xref: /netbsd-src/sys/uvm/uvm_fault.c (revision 3cec974c61d7fac0a37c0377723a33214a458c8b)
1 /*	$NetBSD: uvm_fault.c,v 1.57 2001/03/10 22:46:49 chs Exp $	*/
2 
3 /*
4  *
5  * Copyright (c) 1997 Charles D. Cranor and Washington University.
6  * All rights reserved.
7  *
8  * Redistribution and use in source and binary forms, with or without
9  * modification, are permitted provided that the following conditions
10  * are met:
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in the
15  *    documentation and/or other materials provided with the distribution.
16  * 3. All advertising materials mentioning features or use of this software
17  *    must display the following acknowledgement:
18  *      This product includes software developed by Charles D. Cranor and
19  *      Washington University.
20  * 4. The name of the author may not be used to endorse or promote products
21  *    derived from this software without specific prior written permission.
22  *
23  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
24  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
25  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
26  * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
27  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
28  * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
29  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
30  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
31  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
32  * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
33  *
34  * from: Id: uvm_fault.c,v 1.1.2.23 1998/02/06 05:29:05 chs Exp
35  */
36 
37 #include "opt_uvmhist.h"
38 
39 /*
40  * uvm_fault.c: fault handler
41  */
42 
43 #include <sys/param.h>
44 #include <sys/systm.h>
45 #include <sys/kernel.h>
46 #include <sys/proc.h>
47 #include <sys/malloc.h>
48 #include <sys/mman.h>
49 #include <sys/user.h>
50 
51 #include <uvm/uvm.h>
52 
53 /*
54  *
55  * a word on page faults:
56  *
57  * types of page faults we handle:
58  *
59  * CASE 1: upper layer faults                   CASE 2: lower layer faults
60  *
61  *    CASE 1A         CASE 1B                  CASE 2A        CASE 2B
62  *    read/write1     write>1                  read/write   +-cow_write/zero
63  *         |             |                         |        |
64  *      +--|--+       +--|--+     +-----+       +  |  +     | +-----+
65  * amap |  V  |       |  ----------->new|          |        | |  ^  |
66  *      +-----+       +-----+     +-----+       +  |  +     | +--|--+
67  *                                                 |        |    |
68  *      +-----+       +-----+                   +--|--+     | +--|--+
69  * uobj | d/c |       | d/c |                   |  V  |     +----|  |
70  *      +-----+       +-----+                   +-----+       +-----+
71  *
72  * d/c = don't care
73  *
74  *   case [0]: layerless fault
75  *	no amap or uobj is present.   this is an error.
76  *
77  *   case [1]: upper layer fault [anon active]
78  *     1A: [read] or [write with anon->an_ref == 1]
79  *		I/O takes place in top level anon and uobj is not touched.
80  *     1B: [write with anon->an_ref > 1]
81  *		new anon is alloc'd and data is copied off ["COW"]
82  *
83  *   case [2]: lower layer fault [uobj]
84  *     2A: [read on non-NULL uobj] or [write to non-copy_on_write area]
85  *		I/O takes place directly in object.
86  *     2B: [write to copy_on_write] or [read on NULL uobj]
87  *		data is "promoted" from uobj to a new anon.
88  *		if uobj is null, then we zero fill.
89  *
90  * we follow the standard UVM locking protocol ordering:
91  *
92  * MAPS => AMAP => UOBJ => ANON => PAGE QUEUES (PQ)
93  * we hold a PG_BUSY page if we unlock for I/O
94  *
95  *
96  * the code is structured as follows:
97  *
98  *     - init the "IN" params in the ufi structure
99  *   ReFault:
100  *     - do lookups [locks maps], check protection, handle needs_copy
101  *     - check for case 0 fault (error)
102  *     - establish "range" of fault
103  *     - if we have an amap lock it and extract the anons
104  *     - if sequential advice deactivate pages behind us
105  *     - at the same time check pmap for unmapped areas and anon for pages
106  *	 that we could map in (and do map it if found)
107  *     - check object for resident pages that we could map in
108  *     - if (case 2) goto Case2
109  *     - >>> handle case 1
110  *           - ensure source anon is resident in RAM
111  *           - if case 1B alloc new anon and copy from source
112  *           - map the correct page in
113  *   Case2:
114  *     - >>> handle case 2
115  *           - ensure source page is resident (if uobj)
116  *           - if case 2B alloc new anon and copy from source (could be zero
117  *		fill if uobj == NULL)
118  *           - map the correct page in
119  *     - done!
120  *
121  * note on paging:
122  *   if we have to do I/O we place a PG_BUSY page in the correct object,
123  * unlock everything, and do the I/O.   when I/O is done we must reverify
124  * the state of the world before assuming that our data structures are
125  * valid.   [because mappings could change while the map is unlocked]
126  *
127  *  alternative 1: unbusy the page in question and restart the page fault
128  *    from the top (ReFault).   this is easy but does not take advantage
129  *    of the information that we already have from our previous lookup,
130  *    although it is possible that the "hints" in the vm_map will help here.
131  *
132  * alternative 2: the system already keeps track of a "version" number of
133  *    a map.   [i.e. every time you write-lock a map (e.g. to change a
134  *    mapping) you bump the version number up by one...]   so, we can save
135  *    the version number of the map before we release the lock and start I/O.
136  *    then when I/O is done we can relock and check the version numbers
137  *    to see if anything changed.    this might save us some over 1 because
138  *    we don't have to unbusy the page and may be less compares(?).
139  *
140  * alternative 3: put in backpointers or a way to "hold" part of a map
141  *    in place while I/O is in progress.   this could be complex to
142  *    implement (especially with structures like amap that can be referenced
143  *    by multiple map entries, and figuring out what should wait could be
144  *    complex as well...).
145  *
146  * given that we are not currently multiprocessor or multithreaded we might
147  * as well choose alternative 2 now.   maybe alternative 3 would be useful
148  * in the future.    XXX keep in mind for future consideration//rechecking.
149  */
150 
151 /*
152  * local data structures
153  */
154 
155 struct uvm_advice {
156 	int advice;
157 	int nback;
158 	int nforw;
159 };
160 
161 /*
162  * page range array:
163  * note: index in array must match "advice" value
164  * XXX: borrowed numbers from freebsd.   do they work well for us?
165  */
166 
167 static struct uvm_advice uvmadvice[] = {
168 	{ MADV_NORMAL, 3, 4 },
169 	{ MADV_RANDOM, 0, 0 },
170 	{ MADV_SEQUENTIAL, 8, 7},
171 };
172 
173 #define UVM_MAXRANGE 16	/* must be max() of nback+nforw+1 */
174 
175 /*
176  * private prototypes
177  */
178 
179 static void uvmfault_amapcopy __P((struct uvm_faultinfo *));
180 static __inline void uvmfault_anonflush __P((struct vm_anon **, int));
181 
182 /*
183  * inline functions
184  */
185 
186 /*
187  * uvmfault_anonflush: try and deactivate pages in specified anons
188  *
189  * => does not have to deactivate page if it is busy
190  */
191 
192 static __inline void
193 uvmfault_anonflush(anons, n)
194 	struct vm_anon **anons;
195 	int n;
196 {
197 	int lcv;
198 	struct vm_page *pg;
199 
200 	for (lcv = 0 ; lcv < n ; lcv++) {
201 		if (anons[lcv] == NULL)
202 			continue;
203 		simple_lock(&anons[lcv]->an_lock);
204 		pg = anons[lcv]->u.an_page;
205 		if (pg && (pg->flags & PG_BUSY) == 0 && pg->loan_count == 0) {
206 			uvm_lock_pageq();
207 			if (pg->wire_count == 0) {
208 				pmap_clear_reference(pg);
209 				uvm_pagedeactivate(pg);
210 			}
211 			uvm_unlock_pageq();
212 		}
213 		simple_unlock(&anons[lcv]->an_lock);
214 	}
215 }
216 
217 /*
218  * normal functions
219  */
220 
221 /*
222  * uvmfault_amapcopy: clear "needs_copy" in a map.
223  *
224  * => called with VM data structures unlocked (usually, see below)
225  * => we get a write lock on the maps and clear needs_copy for a VA
226  * => if we are out of RAM we sleep (waiting for more)
227  */
228 
229 static void
230 uvmfault_amapcopy(ufi)
231 	struct uvm_faultinfo *ufi;
232 {
233 
234 	/*
235 	 * while we haven't done the job
236 	 */
237 
238 	while (1) {
239 
240 		/*
241 		 * no mapping?  give up.
242 		 */
243 
244 		if (uvmfault_lookup(ufi, TRUE) == FALSE)
245 			return;
246 
247 		/*
248 		 * copy if needed.
249 		 */
250 
251 		if (UVM_ET_ISNEEDSCOPY(ufi->entry))
252 			amap_copy(ufi->map, ufi->entry, M_NOWAIT, TRUE,
253 				ufi->orig_rvaddr, ufi->orig_rvaddr + 1);
254 
255 		/*
256 		 * didn't work?  must be out of RAM.   unlock and sleep.
257 		 */
258 
259 		if (UVM_ET_ISNEEDSCOPY(ufi->entry)) {
260 			uvmfault_unlockmaps(ufi, TRUE);
261 			uvm_wait("fltamapcopy");
262 			continue;
263 		}
264 
265 		/*
266 		 * got it!   unlock and return.
267 		 */
268 
269 		uvmfault_unlockmaps(ufi, TRUE);
270 		return;
271 	}
272 	/*NOTREACHED*/
273 }
274 
275 /*
276  * uvmfault_anonget: get data in an anon into a non-busy, non-released
277  * page in that anon.
278  *
279  * => maps, amap, and anon locked by caller.
280  * => if we fail (result != 0) we unlock everything.
281  * => if we are successful, we return with everything still locked.
282  * => we don't move the page on the queues [gets moved later]
283  * => if we allocate a new page [we_own], it gets put on the queues.
284  *    either way, the result is that the page is on the queues at return time
285  * => for pages which are on loan from a uvm_object (and thus are not
286  *    owned by the anon): if successful, we return with the owning object
287  *    locked.   the caller must unlock this object when it unlocks everything
288  *    else.
289  */
290 
291 int
292 uvmfault_anonget(ufi, amap, anon)
293 	struct uvm_faultinfo *ufi;
294 	struct vm_amap *amap;
295 	struct vm_anon *anon;
296 {
297 	boolean_t we_own;	/* we own anon's page? */
298 	boolean_t locked;	/* did we relock? */
299 	struct vm_page *pg;
300 	int result;
301 	UVMHIST_FUNC("uvmfault_anonget"); UVMHIST_CALLED(maphist);
302 
303 	LOCK_ASSERT(simple_lock_held(&anon->an_lock));
304 
305 	result = 0;		/* XXX shut up gcc */
306 	uvmexp.fltanget++;
307         /* bump rusage counters */
308 	if (anon->u.an_page)
309 		curproc->p_addr->u_stats.p_ru.ru_minflt++;
310 	else
311 		curproc->p_addr->u_stats.p_ru.ru_majflt++;
312 
313 	/*
314 	 * loop until we get it, or fail.
315 	 */
316 
317 	while (1) {
318 
319 		we_own = FALSE;		/* TRUE if we set PG_BUSY on a page */
320 		pg = anon->u.an_page;
321 
322 		/*
323 		 * if there is a resident page and it is loaned, then anon
324 		 * may not own it.   call out to uvm_anon_lockpage() to ensure
325 		 * the real owner of the page has been identified and locked.
326 		 */
327 
328 		if (pg && pg->loan_count)
329 			pg = uvm_anon_lockloanpg(anon);
330 
331 		/*
332 		 * page there?   make sure it is not busy/released.
333 		 */
334 
335 		if (pg) {
336 
337 			/*
338 			 * at this point, if the page has a uobject [meaning
339 			 * we have it on loan], then that uobject is locked
340 			 * by us!   if the page is busy, we drop all the
341 			 * locks (including uobject) and try again.
342 			 */
343 
344 			if ((pg->flags & (PG_BUSY|PG_RELEASED)) == 0) {
345 				UVMHIST_LOG(maphist, "<- OK",0,0,0,0);
346 				return (0);
347 			}
348 			pg->flags |= PG_WANTED;
349 			uvmexp.fltpgwait++;
350 
351 			/*
352 			 * the last unlock must be an atomic unlock+wait on
353 			 * the owner of page
354 			 */
355 			if (pg->uobject) {	/* owner is uobject ? */
356 				uvmfault_unlockall(ufi, amap, NULL, anon);
357 				UVMHIST_LOG(maphist, " unlock+wait on uobj",0,
358 				    0,0,0);
359 				UVM_UNLOCK_AND_WAIT(pg,
360 				    &pg->uobject->vmobjlock,
361 				    FALSE, "anonget1",0);
362 			} else {
363 				/* anon owns page */
364 				uvmfault_unlockall(ufi, amap, NULL, NULL);
365 				UVMHIST_LOG(maphist, " unlock+wait on anon",0,
366 				    0,0,0);
367 				UVM_UNLOCK_AND_WAIT(pg,&anon->an_lock,0,
368 				    "anonget2",0);
369 			}
370 			/* ready to relock and try again */
371 
372 		} else {
373 
374 			/*
375 			 * no page, we must try and bring it in.
376 			 */
377 			pg = uvm_pagealloc(NULL, 0, anon, 0);
378 
379 			if (pg == NULL) {		/* out of RAM.  */
380 
381 				uvmfault_unlockall(ufi, amap, NULL, anon);
382 				uvmexp.fltnoram++;
383 				UVMHIST_LOG(maphist, "  noram -- UVM_WAIT",0,
384 				    0,0,0);
385 				uvm_wait("flt_noram1");
386 				/* ready to relock and try again */
387 
388 			} else {
389 
390 				/* we set the PG_BUSY bit */
391 				we_own = TRUE;
392 				uvmfault_unlockall(ufi, amap, NULL, anon);
393 
394 				/*
395 				 * we are passing a PG_BUSY+PG_FAKE+PG_CLEAN
396 				 * page into the uvm_swap_get function with
397 				 * all data structures unlocked.  note that
398 				 * it is ok to read an_swslot here because
399 				 * we hold PG_BUSY on the page.
400 				 */
401 				uvmexp.pageins++;
402 				result = uvm_swap_get(pg, anon->an_swslot,
403 				    PGO_SYNCIO);
404 
405 				/*
406 				 * we clean up after the i/o below in the
407 				 * "we_own" case
408 				 */
409 				/* ready to relock and try again */
410 			}
411 		}
412 
413 		/*
414 		 * now relock and try again
415 		 */
416 
417 		locked = uvmfault_relock(ufi);
418 		if (locked && amap != NULL) {
419 			amap_lock(amap);
420 		}
421 		if (locked || we_own)
422 			simple_lock(&anon->an_lock);
423 
424 		/*
425 		 * if we own the page (i.e. we set PG_BUSY), then we need
426 		 * to clean up after the I/O. there are three cases to
427 		 * consider:
428 		 *   [1] page released during I/O: free anon and ReFault.
429 		 *   [2] I/O not OK.   free the page and cause the fault
430 		 *       to fail.
431 		 *   [3] I/O OK!   activate the page and sync with the
432 		 *       non-we_own case (i.e. drop anon lock if not locked).
433 		 */
434 
435 		if (we_own) {
436 
437 			if (pg->flags & PG_WANTED) {
438 				/* still holding object lock */
439 				wakeup(pg);
440 			}
441 			/* un-busy! */
442 			pg->flags &= ~(PG_WANTED|PG_BUSY|PG_FAKE);
443 			UVM_PAGE_OWN(pg, NULL);
444 
445 			/*
446 			 * if we were RELEASED during I/O, then our anon is
447 			 * no longer part of an amap.   we need to free the
448 			 * anon and try again.
449 			 */
450 			if (pg->flags & PG_RELEASED) {
451 				pmap_page_protect(pg, VM_PROT_NONE);
452 				simple_unlock(&anon->an_lock);
453 				uvm_anfree(anon);	/* frees page for us */
454 				if (locked)
455 					uvmfault_unlockall(ufi, amap, NULL,
456 							   NULL);
457 				uvmexp.fltpgrele++;
458 				UVMHIST_LOG(maphist, "<- REFAULT", 0,0,0,0);
459 				return (ERESTART);	/* refault! */
460 			}
461 
462 			if (result != 0) {
463 				/* remove page from anon */
464 				anon->u.an_page = NULL;
465 
466 				/*
467 				 * remove the swap slot from the anon
468 				 * and mark the anon as having no real slot.
469 				 * don't free the swap slot, thus preventing
470 				 * it from being used again.
471 				 */
472 				uvm_swap_markbad(anon->an_swslot, 1);
473 				anon->an_swslot = SWSLOT_BAD;
474 
475 				/*
476 				 * note: page was never !PG_BUSY, so it
477 				 * can't be mapped and thus no need to
478 				 * pmap_page_protect it...
479 				 */
480 				uvm_lock_pageq();
481 				uvm_pagefree(pg);
482 				uvm_unlock_pageq();
483 
484 				if (locked)
485 					uvmfault_unlockall(ufi, amap, NULL,
486 					    anon);
487 				else
488 					simple_unlock(&anon->an_lock);
489 				UVMHIST_LOG(maphist, "<- ERROR", 0,0,0,0);
490 				return (EIO);
491 			}
492 
493 			/*
494 			 * must be OK, clear modify (already PG_CLEAN)
495 			 * and activate
496 			 */
497 			pmap_clear_modify(pg);
498 			uvm_lock_pageq();
499 			uvm_pageactivate(pg);
500 			uvm_unlock_pageq();
501 			if (!locked)
502 				simple_unlock(&anon->an_lock);
503 		}
504 
505 		/*
506 		 * we were not able to relock.   restart fault.
507 		 */
508 
509 		if (!locked) {
510 			UVMHIST_LOG(maphist, "<- REFAULT", 0,0,0,0);
511 			return (ERESTART);
512 		}
513 
514 		/*
515 		 * verify no one has touched the amap and moved the anon on us.
516 		 */
517 
518 		if (ufi != NULL &&
519 		    amap_lookup(&ufi->entry->aref,
520 				ufi->orig_rvaddr - ufi->entry->start) != anon) {
521 
522 			uvmfault_unlockall(ufi, amap, NULL, anon);
523 			UVMHIST_LOG(maphist, "<- REFAULT", 0,0,0,0);
524 			return (ERESTART);
525 		}
526 
527 		/*
528 		 * try it again!
529 		 */
530 
531 		uvmexp.fltanretry++;
532 		continue;
533 
534 	} /* while (1) */
535 
536 	/*NOTREACHED*/
537 }
538 
539 /*
540  *   F A U L T   -   m a i n   e n t r y   p o i n t
541  */
542 
543 /*
544  * uvm_fault: page fault handler
545  *
546  * => called from MD code to resolve a page fault
547  * => VM data structures usually should be unlocked.   however, it is
548  *	possible to call here with the main map locked if the caller
549  *	gets a write lock, sets it recusive, and then calls us (c.f.
550  *	uvm_map_pageable).   this should be avoided because it keeps
551  *	the map locked off during I/O.
552  */
553 
554 #define MASK(entry)     (UVM_ET_ISCOPYONWRITE(entry) ? \
555 			 ~VM_PROT_WRITE : VM_PROT_ALL)
556 
557 int
558 uvm_fault(orig_map, vaddr, fault_type, access_type)
559 	vm_map_t orig_map;
560 	vaddr_t vaddr;
561 	vm_fault_t fault_type;
562 	vm_prot_t access_type;
563 {
564 	struct uvm_faultinfo ufi;
565 	vm_prot_t enter_prot;
566 	boolean_t wired, narrow, promote, locked, shadowed;
567 	int npages, nback, nforw, centeridx, result, lcv, gotpages;
568 	vaddr_t startva, objaddr, currva, offset, uoff;
569 	paddr_t pa;
570 	struct vm_amap *amap;
571 	struct uvm_object *uobj;
572 	struct vm_anon *anons_store[UVM_MAXRANGE], **anons, *anon, *oanon;
573 	struct vm_page *pages[UVM_MAXRANGE], *pg, *uobjpage;
574 	UVMHIST_FUNC("uvm_fault"); UVMHIST_CALLED(maphist);
575 
576 	UVMHIST_LOG(maphist, "(map=0x%x, vaddr=0x%x, ft=%d, at=%d)",
577 	      orig_map, vaddr, fault_type, access_type);
578 
579 	anon = NULL;
580 	pg = NULL;
581 
582 	uvmexp.faults++;	/* XXX: locking? */
583 
584 	/*
585 	 * init the IN parameters in the ufi
586 	 */
587 
588 	ufi.orig_map = orig_map;
589 	ufi.orig_rvaddr = trunc_page(vaddr);
590 	ufi.orig_size = PAGE_SIZE;	/* can't get any smaller than this */
591 	if (fault_type == VM_FAULT_WIRE)
592 		narrow = TRUE;		/* don't look for neighborhood
593 					 * pages on wire */
594 	else
595 		narrow = FALSE;		/* normal fault */
596 
597 	/*
598 	 * before we do anything else, if this is a fault on a kernel
599 	 * address, check to see if the address is managed by an
600 	 * interrupt-safe map.  If it is, we fail immediately.  Intrsafe
601 	 * maps are never pageable, and this approach avoids an evil
602 	 * locking mess.
603 	 */
604 	if (orig_map == kernel_map && uvmfault_check_intrsafe(&ufi)) {
605 		UVMHIST_LOG(maphist, "<- VA 0x%lx in intrsafe map %p",
606 		    ufi.orig_rvaddr, ufi.map, 0, 0);
607 		return (KERN_FAILURE);
608 	}
609 
610 	/*
611 	 * "goto ReFault" means restart the page fault from ground zero.
612 	 */
613 ReFault:
614 
615 	/*
616 	 * lookup and lock the maps
617 	 */
618 
619 	if (uvmfault_lookup(&ufi, FALSE) == FALSE) {
620 		UVMHIST_LOG(maphist, "<- no mapping @ 0x%x", vaddr, 0,0,0);
621 		return (KERN_INVALID_ADDRESS);
622 	}
623 	/* locked: maps(read) */
624 
625 	/*
626 	 * check protection
627 	 */
628 
629 	if ((ufi.entry->protection & access_type) != access_type) {
630 		UVMHIST_LOG(maphist,
631 		    "<- protection failure (prot=0x%x, access=0x%x)",
632 		    ufi.entry->protection, access_type, 0, 0);
633 		uvmfault_unlockmaps(&ufi, FALSE);
634 		return (KERN_PROTECTION_FAILURE);
635 	}
636 
637 	/*
638 	 * if the map is not a pageable map, a page fault always fails.
639 	 */
640 
641 	if ((ufi.map->flags & VM_MAP_PAGEABLE) == 0) {
642 		UVMHIST_LOG(maphist,
643 		    "<- map %p not pageable", ufi.map, 0, 0, 0);
644 		uvmfault_unlockmaps(&ufi, FALSE);
645 		return (KERN_FAILURE);
646 	}
647 
648 	/*
649 	 * "enter_prot" is the protection we want to enter the page in at.
650 	 * for certain pages (e.g. copy-on-write pages) this protection can
651 	 * be more strict than ufi.entry->protection.  "wired" means either
652 	 * the entry is wired or we are fault-wiring the pg.
653 	 */
654 
655 	enter_prot = ufi.entry->protection;
656 	wired = VM_MAPENT_ISWIRED(ufi.entry) || (fault_type == VM_FAULT_WIRE);
657 	if (wired)
658 		access_type = enter_prot; /* full access for wired */
659 
660 	/*
661 	 * handle "needs_copy" case.   if we need to copy the amap we will
662 	 * have to drop our readlock and relock it with a write lock.  (we
663 	 * need a write lock to change anything in a map entry [e.g.
664 	 * needs_copy]).
665 	 */
666 
667 	if (UVM_ET_ISNEEDSCOPY(ufi.entry)) {
668 		if ((access_type & VM_PROT_WRITE) ||
669 		    (ufi.entry->object.uvm_obj == NULL)) {
670 			/* need to clear */
671 			UVMHIST_LOG(maphist,
672 			    "  need to clear needs_copy and refault",0,0,0,0);
673 			uvmfault_unlockmaps(&ufi, FALSE);
674 			uvmfault_amapcopy(&ufi);
675 			uvmexp.fltamcopy++;
676 			goto ReFault;
677 
678 		} else {
679 
680 			/*
681 			 * ensure that we pmap_enter page R/O since
682 			 * needs_copy is still true
683 			 */
684 			enter_prot &= ~VM_PROT_WRITE;
685 
686 		}
687 	}
688 
689 	/*
690 	 * identify the players
691 	 */
692 
693 	amap = ufi.entry->aref.ar_amap;		/* top layer */
694 	uobj = ufi.entry->object.uvm_obj;	/* bottom layer */
695 
696 	/*
697 	 * check for a case 0 fault.  if nothing backing the entry then
698 	 * error now.
699 	 */
700 
701 	if (amap == NULL && uobj == NULL) {
702 		uvmfault_unlockmaps(&ufi, FALSE);
703 		UVMHIST_LOG(maphist,"<- no backing store, no overlay",0,0,0,0);
704 		return (KERN_INVALID_ADDRESS);
705 	}
706 
707 	/*
708 	 * establish range of interest based on advice from mapper
709 	 * and then clip to fit map entry.   note that we only want
710 	 * to do this the first time through the fault.   if we
711 	 * ReFault we will disable this by setting "narrow" to true.
712 	 */
713 
714 	if (narrow == FALSE) {
715 
716 		/* wide fault (!narrow) */
717 		KASSERT(uvmadvice[ufi.entry->advice].advice ==
718 			 ufi.entry->advice);
719 		nback = min(uvmadvice[ufi.entry->advice].nback,
720 			    (ufi.orig_rvaddr - ufi.entry->start) >> PAGE_SHIFT);
721 		startva = ufi.orig_rvaddr - (nback << PAGE_SHIFT);
722 		nforw = min(uvmadvice[ufi.entry->advice].nforw,
723 			    ((ufi.entry->end - ufi.orig_rvaddr) >>
724 			     PAGE_SHIFT) - 1);
725 		/*
726 		 * note: "-1" because we don't want to count the
727 		 * faulting page as forw
728 		 */
729 		npages = nback + nforw + 1;
730 		centeridx = nback;
731 
732 		narrow = TRUE;	/* ensure only once per-fault */
733 
734 	} else {
735 
736 		/* narrow fault! */
737 		nback = nforw = 0;
738 		startva = ufi.orig_rvaddr;
739 		npages = 1;
740 		centeridx = 0;
741 
742 	}
743 
744 	/* locked: maps(read) */
745 	UVMHIST_LOG(maphist, "  narrow=%d, back=%d, forw=%d, startva=0x%x",
746 		    narrow, nback, nforw, startva);
747 	UVMHIST_LOG(maphist, "  entry=0x%x, amap=0x%x, obj=0x%x", ufi.entry,
748 		    amap, uobj, 0);
749 
750 	/*
751 	 * if we've got an amap, lock it and extract current anons.
752 	 */
753 
754 	if (amap) {
755 		amap_lock(amap);
756 		anons = anons_store;
757 		amap_lookups(&ufi.entry->aref, startva - ufi.entry->start,
758 		    anons, npages);
759 	} else {
760 		anons = NULL;	/* to be safe */
761 	}
762 
763 	/* locked: maps(read), amap(if there) */
764 
765 	/*
766 	 * for MADV_SEQUENTIAL mappings we want to deactivate the back pages
767 	 * now and then forget about them (for the rest of the fault).
768 	 */
769 
770 	if (ufi.entry->advice == MADV_SEQUENTIAL) {
771 
772 		UVMHIST_LOG(maphist, "  MADV_SEQUENTIAL: flushing backpages",
773 		    0,0,0,0);
774 		/* flush back-page anons? */
775 		if (amap)
776 			uvmfault_anonflush(anons, nback);
777 
778 		/* flush object? */
779 		if (uobj) {
780 			objaddr =
781 			    (startva - ufi.entry->start) + ufi.entry->offset;
782 			simple_lock(&uobj->vmobjlock);
783 			(void) uobj->pgops->pgo_flush(uobj, objaddr, objaddr +
784 				    (nback << PAGE_SHIFT), PGO_DEACTIVATE);
785 			simple_unlock(&uobj->vmobjlock);
786 		}
787 
788 		/* now forget about the backpages */
789 		if (amap)
790 			anons += nback;
791 		startva += (nback << PAGE_SHIFT);
792 		npages -= nback;
793 		nback = centeridx = 0;
794 	}
795 
796 	/* locked: maps(read), amap(if there) */
797 
798 	/*
799 	 * map in the backpages and frontpages we found in the amap in hopes
800 	 * of preventing future faults.    we also init the pages[] array as
801 	 * we go.
802 	 */
803 
804 	currva = startva;
805 	shadowed = FALSE;
806 	for (lcv = 0 ; lcv < npages ; lcv++, currva += PAGE_SIZE) {
807 
808 		/*
809 		 * dont play with VAs that are already mapped
810 		 * except for center)
811 		 */
812 		if (lcv != centeridx &&
813 		    pmap_extract(ufi.orig_map->pmap, currva, &pa)) {
814 			pages[lcv] = PGO_DONTCARE;
815 			continue;
816 		}
817 
818 		/*
819 		 * unmapped or center page.   check if any anon at this level.
820 		 */
821 		if (amap == NULL || anons[lcv] == NULL) {
822 			pages[lcv] = NULL;
823 			continue;
824 		}
825 
826 		/*
827 		 * check for present page and map if possible.   re-activate it.
828 		 */
829 
830 		pages[lcv] = PGO_DONTCARE;
831 		if (lcv == centeridx) {		/* save center for later! */
832 			shadowed = TRUE;
833 			continue;
834 		}
835 		anon = anons[lcv];
836 		simple_lock(&anon->an_lock);
837 		/* ignore loaned pages */
838 		if (anon->u.an_page && anon->u.an_page->loan_count == 0 &&
839 			(anon->u.an_page->flags & (PG_RELEASED|PG_BUSY)) == 0) {
840 			uvm_lock_pageq();
841 			uvm_pageactivate(anon->u.an_page);	/* reactivate */
842 			uvm_unlock_pageq();
843 			UVMHIST_LOG(maphist,
844 			    "  MAPPING: n anon: pm=0x%x, va=0x%x, pg=0x%x",
845 			    ufi.orig_map->pmap, currva, anon->u.an_page, 0);
846 			uvmexp.fltnamap++;
847 
848 			/*
849 			 * Since this isn't the page that's actually faulting,
850 			 * ignore pmap_enter() failures; it's not critical
851 			 * that we enter these right now.
852 			 */
853 
854 			(void) pmap_enter(ufi.orig_map->pmap, currva,
855 			    VM_PAGE_TO_PHYS(anon->u.an_page),
856 			    (anon->an_ref > 1) ? (enter_prot & ~VM_PROT_WRITE) :
857 			    enter_prot,
858 			    PMAP_CANFAIL |
859 			     (VM_MAPENT_ISWIRED(ufi.entry) ? PMAP_WIRED : 0));
860 		}
861 		simple_unlock(&anon->an_lock);
862 	}
863 
864 	/* locked: maps(read), amap(if there) */
865 	/* (shadowed == TRUE) if there is an anon at the faulting address */
866 	UVMHIST_LOG(maphist, "  shadowed=%d, will_get=%d", shadowed,
867 	    (uobj && shadowed == FALSE),0,0);
868 
869 	/*
870 	 * note that if we are really short of RAM we could sleep in the above
871 	 * call to pmap_enter with everything locked.   bad?
872 	 *
873 	 * XXX Actually, that is bad; pmap_enter() should just fail in that
874 	 * XXX case.  --thorpej
875 	 */
876 
877 	/*
878 	 * if the desired page is not shadowed by the amap and we have a
879 	 * backing object, then we check to see if the backing object would
880 	 * prefer to handle the fault itself (rather than letting us do it
881 	 * with the usual pgo_get hook).  the backing object signals this by
882 	 * providing a pgo_fault routine.
883 	 */
884 
885 	if (uobj && shadowed == FALSE && uobj->pgops->pgo_fault != NULL) {
886 		simple_lock(&uobj->vmobjlock);
887 
888 		/* locked: maps(read), amap (if there), uobj */
889 		result = uobj->pgops->pgo_fault(&ufi, startva, pages, npages,
890 				    centeridx, fault_type, access_type,
891 				    PGO_LOCKED|PGO_SYNCIO);
892 
893 		/* locked: nothing, pgo_fault has unlocked everything */
894 
895 		if (result == 0)
896 			return (KERN_SUCCESS);	/* pgo_fault did pmap enter */
897 		else if (result == ERESTART)
898 			goto ReFault;		/* try again! */
899 		else
900 			return (KERN_PROTECTION_FAILURE);
901 	}
902 
903 	/*
904 	 * now, if the desired page is not shadowed by the amap and we have
905 	 * a backing object that does not have a special fault routine, then
906 	 * we ask (with pgo_get) the object for resident pages that we care
907 	 * about and attempt to map them in.  we do not let pgo_get block
908 	 * (PGO_LOCKED).
909 	 *
910 	 * ("get" has the option of doing a pmap_enter for us)
911 	 */
912 
913 	if (uobj && shadowed == FALSE) {
914 		simple_lock(&uobj->vmobjlock);
915 
916 		/* locked (!shadowed): maps(read), amap (if there), uobj */
917 		/*
918 		 * the following call to pgo_get does _not_ change locking state
919 		 */
920 
921 		uvmexp.fltlget++;
922 		gotpages = npages;
923 		(void) uobj->pgops->pgo_get(uobj, ufi.entry->offset +
924 				(startva - ufi.entry->start),
925 				pages, &gotpages, centeridx,
926 				access_type & MASK(ufi.entry),
927 				ufi.entry->advice, PGO_LOCKED);
928 
929 		/*
930 		 * check for pages to map, if we got any
931 		 */
932 
933 		uobjpage = NULL;
934 
935 		if (gotpages) {
936 			currva = startva;
937 			for (lcv = 0 ; lcv < npages ;
938 			    lcv++, currva += PAGE_SIZE) {
939 
940 				if (pages[lcv] == NULL ||
941 				    pages[lcv] == PGO_DONTCARE)
942 					continue;
943 
944 				KASSERT((pages[lcv]->flags & PG_RELEASED) == 0);
945 
946 				/*
947 				 * if center page is resident and not
948 				 * PG_BUSY|PG_RELEASED then pgo_get
949 				 * made it PG_BUSY for us and gave
950 				 * us a handle to it.   remember this
951 				 * page as "uobjpage." (for later use).
952 				 */
953 
954 				if (lcv == centeridx) {
955 					uobjpage = pages[lcv];
956 					UVMHIST_LOG(maphist, "  got uobjpage "
957 					    "(0x%x) with locked get",
958 					    uobjpage, 0,0,0);
959 					continue;
960 				}
961 
962 				/*
963 				 * note: calling pgo_get with locked data
964 				 * structures returns us pages which are
965 				 * neither busy nor released, so we don't
966 				 * need to check for this.   we can just
967 				 * directly enter the page (after moving it
968 				 * to the head of the active queue [useful?]).
969 				 */
970 
971 				uvm_lock_pageq();
972 				uvm_pageactivate(pages[lcv]);	/* reactivate */
973 				uvm_unlock_pageq();
974 				UVMHIST_LOG(maphist,
975 				  "  MAPPING: n obj: pm=0x%x, va=0x%x, pg=0x%x",
976 				  ufi.orig_map->pmap, currva, pages[lcv], 0);
977 				uvmexp.fltnomap++;
978 
979 				/*
980 				 * Since this page isn't the page that's
981 				 * actually fauling, ignore pmap_enter()
982 				 * failures; it's not critical that we
983 				 * enter these right now.
984 				 */
985 
986 				(void) pmap_enter(ufi.orig_map->pmap, currva,
987 				    VM_PAGE_TO_PHYS(pages[lcv]),
988 				    pages[lcv]->flags & PG_RDONLY ?
989 				    VM_PROT_READ : enter_prot & MASK(ufi.entry),
990 				    PMAP_CANFAIL |
991 				     (wired ? PMAP_WIRED : 0));
992 
993 				/*
994 				 * NOTE: page can't be PG_WANTED or PG_RELEASED
995 				 * because we've held the lock the whole time
996 				 * we've had the handle.
997 				 */
998 
999 				pages[lcv]->flags &= ~(PG_BUSY); /* un-busy! */
1000 				UVM_PAGE_OWN(pages[lcv], NULL);
1001 			}	/* for "lcv" loop */
1002 		}   /* "gotpages" != 0 */
1003 		/* note: object still _locked_ */
1004 	} else {
1005 		uobjpage = NULL;
1006 	}
1007 
1008 	/* locked (shadowed): maps(read), amap */
1009 	/* locked (!shadowed): maps(read), amap(if there),
1010 		 uobj(if !null), uobjpage(if !null) */
1011 
1012 	/*
1013 	 * note that at this point we are done with any front or back pages.
1014 	 * we are now going to focus on the center page (i.e. the one we've
1015 	 * faulted on).  if we have faulted on the top (anon) layer
1016 	 * [i.e. case 1], then the anon we want is anons[centeridx] (we have
1017 	 * not touched it yet).  if we have faulted on the bottom (uobj)
1018 	 * layer [i.e. case 2] and the page was both present and available,
1019 	 * then we've got a pointer to it as "uobjpage" and we've already
1020 	 * made it BUSY.
1021 	 */
1022 
1023 	/*
1024 	 * there are four possible cases we must address: 1A, 1B, 2A, and 2B
1025 	 */
1026 
1027 	/*
1028 	 * redirect case 2: if we are not shadowed, go to case 2.
1029 	 */
1030 
1031 	if (shadowed == FALSE)
1032 		goto Case2;
1033 
1034 	/* locked: maps(read), amap */
1035 
1036 	/*
1037 	 * handle case 1: fault on an anon in our amap
1038 	 */
1039 
1040 	anon = anons[centeridx];
1041 	UVMHIST_LOG(maphist, "  case 1 fault: anon=0x%x", anon, 0,0,0);
1042 	simple_lock(&anon->an_lock);
1043 
1044 	/* locked: maps(read), amap, anon */
1045 
1046 	/*
1047 	 * no matter if we have case 1A or case 1B we are going to need to
1048 	 * have the anon's memory resident.   ensure that now.
1049 	 */
1050 
1051 	/*
1052 	 * let uvmfault_anonget do the dirty work.
1053 	 * if it fails (!OK) it will unlock everything for us.
1054 	 * if it succeeds, locks are still valid and locked.
1055 	 * also, if it is OK, then the anon's page is on the queues.
1056 	 * if the page is on loan from a uvm_object, then anonget will
1057 	 * lock that object for us if it does not fail.
1058 	 */
1059 
1060 	result = uvmfault_anonget(&ufi, amap, anon);
1061 	switch (result) {
1062 	case 0:
1063 		break;
1064 
1065 	case ERESTART:
1066 		goto ReFault;
1067 
1068 	case EAGAIN:
1069 		tsleep(&lbolt, PVM, "fltagain1", 0);
1070 		goto ReFault;
1071 
1072 	default:
1073 #ifdef DIAGNOSTIC
1074 		panic("uvm_fault: uvmfault_anonget -> %d", result);
1075 #else
1076 		return (KERN_PROTECTION_FAILURE);
1077 #endif
1078 	}
1079 
1080 	/*
1081 	 * uobj is non null if the page is on loan from an object (i.e. uobj)
1082 	 */
1083 
1084 	uobj = anon->u.an_page->uobject;	/* locked by anonget if !NULL */
1085 
1086 	/* locked: maps(read), amap, anon, uobj(if one) */
1087 
1088 	/*
1089 	 * special handling for loaned pages
1090 	 */
1091 
1092 	if (anon->u.an_page->loan_count) {
1093 
1094 		if ((access_type & VM_PROT_WRITE) == 0) {
1095 
1096 			/*
1097 			 * for read faults on loaned pages we just cap the
1098 			 * protection at read-only.
1099 			 */
1100 
1101 			enter_prot = enter_prot & ~VM_PROT_WRITE;
1102 
1103 		} else {
1104 			/*
1105 			 * note that we can't allow writes into a loaned page!
1106 			 *
1107 			 * if we have a write fault on a loaned page in an
1108 			 * anon then we need to look at the anon's ref count.
1109 			 * if it is greater than one then we are going to do
1110 			 * a normal copy-on-write fault into a new anon (this
1111 			 * is not a problem).  however, if the reference count
1112 			 * is one (a case where we would normally allow a
1113 			 * write directly to the page) then we need to kill
1114 			 * the loan before we continue.
1115 			 */
1116 
1117 			/* >1 case is already ok */
1118 			if (anon->an_ref == 1) {
1119 
1120 				/* get new un-owned replacement page */
1121 				pg = uvm_pagealloc(NULL, 0, NULL, 0);
1122 				if (pg == NULL) {
1123 					uvmfault_unlockall(&ufi, amap, uobj,
1124 					    anon);
1125 					uvm_wait("flt_noram2");
1126 					goto ReFault;
1127 				}
1128 
1129 				/*
1130 				 * copy data, kill loan, and drop uobj lock
1131 				 * (if any)
1132 				 */
1133 				/* copy old -> new */
1134 				uvm_pagecopy(anon->u.an_page, pg);
1135 
1136 				/* force reload */
1137 				pmap_page_protect(anon->u.an_page,
1138 						  VM_PROT_NONE);
1139 				uvm_lock_pageq();	  /* KILL loan */
1140 				if (uobj)
1141 					/* if we were loaning */
1142 					anon->u.an_page->loan_count--;
1143 				anon->u.an_page->uanon = NULL;
1144 				/* in case we owned */
1145 				anon->u.an_page->pqflags &= ~PQ_ANON;
1146 				uvm_unlock_pageq();
1147 				if (uobj) {
1148 					simple_unlock(&uobj->vmobjlock);
1149 					uobj = NULL;
1150 				}
1151 
1152 				/* install new page in anon */
1153 				anon->u.an_page = pg;
1154 				pg->uanon = anon;
1155 				pg->pqflags |= PQ_ANON;
1156 				pg->flags &= ~(PG_BUSY|PG_FAKE);
1157 				UVM_PAGE_OWN(pg, NULL);
1158 
1159 				/* done! */
1160 			}     /* ref == 1 */
1161 		}       /* write fault */
1162 	}         /* loan count */
1163 
1164 	/*
1165 	 * if we are case 1B then we will need to allocate a new blank
1166 	 * anon to transfer the data into.   note that we have a lock
1167 	 * on anon, so no one can busy or release the page until we are done.
1168 	 * also note that the ref count can't drop to zero here because
1169 	 * it is > 1 and we are only dropping one ref.
1170 	 *
1171 	 * in the (hopefully very rare) case that we are out of RAM we
1172 	 * will unlock, wait for more RAM, and refault.
1173 	 *
1174 	 * if we are out of anon VM we kill the process (XXX: could wait?).
1175 	 */
1176 
1177 	if ((access_type & VM_PROT_WRITE) != 0 && anon->an_ref > 1) {
1178 
1179 		UVMHIST_LOG(maphist, "  case 1B: COW fault",0,0,0,0);
1180 		uvmexp.flt_acow++;
1181 		oanon = anon;		/* oanon = old, locked anon */
1182 		anon = uvm_analloc();
1183 		if (anon) {
1184 			/* new anon is locked! */
1185 			pg = uvm_pagealloc(NULL, 0, anon, 0);
1186 		}
1187 
1188 		/* check for out of RAM */
1189 		if (anon == NULL || pg == NULL) {
1190 			if (anon) {
1191 				anon->an_ref--;
1192 				simple_unlock(&anon->an_lock);
1193 				uvm_anfree(anon);
1194 			}
1195 			uvmfault_unlockall(&ufi, amap, uobj, oanon);
1196 			KASSERT(uvmexp.swpgonly <= uvmexp.swpages);
1197 			if (anon == NULL || uvmexp.swpgonly == uvmexp.swpages) {
1198 				UVMHIST_LOG(maphist,
1199 				    "<- failed.  out of VM",0,0,0,0);
1200 				uvmexp.fltnoanon++;
1201 				return (KERN_RESOURCE_SHORTAGE);
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 
1211 		uvm_pagecopy(oanon->u.an_page, pg);	/* pg now !PG_CLEAN */
1212 		pg->flags &= ~(PG_BUSY|PG_FAKE);	/* un-busy! new page */
1213 		UVM_PAGE_OWN(pg, NULL);
1214 		amap_add(&ufi.entry->aref, ufi.orig_rvaddr - ufi.entry->start,
1215 		    anon, 1);
1216 
1217 		/* deref: can not drop to zero here by defn! */
1218 		oanon->an_ref--;
1219 
1220 		/*
1221 		 * note: oanon is still locked, as is the new anon.  we
1222 		 * need to check for this later when we unlock oanon; if
1223 		 * oanon != anon, we'll have to unlock anon, too.
1224 		 */
1225 
1226 	} else {
1227 
1228 		uvmexp.flt_anon++;
1229 		oanon = anon;		/* old, locked anon is same as anon */
1230 		pg = anon->u.an_page;
1231 		if (anon->an_ref > 1)     /* disallow writes to ref > 1 anons */
1232 			enter_prot = enter_prot & ~VM_PROT_WRITE;
1233 
1234 	}
1235 
1236 	/* locked: maps(read), amap, oanon, anon (if different from oanon) */
1237 
1238 	/*
1239 	 * now map the page in ...
1240 	 * XXX: old fault unlocks object before pmap_enter.  this seems
1241 	 * suspect since some other thread could blast the page out from
1242 	 * under us between the unlock and the pmap_enter.
1243 	 */
1244 
1245 	UVMHIST_LOG(maphist, "  MAPPING: anon: pm=0x%x, va=0x%x, pg=0x%x",
1246 	    ufi.orig_map->pmap, ufi.orig_rvaddr, pg, 0);
1247 	if (pmap_enter(ufi.orig_map->pmap, ufi.orig_rvaddr, VM_PAGE_TO_PHYS(pg),
1248 	    enter_prot, access_type | PMAP_CANFAIL | (wired ? PMAP_WIRED : 0))
1249 	    != KERN_SUCCESS) {
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 		if (anon != oanon)
1258 			simple_unlock(&anon->an_lock);
1259 		uvmfault_unlockall(&ufi, amap, uobj, oanon);
1260 		KASSERT(uvmexp.swpgonly <= uvmexp.swpages);
1261 		if (uvmexp.swpgonly == uvmexp.swpages) {
1262 			UVMHIST_LOG(maphist,
1263 			    "<- failed.  out of VM",0,0,0,0);
1264 			/* XXX instrumentation */
1265 			return (KERN_RESOURCE_SHORTAGE);
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 
1278 	if (fault_type == VM_FAULT_WIRE) {
1279 		uvm_pagewire(pg);
1280 
1281 		/*
1282 		 * since the now-wired page cannot be paged out,
1283 		 * release its swap resources for others to use.
1284 		 * since an anon with no swap cannot be PG_CLEAN,
1285 		 * clear its clean flag now.
1286 		 */
1287 
1288 		pg->flags &= ~(PG_CLEAN);
1289 		uvm_anon_dropswap(anon);
1290 	} else {
1291 		/* activate it */
1292 		uvm_pageactivate(pg);
1293 	}
1294 
1295 	uvm_unlock_pageq();
1296 
1297 	/*
1298 	 * done case 1!  finish up by unlocking everything and returning success
1299 	 */
1300 
1301 	if (anon != oanon)
1302 		simple_unlock(&anon->an_lock);
1303 	uvmfault_unlockall(&ufi, amap, uobj, oanon);
1304 	return (KERN_SUCCESS);
1305 
1306 
1307 Case2:
1308 	/*
1309 	 * handle case 2: faulting on backing object or zero fill
1310 	 */
1311 
1312 	/*
1313 	 * locked:
1314 	 * maps(read), amap(if there), uobj(if !null), uobjpage(if !null)
1315 	 */
1316 
1317 	/*
1318 	 * note that uobjpage can not be PGO_DONTCARE at this point.  we now
1319 	 * set uobjpage to PGO_DONTCARE if we are doing a zero fill.  if we
1320 	 * have a backing object, check and see if we are going to promote
1321 	 * the data up to an anon during the fault.
1322 	 */
1323 
1324 	if (uobj == NULL) {
1325 		uobjpage = PGO_DONTCARE;
1326 		promote = TRUE;		/* always need anon here */
1327 	} else {
1328 		KASSERT(uobjpage != PGO_DONTCARE);
1329 		promote = (access_type & VM_PROT_WRITE) &&
1330 		     UVM_ET_ISCOPYONWRITE(ufi.entry);
1331 	}
1332 	UVMHIST_LOG(maphist, "  case 2 fault: promote=%d, zfill=%d",
1333 	    promote, (uobj == NULL), 0,0);
1334 
1335 	/*
1336 	 * if uobjpage is not null then we do not need to do I/O to get the
1337 	 * uobjpage.
1338 	 *
1339 	 * if uobjpage is null, then we need to unlock and ask the pager to
1340 	 * get the data for us.   once we have the data, we need to reverify
1341 	 * the state the world.   we are currently not holding any resources.
1342 	 */
1343 
1344 	if (uobjpage) {
1345 		/* update rusage counters */
1346 		curproc->p_addr->u_stats.p_ru.ru_minflt++;
1347 	} else {
1348 		/* update rusage counters */
1349 		curproc->p_addr->u_stats.p_ru.ru_majflt++;
1350 
1351 		/* locked: maps(read), amap(if there), uobj */
1352 		uvmfault_unlockall(&ufi, amap, NULL, NULL);
1353 		/* locked: uobj */
1354 
1355 		uvmexp.fltget++;
1356 		gotpages = 1;
1357 		uoff = (ufi.orig_rvaddr - ufi.entry->start) + ufi.entry->offset;
1358 		result = uobj->pgops->pgo_get(uobj, uoff, &uobjpage, &gotpages,
1359 		    0, access_type & MASK(ufi.entry), ufi.entry->advice,
1360 		    PGO_SYNCIO);
1361 
1362 		/* locked: uobjpage(if result OK) */
1363 
1364 		/*
1365 		 * recover from I/O
1366 		 */
1367 
1368 		if (result != 0) {
1369 			if (result == EAGAIN) {
1370 				UVMHIST_LOG(maphist,
1371 				    "  pgo_get says TRY AGAIN!",0,0,0,0);
1372 				tsleep(&lbolt, PVM, "fltagain2", 0);
1373 				goto ReFault;
1374 			}
1375 
1376 			UVMHIST_LOG(maphist, "<- pgo_get failed (code %d)",
1377 			    result, 0,0,0);
1378 			return (KERN_PROTECTION_FAILURE); /* XXX i/o error */
1379 		}
1380 
1381 		/* locked: uobjpage */
1382 
1383 		/*
1384 		 * re-verify the state of the world by first trying to relock
1385 		 * the maps.  always relock the object.
1386 		 */
1387 
1388 		locked = uvmfault_relock(&ufi);
1389 		if (locked && amap)
1390 			amap_lock(amap);
1391 		simple_lock(&uobj->vmobjlock);
1392 
1393 		/* locked(locked): maps(read), amap(if !null), uobj, uobjpage */
1394 		/* locked(!locked): uobj, uobjpage */
1395 
1396 		/*
1397 		 * verify that the page has not be released and re-verify
1398 		 * that amap slot is still free.   if there is a problem,
1399 		 * we unlock and clean up.
1400 		 */
1401 
1402 		if ((uobjpage->flags & PG_RELEASED) != 0 ||
1403 		    (locked && amap &&
1404 		    amap_lookup(&ufi.entry->aref,
1405 		      ufi.orig_rvaddr - ufi.entry->start))) {
1406 			if (locked)
1407 				uvmfault_unlockall(&ufi, amap, NULL, NULL);
1408 			locked = FALSE;
1409 		}
1410 
1411 		/*
1412 		 * didn't get the lock?   release the page and retry.
1413 		 */
1414 
1415 		if (locked == FALSE) {
1416 
1417 			UVMHIST_LOG(maphist,
1418 			    "  wasn't able to relock after fault: retry",
1419 			    0,0,0,0);
1420 			if (uobjpage->flags & PG_WANTED)
1421 				/* still holding object lock */
1422 				wakeup(uobjpage);
1423 
1424 			if (uobjpage->flags & PG_RELEASED) {
1425 				uvmexp.fltpgrele++;
1426 				KASSERT(uobj->pgops->pgo_releasepg != NULL);
1427 
1428 				/* frees page */
1429 				if (uobj->pgops->pgo_releasepg(uobjpage,NULL))
1430 					/* unlock if still alive */
1431 					simple_unlock(&uobj->vmobjlock);
1432 				goto ReFault;
1433 			}
1434 
1435 			uvm_lock_pageq();
1436 			/* make sure it is in queues */
1437 			uvm_pageactivate(uobjpage);
1438 
1439 			uvm_unlock_pageq();
1440 			uobjpage->flags &= ~(PG_BUSY|PG_WANTED);
1441 			UVM_PAGE_OWN(uobjpage, NULL);
1442 			simple_unlock(&uobj->vmobjlock);
1443 			goto ReFault;
1444 
1445 		}
1446 
1447 		/*
1448 		 * we have the data in uobjpage which is PG_BUSY and
1449 		 * !PG_RELEASED.  we are holding object lock (so the page
1450 		 * can't be released on us).
1451 		 */
1452 
1453 		/* locked: maps(read), amap(if !null), uobj, uobjpage */
1454 	}
1455 
1456 	/*
1457 	 * locked:
1458 	 * maps(read), amap(if !null), uobj(if !null), uobjpage(if uobj)
1459 	 */
1460 
1461 	/*
1462 	 * notes:
1463 	 *  - at this point uobjpage can not be NULL
1464 	 *  - at this point uobjpage can not be PG_RELEASED (since we checked
1465 	 *  for it above)
1466 	 *  - at this point uobjpage could be PG_WANTED (handle later)
1467 	 */
1468 
1469 	if (promote == FALSE) {
1470 
1471 		/*
1472 		 * we are not promoting.   if the mapping is COW ensure that we
1473 		 * don't give more access than we should (e.g. when doing a read
1474 		 * fault on a COPYONWRITE mapping we want to map the COW page in
1475 		 * R/O even though the entry protection could be R/W).
1476 		 *
1477 		 * set "pg" to the page we want to map in (uobjpage, usually)
1478 		 */
1479 
1480 		/* no anon in this case. */
1481 		anon = NULL;
1482 
1483 		uvmexp.flt_obj++;
1484 		if (UVM_ET_ISCOPYONWRITE(ufi.entry))
1485 			enter_prot &= ~VM_PROT_WRITE;
1486 		pg = uobjpage;		/* map in the actual object */
1487 
1488 		/* assert(uobjpage != PGO_DONTCARE) */
1489 
1490 		/*
1491 		 * we are faulting directly on the page.   be careful
1492 		 * about writing to loaned pages...
1493 		 */
1494 		if (uobjpage->loan_count) {
1495 
1496 			if ((access_type & VM_PROT_WRITE) == 0) {
1497 				/* read fault: cap the protection at readonly */
1498 				/* cap! */
1499 				enter_prot = enter_prot & ~VM_PROT_WRITE;
1500 			} else {
1501 				/* write fault: must break the loan here */
1502 
1503 				/* alloc new un-owned page */
1504 				pg = uvm_pagealloc(NULL, 0, NULL, 0);
1505 
1506 				if (pg == NULL) {
1507 					/*
1508 					 * drop ownership of page, it can't
1509 					 * be released
1510 					 */
1511 					if (uobjpage->flags & PG_WANTED)
1512 						wakeup(uobjpage);
1513 					uobjpage->flags &= ~(PG_BUSY|PG_WANTED);
1514 					UVM_PAGE_OWN(uobjpage, NULL);
1515 
1516 					uvm_lock_pageq();
1517 					/* activate: we will need it later */
1518 					uvm_pageactivate(uobjpage);
1519 
1520 					uvm_unlock_pageq();
1521 					uvmfault_unlockall(&ufi, amap, uobj,
1522 					  NULL);
1523 					UVMHIST_LOG(maphist,
1524 					  "  out of RAM breaking loan, waiting",
1525 					  0,0,0,0);
1526 					uvmexp.fltnoram++;
1527 					uvm_wait("flt_noram4");
1528 					goto ReFault;
1529 				}
1530 
1531 				/*
1532 				 * copy the data from the old page to the new
1533 				 * one and clear the fake/clean flags on the
1534 				 * new page (keep it busy).  force a reload
1535 				 * of the old page by clearing it from all
1536 				 * pmaps.  then lock the page queues to
1537 				 * rename the pages.
1538 				 */
1539 				uvm_pagecopy(uobjpage, pg);	/* old -> new */
1540 				pg->flags &= ~(PG_FAKE|PG_CLEAN);
1541 				pmap_page_protect(uobjpage, VM_PROT_NONE);
1542 				if (uobjpage->flags & PG_WANTED)
1543 					wakeup(uobjpage);
1544 				/* uobj still locked */
1545 				uobjpage->flags &= ~(PG_WANTED|PG_BUSY);
1546 				UVM_PAGE_OWN(uobjpage, NULL);
1547 
1548 				uvm_lock_pageq();
1549 				offset = uobjpage->offset;
1550 				/* remove old page */
1551 				uvm_pagerealloc(uobjpage, NULL, 0);
1552 
1553 				/*
1554 				 * at this point we have absolutely no
1555 				 * control over uobjpage
1556 				 */
1557 				/* install new page */
1558 				uvm_pagerealloc(pg, uobj, offset);
1559 				uvm_unlock_pageq();
1560 
1561 				/*
1562 				 * done!  loan is broken and "pg" is
1563 				 * PG_BUSY.   it can now replace uobjpage.
1564 				 */
1565 
1566 				uobjpage = pg;
1567 
1568 			}		/* write fault case */
1569 		}		/* if loan_count */
1570 
1571 	} else {
1572 
1573 		/*
1574 		 * if we are going to promote the data to an anon we
1575 		 * allocate a blank anon here and plug it into our amap.
1576 		 */
1577 #if DIAGNOSTIC
1578 		if (amap == NULL)
1579 			panic("uvm_fault: want to promote data, but no anon");
1580 #endif
1581 
1582 		anon = uvm_analloc();
1583 		if (anon) {
1584 			/*
1585 			 * The new anon is locked.
1586 			 *
1587 			 * In `Fill in data...' below, if
1588 			 * uobjpage == PGO_DONTCARE, we want
1589 			 * a zero'd, dirty page, so have
1590 			 * uvm_pagealloc() do that for us.
1591 			 */
1592 			pg = uvm_pagealloc(NULL, 0, anon,
1593 			    (uobjpage == PGO_DONTCARE) ? UVM_PGA_ZERO : 0);
1594 		}
1595 
1596 		/*
1597 		 * out of memory resources?
1598 		 */
1599 		if (anon == NULL || pg == NULL) {
1600 
1601 			if (anon != NULL) {
1602 				anon->an_ref--;
1603 				simple_unlock(&anon->an_lock);
1604 				uvm_anfree(anon);
1605 			}
1606 
1607 			/*
1608 			 * arg!  must unbusy our page and fail or sleep.
1609 			 */
1610 			if (uobjpage != PGO_DONTCARE) {
1611 				if (uobjpage->flags & PG_WANTED)
1612 					/* still holding object lock */
1613 					wakeup(uobjpage);
1614 
1615 				uvm_lock_pageq();
1616 				uvm_pageactivate(uobjpage);
1617 				uvm_unlock_pageq();
1618 				uobjpage->flags &= ~(PG_BUSY|PG_WANTED);
1619 				UVM_PAGE_OWN(uobjpage, NULL);
1620 			}
1621 
1622 			/* unlock and fail ... */
1623 			uvmfault_unlockall(&ufi, amap, uobj, NULL);
1624 			KASSERT(uvmexp.swpgonly <= uvmexp.swpages);
1625 			if (anon == NULL || uvmexp.swpgonly == uvmexp.swpages) {
1626 				UVMHIST_LOG(maphist, "  promote: out of VM",
1627 				    0,0,0,0);
1628 				uvmexp.fltnoanon++;
1629 				return (KERN_RESOURCE_SHORTAGE);
1630 			}
1631 
1632 			UVMHIST_LOG(maphist, "  out of RAM, waiting for more",
1633 			    0,0,0,0);
1634 			uvmexp.fltnoram++;
1635 			uvm_wait("flt_noram5");
1636 			goto ReFault;
1637 		}
1638 
1639 		/*
1640 		 * fill in the data
1641 		 */
1642 
1643 		if (uobjpage != PGO_DONTCARE) {
1644 			uvmexp.flt_prcopy++;
1645 			/* copy page [pg now dirty] */
1646 			uvm_pagecopy(uobjpage, pg);
1647 
1648 			/*
1649 			 * promote to shared amap?  make sure all sharing
1650 			 * procs see it
1651 			 */
1652 			if ((amap_flags(amap) & AMAP_SHARED) != 0) {
1653 				pmap_page_protect(uobjpage, VM_PROT_NONE);
1654 			}
1655 
1656 			/*
1657 			 * dispose of uobjpage.  it can't be PG_RELEASED
1658 			 * since we still hold the object lock.
1659 			 * drop handle to uobj as well.
1660 			 */
1661 
1662 			if (uobjpage->flags & PG_WANTED)
1663 				/* still have the obj lock */
1664 				wakeup(uobjpage);
1665 			uobjpage->flags &= ~(PG_BUSY|PG_WANTED);
1666 			UVM_PAGE_OWN(uobjpage, NULL);
1667 			uvm_lock_pageq();
1668 			uvm_pageactivate(uobjpage);
1669 			uvm_unlock_pageq();
1670 			simple_unlock(&uobj->vmobjlock);
1671 			uobj = NULL;
1672 
1673 			UVMHIST_LOG(maphist,
1674 			    "  promote uobjpage 0x%x to anon/page 0x%x/0x%x",
1675 			    uobjpage, anon, pg, 0);
1676 
1677 		} else {
1678 			uvmexp.flt_przero++;
1679 			/*
1680 			 * Page is zero'd and marked dirty by uvm_pagealloc()
1681 			 * above.
1682 			 */
1683 			UVMHIST_LOG(maphist,"  zero fill anon/page 0x%x/0%x",
1684 			    anon, pg, 0, 0);
1685 		}
1686 
1687 		amap_add(&ufi.entry->aref, ufi.orig_rvaddr - ufi.entry->start,
1688 		    anon, 0);
1689 	}
1690 
1691 	/*
1692 	 * locked:
1693 	 * maps(read), amap(if !null), uobj(if !null), uobjpage(if uobj),
1694 	 *   anon(if !null), pg(if anon)
1695 	 *
1696 	 * note: pg is either the uobjpage or the new page in the new anon
1697 	 */
1698 
1699 	/*
1700 	 * all resources are present.   we can now map it in and free our
1701 	 * resources.
1702 	 */
1703 
1704 	UVMHIST_LOG(maphist,
1705 	    "  MAPPING: case2: pm=0x%x, va=0x%x, pg=0x%x, promote=%d",
1706 	    ufi.orig_map->pmap, ufi.orig_rvaddr, pg, promote);
1707 	KASSERT(access_type == VM_PROT_READ || (pg->flags & PG_RDONLY) == 0);
1708 	if (pmap_enter(ufi.orig_map->pmap, ufi.orig_rvaddr, VM_PAGE_TO_PHYS(pg),
1709 	    pg->flags & PG_RDONLY ? VM_PROT_READ : enter_prot,
1710 	    access_type | PMAP_CANFAIL | (wired ? PMAP_WIRED : 0))
1711 	    != KERN_SUCCESS) {
1712 
1713 		/*
1714 		 * No need to undo what we did; we can simply think of
1715 		 * this as the pmap throwing away the mapping information.
1716 		 *
1717 		 * We do, however, have to go through the ReFault path,
1718 		 * as the map may change while we're asleep.
1719 		 */
1720 
1721 		if (pg->flags & PG_WANTED)
1722 			wakeup(pg);		/* lock still held */
1723 
1724 		/*
1725 		 * note that pg can't be PG_RELEASED since we did not drop
1726 		 * the object lock since the last time we checked.
1727 		 */
1728 
1729 		pg->flags &= ~(PG_BUSY|PG_FAKE|PG_WANTED);
1730 		UVM_PAGE_OWN(pg, NULL);
1731 		uvmfault_unlockall(&ufi, amap, uobj, anon);
1732 		KASSERT(uvmexp.swpgonly <= uvmexp.swpages);
1733 		if (uvmexp.swpgonly == uvmexp.swpages) {
1734 			UVMHIST_LOG(maphist,
1735 			    "<- failed.  out of VM",0,0,0,0);
1736 			/* XXX instrumentation */
1737 			return (KERN_RESOURCE_SHORTAGE);
1738 		}
1739 		/* XXX instrumentation */
1740 		uvm_wait("flt_pmfail2");
1741 		goto ReFault;
1742 	}
1743 
1744 	uvm_lock_pageq();
1745 
1746 	if (fault_type == VM_FAULT_WIRE) {
1747 		uvm_pagewire(pg);
1748 		if (pg->pqflags & PQ_AOBJ) {
1749 
1750 			/*
1751 			 * since the now-wired page cannot be paged out,
1752 			 * release its swap resources for others to use.
1753 			 * since an aobj page with no swap cannot be PG_CLEAN,
1754 			 * clear its clean flag now.
1755 			 */
1756 
1757 			pg->flags &= ~(PG_CLEAN);
1758 			uao_dropswap(uobj, pg->offset >> PAGE_SHIFT);
1759 		}
1760 	} else {
1761 		/* activate it */
1762 		uvm_pageactivate(pg);
1763 	}
1764 	uvm_unlock_pageq();
1765 
1766 	if (pg->flags & PG_WANTED)
1767 		wakeup(pg);		/* lock still held */
1768 
1769 	/*
1770 	 * note that pg can't be PG_RELEASED since we did not drop the object
1771 	 * lock since the last time we checked.
1772 	 */
1773 
1774 	pg->flags &= ~(PG_BUSY|PG_FAKE|PG_WANTED);
1775 	UVM_PAGE_OWN(pg, NULL);
1776 	uvmfault_unlockall(&ufi, amap, uobj, anon);
1777 
1778 	UVMHIST_LOG(maphist, "<- done (SUCCESS!)",0,0,0,0);
1779 	return (KERN_SUCCESS);
1780 }
1781 
1782 
1783 /*
1784  * uvm_fault_wire: wire down a range of virtual addresses in a map.
1785  *
1786  * => map may be read-locked by caller, but MUST NOT be write-locked.
1787  * => if map is read-locked, any operations which may cause map to
1788  *	be write-locked in uvm_fault() must be taken care of by
1789  *	the caller.  See uvm_map_pageable().
1790  */
1791 
1792 int
1793 uvm_fault_wire(map, start, end, access_type)
1794 	vm_map_t map;
1795 	vaddr_t start, end;
1796 	vm_prot_t access_type;
1797 {
1798 	vaddr_t va;
1799 	pmap_t  pmap;
1800 	int rv;
1801 
1802 	pmap = vm_map_pmap(map);
1803 
1804 	/*
1805 	 * now fault it in a page at a time.   if the fault fails then we have
1806 	 * to undo what we have done.   note that in uvm_fault VM_PROT_NONE
1807 	 * is replaced with the max protection if fault_type is VM_FAULT_WIRE.
1808 	 */
1809 
1810 	for (va = start ; va < end ; va += PAGE_SIZE) {
1811 		rv = uvm_fault(map, va, VM_FAULT_WIRE, access_type);
1812 		if (rv) {
1813 			if (va != start) {
1814 				uvm_fault_unwire(map, start, va);
1815 			}
1816 			return (rv);
1817 		}
1818 	}
1819 
1820 	return (KERN_SUCCESS);
1821 }
1822 
1823 /*
1824  * uvm_fault_unwire(): unwire range of virtual space.
1825  */
1826 
1827 void
1828 uvm_fault_unwire(map, start, end)
1829 	vm_map_t map;
1830 	vaddr_t start, end;
1831 {
1832 
1833 	vm_map_lock_read(map);
1834 	uvm_fault_unwire_locked(map, start, end);
1835 	vm_map_unlock_read(map);
1836 }
1837 
1838 /*
1839  * uvm_fault_unwire_locked(): the guts of uvm_fault_unwire().
1840  *
1841  * => map must be at least read-locked.
1842  */
1843 
1844 void
1845 uvm_fault_unwire_locked(map, start, end)
1846 	vm_map_t map;
1847 	vaddr_t start, end;
1848 {
1849 	vm_map_entry_t entry;
1850 	pmap_t pmap = vm_map_pmap(map);
1851 	vaddr_t va;
1852 	paddr_t pa;
1853 	struct vm_page *pg;
1854 
1855 	KASSERT((map->flags & VM_MAP_INTRSAFE) == 0);
1856 
1857 	/*
1858 	 * we assume that the area we are unwiring has actually been wired
1859 	 * in the first place.   this means that we should be able to extract
1860 	 * the PAs from the pmap.   we also lock out the page daemon so that
1861 	 * we can call uvm_pageunwire.
1862 	 */
1863 
1864 	uvm_lock_pageq();
1865 
1866 	/*
1867 	 * find the beginning map entry for the region.
1868 	 */
1869 	KASSERT(start >= vm_map_min(map) && end <= vm_map_max(map));
1870 	if (uvm_map_lookup_entry(map, start, &entry) == FALSE)
1871 		panic("uvm_fault_unwire_locked: address not in map");
1872 
1873 	for (va = start; va < end ; va += PAGE_SIZE) {
1874 		if (pmap_extract(pmap, va, &pa) == FALSE)
1875 			panic("uvm_fault_unwire_locked: unwiring "
1876 			    "non-wired memory");
1877 
1878 		/*
1879 		 * make sure the current entry is for the address we're
1880 		 * dealing with.  if not, grab the next entry.
1881 		 */
1882 
1883 		KASSERT(va >= entry->start);
1884 		if (va >= entry->end) {
1885 			KASSERT(entry->next != &map->header &&
1886 				entry->next->start <= entry->end);
1887 			entry = entry->next;
1888 		}
1889 
1890 		/*
1891 		 * if the entry is no longer wired, tell the pmap.
1892 		 */
1893 		if (VM_MAPENT_ISWIRED(entry) == 0)
1894 			pmap_unwire(pmap, va);
1895 
1896 		pg = PHYS_TO_VM_PAGE(pa);
1897 		if (pg)
1898 			uvm_pageunwire(pg);
1899 	}
1900 
1901 	uvm_unlock_pageq();
1902 }
1903