xref: /netbsd-src/sys/uvm/uvm_fault.c (revision 46f5119e40af2e51998f686b2fdcc76b5488f7f3)
1 /*	$NetBSD: uvm_fault.c,v 1.184 2011/04/23 18:14:12 rmind Exp $	*/
2 
3 /*
4  * Copyright (c) 1997 Charles D. Cranor and Washington University.
5  * All rights reserved.
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  * 1. Redistributions of source code must retain the above copyright
11  *    notice, this list of conditions and the following disclaimer.
12  * 2. Redistributions in binary form must reproduce the above copyright
13  *    notice, this list of conditions and the following disclaimer in the
14  *    documentation and/or other materials provided with the distribution.
15  *
16  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
17  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
18  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
19  * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
20  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
21  * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
22  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
23  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
24  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
25  * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
26  *
27  * from: Id: uvm_fault.c,v 1.1.2.23 1998/02/06 05:29:05 chs Exp
28  */
29 
30 /*
31  * uvm_fault.c: fault handler
32  */
33 
34 #include <sys/cdefs.h>
35 __KERNEL_RCSID(0, "$NetBSD: uvm_fault.c,v 1.184 2011/04/23 18:14:12 rmind Exp $");
36 
37 #include "opt_uvmhist.h"
38 
39 #include <sys/param.h>
40 #include <sys/systm.h>
41 #include <sys/kernel.h>
42 #include <sys/mman.h>
43 
44 #include <uvm/uvm.h>
45 
46 /*
47  *
48  * a word on page faults:
49  *
50  * types of page faults we handle:
51  *
52  * CASE 1: upper layer faults                   CASE 2: lower layer faults
53  *
54  *    CASE 1A         CASE 1B                  CASE 2A        CASE 2B
55  *    read/write1     write>1                  read/write   +-cow_write/zero
56  *         |             |                         |        |
57  *      +--|--+       +--|--+     +-----+       +  |  +     | +-----+
58  * amap |  V  |       |  ---------> new |          |        | |  ^  |
59  *      +-----+       +-----+     +-----+       +  |  +     | +--|--+
60  *                                                 |        |    |
61  *      +-----+       +-----+                   +--|--+     | +--|--+
62  * uobj | d/c |       | d/c |                   |  V  |     +----+  |
63  *      +-----+       +-----+                   +-----+       +-----+
64  *
65  * d/c = don't care
66  *
67  *   case [0]: layerless fault
68  *	no amap or uobj is present.   this is an error.
69  *
70  *   case [1]: upper layer fault [anon active]
71  *     1A: [read] or [write with anon->an_ref == 1]
72  *		I/O takes place in upper level anon and uobj is not touched.
73  *     1B: [write with anon->an_ref > 1]
74  *		new anon is alloc'd and data is copied off ["COW"]
75  *
76  *   case [2]: lower layer fault [uobj]
77  *     2A: [read on non-NULL uobj] or [write to non-copy_on_write area]
78  *		I/O takes place directly in object.
79  *     2B: [write to copy_on_write] or [read on NULL uobj]
80  *		data is "promoted" from uobj to a new anon.
81  *		if uobj is null, then we zero fill.
82  *
83  * we follow the standard UVM locking protocol ordering:
84  *
85  * MAPS => AMAP => UOBJ => ANON => PAGE QUEUES (PQ)
86  * we hold a PG_BUSY page if we unlock for I/O
87  *
88  *
89  * the code is structured as follows:
90  *
91  *     - init the "IN" params in the ufi structure
92  *   ReFault: (ERESTART returned to the loop in uvm_fault_internal)
93  *     - do lookups [locks maps], check protection, handle needs_copy
94  *     - check for case 0 fault (error)
95  *     - establish "range" of fault
96  *     - if we have an amap lock it and extract the anons
97  *     - if sequential advice deactivate pages behind us
98  *     - at the same time check pmap for unmapped areas and anon for pages
99  *	 that we could map in (and do map it if found)
100  *     - check object for resident pages that we could map in
101  *     - if (case 2) goto Case2
102  *     - >>> handle case 1
103  *           - ensure source anon is resident in RAM
104  *           - if case 1B alloc new anon and copy from source
105  *           - map the correct page in
106  *   Case2:
107  *     - >>> handle case 2
108  *           - ensure source page is resident (if uobj)
109  *           - if case 2B alloc new anon and copy from source (could be zero
110  *		fill if uobj == NULL)
111  *           - map the correct page in
112  *     - done!
113  *
114  * note on paging:
115  *   if we have to do I/O we place a PG_BUSY page in the correct object,
116  * unlock everything, and do the I/O.   when I/O is done we must reverify
117  * the state of the world before assuming that our data structures are
118  * valid.   [because mappings could change while the map is unlocked]
119  *
120  *  alternative 1: unbusy the page in question and restart the page fault
121  *    from the top (ReFault).   this is easy but does not take advantage
122  *    of the information that we already have from our previous lookup,
123  *    although it is possible that the "hints" in the vm_map will help here.
124  *
125  * alternative 2: the system already keeps track of a "version" number of
126  *    a map.   [i.e. every time you write-lock a map (e.g. to change a
127  *    mapping) you bump the version number up by one...]   so, we can save
128  *    the version number of the map before we release the lock and start I/O.
129  *    then when I/O is done we can relock and check the version numbers
130  *    to see if anything changed.    this might save us some over 1 because
131  *    we don't have to unbusy the page and may be less compares(?).
132  *
133  * alternative 3: put in backpointers or a way to "hold" part of a map
134  *    in place while I/O is in progress.   this could be complex to
135  *    implement (especially with structures like amap that can be referenced
136  *    by multiple map entries, and figuring out what should wait could be
137  *    complex as well...).
138  *
139  * we use alternative 2.  given that we are multi-threaded now we may want
140  * to reconsider the choice.
141  */
142 
143 /*
144  * local data structures
145  */
146 
147 struct uvm_advice {
148 	int advice;
149 	int nback;
150 	int nforw;
151 };
152 
153 /*
154  * page range array:
155  * note: index in array must match "advice" value
156  * XXX: borrowed numbers from freebsd.   do they work well for us?
157  */
158 
159 static const struct uvm_advice uvmadvice[] = {
160 	{ MADV_NORMAL, 3, 4 },
161 	{ MADV_RANDOM, 0, 0 },
162 	{ MADV_SEQUENTIAL, 8, 7},
163 };
164 
165 #define UVM_MAXRANGE 16	/* must be MAX() of nback+nforw+1 */
166 
167 /*
168  * private prototypes
169  */
170 
171 /*
172  * inline functions
173  */
174 
175 /*
176  * uvmfault_anonflush: try and deactivate pages in specified anons
177  *
178  * => does not have to deactivate page if it is busy
179  */
180 
181 static inline void
182 uvmfault_anonflush(struct vm_anon **anons, int n)
183 {
184 	int lcv;
185 	struct vm_page *pg;
186 
187 	for (lcv = 0; lcv < n; lcv++) {
188 		if (anons[lcv] == NULL)
189 			continue;
190 		mutex_enter(&anons[lcv]->an_lock);
191 		pg = anons[lcv]->an_page;
192 		if (pg && (pg->flags & PG_BUSY) == 0) {
193 			mutex_enter(&uvm_pageqlock);
194 			if (pg->wire_count == 0) {
195 				uvm_pagedeactivate(pg);
196 			}
197 			mutex_exit(&uvm_pageqlock);
198 		}
199 		mutex_exit(&anons[lcv]->an_lock);
200 	}
201 }
202 
203 /*
204  * normal functions
205  */
206 
207 /*
208  * uvmfault_amapcopy: clear "needs_copy" in a map.
209  *
210  * => called with VM data structures unlocked (usually, see below)
211  * => we get a write lock on the maps and clear needs_copy for a VA
212  * => if we are out of RAM we sleep (waiting for more)
213  */
214 
215 static void
216 uvmfault_amapcopy(struct uvm_faultinfo *ufi)
217 {
218 	for (;;) {
219 
220 		/*
221 		 * no mapping?  give up.
222 		 */
223 
224 		if (uvmfault_lookup(ufi, true) == false)
225 			return;
226 
227 		/*
228 		 * copy if needed.
229 		 */
230 
231 		if (UVM_ET_ISNEEDSCOPY(ufi->entry))
232 			amap_copy(ufi->map, ufi->entry, AMAP_COPY_NOWAIT,
233 				ufi->orig_rvaddr, ufi->orig_rvaddr + 1);
234 
235 		/*
236 		 * didn't work?  must be out of RAM.   unlock and sleep.
237 		 */
238 
239 		if (UVM_ET_ISNEEDSCOPY(ufi->entry)) {
240 			uvmfault_unlockmaps(ufi, true);
241 			uvm_wait("fltamapcopy");
242 			continue;
243 		}
244 
245 		/*
246 		 * got it!   unlock and return.
247 		 */
248 
249 		uvmfault_unlockmaps(ufi, true);
250 		return;
251 	}
252 	/*NOTREACHED*/
253 }
254 
255 /*
256  * uvmfault_anonget: get data in an anon into a non-busy, non-released
257  * page in that anon.
258  *
259  * => maps, amap, and anon locked by caller.
260  * => if we fail (result != 0) we unlock everything.
261  * => if we are successful, we return with everything still locked.
262  * => we don't move the page on the queues [gets moved later]
263  * => if we allocate a new page [we_own], it gets put on the queues.
264  *    either way, the result is that the page is on the queues at return time
265  * => for pages which are on loan from a uvm_object (and thus are not
266  *    owned by the anon): if successful, we return with the owning object
267  *    locked.   the caller must unlock this object when it unlocks everything
268  *    else.
269  */
270 
271 int
272 uvmfault_anonget(struct uvm_faultinfo *ufi, struct vm_amap *amap,
273     struct vm_anon *anon)
274 {
275 	bool we_own;	/* we own anon's page? */
276 	bool locked;	/* did we relock? */
277 	struct vm_page *pg;
278 	int error;
279 	UVMHIST_FUNC("uvmfault_anonget"); UVMHIST_CALLED(maphist);
280 
281 	KASSERT(mutex_owned(&anon->an_lock));
282 
283 	error = 0;
284 	uvmexp.fltanget++;
285         /* bump rusage counters */
286 	if (anon->an_page)
287 		curlwp->l_ru.ru_minflt++;
288 	else
289 		curlwp->l_ru.ru_majflt++;
290 
291 	/*
292 	 * loop until we get it, or fail.
293 	 */
294 
295 	for (;;) {
296 		we_own = false;		/* true if we set PG_BUSY on a page */
297 		pg = anon->an_page;
298 
299 		/*
300 		 * if there is a resident page and it is loaned, then anon
301 		 * may not own it.   call out to uvm_anon_lockpage() to ensure
302 		 * the real owner of the page has been identified and locked.
303 		 */
304 
305 		if (pg && pg->loan_count)
306 			pg = uvm_anon_lockloanpg(anon);
307 
308 		/*
309 		 * page there?   make sure it is not busy/released.
310 		 */
311 
312 		if (pg) {
313 
314 			/*
315 			 * at this point, if the page has a uobject [meaning
316 			 * we have it on loan], then that uobject is locked
317 			 * by us!   if the page is busy, we drop all the
318 			 * locks (including uobject) and try again.
319 			 */
320 
321 			if ((pg->flags & PG_BUSY) == 0) {
322 				UVMHIST_LOG(maphist, "<- OK",0,0,0,0);
323 				return (0);
324 			}
325 			pg->flags |= PG_WANTED;
326 			uvmexp.fltpgwait++;
327 
328 			/*
329 			 * the last unlock must be an atomic unlock+wait on
330 			 * the owner of page
331 			 */
332 
333 			if (pg->uobject) {	/* owner is uobject ? */
334 				uvmfault_unlockall(ufi, amap, NULL, anon);
335 				UVMHIST_LOG(maphist, " unlock+wait on uobj",0,
336 				    0,0,0);
337 				UVM_UNLOCK_AND_WAIT(pg,
338 				    &pg->uobject->vmobjlock,
339 				    false, "anonget1",0);
340 			} else {
341 				/* anon owns page */
342 				uvmfault_unlockall(ufi, amap, NULL, NULL);
343 				UVMHIST_LOG(maphist, " unlock+wait on anon",0,
344 				    0,0,0);
345 				UVM_UNLOCK_AND_WAIT(pg,&anon->an_lock,0,
346 				    "anonget2",0);
347 			}
348 		} else {
349 #if defined(VMSWAP)
350 
351 			/*
352 			 * no page, we must try and bring it in.
353 			 */
354 
355 			pg = uvm_pagealloc(NULL,
356 			    ufi != NULL ? ufi->orig_rvaddr : 0,
357 			    anon, UVM_FLAG_COLORMATCH);
358 			if (pg == NULL) {		/* out of RAM.  */
359 				uvmfault_unlockall(ufi, amap, NULL, anon);
360 				uvmexp.fltnoram++;
361 				UVMHIST_LOG(maphist, "  noram -- UVM_WAIT",0,
362 				    0,0,0);
363 				if (!uvm_reclaimable()) {
364 					return ENOMEM;
365 				}
366 				uvm_wait("flt_noram1");
367 			} else {
368 				/* we set the PG_BUSY bit */
369 				we_own = true;
370 				uvmfault_unlockall(ufi, amap, NULL, anon);
371 
372 				/*
373 				 * we are passing a PG_BUSY+PG_FAKE+PG_CLEAN
374 				 * page into the uvm_swap_get function with
375 				 * all data structures unlocked.  note that
376 				 * it is ok to read an_swslot here because
377 				 * we hold PG_BUSY on the page.
378 				 */
379 				uvmexp.pageins++;
380 				error = uvm_swap_get(pg, anon->an_swslot,
381 				    PGO_SYNCIO);
382 
383 				/*
384 				 * we clean up after the i/o below in the
385 				 * "we_own" case
386 				 */
387 			}
388 #else /* defined(VMSWAP) */
389 			panic("%s: no page", __func__);
390 #endif /* defined(VMSWAP) */
391 		}
392 
393 		/*
394 		 * now relock and try again
395 		 */
396 
397 		locked = uvmfault_relock(ufi);
398 		if (locked && amap != NULL) {
399 			amap_lock(amap);
400 		}
401 		if (locked || we_own)
402 			mutex_enter(&anon->an_lock);
403 
404 		/*
405 		 * if we own the page (i.e. we set PG_BUSY), then we need
406 		 * to clean up after the I/O. there are three cases to
407 		 * consider:
408 		 *   [1] page released during I/O: free anon and ReFault.
409 		 *   [2] I/O not OK.   free the page and cause the fault
410 		 *       to fail.
411 		 *   [3] I/O OK!   activate the page and sync with the
412 		 *       non-we_own case (i.e. drop anon lock if not locked).
413 		 */
414 
415 		if (we_own) {
416 #if defined(VMSWAP)
417 			if (pg->flags & PG_WANTED) {
418 				wakeup(pg);
419 			}
420 			if (error) {
421 
422 				/*
423 				 * remove the swap slot from the anon
424 				 * and mark the anon as having no real slot.
425 				 * don't free the swap slot, thus preventing
426 				 * it from being used again.
427 				 */
428 
429 				if (anon->an_swslot > 0)
430 					uvm_swap_markbad(anon->an_swslot, 1);
431 				anon->an_swslot = SWSLOT_BAD;
432 
433 				if ((pg->flags & PG_RELEASED) != 0)
434 					goto released;
435 
436 				/*
437 				 * note: page was never !PG_BUSY, so it
438 				 * can't be mapped and thus no need to
439 				 * pmap_page_protect it...
440 				 */
441 
442 				mutex_enter(&uvm_pageqlock);
443 				uvm_pagefree(pg);
444 				mutex_exit(&uvm_pageqlock);
445 
446 				if (locked)
447 					uvmfault_unlockall(ufi, amap, NULL,
448 					    anon);
449 				else
450 					mutex_exit(&anon->an_lock);
451 				UVMHIST_LOG(maphist, "<- ERROR", 0,0,0,0);
452 				return error;
453 			}
454 
455 			if ((pg->flags & PG_RELEASED) != 0) {
456 released:
457 				KASSERT(anon->an_ref == 0);
458 
459 				/*
460 				 * released while we unlocked amap.
461 				 */
462 
463 				if (locked)
464 					uvmfault_unlockall(ufi, amap, NULL,
465 					    NULL);
466 
467 				uvm_anon_release(anon);
468 
469 				if (error) {
470 					UVMHIST_LOG(maphist,
471 					    "<- ERROR/RELEASED", 0,0,0,0);
472 					return error;
473 				}
474 
475 				UVMHIST_LOG(maphist, "<- RELEASED", 0,0,0,0);
476 				return ERESTART;
477 			}
478 
479 			/*
480 			 * we've successfully read the page, activate it.
481 			 */
482 
483 			mutex_enter(&uvm_pageqlock);
484 			uvm_pageactivate(pg);
485 			mutex_exit(&uvm_pageqlock);
486 			pg->flags &= ~(PG_WANTED|PG_BUSY|PG_FAKE);
487 			UVM_PAGE_OWN(pg, NULL);
488 			if (!locked)
489 				mutex_exit(&anon->an_lock);
490 #else /* defined(VMSWAP) */
491 			panic("%s: we_own", __func__);
492 #endif /* defined(VMSWAP) */
493 		}
494 
495 		/*
496 		 * we were not able to relock.   restart fault.
497 		 */
498 
499 		if (!locked) {
500 			UVMHIST_LOG(maphist, "<- REFAULT", 0,0,0,0);
501 			return (ERESTART);
502 		}
503 
504 		/*
505 		 * verify no one has touched the amap and moved the anon on us.
506 		 */
507 
508 		if (ufi != NULL &&
509 		    amap_lookup(&ufi->entry->aref,
510 				ufi->orig_rvaddr - ufi->entry->start) != anon) {
511 
512 			uvmfault_unlockall(ufi, amap, NULL, anon);
513 			UVMHIST_LOG(maphist, "<- REFAULT", 0,0,0,0);
514 			return (ERESTART);
515 		}
516 
517 		/*
518 		 * try it again!
519 		 */
520 
521 		uvmexp.fltanretry++;
522 		continue;
523 	}
524 	/*NOTREACHED*/
525 }
526 
527 /*
528  * uvmfault_promote: promote data to a new anon.  used for 1B and 2B.
529  *
530  *	1. allocate an anon and a page.
531  *	2. fill its contents.
532  *	3. put it into amap.
533  *
534  * => if we fail (result != 0) we unlock everything.
535  * => on success, return a new locked anon via 'nanon'.
536  *    (*nanon)->an_page will be a resident, locked, dirty page.
537  * => it's caller's responsibility to put the promoted nanon->an_page to the
538  *    page queue.
539  */
540 
541 static int
542 uvmfault_promote(struct uvm_faultinfo *ufi,
543     struct vm_anon *oanon,
544     struct vm_page *uobjpage,
545     struct vm_anon **nanon, /* OUT: allocated anon */
546     struct vm_anon **spare)
547 {
548 	struct vm_amap *amap = ufi->entry->aref.ar_amap;
549 	struct uvm_object *uobj;
550 	struct vm_anon *anon;
551 	struct vm_page *pg;
552 	struct vm_page *opg;
553 	int error;
554 	UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist);
555 
556 	if (oanon) {
557 		/* anon COW */
558 		opg = oanon->an_page;
559 		KASSERT(opg != NULL);
560 		KASSERT(opg->uobject == NULL || opg->loan_count > 0);
561 	} else if (uobjpage != PGO_DONTCARE) {
562 		/* object-backed COW */
563 		opg = uobjpage;
564 	} else {
565 		/* ZFOD */
566 		opg = NULL;
567 	}
568 	if (opg != NULL) {
569 		uobj = opg->uobject;
570 	} else {
571 		uobj = NULL;
572 	}
573 
574 	KASSERT(amap != NULL);
575 	KASSERT(uobjpage != NULL);
576 	KASSERT(uobjpage == PGO_DONTCARE || (uobjpage->flags & PG_BUSY) != 0);
577 	KASSERT(mutex_owned(&amap->am_l));
578 	KASSERT(oanon == NULL || mutex_owned(&oanon->an_lock));
579 	KASSERT(uobj == NULL || mutex_owned(&uobj->vmobjlock));
580 #if 0
581 	KASSERT(*spare == NULL || !mutex_owned(&(*spare)->an_lock));
582 #endif
583 
584 	if (*spare != NULL) {
585 		anon = *spare;
586 		*spare = NULL;
587 		mutex_enter(&anon->an_lock);
588 	} else if (ufi->map != kernel_map) {
589 		anon = uvm_analloc();
590 	} else {
591 		UVMHIST_LOG(maphist, "kernel_map, unlock and retry", 0,0,0,0);
592 
593 		/*
594 		 * we can't allocate anons with kernel_map locked.
595 		 */
596 
597 		uvm_page_unbusy(&uobjpage, 1);
598 		uvmfault_unlockall(ufi, amap, uobj, oanon);
599 
600 		*spare = uvm_analloc();
601 		if (*spare == NULL) {
602 			goto nomem;
603 		}
604 		mutex_exit(&(*spare)->an_lock);
605 		error = ERESTART;
606 		goto done;
607 	}
608 	if (anon) {
609 
610 		/*
611 		 * The new anon is locked.
612 		 *
613 		 * if opg == NULL, we want a zero'd, dirty page,
614 		 * so have uvm_pagealloc() do that for us.
615 		 */
616 
617 		pg = uvm_pagealloc(NULL, ufi->orig_rvaddr, anon,
618 		    UVM_FLAG_COLORMATCH | (opg == NULL ? UVM_PGA_ZERO : 0));
619 	} else {
620 		pg = NULL;
621 	}
622 
623 	/*
624 	 * out of memory resources?
625 	 */
626 
627 	if (pg == NULL) {
628 		/* save anon for the next try. */
629 		if (anon != NULL) {
630 			mutex_exit(&anon->an_lock);
631 			*spare = anon;
632 		}
633 
634 		/* unlock and fail ... */
635 		uvm_page_unbusy(&uobjpage, 1);
636 		uvmfault_unlockall(ufi, amap, uobj, oanon);
637 nomem:
638 		if (!uvm_reclaimable()) {
639 			UVMHIST_LOG(maphist, "out of VM", 0,0,0,0);
640 			uvmexp.fltnoanon++;
641 			error = ENOMEM;
642 			goto done;
643 		}
644 
645 		UVMHIST_LOG(maphist, "out of RAM, waiting for more", 0,0,0,0);
646 		uvmexp.fltnoram++;
647 		uvm_wait("flt_noram5");
648 		error = ERESTART;
649 		goto done;
650 	}
651 
652 	/* copy page [pg now dirty] */
653 	if (opg) {
654 		uvm_pagecopy(opg, pg);
655 	}
656 
657 	amap_add(&ufi->entry->aref, ufi->orig_rvaddr - ufi->entry->start, anon,
658 	    oanon != NULL);
659 
660 	*nanon = anon;
661 	error = 0;
662 done:
663 	return error;
664 }
665 
666 
667 /*
668  *   F A U L T   -   m a i n   e n t r y   p o i n t
669  */
670 
671 /*
672  * uvm_fault: page fault handler
673  *
674  * => called from MD code to resolve a page fault
675  * => VM data structures usually should be unlocked.   however, it is
676  *	possible to call here with the main map locked if the caller
677  *	gets a write lock, sets it recusive, and then calls us (c.f.
678  *	uvm_map_pageable).   this should be avoided because it keeps
679  *	the map locked off during I/O.
680  * => MUST NEVER BE CALLED IN INTERRUPT CONTEXT
681  */
682 
683 #define MASK(entry)     (UVM_ET_ISCOPYONWRITE(entry) ? \
684 			 ~VM_PROT_WRITE : VM_PROT_ALL)
685 
686 /* fault_flag values passed from uvm_fault_wire to uvm_fault_internal */
687 #define UVM_FAULT_WIRE		(1 << 0)
688 #define UVM_FAULT_MAXPROT	(1 << 1)
689 
690 struct uvm_faultctx {
691 	vm_prot_t access_type;
692 	vm_prot_t enter_prot;
693 	vaddr_t startva;
694 	int npages;
695 	int centeridx;
696 	struct vm_anon *anon_spare;
697 	bool wire_mapping;
698 	bool narrow;
699 	bool wire_paging;
700 	bool cow_now;
701 	bool promote;
702 };
703 
704 static inline int	uvm_fault_check(
705 			    struct uvm_faultinfo *, struct uvm_faultctx *,
706 			    struct vm_anon ***, bool);
707 
708 static int		uvm_fault_upper(
709 			    struct uvm_faultinfo *, struct uvm_faultctx *,
710 			    struct vm_anon **);
711 static inline int	uvm_fault_upper_lookup(
712 			    struct uvm_faultinfo *, const struct uvm_faultctx *,
713 			    struct vm_anon **, struct vm_page **);
714 static inline void	uvm_fault_upper_neighbor(
715 			    struct uvm_faultinfo *, const struct uvm_faultctx *,
716 			    vaddr_t, struct vm_page *, bool);
717 static inline int	uvm_fault_upper_loan(
718 			    struct uvm_faultinfo *, struct uvm_faultctx *,
719 			    struct vm_anon *, struct uvm_object **);
720 static inline int	uvm_fault_upper_promote(
721 			    struct uvm_faultinfo *, struct uvm_faultctx *,
722 			    struct uvm_object *, struct vm_anon *);
723 static inline int	uvm_fault_upper_direct(
724 			    struct uvm_faultinfo *, struct uvm_faultctx *,
725 			    struct uvm_object *, struct vm_anon *);
726 static int		uvm_fault_upper_enter(
727 			    struct uvm_faultinfo *, const struct uvm_faultctx *,
728 			    struct uvm_object *, struct vm_anon *,
729 			    struct vm_page *, struct vm_anon *);
730 static inline void	uvm_fault_upper_done(
731 			    struct uvm_faultinfo *, const struct uvm_faultctx *,
732 			    struct vm_anon *, struct vm_page *);
733 
734 static int		uvm_fault_lower(
735 			    struct uvm_faultinfo *, struct uvm_faultctx *,
736 			    struct vm_page **);
737 static inline void	uvm_fault_lower_lookup(
738 			    struct uvm_faultinfo *, const struct uvm_faultctx *,
739 			    struct vm_page **);
740 static inline void	uvm_fault_lower_neighbor(
741 			    struct uvm_faultinfo *, const struct uvm_faultctx *,
742 			    vaddr_t, struct vm_page *, bool);
743 static inline int	uvm_fault_lower_io(
744 			    struct uvm_faultinfo *, const struct uvm_faultctx *,
745 			    struct uvm_object **, struct vm_page **);
746 static inline int	uvm_fault_lower_direct(
747 			    struct uvm_faultinfo *, struct uvm_faultctx *,
748 			    struct uvm_object *, struct vm_page *);
749 static inline int	uvm_fault_lower_direct_loan(
750 			    struct uvm_faultinfo *, struct uvm_faultctx *,
751 			    struct uvm_object *, struct vm_page **,
752 			    struct vm_page **);
753 static inline int	uvm_fault_lower_promote(
754 			    struct uvm_faultinfo *, struct uvm_faultctx *,
755 			    struct uvm_object *, struct vm_page *);
756 static int		uvm_fault_lower_enter(
757 			    struct uvm_faultinfo *, const struct uvm_faultctx *,
758 			    struct uvm_object *,
759 			    struct vm_anon *, struct vm_page *);
760 static inline void	uvm_fault_lower_done(
761 			    struct uvm_faultinfo *, const struct uvm_faultctx *,
762 			    struct uvm_object *, struct vm_page *);
763 
764 int
765 uvm_fault_internal(struct vm_map *orig_map, vaddr_t vaddr,
766     vm_prot_t access_type, int fault_flag)
767 {
768 	struct uvm_faultinfo ufi;
769 	struct uvm_faultctx flt = {
770 		.access_type = access_type,
771 
772 		/* don't look for neighborhood * pages on "wire" fault */
773 		.narrow = (fault_flag & UVM_FAULT_WIRE) != 0,
774 
775 		/* "wire" fault causes wiring of both mapping and paging */
776 		.wire_mapping = (fault_flag & UVM_FAULT_WIRE) != 0,
777 		.wire_paging = (fault_flag & UVM_FAULT_WIRE) != 0,
778 	};
779 	const bool maxprot = (fault_flag & UVM_FAULT_MAXPROT) != 0;
780 	struct vm_anon *anons_store[UVM_MAXRANGE], **anons;
781 	struct vm_page *pages_store[UVM_MAXRANGE], **pages;
782 	int error;
783 	UVMHIST_FUNC("uvm_fault"); UVMHIST_CALLED(maphist);
784 
785 	UVMHIST_LOG(maphist, "(map=0x%x, vaddr=0x%x, at=%d, ff=%d)",
786 	      orig_map, vaddr, access_type, fault_flag);
787 
788 	curcpu()->ci_data.cpu_nfault++;
789 
790 	/*
791 	 * init the IN parameters in the ufi
792 	 */
793 
794 	ufi.orig_map = orig_map;
795 	ufi.orig_rvaddr = trunc_page(vaddr);
796 	ufi.orig_size = PAGE_SIZE;	/* can't get any smaller than this */
797 
798 	error = ERESTART;
799 	while (error == ERESTART) { /* ReFault: */
800 		anons = anons_store;
801 		pages = pages_store;
802 
803 		error = uvm_fault_check(&ufi, &flt, &anons, maxprot);
804 		if (error != 0)
805 			continue;
806 
807 		error = uvm_fault_upper_lookup(&ufi, &flt, anons, pages);
808 		if (error != 0)
809 			continue;
810 
811 		if (pages[flt.centeridx] == PGO_DONTCARE)
812 			error = uvm_fault_upper(&ufi, &flt, anons);
813 		else {
814 			struct uvm_object * const uobj =
815 			    ufi.entry->object.uvm_obj;
816 
817 			if (uobj && uobj->pgops->pgo_fault != NULL) {
818 				/*
819 				 * invoke "special" fault routine.
820 				 */
821 				mutex_enter(&uobj->vmobjlock);
822 				/* locked: maps(read), amap(if there), uobj */
823 				error = uobj->pgops->pgo_fault(&ufi,
824 				    flt.startva, pages, flt.npages,
825 				    flt.centeridx, flt.access_type,
826 				    PGO_LOCKED|PGO_SYNCIO);
827 
828 				/*
829 				 * locked: nothing, pgo_fault has unlocked
830 				 * everything
831 				 */
832 
833 				/*
834 				 * object fault routine responsible for
835 				 * pmap_update().
836 				 */
837 			} else {
838 				error = uvm_fault_lower(&ufi, &flt, pages);
839 			}
840 		}
841 	}
842 
843 	if (flt.anon_spare != NULL) {
844 		flt.anon_spare->an_ref--;
845 		uvm_anfree(flt.anon_spare);
846 	}
847 	return error;
848 }
849 
850 /*
851  * uvm_fault_check: check prot, handle needs-copy, etc.
852  *
853  *	1. lookup entry.
854  *	2. check protection.
855  *	3. adjust fault condition (mainly for simulated fault).
856  *	4. handle needs-copy (lazy amap copy).
857  *	5. establish range of interest for neighbor fault (aka pre-fault).
858  *	6. look up anons (if amap exists).
859  *	7. flush pages (if MADV_SEQUENTIAL)
860  *
861  * => called with nothing locked.
862  * => if we fail (result != 0) we unlock everything.
863  * => initialize/adjust many members of flt.
864  */
865 
866 static int
867 uvm_fault_check(
868 	struct uvm_faultinfo *ufi, struct uvm_faultctx *flt,
869 	struct vm_anon ***ranons, bool maxprot)
870 {
871 	struct vm_amap *amap;
872 	struct uvm_object *uobj;
873 	vm_prot_t check_prot;
874 	int nback, nforw;
875 	UVMHIST_FUNC("uvm_fault_check"); UVMHIST_CALLED(maphist);
876 
877 	/*
878 	 * lookup and lock the maps
879 	 */
880 
881 	if (uvmfault_lookup(ufi, false) == false) {
882 		UVMHIST_LOG(maphist, "<- no mapping @ 0x%x", ufi->orig_rvaddr,
883 		    0,0,0);
884 		return EFAULT;
885 	}
886 	/* locked: maps(read) */
887 
888 #ifdef DIAGNOSTIC
889 	if ((ufi->map->flags & VM_MAP_PAGEABLE) == 0) {
890 		printf("Page fault on non-pageable map:\n");
891 		printf("ufi->map = %p\n", ufi->map);
892 		printf("ufi->orig_map = %p\n", ufi->orig_map);
893 		printf("ufi->orig_rvaddr = 0x%lx\n", (u_long) ufi->orig_rvaddr);
894 		panic("uvm_fault: (ufi->map->flags & VM_MAP_PAGEABLE) == 0");
895 	}
896 #endif
897 
898 	/*
899 	 * check protection
900 	 */
901 
902 	check_prot = maxprot ?
903 	    ufi->entry->max_protection : ufi->entry->protection;
904 	if ((check_prot & flt->access_type) != flt->access_type) {
905 		UVMHIST_LOG(maphist,
906 		    "<- protection failure (prot=0x%x, access=0x%x)",
907 		    ufi->entry->protection, flt->access_type, 0, 0);
908 		uvmfault_unlockmaps(ufi, false);
909 		return EACCES;
910 	}
911 
912 	/*
913 	 * "enter_prot" is the protection we want to enter the page in at.
914 	 * for certain pages (e.g. copy-on-write pages) this protection can
915 	 * be more strict than ufi->entry->protection.  "wired" means either
916 	 * the entry is wired or we are fault-wiring the pg.
917 	 */
918 
919 	flt->enter_prot = ufi->entry->protection;
920 	if (VM_MAPENT_ISWIRED(ufi->entry))
921 		flt->wire_mapping = true;
922 
923 	if (flt->wire_mapping) {
924 		flt->access_type = flt->enter_prot; /* full access for wired */
925 		flt->cow_now = (check_prot & VM_PROT_WRITE) != 0;
926 	} else {
927 		flt->cow_now = (flt->access_type & VM_PROT_WRITE) != 0;
928 	}
929 
930 	flt->promote = false;
931 
932 	/*
933 	 * handle "needs_copy" case.   if we need to copy the amap we will
934 	 * have to drop our readlock and relock it with a write lock.  (we
935 	 * need a write lock to change anything in a map entry [e.g.
936 	 * needs_copy]).
937 	 */
938 
939 	if (UVM_ET_ISNEEDSCOPY(ufi->entry)) {
940 		if (flt->cow_now || (ufi->entry->object.uvm_obj == NULL)) {
941 			KASSERT(!maxprot);
942 			/* need to clear */
943 			UVMHIST_LOG(maphist,
944 			    "  need to clear needs_copy and refault",0,0,0,0);
945 			uvmfault_unlockmaps(ufi, false);
946 			uvmfault_amapcopy(ufi);
947 			uvmexp.fltamcopy++;
948 			return ERESTART;
949 
950 		} else {
951 
952 			/*
953 			 * ensure that we pmap_enter page R/O since
954 			 * needs_copy is still true
955 			 */
956 
957 			flt->enter_prot &= ~VM_PROT_WRITE;
958 		}
959 	}
960 
961 	/*
962 	 * identify the players
963 	 */
964 
965 	amap = ufi->entry->aref.ar_amap;	/* upper layer */
966 	uobj = ufi->entry->object.uvm_obj;	/* lower layer */
967 
968 	/*
969 	 * check for a case 0 fault.  if nothing backing the entry then
970 	 * error now.
971 	 */
972 
973 	if (amap == NULL && uobj == NULL) {
974 		uvmfault_unlockmaps(ufi, false);
975 		UVMHIST_LOG(maphist,"<- no backing store, no overlay",0,0,0,0);
976 		return EFAULT;
977 	}
978 
979 	/*
980 	 * establish range of interest based on advice from mapper
981 	 * and then clip to fit map entry.   note that we only want
982 	 * to do this the first time through the fault.   if we
983 	 * ReFault we will disable this by setting "narrow" to true.
984 	 */
985 
986 	if (flt->narrow == false) {
987 
988 		/* wide fault (!narrow) */
989 		KASSERT(uvmadvice[ufi->entry->advice].advice ==
990 			 ufi->entry->advice);
991 		nback = MIN(uvmadvice[ufi->entry->advice].nback,
992 		    (ufi->orig_rvaddr - ufi->entry->start) >> PAGE_SHIFT);
993 		flt->startva = ufi->orig_rvaddr - (nback << PAGE_SHIFT);
994 		/*
995 		 * note: "-1" because we don't want to count the
996 		 * faulting page as forw
997 		 */
998 		nforw = MIN(uvmadvice[ufi->entry->advice].nforw,
999 			    ((ufi->entry->end - ufi->orig_rvaddr) >>
1000 			     PAGE_SHIFT) - 1);
1001 		flt->npages = nback + nforw + 1;
1002 		flt->centeridx = nback;
1003 
1004 		flt->narrow = true;	/* ensure only once per-fault */
1005 
1006 	} else {
1007 
1008 		/* narrow fault! */
1009 		nback = nforw = 0;
1010 		flt->startva = ufi->orig_rvaddr;
1011 		flt->npages = 1;
1012 		flt->centeridx = 0;
1013 
1014 	}
1015 	/* offset from entry's start to pgs' start */
1016 	const voff_t eoff = flt->startva - ufi->entry->start;
1017 
1018 	/* locked: maps(read) */
1019 	UVMHIST_LOG(maphist, "  narrow=%d, back=%d, forw=%d, startva=0x%x",
1020 		    flt->narrow, nback, nforw, flt->startva);
1021 	UVMHIST_LOG(maphist, "  entry=0x%x, amap=0x%x, obj=0x%x", ufi->entry,
1022 		    amap, uobj, 0);
1023 
1024 	/*
1025 	 * if we've got an amap, lock it and extract current anons.
1026 	 */
1027 
1028 	if (amap) {
1029 		amap_lock(amap);
1030 		amap_lookups(&ufi->entry->aref, eoff, *ranons, flt->npages);
1031 	} else {
1032 		*ranons = NULL;	/* to be safe */
1033 	}
1034 
1035 	/* locked: maps(read), amap(if there) */
1036 	KASSERT(amap == NULL || mutex_owned(&amap->am_l));
1037 
1038 	/*
1039 	 * for MADV_SEQUENTIAL mappings we want to deactivate the back pages
1040 	 * now and then forget about them (for the rest of the fault).
1041 	 */
1042 
1043 	if (ufi->entry->advice == MADV_SEQUENTIAL && nback != 0) {
1044 
1045 		UVMHIST_LOG(maphist, "  MADV_SEQUENTIAL: flushing backpages",
1046 		    0,0,0,0);
1047 		/* flush back-page anons? */
1048 		if (amap)
1049 			uvmfault_anonflush(*ranons, nback);
1050 
1051 		/* flush object? */
1052 		if (uobj) {
1053 			voff_t uoff;
1054 
1055 			uoff = ufi->entry->offset + eoff;
1056 			mutex_enter(&uobj->vmobjlock);
1057 			(void) (uobj->pgops->pgo_put)(uobj, uoff, uoff +
1058 				    (nback << PAGE_SHIFT), PGO_DEACTIVATE);
1059 		}
1060 
1061 		/* now forget about the backpages */
1062 		if (amap)
1063 			*ranons += nback;
1064 		flt->startva += (nback << PAGE_SHIFT);
1065 		flt->npages -= nback;
1066 		flt->centeridx = 0;
1067 	}
1068 	/*
1069 	 * => startva is fixed
1070 	 * => npages is fixed
1071 	 */
1072 	KASSERT(flt->startva <= ufi->orig_rvaddr);
1073 	KASSERT(ufi->orig_rvaddr + ufi->orig_size <=
1074 	    flt->startva + (flt->npages << PAGE_SHIFT));
1075 	return 0;
1076 }
1077 
1078 /*
1079  * uvm_fault_upper_lookup: look up existing h/w mapping and amap.
1080  *
1081  * iterate range of interest:
1082  *	1. check if h/w mapping exists.  if yes, we don't care
1083  *	2. check if anon exists.  if not, page is lower.
1084  *	3. if anon exists, enter h/w mapping for neighbors.
1085  *
1086  * => called with amap locked (if exists).
1087  */
1088 
1089 static int
1090 uvm_fault_upper_lookup(
1091 	struct uvm_faultinfo *ufi, const struct uvm_faultctx *flt,
1092 	struct vm_anon **anons, struct vm_page **pages)
1093 {
1094 	struct vm_amap *amap = ufi->entry->aref.ar_amap;
1095 	int lcv;
1096 	vaddr_t currva;
1097 	bool shadowed;
1098 	UVMHIST_FUNC("uvm_fault_upper_lookup"); UVMHIST_CALLED(maphist);
1099 
1100 	/* locked: maps(read), amap(if there) */
1101 	KASSERT(amap == NULL || mutex_owned(&amap->am_l));
1102 
1103 	/*
1104 	 * map in the backpages and frontpages we found in the amap in hopes
1105 	 * of preventing future faults.    we also init the pages[] array as
1106 	 * we go.
1107 	 */
1108 
1109 	currva = flt->startva;
1110 	shadowed = false;
1111 	for (lcv = 0; lcv < flt->npages; lcv++, currva += PAGE_SIZE) {
1112 		/*
1113 		 * don't play with VAs that are already mapped
1114 		 * (except for center)
1115 		 */
1116 		if (lcv != flt->centeridx &&
1117 		    pmap_extract(ufi->orig_map->pmap, currva, NULL)) {
1118 			pages[lcv] = PGO_DONTCARE;
1119 			continue;
1120 		}
1121 
1122 		/*
1123 		 * unmapped or center page.   check if any anon at this level.
1124 		 */
1125 		if (amap == NULL || anons[lcv] == NULL) {
1126 			pages[lcv] = NULL;
1127 			continue;
1128 		}
1129 
1130 		/*
1131 		 * check for present page and map if possible.   re-activate it.
1132 		 */
1133 
1134 		pages[lcv] = PGO_DONTCARE;
1135 		if (lcv == flt->centeridx) {	/* save center for later! */
1136 			shadowed = true;
1137 		} else {
1138 			struct vm_anon *anon = anons[lcv];
1139 
1140 			mutex_enter(&anon->an_lock);
1141 			struct vm_page *pg = anon->an_page;
1142 
1143 			/* ignore loaned and busy pages */
1144 			if (pg != NULL && pg->loan_count == 0 &&
1145 			    (pg->flags & PG_BUSY) == 0)
1146 				uvm_fault_upper_neighbor(ufi, flt, currva,
1147 				    pg, anon->an_ref > 1);
1148 			mutex_exit(&anon->an_lock);
1149 		}
1150 	}
1151 
1152 	/* locked: maps(read), amap(if there) */
1153 	KASSERT(amap == NULL || mutex_owned(&amap->am_l));
1154 	/* (shadowed == true) if there is an anon at the faulting address */
1155 	UVMHIST_LOG(maphist, "  shadowed=%d, will_get=%d", shadowed,
1156 	    (ufi->entry->object.uvm_obj && shadowed != false),0,0);
1157 
1158 	/*
1159 	 * note that if we are really short of RAM we could sleep in the above
1160 	 * call to pmap_enter with everything locked.   bad?
1161 	 *
1162 	 * XXX Actually, that is bad; pmap_enter() should just fail in that
1163 	 * XXX case.  --thorpej
1164 	 */
1165 
1166 	return 0;
1167 }
1168 
1169 /*
1170  * uvm_fault_upper_neighbor: enter single lower neighbor page.
1171  *
1172  * => called with amap and anon locked.
1173  */
1174 
1175 static void
1176 uvm_fault_upper_neighbor(
1177 	struct uvm_faultinfo *ufi, const struct uvm_faultctx *flt,
1178 	vaddr_t currva, struct vm_page *pg, bool readonly)
1179 {
1180 	UVMHIST_FUNC("uvm_fault_upper_neighbor"); UVMHIST_CALLED(maphist);
1181 
1182 	/* locked: amap, anon */
1183 
1184 	mutex_enter(&uvm_pageqlock);
1185 	uvm_pageenqueue(pg);
1186 	mutex_exit(&uvm_pageqlock);
1187 	UVMHIST_LOG(maphist,
1188 	    "  MAPPING: n anon: pm=0x%x, va=0x%x, pg=0x%x",
1189 	    ufi->orig_map->pmap, currva, pg, 0);
1190 	uvmexp.fltnamap++;
1191 
1192 	/*
1193 	 * Since this page isn't the page that's actually faulting,
1194 	 * ignore pmap_enter() failures; it's not critical that we
1195 	 * enter these right now.
1196 	 */
1197 
1198 	(void) pmap_enter(ufi->orig_map->pmap, currva,
1199 	    VM_PAGE_TO_PHYS(pg),
1200 	    readonly ? (flt->enter_prot & ~VM_PROT_WRITE) :
1201 	    flt->enter_prot,
1202 	    PMAP_CANFAIL | (flt->wire_mapping ? PMAP_WIRED : 0));
1203 
1204 	pmap_update(ufi->orig_map->pmap);
1205 }
1206 
1207 /*
1208  * uvm_fault_upper: handle upper fault.
1209  *
1210  *	1. acquire anon lock.
1211  *	2. get anon.  let uvmfault_anonget do the dirty work.
1212  *	3. handle loan.
1213  *	4. dispatch direct or promote handlers.
1214  */
1215 
1216 static int
1217 uvm_fault_upper(
1218 	struct uvm_faultinfo *ufi, struct uvm_faultctx *flt,
1219 	struct vm_anon **anons)
1220 {
1221 	struct vm_amap * const amap = ufi->entry->aref.ar_amap;
1222 	struct vm_anon * const anon = anons[flt->centeridx];
1223 	struct uvm_object *uobj;
1224 	int error;
1225 	UVMHIST_FUNC("uvm_fault_upper"); UVMHIST_CALLED(maphist);
1226 
1227 	/* locked: maps(read), amap */
1228 	KASSERT(mutex_owned(&amap->am_l));
1229 
1230 	/*
1231 	 * handle case 1: fault on an anon in our amap
1232 	 */
1233 
1234 	UVMHIST_LOG(maphist, "  case 1 fault: anon=0x%x", anon, 0,0,0);
1235 	mutex_enter(&anon->an_lock);
1236 
1237 	/* locked: maps(read), amap, anon */
1238 	KASSERT(mutex_owned(&amap->am_l));
1239 	KASSERT(mutex_owned(&anon->an_lock));
1240 
1241 	/*
1242 	 * no matter if we have case 1A or case 1B we are going to need to
1243 	 * have the anon's memory resident.   ensure that now.
1244 	 */
1245 
1246 	/*
1247 	 * let uvmfault_anonget do the dirty work.
1248 	 * if it fails (!OK) it will unlock everything for us.
1249 	 * if it succeeds, locks are still valid and locked.
1250 	 * also, if it is OK, then the anon's page is on the queues.
1251 	 * if the page is on loan from a uvm_object, then anonget will
1252 	 * lock that object for us if it does not fail.
1253 	 */
1254 
1255 	error = uvmfault_anonget(ufi, amap, anon);
1256 	switch (error) {
1257 	case 0:
1258 		break;
1259 
1260 	case ERESTART:
1261 		return ERESTART;
1262 
1263 	case EAGAIN:
1264 		kpause("fltagain1", false, hz/2, NULL);
1265 		return ERESTART;
1266 
1267 	default:
1268 		return error;
1269 	}
1270 
1271 	/*
1272 	 * uobj is non null if the page is on loan from an object (i.e. uobj)
1273 	 */
1274 
1275 	uobj = anon->an_page->uobject;	/* locked by anonget if !NULL */
1276 
1277 	/* locked: maps(read), amap, anon, uobj(if one) */
1278 	KASSERT(mutex_owned(&amap->am_l));
1279 	KASSERT(mutex_owned(&anon->an_lock));
1280 	KASSERT(uobj == NULL || mutex_owned(&uobj->vmobjlock));
1281 
1282 	/*
1283 	 * special handling for loaned pages
1284 	 */
1285 
1286 	if (anon->an_page->loan_count) {
1287 		error = uvm_fault_upper_loan(ufi, flt, anon, &uobj);
1288 		if (error != 0)
1289 			return error;
1290 	}
1291 
1292 	/*
1293 	 * if we are case 1B then we will need to allocate a new blank
1294 	 * anon to transfer the data into.   note that we have a lock
1295 	 * on anon, so no one can busy or release the page until we are done.
1296 	 * also note that the ref count can't drop to zero here because
1297 	 * it is > 1 and we are only dropping one ref.
1298 	 *
1299 	 * in the (hopefully very rare) case that we are out of RAM we
1300 	 * will unlock, wait for more RAM, and refault.
1301 	 *
1302 	 * if we are out of anon VM we kill the process (XXX: could wait?).
1303 	 */
1304 
1305 	if (flt->cow_now && anon->an_ref > 1) {
1306 		flt->promote = true;
1307 		error = uvm_fault_upper_promote(ufi, flt, uobj, anon);
1308 	} else {
1309 		error = uvm_fault_upper_direct(ufi, flt, uobj, anon);
1310 	}
1311 	return error;
1312 }
1313 
1314 /*
1315  * uvm_fault_upper_loan: handle loaned upper page.
1316  *
1317  *	1. if not cow'ing now, simply adjust flt->enter_prot.
1318  *	2. if cow'ing now, and if ref count is 1, break loan.
1319  */
1320 
1321 static int
1322 uvm_fault_upper_loan(
1323 	struct uvm_faultinfo *ufi, struct uvm_faultctx *flt,
1324 	struct vm_anon *anon, struct uvm_object **ruobj)
1325 {
1326 	struct vm_amap * const amap = ufi->entry->aref.ar_amap;
1327 	int error = 0;
1328 	UVMHIST_FUNC("uvm_fault_upper_loan"); UVMHIST_CALLED(maphist);
1329 
1330 	if (!flt->cow_now) {
1331 
1332 		/*
1333 		 * for read faults on loaned pages we just cap the
1334 		 * protection at read-only.
1335 		 */
1336 
1337 		flt->enter_prot = flt->enter_prot & ~VM_PROT_WRITE;
1338 
1339 	} else {
1340 		/*
1341 		 * note that we can't allow writes into a loaned page!
1342 		 *
1343 		 * if we have a write fault on a loaned page in an
1344 		 * anon then we need to look at the anon's ref count.
1345 		 * if it is greater than one then we are going to do
1346 		 * a normal copy-on-write fault into a new anon (this
1347 		 * is not a problem).  however, if the reference count
1348 		 * is one (a case where we would normally allow a
1349 		 * write directly to the page) then we need to kill
1350 		 * the loan before we continue.
1351 		 */
1352 
1353 		/* >1 case is already ok */
1354 		if (anon->an_ref == 1) {
1355 			error = uvm_loanbreak_anon(anon, *ruobj);
1356 			if (error != 0) {
1357 				uvmfault_unlockall(ufi, amap, *ruobj, anon);
1358 				uvm_wait("flt_noram2");
1359 				return ERESTART;
1360 			}
1361 			/* if we were a loan reciever uobj is gone */
1362 			if (*ruobj)
1363 				*ruobj = NULL;
1364 		}
1365 	}
1366 	return error;
1367 }
1368 
1369 /*
1370  * uvm_fault_upper_promote: promote upper page.
1371  *
1372  *	1. call uvmfault_promote.
1373  *	2. enqueue page.
1374  *	3. deref.
1375  *	4. pass page to uvm_fault_upper_enter.
1376  */
1377 
1378 static int
1379 uvm_fault_upper_promote(
1380 	struct uvm_faultinfo *ufi, struct uvm_faultctx *flt,
1381 	struct uvm_object *uobj, struct vm_anon *anon)
1382 {
1383 	struct vm_anon * const oanon = anon;
1384 	struct vm_page *pg;
1385 	int error;
1386 	UVMHIST_FUNC("uvm_fault_upper_promote"); UVMHIST_CALLED(maphist);
1387 
1388 	UVMHIST_LOG(maphist, "  case 1B: COW fault",0,0,0,0);
1389 	uvmexp.flt_acow++;
1390 
1391 	error = uvmfault_promote(ufi, oanon, PGO_DONTCARE, &anon,
1392 	    &flt->anon_spare);
1393 	switch (error) {
1394 	case 0:
1395 		break;
1396 	case ERESTART:
1397 		return ERESTART;
1398 	default:
1399 		return error;
1400 	}
1401 
1402 	pg = anon->an_page;
1403 	mutex_enter(&uvm_pageqlock);
1404 	uvm_pageenqueue(pg); /* uvm_fault_upper_done will activate the page */
1405 	mutex_exit(&uvm_pageqlock);
1406 	pg->flags &= ~(PG_BUSY|PG_FAKE);
1407 	UVM_PAGE_OWN(pg, NULL);
1408 
1409 	/* deref: can not drop to zero here by defn! */
1410 	KASSERT(oanon->an_ref > 1);
1411 	oanon->an_ref--;
1412 
1413 	/*
1414 	 * note: oanon is still locked, as is the new anon.  we
1415 	 * need to check for this later when we unlock oanon; if
1416 	 * oanon != anon, we'll have to unlock anon, too.
1417 	 */
1418 
1419 	return uvm_fault_upper_enter(ufi, flt, uobj, anon, pg, oanon);
1420 }
1421 
1422 /*
1423  * uvm_fault_upper_direct: handle direct fault.
1424  */
1425 
1426 static int
1427 uvm_fault_upper_direct(
1428 	struct uvm_faultinfo *ufi, struct uvm_faultctx *flt,
1429 	struct uvm_object *uobj, struct vm_anon *anon)
1430 {
1431 	struct vm_anon * const oanon = anon;
1432 	struct vm_page *pg;
1433 	UVMHIST_FUNC("uvm_fault_upper_direct"); UVMHIST_CALLED(maphist);
1434 
1435 	uvmexp.flt_anon++;
1436 	pg = anon->an_page;
1437 	if (anon->an_ref > 1)     /* disallow writes to ref > 1 anons */
1438 		flt->enter_prot = flt->enter_prot & ~VM_PROT_WRITE;
1439 
1440 	return uvm_fault_upper_enter(ufi, flt, uobj, anon, pg, oanon);
1441 }
1442 
1443 /*
1444  * uvm_fault_upper_enter: enter h/w mapping of upper page.
1445  */
1446 
1447 static int
1448 uvm_fault_upper_enter(
1449 	struct uvm_faultinfo *ufi, const struct uvm_faultctx *flt,
1450 	struct uvm_object *uobj, struct vm_anon *anon, struct vm_page *pg,
1451 	struct vm_anon *oanon)
1452 {
1453 	struct vm_amap * const amap = ufi->entry->aref.ar_amap;
1454 	UVMHIST_FUNC("uvm_fault_upper_enter"); UVMHIST_CALLED(maphist);
1455 
1456 	/* locked: maps(read), amap, oanon, anon(if different from oanon) */
1457 	KASSERT(mutex_owned(&amap->am_l));
1458 	KASSERT(mutex_owned(&anon->an_lock));
1459 	KASSERT(mutex_owned(&oanon->an_lock));
1460 
1461 	/*
1462 	 * now map the page in.
1463 	 */
1464 
1465 	UVMHIST_LOG(maphist,
1466 	    "  MAPPING: anon: pm=0x%x, va=0x%x, pg=0x%x, promote=%d",
1467 	    ufi->orig_map->pmap, ufi->orig_rvaddr, pg, flt->promote);
1468 	if (pmap_enter(ufi->orig_map->pmap, ufi->orig_rvaddr,
1469 	    VM_PAGE_TO_PHYS(pg),
1470 	    flt->enter_prot, flt->access_type | PMAP_CANFAIL |
1471 	    (flt->wire_mapping ? PMAP_WIRED : 0)) != 0) {
1472 
1473 		/*
1474 		 * No need to undo what we did; we can simply think of
1475 		 * this as the pmap throwing away the mapping information.
1476 		 *
1477 		 * We do, however, have to go through the ReFault path,
1478 		 * as the map may change while we're asleep.
1479 		 */
1480 
1481 		if (anon != oanon)
1482 			mutex_exit(&anon->an_lock);
1483 		uvmfault_unlockall(ufi, amap, uobj, oanon);
1484 		if (!uvm_reclaimable()) {
1485 			UVMHIST_LOG(maphist,
1486 			    "<- failed.  out of VM",0,0,0,0);
1487 			/* XXX instrumentation */
1488 			return ENOMEM;
1489 		}
1490 		/* XXX instrumentation */
1491 		uvm_wait("flt_pmfail1");
1492 		return ERESTART;
1493 	}
1494 
1495 	uvm_fault_upper_done(ufi, flt, anon, pg);
1496 
1497 	/*
1498 	 * done case 1!  finish up by unlocking everything and returning success
1499 	 */
1500 
1501 	if (anon != oanon) {
1502 		mutex_exit(&anon->an_lock);
1503 	}
1504 	pmap_update(ufi->orig_map->pmap);
1505 	uvmfault_unlockall(ufi, amap, uobj, oanon);
1506 	return 0;
1507 }
1508 
1509 /*
1510  * uvm_fault_upper_done: queue upper center page.
1511  */
1512 
1513 static void
1514 uvm_fault_upper_done(
1515 	struct uvm_faultinfo *ufi, const struct uvm_faultctx *flt,
1516 	struct vm_anon *anon, struct vm_page *pg)
1517 {
1518 	const bool wire_paging = flt->wire_paging;
1519 
1520 	UVMHIST_FUNC("uvm_fault_upper_done"); UVMHIST_CALLED(maphist);
1521 
1522 	/*
1523 	 * ... update the page queues.
1524 	 */
1525 
1526 	mutex_enter(&uvm_pageqlock);
1527 	if (wire_paging) {
1528 		uvm_pagewire(pg);
1529 
1530 		/*
1531 		 * since the now-wired page cannot be paged out,
1532 		 * release its swap resources for others to use.
1533 		 * since an anon with no swap cannot be PG_CLEAN,
1534 		 * clear its clean flag now.
1535 		 */
1536 
1537 		pg->flags &= ~(PG_CLEAN);
1538 
1539 	} else {
1540 		uvm_pageactivate(pg);
1541 	}
1542 	mutex_exit(&uvm_pageqlock);
1543 
1544 	if (wire_paging) {
1545 		uvm_anon_dropswap(anon);
1546 	}
1547 }
1548 
1549 /*
1550  * uvm_fault_lower: handle lower fault.
1551  *
1552  *	1. check uobj
1553  *	1.1. if null, ZFOD.
1554  *	1.2. if not null, look up unnmapped neighbor pages.
1555  *	2. for center page, check if promote.
1556  *	2.1. ZFOD always needs promotion.
1557  *	2.2. other uobjs, when entry is marked COW (usually MAP_PRIVATE vnode).
1558  *	3. if uobj is not ZFOD and page is not found, do i/o.
1559  *	4. dispatch either direct / promote fault.
1560  */
1561 
1562 static int
1563 uvm_fault_lower(
1564 	struct uvm_faultinfo *ufi, struct uvm_faultctx *flt,
1565 	struct vm_page **pages)
1566 {
1567 #ifdef DIAGNOSTIC
1568 	struct vm_amap *amap = ufi->entry->aref.ar_amap;
1569 #endif
1570 	struct uvm_object *uobj = ufi->entry->object.uvm_obj;
1571 	struct vm_page *uobjpage;
1572 	int error;
1573 	UVMHIST_FUNC("uvm_fault_lower"); UVMHIST_CALLED(maphist);
1574 
1575 	/* locked: maps(read), amap(if there), uobj(if !null) */
1576 
1577 	/*
1578 	 * now, if the desired page is not shadowed by the amap and we have
1579 	 * a backing object that does not have a special fault routine, then
1580 	 * we ask (with pgo_get) the object for resident pages that we care
1581 	 * about and attempt to map them in.  we do not let pgo_get block
1582 	 * (PGO_LOCKED).
1583 	 */
1584 
1585 	if (uobj == NULL) {
1586 		/* zero fill; don't care neighbor pages */
1587 		uobjpage = NULL;
1588 	} else {
1589 		uvm_fault_lower_lookup(ufi, flt, pages);
1590 		uobjpage = pages[flt->centeridx];
1591 	}
1592 
1593 	/*
1594 	 * note that at this point we are done with any front or back pages.
1595 	 * we are now going to focus on the center page (i.e. the one we've
1596 	 * faulted on).  if we have faulted on the upper (anon) layer
1597 	 * [i.e. case 1], then the anon we want is anons[centeridx] (we have
1598 	 * not touched it yet).  if we have faulted on the bottom (uobj)
1599 	 * layer [i.e. case 2] and the page was both present and available,
1600 	 * then we've got a pointer to it as "uobjpage" and we've already
1601 	 * made it BUSY.
1602 	 */
1603 
1604 	/*
1605 	 * locked:
1606 	 * maps(read), amap(if there), uobj(if !null), uobjpage(if !null)
1607 	 */
1608 	KASSERT(amap == NULL || mutex_owned(&amap->am_l));
1609 	KASSERT(uobj == NULL || mutex_owned(&uobj->vmobjlock));
1610 	KASSERT(uobjpage == NULL || (uobjpage->flags & PG_BUSY) != 0);
1611 
1612 	/*
1613 	 * note that uobjpage can not be PGO_DONTCARE at this point.  we now
1614 	 * set uobjpage to PGO_DONTCARE if we are doing a zero fill.  if we
1615 	 * have a backing object, check and see if we are going to promote
1616 	 * the data up to an anon during the fault.
1617 	 */
1618 
1619 	if (uobj == NULL) {
1620 		uobjpage = PGO_DONTCARE;
1621 		flt->promote = true;		/* always need anon here */
1622 	} else {
1623 		KASSERT(uobjpage != PGO_DONTCARE);
1624 		flt->promote = flt->cow_now && UVM_ET_ISCOPYONWRITE(ufi->entry);
1625 	}
1626 	UVMHIST_LOG(maphist, "  case 2 fault: promote=%d, zfill=%d",
1627 	    flt->promote, (uobj == NULL), 0,0);
1628 
1629 	/*
1630 	 * if uobjpage is not null then we do not need to do I/O to get the
1631 	 * uobjpage.
1632 	 *
1633 	 * if uobjpage is null, then we need to unlock and ask the pager to
1634 	 * get the data for us.   once we have the data, we need to reverify
1635 	 * the state the world.   we are currently not holding any resources.
1636 	 */
1637 
1638 	if (uobjpage) {
1639 		/* update rusage counters */
1640 		curlwp->l_ru.ru_minflt++;
1641 	} else {
1642 		error = uvm_fault_lower_io(ufi, flt, &uobj, &uobjpage);
1643 		if (error != 0)
1644 			return error;
1645 	}
1646 
1647 	/*
1648 	 * locked:
1649 	 * maps(read), amap(if !null), uobj(if !null), uobjpage(if uobj)
1650 	 */
1651 	KASSERT(amap == NULL || mutex_owned(&amap->am_l));
1652 	KASSERT(uobj == NULL || mutex_owned(&uobj->vmobjlock));
1653 	KASSERT(uobj == NULL || (uobjpage->flags & PG_BUSY) != 0);
1654 
1655 	/*
1656 	 * notes:
1657 	 *  - at this point uobjpage can not be NULL
1658 	 *  - at this point uobjpage can not be PG_RELEASED (since we checked
1659 	 *  for it above)
1660 	 *  - at this point uobjpage could be PG_WANTED (handle later)
1661 	 */
1662 
1663 	KASSERT(uobjpage != NULL);
1664 	KASSERT(uobj == NULL || uobj == uobjpage->uobject);
1665 	KASSERT(uobj == NULL || !UVM_OBJ_IS_CLEAN(uobjpage->uobject) ||
1666 	    (uobjpage->flags & PG_CLEAN) != 0);
1667 
1668 	if (!flt->promote) {
1669 		error = uvm_fault_lower_direct(ufi, flt, uobj, uobjpage);
1670 	} else {
1671 		error = uvm_fault_lower_promote(ufi, flt, uobj, uobjpage);
1672 	}
1673 	return error;
1674 }
1675 
1676 /*
1677  * uvm_fault_lower_lookup: look up on-memory uobj pages.
1678  *
1679  *	1. get on-memory pages.
1680  *	2. if failed, give up (get only center page later).
1681  *	3. if succeeded, enter h/w mapping of neighbor pages.
1682  */
1683 
1684 static void
1685 uvm_fault_lower_lookup(
1686 	struct uvm_faultinfo *ufi, const struct uvm_faultctx *flt,
1687 	struct vm_page **pages)
1688 {
1689 	struct uvm_object *uobj = ufi->entry->object.uvm_obj;
1690 	int lcv, gotpages;
1691 	vaddr_t currva;
1692 	UVMHIST_FUNC("uvm_fault_lower_lookup"); UVMHIST_CALLED(maphist);
1693 
1694 	mutex_enter(&uobj->vmobjlock);
1695 	/* locked: maps(read), amap(if there), uobj */
1696 	/*
1697 	 * the following call to pgo_get does _not_ change locking state
1698 	 */
1699 
1700 	uvmexp.fltlget++;
1701 	gotpages = flt->npages;
1702 	(void) uobj->pgops->pgo_get(uobj,
1703 	    ufi->entry->offset + flt->startva - ufi->entry->start,
1704 	    pages, &gotpages, flt->centeridx,
1705 	    flt->access_type & MASK(ufi->entry), ufi->entry->advice, PGO_LOCKED);
1706 
1707 	/*
1708 	 * check for pages to map, if we got any
1709 	 */
1710 
1711 	if (gotpages == 0) {
1712 		pages[flt->centeridx] = NULL;
1713 		return;
1714 	}
1715 
1716 	currva = flt->startva;
1717 	for (lcv = 0; lcv < flt->npages; lcv++, currva += PAGE_SIZE) {
1718 		struct vm_page *curpg;
1719 
1720 		curpg = pages[lcv];
1721 		if (curpg == NULL || curpg == PGO_DONTCARE) {
1722 			continue;
1723 		}
1724 		KASSERT(curpg->uobject == uobj);
1725 
1726 		/*
1727 		 * if center page is resident and not PG_BUSY|PG_RELEASED
1728 		 * then pgo_get made it PG_BUSY for us and gave us a handle
1729 		 * to it.
1730 		 */
1731 
1732 		if (lcv == flt->centeridx) {
1733 			UVMHIST_LOG(maphist, "  got uobjpage "
1734 			    "(0x%x) with locked get",
1735 			    curpg, 0,0,0);
1736 		} else {
1737 			bool readonly = (curpg->flags & PG_RDONLY)
1738 			    || (curpg->loan_count > 0)
1739 			    || UVM_OBJ_NEEDS_WRITEFAULT(curpg->uobject);
1740 
1741 			uvm_fault_lower_neighbor(ufi, flt,
1742 			    currva, curpg, readonly);
1743 		}
1744 	}
1745 	pmap_update(ufi->orig_map->pmap);
1746 }
1747 
1748 /*
1749  * uvm_fault_lower_neighbor: enter h/w mapping of lower neighbor page.
1750  */
1751 
1752 static void
1753 uvm_fault_lower_neighbor(
1754 	struct uvm_faultinfo *ufi, const struct uvm_faultctx *flt,
1755 	vaddr_t currva, struct vm_page *pg, bool readonly)
1756 {
1757 	UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist);
1758 
1759 	/* locked: maps(read), amap(if there), uobj */
1760 
1761 	/*
1762 	 * calling pgo_get with PGO_LOCKED returns us pages which
1763 	 * are neither busy nor released, so we don't need to check
1764 	 * for this.  we can just directly enter the pages.
1765 	 */
1766 
1767 	mutex_enter(&uvm_pageqlock);
1768 	uvm_pageenqueue(pg);
1769 	mutex_exit(&uvm_pageqlock);
1770 	UVMHIST_LOG(maphist,
1771 	    "  MAPPING: n obj: pm=0x%x, va=0x%x, pg=0x%x",
1772 	    ufi->orig_map->pmap, currva, pg, 0);
1773 	uvmexp.fltnomap++;
1774 
1775 	/*
1776 	 * Since this page isn't the page that's actually faulting,
1777 	 * ignore pmap_enter() failures; it's not critical that we
1778 	 * enter these right now.
1779 	 * NOTE: page can't be PG_WANTED or PG_RELEASED because we've
1780 	 * held the lock the whole time we've had the handle.
1781 	 */
1782 	KASSERT((pg->flags & PG_PAGEOUT) == 0);
1783 	KASSERT((pg->flags & PG_RELEASED) == 0);
1784 	KASSERT((pg->flags & PG_WANTED) == 0);
1785 	KASSERT(!UVM_OBJ_IS_CLEAN(pg->uobject) ||
1786 	    (pg->flags & PG_CLEAN) != 0);
1787 	pg->flags &= ~(PG_BUSY);
1788 	UVM_PAGE_OWN(pg, NULL);
1789 
1790 	(void) pmap_enter(ufi->orig_map->pmap, currva,
1791 	    VM_PAGE_TO_PHYS(pg),
1792 	    readonly ? (flt->enter_prot & ~VM_PROT_WRITE) :
1793 	    flt->enter_prot & MASK(ufi->entry),
1794 	    PMAP_CANFAIL | (flt->wire_mapping ? PMAP_WIRED : 0));
1795 }
1796 
1797 /*
1798  * uvm_fault_lower_io: get lower page from backing store.
1799  *
1800  *	1. unlock everything, because i/o will block.
1801  *	2. call pgo_get.
1802  *	3. if failed, recover.
1803  *	4. if succeeded, relock everything and verify things.
1804  */
1805 
1806 static int
1807 uvm_fault_lower_io(
1808 	struct uvm_faultinfo *ufi, const struct uvm_faultctx *flt,
1809 	struct uvm_object **ruobj, struct vm_page **ruobjpage)
1810 {
1811 	struct vm_amap * const amap = ufi->entry->aref.ar_amap;
1812 	struct uvm_object *uobj = *ruobj;
1813 	struct vm_page *pg;
1814 	bool locked;
1815 	int gotpages;
1816 	int error;
1817 	voff_t uoff;
1818 	UVMHIST_FUNC("uvm_fault_lower_io"); UVMHIST_CALLED(maphist);
1819 
1820 	/* update rusage counters */
1821 	curlwp->l_ru.ru_majflt++;
1822 
1823 	/* locked: maps(read), amap(if there), uobj */
1824 	uvmfault_unlockall(ufi, amap, NULL, NULL);
1825 	/* locked: uobj */
1826 
1827 	uvmexp.fltget++;
1828 	gotpages = 1;
1829 	pg = NULL;
1830 	uoff = (ufi->orig_rvaddr - ufi->entry->start) + ufi->entry->offset;
1831 	error = uobj->pgops->pgo_get(uobj, uoff, &pg, &gotpages,
1832 	    0, flt->access_type & MASK(ufi->entry), ufi->entry->advice,
1833 	    PGO_SYNCIO);
1834 	/* locked: pg(if no error) */
1835 
1836 	/*
1837 	 * recover from I/O
1838 	 */
1839 
1840 	if (error) {
1841 		if (error == EAGAIN) {
1842 			UVMHIST_LOG(maphist,
1843 			    "  pgo_get says TRY AGAIN!",0,0,0,0);
1844 			kpause("fltagain2", false, hz/2, NULL);
1845 			return ERESTART;
1846 		}
1847 
1848 #if 0
1849 		KASSERT(error != ERESTART);
1850 #else
1851 		/* XXXUEBS don't re-fault? */
1852 		if (error == ERESTART)
1853 			error = EIO;
1854 #endif
1855 
1856 		UVMHIST_LOG(maphist, "<- pgo_get failed (code %d)",
1857 		    error, 0,0,0);
1858 		return error;
1859 	}
1860 
1861 	/* locked: pg */
1862 
1863 	KASSERT((pg->flags & PG_BUSY) != 0);
1864 
1865 	mutex_enter(&uvm_pageqlock);
1866 	uvm_pageactivate(pg);
1867 	mutex_exit(&uvm_pageqlock);
1868 
1869 	/*
1870 	 * re-verify the state of the world by first trying to relock
1871 	 * the maps.  always relock the object.
1872 	 */
1873 
1874 	locked = uvmfault_relock(ufi);
1875 	if (locked && amap)
1876 		amap_lock(amap);
1877 
1878 	/* might be changed */
1879 	uobj = pg->uobject;
1880 
1881 	mutex_enter(&uobj->vmobjlock);
1882 
1883 	/* locked(locked): maps(read), amap(if !null), uobj, pg */
1884 	/* locked(!locked): uobj, pg */
1885 
1886 	/*
1887 	 * verify that the page has not be released and re-verify
1888 	 * that amap slot is still free.   if there is a problem,
1889 	 * we unlock and clean up.
1890 	 */
1891 
1892 	if ((pg->flags & PG_RELEASED) != 0 ||
1893 	    (locked && amap && amap_lookup(&ufi->entry->aref,
1894 	      ufi->orig_rvaddr - ufi->entry->start))) {
1895 		if (locked)
1896 			uvmfault_unlockall(ufi, amap, NULL, NULL);
1897 		locked = false;
1898 	}
1899 
1900 	/*
1901 	 * didn't get the lock?   release the page and retry.
1902 	 */
1903 
1904 	if (locked == false) {
1905 		UVMHIST_LOG(maphist,
1906 		    "  wasn't able to relock after fault: retry",
1907 		    0,0,0,0);
1908 		if (pg->flags & PG_WANTED) {
1909 			wakeup(pg);
1910 		}
1911 		if (pg->flags & PG_RELEASED) {
1912 			uvmexp.fltpgrele++;
1913 			uvm_pagefree(pg);
1914 			mutex_exit(&uobj->vmobjlock);
1915 			return ERESTART;
1916 		}
1917 		pg->flags &= ~(PG_BUSY|PG_WANTED);
1918 		UVM_PAGE_OWN(pg, NULL);
1919 		mutex_exit(&uobj->vmobjlock);
1920 		return ERESTART;
1921 	}
1922 
1923 	/*
1924 	 * we have the data in pg which is busy and
1925 	 * not released.  we are holding object lock (so the page
1926 	 * can't be released on us).
1927 	 */
1928 
1929 	/* locked: maps(read), amap(if !null), uobj, pg */
1930 
1931 	*ruobj = uobj;
1932 	*ruobjpage = pg;
1933 	return 0;
1934 }
1935 
1936 /*
1937  * uvm_fault_lower_direct: fault lower center page
1938  *
1939  *	1. adjust flt->enter_prot.
1940  *	2. if page is loaned, resolve.
1941  */
1942 
1943 int
1944 uvm_fault_lower_direct(
1945 	struct uvm_faultinfo *ufi, struct uvm_faultctx *flt,
1946 	struct uvm_object *uobj, struct vm_page *uobjpage)
1947 {
1948 	struct vm_page *pg;
1949 	UVMHIST_FUNC("uvm_fault_lower_direct"); UVMHIST_CALLED(maphist);
1950 
1951 	/*
1952 	 * we are not promoting.   if the mapping is COW ensure that we
1953 	 * don't give more access than we should (e.g. when doing a read
1954 	 * fault on a COPYONWRITE mapping we want to map the COW page in
1955 	 * R/O even though the entry protection could be R/W).
1956 	 *
1957 	 * set "pg" to the page we want to map in (uobjpage, usually)
1958 	 */
1959 
1960 	uvmexp.flt_obj++;
1961 	if (UVM_ET_ISCOPYONWRITE(ufi->entry) ||
1962 	    UVM_OBJ_NEEDS_WRITEFAULT(uobjpage->uobject))
1963 		flt->enter_prot &= ~VM_PROT_WRITE;
1964 	pg = uobjpage;		/* map in the actual object */
1965 
1966 	KASSERT(uobjpage != PGO_DONTCARE);
1967 
1968 	/*
1969 	 * we are faulting directly on the page.   be careful
1970 	 * about writing to loaned pages...
1971 	 */
1972 
1973 	if (uobjpage->loan_count) {
1974 		uvm_fault_lower_direct_loan(ufi, flt, uobj, &pg, &uobjpage);
1975 	}
1976 	KASSERT(pg == uobjpage);
1977 
1978 	KASSERT(uobj == NULL || (uobjpage->flags & PG_BUSY) != 0);
1979 	return uvm_fault_lower_enter(ufi, flt, uobj, NULL, pg);
1980 }
1981 
1982 /*
1983  * uvm_fault_lower_direct_loan: resolve loaned page.
1984  *
1985  *	1. if not cow'ing, adjust flt->enter_prot.
1986  *	2. if cow'ing, break loan.
1987  */
1988 
1989 static int
1990 uvm_fault_lower_direct_loan(
1991 	struct uvm_faultinfo *ufi, struct uvm_faultctx *flt,
1992 	struct uvm_object *uobj, struct vm_page **rpg,
1993 	struct vm_page **ruobjpage)
1994 {
1995 	struct vm_amap * const amap = ufi->entry->aref.ar_amap;
1996 	struct vm_page *pg;
1997 	struct vm_page *uobjpage = *ruobjpage;
1998 	UVMHIST_FUNC("uvm_fault_lower_direct_loan"); UVMHIST_CALLED(maphist);
1999 
2000 	if (!flt->cow_now) {
2001 		/* read fault: cap the protection at readonly */
2002 		/* cap! */
2003 		flt->enter_prot = flt->enter_prot & ~VM_PROT_WRITE;
2004 	} else {
2005 		/* write fault: must break the loan here */
2006 
2007 		pg = uvm_loanbreak(uobjpage);
2008 		if (pg == NULL) {
2009 
2010 			/*
2011 			 * drop ownership of page, it can't be released
2012 			 */
2013 
2014 			if (uobjpage->flags & PG_WANTED)
2015 				wakeup(uobjpage);
2016 			uobjpage->flags &= ~(PG_BUSY|PG_WANTED);
2017 			UVM_PAGE_OWN(uobjpage, NULL);
2018 
2019 			uvmfault_unlockall(ufi, amap, uobj, NULL);
2020 			UVMHIST_LOG(maphist,
2021 			  "  out of RAM breaking loan, waiting",
2022 			  0,0,0,0);
2023 			uvmexp.fltnoram++;
2024 			uvm_wait("flt_noram4");
2025 			return ERESTART;
2026 		}
2027 		*rpg = pg;
2028 		*ruobjpage = pg;
2029 	}
2030 	return 0;
2031 }
2032 
2033 /*
2034  * uvm_fault_lower_promote: promote lower page.
2035  *
2036  *	1. call uvmfault_promote.
2037  *	2. fill in data.
2038  *	3. if not ZFOD, dispose old page.
2039  */
2040 
2041 int
2042 uvm_fault_lower_promote(
2043 	struct uvm_faultinfo *ufi, struct uvm_faultctx *flt,
2044 	struct uvm_object *uobj, struct vm_page *uobjpage)
2045 {
2046 	struct vm_amap * const amap = ufi->entry->aref.ar_amap;
2047 	struct vm_anon *anon;
2048 	struct vm_page *pg;
2049 	int error;
2050 	UVMHIST_FUNC("uvm_fault_lower_promote"); UVMHIST_CALLED(maphist);
2051 
2052 	/*
2053 	 * if we are going to promote the data to an anon we
2054 	 * allocate a blank anon here and plug it into our amap.
2055 	 */
2056 #if DIAGNOSTIC
2057 	if (amap == NULL)
2058 		panic("uvm_fault: want to promote data, but no anon");
2059 #endif
2060 	error = uvmfault_promote(ufi, NULL, uobjpage,
2061 	    &anon, &flt->anon_spare);
2062 	switch (error) {
2063 	case 0:
2064 		break;
2065 	case ERESTART:
2066 		return ERESTART;
2067 	default:
2068 		return error;
2069 	}
2070 
2071 	pg = anon->an_page;
2072 
2073 	/*
2074 	 * fill in the data
2075 	 */
2076 
2077 	if (uobjpage != PGO_DONTCARE) {
2078 		uvmexp.flt_prcopy++;
2079 
2080 		/*
2081 		 * promote to shared amap?  make sure all sharing
2082 		 * procs see it
2083 		 */
2084 
2085 		if ((amap_flags(amap) & AMAP_SHARED) != 0) {
2086 			pmap_page_protect(uobjpage, VM_PROT_NONE);
2087 			/*
2088 			 * XXX: PAGE MIGHT BE WIRED!
2089 			 */
2090 		}
2091 
2092 		/*
2093 		 * dispose of uobjpage.  it can't be PG_RELEASED
2094 		 * since we still hold the object lock.
2095 		 * drop handle to uobj as well.
2096 		 */
2097 
2098 		if (uobjpage->flags & PG_WANTED)
2099 			/* still have the obj lock */
2100 			wakeup(uobjpage);
2101 		uobjpage->flags &= ~(PG_BUSY|PG_WANTED);
2102 		UVM_PAGE_OWN(uobjpage, NULL);
2103 		mutex_exit(&uobj->vmobjlock);
2104 		uobj = NULL;
2105 
2106 		UVMHIST_LOG(maphist,
2107 		    "  promote uobjpage 0x%x to anon/page 0x%x/0x%x",
2108 		    uobjpage, anon, pg, 0);
2109 
2110 	} else {
2111 		uvmexp.flt_przero++;
2112 
2113 		/*
2114 		 * Page is zero'd and marked dirty by
2115 		 * uvmfault_promote().
2116 		 */
2117 
2118 		UVMHIST_LOG(maphist,"  zero fill anon/page 0x%x/0%x",
2119 		    anon, pg, 0, 0);
2120 	}
2121 
2122 	KASSERT(uobj == NULL || (uobjpage->flags & PG_BUSY) != 0);
2123 	return uvm_fault_lower_enter(ufi, flt, uobj, anon, pg);
2124 }
2125 
2126 /*
2127  * uvm_fault_lower_enter: enter h/w mapping of lower page or anon page promoted
2128  * from the lower page.
2129  */
2130 
2131 int
2132 uvm_fault_lower_enter(
2133 	struct uvm_faultinfo *ufi, const struct uvm_faultctx *flt,
2134 	struct uvm_object *uobj,
2135 	struct vm_anon *anon, struct vm_page *pg)
2136 {
2137 	struct vm_amap * const amap = ufi->entry->aref.ar_amap;
2138 	int error;
2139 	UVMHIST_FUNC("uvm_fault_lower_enter"); UVMHIST_CALLED(maphist);
2140 
2141 	/*
2142 	 * locked:
2143 	 * maps(read), amap(if !null), uobj(if !null),
2144 	 *   anon(if !null), pg(if anon)
2145 	 *
2146 	 * note: pg is either the uobjpage or the new page in the new anon
2147 	 */
2148 	KASSERT(amap == NULL || mutex_owned(&amap->am_l));
2149 	KASSERT(uobj == NULL || mutex_owned(&uobj->vmobjlock));
2150 	KASSERT(anon == NULL || mutex_owned(&anon->an_lock));
2151 	KASSERT((pg->flags & PG_BUSY) != 0);
2152 
2153 	/*
2154 	 * all resources are present.   we can now map it in and free our
2155 	 * resources.
2156 	 */
2157 
2158 	UVMHIST_LOG(maphist,
2159 	    "  MAPPING: case2: pm=0x%x, va=0x%x, pg=0x%x, promote=%d",
2160 	    ufi->orig_map->pmap, ufi->orig_rvaddr, pg, flt->promote);
2161 	KASSERT((flt->access_type & VM_PROT_WRITE) == 0 ||
2162 		(pg->flags & PG_RDONLY) == 0);
2163 	if (pmap_enter(ufi->orig_map->pmap, ufi->orig_rvaddr,
2164 	    VM_PAGE_TO_PHYS(pg),
2165 	    (pg->flags & PG_RDONLY) != 0 ?
2166 	    flt->enter_prot & ~VM_PROT_WRITE : flt->enter_prot,
2167 	    flt->access_type | PMAP_CANFAIL |
2168 	    (flt->wire_mapping ? PMAP_WIRED : 0)) != 0) {
2169 
2170 		/*
2171 		 * No need to undo what we did; we can simply think of
2172 		 * this as the pmap throwing away the mapping information.
2173 		 *
2174 		 * We do, however, have to go through the ReFault path,
2175 		 * as the map may change while we're asleep.
2176 		 */
2177 
2178 		/*
2179 		 * ensure that the page is queued in the case that
2180 		 * we just promoted the page.
2181 		 */
2182 
2183 		mutex_enter(&uvm_pageqlock);
2184 		uvm_pageenqueue(pg);
2185 		mutex_exit(&uvm_pageqlock);
2186 
2187 		if (pg->flags & PG_WANTED)
2188 			wakeup(pg);
2189 
2190 		/*
2191 		 * note that pg can't be PG_RELEASED since we did not drop
2192 		 * the object lock since the last time we checked.
2193 		 */
2194 		KASSERT((pg->flags & PG_RELEASED) == 0);
2195 
2196 		pg->flags &= ~(PG_BUSY|PG_FAKE|PG_WANTED);
2197 		UVM_PAGE_OWN(pg, NULL);
2198 
2199 		uvmfault_unlockall(ufi, amap, uobj, anon);
2200 		if (!uvm_reclaimable()) {
2201 			UVMHIST_LOG(maphist,
2202 			    "<- failed.  out of VM",0,0,0,0);
2203 			/* XXX instrumentation */
2204 			error = ENOMEM;
2205 			return error;
2206 		}
2207 		/* XXX instrumentation */
2208 		uvm_wait("flt_pmfail2");
2209 		return ERESTART;
2210 	}
2211 
2212 	uvm_fault_lower_done(ufi, flt, uobj, pg);
2213 
2214 	/*
2215 	 * note that pg can't be PG_RELEASED since we did not drop the object
2216 	 * lock since the last time we checked.
2217 	 */
2218 	KASSERT((pg->flags & PG_RELEASED) == 0);
2219 	if (pg->flags & PG_WANTED)
2220 		wakeup(pg);
2221 	pg->flags &= ~(PG_BUSY|PG_FAKE|PG_WANTED);
2222 	UVM_PAGE_OWN(pg, NULL);
2223 
2224 	pmap_update(ufi->orig_map->pmap);
2225 	uvmfault_unlockall(ufi, amap, uobj, anon);
2226 
2227 	UVMHIST_LOG(maphist, "<- done (SUCCESS!)",0,0,0,0);
2228 	return 0;
2229 }
2230 
2231 /*
2232  * uvm_fault_lower_done: queue lower center page.
2233  */
2234 
2235 void
2236 uvm_fault_lower_done(
2237 	struct uvm_faultinfo *ufi, const struct uvm_faultctx *flt,
2238 	struct uvm_object *uobj, struct vm_page *pg)
2239 {
2240 	bool dropswap = false;
2241 
2242 	UVMHIST_FUNC("uvm_fault_lower_done"); UVMHIST_CALLED(maphist);
2243 
2244 	mutex_enter(&uvm_pageqlock);
2245 	if (flt->wire_paging) {
2246 		uvm_pagewire(pg);
2247 		if (pg->pqflags & PQ_AOBJ) {
2248 
2249 			/*
2250 			 * since the now-wired page cannot be paged out,
2251 			 * release its swap resources for others to use.
2252 			 * since an aobj page with no swap cannot be PG_CLEAN,
2253 			 * clear its clean flag now.
2254 			 */
2255 
2256 			KASSERT(uobj != NULL);
2257 			pg->flags &= ~(PG_CLEAN);
2258 			dropswap = true;
2259 		}
2260 	} else {
2261 		uvm_pageactivate(pg);
2262 	}
2263 	mutex_exit(&uvm_pageqlock);
2264 
2265 	if (dropswap) {
2266 		uao_dropswap(uobj, pg->offset >> PAGE_SHIFT);
2267 	}
2268 }
2269 
2270 
2271 /*
2272  * uvm_fault_wire: wire down a range of virtual addresses in a map.
2273  *
2274  * => map may be read-locked by caller, but MUST NOT be write-locked.
2275  * => if map is read-locked, any operations which may cause map to
2276  *	be write-locked in uvm_fault() must be taken care of by
2277  *	the caller.  See uvm_map_pageable().
2278  */
2279 
2280 int
2281 uvm_fault_wire(struct vm_map *map, vaddr_t start, vaddr_t end,
2282     vm_prot_t access_type, int maxprot)
2283 {
2284 	vaddr_t va;
2285 	int error;
2286 
2287 	/*
2288 	 * now fault it in a page at a time.   if the fault fails then we have
2289 	 * to undo what we have done.   note that in uvm_fault VM_PROT_NONE
2290 	 * is replaced with the max protection if fault_type is VM_FAULT_WIRE.
2291 	 */
2292 
2293 	/*
2294 	 * XXX work around overflowing a vaddr_t.  this prevents us from
2295 	 * wiring the last page in the address space, though.
2296 	 */
2297 	if (start > end) {
2298 		return EFAULT;
2299 	}
2300 
2301 	for (va = start; va < end; va += PAGE_SIZE) {
2302 		error = uvm_fault_internal(map, va, access_type,
2303 		    (maxprot ? UVM_FAULT_MAXPROT : 0) | UVM_FAULT_WIRE);
2304 		if (error) {
2305 			if (va != start) {
2306 				uvm_fault_unwire(map, start, va);
2307 			}
2308 			return error;
2309 		}
2310 	}
2311 	return 0;
2312 }
2313 
2314 /*
2315  * uvm_fault_unwire(): unwire range of virtual space.
2316  */
2317 
2318 void
2319 uvm_fault_unwire(struct vm_map *map, vaddr_t start, vaddr_t end)
2320 {
2321 	vm_map_lock_read(map);
2322 	uvm_fault_unwire_locked(map, start, end);
2323 	vm_map_unlock_read(map);
2324 }
2325 
2326 /*
2327  * uvm_fault_unwire_locked(): the guts of uvm_fault_unwire().
2328  *
2329  * => map must be at least read-locked.
2330  */
2331 
2332 void
2333 uvm_fault_unwire_locked(struct vm_map *map, vaddr_t start, vaddr_t end)
2334 {
2335 	struct vm_map_entry *entry;
2336 	pmap_t pmap = vm_map_pmap(map);
2337 	vaddr_t va;
2338 	paddr_t pa;
2339 	struct vm_page *pg;
2340 
2341 	KASSERT((map->flags & VM_MAP_INTRSAFE) == 0);
2342 
2343 	/*
2344 	 * we assume that the area we are unwiring has actually been wired
2345 	 * in the first place.   this means that we should be able to extract
2346 	 * the PAs from the pmap.   we also lock out the page daemon so that
2347 	 * we can call uvm_pageunwire.
2348 	 */
2349 
2350 	mutex_enter(&uvm_pageqlock);
2351 
2352 	/*
2353 	 * find the beginning map entry for the region.
2354 	 */
2355 
2356 	KASSERT(start >= vm_map_min(map) && end <= vm_map_max(map));
2357 	if (uvm_map_lookup_entry(map, start, &entry) == false)
2358 		panic("uvm_fault_unwire_locked: address not in map");
2359 
2360 	for (va = start; va < end; va += PAGE_SIZE) {
2361 		if (pmap_extract(pmap, va, &pa) == false)
2362 			continue;
2363 
2364 		/*
2365 		 * find the map entry for the current address.
2366 		 */
2367 
2368 		KASSERT(va >= entry->start);
2369 		while (va >= entry->end) {
2370 			KASSERT(entry->next != &map->header &&
2371 				entry->next->start <= entry->end);
2372 			entry = entry->next;
2373 		}
2374 
2375 		/*
2376 		 * if the entry is no longer wired, tell the pmap.
2377 		 */
2378 
2379 		if (VM_MAPENT_ISWIRED(entry) == 0)
2380 			pmap_unwire(pmap, va);
2381 
2382 		pg = PHYS_TO_VM_PAGE(pa);
2383 		if (pg)
2384 			uvm_pageunwire(pg);
2385 	}
2386 
2387 	mutex_exit(&uvm_pageqlock);
2388 }
2389