xref: /netbsd-src/sys/uvm/uvm_fault.c (revision eb961d0e02b7a46a9acfa877b02df48df6637278)
1 /*	$NetBSD: uvm_fault.c,v 1.109 2006/02/22 22:28:18 drochner 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 1.109 2006/02/22 22:28:18 drochner Exp $");
43 
44 #include "opt_uvmhist.h"
45 
46 #include <sys/param.h>
47 #include <sys/systm.h>
48 #include <sys/kernel.h>
49 #include <sys/proc.h>
50 #include <sys/malloc.h>
51 #include <sys/mman.h>
52 #include <sys/user.h>
53 #include <sys/vnode.h>
54 
55 #include <uvm/uvm.h>
56 
57 /*
58  *
59  * a word on page faults:
60  *
61  * types of page faults we handle:
62  *
63  * CASE 1: upper layer faults                   CASE 2: lower layer faults
64  *
65  *    CASE 1A         CASE 1B                  CASE 2A        CASE 2B
66  *    read/write1     write>1                  read/write   +-cow_write/zero
67  *         |             |                         |        |
68  *      +--|--+       +--|--+     +-----+       +  |  +     | +-----+
69  * amap |  V  |       |  ----------->new|          |        | |  ^  |
70  *      +-----+       +-----+     +-----+       +  |  +     | +--|--+
71  *                                                 |        |    |
72  *      +-----+       +-----+                   +--|--+     | +--|--+
73  * uobj | d/c |       | d/c |                   |  V  |     +----|  |
74  *      +-----+       +-----+                   +-----+       +-----+
75  *
76  * d/c = don't care
77  *
78  *   case [0]: layerless fault
79  *	no amap or uobj is present.   this is an error.
80  *
81  *   case [1]: upper layer fault [anon active]
82  *     1A: [read] or [write with anon->an_ref == 1]
83  *		I/O takes place in top level anon and uobj is not touched.
84  *     1B: [write with anon->an_ref > 1]
85  *		new anon is alloc'd and data is copied off ["COW"]
86  *
87  *   case [2]: lower layer fault [uobj]
88  *     2A: [read on non-NULL uobj] or [write to non-copy_on_write area]
89  *		I/O takes place directly in object.
90  *     2B: [write to copy_on_write] or [read on NULL uobj]
91  *		data is "promoted" from uobj to a new anon.
92  *		if uobj is null, then we zero fill.
93  *
94  * we follow the standard UVM locking protocol ordering:
95  *
96  * MAPS => AMAP => UOBJ => ANON => PAGE QUEUES (PQ)
97  * we hold a PG_BUSY page if we unlock for I/O
98  *
99  *
100  * the code is structured as follows:
101  *
102  *     - init the "IN" params in the ufi structure
103  *   ReFault:
104  *     - do lookups [locks maps], check protection, handle needs_copy
105  *     - check for case 0 fault (error)
106  *     - establish "range" of fault
107  *     - if we have an amap lock it and extract the anons
108  *     - if sequential advice deactivate pages behind us
109  *     - at the same time check pmap for unmapped areas and anon for pages
110  *	 that we could map in (and do map it if found)
111  *     - check object for resident pages that we could map in
112  *     - if (case 2) goto Case2
113  *     - >>> handle case 1
114  *           - ensure source anon is resident in RAM
115  *           - if case 1B alloc new anon and copy from source
116  *           - map the correct page in
117  *   Case2:
118  *     - >>> handle case 2
119  *           - ensure source page is resident (if uobj)
120  *           - if case 2B alloc new anon and copy from source (could be zero
121  *		fill if uobj == NULL)
122  *           - map the correct page in
123  *     - done!
124  *
125  * note on paging:
126  *   if we have to do I/O we place a PG_BUSY page in the correct object,
127  * unlock everything, and do the I/O.   when I/O is done we must reverify
128  * the state of the world before assuming that our data structures are
129  * valid.   [because mappings could change while the map is unlocked]
130  *
131  *  alternative 1: unbusy the page in question and restart the page fault
132  *    from the top (ReFault).   this is easy but does not take advantage
133  *    of the information that we already have from our previous lookup,
134  *    although it is possible that the "hints" in the vm_map will help here.
135  *
136  * alternative 2: the system already keeps track of a "version" number of
137  *    a map.   [i.e. every time you write-lock a map (e.g. to change a
138  *    mapping) you bump the version number up by one...]   so, we can save
139  *    the version number of the map before we release the lock and start I/O.
140  *    then when I/O is done we can relock and check the version numbers
141  *    to see if anything changed.    this might save us some over 1 because
142  *    we don't have to unbusy the page and may be less compares(?).
143  *
144  * alternative 3: put in backpointers or a way to "hold" part of a map
145  *    in place while I/O is in progress.   this could be complex to
146  *    implement (especially with structures like amap that can be referenced
147  *    by multiple map entries, and figuring out what should wait could be
148  *    complex as well...).
149  *
150  * given that we are not currently multiprocessor or multithreaded we might
151  * as well choose alternative 2 now.   maybe alternative 3 would be useful
152  * in the future.    XXX keep in mind for future consideration//rechecking.
153  */
154 
155 /*
156  * local data structures
157  */
158 
159 struct uvm_advice {
160 	int advice;
161 	int nback;
162 	int nforw;
163 };
164 
165 /*
166  * page range array:
167  * note: index in array must match "advice" value
168  * XXX: borrowed numbers from freebsd.   do they work well for us?
169  */
170 
171 static const struct uvm_advice uvmadvice[] = {
172 	{ MADV_NORMAL, 3, 4 },
173 	{ MADV_RANDOM, 0, 0 },
174 	{ MADV_SEQUENTIAL, 8, 7},
175 };
176 
177 #define UVM_MAXRANGE 16	/* must be MAX() of nback+nforw+1 */
178 
179 /*
180  * private prototypes
181  */
182 
183 /*
184  * inline functions
185  */
186 
187 /*
188  * uvmfault_anonflush: try and deactivate pages in specified anons
189  *
190  * => does not have to deactivate page if it is busy
191  */
192 
193 static inline void
194 uvmfault_anonflush(struct vm_anon **anons, int n)
195 {
196 	int lcv;
197 	struct vm_page *pg;
198 
199 	for (lcv = 0 ; lcv < n ; lcv++) {
200 		if (anons[lcv] == NULL)
201 			continue;
202 		simple_lock(&anons[lcv]->an_lock);
203 		pg = anons[lcv]->an_page;
204 		if (pg && (pg->flags & PG_BUSY) == 0 && pg->loan_count == 0) {
205 			uvm_lock_pageq();
206 			if (pg->wire_count == 0) {
207 				pmap_clear_reference(pg);
208 				uvm_pagedeactivate(pg);
209 			}
210 			uvm_unlock_pageq();
211 		}
212 		simple_unlock(&anons[lcv]->an_lock);
213 	}
214 }
215 
216 /*
217  * normal functions
218  */
219 
220 /*
221  * uvmfault_amapcopy: clear "needs_copy" in a map.
222  *
223  * => called with VM data structures unlocked (usually, see below)
224  * => we get a write lock on the maps and clear needs_copy for a VA
225  * => if we are out of RAM we sleep (waiting for more)
226  */
227 
228 static void
229 uvmfault_amapcopy(struct uvm_faultinfo *ufi)
230 {
231 	for (;;) {
232 
233 		/*
234 		 * no mapping?  give up.
235 		 */
236 
237 		if (uvmfault_lookup(ufi, TRUE) == FALSE)
238 			return;
239 
240 		/*
241 		 * copy if needed.
242 		 */
243 
244 		if (UVM_ET_ISNEEDSCOPY(ufi->entry))
245 			amap_copy(ufi->map, ufi->entry, AMAP_COPY_NOWAIT,
246 				ufi->orig_rvaddr, ufi->orig_rvaddr + 1);
247 
248 		/*
249 		 * didn't work?  must be out of RAM.   unlock and sleep.
250 		 */
251 
252 		if (UVM_ET_ISNEEDSCOPY(ufi->entry)) {
253 			uvmfault_unlockmaps(ufi, TRUE);
254 			uvm_wait("fltamapcopy");
255 			continue;
256 		}
257 
258 		/*
259 		 * got it!   unlock and return.
260 		 */
261 
262 		uvmfault_unlockmaps(ufi, TRUE);
263 		return;
264 	}
265 	/*NOTREACHED*/
266 }
267 
268 /*
269  * uvmfault_anonget: get data in an anon into a non-busy, non-released
270  * page in that anon.
271  *
272  * => maps, amap, and anon locked by caller.
273  * => if we fail (result != 0) we unlock everything.
274  * => if we are successful, we return with everything still locked.
275  * => we don't move the page on the queues [gets moved later]
276  * => if we allocate a new page [we_own], it gets put on the queues.
277  *    either way, the result is that the page is on the queues at return time
278  * => for pages which are on loan from a uvm_object (and thus are not
279  *    owned by the anon): if successful, we return with the owning object
280  *    locked.   the caller must unlock this object when it unlocks everything
281  *    else.
282  */
283 
284 int
285 uvmfault_anonget(struct uvm_faultinfo *ufi, struct vm_amap *amap,
286     struct vm_anon *anon)
287 {
288 	boolean_t we_own;	/* we own anon's page? */
289 	boolean_t locked;	/* did we relock? */
290 	struct vm_page *pg;
291 	int error;
292 	UVMHIST_FUNC("uvmfault_anonget"); UVMHIST_CALLED(maphist);
293 
294 	LOCK_ASSERT(simple_lock_held(&anon->an_lock));
295 
296 	error = 0;
297 	uvmexp.fltanget++;
298         /* bump rusage counters */
299 	if (anon->an_page)
300 		curproc->p_stats->p_ru.ru_minflt++;
301 	else
302 		curproc->p_stats->p_ru.ru_majflt++;
303 
304 	/*
305 	 * loop until we get it, or fail.
306 	 */
307 
308 	for (;;) {
309 		we_own = FALSE;		/* TRUE if we set PG_BUSY on a page */
310 		pg = anon->an_page;
311 
312 		/*
313 		 * if there is a resident page and it is loaned, then anon
314 		 * may not own it.   call out to uvm_anon_lockpage() to ensure
315 		 * the real owner of the page has been identified and locked.
316 		 */
317 
318 		if (pg && pg->loan_count)
319 			pg = uvm_anon_lockloanpg(anon);
320 
321 		/*
322 		 * page there?   make sure it is not busy/released.
323 		 */
324 
325 		if (pg) {
326 
327 			/*
328 			 * at this point, if the page has a uobject [meaning
329 			 * we have it on loan], then that uobject is locked
330 			 * by us!   if the page is busy, we drop all the
331 			 * locks (including uobject) and try again.
332 			 */
333 
334 			if ((pg->flags & PG_BUSY) == 0) {
335 				UVMHIST_LOG(maphist, "<- OK",0,0,0,0);
336 				return (0);
337 			}
338 			pg->flags |= PG_WANTED;
339 			uvmexp.fltpgwait++;
340 
341 			/*
342 			 * the last unlock must be an atomic unlock+wait on
343 			 * the owner of page
344 			 */
345 
346 			if (pg->uobject) {	/* owner is uobject ? */
347 				uvmfault_unlockall(ufi, amap, NULL, anon);
348 				UVMHIST_LOG(maphist, " unlock+wait on uobj",0,
349 				    0,0,0);
350 				UVM_UNLOCK_AND_WAIT(pg,
351 				    &pg->uobject->vmobjlock,
352 				    FALSE, "anonget1",0);
353 			} else {
354 				/* anon owns page */
355 				uvmfault_unlockall(ufi, amap, NULL, NULL);
356 				UVMHIST_LOG(maphist, " unlock+wait on anon",0,
357 				    0,0,0);
358 				UVM_UNLOCK_AND_WAIT(pg,&anon->an_lock,0,
359 				    "anonget2",0);
360 			}
361 		} else {
362 #if defined(VMSWAP)
363 
364 			/*
365 			 * no page, we must try and bring it in.
366 			 */
367 
368 			pg = uvm_pagealloc(NULL, 0, anon, 0);
369 			if (pg == NULL) {		/* out of RAM.  */
370 				uvmfault_unlockall(ufi, amap, NULL, anon);
371 				uvmexp.fltnoram++;
372 				UVMHIST_LOG(maphist, "  noram -- UVM_WAIT",0,
373 				    0,0,0);
374 				if (!uvm_reclaimable()) {
375 					return ENOMEM;
376 				}
377 				uvm_wait("flt_noram1");
378 			} else {
379 				/* we set the PG_BUSY bit */
380 				we_own = TRUE;
381 				uvmfault_unlockall(ufi, amap, NULL, anon);
382 
383 				/*
384 				 * we are passing a PG_BUSY+PG_FAKE+PG_CLEAN
385 				 * page into the uvm_swap_get function with
386 				 * all data structures unlocked.  note that
387 				 * it is ok to read an_swslot here because
388 				 * we hold PG_BUSY on the page.
389 				 */
390 				uvmexp.pageins++;
391 				error = uvm_swap_get(pg, anon->an_swslot,
392 				    PGO_SYNCIO);
393 
394 				/*
395 				 * we clean up after the i/o below in the
396 				 * "we_own" case
397 				 */
398 			}
399 #else /* defined(VMSWAP) */
400 			panic("%s: no page", __func__);
401 #endif /* defined(VMSWAP) */
402 		}
403 
404 		/*
405 		 * now relock and try again
406 		 */
407 
408 		locked = uvmfault_relock(ufi);
409 		if (locked && amap != NULL) {
410 			amap_lock(amap);
411 		}
412 		if (locked || we_own)
413 			simple_lock(&anon->an_lock);
414 
415 		/*
416 		 * if we own the page (i.e. we set PG_BUSY), then we need
417 		 * to clean up after the I/O. there are three cases to
418 		 * consider:
419 		 *   [1] page released during I/O: free anon and ReFault.
420 		 *   [2] I/O not OK.   free the page and cause the fault
421 		 *       to fail.
422 		 *   [3] I/O OK!   activate the page and sync with the
423 		 *       non-we_own case (i.e. drop anon lock if not locked).
424 		 */
425 
426 		if (we_own) {
427 #if defined(VMSWAP)
428 			if (pg->flags & PG_WANTED) {
429 				wakeup(pg);
430 			}
431 			if (error) {
432 
433 				/*
434 				 * remove the swap slot from the anon
435 				 * and mark the anon as having no real slot.
436 				 * don't free the swap slot, thus preventing
437 				 * it from being used again.
438 				 */
439 
440 				if (anon->an_swslot > 0)
441 					uvm_swap_markbad(anon->an_swslot, 1);
442 				anon->an_swslot = SWSLOT_BAD;
443 
444 				if ((pg->flags & PG_RELEASED) != 0)
445 					goto released;
446 
447 				/*
448 				 * note: page was never !PG_BUSY, so it
449 				 * can't be mapped and thus no need to
450 				 * pmap_page_protect it...
451 				 */
452 
453 				uvm_lock_pageq();
454 				uvm_pagefree(pg);
455 				uvm_unlock_pageq();
456 
457 				if (locked)
458 					uvmfault_unlockall(ufi, amap, NULL,
459 					    anon);
460 				else
461 					simple_unlock(&anon->an_lock);
462 				UVMHIST_LOG(maphist, "<- ERROR", 0,0,0,0);
463 				return error;
464 			}
465 
466 			if ((pg->flags & PG_RELEASED) != 0) {
467 released:
468 				KASSERT(anon->an_ref == 0);
469 
470 				/*
471 				 * released while we unlocked amap.
472 				 */
473 
474 				if (locked)
475 					uvmfault_unlockall(ufi, amap, NULL,
476 					    NULL);
477 
478 				uvm_anon_release(anon);
479 
480 				if (error) {
481 					UVMHIST_LOG(maphist,
482 					    "<- ERROR/RELEASED", 0,0,0,0);
483 					return error;
484 				}
485 
486 				UVMHIST_LOG(maphist, "<- RELEASED", 0,0,0,0);
487 				return ERESTART;
488 			}
489 
490 			/*
491 			 * we've successfully read the page, activate it.
492 			 */
493 
494 			uvm_lock_pageq();
495 			uvm_pageactivate(pg);
496 			uvm_unlock_pageq();
497 			pg->flags &= ~(PG_WANTED|PG_BUSY|PG_FAKE);
498 			UVM_PAGE_OWN(pg, NULL);
499 			if (!locked)
500 				simple_unlock(&anon->an_lock);
501 #else /* defined(VMSWAP) */
502 			panic("%s: we_own", __func__);
503 #endif /* defined(VMSWAP) */
504 		}
505 
506 		/*
507 		 * we were not able to relock.   restart fault.
508 		 */
509 
510 		if (!locked) {
511 			UVMHIST_LOG(maphist, "<- REFAULT", 0,0,0,0);
512 			return (ERESTART);
513 		}
514 
515 		/*
516 		 * verify no one has touched the amap and moved the anon on us.
517 		 */
518 
519 		if (ufi != NULL &&
520 		    amap_lookup(&ufi->entry->aref,
521 				ufi->orig_rvaddr - ufi->entry->start) != anon) {
522 
523 			uvmfault_unlockall(ufi, amap, NULL, anon);
524 			UVMHIST_LOG(maphist, "<- REFAULT", 0,0,0,0);
525 			return (ERESTART);
526 		}
527 
528 		/*
529 		 * try it again!
530 		 */
531 
532 		uvmexp.fltanretry++;
533 		continue;
534 	}
535 	/*NOTREACHED*/
536 }
537 
538 /*
539  * uvmfault_promote: promote data to a new anon.  used for 1B and 2B.
540  *
541  *	1. allocate an anon and a page.
542  *	2. fill its contents.
543  *	3. put it into amap.
544  *
545  * => if we fail (result != 0) we unlock everything.
546  * => on success, return a new locked anon via 'nanon'.
547  *    (*nanon)->an_page will be a resident, locked, dirty page.
548  */
549 
550 static int
551 uvmfault_promote(struct uvm_faultinfo *ufi,
552     struct vm_anon *oanon,
553     struct vm_page *uobjpage,
554     struct vm_anon **nanon, /* OUT: allocated anon */
555     struct vm_anon **spare)
556 {
557 	struct vm_amap *amap = ufi->entry->aref.ar_amap;
558 	struct uvm_object *uobj;
559 	struct vm_anon *anon;
560 	struct vm_page *pg;
561 	struct vm_page *opg;
562 	int error;
563 	UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist);
564 
565 	if (oanon) {
566 		/* anon COW */
567 		opg = oanon->an_page;
568 		KASSERT(opg != NULL);
569 		KASSERT(opg->uobject == NULL || opg->loan_count > 0);
570 	} else if (uobjpage != PGO_DONTCARE) {
571 		/* object-backed COW */
572 		opg = uobjpage;
573 	} else {
574 		/* ZFOD */
575 		opg = NULL;
576 	}
577 	if (opg != NULL) {
578 		uobj = opg->uobject;
579 	} else {
580 		uobj = NULL;
581 	}
582 
583 	KASSERT(amap != NULL);
584 	KASSERT(uobjpage != NULL);
585 	KASSERT(uobjpage == PGO_DONTCARE || (uobjpage->flags & PG_BUSY) != 0);
586 	LOCK_ASSERT(simple_lock_held(&amap->am_l));
587 	LOCK_ASSERT(oanon == NULL || simple_lock_held(&oanon->an_lock));
588 	LOCK_ASSERT(uobj == NULL || simple_lock_held(&uobj->vmobjlock));
589 	LOCK_ASSERT(*spare == NULL || !simple_lock_held(&(*spare)->an_lock));
590 
591 	if (*spare != NULL) {
592 		anon = *spare;
593 		*spare = NULL;
594 		simple_lock(&anon->an_lock);
595 	} else if (ufi->map != kernel_map) {
596 		anon = uvm_analloc();
597 	} else {
598 		UVMHIST_LOG(maphist, "kernel_map, unlock and retry", 0,0,0,0);
599 
600 		/*
601 		 * we can't allocate anons with kernel_map locked.
602 		 */
603 
604 		uvm_page_unbusy(&uobjpage, 1);
605 		uvmfault_unlockall(ufi, amap, uobj, oanon);
606 
607 		*spare = uvm_analloc();
608 		if (*spare == NULL) {
609 			goto nomem;
610 		}
611 		simple_unlock(&(*spare)->an_lock);
612 		error = ERESTART;
613 		goto done;
614 	}
615 	if (anon) {
616 
617 		/*
618 		 * The new anon is locked.
619 		 *
620 		 * if opg == NULL, we want a zero'd, dirty page,
621 		 * so have uvm_pagealloc() do that for us.
622 		 */
623 
624 		pg = uvm_pagealloc(NULL, 0, anon,
625 		    (opg == NULL) ? UVM_PGA_ZERO : 0);
626 	} else {
627 		pg = NULL;
628 	}
629 
630 	/*
631 	 * out of memory resources?
632 	 */
633 
634 	if (pg == NULL) {
635 		/* save anon for the next try. */
636 		if (anon != NULL) {
637 			simple_unlock(&anon->an_lock);
638 			*spare = anon;
639 		}
640 
641 		/* unlock and fail ... */
642 		uvm_page_unbusy(&uobjpage, 1);
643 		uvmfault_unlockall(ufi, amap, uobj, oanon);
644 nomem:
645 		if (!uvm_reclaimable()) {
646 			UVMHIST_LOG(maphist, "out of VM", 0,0,0,0);
647 			uvmexp.fltnoanon++;
648 			error = ENOMEM;
649 			goto done;
650 		}
651 
652 		UVMHIST_LOG(maphist, "out of RAM, waiting for more", 0,0,0,0);
653 		uvmexp.fltnoram++;
654 		uvm_wait("flt_noram5");
655 		error = ERESTART;
656 		goto done;
657 	}
658 
659 	/* copy page [pg now dirty] */
660 	if (opg) {
661 		uvm_pagecopy(opg, pg);
662 	}
663 
664 	amap_add(&ufi->entry->aref, ufi->orig_rvaddr - ufi->entry->start, anon,
665 	    oanon != NULL);
666 
667 	*nanon = anon;
668 	error = 0;
669 done:
670 	return error;
671 }
672 
673 
674 /*
675  *   F A U L T   -   m a i n   e n t r y   p o i n t
676  */
677 
678 /*
679  * uvm_fault: page fault handler
680  *
681  * => called from MD code to resolve a page fault
682  * => VM data structures usually should be unlocked.   however, it is
683  *	possible to call here with the main map locked if the caller
684  *	gets a write lock, sets it recusive, and then calls us (c.f.
685  *	uvm_map_pageable).   this should be avoided because it keeps
686  *	the map locked off during I/O.
687  * => MUST NEVER BE CALLED IN INTERRUPT CONTEXT
688  */
689 
690 #define MASK(entry)     (UVM_ET_ISCOPYONWRITE(entry) ? \
691 			 ~VM_PROT_WRITE : VM_PROT_ALL)
692 
693 int
694 uvm_fault(struct vm_map *orig_map, vaddr_t vaddr, vm_fault_t fault_type,
695     vm_prot_t access_type)
696 {
697 	struct uvm_faultinfo ufi;
698 	vm_prot_t enter_prot, check_prot;
699 	boolean_t wired, narrow, promote, locked, shadowed, wire_fault, cow_now;
700 	int npages, nback, nforw, centeridx, error, lcv, gotpages;
701 	vaddr_t startva, currva;
702 	voff_t uoff;
703 	struct vm_amap *amap;
704 	struct uvm_object *uobj;
705 	struct vm_anon *anons_store[UVM_MAXRANGE], **anons, *anon, *oanon;
706 	struct vm_anon *anon_spare;
707 	struct vm_page *pages[UVM_MAXRANGE], *pg, *uobjpage;
708 	UVMHIST_FUNC("uvm_fault"); UVMHIST_CALLED(maphist);
709 
710 	UVMHIST_LOG(maphist, "(map=0x%x, vaddr=0x%x, ft=%d, at=%d)",
711 	      orig_map, vaddr, fault_type, access_type);
712 
713 	anon = anon_spare = NULL;
714 	pg = NULL;
715 
716 	uvmexp.faults++;	/* XXX: locking? */
717 
718 	/*
719 	 * init the IN parameters in the ufi
720 	 */
721 
722 	ufi.orig_map = orig_map;
723 	ufi.orig_rvaddr = trunc_page(vaddr);
724 	ufi.orig_size = PAGE_SIZE;	/* can't get any smaller than this */
725 	wire_fault = fault_type == VM_FAULT_WIRE ||
726 	    fault_type == VM_FAULT_WIREMAX;
727 	if (wire_fault)
728 		narrow = TRUE;		/* don't look for neighborhood
729 					 * pages on wire */
730 	else
731 		narrow = FALSE;		/* normal fault */
732 
733 	/*
734 	 * "goto ReFault" means restart the page fault from ground zero.
735 	 */
736 ReFault:
737 
738 	/*
739 	 * lookup and lock the maps
740 	 */
741 
742 	if (uvmfault_lookup(&ufi, FALSE) == FALSE) {
743 		UVMHIST_LOG(maphist, "<- no mapping @ 0x%x", vaddr, 0,0,0);
744 		error = EFAULT;
745 		goto done;
746 	}
747 	/* locked: maps(read) */
748 
749 #ifdef DIAGNOSTIC
750 	if ((ufi.map->flags & VM_MAP_PAGEABLE) == 0) {
751 		printf("Page fault on non-pageable map:\n");
752 		printf("ufi.map = %p\n", ufi.map);
753 		printf("ufi.orig_map = %p\n", ufi.orig_map);
754 		printf("ufi.orig_rvaddr = 0x%lx\n", (u_long) ufi.orig_rvaddr);
755 		panic("uvm_fault: (ufi.map->flags & VM_MAP_PAGEABLE) == 0");
756 	}
757 #endif
758 
759 	/*
760 	 * check protection
761 	 */
762 
763 	check_prot = fault_type == VM_FAULT_WIREMAX ?
764 	    ufi.entry->max_protection : ufi.entry->protection;
765 	if ((check_prot & access_type) != access_type) {
766 		UVMHIST_LOG(maphist,
767 		    "<- protection failure (prot=0x%x, access=0x%x)",
768 		    ufi.entry->protection, access_type, 0, 0);
769 		uvmfault_unlockmaps(&ufi, FALSE);
770 		error = EACCES;
771 		goto done;
772 	}
773 
774 	/*
775 	 * "enter_prot" is the protection we want to enter the page in at.
776 	 * for certain pages (e.g. copy-on-write pages) this protection can
777 	 * be more strict than ufi.entry->protection.  "wired" means either
778 	 * the entry is wired or we are fault-wiring the pg.
779 	 */
780 
781 	enter_prot = ufi.entry->protection;
782 	wired = VM_MAPENT_ISWIRED(ufi.entry) || wire_fault;
783 	if (wired) {
784 		access_type = enter_prot; /* full access for wired */
785 		cow_now = (check_prot & VM_PROT_WRITE) != 0;
786 	} else {
787 		cow_now = (access_type & VM_PROT_WRITE) != 0;
788 	}
789 
790 	/*
791 	 * handle "needs_copy" case.   if we need to copy the amap we will
792 	 * have to drop our readlock and relock it with a write lock.  (we
793 	 * need a write lock to change anything in a map entry [e.g.
794 	 * needs_copy]).
795 	 */
796 
797 	if (UVM_ET_ISNEEDSCOPY(ufi.entry)) {
798 		KASSERT(fault_type != VM_FAULT_WIREMAX);
799 		if (cow_now || (ufi.entry->object.uvm_obj == NULL)) {
800 			/* need to clear */
801 			UVMHIST_LOG(maphist,
802 			    "  need to clear needs_copy and refault",0,0,0,0);
803 			uvmfault_unlockmaps(&ufi, FALSE);
804 			uvmfault_amapcopy(&ufi);
805 			uvmexp.fltamcopy++;
806 			goto ReFault;
807 
808 		} else {
809 
810 			/*
811 			 * ensure that we pmap_enter page R/O since
812 			 * needs_copy is still true
813 			 */
814 
815 			enter_prot &= ~VM_PROT_WRITE;
816 		}
817 	}
818 
819 	/*
820 	 * identify the players
821 	 */
822 
823 	amap = ufi.entry->aref.ar_amap;		/* top layer */
824 	uobj = ufi.entry->object.uvm_obj;	/* bottom layer */
825 
826 	/*
827 	 * check for a case 0 fault.  if nothing backing the entry then
828 	 * error now.
829 	 */
830 
831 	if (amap == NULL && uobj == NULL) {
832 		uvmfault_unlockmaps(&ufi, FALSE);
833 		UVMHIST_LOG(maphist,"<- no backing store, no overlay",0,0,0,0);
834 		error = EFAULT;
835 		goto done;
836 	}
837 
838 	/*
839 	 * establish range of interest based on advice from mapper
840 	 * and then clip to fit map entry.   note that we only want
841 	 * to do this the first time through the fault.   if we
842 	 * ReFault we will disable this by setting "narrow" to true.
843 	 */
844 
845 	if (narrow == FALSE) {
846 
847 		/* wide fault (!narrow) */
848 		KASSERT(uvmadvice[ufi.entry->advice].advice ==
849 			 ufi.entry->advice);
850 		nback = MIN(uvmadvice[ufi.entry->advice].nback,
851 			    (ufi.orig_rvaddr - ufi.entry->start) >> PAGE_SHIFT);
852 		startva = ufi.orig_rvaddr - (nback << PAGE_SHIFT);
853 		nforw = MIN(uvmadvice[ufi.entry->advice].nforw,
854 			    ((ufi.entry->end - ufi.orig_rvaddr) >>
855 			     PAGE_SHIFT) - 1);
856 		/*
857 		 * note: "-1" because we don't want to count the
858 		 * faulting page as forw
859 		 */
860 		npages = nback + nforw + 1;
861 		centeridx = nback;
862 
863 		narrow = TRUE;	/* ensure only once per-fault */
864 
865 	} else {
866 
867 		/* narrow fault! */
868 		nback = nforw = 0;
869 		startva = ufi.orig_rvaddr;
870 		npages = 1;
871 		centeridx = 0;
872 
873 	}
874 
875 	/* locked: maps(read) */
876 	UVMHIST_LOG(maphist, "  narrow=%d, back=%d, forw=%d, startva=0x%x",
877 		    narrow, nback, nforw, startva);
878 	UVMHIST_LOG(maphist, "  entry=0x%x, amap=0x%x, obj=0x%x", ufi.entry,
879 		    amap, uobj, 0);
880 
881 	/*
882 	 * if we've got an amap, lock it and extract current anons.
883 	 */
884 
885 	if (amap) {
886 		amap_lock(amap);
887 		anons = anons_store;
888 		amap_lookups(&ufi.entry->aref, startva - ufi.entry->start,
889 		    anons, npages);
890 	} else {
891 		anons = NULL;	/* to be safe */
892 	}
893 
894 	/* locked: maps(read), amap(if there) */
895 
896 	/*
897 	 * for MADV_SEQUENTIAL mappings we want to deactivate the back pages
898 	 * now and then forget about them (for the rest of the fault).
899 	 */
900 
901 	if (ufi.entry->advice == MADV_SEQUENTIAL && nback != 0) {
902 
903 		UVMHIST_LOG(maphist, "  MADV_SEQUENTIAL: flushing backpages",
904 		    0,0,0,0);
905 		/* flush back-page anons? */
906 		if (amap)
907 			uvmfault_anonflush(anons, nback);
908 
909 		/* flush object? */
910 		if (uobj) {
911 			uoff = (startva - ufi.entry->start) + ufi.entry->offset;
912 			simple_lock(&uobj->vmobjlock);
913 			(void) (uobj->pgops->pgo_put)(uobj, uoff, uoff +
914 				    (nback << PAGE_SHIFT), PGO_DEACTIVATE);
915 		}
916 
917 		/* now forget about the backpages */
918 		if (amap)
919 			anons += nback;
920 		startva += (nback << PAGE_SHIFT);
921 		npages -= nback;
922 		nback = centeridx = 0;
923 	}
924 
925 	/* locked: maps(read), amap(if there) */
926 
927 	/*
928 	 * map in the backpages and frontpages we found in the amap in hopes
929 	 * of preventing future faults.    we also init the pages[] array as
930 	 * we go.
931 	 */
932 
933 	currva = startva;
934 	shadowed = FALSE;
935 	for (lcv = 0 ; lcv < npages ; lcv++, currva += PAGE_SIZE) {
936 
937 		/*
938 		 * dont play with VAs that are already mapped
939 		 * except for center)
940 		 */
941 		if (lcv != centeridx &&
942 		    pmap_extract(ufi.orig_map->pmap, currva, NULL)) {
943 			pages[lcv] = PGO_DONTCARE;
944 			continue;
945 		}
946 
947 		/*
948 		 * unmapped or center page.   check if any anon at this level.
949 		 */
950 		if (amap == NULL || anons[lcv] == NULL) {
951 			pages[lcv] = NULL;
952 			continue;
953 		}
954 
955 		/*
956 		 * check for present page and map if possible.   re-activate it.
957 		 */
958 
959 		pages[lcv] = PGO_DONTCARE;
960 		if (lcv == centeridx) {		/* save center for later! */
961 			shadowed = TRUE;
962 			continue;
963 		}
964 		anon = anons[lcv];
965 		simple_lock(&anon->an_lock);
966 		/* ignore loaned pages */
967 		if (anon->an_page && anon->an_page->loan_count == 0 &&
968 		    (anon->an_page->flags & PG_BUSY) == 0) {
969 			uvm_lock_pageq();
970 			uvm_pageactivate(anon->an_page);
971 			uvm_unlock_pageq();
972 			UVMHIST_LOG(maphist,
973 			    "  MAPPING: n anon: pm=0x%x, va=0x%x, pg=0x%x",
974 			    ufi.orig_map->pmap, currva, anon->an_page, 0);
975 			uvmexp.fltnamap++;
976 
977 			/*
978 			 * Since this isn't the page that's actually faulting,
979 			 * ignore pmap_enter() failures; it's not critical
980 			 * that we enter these right now.
981 			 */
982 
983 			(void) pmap_enter(ufi.orig_map->pmap, currva,
984 			    VM_PAGE_TO_PHYS(anon->an_page),
985 			    (anon->an_ref > 1) ? (enter_prot & ~VM_PROT_WRITE) :
986 			    enter_prot,
987 			    PMAP_CANFAIL |
988 			     (VM_MAPENT_ISWIRED(ufi.entry) ? PMAP_WIRED : 0));
989 		}
990 		simple_unlock(&anon->an_lock);
991 		pmap_update(ufi.orig_map->pmap);
992 	}
993 
994 	/* locked: maps(read), amap(if there) */
995 	/* (shadowed == TRUE) if there is an anon at the faulting address */
996 	UVMHIST_LOG(maphist, "  shadowed=%d, will_get=%d", shadowed,
997 	    (uobj && shadowed == FALSE),0,0);
998 
999 	/*
1000 	 * note that if we are really short of RAM we could sleep in the above
1001 	 * call to pmap_enter with everything locked.   bad?
1002 	 *
1003 	 * XXX Actually, that is bad; pmap_enter() should just fail in that
1004 	 * XXX case.  --thorpej
1005 	 */
1006 
1007 	/*
1008 	 * if the desired page is not shadowed by the amap and we have a
1009 	 * backing object, then we check to see if the backing object would
1010 	 * prefer to handle the fault itself (rather than letting us do it
1011 	 * with the usual pgo_get hook).  the backing object signals this by
1012 	 * providing a pgo_fault routine.
1013 	 */
1014 
1015 	if (uobj && shadowed == FALSE && uobj->pgops->pgo_fault != NULL) {
1016 		simple_lock(&uobj->vmobjlock);
1017 
1018 		/* locked: maps(read), amap (if there), uobj */
1019 		error = uobj->pgops->pgo_fault(&ufi, startva, pages, npages,
1020 		    centeridx, access_type, PGO_LOCKED|PGO_SYNCIO);
1021 
1022 		/* locked: nothing, pgo_fault has unlocked everything */
1023 
1024 		if (error == ERESTART)
1025 			goto ReFault;		/* try again! */
1026 		/*
1027 		 * object fault routine responsible for pmap_update().
1028 		 */
1029 		goto done;
1030 	}
1031 
1032 	/*
1033 	 * now, if the desired page is not shadowed by the amap and we have
1034 	 * a backing object that does not have a special fault routine, then
1035 	 * we ask (with pgo_get) the object for resident pages that we care
1036 	 * about and attempt to map them in.  we do not let pgo_get block
1037 	 * (PGO_LOCKED).
1038 	 */
1039 
1040 	if (uobj && shadowed == FALSE) {
1041 		simple_lock(&uobj->vmobjlock);
1042 
1043 		/* locked (!shadowed): maps(read), amap (if there), uobj */
1044 		/*
1045 		 * the following call to pgo_get does _not_ change locking state
1046 		 */
1047 
1048 		uvmexp.fltlget++;
1049 		gotpages = npages;
1050 		(void) uobj->pgops->pgo_get(uobj, ufi.entry->offset +
1051 				(startva - ufi.entry->start),
1052 				pages, &gotpages, centeridx,
1053 				access_type & MASK(ufi.entry),
1054 				ufi.entry->advice, PGO_LOCKED);
1055 
1056 		/*
1057 		 * check for pages to map, if we got any
1058 		 */
1059 
1060 		uobjpage = NULL;
1061 
1062 		if (gotpages) {
1063 			currva = startva;
1064 			for (lcv = 0; lcv < npages;
1065 			     lcv++, currva += PAGE_SIZE) {
1066 				struct vm_page *curpg;
1067 				boolean_t readonly;
1068 
1069 				curpg = pages[lcv];
1070 				if (curpg == NULL || curpg == PGO_DONTCARE) {
1071 					continue;
1072 				}
1073 				KASSERT(curpg->uobject == uobj);
1074 
1075 				/*
1076 				 * if center page is resident and not
1077 				 * PG_BUSY|PG_RELEASED then pgo_get
1078 				 * made it PG_BUSY for us and gave
1079 				 * us a handle to it.   remember this
1080 				 * page as "uobjpage." (for later use).
1081 				 */
1082 
1083 				if (lcv == centeridx) {
1084 					uobjpage = curpg;
1085 					UVMHIST_LOG(maphist, "  got uobjpage "
1086 					    "(0x%x) with locked get",
1087 					    uobjpage, 0,0,0);
1088 					continue;
1089 				}
1090 
1091 				/*
1092 				 * calling pgo_get with PGO_LOCKED returns us
1093 				 * pages which are neither busy nor released,
1094 				 * so we don't need to check for this.
1095 				 * we can just directly enter the pages.
1096 				 */
1097 
1098 				uvm_lock_pageq();
1099 				uvm_pageactivate(curpg);
1100 				uvm_unlock_pageq();
1101 				UVMHIST_LOG(maphist,
1102 				  "  MAPPING: n obj: pm=0x%x, va=0x%x, pg=0x%x",
1103 				  ufi.orig_map->pmap, currva, curpg, 0);
1104 				uvmexp.fltnomap++;
1105 
1106 				/*
1107 				 * Since this page isn't the page that's
1108 				 * actually faulting, ignore pmap_enter()
1109 				 * failures; it's not critical that we
1110 				 * enter these right now.
1111 				 */
1112 				KASSERT((curpg->flags & PG_PAGEOUT) == 0);
1113 				KASSERT((curpg->flags & PG_RELEASED) == 0);
1114 				KASSERT(!UVM_OBJ_IS_CLEAN(curpg->uobject) ||
1115 				    (curpg->flags & PG_CLEAN) != 0);
1116 				readonly = (curpg->flags & PG_RDONLY)
1117 				    || (curpg->loan_count > 0)
1118 				    || UVM_OBJ_NEEDS_WRITEFAULT(curpg->uobject);
1119 
1120 				(void) pmap_enter(ufi.orig_map->pmap, currva,
1121 				    VM_PAGE_TO_PHYS(curpg),
1122 				    readonly ?
1123 				    enter_prot & ~VM_PROT_WRITE :
1124 				    enter_prot & MASK(ufi.entry),
1125 				    PMAP_CANFAIL |
1126 				     (wired ? PMAP_WIRED : 0));
1127 
1128 				/*
1129 				 * NOTE: page can't be PG_WANTED or PG_RELEASED
1130 				 * because we've held the lock the whole time
1131 				 * we've had the handle.
1132 				 */
1133 
1134 				curpg->flags &= ~(PG_BUSY);
1135 				UVM_PAGE_OWN(curpg, NULL);
1136 			}
1137 			pmap_update(ufi.orig_map->pmap);
1138 		}
1139 	} else {
1140 		uobjpage = NULL;
1141 	}
1142 
1143 	/* locked (shadowed): maps(read), amap */
1144 	/* locked (!shadowed): maps(read), amap(if there),
1145 		 uobj(if !null), uobjpage(if !null) */
1146 
1147 	/*
1148 	 * note that at this point we are done with any front or back pages.
1149 	 * we are now going to focus on the center page (i.e. the one we've
1150 	 * faulted on).  if we have faulted on the top (anon) layer
1151 	 * [i.e. case 1], then the anon we want is anons[centeridx] (we have
1152 	 * not touched it yet).  if we have faulted on the bottom (uobj)
1153 	 * layer [i.e. case 2] and the page was both present and available,
1154 	 * then we've got a pointer to it as "uobjpage" and we've already
1155 	 * made it BUSY.
1156 	 */
1157 
1158 	/*
1159 	 * there are four possible cases we must address: 1A, 1B, 2A, and 2B
1160 	 */
1161 
1162 	/*
1163 	 * redirect case 2: if we are not shadowed, go to case 2.
1164 	 */
1165 
1166 	if (shadowed == FALSE)
1167 		goto Case2;
1168 
1169 	/* locked: maps(read), amap */
1170 
1171 	/*
1172 	 * handle case 1: fault on an anon in our amap
1173 	 */
1174 
1175 	anon = anons[centeridx];
1176 	UVMHIST_LOG(maphist, "  case 1 fault: anon=0x%x", anon, 0,0,0);
1177 	simple_lock(&anon->an_lock);
1178 
1179 	/* locked: maps(read), amap, anon */
1180 
1181 	/*
1182 	 * no matter if we have case 1A or case 1B we are going to need to
1183 	 * have the anon's memory resident.   ensure that now.
1184 	 */
1185 
1186 	/*
1187 	 * let uvmfault_anonget do the dirty work.
1188 	 * if it fails (!OK) it will unlock everything for us.
1189 	 * if it succeeds, locks are still valid and locked.
1190 	 * also, if it is OK, then the anon's page is on the queues.
1191 	 * if the page is on loan from a uvm_object, then anonget will
1192 	 * lock that object for us if it does not fail.
1193 	 */
1194 
1195 	error = uvmfault_anonget(&ufi, amap, anon);
1196 	switch (error) {
1197 	case 0:
1198 		break;
1199 
1200 	case ERESTART:
1201 		goto ReFault;
1202 
1203 	case EAGAIN:
1204 		tsleep(&lbolt, PVM, "fltagain1", 0);
1205 		goto ReFault;
1206 
1207 	default:
1208 		goto done;
1209 	}
1210 
1211 	/*
1212 	 * uobj is non null if the page is on loan from an object (i.e. uobj)
1213 	 */
1214 
1215 	uobj = anon->an_page->uobject;	/* locked by anonget if !NULL */
1216 
1217 	/* locked: maps(read), amap, anon, uobj(if one) */
1218 
1219 	/*
1220 	 * special handling for loaned pages
1221 	 */
1222 
1223 	if (anon->an_page->loan_count) {
1224 
1225 		if (!cow_now) {
1226 
1227 			/*
1228 			 * for read faults on loaned pages we just cap the
1229 			 * protection at read-only.
1230 			 */
1231 
1232 			enter_prot = enter_prot & ~VM_PROT_WRITE;
1233 
1234 		} else {
1235 			/*
1236 			 * note that we can't allow writes into a loaned page!
1237 			 *
1238 			 * if we have a write fault on a loaned page in an
1239 			 * anon then we need to look at the anon's ref count.
1240 			 * if it is greater than one then we are going to do
1241 			 * a normal copy-on-write fault into a new anon (this
1242 			 * is not a problem).  however, if the reference count
1243 			 * is one (a case where we would normally allow a
1244 			 * write directly to the page) then we need to kill
1245 			 * the loan before we continue.
1246 			 */
1247 
1248 			/* >1 case is already ok */
1249 			if (anon->an_ref == 1) {
1250 
1251 				/* get new un-owned replacement page */
1252 				pg = uvm_pagealloc(NULL, 0, NULL, 0);
1253 				if (pg == NULL) {
1254 					uvmfault_unlockall(&ufi, amap, uobj,
1255 					    anon);
1256 					uvm_wait("flt_noram2");
1257 					goto ReFault;
1258 				}
1259 
1260 				/*
1261 				 * copy data, kill loan, and drop uobj lock
1262 				 * (if any)
1263 				 */
1264 				/* copy old -> new */
1265 				uvm_pagecopy(anon->an_page, pg);
1266 
1267 				/* force reload */
1268 				pmap_page_protect(anon->an_page,
1269 						  VM_PROT_NONE);
1270 				uvm_lock_pageq();	  /* KILL loan */
1271 
1272 				anon->an_page->uanon = NULL;
1273 				/* in case we owned */
1274 				anon->an_page->pqflags &= ~PQ_ANON;
1275 
1276 				if (uobj) {
1277 					/* if we were receiver of loan */
1278 					anon->an_page->loan_count--;
1279 				} else {
1280 					/*
1281 					 * we were the lender (A->K); need
1282 					 * to remove the page from pageq's.
1283 					 */
1284 					uvm_pagedequeue(anon->an_page);
1285 				}
1286 
1287 				uvm_pageactivate(pg);
1288 				uvm_unlock_pageq();
1289 				if (uobj) {
1290 					simple_unlock(&uobj->vmobjlock);
1291 					uobj = NULL;
1292 				}
1293 
1294 				/* install new page in anon */
1295 				anon->an_page = pg;
1296 				pg->uanon = anon;
1297 				pg->pqflags |= PQ_ANON;
1298 				pg->flags &= ~(PG_BUSY|PG_FAKE);
1299 				UVM_PAGE_OWN(pg, NULL);
1300 
1301 				/* done! */
1302 			}     /* ref == 1 */
1303 		}       /* write fault */
1304 	}         /* loan count */
1305 
1306 	/*
1307 	 * if we are case 1B then we will need to allocate a new blank
1308 	 * anon to transfer the data into.   note that we have a lock
1309 	 * on anon, so no one can busy or release the page until we are done.
1310 	 * also note that the ref count can't drop to zero here because
1311 	 * it is > 1 and we are only dropping one ref.
1312 	 *
1313 	 * in the (hopefully very rare) case that we are out of RAM we
1314 	 * will unlock, wait for more RAM, and refault.
1315 	 *
1316 	 * if we are out of anon VM we kill the process (XXX: could wait?).
1317 	 */
1318 
1319 	if (cow_now && anon->an_ref > 1) {
1320 
1321 		UVMHIST_LOG(maphist, "  case 1B: COW fault",0,0,0,0);
1322 		uvmexp.flt_acow++;
1323 		oanon = anon;		/* oanon = old, locked anon */
1324 
1325 		error = uvmfault_promote(&ufi, oanon, PGO_DONTCARE,
1326 		    &anon, &anon_spare);
1327 		switch (error) {
1328 		case 0:
1329 			break;
1330 		case ERESTART:
1331 			goto ReFault;
1332 		default:
1333 			goto done;
1334 		}
1335 
1336 		pg = anon->an_page;
1337 		uvm_lock_pageq();
1338 		uvm_pageactivate(pg);
1339 		uvm_unlock_pageq();
1340 		pg->flags &= ~(PG_BUSY|PG_FAKE);
1341 		UVM_PAGE_OWN(pg, NULL);
1342 
1343 		/* deref: can not drop to zero here by defn! */
1344 		oanon->an_ref--;
1345 
1346 		/*
1347 		 * note: oanon is still locked, as is the new anon.  we
1348 		 * need to check for this later when we unlock oanon; if
1349 		 * oanon != anon, we'll have to unlock anon, too.
1350 		 */
1351 
1352 	} else {
1353 
1354 		uvmexp.flt_anon++;
1355 		oanon = anon;		/* old, locked anon is same as anon */
1356 		pg = anon->an_page;
1357 		if (anon->an_ref > 1)     /* disallow writes to ref > 1 anons */
1358 			enter_prot = enter_prot & ~VM_PROT_WRITE;
1359 
1360 	}
1361 
1362 	/* locked: maps(read), amap, oanon, anon (if different from oanon) */
1363 
1364 	/*
1365 	 * now map the page in.
1366 	 */
1367 
1368 	UVMHIST_LOG(maphist, "  MAPPING: anon: pm=0x%x, va=0x%x, pg=0x%x",
1369 	    ufi.orig_map->pmap, ufi.orig_rvaddr, pg, 0);
1370 	if (pmap_enter(ufi.orig_map->pmap, ufi.orig_rvaddr, VM_PAGE_TO_PHYS(pg),
1371 	    enter_prot, access_type | PMAP_CANFAIL | (wired ? PMAP_WIRED : 0))
1372 	    != 0) {
1373 
1374 		/*
1375 		 * No need to undo what we did; we can simply think of
1376 		 * this as the pmap throwing away the mapping information.
1377 		 *
1378 		 * We do, however, have to go through the ReFault path,
1379 		 * as the map may change while we're asleep.
1380 		 */
1381 
1382 		if (anon != oanon)
1383 			simple_unlock(&anon->an_lock);
1384 		uvmfault_unlockall(&ufi, amap, uobj, oanon);
1385 		if (!uvm_reclaimable()) {
1386 			UVMHIST_LOG(maphist,
1387 			    "<- failed.  out of VM",0,0,0,0);
1388 			/* XXX instrumentation */
1389 			error = ENOMEM;
1390 			goto done;
1391 		}
1392 		/* XXX instrumentation */
1393 		uvm_wait("flt_pmfail1");
1394 		goto ReFault;
1395 	}
1396 
1397 	/*
1398 	 * ... update the page queues.
1399 	 */
1400 
1401 	uvm_lock_pageq();
1402 	if (wire_fault) {
1403 		uvm_pagewire(pg);
1404 
1405 		/*
1406 		 * since the now-wired page cannot be paged out,
1407 		 * release its swap resources for others to use.
1408 		 * since an anon with no swap cannot be PG_CLEAN,
1409 		 * clear its clean flag now.
1410 		 */
1411 
1412 		pg->flags &= ~(PG_CLEAN);
1413 		uvm_anon_dropswap(anon);
1414 	} else {
1415 		uvm_pageactivate(pg);
1416 	}
1417 	uvm_unlock_pageq();
1418 
1419 	/*
1420 	 * done case 1!  finish up by unlocking everything and returning success
1421 	 */
1422 
1423 	if (anon != oanon)
1424 		simple_unlock(&anon->an_lock);
1425 	uvmfault_unlockall(&ufi, amap, uobj, oanon);
1426 	pmap_update(ufi.orig_map->pmap);
1427 	error = 0;
1428 	goto done;
1429 
1430 Case2:
1431 	/*
1432 	 * handle case 2: faulting on backing object or zero fill
1433 	 */
1434 
1435 	/*
1436 	 * locked:
1437 	 * maps(read), amap(if there), uobj(if !null), uobjpage(if !null)
1438 	 */
1439 
1440 	/*
1441 	 * note that uobjpage can not be PGO_DONTCARE at this point.  we now
1442 	 * set uobjpage to PGO_DONTCARE if we are doing a zero fill.  if we
1443 	 * have a backing object, check and see if we are going to promote
1444 	 * the data up to an anon during the fault.
1445 	 */
1446 
1447 	if (uobj == NULL) {
1448 		uobjpage = PGO_DONTCARE;
1449 		promote = TRUE;		/* always need anon here */
1450 	} else {
1451 		KASSERT(uobjpage != PGO_DONTCARE);
1452 		promote = cow_now && UVM_ET_ISCOPYONWRITE(ufi.entry);
1453 	}
1454 	UVMHIST_LOG(maphist, "  case 2 fault: promote=%d, zfill=%d",
1455 	    promote, (uobj == NULL), 0,0);
1456 
1457 	/*
1458 	 * if uobjpage is not null then we do not need to do I/O to get the
1459 	 * uobjpage.
1460 	 *
1461 	 * if uobjpage is null, then we need to unlock and ask the pager to
1462 	 * get the data for us.   once we have the data, we need to reverify
1463 	 * the state the world.   we are currently not holding any resources.
1464 	 */
1465 
1466 	if (uobjpage) {
1467 		/* update rusage counters */
1468 		curproc->p_stats->p_ru.ru_minflt++;
1469 	} else {
1470 		/* update rusage counters */
1471 		curproc->p_stats->p_ru.ru_majflt++;
1472 
1473 		/* locked: maps(read), amap(if there), uobj */
1474 		uvmfault_unlockall(&ufi, amap, NULL, NULL);
1475 		/* locked: uobj */
1476 
1477 		uvmexp.fltget++;
1478 		gotpages = 1;
1479 		uoff = (ufi.orig_rvaddr - ufi.entry->start) + ufi.entry->offset;
1480 		error = uobj->pgops->pgo_get(uobj, uoff, &uobjpage, &gotpages,
1481 		    0, access_type & MASK(ufi.entry), ufi.entry->advice,
1482 		    PGO_SYNCIO);
1483 		/* locked: uobjpage(if no error) */
1484 
1485 		/*
1486 		 * recover from I/O
1487 		 */
1488 
1489 		if (error) {
1490 			if (error == EAGAIN) {
1491 				UVMHIST_LOG(maphist,
1492 				    "  pgo_get says TRY AGAIN!",0,0,0,0);
1493 				tsleep(&lbolt, PVM, "fltagain2", 0);
1494 				goto ReFault;
1495 			}
1496 
1497 			UVMHIST_LOG(maphist, "<- pgo_get failed (code %d)",
1498 			    error, 0,0,0);
1499 			goto done;
1500 		}
1501 
1502 		/* locked: uobjpage */
1503 
1504 		uvm_lock_pageq();
1505 		uvm_pageactivate(uobjpage);
1506 		uvm_unlock_pageq();
1507 
1508 		/*
1509 		 * re-verify the state of the world by first trying to relock
1510 		 * the maps.  always relock the object.
1511 		 */
1512 
1513 		locked = uvmfault_relock(&ufi);
1514 		if (locked && amap)
1515 			amap_lock(amap);
1516 		uobj = uobjpage->uobject;
1517 		simple_lock(&uobj->vmobjlock);
1518 
1519 		/* locked(locked): maps(read), amap(if !null), uobj, uobjpage */
1520 		/* locked(!locked): uobj, uobjpage */
1521 
1522 		/*
1523 		 * verify that the page has not be released and re-verify
1524 		 * that amap slot is still free.   if there is a problem,
1525 		 * we unlock and clean up.
1526 		 */
1527 
1528 		if ((uobjpage->flags & PG_RELEASED) != 0 ||
1529 		    (locked && amap &&
1530 		    amap_lookup(&ufi.entry->aref,
1531 		      ufi.orig_rvaddr - ufi.entry->start))) {
1532 			if (locked)
1533 				uvmfault_unlockall(&ufi, amap, NULL, NULL);
1534 			locked = FALSE;
1535 		}
1536 
1537 		/*
1538 		 * didn't get the lock?   release the page and retry.
1539 		 */
1540 
1541 		if (locked == FALSE) {
1542 			UVMHIST_LOG(maphist,
1543 			    "  wasn't able to relock after fault: retry",
1544 			    0,0,0,0);
1545 			if (uobjpage->flags & PG_WANTED)
1546 				wakeup(uobjpage);
1547 			if (uobjpage->flags & PG_RELEASED) {
1548 				uvmexp.fltpgrele++;
1549 				uvm_pagefree(uobjpage);
1550 				goto ReFault;
1551 			}
1552 			uobjpage->flags &= ~(PG_BUSY|PG_WANTED);
1553 			UVM_PAGE_OWN(uobjpage, NULL);
1554 			simple_unlock(&uobj->vmobjlock);
1555 			goto ReFault;
1556 		}
1557 
1558 		/*
1559 		 * we have the data in uobjpage which is busy and
1560 		 * not released.  we are holding object lock (so the page
1561 		 * can't be released on us).
1562 		 */
1563 
1564 		/* locked: maps(read), amap(if !null), uobj, uobjpage */
1565 	}
1566 
1567 	/*
1568 	 * locked:
1569 	 * maps(read), amap(if !null), uobj(if !null), uobjpage(if uobj)
1570 	 */
1571 
1572 	/*
1573 	 * notes:
1574 	 *  - at this point uobjpage can not be NULL
1575 	 *  - at this point uobjpage can not be PG_RELEASED (since we checked
1576 	 *  for it above)
1577 	 *  - at this point uobjpage could be PG_WANTED (handle later)
1578 	 */
1579 
1580 	KASSERT(uobj == NULL || uobj == uobjpage->uobject);
1581 	KASSERT(uobj == NULL || !UVM_OBJ_IS_CLEAN(uobjpage->uobject) ||
1582 	    (uobjpage->flags & PG_CLEAN) != 0);
1583 	if (promote == FALSE) {
1584 
1585 		/*
1586 		 * we are not promoting.   if the mapping is COW ensure that we
1587 		 * don't give more access than we should (e.g. when doing a read
1588 		 * fault on a COPYONWRITE mapping we want to map the COW page in
1589 		 * R/O even though the entry protection could be R/W).
1590 		 *
1591 		 * set "pg" to the page we want to map in (uobjpage, usually)
1592 		 */
1593 
1594 		/* no anon in this case. */
1595 		anon = NULL;
1596 
1597 		uvmexp.flt_obj++;
1598 		if (UVM_ET_ISCOPYONWRITE(ufi.entry) ||
1599 		    UVM_OBJ_NEEDS_WRITEFAULT(uobjpage->uobject))
1600 			enter_prot &= ~VM_PROT_WRITE;
1601 		pg = uobjpage;		/* map in the actual object */
1602 
1603 		/* assert(uobjpage != PGO_DONTCARE) */
1604 
1605 		/*
1606 		 * we are faulting directly on the page.   be careful
1607 		 * about writing to loaned pages...
1608 		 */
1609 
1610 		if (uobjpage->loan_count) {
1611 			if (!cow_now) {
1612 				/* read fault: cap the protection at readonly */
1613 				/* cap! */
1614 				enter_prot = enter_prot & ~VM_PROT_WRITE;
1615 			} else {
1616 				/* write fault: must break the loan here */
1617 
1618 				pg = uvm_loanbreak(uobjpage);
1619 				if (pg == NULL) {
1620 
1621 					/*
1622 					 * drop ownership of page, it can't
1623 					 * be released
1624 					 */
1625 
1626 					if (uobjpage->flags & PG_WANTED)
1627 						wakeup(uobjpage);
1628 					uobjpage->flags &= ~(PG_BUSY|PG_WANTED);
1629 					UVM_PAGE_OWN(uobjpage, NULL);
1630 
1631 					uvmfault_unlockall(&ufi, amap, uobj,
1632 					  NULL);
1633 					UVMHIST_LOG(maphist,
1634 					  "  out of RAM breaking loan, waiting",
1635 					  0,0,0,0);
1636 					uvmexp.fltnoram++;
1637 					uvm_wait("flt_noram4");
1638 					goto ReFault;
1639 				}
1640 				uobjpage = pg;
1641 			}
1642 		}
1643 	} else {
1644 
1645 		/*
1646 		 * if we are going to promote the data to an anon we
1647 		 * allocate a blank anon here and plug it into our amap.
1648 		 */
1649 #if DIAGNOSTIC
1650 		if (amap == NULL)
1651 			panic("uvm_fault: want to promote data, but no anon");
1652 #endif
1653 		error = uvmfault_promote(&ufi, NULL, uobjpage,
1654 		    &anon, &anon_spare);
1655 		switch (error) {
1656 		case 0:
1657 			break;
1658 		case ERESTART:
1659 			goto ReFault;
1660 		default:
1661 			goto done;
1662 		}
1663 
1664 		pg = anon->an_page;
1665 
1666 		/*
1667 		 * fill in the data
1668 		 */
1669 
1670 		if (uobjpage != PGO_DONTCARE) {
1671 			uvmexp.flt_prcopy++;
1672 
1673 			/*
1674 			 * promote to shared amap?  make sure all sharing
1675 			 * procs see it
1676 			 */
1677 
1678 			if ((amap_flags(amap) & AMAP_SHARED) != 0) {
1679 				pmap_page_protect(uobjpage, VM_PROT_NONE);
1680 				/*
1681 				 * XXX: PAGE MIGHT BE WIRED!
1682 				 */
1683 			}
1684 
1685 			/*
1686 			 * dispose of uobjpage.  it can't be PG_RELEASED
1687 			 * since we still hold the object lock.
1688 			 * drop handle to uobj as well.
1689 			 */
1690 
1691 			if (uobjpage->flags & PG_WANTED)
1692 				/* still have the obj lock */
1693 				wakeup(uobjpage);
1694 			uobjpage->flags &= ~(PG_BUSY|PG_WANTED);
1695 			UVM_PAGE_OWN(uobjpage, NULL);
1696 			simple_unlock(&uobj->vmobjlock);
1697 			uobj = NULL;
1698 
1699 			UVMHIST_LOG(maphist,
1700 			    "  promote uobjpage 0x%x to anon/page 0x%x/0x%x",
1701 			    uobjpage, anon, pg, 0);
1702 
1703 		} else {
1704 			uvmexp.flt_przero++;
1705 
1706 			/*
1707 			 * Page is zero'd and marked dirty by
1708 			 * uvmfault_promote().
1709 			 */
1710 
1711 			UVMHIST_LOG(maphist,"  zero fill anon/page 0x%x/0%x",
1712 			    anon, pg, 0, 0);
1713 		}
1714 	}
1715 
1716 	/*
1717 	 * locked:
1718 	 * maps(read), amap(if !null), uobj(if !null), uobjpage(if uobj),
1719 	 *   anon(if !null), pg(if anon)
1720 	 *
1721 	 * note: pg is either the uobjpage or the new page in the new anon
1722 	 */
1723 
1724 	/*
1725 	 * all resources are present.   we can now map it in and free our
1726 	 * resources.
1727 	 */
1728 
1729 	UVMHIST_LOG(maphist,
1730 	    "  MAPPING: case2: pm=0x%x, va=0x%x, pg=0x%x, promote=%d",
1731 	    ufi.orig_map->pmap, ufi.orig_rvaddr, pg, promote);
1732 	KASSERT((access_type & VM_PROT_WRITE) == 0 ||
1733 		(pg->flags & PG_RDONLY) == 0);
1734 	if (pmap_enter(ufi.orig_map->pmap, ufi.orig_rvaddr, VM_PAGE_TO_PHYS(pg),
1735 	    pg->flags & PG_RDONLY ? enter_prot & ~VM_PROT_WRITE : enter_prot,
1736 	    access_type | PMAP_CANFAIL | (wired ? PMAP_WIRED : 0)) != 0) {
1737 
1738 		/*
1739 		 * No need to undo what we did; we can simply think of
1740 		 * this as the pmap throwing away the mapping information.
1741 		 *
1742 		 * We do, however, have to go through the ReFault path,
1743 		 * as the map may change while we're asleep.
1744 		 */
1745 
1746 		if (pg->flags & PG_WANTED)
1747 			wakeup(pg);
1748 
1749 		/*
1750 		 * note that pg can't be PG_RELEASED since we did not drop
1751 		 * the object lock since the last time we checked.
1752 		 */
1753 
1754 		pg->flags &= ~(PG_BUSY|PG_FAKE|PG_WANTED);
1755 		UVM_PAGE_OWN(pg, NULL);
1756 		uvmfault_unlockall(&ufi, amap, uobj, anon);
1757 		if (!uvm_reclaimable()) {
1758 			UVMHIST_LOG(maphist,
1759 			    "<- failed.  out of VM",0,0,0,0);
1760 			/* XXX instrumentation */
1761 			error = ENOMEM;
1762 			goto done;
1763 		}
1764 		/* XXX instrumentation */
1765 		uvm_wait("flt_pmfail2");
1766 		goto ReFault;
1767 	}
1768 
1769 	uvm_lock_pageq();
1770 	if (wire_fault) {
1771 		uvm_pagewire(pg);
1772 		if (pg->pqflags & PQ_AOBJ) {
1773 
1774 			/*
1775 			 * since the now-wired page cannot be paged out,
1776 			 * release its swap resources for others to use.
1777 			 * since an aobj page with no swap cannot be PG_CLEAN,
1778 			 * clear its clean flag now.
1779 			 */
1780 
1781 			pg->flags &= ~(PG_CLEAN);
1782 			uao_dropswap(uobj, pg->offset >> PAGE_SHIFT);
1783 		}
1784 	} else {
1785 		uvm_pageactivate(pg);
1786 	}
1787 	uvm_unlock_pageq();
1788 	if (pg->flags & PG_WANTED)
1789 		wakeup(pg);
1790 
1791 	/*
1792 	 * note that pg can't be PG_RELEASED since we did not drop the object
1793 	 * lock since the last time we checked.
1794 	 */
1795 
1796 	pg->flags &= ~(PG_BUSY|PG_FAKE|PG_WANTED);
1797 	UVM_PAGE_OWN(pg, NULL);
1798 	uvmfault_unlockall(&ufi, amap, uobj, anon);
1799 	pmap_update(ufi.orig_map->pmap);
1800 	UVMHIST_LOG(maphist, "<- done (SUCCESS!)",0,0,0,0);
1801 	error = 0;
1802 done:
1803 	if (anon_spare != NULL) {
1804 		anon_spare->an_ref--;
1805 		uvm_anfree(anon_spare);
1806 	}
1807 	return error;
1808 }
1809 
1810 /*
1811  * uvm_fault_wire: wire down a range of virtual addresses in a map.
1812  *
1813  * => map may be read-locked by caller, but MUST NOT be write-locked.
1814  * => if map is read-locked, any operations which may cause map to
1815  *	be write-locked in uvm_fault() must be taken care of by
1816  *	the caller.  See uvm_map_pageable().
1817  */
1818 
1819 int
1820 uvm_fault_wire(struct vm_map *map, vaddr_t start, vaddr_t end,
1821     vm_fault_t fault_type, vm_prot_t access_type)
1822 {
1823 	vaddr_t va;
1824 	int error;
1825 
1826 	/*
1827 	 * now fault it in a page at a time.   if the fault fails then we have
1828 	 * to undo what we have done.   note that in uvm_fault VM_PROT_NONE
1829 	 * is replaced with the max protection if fault_type is VM_FAULT_WIRE.
1830 	 */
1831 
1832 	/*
1833 	 * XXX work around overflowing a vaddr_t.  this prevents us from
1834 	 * wiring the last page in the address space, though.
1835 	 */
1836 	if (start > end) {
1837 		return EFAULT;
1838 	}
1839 
1840 	for (va = start ; va < end ; va += PAGE_SIZE) {
1841 		error = uvm_fault(map, va, fault_type, access_type);
1842 		if (error) {
1843 			if (va != start) {
1844 				uvm_fault_unwire(map, start, va);
1845 			}
1846 			return error;
1847 		}
1848 	}
1849 	return 0;
1850 }
1851 
1852 /*
1853  * uvm_fault_unwire(): unwire range of virtual space.
1854  */
1855 
1856 void
1857 uvm_fault_unwire(struct vm_map *map, vaddr_t start, vaddr_t end)
1858 {
1859 	vm_map_lock_read(map);
1860 	uvm_fault_unwire_locked(map, start, end);
1861 	vm_map_unlock_read(map);
1862 }
1863 
1864 /*
1865  * uvm_fault_unwire_locked(): the guts of uvm_fault_unwire().
1866  *
1867  * => map must be at least read-locked.
1868  */
1869 
1870 void
1871 uvm_fault_unwire_locked(struct vm_map *map, vaddr_t start, vaddr_t end)
1872 {
1873 	struct vm_map_entry *entry;
1874 	pmap_t pmap = vm_map_pmap(map);
1875 	vaddr_t va;
1876 	paddr_t pa;
1877 	struct vm_page *pg;
1878 
1879 	KASSERT((map->flags & VM_MAP_INTRSAFE) == 0);
1880 
1881 	/*
1882 	 * we assume that the area we are unwiring has actually been wired
1883 	 * in the first place.   this means that we should be able to extract
1884 	 * the PAs from the pmap.   we also lock out the page daemon so that
1885 	 * we can call uvm_pageunwire.
1886 	 */
1887 
1888 	uvm_lock_pageq();
1889 
1890 	/*
1891 	 * find the beginning map entry for the region.
1892 	 */
1893 
1894 	KASSERT(start >= vm_map_min(map) && end <= vm_map_max(map));
1895 	if (uvm_map_lookup_entry(map, start, &entry) == FALSE)
1896 		panic("uvm_fault_unwire_locked: address not in map");
1897 
1898 	for (va = start; va < end; va += PAGE_SIZE) {
1899 		if (pmap_extract(pmap, va, &pa) == FALSE)
1900 			continue;
1901 
1902 		/*
1903 		 * find the map entry for the current address.
1904 		 */
1905 
1906 		KASSERT(va >= entry->start);
1907 		while (va >= entry->end) {
1908 			KASSERT(entry->next != &map->header &&
1909 				entry->next->start <= entry->end);
1910 			entry = entry->next;
1911 		}
1912 
1913 		/*
1914 		 * if the entry is no longer wired, tell the pmap.
1915 		 */
1916 
1917 		if (VM_MAPENT_ISWIRED(entry) == 0)
1918 			pmap_unwire(pmap, va);
1919 
1920 		pg = PHYS_TO_VM_PAGE(pa);
1921 		if (pg)
1922 			uvm_pageunwire(pg);
1923 	}
1924 
1925 	uvm_unlock_pageq();
1926 }
1927