xref: /netbsd-src/sys/uvm/uvm_fault.c (revision a536ee5124e62c9a0051a252f7833dc8f50f44c9)
1 /*	$NetBSD: uvm_fault.c,v 1.194 2012/02/19 00:05:55 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 1.194 2012/02/19 00:05:55 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  * externs from other modules
173  */
174 
175 extern int start_init_exec;	/* Is init_main() done / init running? */
176 
177 /*
178  * inline functions
179  */
180 
181 /*
182  * uvmfault_anonflush: try and deactivate pages in specified anons
183  *
184  * => does not have to deactivate page if it is busy
185  */
186 
187 static inline void
188 uvmfault_anonflush(struct vm_anon **anons, int n)
189 {
190 	int lcv;
191 	struct vm_page *pg;
192 
193 	for (lcv = 0; lcv < n; lcv++) {
194 		if (anons[lcv] == NULL)
195 			continue;
196 		KASSERT(mutex_owned(anons[lcv]->an_lock));
197 		pg = anons[lcv]->an_page;
198 		if (pg && (pg->flags & PG_BUSY) == 0) {
199 			mutex_enter(&uvm_pageqlock);
200 			if (pg->wire_count == 0) {
201 				uvm_pagedeactivate(pg);
202 			}
203 			mutex_exit(&uvm_pageqlock);
204 		}
205 	}
206 }
207 
208 /*
209  * normal functions
210  */
211 
212 /*
213  * uvmfault_amapcopy: clear "needs_copy" in a map.
214  *
215  * => called with VM data structures unlocked (usually, see below)
216  * => we get a write lock on the maps and clear needs_copy for a VA
217  * => if we are out of RAM we sleep (waiting for more)
218  */
219 
220 static void
221 uvmfault_amapcopy(struct uvm_faultinfo *ufi)
222 {
223 	for (;;) {
224 
225 		/*
226 		 * no mapping?  give up.
227 		 */
228 
229 		if (uvmfault_lookup(ufi, true) == false)
230 			return;
231 
232 		/*
233 		 * copy if needed.
234 		 */
235 
236 		if (UVM_ET_ISNEEDSCOPY(ufi->entry))
237 			amap_copy(ufi->map, ufi->entry, AMAP_COPY_NOWAIT,
238 				ufi->orig_rvaddr, ufi->orig_rvaddr + 1);
239 
240 		/*
241 		 * didn't work?  must be out of RAM.   unlock and sleep.
242 		 */
243 
244 		if (UVM_ET_ISNEEDSCOPY(ufi->entry)) {
245 			uvmfault_unlockmaps(ufi, true);
246 			uvm_wait("fltamapcopy");
247 			continue;
248 		}
249 
250 		/*
251 		 * got it!   unlock and return.
252 		 */
253 
254 		uvmfault_unlockmaps(ufi, true);
255 		return;
256 	}
257 	/*NOTREACHED*/
258 }
259 
260 /*
261  * uvmfault_anonget: get data in an anon into a non-busy, non-released
262  * page in that anon.
263  *
264  * => Map, amap and thus anon should be locked by caller.
265  * => If we fail, we unlock everything and error is returned.
266  * => If we are successful, return with everything still locked.
267  * => We do not move the page on the queues [gets moved later].  If we
268  *    allocate a new page [we_own], it gets put on the queues.  Either way,
269  *    the result is that the page is on the queues at return time
270  * => For pages which are on loan from a uvm_object (and thus are not owned
271  *    by the anon): if successful, return with the owning object locked.
272  *    The caller must unlock this object when it unlocks everything else.
273  */
274 
275 int
276 uvmfault_anonget(struct uvm_faultinfo *ufi, struct vm_amap *amap,
277     struct vm_anon *anon)
278 {
279 	struct vm_page *pg;
280 	int error;
281 
282 	UVMHIST_FUNC("uvmfault_anonget"); UVMHIST_CALLED(maphist);
283 	KASSERT(mutex_owned(anon->an_lock));
284 	KASSERT(anon->an_lock == amap->am_lock);
285 
286 	/* Increment the counters.*/
287 	uvmexp.fltanget++;
288 	if (anon->an_page) {
289 		curlwp->l_ru.ru_minflt++;
290 	} else {
291 		curlwp->l_ru.ru_majflt++;
292 	}
293 	error = 0;
294 
295 	/*
296 	 * Loop until we get the anon data, or fail.
297 	 */
298 
299 	for (;;) {
300 		bool we_own, locked;
301 		/*
302 		 * Note: 'we_own' will become true if we set PG_BUSY on a page.
303 		 */
304 		we_own = false;
305 		pg = anon->an_page;
306 
307 		/*
308 		 * If there is a resident page and it is loaned, then anon
309 		 * may not own it.  Call out to uvm_anon_lockloanpg() to
310 		 * identify and lock the real owner of the page.
311 		 */
312 
313 		if (pg && pg->loan_count)
314 			pg = uvm_anon_lockloanpg(anon);
315 
316 		/*
317 		 * Is page resident?  Make sure it is not busy/released.
318 		 */
319 
320 		if (pg) {
321 
322 			/*
323 			 * at this point, if the page has a uobject [meaning
324 			 * we have it on loan], then that uobject is locked
325 			 * by us!   if the page is busy, we drop all the
326 			 * locks (including uobject) and try again.
327 			 */
328 
329 			if ((pg->flags & PG_BUSY) == 0) {
330 				UVMHIST_LOG(maphist, "<- OK",0,0,0,0);
331 				return 0;
332 			}
333 			pg->flags |= PG_WANTED;
334 			uvmexp.fltpgwait++;
335 
336 			/*
337 			 * The last unlock must be an atomic unlock and wait
338 			 * on the owner of page.
339 			 */
340 
341 			if (pg->uobject) {
342 				/* Owner of page is UVM object. */
343 				uvmfault_unlockall(ufi, amap, NULL);
344 				UVMHIST_LOG(maphist, " unlock+wait on uobj",0,
345 				    0,0,0);
346 				UVM_UNLOCK_AND_WAIT(pg,
347 				    pg->uobject->vmobjlock,
348 				    false, "anonget1", 0);
349 			} else {
350 				/* Owner of page is anon. */
351 				uvmfault_unlockall(ufi, NULL, NULL);
352 				UVMHIST_LOG(maphist, " unlock+wait on anon",0,
353 				    0,0,0);
354 				UVM_UNLOCK_AND_WAIT(pg, anon->an_lock,
355 				    false, "anonget2", 0);
356 			}
357 		} else {
358 #if defined(VMSWAP)
359 			/*
360 			 * No page, therefore allocate one.
361 			 */
362 
363 			pg = uvm_pagealloc(NULL,
364 			    ufi != NULL ? ufi->orig_rvaddr : 0,
365 			    anon, ufi != NULL ? UVM_FLAG_COLORMATCH : 0);
366 			if (pg == NULL) {
367 				/* Out of memory.  Wait a little. */
368 				uvmfault_unlockall(ufi, amap, NULL);
369 				uvmexp.fltnoram++;
370 				UVMHIST_LOG(maphist, "  noram -- UVM_WAIT",0,
371 				    0,0,0);
372 				if (!uvm_reclaimable()) {
373 					return ENOMEM;
374 				}
375 				uvm_wait("flt_noram1");
376 			} else {
377 				/* PG_BUSY bit is set. */
378 				we_own = true;
379 				uvmfault_unlockall(ufi, amap, NULL);
380 
381 				/*
382 				 * Pass a PG_BUSY+PG_FAKE+PG_CLEAN page into
383 				 * the uvm_swap_get() function with all data
384 				 * structures unlocked.  Note that it is OK
385 				 * to read an_swslot here, because we hold
386 				 * PG_BUSY on the page.
387 				 */
388 				uvmexp.pageins++;
389 				error = uvm_swap_get(pg, anon->an_swslot,
390 				    PGO_SYNCIO);
391 
392 				/*
393 				 * We clean up after the I/O below in the
394 				 * 'we_own' case.
395 				 */
396 			}
397 #else
398 			panic("%s: no page", __func__);
399 #endif /* defined(VMSWAP) */
400 		}
401 
402 		/*
403 		 * Re-lock the map and anon.
404 		 */
405 
406 		locked = uvmfault_relock(ufi);
407 		if (locked || we_own) {
408 			mutex_enter(anon->an_lock);
409 		}
410 
411 		/*
412 		 * If we own the page (i.e. we set PG_BUSY), then we need
413 		 * to clean up after the I/O.  There are three cases to
414 		 * consider:
415 		 *
416 		 * 1) Page was released during I/O: free anon and ReFault.
417 		 * 2) I/O not OK.  Free the page and cause the fault to fail.
418 		 * 3) I/O OK!  Activate the page and sync with the non-we_own
419 		 *    case (i.e. drop anon lock if not locked).
420 		 */
421 
422 		if (we_own) {
423 #if defined(VMSWAP)
424 			if (pg->flags & PG_WANTED) {
425 				wakeup(pg);
426 			}
427 			if (error) {
428 
429 				/*
430 				 * Remove the swap slot from the anon and
431 				 * mark the anon as having no real slot.
432 				 * Do not free the swap slot, thus preventing
433 				 * it from being used again.
434 				 */
435 
436 				if (anon->an_swslot > 0) {
437 					uvm_swap_markbad(anon->an_swslot, 1);
438 				}
439 				anon->an_swslot = SWSLOT_BAD;
440 
441 				if ((pg->flags & PG_RELEASED) != 0) {
442 					goto released;
443 				}
444 
445 				/*
446 				 * Note: page was never !PG_BUSY, so it
447 				 * cannot be mapped and thus no need to
448 				 * pmap_page_protect() it.
449 				 */
450 
451 				mutex_enter(&uvm_pageqlock);
452 				uvm_pagefree(pg);
453 				mutex_exit(&uvm_pageqlock);
454 
455 				if (locked) {
456 					uvmfault_unlockall(ufi, NULL, NULL);
457 				}
458 				mutex_exit(anon->an_lock);
459 				UVMHIST_LOG(maphist, "<- ERROR", 0,0,0,0);
460 				return error;
461 			}
462 
463 			if ((pg->flags & PG_RELEASED) != 0) {
464 released:
465 				KASSERT(anon->an_ref == 0);
466 
467 				/*
468 				 * Released while we had unlocked amap.
469 				 */
470 
471 				if (locked) {
472 					uvmfault_unlockall(ufi, NULL, NULL);
473 				}
474 				uvm_anon_release(anon);
475 
476 				if (error) {
477 					UVMHIST_LOG(maphist,
478 					    "<- ERROR/RELEASED", 0,0,0,0);
479 					return error;
480 				}
481 
482 				UVMHIST_LOG(maphist, "<- RELEASED", 0,0,0,0);
483 				return ERESTART;
484 			}
485 
486 			/*
487 			 * We have successfully read the page, activate it.
488 			 */
489 
490 			mutex_enter(&uvm_pageqlock);
491 			uvm_pageactivate(pg);
492 			mutex_exit(&uvm_pageqlock);
493 			pg->flags &= ~(PG_WANTED|PG_BUSY|PG_FAKE);
494 			UVM_PAGE_OWN(pg, NULL);
495 #else
496 			panic("%s: we_own", __func__);
497 #endif /* defined(VMSWAP) */
498 		}
499 
500 		/*
501 		 * We were not able to re-lock the map - restart the fault.
502 		 */
503 
504 		if (!locked) {
505 			if (we_own) {
506 				mutex_exit(anon->an_lock);
507 			}
508 			UVMHIST_LOG(maphist, "<- REFAULT", 0,0,0,0);
509 			return ERESTART;
510 		}
511 
512 		/*
513 		 * Verify that no one has touched the amap and moved
514 		 * the anon on us.
515 		 */
516 
517 		if (ufi != NULL && amap_lookup(&ufi->entry->aref,
518 		    ufi->orig_rvaddr - ufi->entry->start) != anon) {
519 
520 			uvmfault_unlockall(ufi, amap, NULL);
521 			UVMHIST_LOG(maphist, "<- REFAULT", 0,0,0,0);
522 			return ERESTART;
523 		}
524 
525 		/*
526 		 * Retry..
527 		 */
528 
529 		uvmexp.fltanretry++;
530 		continue;
531 	}
532 	/*NOTREACHED*/
533 }
534 
535 /*
536  * uvmfault_promote: promote data to a new anon.  used for 1B and 2B.
537  *
538  *	1. allocate an anon and a page.
539  *	2. fill its contents.
540  *	3. put it into amap.
541  *
542  * => if we fail (result != 0) we unlock everything.
543  * => on success, return a new locked anon via 'nanon'.
544  *    (*nanon)->an_page will be a resident, locked, dirty page.
545  * => it's caller's responsibility to put the promoted nanon->an_page to the
546  *    page queue.
547  */
548 
549 static int
550 uvmfault_promote(struct uvm_faultinfo *ufi,
551     struct vm_anon *oanon,
552     struct vm_page *uobjpage,
553     struct vm_anon **nanon, /* OUT: allocated anon */
554     struct vm_anon **spare)
555 {
556 	struct vm_amap *amap = ufi->entry->aref.ar_amap;
557 	struct uvm_object *uobj;
558 	struct vm_anon *anon;
559 	struct vm_page *pg;
560 	struct vm_page *opg;
561 	int error;
562 	UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist);
563 
564 	if (oanon) {
565 		/* anon COW */
566 		opg = oanon->an_page;
567 		KASSERT(opg != NULL);
568 		KASSERT(opg->uobject == NULL || opg->loan_count > 0);
569 	} else if (uobjpage != PGO_DONTCARE) {
570 		/* object-backed COW */
571 		opg = uobjpage;
572 	} else {
573 		/* ZFOD */
574 		opg = NULL;
575 	}
576 	if (opg != NULL) {
577 		uobj = opg->uobject;
578 	} else {
579 		uobj = NULL;
580 	}
581 
582 	KASSERT(amap != NULL);
583 	KASSERT(uobjpage != NULL);
584 	KASSERT(uobjpage == PGO_DONTCARE || (uobjpage->flags & PG_BUSY) != 0);
585 	KASSERT(mutex_owned(amap->am_lock));
586 	KASSERT(oanon == NULL || amap->am_lock == oanon->an_lock);
587 	KASSERT(uobj == NULL || mutex_owned(uobj->vmobjlock));
588 
589 	if (*spare != NULL) {
590 		anon = *spare;
591 		*spare = NULL;
592 	} else {
593 		anon = uvm_analloc();
594 	}
595 	if (anon) {
596 
597 		/*
598 		 * The new anon is locked.
599 		 *
600 		 * if opg == NULL, we want a zero'd, dirty page,
601 		 * so have uvm_pagealloc() do that for us.
602 		 */
603 
604 		KASSERT(anon->an_lock == NULL);
605 		anon->an_lock = amap->am_lock;
606 		pg = uvm_pagealloc(NULL, ufi->orig_rvaddr, anon,
607 		    UVM_FLAG_COLORMATCH | (opg == NULL ? UVM_PGA_ZERO : 0));
608 		if (pg == NULL) {
609 			anon->an_lock = NULL;
610 		}
611 	} else {
612 		pg = NULL;
613 	}
614 
615 	/*
616 	 * out of memory resources?
617 	 */
618 
619 	if (pg == NULL) {
620 		/* save anon for the next try. */
621 		if (anon != NULL) {
622 			*spare = anon;
623 		}
624 
625 		/* unlock and fail ... */
626 		uvm_page_unbusy(&uobjpage, 1);
627 		uvmfault_unlockall(ufi, amap, uobj);
628 		if (!uvm_reclaimable()) {
629 			UVMHIST_LOG(maphist, "out of VM", 0,0,0,0);
630 			uvmexp.fltnoanon++;
631 			error = ENOMEM;
632 			goto done;
633 		}
634 
635 		UVMHIST_LOG(maphist, "out of RAM, waiting for more", 0,0,0,0);
636 		uvmexp.fltnoram++;
637 		uvm_wait("flt_noram5");
638 		error = ERESTART;
639 		goto done;
640 	}
641 
642 	/* copy page [pg now dirty] */
643 	if (opg) {
644 		uvm_pagecopy(opg, pg);
645 	}
646 
647 	amap_add(&ufi->entry->aref, ufi->orig_rvaddr - ufi->entry->start, anon,
648 	    oanon != NULL);
649 
650 	*nanon = anon;
651 	error = 0;
652 done:
653 	return error;
654 }
655 
656 
657 /*
658  *   F A U L T   -   m a i n   e n t r y   p o i n t
659  */
660 
661 /*
662  * uvm_fault: page fault handler
663  *
664  * => called from MD code to resolve a page fault
665  * => VM data structures usually should be unlocked.   however, it is
666  *	possible to call here with the main map locked if the caller
667  *	gets a write lock, sets it recusive, and then calls us (c.f.
668  *	uvm_map_pageable).   this should be avoided because it keeps
669  *	the map locked off during I/O.
670  * => MUST NEVER BE CALLED IN INTERRUPT CONTEXT
671  */
672 
673 #define MASK(entry)     (UVM_ET_ISCOPYONWRITE(entry) ? \
674 			 ~VM_PROT_WRITE : VM_PROT_ALL)
675 
676 /* fault_flag values passed from uvm_fault_wire to uvm_fault_internal */
677 #define UVM_FAULT_WIRE		(1 << 0)
678 #define UVM_FAULT_MAXPROT	(1 << 1)
679 
680 struct uvm_faultctx {
681 
682 	/*
683 	 * the following members are set up by uvm_fault_check() and
684 	 * read-only after that.
685 	 *
686 	 * note that narrow is used by uvm_fault_check() to change
687 	 * the behaviour after ERESTART.
688 	 *
689 	 * most of them might change after RESTART if the underlying
690 	 * map entry has been changed behind us.  an exception is
691 	 * wire_paging, which does never change.
692 	 */
693 	vm_prot_t access_type;
694 	vaddr_t startva;
695 	int npages;
696 	int centeridx;
697 	bool narrow;		/* work on a single requested page only */
698 	bool wire_mapping;	/* request a PMAP_WIRED mapping
699 				   (UVM_FAULT_WIRE or VM_MAPENT_ISWIRED) */
700 	bool wire_paging;	/* request uvm_pagewire
701 				   (true for UVM_FAULT_WIRE) */
702 	bool cow_now;		/* VM_PROT_WRITE is actually requested
703 				   (ie. should break COW and page loaning) */
704 
705 	/*
706 	 * enter_prot is set up by uvm_fault_check() and clamped
707 	 * (ie. drop the VM_PROT_WRITE bit) in various places in case
708 	 * of !cow_now.
709 	 */
710 	vm_prot_t enter_prot;	/* prot at which we want to enter pages in */
711 
712 	/*
713 	 * the following member is for uvmfault_promote() and ERESTART.
714 	 */
715 	struct vm_anon *anon_spare;
716 
717 	/*
718 	 * the folloing is actually a uvm_fault_lower() internal.
719 	 * it's here merely for debugging.
720 	 * (or due to the mechanical separation of the function?)
721 	 */
722 	bool promote;
723 };
724 
725 static inline int	uvm_fault_check(
726 			    struct uvm_faultinfo *, struct uvm_faultctx *,
727 			    struct vm_anon ***, bool);
728 
729 static int		uvm_fault_upper(
730 			    struct uvm_faultinfo *, struct uvm_faultctx *,
731 			    struct vm_anon **);
732 static inline int	uvm_fault_upper_lookup(
733 			    struct uvm_faultinfo *, const struct uvm_faultctx *,
734 			    struct vm_anon **, struct vm_page **);
735 static inline void	uvm_fault_upper_neighbor(
736 			    struct uvm_faultinfo *, const struct uvm_faultctx *,
737 			    vaddr_t, struct vm_page *, bool);
738 static inline int	uvm_fault_upper_loan(
739 			    struct uvm_faultinfo *, struct uvm_faultctx *,
740 			    struct vm_anon *, struct uvm_object **);
741 static inline int	uvm_fault_upper_promote(
742 			    struct uvm_faultinfo *, struct uvm_faultctx *,
743 			    struct uvm_object *, struct vm_anon *);
744 static inline int	uvm_fault_upper_direct(
745 			    struct uvm_faultinfo *, struct uvm_faultctx *,
746 			    struct uvm_object *, struct vm_anon *);
747 static int		uvm_fault_upper_enter(
748 			    struct uvm_faultinfo *, const struct uvm_faultctx *,
749 			    struct uvm_object *, struct vm_anon *,
750 			    struct vm_page *, struct vm_anon *);
751 static inline void	uvm_fault_upper_done(
752 			    struct uvm_faultinfo *, const struct uvm_faultctx *,
753 			    struct vm_anon *, struct vm_page *);
754 
755 static int		uvm_fault_lower(
756 			    struct uvm_faultinfo *, struct uvm_faultctx *,
757 			    struct vm_page **);
758 static inline void	uvm_fault_lower_lookup(
759 			    struct uvm_faultinfo *, const struct uvm_faultctx *,
760 			    struct vm_page **);
761 static inline void	uvm_fault_lower_neighbor(
762 			    struct uvm_faultinfo *, const struct uvm_faultctx *,
763 			    vaddr_t, struct vm_page *, bool);
764 static inline int	uvm_fault_lower_io(
765 			    struct uvm_faultinfo *, const struct uvm_faultctx *,
766 			    struct uvm_object **, struct vm_page **);
767 static inline int	uvm_fault_lower_direct(
768 			    struct uvm_faultinfo *, struct uvm_faultctx *,
769 			    struct uvm_object *, struct vm_page *);
770 static inline int	uvm_fault_lower_direct_loan(
771 			    struct uvm_faultinfo *, struct uvm_faultctx *,
772 			    struct uvm_object *, struct vm_page **,
773 			    struct vm_page **);
774 static inline int	uvm_fault_lower_promote(
775 			    struct uvm_faultinfo *, struct uvm_faultctx *,
776 			    struct uvm_object *, struct vm_page *);
777 static int		uvm_fault_lower_enter(
778 			    struct uvm_faultinfo *, const struct uvm_faultctx *,
779 			    struct uvm_object *,
780 			    struct vm_anon *, struct vm_page *);
781 static inline void	uvm_fault_lower_done(
782 			    struct uvm_faultinfo *, const struct uvm_faultctx *,
783 			    struct uvm_object *, struct vm_page *);
784 
785 int
786 uvm_fault_internal(struct vm_map *orig_map, vaddr_t vaddr,
787     vm_prot_t access_type, int fault_flag)
788 {
789 	struct cpu_data *cd;
790 	struct uvm_cpu *ucpu;
791 	struct uvm_faultinfo ufi;
792 	struct uvm_faultctx flt = {
793 		.access_type = access_type,
794 
795 		/* don't look for neighborhood * pages on "wire" fault */
796 		.narrow = (fault_flag & UVM_FAULT_WIRE) != 0,
797 
798 		/* "wire" fault causes wiring of both mapping and paging */
799 		.wire_mapping = (fault_flag & UVM_FAULT_WIRE) != 0,
800 		.wire_paging = (fault_flag & UVM_FAULT_WIRE) != 0,
801 	};
802 	const bool maxprot = (fault_flag & UVM_FAULT_MAXPROT) != 0;
803 	struct vm_anon *anons_store[UVM_MAXRANGE], **anons;
804 	struct vm_page *pages_store[UVM_MAXRANGE], **pages;
805 	int error;
806 #if 0
807 	uintptr_t delta, delta2, delta3;
808 #endif
809 	UVMHIST_FUNC("uvm_fault"); UVMHIST_CALLED(maphist);
810 
811 	UVMHIST_LOG(maphist, "(map=0x%x, vaddr=0x%x, at=%d, ff=%d)",
812 	      orig_map, vaddr, access_type, fault_flag);
813 
814 	cd = &(curcpu()->ci_data);
815 	cd->cpu_nfault++;
816 	ucpu = cd->cpu_uvm;
817 
818 	/* Don't flood RNG subsystem with samples. */
819 	if (cd->cpu_nfault % 503)
820 		goto norng;
821 #if 0
822 	/*
823 	 * Avoid trying to count "entropy" for accesses of regular
824 	 * stride, by checking the 1st, 2nd, 3rd order differentials
825 	 * of vaddr, like the rnd code does internally with sample times.
826 	 *
827 	 * XXX If the selection of only every 503rd fault above is
828 	 * XXX removed, this code should exclude most samples, but
829 	 * XXX does not, and is therefore disabled.
830 	 */
831 	if (ucpu->last_fltaddr > (uintptr_t)trunc_page(vaddr))
832 		delta = ucpu->last_fltaddr - (uintptr_t)trunc_page(vaddr);
833 	else
834 		delta = (uintptr_t)trunc_page(vaddr) - ucpu->last_fltaddr;
835 
836 	if (ucpu->last_delta > delta)
837 		delta2 = ucpu->last_delta - delta;
838 	else
839 		delta2 = delta - ucpu->last_delta;
840 
841 	if (ucpu->last_delta2 > delta2)
842 		delta3 = ucpu->last_delta2 - delta2;
843 	else
844 		delta3 = delta2 - ucpu->last_delta2;
845 
846 	ucpu->last_fltaddr = (uintptr_t)vaddr;
847 	ucpu->last_delta = delta;
848 	ucpu->last_delta2 = delta2;
849 
850 	if (delta != 0 && delta2 != 0 && delta3 != 0)
851 #endif
852 	/* Don't count anything until user interaction is possible */
853 	if (__predict_true(start_init_exec)) {
854 		kpreempt_disable();
855 		rnd_add_uint32(&ucpu->rs,
856 			       sizeof(vaddr_t) == sizeof(uint32_t) ?
857 			       (uint32_t)vaddr : sizeof(vaddr_t) ==
858 			       sizeof(uint64_t) ?
859 			       (uint32_t)(vaddr & 0x00000000ffffffff) :
860 			       (uint32_t)(cd->cpu_nfault & 0x00000000ffffffff));
861 		kpreempt_enable();
862 	}
863 norng:
864 	/*
865 	 * init the IN parameters in the ufi
866 	 */
867 
868 	ufi.orig_map = orig_map;
869 	ufi.orig_rvaddr = trunc_page(vaddr);
870 	ufi.orig_size = PAGE_SIZE;	/* can't get any smaller than this */
871 
872 	error = ERESTART;
873 	while (error == ERESTART) { /* ReFault: */
874 		anons = anons_store;
875 		pages = pages_store;
876 
877 		error = uvm_fault_check(&ufi, &flt, &anons, maxprot);
878 		if (error != 0)
879 			continue;
880 
881 		error = uvm_fault_upper_lookup(&ufi, &flt, anons, pages);
882 		if (error != 0)
883 			continue;
884 
885 		if (pages[flt.centeridx] == PGO_DONTCARE)
886 			error = uvm_fault_upper(&ufi, &flt, anons);
887 		else {
888 			struct uvm_object * const uobj =
889 			    ufi.entry->object.uvm_obj;
890 
891 			if (uobj && uobj->pgops->pgo_fault != NULL) {
892 				/*
893 				 * invoke "special" fault routine.
894 				 */
895 				mutex_enter(uobj->vmobjlock);
896 				/* locked: maps(read), amap(if there), uobj */
897 				error = uobj->pgops->pgo_fault(&ufi,
898 				    flt.startva, pages, flt.npages,
899 				    flt.centeridx, flt.access_type,
900 				    PGO_LOCKED|PGO_SYNCIO);
901 
902 				/*
903 				 * locked: nothing, pgo_fault has unlocked
904 				 * everything
905 				 */
906 
907 				/*
908 				 * object fault routine responsible for
909 				 * pmap_update().
910 				 */
911 			} else {
912 				error = uvm_fault_lower(&ufi, &flt, pages);
913 			}
914 		}
915 	}
916 
917 	if (flt.anon_spare != NULL) {
918 		flt.anon_spare->an_ref--;
919 		KASSERT(flt.anon_spare->an_ref == 0);
920 		KASSERT(flt.anon_spare->an_lock == NULL);
921 		uvm_anon_free(flt.anon_spare);
922 	}
923 	return error;
924 }
925 
926 /*
927  * uvm_fault_check: check prot, handle needs-copy, etc.
928  *
929  *	1. lookup entry.
930  *	2. check protection.
931  *	3. adjust fault condition (mainly for simulated fault).
932  *	4. handle needs-copy (lazy amap copy).
933  *	5. establish range of interest for neighbor fault (aka pre-fault).
934  *	6. look up anons (if amap exists).
935  *	7. flush pages (if MADV_SEQUENTIAL)
936  *
937  * => called with nothing locked.
938  * => if we fail (result != 0) we unlock everything.
939  * => initialize/adjust many members of flt.
940  */
941 
942 static int
943 uvm_fault_check(
944 	struct uvm_faultinfo *ufi, struct uvm_faultctx *flt,
945 	struct vm_anon ***ranons, bool maxprot)
946 {
947 	struct vm_amap *amap;
948 	struct uvm_object *uobj;
949 	vm_prot_t check_prot;
950 	int nback, nforw;
951 	UVMHIST_FUNC("uvm_fault_check"); UVMHIST_CALLED(maphist);
952 
953 	/*
954 	 * lookup and lock the maps
955 	 */
956 
957 	if (uvmfault_lookup(ufi, false) == false) {
958 		UVMHIST_LOG(maphist, "<- no mapping @ 0x%x", ufi->orig_rvaddr,
959 		    0,0,0);
960 		return EFAULT;
961 	}
962 	/* locked: maps(read) */
963 
964 #ifdef DIAGNOSTIC
965 	if ((ufi->map->flags & VM_MAP_PAGEABLE) == 0) {
966 		printf("Page fault on non-pageable map:\n");
967 		printf("ufi->map = %p\n", ufi->map);
968 		printf("ufi->orig_map = %p\n", ufi->orig_map);
969 		printf("ufi->orig_rvaddr = 0x%lx\n", (u_long) ufi->orig_rvaddr);
970 		panic("uvm_fault: (ufi->map->flags & VM_MAP_PAGEABLE) == 0");
971 	}
972 #endif
973 
974 	/*
975 	 * check protection
976 	 */
977 
978 	check_prot = maxprot ?
979 	    ufi->entry->max_protection : ufi->entry->protection;
980 	if ((check_prot & flt->access_type) != flt->access_type) {
981 		UVMHIST_LOG(maphist,
982 		    "<- protection failure (prot=0x%x, access=0x%x)",
983 		    ufi->entry->protection, flt->access_type, 0, 0);
984 		uvmfault_unlockmaps(ufi, false);
985 		return EACCES;
986 	}
987 
988 	/*
989 	 * "enter_prot" is the protection we want to enter the page in at.
990 	 * for certain pages (e.g. copy-on-write pages) this protection can
991 	 * be more strict than ufi->entry->protection.  "wired" means either
992 	 * the entry is wired or we are fault-wiring the pg.
993 	 */
994 
995 	flt->enter_prot = ufi->entry->protection;
996 	if (VM_MAPENT_ISWIRED(ufi->entry))
997 		flt->wire_mapping = true;
998 
999 	if (flt->wire_mapping) {
1000 		flt->access_type = flt->enter_prot; /* full access for wired */
1001 		flt->cow_now = (check_prot & VM_PROT_WRITE) != 0;
1002 	} else {
1003 		flt->cow_now = (flt->access_type & VM_PROT_WRITE) != 0;
1004 	}
1005 
1006 	flt->promote = false;
1007 
1008 	/*
1009 	 * handle "needs_copy" case.   if we need to copy the amap we will
1010 	 * have to drop our readlock and relock it with a write lock.  (we
1011 	 * need a write lock to change anything in a map entry [e.g.
1012 	 * needs_copy]).
1013 	 */
1014 
1015 	if (UVM_ET_ISNEEDSCOPY(ufi->entry)) {
1016 		if (flt->cow_now || (ufi->entry->object.uvm_obj == NULL)) {
1017 			KASSERT(!maxprot);
1018 			/* need to clear */
1019 			UVMHIST_LOG(maphist,
1020 			    "  need to clear needs_copy and refault",0,0,0,0);
1021 			uvmfault_unlockmaps(ufi, false);
1022 			uvmfault_amapcopy(ufi);
1023 			uvmexp.fltamcopy++;
1024 			return ERESTART;
1025 
1026 		} else {
1027 
1028 			/*
1029 			 * ensure that we pmap_enter page R/O since
1030 			 * needs_copy is still true
1031 			 */
1032 
1033 			flt->enter_prot &= ~VM_PROT_WRITE;
1034 		}
1035 	}
1036 
1037 	/*
1038 	 * identify the players
1039 	 */
1040 
1041 	amap = ufi->entry->aref.ar_amap;	/* upper layer */
1042 	uobj = ufi->entry->object.uvm_obj;	/* lower layer */
1043 
1044 	/*
1045 	 * check for a case 0 fault.  if nothing backing the entry then
1046 	 * error now.
1047 	 */
1048 
1049 	if (amap == NULL && uobj == NULL) {
1050 		uvmfault_unlockmaps(ufi, false);
1051 		UVMHIST_LOG(maphist,"<- no backing store, no overlay",0,0,0,0);
1052 		return EFAULT;
1053 	}
1054 
1055 	/*
1056 	 * establish range of interest based on advice from mapper
1057 	 * and then clip to fit map entry.   note that we only want
1058 	 * to do this the first time through the fault.   if we
1059 	 * ReFault we will disable this by setting "narrow" to true.
1060 	 */
1061 
1062 	if (flt->narrow == false) {
1063 
1064 		/* wide fault (!narrow) */
1065 		KASSERT(uvmadvice[ufi->entry->advice].advice ==
1066 			 ufi->entry->advice);
1067 		nback = MIN(uvmadvice[ufi->entry->advice].nback,
1068 		    (ufi->orig_rvaddr - ufi->entry->start) >> PAGE_SHIFT);
1069 		flt->startva = ufi->orig_rvaddr - (nback << PAGE_SHIFT);
1070 		/*
1071 		 * note: "-1" because we don't want to count the
1072 		 * faulting page as forw
1073 		 */
1074 		nforw = MIN(uvmadvice[ufi->entry->advice].nforw,
1075 			    ((ufi->entry->end - ufi->orig_rvaddr) >>
1076 			     PAGE_SHIFT) - 1);
1077 		flt->npages = nback + nforw + 1;
1078 		flt->centeridx = nback;
1079 
1080 		flt->narrow = true;	/* ensure only once per-fault */
1081 
1082 	} else {
1083 
1084 		/* narrow fault! */
1085 		nback = nforw = 0;
1086 		flt->startva = ufi->orig_rvaddr;
1087 		flt->npages = 1;
1088 		flt->centeridx = 0;
1089 
1090 	}
1091 	/* offset from entry's start to pgs' start */
1092 	const voff_t eoff = flt->startva - ufi->entry->start;
1093 
1094 	/* locked: maps(read) */
1095 	UVMHIST_LOG(maphist, "  narrow=%d, back=%d, forw=%d, startva=0x%x",
1096 		    flt->narrow, nback, nforw, flt->startva);
1097 	UVMHIST_LOG(maphist, "  entry=0x%x, amap=0x%x, obj=0x%x", ufi->entry,
1098 		    amap, uobj, 0);
1099 
1100 	/*
1101 	 * if we've got an amap, lock it and extract current anons.
1102 	 */
1103 
1104 	if (amap) {
1105 		amap_lock(amap);
1106 		amap_lookups(&ufi->entry->aref, eoff, *ranons, flt->npages);
1107 	} else {
1108 		*ranons = NULL;	/* to be safe */
1109 	}
1110 
1111 	/* locked: maps(read), amap(if there) */
1112 	KASSERT(amap == NULL || mutex_owned(amap->am_lock));
1113 
1114 	/*
1115 	 * for MADV_SEQUENTIAL mappings we want to deactivate the back pages
1116 	 * now and then forget about them (for the rest of the fault).
1117 	 */
1118 
1119 	if (ufi->entry->advice == MADV_SEQUENTIAL && nback != 0) {
1120 
1121 		UVMHIST_LOG(maphist, "  MADV_SEQUENTIAL: flushing backpages",
1122 		    0,0,0,0);
1123 		/* flush back-page anons? */
1124 		if (amap)
1125 			uvmfault_anonflush(*ranons, nback);
1126 
1127 		/* flush object? */
1128 		if (uobj) {
1129 			voff_t uoff;
1130 
1131 			uoff = ufi->entry->offset + eoff;
1132 			mutex_enter(uobj->vmobjlock);
1133 			(void) (uobj->pgops->pgo_put)(uobj, uoff, uoff +
1134 				    (nback << PAGE_SHIFT), PGO_DEACTIVATE);
1135 		}
1136 
1137 		/* now forget about the backpages */
1138 		if (amap)
1139 			*ranons += nback;
1140 		flt->startva += (nback << PAGE_SHIFT);
1141 		flt->npages -= nback;
1142 		flt->centeridx = 0;
1143 	}
1144 	/*
1145 	 * => startva is fixed
1146 	 * => npages is fixed
1147 	 */
1148 	KASSERT(flt->startva <= ufi->orig_rvaddr);
1149 	KASSERT(ufi->orig_rvaddr + ufi->orig_size <=
1150 	    flt->startva + (flt->npages << PAGE_SHIFT));
1151 	return 0;
1152 }
1153 
1154 /*
1155  * uvm_fault_upper_lookup: look up existing h/w mapping and amap.
1156  *
1157  * iterate range of interest:
1158  *	1. check if h/w mapping exists.  if yes, we don't care
1159  *	2. check if anon exists.  if not, page is lower.
1160  *	3. if anon exists, enter h/w mapping for neighbors.
1161  *
1162  * => called with amap locked (if exists).
1163  */
1164 
1165 static int
1166 uvm_fault_upper_lookup(
1167 	struct uvm_faultinfo *ufi, const struct uvm_faultctx *flt,
1168 	struct vm_anon **anons, struct vm_page **pages)
1169 {
1170 	struct vm_amap *amap = ufi->entry->aref.ar_amap;
1171 	int lcv;
1172 	vaddr_t currva;
1173 	bool shadowed;
1174 	UVMHIST_FUNC("uvm_fault_upper_lookup"); UVMHIST_CALLED(maphist);
1175 
1176 	/* locked: maps(read), amap(if there) */
1177 	KASSERT(amap == NULL || mutex_owned(amap->am_lock));
1178 
1179 	/*
1180 	 * map in the backpages and frontpages we found in the amap in hopes
1181 	 * of preventing future faults.    we also init the pages[] array as
1182 	 * we go.
1183 	 */
1184 
1185 	currva = flt->startva;
1186 	shadowed = false;
1187 	for (lcv = 0; lcv < flt->npages; lcv++, currva += PAGE_SIZE) {
1188 		/*
1189 		 * don't play with VAs that are already mapped
1190 		 * (except for center)
1191 		 */
1192 		if (lcv != flt->centeridx &&
1193 		    pmap_extract(ufi->orig_map->pmap, currva, NULL)) {
1194 			pages[lcv] = PGO_DONTCARE;
1195 			continue;
1196 		}
1197 
1198 		/*
1199 		 * unmapped or center page.   check if any anon at this level.
1200 		 */
1201 		if (amap == NULL || anons[lcv] == NULL) {
1202 			pages[lcv] = NULL;
1203 			continue;
1204 		}
1205 
1206 		/*
1207 		 * check for present page and map if possible.   re-activate it.
1208 		 */
1209 
1210 		pages[lcv] = PGO_DONTCARE;
1211 		if (lcv == flt->centeridx) {	/* save center for later! */
1212 			shadowed = true;
1213 			continue;
1214 		}
1215 
1216 		struct vm_anon *anon = anons[lcv];
1217 		struct vm_page *pg = anon->an_page;
1218 
1219 		KASSERT(anon->an_lock == amap->am_lock);
1220 
1221 		/* Ignore loaned and busy pages. */
1222 		if (pg && pg->loan_count == 0 && (pg->flags & PG_BUSY) == 0) {
1223 			uvm_fault_upper_neighbor(ufi, flt, currva,
1224 			    pg, anon->an_ref > 1);
1225 		}
1226 	}
1227 
1228 	/* locked: maps(read), amap(if there) */
1229 	KASSERT(amap == NULL || mutex_owned(amap->am_lock));
1230 	/* (shadowed == true) if there is an anon at the faulting address */
1231 	UVMHIST_LOG(maphist, "  shadowed=%d, will_get=%d", shadowed,
1232 	    (ufi->entry->object.uvm_obj && shadowed != false),0,0);
1233 
1234 	/*
1235 	 * note that if we are really short of RAM we could sleep in the above
1236 	 * call to pmap_enter with everything locked.   bad?
1237 	 *
1238 	 * XXX Actually, that is bad; pmap_enter() should just fail in that
1239 	 * XXX case.  --thorpej
1240 	 */
1241 
1242 	return 0;
1243 }
1244 
1245 /*
1246  * uvm_fault_upper_neighbor: enter single lower neighbor page.
1247  *
1248  * => called with amap and anon locked.
1249  */
1250 
1251 static void
1252 uvm_fault_upper_neighbor(
1253 	struct uvm_faultinfo *ufi, const struct uvm_faultctx *flt,
1254 	vaddr_t currva, struct vm_page *pg, bool readonly)
1255 {
1256 	UVMHIST_FUNC("uvm_fault_upper_neighbor"); UVMHIST_CALLED(maphist);
1257 
1258 	/* locked: amap, anon */
1259 
1260 	mutex_enter(&uvm_pageqlock);
1261 	uvm_pageenqueue(pg);
1262 	mutex_exit(&uvm_pageqlock);
1263 	UVMHIST_LOG(maphist,
1264 	    "  MAPPING: n anon: pm=0x%x, va=0x%x, pg=0x%x",
1265 	    ufi->orig_map->pmap, currva, pg, 0);
1266 	uvmexp.fltnamap++;
1267 
1268 	/*
1269 	 * Since this page isn't the page that's actually faulting,
1270 	 * ignore pmap_enter() failures; it's not critical that we
1271 	 * enter these right now.
1272 	 */
1273 
1274 	(void) pmap_enter(ufi->orig_map->pmap, currva,
1275 	    VM_PAGE_TO_PHYS(pg),
1276 	    readonly ? (flt->enter_prot & ~VM_PROT_WRITE) :
1277 	    flt->enter_prot,
1278 	    PMAP_CANFAIL | (flt->wire_mapping ? PMAP_WIRED : 0));
1279 
1280 	pmap_update(ufi->orig_map->pmap);
1281 }
1282 
1283 /*
1284  * uvm_fault_upper: handle upper fault.
1285  *
1286  *	1. acquire anon lock.
1287  *	2. get anon.  let uvmfault_anonget do the dirty work.
1288  *	3. handle loan.
1289  *	4. dispatch direct or promote handlers.
1290  */
1291 
1292 static int
1293 uvm_fault_upper(
1294 	struct uvm_faultinfo *ufi, struct uvm_faultctx *flt,
1295 	struct vm_anon **anons)
1296 {
1297 	struct vm_amap * const amap = ufi->entry->aref.ar_amap;
1298 	struct vm_anon * const anon = anons[flt->centeridx];
1299 	struct uvm_object *uobj;
1300 	int error;
1301 	UVMHIST_FUNC("uvm_fault_upper"); UVMHIST_CALLED(maphist);
1302 
1303 	/* locked: maps(read), amap, anon */
1304 	KASSERT(mutex_owned(amap->am_lock));
1305 	KASSERT(anon->an_lock == amap->am_lock);
1306 
1307 	/*
1308 	 * handle case 1: fault on an anon in our amap
1309 	 */
1310 
1311 	UVMHIST_LOG(maphist, "  case 1 fault: anon=0x%x", anon, 0,0,0);
1312 
1313 	/*
1314 	 * no matter if we have case 1A or case 1B we are going to need to
1315 	 * have the anon's memory resident.   ensure that now.
1316 	 */
1317 
1318 	/*
1319 	 * let uvmfault_anonget do the dirty work.
1320 	 * if it fails (!OK) it will unlock everything for us.
1321 	 * if it succeeds, locks are still valid and locked.
1322 	 * also, if it is OK, then the anon's page is on the queues.
1323 	 * if the page is on loan from a uvm_object, then anonget will
1324 	 * lock that object for us if it does not fail.
1325 	 */
1326 
1327 	error = uvmfault_anonget(ufi, amap, anon);
1328 	switch (error) {
1329 	case 0:
1330 		break;
1331 
1332 	case ERESTART:
1333 		return ERESTART;
1334 
1335 	case EAGAIN:
1336 		kpause("fltagain1", false, hz/2, NULL);
1337 		return ERESTART;
1338 
1339 	default:
1340 		return error;
1341 	}
1342 
1343 	/*
1344 	 * uobj is non null if the page is on loan from an object (i.e. uobj)
1345 	 */
1346 
1347 	uobj = anon->an_page->uobject;	/* locked by anonget if !NULL */
1348 
1349 	/* locked: maps(read), amap, anon, uobj(if one) */
1350 	KASSERT(mutex_owned(amap->am_lock));
1351 	KASSERT(anon->an_lock == amap->am_lock);
1352 	KASSERT(uobj == NULL || mutex_owned(uobj->vmobjlock));
1353 
1354 	/*
1355 	 * special handling for loaned pages
1356 	 */
1357 
1358 	if (anon->an_page->loan_count) {
1359 		error = uvm_fault_upper_loan(ufi, flt, anon, &uobj);
1360 		if (error != 0)
1361 			return error;
1362 	}
1363 
1364 	/*
1365 	 * if we are case 1B then we will need to allocate a new blank
1366 	 * anon to transfer the data into.   note that we have a lock
1367 	 * on anon, so no one can busy or release the page until we are done.
1368 	 * also note that the ref count can't drop to zero here because
1369 	 * it is > 1 and we are only dropping one ref.
1370 	 *
1371 	 * in the (hopefully very rare) case that we are out of RAM we
1372 	 * will unlock, wait for more RAM, and refault.
1373 	 *
1374 	 * if we are out of anon VM we kill the process (XXX: could wait?).
1375 	 */
1376 
1377 	if (flt->cow_now && anon->an_ref > 1) {
1378 		flt->promote = true;
1379 		error = uvm_fault_upper_promote(ufi, flt, uobj, anon);
1380 	} else {
1381 		error = uvm_fault_upper_direct(ufi, flt, uobj, anon);
1382 	}
1383 	return error;
1384 }
1385 
1386 /*
1387  * uvm_fault_upper_loan: handle loaned upper page.
1388  *
1389  *	1. if not cow'ing now, simply adjust flt->enter_prot.
1390  *	2. if cow'ing now, and if ref count is 1, break loan.
1391  */
1392 
1393 static int
1394 uvm_fault_upper_loan(
1395 	struct uvm_faultinfo *ufi, struct uvm_faultctx *flt,
1396 	struct vm_anon *anon, struct uvm_object **ruobj)
1397 {
1398 	struct vm_amap * const amap = ufi->entry->aref.ar_amap;
1399 	int error = 0;
1400 	UVMHIST_FUNC("uvm_fault_upper_loan"); UVMHIST_CALLED(maphist);
1401 
1402 	if (!flt->cow_now) {
1403 
1404 		/*
1405 		 * for read faults on loaned pages we just cap the
1406 		 * protection at read-only.
1407 		 */
1408 
1409 		flt->enter_prot = flt->enter_prot & ~VM_PROT_WRITE;
1410 
1411 	} else {
1412 		/*
1413 		 * note that we can't allow writes into a loaned page!
1414 		 *
1415 		 * if we have a write fault on a loaned page in an
1416 		 * anon then we need to look at the anon's ref count.
1417 		 * if it is greater than one then we are going to do
1418 		 * a normal copy-on-write fault into a new anon (this
1419 		 * is not a problem).  however, if the reference count
1420 		 * is one (a case where we would normally allow a
1421 		 * write directly to the page) then we need to kill
1422 		 * the loan before we continue.
1423 		 */
1424 
1425 		/* >1 case is already ok */
1426 		if (anon->an_ref == 1) {
1427 			error = uvm_loanbreak_anon(anon, *ruobj);
1428 			if (error != 0) {
1429 				uvmfault_unlockall(ufi, amap, *ruobj);
1430 				uvm_wait("flt_noram2");
1431 				return ERESTART;
1432 			}
1433 			/* if we were a loan reciever uobj is gone */
1434 			if (*ruobj)
1435 				*ruobj = NULL;
1436 		}
1437 	}
1438 	return error;
1439 }
1440 
1441 /*
1442  * uvm_fault_upper_promote: promote upper page.
1443  *
1444  *	1. call uvmfault_promote.
1445  *	2. enqueue page.
1446  *	3. deref.
1447  *	4. pass page to uvm_fault_upper_enter.
1448  */
1449 
1450 static int
1451 uvm_fault_upper_promote(
1452 	struct uvm_faultinfo *ufi, struct uvm_faultctx *flt,
1453 	struct uvm_object *uobj, struct vm_anon *anon)
1454 {
1455 	struct vm_anon * const oanon = anon;
1456 	struct vm_page *pg;
1457 	int error;
1458 	UVMHIST_FUNC("uvm_fault_upper_promote"); UVMHIST_CALLED(maphist);
1459 
1460 	UVMHIST_LOG(maphist, "  case 1B: COW fault",0,0,0,0);
1461 	uvmexp.flt_acow++;
1462 
1463 	error = uvmfault_promote(ufi, oanon, PGO_DONTCARE, &anon,
1464 	    &flt->anon_spare);
1465 	switch (error) {
1466 	case 0:
1467 		break;
1468 	case ERESTART:
1469 		return ERESTART;
1470 	default:
1471 		return error;
1472 	}
1473 
1474 	KASSERT(anon == NULL || anon->an_lock == oanon->an_lock);
1475 
1476 	pg = anon->an_page;
1477 	mutex_enter(&uvm_pageqlock);
1478 	uvm_pageenqueue(pg); /* uvm_fault_upper_done will activate the page */
1479 	mutex_exit(&uvm_pageqlock);
1480 	pg->flags &= ~(PG_BUSY|PG_FAKE);
1481 	UVM_PAGE_OWN(pg, NULL);
1482 
1483 	/* deref: can not drop to zero here by defn! */
1484 	KASSERT(oanon->an_ref > 1);
1485 	oanon->an_ref--;
1486 
1487 	/*
1488 	 * note: oanon is still locked, as is the new anon.  we
1489 	 * need to check for this later when we unlock oanon; if
1490 	 * oanon != anon, we'll have to unlock anon, too.
1491 	 */
1492 
1493 	return uvm_fault_upper_enter(ufi, flt, uobj, anon, pg, oanon);
1494 }
1495 
1496 /*
1497  * uvm_fault_upper_direct: handle direct fault.
1498  */
1499 
1500 static int
1501 uvm_fault_upper_direct(
1502 	struct uvm_faultinfo *ufi, struct uvm_faultctx *flt,
1503 	struct uvm_object *uobj, struct vm_anon *anon)
1504 {
1505 	struct vm_anon * const oanon = anon;
1506 	struct vm_page *pg;
1507 	UVMHIST_FUNC("uvm_fault_upper_direct"); UVMHIST_CALLED(maphist);
1508 
1509 	uvmexp.flt_anon++;
1510 	pg = anon->an_page;
1511 	if (anon->an_ref > 1)     /* disallow writes to ref > 1 anons */
1512 		flt->enter_prot = flt->enter_prot & ~VM_PROT_WRITE;
1513 
1514 	return uvm_fault_upper_enter(ufi, flt, uobj, anon, pg, oanon);
1515 }
1516 
1517 /*
1518  * uvm_fault_upper_enter: enter h/w mapping of upper page.
1519  */
1520 
1521 static int
1522 uvm_fault_upper_enter(
1523 	struct uvm_faultinfo *ufi, const struct uvm_faultctx *flt,
1524 	struct uvm_object *uobj, struct vm_anon *anon, struct vm_page *pg,
1525 	struct vm_anon *oanon)
1526 {
1527 	struct vm_amap * const amap = ufi->entry->aref.ar_amap;
1528 	UVMHIST_FUNC("uvm_fault_upper_enter"); UVMHIST_CALLED(maphist);
1529 
1530 	/* locked: maps(read), amap, oanon, anon(if different from oanon) */
1531 	KASSERT(mutex_owned(amap->am_lock));
1532 	KASSERT(anon->an_lock == amap->am_lock);
1533 	KASSERT(oanon->an_lock == amap->am_lock);
1534 	KASSERT(uobj == NULL || mutex_owned(uobj->vmobjlock));
1535 
1536 	/*
1537 	 * now map the page in.
1538 	 */
1539 
1540 	UVMHIST_LOG(maphist,
1541 	    "  MAPPING: anon: pm=0x%x, va=0x%x, pg=0x%x, promote=%d",
1542 	    ufi->orig_map->pmap, ufi->orig_rvaddr, pg, flt->promote);
1543 	if (pmap_enter(ufi->orig_map->pmap, ufi->orig_rvaddr,
1544 	    VM_PAGE_TO_PHYS(pg),
1545 	    flt->enter_prot, flt->access_type | PMAP_CANFAIL |
1546 	    (flt->wire_mapping ? PMAP_WIRED : 0)) != 0) {
1547 
1548 		/*
1549 		 * No need to undo what we did; we can simply think of
1550 		 * this as the pmap throwing away the mapping information.
1551 		 *
1552 		 * We do, however, have to go through the ReFault path,
1553 		 * as the map may change while we're asleep.
1554 		 */
1555 
1556 		uvmfault_unlockall(ufi, amap, uobj);
1557 		if (!uvm_reclaimable()) {
1558 			UVMHIST_LOG(maphist,
1559 			    "<- failed.  out of VM",0,0,0,0);
1560 			/* XXX instrumentation */
1561 			return ENOMEM;
1562 		}
1563 		/* XXX instrumentation */
1564 		uvm_wait("flt_pmfail1");
1565 		return ERESTART;
1566 	}
1567 
1568 	uvm_fault_upper_done(ufi, flt, anon, pg);
1569 
1570 	/*
1571 	 * done case 1!  finish up by unlocking everything and returning success
1572 	 */
1573 
1574 	pmap_update(ufi->orig_map->pmap);
1575 	uvmfault_unlockall(ufi, amap, uobj);
1576 	return 0;
1577 }
1578 
1579 /*
1580  * uvm_fault_upper_done: queue upper center page.
1581  */
1582 
1583 static void
1584 uvm_fault_upper_done(
1585 	struct uvm_faultinfo *ufi, const struct uvm_faultctx *flt,
1586 	struct vm_anon *anon, struct vm_page *pg)
1587 {
1588 	const bool wire_paging = flt->wire_paging;
1589 
1590 	UVMHIST_FUNC("uvm_fault_upper_done"); UVMHIST_CALLED(maphist);
1591 
1592 	/*
1593 	 * ... update the page queues.
1594 	 */
1595 
1596 	mutex_enter(&uvm_pageqlock);
1597 	if (wire_paging) {
1598 		uvm_pagewire(pg);
1599 
1600 		/*
1601 		 * since the now-wired page cannot be paged out,
1602 		 * release its swap resources for others to use.
1603 		 * since an anon with no swap cannot be PG_CLEAN,
1604 		 * clear its clean flag now.
1605 		 */
1606 
1607 		pg->flags &= ~(PG_CLEAN);
1608 
1609 	} else {
1610 		uvm_pageactivate(pg);
1611 	}
1612 	mutex_exit(&uvm_pageqlock);
1613 
1614 	if (wire_paging) {
1615 		uvm_anon_dropswap(anon);
1616 	}
1617 }
1618 
1619 /*
1620  * uvm_fault_lower: handle lower fault.
1621  *
1622  *	1. check uobj
1623  *	1.1. if null, ZFOD.
1624  *	1.2. if not null, look up unnmapped neighbor pages.
1625  *	2. for center page, check if promote.
1626  *	2.1. ZFOD always needs promotion.
1627  *	2.2. other uobjs, when entry is marked COW (usually MAP_PRIVATE vnode).
1628  *	3. if uobj is not ZFOD and page is not found, do i/o.
1629  *	4. dispatch either direct / promote fault.
1630  */
1631 
1632 static int
1633 uvm_fault_lower(
1634 	struct uvm_faultinfo *ufi, struct uvm_faultctx *flt,
1635 	struct vm_page **pages)
1636 {
1637 #ifdef DIAGNOSTIC
1638 	struct vm_amap *amap = ufi->entry->aref.ar_amap;
1639 #endif
1640 	struct uvm_object *uobj = ufi->entry->object.uvm_obj;
1641 	struct vm_page *uobjpage;
1642 	int error;
1643 	UVMHIST_FUNC("uvm_fault_lower"); UVMHIST_CALLED(maphist);
1644 
1645 	/*
1646 	 * now, if the desired page is not shadowed by the amap and we have
1647 	 * a backing object that does not have a special fault routine, then
1648 	 * we ask (with pgo_get) the object for resident pages that we care
1649 	 * about and attempt to map them in.  we do not let pgo_get block
1650 	 * (PGO_LOCKED).
1651 	 */
1652 
1653 	if (uobj == NULL) {
1654 		/* zero fill; don't care neighbor pages */
1655 		uobjpage = NULL;
1656 	} else {
1657 		uvm_fault_lower_lookup(ufi, flt, pages);
1658 		uobjpage = pages[flt->centeridx];
1659 	}
1660 
1661 	/*
1662 	 * note that at this point we are done with any front or back pages.
1663 	 * we are now going to focus on the center page (i.e. the one we've
1664 	 * faulted on).  if we have faulted on the upper (anon) layer
1665 	 * [i.e. case 1], then the anon we want is anons[centeridx] (we have
1666 	 * not touched it yet).  if we have faulted on the bottom (uobj)
1667 	 * layer [i.e. case 2] and the page was both present and available,
1668 	 * then we've got a pointer to it as "uobjpage" and we've already
1669 	 * made it BUSY.
1670 	 */
1671 
1672 	/*
1673 	 * locked:
1674 	 * maps(read), amap(if there), uobj(if !null), uobjpage(if !null)
1675 	 */
1676 	KASSERT(amap == NULL || mutex_owned(amap->am_lock));
1677 	KASSERT(uobj == NULL || mutex_owned(uobj->vmobjlock));
1678 	KASSERT(uobjpage == NULL || (uobjpage->flags & PG_BUSY) != 0);
1679 
1680 	/*
1681 	 * note that uobjpage can not be PGO_DONTCARE at this point.  we now
1682 	 * set uobjpage to PGO_DONTCARE if we are doing a zero fill.  if we
1683 	 * have a backing object, check and see if we are going to promote
1684 	 * the data up to an anon during the fault.
1685 	 */
1686 
1687 	if (uobj == NULL) {
1688 		uobjpage = PGO_DONTCARE;
1689 		flt->promote = true;		/* always need anon here */
1690 	} else {
1691 		KASSERT(uobjpage != PGO_DONTCARE);
1692 		flt->promote = flt->cow_now && UVM_ET_ISCOPYONWRITE(ufi->entry);
1693 	}
1694 	UVMHIST_LOG(maphist, "  case 2 fault: promote=%d, zfill=%d",
1695 	    flt->promote, (uobj == NULL), 0,0);
1696 
1697 	/*
1698 	 * if uobjpage is not null then we do not need to do I/O to get the
1699 	 * uobjpage.
1700 	 *
1701 	 * if uobjpage is null, then we need to unlock and ask the pager to
1702 	 * get the data for us.   once we have the data, we need to reverify
1703 	 * the state the world.   we are currently not holding any resources.
1704 	 */
1705 
1706 	if (uobjpage) {
1707 		/* update rusage counters */
1708 		curlwp->l_ru.ru_minflt++;
1709 	} else {
1710 		error = uvm_fault_lower_io(ufi, flt, &uobj, &uobjpage);
1711 		if (error != 0)
1712 			return error;
1713 	}
1714 
1715 	/*
1716 	 * locked:
1717 	 * maps(read), amap(if !null), uobj(if !null), uobjpage(if uobj)
1718 	 */
1719 	KASSERT(amap == NULL || mutex_owned(amap->am_lock));
1720 	KASSERT(uobj == NULL || mutex_owned(uobj->vmobjlock));
1721 	KASSERT(uobj == NULL || (uobjpage->flags & PG_BUSY) != 0);
1722 
1723 	/*
1724 	 * notes:
1725 	 *  - at this point uobjpage can not be NULL
1726 	 *  - at this point uobjpage can not be PG_RELEASED (since we checked
1727 	 *  for it above)
1728 	 *  - at this point uobjpage could be PG_WANTED (handle later)
1729 	 */
1730 
1731 	KASSERT(uobjpage != NULL);
1732 	KASSERT(uobj == NULL || uobj == uobjpage->uobject);
1733 	KASSERT(uobj == NULL || !UVM_OBJ_IS_CLEAN(uobjpage->uobject) ||
1734 	    (uobjpage->flags & PG_CLEAN) != 0);
1735 
1736 	if (!flt->promote) {
1737 		error = uvm_fault_lower_direct(ufi, flt, uobj, uobjpage);
1738 	} else {
1739 		error = uvm_fault_lower_promote(ufi, flt, uobj, uobjpage);
1740 	}
1741 	return error;
1742 }
1743 
1744 /*
1745  * uvm_fault_lower_lookup: look up on-memory uobj pages.
1746  *
1747  *	1. get on-memory pages.
1748  *	2. if failed, give up (get only center page later).
1749  *	3. if succeeded, enter h/w mapping of neighbor pages.
1750  */
1751 
1752 static void
1753 uvm_fault_lower_lookup(
1754 	struct uvm_faultinfo *ufi, const struct uvm_faultctx *flt,
1755 	struct vm_page **pages)
1756 {
1757 	struct uvm_object *uobj = ufi->entry->object.uvm_obj;
1758 	int lcv, gotpages;
1759 	vaddr_t currva;
1760 	UVMHIST_FUNC("uvm_fault_lower_lookup"); UVMHIST_CALLED(maphist);
1761 
1762 	mutex_enter(uobj->vmobjlock);
1763 	/* Locked: maps(read), amap(if there), uobj */
1764 
1765 	uvmexp.fltlget++;
1766 	gotpages = flt->npages;
1767 	(void) uobj->pgops->pgo_get(uobj,
1768 	    ufi->entry->offset + flt->startva - ufi->entry->start,
1769 	    pages, &gotpages, flt->centeridx,
1770 	    flt->access_type & MASK(ufi->entry), ufi->entry->advice, PGO_LOCKED);
1771 
1772 	KASSERT(mutex_owned(uobj->vmobjlock));
1773 
1774 	/*
1775 	 * check for pages to map, if we got any
1776 	 */
1777 
1778 	if (gotpages == 0) {
1779 		pages[flt->centeridx] = NULL;
1780 		return;
1781 	}
1782 
1783 	currva = flt->startva;
1784 	for (lcv = 0; lcv < flt->npages; lcv++, currva += PAGE_SIZE) {
1785 		struct vm_page *curpg;
1786 
1787 		curpg = pages[lcv];
1788 		if (curpg == NULL || curpg == PGO_DONTCARE) {
1789 			continue;
1790 		}
1791 		KASSERT(curpg->uobject == uobj);
1792 
1793 		/*
1794 		 * if center page is resident and not PG_BUSY|PG_RELEASED
1795 		 * then pgo_get made it PG_BUSY for us and gave us a handle
1796 		 * to it.
1797 		 */
1798 
1799 		if (lcv == flt->centeridx) {
1800 			UVMHIST_LOG(maphist, "  got uobjpage "
1801 			    "(0x%x) with locked get",
1802 			    curpg, 0,0,0);
1803 		} else {
1804 			bool readonly = (curpg->flags & PG_RDONLY)
1805 			    || (curpg->loan_count > 0)
1806 			    || UVM_OBJ_NEEDS_WRITEFAULT(curpg->uobject);
1807 
1808 			uvm_fault_lower_neighbor(ufi, flt,
1809 			    currva, curpg, readonly);
1810 		}
1811 	}
1812 	pmap_update(ufi->orig_map->pmap);
1813 }
1814 
1815 /*
1816  * uvm_fault_lower_neighbor: enter h/w mapping of lower neighbor page.
1817  */
1818 
1819 static void
1820 uvm_fault_lower_neighbor(
1821 	struct uvm_faultinfo *ufi, const struct uvm_faultctx *flt,
1822 	vaddr_t currva, struct vm_page *pg, bool readonly)
1823 {
1824 	UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist);
1825 
1826 	/* locked: maps(read), amap(if there), uobj */
1827 
1828 	/*
1829 	 * calling pgo_get with PGO_LOCKED returns us pages which
1830 	 * are neither busy nor released, so we don't need to check
1831 	 * for this.  we can just directly enter the pages.
1832 	 */
1833 
1834 	mutex_enter(&uvm_pageqlock);
1835 	uvm_pageenqueue(pg);
1836 	mutex_exit(&uvm_pageqlock);
1837 	UVMHIST_LOG(maphist,
1838 	    "  MAPPING: n obj: pm=0x%x, va=0x%x, pg=0x%x",
1839 	    ufi->orig_map->pmap, currva, pg, 0);
1840 	uvmexp.fltnomap++;
1841 
1842 	/*
1843 	 * Since this page isn't the page that's actually faulting,
1844 	 * ignore pmap_enter() failures; it's not critical that we
1845 	 * enter these right now.
1846 	 * NOTE: page can't be PG_WANTED or PG_RELEASED because we've
1847 	 * held the lock the whole time we've had the handle.
1848 	 */
1849 	KASSERT((pg->flags & PG_PAGEOUT) == 0);
1850 	KASSERT((pg->flags & PG_RELEASED) == 0);
1851 	KASSERT((pg->flags & PG_WANTED) == 0);
1852 	KASSERT(!UVM_OBJ_IS_CLEAN(pg->uobject) || (pg->flags & PG_CLEAN) != 0);
1853 	pg->flags &= ~(PG_BUSY);
1854 	UVM_PAGE_OWN(pg, NULL);
1855 
1856 	KASSERT(mutex_owned(pg->uobject->vmobjlock));
1857 	(void) pmap_enter(ufi->orig_map->pmap, currva,
1858 	    VM_PAGE_TO_PHYS(pg),
1859 	    readonly ? (flt->enter_prot & ~VM_PROT_WRITE) :
1860 	    flt->enter_prot & MASK(ufi->entry),
1861 	    PMAP_CANFAIL | (flt->wire_mapping ? PMAP_WIRED : 0));
1862 }
1863 
1864 /*
1865  * uvm_fault_lower_io: get lower page from backing store.
1866  *
1867  *	1. unlock everything, because i/o will block.
1868  *	2. call pgo_get.
1869  *	3. if failed, recover.
1870  *	4. if succeeded, relock everything and verify things.
1871  */
1872 
1873 static int
1874 uvm_fault_lower_io(
1875 	struct uvm_faultinfo *ufi, const struct uvm_faultctx *flt,
1876 	struct uvm_object **ruobj, struct vm_page **ruobjpage)
1877 {
1878 	struct vm_amap * const amap = ufi->entry->aref.ar_amap;
1879 	struct uvm_object *uobj = *ruobj;
1880 	struct vm_page *pg;
1881 	bool locked;
1882 	int gotpages;
1883 	int error;
1884 	voff_t uoff;
1885 	UVMHIST_FUNC("uvm_fault_lower_io"); UVMHIST_CALLED(maphist);
1886 
1887 	/* update rusage counters */
1888 	curlwp->l_ru.ru_majflt++;
1889 
1890 	/* Locked: maps(read), amap(if there), uobj */
1891 	uvmfault_unlockall(ufi, amap, NULL);
1892 
1893 	/* Locked: uobj */
1894 	KASSERT(uobj == NULL || mutex_owned(uobj->vmobjlock));
1895 
1896 	uvmexp.fltget++;
1897 	gotpages = 1;
1898 	pg = NULL;
1899 	uoff = (ufi->orig_rvaddr - ufi->entry->start) + ufi->entry->offset;
1900 	error = uobj->pgops->pgo_get(uobj, uoff, &pg, &gotpages,
1901 	    0, flt->access_type & MASK(ufi->entry), ufi->entry->advice,
1902 	    PGO_SYNCIO);
1903 	/* locked: pg(if no error) */
1904 
1905 	/*
1906 	 * recover from I/O
1907 	 */
1908 
1909 	if (error) {
1910 		if (error == EAGAIN) {
1911 			UVMHIST_LOG(maphist,
1912 			    "  pgo_get says TRY AGAIN!",0,0,0,0);
1913 			kpause("fltagain2", false, hz/2, NULL);
1914 			return ERESTART;
1915 		}
1916 
1917 #if 0
1918 		KASSERT(error != ERESTART);
1919 #else
1920 		/* XXXUEBS don't re-fault? */
1921 		if (error == ERESTART)
1922 			error = EIO;
1923 #endif
1924 
1925 		UVMHIST_LOG(maphist, "<- pgo_get failed (code %d)",
1926 		    error, 0,0,0);
1927 		return error;
1928 	}
1929 
1930 	/*
1931 	 * re-verify the state of the world by first trying to relock
1932 	 * the maps.  always relock the object.
1933 	 */
1934 
1935 	locked = uvmfault_relock(ufi);
1936 	if (locked && amap)
1937 		amap_lock(amap);
1938 
1939 	/* might be changed */
1940 	uobj = pg->uobject;
1941 
1942 	mutex_enter(uobj->vmobjlock);
1943 	KASSERT((pg->flags & PG_BUSY) != 0);
1944 
1945 	mutex_enter(&uvm_pageqlock);
1946 	uvm_pageactivate(pg);
1947 	mutex_exit(&uvm_pageqlock);
1948 
1949 	/* locked(locked): maps(read), amap(if !null), uobj, pg */
1950 	/* locked(!locked): uobj, pg */
1951 
1952 	/*
1953 	 * verify that the page has not be released and re-verify
1954 	 * that amap slot is still free.   if there is a problem,
1955 	 * we unlock and clean up.
1956 	 */
1957 
1958 	if ((pg->flags & PG_RELEASED) != 0 ||
1959 	    (locked && amap && amap_lookup(&ufi->entry->aref,
1960 	      ufi->orig_rvaddr - ufi->entry->start))) {
1961 		if (locked)
1962 			uvmfault_unlockall(ufi, amap, NULL);
1963 		locked = false;
1964 	}
1965 
1966 	/*
1967 	 * didn't get the lock?   release the page and retry.
1968 	 */
1969 
1970 	if (locked == false) {
1971 		UVMHIST_LOG(maphist,
1972 		    "  wasn't able to relock after fault: retry",
1973 		    0,0,0,0);
1974 		if (pg->flags & PG_WANTED) {
1975 			wakeup(pg);
1976 		}
1977 		if ((pg->flags & PG_RELEASED) == 0) {
1978 			pg->flags &= ~(PG_BUSY | PG_WANTED);
1979 			UVM_PAGE_OWN(pg, NULL);
1980 		} else {
1981 			uvmexp.fltpgrele++;
1982 			uvm_pagefree(pg);
1983 		}
1984 		mutex_exit(uobj->vmobjlock);
1985 		return ERESTART;
1986 	}
1987 
1988 	/*
1989 	 * we have the data in pg which is busy and
1990 	 * not released.  we are holding object lock (so the page
1991 	 * can't be released on us).
1992 	 */
1993 
1994 	/* locked: maps(read), amap(if !null), uobj, pg */
1995 
1996 	*ruobj = uobj;
1997 	*ruobjpage = pg;
1998 	return 0;
1999 }
2000 
2001 /*
2002  * uvm_fault_lower_direct: fault lower center page
2003  *
2004  *	1. adjust flt->enter_prot.
2005  *	2. if page is loaned, resolve.
2006  */
2007 
2008 int
2009 uvm_fault_lower_direct(
2010 	struct uvm_faultinfo *ufi, struct uvm_faultctx *flt,
2011 	struct uvm_object *uobj, struct vm_page *uobjpage)
2012 {
2013 	struct vm_page *pg;
2014 	UVMHIST_FUNC("uvm_fault_lower_direct"); UVMHIST_CALLED(maphist);
2015 
2016 	/*
2017 	 * we are not promoting.   if the mapping is COW ensure that we
2018 	 * don't give more access than we should (e.g. when doing a read
2019 	 * fault on a COPYONWRITE mapping we want to map the COW page in
2020 	 * R/O even though the entry protection could be R/W).
2021 	 *
2022 	 * set "pg" to the page we want to map in (uobjpage, usually)
2023 	 */
2024 
2025 	uvmexp.flt_obj++;
2026 	if (UVM_ET_ISCOPYONWRITE(ufi->entry) ||
2027 	    UVM_OBJ_NEEDS_WRITEFAULT(uobjpage->uobject))
2028 		flt->enter_prot &= ~VM_PROT_WRITE;
2029 	pg = uobjpage;		/* map in the actual object */
2030 
2031 	KASSERT(uobjpage != PGO_DONTCARE);
2032 
2033 	/*
2034 	 * we are faulting directly on the page.   be careful
2035 	 * about writing to loaned pages...
2036 	 */
2037 
2038 	if (uobjpage->loan_count) {
2039 		uvm_fault_lower_direct_loan(ufi, flt, uobj, &pg, &uobjpage);
2040 	}
2041 	KASSERT(pg == uobjpage);
2042 
2043 	KASSERT(uobj == NULL || (uobjpage->flags & PG_BUSY) != 0);
2044 	return uvm_fault_lower_enter(ufi, flt, uobj, NULL, pg);
2045 }
2046 
2047 /*
2048  * uvm_fault_lower_direct_loan: resolve loaned page.
2049  *
2050  *	1. if not cow'ing, adjust flt->enter_prot.
2051  *	2. if cow'ing, break loan.
2052  */
2053 
2054 static int
2055 uvm_fault_lower_direct_loan(
2056 	struct uvm_faultinfo *ufi, struct uvm_faultctx *flt,
2057 	struct uvm_object *uobj, struct vm_page **rpg,
2058 	struct vm_page **ruobjpage)
2059 {
2060 	struct vm_amap * const amap = ufi->entry->aref.ar_amap;
2061 	struct vm_page *pg;
2062 	struct vm_page *uobjpage = *ruobjpage;
2063 	UVMHIST_FUNC("uvm_fault_lower_direct_loan"); UVMHIST_CALLED(maphist);
2064 
2065 	if (!flt->cow_now) {
2066 		/* read fault: cap the protection at readonly */
2067 		/* cap! */
2068 		flt->enter_prot = flt->enter_prot & ~VM_PROT_WRITE;
2069 	} else {
2070 		/* write fault: must break the loan here */
2071 
2072 		pg = uvm_loanbreak(uobjpage);
2073 		if (pg == NULL) {
2074 
2075 			/*
2076 			 * drop ownership of page, it can't be released
2077 			 */
2078 
2079 			if (uobjpage->flags & PG_WANTED)
2080 				wakeup(uobjpage);
2081 			uobjpage->flags &= ~(PG_BUSY|PG_WANTED);
2082 			UVM_PAGE_OWN(uobjpage, NULL);
2083 
2084 			uvmfault_unlockall(ufi, amap, uobj);
2085 			UVMHIST_LOG(maphist,
2086 			  "  out of RAM breaking loan, waiting",
2087 			  0,0,0,0);
2088 			uvmexp.fltnoram++;
2089 			uvm_wait("flt_noram4");
2090 			return ERESTART;
2091 		}
2092 		*rpg = pg;
2093 		*ruobjpage = pg;
2094 	}
2095 	return 0;
2096 }
2097 
2098 /*
2099  * uvm_fault_lower_promote: promote lower page.
2100  *
2101  *	1. call uvmfault_promote.
2102  *	2. fill in data.
2103  *	3. if not ZFOD, dispose old page.
2104  */
2105 
2106 int
2107 uvm_fault_lower_promote(
2108 	struct uvm_faultinfo *ufi, struct uvm_faultctx *flt,
2109 	struct uvm_object *uobj, struct vm_page *uobjpage)
2110 {
2111 	struct vm_amap * const amap = ufi->entry->aref.ar_amap;
2112 	struct vm_anon *anon;
2113 	struct vm_page *pg;
2114 	int error;
2115 	UVMHIST_FUNC("uvm_fault_lower_promote"); UVMHIST_CALLED(maphist);
2116 
2117 	KASSERT(amap != NULL);
2118 
2119 	/*
2120 	 * If we are going to promote the data to an anon we
2121 	 * allocate a blank anon here and plug it into our amap.
2122 	 */
2123 	error = uvmfault_promote(ufi, NULL, uobjpage,
2124 	    &anon, &flt->anon_spare);
2125 	switch (error) {
2126 	case 0:
2127 		break;
2128 	case ERESTART:
2129 		return ERESTART;
2130 	default:
2131 		return error;
2132 	}
2133 
2134 	pg = anon->an_page;
2135 
2136 	/*
2137 	 * Fill in the data.
2138 	 */
2139 	KASSERT(uobj == NULL || (uobjpage->flags & PG_BUSY) != 0);
2140 
2141 	if (uobjpage != PGO_DONTCARE) {
2142 		uvmexp.flt_prcopy++;
2143 
2144 		/*
2145 		 * promote to shared amap?  make sure all sharing
2146 		 * procs see it
2147 		 */
2148 
2149 		if ((amap_flags(amap) & AMAP_SHARED) != 0) {
2150 			pmap_page_protect(uobjpage, VM_PROT_NONE);
2151 			/*
2152 			 * XXX: PAGE MIGHT BE WIRED!
2153 			 */
2154 		}
2155 
2156 		/*
2157 		 * dispose of uobjpage.  it can't be PG_RELEASED
2158 		 * since we still hold the object lock.
2159 		 */
2160 
2161 		if (uobjpage->flags & PG_WANTED) {
2162 			/* still have the obj lock */
2163 			wakeup(uobjpage);
2164 		}
2165 		uobjpage->flags &= ~(PG_BUSY|PG_WANTED);
2166 		UVM_PAGE_OWN(uobjpage, NULL);
2167 
2168 		UVMHIST_LOG(maphist,
2169 		    "  promote uobjpage 0x%x to anon/page 0x%x/0x%x",
2170 		    uobjpage, anon, pg, 0);
2171 
2172 	} else {
2173 		uvmexp.flt_przero++;
2174 
2175 		/*
2176 		 * Page is zero'd and marked dirty by
2177 		 * uvmfault_promote().
2178 		 */
2179 
2180 		UVMHIST_LOG(maphist,"  zero fill anon/page 0x%x/0%x",
2181 		    anon, pg, 0, 0);
2182 	}
2183 
2184 	return uvm_fault_lower_enter(ufi, flt, uobj, anon, pg);
2185 }
2186 
2187 /*
2188  * uvm_fault_lower_enter: enter h/w mapping of lower page or anon page promoted
2189  * from the lower page.
2190  */
2191 
2192 int
2193 uvm_fault_lower_enter(
2194 	struct uvm_faultinfo *ufi, const struct uvm_faultctx *flt,
2195 	struct uvm_object *uobj,
2196 	struct vm_anon *anon, struct vm_page *pg)
2197 {
2198 	struct vm_amap * const amap = ufi->entry->aref.ar_amap;
2199 	int error;
2200 	UVMHIST_FUNC("uvm_fault_lower_enter"); UVMHIST_CALLED(maphist);
2201 
2202 	/*
2203 	 * Locked:
2204 	 *
2205 	 *	maps(read), amap(if !null), uobj(if !null),
2206 	 *	anon(if !null), pg(if anon), unlock_uobj(if !null)
2207 	 *
2208 	 * Note: pg is either the uobjpage or the new page in the new anon.
2209 	 */
2210 	KASSERT(amap == NULL || mutex_owned(amap->am_lock));
2211 	KASSERT(uobj == NULL || mutex_owned(uobj->vmobjlock));
2212 	KASSERT(anon == NULL || anon->an_lock == amap->am_lock);
2213 	KASSERT((pg->flags & PG_BUSY) != 0);
2214 
2215 	/*
2216 	 * all resources are present.   we can now map it in and free our
2217 	 * resources.
2218 	 */
2219 
2220 	UVMHIST_LOG(maphist,
2221 	    "  MAPPING: case2: pm=0x%x, va=0x%x, pg=0x%x, promote=%d",
2222 	    ufi->orig_map->pmap, ufi->orig_rvaddr, pg, flt->promote);
2223 	KASSERT((flt->access_type & VM_PROT_WRITE) == 0 ||
2224 		(pg->flags & PG_RDONLY) == 0);
2225 	if (pmap_enter(ufi->orig_map->pmap, ufi->orig_rvaddr,
2226 	    VM_PAGE_TO_PHYS(pg),
2227 	    (pg->flags & PG_RDONLY) != 0 ?
2228 	    flt->enter_prot & ~VM_PROT_WRITE : flt->enter_prot,
2229 	    flt->access_type | PMAP_CANFAIL |
2230 	    (flt->wire_mapping ? PMAP_WIRED : 0)) != 0) {
2231 
2232 		/*
2233 		 * No need to undo what we did; we can simply think of
2234 		 * this as the pmap throwing away the mapping information.
2235 		 *
2236 		 * We do, however, have to go through the ReFault path,
2237 		 * as the map may change while we're asleep.
2238 		 */
2239 
2240 		/*
2241 		 * ensure that the page is queued in the case that
2242 		 * we just promoted the page.
2243 		 */
2244 
2245 		mutex_enter(&uvm_pageqlock);
2246 		uvm_pageenqueue(pg);
2247 		mutex_exit(&uvm_pageqlock);
2248 
2249 		if (pg->flags & PG_WANTED)
2250 			wakeup(pg);
2251 
2252 		/*
2253 		 * note that pg can't be PG_RELEASED since we did not drop
2254 		 * the object lock since the last time we checked.
2255 		 */
2256 		KASSERT((pg->flags & PG_RELEASED) == 0);
2257 
2258 		pg->flags &= ~(PG_BUSY|PG_FAKE|PG_WANTED);
2259 		UVM_PAGE_OWN(pg, NULL);
2260 
2261 		uvmfault_unlockall(ufi, amap, uobj);
2262 		if (!uvm_reclaimable()) {
2263 			UVMHIST_LOG(maphist,
2264 			    "<- failed.  out of VM",0,0,0,0);
2265 			/* XXX instrumentation */
2266 			error = ENOMEM;
2267 			return error;
2268 		}
2269 		/* XXX instrumentation */
2270 		uvm_wait("flt_pmfail2");
2271 		return ERESTART;
2272 	}
2273 
2274 	uvm_fault_lower_done(ufi, flt, uobj, pg);
2275 
2276 	/*
2277 	 * note that pg can't be PG_RELEASED since we did not drop the object
2278 	 * lock since the last time we checked.
2279 	 */
2280 	KASSERT((pg->flags & PG_RELEASED) == 0);
2281 	if (pg->flags & PG_WANTED)
2282 		wakeup(pg);
2283 	pg->flags &= ~(PG_BUSY|PG_FAKE|PG_WANTED);
2284 	UVM_PAGE_OWN(pg, NULL);
2285 
2286 	pmap_update(ufi->orig_map->pmap);
2287 	uvmfault_unlockall(ufi, amap, uobj);
2288 
2289 	UVMHIST_LOG(maphist, "<- done (SUCCESS!)",0,0,0,0);
2290 	return 0;
2291 }
2292 
2293 /*
2294  * uvm_fault_lower_done: queue lower center page.
2295  */
2296 
2297 void
2298 uvm_fault_lower_done(
2299 	struct uvm_faultinfo *ufi, const struct uvm_faultctx *flt,
2300 	struct uvm_object *uobj, struct vm_page *pg)
2301 {
2302 	bool dropswap = false;
2303 
2304 	UVMHIST_FUNC("uvm_fault_lower_done"); UVMHIST_CALLED(maphist);
2305 
2306 	mutex_enter(&uvm_pageqlock);
2307 	if (flt->wire_paging) {
2308 		uvm_pagewire(pg);
2309 		if (pg->pqflags & PQ_AOBJ) {
2310 
2311 			/*
2312 			 * since the now-wired page cannot be paged out,
2313 			 * release its swap resources for others to use.
2314 			 * since an aobj page with no swap cannot be PG_CLEAN,
2315 			 * clear its clean flag now.
2316 			 */
2317 
2318 			KASSERT(uobj != NULL);
2319 			pg->flags &= ~(PG_CLEAN);
2320 			dropswap = true;
2321 		}
2322 	} else {
2323 		uvm_pageactivate(pg);
2324 	}
2325 	mutex_exit(&uvm_pageqlock);
2326 
2327 	if (dropswap) {
2328 		uao_dropswap(uobj, pg->offset >> PAGE_SHIFT);
2329 	}
2330 }
2331 
2332 
2333 /*
2334  * uvm_fault_wire: wire down a range of virtual addresses in a map.
2335  *
2336  * => map may be read-locked by caller, but MUST NOT be write-locked.
2337  * => if map is read-locked, any operations which may cause map to
2338  *	be write-locked in uvm_fault() must be taken care of by
2339  *	the caller.  See uvm_map_pageable().
2340  */
2341 
2342 int
2343 uvm_fault_wire(struct vm_map *map, vaddr_t start, vaddr_t end,
2344     vm_prot_t access_type, int maxprot)
2345 {
2346 	vaddr_t va;
2347 	int error;
2348 
2349 	/*
2350 	 * now fault it in a page at a time.   if the fault fails then we have
2351 	 * to undo what we have done.   note that in uvm_fault VM_PROT_NONE
2352 	 * is replaced with the max protection if fault_type is VM_FAULT_WIRE.
2353 	 */
2354 
2355 	/*
2356 	 * XXX work around overflowing a vaddr_t.  this prevents us from
2357 	 * wiring the last page in the address space, though.
2358 	 */
2359 	if (start > end) {
2360 		return EFAULT;
2361 	}
2362 
2363 	for (va = start; va < end; va += PAGE_SIZE) {
2364 		error = uvm_fault_internal(map, va, access_type,
2365 		    (maxprot ? UVM_FAULT_MAXPROT : 0) | UVM_FAULT_WIRE);
2366 		if (error) {
2367 			if (va != start) {
2368 				uvm_fault_unwire(map, start, va);
2369 			}
2370 			return error;
2371 		}
2372 	}
2373 	return 0;
2374 }
2375 
2376 /*
2377  * uvm_fault_unwire(): unwire range of virtual space.
2378  */
2379 
2380 void
2381 uvm_fault_unwire(struct vm_map *map, vaddr_t start, vaddr_t end)
2382 {
2383 	vm_map_lock_read(map);
2384 	uvm_fault_unwire_locked(map, start, end);
2385 	vm_map_unlock_read(map);
2386 }
2387 
2388 /*
2389  * uvm_fault_unwire_locked(): the guts of uvm_fault_unwire().
2390  *
2391  * => map must be at least read-locked.
2392  */
2393 
2394 void
2395 uvm_fault_unwire_locked(struct vm_map *map, vaddr_t start, vaddr_t end)
2396 {
2397 	struct vm_map_entry *entry, *oentry;
2398 	pmap_t pmap = vm_map_pmap(map);
2399 	vaddr_t va;
2400 	paddr_t pa;
2401 	struct vm_page *pg;
2402 
2403 	/*
2404 	 * we assume that the area we are unwiring has actually been wired
2405 	 * in the first place.   this means that we should be able to extract
2406 	 * the PAs from the pmap.   we also lock out the page daemon so that
2407 	 * we can call uvm_pageunwire.
2408 	 */
2409 
2410 	/*
2411 	 * find the beginning map entry for the region.
2412 	 */
2413 
2414 	KASSERT(start >= vm_map_min(map) && end <= vm_map_max(map));
2415 	if (uvm_map_lookup_entry(map, start, &entry) == false)
2416 		panic("uvm_fault_unwire_locked: address not in map");
2417 
2418 	oentry = NULL;
2419 	for (va = start; va < end; va += PAGE_SIZE) {
2420 		if (pmap_extract(pmap, va, &pa) == false)
2421 			continue;
2422 
2423 		/*
2424 		 * find the map entry for the current address.
2425 		 */
2426 
2427 		KASSERT(va >= entry->start);
2428 		while (va >= entry->end) {
2429 			KASSERT(entry->next != &map->header &&
2430 				entry->next->start <= entry->end);
2431 			entry = entry->next;
2432 		}
2433 
2434 		/*
2435 		 * lock it.
2436 		 */
2437 
2438 		if (entry != oentry) {
2439 			if (oentry != NULL) {
2440 				mutex_exit(&uvm_pageqlock);
2441 				uvm_map_unlock_entry(oentry);
2442 			}
2443 			uvm_map_lock_entry(entry);
2444 			mutex_enter(&uvm_pageqlock);
2445 			oentry = entry;
2446 		}
2447 
2448 		/*
2449 		 * if the entry is no longer wired, tell the pmap.
2450 		 */
2451 
2452 		if (VM_MAPENT_ISWIRED(entry) == 0)
2453 			pmap_unwire(pmap, va);
2454 
2455 		pg = PHYS_TO_VM_PAGE(pa);
2456 		if (pg)
2457 			uvm_pageunwire(pg);
2458 	}
2459 
2460 	if (oentry != NULL) {
2461 		mutex_exit(&uvm_pageqlock);
2462 		uvm_map_unlock_entry(entry);
2463 	}
2464 }
2465