xref: /netbsd-src/sys/uvm/uvm_fault.c (revision d25ffa98a4bfca1fe272f3c182496ec9934faac7)
1 /*	$NetBSD: uvm_fault.c,v 1.186 2011/06/12 03:36:02 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 1.186 2011/06/12 03:36:02 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 	{ UVM_ADV_NORMAL, 3, 4 },
161 	{ UVM_ADV_RANDOM, 0, 0 },
162 	{ UVM_ADV_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 		KASSERT(mutex_owned(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 	}
200 }
201 
202 /*
203  * normal functions
204  */
205 
206 /*
207  * uvmfault_amapcopy: clear "needs_copy" in a map.
208  *
209  * => called with VM data structures unlocked (usually, see below)
210  * => we get a write lock on the maps and clear needs_copy for a VA
211  * => if we are out of RAM we sleep (waiting for more)
212  */
213 
214 static void
215 uvmfault_amapcopy(struct uvm_faultinfo *ufi)
216 {
217 	for (;;) {
218 
219 		/*
220 		 * no mapping?  give up.
221 		 */
222 
223 		if (uvmfault_lookup(ufi, true) == false)
224 			return;
225 
226 		/*
227 		 * copy if needed.
228 		 */
229 
230 		if (UVM_ET_ISNEEDSCOPY(ufi->entry))
231 			amap_copy(ufi->map, ufi->entry, AMAP_COPY_NOWAIT,
232 				ufi->orig_rvaddr, ufi->orig_rvaddr + 1);
233 
234 		/*
235 		 * didn't work?  must be out of RAM.   unlock and sleep.
236 		 */
237 
238 		if (UVM_ET_ISNEEDSCOPY(ufi->entry)) {
239 			uvmfault_unlockmaps(ufi, true);
240 			uvm_wait("fltamapcopy");
241 			continue;
242 		}
243 
244 		/*
245 		 * got it!   unlock and return.
246 		 */
247 
248 		uvmfault_unlockmaps(ufi, true);
249 		return;
250 	}
251 	/*NOTREACHED*/
252 }
253 
254 /*
255  * uvmfault_anonget: get data in an anon into a non-busy, non-released
256  * page in that anon.
257  *
258  * => maps, amap, and anon locked by caller.
259  * => if we fail (result != 0) we unlock everything.
260  * => if we are successful, we return with everything still locked.
261  * => we don't move the page on the queues [gets moved later]
262  * => if we allocate a new page [we_own], it gets put on the queues.
263  *    either way, the result is that the page is on the queues at return time
264  * => for pages which are on loan from a uvm_object (and thus are not
265  *    owned by the anon): if successful, we return with the owning object
266  *    locked.   the caller must unlock this object when it unlocks everything
267  *    else.
268  */
269 
270 int
271 uvmfault_anonget(struct uvm_faultinfo *ufi, struct vm_amap *amap,
272     struct vm_anon *anon)
273 {
274 	bool we_own;	/* we own anon's page? */
275 	bool locked;	/* did we relock? */
276 	struct vm_page *pg;
277 	int error;
278 	UVMHIST_FUNC("uvmfault_anonget"); UVMHIST_CALLED(maphist);
279 
280 	KASSERT(mutex_owned(anon->an_lock));
281 	KASSERT(amap == NULL || anon->an_lock == amap->am_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);
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, 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, ufi != NULL ? UVM_FLAG_COLORMATCH : 0);
358 			if (pg == NULL) {		/* out of RAM.  */
359 				uvmfault_unlockall(ufi, amap, NULL);
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);
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 || we_own) {
399 			mutex_enter(anon->an_lock);
400 		}
401 
402 		/*
403 		 * if we own the page (i.e. we set PG_BUSY), then we need
404 		 * to clean up after the I/O. there are three cases to
405 		 * consider:
406 		 *   [1] page released during I/O: free anon and ReFault.
407 		 *   [2] I/O not OK.   free the page and cause the fault
408 		 *       to fail.
409 		 *   [3] I/O OK!   activate the page and sync with the
410 		 *       non-we_own case (i.e. drop anon lock if not locked).
411 		 */
412 
413 		if (we_own) {
414 #if defined(VMSWAP)
415 			if (pg->flags & PG_WANTED) {
416 				wakeup(pg);
417 			}
418 			if (error) {
419 
420 				/*
421 				 * remove the swap slot from the anon
422 				 * and mark the anon as having no real slot.
423 				 * don't free the swap slot, thus preventing
424 				 * it from being used again.
425 				 */
426 
427 				if (anon->an_swslot > 0)
428 					uvm_swap_markbad(anon->an_swslot, 1);
429 				anon->an_swslot = SWSLOT_BAD;
430 
431 				if ((pg->flags & PG_RELEASED) != 0)
432 					goto released;
433 
434 				/*
435 				 * note: page was never !PG_BUSY, so it
436 				 * can't be mapped and thus no need to
437 				 * pmap_page_protect it...
438 				 */
439 
440 				mutex_enter(&uvm_pageqlock);
441 				uvm_pagefree(pg);
442 				mutex_exit(&uvm_pageqlock);
443 
444 				if (locked)
445 					uvmfault_unlockall(ufi, NULL, NULL);
446 				mutex_exit(anon->an_lock);
447 				UVMHIST_LOG(maphist, "<- ERROR", 0,0,0,0);
448 				return error;
449 			}
450 
451 			if ((pg->flags & PG_RELEASED) != 0) {
452 released:
453 				KASSERT(anon->an_ref == 0);
454 
455 				/*
456 				 * released while we unlocked amap.
457 				 */
458 
459 				if (locked)
460 					uvmfault_unlockall(ufi, NULL, NULL);
461 
462 				uvm_anon_release(anon);
463 
464 				if (error) {
465 					UVMHIST_LOG(maphist,
466 					    "<- ERROR/RELEASED", 0,0,0,0);
467 					return error;
468 				}
469 
470 				UVMHIST_LOG(maphist, "<- RELEASED", 0,0,0,0);
471 				return ERESTART;
472 			}
473 
474 			/*
475 			 * we've successfully read the page, activate it.
476 			 */
477 
478 			mutex_enter(&uvm_pageqlock);
479 			uvm_pageactivate(pg);
480 			mutex_exit(&uvm_pageqlock);
481 			pg->flags &= ~(PG_WANTED|PG_BUSY|PG_FAKE);
482 			UVM_PAGE_OWN(pg, NULL);
483 #else /* defined(VMSWAP) */
484 			panic("%s: we_own", __func__);
485 #endif /* defined(VMSWAP) */
486 		}
487 
488 		/*
489 		 * we were not able to relock.   restart fault.
490 		 */
491 
492 		if (!locked) {
493 			if (we_own) {
494 				mutex_exit(anon->an_lock);
495 			}
496 			UVMHIST_LOG(maphist, "<- REFAULT", 0,0,0,0);
497 			return (ERESTART);
498 		}
499 
500 		/*
501 		 * verify no one has touched the amap and moved the anon on us.
502 		 */
503 
504 		if (ufi != NULL && amap_lookup(&ufi->entry->aref,
505 		    ufi->orig_rvaddr - ufi->entry->start) != anon) {
506 
507 			uvmfault_unlockall(ufi, amap, NULL);
508 			UVMHIST_LOG(maphist, "<- REFAULT", 0,0,0,0);
509 			return (ERESTART);
510 		}
511 
512 		/*
513 		 * try it again!
514 		 */
515 
516 		uvmexp.fltanretry++;
517 		continue;
518 	}
519 	/*NOTREACHED*/
520 }
521 
522 /*
523  * uvmfault_promote: promote data to a new anon.  used for 1B and 2B.
524  *
525  *	1. allocate an anon and a page.
526  *	2. fill its contents.
527  *	3. put it into amap.
528  *
529  * => if we fail (result != 0) we unlock everything.
530  * => on success, return a new locked anon via 'nanon'.
531  *    (*nanon)->an_page will be a resident, locked, dirty page.
532  * => it's caller's responsibility to put the promoted nanon->an_page to the
533  *    page queue.
534  */
535 
536 static int
537 uvmfault_promote(struct uvm_faultinfo *ufi,
538     struct vm_anon *oanon,
539     struct vm_page *uobjpage,
540     struct vm_anon **nanon, /* OUT: allocated anon */
541     struct vm_anon **spare)
542 {
543 	struct vm_amap *amap = ufi->entry->aref.ar_amap;
544 	struct uvm_object *uobj;
545 	struct vm_anon *anon;
546 	struct vm_page *pg;
547 	struct vm_page *opg;
548 	int error;
549 	UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist);
550 
551 	if (oanon) {
552 		/* anon COW */
553 		opg = oanon->an_page;
554 		KASSERT(opg != NULL);
555 		KASSERT(opg->uobject == NULL || opg->loan_count > 0);
556 	} else if (uobjpage != PGO_DONTCARE) {
557 		/* object-backed COW */
558 		opg = uobjpage;
559 	} else {
560 		/* ZFOD */
561 		opg = NULL;
562 	}
563 	if (opg != NULL) {
564 		uobj = opg->uobject;
565 	} else {
566 		uobj = NULL;
567 	}
568 
569 	KASSERT(amap != NULL);
570 	KASSERT(uobjpage != NULL);
571 	KASSERT(uobjpage == PGO_DONTCARE || (uobjpage->flags & PG_BUSY) != 0);
572 	KASSERT(mutex_owned(amap->am_lock));
573 	KASSERT(oanon == NULL || amap->am_lock == oanon->an_lock);
574 	KASSERT(uobj == NULL || mutex_owned(uobj->vmobjlock));
575 
576 	if (*spare != NULL) {
577 		anon = *spare;
578 		*spare = NULL;
579 	} else if (ufi->map != kernel_map) {
580 		anon = uvm_analloc();
581 	} else {
582 		UVMHIST_LOG(maphist, "kernel_map, unlock and retry", 0,0,0,0);
583 
584 		/*
585 		 * we can't allocate anons with kernel_map locked.
586 		 */
587 
588 		uvm_page_unbusy(&uobjpage, 1);
589 		uvmfault_unlockall(ufi, amap, uobj);
590 
591 		*spare = uvm_analloc();
592 		if (*spare == NULL) {
593 			goto nomem;
594 		}
595 		KASSERT((*spare)->an_lock == NULL);
596 		error = ERESTART;
597 		goto done;
598 	}
599 	if (anon) {
600 
601 		/*
602 		 * The new anon is locked.
603 		 *
604 		 * if opg == NULL, we want a zero'd, dirty page,
605 		 * so have uvm_pagealloc() do that for us.
606 		 */
607 
608 		KASSERT(anon->an_lock == NULL);
609 		anon->an_lock = amap->am_lock;
610 		mutex_obj_hold(anon->an_lock);
611 		pg = uvm_pagealloc(NULL, ufi->orig_rvaddr, anon,
612 		    UVM_FLAG_COLORMATCH | (opg == NULL ? UVM_PGA_ZERO : 0));
613 		if (pg == NULL) {
614 			mutex_obj_free(anon->an_lock);
615 			anon->an_lock = NULL;
616 		}
617 	} else {
618 		pg = NULL;
619 	}
620 
621 	/*
622 	 * out of memory resources?
623 	 */
624 
625 	if (pg == NULL) {
626 		/* save anon for the next try. */
627 		if (anon != NULL) {
628 			*spare = anon;
629 		}
630 
631 		/* unlock and fail ... */
632 		uvm_page_unbusy(&uobjpage, 1);
633 		uvmfault_unlockall(ufi, amap, uobj);
634 nomem:
635 		if (!uvm_reclaimable()) {
636 			UVMHIST_LOG(maphist, "out of VM", 0,0,0,0);
637 			uvmexp.fltnoanon++;
638 			error = ENOMEM;
639 			goto done;
640 		}
641 
642 		UVMHIST_LOG(maphist, "out of RAM, waiting for more", 0,0,0,0);
643 		uvmexp.fltnoram++;
644 		uvm_wait("flt_noram5");
645 		error = ERESTART;
646 		goto done;
647 	}
648 
649 	/* copy page [pg now dirty] */
650 	if (opg) {
651 		uvm_pagecopy(opg, pg);
652 	}
653 
654 	amap_add(&ufi->entry->aref, ufi->orig_rvaddr - ufi->entry->start, anon,
655 	    oanon != NULL);
656 
657 	*nanon = anon;
658 	error = 0;
659 done:
660 	return error;
661 }
662 
663 
664 /*
665  *   F A U L T   -   m a i n   e n t r y   p o i n t
666  */
667 
668 /*
669  * uvm_fault: page fault handler
670  *
671  * => called from MD code to resolve a page fault
672  * => VM data structures usually should be unlocked.   however, it is
673  *	possible to call here with the main map locked if the caller
674  *	gets a write lock, sets it recusive, and then calls us (c.f.
675  *	uvm_map_pageable).   this should be avoided because it keeps
676  *	the map locked off during I/O.
677  * => MUST NEVER BE CALLED IN INTERRUPT CONTEXT
678  */
679 
680 #define MASK(entry)     (UVM_ET_ISCOPYONWRITE(entry) ? \
681 			 ~VM_PROT_WRITE : VM_PROT_ALL)
682 
683 /* fault_flag values passed from uvm_fault_wire to uvm_fault_internal */
684 #define UVM_FAULT_WIRE		(1 << 0)
685 #define UVM_FAULT_MAXPROT	(1 << 1)
686 
687 struct uvm_faultctx {
688 	vm_prot_t access_type;
689 	vm_prot_t enter_prot;
690 	vaddr_t startva;
691 	int npages;
692 	int centeridx;
693 	struct vm_anon *anon_spare;
694 	bool wire_mapping;
695 	bool narrow;
696 	bool wire_paging;
697 	bool cow_now;
698 	bool promote;
699 };
700 
701 static inline int	uvm_fault_check(
702 			    struct uvm_faultinfo *, struct uvm_faultctx *,
703 			    struct vm_anon ***, bool);
704 
705 static int		uvm_fault_upper(
706 			    struct uvm_faultinfo *, struct uvm_faultctx *,
707 			    struct vm_anon **);
708 static inline int	uvm_fault_upper_lookup(
709 			    struct uvm_faultinfo *, const struct uvm_faultctx *,
710 			    struct vm_anon **, struct vm_page **);
711 static inline void	uvm_fault_upper_neighbor(
712 			    struct uvm_faultinfo *, const struct uvm_faultctx *,
713 			    vaddr_t, struct vm_page *, bool);
714 static inline int	uvm_fault_upper_loan(
715 			    struct uvm_faultinfo *, struct uvm_faultctx *,
716 			    struct vm_anon *, struct uvm_object **);
717 static inline int	uvm_fault_upper_promote(
718 			    struct uvm_faultinfo *, struct uvm_faultctx *,
719 			    struct uvm_object *, struct vm_anon *);
720 static inline int	uvm_fault_upper_direct(
721 			    struct uvm_faultinfo *, struct uvm_faultctx *,
722 			    struct uvm_object *, struct vm_anon *);
723 static int		uvm_fault_upper_enter(
724 			    struct uvm_faultinfo *, const struct uvm_faultctx *,
725 			    struct uvm_object *, struct vm_anon *,
726 			    struct vm_page *, struct vm_anon *);
727 static inline void	uvm_fault_upper_done(
728 			    struct uvm_faultinfo *, const struct uvm_faultctx *,
729 			    struct vm_anon *, struct vm_page *);
730 
731 static int		uvm_fault_lower(
732 			    struct uvm_faultinfo *, struct uvm_faultctx *,
733 			    struct vm_page **);
734 static inline void	uvm_fault_lower_lookup(
735 			    struct uvm_faultinfo *, const struct uvm_faultctx *,
736 			    struct vm_page **);
737 static inline void	uvm_fault_lower_neighbor(
738 			    struct uvm_faultinfo *, const struct uvm_faultctx *,
739 			    vaddr_t, struct vm_page *, bool);
740 static inline int	uvm_fault_lower_io(
741 			    struct uvm_faultinfo *, const struct uvm_faultctx *,
742 			    struct uvm_object **, struct vm_page **);
743 static inline int	uvm_fault_lower_direct(
744 			    struct uvm_faultinfo *, struct uvm_faultctx *,
745 			    struct uvm_object *, struct vm_page *);
746 static inline int	uvm_fault_lower_direct_loan(
747 			    struct uvm_faultinfo *, struct uvm_faultctx *,
748 			    struct uvm_object *, struct vm_page **,
749 			    struct vm_page **);
750 static inline int	uvm_fault_lower_promote(
751 			    struct uvm_faultinfo *, struct uvm_faultctx *,
752 			    struct uvm_object *, struct vm_page *);
753 static int		uvm_fault_lower_enter(
754 			    struct uvm_faultinfo *, const struct uvm_faultctx *,
755 			    struct uvm_object *,
756 			    struct vm_anon *, struct vm_page *);
757 static inline void	uvm_fault_lower_done(
758 			    struct uvm_faultinfo *, const struct uvm_faultctx *,
759 			    struct uvm_object *, struct vm_page *);
760 
761 int
762 uvm_fault_internal(struct vm_map *orig_map, vaddr_t vaddr,
763     vm_prot_t access_type, int fault_flag)
764 {
765 	struct uvm_faultinfo ufi;
766 	struct uvm_faultctx flt = {
767 		.access_type = access_type,
768 
769 		/* don't look for neighborhood * pages on "wire" fault */
770 		.narrow = (fault_flag & UVM_FAULT_WIRE) != 0,
771 
772 		/* "wire" fault causes wiring of both mapping and paging */
773 		.wire_mapping = (fault_flag & UVM_FAULT_WIRE) != 0,
774 		.wire_paging = (fault_flag & UVM_FAULT_WIRE) != 0,
775 	};
776 	const bool maxprot = (fault_flag & UVM_FAULT_MAXPROT) != 0;
777 	struct vm_anon *anons_store[UVM_MAXRANGE], **anons;
778 	struct vm_page *pages_store[UVM_MAXRANGE], **pages;
779 	int error;
780 	UVMHIST_FUNC("uvm_fault"); UVMHIST_CALLED(maphist);
781 
782 	UVMHIST_LOG(maphist, "(map=0x%x, vaddr=0x%x, at=%d, ff=%d)",
783 	      orig_map, vaddr, access_type, fault_flag);
784 
785 	curcpu()->ci_data.cpu_nfault++;
786 
787 	/*
788 	 * init the IN parameters in the ufi
789 	 */
790 
791 	ufi.orig_map = orig_map;
792 	ufi.orig_rvaddr = trunc_page(vaddr);
793 	ufi.orig_size = PAGE_SIZE;	/* can't get any smaller than this */
794 
795 	error = ERESTART;
796 	while (error == ERESTART) { /* ReFault: */
797 		anons = anons_store;
798 		pages = pages_store;
799 
800 		error = uvm_fault_check(&ufi, &flt, &anons, maxprot);
801 		if (error != 0)
802 			continue;
803 
804 		error = uvm_fault_upper_lookup(&ufi, &flt, anons, pages);
805 		if (error != 0)
806 			continue;
807 
808 		if (pages[flt.centeridx] == PGO_DONTCARE)
809 			error = uvm_fault_upper(&ufi, &flt, anons);
810 		else {
811 			struct uvm_object * const uobj =
812 			    ufi.entry->object.uvm_obj;
813 
814 			if (uobj && uobj->pgops->pgo_fault != NULL) {
815 				/*
816 				 * invoke "special" fault routine.
817 				 */
818 				mutex_enter(uobj->vmobjlock);
819 				/* locked: maps(read), amap(if there), uobj */
820 				error = uobj->pgops->pgo_fault(&ufi,
821 				    flt.startva, pages, flt.npages,
822 				    flt.centeridx, flt.access_type,
823 				    PGO_LOCKED|PGO_SYNCIO);
824 
825 				/*
826 				 * locked: nothing, pgo_fault has unlocked
827 				 * everything
828 				 */
829 
830 				/*
831 				 * object fault routine responsible for
832 				 * pmap_update().
833 				 */
834 			} else {
835 				error = uvm_fault_lower(&ufi, &flt, pages);
836 			}
837 		}
838 	}
839 
840 	if (flt.anon_spare != NULL) {
841 		flt.anon_spare->an_ref--;
842 		KASSERT(flt.anon_spare->an_ref == 0);
843 		KASSERT(flt.anon_spare->an_lock == NULL);
844 		uvm_anfree(flt.anon_spare);
845 	}
846 	return error;
847 }
848 
849 /*
850  * uvm_fault_check: check prot, handle needs-copy, etc.
851  *
852  *	1. lookup entry.
853  *	2. check protection.
854  *	3. adjust fault condition (mainly for simulated fault).
855  *	4. handle needs-copy (lazy amap copy).
856  *	5. establish range of interest for neighbor fault (aka pre-fault).
857  *	6. look up anons (if amap exists).
858  *	7. flush pages (if MADV_SEQUENTIAL)
859  *
860  * => called with nothing locked.
861  * => if we fail (result != 0) we unlock everything.
862  * => initialize/adjust many members of flt.
863  */
864 
865 static int
866 uvm_fault_check(
867 	struct uvm_faultinfo *ufi, struct uvm_faultctx *flt,
868 	struct vm_anon ***ranons, bool maxprot)
869 {
870 	struct vm_amap *amap;
871 	struct uvm_object *uobj;
872 	vm_prot_t check_prot;
873 	int nback, nforw;
874 	UVMHIST_FUNC("uvm_fault_check"); UVMHIST_CALLED(maphist);
875 
876 	/*
877 	 * lookup and lock the maps
878 	 */
879 
880 	if (uvmfault_lookup(ufi, false) == false) {
881 		UVMHIST_LOG(maphist, "<- no mapping @ 0x%x", ufi->orig_rvaddr,
882 		    0,0,0);
883 		return EFAULT;
884 	}
885 	/* locked: maps(read) */
886 
887 #ifdef DIAGNOSTIC
888 	if ((ufi->map->flags & VM_MAP_PAGEABLE) == 0) {
889 		printf("Page fault on non-pageable map:\n");
890 		printf("ufi->map = %p\n", ufi->map);
891 		printf("ufi->orig_map = %p\n", ufi->orig_map);
892 		printf("ufi->orig_rvaddr = 0x%lx\n", (u_long) ufi->orig_rvaddr);
893 		panic("uvm_fault: (ufi->map->flags & VM_MAP_PAGEABLE) == 0");
894 	}
895 #endif
896 
897 	/*
898 	 * check protection
899 	 */
900 
901 	check_prot = maxprot ?
902 	    ufi->entry->max_protection : ufi->entry->protection;
903 	if ((check_prot & flt->access_type) != flt->access_type) {
904 		UVMHIST_LOG(maphist,
905 		    "<- protection failure (prot=0x%x, access=0x%x)",
906 		    ufi->entry->protection, flt->access_type, 0, 0);
907 		uvmfault_unlockmaps(ufi, false);
908 		return EACCES;
909 	}
910 
911 	/*
912 	 * "enter_prot" is the protection we want to enter the page in at.
913 	 * for certain pages (e.g. copy-on-write pages) this protection can
914 	 * be more strict than ufi->entry->protection.  "wired" means either
915 	 * the entry is wired or we are fault-wiring the pg.
916 	 */
917 
918 	flt->enter_prot = ufi->entry->protection;
919 	if (VM_MAPENT_ISWIRED(ufi->entry))
920 		flt->wire_mapping = true;
921 
922 	if (flt->wire_mapping) {
923 		flt->access_type = flt->enter_prot; /* full access for wired */
924 		flt->cow_now = (check_prot & VM_PROT_WRITE) != 0;
925 	} else {
926 		flt->cow_now = (flt->access_type & VM_PROT_WRITE) != 0;
927 	}
928 
929 	flt->promote = false;
930 
931 	/*
932 	 * handle "needs_copy" case.   if we need to copy the amap we will
933 	 * have to drop our readlock and relock it with a write lock.  (we
934 	 * need a write lock to change anything in a map entry [e.g.
935 	 * needs_copy]).
936 	 */
937 
938 	if (UVM_ET_ISNEEDSCOPY(ufi->entry)) {
939 		if (flt->cow_now || (ufi->entry->object.uvm_obj == NULL)) {
940 			KASSERT(!maxprot);
941 			/* need to clear */
942 			UVMHIST_LOG(maphist,
943 			    "  need to clear needs_copy and refault",0,0,0,0);
944 			uvmfault_unlockmaps(ufi, false);
945 			uvmfault_amapcopy(ufi);
946 			uvmexp.fltamcopy++;
947 			return ERESTART;
948 
949 		} else {
950 
951 			/*
952 			 * ensure that we pmap_enter page R/O since
953 			 * needs_copy is still true
954 			 */
955 
956 			flt->enter_prot &= ~VM_PROT_WRITE;
957 		}
958 	}
959 
960 	/*
961 	 * identify the players
962 	 */
963 
964 	amap = ufi->entry->aref.ar_amap;	/* upper layer */
965 	uobj = ufi->entry->object.uvm_obj;	/* lower layer */
966 
967 	/*
968 	 * check for a case 0 fault.  if nothing backing the entry then
969 	 * error now.
970 	 */
971 
972 	if (amap == NULL && uobj == NULL) {
973 		uvmfault_unlockmaps(ufi, false);
974 		UVMHIST_LOG(maphist,"<- no backing store, no overlay",0,0,0,0);
975 		return EFAULT;
976 	}
977 
978 	/*
979 	 * establish range of interest based on advice from mapper
980 	 * and then clip to fit map entry.   note that we only want
981 	 * to do this the first time through the fault.   if we
982 	 * ReFault we will disable this by setting "narrow" to true.
983 	 */
984 
985 	if (flt->narrow == false) {
986 
987 		/* wide fault (!narrow) */
988 		KASSERT(uvmadvice[ufi->entry->advice].advice ==
989 			 ufi->entry->advice);
990 		nback = MIN(uvmadvice[ufi->entry->advice].nback,
991 		    (ufi->orig_rvaddr - ufi->entry->start) >> PAGE_SHIFT);
992 		flt->startva = ufi->orig_rvaddr - (nback << PAGE_SHIFT);
993 		/*
994 		 * note: "-1" because we don't want to count the
995 		 * faulting page as forw
996 		 */
997 		nforw = MIN(uvmadvice[ufi->entry->advice].nforw,
998 			    ((ufi->entry->end - ufi->orig_rvaddr) >>
999 			     PAGE_SHIFT) - 1);
1000 		flt->npages = nback + nforw + 1;
1001 		flt->centeridx = nback;
1002 
1003 		flt->narrow = true;	/* ensure only once per-fault */
1004 
1005 	} else {
1006 
1007 		/* narrow fault! */
1008 		nback = nforw = 0;
1009 		flt->startva = ufi->orig_rvaddr;
1010 		flt->npages = 1;
1011 		flt->centeridx = 0;
1012 
1013 	}
1014 	/* offset from entry's start to pgs' start */
1015 	const voff_t eoff = flt->startva - ufi->entry->start;
1016 
1017 	/* locked: maps(read) */
1018 	UVMHIST_LOG(maphist, "  narrow=%d, back=%d, forw=%d, startva=0x%x",
1019 		    flt->narrow, nback, nforw, flt->startva);
1020 	UVMHIST_LOG(maphist, "  entry=0x%x, amap=0x%x, obj=0x%x", ufi->entry,
1021 		    amap, uobj, 0);
1022 
1023 	/*
1024 	 * if we've got an amap, lock it and extract current anons.
1025 	 */
1026 
1027 	if (amap) {
1028 		amap_lock(amap);
1029 		amap_lookups(&ufi->entry->aref, eoff, *ranons, flt->npages);
1030 	} else {
1031 		*ranons = NULL;	/* to be safe */
1032 	}
1033 
1034 	/* locked: maps(read), amap(if there) */
1035 	KASSERT(amap == NULL || mutex_owned(amap->am_lock));
1036 
1037 	/*
1038 	 * for MADV_SEQUENTIAL mappings we want to deactivate the back pages
1039 	 * now and then forget about them (for the rest of the fault).
1040 	 */
1041 
1042 	if (ufi->entry->advice == MADV_SEQUENTIAL && nback != 0) {
1043 
1044 		UVMHIST_LOG(maphist, "  MADV_SEQUENTIAL: flushing backpages",
1045 		    0,0,0,0);
1046 		/* flush back-page anons? */
1047 		if (amap)
1048 			uvmfault_anonflush(*ranons, nback);
1049 
1050 		/* flush object? */
1051 		if (uobj) {
1052 			voff_t uoff;
1053 
1054 			uoff = ufi->entry->offset + eoff;
1055 			mutex_enter(uobj->vmobjlock);
1056 			(void) (uobj->pgops->pgo_put)(uobj, uoff, uoff +
1057 				    (nback << PAGE_SHIFT), PGO_DEACTIVATE);
1058 		}
1059 
1060 		/* now forget about the backpages */
1061 		if (amap)
1062 			*ranons += nback;
1063 		flt->startva += (nback << PAGE_SHIFT);
1064 		flt->npages -= nback;
1065 		flt->centeridx = 0;
1066 	}
1067 	/*
1068 	 * => startva is fixed
1069 	 * => npages is fixed
1070 	 */
1071 	KASSERT(flt->startva <= ufi->orig_rvaddr);
1072 	KASSERT(ufi->orig_rvaddr + ufi->orig_size <=
1073 	    flt->startva + (flt->npages << PAGE_SHIFT));
1074 	return 0;
1075 }
1076 
1077 /*
1078  * uvm_fault_upper_lookup: look up existing h/w mapping and amap.
1079  *
1080  * iterate range of interest:
1081  *	1. check if h/w mapping exists.  if yes, we don't care
1082  *	2. check if anon exists.  if not, page is lower.
1083  *	3. if anon exists, enter h/w mapping for neighbors.
1084  *
1085  * => called with amap locked (if exists).
1086  */
1087 
1088 static int
1089 uvm_fault_upper_lookup(
1090 	struct uvm_faultinfo *ufi, const struct uvm_faultctx *flt,
1091 	struct vm_anon **anons, struct vm_page **pages)
1092 {
1093 	struct vm_amap *amap = ufi->entry->aref.ar_amap;
1094 	int lcv;
1095 	vaddr_t currva;
1096 	bool shadowed;
1097 	UVMHIST_FUNC("uvm_fault_upper_lookup"); UVMHIST_CALLED(maphist);
1098 
1099 	/* locked: maps(read), amap(if there) */
1100 	KASSERT(amap == NULL || mutex_owned(amap->am_lock));
1101 
1102 	/*
1103 	 * map in the backpages and frontpages we found in the amap in hopes
1104 	 * of preventing future faults.    we also init the pages[] array as
1105 	 * we go.
1106 	 */
1107 
1108 	currva = flt->startva;
1109 	shadowed = false;
1110 	for (lcv = 0; lcv < flt->npages; lcv++, currva += PAGE_SIZE) {
1111 		/*
1112 		 * don't play with VAs that are already mapped
1113 		 * (except for center)
1114 		 */
1115 		if (lcv != flt->centeridx &&
1116 		    pmap_extract(ufi->orig_map->pmap, currva, NULL)) {
1117 			pages[lcv] = PGO_DONTCARE;
1118 			continue;
1119 		}
1120 
1121 		/*
1122 		 * unmapped or center page.   check if any anon at this level.
1123 		 */
1124 		if (amap == NULL || anons[lcv] == NULL) {
1125 			pages[lcv] = NULL;
1126 			continue;
1127 		}
1128 
1129 		/*
1130 		 * check for present page and map if possible.   re-activate it.
1131 		 */
1132 
1133 		pages[lcv] = PGO_DONTCARE;
1134 		if (lcv == flt->centeridx) {	/* save center for later! */
1135 			shadowed = true;
1136 			continue;
1137 		}
1138 
1139 		struct vm_anon *anon = anons[lcv];
1140 		struct vm_page *pg = anon->an_page;
1141 
1142 		KASSERT(anon->an_lock == amap->am_lock);
1143 
1144 		/* Ignore loaned and busy pages. */
1145 		if (pg && pg->loan_count == 0 && (pg->flags & PG_BUSY) == 0) {
1146 			uvm_fault_upper_neighbor(ufi, flt, currva,
1147 			    pg, anon->an_ref > 1);
1148 		}
1149 	}
1150 
1151 	/* locked: maps(read), amap(if there) */
1152 	KASSERT(amap == NULL || mutex_owned(amap->am_lock));
1153 	/* (shadowed == true) if there is an anon at the faulting address */
1154 	UVMHIST_LOG(maphist, "  shadowed=%d, will_get=%d", shadowed,
1155 	    (ufi->entry->object.uvm_obj && shadowed != false),0,0);
1156 
1157 	/*
1158 	 * note that if we are really short of RAM we could sleep in the above
1159 	 * call to pmap_enter with everything locked.   bad?
1160 	 *
1161 	 * XXX Actually, that is bad; pmap_enter() should just fail in that
1162 	 * XXX case.  --thorpej
1163 	 */
1164 
1165 	return 0;
1166 }
1167 
1168 /*
1169  * uvm_fault_upper_neighbor: enter single lower neighbor page.
1170  *
1171  * => called with amap and anon locked.
1172  */
1173 
1174 static void
1175 uvm_fault_upper_neighbor(
1176 	struct uvm_faultinfo *ufi, const struct uvm_faultctx *flt,
1177 	vaddr_t currva, struct vm_page *pg, bool readonly)
1178 {
1179 	UVMHIST_FUNC("uvm_fault_upper_neighbor"); UVMHIST_CALLED(maphist);
1180 
1181 	/* locked: amap, anon */
1182 
1183 	mutex_enter(&uvm_pageqlock);
1184 	uvm_pageenqueue(pg);
1185 	mutex_exit(&uvm_pageqlock);
1186 	UVMHIST_LOG(maphist,
1187 	    "  MAPPING: n anon: pm=0x%x, va=0x%x, pg=0x%x",
1188 	    ufi->orig_map->pmap, currva, pg, 0);
1189 	uvmexp.fltnamap++;
1190 
1191 	/*
1192 	 * Since this page isn't the page that's actually faulting,
1193 	 * ignore pmap_enter() failures; it's not critical that we
1194 	 * enter these right now.
1195 	 */
1196 
1197 	(void) pmap_enter(ufi->orig_map->pmap, currva,
1198 	    VM_PAGE_TO_PHYS(pg),
1199 	    readonly ? (flt->enter_prot & ~VM_PROT_WRITE) :
1200 	    flt->enter_prot,
1201 	    PMAP_CANFAIL | (flt->wire_mapping ? PMAP_WIRED : 0));
1202 
1203 	pmap_update(ufi->orig_map->pmap);
1204 }
1205 
1206 /*
1207  * uvm_fault_upper: handle upper fault.
1208  *
1209  *	1. acquire anon lock.
1210  *	2. get anon.  let uvmfault_anonget do the dirty work.
1211  *	3. handle loan.
1212  *	4. dispatch direct or promote handlers.
1213  */
1214 
1215 static int
1216 uvm_fault_upper(
1217 	struct uvm_faultinfo *ufi, struct uvm_faultctx *flt,
1218 	struct vm_anon **anons)
1219 {
1220 	struct vm_amap * const amap = ufi->entry->aref.ar_amap;
1221 	struct vm_anon * const anon = anons[flt->centeridx];
1222 	struct uvm_object *uobj;
1223 	int error;
1224 	UVMHIST_FUNC("uvm_fault_upper"); UVMHIST_CALLED(maphist);
1225 
1226 	/* locked: maps(read), amap, anon */
1227 	KASSERT(mutex_owned(amap->am_lock));
1228 	KASSERT(anon->an_lock == amap->am_lock);
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 
1236 	/*
1237 	 * no matter if we have case 1A or case 1B we are going to need to
1238 	 * have the anon's memory resident.   ensure that now.
1239 	 */
1240 
1241 	/*
1242 	 * let uvmfault_anonget do the dirty work.
1243 	 * if it fails (!OK) it will unlock everything for us.
1244 	 * if it succeeds, locks are still valid and locked.
1245 	 * also, if it is OK, then the anon's page is on the queues.
1246 	 * if the page is on loan from a uvm_object, then anonget will
1247 	 * lock that object for us if it does not fail.
1248 	 */
1249 
1250 	error = uvmfault_anonget(ufi, amap, anon);
1251 	switch (error) {
1252 	case 0:
1253 		break;
1254 
1255 	case ERESTART:
1256 		return ERESTART;
1257 
1258 	case EAGAIN:
1259 		kpause("fltagain1", false, hz/2, NULL);
1260 		return ERESTART;
1261 
1262 	default:
1263 		return error;
1264 	}
1265 
1266 	/*
1267 	 * uobj is non null if the page is on loan from an object (i.e. uobj)
1268 	 */
1269 
1270 	uobj = anon->an_page->uobject;	/* locked by anonget if !NULL */
1271 
1272 	/* locked: maps(read), amap, anon, uobj(if one) */
1273 	KASSERT(mutex_owned(amap->am_lock));
1274 	KASSERT(anon->an_lock == amap->am_lock);
1275 	KASSERT(uobj == NULL || mutex_owned(uobj->vmobjlock));
1276 
1277 	/*
1278 	 * special handling for loaned pages
1279 	 */
1280 
1281 	if (anon->an_page->loan_count) {
1282 		error = uvm_fault_upper_loan(ufi, flt, anon, &uobj);
1283 		if (error != 0)
1284 			return error;
1285 	}
1286 
1287 	/*
1288 	 * if we are case 1B then we will need to allocate a new blank
1289 	 * anon to transfer the data into.   note that we have a lock
1290 	 * on anon, so no one can busy or release the page until we are done.
1291 	 * also note that the ref count can't drop to zero here because
1292 	 * it is > 1 and we are only dropping one ref.
1293 	 *
1294 	 * in the (hopefully very rare) case that we are out of RAM we
1295 	 * will unlock, wait for more RAM, and refault.
1296 	 *
1297 	 * if we are out of anon VM we kill the process (XXX: could wait?).
1298 	 */
1299 
1300 	if (flt->cow_now && anon->an_ref > 1) {
1301 		flt->promote = true;
1302 		error = uvm_fault_upper_promote(ufi, flt, uobj, anon);
1303 	} else {
1304 		error = uvm_fault_upper_direct(ufi, flt, uobj, anon);
1305 	}
1306 	return error;
1307 }
1308 
1309 /*
1310  * uvm_fault_upper_loan: handle loaned upper page.
1311  *
1312  *	1. if not cow'ing now, simply adjust flt->enter_prot.
1313  *	2. if cow'ing now, and if ref count is 1, break loan.
1314  */
1315 
1316 static int
1317 uvm_fault_upper_loan(
1318 	struct uvm_faultinfo *ufi, struct uvm_faultctx *flt,
1319 	struct vm_anon *anon, struct uvm_object **ruobj)
1320 {
1321 	struct vm_amap * const amap = ufi->entry->aref.ar_amap;
1322 	int error = 0;
1323 	UVMHIST_FUNC("uvm_fault_upper_loan"); UVMHIST_CALLED(maphist);
1324 
1325 	if (!flt->cow_now) {
1326 
1327 		/*
1328 		 * for read faults on loaned pages we just cap the
1329 		 * protection at read-only.
1330 		 */
1331 
1332 		flt->enter_prot = flt->enter_prot & ~VM_PROT_WRITE;
1333 
1334 	} else {
1335 		/*
1336 		 * note that we can't allow writes into a loaned page!
1337 		 *
1338 		 * if we have a write fault on a loaned page in an
1339 		 * anon then we need to look at the anon's ref count.
1340 		 * if it is greater than one then we are going to do
1341 		 * a normal copy-on-write fault into a new anon (this
1342 		 * is not a problem).  however, if the reference count
1343 		 * is one (a case where we would normally allow a
1344 		 * write directly to the page) then we need to kill
1345 		 * the loan before we continue.
1346 		 */
1347 
1348 		/* >1 case is already ok */
1349 		if (anon->an_ref == 1) {
1350 			error = uvm_loanbreak_anon(anon, *ruobj);
1351 			if (error != 0) {
1352 				uvmfault_unlockall(ufi, amap, *ruobj);
1353 				uvm_wait("flt_noram2");
1354 				return ERESTART;
1355 			}
1356 			/* if we were a loan reciever uobj is gone */
1357 			if (*ruobj)
1358 				*ruobj = NULL;
1359 		}
1360 	}
1361 	return error;
1362 }
1363 
1364 /*
1365  * uvm_fault_upper_promote: promote upper page.
1366  *
1367  *	1. call uvmfault_promote.
1368  *	2. enqueue page.
1369  *	3. deref.
1370  *	4. pass page to uvm_fault_upper_enter.
1371  */
1372 
1373 static int
1374 uvm_fault_upper_promote(
1375 	struct uvm_faultinfo *ufi, struct uvm_faultctx *flt,
1376 	struct uvm_object *uobj, struct vm_anon *anon)
1377 {
1378 	struct vm_anon * const oanon = anon;
1379 	struct vm_page *pg;
1380 	int error;
1381 	UVMHIST_FUNC("uvm_fault_upper_promote"); UVMHIST_CALLED(maphist);
1382 
1383 	UVMHIST_LOG(maphist, "  case 1B: COW fault",0,0,0,0);
1384 	uvmexp.flt_acow++;
1385 
1386 	error = uvmfault_promote(ufi, oanon, PGO_DONTCARE, &anon,
1387 	    &flt->anon_spare);
1388 	switch (error) {
1389 	case 0:
1390 		break;
1391 	case ERESTART:
1392 		return ERESTART;
1393 	default:
1394 		return error;
1395 	}
1396 
1397 	KASSERT(anon == NULL || anon->an_lock == oanon->an_lock);
1398 
1399 	pg = anon->an_page;
1400 	mutex_enter(&uvm_pageqlock);
1401 	uvm_pageenqueue(pg); /* uvm_fault_upper_done will activate the page */
1402 	mutex_exit(&uvm_pageqlock);
1403 	pg->flags &= ~(PG_BUSY|PG_FAKE);
1404 	UVM_PAGE_OWN(pg, NULL);
1405 
1406 	/* deref: can not drop to zero here by defn! */
1407 	KASSERT(oanon->an_ref > 1);
1408 	oanon->an_ref--;
1409 
1410 	/*
1411 	 * note: oanon is still locked, as is the new anon.  we
1412 	 * need to check for this later when we unlock oanon; if
1413 	 * oanon != anon, we'll have to unlock anon, too.
1414 	 */
1415 
1416 	return uvm_fault_upper_enter(ufi, flt, uobj, anon, pg, oanon);
1417 }
1418 
1419 /*
1420  * uvm_fault_upper_direct: handle direct fault.
1421  */
1422 
1423 static int
1424 uvm_fault_upper_direct(
1425 	struct uvm_faultinfo *ufi, struct uvm_faultctx *flt,
1426 	struct uvm_object *uobj, struct vm_anon *anon)
1427 {
1428 	struct vm_anon * const oanon = anon;
1429 	struct vm_page *pg;
1430 	UVMHIST_FUNC("uvm_fault_upper_direct"); UVMHIST_CALLED(maphist);
1431 
1432 	uvmexp.flt_anon++;
1433 	pg = anon->an_page;
1434 	if (anon->an_ref > 1)     /* disallow writes to ref > 1 anons */
1435 		flt->enter_prot = flt->enter_prot & ~VM_PROT_WRITE;
1436 
1437 	return uvm_fault_upper_enter(ufi, flt, uobj, anon, pg, oanon);
1438 }
1439 
1440 /*
1441  * uvm_fault_upper_enter: enter h/w mapping of upper page.
1442  */
1443 
1444 static int
1445 uvm_fault_upper_enter(
1446 	struct uvm_faultinfo *ufi, const struct uvm_faultctx *flt,
1447 	struct uvm_object *uobj, struct vm_anon *anon, struct vm_page *pg,
1448 	struct vm_anon *oanon)
1449 {
1450 	struct vm_amap * const amap = ufi->entry->aref.ar_amap;
1451 	UVMHIST_FUNC("uvm_fault_upper_enter"); UVMHIST_CALLED(maphist);
1452 
1453 	/* locked: maps(read), amap, oanon, anon(if different from oanon) */
1454 	KASSERT(mutex_owned(amap->am_lock));
1455 	KASSERT(anon->an_lock == amap->am_lock);
1456 	KASSERT(oanon->an_lock == amap->am_lock);
1457 	KASSERT(uobj == NULL || mutex_owned(uobj->vmobjlock));
1458 
1459 	/*
1460 	 * now map the page in.
1461 	 */
1462 
1463 	UVMHIST_LOG(maphist,
1464 	    "  MAPPING: anon: pm=0x%x, va=0x%x, pg=0x%x, promote=%d",
1465 	    ufi->orig_map->pmap, ufi->orig_rvaddr, pg, flt->promote);
1466 	if (pmap_enter(ufi->orig_map->pmap, ufi->orig_rvaddr,
1467 	    VM_PAGE_TO_PHYS(pg),
1468 	    flt->enter_prot, flt->access_type | PMAP_CANFAIL |
1469 	    (flt->wire_mapping ? PMAP_WIRED : 0)) != 0) {
1470 
1471 		/*
1472 		 * No need to undo what we did; we can simply think of
1473 		 * this as the pmap throwing away the mapping information.
1474 		 *
1475 		 * We do, however, have to go through the ReFault path,
1476 		 * as the map may change while we're asleep.
1477 		 */
1478 
1479 		uvmfault_unlockall(ufi, amap, uobj);
1480 		if (!uvm_reclaimable()) {
1481 			UVMHIST_LOG(maphist,
1482 			    "<- failed.  out of VM",0,0,0,0);
1483 			/* XXX instrumentation */
1484 			return ENOMEM;
1485 		}
1486 		/* XXX instrumentation */
1487 		uvm_wait("flt_pmfail1");
1488 		return ERESTART;
1489 	}
1490 
1491 	uvm_fault_upper_done(ufi, flt, anon, pg);
1492 
1493 	/*
1494 	 * done case 1!  finish up by unlocking everything and returning success
1495 	 */
1496 
1497 	pmap_update(ufi->orig_map->pmap);
1498 	uvmfault_unlockall(ufi, amap, uobj);
1499 	return 0;
1500 }
1501 
1502 /*
1503  * uvm_fault_upper_done: queue upper center page.
1504  */
1505 
1506 static void
1507 uvm_fault_upper_done(
1508 	struct uvm_faultinfo *ufi, const struct uvm_faultctx *flt,
1509 	struct vm_anon *anon, struct vm_page *pg)
1510 {
1511 	const bool wire_paging = flt->wire_paging;
1512 
1513 	UVMHIST_FUNC("uvm_fault_upper_done"); UVMHIST_CALLED(maphist);
1514 
1515 	/*
1516 	 * ... update the page queues.
1517 	 */
1518 
1519 	mutex_enter(&uvm_pageqlock);
1520 	if (wire_paging) {
1521 		uvm_pagewire(pg);
1522 
1523 		/*
1524 		 * since the now-wired page cannot be paged out,
1525 		 * release its swap resources for others to use.
1526 		 * since an anon with no swap cannot be PG_CLEAN,
1527 		 * clear its clean flag now.
1528 		 */
1529 
1530 		pg->flags &= ~(PG_CLEAN);
1531 
1532 	} else {
1533 		uvm_pageactivate(pg);
1534 	}
1535 	mutex_exit(&uvm_pageqlock);
1536 
1537 	if (wire_paging) {
1538 		uvm_anon_dropswap(anon);
1539 	}
1540 }
1541 
1542 /*
1543  * uvm_fault_lower: handle lower fault.
1544  *
1545  *	1. check uobj
1546  *	1.1. if null, ZFOD.
1547  *	1.2. if not null, look up unnmapped neighbor pages.
1548  *	2. for center page, check if promote.
1549  *	2.1. ZFOD always needs promotion.
1550  *	2.2. other uobjs, when entry is marked COW (usually MAP_PRIVATE vnode).
1551  *	3. if uobj is not ZFOD and page is not found, do i/o.
1552  *	4. dispatch either direct / promote fault.
1553  */
1554 
1555 static int
1556 uvm_fault_lower(
1557 	struct uvm_faultinfo *ufi, struct uvm_faultctx *flt,
1558 	struct vm_page **pages)
1559 {
1560 #ifdef DIAGNOSTIC
1561 	struct vm_amap *amap = ufi->entry->aref.ar_amap;
1562 #endif
1563 	struct uvm_object *uobj = ufi->entry->object.uvm_obj;
1564 	struct vm_page *uobjpage;
1565 	int error;
1566 	UVMHIST_FUNC("uvm_fault_lower"); UVMHIST_CALLED(maphist);
1567 
1568 	/*
1569 	 * now, if the desired page is not shadowed by the amap and we have
1570 	 * a backing object that does not have a special fault routine, then
1571 	 * we ask (with pgo_get) the object for resident pages that we care
1572 	 * about and attempt to map them in.  we do not let pgo_get block
1573 	 * (PGO_LOCKED).
1574 	 */
1575 
1576 	if (uobj == NULL) {
1577 		/* zero fill; don't care neighbor pages */
1578 		uobjpage = NULL;
1579 	} else {
1580 		uvm_fault_lower_lookup(ufi, flt, pages);
1581 		uobjpage = pages[flt->centeridx];
1582 	}
1583 
1584 	/*
1585 	 * note that at this point we are done with any front or back pages.
1586 	 * we are now going to focus on the center page (i.e. the one we've
1587 	 * faulted on).  if we have faulted on the upper (anon) layer
1588 	 * [i.e. case 1], then the anon we want is anons[centeridx] (we have
1589 	 * not touched it yet).  if we have faulted on the bottom (uobj)
1590 	 * layer [i.e. case 2] and the page was both present and available,
1591 	 * then we've got a pointer to it as "uobjpage" and we've already
1592 	 * made it BUSY.
1593 	 */
1594 
1595 	/*
1596 	 * locked:
1597 	 * maps(read), amap(if there), uobj(if !null), uobjpage(if !null)
1598 	 */
1599 	KASSERT(amap == NULL || mutex_owned(amap->am_lock));
1600 	KASSERT(uobj == NULL || mutex_owned(uobj->vmobjlock));
1601 	KASSERT(uobjpage == NULL || (uobjpage->flags & PG_BUSY) != 0);
1602 
1603 	/*
1604 	 * note that uobjpage can not be PGO_DONTCARE at this point.  we now
1605 	 * set uobjpage to PGO_DONTCARE if we are doing a zero fill.  if we
1606 	 * have a backing object, check and see if we are going to promote
1607 	 * the data up to an anon during the fault.
1608 	 */
1609 
1610 	if (uobj == NULL) {
1611 		uobjpage = PGO_DONTCARE;
1612 		flt->promote = true;		/* always need anon here */
1613 	} else {
1614 		KASSERT(uobjpage != PGO_DONTCARE);
1615 		flt->promote = flt->cow_now && UVM_ET_ISCOPYONWRITE(ufi->entry);
1616 	}
1617 	UVMHIST_LOG(maphist, "  case 2 fault: promote=%d, zfill=%d",
1618 	    flt->promote, (uobj == NULL), 0,0);
1619 
1620 	/*
1621 	 * if uobjpage is not null then we do not need to do I/O to get the
1622 	 * uobjpage.
1623 	 *
1624 	 * if uobjpage is null, then we need to unlock and ask the pager to
1625 	 * get the data for us.   once we have the data, we need to reverify
1626 	 * the state the world.   we are currently not holding any resources.
1627 	 */
1628 
1629 	if (uobjpage) {
1630 		/* update rusage counters */
1631 		curlwp->l_ru.ru_minflt++;
1632 	} else {
1633 		error = uvm_fault_lower_io(ufi, flt, &uobj, &uobjpage);
1634 		if (error != 0)
1635 			return error;
1636 	}
1637 
1638 	/*
1639 	 * locked:
1640 	 * maps(read), amap(if !null), uobj(if !null), uobjpage(if uobj)
1641 	 */
1642 	KASSERT(amap == NULL || mutex_owned(amap->am_lock));
1643 	KASSERT(uobj == NULL || mutex_owned(uobj->vmobjlock));
1644 	KASSERT(uobj == NULL || (uobjpage->flags & PG_BUSY) != 0);
1645 
1646 	/*
1647 	 * notes:
1648 	 *  - at this point uobjpage can not be NULL
1649 	 *  - at this point uobjpage can not be PG_RELEASED (since we checked
1650 	 *  for it above)
1651 	 *  - at this point uobjpage could be PG_WANTED (handle later)
1652 	 */
1653 
1654 	KASSERT(uobjpage != NULL);
1655 	KASSERT(uobj == NULL || uobj == uobjpage->uobject);
1656 	KASSERT(uobj == NULL || !UVM_OBJ_IS_CLEAN(uobjpage->uobject) ||
1657 	    (uobjpage->flags & PG_CLEAN) != 0);
1658 
1659 	if (!flt->promote) {
1660 		error = uvm_fault_lower_direct(ufi, flt, uobj, uobjpage);
1661 	} else {
1662 		error = uvm_fault_lower_promote(ufi, flt, uobj, uobjpage);
1663 	}
1664 	return error;
1665 }
1666 
1667 /*
1668  * uvm_fault_lower_lookup: look up on-memory uobj pages.
1669  *
1670  *	1. get on-memory pages.
1671  *	2. if failed, give up (get only center page later).
1672  *	3. if succeeded, enter h/w mapping of neighbor pages.
1673  */
1674 
1675 static void
1676 uvm_fault_lower_lookup(
1677 	struct uvm_faultinfo *ufi, const struct uvm_faultctx *flt,
1678 	struct vm_page **pages)
1679 {
1680 	struct uvm_object *uobj = ufi->entry->object.uvm_obj;
1681 	int lcv, gotpages;
1682 	vaddr_t currva;
1683 	UVMHIST_FUNC("uvm_fault_lower_lookup"); UVMHIST_CALLED(maphist);
1684 
1685 	mutex_enter(uobj->vmobjlock);
1686 	/* Locked: maps(read), amap(if there), uobj */
1687 
1688 	uvmexp.fltlget++;
1689 	gotpages = flt->npages;
1690 	(void) uobj->pgops->pgo_get(uobj,
1691 	    ufi->entry->offset + flt->startva - ufi->entry->start,
1692 	    pages, &gotpages, flt->centeridx,
1693 	    flt->access_type & MASK(ufi->entry), ufi->entry->advice, PGO_LOCKED);
1694 
1695 	KASSERT(mutex_owned(uobj->vmobjlock));
1696 
1697 	/*
1698 	 * check for pages to map, if we got any
1699 	 */
1700 
1701 	if (gotpages == 0) {
1702 		pages[flt->centeridx] = NULL;
1703 		return;
1704 	}
1705 
1706 	currva = flt->startva;
1707 	for (lcv = 0; lcv < flt->npages; lcv++, currva += PAGE_SIZE) {
1708 		struct vm_page *curpg;
1709 
1710 		curpg = pages[lcv];
1711 		if (curpg == NULL || curpg == PGO_DONTCARE) {
1712 			continue;
1713 		}
1714 		KASSERT(curpg->uobject == uobj);
1715 
1716 		/*
1717 		 * if center page is resident and not PG_BUSY|PG_RELEASED
1718 		 * then pgo_get made it PG_BUSY for us and gave us a handle
1719 		 * to it.
1720 		 */
1721 
1722 		if (lcv == flt->centeridx) {
1723 			UVMHIST_LOG(maphist, "  got uobjpage "
1724 			    "(0x%x) with locked get",
1725 			    curpg, 0,0,0);
1726 		} else {
1727 			bool readonly = (curpg->flags & PG_RDONLY)
1728 			    || (curpg->loan_count > 0)
1729 			    || UVM_OBJ_NEEDS_WRITEFAULT(curpg->uobject);
1730 
1731 			uvm_fault_lower_neighbor(ufi, flt,
1732 			    currva, curpg, readonly);
1733 		}
1734 	}
1735 	pmap_update(ufi->orig_map->pmap);
1736 }
1737 
1738 /*
1739  * uvm_fault_lower_neighbor: enter h/w mapping of lower neighbor page.
1740  */
1741 
1742 static void
1743 uvm_fault_lower_neighbor(
1744 	struct uvm_faultinfo *ufi, const struct uvm_faultctx *flt,
1745 	vaddr_t currva, struct vm_page *pg, bool readonly)
1746 {
1747 	UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist);
1748 
1749 	/* locked: maps(read), amap(if there), uobj */
1750 
1751 	/*
1752 	 * calling pgo_get with PGO_LOCKED returns us pages which
1753 	 * are neither busy nor released, so we don't need to check
1754 	 * for this.  we can just directly enter the pages.
1755 	 */
1756 
1757 	mutex_enter(&uvm_pageqlock);
1758 	uvm_pageenqueue(pg);
1759 	mutex_exit(&uvm_pageqlock);
1760 	UVMHIST_LOG(maphist,
1761 	    "  MAPPING: n obj: pm=0x%x, va=0x%x, pg=0x%x",
1762 	    ufi->orig_map->pmap, currva, pg, 0);
1763 	uvmexp.fltnomap++;
1764 
1765 	/*
1766 	 * Since this page isn't the page that's actually faulting,
1767 	 * ignore pmap_enter() failures; it's not critical that we
1768 	 * enter these right now.
1769 	 * NOTE: page can't be PG_WANTED or PG_RELEASED because we've
1770 	 * held the lock the whole time we've had the handle.
1771 	 */
1772 	KASSERT((pg->flags & PG_PAGEOUT) == 0);
1773 	KASSERT((pg->flags & PG_RELEASED) == 0);
1774 	KASSERT((pg->flags & PG_WANTED) == 0);
1775 	KASSERT(!UVM_OBJ_IS_CLEAN(pg->uobject) || (pg->flags & PG_CLEAN) != 0);
1776 	pg->flags &= ~(PG_BUSY);
1777 	UVM_PAGE_OWN(pg, NULL);
1778 
1779 	KASSERT(mutex_owned(pg->uobject->vmobjlock));
1780 	(void) pmap_enter(ufi->orig_map->pmap, currva,
1781 	    VM_PAGE_TO_PHYS(pg),
1782 	    readonly ? (flt->enter_prot & ~VM_PROT_WRITE) :
1783 	    flt->enter_prot & MASK(ufi->entry),
1784 	    PMAP_CANFAIL | (flt->wire_mapping ? PMAP_WIRED : 0));
1785 }
1786 
1787 /*
1788  * uvm_fault_lower_io: get lower page from backing store.
1789  *
1790  *	1. unlock everything, because i/o will block.
1791  *	2. call pgo_get.
1792  *	3. if failed, recover.
1793  *	4. if succeeded, relock everything and verify things.
1794  */
1795 
1796 static int
1797 uvm_fault_lower_io(
1798 	struct uvm_faultinfo *ufi, const struct uvm_faultctx *flt,
1799 	struct uvm_object **ruobj, struct vm_page **ruobjpage)
1800 {
1801 	struct vm_amap * const amap = ufi->entry->aref.ar_amap;
1802 	struct uvm_object *uobj = *ruobj;
1803 	struct vm_page *pg;
1804 	bool locked;
1805 	int gotpages;
1806 	int error;
1807 	voff_t uoff;
1808 	UVMHIST_FUNC("uvm_fault_lower_io"); UVMHIST_CALLED(maphist);
1809 
1810 	/* update rusage counters */
1811 	curlwp->l_ru.ru_majflt++;
1812 
1813 	/* Locked: maps(read), amap(if there), uobj */
1814 	uvmfault_unlockall(ufi, amap, NULL);
1815 
1816 	/* Locked: uobj */
1817 	KASSERT(uobj == NULL || mutex_owned(uobj->vmobjlock));
1818 
1819 	uvmexp.fltget++;
1820 	gotpages = 1;
1821 	pg = NULL;
1822 	uoff = (ufi->orig_rvaddr - ufi->entry->start) + ufi->entry->offset;
1823 	error = uobj->pgops->pgo_get(uobj, uoff, &pg, &gotpages,
1824 	    0, flt->access_type & MASK(ufi->entry), ufi->entry->advice,
1825 	    PGO_SYNCIO);
1826 	/* locked: pg(if no error) */
1827 
1828 	/*
1829 	 * recover from I/O
1830 	 */
1831 
1832 	if (error) {
1833 		if (error == EAGAIN) {
1834 			UVMHIST_LOG(maphist,
1835 			    "  pgo_get says TRY AGAIN!",0,0,0,0);
1836 			kpause("fltagain2", false, hz/2, NULL);
1837 			return ERESTART;
1838 		}
1839 
1840 #if 0
1841 		KASSERT(error != ERESTART);
1842 #else
1843 		/* XXXUEBS don't re-fault? */
1844 		if (error == ERESTART)
1845 			error = EIO;
1846 #endif
1847 
1848 		UVMHIST_LOG(maphist, "<- pgo_get failed (code %d)",
1849 		    error, 0,0,0);
1850 		return error;
1851 	}
1852 
1853 	/*
1854 	 * re-verify the state of the world by first trying to relock
1855 	 * the maps.  always relock the object.
1856 	 */
1857 
1858 	locked = uvmfault_relock(ufi);
1859 	if (locked && amap)
1860 		amap_lock(amap);
1861 
1862 	/* might be changed */
1863 	uobj = pg->uobject;
1864 
1865 	mutex_enter(uobj->vmobjlock);
1866 	KASSERT((pg->flags & PG_BUSY) != 0);
1867 
1868 	mutex_enter(&uvm_pageqlock);
1869 	uvm_pageactivate(pg);
1870 	mutex_exit(&uvm_pageqlock);
1871 
1872 	/* locked(locked): maps(read), amap(if !null), uobj, pg */
1873 	/* locked(!locked): uobj, pg */
1874 
1875 	/*
1876 	 * verify that the page has not be released and re-verify
1877 	 * that amap slot is still free.   if there is a problem,
1878 	 * we unlock and clean up.
1879 	 */
1880 
1881 	if ((pg->flags & PG_RELEASED) != 0 ||
1882 	    (locked && amap && amap_lookup(&ufi->entry->aref,
1883 	      ufi->orig_rvaddr - ufi->entry->start))) {
1884 		if (locked)
1885 			uvmfault_unlockall(ufi, amap, NULL);
1886 		locked = false;
1887 	}
1888 
1889 	/*
1890 	 * didn't get the lock?   release the page and retry.
1891 	 */
1892 
1893 	if (locked == false) {
1894 		UVMHIST_LOG(maphist,
1895 		    "  wasn't able to relock after fault: retry",
1896 		    0,0,0,0);
1897 		if (pg->flags & PG_WANTED) {
1898 			wakeup(pg);
1899 		}
1900 		if ((pg->flags & PG_RELEASED) == 0) {
1901 			pg->flags &= ~(PG_BUSY | PG_WANTED);
1902 			UVM_PAGE_OWN(pg, NULL);
1903 		} else {
1904 			uvmexp.fltpgrele++;
1905 			uvm_pagefree(pg);
1906 		}
1907 		mutex_exit(uobj->vmobjlock);
1908 		return ERESTART;
1909 	}
1910 
1911 	/*
1912 	 * we have the data in pg which is busy and
1913 	 * not released.  we are holding object lock (so the page
1914 	 * can't be released on us).
1915 	 */
1916 
1917 	/* locked: maps(read), amap(if !null), uobj, pg */
1918 
1919 	*ruobj = uobj;
1920 	*ruobjpage = pg;
1921 	return 0;
1922 }
1923 
1924 /*
1925  * uvm_fault_lower_direct: fault lower center page
1926  *
1927  *	1. adjust flt->enter_prot.
1928  *	2. if page is loaned, resolve.
1929  */
1930 
1931 int
1932 uvm_fault_lower_direct(
1933 	struct uvm_faultinfo *ufi, struct uvm_faultctx *flt,
1934 	struct uvm_object *uobj, struct vm_page *uobjpage)
1935 {
1936 	struct vm_page *pg;
1937 	UVMHIST_FUNC("uvm_fault_lower_direct"); UVMHIST_CALLED(maphist);
1938 
1939 	/*
1940 	 * we are not promoting.   if the mapping is COW ensure that we
1941 	 * don't give more access than we should (e.g. when doing a read
1942 	 * fault on a COPYONWRITE mapping we want to map the COW page in
1943 	 * R/O even though the entry protection could be R/W).
1944 	 *
1945 	 * set "pg" to the page we want to map in (uobjpage, usually)
1946 	 */
1947 
1948 	uvmexp.flt_obj++;
1949 	if (UVM_ET_ISCOPYONWRITE(ufi->entry) ||
1950 	    UVM_OBJ_NEEDS_WRITEFAULT(uobjpage->uobject))
1951 		flt->enter_prot &= ~VM_PROT_WRITE;
1952 	pg = uobjpage;		/* map in the actual object */
1953 
1954 	KASSERT(uobjpage != PGO_DONTCARE);
1955 
1956 	/*
1957 	 * we are faulting directly on the page.   be careful
1958 	 * about writing to loaned pages...
1959 	 */
1960 
1961 	if (uobjpage->loan_count) {
1962 		uvm_fault_lower_direct_loan(ufi, flt, uobj, &pg, &uobjpage);
1963 	}
1964 	KASSERT(pg == uobjpage);
1965 
1966 	KASSERT(uobj == NULL || (uobjpage->flags & PG_BUSY) != 0);
1967 	return uvm_fault_lower_enter(ufi, flt, uobj, NULL, pg);
1968 }
1969 
1970 /*
1971  * uvm_fault_lower_direct_loan: resolve loaned page.
1972  *
1973  *	1. if not cow'ing, adjust flt->enter_prot.
1974  *	2. if cow'ing, break loan.
1975  */
1976 
1977 static int
1978 uvm_fault_lower_direct_loan(
1979 	struct uvm_faultinfo *ufi, struct uvm_faultctx *flt,
1980 	struct uvm_object *uobj, struct vm_page **rpg,
1981 	struct vm_page **ruobjpage)
1982 {
1983 	struct vm_amap * const amap = ufi->entry->aref.ar_amap;
1984 	struct vm_page *pg;
1985 	struct vm_page *uobjpage = *ruobjpage;
1986 	UVMHIST_FUNC("uvm_fault_lower_direct_loan"); UVMHIST_CALLED(maphist);
1987 
1988 	if (!flt->cow_now) {
1989 		/* read fault: cap the protection at readonly */
1990 		/* cap! */
1991 		flt->enter_prot = flt->enter_prot & ~VM_PROT_WRITE;
1992 	} else {
1993 		/* write fault: must break the loan here */
1994 
1995 		pg = uvm_loanbreak(uobjpage);
1996 		if (pg == NULL) {
1997 
1998 			/*
1999 			 * drop ownership of page, it can't be released
2000 			 */
2001 
2002 			if (uobjpage->flags & PG_WANTED)
2003 				wakeup(uobjpage);
2004 			uobjpage->flags &= ~(PG_BUSY|PG_WANTED);
2005 			UVM_PAGE_OWN(uobjpage, NULL);
2006 
2007 			uvmfault_unlockall(ufi, amap, uobj);
2008 			UVMHIST_LOG(maphist,
2009 			  "  out of RAM breaking loan, waiting",
2010 			  0,0,0,0);
2011 			uvmexp.fltnoram++;
2012 			uvm_wait("flt_noram4");
2013 			return ERESTART;
2014 		}
2015 		*rpg = pg;
2016 		*ruobjpage = pg;
2017 	}
2018 	return 0;
2019 }
2020 
2021 /*
2022  * uvm_fault_lower_promote: promote lower page.
2023  *
2024  *	1. call uvmfault_promote.
2025  *	2. fill in data.
2026  *	3. if not ZFOD, dispose old page.
2027  */
2028 
2029 int
2030 uvm_fault_lower_promote(
2031 	struct uvm_faultinfo *ufi, struct uvm_faultctx *flt,
2032 	struct uvm_object *uobj, struct vm_page *uobjpage)
2033 {
2034 	struct vm_amap * const amap = ufi->entry->aref.ar_amap;
2035 	struct vm_anon *anon;
2036 	struct vm_page *pg;
2037 	int error;
2038 	UVMHIST_FUNC("uvm_fault_lower_promote"); UVMHIST_CALLED(maphist);
2039 
2040 	KASSERT(amap != NULL);
2041 
2042 	/*
2043 	 * If we are going to promote the data to an anon we
2044 	 * allocate a blank anon here and plug it into our amap.
2045 	 */
2046 	error = uvmfault_promote(ufi, NULL, uobjpage,
2047 	    &anon, &flt->anon_spare);
2048 	switch (error) {
2049 	case 0:
2050 		break;
2051 	case ERESTART:
2052 		return ERESTART;
2053 	default:
2054 		return error;
2055 	}
2056 
2057 	pg = anon->an_page;
2058 
2059 	/*
2060 	 * Fill in the data.
2061 	 */
2062 	KASSERT(uobj == NULL || (uobjpage->flags & PG_BUSY) != 0);
2063 
2064 	if (uobjpage != PGO_DONTCARE) {
2065 		uvmexp.flt_prcopy++;
2066 
2067 		/*
2068 		 * promote to shared amap?  make sure all sharing
2069 		 * procs see it
2070 		 */
2071 
2072 		if ((amap_flags(amap) & AMAP_SHARED) != 0) {
2073 			pmap_page_protect(uobjpage, VM_PROT_NONE);
2074 			/*
2075 			 * XXX: PAGE MIGHT BE WIRED!
2076 			 */
2077 		}
2078 
2079 		/*
2080 		 * dispose of uobjpage.  it can't be PG_RELEASED
2081 		 * since we still hold the object lock.
2082 		 */
2083 
2084 		if (uobjpage->flags & PG_WANTED) {
2085 			/* still have the obj lock */
2086 			wakeup(uobjpage);
2087 		}
2088 		uobjpage->flags &= ~(PG_BUSY|PG_WANTED);
2089 		UVM_PAGE_OWN(uobjpage, NULL);
2090 
2091 		UVMHIST_LOG(maphist,
2092 		    "  promote uobjpage 0x%x to anon/page 0x%x/0x%x",
2093 		    uobjpage, anon, pg, 0);
2094 
2095 	} else {
2096 		uvmexp.flt_przero++;
2097 
2098 		/*
2099 		 * Page is zero'd and marked dirty by
2100 		 * uvmfault_promote().
2101 		 */
2102 
2103 		UVMHIST_LOG(maphist,"  zero fill anon/page 0x%x/0%x",
2104 		    anon, pg, 0, 0);
2105 	}
2106 
2107 	return uvm_fault_lower_enter(ufi, flt, uobj, anon, pg);
2108 }
2109 
2110 /*
2111  * uvm_fault_lower_enter: enter h/w mapping of lower page or anon page promoted
2112  * from the lower page.
2113  */
2114 
2115 int
2116 uvm_fault_lower_enter(
2117 	struct uvm_faultinfo *ufi, const struct uvm_faultctx *flt,
2118 	struct uvm_object *uobj,
2119 	struct vm_anon *anon, struct vm_page *pg)
2120 {
2121 	struct vm_amap * const amap = ufi->entry->aref.ar_amap;
2122 	int error;
2123 	UVMHIST_FUNC("uvm_fault_lower_enter"); UVMHIST_CALLED(maphist);
2124 
2125 	/*
2126 	 * Locked:
2127 	 *
2128 	 *	maps(read), amap(if !null), uobj(if !null),
2129 	 *	anon(if !null), pg(if anon), unlock_uobj(if !null)
2130 	 *
2131 	 * Note: pg is either the uobjpage or the new page in the new anon.
2132 	 */
2133 	KASSERT(amap == NULL || mutex_owned(amap->am_lock));
2134 	KASSERT(uobj == NULL || mutex_owned(uobj->vmobjlock));
2135 	KASSERT(anon == NULL || anon->an_lock == amap->am_lock);
2136 	KASSERT((pg->flags & PG_BUSY) != 0);
2137 
2138 	/*
2139 	 * all resources are present.   we can now map it in and free our
2140 	 * resources.
2141 	 */
2142 
2143 	UVMHIST_LOG(maphist,
2144 	    "  MAPPING: case2: pm=0x%x, va=0x%x, pg=0x%x, promote=%d",
2145 	    ufi->orig_map->pmap, ufi->orig_rvaddr, pg, flt->promote);
2146 	KASSERT((flt->access_type & VM_PROT_WRITE) == 0 ||
2147 		(pg->flags & PG_RDONLY) == 0);
2148 	if (pmap_enter(ufi->orig_map->pmap, ufi->orig_rvaddr,
2149 	    VM_PAGE_TO_PHYS(pg),
2150 	    (pg->flags & PG_RDONLY) != 0 ?
2151 	    flt->enter_prot & ~VM_PROT_WRITE : flt->enter_prot,
2152 	    flt->access_type | PMAP_CANFAIL |
2153 	    (flt->wire_mapping ? PMAP_WIRED : 0)) != 0) {
2154 
2155 		/*
2156 		 * No need to undo what we did; we can simply think of
2157 		 * this as the pmap throwing away the mapping information.
2158 		 *
2159 		 * We do, however, have to go through the ReFault path,
2160 		 * as the map may change while we're asleep.
2161 		 */
2162 
2163 		/*
2164 		 * ensure that the page is queued in the case that
2165 		 * we just promoted the page.
2166 		 */
2167 
2168 		mutex_enter(&uvm_pageqlock);
2169 		uvm_pageenqueue(pg);
2170 		mutex_exit(&uvm_pageqlock);
2171 
2172 		if (pg->flags & PG_WANTED)
2173 			wakeup(pg);
2174 
2175 		/*
2176 		 * note that pg can't be PG_RELEASED since we did not drop
2177 		 * the object lock since the last time we checked.
2178 		 */
2179 		KASSERT((pg->flags & PG_RELEASED) == 0);
2180 
2181 		pg->flags &= ~(PG_BUSY|PG_FAKE|PG_WANTED);
2182 		UVM_PAGE_OWN(pg, NULL);
2183 
2184 		uvmfault_unlockall(ufi, amap, uobj);
2185 		if (!uvm_reclaimable()) {
2186 			UVMHIST_LOG(maphist,
2187 			    "<- failed.  out of VM",0,0,0,0);
2188 			/* XXX instrumentation */
2189 			error = ENOMEM;
2190 			return error;
2191 		}
2192 		/* XXX instrumentation */
2193 		uvm_wait("flt_pmfail2");
2194 		return ERESTART;
2195 	}
2196 
2197 	uvm_fault_lower_done(ufi, flt, uobj, pg);
2198 
2199 	/*
2200 	 * note that pg can't be PG_RELEASED since we did not drop the object
2201 	 * lock since the last time we checked.
2202 	 */
2203 	KASSERT((pg->flags & PG_RELEASED) == 0);
2204 	if (pg->flags & PG_WANTED)
2205 		wakeup(pg);
2206 	pg->flags &= ~(PG_BUSY|PG_FAKE|PG_WANTED);
2207 	UVM_PAGE_OWN(pg, NULL);
2208 
2209 	pmap_update(ufi->orig_map->pmap);
2210 	uvmfault_unlockall(ufi, amap, uobj);
2211 
2212 	UVMHIST_LOG(maphist, "<- done (SUCCESS!)",0,0,0,0);
2213 	return 0;
2214 }
2215 
2216 /*
2217  * uvm_fault_lower_done: queue lower center page.
2218  */
2219 
2220 void
2221 uvm_fault_lower_done(
2222 	struct uvm_faultinfo *ufi, const struct uvm_faultctx *flt,
2223 	struct uvm_object *uobj, struct vm_page *pg)
2224 {
2225 	bool dropswap = false;
2226 
2227 	UVMHIST_FUNC("uvm_fault_lower_done"); UVMHIST_CALLED(maphist);
2228 
2229 	mutex_enter(&uvm_pageqlock);
2230 	if (flt->wire_paging) {
2231 		uvm_pagewire(pg);
2232 		if (pg->pqflags & PQ_AOBJ) {
2233 
2234 			/*
2235 			 * since the now-wired page cannot be paged out,
2236 			 * release its swap resources for others to use.
2237 			 * since an aobj page with no swap cannot be PG_CLEAN,
2238 			 * clear its clean flag now.
2239 			 */
2240 
2241 			KASSERT(uobj != NULL);
2242 			pg->flags &= ~(PG_CLEAN);
2243 			dropswap = true;
2244 		}
2245 	} else {
2246 		uvm_pageactivate(pg);
2247 	}
2248 	mutex_exit(&uvm_pageqlock);
2249 
2250 	if (dropswap) {
2251 		uao_dropswap(uobj, pg->offset >> PAGE_SHIFT);
2252 	}
2253 }
2254 
2255 
2256 /*
2257  * uvm_fault_wire: wire down a range of virtual addresses in a map.
2258  *
2259  * => map may be read-locked by caller, but MUST NOT be write-locked.
2260  * => if map is read-locked, any operations which may cause map to
2261  *	be write-locked in uvm_fault() must be taken care of by
2262  *	the caller.  See uvm_map_pageable().
2263  */
2264 
2265 int
2266 uvm_fault_wire(struct vm_map *map, vaddr_t start, vaddr_t end,
2267     vm_prot_t access_type, int maxprot)
2268 {
2269 	vaddr_t va;
2270 	int error;
2271 
2272 	/*
2273 	 * now fault it in a page at a time.   if the fault fails then we have
2274 	 * to undo what we have done.   note that in uvm_fault VM_PROT_NONE
2275 	 * is replaced with the max protection if fault_type is VM_FAULT_WIRE.
2276 	 */
2277 
2278 	/*
2279 	 * XXX work around overflowing a vaddr_t.  this prevents us from
2280 	 * wiring the last page in the address space, though.
2281 	 */
2282 	if (start > end) {
2283 		return EFAULT;
2284 	}
2285 
2286 	for (va = start; va < end; va += PAGE_SIZE) {
2287 		error = uvm_fault_internal(map, va, access_type,
2288 		    (maxprot ? UVM_FAULT_MAXPROT : 0) | UVM_FAULT_WIRE);
2289 		if (error) {
2290 			if (va != start) {
2291 				uvm_fault_unwire(map, start, va);
2292 			}
2293 			return error;
2294 		}
2295 	}
2296 	return 0;
2297 }
2298 
2299 /*
2300  * uvm_fault_unwire(): unwire range of virtual space.
2301  */
2302 
2303 void
2304 uvm_fault_unwire(struct vm_map *map, vaddr_t start, vaddr_t end)
2305 {
2306 	vm_map_lock_read(map);
2307 	uvm_fault_unwire_locked(map, start, end);
2308 	vm_map_unlock_read(map);
2309 }
2310 
2311 /*
2312  * uvm_fault_unwire_locked(): the guts of uvm_fault_unwire().
2313  *
2314  * => map must be at least read-locked.
2315  */
2316 
2317 void
2318 uvm_fault_unwire_locked(struct vm_map *map, vaddr_t start, vaddr_t end)
2319 {
2320 	struct vm_map_entry *entry, *oentry;
2321 	pmap_t pmap = vm_map_pmap(map);
2322 	vaddr_t va;
2323 	paddr_t pa;
2324 	struct vm_page *pg;
2325 
2326 	KASSERT((map->flags & VM_MAP_INTRSAFE) == 0);
2327 
2328 	/*
2329 	 * we assume that the area we are unwiring has actually been wired
2330 	 * in the first place.   this means that we should be able to extract
2331 	 * the PAs from the pmap.   we also lock out the page daemon so that
2332 	 * we can call uvm_pageunwire.
2333 	 */
2334 
2335 	/*
2336 	 * find the beginning map entry for the region.
2337 	 */
2338 
2339 	KASSERT(start >= vm_map_min(map) && end <= vm_map_max(map));
2340 	if (uvm_map_lookup_entry(map, start, &entry) == false)
2341 		panic("uvm_fault_unwire_locked: address not in map");
2342 
2343 	oentry = NULL;
2344 	for (va = start; va < end; va += PAGE_SIZE) {
2345 		if (pmap_extract(pmap, va, &pa) == false)
2346 			continue;
2347 
2348 		/*
2349 		 * find the map entry for the current address.
2350 		 */
2351 
2352 		KASSERT(va >= entry->start);
2353 		while (va >= entry->end) {
2354 			KASSERT(entry->next != &map->header &&
2355 				entry->next->start <= entry->end);
2356 			entry = entry->next;
2357 		}
2358 
2359 		/*
2360 		 * lock it.
2361 		 */
2362 
2363 		if (entry != oentry) {
2364 			if (oentry != NULL) {
2365 				mutex_exit(&uvm_pageqlock);
2366 				uvm_map_unlock_entry(oentry);
2367 			}
2368 			uvm_map_lock_entry(entry);
2369 			mutex_enter(&uvm_pageqlock);
2370 			oentry = entry;
2371 		}
2372 
2373 		/*
2374 		 * if the entry is no longer wired, tell the pmap.
2375 		 */
2376 
2377 		if (VM_MAPENT_ISWIRED(entry) == 0)
2378 			pmap_unwire(pmap, va);
2379 
2380 		pg = PHYS_TO_VM_PAGE(pa);
2381 		if (pg)
2382 			uvm_pageunwire(pg);
2383 	}
2384 
2385 	if (oentry != NULL) {
2386 		mutex_exit(&uvm_pageqlock);
2387 		uvm_map_unlock_entry(entry);
2388 	}
2389 }
2390