uobjpage == PGO_DONTCARE, we want
1534 			 * a zero'd, dirty page, so have
1535 			 * uvm_pagealloc() do that for us.
1536 			 */
1537 
1538 			pg = uvm_pagealloc(NULL, 0, anon,
1539 			    (uobjpage == PGO_DONTCARE) ? UVM_PGA_ZERO : 0);
1540 		}
1541 
1542 		/*
1543 		 * out of memory resources?
1544 		 */
1545 
1546 		if (anon == NULL || pg == NULL) {
1547 			if (anon != NULL) {
1548 				anon->an_ref--;
1549 				simple_unlock(&anon->an_lock);
1550 				uvm_anfree(anon);
1551 			}
1552 
1553 			/*
1554 			 * arg!  must unbusy our page and fail or sleep.
1555 			 */
1556 
1557 			if (uobjpage != PGO_DONTCARE) {
1558 				if (uobjpage->flags & PG_WANTED)
1559 					/* still holding object lock */
1560 					wakeup(uobjpage);
1561 
1562 				uobjpage->flags &= ~(PG_BUSY|PG_WANTED);
1563 				UVM_PAGE_OWN(uobjpage, NULL);
1564 			}
1565 
1566 			/* unlock and fail ... */
1567 			uvmfault_unlockall(&ufi, amap, uobj, NULL);
1568 			if (!uvm_reclaimable()) {
1569 				UVMHIST_LOG(maphist, "  promote: out of VM",
1570 				    0,0,0,0);
1571 				uvmexp.fltnoanon++;
1572 				return ENOMEM;
1573 			}
1574 
1575 			UVMHIST_LOG(maphist, "  out of RAM, waiting for more",
1576 			    0,0,0,0);
1577 			uvmexp.fltnoram++;
1578 			uvm_wait("flt_noram5");
1579 			goto ReFault;
1580 		}
1581 
1582 		/*
1583 		 * fill in the data
1584 		 */
1585 
1586 		if (uobjpage != PGO_DONTCARE) {
1587 			uvmexp.flt_prcopy++;
1588 			/* copy page [pg now dirty] */
1589 			uvm_pagecopy(uobjpage, pg);
1590 
1591 			/*
1592 			 * promote to shared amap?  make sure all sharing
1593 			 * procs see it
1594 			 */
1595 
1596 			if ((amap_flags(amap) & AMAP_SHARED) != 0) {
1597 				pmap_page_protect(uobjpage, VM_PROT_NONE);
1598 				/*
1599 				 * XXX: PAGE MIGHT BE WIRED!
1600 				 */
1601 			}
1602 
1603 			/*
1604 			 * dispose of uobjpage.  it can't be PG_RELEASED
1605 			 * since we still hold the object lock.
1606 			 * drop handle to uobj as well.
1607 			 */
1608 
1609 			if (uobjpage->flags & PG_WANTED)
1610 				/* still have the obj lock */
1611 				wakeup(uobjpage);
1612 			uobjpage->flags &= ~(PG_BUSY|PG_WANTED);
1613 			UVM_PAGE_OWN(uobjpage, NULL);
1614 			simple_unlock(&uobj->vmobjlock);
1615 			uobj = NULL;
1616 
1617 			UVMHIST_LOG(maphist,
1618 			    "  promote uobjpage 0x%x to anon/page 0x%x/0x%x",
1619 			    uobjpage, anon, pg, 0);
1620 
1621 		} else {
1622 			uvmexp.flt_przero++;
1623 
1624 			/*
1625 			 * Page is zero'd and marked dirty by uvm_pagealloc()
1626 			 * above.
1627 			 */
1628 
1629 			UVMHIST_LOG(maphist,"  zero fill anon/page 0x%x/0%x",
1630 			    anon, pg, 0, 0);
1631 		}
1632 		amap_add(&ufi.entry->aref, ufi.orig_rvaddr - ufi.entry->start,
1633 		    anon, FALSE);
1634 	}
1635 
1636 	/*
1637 	 * locked:
1638 	 * maps(read), amap(if !null), uobj(if !null), uobjpage(if uobj),
1639 	 *   anon(if !null), pg(if anon)
1640 	 *
1641 	 * note: pg is either the uobjpage or the new page in the new anon
1642 	 */
1643 
1644 	/*
1645 	 * all resources are present.   we can now map it in and free our
1646 	 * resources.
1647 	 */
1648 
1649 	UVMHIST_LOG(maphist,
1650 	    "  MAPPING: case2: pm=0x%x, va=0x%x, pg=0x%x, promote=%d",
1651 	    ufi.orig_map->pmap, ufi.orig_rvaddr, pg, promote);
1652 	KASSERT((access_type & VM_PROT_WRITE) == 0 ||
1653 		(pg->flags & PG_RDONLY) == 0);
1654 	if (pmap_enter(ufi.orig_map->pmap, ufi.orig_rvaddr, VM_PAGE_TO_PHYS(pg),
1655 	    pg->flags & PG_RDONLY ? enter_prot & ~VM_PROT_WRITE : enter_prot,
1656 	    access_type | PMAP_CANFAIL | (wired ? PMAP_WIRED : 0)) != 0) {
1657 
1658 		/*
1659 		 * No need to undo what we did; we can simply think of
1660 		 * this as the pmap throwing away the mapping information.
1661 		 *
1662 		 * We do, however, have to go through the ReFault path,
1663 		 * as the map may change while we're asleep.
1664 		 */
1665 
1666 		if (pg->flags & PG_WANTED)
1667 			wakeup(pg);
1668 
1669 		/*
1670 		 * note that pg can't be PG_RELEASED since we did not drop
1671 		 * the object lock since the last time we checked.
1672 		 */
1673 
1674 		pg->flags &= ~(PG_BUSY|PG_FAKE|PG_WANTED);
1675 		UVM_PAGE_OWN(pg, NULL);
1676 		uvmfault_unlockall(&ufi, amap, uobj, anon);
1677 		if (!uvm_reclaimable()) {
1678 			UVMHIST_LOG(maphist,
1679 			    "<- failed.  out of VM",0,0,0,0);
1680 			/* XXX instrumentation */
1681 			return ENOMEM;
1682 		}
1683 		/* XXX instrumentation */
1684 		uvm_wait("flt_pmfail2");
1685 		goto ReFault;
1686 	}
1687 
1688 	uvm_lock_pageq();
1689 	if (wire_fault) {
1690 		uvm_pagewire(pg);
1691 		if (pg->pqflags & PQ_AOBJ) {
1692 
1693 			/*
1694 			 * since the now-wired page cannot be paged out,
1695 			 * release its swap resources for others to use.
1696 			 * since an aobj page with no swap cannot be PG_CLEAN,
1697 			 * clear its clean flag now.
1698 			 */
1699 
1700 			pg->flags &= ~(PG_CLEAN);
1701 			uao_dropswap(uobj, pg->offset >> PAGE_SHIFT);
1702 		}
1703 	} else {
1704 		uvm_pageactivate(pg);
1705 	}
1706 	uvm_unlock_pageq();
1707 	if (pg->flags & PG_WANTED)
1708 		wakeup(pg);
1709 
1710 	/*
1711 	 * note that pg can't be PG_RELEASED since we did not drop the object
1712 	 * lock since the last time we checked.
1713 	 */
1714 
1715 	pg->flags &= ~(PG_BUSY|PG_FAKE|PG_WANTED);
1716 	UVM_PAGE_OWN(pg, NULL);
1717 	uvmfault_unlockall(&ufi, amap, uobj, anon);
1718 	pmap_update(ufi.orig_map->pmap);
1719 	UVMHIST_LOG(maphist, "<- done (SUCCESS!)",0,0,0,0);
1720 	return 0;
1721 }
1722 
1723 /*
1724  * uvm_fault_wire: wire down a range of virtual addresses in a map.
1725  *
1726  * => map may be read-locked by caller, but MUST NOT be write-locked.
1727  * => if map is read-locked, any operations which may cause map to
1728  *	be write-locked in uvm_fault() must be taken care of by
1729  *	the caller.  See uvm_map_pageable().
1730  */
1731 
1732 int
1733 uvm_fault_wire(struct vm_map *map, vaddr_t start, vaddr_t end,
1734     vm_fault_t fault_type, vm_prot_t access_type)
1735 {
1736 	vaddr_t va;
1737 	int error;
1738 
1739 	/*
1740 	 * now fault it in a page at a time.   if the fault fails then we have
1741 	 * to undo what we have done.   note that in uvm_fault VM_PROT_NONE
1742 	 * is replaced with the max protection if fault_type is VM_FAULT_WIRE.
1743 	 */
1744 
1745 	/*
1746 	 * XXX work around overflowing a vaddr_t.  this prevents us from
1747 	 * wiring the last page in the address space, though.
1748 	 */
1749 	if (start > end) {
1750 		return EFAULT;
1751 	}
1752 
1753 	for (va = start ; va < end ; va += PAGE_SIZE) {
1754 		error = uvm_fault(map, va, fault_type, access_type);
1755 		if (error) {
1756 			if (va != start) {
1757 				uvm_fault_unwire(map, start, va);
1758 			}
1759 			return error;
1760 		}
1761 	}
1762 	return 0;
1763 }
1764 
1765 /*
1766  * uvm_fault_unwire(): unwire range of virtual space.
1767  */
1768 
1769 void
1770 uvm_fault_unwire(struct vm_map *map, vaddr_t start, vaddr_t end)
1771 {
1772 	vm_map_lock_read(map);
1773 	uvm_fault_unwire_locked(map, start, end);
1774 	vm_map_unlock_read(map);
1775 }
1776 
1777 /*
1778  * uvm_fault_unwire_locked(): the guts of uvm_fault_unwire().
1779  *
1780  * => map must be at least read-locked.
1781  */
1782 
1783 void
1784 uvm_fault_unwire_locked(struct vm_map *map, vaddr_t start, vaddr_t end)
1785 {
1786 	struct vm_map_entry *entry;
1787 	pmap_t pmap = vm_map_pmap(map);
1788 	vaddr_t va;
1789 	paddr_t pa;
1790 	struct vm_page *pg;
1791 
1792 	KASSERT((map->flags & VM_MAP_INTRSAFE) == 0);
1793 
1794 	/*
1795 	 * we assume that the area we are unwiring has actually been wired
1796 	 * in the first place.   this means that we should be able to extract
1797 	 * the PAs from the pmap.   we also lock out the page daemon so that
1798 	 * we can call uvm_pageunwire.
1799 	 */
1800 
1801 	uvm_lock_pageq();
1802 
1803 	/*
1804 	 * find the beginning map entry for the region.
1805 	 */
1806 
1807 	KASSERT(start >= vm_map_min(map) && end <= vm_map_max(map));
1808 	if (uvm_map_lookup_entry(map, start, &entry) == FALSE)
1809 		panic("uvm_fault_unwire_locked: address not in map");
1810 
1811 	for (va = start; va < end; va += PAGE_SIZE) {
1812 		if (pmap_extract(pmap, va, &pa) == FALSE)
1813 			continue;
1814 
1815 		/*
1816 		 * find the map entry for the current address.
1817 		 */
1818 
1819 		KASSERT(va >= entry->start);
1820 		while (va >= entry->end) {
1821 			KASSERT(entry->next != &map->header &&
1822 				entry->next->start <= entry->end);
1823 			entry = entry->next;
1824 		}
1825 
1826 		/*
1827 		 * if the entry is no longer wired, tell the pmap.
1828 		 */
1829 
1830 		if (VM_MAPENT_ISWIRED(entry) == 0)
1831 			pmap_unwire(pmap, va);
1832 
1833 		pg = PHYS_TO_VM_PAGE(pa);
1834 		if (pg)
1835 			uvm_pageunwire(pg);
1836 	}
1837 
1838 	uvm_unlock_pageq();
1839 }
1840