xref: /netbsd-src/sys/uvm/uvm_fault.c (revision da5f4674a3fc214be3572d358b66af40ab9401e7)
1 /*	$NetBSD: uvm_fault.c,v 1.84 2003/08/11 16:45:43 pk 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 1.84 2003/08/11 16:45:43 pk 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 
54 #include <uvm/uvm.h>
55 
56 /*
57  *
58  * a word on page faults:
59  *
60  * types of page faults we handle:
61  *
62  * CASE 1: upper layer faults                   CASE 2: lower layer faults
63  *
64  *    CASE 1A         CASE 1B                  CASE 2A        CASE 2B
65  *    read/write1     write>1                  read/write   +-cow_write/zero
66  *         |             |                         |        |
67  *      +--|--+       +--|--+     +-----+       +  |  +     | +-----+
68  * amap |  V  |       |  ----------->new|          |        | |  ^  |
69  *      +-----+       +-----+     +-----+       +  |  +     | +--|--+
70  *                                                 |        |    |
71  *      +-----+       +-----+                   +--|--+     | +--|--+
72  * uobj | d/c |       | d/c |                   |  V  |     +----|  |
73  *      +-----+       +-----+                   +-----+       +-----+
74  *
75  * d/c = don't care
76  *
77  *   case [0]: layerless fault
78  *	no amap or uobj is present.   this is an error.
79  *
80  *   case [1]: upper layer fault [anon active]
81  *     1A: [read] or [write with anon->an_ref == 1]
82  *		I/O takes place in top level anon and uobj is not touched.
83  *     1B: [write with anon->an_ref > 1]
84  *		new anon is alloc'd and data is copied off ["COW"]
85  *
86  *   case [2]: lower layer fault [uobj]
87  *     2A: [read on non-NULL uobj] or [write to non-copy_on_write area]
88  *		I/O takes place directly in object.
89  *     2B: [write to copy_on_write] or [read on NULL uobj]
90  *		data is "promoted" from uobj to a new anon.
91  *		if uobj is null, then we zero fill.
92  *
93  * we follow the standard UVM locking protocol ordering:
94  *
95  * MAPS => AMAP => UOBJ => ANON => PAGE QUEUES (PQ)
96  * we hold a PG_BUSY page if we unlock for I/O
97  *
98  *
99  * the code is structured as follows:
100  *
101  *     - init the "IN" params in the ufi structure
102  *   ReFault:
103  *     - do lookups [locks maps], check protection, handle needs_copy
104  *     - check for case 0 fault (error)
105  *     - establish "range" of fault
106  *     - if we have an amap lock it and extract the anons
107  *     - if sequential advice deactivate pages behind us
108  *     - at the same time check pmap for unmapped areas and anon for pages
109  *	 that we could map in (and do map it if found)
110  *     - check object for resident pages that we could map in
111  *     - if (case 2) goto Case2
112  *     - >>> handle case 1
113  *           - ensure source anon is resident in RAM
114  *           - if case 1B alloc new anon and copy from source
115  *           - map the correct page in
116  *   Case2:
117  *     - >>> handle case 2
118  *           - ensure source page is resident (if uobj)
119  *           - if case 2B alloc new anon and copy from source (could be zero
120  *		fill if uobj == NULL)
121  *           - map the correct page in
122  *     - done!
123  *
124  * note on paging:
125  *   if we have to do I/O we place a PG_BUSY page in the correct object,
126  * unlock everything, and do the I/O.   when I/O is done we must reverify
127  * the state of the world before assuming that our data structures are
128  * valid.   [because mappings could change while the map is unlocked]
129  *
130  *  alternative 1: unbusy the page in question and restart the page fault
131  *    from the top (ReFault).   this is easy but does not take advantage
132  *    of the information that we already have from our previous lookup,
133  *    although it is possible that the "hints" in the vm_map will help here.
134  *
135  * alternative 2: the system already keeps track of a "version" number of
136  *    a map.   [i.e. every time you write-lock a map (e.g. to change a
137  *    mapping) you bump the version number up by one...]   so, we can save
138  *    the version number of the map before we release the lock and start I/O.
139  *    then when I/O is done we can relock and check the version numbers
140  *    to see if anything changed.    this might save us some over 1 because
141  *    we don't have to unbusy the page and may be less compares(?).
142  *
143  * alternative 3: put in backpointers or a way to "hold" part of a map
144  *    in place while I/O is in progress.   this could be complex to
145  *    implement (especially with structures like amap that can be referenced
146  *    by multiple map entries, and figuring out what should wait could be
147  *    complex as well...).
148  *
149  * given that we are not currently multiprocessor or multithreaded we might
150  * as well choose alternative 2 now.   maybe alternative 3 would be useful
151  * in the future.    XXX keep in mind for future consideration//rechecking.
152  */
153 
154 /*
155  * local data structures
156  */
157 
158 struct uvm_advice {
159 	int advice;
160 	int nback;
161 	int nforw;
162 };
163 
164 /*
165  * page range array:
166  * note: index in array must match "advice" value
167  * XXX: borrowed numbers from freebsd.   do they work well for us?
168  */
169 
170 static struct uvm_advice uvmadvice[] = {
171 	{ MADV_NORMAL, 3, 4 },
172 	{ MADV_RANDOM, 0, 0 },
173 	{ MADV_SEQUENTIAL, 8, 7},
174 };
175 
176 #define UVM_MAXRANGE 16	/* must be MAX() of nback+nforw+1 */
177 
178 /*
179  * private prototypes
180  */
181 
182 static void uvmfault_amapcopy __P((struct uvm_faultinfo *));
183 static __inline void uvmfault_anonflush __P((struct vm_anon **, int));
184 
185 /*
186  * inline functions
187  */
188 
189 /*
190  * uvmfault_anonflush: try and deactivate pages in specified anons
191  *
192  * => does not have to deactivate page if it is busy
193  */
194 
195 static __inline void
196 uvmfault_anonflush(anons, n)
197 	struct vm_anon **anons;
198 	int n;
199 {
200 	int lcv;
201 	struct vm_page *pg;
202 
203 	for (lcv = 0 ; lcv < n ; lcv++) {
204 		if (anons[lcv] == NULL)
205 			continue;
206 		simple_lock(&anons[lcv]->an_lock);
207 		pg = anons[lcv]->u.an_page;
208 		if (pg && (pg->flags & PG_BUSY) == 0 && pg->loan_count == 0) {
209 			uvm_lock_pageq();
210 			if (pg->wire_count == 0) {
211 				pmap_clear_reference(pg);
212 				uvm_pagedeactivate(pg);
213 			}
214 			uvm_unlock_pageq();
215 		}
216 		simple_unlock(&anons[lcv]->an_lock);
217 	}
218 }
219 
220 /*
221  * normal functions
222  */
223 
224 /*
225  * uvmfault_amapcopy: clear "needs_copy" in a map.
226  *
227  * => called with VM data structures unlocked (usually, see below)
228  * => we get a write lock on the maps and clear needs_copy for a VA
229  * => if we are out of RAM we sleep (waiting for more)
230  */
231 
232 static void
233 uvmfault_amapcopy(ufi)
234 	struct uvm_faultinfo *ufi;
235 {
236 	for (;;) {
237 
238 		/*
239 		 * no mapping?  give up.
240 		 */
241 
242 		if (uvmfault_lookup(ufi, TRUE) == FALSE)
243 			return;
244 
245 		/*
246 		 * copy if needed.
247 		 */
248 
249 		if (UVM_ET_ISNEEDSCOPY(ufi->entry))
250 			amap_copy(ufi->map, ufi->entry, M_NOWAIT, TRUE,
251 				ufi->orig_rvaddr, ufi->orig_rvaddr + 1);
252 
253 		/*
254 		 * didn't work?  must be out of RAM.   unlock and sleep.
255 		 */
256 
257 		if (UVM_ET_ISNEEDSCOPY(ufi->entry)) {
258 			uvmfault_unlockmaps(ufi, TRUE);
259 			uvm_wait("fltamapcopy");
260 			continue;
261 		}
262 
263 		/*
264 		 * got it!   unlock and return.
265 		 */
266 
267 		uvmfault_unlockmaps(ufi, TRUE);
268 		return;
269 	}
270 	/*NOTREACHED*/
271 }
272 
273 /*
274  * uvmfault_anonget: get data in an anon into a non-busy, non-released
275  * page in that anon.
276  *
277  * => maps, amap, and anon locked by caller.
278  * => if we fail (result != 0) we unlock everything.
279  * => if we are successful, we return with everything still locked.
280  * => we don't move the page on the queues [gets moved later]
281  * => if we allocate a new page [we_own], it gets put on the queues.
282  *    either way, the result is that the page is on the queues at return time
283  * => for pages which are on loan from a uvm_object (and thus are not
284  *    owned by the anon): if successful, we return with the owning object
285  *    locked.   the caller must unlock this object when it unlocks everything
286  *    else.
287  */
288 
289 int
290 uvmfault_anonget(ufi, amap, anon)
291 	struct uvm_faultinfo *ufi;
292 	struct vm_amap *amap;
293 	struct vm_anon *anon;
294 {
295 	boolean_t we_own;	/* we own anon's page? */
296 	boolean_t locked;	/* did we relock? */
297 	struct vm_page *pg;
298 	int error;
299 	UVMHIST_FUNC("uvmfault_anonget"); UVMHIST_CALLED(maphist);
300 
301 	LOCK_ASSERT(simple_lock_held(&anon->an_lock));
302 
303 	error = 0;
304 	uvmexp.fltanget++;
305         /* bump rusage counters */
306 	if (anon->u.an_page)
307 		curproc->p_stats->p_ru.ru_minflt++;
308 	else
309 		curproc->p_stats->p_ru.ru_majflt++;
310 
311 	/*
312 	 * loop until we get it, or fail.
313 	 */
314 
315 	for (;;) {
316 		we_own = FALSE;		/* TRUE if we set PG_BUSY on a page */
317 		pg = anon->u.an_page;
318 
319 		/*
320 		 * if there is a resident page and it is loaned, then anon
321 		 * may not own it.   call out to uvm_anon_lockpage() to ensure
322 		 * the real owner of the page has been identified and locked.
323 		 */
324 
325 		if (pg && pg->loan_count)
326 			pg = uvm_anon_lockloanpg(anon);
327 
328 		/*
329 		 * page there?   make sure it is not busy/released.
330 		 */
331 
332 		if (pg) {
333 
334 			/*
335 			 * at this point, if the page has a uobject [meaning
336 			 * we have it on loan], then that uobject is locked
337 			 * by us!   if the page is busy, we drop all the
338 			 * locks (including uobject) and try again.
339 			 */
340 
341 			if ((pg->flags & PG_BUSY) == 0) {
342 				UVMHIST_LOG(maphist, "<- OK",0,0,0,0);
343 				return (0);
344 			}
345 			pg->flags |= PG_WANTED;
346 			uvmexp.fltpgwait++;
347 
348 			/*
349 			 * the last unlock must be an atomic unlock+wait on
350 			 * the owner of page
351 			 */
352 
353 			if (pg->uobject) {	/* owner is uobject ? */
354 				uvmfault_unlockall(ufi, amap, NULL, anon);
355 				UVMHIST_LOG(maphist, " unlock+wait on uobj",0,
356 				    0,0,0);
357 				UVM_UNLOCK_AND_WAIT(pg,
358 				    &pg->uobject->vmobjlock,
359 				    FALSE, "anonget1",0);
360 			} else {
361 				/* anon owns page */
362 				uvmfault_unlockall(ufi, amap, NULL, NULL);
363 				UVMHIST_LOG(maphist, " unlock+wait on anon",0,
364 				    0,0,0);
365 				UVM_UNLOCK_AND_WAIT(pg,&anon->an_lock,0,
366 				    "anonget2",0);
367 			}
368 		} else {
369 
370 			/*
371 			 * no page, we must try and bring it in.
372 			 */
373 
374 			pg = uvm_pagealloc(NULL, 0, anon, 0);
375 			if (pg == NULL) {		/* out of RAM.  */
376 				uvmfault_unlockall(ufi, amap, NULL, anon);
377 				uvmexp.fltnoram++;
378 				UVMHIST_LOG(maphist, "  noram -- UVM_WAIT",0,
379 				    0,0,0);
380 				uvm_wait("flt_noram1");
381 			} else {
382 				/* we set the PG_BUSY bit */
383 				we_own = TRUE;
384 				uvmfault_unlockall(ufi, amap, NULL, anon);
385 
386 				/*
387 				 * we are passing a PG_BUSY+PG_FAKE+PG_CLEAN
388 				 * page into the uvm_swap_get function with
389 				 * all data structures unlocked.  note that
390 				 * it is ok to read an_swslot here because
391 				 * we hold PG_BUSY on the page.
392 				 */
393 				uvmexp.pageins++;
394 				error = uvm_swap_get(pg, anon->an_swslot,
395 				    PGO_SYNCIO);
396 
397 				/*
398 				 * we clean up after the i/o below in the
399 				 * "we_own" case
400 				 */
401 			}
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 (pg->flags & PG_WANTED) {
428 				wakeup(pg);
429 			}
430 			if (error) {
431 				/* remove page from anon */
432 				anon->u.an_page = NULL;
433 
434 				/*
435 				 * remove the swap slot from the anon
436 				 * and mark the anon as having no real slot.
437 				 * don't free the swap slot, thus preventing
438 				 * it from being used again.
439 				 */
440 
441 				if (anon->an_swslot > 0)
442 					uvm_swap_markbad(anon->an_swslot, 1);
443 				anon->an_swslot = SWSLOT_BAD;
444 
445 				/*
446 				 * note: page was never !PG_BUSY, so it
447 				 * can't be mapped and thus no need to
448 				 * pmap_page_protect it...
449 				 */
450 
451 				uvm_lock_pageq();
452 				uvm_pagefree(pg);
453 				uvm_unlock_pageq();
454 
455 				if (locked)
456 					uvmfault_unlockall(ufi, amap, NULL,
457 					    anon);
458 				else
459 					simple_unlock(&anon->an_lock);
460 				UVMHIST_LOG(maphist, "<- ERROR", 0,0,0,0);
461 				return error;
462 			}
463 
464 			/*
465 			 * we've successfully read the page, activate it.
466 			 */
467 
468 			uvm_lock_pageq();
469 			uvm_pageactivate(pg);
470 			uvm_unlock_pageq();
471 			pg->flags &= ~(PG_WANTED|PG_BUSY|PG_FAKE);
472 			UVM_PAGE_OWN(pg, NULL);
473 			if (!locked)
474 				simple_unlock(&anon->an_lock);
475 		}
476 
477 		/*
478 		 * we were not able to relock.   restart fault.
479 		 */
480 
481 		if (!locked) {
482 			UVMHIST_LOG(maphist, "<- REFAULT", 0,0,0,0);
483 			return (ERESTART);
484 		}
485 
486 		/*
487 		 * verify no one has touched the amap and moved the anon on us.
488 		 */
489 
490 		if (ufi != NULL &&
491 		    amap_lookup(&ufi->entry->aref,
492 				ufi->orig_rvaddr - ufi->entry->start) != anon) {
493 
494 			uvmfault_unlockall(ufi, amap, NULL, anon);
495 			UVMHIST_LOG(maphist, "<- REFAULT", 0,0,0,0);
496 			return (ERESTART);
497 		}
498 
499 		/*
500 		 * try it again!
501 		 */
502 
503 		uvmexp.fltanretry++;
504 		continue;
505 	}
506 	/*NOTREACHED*/
507 }
508 
509 /*
510  *   F A U L T   -   m a i n   e n t r y   p o i n t
511  */
512 
513 /*
514  * uvm_fault: page fault handler
515  *
516  * => called from MD code to resolve a page fault
517  * => VM data structures usually should be unlocked.   however, it is
518  *	possible to call here with the main map locked if the caller
519  *	gets a write lock, sets it recusive, and then calls us (c.f.
520  *	uvm_map_pageable).   this should be avoided because it keeps
521  *	the map locked off during I/O.
522  * => MUST NEVER BE CALLED IN INTERRUPT CONTEXT
523  */
524 
525 #define MASK(entry)     (UVM_ET_ISCOPYONWRITE(entry) ? \
526 			 ~VM_PROT_WRITE : VM_PROT_ALL)
527 
528 int
529 uvm_fault(orig_map, vaddr, fault_type, access_type)
530 	struct vm_map *orig_map;
531 	vaddr_t vaddr;
532 	vm_fault_t fault_type;
533 	vm_prot_t access_type;
534 {
535 	struct uvm_faultinfo ufi;
536 	vm_prot_t enter_prot, check_prot;
537 	boolean_t wired, narrow, promote, locked, shadowed, wire_fault, cow_now;
538 	int npages, nback, nforw, centeridx, error, lcv, gotpages;
539 	vaddr_t startva, objaddr, currva;
540 	voff_t uoff;
541 	paddr_t pa;
542 	struct vm_amap *amap;
543 	struct uvm_object *uobj;
544 	struct vm_anon *anons_store[UVM_MAXRANGE], **anons, *anon, *oanon;
545 	struct vm_page *pages[UVM_MAXRANGE], *pg, *uobjpage;
546 	UVMHIST_FUNC("uvm_fault"); UVMHIST_CALLED(maphist);
547 
548 	UVMHIST_LOG(maphist, "(map=0x%x, vaddr=0x%x, ft=%d, at=%d)",
549 	      orig_map, vaddr, fault_type, access_type);
550 
551 	anon = NULL;
552 	pg = NULL;
553 
554 	uvmexp.faults++;	/* XXX: locking? */
555 
556 	/*
557 	 * init the IN parameters in the ufi
558 	 */
559 
560 	ufi.orig_map = orig_map;
561 	ufi.orig_rvaddr = trunc_page(vaddr);
562 	ufi.orig_size = PAGE_SIZE;	/* can't get any smaller than this */
563 	wire_fault = fault_type == VM_FAULT_WIRE ||
564 	    fault_type == VM_FAULT_WIREMAX;
565 	if (wire_fault)
566 		narrow = TRUE;		/* don't look for neighborhood
567 					 * pages on wire */
568 	else
569 		narrow = FALSE;		/* normal fault */
570 
571 	/*
572 	 * "goto ReFault" means restart the page fault from ground zero.
573 	 */
574 ReFault:
575 
576 	/*
577 	 * lookup and lock the maps
578 	 */
579 
580 	if (uvmfault_lookup(&ufi, FALSE) == FALSE) {
581 		UVMHIST_LOG(maphist, "<- no mapping @ 0x%x", vaddr, 0,0,0);
582 		return (EFAULT);
583 	}
584 	/* locked: maps(read) */
585 
586 #ifdef DIAGNOSTIC
587 	if ((ufi.map->flags & VM_MAP_PAGEABLE) == 0) {
588 		printf("Page fault on non-pageable map:\n");
589 		printf("ufi.map = %p\n", ufi.map);
590 		printf("ufi.orig_map = %p\n", ufi.orig_map);
591 		printf("ufi.orig_rvaddr = 0x%lx\n", (u_long) ufi.orig_rvaddr);
592 		panic("uvm_fault: (ufi.map->flags & VM_MAP_PAGEABLE) == 0");
593 	}
594 #endif
595 
596 	/*
597 	 * check protection
598 	 */
599 
600 	check_prot = fault_type == VM_FAULT_WIREMAX ?
601 	    ufi.entry->max_protection : ufi.entry->protection;
602 	if ((check_prot & access_type) != access_type) {
603 		UVMHIST_LOG(maphist,
604 		    "<- protection failure (prot=0x%x, access=0x%x)",
605 		    ufi.entry->protection, access_type, 0, 0);
606 		uvmfault_unlockmaps(&ufi, FALSE);
607 		return EACCES;
608 	}
609 
610 	/*
611 	 * "enter_prot" is the protection we want to enter the page in at.
612 	 * for certain pages (e.g. copy-on-write pages) this protection can
613 	 * be more strict than ufi.entry->protection.  "wired" means either
614 	 * the entry is wired or we are fault-wiring the pg.
615 	 */
616 
617 	enter_prot = ufi.entry->protection;
618 	wired = VM_MAPENT_ISWIRED(ufi.entry) || wire_fault;
619 	if (wired) {
620 		access_type = enter_prot; /* full access for wired */
621 		cow_now = (check_prot & VM_PROT_WRITE) != 0;
622 	} else {
623 		cow_now = (access_type & VM_PROT_WRITE) != 0;
624 	}
625 
626 	/*
627 	 * handle "needs_copy" case.   if we need to copy the amap we will
628 	 * have to drop our readlock and relock it with a write lock.  (we
629 	 * need a write lock to change anything in a map entry [e.g.
630 	 * needs_copy]).
631 	 */
632 
633 	if (UVM_ET_ISNEEDSCOPY(ufi.entry)) {
634 		KASSERT(fault_type != VM_FAULT_WIREMAX);
635 		if (cow_now || (ufi.entry->object.uvm_obj == NULL)) {
636 			/* need to clear */
637 			UVMHIST_LOG(maphist,
638 			    "  need to clear needs_copy and refault",0,0,0,0);
639 			uvmfault_unlockmaps(&ufi, FALSE);
640 			uvmfault_amapcopy(&ufi);
641 			uvmexp.fltamcopy++;
642 			goto ReFault;
643 
644 		} else {
645 
646 			/*
647 			 * ensure that we pmap_enter page R/O since
648 			 * needs_copy is still true
649 			 */
650 
651 			enter_prot &= ~VM_PROT_WRITE;
652 		}
653 	}
654 
655 	/*
656 	 * identify the players
657 	 */
658 
659 	amap = ufi.entry->aref.ar_amap;		/* top layer */
660 	uobj = ufi.entry->object.uvm_obj;	/* bottom layer */
661 
662 	/*
663 	 * check for a case 0 fault.  if nothing backing the entry then
664 	 * error now.
665 	 */
666 
667 	if (amap == NULL && uobj == NULL) {
668 		uvmfault_unlockmaps(&ufi, FALSE);
669 		UVMHIST_LOG(maphist,"<- no backing store, no overlay",0,0,0,0);
670 		return (EFAULT);
671 	}
672 
673 	/*
674 	 * establish range of interest based on advice from mapper
675 	 * and then clip to fit map entry.   note that we only want
676 	 * to do this the first time through the fault.   if we
677 	 * ReFault we will disable this by setting "narrow" to true.
678 	 */
679 
680 	if (narrow == FALSE) {
681 
682 		/* wide fault (!narrow) */
683 		KASSERT(uvmadvice[ufi.entry->advice].advice ==
684 			 ufi.entry->advice);
685 		nback = MIN(uvmadvice[ufi.entry->advice].nback,
686 			    (ufi.orig_rvaddr - ufi.entry->start) >> PAGE_SHIFT);
687 		startva = ufi.orig_rvaddr - (nback << PAGE_SHIFT);
688 		nforw = MIN(uvmadvice[ufi.entry->advice].nforw,
689 			    ((ufi.entry->end - ufi.orig_rvaddr) >>
690 			     PAGE_SHIFT) - 1);
691 		/*
692 		 * note: "-1" because we don't want to count the
693 		 * faulting page as forw
694 		 */
695 		npages = nback + nforw + 1;
696 		centeridx = nback;
697 
698 		narrow = TRUE;	/* ensure only once per-fault */
699 
700 	} else {
701 
702 		/* narrow fault! */
703 		nback = nforw = 0;
704 		startva = ufi.orig_rvaddr;
705 		npages = 1;
706 		centeridx = 0;
707 
708 	}
709 
710 	/* locked: maps(read) */
711 	UVMHIST_LOG(maphist, "  narrow=%d, back=%d, forw=%d, startva=0x%x",
712 		    narrow, nback, nforw, startva);
713 	UVMHIST_LOG(maphist, "  entry=0x%x, amap=0x%x, obj=0x%x", ufi.entry,
714 		    amap, uobj, 0);
715 
716 	/*
717 	 * if we've got an amap, lock it and extract current anons.
718 	 */
719 
720 	if (amap) {
721 		amap_lock(amap);
722 		anons = anons_store;
723 		amap_lookups(&ufi.entry->aref, startva - ufi.entry->start,
724 		    anons, npages);
725 	} else {
726 		anons = NULL;	/* to be safe */
727 	}
728 
729 	/* locked: maps(read), amap(if there) */
730 
731 	/*
732 	 * for MADV_SEQUENTIAL mappings we want to deactivate the back pages
733 	 * now and then forget about them (for the rest of the fault).
734 	 */
735 
736 	if (ufi.entry->advice == MADV_SEQUENTIAL && nback != 0) {
737 
738 		UVMHIST_LOG(maphist, "  MADV_SEQUENTIAL: flushing backpages",
739 		    0,0,0,0);
740 		/* flush back-page anons? */
741 		if (amap)
742 			uvmfault_anonflush(anons, nback);
743 
744 		/* flush object? */
745 		if (uobj) {
746 			objaddr =
747 			    (startva - ufi.entry->start) + ufi.entry->offset;
748 			simple_lock(&uobj->vmobjlock);
749 			(void) (uobj->pgops->pgo_put)(uobj, objaddr, objaddr +
750 				    (nback << PAGE_SHIFT), PGO_DEACTIVATE);
751 		}
752 
753 		/* now forget about the backpages */
754 		if (amap)
755 			anons += nback;
756 		startva += (nback << PAGE_SHIFT);
757 		npages -= nback;
758 		nback = centeridx = 0;
759 	}
760 
761 	/* locked: maps(read), amap(if there) */
762 
763 	/*
764 	 * map in the backpages and frontpages we found in the amap in hopes
765 	 * of preventing future faults.    we also init the pages[] array as
766 	 * we go.
767 	 */
768 
769 	currva = startva;
770 	shadowed = FALSE;
771 	for (lcv = 0 ; lcv < npages ; lcv++, currva += PAGE_SIZE) {
772 
773 		/*
774 		 * dont play with VAs that are already mapped
775 		 * except for center)
776 		 */
777 		if (lcv != centeridx &&
778 		    pmap_extract(ufi.orig_map->pmap, currva, &pa)) {
779 			pages[lcv] = PGO_DONTCARE;
780 			continue;
781 		}
782 
783 		/*
784 		 * unmapped or center page.   check if any anon at this level.
785 		 */
786 		if (amap == NULL || anons[lcv] == NULL) {
787 			pages[lcv] = NULL;
788 			continue;
789 		}
790 
791 		/*
792 		 * check for present page and map if possible.   re-activate it.
793 		 */
794 
795 		pages[lcv] = PGO_DONTCARE;
796 		if (lcv == centeridx) {		/* save center for later! */
797 			shadowed = TRUE;
798 			continue;
799 		}
800 		anon = anons[lcv];
801 		simple_lock(&anon->an_lock);
802 		/* ignore loaned pages */
803 		if (anon->u.an_page && anon->u.an_page->loan_count == 0 &&
804 		    (anon->u.an_page->flags & PG_BUSY) == 0) {
805 			uvm_lock_pageq();
806 			uvm_pageactivate(anon->u.an_page);
807 			uvm_unlock_pageq();
808 			UVMHIST_LOG(maphist,
809 			    "  MAPPING: n anon: pm=0x%x, va=0x%x, pg=0x%x",
810 			    ufi.orig_map->pmap, currva, anon->u.an_page, 0);
811 			uvmexp.fltnamap++;
812 
813 			/*
814 			 * Since this isn't the page that's actually faulting,
815 			 * ignore pmap_enter() failures; it's not critical
816 			 * that we enter these right now.
817 			 */
818 
819 			(void) pmap_enter(ufi.orig_map->pmap, currva,
820 			    VM_PAGE_TO_PHYS(anon->u.an_page),
821 			    (anon->an_ref > 1) ? (enter_prot & ~VM_PROT_WRITE) :
822 			    enter_prot,
823 			    PMAP_CANFAIL |
824 			     (VM_MAPENT_ISWIRED(ufi.entry) ? PMAP_WIRED : 0));
825 		}
826 		simple_unlock(&anon->an_lock);
827 		pmap_update(ufi.orig_map->pmap);
828 	}
829 
830 	/* locked: maps(read), amap(if there) */
831 	/* (shadowed == TRUE) if there is an anon at the faulting address */
832 	UVMHIST_LOG(maphist, "  shadowed=%d, will_get=%d", shadowed,
833 	    (uobj && shadowed == FALSE),0,0);
834 
835 	/*
836 	 * note that if we are really short of RAM we could sleep in the above
837 	 * call to pmap_enter with everything locked.   bad?
838 	 *
839 	 * XXX Actually, that is bad; pmap_enter() should just fail in that
840 	 * XXX case.  --thorpej
841 	 */
842 
843 	/*
844 	 * if the desired page is not shadowed by the amap and we have a
845 	 * backing object, then we check to see if the backing object would
846 	 * prefer to handle the fault itself (rather than letting us do it
847 	 * with the usual pgo_get hook).  the backing object signals this by
848 	 * providing a pgo_fault routine.
849 	 */
850 
851 	if (uobj && shadowed == FALSE && uobj->pgops->pgo_fault != NULL) {
852 		simple_lock(&uobj->vmobjlock);
853 
854 		/* locked: maps(read), amap (if there), uobj */
855 		error = uobj->pgops->pgo_fault(&ufi, startva, pages, npages,
856 		    centeridx, fault_type, access_type, PGO_LOCKED|PGO_SYNCIO);
857 
858 		/* locked: nothing, pgo_fault has unlocked everything */
859 
860 		if (error == ERESTART)
861 			goto ReFault;		/* try again! */
862 		/*
863 		 * object fault routine responsible for pmap_update().
864 		 */
865 		return error;
866 	}
867 
868 	/*
869 	 * now, if the desired page is not shadowed by the amap and we have
870 	 * a backing object that does not have a special fault routine, then
871 	 * we ask (with pgo_get) the object for resident pages that we care
872 	 * about and attempt to map them in.  we do not let pgo_get block
873 	 * (PGO_LOCKED).
874 	 */
875 
876 	if (uobj && shadowed == FALSE) {
877 		simple_lock(&uobj->vmobjlock);
878 
879 		/* locked (!shadowed): maps(read), amap (if there), uobj */
880 		/*
881 		 * the following call to pgo_get does _not_ change locking state
882 		 */
883 
884 		uvmexp.fltlget++;
885 		gotpages = npages;
886 		(void) uobj->pgops->pgo_get(uobj, ufi.entry->offset +
887 				(startva - ufi.entry->start),
888 				pages, &gotpages, centeridx,
889 				access_type & MASK(ufi.entry),
890 				ufi.entry->advice, PGO_LOCKED);
891 
892 		/*
893 		 * check for pages to map, if we got any
894 		 */
895 
896 		uobjpage = NULL;
897 
898 		if (gotpages) {
899 			currva = startva;
900 			for (lcv = 0; lcv < npages;
901 			     lcv++, currva += PAGE_SIZE) {
902 				if (pages[lcv] == NULL ||
903 				    pages[lcv] == PGO_DONTCARE) {
904 					continue;
905 				}
906 
907 				/*
908 				 * if center page is resident and not
909 				 * PG_BUSY|PG_RELEASED then pgo_get
910 				 * made it PG_BUSY for us and gave
911 				 * us a handle to it.   remember this
912 				 * page as "uobjpage." (for later use).
913 				 */
914 
915 				if (lcv == centeridx) {
916 					uobjpage = pages[lcv];
917 					UVMHIST_LOG(maphist, "  got uobjpage "
918 					    "(0x%x) with locked get",
919 					    uobjpage, 0,0,0);
920 					continue;
921 				}
922 
923 				/*
924 				 * calling pgo_get with PGO_LOCKED returns us
925 				 * pages which are neither busy nor released,
926 				 * so we don't need to check for this.
927 				 * we can just directly enter the pages.
928 				 */
929 
930 				uvm_lock_pageq();
931 				uvm_pageactivate(pages[lcv]);
932 				uvm_unlock_pageq();
933 				UVMHIST_LOG(maphist,
934 				  "  MAPPING: n obj: pm=0x%x, va=0x%x, pg=0x%x",
935 				  ufi.orig_map->pmap, currva, pages[lcv], 0);
936 				uvmexp.fltnomap++;
937 
938 				/*
939 				 * Since this page isn't the page that's
940 				 * actually fauling, ignore pmap_enter()
941 				 * failures; it's not critical that we
942 				 * enter these right now.
943 				 */
944 
945 				(void) pmap_enter(ufi.orig_map->pmap, currva,
946 				    VM_PAGE_TO_PHYS(pages[lcv]),
947 				    pages[lcv]->flags & PG_RDONLY ?
948 				    enter_prot & ~VM_PROT_WRITE :
949 				    enter_prot & MASK(ufi.entry),
950 				    PMAP_CANFAIL |
951 				     (wired ? PMAP_WIRED : 0));
952 
953 				/*
954 				 * NOTE: page can't be PG_WANTED or PG_RELEASED
955 				 * because we've held the lock the whole time
956 				 * we've had the handle.
957 				 */
958 
959 				pages[lcv]->flags &= ~(PG_BUSY);
960 				UVM_PAGE_OWN(pages[lcv], NULL);
961 			}
962 			pmap_update(ufi.orig_map->pmap);
963 		}
964 	} else {
965 		uobjpage = NULL;
966 	}
967 
968 	/* locked (shadowed): maps(read), amap */
969 	/* locked (!shadowed): maps(read), amap(if there),
970 		 uobj(if !null), uobjpage(if !null) */
971 
972 	/*
973 	 * note that at this point we are done with any front or back pages.
974 	 * we are now going to focus on the center page (i.e. the one we've
975 	 * faulted on).  if we have faulted on the top (anon) layer
976 	 * [i.e. case 1], then the anon we want is anons[centeridx] (we have
977 	 * not touched it yet).  if we have faulted on the bottom (uobj)
978 	 * layer [i.e. case 2] and the page was both present and available,
979 	 * then we've got a pointer to it as "uobjpage" and we've already
980 	 * made it BUSY.
981 	 */
982 
983 	/*
984 	 * there are four possible cases we must address: 1A, 1B, 2A, and 2B
985 	 */
986 
987 	/*
988 	 * redirect case 2: if we are not shadowed, go to case 2.
989 	 */
990 
991 	if (shadowed == FALSE)
992 		goto Case2;
993 
994 	/* locked: maps(read), amap */
995 
996 	/*
997 	 * handle case 1: fault on an anon in our amap
998 	 */
999 
1000 	anon = anons[centeridx];
1001 	UVMHIST_LOG(maphist, "  case 1 fault: anon=0x%x", anon, 0,0,0);
1002 	simple_lock(&anon->an_lock);
1003 
1004 	/* locked: maps(read), amap, anon */
1005 
1006 	/*
1007 	 * no matter if we have case 1A or case 1B we are going to need to
1008 	 * have the anon's memory resident.   ensure that now.
1009 	 */
1010 
1011 	/*
1012 	 * let uvmfault_anonget do the dirty work.
1013 	 * if it fails (!OK) it will unlock everything for us.
1014 	 * if it succeeds, locks are still valid and locked.
1015 	 * also, if it is OK, then the anon's page is on the queues.
1016 	 * if the page is on loan from a uvm_object, then anonget will
1017 	 * lock that object for us if it does not fail.
1018 	 */
1019 
1020 	error = uvmfault_anonget(&ufi, amap, anon);
1021 	switch (error) {
1022 	case 0:
1023 		break;
1024 
1025 	case ERESTART:
1026 		goto ReFault;
1027 
1028 	case EAGAIN:
1029 		tsleep(&lbolt, PVM, "fltagain1", 0);
1030 		goto ReFault;
1031 
1032 	default:
1033 		return error;
1034 	}
1035 
1036 	/*
1037 	 * uobj is non null if the page is on loan from an object (i.e. uobj)
1038 	 */
1039 
1040 	uobj = anon->u.an_page->uobject;	/* locked by anonget if !NULL */
1041 
1042 	/* locked: maps(read), amap, anon, uobj(if one) */
1043 
1044 	/*
1045 	 * special handling for loaned pages
1046 	 */
1047 
1048 	if (anon->u.an_page->loan_count) {
1049 
1050 		if (!cow_now) {
1051 
1052 			/*
1053 			 * for read faults on loaned pages we just cap the
1054 			 * protection at read-only.
1055 			 */
1056 
1057 			enter_prot = enter_prot & ~VM_PROT_WRITE;
1058 
1059 		} else {
1060 			/*
1061 			 * note that we can't allow writes into a loaned page!
1062 			 *
1063 			 * if we have a write fault on a loaned page in an
1064 			 * anon then we need to look at the anon's ref count.
1065 			 * if it is greater than one then we are going to do
1066 			 * a normal copy-on-write fault into a new anon (this
1067 			 * is not a problem).  however, if the reference count
1068 			 * is one (a case where we would normally allow a
1069 			 * write directly to the page) then we need to kill
1070 			 * the loan before we continue.
1071 			 */
1072 
1073 			/* >1 case is already ok */
1074 			if (anon->an_ref == 1) {
1075 
1076 				/* get new un-owned replacement page */
1077 				pg = uvm_pagealloc(NULL, 0, NULL, 0);
1078 				if (pg == NULL) {
1079 					uvmfault_unlockall(&ufi, amap, uobj,
1080 					    anon);
1081 					uvm_wait("flt_noram2");
1082 					goto ReFault;
1083 				}
1084 
1085 				/*
1086 				 * copy data, kill loan, and drop uobj lock
1087 				 * (if any)
1088 				 */
1089 				/* copy old -> new */
1090 				uvm_pagecopy(anon->u.an_page, pg);
1091 
1092 				/* force reload */
1093 				pmap_page_protect(anon->u.an_page,
1094 						  VM_PROT_NONE);
1095 				uvm_lock_pageq();	  /* KILL loan */
1096 
1097 				anon->u.an_page->uanon = NULL;
1098 				/* in case we owned */
1099 				anon->u.an_page->pqflags &= ~PQ_ANON;
1100 
1101 				if (uobj) {
1102 					/* if we were receiver of loan */
1103 					anon->u.an_page->loan_count--;
1104 				} else {
1105 					/*
1106 					 * we were the lender (A->K); need
1107 					 * to remove the page from pageq's.
1108 					 */
1109 					uvm_pagedequeue(anon->u.an_page);
1110 				}
1111 
1112 				uvm_pageactivate(pg);
1113 				uvm_unlock_pageq();
1114 				if (uobj) {
1115 					simple_unlock(&uobj->vmobjlock);
1116 					uobj = NULL;
1117 				}
1118 
1119 				/* install new page in anon */
1120 				anon->u.an_page = pg;
1121 				pg->uanon = anon;
1122 				pg->pqflags |= PQ_ANON;
1123 				pg->flags &= ~(PG_BUSY|PG_FAKE);
1124 				UVM_PAGE_OWN(pg, NULL);
1125 
1126 				/* done! */
1127 			}     /* ref == 1 */
1128 		}       /* write fault */
1129 	}         /* loan count */
1130 
1131 	/*
1132 	 * if we are case 1B then we will need to allocate a new blank
1133 	 * anon to transfer the data into.   note that we have a lock
1134 	 * on anon, so no one can busy or release the page until we are done.
1135 	 * also note that the ref count can't drop to zero here because
1136 	 * it is > 1 and we are only dropping one ref.
1137 	 *
1138 	 * in the (hopefully very rare) case that we are out of RAM we
1139 	 * will unlock, wait for more RAM, and refault.
1140 	 *
1141 	 * if we are out of anon VM we kill the process (XXX: could wait?).
1142 	 */
1143 
1144 	if (cow_now && anon->an_ref > 1) {
1145 
1146 		UVMHIST_LOG(maphist, "  case 1B: COW fault",0,0,0,0);
1147 		uvmexp.flt_acow++;
1148 		oanon = anon;		/* oanon = old, locked anon */
1149 		anon = uvm_analloc();
1150 		if (anon) {
1151 			/* new anon is locked! */
1152 			pg = uvm_pagealloc(NULL, 0, anon, 0);
1153 		}
1154 
1155 		/* check for out of RAM */
1156 		if (anon == NULL || pg == NULL) {
1157 			if (anon) {
1158 				anon->an_ref--;
1159 				simple_unlock(&anon->an_lock);
1160 				uvm_anfree(anon);
1161 			}
1162 			uvmfault_unlockall(&ufi, amap, uobj, oanon);
1163 			if (anon == NULL || uvm_swapisfull()) {
1164 				UVMHIST_LOG(maphist,
1165 				    "<- failed.  out of VM",0,0,0,0);
1166 				uvmexp.fltnoanon++;
1167 				return ENOMEM;
1168 			}
1169 
1170 			uvmexp.fltnoram++;
1171 			uvm_wait("flt_noram3");	/* out of RAM, wait for more */
1172 			goto ReFault;
1173 		}
1174 
1175 		/* got all resources, replace anon with nanon */
1176 		uvm_pagecopy(oanon->u.an_page, pg);
1177 		uvm_lock_pageq();
1178 		uvm_pageactivate(pg);
1179 		pg->flags &= ~(PG_BUSY|PG_FAKE);
1180 		uvm_unlock_pageq();
1181 		UVM_PAGE_OWN(pg, NULL);
1182 		amap_add(&ufi.entry->aref, ufi.orig_rvaddr - ufi.entry->start,
1183 		    anon, 1);
1184 
1185 		/* deref: can not drop to zero here by defn! */
1186 		oanon->an_ref--;
1187 
1188 		/*
1189 		 * note: oanon is still locked, as is the new anon.  we
1190 		 * need to check for this later when we unlock oanon; if
1191 		 * oanon != anon, we'll have to unlock anon, too.
1192 		 */
1193 
1194 	} else {
1195 
1196 		uvmexp.flt_anon++;
1197 		oanon = anon;		/* old, locked anon is same as anon */
1198 		pg = anon->u.an_page;
1199 		if (anon->an_ref > 1)     /* disallow writes to ref > 1 anons */
1200 			enter_prot = enter_prot & ~VM_PROT_WRITE;
1201 
1202 	}
1203 
1204 	/* locked: maps(read), amap, oanon, anon (if different from oanon) */
1205 
1206 	/*
1207 	 * now map the page in.
1208 	 */
1209 
1210 	UVMHIST_LOG(maphist, "  MAPPING: anon: pm=0x%x, va=0x%x, pg=0x%x",
1211 	    ufi.orig_map->pmap, ufi.orig_rvaddr, pg, 0);
1212 	if (pmap_enter(ufi.orig_map->pmap, ufi.orig_rvaddr, VM_PAGE_TO_PHYS(pg),
1213 	    enter_prot, access_type | PMAP_CANFAIL | (wired ? PMAP_WIRED : 0))
1214 	    != 0) {
1215 
1216 		/*
1217 		 * No need to undo what we did; we can simply think of
1218 		 * this as the pmap throwing away the mapping information.
1219 		 *
1220 		 * We do, however, have to go through the ReFault path,
1221 		 * as the map may change while we're asleep.
1222 		 */
1223 
1224 		if (anon != oanon)
1225 			simple_unlock(&anon->an_lock);
1226 		uvmfault_unlockall(&ufi, amap, uobj, oanon);
1227 		if (uvm_swapisfull()) {
1228 			UVMHIST_LOG(maphist,
1229 			    "<- failed.  out of VM",0,0,0,0);
1230 			/* XXX instrumentation */
1231 			return ENOMEM;
1232 		}
1233 		/* XXX instrumentation */
1234 		uvm_wait("flt_pmfail1");
1235 		goto ReFault;
1236 	}
1237 
1238 	/*
1239 	 * ... update the page queues.
1240 	 */
1241 
1242 	uvm_lock_pageq();
1243 	if (wire_fault) {
1244 		uvm_pagewire(pg);
1245 
1246 		/*
1247 		 * since the now-wired page cannot be paged out,
1248 		 * release its swap resources for others to use.
1249 		 * since an anon with no swap cannot be PG_CLEAN,
1250 		 * clear its clean flag now.
1251 		 */
1252 
1253 		pg->flags &= ~(PG_CLEAN);
1254 		uvm_anon_dropswap(anon);
1255 	} else {
1256 		uvm_pageactivate(pg);
1257 	}
1258 	uvm_unlock_pageq();
1259 
1260 	/*
1261 	 * done case 1!  finish up by unlocking everything and returning success
1262 	 */
1263 
1264 	if (anon != oanon)
1265 		simple_unlock(&anon->an_lock);
1266 	uvmfault_unlockall(&ufi, amap, uobj, oanon);
1267 	pmap_update(ufi.orig_map->pmap);
1268 	return 0;
1269 
1270 Case2:
1271 	/*
1272 	 * handle case 2: faulting on backing object or zero fill
1273 	 */
1274 
1275 	/*
1276 	 * locked:
1277 	 * maps(read), amap(if there), uobj(if !null), uobjpage(if !null)
1278 	 */
1279 
1280 	/*
1281 	 * note that uobjpage can not be PGO_DONTCARE at this point.  we now
1282 	 * set uobjpage to PGO_DONTCARE if we are doing a zero fill.  if we
1283 	 * have a backing object, check and see if we are going to promote
1284 	 * the data up to an anon during the fault.
1285 	 */
1286 
1287 	if (uobj == NULL) {
1288 		uobjpage = PGO_DONTCARE;
1289 		promote = TRUE;		/* always need anon here */
1290 	} else {
1291 		KASSERT(uobjpage != PGO_DONTCARE);
1292 		promote = cow_now && UVM_ET_ISCOPYONWRITE(ufi.entry);
1293 	}
1294 	UVMHIST_LOG(maphist, "  case 2 fault: promote=%d, zfill=%d",
1295 	    promote, (uobj == NULL), 0,0);
1296 
1297 	/*
1298 	 * if uobjpage is not null then we do not need to do I/O to get the
1299 	 * uobjpage.
1300 	 *
1301 	 * if uobjpage is null, then we need to unlock and ask the pager to
1302 	 * get the data for us.   once we have the data, we need to reverify
1303 	 * the state the world.   we are currently not holding any resources.
1304 	 */
1305 
1306 	if (uobjpage) {
1307 		/* update rusage counters */
1308 		curproc->p_stats->p_ru.ru_minflt++;
1309 	} else {
1310 		/* update rusage counters */
1311 		curproc->p_stats->p_ru.ru_majflt++;
1312 
1313 		/* locked: maps(read), amap(if there), uobj */
1314 		uvmfault_unlockall(&ufi, amap, NULL, NULL);
1315 		/* locked: uobj */
1316 
1317 		uvmexp.fltget++;
1318 		gotpages = 1;
1319 		uoff = (ufi.orig_rvaddr - ufi.entry->start) + ufi.entry->offset;
1320 		error = uobj->pgops->pgo_get(uobj, uoff, &uobjpage, &gotpages,
1321 		    0, access_type & MASK(ufi.entry), ufi.entry->advice,
1322 		    PGO_SYNCIO);
1323 		/* locked: uobjpage(if no error) */
1324 
1325 		/*
1326 		 * recover from I/O
1327 		 */
1328 
1329 		if (error) {
1330 			if (error == EAGAIN) {
1331 				UVMHIST_LOG(maphist,
1332 				    "  pgo_get says TRY AGAIN!",0,0,0,0);
1333 				tsleep(&lbolt, PVM, "fltagain2", 0);
1334 				goto ReFault;
1335 			}
1336 
1337 			UVMHIST_LOG(maphist, "<- pgo_get failed (code %d)",
1338 			    error, 0,0,0);
1339 			return error;
1340 		}
1341 
1342 		/* locked: uobjpage */
1343 
1344 		uvm_lock_pageq();
1345 		uvm_pageactivate(uobjpage);
1346 		uvm_unlock_pageq();
1347 
1348 		/*
1349 		 * re-verify the state of the world by first trying to relock
1350 		 * the maps.  always relock the object.
1351 		 */
1352 
1353 		locked = uvmfault_relock(&ufi);
1354 		if (locked && amap)
1355 			amap_lock(amap);
1356 		simple_lock(&uobj->vmobjlock);
1357 
1358 		/* locked(locked): maps(read), amap(if !null), uobj, uobjpage */
1359 		/* locked(!locked): uobj, uobjpage */
1360 
1361 		/*
1362 		 * verify that the page has not be released and re-verify
1363 		 * that amap slot is still free.   if there is a problem,
1364 		 * we unlock and clean up.
1365 		 */
1366 
1367 		if ((uobjpage->flags & PG_RELEASED) != 0 ||
1368 		    (locked && amap &&
1369 		    amap_lookup(&ufi.entry->aref,
1370 		      ufi.orig_rvaddr - ufi.entry->start))) {
1371 			if (locked)
1372 				uvmfault_unlockall(&ufi, amap, NULL, NULL);
1373 			locked = FALSE;
1374 		}
1375 
1376 		/*
1377 		 * didn't get the lock?   release the page and retry.
1378 		 */
1379 
1380 		if (locked == FALSE) {
1381 			UVMHIST_LOG(maphist,
1382 			    "  wasn't able to relock after fault: retry",
1383 			    0,0,0,0);
1384 			if (uobjpage->flags & PG_WANTED)
1385 				wakeup(uobjpage);
1386 			if (uobjpage->flags & PG_RELEASED) {
1387 				uvmexp.fltpgrele++;
1388 				uvm_pagefree(uobjpage);
1389 				goto ReFault;
1390 			}
1391 			uobjpage->flags &= ~(PG_BUSY|PG_WANTED);
1392 			UVM_PAGE_OWN(uobjpage, NULL);
1393 			simple_unlock(&uobj->vmobjlock);
1394 			goto ReFault;
1395 		}
1396 
1397 		/*
1398 		 * we have the data in uobjpage which is busy and
1399 		 * not released.  we are holding object lock (so the page
1400 		 * can't be released on us).
1401 		 */
1402 
1403 		/* locked: maps(read), amap(if !null), uobj, uobjpage */
1404 	}
1405 
1406 	/*
1407 	 * locked:
1408 	 * maps(read), amap(if !null), uobj(if !null), uobjpage(if uobj)
1409 	 */
1410 
1411 	/*
1412 	 * notes:
1413 	 *  - at this point uobjpage can not be NULL
1414 	 *  - at this point uobjpage can not be PG_RELEASED (since we checked
1415 	 *  for it above)
1416 	 *  - at this point uobjpage could be PG_WANTED (handle later)
1417 	 */
1418 
1419 	if (promote == FALSE) {
1420 
1421 		/*
1422 		 * we are not promoting.   if the mapping is COW ensure that we
1423 		 * don't give more access than we should (e.g. when doing a read
1424 		 * fault on a COPYONWRITE mapping we want to map the COW page in
1425 		 * R/O even though the entry protection could be R/W).
1426 		 *
1427 		 * set "pg" to the page we want to map in (uobjpage, usually)
1428 		 */
1429 
1430 		/* no anon in this case. */
1431 		anon = NULL;
1432 
1433 		uvmexp.flt_obj++;
1434 		if (UVM_ET_ISCOPYONWRITE(ufi.entry))
1435 			enter_prot &= ~VM_PROT_WRITE;
1436 		pg = uobjpage;		/* map in the actual object */
1437 
1438 		/* assert(uobjpage != PGO_DONTCARE) */
1439 
1440 		/*
1441 		 * we are faulting directly on the page.   be careful
1442 		 * about writing to loaned pages...
1443 		 */
1444 
1445 		if (uobjpage->loan_count) {
1446 			if (!cow_now) {
1447 				/* read fault: cap the protection at readonly */
1448 				/* cap! */
1449 				enter_prot = enter_prot & ~VM_PROT_WRITE;
1450 			} else {
1451 				/* write fault: must break the loan here */
1452 
1453 				pg = uvm_loanbreak(uobjpage);
1454 				if (pg == NULL) {
1455 
1456 					/*
1457 					 * drop ownership of page, it can't
1458 					 * be released
1459 					 */
1460 
1461 					if (uobjpage->flags & PG_WANTED)
1462 						wakeup(uobjpage);
1463 					uobjpage->flags &= ~(PG_BUSY|PG_WANTED);
1464 					UVM_PAGE_OWN(uobjpage, NULL);
1465 
1466 					uvmfault_unlockall(&ufi, amap, uobj,
1467 					  NULL);
1468 					UVMHIST_LOG(maphist,
1469 					  "  out of RAM breaking loan, waiting",
1470 					  0,0,0,0);
1471 					uvmexp.fltnoram++;
1472 					uvm_wait("flt_noram4");
1473 					goto ReFault;
1474 				}
1475 				uobjpage = pg;
1476 			}
1477 		}
1478 	} else {
1479 
1480 		/*
1481 		 * if we are going to promote the data to an anon we
1482 		 * allocate a blank anon here and plug it into our amap.
1483 		 */
1484 #if DIAGNOSTIC
1485 		if (amap == NULL)
1486 			panic("uvm_fault: want to promote data, but no anon");
1487 #endif
1488 
1489 		anon = uvm_analloc();
1490 		if (anon) {
1491 
1492 			/*
1493 			 * The new anon is locked.
1494 			 *
1495 			 * In `Fill in data...' below, if
1496 			 * uobjpage == PGO_DONTCARE, we want
1497 			 * a zero'd, dirty page, so have
1498 			 * uvm_pagealloc() do that for us.
1499 			 */
1500 
1501 			pg = uvm_pagealloc(NULL, 0, anon,
1502 			    (uobjpage == PGO_DONTCARE) ? UVM_PGA_ZERO : 0);
1503 		}
1504 
1505 		/*
1506 		 * out of memory resources?
1507 		 */
1508 
1509 		if (anon == NULL || pg == NULL) {
1510 			if (anon != NULL) {
1511 				anon->an_ref--;
1512 				simple_unlock(&anon->an_lock);
1513 				uvm_anfree(anon);
1514 			}
1515 
1516 			/*
1517 			 * arg!  must unbusy our page and fail or sleep.
1518 			 */
1519 
1520 			if (uobjpage != PGO_DONTCARE) {
1521 				if (uobjpage->flags & PG_WANTED)
1522 					/* still holding object lock */
1523 					wakeup(uobjpage);
1524 
1525 				uobjpage->flags &= ~(PG_BUSY|PG_WANTED);
1526 				UVM_PAGE_OWN(uobjpage, NULL);
1527 			}
1528 
1529 			/* unlock and fail ... */
1530 			uvmfault_unlockall(&ufi, amap, uobj, NULL);
1531 			if (anon == NULL || uvm_swapisfull()) {
1532 				UVMHIST_LOG(maphist, "  promote: out of VM",
1533 				    0,0,0,0);
1534 				uvmexp.fltnoanon++;
1535 				return ENOMEM;
1536 			}
1537 
1538 			UVMHIST_LOG(maphist, "  out of RAM, waiting for more",
1539 			    0,0,0,0);
1540 			uvmexp.fltnoram++;
1541 			uvm_wait("flt_noram5");
1542 			goto ReFault;
1543 		}
1544 
1545 		/*
1546 		 * fill in the data
1547 		 */
1548 
1549 		if (uobjpage != PGO_DONTCARE) {
1550 			uvmexp.flt_prcopy++;
1551 			/* copy page [pg now dirty] */
1552 			uvm_pagecopy(uobjpage, pg);
1553 
1554 			/*
1555 			 * promote to shared amap?  make sure all sharing
1556 			 * procs see it
1557 			 */
1558 
1559 			if ((amap_flags(amap) & AMAP_SHARED) != 0) {
1560 				pmap_page_protect(uobjpage, VM_PROT_NONE);
1561 				/*
1562 				 * XXX: PAGE MIGHT BE WIRED!
1563 				 */
1564 			}
1565 
1566 			/*
1567 			 * dispose of uobjpage.  it can't be PG_RELEASED
1568 			 * since we still hold the object lock.
1569 			 * drop handle to uobj as well.
1570 			 */
1571 
1572 			if (uobjpage->flags & PG_WANTED)
1573 				/* still have the obj lock */
1574 				wakeup(uobjpage);
1575 			uobjpage->flags &= ~(PG_BUSY|PG_WANTED);
1576 			UVM_PAGE_OWN(uobjpage, NULL);
1577 			simple_unlock(&uobj->vmobjlock);
1578 			uobj = NULL;
1579 
1580 			UVMHIST_LOG(maphist,
1581 			    "  promote uobjpage 0x%x to anon/page 0x%x/0x%x",
1582 			    uobjpage, anon, pg, 0);
1583 
1584 		} else {
1585 			uvmexp.flt_przero++;
1586 
1587 			/*
1588 			 * Page is zero'd and marked dirty by uvm_pagealloc()
1589 			 * above.
1590 			 */
1591 
1592 			UVMHIST_LOG(maphist,"  zero fill anon/page 0x%x/0%x",
1593 			    anon, pg, 0, 0);
1594 		}
1595 		amap_add(&ufi.entry->aref, ufi.orig_rvaddr - ufi.entry->start,
1596 		    anon, 0);
1597 	}
1598 
1599 	/*
1600 	 * locked:
1601 	 * maps(read), amap(if !null), uobj(if !null), uobjpage(if uobj),
1602 	 *   anon(if !null), pg(if anon)
1603 	 *
1604 	 * note: pg is either the uobjpage or the new page in the new anon
1605 	 */
1606 
1607 	/*
1608 	 * all resources are present.   we can now map it in and free our
1609 	 * resources.
1610 	 */
1611 
1612 	UVMHIST_LOG(maphist,
1613 	    "  MAPPING: case2: pm=0x%x, va=0x%x, pg=0x%x, promote=%d",
1614 	    ufi.orig_map->pmap, ufi.orig_rvaddr, pg, promote);
1615 	KASSERT((access_type & VM_PROT_WRITE) == 0 ||
1616 		(pg->flags & PG_RDONLY) == 0);
1617 	if (pmap_enter(ufi.orig_map->pmap, ufi.orig_rvaddr, VM_PAGE_TO_PHYS(pg),
1618 	    pg->flags & PG_RDONLY ? enter_prot & ~VM_PROT_WRITE : enter_prot,
1619 	    access_type | PMAP_CANFAIL | (wired ? PMAP_WIRED : 0)) != 0) {
1620 
1621 		/*
1622 		 * No need to undo what we did; we can simply think of
1623 		 * this as the pmap throwing away the mapping information.
1624 		 *
1625 		 * We do, however, have to go through the ReFault path,
1626 		 * as the map may change while we're asleep.
1627 		 */
1628 
1629 		if (pg->flags & PG_WANTED)
1630 			wakeup(pg);
1631 
1632 		/*
1633 		 * note that pg can't be PG_RELEASED since we did not drop
1634 		 * the object lock since the last time we checked.
1635 		 */
1636 
1637 		pg->flags &= ~(PG_BUSY|PG_FAKE|PG_WANTED);
1638 		UVM_PAGE_OWN(pg, NULL);
1639 		uvmfault_unlockall(&ufi, amap, uobj, anon);
1640 		if (uvm_swapisfull()) {
1641 			UVMHIST_LOG(maphist,
1642 			    "<- failed.  out of VM",0,0,0,0);
1643 			/* XXX instrumentation */
1644 			return ENOMEM;
1645 		}
1646 		/* XXX instrumentation */
1647 		uvm_wait("flt_pmfail2");
1648 		goto ReFault;
1649 	}
1650 
1651 	uvm_lock_pageq();
1652 	if (wire_fault) {
1653 		uvm_pagewire(pg);
1654 		if (pg->pqflags & PQ_AOBJ) {
1655 
1656 			/*
1657 			 * since the now-wired page cannot be paged out,
1658 			 * release its swap resources for others to use.
1659 			 * since an aobj page with no swap cannot be PG_CLEAN,
1660 			 * clear its clean flag now.
1661 			 */
1662 
1663 			pg->flags &= ~(PG_CLEAN);
1664 			uao_dropswap(uobj, pg->offset >> PAGE_SHIFT);
1665 		}
1666 	} else {
1667 		uvm_pageactivate(pg);
1668 	}
1669 	uvm_unlock_pageq();
1670 	if (pg->flags & PG_WANTED)
1671 		wakeup(pg);
1672 
1673 	/*
1674 	 * note that pg can't be PG_RELEASED since we did not drop the object
1675 	 * lock since the last time we checked.
1676 	 */
1677 
1678 	pg->flags &= ~(PG_BUSY|PG_FAKE|PG_WANTED);
1679 	UVM_PAGE_OWN(pg, NULL);
1680 	uvmfault_unlockall(&ufi, amap, uobj, anon);
1681 	pmap_update(ufi.orig_map->pmap);
1682 	UVMHIST_LOG(maphist, "<- done (SUCCESS!)",0,0,0,0);
1683 	return 0;
1684 }
1685 
1686 /*
1687  * uvm_fault_wire: wire down a range of virtual addresses in a map.
1688  *
1689  * => map may be read-locked by caller, but MUST NOT be write-locked.
1690  * => if map is read-locked, any operations which may cause map to
1691  *	be write-locked in uvm_fault() must be taken care of by
1692  *	the caller.  See uvm_map_pageable().
1693  */
1694 
1695 int
1696 uvm_fault_wire(map, start, end, fault_type, access_type)
1697 	struct vm_map *map;
1698 	vaddr_t start, end;
1699 	vm_fault_t fault_type;
1700 	vm_prot_t access_type;
1701 {
1702 	vaddr_t va;
1703 	int error;
1704 
1705 	/*
1706 	 * now fault it in a page at a time.   if the fault fails then we have
1707 	 * to undo what we have done.   note that in uvm_fault VM_PROT_NONE
1708 	 * is replaced with the max protection if fault_type is VM_FAULT_WIRE.
1709 	 */
1710 
1711 	/*
1712 	 * XXX work around overflowing a vaddr_t.  this prevents us from
1713 	 * wiring the last page in the address space, though.
1714 	 */
1715 	if (start > end) {
1716 		return EFAULT;
1717 	}
1718 
1719 	for (va = start ; va < end ; va += PAGE_SIZE) {
1720 		error = uvm_fault(map, va, fault_type, access_type);
1721 		if (error) {
1722 			if (va != start) {
1723 				uvm_fault_unwire(map, start, va);
1724 			}
1725 			return error;
1726 		}
1727 	}
1728 	return 0;
1729 }
1730 
1731 /*
1732  * uvm_fault_unwire(): unwire range of virtual space.
1733  */
1734 
1735 void
1736 uvm_fault_unwire(map, start, end)
1737 	struct vm_map *map;
1738 	vaddr_t start, end;
1739 {
1740 	vm_map_lock_read(map);
1741 	uvm_fault_unwire_locked(map, start, end);
1742 	vm_map_unlock_read(map);
1743 }
1744 
1745 /*
1746  * uvm_fault_unwire_locked(): the guts of uvm_fault_unwire().
1747  *
1748  * => map must be at least read-locked.
1749  */
1750 
1751 void
1752 uvm_fault_unwire_locked(map, start, end)
1753 	struct vm_map *map;
1754 	vaddr_t start, end;
1755 {
1756 	struct vm_map_entry *entry;
1757 	pmap_t pmap = vm_map_pmap(map);
1758 	vaddr_t va;
1759 	paddr_t pa;
1760 	struct vm_page *pg;
1761 
1762 	KASSERT((map->flags & VM_MAP_INTRSAFE) == 0);
1763 
1764 	/*
1765 	 * we assume that the area we are unwiring has actually been wired
1766 	 * in the first place.   this means that we should be able to extract
1767 	 * the PAs from the pmap.   we also lock out the page daemon so that
1768 	 * we can call uvm_pageunwire.
1769 	 */
1770 
1771 	uvm_lock_pageq();
1772 
1773 	/*
1774 	 * find the beginning map entry for the region.
1775 	 */
1776 
1777 	KASSERT(start >= vm_map_min(map) && end <= vm_map_max(map));
1778 	if (uvm_map_lookup_entry(map, start, &entry) == FALSE)
1779 		panic("uvm_fault_unwire_locked: address not in map");
1780 
1781 	for (va = start; va < end; va += PAGE_SIZE) {
1782 		if (pmap_extract(pmap, va, &pa) == FALSE)
1783 			continue;
1784 
1785 		/*
1786 		 * find the map entry for the current address.
1787 		 */
1788 
1789 		KASSERT(va >= entry->start);
1790 		while (va >= entry->end) {
1791 			KASSERT(entry->next != &map->header &&
1792 				entry->next->start <= entry->end);
1793 			entry = entry->next;
1794 		}
1795 
1796 		/*
1797 		 * if the entry is no longer wired, tell the pmap.
1798 		 */
1799 
1800 		if (VM_MAPENT_ISWIRED(entry) == 0)
1801 			pmap_unwire(pmap, va);
1802 
1803 		pg = PHYS_TO_VM_PAGE(pa);
1804 		if (pg)
1805 			uvm_pageunwire(pg);
1806 	}
1807 
1808 	uvm_unlock_pageq();
1809 }
1810